17 #include "lib/axmap.h"
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
25 static FLIST_HEAD(filename_list);
27 static inline void clear_error(struct thread_data *td)
34 * Leaves f->fd open on success, caller must close
36 static int extend_file(struct thread_data *td, struct fio_file *f)
38 int r, new_layout = 0, unlink_file = 0, flags;
39 unsigned long long left;
44 log_err("fio: refusing extend of file due to read-only\n");
49 * check if we need to lay the file out complete again. fio
50 * does that for operations involving reads, or for writes
51 * where overwrite is set
54 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
55 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
57 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
60 if (unlink_file || new_layout) {
63 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
65 ret = td_io_unlink_file(td, f);
66 if (ret != 0 && ret != ENOENT) {
67 td_verror(td, errno, "unlink");
73 if (td->o.allow_create)
82 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
83 f->fd = open(f->file_name, flags, 0644);
87 if (err == ENOENT && !td->o.allow_create)
88 log_err("fio: file creation disallowed by "
89 "allow_file_create=0\n");
91 td_verror(td, err, "open");
95 #ifdef CONFIG_POSIX_FALLOCATE
96 if (!td->o.fill_device) {
97 switch (td->o.fallocate_mode) {
98 case FIO_FALLOCATE_NONE:
100 case FIO_FALLOCATE_POSIX:
101 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
103 (unsigned long long) f->real_file_size);
105 r = posix_fallocate(f->fd, 0, f->real_file_size);
107 log_err("fio: posix_fallocate fails: %s\n",
111 #ifdef CONFIG_LINUX_FALLOCATE
112 case FIO_FALLOCATE_KEEP_SIZE:
114 "fallocate(FALLOC_FL_KEEP_SIZE) "
115 "file %s size %llu\n", f->file_name,
116 (unsigned long long) f->real_file_size);
118 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
121 td_verror(td, errno, "fallocate");
124 #endif /* CONFIG_LINUX_FALLOCATE */
126 log_err("fio: unknown fallocate mode: %d\n",
127 td->o.fallocate_mode);
131 #endif /* CONFIG_POSIX_FALLOCATE */
134 * If our jobs don't require regular files initially, we're done.
140 * The size will be -1ULL when fill_device is used, so don't truncate
141 * or fallocate this file, just write it
143 if (!td->o.fill_device) {
144 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
145 (unsigned long long) f->real_file_size);
146 if (ftruncate(f->fd, f->real_file_size) == -1) {
147 if (errno != EFBIG) {
148 td_verror(td, errno, "ftruncate");
154 b = malloc(td->o.max_bs[DDIR_WRITE]);
156 left = f->real_file_size;
157 while (left && !td->terminate) {
158 bs = td->o.max_bs[DDIR_WRITE];
162 fill_io_buffer(td, b, bs, bs);
164 r = write(f->fd, b, bs);
174 if (td->o.fill_device)
176 log_info("fio: ENOSPC on laying out "
180 td_verror(td, errno, "write");
182 td_verror(td, EIO, "write");
189 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
190 td_io_unlink_file(td, f);
191 } else if (td->o.create_fsync) {
192 if (fsync(f->fd) < 0) {
193 td_verror(td, errno, "fsync");
197 if (td->o.fill_device && !td_write(td)) {
198 fio_file_clear_size_known(f);
199 if (td_io_get_file_size(td, f))
201 if (f->io_size > f->real_file_size)
202 f->io_size = f->real_file_size;
216 static int pre_read_file(struct thread_data *td, struct fio_file *f)
218 int ret = 0, r, did_open = 0, old_runstate;
219 unsigned long long left;
223 if (td_ioengine_flagged(td, FIO_PIPEIO))
226 if (!fio_file_open(f)) {
227 if (td->io_ops->open_file(td, f)) {
228 log_err("fio: cannot pre-read, failed to open file\n");
234 old_runstate = td_bump_runstate(td, TD_PRE_READING);
236 bs = td->o.max_bs[DDIR_READ];
240 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
241 td_verror(td, errno, "lseek");
242 log_err("fio: failed to lseek pre-read file\n");
249 while (left && !td->terminate) {
253 r = read(f->fd, b, bs);
259 td_verror(td, EIO, "pre_read");
265 td_restore_runstate(td, old_runstate);
