15 #include "lib/axmap.h"
18 #ifdef CONFIG_LINUX_FALLOCATE
19 #include <linux/falloc.h>
22 static FLIST_HEAD(filename_list);
25 * List entry for filename_list
28 struct flist_head list;
32 static inline void clear_error(struct thread_data *td)
38 static int native_fallocate(struct thread_data *td, struct fio_file *f)
42 success = fio_fallocate(f, 0, f->real_file_size);
43 dprint(FD_FILE, "native fallocate of file %s size %llu was "
44 "%ssuccessful\n", f->file_name,
45 (unsigned long long) f->real_file_size,
52 dprint(FD_FILE, "native fallocate is not implemented\n");
57 static void fallocate_file(struct thread_data *td, struct fio_file *f)
59 if (td->o.fill_device)
62 switch (td->o.fallocate_mode) {
63 case FIO_FALLOCATE_NATIVE:
64 native_fallocate(td, f);
66 case FIO_FALLOCATE_NONE:
68 #ifdef CONFIG_POSIX_FALLOCATE
69 case FIO_FALLOCATE_POSIX: {
72 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
74 (unsigned long long) f->real_file_size);
76 r = posix_fallocate(f->fd, 0, f->real_file_size);
78 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
81 #endif /* CONFIG_POSIX_FALLOCATE */
82 #ifdef CONFIG_LINUX_FALLOCATE
83 case FIO_FALLOCATE_KEEP_SIZE: {
86 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
87 "file %s size %llu\n", f->file_name,
88 (unsigned long long) f->real_file_size);
90 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
92 td_verror(td, errno, "fallocate");
96 #endif /* CONFIG_LINUX_FALLOCATE */
98 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
104 * Leaves f->fd open on success, caller must close
106 static int extend_file(struct thread_data *td, struct fio_file *f)
108 int new_layout = 0, unlink_file = 0, flags;
109 unsigned long long left;
114 log_err("fio: refusing extend of file due to read-only\n");
119 * check if we need to lay the file out complete again. fio
120 * does that for operations involving reads, or for writes
121 * where overwrite is set
124 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
125 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
127 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
130 if (unlink_file || new_layout) {
133 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
135 ret = td_io_unlink_file(td, f);
136 if (ret != 0 && ret != ENOENT) {
137 td_verror(td, errno, "unlink");
143 if (td->o.allow_create)
152 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
153 f->fd = open(f->file_name, flags, 0644);
157 if (err == ENOENT && !td->o.allow_create)
158 log_err("fio: file creation disallowed by "
159 "allow_file_create=0\n");
161 td_verror(td, err, "open");
165 fallocate_file(td, f);
168 * If our jobs don't require regular files initially, we're done.
174 * The size will be -1ULL when fill_device is used, so don't truncate
175 * or fallocate this file, just write it
177 if (!td->o.fill_device) {
178 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
179 (unsigned long long) f->real_file_size);
180 if (ftruncate(f->fd, f->real_file_size) == -1) {
181 if (errno != EFBIG) {
182 td_verror(td, errno, "ftruncate");
188 left = f->real_file_size;
189 bs = td->o.max_bs[DDIR_WRITE];
195 td_verror(td, errno, "malloc");
199 while (left && !td->terminate) {
205 fill_io_buffer(td, b, bs, bs);
207 r = write(f->fd, b, bs);
217 if (td->o.fill_device)
219 log_info("fio: ENOSPC on laying out "
223 td_verror(td, errno, "write");
225 td_verror(td, EIO, "write");
232 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
233 td_io_unlink_file(td, f);
234 } else if (td->o.create_fsync) {
235 if (fsync(f->fd) < 0) {
236 td_verror(td, errno, "fsync");
240 if (td->o.fill_device && !td_write(td)) {
241 fio_file_clear_size_known(f);
242 if (td_io_get_file_size(td, f))
244 if (f->io_size > f->real_file_size)
245 f->io_size = f->real_file_size;
259 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
261 int r, did_open = 0, old_runstate;
262 unsigned long long left;
267 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
268 td_ioengine_flagged(td, FIO_NOIO))
271 if (f->filetype == FIO_TYPE_CHAR)
274 if (!fio_file_open(f)) {
275 if (td->io_ops->open_file(td, f)) {
