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
741 flags |= OS_O_DIRECT;
742 flags |= td->o.sync_io;
743 if (td->o.create_on_open && td->o.allow_create)
745 if (f->filetype != FIO_TYPE_FILE)
746 flags |= FIO_O_NOATIME;
753 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
757 f->fd = dup(STDOUT_FILENO);
759 from_hash = file_lookup_open(f, flags);
760 } else if (td_read(td)) {
761 if (td_ioengine_flagged(td, FIO_RO_NEEDS_RW_OPEN) && !read_only)
767 f->fd = dup(STDIN_FILENO);
769 from_hash = file_lookup_open(f, flags);
770 } else if (td_trim(td)) {
771 assert(!td_rw(td)); /* should have matched above */
774 from_hash = file_lookup_open(f, flags);
778 char buf[FIO_VERROR_SIZE];
781 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
782 flags &= ~FIO_O_NOATIME;
785 if (__e == EMFILE && file_close_shadow_fds(td))
788 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
790 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
791 log_err("fio: looks like your file system does not " \
792 "support direct=1/buffered=0\n");
795 td_verror(td, __e, buf);
799 if (!from_hash && f->fd != -1) {
800 if (add_file_hash(f)) {
804 * Stash away descriptor for later close. This is to
805 * work-around a "feature" on Linux, where a close of
806 * an fd that has been opened for write will trigger
807 * udev to call blkid to check partitions, fs id, etc.
808 * That pollutes the device cache, which can slow down
809 * unbuffered accesses.
811 if (f->shadow_fd == -1)
812 f->shadow_fd = f->fd;
815 * OK to ignore, we haven't done anything
818 ret = generic_close_file(td, f);
828 * This function i.e. get_file_size() is the default .get_file_size
829 * implementation of majority of I/O engines.
831 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
833 return get_file_size(td, f);
837 * open/close all files, so that ->real_file_size gets set
839 static int get_file_sizes(struct thread_data *td)
845 for_each_file(td, f, i) {
846 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
849 if (td_io_get_file_size(td, f)) {
850 if (td->error != ENOENT) {
851 log_err("%s\n", td->verror);
859 * There are corner cases where we end up with -1 for
860 * ->real_file_size due to unsupported file type, etc.
861 * We then just set to size option value divided by number
862 * of files, similar to the way file ->io_size is set.
863 * stat(2) failure doesn't set ->real_file_size to -1.
865 if (f->real_file_size == -1ULL && td->o.size)
866 f->real_file_size = td->o.size / td->o.nr_files;
873 struct flist_head list;
880 * Get free number of bytes for each file on each unique mount.
882 static unsigned long long get_fs_free_counts(struct thread_data *td)
884 struct flist_head *n, *tmp;
885 unsigned long long ret = 0;
886 struct fio_mount *fm;
891 for_each_file(td, f, i) {
895 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
896 if (f->real_file_size != -1ULL)
897 ret += f->real_file_size;
899 } else if (f->filetype != FIO_TYPE_FILE)
902 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
904 if (stat(buf, &sb) < 0) {
908 if (stat(buf, &sb) < 0)
913 flist_for_each(n, &list) {
914 fm = flist_entry(n, struct fio_mount, list);
915 if (fm->key == sb.st_dev)
924 fm = calloc(1, sizeof(*fm));
925 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
926 fm->base = basename(fm->__base);
928 flist_add(&fm->list, &list);
931 flist_for_each_safe(n, tmp, &list) {
932 unsigned long long sz;
934 fm = flist_entry(n, struct fio_mount, list);
935 flist_del(&fm->list);
937 sz = get_fs_free_size(fm->base);
938 if (sz && sz != -1ULL)
947 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
950 struct thread_options *o = &td->o;
951 unsigned long long align_bs;
952 unsigned long long offset;
953 unsigned long long increment;
955 if (o->file_append && f->filetype == FIO_TYPE_FILE)
956 return f->real_file_size;
958 if (o->offset_increment_percent) {
959 assert(!o->offset_increment);
960 increment = o->offset_increment_percent * f->real_file_size / 100;
963 increment = o->offset_increment;
965 if (o->start_offset_percent > 0) {
