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 offset_align is provided, set initial offset
874 if (fio_option_is_set(o, start_offset_align)) {
875 align_bs = o->start_offset_align;
877 /* else take the minimum block size */
878 align_bs = td_min_bs(td);
881 /* calculate the raw offset */
882 offset = (f->real_file_size * o->start_offset_percent / 100) +
883 (td->subjob_number * o->offset_increment);
886 * block align the offset at the next available boundary at
887 * ceiling(offset / align_bs) * align_bs
889 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
892 /* start_offset_percent not set */
893 offset = o->start_offset +
894 td->subjob_number * o->offset_increment;
901 * Open the files and setup files sizes, creating files if necessary.
903 int setup_files(struct thread_data *td)
905 unsigned long long total_size, extend_size;
906 struct thread_options *o = &td->o;
908 unsigned int i, nr_fs_extra = 0;
909 int err = 0, need_extend;
911 const unsigned int bs = td_min_bs(td);
914 dprint(FD_FILE, "setup files\n");
916 old_state = td_bump_runstate(td, TD_SETTING_UP);
918 if (o->read_iolog_file)
922 * Find out physical size of files or devices for this thread,
923 * before we determine I/O size and range of our targets.
924 * If ioengine defines a setup() method, it's responsible for
925 * opening the files and setting f->real_file_size to indicate
926 * the valid range for that file.
928 if (td->io_ops->setup)
929 err = td->io_ops->setup(td);
931 err = get_file_sizes(td);
937 * check sizes. if the files/devices do not exist and the size
938 * isn't passed to fio, abort.
941 for_each_file(td, f, i) {
943 if (f->real_file_size == -1ULL)
946 total_size += f->real_file_size;
950 td->fill_device_size = get_fs_free_counts(td);
953 * device/file sizes are zero and no size given, punt
955 if ((!total_size || total_size == -1ULL) && !o->size &&
956 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
957 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
958 log_err("%s: you need to specify size=\n", o->name);
959 td_verror(td, EINVAL, "total_file_size");
964 * Calculate per-file size and potential extra size for the
965 * first files, if needed (i.e. if we don't have a fixed size).
967 if (!o->file_size_low && o->nr_files) {
970 fs = o->size / o->nr_files;
971 all_fs = fs * o->nr_files;
973 if (all_fs < o->size)
974 nr_fs_extra = (o->size - all_fs) / bs;
978 * now file sizes are known, so we can set ->io_size. if size= is
979 * not given, ->io_size is just equal to ->real_file_size. if size
980 * is given, ->io_size is size / nr_files.
982 extend_size = total_size = 0;
984 for_each_file(td, f, i) {
985 f->file_offset = get_start_offset(td, f);
988 * Update ->io_size depending on options specified.
989 * ->file_size_low being 0 means filesize option isn't set.
990 * Non zero ->file_size_low equals ->file_size_high means
991 * filesize option is set in a fixed size format.
992 * Non zero ->file_size_low not equals ->file_size_high means
993 * filesize option is set in a range format.
995 if (!o->file_size_low) {
997 * no file size or range given, file size is equal to
998 * total size divided by number of files. If the size
999 * doesn't divide nicely with the min blocksize,
1000 * make the first files bigger.
1009 * We normally don't come here for regular files, but
1010 * if the result is 0 for a regular file, set it to the
1011 * real file size. This could be size of the existing
1012 * one if it already exists, but otherwise will be set
1013 * to 0. A new file won't be created because
1014 * ->io_size + ->file_offset equals ->real_file_size.
1017 if (f->file_offset > f->real_file_size)
1019 f->io_size = f->real_file_size - f->file_offset;
1021 log_info("fio: file %s may be ignored\n",
1024 } else if (f->real_file_size < o->file_size_low ||
1025 f->real_file_size > o->file_size_high) {
1026 if (f->file_offset > o->file_size_low)
1029 * file size given. if it's fixed, use that. if it's a
1030 * range, generate a random size in-between.