276 log_err("fio: cannot pre-read, failed to open file\n");
282 old_runstate = td_bump_runstate(td, TD_PRE_READING);
285 bs = td->o.max_bs[DDIR_READ];
291 td_verror(td, errno, "malloc");
297 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
298 td_verror(td, errno, "lseek");
299 log_err("fio: failed to lseek pre-read file\n");
304 while (left && !td->terminate) {
308 r = read(f->fd, b, bs);
314 td_verror(td, EIO, "pre_read");
320 td_restore_runstate(td, old_runstate);
323 td->io_ops->close_file(td, f);
329 unsigned long long get_rand_file_size(struct thread_data *td)
331 unsigned long long ret, sized;
335 frand_max = rand_max(&td->file_size_state);
336 r = __rand(&td->file_size_state);
337 sized = td->o.file_size_high - td->o.file_size_low;
338 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
339 ret += td->o.file_size_low;
340 ret -= (ret % td->o.rw_min_bs);
344 static int file_size(struct thread_data *td, struct fio_file *f)
348 if (stat(f->file_name, &st) == -1) {
349 td_verror(td, errno, "fstat");
353 f->real_file_size = st.st_size;
357 static int bdev_size(struct thread_data *td, struct fio_file *f)
359 unsigned long long bytes = 0;
362 if (td->io_ops->open_file(td, f)) {
363 log_err("fio: failed opening blockdev %s for size check\n",
368 r = blockdev_size(f, &bytes);
370 td_verror(td, r, "blockdev_size");
375 log_err("%s: zero sized block device?\n", f->file_name);
379 f->real_file_size = bytes;
380 td->io_ops->close_file(td, f);
383 td->io_ops->close_file(td, f);
387 static int char_size(struct thread_data *td, struct fio_file *f)
389 #ifdef FIO_HAVE_CHARDEV_SIZE
390 unsigned long long bytes = 0;
393 if (td->io_ops->open_file(td, f)) {
394 log_err("fio: failed opening chardev %s for size check\n",
399 r = chardev_size(f, &bytes);
401 td_verror(td, r, "chardev_size");
406 log_err("%s: zero sized char device?\n", f->file_name);
410 f->real_file_size = bytes;
411 td->io_ops->close_file(td, f);
414 td->io_ops->close_file(td, f);
417 f->real_file_size = -1ULL;
422 static int get_file_size(struct thread_data *td, struct fio_file *f)
426 if (fio_file_size_known(f))
429 if (f->filetype == FIO_TYPE_FILE)
430 ret = file_size(td, f);
431 else if (f->filetype == FIO_TYPE_BLOCK)
432 ret = bdev_size(td, f);
433 else if (f->filetype == FIO_TYPE_CHAR)
434 ret = char_size(td, f);
436 f->real_file_size = -1;
437 log_info("%s: failed to get file size of %s\n", td->o.name,
439 return 1; /* avoid offset extends end error message */
443 * Leave ->real_file_size with 0 since it could be expectation
444 * of initial setup for regular files.
450 * ->file_offset normally hasn't been initialized yet, so this
451 * is basically always false unless ->real_file_size is -1, but
452 * if ->real_file_size is -1 this message doesn't make sense.
453 * As a result, this message is basically useless.
455 if (f->file_offset > f->real_file_size) {
456 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
457 (unsigned long long) f->file_offset,
458 (unsigned long long) f->real_file_size);
462 fio_file_set_size_known(f);
466 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
467 unsigned long long off,
468 unsigned long long len)
470 int errval = 0, ret = 0;
479 off = f->file_offset;
481 if (len == -1ULL || off == -1ULL)
484 if (td->io_ops->invalidate) {
485 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
487 ret = td->io_ops->invalidate(td, f);
490 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
491 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
493 } else if (f->filetype == FIO_TYPE_FILE) {
494 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
495 f->file_name, off, len);
496 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
499 } else if (f->filetype == FIO_TYPE_BLOCK) {
502 dprint(FD_IO, "drop page cache %s\n", f->file_name);
503 ret = blockdev_invalidate_cache(f);
504 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
506 * Linux multipath devices reject ioctl while
507 * the maps are being updated. That window can
508 * last tens of milliseconds; we'll try up to
509 * a quarter of a second.