966 /* calculate the raw offset */
967 offset = (f->real_file_size * o->start_offset_percent / 100) +
968 (td->subjob_number * increment);
972 /* start_offset_percent not set */
973 offset = o->start_offset +
974 td->subjob_number * increment;
979 * if offset_align is provided, use it
981 if (fio_option_is_set(o, start_offset_align)) {
982 align_bs = o->start_offset_align;
984 /* else take the minimum block size */
985 align_bs = td_min_bs(td);
989 * block align the offset at the next available boundary at
990 * ceiling(offset / align_bs) * align_bs
992 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
999 * Find longest path component that exists and return its length
1001 int longest_existing_path(char *path) {
1012 sprintf(buf, "%s", path);
1015 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1021 *(buf_pos + 1) = '\0';
1024 dwAttr = GetFileAttributesA(buf);
1025 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1029 if (stat(buf, &sb) == 0)
1033 offset = buf_pos - buf;
1041 static bool create_work_dirs(struct thread_data *td, const char *fname)
1043 char path[PATH_MAX];
1047 snprintf(path, PATH_MAX, "%s", fname);
1050 offset = longest_existing_path(path);
1051 end = start + offset;
1052 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1059 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1060 log_err("fio: failed to create dir (%s): %s\n",
1061 start, strerror(errno));
1064 *end = FIO_OS_PATH_SEPARATOR;
1067 td->flags |= TD_F_DIRS_CREATED;
1072 * Open the files and setup files sizes, creating files if necessary.
1074 int setup_files(struct thread_data *td)
1076 unsigned long long total_size, extend_size;
1077 struct thread_options *o = &td->o;
1079 unsigned int i, nr_fs_extra = 0;
1080 int err = 0, need_extend;
1082 const unsigned long long bs = td_min_bs(td);
1085 dprint(FD_FILE, "setup files\n");
1087 old_state = td_bump_runstate(td, TD_SETTING_UP);
1089 for_each_file(td, f, i) {
1090 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1091 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1092 !(td->flags & TD_F_DIRS_CREATED) &&
1093 !create_work_dirs(td, f->file_name))
1098 * Find out physical size of files or devices for this thread,
1099 * before we determine I/O size and range of our targets.
1100 * If ioengine defines a setup() method, it's responsible for
1101 * opening the files and setting f->real_file_size to indicate
1102 * the valid range for that file.
1104 if (td->io_ops->setup)
1105 err = td->io_ops->setup(td);
1107 err = get_file_sizes(td);
1112 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1113 err = zbd_init_files(td);
1117 zbd_recalc_options_with_zone_granularity(td);
1119 if (o->read_iolog_file)
1123 * check sizes. if the files/devices do not exist and the size
1124 * isn't passed to fio, abort.
1127 for_each_file(td, f, i) {
1129 if (f->real_file_size == -1ULL)
1132 total_size += f->real_file_size;
1136 td->fill_device_size = get_fs_free_counts(td);
1139 * device/file sizes are zero and no size given, punt
1141 if ((!total_size || total_size == -1ULL) && !o->size &&
1142 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1143 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1144 log_err("%s: you need to specify size=\n", o->name);
1145 td_verror(td, EINVAL, "total_file_size");
1150 * Calculate per-file size and potential extra size for the
1151 * first files, if needed (i.e. if we don't have a fixed size).
1153 if (!o->file_size_low && o->nr_files) {
1156 fs = o->size / o->nr_files;
1157 all_fs = fs * o->nr_files;
1159 if (all_fs < o->size)
1160 nr_fs_extra = (o->size - all_fs) / bs;
1164 * now file sizes are known, so we can set ->io_size. if size= is
1165 * not given, ->io_size is just equal to ->real_file_size. if size
1166 * is given, ->io_size is size / nr_files.
1168 extend_size = total_size = 0;
1170 for_each_file(td, f, i) {
1171 f->file_offset = get_start_offset(td, f);
1174 * Update ->io_size depending on options specified.
1175 * ->file_size_low being 0 means filesize option isn't set.
