17 #include "lib/axmap.h"
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
25 static FLIST_HEAD(filename_list);
28 * List entry for filename_list
31 struct flist_head list;
35 static inline void clear_error(struct thread_data *td)
41 static inline int native_fallocate(struct thread_data *td, struct fio_file *f)
45 success = fio_fallocate(f, 0, f->real_file_size);
46 dprint(FD_FILE, "native fallocate of file %s size %llu was "
47 "%ssuccessful\n", f->file_name,
48 (unsigned long long) f->real_file_size,
55 dprint(FD_FILE, "native fallocate is not implemented\n");
60 static void fallocate_file(struct thread_data *td, struct fio_file *f)
64 if (td->o.fill_device)
67 switch (td->o.fallocate_mode) {
68 case FIO_FALLOCATE_NATIVE:
69 r = native_fallocate(td, f);
70 if (r != 0 && errno != ENOSYS)
71 log_err("fio: native_fallocate call failed: %s\n",
74 case FIO_FALLOCATE_NONE:
76 #ifdef CONFIG_POSIX_FALLOCATE
77 case FIO_FALLOCATE_POSIX:
78 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
80 (unsigned long long) f->real_file_size);
82 r = posix_fallocate(f->fd, 0, f->real_file_size);
84 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
86 #endif /* CONFIG_POSIX_FALLOCATE */
87 #ifdef CONFIG_LINUX_FALLOCATE
88 case FIO_FALLOCATE_KEEP_SIZE:
89 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
90 "file %s size %llu\n", f->file_name,
91 (unsigned long long) f->real_file_size);
93 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
95 td_verror(td, errno, "fallocate");
98 #endif /* CONFIG_LINUX_FALLOCATE */
100 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
106 * Leaves f->fd open on success, caller must close
108 static int extend_file(struct thread_data *td, struct fio_file *f)
110 int new_layout = 0, unlink_file = 0, flags;
111 unsigned long long left;
116 log_err("fio: refusing extend of file due to read-only\n");
121 * check if we need to lay the file out complete again. fio
122 * does that for operations involving reads, or for writes
123 * where overwrite is set
126 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
127 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
129 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
132 if (unlink_file || new_layout) {
135 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
137 ret = td_io_unlink_file(td, f);
138 if (ret != 0 && ret != ENOENT) {
139 td_verror(td, errno, "unlink");
145 if (td->o.allow_create)
150 flags |= OS_O_DIRECT;
156 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
157 f->fd = open(f->file_name, flags, 0644);
161 if (err == ENOENT && !td->o.allow_create)
162 log_err("fio: file creation disallowed by "
163 "allow_file_create=0\n");
165 td_verror(td, err, "open");
169 fallocate_file(td, f);
172 * If our jobs don't require regular files initially, we're done.
178 * The size will be -1ULL when fill_device is used, so don't truncate
179 * or fallocate this file, just write it
181 if (!td->o.fill_device) {
182 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
183 (unsigned long long) f->real_file_size);
184 if (ftruncate(f->fd, f->real_file_size) == -1) {
185 if (errno != EFBIG) {
186 td_verror(td, errno, "ftruncate");
192 left = f->real_file_size;
193 bs = td->o.max_bs[DDIR_WRITE];
199 td_verror(td, errno, "malloc");
203 while (left && !td->terminate) {
209 fill_io_buffer(td, b, bs, bs);
211 r = write(f->fd, b, bs);
221 if (td->o.fill_device)
223 log_info("fio: ENOSPC on laying out "
227 td_verror(td, errno, "write");
229 td_verror(td, EIO, "write");
236 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
237 td_io_unlink_file(td, f);
