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);
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);
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 int pre_read_file(struct thread_data *td, struct fio_file *f)
264 int ret = 0, r, did_open = 0, old_runstate;
265 unsigned long long left;
269 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
270 td_ioengine_flagged(td, FIO_NOIO))
273 if (f->filetype == FIO_TYPE_CHAR)
276 if (!fio_file_open(f)) {
277 if (td->io_ops->open_file(td, f)) {
278 log_err("fio: cannot pre-read, failed to open file\n");
284 old_runstate = td_bump_runstate(td, TD_PRE_READING);
287 bs = td->o.max_bs[DDIR_READ];
293 td_verror(td, errno, "malloc");
299 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
300 td_verror(td, errno, "lseek");
301 log_err("fio: failed to lseek pre-read file\n");
306 while (left && !td->terminate) {
310 r = read(f->fd, b, bs);
316 td_verror(td, EIO, "pre_read");
322 td_restore_runstate(td, old_runstate);
325 td->io_ops->close_file(td, f);
331 unsigned long long get_rand_file_size(struct thread_data *td)
333 unsigned long long ret, sized;
337 frand_max = rand_max(&td->file_size_state);
338 r = __rand(&td->file_size_state);
339 sized = td->o.file_size_high - td->o.file_size_low;
340 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
341 ret += td->o.file_size_low;
342 ret -= (ret % td->o.rw_min_bs);
346 static int file_size(struct thread_data *td, struct fio_file *f)
350 if (stat(f->file_name, &st) == -1) {
351 td_verror(td, errno, "fstat");
355 f->real_file_size = st.st_size;
359 static int bdev_size(struct thread_data *td, struct fio_file *f)
361 unsigned long long bytes = 0;
364 if (td->io_ops->open_file(td, f)) {
365 log_err("fio: failed opening blockdev %s for size check\n",
370 r = blockdev_size(f, &bytes);
372 td_verror(td, r, "blockdev_size");
377 log_err("%s: zero sized block device?\n", f->file_name);
381 f->real_file_size = bytes;
382 td->io_ops->close_file(td, f);
385 td->io_ops->close_file(td, f);
389 static int char_size(struct thread_data *td, struct fio_file *f)
391 #ifdef FIO_HAVE_CHARDEV_SIZE
392 unsigned long long bytes = 0;
395 if (td->io_ops->open_file(td, f)) {
396 log_err("fio: failed opening chardev %s for size check\n",
401 r = chardev_size(f, &bytes);
403 td_verror(td, r, "chardev_size");
408 log_err("%s: zero sized char device?\n", f->file_name);
412 f->real_file_size = bytes;
413 td->io_ops->close_file(td, f);
416 td->io_ops->close_file(td, f);
419 f->real_file_size = -1ULL;
424 static int get_file_size(struct thread_data *td, struct fio_file *f)
428 if (fio_file_size_known(f))
431 if (f->filetype == FIO_TYPE_FILE)
432 ret = file_size(td, f);
433 else if (f->filetype == FIO_TYPE_BLOCK)
434 ret = bdev_size(td, f);
435 else if (f->filetype == FIO_TYPE_CHAR)
436 ret = char_size(td, f);
438 f->real_file_size = -1ULL;
441 * Leave ->real_file_size with 0 since it could be expectation
442 * of initial setup for regular files.
448 * If ->real_file_size is -1, a conditional for the message
449 * "offset extends end" is always true, but it makes no sense,
450 * so just return the same value here.
452 if (f->real_file_size == -1ULL) {
453 log_info("%s: failed to get file size of %s\n", td->o.name,
458 if (td->o.start_offset && f->file_offset == 0)
459 dprint(FD_FILE, "offset of file %s not initialized yet\n",
462 * ->file_offset normally hasn't been initialized yet, so this
463 * is basically always false.
