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);
230 if (td->o.fill_device)
232 log_info("fio: ENOSPC on laying out "
236 td_verror(td, errno, "write");
238 td_verror(td, EIO, "write");
245 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
246 td_io_unlink_file(td, f);
247 } else if (td->o.create_fsync) {
248 if (fsync(f->fd) < 0) {
249 td_verror(td, errno, "fsync");
253 if (td->o.fill_device && !td_write(td)) {
254 fio_file_clear_size_known(f);
255 if (td_io_get_file_size(td, f))
257 if (f->io_size > f->real_file_size)
258 f->io_size = f->real_file_size;
272 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
274 int r, did_open = 0, old_runstate;
275 unsigned long long left;
276 unsigned long long bs;
280 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
281 td_ioengine_flagged(td, FIO_NOIO))
284 if (f->filetype == FIO_TYPE_CHAR)
287 if (!fio_file_open(f)) {
288 if (td->io_ops->open_file(td, f)) {
289 log_err("fio: cannot pre-read, failed to open file\n");
295 old_runstate = td_bump_runstate(td, TD_PRE_READING);
298 bs = td->o.max_bs[DDIR_READ];
304 td_verror(td, errno, "malloc");
310 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
311 td_verror(td, errno, "lseek");
312 log_err("fio: failed to lseek pre-read file\n");
317 while (left && !td->terminate) {
321 r = read(f->fd, b, bs);
327 td_verror(td, EIO, "pre_read");
333 td_restore_runstate(td, old_runstate);
336 td->io_ops->close_file(td, f);
342 unsigned long long get_rand_file_size(struct thread_data *td)
344 unsigned long long ret, sized;
348 frand_max = rand_max(&td->file_size_state);
349 r = __rand(&td->file_size_state);
350 sized = td->o.file_size_high - td->o.file_size_low;
351 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
352 ret += td->o.file_size_low;
353 ret -= (ret % td->o.rw_min_bs);
357 static int file_size(struct thread_data *td, struct fio_file *f)
361 if (stat(f->file_name, &st) == -1) {
362 td_verror(td, errno, "fstat");
366 f->real_file_size = st.st_size;
370 static int bdev_size(struct thread_data *td, struct fio_file *f)
372 unsigned long long bytes = 0;
375 if (td->io_ops->open_file(td, f)) {
376 log_err("fio: failed opening blockdev %s for size check\n",
381 r = blockdev_size(f, &bytes);
383 td_verror(td, r, "blockdev_size");
388 log_err("%s: zero sized block device?\n", f->file_name);
392 f->real_file_size = bytes;
393 td->io_ops->close_file(td, f);
396 td->io_ops->close_file(td, f);
400 static int char_size(struct thread_data *td, struct fio_file *f)
402 #ifdef FIO_HAVE_CHARDEV_SIZE
403 unsigned long long bytes = 0;
406 if (td->io_ops->open_file(td, f)) {
407 log_err("fio: failed opening chardev %s for size check\n",
412 r = chardev_size(f, &bytes);
414 td_verror(td, r, "chardev_size");
419 log_err("%s: zero sized char device?\n", f->file_name);
423 f->real_file_size = bytes;
424 td->io_ops->close_file(td, f);
427 td->io_ops->close_file(td, f);
430 f->real_file_size = -1ULL;
435 static int get_file_size(struct thread_data *td, struct fio_file *f)
439 if (fio_file_size_known(f))
442 if (f->filetype == FIO_TYPE_FILE)
443 ret = file_size(td, f);
444 else if (f->filetype == FIO_TYPE_BLOCK)
445 ret = bdev_size(td, f);
446 else if (f->filetype == FIO_TYPE_CHAR)
447 ret = char_size(td, f);
449 f->real_file_size = -1;
450 log_info("%s: failed to get file size of %s\n", td->o.name,
452 return 1; /* avoid offset extends end error message */
456 * Leave ->real_file_size with 0 since it could be expectation
457 * of initial setup for regular files.
