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)
154 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
155 f->fd = open(f->file_name, flags, 0644);
159 if (err == ENOENT && !td->o.allow_create)
160 log_err("fio: file creation disallowed by "
161 "allow_file_create=0\n");
163 td_verror(td, err, "open");
167 fallocate_file(td, f);
170 * If our jobs don't require regular files initially, we're done.
176 * The size will be -1ULL when fill_device is used, so don't truncate
177 * or fallocate this file, just write it
179 if (!td->o.fill_device) {
180 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
181 (unsigned long long) f->real_file_size);
182 if (ftruncate(f->fd, f->real_file_size) == -1) {
183 if (errno != EFBIG) {
184 td_verror(td, errno, "ftruncate");
190 left = f->real_file_size;
191 bs = td->o.max_bs[DDIR_WRITE];
197 td_verror(td, errno, "malloc");
201 while (left && !td->terminate) {
207 fill_io_buffer(td, b, bs, bs);
209 r = write(f->fd, b, bs);
219 if (td->o.fill_device)
221 log_info("fio: ENOSPC on laying out "
225 td_verror(td, errno, "write");
227 td_verror(td, EIO, "write");
234 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
235 td_io_unlink_file(td, f);
236 } else if (td->o.create_fsync) {
237 if (fsync(f->fd) < 0) {
238 td_verror(td, errno, "fsync");
242 if (td->o.fill_device && !td_write(td)) {
243 fio_file_clear_size_known(f);
244 if (td_io_get_file_size(td, f))
246 if (f->io_size > f->real_file_size)
247 f->io_size = f->real_file_size;
261 static int pre_read_file(struct thread_data *td, struct fio_file *f)
263 int ret = 0, r, did_open = 0, old_runstate;
264 unsigned long long left;
268 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
269 td_ioengine_flagged(td, FIO_NOIO))
272 if (f->filetype == FIO_TYPE_CHAR)
275 if (!fio_file_open(f)) {
276 if (td->io_ops->open_file(td, f)) {
277 log_err("fio: cannot pre-read, failed to open file\n");
283 old_runstate = td_bump_runstate(td, TD_PRE_READING);
286 bs = td->o.max_bs[DDIR_READ];
292 td_verror(td, errno, "malloc");
298 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
299 td_verror(td, errno, "lseek");
300 log_err("fio: failed to lseek pre-read file\n");
305 while (left && !td->terminate) {
309 r = read(f->fd, b, bs);
315 td_verror(td, EIO, "pre_read");
321 td_restore_runstate(td, old_runstate);
324 td->io_ops->close_file(td, f);
330 unsigned long long get_rand_file_size(struct thread_data *td)
332 unsigned long long ret, sized;
336 frand_max = rand_max(&td->file_size_state);
337 r = __rand(&td->file_size_state);
338 sized = td->o.file_size_high - td->o.file_size_low;
339 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
340 ret += td->o.file_size_low;
341 ret -= (ret % td->o.rw_min_bs);
345 static int file_size(struct thread_data *td, struct fio_file *f)
349 if (stat(f->file_name, &st) == -1) {
350 td_verror(td, errno, "fstat");
354 f->real_file_size = st.st_size;
358 static int bdev_size(struct thread_data *td, struct fio_file *f)
360 unsigned long long bytes = 0;
363 if (td->io_ops->open_file(td, f)) {
364 log_err("fio: failed opening blockdev %s for size check\n",
369 r = blockdev_size(f, &bytes);
371 td_verror(td, r, "blockdev_size");
376 log_err("%s: zero sized block device?\n", f->file_name);
380 f->real_file_size = bytes;
381 td->io_ops->close_file(td, f);
384 td->io_ops->close_file(td, f);
388 static int char_size(struct thread_data *td, struct fio_file *f)
390 #ifdef FIO_HAVE_CHARDEV_SIZE
391 unsigned long long bytes = 0;
394 if (td->io_ops->open_file(td, f)) {
395 log_err("fio: failed opening chardev %s for size check\n",
400 r = chardev_size(f, &bytes);
402 td_verror(td, r, "chardev_size");
407 log_err("%s: zero sized char device?\n", f->file_name);
411 f->real_file_size = bytes;
412 td->io_ops->close_file(td, f);
415 td->io_ops->close_file(td, f);
418 f->real_file_size = -1ULL;
423 static int get_file_size(struct thread_data *td, struct fio_file *f)
427 if (fio_file_size_known(f))
430 if (f->filetype == FIO_TYPE_FILE)
431 ret = file_size(td, f);
432 else if (f->filetype == FIO_TYPE_BLOCK)
433 ret = bdev_size(td, f);
434 else if (f->filetype == FIO_TYPE_CHAR)
435 ret = char_size(td, f);
437 f->real_file_size = -1ULL;
440 * Leave ->real_file_size with 0 since it could be expectation
441 * of initial setup for regular files.