465 if (f->file_offset > f->real_file_size) {
466 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
467 (unsigned long long) f->file_offset,
468 (unsigned long long) f->real_file_size);
472 fio_file_set_size_known(f);
476 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
477 unsigned long long off,
478 unsigned long long len)
480 int errval = 0, ret = 0;
489 off = f->file_offset;
491 if (len == -1ULL || off == -1ULL)
494 if (td->io_ops->invalidate) {
495 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
497 ret = td->io_ops->invalidate(td, f);
500 } else if (f->filetype == FIO_TYPE_FILE) {
501 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
502 f->file_name, off, len);
503 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
506 } else if (f->filetype == FIO_TYPE_BLOCK) {
509 dprint(FD_IO, "drop page cache %s\n", f->file_name);
510 ret = blockdev_invalidate_cache(f);
511 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
513 * Linux multipath devices reject ioctl while
514 * the maps are being updated. That window can
515 * last tens of milliseconds; we'll try up to
516 * a quarter of a second.
519 ret = blockdev_invalidate_cache(f);
521 if (ret < 0 && errno == EACCES && geteuid()) {
523 log_err("fio: only root may flush block "
524 "devices. Cache flush bypassed!\n");
531 else if (ret) /* probably not supported */
533 } else if (f->filetype == FIO_TYPE_CHAR ||
534 f->filetype == FIO_TYPE_PIPE) {
535 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
540 * Cache flushing isn't a fatal condition, and we know it will
541 * happen on some platforms where we don't have the proper
542 * function to flush eg block device caches. So just warn and
543 * continue on our way.
546 log_info("fio: cache invalidation of %s failed: %s\n",
547 f->file_name, strerror(errval));
553 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
555 if (!fio_file_open(f))
558 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
561 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
565 dprint(FD_FILE, "fd close %s\n", f->file_name);
569 if (close(f->fd) < 0)
574 if (f->shadow_fd != -1) {
583 int file_lookup_open(struct fio_file *f, int flags)
585 struct fio_file *__f;
588 __f = lookup_file_hash(f->file_name);
590 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
594 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
602 f->fd = open(f->file_name, flags, 0600);
606 static int file_close_shadow_fds(struct thread_data *td)
612 for_each_file(td, f, i) {
613 if (f->shadow_fd == -1)
624 int generic_open_file(struct thread_data *td, struct fio_file *f)
630 dprint(FD_FILE, "fd open %s\n", f->file_name);
632 if (!strcmp(f->file_name, "-")) {
634 log_err("fio: can't read/write to stdin/out\n");
640 * move output logging to stderr, if we are writing to stdout
649 flags |= OS_O_DIRECT;
652 td_verror(td, EINVAL, "OS does not support atomic IO");
655 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
659 if (td->o.create_on_open && td->o.allow_create)
662 if (f->filetype != FIO_TYPE_FILE)
663 flags |= FIO_O_NOATIME;
670 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
674 f->fd = dup(STDOUT_FILENO);
676 from_hash = file_lookup_open(f, flags);
677 } else if (td_read(td)) {
678 if (f->filetype == FIO_TYPE_CHAR && !read_only)
684 f->fd = dup(STDIN_FILENO);
686 from_hash = file_lookup_open(f, flags);
687 } else if (td_trim(td)) {
688 assert(!td_rw(td)); /* should have matched above */
690 from_hash = file_lookup_open(f, flags);
694 char buf[FIO_VERROR_SIZE];
697 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
698 flags &= ~FIO_O_NOATIME;
701 if (__e == EMFILE && file_close_shadow_fds(td))
704 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
706 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
707 log_err("fio: looks like your file system does not " \
708 "support direct=1/buffered=0\n");
711 td_verror(td, __e, buf);
715 if (!from_hash && f->fd != -1) {
716 if (add_file_hash(f)) {
720 * Stash away descriptor for later close. This is to
721 * work-around a "feature" on Linux, where a close of
722 * an fd that has been opened for write will trigger
723 * udev to call blkid to check partitions, fs id, etc.
724 * That pollutes the device cache, which can slow down
725 * unbuffered accesses.
727 if (f->shadow_fd == -1)
728 f->shadow_fd = f->fd;
731 * OK to ignore, we haven't done anything
734 ret = generic_close_file(td, f);
744 * This function i.e. get_file_size() is the default .get_file_size
745 * implementation of majority of I/O engines.