1176 * Non zero ->file_size_low equals ->file_size_high means
1177 * filesize option is set in a fixed size format.
1178 * Non zero ->file_size_low not equals ->file_size_high means
1179 * filesize option is set in a range format.
1181 if (!o->file_size_low) {
1183 * no file size or range given, file size is equal to
1184 * total size divided by number of files. If the size
1185 * doesn't divide nicely with the min blocksize,
1186 * make the first files bigger.
1195 * We normally don't come here for regular files, but
1196 * if the result is 0 for a regular file, set it to the
1197 * real file size. This could be size of the existing
1198 * one if it already exists, but otherwise will be set
1199 * to 0. A new file won't be created because
1200 * ->io_size + ->file_offset equals ->real_file_size.
1203 if (f->file_offset > f->real_file_size)
1205 f->io_size = f->real_file_size - f->file_offset;
1207 log_info("fio: file %s may be ignored\n",
1210 } else if (f->real_file_size < o->file_size_low ||
1211 f->real_file_size > o->file_size_high) {
1212 if (f->file_offset > o->file_size_low)
1215 * file size given. if it's fixed, use that. if it's a
1216 * range, generate a random size in-between.
1218 if (o->file_size_low == o->file_size_high)
1219 f->io_size = o->file_size_low - f->file_offset;
1221 f->io_size = get_rand_file_size(td)
1225 f->io_size = f->real_file_size - f->file_offset;
1227 if (f->io_size == -1ULL)
1232 if (o->size_percent && o->size_percent != 100) {
1235 file_size = f->io_size + f->file_offset;
1236 f->io_size = (file_size *
1237 o->size_percent) / 100;
1238 if (f->io_size > (file_size - f->file_offset))
1239 f->io_size = file_size - f->file_offset;
1241 f->io_size -= (f->io_size % td_min_bs(td));
1244 io_size = f->io_size;
1245 if (o->io_size_percent && o->io_size_percent != 100) {
1246 io_size *= o->io_size_percent;
1250 total_size += io_size;
1253 if (f->filetype == FIO_TYPE_FILE &&
1254 (f->io_size + f->file_offset) > f->real_file_size) {
1255 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1256 !o->create_on_open) {
1258 extend_size += (f->io_size + f->file_offset);
1259 fio_file_set_extend(f);
1260 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1261 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1262 td_ioengine_flagged(td, FIO_FAKEIO)))
1263 f->real_file_size = f->io_size + f->file_offset;
1267 if (td->o.block_error_hist) {
1270 assert(td->o.nr_files == 1); /* checked in fixup_options */
1272 len = f->io_size / td->o.bs[DDIR_TRIM];
1273 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1274 log_err("fio: cannot calculate block histogram with "
1275 "%d trim blocks, maximum %d\n",
1276 len, MAX_NR_BLOCK_INFOS);
1277 td_verror(td, EINVAL, "block_error_hist");
1281 td->ts.nr_block_infos = len;
1282 for (i = 0; i < len; i++)
1283 td->ts.block_infos[i] =
1284 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1286 td->ts.nr_block_infos = 0;
1288 if (!o->size || (total_size && o->size > total_size))
1289 o->size = total_size;
1291 if (o->size < td_min_bs(td)) {
1292 log_err("fio: blocksize is larger than data set range\n");
1297 * See if we need to extend some files, typically needed when our
1298 * target regular files don't exist yet, but our jobs require them
1299 * initially due to read I/Os.
1303 if (output_format & FIO_OUTPUT_NORMAL) {
1304 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1306 need_extend > 1 ? "s" : "",
1308 need_extend > 1 ? "s" : "",
1309 need_extend > 1 ? "total " : "",
1313 for_each_file(td, f, i) {
1314 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1316 if (!fio_file_extend(f))
1319 assert(f->filetype == FIO_TYPE_FILE);
1320 fio_file_clear_extend(f);
1321 if (!o->fill_device) {
1322 old_len = f->real_file_size;
1323 extend_len = f->io_size + f->file_offset -
1326 f->real_file_size = (f->io_size + f->file_offset);
1327 err = extend_file(td, f);
1331 err = __file_invalidate_cache(td, f, old_len,
1335 * Shut up static checker
1351 * Prepopulate files with data. It might be expected to read some
1352 * "real" data instead of zero'ed files (if no writes to file occurred
1353 * prior to a read job). Engine has to provide a way to do that.
