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)
42 * Leaves f->fd open on success, caller must close
44 static int extend_file(struct thread_data *td, struct fio_file *f)
46 int r, new_layout = 0, unlink_file = 0, flags;
47 unsigned long long left;
52 log_err("fio: refusing extend of file due to read-only\n");
57 * check if we need to lay the file out complete again. fio
58 * does that for operations involving reads, or for writes
59 * where overwrite is set
62 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
63 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
65 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
68 if (unlink_file || new_layout) {
71 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
73 ret = td_io_unlink_file(td, f);
74 if (ret != 0 && ret != ENOENT) {
75 td_verror(td, errno, "unlink");
81 if (td->o.allow_create)
90 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
91 f->fd = open(f->file_name, flags, 0644);
95 if (err == ENOENT && !td->o.allow_create)
96 log_err("fio: file creation disallowed by "
97 "allow_file_create=0\n");
99 td_verror(td, err, "open");
103 #ifdef CONFIG_POSIX_FALLOCATE
104 if (!td->o.fill_device) {
105 switch (td->o.fallocate_mode) {
106 case FIO_FALLOCATE_NONE:
108 case FIO_FALLOCATE_POSIX:
109 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
111 (unsigned long long) f->real_file_size);
113 r = posix_fallocate(f->fd, 0, f->real_file_size);
115 log_err("fio: posix_fallocate fails: %s\n",
119 #ifdef CONFIG_LINUX_FALLOCATE
120 case FIO_FALLOCATE_KEEP_SIZE:
122 "fallocate(FALLOC_FL_KEEP_SIZE) "
123 "file %s size %llu\n", f->file_name,
124 (unsigned long long) f->real_file_size);
126 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
129 td_verror(td, errno, "fallocate");
132 #endif /* CONFIG_LINUX_FALLOCATE */
134 log_err("fio: unknown fallocate mode: %d\n",
135 td->o.fallocate_mode);
139 #endif /* CONFIG_POSIX_FALLOCATE */
142 * If our jobs don't require regular files initially, we're done.
148 * The size will be -1ULL when fill_device is used, so don't truncate
149 * or fallocate this file, just write it
151 if (!td->o.fill_device) {
152 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
153 (unsigned long long) f->real_file_size);
154 if (ftruncate(f->fd, f->real_file_size) == -1) {
155 if (errno != EFBIG) {
156 td_verror(td, errno, "ftruncate");
162 left = f->real_file_size;
163 bs = td->o.max_bs[DDIR_WRITE];
169 td_verror(td, errno, "malloc");
173 while (left && !td->terminate) {
177 fill_io_buffer(td, b, bs, bs);
179 r = write(f->fd, b, bs);
189 if (td->o.fill_device)
191 log_info("fio: ENOSPC on laying out "
195 td_verror(td, errno, "write");
197 td_verror(td, EIO, "write");
204 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
205 td_io_unlink_file(td, f);
206 } else if (td->o.create_fsync) {
207 if (fsync(f->fd) < 0) {
208 td_verror(td, errno, "fsync");
212 if (td->o.fill_device && !td_write(td)) {
213 fio_file_clear_size_known(f);
214 if (td_io_get_file_size(td, f))
216 if (f->io_size > f->real_file_size)
217 f->io_size = f->real_file_size;
231 static int pre_read_file(struct thread_data *td, struct fio_file *f)
233 int ret = 0, r, did_open = 0, old_runstate;
234 unsigned long long left;
238 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
239 td_ioengine_flagged(td, FIO_NOIO))
242 if (f->filetype == FIO_TYPE_CHAR)
245 if (!fio_file_open(f)) {
246 if (td->io_ops->open_file(td, f)) {
247 log_err("fio: cannot pre-read, failed to open file\n");
253 old_runstate = td_bump_runstate(td, TD_PRE_READING);
256 bs = td->o.max_bs[DDIR_READ];
262 td_verror(td, errno, "malloc");
268 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
269 td_verror(td, errno, "lseek");
270 log_err("fio: failed to lseek pre-read file\n");