238 } else if (td->o.create_fsync) {
239 if (fsync(f->fd) < 0) {
240 td_verror(td, errno, "fsync");
244 if (td->o.fill_device && !td_write(td)) {
245 fio_file_clear_size_known(f);
246 if (td_io_get_file_size(td, f))
248 if (f->io_size > f->real_file_size)
249 f->io_size = f->real_file_size;
263 static int pre_read_file(struct thread_data *td, struct fio_file *f)
265 int ret = 0, r, did_open = 0, old_runstate;
266 unsigned long long left;
270 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
271 td_ioengine_flagged(td, FIO_NOIO))
274 if (f->filetype == FIO_TYPE_CHAR)
277 if (!fio_file_open(f)) {
278 if (td->io_ops->open_file(td, f)) {
279 log_err("fio: cannot pre-read, failed to open file\n");
285 old_runstate = td_bump_runstate(td, TD_PRE_READING);
288 bs = td->o.max_bs[DDIR_READ];
294 td_verror(td, errno, "malloc");
300 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
301 td_verror(td, errno, "lseek");
302 log_err("fio: failed to lseek pre-read file\n");
307 while (left && !td->terminate) {
311 r = read(f->fd, b, bs);
317 td_verror(td, EIO, "pre_read");
323 td_restore_runstate(td, old_runstate);
326 td->io_ops->close_file(td, f);
332 unsigned long long get_rand_file_size(struct thread_data *td)
334 unsigned long long ret, sized;
338 frand_max = rand_max(&td->file_size_state);
339 r = __rand(&td->file_size_state);
340 sized = td->o.file_size_high - td->o.file_size_low;
341 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
342 ret += td->o.file_size_low;
343 ret -= (ret % td->o.rw_min_bs);
347 static int file_size(struct thread_data *td, struct fio_file *f)
351 if (stat(f->file_name, &st) == -1) {
352 td_verror(td, errno, "fstat");
356 f->real_file_size = st.st_size;
360 static int bdev_size(struct thread_data *td, struct fio_file *f)
362 unsigned long long bytes = 0;
365 if (td->io_ops->open_file(td, f)) {
366 log_err("fio: failed opening blockdev %s for size check\n",
371 r = blockdev_size(f, &bytes);
373 td_verror(td, r, "blockdev_size");
378 log_err("%s: zero sized block device?\n", f->file_name);
382 f->real_file_size = bytes;
383 td->io_ops->close_file(td, f);
386 td->io_ops->close_file(td, f);
390 static int char_size(struct thread_data *td, struct fio_file *f)
392 #ifdef FIO_HAVE_CHARDEV_SIZE
393 unsigned long long bytes = 0;
396 if (td->io_ops->open_file(td, f)) {
397 log_err("fio: failed opening chardev %s for size check\n",
402 r = chardev_size(f, &bytes);
404 td_verror(td, r, "chardev_size");
409 log_err("%s: zero sized char device?\n", f->file_name);
413 f->real_file_size = bytes;
414 td->io_ops->close_file(td, f);
417 td->io_ops->close_file(td, f);
420 f->real_file_size = -1ULL;
425 static int get_file_size(struct thread_data *td, struct fio_file *f)
429 if (fio_file_size_known(f))
432 if (f->filetype == FIO_TYPE_FILE)
433 ret = file_size(td, f);
434 else if (f->filetype == FIO_TYPE_BLOCK)
435 ret = bdev_size(td, f);
436 else if (f->filetype == FIO_TYPE_CHAR)
437 ret = char_size(td, f);
439 f->real_file_size = -1ULL;
442 * Leave ->real_file_size with 0 since it could be expectation
443 * of initial setup for regular files.
449 * If ->real_file_size is -1, a conditional for the message
450 * "offset extends end" is always true, but it makes no sense,
451 * so just return the same value here.
453 if (f->real_file_size == -1ULL) {
454 log_info("%s: failed to get file size of %s\n", td->o.name,
459 if (td->o.start_offset && f->file_offset == 0)
460 dprint(FD_FILE, "offset of file %s not initialized yet\n",
463 * ->file_offset normally hasn't been initialized yet, so this
464 * is basically always false.
