17 #include "lib/axmap.h"
18 #include "lib/memalign.h"
20 #ifdef CONFIG_LINUX_FALLOCATE
21 #include <linux/falloc.h>
26 static FLIST_HEAD(filename_list);
29 * List entry for filename_list
32 struct flist_head list;
36 static inline void clear_error(struct thread_data *td)
42 static inline int native_fallocate(struct thread_data *td, struct fio_file *f)
46 success = fio_fallocate(f, 0, f->real_file_size);
47 dprint(FD_FILE, "native fallocate of file %s size %llu was "
48 "%ssuccessful\n", f->file_name,
49 (unsigned long long) f->real_file_size,
56 dprint(FD_FILE, "native fallocate is not implemented\n");
61 static void fallocate_file(struct thread_data *td, struct fio_file *f)
65 if (td->o.fill_device)
68 switch (td->o.fallocate_mode) {
69 case FIO_FALLOCATE_NATIVE:
70 r = native_fallocate(td, f);
71 if (r != 0 && errno != ENOSYS)
72 log_err("fio: native_fallocate call failed: %s\n",
75 case FIO_FALLOCATE_NONE:
77 #ifdef CONFIG_POSIX_FALLOCATE
78 case FIO_FALLOCATE_POSIX:
79 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
81 (unsigned long long) f->real_file_size);
83 r = posix_fallocate(f->fd, 0, f->real_file_size);
85 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
87 #endif /* CONFIG_POSIX_FALLOCATE */
88 #ifdef CONFIG_LINUX_FALLOCATE
89 case FIO_FALLOCATE_KEEP_SIZE:
90 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
91 "file %s size %llu\n", f->file_name,
92 (unsigned long long) f->real_file_size);
94 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
96 td_verror(td, errno, "fallocate");
99 #endif /* CONFIG_LINUX_FALLOCATE */
101 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;
113 unsigned int bs, alloc_size = 0;
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)
151 flags |= OS_O_DIRECT;
157 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
158 f->fd = open(f->file_name, flags, 0644);
162 if (err == ENOENT && !td->o.allow_create)
163 log_err("fio: file creation disallowed by "
164 "allow_file_create=0\n");
166 if (err == EINVAL && (flags & OS_O_DIRECT))
167 log_err("fio: looks like your filesystem "
169 "direct=1/buffered=0\n");
171 td_verror(td, err, "open");
176 fallocate_file(td, f);
179 * If our jobs don't require regular files initially, we're done.
185 * The size will be -1ULL when fill_device is used, so don't truncate
186 * or fallocate this file, just write it
188 if (!td->o.fill_device) {
189 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
190 (unsigned long long) f->real_file_size);
191 if (ftruncate(f->fd, f->real_file_size) == -1) {
192 if (errno != EFBIG) {
193 td_verror(td, errno, "ftruncate");
199 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
202 left = f->real_file_size;
203 bs = td->o.max_bs[DDIR_WRITE];
208 b = fio_memalign(page_size, alloc_size);
210 td_verror(td, errno, "fio_memalign");
214 while (left && !td->terminate) {
220 fill_io_buffer(td, b, bs, bs);
222 r = write(f->fd, b, bs);
232 if (td->o.fill_device)
234 log_info("fio: ENOSPC on laying out "
238 td_verror(td, errno, "write");
240 td_verror(td, EIO, "write");
247 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
248 td_io_unlink_file(td, f);
249 } else if (td->o.create_fsync) {
250 if (fsync(f->fd) < 0) {
251 td_verror(td, errno, "fsync");
255 if (td->o.fill_device && !td_write(td)) {
256 fio_file_clear_size_known(f);
257 if (td_io_get_file_size(td, f))
259 if (f->io_size > f->real_file_size)
260 f->io_size = f->real_file_size;
263 fio_memfree(b, alloc_size);
270 fio_memfree(b, alloc_size);
274 static int pre_read_file(struct thread_data *td, struct fio_file *f)
276 int ret = 0, r, did_open = 0, old_runstate;
277 unsigned long long left;
281 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
282 td_ioengine_flagged(td, FIO_NOIO))
285 if (f->filetype == FIO_TYPE_CHAR)
288 if (!fio_file_open(f)) {
289 if (td->io_ops->open_file(td, f)) {
290 log_err("fio: cannot pre-read, failed to open file\n");
296 old_runstate = td_bump_runstate(td, TD_PRE_READING);
