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 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)
62 if (td->o.fill_device)
65 switch (td->o.fallocate_mode) {
66 case FIO_FALLOCATE_NATIVE:
67 native_fallocate(td, f);
69 case FIO_FALLOCATE_NONE:
71 #ifdef CONFIG_POSIX_FALLOCATE
72 case FIO_FALLOCATE_POSIX: {
75 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
77 (unsigned long long) f->real_file_size);
79 r = posix_fallocate(f->fd, 0, f->real_file_size);
81 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
84 #endif /* CONFIG_POSIX_FALLOCATE */
85 #ifdef CONFIG_LINUX_FALLOCATE
86 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");
99 #endif /* CONFIG_LINUX_FALLOCATE */
101 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
107 * Leaves f->fd open on success, caller must close
109 static int extend_file(struct thread_data *td, struct fio_file *f)
111 int new_layout = 0, unlink_file = 0, flags;
112 unsigned long long left;
117 log_err("fio: refusing extend of file due to read-only\n");
122 * check if we need to lay the file out complete again. fio
123 * does that for operations involving reads, or for writes
124 * where overwrite is set
127 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
128 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
130 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
133 if (unlink_file || new_layout) {
136 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
138 ret = td_io_unlink_file(td, f);
139 if (ret != 0 && ret != ENOENT) {
140 td_verror(td, errno, "unlink");
146 if (td->o.allow_create)
155 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
156 f->fd = open(f->file_name, flags, 0644);
160 if (err == ENOENT && !td->o.allow_create)
161 log_err("fio: file creation disallowed by "
162 "allow_file_create=0\n");
164 td_verror(td, err, "open");
168 fallocate_file(td, f);
171 * If our jobs don't require regular files initially, we're done.
177 * The size will be -1ULL when fill_device is used, so don't truncate
178 * or fallocate this file, just write it
180 if (!td->o.fill_device) {
181 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
182 (unsigned long long) f->real_file_size);
183 if (ftruncate(f->fd, f->real_file_size) == -1) {
184 if (errno != EFBIG) {
185 td_verror(td, errno, "ftruncate");
191 left = f->real_file_size;
192 bs = td->o.max_bs[DDIR_WRITE];
198 td_verror(td, errno, "malloc");
202 while (left && !td->terminate) {
208 fill_io_buffer(td, b, bs, bs);
210 r = write(f->fd, b, bs);
220 if (td->o.fill_device)
222 log_info("fio: ENOSPC on laying out "
226 td_verror(td, errno, "write");
228 td_verror(td, EIO, "write");
235 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
236 td_io_unlink_file(td, f);
237 } else if (td->o.create_fsync) {
238 if (fsync(f->fd) < 0) {
239 td_verror(td, errno, "fsync");
243 if (td->o.fill_device && !td_write(td)) {
244 fio_file_clear_size_known(f);
245 if (td_io_get_file_size(td, f))
247 if (f->io_size > f->real_file_size)
248 f->io_size = f->real_file_size;
262 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
264 int r, did_open = 0, old_runstate;
265 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 offset_align is provided, set initial offset
876 if (fio_option_is_set(o, start_offset_align)) {
877 align_bs = o->start_offset_align;
879 /* else take the minimum block size */
880 align_bs = td_min_bs(td);
883 /* calculate the raw offset */
884 offset = (f->real_file_size * o->start_offset_percent / 100) +
885 (td->subjob_number * o->offset_increment);
888 * block align the offset at the next available boundary at
889 * ceiling(offset / align_bs) * align_bs
891 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
894 /* start_offset_percent not set */
895 offset = o->start_offset +
896 td->subjob_number * o->offset_increment;
903 * Open the files and setup files sizes, creating files if necessary.