463 * ->file_offset normally hasn't been initialized yet, so this
464 * is basically always false unless ->real_file_size is -1, but
465 * if ->real_file_size is -1 this message doesn't make sense.
466 * As a result, this message is basically useless.
468 if (f->file_offset > f->real_file_size) {
469 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
470 (unsigned long long) f->file_offset,
471 (unsigned long long) f->real_file_size);
475 fio_file_set_size_known(f);
479 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
480 unsigned long long off,
481 unsigned long long len)
483 int errval = 0, ret = 0;
492 off = f->file_offset;
494 if (len == -1ULL || off == -1ULL)
497 if (td->io_ops->invalidate) {
498 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
500 ret = td->io_ops->invalidate(td, f);
503 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
504 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
506 } else if (f->filetype == FIO_TYPE_FILE) {
507 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
508 f->file_name, off, len);
509 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
512 } else if (f->filetype == FIO_TYPE_BLOCK) {
515 dprint(FD_IO, "drop page cache %s\n", f->file_name);
516 ret = blockdev_invalidate_cache(f);
517 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
519 * Linux multipath devices reject ioctl while
520 * the maps are being updated. That window can
521 * last tens of milliseconds; we'll try up to
522 * a quarter of a second.
525 ret = blockdev_invalidate_cache(f);
527 if (ret < 0 && errno == EACCES && geteuid()) {
528 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
529 log_err("fio: only root may flush block "
530 "devices. Cache flush bypassed!\n");
535 } else if (f->filetype == FIO_TYPE_CHAR ||
536 f->filetype == FIO_TYPE_PIPE) {
537 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
541 * Cache flushing isn't a fatal condition, and we know it will
542 * happen on some platforms where we don't have the proper
543 * function to flush eg block device caches. So just warn and
544 * continue on our way.
547 log_info("fio: cache invalidation of %s failed: %s\n",
548 f->file_name, strerror(errval));
554 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
556 if (!fio_file_open(f))
559 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
562 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
566 dprint(FD_FILE, "fd close %s\n", f->file_name);
570 if (close(f->fd) < 0)
575 if (f->shadow_fd != -1) {
584 int file_lookup_open(struct fio_file *f, int flags)
586 struct fio_file *__f;
589 __f = lookup_file_hash(f->file_name);
591 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
595 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
603 f->fd = open(f->file_name, flags, 0600);
607 static int file_close_shadow_fds(struct thread_data *td)
613 for_each_file(td, f, i) {
614 if (f->shadow_fd == -1)
625 int generic_open_file(struct thread_data *td, struct fio_file *f)
631 dprint(FD_FILE, "fd open %s\n", f->file_name);
633 if (!strcmp(f->file_name, "-")) {
635 log_err("fio: can't read/write to stdin/out\n");
641 * move output logging to stderr, if we are writing to stdout
650 flags |= OS_O_DIRECT;
653 td_verror(td, EINVAL, "OS does not support atomic IO");
656 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
660 if (td->o.create_on_open && td->o.allow_create)
663 if (f->filetype != FIO_TYPE_FILE)
664 flags |= FIO_O_NOATIME;
671 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
675 f->fd = dup(STDOUT_FILENO);
677 from_hash = file_lookup_open(f, flags);
678 } else if (td_read(td)) {
679 if (f->filetype == FIO_TYPE_CHAR && !read_only)
685 f->fd = dup(STDIN_FILENO);
687 from_hash = file_lookup_open(f, flags);
688 } else if (td_trim(td)) {
689 assert(!td_rw(td)); /* should have matched above */
692 from_hash = file_lookup_open(f, flags);
696 char buf[FIO_VERROR_SIZE];
699 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
700 flags &= ~FIO_O_NOATIME;
703 if (__e == EMFILE && file_close_shadow_fds(td))
706 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
708 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
709 log_err("fio: looks like your file system does not " \
710 "support direct=1/buffered=0\n");
713 td_verror(td, __e, buf);
717 if (!from_hash && f->fd != -1) {
718 if (add_file_hash(f)) {
722 * Stash away descriptor for later close. This is to
723 * work-around a "feature" on Linux, where a close of
724 * an fd that has been opened for write will trigger
725 * udev to call blkid to check partitions, fs id, etc.