275 while (left && !td->terminate) {
279 r = read(f->fd, b, bs);
285 td_verror(td, EIO, "pre_read");
291 td_restore_runstate(td, old_runstate);
294 td->io_ops->close_file(td, f);
300 unsigned long long get_rand_file_size(struct thread_data *td)
302 unsigned long long ret, sized;
306 frand_max = rand_max(&td->file_size_state);
307 r = __rand(&td->file_size_state);
308 sized = td->o.file_size_high - td->o.file_size_low;
309 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
310 ret += td->o.file_size_low;
311 ret -= (ret % td->o.rw_min_bs);
315 static int file_size(struct thread_data *td, struct fio_file *f)
319 if (stat(f->file_name, &st) == -1) {
320 td_verror(td, errno, "fstat");
324 f->real_file_size = st.st_size;
328 static int bdev_size(struct thread_data *td, struct fio_file *f)
330 unsigned long long bytes = 0;
333 if (td->io_ops->open_file(td, f)) {
334 log_err("fio: failed opening blockdev %s for size check\n",
339 r = blockdev_size(f, &bytes);
341 td_verror(td, r, "blockdev_size");
346 log_err("%s: zero sized block device?\n", f->file_name);
350 f->real_file_size = bytes;
351 td->io_ops->close_file(td, f);
354 td->io_ops->close_file(td, f);
358 static int char_size(struct thread_data *td, struct fio_file *f)
360 #ifdef FIO_HAVE_CHARDEV_SIZE
361 unsigned long long bytes = 0;
364 if (td->io_ops->open_file(td, f)) {
365 log_err("fio: failed opening chardev %s for size check\n",
370 r = chardev_size(f, &bytes);
372 td_verror(td, r, "chardev_size");
377 log_err("%s: zero sized char device?\n", f->file_name);
381 f->real_file_size = bytes;
382 td->io_ops->close_file(td, f);
385 td->io_ops->close_file(td, f);
388 f->real_file_size = -1ULL;
393 static int get_file_size(struct thread_data *td, struct fio_file *f)
397 if (fio_file_size_known(f))
400 if (f->filetype == FIO_TYPE_FILE)
401 ret = file_size(td, f);
402 else if (f->filetype == FIO_TYPE_BLOCK)
403 ret = bdev_size(td, f);
404 else if (f->filetype == FIO_TYPE_CHAR)
405 ret = char_size(td, f);
407 f->real_file_size = -1ULL;
410 * Leave ->real_file_size with 0 since it could be expectation
411 * of initial setup for regular files.
417 * If ->real_file_size is -1, a conditional for the message
418 * "offset extends end" is always true, but it makes no sense,
419 * so just return the same value here.
421 if (f->real_file_size == -1ULL) {
422 log_info("%s: failed to get file size of %s\n", td->o.name,
427 if (td->o.start_offset && f->file_offset == 0)
428 dprint(FD_FILE, "offset of file %s not initialized yet\n",
431 * ->file_offset normally hasn't been initialized yet, so this
432 * is basically always false.
434 if (f->file_offset > f->real_file_size) {
435 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
436 (unsigned long long) f->file_offset,
437 (unsigned long long) f->real_file_size);
441 fio_file_set_size_known(f);
445 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
446 unsigned long long off,
447 unsigned long long len)
449 int errval = 0, ret = 0;
458 off = f->file_offset;
460 if (len == -1ULL || off == -1ULL)
463 if (td->io_ops->invalidate) {
464 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
466 ret = td->io_ops->invalidate(td, f);
469 } else if (f->filetype == FIO_TYPE_FILE) {
470 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
471 f->file_name, off, len);
472 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
475 } else if (f->filetype == FIO_TYPE_BLOCK) {
478 dprint(FD_IO, "drop page cache %s\n", f->file_name);
479 ret = blockdev_invalidate_cache(f);
480 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
482 * Linux multipath devices reject ioctl while
483 * the maps are being updated. That window can
484 * last tens of milliseconds; we'll try up to
485 * a quarter of a second.