905 int setup_files(struct thread_data *td)
907 unsigned long long total_size, extend_size;
908 struct thread_options *o = &td->o;
910 unsigned int i, nr_fs_extra = 0;
911 int err = 0, need_extend;
913 const unsigned int bs = td_min_bs(td);
916 dprint(FD_FILE, "setup files\n");
918 old_state = td_bump_runstate(td, TD_SETTING_UP);
920 if (o->read_iolog_file)
924 * Find out physical size of files or devices for this thread,
925 * before we determine I/O size and range of our targets.
926 * If ioengine defines a setup() method, it's responsible for
927 * opening the files and setting f->real_file_size to indicate
928 * the valid range for that file.
930 if (td->io_ops->setup)
931 err = td->io_ops->setup(td);
933 err = get_file_sizes(td);
939 * check sizes. if the files/devices do not exist and the size
940 * isn't passed to fio, abort.
943 for_each_file(td, f, i) {
945 if (f->real_file_size == -1ULL)
948 total_size += f->real_file_size;
952 td->fill_device_size = get_fs_free_counts(td);
955 * device/file sizes are zero and no size given, punt
957 if ((!total_size || total_size == -1ULL) && !o->size &&
958 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
959 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
960 log_err("%s: you need to specify size=\n", o->name);
961 td_verror(td, EINVAL, "total_file_size");
966 * Calculate per-file size and potential extra size for the
967 * first files, if needed (i.e. if we don't have a fixed size).
969 if (!o->file_size_low && o->nr_files) {
972 fs = o->size / o->nr_files;
973 all_fs = fs * o->nr_files;
975 if (all_fs < o->size)
976 nr_fs_extra = (o->size - all_fs) / bs;
980 * now file sizes are known, so we can set ->io_size. if size= is
981 * not given, ->io_size is just equal to ->real_file_size. if size
982 * is given, ->io_size is size / nr_files.
984 extend_size = total_size = 0;
986 for_each_file(td, f, i) {
987 f->file_offset = get_start_offset(td, f);
990 * Update ->io_size depending on options specified.
991 * ->file_size_low being 0 means filesize option isn't set.
992 * Non zero ->file_size_low equals ->file_size_high means
993 * filesize option is set in a fixed size format.
994 * Non zero ->file_size_low not equals ->file_size_high means
995 * filesize option is set in a range format.
997 if (!o->file_size_low) {
999 * no file size or range given, file size is equal to
1000 * total size divided by number of files. If the size
1001 * doesn't divide nicely with the min blocksize,
1002 * make the first files bigger.
1011 * We normally don't come here for regular files, but
1012 * if the result is 0 for a regular file, set it to the
1013 * real file size. This could be size of the existing
1014 * one if it already exists, but otherwise will be set
1015 * to 0. A new file won't be created because
1016 * ->io_size + ->file_offset equals ->real_file_size.
1019 if (f->file_offset > f->real_file_size)
1021 f->io_size = f->real_file_size - f->file_offset;
1023 log_info("fio: file %s may be ignored\n",
1026 } else if (f->real_file_size < o->file_size_low ||
1027 f->real_file_size > o->file_size_high) {
1028 if (f->file_offset > o->file_size_low)
1031 * file size given. if it's fixed, use that. if it's a
1032 * range, generate a random size in-between.