726 * That pollutes the device cache, which can slow down
727 * unbuffered accesses.
729 if (f->shadow_fd == -1)
730 f->shadow_fd = f->fd;
733 * OK to ignore, we haven't done anything
736 ret = generic_close_file(td, f);
746 * This function i.e. get_file_size() is the default .get_file_size
747 * implementation of majority of I/O engines.
749 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
751 return get_file_size(td, f);
755 * open/close all files, so that ->real_file_size gets set
757 static int get_file_sizes(struct thread_data *td)
763 for_each_file(td, f, i) {
764 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
767 if (td_io_get_file_size(td, f)) {
768 if (td->error != ENOENT) {
769 log_err("%s\n", td->verror);
777 * There are corner cases where we end up with -1 for
778 * ->real_file_size due to unsupported file type, etc.
779 * We then just set to size option value divided by number
780 * of files, similar to the way file ->io_size is set.
781 * stat(2) failure doesn't set ->real_file_size to -1.
783 if (f->real_file_size == -1ULL && td->o.size)
784 f->real_file_size = td->o.size / td->o.nr_files;
791 struct flist_head list;
798 * Get free number of bytes for each file on each unique mount.
800 static unsigned long long get_fs_free_counts(struct thread_data *td)
802 struct flist_head *n, *tmp;
803 unsigned long long ret = 0;
804 struct fio_mount *fm;
809 for_each_file(td, f, i) {
813 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
814 if (f->real_file_size != -1ULL)
815 ret += f->real_file_size;
817 } else if (f->filetype != FIO_TYPE_FILE)
820 snprintf(buf, ARRAY_SIZE(buf), "%s", f->file_name);
822 if (stat(buf, &sb) < 0) {
826 if (stat(buf, &sb) < 0)
831 flist_for_each(n, &list) {
832 fm = flist_entry(n, struct fio_mount, list);
833 if (fm->key == sb.st_dev)
842 fm = calloc(1, sizeof(*fm));
843 snprintf(fm->__base, ARRAY_SIZE(fm->__base), "%s", buf);
844 fm->base = basename(fm->__base);
846 flist_add(&fm->list, &list);
849 flist_for_each_safe(n, tmp, &list) {
850 unsigned long long sz;
852 fm = flist_entry(n, struct fio_mount, list);
853 flist_del(&fm->list);
855 sz = get_fs_free_size(fm->base);
856 if (sz && sz != -1ULL)
865 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
868 struct thread_options *o = &td->o;
869 unsigned long long align_bs;
870 unsigned long long offset;
871 unsigned long long increment;
873 if (o->file_append && f->filetype == FIO_TYPE_FILE)
874 return f->real_file_size;
876 if (o->offset_increment_percent) {
877 assert(!o->offset_increment);
878 increment = o->offset_increment_percent * f->real_file_size / 100;
881 increment = o->offset_increment;
883 if (o->start_offset_percent > 0) {
884 /* calculate the raw offset */
885 offset = (f->real_file_size * o->start_offset_percent / 100) +
886 (td->subjob_number * increment);
890 /* start_offset_percent not set */
891 offset = o->start_offset +
892 td->subjob_number * increment;
897 * if offset_align is provided, use it
899 if (fio_option_is_set(o, start_offset_align)) {
900 align_bs = o->start_offset_align;
902 /* else take the minimum block size */
903 align_bs = td_min_bs(td);
907 * block align the offset at the next available boundary at
908 * ceiling(offset / align_bs) * align_bs
910 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
917 * Find longest path component that exists and return its length
919 int longest_existing_path(char *path) {
930 sprintf(buf, "%s", path);
933 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
940 *(buf_pos + 1) = '\0';
943 dwAttr = GetFileAttributesA(buf);
944 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
948 if (stat(buf, &sb) == 0)
952 offset = buf_pos - buf;
960 static bool create_work_dirs(struct thread_data *td, const char *fname)
966 snprintf(path, PATH_MAX, "%s", fname);
969 offset = longest_existing_path(path);
970 end = start + offset;
971 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
978 if (fio_mkdir(path, 0700) && errno != EEXIST) {
979 log_err("fio: failed to create dir (%s): %s\n",
980 start, strerror(errno));
983 *end = FIO_OS_PATH_SEPARATOR;
986 td->flags |= TD_F_DIRS_CREATED;
991 * Open the files and setup files sizes, creating files if necessary.