512 ret = blockdev_invalidate_cache(f);
514 if (ret < 0 && errno == EACCES && geteuid()) {
515 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
516 log_err("fio: only root may flush block "
517 "devices. Cache flush bypassed!\n");
522 } else if (f->filetype == FIO_TYPE_CHAR ||
523 f->filetype == FIO_TYPE_PIPE) {
524 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
528 * Cache flushing isn't a fatal condition, and we know it will
529 * happen on some platforms where we don't have the proper
530 * function to flush eg block device caches. So just warn and
531 * continue on our way.
534 log_info("fio: cache invalidation of %s failed: %s\n",
535 f->file_name, strerror(errval));
541 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
543 if (!fio_file_open(f))
546 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
549 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
553 dprint(FD_FILE, "fd close %s\n", f->file_name);
557 if (close(f->fd) < 0)
562 if (f->shadow_fd != -1) {
571 int file_lookup_open(struct fio_file *f, int flags)
573 struct fio_file *__f;
576 __f = lookup_file_hash(f->file_name);
578 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
582 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
590 f->fd = open(f->file_name, flags, 0600);
594 static int file_close_shadow_fds(struct thread_data *td)
600 for_each_file(td, f, i) {
601 if (f->shadow_fd == -1)
612 int generic_open_file(struct thread_data *td, struct fio_file *f)
618 dprint(FD_FILE, "fd open %s\n", f->file_name);
620 if (!strcmp(f->file_name, "-")) {
622 log_err("fio: can't read/write to stdin/out\n");
628 * move output logging to stderr, if we are writing to stdout
637 flags |= OS_O_DIRECT;
640 td_verror(td, EINVAL, "OS does not support atomic IO");
643 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
647 if (td->o.create_on_open && td->o.allow_create)
650 if (f->filetype != FIO_TYPE_FILE)
651 flags |= FIO_O_NOATIME;
658 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
662 f->fd = dup(STDOUT_FILENO);
664 from_hash = file_lookup_open(f, flags);
665 } else if (td_read(td)) {
666 if (f->filetype == FIO_TYPE_CHAR && !read_only)
672 f->fd = dup(STDIN_FILENO);
674 from_hash = file_lookup_open(f, flags);
675 } else if (td_trim(td)) {
676 assert(!td_rw(td)); /* should have matched above */
678 from_hash = file_lookup_open(f, flags);
682 char buf[FIO_VERROR_SIZE];
685 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
686 flags &= ~FIO_O_NOATIME;
689 if (__e == EMFILE && file_close_shadow_fds(td))
692 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
694 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
695 log_err("fio: looks like your file system does not " \
696 "support direct=1/buffered=0\n");
699 td_verror(td, __e, buf);
703 if (!from_hash && f->fd != -1) {
704 if (add_file_hash(f)) {
708 * Stash away descriptor for later close. This is to
709 * work-around a "feature" on Linux, where a close of
710 * an fd that has been opened for write will trigger
711 * udev to call blkid to check partitions, fs id, etc.
712 * That pollutes the device cache, which can slow down
713 * unbuffered accesses.
715 if (f->shadow_fd == -1)
716 f->shadow_fd = f->fd;
719 * OK to ignore, we haven't done anything
722 ret = generic_close_file(td, f);
732 * This function i.e. get_file_size() is the default .get_file_size
733 * implementation of majority of I/O engines.
735 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
737 return get_file_size(td, f);
741 * open/close all files, so that ->real_file_size gets set
743 static int get_file_sizes(struct thread_data *td)
749 for_each_file(td, f, i) {
750 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
753 if (td_io_get_file_size(td, f)) {
754 if (td->error != ENOENT) {
755 log_err("%s\n", td->verror);
763 * There are corner cases where we end up with -1 for
764 * ->real_file_size due to unsupported file type, etc.
765 * We then just set to size option value divided by number
766 * of files, similar to the way file ->io_size is set.
767 * stat(2) failure doesn't set ->real_file_size to -1.
769 if (f->real_file_size == -1ULL && td->o.size)
770 f->real_file_size = td->o.size / td->o.nr_files;
777 struct flist_head list;
784 * Get free number of bytes for each file on each unique mount.
