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 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 native_fallocate(td, f);
71 case FIO_FALLOCATE_NONE:
73 #ifdef CONFIG_POSIX_FALLOCATE
74 case FIO_FALLOCATE_POSIX:
75 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
77 (unsigned long long) f->real_file_size);
79 r = posix_fallocate(f->fd, 0, f->real_file_size);
81 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
83 #endif /* CONFIG_POSIX_FALLOCATE */
84 #ifdef CONFIG_LINUX_FALLOCATE
85 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");
95 #endif /* CONFIG_LINUX_FALLOCATE */
97 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
103 * Leaves f->fd open on success, caller must close
105 static int extend_file(struct thread_data *td, struct fio_file *f)
107 int new_layout = 0, unlink_file = 0, flags;
108 unsigned long long left;
113 log_err("fio: refusing extend of file due to read-only\n");
118 * check if we need to lay the file out complete again. fio
119 * does that for operations involving reads, or for writes
120 * where overwrite is set
123 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
124 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
126 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
129 if (unlink_file || new_layout) {
132 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
134 ret = td_io_unlink_file(td, f);
135 if (ret != 0 && ret != ENOENT) {
136 td_verror(td, errno, "unlink");
142 if (td->o.allow_create)
151 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
152 f->fd = open(f->file_name, flags, 0644);
156 if (err == ENOENT && !td->o.allow_create)
157 log_err("fio: file creation disallowed by "
158 "allow_file_create=0\n");
160 td_verror(td, err, "open");
164 fallocate_file(td, f);
167 * If our jobs don't require regular files initially, we're done.
173 * The size will be -1ULL when fill_device is used, so don't truncate
174 * or fallocate this file, just write it
176 if (!td->o.fill_device) {
177 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
178 (unsigned long long) f->real_file_size);
179 if (ftruncate(f->fd, f->real_file_size) == -1) {
180 if (errno != EFBIG) {
181 td_verror(td, errno, "ftruncate");
187 left = f->real_file_size;
188 bs = td->o.max_bs[DDIR_WRITE];
194 td_verror(td, errno, "malloc");
198 while (left && !td->terminate) {
204 fill_io_buffer(td, b, bs, bs);
206 r = write(f->fd, b, bs);
216 if (td->o.fill_device)
218 log_info("fio: ENOSPC on laying out "
222 td_verror(td, errno, "write");
224 td_verror(td, EIO, "write");
231 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
232 td_io_unlink_file(td, f);
233 } else if (td->o.create_fsync) {
234 if (fsync(f->fd) < 0) {
235 td_verror(td, errno, "fsync");
239 if (td->o.fill_device && !td_write(td)) {
240 fio_file_clear_size_known(f);
241 if (td_io_get_file_size(td, f))
243 if (f->io_size > f->real_file_size)
244 f->io_size = f->real_file_size;
258 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
260 int r, did_open = 0, old_runstate;
261 unsigned long long left;
266 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
267 td_ioengine_flagged(td, FIO_NOIO))
270 if (f->filetype == FIO_TYPE_CHAR)
273 if (!fio_file_open(f)) {
274 if (td->io_ops->open_file(td, f)) {
275 log_err("fio: cannot pre-read, failed to open file\n");
281 old_runstate = td_bump_runstate(td, TD_PRE_READING);
284 bs = td->o.max_bs[DDIR_READ];
290 td_verror(td, errno, "malloc");
296 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
297 td_verror(td, errno, "lseek");
298 log_err("fio: failed to lseek pre-read file\n");
303 while (left && !td->terminate) {
307 r = read(f->fd, b, bs);
313 td_verror(td, EIO, "pre_read");
319 td_restore_runstate(td, old_runstate);
322 td->io_ops->close_file(td, f);
328 unsigned long long get_rand_file_size(struct thread_data *td)
330 unsigned long long ret, sized;
334 frand_max = rand_max(&td->file_size_state);