993 int setup_files(struct thread_data *td)
995 unsigned long long total_size, extend_size;
996 struct thread_options *o = &td->o;
998 unsigned int i, nr_fs_extra = 0;
999 int err = 0, need_extend;
1001 const unsigned long long bs = td_min_bs(td);
1004 dprint(FD_FILE, "setup files\n");
1006 old_state = td_bump_runstate(td, TD_SETTING_UP);
1008 for_each_file(td, f, i) {
1009 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1010 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1011 !(td->flags & TD_F_DIRS_CREATED) &&
1012 !create_work_dirs(td, f->file_name))
1017 * Find out physical size of files or devices for this thread,
1018 * before we determine I/O size and range of our targets.
1019 * If ioengine defines a setup() method, it's responsible for
1020 * opening the files and setting f->real_file_size to indicate
1021 * the valid range for that file.
1023 if (td->io_ops->setup)
1024 err = td->io_ops->setup(td);
1026 err = get_file_sizes(td);
1031 if (o->read_iolog_file)
1035 * check sizes. if the files/devices do not exist and the size
1036 * isn't passed to fio, abort.
1039 for_each_file(td, f, i) {
1041 if (f->real_file_size == -1ULL)
1044 total_size += f->real_file_size;
1048 td->fill_device_size = get_fs_free_counts(td);
1051 * device/file sizes are zero and no size given, punt
1053 if ((!total_size || total_size == -1ULL) && !o->size &&
1054 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1055 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1056 log_err("%s: you need to specify size=\n", o->name);
1057 td_verror(td, EINVAL, "total_file_size");
1062 * Calculate per-file size and potential extra size for the
1063 * first files, if needed (i.e. if we don't have a fixed size).
1065 if (!o->file_size_low && o->nr_files) {
1068 fs = o->size / o->nr_files;
1069 all_fs = fs * o->nr_files;
1071 if (all_fs < o->size)
1072 nr_fs_extra = (o->size - all_fs) / bs;
1076 * now file sizes are known, so we can set ->io_size. if size= is
1077 * not given, ->io_size is just equal to ->real_file_size. if size
1078 * is given, ->io_size is size / nr_files.
1080 extend_size = total_size = 0;
1082 for_each_file(td, f, i) {
1083 f->file_offset = get_start_offset(td, f);
1086 * Update ->io_size depending on options specified.
1087 * ->file_size_low being 0 means filesize option isn't set.
1088 * Non zero ->file_size_low equals ->file_size_high means
1089 * filesize option is set in a fixed size format.
1090 * Non zero ->file_size_low not equals ->file_size_high means
1091 * filesize option is set in a range format.
1093 if (!o->file_size_low) {
1095 * no file size or range given, file size is equal to
1096 * total size divided by number of files. If the size
1097 * doesn't divide nicely with the min blocksize,
1098 * make the first files bigger.
1107 * We normally don't come here for regular files, but
1108 * if the result is 0 for a regular file, set it to the
1109 * real file size. This could be size of the existing
1110 * one if it already exists, but otherwise will be set
1111 * to 0. A new file won't be created because
1112 * ->io_size + ->file_offset equals ->real_file_size.
1115 if (f->file_offset > f->real_file_size)
1117 f->io_size = f->real_file_size - f->file_offset;
1119 log_info("fio: file %s may be ignored\n",
1122 } else if (f->real_file_size < o->file_size_low ||
1123 f->real_file_size > o->file_size_high) {
1124 if (f->file_offset > o->file_size_low)
1127 * file size given. if it's fixed, use that. if it's a
1128 * range, generate a random size in-between.