299 bs = td->o.max_bs[DDIR_READ];
305 td_verror(td, errno, "malloc");
311 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
312 td_verror(td, errno, "lseek");
313 log_err("fio: failed to lseek pre-read file\n");
318 while (left && !td->terminate) {
322 r = read(f->fd, b, bs);
328 td_verror(td, EIO, "pre_read");
334 td_restore_runstate(td, old_runstate);
337 td->io_ops->close_file(td, f);
343 unsigned long long get_rand_file_size(struct thread_data *td)
345 unsigned long long ret, sized;
349 frand_max = rand_max(&td->file_size_state);
350 r = __rand(&td->file_size_state);
351 sized = td->o.file_size_high - td->o.file_size_low;
352 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
353 ret += td->o.file_size_low;
354 ret -= (ret % td->o.rw_min_bs);
358 static int file_size(struct thread_data *td, struct fio_file *f)
362 if (stat(f->file_name, &st) == -1) {
363 td_verror(td, errno, "fstat");
367 f->real_file_size = st.st_size;
371 static int bdev_size(struct thread_data *td, struct fio_file *f)
373 unsigned long long bytes = 0;
376 if (td->io_ops->open_file(td, f)) {
377 log_err("fio: failed opening blockdev %s for size check\n",
382 r = blockdev_size(f, &bytes);
384 td_verror(td, r, "blockdev_size");
389 log_err("%s: zero sized block device?\n", f->file_name);
393 f->real_file_size = bytes;
394 td->io_ops->close_file(td, f);
397 td->io_ops->close_file(td, f);
401 static int char_size(struct thread_data *td, struct fio_file *f)
403 #ifdef FIO_HAVE_CHARDEV_SIZE
404 unsigned long long bytes = 0;
407 if (td->io_ops->open_file(td, f)) {
408 log_err("fio: failed opening chardev %s for size check\n",
413 r = chardev_size(f, &bytes);
415 td_verror(td, r, "chardev_size");
420 log_err("%s: zero sized char device?\n", f->file_name);
424 f->real_file_size = bytes;
425 td->io_ops->close_file(td, f);
428 td->io_ops->close_file(td, f);
431 f->real_file_size = -1ULL;
436 static int get_file_size(struct thread_data *td, struct fio_file *f)
440 if (fio_file_size_known(f))
443 if (f->filetype == FIO_TYPE_FILE)
444 ret = file_size(td, f);
445 else if (f->filetype == FIO_TYPE_BLOCK)
446 ret = bdev_size(td, f);
447 else if (f->filetype == FIO_TYPE_CHAR)
448 ret = char_size(td, f);
450 f->real_file_size = -1ULL;
453 * Leave ->real_file_size with 0 since it could be expectation
454 * of initial setup for regular files.
460 * If ->real_file_size is -1, a conditional for the message
461 * "offset extends end" is always true, but it makes no sense,
462 * so just return the same value here.
464 if (f->real_file_size == -1ULL) {
465 log_info("%s: failed to get file size of %s\n", td->o.name,
470 if (td->o.start_offset && f->file_offset == 0)
471 dprint(FD_FILE, "offset of file %s not initialized yet\n",
474 * ->file_offset normally hasn't been initialized yet, so this
475 * is basically always false.
477 if (f->file_offset > f->real_file_size) {
478 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
479 (unsigned long long) f->file_offset,
480 (unsigned long long) f->real_file_size);
484 fio_file_set_size_known(f);
488 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
489 unsigned long long off,
490 unsigned long long len)
492 int errval = 0, ret = 0;
501 off = f->file_offset;
503 if (len == -1ULL || off == -1ULL)
506 if (td->io_ops->invalidate) {
507 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
509 ret = td->io_ops->invalidate(td, f);
512 } else if (f->filetype == FIO_TYPE_FILE) {
513 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
514 f->file_name, off, len);
515 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
518 } else if (f->filetype == FIO_TYPE_BLOCK) {
521 dprint(FD_IO, "drop page cache %s\n", f->file_name);
522 ret = blockdev_invalidate_cache(f);
523 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
525 * Linux multipath devices reject ioctl while
526 * the maps are being updated. That window can
527 * last tens of milliseconds; we'll try up to
528 * a quarter of a second.