335 r = __rand(&td->file_size_state);
336 sized = td->o.file_size_high - td->o.file_size_low;
337 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
338 ret += td->o.file_size_low;
339 ret -= (ret % td->o.rw_min_bs);
343 static int file_size(struct thread_data *td, struct fio_file *f)
347 if (stat(f->file_name, &st) == -1) {
348 td_verror(td, errno, "fstat");
352 f->real_file_size = st.st_size;
356 static int bdev_size(struct thread_data *td, struct fio_file *f)
358 unsigned long long bytes = 0;
361 if (td->io_ops->open_file(td, f)) {
362 log_err("fio: failed opening blockdev %s for size check\n",
367 r = blockdev_size(f, &bytes);
369 td_verror(td, r, "blockdev_size");
374 log_err("%s: zero sized block device?\n", f->file_name);
378 f->real_file_size = bytes;
379 td->io_ops->close_file(td, f);
382 td->io_ops->close_file(td, f);
386 static int char_size(struct thread_data *td, struct fio_file *f)
388 #ifdef FIO_HAVE_CHARDEV_SIZE
389 unsigned long long bytes = 0;
392 if (td->io_ops->open_file(td, f)) {
393 log_err("fio: failed opening chardev %s for size check\n",
398 r = chardev_size(f, &bytes);
400 td_verror(td, r, "chardev_size");
405 log_err("%s: zero sized char device?\n", f->file_name);
409 f->real_file_size = bytes;
410 td->io_ops->close_file(td, f);
413 td->io_ops->close_file(td, f);
416 f->real_file_size = -1ULL;
421 static int get_file_size(struct thread_data *td, struct fio_file *f)
425 if (fio_file_size_known(f))
428 if (f->filetype == FIO_TYPE_FILE)
429 ret = file_size(td, f);
430 else if (f->filetype == FIO_TYPE_BLOCK)
431 ret = bdev_size(td, f);
432 else if (f->filetype == FIO_TYPE_CHAR)
433 ret = char_size(td, f);
435 f->real_file_size = -1ULL;
438 * Leave ->real_file_size with 0 since it could be expectation
439 * of initial setup for regular files.
445 * If ->real_file_size is -1, a conditional for the message
446 * "offset extends end" is always true, but it makes no sense,
447 * so just return the same value here.
449 if (f->real_file_size == -1ULL) {
450 log_info("%s: failed to get file size of %s\n", td->o.name,
455 if (td->o.start_offset && f->file_offset == 0)
456 dprint(FD_FILE, "offset of file %s not initialized yet\n",
459 * ->file_offset normally hasn't been initialized yet, so this
460 * is basically always false.
462 if (f->file_offset > f->real_file_size) {
463 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
464 (unsigned long long) f->file_offset,
465 (unsigned long long) f->real_file_size);
469 fio_file_set_size_known(f);
473 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
474 unsigned long long off,
475 unsigned long long len)
477 int errval = 0, ret = 0;
486 off = f->file_offset;
488 if (len == -1ULL || off == -1ULL)
491 if (td->io_ops->invalidate) {
492 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
494 ret = td->io_ops->invalidate(td, f);
497 } else if (f->filetype == FIO_TYPE_FILE) {
498 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
499 f->file_name, off, len);
500 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
503 } else if (f->filetype == FIO_TYPE_BLOCK) {
506 dprint(FD_IO, "drop page cache %s\n", f->file_name);
507 ret = blockdev_invalidate_cache(f);
508 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
510 * Linux multipath devices reject ioctl while
511 * the maps are being updated. That window can
512 * last tens of milliseconds; we'll try up to
513 * a quarter of a second.
516 ret = blockdev_invalidate_cache(f);
518 if (ret < 0 && errno == EACCES && geteuid()) {
520 log_err("fio: only root may flush block "
521 "devices. Cache flush bypassed!\n");
528 } else if (f->filetype == FIO_TYPE_CHAR ||
529 f->filetype == FIO_TYPE_PIPE) {
530 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
535 * Cache flushing isn't a fatal condition, and we know it will
536 * happen on some platforms where we don't have the proper
537 * function to flush eg block device caches. So just warn and
538 * continue on our way.
