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 (f->filetype == FIO_TYPE_CHAR ||
532 f->filetype == FIO_TYPE_PIPE) {
533 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
538 * Cache flushing isn't a fatal condition, and we know it will
539 * happen on some platforms where we don't have the proper
540 * function to flush eg block device caches. So just warn and
541 * continue on our way.
544 log_info("fio: cache invalidation of %s failed: %s\n",
545 f->file_name, strerror(errval));
551 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
553 if (!fio_file_open(f))
556 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
559 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
563 dprint(FD_FILE, "fd close %s\n", f->file_name);
567 if (close(f->fd) < 0)
572 if (f->shadow_fd != -1) {
581 int file_lookup_open(struct fio_file *f, int flags)
583 struct fio_file *__f;
586 __f = lookup_file_hash(f->file_name);
588 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
592 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
600 f->fd = open(f->file_name, flags, 0600);
604 static int file_close_shadow_fds(struct thread_data *td)
610 for_each_file(td, f, i) {
611 if (f->shadow_fd == -1)
622 int generic_open_file(struct thread_data *td, struct fio_file *f)
628 dprint(FD_FILE, "fd open %s\n", f->file_name);
630 if (!strcmp(f->file_name, "-")) {
632 log_err("fio: can't read/write to stdin/out\n");
638 * move output logging to stderr, if we are writing to stdout
647 flags |= OS_O_DIRECT;
650 td_verror(td, EINVAL, "OS does not support atomic IO");
653 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
657 if (td->o.create_on_open && td->o.allow_create)
660 if (f->filetype != FIO_TYPE_FILE)
661 flags |= FIO_O_NOATIME;
668 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
672 f->fd = dup(STDOUT_FILENO);
674 from_hash = file_lookup_open(f, flags);
675 } else if (td_read(td)) {
676 if (f->filetype == FIO_TYPE_CHAR && !read_only)
682 f->fd = dup(STDIN_FILENO);
684 from_hash = file_lookup_open(f, flags);
685 } else if (td_trim(td)) {
686 assert(!td_rw(td)); /* should have matched above */
688 from_hash = file_lookup_open(f, flags);
692 char buf[FIO_VERROR_SIZE];
695 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
696 flags &= ~FIO_O_NOATIME;
699 if (__e == EMFILE && file_close_shadow_fds(td))
702 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
704 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
705 log_err("fio: looks like your file system does not " \
706 "support direct=1/buffered=0\n");
709 td_verror(td, __e, buf);
713 if (!from_hash && f->fd != -1) {
714 if (add_file_hash(f)) {
718 * Stash away descriptor for later close. This is to
719 * work-around a "feature" on Linux, where a close of
720 * an fd that has been opened for write will trigger
721 * udev to call blkid to check partitions, fs id, etc.
722 * That pollutes the device cache, which can slow down
723 * unbuffered accesses.
725 if (f->shadow_fd == -1)
726 f->shadow_fd = f->fd;
729 * OK to ignore, we haven't done anything
732 ret = generic_close_file(td, f);
742 * This function i.e. get_file_size() is the default .get_file_size
743 * implementation of majority of I/O engines.
745 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
747 return get_file_size(td, f);
751 * open/close all files, so that ->real_file_size gets set
753 static int get_file_sizes(struct thread_data *td)
759 for_each_file(td, f, i) {
760 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
763 if (td_io_get_file_size(td, f)) {
764 if (td->error != ENOENT) {
765 log_err("%s\n", td->verror);
773 * There are corner cases where we end up with -1 for
774 * ->real_file_size due to unsupported file type, etc.
775 * We then just set to size option value divided by number
776 * of files, similar to the way file ->io_size is set.
777 * stat(2) failure doesn't set ->real_file_size to -1.
779 if (f->real_file_size == -1ULL && td->o.size)
780 f->real_file_size = td->o.size / td->o.nr_files;
787 struct flist_head list;
794 * Get free number of bytes for each file on each unique mount.