531 ret = blockdev_invalidate_cache(f);
533 if (ret < 0 && errno == EACCES && geteuid()) {
535 log_err("fio: only root may flush block "
536 "devices. Cache flush bypassed!\n");
543 } else if (f->filetype == FIO_TYPE_CHAR ||
544 f->filetype == FIO_TYPE_PIPE) {
545 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
550 * Cache flushing isn't a fatal condition, and we know it will
551 * happen on some platforms where we don't have the proper
552 * function to flush eg block device caches. So just warn and
553 * continue on our way.
556 log_info("fio: cache invalidation of %s failed: %s\n",
557 f->file_name, strerror(errval));
563 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
565 if (!fio_file_open(f))
568 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
571 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
575 dprint(FD_FILE, "fd close %s\n", f->file_name);
579 if (close(f->fd) < 0)
584 if (f->shadow_fd != -1) {
593 int file_lookup_open(struct fio_file *f, int flags)
595 struct fio_file *__f;
598 __f = lookup_file_hash(f->file_name);
600 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
604 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
612 f->fd = open(f->file_name, flags, 0600);
616 static int file_close_shadow_fds(struct thread_data *td)
622 for_each_file(td, f, i) {
623 if (f->shadow_fd == -1)
634 int generic_open_file(struct thread_data *td, struct fio_file *f)
640 dprint(FD_FILE, "fd open %s\n", f->file_name);
642 if (!strcmp(f->file_name, "-")) {
644 log_err("fio: can't read/write to stdin/out\n");
650 * move output logging to stderr, if we are writing to stdout
659 flags |= OS_O_DIRECT;
662 td_verror(td, EINVAL, "OS does not support atomic IO");
665 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
669 if (td->o.create_on_open && td->o.allow_create)
672 if (f->filetype != FIO_TYPE_FILE)
673 flags |= FIO_O_NOATIME;
680 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
684 f->fd = dup(STDOUT_FILENO);
686 from_hash = file_lookup_open(f, flags);
687 } else if (td_read(td)) {
688 if (f->filetype == FIO_TYPE_CHAR && !read_only)
694 f->fd = dup(STDIN_FILENO);
696 from_hash = file_lookup_open(f, flags);
697 } else if (td_trim(td)) {
698 assert(!td_rw(td)); /* should have matched above */
700 from_hash = file_lookup_open(f, flags);
704 char buf[FIO_VERROR_SIZE];
707 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
708 flags &= ~FIO_O_NOATIME;
711 if (__e == EMFILE && file_close_shadow_fds(td))
714 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
716 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
717 log_err("fio: looks like your file system does not " \
718 "support direct=1/buffered=0\n");
721 td_verror(td, __e, buf);
725 if (!from_hash && f->fd != -1) {
726 if (add_file_hash(f)) {
730 * Stash away descriptor for later close. This is to
731 * work-around a "feature" on Linux, where a close of
732 * an fd that has been opened for write will trigger
733 * udev to call blkid to check partitions, fs id, etc.
734 * That pollutes the device cache, which can slow down
735 * unbuffered accesses.
737 if (f->shadow_fd == -1)
738 f->shadow_fd = f->fd;
741 * OK to ignore, we haven't done anything
744 ret = generic_close_file(td, f);
754 * This function i.e. get_file_size() is the default .get_file_size
755 * implementation of majority of I/O engines.
757 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
759 return get_file_size(td, f);
763 * open/close all files, so that ->real_file_size gets set
765 static int get_file_sizes(struct thread_data *td)
771 for_each_file(td, f, i) {
772 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
775 if (td_io_get_file_size(td, f)) {
776 if (td->error != ENOENT) {
777 log_err("%s\n", td->verror);
785 * There are corner cases where we end up with -1 for
786 * ->real_file_size due to unsupported file type, etc.
787 * We then just set to size option value divided by number
788 * of files, similar to the way file ->io_size is set.