447 * If ->real_file_size is -1, a conditional for the message
448 * "offset extends end" is always true, but it makes no sense,
449 * so just return the same value here.
451 if (f->real_file_size == -1ULL) {
452 log_info("%s: failed to get file size of %s\n", td->o.name,
457 if (td->o.start_offset && f->file_offset == 0)
458 dprint(FD_FILE, "offset of file %s not initialized yet\n",
461 * ->file_offset normally hasn't been initialized yet, so this
462 * is basically always false.
464 if (f->file_offset > f->real_file_size) {
465 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
466 (unsigned long long) f->file_offset,
467 (unsigned long long) f->real_file_size);
471 fio_file_set_size_known(f);
475 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
476 unsigned long long off,
477 unsigned long long len)
479 int errval = 0, ret = 0;
488 off = f->file_offset;
490 if (len == -1ULL || off == -1ULL)
493 if (td->io_ops->invalidate) {
494 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
496 ret = td->io_ops->invalidate(td, f);
499 } else if (f->filetype == FIO_TYPE_FILE) {
500 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
501 f->file_name, off, len);
502 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
505 } else if (f->filetype == FIO_TYPE_BLOCK) {
508 dprint(FD_IO, "drop page cache %s\n", f->file_name);
509 ret = blockdev_invalidate_cache(f);
510 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
512 * Linux multipath devices reject ioctl while
513 * the maps are being updated. That window can
514 * last tens of milliseconds; we'll try up to
515 * a quarter of a second.
518 ret = blockdev_invalidate_cache(f);
520 if (ret < 0 && errno == EACCES && geteuid()) {
522 log_err("fio: only root may flush block "
523 "devices. Cache flush bypassed!\n");
530 } else if (f->filetype == FIO_TYPE_CHAR ||
531 f->filetype == FIO_TYPE_PIPE) {
532 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
537 * Cache flushing isn't a fatal condition, and we know it will
538 * happen on some platforms where we don't have the proper
539 * function to flush eg block device caches. So just warn and
540 * continue on our way.