268 td->io_ops->close_file(td, f);
274 unsigned long long get_rand_file_size(struct thread_data *td)
276 unsigned long long ret, sized;
280 frand_max = rand_max(&td->file_size_state);
281 r = __rand(&td->file_size_state);
282 sized = td->o.file_size_high - td->o.file_size_low;
283 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
284 ret += td->o.file_size_low;
285 ret -= (ret % td->o.rw_min_bs);
289 static int file_size(struct thread_data *td, struct fio_file *f)
293 if (stat(f->file_name, &st) == -1) {
294 td_verror(td, errno, "fstat");
298 f->real_file_size = st.st_size;
302 static int bdev_size(struct thread_data *td, struct fio_file *f)
304 unsigned long long bytes = 0;
307 if (td->io_ops->open_file(td, f)) {
308 log_err("fio: failed opening blockdev %s for size check\n",
313 r = blockdev_size(f, &bytes);
315 td_verror(td, r, "blockdev_size");
320 log_err("%s: zero sized block device?\n", f->file_name);
324 f->real_file_size = bytes;
325 td->io_ops->close_file(td, f);
328 td->io_ops->close_file(td, f);
332 static int char_size(struct thread_data *td, struct fio_file *f)
334 #ifdef FIO_HAVE_CHARDEV_SIZE
335 unsigned long long bytes = 0;
338 if (td->io_ops->open_file(td, f)) {
339 log_err("fio: failed opening chardev %s for size check\n",
344 r = chardev_size(f, &bytes);
346 td_verror(td, r, "chardev_size");
351 log_err("%s: zero sized char device?\n", f->file_name);
355 f->real_file_size = bytes;
356 td->io_ops->close_file(td, f);
359 td->io_ops->close_file(td, f);
362 f->real_file_size = -1ULL;
367 static int get_file_size(struct thread_data *td, struct fio_file *f)
371 if (fio_file_size_known(f))
374 if (f->filetype == FIO_TYPE_FILE)
375 ret = file_size(td, f);
376 else if (f->filetype == FIO_TYPE_BLOCK)
377 ret = bdev_size(td, f);
378 else if (f->filetype == FIO_TYPE_CHAR)
379 ret = char_size(td, f);
381 f->real_file_size = -1ULL;
384 * Leave ->real_file_size with 0 since it could be expectation
385 * of initial setup for regular files.
391 * If ->real_file_size is -1, a conditional for the message
392 * "offset extends end" is always true, but it makes no sense,
393 * so just return the same value here.
395 if (f->real_file_size == -1ULL) {
396 log_info("%s: failed to get file size of %s\n", td->o.name,
401 if (td->o.start_offset && f->file_offset == 0)
402 dprint(FD_FILE, "offset of file %s not initialized yet\n",
405 * ->file_offset normally hasn't been initialized yet, so this
406 * is basically always false.
408 if (f->file_offset > f->real_file_size) {
409 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
410 (unsigned long long) f->file_offset,
411 (unsigned long long) f->real_file_size);
415 fio_file_set_size_known(f);
419 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
420 unsigned long long off,
421 unsigned long long len)
423 int errval = 0, ret = 0;
432 off = f->file_offset;
434 if (len == -1ULL || off == -1ULL)
437 dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
440 if (td->io_ops->invalidate) {
441 ret = td->io_ops->invalidate(td, f);
444 } else if (f->filetype == FIO_TYPE_FILE) {
445 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
448 } else if (f->filetype == FIO_TYPE_BLOCK) {
451 ret = blockdev_invalidate_cache(f);
452 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
454 * Linux multipath devices reject ioctl while
455 * the maps are being updated. That window can
456 * last tens of milliseconds; we'll try up to
457 * a quarter of a second.
460 ret = blockdev_invalidate_cache(f);
462 if (ret < 0 && errno == EACCES && geteuid()) {
464 log_err("fio: only root may flush block "
465 "devices. Cache flush bypassed!\n");
472 } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
476 * Cache flushing isn't a fatal condition, and we know it will
477 * happen on some platforms where we don't have the proper
478 * function to flush eg block device caches. So just warn and
479 * continue on our way.