541 log_info("fio: cache invalidation of %s failed: %s\n",
542 f->file_name, strerror(errval));
548 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
550 if (!fio_file_open(f))
553 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
556 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
560 dprint(FD_FILE, "fd close %s\n", f->file_name);
564 if (close(f->fd) < 0)
569 if (f->shadow_fd != -1) {
578 int file_lookup_open(struct fio_file *f, int flags)
580 struct fio_file *__f;
583 __f = lookup_file_hash(f->file_name);
585 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
589 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
597 f->fd = open(f->file_name, flags, 0600);
601 static int file_close_shadow_fds(struct thread_data *td)
607 for_each_file(td, f, i) {
608 if (f->shadow_fd == -1)
619 int generic_open_file(struct thread_data *td, struct fio_file *f)
625 dprint(FD_FILE, "fd open %s\n", f->file_name);
627 if (!strcmp(f->file_name, "-")) {
629 log_err("fio: can't read/write to stdin/out\n");
635 * move output logging to stderr, if we are writing to stdout
644 flags |= OS_O_DIRECT;
647 td_verror(td, EINVAL, "OS does not support atomic IO");
650 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
654 if (td->o.create_on_open && td->o.allow_create)
657 if (f->filetype != FIO_TYPE_FILE)
658 flags |= FIO_O_NOATIME;
665 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
669 f->fd = dup(STDOUT_FILENO);
671 from_hash = file_lookup_open(f, flags);
672 } else if (td_read(td)) {
673 if (f->filetype == FIO_TYPE_CHAR && !read_only)
679 f->fd = dup(STDIN_FILENO);
681 from_hash = file_lookup_open(f, flags);
682 } else if (td_trim(td)) {
683 assert(!td_rw(td)); /* should have matched above */
685 from_hash = file_lookup_open(f, flags);
689 char buf[FIO_VERROR_SIZE];
692 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
693 flags &= ~FIO_O_NOATIME;
696 if (__e == EMFILE && file_close_shadow_fds(td))
699 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
701 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
702 log_err("fio: looks like your file system does not " \
703 "support direct=1/buffered=0\n");
706 td_verror(td, __e, buf);
710 if (!from_hash && f->fd != -1) {
711 if (add_file_hash(f)) {
715 * Stash away descriptor for later close. This is to
716 * work-around a "feature" on Linux, where a close of
717 * an fd that has been opened for write will trigger
718 * udev to call blkid to check partitions, fs id, etc.
719 * That pollutes the device cache, which can slow down
720 * unbuffered accesses.
722 if (f->shadow_fd == -1)
723 f->shadow_fd = f->fd;
726 * OK to ignore, we haven't done anything
729 ret = generic_close_file(td, f);
739 * This function i.e. get_file_size() is the default .get_file_size
740 * implementation of majority of I/O engines.
742 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
744 return get_file_size(td, f);
748 * open/close all files, so that ->real_file_size gets set
750 static int get_file_sizes(struct thread_data *td)
756 for_each_file(td, f, i) {
757 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
760 if (td_io_get_file_size(td, f)) {
761 if (td->error != ENOENT) {
762 log_err("%s\n", td->verror);
770 * There are corner cases where we end up with -1 for
771 * ->real_file_size due to unsupported file type, etc.
772 * We then just set to size option value divided by number
773 * of files, similar to the way file ->io_size is set.
774 * stat(2) failure doesn't set ->real_file_size to -1.
776 if (f->real_file_size == -1ULL && td->o.size)
777 f->real_file_size = td->o.size / td->o.nr_files;
784 struct flist_head list;
791 * Get free number of bytes for each file on each unique mount.