543 log_info("fio: cache invalidation of %s failed: %s\n",
544 f->file_name, strerror(errval));
550 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
552 if (!fio_file_open(f))
555 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
558 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
562 dprint(FD_FILE, "fd close %s\n", f->file_name);
566 if (close(f->fd) < 0)
571 if (f->shadow_fd != -1) {
580 int file_lookup_open(struct fio_file *f, int flags)
582 struct fio_file *__f;
585 __f = lookup_file_hash(f->file_name);
587 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
591 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
599 f->fd = open(f->file_name, flags, 0600);
603 static int file_close_shadow_fds(struct thread_data *td)
609 for_each_file(td, f, i) {
610 if (f->shadow_fd == -1)
621 int generic_open_file(struct thread_data *td, struct fio_file *f)
627 dprint(FD_FILE, "fd open %s\n", f->file_name);
629 if (!strcmp(f->file_name, "-")) {
631 log_err("fio: can't read/write to stdin/out\n");
637 * move output logging to stderr, if we are writing to stdout
646 flags |= OS_O_DIRECT;
649 td_verror(td, EINVAL, "OS does not support atomic IO");
652 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
656 if (td->o.create_on_open && td->o.allow_create)
659 if (f->filetype != FIO_TYPE_FILE)
660 flags |= FIO_O_NOATIME;
667 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
671 f->fd = dup(STDOUT_FILENO);
673 from_hash = file_lookup_open(f, flags);
674 } else if (td_read(td)) {
675 if (f->filetype == FIO_TYPE_CHAR && !read_only)
681 f->fd = dup(STDIN_FILENO);
683 from_hash = file_lookup_open(f, flags);
684 } else if (td_trim(td)) {
685 assert(!td_rw(td)); /* should have matched above */
687 from_hash = file_lookup_open(f, flags);
691 char buf[FIO_VERROR_SIZE];
694 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
695 flags &= ~FIO_O_NOATIME;
698 if (__e == EMFILE && file_close_shadow_fds(td))
701 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
703 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
704 log_err("fio: looks like your file system does not " \
705 "support direct=1/buffered=0\n");
708 td_verror(td, __e, buf);
712 if (!from_hash && f->fd != -1) {
713 if (add_file_hash(f)) {
717 * Stash away descriptor for later close. This is to
718 * work-around a "feature" on Linux, where a close of
719 * an fd that has been opened for write will trigger
720 * udev to call blkid to check partitions, fs id, etc.
721 * That pollutes the device cache, which can slow down
722 * unbuffered accesses.
724 if (f->shadow_fd == -1)
725 f->shadow_fd = f->fd;
728 * OK to ignore, we haven't done anything
731 ret = generic_close_file(td, f);
741 * This function i.e. get_file_size() is the default .get_file_size
742 * implementation of majority of I/O engines.
744 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
746 return get_file_size(td, f);
750 * open/close all files, so that ->real_file_size gets set
752 static int get_file_sizes(struct thread_data *td)
758 for_each_file(td, f, i) {
759 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
762 if (td_io_get_file_size(td, f)) {
763 if (td->error != ENOENT) {
764 log_err("%s\n", td->verror);
772 * There are corner cases where we end up with -1 for
773 * ->real_file_size due to unsupported file type, etc.
774 * We then just set to size option value divided by number
775 * of files, similar to the way file ->io_size is set.
776 * stat(2) failure doesn't set ->real_file_size to -1.
778 if (f->real_file_size == -1ULL && td->o.size)
779 f->real_file_size = td->o.size / td->o.nr_files;
786 struct flist_head list;
793 * Get free number of bytes for each file on each unique mount.
795 static unsigned long long get_fs_free_counts(struct thread_data *td)
797 struct flist_head *n, *tmp;
798 unsigned long long ret = 0;
799 struct fio_mount *fm;
804 for_each_file(td, f, i) {
808 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
809 if (f->real_file_size != -1ULL)
810 ret += f->real_file_size;
812 } else if (f->filetype != FIO_TYPE_FILE)
816 strncpy(buf, f->file_name, 255);
818 if (stat(buf, &sb) < 0) {
822 if (stat(buf, &sb) < 0)
827 flist_for_each(n, &list) {
828 fm = flist_entry(n, struct fio_mount, list);
829 if (fm->key == sb.st_dev)
838 fm = calloc(1, sizeof(*fm));
839 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
840 fm->base = basename(fm->__base);
842 flist_add(&fm->list, &list);
845 flist_for_each_safe(n, tmp, &list) {
846 unsigned long long sz;
848 fm = flist_entry(n, struct fio_mount, list);
849 flist_del(&fm->list);
851 sz = get_fs_free_size(fm->base);
852 if (sz && sz != -1ULL)
861 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
863 struct thread_options *o = &td->o;
864 unsigned long long align_bs;
865 unsigned long long offset;
867 if (o->file_append && f->filetype == FIO_TYPE_FILE)
868 return f->real_file_size;
870 if (o->start_offset_percent > 0) {
872 * if blockalign is provided, find the min across read, write,
875 if (fio_option_is_set(o, ba)) {
876 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
877 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
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 int 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 int __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);
1222 static int init_rand_distribution(struct thread_data *td)
1228 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1231 state = td_bump_runstate(td, TD_SETTING_UP);
1233 for_each_file(td, f, i)
1234 __init_rand_distribution(td, f);
1236 td_restore_runstate(td, state);
1242 * Check if the number of blocks exceeds the randomness capability of
1243 * the selected generator. Tausworthe is 32-bit, the others are fullly
1246 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1249 if (blocks <= FRAND32_MAX)
1251 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1255 * If the user hasn't specified a random generator, switch
1256 * to tausworthe64 with informational warning. If the user did
1257 * specify one, just warn.