789 * stat(2) failure doesn't set ->real_file_size to -1.
791 if (f->real_file_size == -1ULL && td->o.size)
792 f->real_file_size = td->o.size / td->o.nr_files;
799 struct flist_head list;
806 * Get free number of bytes for each file on each unique mount.
808 static unsigned long long get_fs_free_counts(struct thread_data *td)
810 struct flist_head *n, *tmp;
811 unsigned long long ret = 0;
812 struct fio_mount *fm;
817 for_each_file(td, f, i) {
821 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
822 if (f->real_file_size != -1ULL)
823 ret += f->real_file_size;
825 } else if (f->filetype != FIO_TYPE_FILE)
829 strncpy(buf, f->file_name, 255);
831 if (stat(buf, &sb) < 0) {
835 if (stat(buf, &sb) < 0)
840 flist_for_each(n, &list) {
841 fm = flist_entry(n, struct fio_mount, list);
842 if (fm->key == sb.st_dev)
851 fm = calloc(1, sizeof(*fm));
852 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
853 fm->base = basename(fm->__base);
855 flist_add(&fm->list, &list);
858 flist_for_each_safe(n, tmp, &list) {
859 unsigned long long sz;
861 fm = flist_entry(n, struct fio_mount, list);
862 flist_del(&fm->list);
864 sz = get_fs_free_size(fm->base);
865 if (sz && sz != -1ULL)
874 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
876 struct thread_options *o = &td->o;
877 unsigned long long align_bs;
878 unsigned long long offset;
880 if (o->file_append && f->filetype == FIO_TYPE_FILE)
881 return f->real_file_size;
883 if (o->start_offset_percent > 0) {
885 * if blockalign is provided, find the min across read, write,
888 if (fio_option_is_set(o, ba)) {
889 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
890 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
892 /* else take the minimum block size */
893 align_bs = td_min_bs(td);
896 /* calculate the raw offset */
897 offset = (f->real_file_size * o->start_offset_percent / 100) +
898 (td->subjob_number * o->offset_increment);
901 * block align the offset at the next available boundary at
902 * ceiling(offset / align_bs) * align_bs
904 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
907 /* start_offset_percent not set */
908 offset = o->start_offset +
909 td->subjob_number * o->offset_increment;
916 * Open the files and setup files sizes, creating files if necessary.
918 int setup_files(struct thread_data *td)
920 unsigned long long total_size, extend_size;
921 struct thread_options *o = &td->o;
923 unsigned int i, nr_fs_extra = 0;
924 int err = 0, need_extend;
926 const unsigned int bs = td_min_bs(td);
929 dprint(FD_FILE, "setup files\n");
931 old_state = td_bump_runstate(td, TD_SETTING_UP);
933 if (o->read_iolog_file)
937 * Find out physical size of files or devices for this thread,
938 * before we determine I/O size and range of our targets.
939 * If ioengine defines a setup() method, it's responsible for
940 * opening the files and setting f->real_file_size to indicate
941 * the valid range for that file.
943 if (td->io_ops->setup)
944 err = td->io_ops->setup(td);
946 err = get_file_sizes(td);
952 * check sizes. if the files/devices do not exist and the size
953 * isn't passed to fio, abort.
956 for_each_file(td, f, i) {
958 if (f->real_file_size == -1ULL)
961 total_size += f->real_file_size;
965 td->fill_device_size = get_fs_free_counts(td);
968 * device/file sizes are zero and no size given, punt
970 if ((!total_size || total_size == -1ULL) && !o->size &&
971 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
972 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
973 log_err("%s: you need to specify size=\n", o->name);
974 td_verror(td, EINVAL, "total_file_size");
979 * Calculate per-file size and potential extra size for the
980 * first files, if needed (i.e. if we don't have a fixed size).
982 if (!o->file_size_low && o->nr_files) {
985 fs = o->size / o->nr_files;
986 all_fs = fs * o->nr_files;
988 if (all_fs < o->size)
989 nr_fs_extra = (o->size - all_fs) / bs;
993 * now file sizes are known, so we can set ->io_size. if size= is
994 * not given, ->io_size is just equal to ->real_file_size. if size
995 * is given, ->io_size is size / nr_files.