488 ret = blockdev_invalidate_cache(f);
490 if (ret < 0 && errno == EACCES && geteuid()) {
492 log_err("fio: only root may flush block "
493 "devices. Cache flush bypassed!\n");
500 else if (ret) /* probably not supported */
502 } else if (f->filetype == FIO_TYPE_CHAR ||
503 f->filetype == FIO_TYPE_PIPE) {
504 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
509 * Cache flushing isn't a fatal condition, and we know it will
510 * happen on some platforms where we don't have the proper
511 * function to flush eg block device caches. So just warn and
512 * continue on our way.
515 log_info("fio: cache invalidation of %s failed: %s\n",
516 f->file_name, strerror(errval));
522 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
524 if (!fio_file_open(f))
527 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
530 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
534 dprint(FD_FILE, "fd close %s\n", f->file_name);
538 if (close(f->fd) < 0)
543 if (f->shadow_fd != -1) {
552 int file_lookup_open(struct fio_file *f, int flags)
554 struct fio_file *__f;
557 __f = lookup_file_hash(f->file_name);
559 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
563 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
571 f->fd = open(f->file_name, flags, 0600);
575 static int file_close_shadow_fds(struct thread_data *td)
581 for_each_file(td, f, i) {
582 if (f->shadow_fd == -1)
593 int generic_open_file(struct thread_data *td, struct fio_file *f)
599 dprint(FD_FILE, "fd open %s\n", f->file_name);
601 if (!strcmp(f->file_name, "-")) {
603 log_err("fio: can't read/write to stdin/out\n");
609 * move output logging to stderr, if we are writing to stdout
618 flags |= OS_O_DIRECT;
621 td_verror(td, EINVAL, "OS does not support atomic IO");
624 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
628 if (td->o.create_on_open && td->o.allow_create)
631 if (f->filetype != FIO_TYPE_FILE)
632 flags |= FIO_O_NOATIME;
639 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
643 f->fd = dup(STDOUT_FILENO);
645 from_hash = file_lookup_open(f, flags);
646 } else if (td_read(td)) {
647 if (f->filetype == FIO_TYPE_CHAR && !read_only)
653 f->fd = dup(STDIN_FILENO);
655 from_hash = file_lookup_open(f, flags);
656 } else if (td_trim(td)) {
657 assert(!td_rw(td)); /* should have matched above */
659 from_hash = file_lookup_open(f, flags);
663 char buf[FIO_VERROR_SIZE];
666 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
667 flags &= ~FIO_O_NOATIME;
670 if (__e == EMFILE && file_close_shadow_fds(td))
673 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
675 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
676 log_err("fio: looks like your file system does not " \
677 "support direct=1/buffered=0\n");
680 td_verror(td, __e, buf);
684 if (!from_hash && f->fd != -1) {
685 if (add_file_hash(f)) {
689 * Stash away descriptor for later close. This is to
690 * work-around a "feature" on Linux, where a close of
691 * an fd that has been opened for write will trigger
692 * udev to call blkid to check partitions, fs id, etc.
693 * That pollutes the device cache, which can slow down
694 * unbuffered accesses.
696 if (f->shadow_fd == -1)
697 f->shadow_fd = f->fd;
700 * OK to ignore, we haven't done anything
703 ret = generic_close_file(td, f);
713 * This function i.e. get_file_size() is the default .get_file_size
714 * implementation of majority of I/O engines.
716 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
718 return get_file_size(td, f);
722 * open/close all files, so that ->real_file_size gets set
724 static int get_file_sizes(struct thread_data *td)
730 for_each_file(td, f, i) {
731 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
734 if (td_io_get_file_size(td, f)) {
735 if (td->error != ENOENT) {
736 log_err("%s\n", td->verror);
744 * There are corner cases where we end up with -1 for
745 * ->real_file_size due to unsupported file type, etc.
746 * We then just set to size option value divided by number
747 * of files, similar to the way file ->io_size is set.
748 * stat(2) failure doesn't set ->real_file_size to -1.