1130 if (o->file_size_low == o->file_size_high)
1131 f->io_size = o->file_size_low - f->file_offset;
1133 f->io_size = get_rand_file_size(td)
1137 f->io_size = f->real_file_size - f->file_offset;
1139 if (f->io_size == -1ULL)
1144 if (o->size_percent && o->size_percent != 100) {
1147 file_size = f->io_size + f->file_offset;
1148 f->io_size = (file_size *
1149 o->size_percent) / 100;
1150 if (f->io_size > (file_size - f->file_offset))
1151 f->io_size = file_size - f->file_offset;
1153 f->io_size -= (f->io_size % td_min_bs(td));
1156 io_size = f->io_size;
1157 if (o->io_size_percent && o->io_size_percent != 100) {
1158 io_size *= o->io_size_percent;
1162 total_size += io_size;
1165 if (f->filetype == FIO_TYPE_FILE &&
1166 (f->io_size + f->file_offset) > f->real_file_size) {
1167 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1168 !o->create_on_open) {
1170 extend_size += (f->io_size + f->file_offset);
1171 fio_file_set_extend(f);
1172 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1173 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1174 td_ioengine_flagged(td, FIO_FAKEIO)))
1175 f->real_file_size = f->io_size + f->file_offset;
1179 if (td->o.block_error_hist) {
1182 assert(td->o.nr_files == 1); /* checked in fixup_options */
1184 len = f->io_size / td->o.bs[DDIR_TRIM];
1185 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1186 log_err("fio: cannot calculate block histogram with "
1187 "%d trim blocks, maximum %d\n",
1188 len, MAX_NR_BLOCK_INFOS);
1189 td_verror(td, EINVAL, "block_error_hist");
1193 td->ts.nr_block_infos = len;
1194 for (i = 0; i < len; i++)
1195 td->ts.block_infos[i] =
1196 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1198 td->ts.nr_block_infos = 0;
1200 if (!o->size || (total_size && o->size > total_size))
1201 o->size = total_size;
1203 if (o->size < td_min_bs(td)) {
1204 log_err("fio: blocksize too large for data set\n");
1209 * See if we need to extend some files, typically needed when our
1210 * target regular files don't exist yet, but our jobs require them
1211 * initially due to read I/Os.
1215 if (output_format & FIO_OUTPUT_NORMAL) {
1216 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1218 need_extend > 1 ? "s" : "",
1220 need_extend > 1 ? "s" : "",
1221 need_extend > 1 ? "total " : "",
1225 for_each_file(td, f, i) {
1226 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1228 if (!fio_file_extend(f))
1231 assert(f->filetype == FIO_TYPE_FILE);
1232 fio_file_clear_extend(f);
1233 if (!o->fill_device) {
1234 old_len = f->real_file_size;
1235 extend_len = f->io_size + f->file_offset -
1238 f->real_file_size = (f->io_size + f->file_offset);
1239 err = extend_file(td, f);
1243 err = __file_invalidate_cache(td, f, old_len,
1247 * Shut up static checker
1263 * iolog already set the total io size, if we read back
1266 if (!o->read_iolog_file) {
1268 td->total_io_size = o->io_size * o->loops;
1270 td->total_io_size = o->size * o->loops;
1277 td_restore_runstate(td, old_state);
1279 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1280 err = zbd_setup_files(td);
1287 log_err("%s: you need to specify valid offset=\n", o->name);
1289 td_restore_runstate(td, old_state);
1293 bool pre_read_files(struct thread_data *td)
1298 dprint(FD_FILE, "pre_read files\n");
1300 for_each_file(td, f, i) {
1301 if (!pre_read_file(td, f))
1308 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1310 unsigned int range_size, seed;
1314 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1315 fsize = min(f->real_file_size, f->io_size);
1317 nranges = (fsize + range_size - 1ULL) / range_size;
1319 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1320 if (!td->o.rand_repeatable)
1321 seed = td->rand_seeds[4];
1323 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1324 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1325 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1326 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1327 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1328 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1331 static bool init_rand_distribution(struct thread_data *td)
1337 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1338 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1339 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1342 state = td_bump_runstate(td, TD_SETTING_UP);
1344 for_each_file(td, f, i)
1345 __init_rand_distribution(td, f);
1347 td_restore_runstate(td, state);
1352 * Check if the number of blocks exceeds the randomness capability of
1353 * the selected generator. Tausworthe is 32-bit, the others are fullly
1356 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1359 if (blocks <= FRAND32_MAX)
1361 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1365 * If the user hasn't specified a random generator, switch
1366 * to tausworthe64 with informational warning. If the user did
1367 * specify one, just warn.