747 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
749 return get_file_size(td, f);
753 * open/close all files, so that ->real_file_size gets set
755 static int get_file_sizes(struct thread_data *td)
761 for_each_file(td, f, i) {
762 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
765 if (td_io_get_file_size(td, f)) {
766 if (td->error != ENOENT) {
767 log_err("%s\n", td->verror);
775 * There are corner cases where we end up with -1 for
776 * ->real_file_size due to unsupported file type, etc.
777 * We then just set to size option value divided by number
778 * of files, similar to the way file ->io_size is set.
779 * stat(2) failure doesn't set ->real_file_size to -1.
781 if (f->real_file_size == -1ULL && td->o.size)
782 f->real_file_size = td->o.size / td->o.nr_files;
789 struct flist_head list;
796 * Get free number of bytes for each file on each unique mount.
798 static unsigned long long get_fs_free_counts(struct thread_data *td)
800 struct flist_head *n, *tmp;
801 unsigned long long ret = 0;
802 struct fio_mount *fm;
807 for_each_file(td, f, i) {
811 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
812 if (f->real_file_size != -1ULL)
813 ret += f->real_file_size;
815 } else if (f->filetype != FIO_TYPE_FILE)
819 strncpy(buf, f->file_name, 255);
821 if (stat(buf, &sb) < 0) {
825 if (stat(buf, &sb) < 0)
830 flist_for_each(n, &list) {
831 fm = flist_entry(n, struct fio_mount, list);
832 if (fm->key == sb.st_dev)
841 fm = calloc(1, sizeof(*fm));
842 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
843 fm->base = basename(fm->__base);
845 flist_add(&fm->list, &list);
848 flist_for_each_safe(n, tmp, &list) {
849 unsigned long long sz;
851 fm = flist_entry(n, struct fio_mount, list);
852 flist_del(&fm->list);
854 sz = get_fs_free_size(fm->base);
855 if (sz && sz != -1ULL)
864 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
866 struct thread_options *o = &td->o;
867 unsigned long long align_bs;
868 unsigned long long offset;
870 if (o->file_append && f->filetype == FIO_TYPE_FILE)
871 return f->real_file_size;
873 if (o->start_offset_percent > 0) {
875 * if blockalign is provided, find the min across read, write,
878 if (fio_option_is_set(o, ba)) {
879 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
880 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
882 /* else take the minimum block size */
883 align_bs = td_min_bs(td);
886 /* calculate the raw offset */
887 offset = (f->real_file_size * o->start_offset_percent / 100) +
888 (td->subjob_number * o->offset_increment);
891 * block align the offset at the next available boundary at
892 * ceiling(offset / align_bs) * align_bs
894 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
897 /* start_offset_percent not set */
898 offset = o->start_offset +
899 td->subjob_number * o->offset_increment;
906 * Open the files and setup files sizes, creating files if necessary.
908 int setup_files(struct thread_data *td)
910 unsigned long long total_size, extend_size;
911 struct thread_options *o = &td->o;
913 unsigned int i, nr_fs_extra = 0;
914 int err = 0, need_extend;
916 const unsigned int bs = td_min_bs(td);
919 dprint(FD_FILE, "setup files\n");
921 old_state = td_bump_runstate(td, TD_SETTING_UP);
923 if (o->read_iolog_file)
927 * Find out physical size of files or devices for this thread,
928 * before we determine I/O size and range of our targets.
929 * If ioengine defines a setup() method, it's responsible for
930 * opening the files and setting f->real_file_size to indicate
931 * the valid range for that file.
933 if (td->io_ops->setup)
934 err = td->io_ops->setup(td);
936 err = get_file_sizes(td);
942 * check sizes. if the files/devices do not exist and the size
943 * isn't passed to fio, abort.
946 for_each_file(td, f, i) {
948 if (f->real_file_size == -1ULL)
951 total_size += f->real_file_size;
955 td->fill_device_size = get_fs_free_counts(td);
958 * device/file sizes are zero and no size given, punt
960 if ((!total_size || total_size == -1ULL) && !o->size &&
961 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
962 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
963 log_err("%s: you need to specify size=\n", o->name);
964 td_verror(td, EINVAL, "total_file_size");
969 * Calculate per-file size and potential extra size for the
970 * first files, if needed (i.e. if we don't have a fixed size).