786 static unsigned long long get_fs_free_counts(struct thread_data *td)
788 struct flist_head *n, *tmp;
789 unsigned long long ret = 0;
790 struct fio_mount *fm;
795 for_each_file(td, f, i) {
799 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
800 if (f->real_file_size != -1ULL)
801 ret += f->real_file_size;
803 } else if (f->filetype != FIO_TYPE_FILE)
807 strncpy(buf, f->file_name, 255);
809 if (stat(buf, &sb) < 0) {
813 if (stat(buf, &sb) < 0)
818 flist_for_each(n, &list) {
819 fm = flist_entry(n, struct fio_mount, list);
820 if (fm->key == sb.st_dev)
829 fm = calloc(1, sizeof(*fm));
830 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
831 fm->base = basename(fm->__base);
833 flist_add(&fm->list, &list);
836 flist_for_each_safe(n, tmp, &list) {
837 unsigned long long sz;
839 fm = flist_entry(n, struct fio_mount, list);
840 flist_del(&fm->list);
842 sz = get_fs_free_size(fm->base);
843 if (sz && sz != -1ULL)
852 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
854 struct thread_options *o = &td->o;
855 unsigned long long align_bs;
856 unsigned long long offset;
858 if (o->file_append && f->filetype == FIO_TYPE_FILE)
859 return f->real_file_size;
861 if (o->start_offset_percent > 0) {
863 * if offset_align is provided, set initial offset
865 if (fio_option_is_set(o, start_offset_align)) {
866 align_bs = o->start_offset_align;
868 /* else take the minimum block size */
869 align_bs = td_min_bs(td);
872 /* calculate the raw offset */
873 offset = (f->real_file_size * o->start_offset_percent / 100) +
874 (td->subjob_number * o->offset_increment);
877 * block align the offset at the next available boundary at
878 * ceiling(offset / align_bs) * align_bs
880 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
883 /* start_offset_percent not set */
884 offset = o->start_offset +
885 td->subjob_number * o->offset_increment;
892 * Open the files and setup files sizes, creating files if necessary.
894 int setup_files(struct thread_data *td)
896 unsigned long long total_size, extend_size;
897 struct thread_options *o = &td->o;
899 unsigned int i, nr_fs_extra = 0;
900 int err = 0, need_extend;
902 const unsigned int bs = td_min_bs(td);
905 dprint(FD_FILE, "setup files\n");
907 old_state = td_bump_runstate(td, TD_SETTING_UP);
909 if (o->read_iolog_file)
913 * Find out physical size of files or devices for this thread,
914 * before we determine I/O size and range of our targets.
915 * If ioengine defines a setup() method, it's responsible for
916 * opening the files and setting f->real_file_size to indicate
917 * the valid range for that file.
919 if (td->io_ops->setup)
920 err = td->io_ops->setup(td);
922 err = get_file_sizes(td);
928 * check sizes. if the files/devices do not exist and the size
929 * isn't passed to fio, abort.
932 for_each_file(td, f, i) {
934 if (f->real_file_size == -1ULL)
937 total_size += f->real_file_size;
941 td->fill_device_size = get_fs_free_counts(td);
944 * device/file sizes are zero and no size given, punt
946 if ((!total_size || total_size == -1ULL) && !o->size &&
947 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
948 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
949 log_err("%s: you need to specify size=\n", o->name);
950 td_verror(td, EINVAL, "total_file_size");
955 * Calculate per-file size and potential extra size for the
956 * first files, if needed (i.e. if we don't have a fixed size).
958 if (!o->file_size_low && o->nr_files) {
961 fs = o->size / o->nr_files;
962 all_fs = fs * o->nr_files;
964 if (all_fs < o->size)
965 nr_fs_extra = (o->size - all_fs) / bs;
969 * now file sizes are known, so we can set ->io_size. if size= is
970 * not given, ->io_size is just equal to ->real_file_size. if size
971 * is given, ->io_size is size / nr_files.
973 extend_size = total_size = 0;
975 for_each_file(td, f, i) {
976 f->file_offset = get_start_offset(td, f);
979 * Update ->io_size depending on options specified.
980 * ->file_size_low being 0 means filesize option isn't set.
981 * Non zero ->file_size_low equals ->file_size_high means
982 * filesize option is set in a fixed size format.
983 * Non zero ->file_size_low not equals ->file_size_high means
984 * filesize option is set in a range format.
986 if (!o->file_size_low) {
988 * no file size or range given, file size is equal to
989 * total size divided by number of files. If the size
990 * doesn't divide nicely with the min blocksize,
991 * make the first files bigger.