793 static unsigned long long get_fs_free_counts(struct thread_data *td)
795 struct flist_head *n, *tmp;
796 unsigned long long ret = 0;
797 struct fio_mount *fm;
802 for_each_file(td, f, i) {
806 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
807 if (f->real_file_size != -1ULL)
808 ret += f->real_file_size;
810 } else if (f->filetype != FIO_TYPE_FILE)
814 strncpy(buf, f->file_name, 255);
816 if (stat(buf, &sb) < 0) {
820 if (stat(buf, &sb) < 0)
825 flist_for_each(n, &list) {
826 fm = flist_entry(n, struct fio_mount, list);
827 if (fm->key == sb.st_dev)
836 fm = calloc(1, sizeof(*fm));
837 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
838 fm->base = basename(fm->__base);
840 flist_add(&fm->list, &list);
843 flist_for_each_safe(n, tmp, &list) {
844 unsigned long long sz;
846 fm = flist_entry(n, struct fio_mount, list);
847 flist_del(&fm->list);
849 sz = get_fs_free_size(fm->base);
850 if (sz && sz != -1ULL)
859 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
861 struct thread_options *o = &td->o;
862 unsigned long long align_bs;
863 unsigned long long offset;
865 if (o->file_append && f->filetype == FIO_TYPE_FILE)
866 return f->real_file_size;
868 if (o->start_offset_percent > 0) {
870 * if offset_align is provided, set initial offset
872 if (fio_option_is_set(o, start_offset_align)) {
873 align_bs = o->start_offset_align;
875 /* else take the minimum block size */
876 align_bs = td_min_bs(td);
879 /* calculate the raw offset */
880 offset = (f->real_file_size * o->start_offset_percent / 100) +
881 (td->subjob_number * o->offset_increment);
884 * block align the offset at the next available boundary at
885 * ceiling(offset / align_bs) * align_bs
887 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
890 /* start_offset_percent not set */
891 offset = o->start_offset +
892 td->subjob_number * o->offset_increment;
899 * Open the files and setup files sizes, creating files if necessary.
901 int setup_files(struct thread_data *td)
903 unsigned long long total_size, extend_size;
904 struct thread_options *o = &td->o;
906 unsigned int i, nr_fs_extra = 0;
907 int err = 0, need_extend;
909 const unsigned int bs = td_min_bs(td);
912 dprint(FD_FILE, "setup files\n");
914 old_state = td_bump_runstate(td, TD_SETTING_UP);
916 if (o->read_iolog_file)
920 * Find out physical size of files or devices for this thread,
921 * before we determine I/O size and range of our targets.
922 * If ioengine defines a setup() method, it's responsible for
923 * opening the files and setting f->real_file_size to indicate
924 * the valid range for that file.
926 if (td->io_ops->setup)
927 err = td->io_ops->setup(td);
929 err = get_file_sizes(td);
935 * check sizes. if the files/devices do not exist and the size
936 * isn't passed to fio, abort.
939 for_each_file(td, f, i) {
941 if (f->real_file_size == -1ULL)
944 total_size += f->real_file_size;
948 td->fill_device_size = get_fs_free_counts(td);
951 * device/file sizes are zero and no size given, punt
953 if ((!total_size || total_size == -1ULL) && !o->size &&
954 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
955 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
956 log_err("%s: you need to specify size=\n", o->name);
957 td_verror(td, EINVAL, "total_file_size");
962 * Calculate per-file size and potential extra size for the
963 * first files, if needed (i.e. if we don't have a fixed size).
965 if (!o->file_size_low && o->nr_files) {
968 fs = o->size / o->nr_files;
969 all_fs = fs * o->nr_files;
971 if (all_fs < o->size)
972 nr_fs_extra = (o->size - all_fs) / bs;
976 * now file sizes are known, so we can set ->io_size. if size= is
977 * not given, ->io_size is just equal to ->real_file_size. if size
978 * is given, ->io_size is size / nr_files.
980 extend_size = total_size = 0;
982 for_each_file(td, f, i) {
983 f->file_offset = get_start_offset(td, f);
986 * Update ->io_size depending on options specified.
987 * ->file_size_low being 0 means filesize option isn't set.
988 * Non zero ->file_size_low equals ->file_size_high means
989 * filesize option is set in a fixed size format.
990 * Non zero ->file_size_low not equals ->file_size_high means
991 * filesize option is set in a range format.
993 if (!o->file_size_low) {
995 * no file size or range given, file size is equal to
996 * total size divided by number of files. If the size
997 * doesn't divide nicely with the min blocksize,
998 * make the first files bigger.
1007 * We normally don't come here for regular files, but
1008 * if the result is 0 for a regular file, set it to the
1009 * real file size. This could be size of the existing
1010 * one if it already exists, but otherwise will be set
1011 * to 0. A new file won't be created because
1012 * ->io_size + ->file_offset equals ->real_file_size.