1355 if (td->io_ops->prepopulate_file) {
1358 for_each_file(td, f, i) {
1359 if (output_format & FIO_OUTPUT_NORMAL) {
1360 log_info("%s: Prepopulating IO file (%s)\n",
1361 o->name, f->file_name);
1364 err = td->io_ops->prepopulate_file(td, f);
1368 err = __file_invalidate_cache(td, f, f->file_offset,
1372 * Shut up static checker
1388 * iolog already set the total io size, if we read back
1391 if (!o->read_iolog_file) {
1393 td->total_io_size = o->io_size * o->loops;
1395 td->total_io_size = o->size * o->loops;
1399 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1400 err = zbd_setup_files(td);
1408 td_restore_runstate(td, old_state);
1419 log_err("%s: you need to specify valid offset=\n", o->name);
1421 td_restore_runstate(td, old_state);
1425 bool pre_read_files(struct thread_data *td)
1430 dprint(FD_FILE, "pre_read files\n");
1432 for_each_file(td, f, i) {
1433 if (!pre_read_file(td, f))
1440 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1442 unsigned int range_size, seed;
1446 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1447 fsize = min(f->real_file_size, f->io_size);
1449 nranges = (fsize + range_size - 1ULL) / range_size;
1451 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1452 if (!td->o.rand_repeatable)
1453 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1455 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1456 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1457 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1458 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1459 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1460 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1463 static bool init_rand_distribution(struct thread_data *td)
1469 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1470 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1471 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1474 state = td_bump_runstate(td, TD_SETTING_UP);
1476 for_each_file(td, f, i)
1477 __init_rand_distribution(td, f);
1479 td_restore_runstate(td, state);
1484 * Check if the number of blocks exceeds the randomness capability of
1485 * the selected generator. Tausworthe is 32-bit, the others are fully
1488 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1491 if (blocks <= FRAND32_MAX)
1493 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1497 * If the user hasn't specified a random generator, switch
1498 * to tausworthe64 with informational warning. If the user did
1499 * specify one, just warn.
1501 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1504 if (!fio_option_is_set(&td->o, random_generator)) {
1505 log_info("fio: Switching to tausworthe64. Use the "
1506 "random_generator= option to get rid of this "
1508 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1513 * Just make this information to avoid breaking scripts.
1515 log_info("fio: Use the random_generator= option to switch to lfsr or "
1520 bool init_random_map(struct thread_data *td)
1522 unsigned long long blocks;
1526 if (init_rand_distribution(td))
1531 for_each_file(td, f, i) {
1532 uint64_t fsize = min(f->real_file_size, f->io_size);
1534 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1535 fsize = td->o.zone_range;
1537 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1539 if (check_rand_gen_limits(td, f, blocks))
1542 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1545 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1547 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1548 fio_file_set_lfsr(f);
1551 log_err("fio: failed initializing LFSR\n");
1554 } else if (!td->o.norandommap) {
1555 f->io_axmap = axmap_new(blocks);
1557 fio_file_set_axmap(f);
1560 } else if (td->o.norandommap)
1563 if (!td->o.softrandommap) {
1564 log_err("fio: failed allocating random map. If running"
1565 " a large number of jobs, try the 'norandommap'"
1566 " option or set 'softrandommap'. Or give"
1567 " a larger --alloc-size to fio.\n");
1571 log_info("fio: file %s failed allocating random map. Running "
1572 "job without.\n", f->file_name);
1578 void close_files(struct thread_data *td)