1034 if (o->file_size_low == o->file_size_high)
1035 f->io_size = o->file_size_low - f->file_offset;
1037 f->io_size = get_rand_file_size(td)
1041 f->io_size = f->real_file_size - f->file_offset;
1043 if (f->io_size == -1ULL)
1046 if (o->size_percent) {
1049 file_size = f->io_size + f->file_offset;
1050 f->io_size = (file_size *
1051 o->size_percent) / 100;
1052 if (f->io_size > (file_size - f->file_offset))
1053 f->io_size = file_size - f->file_offset;
1055 f->io_size -= (f->io_size % td_min_bs(td));
1057 total_size += f->io_size;
1060 if (f->filetype == FIO_TYPE_FILE &&
1061 (f->io_size + f->file_offset) > f->real_file_size &&
1062 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1063 if (!o->create_on_open) {
1065 extend_size += (f->io_size + f->file_offset);
1066 fio_file_set_extend(f);
1068 f->real_file_size = f->io_size + f->file_offset;
1072 if (td->o.block_error_hist) {
1075 assert(td->o.nr_files == 1); /* checked in fixup_options */
1077 len = f->io_size / td->o.bs[DDIR_TRIM];
1078 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1079 log_err("fio: cannot calculate block histogram with "
1080 "%d trim blocks, maximum %d\n",
1081 len, MAX_NR_BLOCK_INFOS);
1082 td_verror(td, EINVAL, "block_error_hist");
1086 td->ts.nr_block_infos = len;
1087 for (i = 0; i < len; i++)
1088 td->ts.block_infos[i] =
1089 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1091 td->ts.nr_block_infos = 0;
1093 if (!o->size || (total_size && o->size > total_size))
1094 o->size = total_size;
1096 if (o->size < td_min_bs(td)) {
1097 log_err("fio: blocksize too large for data set\n");
1102 * See if we need to extend some files, typically needed when our
1103 * target regular files don't exist yet, but our jobs require them
1104 * initially due to read I/Os.
1108 if (output_format & FIO_OUTPUT_NORMAL) {
1109 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1111 need_extend > 1 ? "s" : "",
1113 need_extend > 1 ? "s" : "",
1114 need_extend > 1 ? "total " : "",
1118 for_each_file(td, f, i) {
1119 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1121 if (!fio_file_extend(f))
1124 assert(f->filetype == FIO_TYPE_FILE);
1125 fio_file_clear_extend(f);
1126 if (!o->fill_device) {
1127 old_len = f->real_file_size;
1128 extend_len = f->io_size + f->file_offset -
1131 f->real_file_size = (f->io_size + f->file_offset);
1132 err = extend_file(td, f);
1136 err = __file_invalidate_cache(td, f, old_len,
1140 * Shut up static checker
1156 o->zone_size = o->size;
1159 * iolog already set the total io size, if we read back
1162 if (!o->read_iolog_file) {
1164 td->total_io_size = o->io_size * o->loops;
1166 td->total_io_size = o->size * o->loops;
1173 td_restore_runstate(td, old_state);
1176 log_err("%s: you need to specify valid offset=\n", o->name);
1178 td_restore_runstate(td, old_state);
1182 bool pre_read_files(struct thread_data *td)
1187 dprint(FD_FILE, "pre_read files\n");
1189 for_each_file(td, f, i) {
1190 if (!pre_read_file(td, f))
1197 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1199 unsigned int range_size, seed;
1200 unsigned long nranges;
1203 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1204 fsize = min(f->real_file_size, f->io_size);
1206 nranges = (fsize + range_size - 1) / range_size;
1208 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1209 if (!td->o.rand_repeatable)
1210 seed = td->rand_seeds[4];
1212 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1213 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1214 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1215 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1216 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1217 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1220 static bool init_rand_distribution(struct thread_data *td)
1226 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1229 state = td_bump_runstate(td, TD_SETTING_UP);
1231 for_each_file(td, f, i)
1232 __init_rand_distribution(td, f);
1234 td_restore_runstate(td, state);
1239 * Check if the number of blocks exceeds the randomness capability of
1240 * the selected generator. Tausworthe is 32-bit, the others are fullly
1243 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1246 if (blocks <= FRAND32_MAX)
1248 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1252 * If the user hasn't specified a random generator, switch
1253 * to tausworthe64 with informational warning. If the user did
1254 * specify one, just warn.
1256 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1259 if (!fio_option_is_set(&td->o, random_generator)) {
1260 log_info("fio: Switching to tausworthe64. Use the "
1261 "random_generator= option to get rid of this "
1263 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1268 * Just make this information to avoid breaking scripts.