972 if (!o->file_size_low && o->nr_files) {
975 fs = o->size / o->nr_files;
976 all_fs = fs * o->nr_files;
978 if (all_fs < o->size)
979 nr_fs_extra = (o->size - all_fs) / bs;
983 * now file sizes are known, so we can set ->io_size. if size= is
984 * not given, ->io_size is just equal to ->real_file_size. if size
985 * is given, ->io_size is size / nr_files.
987 extend_size = total_size = 0;
989 for_each_file(td, f, i) {
990 f->file_offset = get_start_offset(td, f);
993 * Update ->io_size depending on options specified.
994 * ->file_size_low being 0 means filesize option isn't set.
995 * Non zero ->file_size_low equals ->file_size_high means
996 * filesize option is set in a fixed size format.
997 * Non zero ->file_size_low not equals ->file_size_high means
998 * filesize option is set in a range format.
1000 if (!o->file_size_low) {
1002 * no file size or range given, file size is equal to
1003 * total size divided by number of files. If the size
1004 * doesn't divide nicely with the min blocksize,
1005 * make the first files bigger.
1014 * We normally don't come here for regular files, but
1015 * if the result is 0 for a regular file, set it to the
1016 * real file size. This could be size of the existing
1017 * one if it already exists, but otherwise will be set
1018 * to 0. A new file won't be created because
1019 * ->io_size + ->file_offset equals ->real_file_size.
1022 if (f->file_offset > f->real_file_size)
1024 f->io_size = f->real_file_size - f->file_offset;
1026 log_info("fio: file %s may be ignored\n",
1029 } else if (f->real_file_size < o->file_size_low ||
1030 f->real_file_size > o->file_size_high) {
1031 if (f->file_offset > o->file_size_low)
1034 * file size given. if it's fixed, use that. if it's a
1035 * range, generate a random size in-between.
1037 if (o->file_size_low == o->file_size_high)
1038 f->io_size = o->file_size_low - f->file_offset;
1040 f->io_size = get_rand_file_size(td)
1044 f->io_size = f->real_file_size - f->file_offset;
1046 if (f->io_size == -1ULL)
1049 if (o->size_percent) {
1052 file_size = f->io_size + f->file_offset;
1053 f->io_size = (file_size *
1054 o->size_percent) / 100;
1055 if (f->io_size > (file_size - f->file_offset))
1056 f->io_size = file_size - f->file_offset;
1058 f->io_size -= (f->io_size % td_min_bs(td));
1060 total_size += f->io_size;
1063 if (f->filetype == FIO_TYPE_FILE &&
1064 (f->io_size + f->file_offset) > f->real_file_size &&
1065 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1066 if (!o->create_on_open) {
1068 extend_size += (f->io_size + f->file_offset);
1069 fio_file_set_extend(f);
1071 f->real_file_size = f->io_size + f->file_offset;
1075 if (td->o.block_error_hist) {
1078 assert(td->o.nr_files == 1); /* checked in fixup_options */
1080 len = f->io_size / td->o.bs[DDIR_TRIM];
1081 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1082 log_err("fio: cannot calculate block histogram with "
1083 "%d trim blocks, maximum %d\n",
1084 len, MAX_NR_BLOCK_INFOS);
1085 td_verror(td, EINVAL, "block_error_hist");
1089 td->ts.nr_block_infos = len;
1090 for (i = 0; i < len; i++)
1091 td->ts.block_infos[i] =
1092 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1094 td->ts.nr_block_infos = 0;
1096 if (!o->size || (total_size && o->size > total_size))
1097 o->size = total_size;
1099 if (o->size < td_min_bs(td)) {
1100 log_err("fio: blocksize too large for data set\n");
1105 * See if we need to extend some files, typically needed when our
1106 * target regular files don't exist yet, but our jobs require them
1107 * initially due to read I/Os.