1032 if (o->file_size_low == o->file_size_high)
1033 f->io_size = o->file_size_low - f->file_offset;
1035 f->io_size = get_rand_file_size(td)
1039 f->io_size = f->real_file_size - f->file_offset;
1041 if (f->io_size == -1ULL)
1044 if (o->size_percent) {
1047 file_size = f->io_size + f->file_offset;
1048 f->io_size = (file_size *
1049 o->size_percent) / 100;
1050 if (f->io_size > (file_size - f->file_offset))
1051 f->io_size = file_size - f->file_offset;
1053 f->io_size -= (f->io_size % td_min_bs(td));
1055 total_size += f->io_size;
1058 if (f->filetype == FIO_TYPE_FILE &&
1059 (f->io_size + f->file_offset) > f->real_file_size &&
1060 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1061 if (!o->create_on_open) {
1063 extend_size += (f->io_size + f->file_offset);
1064 fio_file_set_extend(f);
1066 f->real_file_size = f->io_size + f->file_offset;
1070 if (td->o.block_error_hist) {
1073 assert(td->o.nr_files == 1); /* checked in fixup_options */
1075 len = f->io_size / td->o.bs[DDIR_TRIM];
1076 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1077 log_err("fio: cannot calculate block histogram with "
1078 "%d trim blocks, maximum %d\n",
1079 len, MAX_NR_BLOCK_INFOS);
1080 td_verror(td, EINVAL, "block_error_hist");
1084 td->ts.nr_block_infos = len;
1085 for (i = 0; i < len; i++)
1086 td->ts.block_infos[i] =
1087 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1089 td->ts.nr_block_infos = 0;
1091 if (!o->size || (total_size && o->size > total_size))
1092 o->size = total_size;
1094 if (o->size < td_min_bs(td)) {
1095 log_err("fio: blocksize too large for data set\n");
1100 * See if we need to extend some files, typically needed when our
1101 * target regular files don't exist yet, but our jobs require them
1102 * initially due to read I/Os.
1106 if (output_format & FIO_OUTPUT_NORMAL) {
1107 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1109 need_extend > 1 ? "s" : "",
1111 need_extend > 1 ? "s" : "",
1112 need_extend > 1 ? "total " : "",
1116 for_each_file(td, f, i) {
1117 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1119 if (!fio_file_extend(f))
1122 assert(f->filetype == FIO_TYPE_FILE);
1123 fio_file_clear_extend(f);
1124 if (!o->fill_device) {
1125 old_len = f->real_file_size;
1126 extend_len = f->io_size + f->file_offset -
1129 f->real_file_size = (f->io_size + f->file_offset);
1130 err = extend_file(td, f);
1134 err = __file_invalidate_cache(td, f, old_len,
1138 * Shut up static checker
1154 o->zone_size = o->size;
1157 * iolog already set the total io size, if we read back
1160 if (!o->read_iolog_file) {
1162 td->total_io_size = o->io_size * o->loops;
1164 td->total_io_size = o->size * o->loops;
1171 td_restore_runstate(td, old_state);
1174 log_err("%s: you need to specify valid offset=\n", o->name);
1176 td_restore_runstate(td, old_state);
1180 int pre_read_files(struct thread_data *td)
1185 dprint(FD_FILE, "pre_read files\n");
1187 for_each_file(td, f, i) {
1188 if (pre_read_file(td, f))
1195 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1197 unsigned int range_size, seed;
1198 unsigned long nranges;
1201 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1202 fsize = min(f->real_file_size, f->io_size);
1204 nranges = (fsize + range_size - 1) / range_size;
1206 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1207 if (!td->o.rand_repeatable)
1208 seed = td->rand_seeds[4];
1210 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1211 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1212 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1213 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1214 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1215 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1220 static int 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);
1240 * Check if the number of blocks exceeds the randomness capability of
1241 * the selected generator. Tausworthe is 32-bit, the others are fullly
1244 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1247 if (blocks <= FRAND32_MAX)
1249 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1253 * If the user hasn't specified a random generator, switch
1254 * to tausworthe64 with informational warning. If the user did
1255 * specify one, just warn.