750 if (f->real_file_size == -1ULL && td->o.size)
751 f->real_file_size = td->o.size / td->o.nr_files;
758 struct flist_head list;
765 * Get free number of bytes for each file on each unique mount.
767 static unsigned long long get_fs_free_counts(struct thread_data *td)
769 struct flist_head *n, *tmp;
770 unsigned long long ret = 0;
771 struct fio_mount *fm;
776 for_each_file(td, f, i) {
780 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
781 if (f->real_file_size != -1ULL)
782 ret += f->real_file_size;
784 } else if (f->filetype != FIO_TYPE_FILE)
788 strncpy(buf, f->file_name, 255);
790 if (stat(buf, &sb) < 0) {
794 if (stat(buf, &sb) < 0)
799 flist_for_each(n, &list) {
800 fm = flist_entry(n, struct fio_mount, list);
801 if (fm->key == sb.st_dev)
810 fm = calloc(1, sizeof(*fm));
811 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
812 fm->base = basename(fm->__base);
814 flist_add(&fm->list, &list);
817 flist_for_each_safe(n, tmp, &list) {
818 unsigned long long sz;
820 fm = flist_entry(n, struct fio_mount, list);
821 flist_del(&fm->list);
823 sz = get_fs_free_size(fm->base);
824 if (sz && sz != -1ULL)
833 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
835 struct thread_options *o = &td->o;
836 unsigned long long align_bs; /* align the offset to this block size */
837 unsigned long long offset; /* align_bs-aligned offset */
839 if (o->file_append && f->filetype == FIO_TYPE_FILE)
840 return f->real_file_size;
842 if (o->start_offset_percent > 0) {
844 /* if blockalign is provided, find the min across read, write, and trim */
845 if (fio_option_is_set(o, ba)) {
846 align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
847 align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
848 } else { /* else take the minimum block size */
849 align_bs = td_min_bs(td);
852 /* calculate the raw offset */
853 offset = (f->real_file_size * o->start_offset_percent / 100) +
854 (td->subjob_number * o->offset_increment);
856 /* block align the offset at the next available boundary at
857 ceiling(offset / align_bs) * align_bs */
858 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
860 } else { /* start_offset_percent not set */
861 offset = o->start_offset + o->start_offset +
862 td->subjob_number * o->offset_increment;
868 * Open the files and setup files sizes, creating files if necessary.
870 int setup_files(struct thread_data *td)
872 unsigned long long total_size, extend_size;
873 struct thread_options *o = &td->o;
875 unsigned int i, nr_fs_extra = 0;
876 int err = 0, need_extend;
878 const unsigned int bs = td_min_bs(td);
881 dprint(FD_FILE, "setup files\n");
883 old_state = td_bump_runstate(td, TD_SETTING_UP);
885 if (o->read_iolog_file)
889 * Find out physical size of files or devices for this thread,
890 * before we determine I/O size and range of our targets.
891 * If ioengine defines a setup() method, it's responsible for
892 * opening the files and setting f->real_file_size to indicate
893 * the valid range for that file.
895 if (td->io_ops->setup)
896 err = td->io_ops->setup(td);
898 err = get_file_sizes(td);
904 * check sizes. if the files/devices do not exist and the size
905 * isn't passed to fio, abort.
908 for_each_file(td, f, i) {
910 if (f->real_file_size == -1ULL)
913 total_size += f->real_file_size;
917 td->fill_device_size = get_fs_free_counts(td);
920 * device/file sizes are zero and no size given, punt
922 if ((!total_size || total_size == -1ULL) && !o->size &&
923 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
924 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
925 log_err("%s: you need to specify size=\n", o->name);
926 td_verror(td, EINVAL, "total_file_size");
931 * Calculate per-file size and potential extra size for the
932 * first files, if needed (i.e. if we don't have a fixed size).
934 if (!o->file_size_low && o->nr_files) {
937 fs = o->size / o->nr_files;
938 all_fs = fs * o->nr_files;
940 if (all_fs < o->size)
941 nr_fs_extra = (o->size - all_fs) / bs;
945 * now file sizes are known, so we can set ->io_size. if size= is
946 * not given, ->io_size is just equal to ->real_file_size. if size
947 * is given, ->io_size is size / nr_files.