1015 if (f->file_offset > f->real_file_size)
1017 f->io_size = f->real_file_size - f->file_offset;
1019 log_info("fio: file %s may be ignored\n",
1022 } else if (f->real_file_size < o->file_size_low ||
1023 f->real_file_size > o->file_size_high) {
1024 if (f->file_offset > o->file_size_low)
1027 * file size given. if it's fixed, use that. if it's a
1028 * range, generate a random size in-between.
1030 if (o->file_size_low == o->file_size_high)
1031 f->io_size = o->file_size_low - f->file_offset;
1033 f->io_size = get_rand_file_size(td)
1037 f->io_size = f->real_file_size - f->file_offset;
1039 if (f->io_size == -1ULL)
1042 if (o->size_percent) {
1045 file_size = f->io_size + f->file_offset;
1046 f->io_size = (file_size *
1047 o->size_percent) / 100;
1048 if (f->io_size > (file_size - f->file_offset))
1049 f->io_size = file_size - f->file_offset;
1051 f->io_size -= (f->io_size % td_min_bs(td));
1053 total_size += f->io_size;
1056 if (f->filetype == FIO_TYPE_FILE &&
1057 (f->io_size + f->file_offset) > f->real_file_size &&
1058 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1059 if (!o->create_on_open) {
1061 extend_size += (f->io_size + f->file_offset);
1062 fio_file_set_extend(f);
1064 f->real_file_size = f->io_size + f->file_offset;
1068 if (td->o.block_error_hist) {
1071 assert(td->o.nr_files == 1); /* checked in fixup_options */
1073 len = f->io_size / td->o.bs[DDIR_TRIM];
1074 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1075 log_err("fio: cannot calculate block histogram with "
1076 "%d trim blocks, maximum %d\n",
1077 len, MAX_NR_BLOCK_INFOS);
1078 td_verror(td, EINVAL, "block_error_hist");
1082 td->ts.nr_block_infos = len;
1083 for (i = 0; i < len; i++)
1084 td->ts.block_infos[i] =
1085 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1087 td->ts.nr_block_infos = 0;
1089 if (!o->size || (total_size && o->size > total_size))
1090 o->size = total_size;
1092 if (o->size < td_min_bs(td)) {
1093 log_err("fio: blocksize too large for data set\n");
1098 * See if we need to extend some files, typically needed when our
1099 * target regular files don't exist yet, but our jobs require them
1100 * initially due to read I/Os.
1104 if (output_format & FIO_OUTPUT_NORMAL) {
1105 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1107 need_extend > 1 ? "s" : "",
1109 need_extend > 1 ? "s" : "",
1110 need_extend > 1 ? "total " : "",
1114 for_each_file(td, f, i) {
1115 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1117 if (!fio_file_extend(f))
1120 assert(f->filetype == FIO_TYPE_FILE);
1121 fio_file_clear_extend(f);
1122 if (!o->fill_device) {
1123 old_len = f->real_file_size;
1124 extend_len = f->io_size + f->file_offset -
1127 f->real_file_size = (f->io_size + f->file_offset);
1128 err = extend_file(td, f);
1132 err = __file_invalidate_cache(td, f, old_len,
1136 * Shut up static checker
1152 o->zone_size = o->size;
1155 * iolog already set the total io size, if we read back
1158 if (!o->read_iolog_file) {
1160 td->total_io_size = o->io_size * o->loops;
1162 td->total_io_size = o->size * o->loops;
1169 td_restore_runstate(td, old_state);
1172 log_err("%s: you need to specify valid offset=\n", o->name);
1174 td_restore_runstate(td, old_state);
1178 bool pre_read_files(struct thread_data *td)
1183 dprint(FD_FILE, "pre_read files\n");
1185 for_each_file(td, f, i) {
1186 if (!pre_read_file(td, f))
1193 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1195 unsigned int range_size, seed;
1196 unsigned long nranges;
1199 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1200 fsize = min(f->real_file_size, f->io_size);
1202 nranges = (fsize + range_size - 1) / range_size;
1204 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1205 if (!td->o.rand_repeatable)
1206 seed = td->rand_seeds[4];
1208 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1209 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1210 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1211 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1212 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1213 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1216 static bool init_rand_distribution(struct thread_data *td)
1222 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1225 state = td_bump_runstate(td, TD_SETTING_UP);
1227 for_each_file(td, f, i)
1228 __init_rand_distribution(td, f);
1230 td_restore_runstate(td, state);
1235 * Check if the number of blocks exceeds the randomness capability of
1236 * the selected generator. Tausworthe is 32-bit, the others are fullly
1239 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1242 if (blocks <= FRAND32_MAX)
1244 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1248 * If the user hasn't specified a random generator, switch
1249 * to tausworthe64 with informational warning. If the user did
1250 * specify one, just warn.