1000 * We normally don't come here for regular files, but
1001 * if the result is 0 for a regular file, set it to the
1002 * real file size. This could be size of the existing
1003 * one if it already exists, but otherwise will be set
1004 * to 0. A new file won't be created because
1005 * ->io_size + ->file_offset equals ->real_file_size.
1008 if (f->file_offset > f->real_file_size)
1010 f->io_size = f->real_file_size - f->file_offset;
1012 log_info("fio: file %s may be ignored\n",
1015 } else if (f->real_file_size < o->file_size_low ||
1016 f->real_file_size > o->file_size_high) {
1017 if (f->file_offset > o->file_size_low)
1020 * file size given. if it's fixed, use that. if it's a
1021 * range, generate a random size in-between.
1023 if (o->file_size_low == o->file_size_high)
1024 f->io_size = o->file_size_low - f->file_offset;
1026 f->io_size = get_rand_file_size(td)
1030 f->io_size = f->real_file_size - f->file_offset;
1032 if (f->io_size == -1ULL)
1035 if (o->size_percent && o->size_percent != 100) {
1038 file_size = f->io_size + f->file_offset;
1039 f->io_size = (file_size *
1040 o->size_percent) / 100;
1041 if (f->io_size > (file_size - f->file_offset))
1042 f->io_size = file_size - f->file_offset;
1044 f->io_size -= (f->io_size % td_min_bs(td));
1046 total_size += f->io_size;
1049 if (f->filetype == FIO_TYPE_FILE &&
1050 (f->io_size + f->file_offset) > f->real_file_size &&
1051 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1052 if (!o->create_on_open) {
1054 extend_size += (f->io_size + f->file_offset);
1055 fio_file_set_extend(f);
1057 f->real_file_size = f->io_size + f->file_offset;
1061 if (td->o.block_error_hist) {
1064 assert(td->o.nr_files == 1); /* checked in fixup_options */
1066 len = f->io_size / td->o.bs[DDIR_TRIM];
1067 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1068 log_err("fio: cannot calculate block histogram with "
1069 "%d trim blocks, maximum %d\n",
1070 len, MAX_NR_BLOCK_INFOS);
1071 td_verror(td, EINVAL, "block_error_hist");
1075 td->ts.nr_block_infos = len;
1076 for (i = 0; i < len; i++)
1077 td->ts.block_infos[i] =
1078 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1080 td->ts.nr_block_infos = 0;
1082 if (!o->size || (total_size && o->size > total_size))
1083 o->size = total_size;
1085 if (o->size < td_min_bs(td)) {
1086 log_err("fio: blocksize too large for data set\n");
1091 * See if we need to extend some files, typically needed when our
1092 * target regular files don't exist yet, but our jobs require them
1093 * initially due to read I/Os.
1097 if (output_format & FIO_OUTPUT_NORMAL) {
1098 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1100 need_extend > 1 ? "s" : "",
1102 need_extend > 1 ? "s" : "",
1103 need_extend > 1 ? "total " : "",
1107 for_each_file(td, f, i) {
1108 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1110 if (!fio_file_extend(f))
1113 assert(f->filetype == FIO_TYPE_FILE);
1114 fio_file_clear_extend(f);
1115 if (!o->fill_device) {
1116 old_len = f->real_file_size;
1117 extend_len = f->io_size + f->file_offset -
1120 f->real_file_size = (f->io_size + f->file_offset);
1121 err = extend_file(td, f);
1125 err = __file_invalidate_cache(td, f, old_len,
1129 * Shut up static checker
1145 o->zone_size = o->size;
1148 * iolog already set the total io size, if we read back
1151 if (!o->read_iolog_file) {
1153 td->total_io_size = o->io_size * o->loops;
1155 td->total_io_size = o->size * o->loops;
1162 td_restore_runstate(td, old_state);
1165 log_err("%s: you need to specify valid offset=\n", o->name);
1167 td_restore_runstate(td, old_state);
1171 bool pre_read_files(struct thread_data *td)
1176 dprint(FD_FILE, "pre_read files\n");
1178 for_each_file(td, f, i) {
1179 if (!pre_read_file(td, f))
1186 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1188 unsigned int range_size, seed;
1189 unsigned long nranges;
1192 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1193 fsize = min(f->real_file_size, f->io_size);
1195 nranges = (fsize + range_size - 1) / range_size;
1197 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1198 if (!td->o.rand_repeatable)
1199 seed = td->rand_seeds[4];
1201 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1202 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1203 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1204 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1205 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1206 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1209 static bool init_rand_distribution(struct thread_data *td)
1215 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1218 state = td_bump_runstate(td, TD_SETTING_UP);
1220 for_each_file(td, f, i)
1221 __init_rand_distribution(td, f);
1223 td_restore_runstate(td, state);
1228 * Check if the number of blocks exceeds the randomness capability of
1229 * the selected generator. Tausworthe is 32-bit, the others are fullly
1232 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1235 if (blocks <= FRAND32_MAX)
1237 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1241 * If the user hasn't specified a random generator, switch
1242 * to tausworthe64 with informational warning. If the user did
1243 * specify one, just warn.