997 extend_size = total_size = 0;
999 for_each_file(td, f, i) {
1000 f->file_offset = get_start_offset(td, f);
1003 * Update ->io_size depending on options specified.
1004 * ->file_size_low being 0 means filesize option isn't set.
1005 * Non zero ->file_size_low equals ->file_size_high means
1006 * filesize option is set in a fixed size format.
1007 * Non zero ->file_size_low not equals ->file_size_high means
1008 * filesize option is set in a range format.
1010 if (!o->file_size_low) {
1012 * no file size or range given, file size is equal to
1013 * total size divided by number of files. If the size
1014 * doesn't divide nicely with the min blocksize,
1015 * make the first files bigger.
1024 * We normally don't come here for regular files, but
1025 * if the result is 0 for a regular file, set it to the
1026 * real file size. This could be size of the existing
1027 * one if it already exists, but otherwise will be set
1028 * to 0. A new file won't be created because
1029 * ->io_size + ->file_offset equals ->real_file_size.
1032 if (f->file_offset > f->real_file_size)
1034 f->io_size = f->real_file_size - f->file_offset;
1036 log_info("fio: file %s may be ignored\n",
1039 } else if (f->real_file_size < o->file_size_low ||
1040 f->real_file_size > o->file_size_high) {
1041 if (f->file_offset > o->file_size_low)
1044 * file size given. if it's fixed, use that. if it's a
1045 * range, generate a random size in-between.
1047 if (o->file_size_low == o->file_size_high)
1048 f->io_size = o->file_size_low - f->file_offset;
1050 f->io_size = get_rand_file_size(td)
1054 f->io_size = f->real_file_size - f->file_offset;
1056 if (f->io_size == -1ULL)
1059 if (o->size_percent) {
1062 file_size = f->io_size + f->file_offset;
1063 f->io_size = (file_size *
1064 o->size_percent) / 100;
1065 if (f->io_size > (file_size - f->file_offset))
1066 f->io_size = file_size - f->file_offset;
1068 f->io_size -= (f->io_size % td_min_bs(td));
1070 total_size += f->io_size;
1073 if (f->filetype == FIO_TYPE_FILE &&
1074 (f->io_size + f->file_offset) > f->real_file_size &&
1075 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1076 if (!o->create_on_open) {
1078 extend_size += (f->io_size + f->file_offset);
1079 fio_file_set_extend(f);
1081 f->real_file_size = f->io_size + f->file_offset;
1085 if (td->o.block_error_hist) {
1088 assert(td->o.nr_files == 1); /* checked in fixup_options */
1090 len = f->io_size / td->o.bs[DDIR_TRIM];
1091 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1092 log_err("fio: cannot calculate block histogram with "
1093 "%d trim blocks, maximum %d\n",
1094 len, MAX_NR_BLOCK_INFOS);
1095 td_verror(td, EINVAL, "block_error_hist");
1099 td->ts.nr_block_infos = len;
1100 for (i = 0; i < len; i++)
1101 td->ts.block_infos[i] =
1102 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1104 td->ts.nr_block_infos = 0;
1106 if (!o->size || (total_size && o->size > total_size))
1107 o->size = total_size;
1109 if (o->size < td_min_bs(td)) {
1110 log_err("fio: blocksize too large for data set\n");
1115 * See if we need to extend some files, typically needed when our
1116 * target regular files don't exist yet, but our jobs require them
1117 * initially due to read I/Os.