1111 if (output_format & FIO_OUTPUT_NORMAL) {
1112 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1114 need_extend > 1 ? "s" : "",
1116 need_extend > 1 ? "s" : "",
1117 need_extend > 1 ? "total " : "",
1121 for_each_file(td, f, i) {
1122 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1124 if (!fio_file_extend(f))
1127 assert(f->filetype == FIO_TYPE_FILE);
1128 fio_file_clear_extend(f);
1129 if (!o->fill_device) {
1130 old_len = f->real_file_size;
1131 extend_len = f->io_size + f->file_offset -
1134 f->real_file_size = (f->io_size + f->file_offset);
1135 err = extend_file(td, f);
1139 err = __file_invalidate_cache(td, f, old_len,
1143 * Shut up static checker
1159 o->zone_size = o->size;
1162 * iolog already set the total io size, if we read back
1165 if (!o->read_iolog_file) {
1167 td->total_io_size = o->io_size * o->loops;
1169 td->total_io_size = o->size * o->loops;
1176 td_restore_runstate(td, old_state);
1179 log_err("%s: you need to specify valid offset=\n", o->name);
1181 td_restore_runstate(td, old_state);
1185 int pre_read_files(struct thread_data *td)
1190 dprint(FD_FILE, "pre_read files\n");
1192 for_each_file(td, f, i) {
1193 if (pre_read_file(td, f))
1200 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1202 unsigned int range_size, seed;
1203 unsigned long nranges;
1206 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1207 fsize = min(f->real_file_size, f->io_size);
1209 nranges = (fsize + range_size - 1) / range_size;
1211 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1212 if (!td->o.rand_repeatable)
1213 seed = td->rand_seeds[4];
1215 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1216 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1217 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1218 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1219 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1220 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1225 static int init_rand_distribution(struct thread_data *td)
1231 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1234 state = td_bump_runstate(td, TD_SETTING_UP);
1236 for_each_file(td, f, i)
1237 __init_rand_distribution(td, f);
1239 td_restore_runstate(td, state);
1245 * Check if the number of blocks exceeds the randomness capability of
1246 * the selected generator. Tausworthe is 32-bit, the others are fullly
1249 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1252 if (blocks <= FRAND32_MAX)
1254 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1258 * If the user hasn't specified a random generator, switch
1259 * to tausworthe64 with informational warning. If the user did
1260 * specify one, just warn.
1262 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1265 if (!fio_option_is_set(&td->o, random_generator)) {
1266 log_info("fio: Switching to tausworthe64. Use the "
1267 "random_generator= option to get rid of this "
1269 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1274 * Just make this information to avoid breaking scripts.
1276 log_info("fio: Use the random_generator= option to switch to lfsr or "
1281 int init_random_map(struct thread_data *td)
1283 unsigned long long blocks;
1287 if (init_rand_distribution(td))
1292 for_each_file(td, f, i) {
1293 uint64_t fsize = min(f->real_file_size, f->io_size);
1295 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1297 if (check_rand_gen_limits(td, f, blocks))
1300 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1303 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1305 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1306 fio_file_set_lfsr(f);
1309 } else if (!td->o.norandommap) {
1310 f->io_axmap = axmap_new(blocks);
1312 fio_file_set_axmap(f);
1315 } else if (td->o.norandommap)
1318 if (!td->o.softrandommap) {
1319 log_err("fio: failed allocating random map. If running"
1320 " a large number of jobs, try the 'norandommap'"
1321 " option or set 'softrandommap'. Or give"
1322 " a larger --alloc-size to fio.\n");
1326 log_info("fio: file %s failed allocating random map. Running "
1327 "job without.\n", f->file_name);
1333 void close_files(struct thread_data *td)
1338 for_each_file(td, f, i) {
1339 if (fio_file_open(f))
1340 td_io_close_file(td, f);
1344 void close_and_free_files(struct thread_data *td)
1349 dprint(FD_FILE, "close files\n");
1351 for_each_file(td, f, i) {