1252 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1255 if (!fio_option_is_set(&td->o, random_generator)) {
1256 log_info("fio: Switching to tausworthe64. Use the "
1257 "random_generator= option to get rid of this "
1259 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1264 * Just make this information to avoid breaking scripts.
1266 log_info("fio: Use the random_generator= option to switch to lfsr or "
1271 bool init_random_map(struct thread_data *td)
1273 unsigned long long blocks;
1277 if (init_rand_distribution(td))
1282 for_each_file(td, f, i) {
1283 uint64_t fsize = min(f->real_file_size, f->io_size);
1285 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1287 if (check_rand_gen_limits(td, f, blocks))
1290 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1293 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1295 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1296 fio_file_set_lfsr(f);
1299 } else if (!td->o.norandommap) {
1300 f->io_axmap = axmap_new(blocks);
1302 fio_file_set_axmap(f);
1305 } else if (td->o.norandommap)
1308 if (!td->o.softrandommap) {
1309 log_err("fio: failed allocating random map. If running"
1310 " a large number of jobs, try the 'norandommap'"
1311 " option or set 'softrandommap'. Or give"
1312 " a larger --alloc-size to fio.\n");
1316 log_info("fio: file %s failed allocating random map. Running "
1317 "job without.\n", f->file_name);
1323 void close_files(struct thread_data *td)
1328 for_each_file(td, f, i) {
1329 if (fio_file_open(f))
1330 td_io_close_file(td, f);
1334 void close_and_free_files(struct thread_data *td)
1338 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1340 dprint(FD_FILE, "close files\n");
1342 for_each_file(td, f, i) {
1343 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1344 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1345 td_io_unlink_file(td, f);
1348 if (fio_file_open(f))
1349 td_io_close_file(td, f);
1351 remove_file_hash(f);
1353 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1354 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1355 td_io_unlink_file(td, f);
1361 sfree(f->file_name);
1362 f->file_name = NULL;
1363 if (fio_file_axmap(f)) {
1364 axmap_free(f->io_axmap);
1373 td->o.filename = NULL;
1375 free(td->file_locks);
1376 td->files_index = 0;
1378 td->file_locks = NULL;
1379 td->o.file_lock_mode = FILE_LOCK_NONE;
1383 static void get_file_type(struct fio_file *f)
1387 if (!strcmp(f->file_name, "-"))
1388 f->filetype = FIO_TYPE_PIPE;
1390 f->filetype = FIO_TYPE_FILE;
1393 /* \\.\ is the device namespace in Windows, where every file is
1395 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1396 f->filetype = FIO_TYPE_BLOCK;
1399 if (!stat(f->file_name, &sb)) {
1400 if (S_ISBLK(sb.st_mode))
1401 f->filetype = FIO_TYPE_BLOCK;
1402 else if (S_ISCHR(sb.st_mode))
1403 f->filetype = FIO_TYPE_CHAR;
1404 else if (S_ISFIFO(sb.st_mode))
1405 f->filetype = FIO_TYPE_PIPE;
1409 static bool __is_already_allocated(const char *fname, bool set)
1411 struct flist_head *entry;
1414 ret = file_bloom_exists(fname, set);
1418 flist_for_each(entry, &filename_list) {
1419 struct file_name *fn;
1421 fn = flist_entry(entry, struct file_name, list);
1423 if (!strcmp(fn->filename, fname))
1430 static bool is_already_allocated(const char *fname)
1434 fio_file_hash_lock();
1435 ret = __is_already_allocated(fname, false);
1436 fio_file_hash_unlock();
1441 static void set_already_allocated(const char *fname)
1443 struct file_name *fn;
1445 fn = malloc(sizeof(struct file_name));
1446 fn->filename = strdup(fname);
1448 fio_file_hash_lock();
1449 if (!__is_already_allocated(fname, true)) {
1450 flist_add_tail(&fn->list, &filename_list);
1453 fio_file_hash_unlock();
1461 static void free_already_allocated(void)
1463 struct flist_head *entry, *tmp;
1464 struct file_name *fn;
1466 if (flist_empty(&filename_list))
1469 fio_file_hash_lock();
1470 flist_for_each_safe(entry, tmp, &filename_list) {