1121 if (output_format & FIO_OUTPUT_NORMAL) {
1122 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1124 need_extend > 1 ? "s" : "",
1126 need_extend > 1 ? "s" : "",
1127 need_extend > 1 ? "total " : "",
1131 for_each_file(td, f, i) {
1132 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1134 if (!fio_file_extend(f))
1137 assert(f->filetype == FIO_TYPE_FILE);
1138 fio_file_clear_extend(f);
1139 if (!o->fill_device) {
1140 old_len = f->real_file_size;
1141 extend_len = f->io_size + f->file_offset -
1144 f->real_file_size = (f->io_size + f->file_offset);
1145 err = extend_file(td, f);
1149 err = __file_invalidate_cache(td, f, old_len,
1153 * Shut up static checker
1169 o->zone_size = o->size;
1172 * iolog already set the total io size, if we read back
1175 if (!o->read_iolog_file) {
1177 td->total_io_size = o->io_size * o->loops;
1179 td->total_io_size = o->size * o->loops;
1186 td_restore_runstate(td, old_state);
1189 log_err("%s: you need to specify valid offset=\n", o->name);
1191 td_restore_runstate(td, old_state);
1195 int pre_read_files(struct thread_data *td)
1200 dprint(FD_FILE, "pre_read files\n");
1202 for_each_file(td, f, i) {
1203 if (pre_read_file(td, f))
1210 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1212 unsigned int range_size, seed;
1213 unsigned long nranges;
1216 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1217 fsize = min(f->real_file_size, f->io_size);
1219 nranges = (fsize + range_size - 1) / range_size;
1221 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1222 if (!td->o.rand_repeatable)
1223 seed = td->rand_seeds[4];
1225 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1226 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1227 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1228 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1229 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1230 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1235 static int init_rand_distribution(struct thread_data *td)
1241 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1244 state = td_bump_runstate(td, TD_SETTING_UP);
1246 for_each_file(td, f, i)
1247 __init_rand_distribution(td, f);
1249 td_restore_runstate(td, state);
1255 * Check if the number of blocks exceeds the randomness capability of
1256 * the selected generator. Tausworthe is 32-bit, the others are fullly
1259 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1262 if (blocks <= FRAND32_MAX)
1264 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1268 * If the user hasn't specified a random generator, switch
1269 * to tausworthe64 with informational warning. If the user did
1270 * specify one, just warn.
1272 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1275 if (!fio_option_is_set(&td->o, random_generator)) {
1276 log_info("fio: Switching to tausworthe64. Use the "
1277 "random_generator= option to get rid of this "
1279 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1284 * Just make this information to avoid breaking scripts.
1286 log_info("fio: Use the random_generator= option to switch to lfsr or "
1291 int init_random_map(struct thread_data *td)
1293 unsigned long long blocks;
1297 if (init_rand_distribution(td))
1302 for_each_file(td, f, i) {
1303 uint64_t fsize = min(f->real_file_size, f->io_size);
1305 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1307 if (check_rand_gen_limits(td, f, blocks))
1310 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1313 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1315 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1316 fio_file_set_lfsr(f);
1319 } else if (!td->o.norandommap) {
1320 f->io_axmap = axmap_new(blocks);
1322 fio_file_set_axmap(f);
1325 } else if (td->o.norandommap)
1328 if (!td->o.softrandommap) {
1329 log_err("fio: failed allocating random map. If running"
1330 " a large number of jobs, try the 'norandommap'"
1331 " option or set 'softrandommap'. Or give"
1332 " a larger --alloc-size to fio.\n");
1336 log_info("fio: file %s failed allocating random map. Running "
1337 "job without.\n", f->file_name);
1343 void close_files(struct thread_data *td)
1348 for_each_file(td, f, i) {
1349 if (fio_file_open(f))
1350 td_io_close_file(td, f);
1354 void close_and_free_files(struct thread_data *td)
1359 dprint(FD_FILE, "close files\n");
1361 for_each_file(td, f, i) {
1362 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1363 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1364 td_io_unlink_file(td, f);
1367 if (fio_file_open(f))