1369 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1372 if (!fio_option_is_set(&td->o, random_generator)) {
1373 log_info("fio: Switching to tausworthe64. Use the "
1374 "random_generator= option to get rid of this "
1376 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1381 * Just make this information to avoid breaking scripts.
1383 log_info("fio: Use the random_generator= option to switch to lfsr or "
1388 bool init_random_map(struct thread_data *td)
1390 unsigned long long blocks;
1394 if (init_rand_distribution(td))
1399 for_each_file(td, f, i) {
1400 uint64_t fsize = min(f->real_file_size, f->io_size);
1402 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1403 fsize = td->o.zone_range;
1405 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1407 if (check_rand_gen_limits(td, f, blocks))
1410 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1413 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1415 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1416 fio_file_set_lfsr(f);
1419 log_err("fio: failed initializing LFSR\n");
1422 } else if (!td->o.norandommap) {
1423 f->io_axmap = axmap_new(blocks);
1425 fio_file_set_axmap(f);
1428 } else if (td->o.norandommap)
1431 if (!td->o.softrandommap) {
1432 log_err("fio: failed allocating random map. If running"
1433 " a large number of jobs, try the 'norandommap'"
1434 " option or set 'softrandommap'. Or give"
1435 " a larger --alloc-size to fio.\n");
1439 log_info("fio: file %s failed allocating random map. Running "
1440 "job without.\n", f->file_name);
1446 void close_files(struct thread_data *td)
1451 for_each_file(td, f, i) {
1452 if (fio_file_open(f))
1453 td_io_close_file(td, f);
1457 void fio_file_free(struct fio_file *f)
1459 if (fio_file_axmap(f))
1460 axmap_free(f->io_axmap);
1461 if (!fio_file_smalloc(f)) {
1465 sfree(f->file_name);
1470 void close_and_free_files(struct thread_data *td)
1475 dprint(FD_FILE, "close files\n");
1477 for_each_file(td, f, i) {
1478 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1479 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1480 td_io_unlink_file(td, f);
1483 if (fio_file_open(f))
1484 td_io_close_file(td, f);
1486 remove_file_hash(f);
1488 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1489 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1490 td_io_unlink_file(td, f);
1497 td->o.filename = NULL;
1499 free(td->file_locks);
1500 td->files_index = 0;
1502 td->file_locks = NULL;
1503 td->o.file_lock_mode = FILE_LOCK_NONE;
1507 static void get_file_type(struct fio_file *f)
1511 if (!strcmp(f->file_name, "-"))
1512 f->filetype = FIO_TYPE_PIPE;
1514 f->filetype = FIO_TYPE_FILE;
1517 /* \\.\ is the device namespace in Windows, where every file is
1519 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1520 f->filetype = FIO_TYPE_BLOCK;
1523 if (!stat(f->file_name, &sb)) {
1524 if (S_ISBLK(sb.st_mode))
1525 f->filetype = FIO_TYPE_BLOCK;
1526 else if (S_ISCHR(sb.st_mode))
1527 f->filetype = FIO_TYPE_CHAR;
1528 else if (S_ISFIFO(sb.st_mode))
1529 f->filetype = FIO_TYPE_PIPE;
1533 static bool __is_already_allocated(const char *fname, bool set)
1535 struct flist_head *entry;
1538 ret = file_bloom_exists(fname, set);
1542 flist_for_each(entry, &filename_list) {
1543 struct file_name *fn;
1545 fn = flist_entry(entry, struct file_name, list);
1547 if (!strcmp(fn->filename, fname))
1554 static bool is_already_allocated(const char *fname)
1558 fio_file_hash_lock();
1559 ret = __is_already_allocated(fname, false);
1560 fio_file_hash_unlock();
1565 static void set_already_allocated(const char *fname)
1567 struct file_name *fn;
1569 fn = malloc(sizeof(struct file_name));
1570 fn->filename = strdup(fname);
1572 fio_file_hash_lock();
1573 if (!__is_already_allocated(fname, true)) {
1574 flist_add_tail(&fn->list, &filename_list);
1577 fio_file_hash_unlock();
1585 static void free_already_allocated(void)
1587 struct flist_head *entry, *tmp;