1270 log_info("fio: Use the random_generator= option to switch to lfsr or "
1275 bool init_random_map(struct thread_data *td)
1277 unsigned long long blocks;
1281 if (init_rand_distribution(td))
1286 for_each_file(td, f, i) {
1287 uint64_t fsize = min(f->real_file_size, f->io_size);
1289 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1291 if (check_rand_gen_limits(td, f, blocks))
1294 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1297 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1299 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1300 fio_file_set_lfsr(f);
1303 } else if (!td->o.norandommap) {
1304 f->io_axmap = axmap_new(blocks);
1306 fio_file_set_axmap(f);
1309 } else if (td->o.norandommap)
1312 if (!td->o.softrandommap) {
1313 log_err("fio: failed allocating random map. If running"
1314 " a large number of jobs, try the 'norandommap'"
1315 " option or set 'softrandommap'. Or give"
1316 " a larger --alloc-size to fio.\n");
1320 log_info("fio: file %s failed allocating random map. Running "
1321 "job without.\n", f->file_name);
1327 void close_files(struct thread_data *td)
1332 for_each_file(td, f, i) {
1333 if (fio_file_open(f))
1334 td_io_close_file(td, f);
1338 void close_and_free_files(struct thread_data *td)
1342 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1344 dprint(FD_FILE, "close files\n");
1346 for_each_file(td, f, i) {
1347 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1348 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1349 td_io_unlink_file(td, f);
1352 if (fio_file_open(f))
1353 td_io_close_file(td, f);
1355 remove_file_hash(f);
1357 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1358 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1359 td_io_unlink_file(td, f);
1365 sfree(f->file_name);
1366 f->file_name = NULL;
1367 if (fio_file_axmap(f)) {
1368 axmap_free(f->io_axmap);
1377 td->o.filename = NULL;
1379 free(td->file_locks);
1380 td->files_index = 0;
1382 td->file_locks = NULL;
1383 td->o.file_lock_mode = FILE_LOCK_NONE;
1387 static void get_file_type(struct fio_file *f)
1391 if (!strcmp(f->file_name, "-"))
1392 f->filetype = FIO_TYPE_PIPE;
1394 f->filetype = FIO_TYPE_FILE;
1397 /* \\.\ is the device namespace in Windows, where every file is
1399 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1400 f->filetype = FIO_TYPE_BLOCK;
1403 if (!stat(f->file_name, &sb)) {
1404 if (S_ISBLK(sb.st_mode))
1405 f->filetype = FIO_TYPE_BLOCK;
1406 else if (S_ISCHR(sb.st_mode))
1407 f->filetype = FIO_TYPE_CHAR;
1408 else if (S_ISFIFO(sb.st_mode))
1409 f->filetype = FIO_TYPE_PIPE;
1413 static bool __is_already_allocated(const char *fname, bool set)
1415 struct flist_head *entry;
1418 ret = file_bloom_exists(fname, set);
1422 flist_for_each(entry, &filename_list) {
1423 struct file_name *fn;
1425 fn = flist_entry(entry, struct file_name, list);
1427 if (!strcmp(fn->filename, fname))
1434 static bool is_already_allocated(const char *fname)
1438 fio_file_hash_lock();
1439 ret = __is_already_allocated(fname, false);
1440 fio_file_hash_unlock();
1445 static void set_already_allocated(const char *fname)
1447 struct file_name *fn;
1449 fn = malloc(sizeof(struct file_name));
1450 fn->filename = strdup(fname);
1452 fio_file_hash_lock();
1453 if (!__is_already_allocated(fname, true)) {
1454 flist_add_tail(&fn->list, &filename_list);
1457 fio_file_hash_unlock();
1465 static void free_already_allocated(void)
1467 struct flist_head *entry, *tmp;
1468 struct file_name *fn;
1470 if (flist_empty(&filename_list))
1473 fio_file_hash_lock();
1474 flist_for_each_safe(entry, tmp, &filename_list) {
1475 fn = flist_entry(entry, struct file_name, list);
1477 flist_del(&fn->list);
1481 fio_file_hash_unlock();
1484 static struct fio_file *alloc_new_file(struct thread_data *td)
1488 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1489 f = calloc(1, sizeof(*f));
1491 f = smalloc(sizeof(*f));
1499 fio_file_reset(td, f);