1368 td_io_close_file(td, f);
1370 remove_file_hash(f);
1372 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1373 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1374 td_io_unlink_file(td, f);
1377 sfree(f->file_name);
1378 f->file_name = NULL;
1379 if (fio_file_axmap(f)) {
1380 axmap_free(f->io_axmap);
1386 td->o.filename = NULL;
1388 free(td->file_locks);
1389 td->files_index = 0;
1391 td->file_locks = NULL;
1392 td->o.file_lock_mode = FILE_LOCK_NONE;
1396 static void get_file_type(struct fio_file *f)
1400 if (!strcmp(f->file_name, "-"))
1401 f->filetype = FIO_TYPE_PIPE;
1403 f->filetype = FIO_TYPE_FILE;
1406 /* \\.\ is the device namespace in Windows, where every file is
1408 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1409 f->filetype = FIO_TYPE_BLOCK;
1412 if (!stat(f->file_name, &sb)) {
1413 if (S_ISBLK(sb.st_mode))
1414 f->filetype = FIO_TYPE_BLOCK;
1415 else if (S_ISCHR(sb.st_mode))
1416 f->filetype = FIO_TYPE_CHAR;
1417 else if (S_ISFIFO(sb.st_mode))
1418 f->filetype = FIO_TYPE_PIPE;
1422 static bool __is_already_allocated(const char *fname, bool set)
1424 struct flist_head *entry;
1427 ret = file_bloom_exists(fname, set);
1431 flist_for_each(entry, &filename_list) {
1432 struct file_name *fn;
1434 fn = flist_entry(entry, struct file_name, list);
1436 if (!strcmp(fn->filename, fname))
1443 static bool is_already_allocated(const char *fname)
1447 fio_file_hash_lock();
1448 ret = __is_already_allocated(fname, false);
1449 fio_file_hash_unlock();
1454 static void set_already_allocated(const char *fname)
1456 struct file_name *fn;
1458 fn = malloc(sizeof(struct file_name));
1459 fn->filename = strdup(fname);
1461 fio_file_hash_lock();
1462 if (!__is_already_allocated(fname, true)) {
1463 flist_add_tail(&fn->list, &filename_list);
1466 fio_file_hash_unlock();
1474 static void free_already_allocated(void)
1476 struct flist_head *entry, *tmp;
1477 struct file_name *fn;
1479 if (flist_empty(&filename_list))
1482 fio_file_hash_lock();
1483 flist_for_each_safe(entry, tmp, &filename_list) {
1484 fn = flist_entry(entry, struct file_name, list);
1486 flist_del(&fn->list);
1490 fio_file_hash_unlock();
1493 static struct fio_file *alloc_new_file(struct thread_data *td)
1497 f = smalloc(sizeof(*f));
1505 fio_file_reset(td, f);
1509 bool exists_and_not_regfile(const char *filename)
1513 if (lstat(filename, &sb) == -1)
1516 #ifndef WIN32 /* NOT Windows */
1517 if (S_ISREG(sb.st_mode))
1520 /* \\.\ is the device namespace in Windows, where every file
1521 * is a device node */
1522 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1529 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1531 int cur_files = td->files_index;
1532 char file_name[PATH_MAX];
1536 dprint(FD_FILE, "add file %s\n", fname);
1538 if (td->o.directory)
1539 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1540 td->o.unique_filename);
1542 sprintf(file_name + len, "%s", fname);
1544 /* clean cloned siblings using existing files */
1545 if (numjob && is_already_allocated(file_name) &&
1546 !exists_and_not_regfile(fname))
1549 f = alloc_new_file(td);
1551 if (td->files_size <= td->files_index) {
1552 unsigned int new_size = td->o.nr_files + 1;
1554 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1556 td->files = realloc(td->files, new_size * sizeof(f));
1557 if (td->files == NULL) {
1558 log_err("fio: realloc OOM\n");
1561 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1562 td->file_locks = realloc(td->file_locks, new_size);
1563 if (!td->file_locks) {
1564 log_err("fio: realloc OOM\n");
1567 td->file_locks[cur_files] = FILE_LOCK_NONE;
1569 td->files_size = new_size;
1571 td->files[cur_files] = f;
1572 f->fileno = cur_files;
1575 * init function, io engine may not be loaded yet
1577 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1578 f->real_file_size = -1ULL;
1580 f->file_name = smalloc_strdup(file_name);
1586 switch (td->o.file_lock_mode) {
1587 case FILE_LOCK_NONE:
1589 case FILE_LOCK_READWRITE:
1590 f->rwlock = fio_rwlock_init();
1592 case FILE_LOCK_EXCLUSIVE:
1593 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1596 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1601 if (f->filetype == FIO_TYPE_FILE)
1602 td->nr_normal_files++;
1604 set_already_allocated(file_name);
1609 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1615 int add_file_exclusive(struct thread_data *td, const char *fname)
1620 for_each_file(td, f, i) {
1621 if (!strcmp(f->file_name, fname))
1625 return add_file(td, fname, 0, 1);
1628 void get_file(struct fio_file *f)