466 if (f->file_offset > f->real_file_size) {
467 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
468 (unsigned long long) f->file_offset,
469 (unsigned long long) f->real_file_size);
473 fio_file_set_size_known(f);
477 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
478 unsigned long long off,
479 unsigned long long len)
481 int errval = 0, ret = 0;
490 off = f->file_offset;
492 if (len == -1ULL || off == -1ULL)
495 if (td->io_ops->invalidate) {
496 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
498 ret = td->io_ops->invalidate(td, f);
501 } else if (f->filetype == FIO_TYPE_FILE) {
502 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
503 f->file_name, off, len);
504 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
507 } else if (f->filetype == FIO_TYPE_BLOCK) {
510 dprint(FD_IO, "drop page cache %s\n", f->file_name);
511 ret = blockdev_invalidate_cache(f);
512 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
514 * Linux multipath devices reject ioctl while
515 * the maps are being updated. That window can
516 * last tens of milliseconds; we'll try up to
517 * a quarter of a second.
520 ret = blockdev_invalidate_cache(f);
522 if (ret < 0 && errno == EACCES && geteuid()) {
524 log_err("fio: only root may flush block "
525 "devices. Cache flush bypassed!\n");
532 } else if (f->filetype == FIO_TYPE_CHAR ||
533 f->filetype == FIO_TYPE_PIPE) {
534 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
539 * Cache flushing isn't a fatal condition, and we know it will
540 * happen on some platforms where we don't have the proper
541 * function to flush eg block device caches. So just warn and
542 * continue on our way.
545 log_info("fio: cache invalidation of %s failed: %s\n",
546 f->file_name, strerror(errval));
552 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
554 if (!fio_file_open(f))
557 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
560 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
564 dprint(FD_FILE, "fd close %s\n", f->file_name);
568 if (close(f->fd) < 0)
573 if (f->shadow_fd != -1) {
582 int file_lookup_open(struct fio_file *f, int flags)
584 struct fio_file *__f;
587 __f = lookup_file_hash(f->file_name);
589 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
593 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
601 f->fd = open(f->file_name, flags, 0600);
605 static int file_close_shadow_fds(struct thread_data *td)
611 for_each_file(td, f, i) {
612 if (f->shadow_fd == -1)
623 int generic_open_file(struct thread_data *td, struct fio_file *f)
629 dprint(FD_FILE, "fd open %s\n", f->file_name);
631 if (!strcmp(f->file_name, "-")) {
633 log_err("fio: can't read/write to stdin/out\n");
639 * move output logging to stderr, if we are writing to stdout
648 flags |= OS_O_DIRECT;
651 td_verror(td, EINVAL, "OS does not support atomic IO");
654 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
658 if (td->o.create_on_open && td->o.allow_create)
661 if (f->filetype != FIO_TYPE_FILE)
662 flags |= FIO_O_NOATIME;
669 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
673 f->fd = dup(STDOUT_FILENO);
675 from_hash = file_lookup_open(f, flags);
676 } else if (td_read(td)) {
677 if (f->filetype == FIO_TYPE_CHAR && !read_only)
683 f->fd = dup(STDIN_FILENO);
685 from_hash = file_lookup_open(f, flags);
686 } else if (td_trim(td)) {
687 assert(!td_rw(td)); /* should have matched above */
689 from_hash = file_lookup_open(f, flags);
693 char buf[FIO_VERROR_SIZE];
696 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
697 flags &= ~FIO_O_NOATIME;
700 if (__e == EMFILE && file_close_shadow_fds(td))
703 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
705 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
706 log_err("fio: looks like your file system does not " \
707 "support direct=1/buffered=0\n");
710 td_verror(td, __e, buf);
714 if (!from_hash && f->fd != -1) {
715 if (add_file_hash(f)) {
719 * Stash away descriptor for later close. This is to
720 * work-around a "feature" on Linux, where a close of
721 * an fd that has been opened for write will trigger
722 * udev to call blkid to check partitions, fs id, etc.
723 * That pollutes the device cache, which can slow down
724 * unbuffered accesses.
726 if (f->shadow_fd == -1)
727 f->shadow_fd = f->fd;
730 * OK to ignore, we haven't done anything
733 ret = generic_close_file(td, f);
743 * This function i.e. get_file_size() is the default .get_file_size
744 * implementation of majority of I/O engines.
746 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
748 return get_file_size(td, f);
752 * open/close all files, so that ->real_file_size gets set
754 static int get_file_sizes(struct thread_data *td)
760 for_each_file(td, f, i) {
761 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
764 if (td_io_get_file_size(td, f)) {
765 if (td->error != ENOENT) {
766 log_err("%s\n", td->verror);
774 * There are corner cases where we end up with -1 for
775 * ->real_file_size due to unsupported file type, etc.