482 log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errval));
488 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
490 if (!fio_file_open(f))
493 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
496 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
500 dprint(FD_FILE, "fd close %s\n", f->file_name);
504 if (close(f->fd) < 0)
509 if (f->shadow_fd != -1) {
518 int file_lookup_open(struct fio_file *f, int flags)
520 struct fio_file *__f;
523 __f = lookup_file_hash(f->file_name);
525 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
529 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
537 f->fd = open(f->file_name, flags, 0600);
541 static int file_close_shadow_fds(struct thread_data *td)
547 for_each_file(td, f, i) {
548 if (f->shadow_fd == -1)
559 int generic_open_file(struct thread_data *td, struct fio_file *f)
565 dprint(FD_FILE, "fd open %s\n", f->file_name);
567 if (!strcmp(f->file_name, "-")) {
569 log_err("fio: can't read/write to stdin/out\n");
575 * move output logging to stderr, if we are writing to stdout
584 flags |= OS_O_DIRECT;
587 td_verror(td, EINVAL, "OS does not support atomic IO");
590 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
594 if (td->o.create_on_open && td->o.allow_create)
597 if (f->filetype != FIO_TYPE_FILE)
598 flags |= FIO_O_NOATIME;
605 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
609 f->fd = dup(STDOUT_FILENO);
611 from_hash = file_lookup_open(f, flags);
612 } else if (td_read(td)) {
613 if (f->filetype == FIO_TYPE_CHAR && !read_only)
619 f->fd = dup(STDIN_FILENO);
621 from_hash = file_lookup_open(f, flags);
624 from_hash = file_lookup_open(f, flags);
628 char buf[FIO_VERROR_SIZE];
631 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
632 flags &= ~FIO_O_NOATIME;
635 if (__e == EMFILE && file_close_shadow_fds(td))
638 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
640 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
641 log_err("fio: looks like your file system does not " \
642 "support direct=1/buffered=0\n");
645 td_verror(td, __e, buf);
649 if (!from_hash && f->fd != -1) {
650 if (add_file_hash(f)) {
654 * Stash away descriptor for later close. This is to
655 * work-around a "feature" on Linux, where a close of
656 * an fd that has been opened for write will trigger
657 * udev to call blkid to check partitions, fs id, etc.
658 * That pollutes the device cache, which can slow down
659 * unbuffered accesses.
661 if (f->shadow_fd == -1)
662 f->shadow_fd = f->fd;
665 * OK to ignore, we haven't done anything
668 ret = generic_close_file(td, f);
678 * This function i.e. get_file_size() is the default .get_file_size
679 * implementation of majority of I/O engines.
681 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
683 return get_file_size(td, f);
687 * open/close all files, so that ->real_file_size gets set
689 static int get_file_sizes(struct thread_data *td)
695 for_each_file(td, f, i) {
696 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
699 if (td_io_get_file_size(td, f)) {
700 if (td->error != ENOENT) {
701 log_err("%s\n", td->verror);
709 * There are corner cases where we end up with -1 for
710 * ->real_file_size due to unsupported file type, etc.
711 * We then just set to size option value divided by number
712 * of files, similar to the way file ->io_size is set.
713 * stat(2) failure doesn't set ->real_file_size to -1.
715 if (f->real_file_size == -1ULL && td->o.size)
716 f->real_file_size = td->o.size / td->o.nr_files;
723 struct flist_head list;
730 * Get free number of bytes for each file on each unique mount.
732 static unsigned long long get_fs_free_counts(struct thread_data *td)
734 struct flist_head *n, *tmp;
735 unsigned long long ret = 0;
736 struct fio_mount *fm;
741 for_each_file(td, f, i) {
745 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
746 if (f->real_file_size != -1ULL)
747 ret += f->real_file_size;
749 } else if (f->filetype != FIO_TYPE_FILE)
753 strncpy(buf, f->file_name, 255);
755 if (stat(buf, &sb) < 0) {
759 if (stat(buf, &sb) < 0)
764 flist_for_each(n, &list) {
765 fm = flist_entry(n, struct fio_mount, list);
766 if (fm->key == sb.st_dev)
775 fm = calloc(1, sizeof(*fm));
776 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
777 fm->base = basename(fm->__base);
779 flist_add(&fm->list, &list);
782 flist_for_each_safe(n, tmp, &list) {
783 unsigned long long sz;
785 fm = flist_entry(n, struct fio_mount, list);
786 flist_del(&fm->list);
788 sz = get_fs_free_size(fm->base);
789 if (sz && sz != -1ULL)
798 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
800 struct thread_options *o = &td->o;
802 if (o->file_append && f->filetype == FIO_TYPE_FILE)
803 return f->real_file_size;
805 return td->o.start_offset +
806 td->subjob_number * td->o.offset_increment;
810 * Open the files and setup files sizes, creating files if necessary.