1588 struct file_name *fn;
1590 if (flist_empty(&filename_list))
1593 fio_file_hash_lock();
1594 flist_for_each_safe(entry, tmp, &filename_list) {
1595 fn = flist_entry(entry, struct file_name, list);
1597 flist_del(&fn->list);
1601 fio_file_hash_unlock();
1604 static struct fio_file *alloc_new_file(struct thread_data *td)
1608 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1609 f = calloc(1, sizeof(*f));
1611 f = scalloc(1, sizeof(*f));
1619 fio_file_reset(td, f);
1620 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1621 fio_file_set_smalloc(f);
1625 bool exists_and_not_regfile(const char *filename)
1629 if (lstat(filename, &sb) == -1)
1632 #ifndef WIN32 /* NOT Windows */
1633 if (S_ISREG(sb.st_mode))
1636 /* \\.\ is the device namespace in Windows, where every file
1637 * is a device node */
1638 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1645 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1647 int cur_files = td->files_index;
1648 char file_name[PATH_MAX];
1652 dprint(FD_FILE, "add file %s\n", fname);
1654 if (td->o.directory)
1655 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1656 td->o.unique_filename);
1658 sprintf(file_name + len, "%s", fname);
1660 /* clean cloned siblings using existing files */
1661 if (numjob && is_already_allocated(file_name) &&
1662 !exists_and_not_regfile(fname))
1665 f = alloc_new_file(td);
1667 if (td->files_size <= td->files_index) {
1668 unsigned int new_size = td->o.nr_files + 1;
1670 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1672 td->files = realloc(td->files, new_size * sizeof(f));
1673 if (td->files == NULL) {
1674 log_err("fio: realloc OOM\n");
1677 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1678 td->file_locks = realloc(td->file_locks, new_size);
1679 if (!td->file_locks) {
1680 log_err("fio: realloc OOM\n");
1683 td->file_locks[cur_files] = FILE_LOCK_NONE;
1685 td->files_size = new_size;
1687 td->files[cur_files] = f;
1688 f->fileno = cur_files;
1691 * init function, io engine may not be loaded yet
1693 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1694 f->real_file_size = -1ULL;
1696 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1697 f->file_name = strdup(file_name);
1699 f->file_name = smalloc_strdup(file_name);
1701 /* can't handle smalloc failure from here */
1702 assert(f->file_name);
1706 switch (td->o.file_lock_mode) {
1707 case FILE_LOCK_NONE:
1709 case FILE_LOCK_READWRITE:
1710 f->rwlock = fio_rwlock_init();
1712 case FILE_LOCK_EXCLUSIVE:
1713 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1716 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1722 if (td->o.numjobs > 1)
1723 set_already_allocated(file_name);
1728 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1734 int add_file_exclusive(struct thread_data *td, const char *fname)
1739 for_each_file(td, f, i) {
1740 if (!strcmp(f->file_name, fname))
1744 return add_file(td, fname, 0, 1);
1747 void get_file(struct fio_file *f)
1749 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1750 assert(fio_file_open(f));
1754 int put_file(struct thread_data *td, struct fio_file *f)
1756 int f_ret = 0, ret = 0;
1758 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1760 if (!fio_file_open(f)) {
1761 assert(f->fd == -1);
1765 assert(f->references);
1766 if (--f->references)
1769 disk_util_dec(f->du);
1771 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1772 unlock_file_all(td, f);
1774 if (should_fsync(td) && td->o.fsync_on_close) {
1775 f_ret = fsync(f->fd);
1780 if (td->io_ops->close_file)
1781 ret = td->io_ops->close_file(td, f);
1786 td->nr_open_files--;
1787 fio_file_clear_closing(f);
1788 fio_file_clear_open(f);
1789 assert(f->fd == -1);
1793 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1795 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1798 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1799 if (ddir == DDIR_READ)
1800 fio_rwlock_read(f->rwlock);