1352 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1353 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1354 td_io_unlink_file(td, f);
1357 if (fio_file_open(f))
1358 td_io_close_file(td, f);
1360 remove_file_hash(f);
1362 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1363 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1364 td_io_unlink_file(td, f);
1367 sfree(f->file_name);
1368 f->file_name = NULL;
1369 if (fio_file_axmap(f)) {
1370 axmap_free(f->io_axmap);
1376 td->o.filename = NULL;
1378 free(td->file_locks);
1379 td->files_index = 0;
1381 td->file_locks = NULL;
1382 td->o.file_lock_mode = FILE_LOCK_NONE;
1386 static void get_file_type(struct fio_file *f)
1390 if (!strcmp(f->file_name, "-"))
1391 f->filetype = FIO_TYPE_PIPE;
1393 f->filetype = FIO_TYPE_FILE;
1396 /* \\.\ is the device namespace in Windows, where every file is
1398 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1399 f->filetype = FIO_TYPE_BLOCK;
1402 if (!stat(f->file_name, &sb)) {
1403 if (S_ISBLK(sb.st_mode))
1404 f->filetype = FIO_TYPE_BLOCK;
1405 else if (S_ISCHR(sb.st_mode))
1406 f->filetype = FIO_TYPE_CHAR;
1407 else if (S_ISFIFO(sb.st_mode))
1408 f->filetype = FIO_TYPE_PIPE;
1412 static bool __is_already_allocated(const char *fname, bool set)
1414 struct flist_head *entry;
1417 ret = file_bloom_exists(fname, set);
1421 flist_for_each(entry, &filename_list) {
1422 struct file_name *fn;
1424 fn = flist_entry(entry, struct file_name, list);
1426 if (!strcmp(fn->filename, fname))
1433 static bool is_already_allocated(const char *fname)
1437 fio_file_hash_lock();
1438 ret = __is_already_allocated(fname, false);
1439 fio_file_hash_unlock();
1444 static void set_already_allocated(const char *fname)
1446 struct file_name *fn;
1448 fn = malloc(sizeof(struct file_name));
1449 fn->filename = strdup(fname);
1451 fio_file_hash_lock();
1452 if (!__is_already_allocated(fname, true)) {
1453 flist_add_tail(&fn->list, &filename_list);
1456 fio_file_hash_unlock();
1464 static void free_already_allocated(void)
1466 struct flist_head *entry, *tmp;
1467 struct file_name *fn;
1469 if (flist_empty(&filename_list))
1472 fio_file_hash_lock();
1473 flist_for_each_safe(entry, tmp, &filename_list) {
1474 fn = flist_entry(entry, struct file_name, list);
1476 flist_del(&fn->list);
1480 fio_file_hash_unlock();
1483 static struct fio_file *alloc_new_file(struct thread_data *td)
1487 f = smalloc(sizeof(*f));
1495 fio_file_reset(td, f);
1499 bool exists_and_not_regfile(const char *filename)
1503 if (lstat(filename, &sb) == -1)
1506 #ifndef WIN32 /* NOT Windows */
1507 if (S_ISREG(sb.st_mode))
1510 /* \\.\ is the device namespace in Windows, where every file
1511 * is a device node */
1512 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1519 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1521 int cur_files = td->files_index;
1522 char file_name[PATH_MAX];
1526 dprint(FD_FILE, "add file %s\n", fname);
1528 if (td->o.directory)
1529 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1530 td->o.unique_filename);
1532 sprintf(file_name + len, "%s", fname);
1534 /* clean cloned siblings using existing files */
1535 if (numjob && is_already_allocated(file_name) &&
1536 !exists_and_not_regfile(fname))
1539 f = alloc_new_file(td);
1541 if (td->files_size <= td->files_index) {
1542 unsigned int new_size = td->o.nr_files + 1;
1544 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1546 td->files = realloc(td->files, new_size * sizeof(f));
1547 if (td->files == NULL) {
1548 log_err("fio: realloc OOM\n");
1551 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1552 td->file_locks = realloc(td->file_locks, new_size);
1553 if (!td->file_locks) {
1554 log_err("fio: realloc OOM\n");
1557 td->file_locks[cur_files] = FILE_LOCK_NONE;
1559 td->files_size = new_size;
1561 td->files[cur_files] = f;
1562 f->fileno = cur_files;
1565 * init function, io engine may not be loaded yet
1567 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1568 f->real_file_size = -1ULL;
1570 f->file_name = smalloc_strdup(file_name);
1576 switch (td->o.file_lock_mode) {
1577 case FILE_LOCK_NONE:
1579 case FILE_LOCK_READWRITE:
1580 f->rwlock = fio_rwlock_init();
1582 case FILE_LOCK_EXCLUSIVE:
1583 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1586 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1591 if (f->filetype == FIO_TYPE_FILE)