812 int setup_files(struct thread_data *td)
814 unsigned long long total_size, extend_size;
815 struct thread_options *o = &td->o;
817 unsigned int i, nr_fs_extra = 0;
818 int err = 0, need_extend;
820 const unsigned int bs = td_min_bs(td);
823 dprint(FD_FILE, "setup files\n");
825 old_state = td_bump_runstate(td, TD_SETTING_UP);
827 if (o->read_iolog_file)
831 * Find out physical size of files or devices for this thread,
832 * before we determine I/O size and range of our targets.
833 * If ioengine defines a setup() method, it's responsible for
834 * opening the files and setting f->real_file_size to indicate
835 * the valid range for that file.
837 if (td->io_ops->setup)
838 err = td->io_ops->setup(td);
840 err = get_file_sizes(td);
846 * check sizes. if the files/devices do not exist and the size
847 * isn't passed to fio, abort.
850 for_each_file(td, f, i) {
852 if (f->real_file_size == -1ULL)
855 total_size += f->real_file_size;
859 td->fill_device_size = get_fs_free_counts(td);
862 * device/file sizes are zero and no size given, punt
864 if ((!total_size || total_size == -1ULL) && !o->size &&
865 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
866 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
867 log_err("%s: you need to specify size=\n", o->name);
868 td_verror(td, EINVAL, "total_file_size");
873 * Calculate per-file size and potential extra size for the
874 * first files, if needed (i.e. if we don't have a fixed size).
876 if (!o->file_size_low && o->nr_files) {
879 fs = o->size / o->nr_files;
880 all_fs = fs * o->nr_files;
882 if (all_fs < o->size)
883 nr_fs_extra = (o->size - all_fs) / bs;
887 * now file sizes are known, so we can set ->io_size. if size= is
888 * not given, ->io_size is just equal to ->real_file_size. if size
889 * is given, ->io_size is size / nr_files.
891 extend_size = total_size = 0;
893 for_each_file(td, f, i) {
894 f->file_offset = get_start_offset(td, f);
897 * Update ->io_size depending on options specified.
898 * ->file_size_low being 0 means filesize option isn't set.
899 * Non zero ->file_size_low equals ->file_size_high means
900 * filesize option is set in a fixed size format.
901 * Non zero ->file_size_low not equals ->file_size_high means
902 * filesize option is set in a range format.
904 if (!o->file_size_low) {
906 * no file size or range given, file size is equal to
907 * total size divided by number of files. If that is
908 * zero, set it to the real file size. If the size
909 * doesn't divide nicely with the min blocksize,
910 * make the first files bigger.
919 f->io_size = f->real_file_size - f->file_offset;
920 } else if (f->real_file_size < o->file_size_low ||
921 f->real_file_size > o->file_size_high) {
922 if (f->file_offset > o->file_size_low)
925 * file size given. if it's fixed, use that. if it's a
926 * range, generate a random size in-between.
928 if (o->file_size_low == o->file_size_high)
929 f->io_size = o->file_size_low - f->file_offset;
931 f->io_size = get_rand_file_size(td)
935 f->io_size = f->real_file_size - f->file_offset;
937 if (f->io_size == -1ULL)
940 if (o->size_percent) {
941 f->io_size = (f->io_size * o->size_percent) / 100;
942 f->io_size -= (f->io_size % td_min_bs(td));
944 total_size += f->io_size;
947 if (f->filetype == FIO_TYPE_FILE &&
948 (f->io_size + f->file_offset) > f->real_file_size &&
949 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
950 if (!o->create_on_open) {
952 extend_size += (f->io_size + f->file_offset);
954 f->real_file_size = f->io_size + f->file_offset;
955 fio_file_set_extend(f);
959 if (td->o.block_error_hist) {
962 assert(td->o.nr_files == 1); /* checked in fixup_options */
964 len = f->io_size / td->o.bs[DDIR_TRIM];
965 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
966 log_err("fio: cannot calculate block histogram with "
967 "%d trim blocks, maximum %d\n",
968 len, MAX_NR_BLOCK_INFOS);
969 td_verror(td, EINVAL, "block_error_hist");
973 td->ts.nr_block_infos = len;
974 for (i = 0; i < len; i++)
975 td->ts.block_infos[i] =
976 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
978 td->ts.nr_block_infos = 0;
980 if (!o->size || (total_size && o->size > total_size))
981 o->size = total_size;
983 if (o->size < td_min_bs(td)) {
984 log_err("fio: blocksize too large for data set\n");
989 * See if we need to extend some files, typically needed when our
990 * target regular files don't exist yet, but our jobs require them
991 * initially due to read I/Os.