1259 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1262 if (!fio_option_is_set(&td->o, random_generator)) {
1263 log_info("fio: Switching to tausworthe64. Use the "
1264 "random_generator= option to get rid of this "
1266 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1271 * Just make this information to avoid breaking scripts.
1273 log_info("fio: Use the random_generator= option to switch to lfsr or "
1278 int init_random_map(struct thread_data *td)
1280 unsigned long long blocks;
1284 if (init_rand_distribution(td))
1289 for_each_file(td, f, i) {
1290 uint64_t fsize = min(f->real_file_size, f->io_size);
1292 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1294 if (check_rand_gen_limits(td, f, blocks))
1297 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1300 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1302 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1303 fio_file_set_lfsr(f);
1306 } else if (!td->o.norandommap) {
1307 f->io_axmap = axmap_new(blocks);
1309 fio_file_set_axmap(f);
1312 } else if (td->o.norandommap)
1315 if (!td->o.softrandommap) {
1316 log_err("fio: failed allocating random map. If running"
1317 " a large number of jobs, try the 'norandommap'"
1318 " option or set 'softrandommap'. Or give"
1319 " a larger --alloc-size to fio.\n");
1323 log_info("fio: file %s failed allocating random map. Running "
1324 "job without.\n", f->file_name);
1330 void close_files(struct thread_data *td)
1335 for_each_file(td, f, i) {
1336 if (fio_file_open(f))
1337 td_io_close_file(td, f);
1341 void close_and_free_files(struct thread_data *td)
1345 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1347 dprint(FD_FILE, "close files\n");
1349 for_each_file(td, f, i) {
1350 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1351 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1352 td_io_unlink_file(td, f);
1355 if (fio_file_open(f))
1356 td_io_close_file(td, f);
1358 remove_file_hash(f);
1360 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1361 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1362 td_io_unlink_file(td, f);
1368 sfree(f->file_name);
1369 f->file_name = NULL;
1370 if (fio_file_axmap(f)) {
1371 axmap_free(f->io_axmap);
1380 td->o.filename = NULL;
1382 free(td->file_locks);
1383 td->files_index = 0;
1385 td->file_locks = NULL;
1386 td->o.file_lock_mode = FILE_LOCK_NONE;
1390 static void get_file_type(struct fio_file *f)
1394 if (!strcmp(f->file_name, "-"))
1395 f->filetype = FIO_TYPE_PIPE;
1397 f->filetype = FIO_TYPE_FILE;
1400 /* \\.\ is the device namespace in Windows, where every file is
1402 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1403 f->filetype = FIO_TYPE_BLOCK;
1406 if (!stat(f->file_name, &sb)) {
1407 if (S_ISBLK(sb.st_mode))
1408 f->filetype = FIO_TYPE_BLOCK;
1409 else if (S_ISCHR(sb.st_mode))
1410 f->filetype = FIO_TYPE_CHAR;
1411 else if (S_ISFIFO(sb.st_mode))
1412 f->filetype = FIO_TYPE_PIPE;
1416 static bool __is_already_allocated(const char *fname, bool set)
1418 struct flist_head *entry;
1421 ret = file_bloom_exists(fname, set);
1425 flist_for_each(entry, &filename_list) {
1426 struct file_name *fn;
1428 fn = flist_entry(entry, struct file_name, list);
1430 if (!strcmp(fn->filename, fname))
1437 static bool is_already_allocated(const char *fname)
1441 fio_file_hash_lock();
1442 ret = __is_already_allocated(fname, false);
1443 fio_file_hash_unlock();
1448 static void set_already_allocated(const char *fname)
1450 struct file_name *fn;
1452 fn = malloc(sizeof(struct file_name));
1453 fn->filename = strdup(fname);
1455 fio_file_hash_lock();
1456 if (!__is_already_allocated(fname, true)) {
1457 flist_add_tail(&fn->list, &filename_list);
1460 fio_file_hash_unlock();
1468 static void free_already_allocated(void)
1470 struct flist_head *entry, *tmp;