1471 fn = flist_entry(entry, struct file_name, list);
1473 flist_del(&fn->list);
1477 fio_file_hash_unlock();
1480 static struct fio_file *alloc_new_file(struct thread_data *td)
1484 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1485 f = calloc(1, sizeof(*f));
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 static bool create_work_dirs(struct thread_data *td, const char *fname)
1521 char path[PATH_MAX];
1524 if (td->o.directory) {
1525 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1526 FIO_OS_PATH_SEPARATOR, fname);
1527 start = strstr(path, fname);
1529 snprintf(path, PATH_MAX, "%s", fname);
1534 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1539 #ifdef CONFIG_HAVE_MKDIR_TWO
1540 if (mkdir(path, 0600) && errno != EEXIST) {
1542 if (mkdir(path) && errno != EEXIST) {
1544 log_err("fio: failed to create dir (%s): %d\n",
1548 *end = FIO_OS_PATH_SEPARATOR;
1551 td->flags |= TD_F_DIRS_CREATED;
1555 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1557 int cur_files = td->files_index;
1558 char file_name[PATH_MAX];
1562 dprint(FD_FILE, "add file %s\n", fname);
1564 if (td->o.directory)
1565 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1566 td->o.unique_filename);
1568 sprintf(file_name + len, "%s", fname);
1570 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1571 !(td->flags & TD_F_DIRS_CREATED) &&
1572 !create_work_dirs(td, fname))
1575 /* clean cloned siblings using existing files */
1576 if (numjob && is_already_allocated(file_name) &&
1577 !exists_and_not_regfile(fname))
1580 f = alloc_new_file(td);
1582 if (td->files_size <= td->files_index) {
1583 unsigned int new_size = td->o.nr_files + 1;
1585 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1587 td->files = realloc(td->files, new_size * sizeof(f));
1588 if (td->files == NULL) {
1589 log_err("fio: realloc OOM\n");
1592 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1593 td->file_locks = realloc(td->file_locks, new_size);
1594 if (!td->file_locks) {
1595 log_err("fio: realloc OOM\n");
1598 td->file_locks[cur_files] = FILE_LOCK_NONE;
1600 td->files_size = new_size;
1602 td->files[cur_files] = f;
1603 f->fileno = cur_files;
1606 * init function, io engine may not be loaded yet
1608 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1609 f->real_file_size = -1ULL;
1611 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1612 f->file_name = strdup(file_name);
1614 f->file_name = smalloc_strdup(file_name);
1620 switch (td->o.file_lock_mode) {
1621 case FILE_LOCK_NONE:
1623 case FILE_LOCK_READWRITE:
1624 f->rwlock = fio_rwlock_init();
1626 case FILE_LOCK_EXCLUSIVE:
1627 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1630 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1635 if (f->filetype == FIO_TYPE_FILE)
1636 td->nr_normal_files++;
1638 if (td->o.numjobs > 1)
1639 set_already_allocated(file_name);
1644 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1650 int add_file_exclusive(struct thread_data *td, const char *fname)
1655 for_each_file(td, f, i) {
1656 if (!strcmp(f->file_name, fname))
1660 return add_file(td, fname, 0, 1);
1663 void get_file(struct fio_file *f)
1665 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1666 assert(fio_file_open(f));
1670 int put_file(struct thread_data *td, struct fio_file *f)
1672 int f_ret = 0, ret = 0;
1674 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1676 if (!fio_file_open(f)) {
1677 assert(f->fd == -1);
1681 assert(f->references);
1682 if (--f->references)
1685 if (should_fsync(td) && td->o.fsync_on_close) {
1686 f_ret = fsync(f->fd);
1691 if (td->io_ops->close_file)
1692 ret = td->io_ops->close_file(td, f);
1697 td->nr_open_files--;
1698 fio_file_clear_open(f);
1699 assert(f->fd == -1);
1703 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1705 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1708 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1709 if (ddir == DDIR_READ)