1630 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1631 assert(fio_file_open(f));
1635 int put_file(struct thread_data *td, struct fio_file *f)
1637 int f_ret = 0, ret = 0;
1639 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1641 if (!fio_file_open(f)) {
1642 assert(f->fd == -1);
1646 assert(f->references);
1647 if (--f->references)
1650 if (should_fsync(td) && td->o.fsync_on_close) {
1651 f_ret = fsync(f->fd);
1656 if (td->io_ops->close_file)
1657 ret = td->io_ops->close_file(td, f);
1662 td->nr_open_files--;
1663 fio_file_clear_open(f);
1664 assert(f->fd == -1);
1668 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1670 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1673 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1674 if (ddir == DDIR_READ)
1675 fio_rwlock_read(f->rwlock);
1677 fio_rwlock_write(f->rwlock);
1678 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1679 fio_mutex_down(f->lock);
1681 td->file_locks[f->fileno] = td->o.file_lock_mode;
1684 void unlock_file(struct thread_data *td, struct fio_file *f)
1686 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1689 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1690 fio_rwlock_unlock(f->rwlock);
1691 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1692 fio_mutex_up(f->lock);
1694 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1697 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1699 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1701 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1705 static int recurse_dir(struct thread_data *td, const char *dirname)
1711 D = opendir(dirname);
1713 char buf[FIO_VERROR_SIZE];
1715 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1716 td_verror(td, errno, buf);
1720 while ((dir = readdir(D)) != NULL) {
1721 char full_path[PATH_MAX];
1724 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1727 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1729 if (lstat(full_path, &sb) == -1) {
1730 if (errno != ENOENT) {
1731 td_verror(td, errno, "stat");
1737 if (S_ISREG(sb.st_mode)) {
1738 add_file(td, full_path, 0, 1);
1741 if (!S_ISDIR(sb.st_mode))
1744 ret = recurse_dir(td, full_path);
1753 int add_dir_files(struct thread_data *td, const char *path)
1755 int ret = recurse_dir(td, path);
1758 log_info("fio: opendir added %d files\n", td->o.nr_files);
1763 void dup_files(struct thread_data *td, struct thread_data *org)
1768 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1773 td->files = malloc(org->files_index * sizeof(f));
1775 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1776 td->file_locks = malloc(org->files_index);
1778 for_each_file(org, f, i) {
1779 struct fio_file *__f;
1781 __f = alloc_new_file(td);
1784 __f->file_name = smalloc_strdup(f->file_name);
1785 if (!__f->file_name)
1788 __f->filetype = f->filetype;
1791 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1792 __f->lock = f->lock;
1793 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1794 __f->rwlock = f->rwlock;
1801 * Returns the index that matches the filename, or -1 if not there
1803 int get_fileno(struct thread_data *td, const char *fname)
1808 for_each_file(td, f, i)
1809 if (!strcmp(f->file_name, fname))
1816 * For log usage, where we add/open/close files automatically
1818 void free_release_files(struct thread_data *td)
1822 td->o.open_files = 0;
1823 td->files_index = 0;
1824 td->nr_normal_files = 0;
1827 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1831 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1832 f->last_pos[i] = f->file_offset;
1833 f->last_start[i] = -1ULL;
1836 if (fio_file_axmap(f))
1837 axmap_reset(f->io_axmap);
1838 else if (fio_file_lfsr(f))
1839 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1842 bool fio_files_done(struct thread_data *td)
1847 for_each_file(td, f, i)
1848 if (!fio_file_done(f))
1854 /* free memory used in initialization phase only */
1855 void filesetup_mem_free(void)
1857 free_already_allocated();
1861 * This function is for platforms which support direct I/O but not O_DIRECT.
1863 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1865 #ifdef FIO_OS_DIRECTIO
1866 int ret = fio_set_odirect(f);
1869 td_verror(td, ret, "fio_set_directio");
1870 #if defined(__sun__)
1871 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1872 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1874 log_err("fio: the file system does not seem to support direct IO\n");
1877 log_err("fio: the file system does not seem to support direct IO\n");