1257 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1260 if (!fio_option_is_set(&td->o, random_generator)) {
1261 log_info("fio: Switching to tausworthe64. Use the "
1262 "random_generator= option to get rid of this "
1264 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1269 * Just make this information to avoid breaking scripts.
1271 log_info("fio: Use the random_generator= option to switch to lfsr or "
1276 int init_random_map(struct thread_data *td)
1278 unsigned long long blocks;
1282 if (init_rand_distribution(td))
1287 for_each_file(td, f, i) {
1288 uint64_t fsize = min(f->real_file_size, f->io_size);
1290 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1292 if (check_rand_gen_limits(td, f, blocks))
1295 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1298 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1300 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1301 fio_file_set_lfsr(f);
1304 } else if (!td->o.norandommap) {
1305 f->io_axmap = axmap_new(blocks);
1307 fio_file_set_axmap(f);
1310 } else if (td->o.norandommap)
1313 if (!td->o.softrandommap) {
1314 log_err("fio: failed allocating random map. If running"
1315 " a large number of jobs, try the 'norandommap'"
1316 " option or set 'softrandommap'. Or give"
1317 " a larger --alloc-size to fio.\n");
1321 log_info("fio: file %s failed allocating random map. Running "
1322 "job without.\n", f->file_name);
1328 void close_files(struct thread_data *td)
1333 for_each_file(td, f, i) {
1334 if (fio_file_open(f))
1335 td_io_close_file(td, f);
1339 void close_and_free_files(struct thread_data *td)
1343 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1345 dprint(FD_FILE, "close files\n");
1347 for_each_file(td, f, i) {
1348 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1349 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1350 td_io_unlink_file(td, f);
1353 if (fio_file_open(f))
1354 td_io_close_file(td, f);
1356 remove_file_hash(f);
1358 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1359 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1360 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);
1378 td->o.filename = NULL;
1380 free(td->file_locks);
1381 td->files_index = 0;
1383 td->file_locks = NULL;
1384 td->o.file_lock_mode = FILE_LOCK_NONE;
1388 static void get_file_type(struct fio_file *f)
1392 if (!strcmp(f->file_name, "-"))
1393 f->filetype = FIO_TYPE_PIPE;
1395 f->filetype = FIO_TYPE_FILE;
1398 /* \\.\ is the device namespace in Windows, where every file is
1400 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1401 f->filetype = FIO_TYPE_BLOCK;
1404 if (!stat(f->file_name, &sb)) {
1405 if (S_ISBLK(sb.st_mode))
1406 f->filetype = FIO_TYPE_BLOCK;
1407 else if (S_ISCHR(sb.st_mode))
1408 f->filetype = FIO_TYPE_CHAR;
1409 else if (S_ISFIFO(sb.st_mode))
1410 f->filetype = FIO_TYPE_PIPE;
1414 static bool __is_already_allocated(const char *fname, bool set)
1416 struct flist_head *entry;
1419 ret = file_bloom_exists(fname, set);
1423 flist_for_each(entry, &filename_list) {
1424 struct file_name *fn;
1426 fn = flist_entry(entry, struct file_name, list);
1428 if (!strcmp(fn->filename, fname))
1435 static bool is_already_allocated(const char *fname)
1439 fio_file_hash_lock();
1440 ret = __is_already_allocated(fname, false);
1441 fio_file_hash_unlock();
1446 static void set_already_allocated(const char *fname)
1448 struct file_name *fn;
1450 fn = malloc(sizeof(struct file_name));
1451 fn->filename = strdup(fname);
1453 fio_file_hash_lock();
1454 if (!__is_already_allocated(fname, true)) {
1455 flist_add_tail(&fn->list, &filename_list);
1458 fio_file_hash_unlock();
1466 static void free_already_allocated(void)
1468 struct flist_head *entry, *tmp;
1469 struct file_name *fn;
1471 if (flist_empty(&filename_list))
1474 fio_file_hash_lock();
1475 flist_for_each_safe(entry, tmp, &filename_list) {
1476 fn = flist_entry(entry, struct file_name, list);
1478 flist_del(&fn->list);
1482 fio_file_hash_unlock();