1471 struct file_name *fn;
1473 if (flist_empty(&filename_list))
1476 fio_file_hash_lock();
1477 flist_for_each_safe(entry, tmp, &filename_list) {
1478 fn = flist_entry(entry, struct file_name, list);
1480 flist_del(&fn->list);
1484 fio_file_hash_unlock();
1487 static struct fio_file *alloc_new_file(struct thread_data *td)
1491 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1492 f = calloc(1, sizeof(*f));
1494 f = smalloc(sizeof(*f));
1502 fio_file_reset(td, f);
1506 bool exists_and_not_regfile(const char *filename)
1510 if (lstat(filename, &sb) == -1)
1513 #ifndef WIN32 /* NOT Windows */
1514 if (S_ISREG(sb.st_mode))
1517 /* \\.\ is the device namespace in Windows, where every file
1518 * is a device node */
1519 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1526 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1528 int cur_files = td->files_index;
1529 char file_name[PATH_MAX];
1533 dprint(FD_FILE, "add file %s\n", fname);
1535 if (td->o.directory)
1536 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1537 td->o.unique_filename);
1539 sprintf(file_name + len, "%s", fname);
1541 /* clean cloned siblings using existing files */
1542 if (numjob && is_already_allocated(file_name) &&
1543 !exists_and_not_regfile(fname))
1546 f = alloc_new_file(td);
1548 if (td->files_size <= td->files_index) {
1549 unsigned int new_size = td->o.nr_files + 1;
1551 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1553 td->files = realloc(td->files, new_size * sizeof(f));
1554 if (td->files == NULL) {
1555 log_err("fio: realloc OOM\n");
1558 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1559 td->file_locks = realloc(td->file_locks, new_size);
1560 if (!td->file_locks) {
1561 log_err("fio: realloc OOM\n");
1564 td->file_locks[cur_files] = FILE_LOCK_NONE;
1566 td->files_size = new_size;
1568 td->files[cur_files] = f;
1569 f->fileno = cur_files;
1572 * init function, io engine may not be loaded yet
1574 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1575 f->real_file_size = -1ULL;
1577 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1578 f->file_name = strdup(file_name);
1580 f->file_name = smalloc_strdup(file_name);
1586 switch (td->o.file_lock_mode) {
1587 case FILE_LOCK_NONE:
1589 case FILE_LOCK_READWRITE:
1590 f->rwlock = fio_rwlock_init();
1592 case FILE_LOCK_EXCLUSIVE:
1593 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1596 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1601 if (f->filetype == FIO_TYPE_FILE)
1602 td->nr_normal_files++;
1604 if (td->o.numjobs > 1)
1605 set_already_allocated(file_name);
1610 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1616 int add_file_exclusive(struct thread_data *td, const char *fname)
1621 for_each_file(td, f, i) {
1622 if (!strcmp(f->file_name, fname))
1626 return add_file(td, fname, 0, 1);
1629 void get_file(struct fio_file *f)
1631 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1632 assert(fio_file_open(f));
1636 int put_file(struct thread_data *td, struct fio_file *f)
1638 int f_ret = 0, ret = 0;
1640 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1642 if (!fio_file_open(f)) {
1643 assert(f->fd == -1);
1647 assert(f->references);
1648 if (--f->references)
1651 if (should_fsync(td) && td->o.fsync_on_close) {
1652 f_ret = fsync(f->fd);
1657 if (td->io_ops->close_file)
1658 ret = td->io_ops->close_file(td, f);
1663 td->nr_open_files--;
1664 fio_file_clear_open(f);
1665 assert(f->fd == -1);
1669 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1671 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1674 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1675 if (ddir == DDIR_READ)
1676 fio_rwlock_read(f->rwlock);
1678 fio_rwlock_write(f->rwlock);
1679 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1680 fio_mutex_down(f->lock);