796 static unsigned long long get_fs_free_counts(struct thread_data *td)
798 struct flist_head *n, *tmp;
799 unsigned long long ret = 0;
800 struct fio_mount *fm;
805 for_each_file(td, f, i) {
809 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
810 if (f->real_file_size != -1ULL)
811 ret += f->real_file_size;
813 } else if (f->filetype != FIO_TYPE_FILE)
817 strncpy(buf, f->file_name, 255);
819 if (stat(buf, &sb) < 0) {
823 if (stat(buf, &sb) < 0)
828 flist_for_each(n, &list) {
829 fm = flist_entry(n, struct fio_mount, list);
830 if (fm->key == sb.st_dev)
839 fm = calloc(1, sizeof(*fm));
840 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
841 fm->base = basename(fm->__base);
843 flist_add(&fm->list, &list);
846 flist_for_each_safe(n, tmp, &list) {
847 unsigned long long sz;
849 fm = flist_entry(n, struct fio_mount, list);
850 flist_del(&fm->list);
852 sz = get_fs_free_size(fm->base);
853 if (sz && sz != -1ULL)
862 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
864 struct thread_options *o = &td->o;
865 unsigned long long align_bs;
866 unsigned long long offset;
868 if (o->file_append && f->filetype == FIO_TYPE_FILE)
869 return f->real_file_size;
871 if (o->start_offset_percent > 0) {
873 * if blockalign is provided, find the min across read, write,
876 if (fio_option_is_set(o, ba)) {
877 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
878 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
880 /* else take the minimum block size */
881 align_bs = td_min_bs(td);
884 /* calculate the raw offset */
885 offset = (f->real_file_size * o->start_offset_percent / 100) +
886 (td->subjob_number * o->offset_increment);
889 * block align the offset at the next available boundary at
890 * ceiling(offset / align_bs) * align_bs
892 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
895 /* start_offset_percent not set */
896 offset = o->start_offset +
897 td->subjob_number * o->offset_increment;
904 * Open the files and setup files sizes, creating files if necessary.
906 int setup_files(struct thread_data *td)
908 unsigned long long total_size, extend_size;
909 struct thread_options *o = &td->o;
911 unsigned int i, nr_fs_extra = 0;
912 int err = 0, need_extend;
914 const unsigned int bs = td_min_bs(td);
917 dprint(FD_FILE, "setup files\n");
919 old_state = td_bump_runstate(td, TD_SETTING_UP);
921 if (o->read_iolog_file)
925 * Find out physical size of files or devices for this thread,
926 * before we determine I/O size and range of our targets.
927 * If ioengine defines a setup() method, it's responsible for
928 * opening the files and setting f->real_file_size to indicate
929 * the valid range for that file.
931 if (td->io_ops->setup)
932 err = td->io_ops->setup(td);
934 err = get_file_sizes(td);
940 * check sizes. if the files/devices do not exist and the size
941 * isn't passed to fio, abort.
944 for_each_file(td, f, i) {
946 if (f->real_file_size == -1ULL)
949 total_size += f->real_file_size;
953 td->fill_device_size = get_fs_free_counts(td);
956 * device/file sizes are zero and no size given, punt
958 if ((!total_size || total_size == -1ULL) && !o->size &&
959 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
960 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
961 log_err("%s: you need to specify size=\n", o->name);
962 td_verror(td, EINVAL, "total_file_size");
967 * Calculate per-file size and potential extra size for the
968 * first files, if needed (i.e. if we don't have a fixed size).
970 if (!o->file_size_low && o->nr_files) {
973 fs = o->size / o->nr_files;
974 all_fs = fs * o->nr_files;
976 if (all_fs < o->size)
977 nr_fs_extra = (o->size - all_fs) / bs;
981 * now file sizes are known, so we can set ->io_size. if size= is
982 * not given, ->io_size is just equal to ->real_file_size. if size
983 * is given, ->io_size is size / nr_files.