949 extend_size = total_size = 0;
951 for_each_file(td, f, i) {
952 f->file_offset = get_start_offset(td, f);
955 * Update ->io_size depending on options specified.
956 * ->file_size_low being 0 means filesize option isn't set.
957 * Non zero ->file_size_low equals ->file_size_high means
958 * filesize option is set in a fixed size format.
959 * Non zero ->file_size_low not equals ->file_size_high means
960 * filesize option is set in a range format.
962 if (!o->file_size_low) {
964 * no file size or range given, file size is equal to
965 * total size divided by number of files. If the size
966 * doesn't divide nicely with the min blocksize,
967 * make the first files bigger.
976 * We normally don't come here for regular files, but
977 * if the result is 0 for a regular file, set it to the
978 * real file size. This could be size of the existing
979 * one if it already exists, but otherwise will be set
980 * to 0. A new file won't be created because
981 * ->io_size + ->file_offset equals ->real_file_size.
984 if (f->file_offset > f->real_file_size)
986 f->io_size = f->real_file_size - f->file_offset;
988 log_info("fio: file %s may be ignored\n",
991 } else if (f->real_file_size < o->file_size_low ||
992 f->real_file_size > o->file_size_high) {
993 if (f->file_offset > o->file_size_low)
996 * file size given. if it's fixed, use that. if it's a
997 * range, generate a random size in-between.
999 if (o->file_size_low == o->file_size_high)
1000 f->io_size = o->file_size_low - f->file_offset;
1002 f->io_size = get_rand_file_size(td)
1006 f->io_size = f->real_file_size - f->file_offset;
1008 if (f->io_size == -1ULL)
1011 if (o->size_percent) {
1012 f->io_size = (f->io_size * o->size_percent) / 100;
1013 f->io_size -= (f->io_size % td_min_bs(td));
1015 total_size += f->io_size;
1018 if (f->filetype == FIO_TYPE_FILE &&
1019 (f->io_size + f->file_offset) > f->real_file_size &&
1020 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1021 if (!o->create_on_open) {
1023 extend_size += (f->io_size + f->file_offset);
1024 fio_file_set_extend(f);
1026 f->real_file_size = f->io_size + f->file_offset;
1030 if (td->o.block_error_hist) {
1033 assert(td->o.nr_files == 1); /* checked in fixup_options */
1035 len = f->io_size / td->o.bs[DDIR_TRIM];
1036 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1037 log_err("fio: cannot calculate block histogram with "
1038 "%d trim blocks, maximum %d\n",
1039 len, MAX_NR_BLOCK_INFOS);
1040 td_verror(td, EINVAL, "block_error_hist");
1044 td->ts.nr_block_infos = len;
1045 for (i = 0; i < len; i++)
1046 td->ts.block_infos[i] =
1047 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1049 td->ts.nr_block_infos = 0;
1051 if (!o->size || (total_size && o->size > total_size))
1052 o->size = total_size;
1054 if (o->size < td_min_bs(td)) {
1055 log_err("fio: blocksize too large for data set\n");
1060 * See if we need to extend some files, typically needed when our
1061 * target regular files don't exist yet, but our jobs require them
1062 * initially due to read I/Os.