776 * We then just set to size option value divided by number
777 * of files, similar to the way file ->io_size is set.
778 * stat(2) failure doesn't set ->real_file_size to -1.
780 if (f->real_file_size == -1ULL && td->o.size)
781 f->real_file_size = td->o.size / td->o.nr_files;
788 struct flist_head list;
795 * Get free number of bytes for each file on each unique mount.
797 static unsigned long long get_fs_free_counts(struct thread_data *td)
799 struct flist_head *n, *tmp;
800 unsigned long long ret = 0;
801 struct fio_mount *fm;
806 for_each_file(td, f, i) {
810 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
811 if (f->real_file_size != -1ULL)
812 ret += f->real_file_size;
814 } else if (f->filetype != FIO_TYPE_FILE)
818 strncpy(buf, f->file_name, 255);
820 if (stat(buf, &sb) < 0) {
824 if (stat(buf, &sb) < 0)
829 flist_for_each(n, &list) {
830 fm = flist_entry(n, struct fio_mount, list);
831 if (fm->key == sb.st_dev)
840 fm = calloc(1, sizeof(*fm));
841 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
842 fm->base = basename(fm->__base);
844 flist_add(&fm->list, &list);
847 flist_for_each_safe(n, tmp, &list) {
848 unsigned long long sz;
850 fm = flist_entry(n, struct fio_mount, list);
851 flist_del(&fm->list);
853 sz = get_fs_free_size(fm->base);
854 if (sz && sz != -1ULL)
863 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
865 struct thread_options *o = &td->o;
866 unsigned long long align_bs;
867 unsigned long long offset;
869 if (o->file_append && f->filetype == FIO_TYPE_FILE)
870 return f->real_file_size;
872 if (o->start_offset_percent > 0) {
874 * if blockalign is provided, find the min across read, write,
877 if (fio_option_is_set(o, ba)) {
878 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
879 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
881 /* else take the minimum block size */
882 align_bs = td_min_bs(td);
885 /* calculate the raw offset */
886 offset = (f->real_file_size * o->start_offset_percent / 100) +
887 (td->subjob_number * o->offset_increment);
890 * block align the offset at the next available boundary at
891 * ceiling(offset / align_bs) * align_bs
893 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
896 /* start_offset_percent not set */
897 offset = o->start_offset +
898 td->subjob_number * o->offset_increment;
905 * Open the files and setup files sizes, creating files if necessary.
907 int setup_files(struct thread_data *td)
909 unsigned long long total_size, extend_size;
910 struct thread_options *o = &td->o;
912 unsigned int i, nr_fs_extra = 0;
913 int err = 0, need_extend;
915 const unsigned int bs = td_min_bs(td);
918 dprint(FD_FILE, "setup files\n");
920 old_state = td_bump_runstate(td, TD_SETTING_UP);
922 if (o->read_iolog_file)
926 * Find out physical size of files or devices for this thread,
927 * before we determine I/O size and range of our targets.
928 * If ioengine defines a setup() method, it's responsible for
929 * opening the files and setting f->real_file_size to indicate
930 * the valid range for that file.
932 if (td->io_ops->setup)
933 err = td->io_ops->setup(td);
935 err = get_file_sizes(td);
941 * check sizes. if the files/devices do not exist and the size
942 * isn't passed to fio, abort.
945 for_each_file(td, f, i) {
947 if (f->real_file_size == -1ULL)
950 total_size += f->real_file_size;
954 td->fill_device_size = get_fs_free_counts(td);
957 * device/file sizes are zero and no size given, punt
959 if ((!total_size || total_size == -1ULL) && !o->size &&
960 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
961 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
962 log_err("%s: you need to specify size=\n", o->name);
963 td_verror(td, EINVAL, "total_file_size");
968 * Calculate per-file size and potential extra size for the
969 * first files, if needed (i.e. if we don't have a fixed size).