1682 td->file_locks[f->fileno] = td->o.file_lock_mode;
1685 void unlock_file(struct thread_data *td, struct fio_file *f)
1687 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1690 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1691 fio_rwlock_unlock(f->rwlock);
1692 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1693 fio_mutex_up(f->lock);
1695 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1698 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1700 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1702 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1706 static int recurse_dir(struct thread_data *td, const char *dirname)
1712 D = opendir(dirname);
1714 char buf[FIO_VERROR_SIZE];
1716 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1717 td_verror(td, errno, buf);
1721 while ((dir = readdir(D)) != NULL) {
1722 char full_path[PATH_MAX];
1725 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1728 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1730 if (lstat(full_path, &sb) == -1) {
1731 if (errno != ENOENT) {
1732 td_verror(td, errno, "stat");
1738 if (S_ISREG(sb.st_mode)) {
1739 add_file(td, full_path, 0, 1);
1742 if (!S_ISDIR(sb.st_mode))
1745 ret = recurse_dir(td, full_path);
1754 int add_dir_files(struct thread_data *td, const char *path)
1756 int ret = recurse_dir(td, path);
1759 log_info("fio: opendir added %d files\n", td->o.nr_files);
1764 void dup_files(struct thread_data *td, struct thread_data *org)
1769 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1774 td->files = malloc(org->files_index * sizeof(f));
1776 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1777 td->file_locks = malloc(org->files_index);
1779 for_each_file(org, f, i) {
1780 struct fio_file *__f;
1782 __f = alloc_new_file(td);
1785 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1786 __f->file_name = strdup(f->file_name);
1788 __f->file_name = smalloc_strdup(f->file_name);
1789 if (!__f->file_name)
1792 __f->filetype = f->filetype;
1795 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1796 __f->lock = f->lock;
1797 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1798 __f->rwlock = f->rwlock;
1805 * Returns the index that matches the filename, or -1 if not there
1807 int get_fileno(struct thread_data *td, const char *fname)
1812 for_each_file(td, f, i)
1813 if (!strcmp(f->file_name, fname))
1820 * For log usage, where we add/open/close files automatically
1822 void free_release_files(struct thread_data *td)
1826 td->o.open_files = 0;
1827 td->files_index = 0;
1828 td->nr_normal_files = 0;
1831 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1835 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1836 f->last_pos[i] = f->file_offset;
1837 f->last_start[i] = -1ULL;
1840 if (fio_file_axmap(f))
1841 axmap_reset(f->io_axmap);
1842 else if (fio_file_lfsr(f))
1843 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1846 bool fio_files_done(struct thread_data *td)
1851 for_each_file(td, f, i)
1852 if (!fio_file_done(f))
1858 /* free memory used in initialization phase only */
1859 void filesetup_mem_free(void)
1861 free_already_allocated();
1865 * This function is for platforms which support direct I/O but not O_DIRECT.
1867 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1869 #ifdef FIO_OS_DIRECTIO
1870 int ret = fio_set_odirect(f);
1873 td_verror(td, ret, "fio_set_directio");
1874 #if defined(__sun__)
1875 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1876 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1878 log_err("fio: the file system does not seem to support direct IO\n");
1881 log_err("fio: the file system does not seem to support direct IO\n");
1888 log_err("fio: direct IO is not supported on this host operating system\n");