995 if (output_format & FIO_OUTPUT_NORMAL)
996 log_info("%s: Laying out IO file(s) (%u file(s) / %lluMiB)\n",
997 o->name, need_extend, extend_size >> 20);
999 for_each_file(td, f, i) {
1000 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1002 if (!fio_file_extend(f))
1005 assert(f->filetype == FIO_TYPE_FILE);
1006 fio_file_clear_extend(f);
1007 if (!o->fill_device) {
1008 old_len = f->real_file_size;
1009 extend_len = f->io_size + f->file_offset -
1012 f->real_file_size = (f->io_size + f->file_offset);
1013 err = extend_file(td, f);
1017 err = __file_invalidate_cache(td, f, old_len,
1021 * Shut up static checker
1037 o->zone_size = o->size;
1040 * iolog already set the total io size, if we read back
1043 if (!o->read_iolog_file) {
1045 td->total_io_size = o->io_size * o->loops;
1047 td->total_io_size = o->size * o->loops;
1054 td_restore_runstate(td, old_state);
1057 log_err("%s: you need to specify valid offset=\n", o->name);
1059 td_restore_runstate(td, old_state);
1063 int pre_read_files(struct thread_data *td)
1068 dprint(FD_FILE, "pre_read files\n");
1070 for_each_file(td, f, i) {
1071 pre_read_file(td, f);
1077 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1079 unsigned int range_size, seed;
1080 unsigned long nranges;
1083 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1084 fsize = min(f->real_file_size, f->io_size);
1086 nranges = (fsize + range_size - 1) / range_size;
1088 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1089 if (!td->o.rand_repeatable)
1090 seed = td->rand_seeds[4];
1092 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1093 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1094 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1095 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1096 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1097 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1102 static int init_rand_distribution(struct thread_data *td)
1108 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1111 state = td_bump_runstate(td, TD_SETTING_UP);
1113 for_each_file(td, f, i)
1114 __init_rand_distribution(td, f);
1116 td_restore_runstate(td, state);
1122 * Check if the number of blocks exceeds the randomness capability of
1123 * the selected generator. Tausworthe is 32-bit, the others are fullly
1126 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1129 if (blocks <= FRAND32_MAX)
1131 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1135 * If the user hasn't specified a random generator, switch
1136 * to tausworthe64 with informational warning. If the user did
1137 * specify one, just warn.
1139 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1142 if (!fio_option_is_set(&td->o, random_generator)) {
1143 log_info("fio: Switching to tausworthe64. Use the "
1144 "random_generator= option to get rid of this "
1146 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1151 * Just make this information to avoid breaking scripts.
1153 log_info("fio: Use the random_generator= option to switch to lfsr or "
1158 int init_random_map(struct thread_data *td)
1160 unsigned long long blocks;
1164 if (init_rand_distribution(td))
1169 for_each_file(td, f, i) {
1170 uint64_t fsize = min(f->real_file_size, f->io_size);
1172 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1174 if (check_rand_gen_limits(td, f, blocks))
1177 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1180 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1182 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1183 fio_file_set_lfsr(f);
1186 } else if (!td->o.norandommap) {
1187 f->io_axmap = axmap_new(blocks);
1189 fio_file_set_axmap(f);
1192 } else if (td->o.norandommap)
1195 if (!td->o.softrandommap) {
1196 log_err("fio: failed allocating random map. If running"
1197 " a large number of jobs, try the 'norandommap'"
1198 " option or set 'softrandommap'. Or give"
1199 " a larger --alloc-size to fio.\n");
1203 log_info("fio: file %s failed allocating random map. Running "
1204 "job without.\n", f->file_name);
1210 void close_files(struct thread_data *td)
1215 for_each_file(td, f, i) {
1216 if (fio_file_open(f))
1217 td_io_close_file(td, f);
1221 void close_and_free_files(struct thread_data *td)
1226 dprint(FD_FILE, "close files\n");
1228 for_each_file(td, f, i) {