971 if (!o->file_size_low && o->nr_files) {
974 fs = o->size / o->nr_files;
975 all_fs = fs * o->nr_files;
977 if (all_fs < o->size)
978 nr_fs_extra = (o->size - all_fs) / bs;
982 * now file sizes are known, so we can set ->io_size. if size= is
983 * not given, ->io_size is just equal to ->real_file_size. if size
984 * is given, ->io_size is size / nr_files.
986 extend_size = total_size = 0;
988 for_each_file(td, f, i) {
989 f->file_offset = get_start_offset(td, f);
992 * Update ->io_size depending on options specified.
993 * ->file_size_low being 0 means filesize option isn't set.
994 * Non zero ->file_size_low equals ->file_size_high means
995 * filesize option is set in a fixed size format.
996 * Non zero ->file_size_low not equals ->file_size_high means
997 * filesize option is set in a range format.
999 if (!o->file_size_low) {
1001 * no file size or range given, file size is equal to
1002 * total size divided by number of files. If the size
1003 * doesn't divide nicely with the min blocksize,
1004 * make the first files bigger.
1013 * We normally don't come here for regular files, but
1014 * if the result is 0 for a regular file, set it to the
1015 * real file size. This could be size of the existing
1016 * one if it already exists, but otherwise will be set
1017 * to 0. A new file won't be created because
1018 * ->io_size + ->file_offset equals ->real_file_size.
1021 if (f->file_offset > f->real_file_size)
1023 f->io_size = f->real_file_size - f->file_offset;
1025 log_info("fio: file %s may be ignored\n",
1028 } else if (f->real_file_size < o->file_size_low ||
1029 f->real_file_size > o->file_size_high) {
1030 if (f->file_offset > o->file_size_low)
1033 * file size given. if it's fixed, use that. if it's a
1034 * range, generate a random size in-between.
1036 if (o->file_size_low == o->file_size_high)
1037 f->io_size = o->file_size_low - f->file_offset;
1039 f->io_size = get_rand_file_size(td)
1043 f->io_size = f->real_file_size - f->file_offset;
1045 if (f->io_size == -1ULL)
1048 if (o->size_percent) {
1051 file_size = f->io_size + f->file_offset;
1052 f->io_size = (file_size *
1053 o->size_percent) / 100;
1054 if (f->io_size > (file_size - f->file_offset))
1055 f->io_size = file_size - f->file_offset;
1057 f->io_size -= (f->io_size % td_min_bs(td));
1059 total_size += f->io_size;
1062 if (f->filetype == FIO_TYPE_FILE &&
1063 (f->io_size + f->file_offset) > f->real_file_size &&
1064 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1065 if (!o->create_on_open) {
1067 extend_size += (f->io_size + f->file_offset);
1068 fio_file_set_extend(f);
1070 f->real_file_size = f->io_size + f->file_offset;
1074 if (td->o.block_error_hist) {
1077 assert(td->o.nr_files == 1); /* checked in fixup_options */
1079 len = f->io_size / td->o.bs[DDIR_TRIM];
1080 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1081 log_err("fio: cannot calculate block histogram with "
1082 "%d trim blocks, maximum %d\n",
1083 len, MAX_NR_BLOCK_INFOS);
1084 td_verror(td, EINVAL, "block_error_hist");
1088 td->ts.nr_block_infos = len;
1089 for (i = 0; i < len; i++)
1090 td->ts.block_infos[i] =
1091 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1093 td->ts.nr_block_infos = 0;
1095 if (!o->size || (total_size && o->size > total_size))
1096 o->size = total_size;
1098 if (o->size < td_min_bs(td)) {
1099 log_err("fio: blocksize too large for data set\n");
1104 * See if we need to extend some files, typically needed when our
1105 * target regular files don't exist yet, but our jobs require them
1106 * initially due to read I/Os.