1503 bool exists_and_not_regfile(const char *filename)
1507 if (lstat(filename, &sb) == -1)
1510 #ifndef WIN32 /* NOT Windows */
1511 if (S_ISREG(sb.st_mode))
1514 /* \\.\ is the device namespace in Windows, where every file
1515 * is a device node */
1516 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1523 static bool create_work_dirs(struct thread_data *td, const char *fname)
1525 char path[PATH_MAX];
1528 if (td->o.directory) {
1529 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1530 FIO_OS_PATH_SEPARATOR, fname);
1531 start = strstr(path, fname);
1533 snprintf(path, PATH_MAX, "%s", fname);
1538 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1543 #ifdef CONFIG_HAVE_MKDIR_TWO
1544 if (mkdir(path, 0600) && errno != EEXIST) {
1546 if (mkdir(path) && errno != EEXIST) {
1548 log_err("fio: failed to create dir (%s): %d\n",
1552 *end = FIO_OS_PATH_SEPARATOR;
1555 td->flags |= TD_F_DIRS_CREATED;
1559 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1561 int cur_files = td->files_index;
1562 char file_name[PATH_MAX];
1566 dprint(FD_FILE, "add file %s\n", fname);
1568 if (td->o.directory)
1569 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1570 td->o.unique_filename);
1572 sprintf(file_name + len, "%s", fname);
1574 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1575 !(td->flags & TD_F_DIRS_CREATED) &&
1576 !create_work_dirs(td, fname))
1579 /* clean cloned siblings using existing files */
1580 if (numjob && is_already_allocated(file_name) &&
1581 !exists_and_not_regfile(fname))
1584 f = alloc_new_file(td);
1586 if (td->files_size <= td->files_index) {
1587 unsigned int new_size = td->o.nr_files + 1;
1589 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1591 td->files = realloc(td->files, new_size * sizeof(f));
1592 if (td->files == NULL) {
1593 log_err("fio: realloc OOM\n");
1596 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1597 td->file_locks = realloc(td->file_locks, new_size);
1598 if (!td->file_locks) {
1599 log_err("fio: realloc OOM\n");
1602 td->file_locks[cur_files] = FILE_LOCK_NONE;
1604 td->files_size = new_size;
1606 td->files[cur_files] = f;
1607 f->fileno = cur_files;
1610 * init function, io engine may not be loaded yet
1612 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1613 f->real_file_size = -1ULL;
1615 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1616 f->file_name = strdup(file_name);
1618 f->file_name = smalloc_strdup(file_name);
1624 switch (td->o.file_lock_mode) {
1625 case FILE_LOCK_NONE:
1627 case FILE_LOCK_READWRITE:
1628 f->rwlock = fio_rwlock_init();
1630 case FILE_LOCK_EXCLUSIVE:
1631 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1634 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1639 if (f->filetype == FIO_TYPE_FILE)
1640 td->nr_normal_files++;
1642 if (td->o.numjobs > 1)
1643 set_already_allocated(file_name);
1648 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1654 int add_file_exclusive(struct thread_data *td, const char *fname)
1659 for_each_file(td, f, i) {
1660 if (!strcmp(f->file_name, fname))
1664 return add_file(td, fname, 0, 1);
1667 void get_file(struct fio_file *f)
1669 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1670 assert(fio_file_open(f));
1674 int put_file(struct thread_data *td, struct fio_file *f)
1676 int f_ret = 0, ret = 0;
1678 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1680 if (!fio_file_open(f)) {
1681 assert(f->fd == -1);
1685 assert(f->references);
1686 if (--f->references)
1689 if (should_fsync(td) && td->o.fsync_on_close) {
1690 f_ret = fsync(f->fd);
1695 if (td->io_ops->close_file)
1696 ret = td->io_ops->close_file(td, f);
1701 td->nr_open_files--;
1702 fio_file_clear_open(f);
1703 assert(f->fd == -1);
1707 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1709 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1712 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1713 if (ddir == DDIR_READ)