1592 td->nr_normal_files++;
1594 set_already_allocated(file_name);
1599 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1605 int add_file_exclusive(struct thread_data *td, const char *fname)
1610 for_each_file(td, f, i) {
1611 if (!strcmp(f->file_name, fname))
1615 return add_file(td, fname, 0, 1);
1618 void get_file(struct fio_file *f)
1620 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1621 assert(fio_file_open(f));
1625 int put_file(struct thread_data *td, struct fio_file *f)
1627 int f_ret = 0, ret = 0;
1629 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1631 if (!fio_file_open(f)) {
1632 assert(f->fd == -1);
1636 assert(f->references);
1637 if (--f->references)
1640 if (should_fsync(td) && td->o.fsync_on_close) {
1641 f_ret = fsync(f->fd);
1646 if (td->io_ops->close_file)
1647 ret = td->io_ops->close_file(td, f);
1652 td->nr_open_files--;
1653 fio_file_clear_open(f);
1654 assert(f->fd == -1);
1658 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1660 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1663 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1664 if (ddir == DDIR_READ)
1665 fio_rwlock_read(f->rwlock);
1667 fio_rwlock_write(f->rwlock);
1668 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1669 fio_mutex_down(f->lock);
1671 td->file_locks[f->fileno] = td->o.file_lock_mode;
1674 void unlock_file(struct thread_data *td, struct fio_file *f)
1676 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1679 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1680 fio_rwlock_unlock(f->rwlock);
1681 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1682 fio_mutex_up(f->lock);
1684 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1687 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1689 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1691 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1695 static int recurse_dir(struct thread_data *td, const char *dirname)
1701 D = opendir(dirname);
1703 char buf[FIO_VERROR_SIZE];
1705 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1706 td_verror(td, errno, buf);
1710 while ((dir = readdir(D)) != NULL) {
1711 char full_path[PATH_MAX];
1714 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1717 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1719 if (lstat(full_path, &sb) == -1) {
1720 if (errno != ENOENT) {
1721 td_verror(td, errno, "stat");
1727 if (S_ISREG(sb.st_mode)) {
1728 add_file(td, full_path, 0, 1);
1731 if (!S_ISDIR(sb.st_mode))
1734 ret = recurse_dir(td, full_path);
1743 int add_dir_files(struct thread_data *td, const char *path)
1745 int ret = recurse_dir(td, path);
1748 log_info("fio: opendir added %d files\n", td->o.nr_files);
1753 void dup_files(struct thread_data *td, struct thread_data *org)
1758 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1763 td->files = malloc(org->files_index * sizeof(f));
1765 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1766 td->file_locks = malloc(org->files_index);
1768 for_each_file(org, f, i) {
1769 struct fio_file *__f;
1771 __f = alloc_new_file(td);
1774 __f->file_name = smalloc_strdup(f->file_name);
1775 if (!__f->file_name)
1778 __f->filetype = f->filetype;
1781 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1782 __f->lock = f->lock;
1783 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1784 __f->rwlock = f->rwlock;
1791 * Returns the index that matches the filename, or -1 if not there
1793 int get_fileno(struct thread_data *td, const char *fname)
1798 for_each_file(td, f, i)
1799 if (!strcmp(f->file_name, fname))
1806 * For log usage, where we add/open/close files automatically
1808 void free_release_files(struct thread_data *td)
1812 td->o.open_files = 0;
1813 td->files_index = 0;
1814 td->nr_normal_files = 0;
1817 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1821 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1822 f->last_pos[i] = f->file_offset;
1823 f->last_start[i] = -1ULL;
1826 if (fio_file_axmap(f))
1827 axmap_reset(f->io_axmap);
1828 else if (fio_file_lfsr(f))
1829 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1832 bool fio_files_done(struct thread_data *td)
1837 for_each_file(td, f, i)
1838 if (!fio_file_done(f))
1844 /* free memory used in initialization phase only */
1845 void filesetup_mem_free(void)
1847 free_already_allocated();