985 extend_size = total_size = 0;
987 for_each_file(td, f, i) {
988 f->file_offset = get_start_offset(td, f);
991 * Update ->io_size depending on options specified.
992 * ->file_size_low being 0 means filesize option isn't set.
993 * Non zero ->file_size_low equals ->file_size_high means
994 * filesize option is set in a fixed size format.
995 * Non zero ->file_size_low not equals ->file_size_high means
996 * filesize option is set in a range format.
998 if (!o->file_size_low) {
1000 * no file size or range given, file size is equal to
1001 * total size divided by number of files. If the size
1002 * doesn't divide nicely with the min blocksize,
1003 * make the first files bigger.
1012 * We normally don't come here for regular files, but
1013 * if the result is 0 for a regular file, set it to the
1014 * real file size. This could be size of the existing
1015 * one if it already exists, but otherwise will be set
1016 * to 0. A new file won't be created because
1017 * ->io_size + ->file_offset equals ->real_file_size.
1020 if (f->file_offset > f->real_file_size)
1022 f->io_size = f->real_file_size - f->file_offset;
1024 log_info("fio: file %s may be ignored\n",
1027 } else if (f->real_file_size < o->file_size_low ||
1028 f->real_file_size > o->file_size_high) {
1029 if (f->file_offset > o->file_size_low)
1032 * file size given. if it's fixed, use that. if it's a
1033 * range, generate a random size in-between.
1035 if (o->file_size_low == o->file_size_high)
1036 f->io_size = o->file_size_low - f->file_offset;
1038 f->io_size = get_rand_file_size(td)
1042 f->io_size = f->real_file_size - f->file_offset;
1044 if (f->io_size == -1ULL)
1047 if (o->size_percent) {
1050 file_size = f->io_size + f->file_offset;
1051 f->io_size = (file_size *
1052 o->size_percent) / 100;
1053 if (f->io_size > (file_size - f->file_offset))
1054 f->io_size = file_size - f->file_offset;
1056 f->io_size -= (f->io_size % td_min_bs(td));
1058 total_size += f->io_size;
1061 if (f->filetype == FIO_TYPE_FILE &&
1062 (f->io_size + f->file_offset) > f->real_file_size &&
1063 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1064 if (!o->create_on_open) {
1066 extend_size += (f->io_size + f->file_offset);
1067 fio_file_set_extend(f);
1069 f->real_file_size = f->io_size + f->file_offset;
1073 if (td->o.block_error_hist) {
1076 assert(td->o.nr_files == 1); /* checked in fixup_options */
1078 len = f->io_size / td->o.bs[DDIR_TRIM];
1079 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1080 log_err("fio: cannot calculate block histogram with "
1081 "%d trim blocks, maximum %d\n",
1082 len, MAX_NR_BLOCK_INFOS);
1083 td_verror(td, EINVAL, "block_error_hist");
1087 td->ts.nr_block_infos = len;
1088 for (i = 0; i < len; i++)
1089 td->ts.block_infos[i] =
1090 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1092 td->ts.nr_block_infos = 0;
1094 if (!o->size || (total_size && o->size > total_size))
1095 o->size = total_size;
1097 if (o->size < td_min_bs(td)) {
1098 log_err("fio: blocksize too large for data set\n");
1103 * See if we need to extend some files, typically needed when our
1104 * target regular files don't exist yet, but our jobs require them
1105 * initially due to read I/Os.