1583 for_each_file(td, f, i) {
1584 if (fio_file_open(f))
1585 td_io_close_file(td, f);
1589 void fio_file_free(struct fio_file *f)
1591 if (fio_file_axmap(f))
1592 axmap_free(f->io_axmap);
1594 sfree(f->ruhs_info);
1595 if (!fio_file_smalloc(f)) {
1599 sfree(f->file_name);
1604 void close_and_free_files(struct thread_data *td)
1609 dprint(FD_FILE, "close files\n");
1611 for_each_file(td, f, i) {
1612 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1613 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1614 td_io_unlink_file(td, f);
1617 if (fio_file_open(f))
1618 td_io_close_file(td, f);
1620 remove_file_hash(f);
1622 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1623 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1624 td_io_unlink_file(td, f);
1628 fdp_free_ruhs_info(f);
1632 td->o.filename = NULL;
1634 free(td->file_locks);
1635 td->files_index = 0;
1637 td->file_locks = NULL;
1638 td->o.file_lock_mode = FILE_LOCK_NONE;
1642 static void get_file_type(struct fio_file *f)
1646 if (!strcmp(f->file_name, "-"))
1647 f->filetype = FIO_TYPE_PIPE;
1649 f->filetype = FIO_TYPE_FILE;
1652 /* \\.\ is the device namespace in Windows, where every file is
1654 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1655 f->filetype = FIO_TYPE_BLOCK;
1658 if (!stat(f->file_name, &sb)) {
1659 if (S_ISBLK(sb.st_mode))
1660 f->filetype = FIO_TYPE_BLOCK;
1661 else if (S_ISCHR(sb.st_mode))
1662 f->filetype = FIO_TYPE_CHAR;
1663 else if (S_ISFIFO(sb.st_mode))
1664 f->filetype = FIO_TYPE_PIPE;
1668 static bool __is_already_allocated(const char *fname, bool set)
1670 struct flist_head *entry;
1673 ret = file_bloom_exists(fname, set);
1677 flist_for_each(entry, &filename_list) {
1678 struct file_name *fn;
1680 fn = flist_entry(entry, struct file_name, list);
1682 if (!strcmp(fn->filename, fname))
1689 static bool is_already_allocated(const char *fname)
1693 fio_file_hash_lock();
1694 ret = __is_already_allocated(fname, false);
1695 fio_file_hash_unlock();
1700 static void set_already_allocated(const char *fname)
1702 struct file_name *fn;
1704 fn = malloc(sizeof(struct file_name));
1705 fn->filename = strdup(fname);
1707 fio_file_hash_lock();
1708 if (!__is_already_allocated(fname, true)) {
1709 flist_add_tail(&fn->list, &filename_list);
1712 fio_file_hash_unlock();
1720 static void free_already_allocated(void)
1722 struct flist_head *entry, *tmp;
1723 struct file_name *fn;
1725 if (flist_empty(&filename_list))
1728 fio_file_hash_lock();
1729 flist_for_each_safe(entry, tmp, &filename_list) {
1730 fn = flist_entry(entry, struct file_name, list);
1732 flist_del(&fn->list);
1736 fio_file_hash_unlock();
1739 static struct fio_file *alloc_new_file(struct thread_data *td)
1743 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1744 f = calloc(1, sizeof(*f));
1746 f = scalloc(1, sizeof(*f));
1754 fio_file_reset(td, f);
1755 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1756 fio_file_set_smalloc(f);
1760 bool exists_and_not_regfile(const char *filename)
1764 if (lstat(filename, &sb) == -1)
1767 #ifndef WIN32 /* NOT Windows */
1768 if (S_ISREG(sb.st_mode))
1771 /* \\.\ is the device namespace in Windows, where every file
1772 * is a device node */
1773 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1780 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1782 int cur_files = td->files_index;
1783 char file_name[PATH_MAX];
1787 dprint(FD_FILE, "add file %s\n", fname);
1789 if (td->o.directory)
1790 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1791 td->o.unique_filename);
1793 sprintf(file_name + len, "%s", fname);
1795 /* clean cloned siblings using existing files */
1796 if (numjob && is_already_allocated(file_name) &&
1797 !exists_and_not_regfile(fname))
1800 f = alloc_new_file(td);
1802 if (td->files_size <= td->files_index) {
1803 unsigned int new_size = td->o.nr_files + 1;
1805 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1807 td->files = realloc(td->files, new_size * sizeof(f));
1808 if (td->files == NULL) {
1809 log_err("fio: realloc OOM\n");
1812 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1813 td->file_locks = realloc(td->file_locks, new_size);