1710 fio_rwlock_read(f->rwlock);
1712 fio_rwlock_write(f->rwlock);
1713 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1714 fio_mutex_down(f->lock);
1716 td->file_locks[f->fileno] = td->o.file_lock_mode;
1719 void unlock_file(struct thread_data *td, struct fio_file *f)
1721 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1724 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1725 fio_rwlock_unlock(f->rwlock);
1726 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1727 fio_mutex_up(f->lock);
1729 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1732 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1734 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1736 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1740 static int recurse_dir(struct thread_data *td, const char *dirname)
1746 D = opendir(dirname);
1748 char buf[FIO_VERROR_SIZE];
1750 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1751 td_verror(td, errno, buf);
1755 while ((dir = readdir(D)) != NULL) {
1756 char full_path[PATH_MAX];
1759 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1762 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1764 if (lstat(full_path, &sb) == -1) {
1765 if (errno != ENOENT) {
1766 td_verror(td, errno, "stat");
1772 if (S_ISREG(sb.st_mode)) {
1773 add_file(td, full_path, 0, 1);
1776 if (!S_ISDIR(sb.st_mode))
1779 ret = recurse_dir(td, full_path);
1788 int add_dir_files(struct thread_data *td, const char *path)
1790 int ret = recurse_dir(td, path);
1793 log_info("fio: opendir added %d files\n", td->o.nr_files);
1798 void dup_files(struct thread_data *td, struct thread_data *org)
1803 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1808 td->files = malloc(org->files_index * sizeof(f));
1810 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1811 td->file_locks = malloc(org->files_index);
1813 for_each_file(org, f, i) {
1814 struct fio_file *__f;
1816 __f = alloc_new_file(td);
1819 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1820 __f->file_name = strdup(f->file_name);
1822 __f->file_name = smalloc_strdup(f->file_name);
1823 if (!__f->file_name)
1826 __f->filetype = f->filetype;
1829 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1830 __f->lock = f->lock;
1831 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1832 __f->rwlock = f->rwlock;
1839 * Returns the index that matches the filename, or -1 if not there
1841 int get_fileno(struct thread_data *td, const char *fname)
1846 for_each_file(td, f, i)
1847 if (!strcmp(f->file_name, fname))
1854 * For log usage, where we add/open/close files automatically
1856 void free_release_files(struct thread_data *td)
1860 td->o.open_files = 0;
1861 td->files_index = 0;
1862 td->nr_normal_files = 0;
1865 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1869 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1870 f->last_pos[i] = f->file_offset;
1871 f->last_start[i] = -1ULL;
1874 if (fio_file_axmap(f))
1875 axmap_reset(f->io_axmap);
1876 else if (fio_file_lfsr(f))
1877 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1880 bool fio_files_done(struct thread_data *td)
1885 for_each_file(td, f, i)
1886 if (!fio_file_done(f))
1892 /* free memory used in initialization phase only */
1893 void filesetup_mem_free(void)
1895 free_already_allocated();
1899 * This function is for platforms which support direct I/O but not O_DIRECT.
1901 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1903 #ifdef FIO_OS_DIRECTIO
1904 int ret = fio_set_odirect(f);
1907 td_verror(td, ret, "fio_set_directio");
1908 #if defined(__sun__)
1909 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1910 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1912 log_err("fio: the file system does not seem to support direct IO\n");
1915 log_err("fio: the file system does not seem to support direct IO\n");
1922 log_err("fio: direct IO is not supported on this host operating system\n");