1485 static struct fio_file *alloc_new_file(struct thread_data *td)
1489 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1490 f = calloc(1, sizeof(*f));
1492 f = smalloc(sizeof(*f));
1500 fio_file_reset(td, f);
1504 bool exists_and_not_regfile(const char *filename)
1508 if (lstat(filename, &sb) == -1)
1511 #ifndef WIN32 /* NOT Windows */
1512 if (S_ISREG(sb.st_mode))
1515 /* \\.\ is the device namespace in Windows, where every file
1516 * is a device node */
1517 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1524 static int create_work_dirs(struct thread_data *td, const char *fname)
1526 char path[PATH_MAX];
1529 if (td->o.directory) {
1530 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1531 FIO_OS_PATH_SEPARATOR, fname);
1532 start = strstr(path, fname);
1534 snprintf(path, PATH_MAX, "%s", fname);
1539 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1544 #ifdef CONFIG_HAVE_MKDIR_TWO
1545 if (mkdir(path, 0600) && errno != EEXIST) {
1547 if (mkdir(path) && errno != EEXIST) {
1549 log_err("fio: failed to create dir (%s): %d\n",
1553 *end = FIO_OS_PATH_SEPARATOR;
1556 td->flags |= TD_F_DIRS_CREATED;
1560 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1562 int cur_files = td->files_index;
1563 char file_name[PATH_MAX];
1567 dprint(FD_FILE, "add file %s\n", fname);
1569 if (td->o.directory)
1570 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1571 td->o.unique_filename);
1573 sprintf(file_name + len, "%s", fname);
1575 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1576 !(td->flags & TD_F_DIRS_CREATED) &&
1577 create_work_dirs(td, fname))
1580 /* clean cloned siblings using existing files */
1581 if (numjob && is_already_allocated(file_name) &&
1582 !exists_and_not_regfile(fname))
1585 f = alloc_new_file(td);
1587 if (td->files_size <= td->files_index) {
1588 unsigned int new_size = td->o.nr_files + 1;
1590 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1592 td->files = realloc(td->files, new_size * sizeof(f));
1593 if (td->files == NULL) {
1594 log_err("fio: realloc OOM\n");
1597 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1598 td->file_locks = realloc(td->file_locks, new_size);
1599 if (!td->file_locks) {
1600 log_err("fio: realloc OOM\n");
1603 td->file_locks[cur_files] = FILE_LOCK_NONE;
1605 td->files_size = new_size;
1607 td->files[cur_files] = f;
1608 f->fileno = cur_files;
1611 * init function, io engine may not be loaded yet
1613 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1614 f->real_file_size = -1ULL;
1616 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1617 f->file_name = strdup(file_name);
1619 f->file_name = smalloc_strdup(file_name);
1625 switch (td->o.file_lock_mode) {
1626 case FILE_LOCK_NONE:
1628 case FILE_LOCK_READWRITE:
1629 f->rwlock = fio_rwlock_init();
1631 case FILE_LOCK_EXCLUSIVE:
1632 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1635 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1640 if (f->filetype == FIO_TYPE_FILE)
1641 td->nr_normal_files++;
1643 if (td->o.numjobs > 1)
1644 set_already_allocated(file_name);
1649 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1655 int add_file_exclusive(struct thread_data *td, const char *fname)
1660 for_each_file(td, f, i) {
1661 if (!strcmp(f->file_name, fname))
1665 return add_file(td, fname, 0, 1);
1668 void get_file(struct fio_file *f)
1670 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1671 assert(fio_file_open(f));
1675 int put_file(struct thread_data *td, struct fio_file *f)
1677 int f_ret = 0, ret = 0;
1679 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1681 if (!fio_file_open(f)) {
1682 assert(f->fd == -1);
1686 assert(f->references);
1687 if (--f->references)
1690 if (should_fsync(td) && td->o.fsync_on_close) {
1691 f_ret = fsync(f->fd);
1696 if (td->io_ops->close_file)
1697 ret = td->io_ops->close_file(td, f);
1702 td->nr_open_files--;
1703 fio_file_clear_open(f);
1704 assert(f->fd == -1);
1708 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1710 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1713 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1714 if (ddir == DDIR_READ)