1110 if (output_format & FIO_OUTPUT_NORMAL) {
1111 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1113 need_extend > 1 ? "s" : "",
1115 need_extend > 1 ? "s" : "",
1116 need_extend > 1 ? "total " : "",
1120 for_each_file(td, f, i) {
1121 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1123 if (!fio_file_extend(f))
1126 assert(f->filetype == FIO_TYPE_FILE);
1127 fio_file_clear_extend(f);
1128 if (!o->fill_device) {
1129 old_len = f->real_file_size;
1130 extend_len = f->io_size + f->file_offset -
1133 f->real_file_size = (f->io_size + f->file_offset);
1134 err = extend_file(td, f);
1138 err = __file_invalidate_cache(td, f, old_len,
1142 * Shut up static checker
1158 o->zone_size = o->size;
1161 * iolog already set the total io size, if we read back
1164 if (!o->read_iolog_file) {
1166 td->total_io_size = o->io_size * o->loops;
1168 td->total_io_size = o->size * o->loops;
1175 td_restore_runstate(td, old_state);
1178 log_err("%s: you need to specify valid offset=\n", o->name);
1180 td_restore_runstate(td, old_state);
1184 int pre_read_files(struct thread_data *td)
1189 dprint(FD_FILE, "pre_read files\n");
1191 for_each_file(td, f, i) {
1192 if (pre_read_file(td, f))
1199 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1201 unsigned int range_size, seed;
1202 unsigned long nranges;
1205 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1206 fsize = min(f->real_file_size, f->io_size);
1208 nranges = (fsize + range_size - 1) / range_size;
1210 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1211 if (!td->o.rand_repeatable)
1212 seed = td->rand_seeds[4];
1214 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1215 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1216 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1217 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1218 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1219 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1224 static int init_rand_distribution(struct thread_data *td)
1230 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1233 state = td_bump_runstate(td, TD_SETTING_UP);
1235 for_each_file(td, f, i)
1236 __init_rand_distribution(td, f);
1238 td_restore_runstate(td, state);
1244 * Check if the number of blocks exceeds the randomness capability of
1245 * the selected generator. Tausworthe is 32-bit, the others are fullly
1248 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1251 if (blocks <= FRAND32_MAX)
1253 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1257 * If the user hasn't specified a random generator, switch
1258 * to tausworthe64 with informational warning. If the user did
1259 * specify one, just warn.
1261 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1264 if (!fio_option_is_set(&td->o, random_generator)) {
1265 log_info("fio: Switching to tausworthe64. Use the "
1266 "random_generator= option to get rid of this "
1268 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1273 * Just make this information to avoid breaking scripts.
1275 log_info("fio: Use the random_generator= option to switch to lfsr or "
1280 int init_random_map(struct thread_data *td)
1282 unsigned long long blocks;
1286 if (init_rand_distribution(td))
1291 for_each_file(td, f, i) {
1292 uint64_t fsize = min(f->real_file_size, f->io_size);
1294 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1296 if (check_rand_gen_limits(td, f, blocks))
1299 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1302 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1304 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1305 fio_file_set_lfsr(f);
1308 } else if (!td->o.norandommap) {
1309 f->io_axmap = axmap_new(blocks);
1311 fio_file_set_axmap(f);
1314 } else if (td->o.norandommap)
1317 if (!td->o.softrandommap) {
1318 log_err("fio: failed allocating random map. If running"
1319 " a large number of jobs, try the 'norandommap'"
1320 " option or set 'softrandommap'. Or give"
1321 " a larger --alloc-size to fio.\n");
1325 log_info("fio: file %s failed allocating random map. Running "
1326 "job without.\n", f->file_name);
1332 void close_files(struct thread_data *td)
1337 for_each_file(td, f, i) {
1338 if (fio_file_open(f))
1339 td_io_close_file(td, f);
1343 void close_and_free_files(struct thread_data *td)
1348 dprint(FD_FILE, "close files\n");
1350 for_each_file(td, f, i) {