1066 if (output_format & FIO_OUTPUT_NORMAL) {
1067 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1069 need_extend > 1 ? "s" : "",
1071 need_extend > 1 ? "s" : "",
1072 need_extend > 1 ? "total " : "",
1076 for_each_file(td, f, i) {
1077 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1079 if (!fio_file_extend(f))
1082 assert(f->filetype == FIO_TYPE_FILE);
1083 fio_file_clear_extend(f);
1084 if (!o->fill_device) {
1085 old_len = f->real_file_size;
1086 extend_len = f->io_size + f->file_offset -
1089 f->real_file_size = (f->io_size + f->file_offset);
1090 err = extend_file(td, f);
1094 err = __file_invalidate_cache(td, f, old_len,
1098 * Shut up static checker
1114 o->zone_size = o->size;
1117 * iolog already set the total io size, if we read back
1120 if (!o->read_iolog_file) {
1122 td->total_io_size = o->io_size * o->loops;
1124 td->total_io_size = o->size * o->loops;
1131 td_restore_runstate(td, old_state);
1134 log_err("%s: you need to specify valid offset=\n", o->name);
1136 td_restore_runstate(td, old_state);
1140 int pre_read_files(struct thread_data *td)
1145 dprint(FD_FILE, "pre_read files\n");
1147 for_each_file(td, f, i) {
1148 if (pre_read_file(td, f))
1155 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1157 unsigned int range_size, seed;
1158 unsigned long nranges;
1161 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1162 fsize = min(f->real_file_size, f->io_size);
1164 nranges = (fsize + range_size - 1) / range_size;
1166 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1167 if (!td->o.rand_repeatable)
1168 seed = td->rand_seeds[4];
1170 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1171 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1172 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1173 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1174 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1175 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1180 static int init_rand_distribution(struct thread_data *td)
1186 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1189 state = td_bump_runstate(td, TD_SETTING_UP);
1191 for_each_file(td, f, i)
1192 __init_rand_distribution(td, f);
1194 td_restore_runstate(td, state);
1200 * Check if the number of blocks exceeds the randomness capability of
1201 * the selected generator. Tausworthe is 32-bit, the others are fullly
1204 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1207 if (blocks <= FRAND32_MAX)
1209 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1213 * If the user hasn't specified a random generator, switch
1214 * to tausworthe64 with informational warning. If the user did
1215 * specify one, just warn.
1217 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1220 if (!fio_option_is_set(&td->o, random_generator)) {
1221 log_info("fio: Switching to tausworthe64. Use the "
1222 "random_generator= option to get rid of this "
1224 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1229 * Just make this information to avoid breaking scripts.
1231 log_info("fio: Use the random_generator= option to switch to lfsr or "
1236 int init_random_map(struct thread_data *td)
1238 unsigned long long blocks;
1242 if (init_rand_distribution(td))
1247 for_each_file(td, f, i) {
1248 uint64_t fsize = min(f->real_file_size, f->io_size);
1250 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1252 if (check_rand_gen_limits(td, f, blocks))
1255 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1258 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1260 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1261 fio_file_set_lfsr(f);
1264 } else if (!td->o.norandommap) {
1265 f->io_axmap = axmap_new(blocks);
1267 fio_file_set_axmap(f);
1270 } else if (td->o.norandommap)
1273 if (!td->o.softrandommap) {
1274 log_err("fio: failed allocating random map. If running"
1275 " a large number of jobs, try the 'norandommap'"
1276 " option or set 'softrandommap'. Or give"
1277 " a larger --alloc-size to fio.\n");
1281 log_info("fio: file %s failed allocating random map. Running "
1282 "job without.\n", f->file_name);
1288 void close_files(struct thread_data *td)
1293 for_each_file(td, f, i) {
1294 if (fio_file_open(f))
1295 td_io_close_file(td, f);
1299 void close_and_free_files(struct thread_data *td)
1304 dprint(FD_FILE, "close files\n");
1306 for_each_file(td, f, i) {