1229 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1230 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1231 td_io_unlink_file(td, f);
1234 if (fio_file_open(f))
1235 td_io_close_file(td, f);
1237 remove_file_hash(f);
1239 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1240 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1241 td_io_unlink_file(td, f);
1244 sfree(f->file_name);
1245 f->file_name = NULL;
1246 if (fio_file_axmap(f)) {
1247 axmap_free(f->io_axmap);
1253 td->o.filename = NULL;
1255 free(td->file_locks);
1256 td->files_index = 0;
1258 td->file_locks = NULL;
1259 td->o.file_lock_mode = FILE_LOCK_NONE;
1263 static void get_file_type(struct fio_file *f)
1267 if (!strcmp(f->file_name, "-"))
1268 f->filetype = FIO_TYPE_PIPE;
1270 f->filetype = FIO_TYPE_FILE;
1273 /* \\.\ is the device namespace in Windows, where every file is
1275 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1276 f->filetype = FIO_TYPE_BLOCK;
1279 if (!stat(f->file_name, &sb)) {
1280 if (S_ISBLK(sb.st_mode))
1281 f->filetype = FIO_TYPE_BLOCK;
1282 else if (S_ISCHR(sb.st_mode))
1283 f->filetype = FIO_TYPE_CHAR;
1284 else if (S_ISFIFO(sb.st_mode))
1285 f->filetype = FIO_TYPE_PIPE;
1289 static bool __is_already_allocated(const char *fname, bool set)
1291 struct flist_head *entry;
1294 ret = file_bloom_exists(fname, set);
1298 flist_for_each(entry, &filename_list) {
1299 struct file_name *fn;
1301 fn = flist_entry(entry, struct file_name, list);
1303 if (!strcmp(fn->filename, fname))
1310 static bool is_already_allocated(const char *fname)
1314 fio_file_hash_lock();
1315 ret = __is_already_allocated(fname, false);
1316 fio_file_hash_unlock();
1321 static void set_already_allocated(const char *fname)
1323 struct file_name *fn;
1325 fn = malloc(sizeof(struct file_name));
1326 fn->filename = strdup(fname);
1328 fio_file_hash_lock();
1329 if (!__is_already_allocated(fname, true)) {
1330 flist_add_tail(&fn->list, &filename_list);
1333 fio_file_hash_unlock();
1341 static void free_already_allocated(void)
1343 struct flist_head *entry, *tmp;
1344 struct file_name *fn;
1346 if (flist_empty(&filename_list))
1349 fio_file_hash_lock();
1350 flist_for_each_safe(entry, tmp, &filename_list) {
1351 fn = flist_entry(entry, struct file_name, list);
1353 flist_del(&fn->list);
1357 fio_file_hash_unlock();
1360 static struct fio_file *alloc_new_file(struct thread_data *td)
1364 f = smalloc(sizeof(*f));
1372 fio_file_reset(td, f);
1376 bool exists_and_not_regfile(const char *filename)
1380 if (lstat(filename, &sb) == -1)
1383 #ifndef WIN32 /* NOT Windows */
1384 if (S_ISREG(sb.st_mode))
1387 /* \\.\ is the device namespace in Windows, where every file
1388 * is a device node */
1389 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1396 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1398 int cur_files = td->files_index;
1399 char file_name[PATH_MAX];
1403 dprint(FD_FILE, "add file %s\n", fname);
1405 if (td->o.directory)
1406 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1407 td->o.unique_filename);
1409 sprintf(file_name + len, "%s", fname);
1411 /* clean cloned siblings using existing files */
1412 if (numjob && is_already_allocated(file_name) &&
1413 !exists_and_not_regfile(fname))
1416 f = alloc_new_file(td);
1418 if (td->files_size <= td->files_index) {
1419 unsigned int new_size = td->o.nr_files + 1;
1421 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1423 td->files = realloc(td->files, new_size * sizeof(f));
1424 if (td->files == NULL) {
1425 log_err("fio: realloc OOM\n");
1428 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1429 td->file_locks = realloc(td->file_locks, new_size);
1430 if (!td->file_locks) {
1431 log_err("fio: realloc OOM\n");
1434 td->file_locks[cur_files] = FILE_LOCK_NONE;
1436 td->files_size = new_size;
1438 td->files[cur_files] = f;
1439 f->fileno = cur_files;
1442 * init function, io engine may not be loaded yet
1444 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1445 f->real_file_size = -1ULL;
1447 f->file_name = smalloc_strdup(file_name);
1453 switch (td->o.file_lock_mode) {
1454 case FILE_LOCK_NONE:
1456 case FILE_LOCK_READWRITE:
1457 f->rwlock = fio_rwlock_init();
1459 case FILE_LOCK_EXCLUSIVE:
1460 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1463 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1468 if (f->filetype == FIO_TYPE_FILE)