1714 fio_rwlock_read(f->rwlock);
1716 fio_rwlock_write(f->rwlock);
1717 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1718 fio_mutex_down(f->lock);
1720 td->file_locks[f->fileno] = td->o.file_lock_mode;
1723 void unlock_file(struct thread_data *td, struct fio_file *f)
1725 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1728 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1729 fio_rwlock_unlock(f->rwlock);
1730 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1731 fio_mutex_up(f->lock);
1733 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1736 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1738 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1740 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1744 static bool recurse_dir(struct thread_data *td, const char *dirname)
1750 D = opendir(dirname);
1752 char buf[FIO_VERROR_SIZE];
1754 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1755 td_verror(td, errno, buf);
1759 while ((dir = readdir(D)) != NULL) {
1760 char full_path[PATH_MAX];
1763 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1766 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1768 if (lstat(full_path, &sb) == -1) {
1769 if (errno != ENOENT) {
1770 td_verror(td, errno, "stat");
1776 if (S_ISREG(sb.st_mode)) {
1777 add_file(td, full_path, 0, 1);
1780 if (!S_ISDIR(sb.st_mode))
1783 ret = recurse_dir(td, full_path);
1792 int add_dir_files(struct thread_data *td, const char *path)
1794 int ret = recurse_dir(td, path);
1797 log_info("fio: opendir added %d files\n", td->o.nr_files);
1802 void dup_files(struct thread_data *td, struct thread_data *org)
1807 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1812 td->files = malloc(org->files_index * sizeof(f));
1814 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1815 td->file_locks = malloc(org->files_index);
1817 for_each_file(org, f, i) {
1818 struct fio_file *__f;
1820 __f = alloc_new_file(td);
1823 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1824 __f->file_name = strdup(f->file_name);
1826 __f->file_name = smalloc_strdup(f->file_name);
1827 if (!__f->file_name)
1830 __f->filetype = f->filetype;
1833 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1834 __f->lock = f->lock;
1835 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1836 __f->rwlock = f->rwlock;
1843 * Returns the index that matches the filename, or -1 if not there
1845 int get_fileno(struct thread_data *td, const char *fname)
1850 for_each_file(td, f, i)
1851 if (!strcmp(f->file_name, fname))
1858 * For log usage, where we add/open/close files automatically
1860 void free_release_files(struct thread_data *td)
1864 td->o.open_files = 0;
1865 td->files_index = 0;
1866 td->nr_normal_files = 0;
1869 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1873 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1874 f->last_pos[i] = f->file_offset;
1875 f->last_start[i] = -1ULL;
1878 if (fio_file_axmap(f))
1879 axmap_reset(f->io_axmap);
1880 else if (fio_file_lfsr(f))
1881 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1884 bool fio_files_done(struct thread_data *td)
1889 for_each_file(td, f, i)
1890 if (!fio_file_done(f))
1896 /* free memory used in initialization phase only */
1897 void filesetup_mem_free(void)
1899 free_already_allocated();
1903 * This function is for platforms which support direct I/O but not O_DIRECT.
1905 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1907 #ifdef FIO_OS_DIRECTIO
1908 int ret = fio_set_odirect(f);
1911 td_verror(td, ret, "fio_set_directio");
1912 #if defined(__sun__)
1913 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1914 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1916 log_err("fio: the file system does not seem to support direct IO\n");
1919 log_err("fio: the file system does not seem to support direct IO\n");
1926 log_err("fio: direct IO is not supported on this host operating system\n");