1245 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1248 if (!fio_option_is_set(&td->o, random_generator)) {
1249 log_info("fio: Switching to tausworthe64. Use the "
1250 "random_generator= option to get rid of this "
1252 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1257 * Just make this information to avoid breaking scripts.
1259 log_info("fio: Use the random_generator= option to switch to lfsr or "
1264 bool init_random_map(struct thread_data *td)
1266 unsigned long long blocks;
1270 if (init_rand_distribution(td))
1275 for_each_file(td, f, i) {
1276 uint64_t fsize = min(f->real_file_size, f->io_size);
1278 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1280 if (check_rand_gen_limits(td, f, blocks))
1283 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1286 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1288 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1289 fio_file_set_lfsr(f);
1292 } else if (!td->o.norandommap) {
1293 f->io_axmap = axmap_new(blocks);
1295 fio_file_set_axmap(f);
1298 } else if (td->o.norandommap)
1301 if (!td->o.softrandommap) {
1302 log_err("fio: failed allocating random map. If running"
1303 " a large number of jobs, try the 'norandommap'"
1304 " option or set 'softrandommap'. Or give"
1305 " a larger --alloc-size to fio.\n");
1309 log_info("fio: file %s failed allocating random map. Running "
1310 "job without.\n", f->file_name);
1316 void close_files(struct thread_data *td)
1321 for_each_file(td, f, i) {
1322 if (fio_file_open(f))
1323 td_io_close_file(td, f);
1327 void close_and_free_files(struct thread_data *td)
1331 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1333 dprint(FD_FILE, "close files\n");
1335 for_each_file(td, f, i) {
1336 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1337 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1338 td_io_unlink_file(td, f);
1341 if (fio_file_open(f))
1342 td_io_close_file(td, f);
1344 remove_file_hash(f);
1346 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1347 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1348 td_io_unlink_file(td, f);
1354 sfree(f->file_name);
1355 f->file_name = NULL;
1356 if (fio_file_axmap(f)) {
1357 axmap_free(f->io_axmap);
1366 td->o.filename = NULL;
1368 free(td->file_locks);
1369 td->files_index = 0;
1371 td->file_locks = NULL;
1372 td->o.file_lock_mode = FILE_LOCK_NONE;
1376 static void get_file_type(struct fio_file *f)
1380 if (!strcmp(f->file_name, "-"))
1381 f->filetype = FIO_TYPE_PIPE;
1383 f->filetype = FIO_TYPE_FILE;
1386 /* \\.\ is the device namespace in Windows, where every file is
1388 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1389 f->filetype = FIO_TYPE_BLOCK;
1392 if (!stat(f->file_name, &sb)) {
1393 if (S_ISBLK(sb.st_mode))
1394 f->filetype = FIO_TYPE_BLOCK;
1395 else if (S_ISCHR(sb.st_mode))
1396 f->filetype = FIO_TYPE_CHAR;
1397 else if (S_ISFIFO(sb.st_mode))
1398 f->filetype = FIO_TYPE_PIPE;
1402 static bool __is_already_allocated(const char *fname, bool set)
1404 struct flist_head *entry;
1407 ret = file_bloom_exists(fname, set);
1411 flist_for_each(entry, &filename_list) {
1412 struct file_name *fn;
1414 fn = flist_entry(entry, struct file_name, list);
1416 if (!strcmp(fn->filename, fname))
1423 static bool is_already_allocated(const char *fname)
1427 fio_file_hash_lock();
1428 ret = __is_already_allocated(fname, false);
1429 fio_file_hash_unlock();
1434 static void set_already_allocated(const char *fname)
1436 struct file_name *fn;
1438 fn = malloc(sizeof(struct file_name));
1439 fn->filename = strdup(fname);
1441 fio_file_hash_lock();
1442 if (!__is_already_allocated(fname, true)) {
1443 flist_add_tail(&fn->list, &filename_list);
1446 fio_file_hash_unlock();
1454 static void free_already_allocated(void)
1456 struct flist_head *entry, *tmp;
1457 struct file_name *fn;
1459 if (flist_empty(&filename_list))
1462 fio_file_hash_lock();
1463 flist_for_each_safe(entry, tmp, &filename_list) {
1464 fn = flist_entry(entry, struct file_name, list);
1466 flist_del(&fn->list);
1470 fio_file_hash_unlock();