1469 td->nr_normal_files++;
1471 set_already_allocated(file_name);
1476 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1482 int add_file_exclusive(struct thread_data *td, const char *fname)
1487 for_each_file(td, f, i) {
1488 if (!strcmp(f->file_name, fname))
1492 return add_file(td, fname, 0, 1);
1495 void get_file(struct fio_file *f)
1497 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1498 assert(fio_file_open(f));
1502 int put_file(struct thread_data *td, struct fio_file *f)
1504 int f_ret = 0, ret = 0;
1506 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1508 if (!fio_file_open(f)) {
1509 assert(f->fd == -1);
1513 assert(f->references);
1514 if (--f->references)
1517 if (should_fsync(td) && td->o.fsync_on_close) {
1518 f_ret = fsync(f->fd);
1523 if (td->io_ops->close_file)
1524 ret = td->io_ops->close_file(td, f);
1529 td->nr_open_files--;
1530 fio_file_clear_open(f);
1531 assert(f->fd == -1);
1535 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1537 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1540 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1541 if (ddir == DDIR_READ)
1542 fio_rwlock_read(f->rwlock);
1544 fio_rwlock_write(f->rwlock);
1545 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1546 fio_mutex_down(f->lock);
1548 td->file_locks[f->fileno] = td->o.file_lock_mode;
1551 void unlock_file(struct thread_data *td, struct fio_file *f)
1553 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1556 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1557 fio_rwlock_unlock(f->rwlock);
1558 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1559 fio_mutex_up(f->lock);
1561 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1564 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1566 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1568 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1572 static int recurse_dir(struct thread_data *td, const char *dirname)
1578 D = opendir(dirname);
1580 char buf[FIO_VERROR_SIZE];
1582 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1583 td_verror(td, errno, buf);
1587 while ((dir = readdir(D)) != NULL) {
1588 char full_path[PATH_MAX];
1591 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1594 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1596 if (lstat(full_path, &sb) == -1) {
1597 if (errno != ENOENT) {
1598 td_verror(td, errno, "stat");
1604 if (S_ISREG(sb.st_mode)) {
1605 add_file(td, full_path, 0, 1);
1608 if (!S_ISDIR(sb.st_mode))
1611 ret = recurse_dir(td, full_path);
1620 int add_dir_files(struct thread_data *td, const char *path)
1622 int ret = recurse_dir(td, path);
1625 log_info("fio: opendir added %d files\n", td->o.nr_files);
1630 void dup_files(struct thread_data *td, struct thread_data *org)
1635 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1640 td->files = malloc(org->files_index * sizeof(f));
1642 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1643 td->file_locks = malloc(org->files_index);
1645 for_each_file(org, f, i) {
1646 struct fio_file *__f;
1648 __f = alloc_new_file(td);
1651 __f->file_name = smalloc_strdup(f->file_name);
1652 if (!__f->file_name)
1655 __f->filetype = f->filetype;
1658 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1659 __f->lock = f->lock;
1660 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1661 __f->rwlock = f->rwlock;
1668 * Returns the index that matches the filename, or -1 if not there
1670 int get_fileno(struct thread_data *td, const char *fname)
1675 for_each_file(td, f, i)
1676 if (!strcmp(f->file_name, fname))
1683 * For log usage, where we add/open/close files automatically
1685 void free_release_files(struct thread_data *td)
1689 td->o.open_files = 0;
1690 td->files_index = 0;
1691 td->nr_normal_files = 0;
1694 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1698 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1699 f->last_pos[i] = f->file_offset;
1700 f->last_start[i] = -1ULL;
1703 if (fio_file_axmap(f))
1704 axmap_reset(f->io_axmap);
1705 else if (fio_file_lfsr(f))
1706 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1709 bool fio_files_done(struct thread_data *td)
1714 for_each_file(td, f, i)
1715 if (!fio_file_done(f))
1721 /* free memory used in initialization phase only */
1722 void filesetup_mem_free(void)
1724 free_already_allocated();