1802 fio_rwlock_write(f->rwlock);
1803 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1804 fio_sem_down(f->lock);
1806 td->file_locks[f->fileno] = td->o.file_lock_mode;
1809 void unlock_file(struct thread_data *td, struct fio_file *f)
1811 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1814 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1815 fio_rwlock_unlock(f->rwlock);
1816 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1817 fio_sem_up(f->lock);
1819 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1822 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1824 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1826 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1830 static bool recurse_dir(struct thread_data *td, const char *dirname)
1836 D = opendir(dirname);
1838 char buf[FIO_VERROR_SIZE];
1840 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1841 td_verror(td, errno, buf);
1845 while ((dir = readdir(D)) != NULL) {
1846 char full_path[PATH_MAX];
1849 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1852 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1854 if (lstat(full_path, &sb) == -1) {
1855 if (errno != ENOENT) {
1856 td_verror(td, errno, "stat");
1862 if (S_ISREG(sb.st_mode)) {
1863 add_file(td, full_path, 0, 1);
1866 if (!S_ISDIR(sb.st_mode))
1869 ret = recurse_dir(td, full_path);
1878 int add_dir_files(struct thread_data *td, const char *path)
1880 int ret = recurse_dir(td, path);
1883 log_info("fio: opendir added %d files\n", td->o.nr_files);
1888 void dup_files(struct thread_data *td, struct thread_data *org)
1893 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1898 td->files = malloc(org->files_index * sizeof(f));
1900 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1901 td->file_locks = malloc(org->files_index);
1903 for_each_file(org, f, i) {
1904 struct fio_file *__f;
1906 __f = alloc_new_file(td);
1909 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1910 __f->file_name = strdup(f->file_name);
1912 __f->file_name = smalloc_strdup(f->file_name);
1914 /* can't handle smalloc failure from here */
1915 assert(__f->file_name);
1916 __f->filetype = f->filetype;
1919 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1920 __f->lock = f->lock;
1921 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1922 __f->rwlock = f->rwlock;
1929 * Returns the index that matches the filename, or -1 if not there
1931 int get_fileno(struct thread_data *td, const char *fname)
1936 for_each_file(td, f, i)
1937 if (!strcmp(f->file_name, fname))
1944 * For log usage, where we add/open/close files automatically
1946 void free_release_files(struct thread_data *td)
1950 td->o.open_files = 0;
1951 td->files_index = 0;
1954 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1958 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1959 f->last_pos[i] = f->file_offset;
1960 f->last_start[i] = -1ULL;
1963 if (fio_file_axmap(f))
1964 axmap_reset(f->io_axmap);
1965 else if (fio_file_lfsr(f))
1966 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1968 zbd_file_reset(td, f);
1971 bool fio_files_done(struct thread_data *td)
1976 for_each_file(td, f, i)
1977 if (!fio_file_done(f))
1983 /* free memory used in initialization phase only */
1984 void filesetup_mem_free(void)
1986 free_already_allocated();
1990 * This function is for platforms which support direct I/O but not O_DIRECT.
1992 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1994 #ifdef FIO_OS_DIRECTIO
1995 int ret = fio_set_odirect(f);
1998 td_verror(td, ret, "fio_set_directio");
1999 #if defined(__sun__)
2000 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2001 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2003 log_err("fio: the file system does not seem to support direct IO\n");
2006 log_err("fio: the file system does not seem to support direct IO\n");
2013 log_err("fio: direct IO is not supported on this host operating system\n");