1814 if (!td->file_locks) {
1815 log_err("fio: realloc OOM\n");
1818 td->file_locks[cur_files] = FILE_LOCK_NONE;
1820 td->files_size = new_size;
1822 td->files[cur_files] = f;
1823 f->fileno = cur_files;
1826 * init function, io engine may not be loaded yet
1828 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1829 f->real_file_size = -1ULL;
1831 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1832 f->file_name = strdup(file_name);
1834 f->file_name = smalloc_strdup(file_name);
1836 /* can't handle smalloc failure from here */
1837 assert(f->file_name);
1841 switch (td->o.file_lock_mode) {
1842 case FILE_LOCK_NONE:
1844 case FILE_LOCK_READWRITE:
1845 f->rwlock = fio_rwlock_init();
1847 case FILE_LOCK_EXCLUSIVE:
1848 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1851 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1857 if (td->o.numjobs > 1)
1858 set_already_allocated(file_name);
1863 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1869 int add_file_exclusive(struct thread_data *td, const char *fname)
1874 for_each_file(td, f, i) {
1875 if (!strcmp(f->file_name, fname))
1879 return add_file(td, fname, 0, 1);
1882 void get_file(struct fio_file *f)
1884 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1885 assert(fio_file_open(f));
1889 int put_file(struct thread_data *td, struct fio_file *f)
1891 int f_ret = 0, ret = 0;
1893 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1895 if (!fio_file_open(f)) {
1896 assert(f->fd == -1);
1900 assert(f->references);
1901 if (--f->references)
1904 disk_util_dec(f->du);
1906 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1907 unlock_file_all(td, f);
1909 if (should_fsync(td) && td->o.fsync_on_close) {
1910 f_ret = fsync(f->fd);
1915 if (td->io_ops->close_file)
1916 ret = td->io_ops->close_file(td, f);
1921 td->nr_open_files--;
1922 fio_file_clear_closing(f);
1923 fio_file_clear_open(f);
1924 assert(f->fd == -1);
1928 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1930 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1933 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1934 if (ddir == DDIR_READ)
1935 fio_rwlock_read(f->rwlock);
1937 fio_rwlock_write(f->rwlock);
1938 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1939 fio_sem_down(f->lock);
1941 td->file_locks[f->fileno] = td->o.file_lock_mode;
1944 void unlock_file(struct thread_data *td, struct fio_file *f)
1946 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1949 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1950 fio_rwlock_unlock(f->rwlock);
1951 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1952 fio_sem_up(f->lock);
1954 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1957 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1959 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1961 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1965 static bool recurse_dir(struct thread_data *td, const char *dirname)
1971 D = opendir(dirname);
1973 char buf[FIO_VERROR_SIZE];
1975 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1976 td_verror(td, errno, buf);
1980 while ((dir = readdir(D)) != NULL) {
1981 char full_path[PATH_MAX];
1984 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1987 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1989 if (lstat(full_path, &sb) == -1) {
1990 if (errno != ENOENT) {
1991 td_verror(td, errno, "stat");
1997 if (S_ISREG(sb.st_mode)) {
1998 add_file(td, full_path, 0, 1);
2001 if (!S_ISDIR(sb.st_mode))
2004 ret = recurse_dir(td, full_path);
2013 int add_dir_files(struct thread_data *td, const char *path)
2015 int ret = recurse_dir(td, path);
2018 log_info("fio: opendir added %d files\n", td->o.nr_files);
2023 void dup_files(struct thread_data *td, struct thread_data *org)
2028 dprint(FD_FILE, "dup files: %d\n", org->files_index);
2033 td->files = calloc(org->files_index, sizeof(f));
2035 if (td->o.file_lock_mode != FILE_LOCK_NONE)
2036 td->file_locks = malloc(org->files_index);
2038 assert(org->files_index >= org->o.nr_files);
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");