1109 if (output_format & FIO_OUTPUT_NORMAL) {
1110 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1112 need_extend > 1 ? "s" : "",
1114 need_extend > 1 ? "s" : "",
1115 need_extend > 1 ? "total " : "",
1119 for_each_file(td, f, i) {
1120 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1122 if (!fio_file_extend(f))
1125 assert(f->filetype == FIO_TYPE_FILE);
1126 fio_file_clear_extend(f);
1127 if (!o->fill_device) {
1128 old_len = f->real_file_size;
1129 extend_len = f->io_size + f->file_offset -
1132 f->real_file_size = (f->io_size + f->file_offset);
1133 err = extend_file(td, f);
1137 err = __file_invalidate_cache(td, f, old_len,
1141 * Shut up static checker
1157 o->zone_size = o->size;
1160 * iolog already set the total io size, if we read back
1163 if (!o->read_iolog_file) {
1165 td->total_io_size = o->io_size * o->loops;
1167 td->total_io_size = o->size * o->loops;
1174 td_restore_runstate(td, old_state);
1177 log_err("%s: you need to specify valid offset=\n", o->name);
1179 td_restore_runstate(td, old_state);
1183 int pre_read_files(struct thread_data *td)
1188 dprint(FD_FILE, "pre_read files\n");
1190 for_each_file(td, f, i) {
1191 if (pre_read_file(td, f))
1198 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1200 unsigned int range_size, seed;
1201 unsigned long nranges;
1204 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1205 fsize = min(f->real_file_size, f->io_size);
1207 nranges = (fsize + range_size - 1) / range_size;
1209 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1210 if (!td->o.rand_repeatable)
1211 seed = td->rand_seeds[4];
1213 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1214 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1215 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1216 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1217 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1218 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1223 static int init_rand_distribution(struct thread_data *td)
1229 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1232 state = td_bump_runstate(td, TD_SETTING_UP);
1234 for_each_file(td, f, i)
1235 __init_rand_distribution(td, f);
1237 td_restore_runstate(td, state);
1243 * Check if the number of blocks exceeds the randomness capability of
1244 * the selected generator. Tausworthe is 32-bit, the others are fullly
1247 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1250 if (blocks <= FRAND32_MAX)
1252 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1256 * If the user hasn't specified a random generator, switch
1257 * to tausworthe64 with informational warning. If the user did
1258 * specify one, just warn.
1260 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1263 if (!fio_option_is_set(&td->o, random_generator)) {
1264 log_info("fio: Switching to tausworthe64. Use the "
1265 "random_generator= option to get rid of this "
1267 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1272 * Just make this information to avoid breaking scripts.
1274 log_info("fio: Use the random_generator= option to switch to lfsr or "
1279 int init_random_map(struct thread_data *td)
1281 unsigned long long blocks;
1285 if (init_rand_distribution(td))
1290 for_each_file(td, f, i) {
1291 uint64_t fsize = min(f->real_file_size, f->io_size);
1293 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1295 if (check_rand_gen_limits(td, f, blocks))
1298 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1301 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1303 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1304 fio_file_set_lfsr(f);
1307 } else if (!td->o.norandommap) {
1308 f->io_axmap = axmap_new(blocks);
1310 fio_file_set_axmap(f);
1313 } else if (td->o.norandommap)
1316 if (!td->o.softrandommap) {
1317 log_err("fio: failed allocating random map. If running"
1318 " a large number of jobs, try the 'norandommap'"
1319 " option or set 'softrandommap'. Or give"
1320 " a larger --alloc-size to fio.\n");
1324 log_info("fio: file %s failed allocating random map. Running "
1325 "job without.\n", f->file_name);
1331 void close_files(struct thread_data *td)
1336 for_each_file(td, f, i) {
1337 if (fio_file_open(f))
1338 td_io_close_file(td, f);