1473 static struct fio_file *alloc_new_file(struct thread_data *td)
1477 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1478 f = calloc(1, sizeof(*f));
1480 f = scalloc(1, sizeof(*f));
1488 fio_file_reset(td, f);
1492 bool exists_and_not_regfile(const char *filename)
1496 if (lstat(filename, &sb) == -1)
1499 #ifndef WIN32 /* NOT Windows */
1500 if (S_ISREG(sb.st_mode))
1503 /* \\.\ is the device namespace in Windows, where every file
1504 * is a device node */
1505 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1512 static bool create_work_dirs(struct thread_data *td, const char *fname)
1514 char path[PATH_MAX];
1517 if (td->o.directory) {
1518 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1519 FIO_OS_PATH_SEPARATOR, fname);
1520 start = strstr(path, fname);
1522 snprintf(path, PATH_MAX, "%s", fname);
1527 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1532 #ifdef CONFIG_HAVE_MKDIR_TWO
1533 if (mkdir(path, 0600) && errno != EEXIST) {
1535 if (mkdir(path) && errno != EEXIST) {
1537 log_err("fio: failed to create dir (%s): %d\n",
1541 *end = FIO_OS_PATH_SEPARATOR;
1544 td->flags |= TD_F_DIRS_CREATED;
1548 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1550 int cur_files = td->files_index;
1551 char file_name[PATH_MAX];
1555 dprint(FD_FILE, "add file %s\n", fname);
1557 if (td->o.directory)
1558 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1559 td->o.unique_filename);
1561 sprintf(file_name + len, "%s", fname);
1563 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1564 !(td->flags & TD_F_DIRS_CREATED) &&
1565 !create_work_dirs(td, fname))
1568 /* clean cloned siblings using existing files */
1569 if (numjob && is_already_allocated(file_name) &&
1570 !exists_and_not_regfile(fname))
1573 f = alloc_new_file(td);
1575 if (td->files_size <= td->files_index) {
1576 unsigned int new_size = td->o.nr_files + 1;
1578 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1580 td->files = realloc(td->files, new_size * sizeof(f));
1581 if (td->files == NULL) {
1582 log_err("fio: realloc OOM\n");
1585 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1586 td->file_locks = realloc(td->file_locks, new_size);
1587 if (!td->file_locks) {
1588 log_err("fio: realloc OOM\n");
1591 td->file_locks[cur_files] = FILE_LOCK_NONE;
1593 td->files_size = new_size;
1595 td->files[cur_files] = f;
1596 f->fileno = cur_files;
1599 * init function, io engine may not be loaded yet
1601 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1602 f->real_file_size = -1ULL;
1604 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1605 f->file_name = strdup(file_name);
1607 f->file_name = smalloc_strdup(file_name);
1609 /* can't handle smalloc failure from here */
1610 assert(f->file_name);
1614 switch (td->o.file_lock_mode) {
1615 case FILE_LOCK_NONE:
1617 case FILE_LOCK_READWRITE:
1618 f->rwlock = fio_rwlock_init();
1620 case FILE_LOCK_EXCLUSIVE:
1621 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1624 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1630 if (td->o.numjobs > 1)
1631 set_already_allocated(file_name);
1636 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1642 int add_file_exclusive(struct thread_data *td, const char *fname)
1647 for_each_file(td, f, i) {
1648 if (!strcmp(f->file_name, fname))
1652 return add_file(td, fname, 0, 1);
1655 void get_file(struct fio_file *f)
1657 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1658 assert(fio_file_open(f));
1662 int put_file(struct thread_data *td, struct fio_file *f)
1664 int f_ret = 0, ret = 0;
1666 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1668 if (!fio_file_open(f)) {
1669 assert(f->fd == -1);
1673 assert(f->references);
1674 if (--f->references)
1677 if (should_fsync(td) && td->o.fsync_on_close) {
1678 f_ret = fsync(f->fd);
1683 if (td->io_ops->close_file)
1684 ret = td->io_ops->close_file(td, f);
1689 td->nr_open_files--;
1690 fio_file_clear_open(f);
1691 assert(f->fd == -1);