1351 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1352 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1353 td_io_unlink_file(td, f);
1356 if (fio_file_open(f))
1357 td_io_close_file(td, f);
1359 remove_file_hash(f);
1361 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1362 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1363 td_io_unlink_file(td, f);
1366 sfree(f->file_name);
1367 f->file_name = NULL;
1368 if (fio_file_axmap(f)) {
1369 axmap_free(f->io_axmap);
1375 td->o.filename = NULL;
1377 free(td->file_locks);
1378 td->files_index = 0;
1380 td->file_locks = NULL;
1381 td->o.file_lock_mode = FILE_LOCK_NONE;
1385 static void get_file_type(struct fio_file *f)
1389 if (!strcmp(f->file_name, "-"))
1390 f->filetype = FIO_TYPE_PIPE;
1392 f->filetype = FIO_TYPE_FILE;
1395 /* \\.\ is the device namespace in Windows, where every file is
1397 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1398 f->filetype = FIO_TYPE_BLOCK;
1401 if (!stat(f->file_name, &sb)) {
1402 if (S_ISBLK(sb.st_mode))
1403 f->filetype = FIO_TYPE_BLOCK;
1404 else if (S_ISCHR(sb.st_mode))
1405 f->filetype = FIO_TYPE_CHAR;
1406 else if (S_ISFIFO(sb.st_mode))
1407 f->filetype = FIO_TYPE_PIPE;
1411 static bool __is_already_allocated(const char *fname, bool set)
1413 struct flist_head *entry;
1416 ret = file_bloom_exists(fname, set);
1420 flist_for_each(entry, &filename_list) {
1421 struct file_name *fn;
1423 fn = flist_entry(entry, struct file_name, list);
1425 if (!strcmp(fn->filename, fname))
1432 static bool is_already_allocated(const char *fname)
1436 fio_file_hash_lock();
1437 ret = __is_already_allocated(fname, false);
1438 fio_file_hash_unlock();
1443 static void set_already_allocated(const char *fname)
1445 struct file_name *fn;
1447 fn = malloc(sizeof(struct file_name));
1448 fn->filename = strdup(fname);
1450 fio_file_hash_lock();
1451 if (!__is_already_allocated(fname, true)) {
1452 flist_add_tail(&fn->list, &filename_list);
1455 fio_file_hash_unlock();
1463 static void free_already_allocated(void)
1465 struct flist_head *entry, *tmp;
1466 struct file_name *fn;
1468 if (flist_empty(&filename_list))
1471 fio_file_hash_lock();
1472 flist_for_each_safe(entry, tmp, &filename_list) {
1473 fn = flist_entry(entry, struct file_name, list);
1475 flist_del(&fn->list);
1479 fio_file_hash_unlock();
1482 static struct fio_file *alloc_new_file(struct thread_data *td)
1486 f = smalloc(sizeof(*f));
1494 fio_file_reset(td, f);
1498 bool exists_and_not_regfile(const char *filename)
1502 if (lstat(filename, &sb) == -1)
1505 #ifndef WIN32 /* NOT Windows */
1506 if (S_ISREG(sb.st_mode))
1509 /* \\.\ is the device namespace in Windows, where every file
1510 * is a device node */
1511 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1518 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1520 int cur_files = td->files_index;
1521 char file_name[PATH_MAX];
1525 dprint(FD_FILE, "add file %s\n", fname);
1527 if (td->o.directory)
1528 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1529 td->o.unique_filename);
1531 sprintf(file_name + len, "%s", fname);
1533 /* clean cloned siblings using existing files */
1534 if (numjob && is_already_allocated(file_name) &&
1535 !exists_and_not_regfile(fname))
1538 f = alloc_new_file(td);
1540 if (td->files_size <= td->files_index) {
1541 unsigned int new_size = td->o.nr_files + 1;
1543 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1545 td->files = realloc(td->files, new_size * sizeof(f));
1546 if (td->files == NULL) {
1547 log_err("fio: realloc OOM\n");
1550 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1551 td->file_locks = realloc(td->file_locks, new_size);
1552 if (!td->file_locks) {
1553 log_err("fio: realloc OOM\n");
1556 td->file_locks[cur_files] = FILE_LOCK_NONE;
1558 td->files_size = new_size;
1560 td->files[cur_files] = f;
1561 f->fileno = cur_files;
1564 * init function, io engine may not be loaded yet
1566 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1567 f->real_file_size = -1ULL;
1569 f->file_name = smalloc_strdup(file_name);
1575 switch (td->o.file_lock_mode) {
1576 case FILE_LOCK_NONE:
1578 case FILE_LOCK_READWRITE:
1579 f->rwlock = fio_rwlock_init();
1581 case FILE_LOCK_EXCLUSIVE:
1582 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1585 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1590 if (f->filetype == FIO_TYPE_FILE)