1307 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1308 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1309 td_io_unlink_file(td, f);
1312 if (fio_file_open(f))
1313 td_io_close_file(td, f);
1315 remove_file_hash(f);
1317 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1318 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1319 td_io_unlink_file(td, f);
1322 sfree(f->file_name);
1323 f->file_name = NULL;
1324 if (fio_file_axmap(f)) {
1325 axmap_free(f->io_axmap);
1331 td->o.filename = NULL;
1333 free(td->file_locks);
1334 td->files_index = 0;
1336 td->file_locks = NULL;
1337 td->o.file_lock_mode = FILE_LOCK_NONE;
1341 static void get_file_type(struct fio_file *f)
1345 if (!strcmp(f->file_name, "-"))
1346 f->filetype = FIO_TYPE_PIPE;
1348 f->filetype = FIO_TYPE_FILE;
1351 /* \\.\ is the device namespace in Windows, where every file is
1353 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1354 f->filetype = FIO_TYPE_BLOCK;
1357 if (!stat(f->file_name, &sb)) {
1358 if (S_ISBLK(sb.st_mode))
1359 f->filetype = FIO_TYPE_BLOCK;
1360 else if (S_ISCHR(sb.st_mode))
1361 f->filetype = FIO_TYPE_CHAR;
1362 else if (S_ISFIFO(sb.st_mode))
1363 f->filetype = FIO_TYPE_PIPE;
1367 static bool __is_already_allocated(const char *fname, bool set)
1369 struct flist_head *entry;
1372 ret = file_bloom_exists(fname, set);
1376 flist_for_each(entry, &filename_list) {
1377 struct file_name *fn;
1379 fn = flist_entry(entry, struct file_name, list);
1381 if (!strcmp(fn->filename, fname))
1388 static bool is_already_allocated(const char *fname)
1392 fio_file_hash_lock();
1393 ret = __is_already_allocated(fname, false);
1394 fio_file_hash_unlock();
1399 static void set_already_allocated(const char *fname)
1401 struct file_name *fn;
1403 fn = malloc(sizeof(struct file_name));
1404 fn->filename = strdup(fname);
1406 fio_file_hash_lock();
1407 if (!__is_already_allocated(fname, true)) {
1408 flist_add_tail(&fn->list, &filename_list);
1411 fio_file_hash_unlock();
1419 static void free_already_allocated(void)
1421 struct flist_head *entry, *tmp;
1422 struct file_name *fn;
1424 if (flist_empty(&filename_list))
1427 fio_file_hash_lock();
1428 flist_for_each_safe(entry, tmp, &filename_list) {
1429 fn = flist_entry(entry, struct file_name, list);
1431 flist_del(&fn->list);
1435 fio_file_hash_unlock();
1438 static struct fio_file *alloc_new_file(struct thread_data *td)
1442 f = smalloc(sizeof(*f));
1450 fio_file_reset(td, f);
1454 bool exists_and_not_regfile(const char *filename)
1458 if (lstat(filename, &sb) == -1)
1461 #ifndef WIN32 /* NOT Windows */
1462 if (S_ISREG(sb.st_mode))
1465 /* \\.\ is the device namespace in Windows, where every file
1466 * is a device node */
1467 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1474 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1476 int cur_files = td->files_index;
1477 char file_name[PATH_MAX];
1481 dprint(FD_FILE, "add file %s\n", fname);
1483 if (td->o.directory)
1484 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1485 td->o.unique_filename);
1487 sprintf(file_name + len, "%s", fname);
1489 /* clean cloned siblings using existing files */
1490 if (numjob && is_already_allocated(file_name) &&
1491 !exists_and_not_regfile(fname))
1494 f = alloc_new_file(td);
1496 if (td->files_size <= td->files_index) {
1497 unsigned int new_size = td->o.nr_files + 1;
1499 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1501 td->files = realloc(td->files, new_size * sizeof(f));
1502 if (td->files == NULL) {
1503 log_err("fio: realloc OOM\n");
1506 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1507 td->file_locks = realloc(td->file_locks, new_size);
1508 if (!td->file_locks) {
1509 log_err("fio: realloc OOM\n");
1512 td->file_locks[cur_files] = FILE_LOCK_NONE;
1514 td->files_size = new_size;
1516 td->files[cur_files] = f;
1517 f->fileno = cur_files;
1520 * init function, io engine may not be loaded yet
1522 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1523 f->real_file_size = -1ULL;
1525 f->file_name = smalloc_strdup(file_name);
1531 switch (td->o.file_lock_mode) {
1532 case FILE_LOCK_NONE:
1534 case FILE_LOCK_READWRITE:
1535 f->rwlock = fio_rwlock_init();
1537 case FILE_LOCK_EXCLUSIVE:
1538 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1541 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1546 if (f->filetype == FIO_TYPE_FILE)