1342 void close_and_free_files(struct thread_data *td)
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);
1365 sfree(f->file_name);
1366 f->file_name = NULL;
1367 if (fio_file_axmap(f)) {
1368 axmap_free(f->io_axmap);
1374 td->o.filename = NULL;
1376 free(td->file_locks);
1377 td->files_index = 0;
1379 td->file_locks = NULL;
1380 td->o.file_lock_mode = FILE_LOCK_NONE;
1384 static void get_file_type(struct fio_file *f)
1388 if (!strcmp(f->file_name, "-"))
1389 f->filetype = FIO_TYPE_PIPE;
1391 f->filetype = FIO_TYPE_FILE;
1394 /* \\.\ is the device namespace in Windows, where every file is
1396 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1397 f->filetype = FIO_TYPE_BLOCK;
1400 if (!stat(f->file_name, &sb)) {
1401 if (S_ISBLK(sb.st_mode))
1402 f->filetype = FIO_TYPE_BLOCK;
1403 else if (S_ISCHR(sb.st_mode))
1404 f->filetype = FIO_TYPE_CHAR;
1405 else if (S_ISFIFO(sb.st_mode))
1406 f->filetype = FIO_TYPE_PIPE;
1410 static bool __is_already_allocated(const char *fname, bool set)
1412 struct flist_head *entry;
1415 ret = file_bloom_exists(fname, set);
1419 flist_for_each(entry, &filename_list) {
1420 struct file_name *fn;
1422 fn = flist_entry(entry, struct file_name, list);
1424 if (!strcmp(fn->filename, fname))
1431 static bool is_already_allocated(const char *fname)
1435 fio_file_hash_lock();
1436 ret = __is_already_allocated(fname, false);
1437 fio_file_hash_unlock();
1442 static void set_already_allocated(const char *fname)
1444 struct file_name *fn;
1446 fn = malloc(sizeof(struct file_name));
1447 fn->filename = strdup(fname);
1449 fio_file_hash_lock();
1450 if (!__is_already_allocated(fname, true)) {
1451 flist_add_tail(&fn->list, &filename_list);
1454 fio_file_hash_unlock();
1462 static void free_already_allocated(void)
1464 struct flist_head *entry, *tmp;
1465 struct file_name *fn;
1467 if (flist_empty(&filename_list))
1470 fio_file_hash_lock();
1471 flist_for_each_safe(entry, tmp, &filename_list) {
1472 fn = flist_entry(entry, struct file_name, list);
1474 flist_del(&fn->list);
1478 fio_file_hash_unlock();
1481 static struct fio_file *alloc_new_file(struct thread_data *td)
1485 f = smalloc(sizeof(*f));
1493 fio_file_reset(td, f);
1497 bool exists_and_not_regfile(const char *filename)
1501 if (lstat(filename, &sb) == -1)
1504 #ifndef WIN32 /* NOT Windows */
1505 if (S_ISREG(sb.st_mode))
1508 /* \\.\ is the device namespace in Windows, where every file
1509 * is a device node */
1510 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1517 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1519 int cur_files = td->files_index;
1520 char file_name[PATH_MAX];
1524 dprint(FD_FILE, "add file %s\n", fname);
1526 if (td->o.directory)
1527 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1528 td->o.unique_filename);
1530 sprintf(file_name + len, "%s", fname);
1532 /* clean cloned siblings using existing files */
1533 if (numjob && is_already_allocated(file_name) &&
1534 !exists_and_not_regfile(fname))
1537 f = alloc_new_file(td);
1539 if (td->files_size <= td->files_index) {
1540 unsigned int new_size = td->o.nr_files + 1;
1542 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1544 td->files = realloc(td->files, new_size * sizeof(f));
1545 if (td->files == NULL) {
1546 log_err("fio: realloc OOM\n");
1549 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1550 td->file_locks = realloc(td->file_locks, new_size);
1551 if (!td->file_locks) {
1552 log_err("fio: realloc OOM\n");
1555 td->file_locks[cur_files] = FILE_LOCK_NONE;
1557 td->files_size = new_size;
1559 td->files[cur_files] = f;
1560 f->fileno = cur_files;
1563 * init function, io engine may not be loaded yet
1565 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1566 f->real_file_size = -1ULL;
1568 f->file_name = smalloc_strdup(file_name);
1574 switch (td->o.file_lock_mode) {
1575 case FILE_LOCK_NONE:
1577 case FILE_LOCK_READWRITE:
1578 f->rwlock = fio_rwlock_init();
1580 case FILE_LOCK_EXCLUSIVE:
1581 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1584 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1589 if (f->filetype == FIO_TYPE_FILE)
1590 td->nr_normal_files++;
1592 set_already_allocated(file_name);