1715 fio_rwlock_read(f->rwlock);
1717 fio_rwlock_write(f->rwlock);
1718 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1719 fio_mutex_down(f->lock);
1721 td->file_locks[f->fileno] = td->o.file_lock_mode;
1724 void unlock_file(struct thread_data *td, struct fio_file *f)
1726 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1729 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1730 fio_rwlock_unlock(f->rwlock);
1731 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1732 fio_mutex_up(f->lock);
1734 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1737 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1739 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1741 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1745 static int recurse_dir(struct thread_data *td, const char *dirname)
1751 D = opendir(dirname);
1753 char buf[FIO_VERROR_SIZE];
1755 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1756 td_verror(td, errno, buf);
1760 while ((dir = readdir(D)) != NULL) {
1761 char full_path[PATH_MAX];
1764 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1767 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1769 if (lstat(full_path, &sb) == -1) {
1770 if (errno != ENOENT) {
1771 td_verror(td, errno, "stat");
1777 if (S_ISREG(sb.st_mode)) {
1778 add_file(td, full_path, 0, 1);
1781 if (!S_ISDIR(sb.st_mode))
1784 ret = recurse_dir(td, full_path);
1793 int add_dir_files(struct thread_data *td, const char *path)
1795 int ret = recurse_dir(td, path);
1798 log_info("fio: opendir added %d files\n", td->o.nr_files);
1803 void dup_files(struct thread_data *td, struct thread_data *org)
1808 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1813 td->files = malloc(org->files_index * sizeof(f));
1815 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1816 td->file_locks = malloc(org->files_index);
1818 for_each_file(org, f, i) {
1819 struct fio_file *__f;
1821 __f = alloc_new_file(td);
1824 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1825 __f->file_name = strdup(f->file_name);
1827 __f->file_name = smalloc_strdup(f->file_name);
1828 if (!__f->file_name)
1831 __f->filetype = f->filetype;
1834 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1835 __f->lock = f->lock;
1836 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1837 __f->rwlock = f->rwlock;
1844 * Returns the index that matches the filename, or -1 if not there
1846 int get_fileno(struct thread_data *td, const char *fname)
1851 for_each_file(td, f, i)
1852 if (!strcmp(f->file_name, fname))
1859 * For log usage, where we add/open/close files automatically
1861 void free_release_files(struct thread_data *td)
1865 td->o.open_files = 0;
1866 td->files_index = 0;
1867 td->nr_normal_files = 0;
1870 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1874 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1875 f->last_pos[i] = f->file_offset;
1876 f->last_start[i] = -1ULL;
1879 if (fio_file_axmap(f))
1880 axmap_reset(f->io_axmap);
1881 else if (fio_file_lfsr(f))
1882 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1885 bool fio_files_done(struct thread_data *td)
1890 for_each_file(td, f, i)
1891 if (!fio_file_done(f))
1897 /* free memory used in initialization phase only */
1898 void filesetup_mem_free(void)
1900 free_already_allocated();
1904 * This function is for platforms which support direct I/O but not O_DIRECT.
1906 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1908 #ifdef FIO_OS_DIRECTIO
1909 int ret = fio_set_odirect(f);
1912 td_verror(td, ret, "fio_set_directio");
1913 #if defined(__sun__)
1914 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1915 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1917 log_err("fio: the file system does not seem to support direct IO\n");
1920 log_err("fio: the file system does not seem to support direct IO\n");
1927 log_err("fio: direct IO is not supported on this host operating system\n");