1547 td->nr_normal_files++;
1549 set_already_allocated(file_name);
1554 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1560 int add_file_exclusive(struct thread_data *td, const char *fname)
1565 for_each_file(td, f, i) {
1566 if (!strcmp(f->file_name, fname))
1570 return add_file(td, fname, 0, 1);
1573 void get_file(struct fio_file *f)
1575 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1576 assert(fio_file_open(f));
1580 int put_file(struct thread_data *td, struct fio_file *f)
1582 int f_ret = 0, ret = 0;
1584 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1586 if (!fio_file_open(f)) {
1587 assert(f->fd == -1);
1591 assert(f->references);
1592 if (--f->references)
1595 if (should_fsync(td) && td->o.fsync_on_close) {
1596 f_ret = fsync(f->fd);
1601 if (td->io_ops->close_file)
1602 ret = td->io_ops->close_file(td, f);
1607 td->nr_open_files--;
1608 fio_file_clear_open(f);
1609 assert(f->fd == -1);
1613 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1615 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1618 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1619 if (ddir == DDIR_READ)
1620 fio_rwlock_read(f->rwlock);
1622 fio_rwlock_write(f->rwlock);
1623 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1624 fio_mutex_down(f->lock);
1626 td->file_locks[f->fileno] = td->o.file_lock_mode;
1629 void unlock_file(struct thread_data *td, struct fio_file *f)
1631 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1634 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1635 fio_rwlock_unlock(f->rwlock);
1636 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1637 fio_mutex_up(f->lock);
1639 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1642 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1644 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1646 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1650 static int recurse_dir(struct thread_data *td, const char *dirname)
1656 D = opendir(dirname);
1658 char buf[FIO_VERROR_SIZE];
1660 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1661 td_verror(td, errno, buf);
1665 while ((dir = readdir(D)) != NULL) {
1666 char full_path[PATH_MAX];
1669 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1672 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1674 if (lstat(full_path, &sb) == -1) {
1675 if (errno != ENOENT) {
1676 td_verror(td, errno, "stat");
1682 if (S_ISREG(sb.st_mode)) {
1683 add_file(td, full_path, 0, 1);
1686 if (!S_ISDIR(sb.st_mode))
1689 ret = recurse_dir(td, full_path);
1698 int add_dir_files(struct thread_data *td, const char *path)
1700 int ret = recurse_dir(td, path);
1703 log_info("fio: opendir added %d files\n", td->o.nr_files);
1708 void dup_files(struct thread_data *td, struct thread_data *org)
1713 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1718 td->files = malloc(org->files_index * sizeof(f));
1720 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1721 td->file_locks = malloc(org->files_index);
1723 for_each_file(org, f, i) {
1724 struct fio_file *__f;
1726 __f = alloc_new_file(td);
1729 __f->file_name = smalloc_strdup(f->file_name);
1730 if (!__f->file_name)
1733 __f->filetype = f->filetype;
1736 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1737 __f->lock = f->lock;
1738 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1739 __f->rwlock = f->rwlock;
1746 * Returns the index that matches the filename, or -1 if not there
1748 int get_fileno(struct thread_data *td, const char *fname)
1753 for_each_file(td, f, i)
1754 if (!strcmp(f->file_name, fname))
1761 * For log usage, where we add/open/close files automatically
1763 void free_release_files(struct thread_data *td)
1767 td->o.open_files = 0;
1768 td->files_index = 0;
1769 td->nr_normal_files = 0;
1772 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1776 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1777 f->last_pos[i] = f->file_offset;
1778 f->last_start[i] = -1ULL;
1781 if (fio_file_axmap(f))
1782 axmap_reset(f->io_axmap);
1783 else if (fio_file_lfsr(f))
1784 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1787 bool fio_files_done(struct thread_data *td)
1792 for_each_file(td, f, i)
1793 if (!fio_file_done(f))
1799 /* free memory used in initialization phase only */
1800 void filesetup_mem_free(void)
1802 free_already_allocated();