1597 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1603 int add_file_exclusive(struct thread_data *td, const char *fname)
1608 for_each_file(td, f, i) {
1609 if (!strcmp(f->file_name, fname))
1613 return add_file(td, fname, 0, 1);
1616 void get_file(struct fio_file *f)
1618 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1619 assert(fio_file_open(f));
1623 int put_file(struct thread_data *td, struct fio_file *f)
1625 int f_ret = 0, ret = 0;
1627 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1629 if (!fio_file_open(f)) {
1630 assert(f->fd == -1);
1634 assert(f->references);
1635 if (--f->references)
1638 if (should_fsync(td) && td->o.fsync_on_close) {
1639 f_ret = fsync(f->fd);
1644 if (td->io_ops->close_file)
1645 ret = td->io_ops->close_file(td, f);
1650 td->nr_open_files--;
1651 fio_file_clear_open(f);
1652 assert(f->fd == -1);
1656 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1658 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1661 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1662 if (ddir == DDIR_READ)
1663 fio_rwlock_read(f->rwlock);
1665 fio_rwlock_write(f->rwlock);
1666 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1667 fio_mutex_down(f->lock);
1669 td->file_locks[f->fileno] = td->o.file_lock_mode;
1672 void unlock_file(struct thread_data *td, struct fio_file *f)
1674 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1677 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1678 fio_rwlock_unlock(f->rwlock);
1679 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1680 fio_mutex_up(f->lock);
1682 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1685 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1687 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1689 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1693 static int recurse_dir(struct thread_data *td, const char *dirname)
1699 D = opendir(dirname);
1701 char buf[FIO_VERROR_SIZE];
1703 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1704 td_verror(td, errno, buf);
1708 while ((dir = readdir(D)) != NULL) {
1709 char full_path[PATH_MAX];
1712 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1715 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1717 if (lstat(full_path, &sb) == -1) {
1718 if (errno != ENOENT) {
1719 td_verror(td, errno, "stat");
1725 if (S_ISREG(sb.st_mode)) {
1726 add_file(td, full_path, 0, 1);
1729 if (!S_ISDIR(sb.st_mode))
1732 ret = recurse_dir(td, full_path);
1741 int add_dir_files(struct thread_data *td, const char *path)
1743 int ret = recurse_dir(td, path);
1746 log_info("fio: opendir added %d files\n", td->o.nr_files);
1751 void dup_files(struct thread_data *td, struct thread_data *org)
1756 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1761 td->files = malloc(org->files_index * sizeof(f));
1763 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1764 td->file_locks = malloc(org->files_index);
1766 for_each_file(org, f, i) {
1767 struct fio_file *__f;
1769 __f = alloc_new_file(td);
1772 __f->file_name = smalloc_strdup(f->file_name);
1773 if (!__f->file_name)
1776 __f->filetype = f->filetype;
1779 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1780 __f->lock = f->lock;
1781 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1782 __f->rwlock = f->rwlock;
1789 * Returns the index that matches the filename, or -1 if not there
1791 int get_fileno(struct thread_data *td, const char *fname)
1796 for_each_file(td, f, i)
1797 if (!strcmp(f->file_name, fname))
1804 * For log usage, where we add/open/close files automatically
1806 void free_release_files(struct thread_data *td)
1810 td->o.open_files = 0;
1811 td->files_index = 0;
1812 td->nr_normal_files = 0;
1815 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1819 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1820 f->last_pos[i] = f->file_offset;
1821 f->last_start[i] = -1ULL;
1824 if (fio_file_axmap(f))
1825 axmap_reset(f->io_axmap);
1826 else if (fio_file_lfsr(f))
1827 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1830 bool fio_files_done(struct thread_data *td)
1835 for_each_file(td, f, i)
1836 if (!fio_file_done(f))
1842 /* free memory used in initialization phase only */
1843 void filesetup_mem_free(void)
1845 free_already_allocated();