1695 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1697 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1700 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1701 if (ddir == DDIR_READ)
1702 fio_rwlock_read(f->rwlock);
1704 fio_rwlock_write(f->rwlock);
1705 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1706 fio_sem_down(f->lock);
1708 td->file_locks[f->fileno] = td->o.file_lock_mode;
1711 void unlock_file(struct thread_data *td, struct fio_file *f)
1713 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1716 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1717 fio_rwlock_unlock(f->rwlock);
1718 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1719 fio_sem_up(f->lock);
1721 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1724 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1726 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1728 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1732 static bool recurse_dir(struct thread_data *td, const char *dirname)
1738 D = opendir(dirname);
1740 char buf[FIO_VERROR_SIZE];
1742 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1743 td_verror(td, errno, buf);
1747 while ((dir = readdir(D)) != NULL) {
1748 char full_path[PATH_MAX];
1751 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1754 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1756 if (lstat(full_path, &sb) == -1) {
1757 if (errno != ENOENT) {
1758 td_verror(td, errno, "stat");
1764 if (S_ISREG(sb.st_mode)) {
1765 add_file(td, full_path, 0, 1);
1768 if (!S_ISDIR(sb.st_mode))
1771 ret = recurse_dir(td, full_path);
1780 int add_dir_files(struct thread_data *td, const char *path)
1782 int ret = recurse_dir(td, path);
1785 log_info("fio: opendir added %d files\n", td->o.nr_files);
1790 void dup_files(struct thread_data *td, struct thread_data *org)
1795 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1800 td->files = malloc(org->files_index * sizeof(f));
1802 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1803 td->file_locks = malloc(org->files_index);
1805 for_each_file(org, f, i) {
1806 struct fio_file *__f;
1808 __f = alloc_new_file(td);
1811 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1812 __f->file_name = strdup(f->file_name);
1814 __f->file_name = smalloc_strdup(f->file_name);
1816 /* can't handle smalloc failure from here */
1817 assert(__f->file_name);
1818 __f->filetype = f->filetype;
1821 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1822 __f->lock = f->lock;
1823 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1824 __f->rwlock = f->rwlock;
1831 * Returns the index that matches the filename, or -1 if not there
1833 int get_fileno(struct thread_data *td, const char *fname)
1838 for_each_file(td, f, i)
1839 if (!strcmp(f->file_name, fname))
1846 * For log usage, where we add/open/close files automatically
1848 void free_release_files(struct thread_data *td)
1852 td->o.open_files = 0;
1853 td->files_index = 0;
1856 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1860 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1861 f->last_pos[i] = f->file_offset;
1862 f->last_start[i] = -1ULL;
1865 if (fio_file_axmap(f))
1866 axmap_reset(f->io_axmap);
1867 else if (fio_file_lfsr(f))
1868 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1871 bool fio_files_done(struct thread_data *td)
1876 for_each_file(td, f, i)
1877 if (!fio_file_done(f))
1883 /* free memory used in initialization phase only */
1884 void filesetup_mem_free(void)
1886 free_already_allocated();
1890 * This function is for platforms which support direct I/O but not O_DIRECT.
1892 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1894 #ifdef FIO_OS_DIRECTIO
1895 int ret = fio_set_odirect(f);
1898 td_verror(td, ret, "fio_set_directio");
1899 #if defined(__sun__)
1900 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1901 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1903 log_err("fio: the file system does not seem to support direct IO\n");
1906 log_err("fio: the file system does not seem to support direct IO\n");
1913 log_err("fio: direct IO is not supported on this host operating system\n");