1591 td->nr_normal_files++;
1593 set_already_allocated(file_name);
1598 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1604 int add_file_exclusive(struct thread_data *td, const char *fname)
1609 for_each_file(td, f, i) {
1610 if (!strcmp(f->file_name, fname))
1614 return add_file(td, fname, 0, 1);
1617 void get_file(struct fio_file *f)
1619 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1620 assert(fio_file_open(f));
1624 int put_file(struct thread_data *td, struct fio_file *f)
1626 int f_ret = 0, ret = 0;
1628 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1630 if (!fio_file_open(f)) {
1631 assert(f->fd == -1);
1635 assert(f->references);
1636 if (--f->references)
1639 if (should_fsync(td) && td->o.fsync_on_close) {
1640 f_ret = fsync(f->fd);
1645 if (td->io_ops->close_file)
1646 ret = td->io_ops->close_file(td, f);
1651 td->nr_open_files--;
1652 fio_file_clear_open(f);
1653 assert(f->fd == -1);
1657 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1659 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1662 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1663 if (ddir == DDIR_READ)
1664 fio_rwlock_read(f->rwlock);
1666 fio_rwlock_write(f->rwlock);
1667 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1668 fio_mutex_down(f->lock);
1670 td->file_locks[f->fileno] = td->o.file_lock_mode;
1673 void unlock_file(struct thread_data *td, struct fio_file *f)
1675 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1678 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1679 fio_rwlock_unlock(f->rwlock);
1680 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1681 fio_mutex_up(f->lock);
1683 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1686 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1688 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1690 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1694 static int recurse_dir(struct thread_data *td, const char *dirname)
1700 D = opendir(dirname);
1702 char buf[FIO_VERROR_SIZE];
1704 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1705 td_verror(td, errno, buf);
1709 while ((dir = readdir(D)) != NULL) {
1710 char full_path[PATH_MAX];
1713 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1716 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1718 if (lstat(full_path, &sb) == -1) {
1719 if (errno != ENOENT) {
1720 td_verror(td, errno, "stat");
1726 if (S_ISREG(sb.st_mode)) {
1727 add_file(td, full_path, 0, 1);
1730 if (!S_ISDIR(sb.st_mode))
1733 ret = recurse_dir(td, full_path);
1742 int add_dir_files(struct thread_data *td, const char *path)
1744 int ret = recurse_dir(td, path);
1747 log_info("fio: opendir added %d files\n", td->o.nr_files);
1752 void dup_files(struct thread_data *td, struct thread_data *org)
1757 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1762 td->files = malloc(org->files_index * sizeof(f));
1764 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1765 td->file_locks = malloc(org->files_index);
1767 for_each_file(org, f, i) {
1768 struct fio_file *__f;
1770 __f = alloc_new_file(td);
1773 __f->file_name = smalloc_strdup(f->file_name);
1774 if (!__f->file_name)
1777 __f->filetype = f->filetype;
1780 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1781 __f->lock = f->lock;
1782 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1783 __f->rwlock = f->rwlock;
1790 * Returns the index that matches the filename, or -1 if not there
1792 int get_fileno(struct thread_data *td, const char *fname)
1797 for_each_file(td, f, i)
1798 if (!strcmp(f->file_name, fname))
1805 * For log usage, where we add/open/close files automatically
1807 void free_release_files(struct thread_data *td)
1811 td->o.open_files = 0;
1812 td->files_index = 0;
1813 td->nr_normal_files = 0;
1816 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1820 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1821 f->last_pos[i] = f->file_offset;
1822 f->last_start[i] = -1ULL;
1825 if (fio_file_axmap(f))
1826 axmap_reset(f->io_axmap);
1827 else if (fio_file_lfsr(f))
1828 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1831 bool fio_files_done(struct thread_data *td)
1836 for_each_file(td, f, i)
1837 if (!fio_file_done(f))
1843 /* free memory used in initialization phase only */
1844 void filesetup_mem_free(void)
1846 free_already_allocated();