Merge branch 'master' of ssh://git.kernel.dk/data/git/fio
[fio.git] / filesetup.c
index 422d706a0331b93730631465ef1ab355f0c661a7..7dceddbbf4dc21077d312dace3fa2321ba04f037 100644 (file)
@@ -11,6 +11,7 @@
 #include "fio.h"
 #include "smalloc.h"
 #include "filehash.h"
+#include "options.h"
 #include "os/os.h"
 #include "hash.h"
 #include "lib/axmap.h"
@@ -21,6 +22,8 @@
 
 static int root_warn;
 
+static FLIST_HEAD(filename_list);
+
 static inline void clear_error(struct thread_data *td)
 {
        td->error = 0;
@@ -35,7 +38,7 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        int r, new_layout = 0, unlink_file = 0, flags;
        unsigned long long left;
        unsigned int bs;
-       char *b;
+       char *b = NULL;
 
        if (read_only) {
                log_err("fio: refusing extend of file due to read-only\n");
@@ -47,15 +50,16 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
         * does that for operations involving reads, or for writes
         * where overwrite is set
         */
-       if (td_read(td) || (td_write(td) && td->o.overwrite) ||
+       if (td_read(td) ||
+          (td_write(td) && td->o.overwrite && !td->o.file_append) ||
            (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
                new_layout = 1;
-       if (td_write(td) && !td->o.overwrite)
+       if (td_write(td) && !td->o.overwrite && !td->o.file_append)
                unlink_file = 1;
 
        if (unlink_file || new_layout) {
                dprint(FD_FILE, "layout unlink %s\n", f->file_name);
-               if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
+               if ((td_io_unlink_file(td, f) < 0) && (errno != ENOENT)) {
                        td_verror(td, errno, "unlink");
                        return 1;
                }
@@ -65,6 +69,10 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        if (new_layout)
                flags |= O_TRUNC;
 
+#ifdef WIN32
+       flags |= _O_BINARY;
+#endif
+
        dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
        f->fd = open(f->file_name, flags, 0644);
        if (f->fd < 0) {
@@ -121,13 +129,14 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
                                        (unsigned long long) f->real_file_size);
                if (ftruncate(f->fd, f->real_file_size) == -1) {
-                       td_verror(td, errno, "ftruncate");
-                       goto err;
+                       if (errno != EFBIG) {
+                               td_verror(td, errno, "ftruncate");
+                               goto err;
+                       }
                }
        }
 
        b = malloc(td->o.max_bs[DDIR_WRITE]);
-       memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 
        left = f->real_file_size;
        while (left && !td->terminate) {
@@ -135,6 +144,8 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                if (bs > left)
                        bs = left;
 
+               fill_io_buffer(td, b, bs, bs);
+
                r = write(f->fd, b, bs);
 
                if (r > 0) {
@@ -161,7 +172,7 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
 
        if (td->terminate) {
                dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
-               unlink(f->file_name);
+               td_io_unlink_file(td, f);
        } else if (td->o.create_fsync) {
                if (fsync(f->fd) < 0) {
                        td_verror(td, errno, "fsync");
@@ -182,12 +193,14 @@ done:
 err:
        close(f->fd);
        f->fd = -1;
+       if (b)
+               free(b);
        return 1;
 }
 
 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 {
-       int r, did_open = 0, old_runstate;
+       int ret = 0, r, did_open = 0, old_runstate;
        unsigned long long left;
        unsigned int bs;
        char *b;
@@ -203,14 +216,19 @@ static int pre_read_file(struct thread_data *td, struct fio_file *f)
                did_open = 1;
        }
 
-       old_runstate = td->runstate;
-       td_set_runstate(td, TD_PRE_READING);
+       old_runstate = td_bump_runstate(td, TD_PRE_READING);
 
        bs = td->o.max_bs[DDIR_READ];
        b = malloc(bs);
        memset(b, 0, bs);
 
-       lseek(f->fd, f->file_offset, SEEK_SET);
+       if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
+               td_verror(td, errno, "lseek");
+               log_err("fio: failed to lseek pre-read file\n");
+               ret = 1;
+               goto error;
+       }
+
        left = f->io_size;
 
        while (left && !td->terminate) {
@@ -228,12 +246,14 @@ static int pre_read_file(struct thread_data *td, struct fio_file *f)
                }
        }
 
-       td_set_runstate(td, old_runstate);
+error:
+       td_restore_runstate(td, old_runstate);
 
        if (did_open)
                td->io_ops->close_file(td, f);
+
        free(b);
-       return 0;
+       return ret;
 }
 
 static unsigned long long get_rand_file_size(struct thread_data *td)
@@ -241,16 +261,9 @@ static unsigned long long get_rand_file_size(struct thread_data *td)
        unsigned long long ret, sized;
        unsigned long r;
 
-       if (td->o.use_os_rand) {
-               r = os_random_long(&td->file_size_state);
-               sized = td->o.file_size_high - td->o.file_size_low;
-               ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
-       } else {
-               r = __rand(&td->__file_size_state);
-               sized = td->o.file_size_high - td->o.file_size_low;
-               ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
-       }
-
+       r = __rand(&td->file_size_state);
+       sized = td->o.file_size_high - td->o.file_size_low;
+       ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
        ret += td->o.file_size_low;
        ret -= (ret % td->o.rw_min_bs);
        return ret;
@@ -355,7 +368,8 @@ static int get_file_size(struct thread_data *td, struct fio_file *f)
 
        if (f->file_offset > f->real_file_size) {
                log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
-                                       f->file_offset, f->real_file_size);
+                                       (unsigned long long) f->file_offset,
+                                       (unsigned long long) f->real_file_size);
                return 1;
        }
 
@@ -369,6 +383,10 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 {
        int ret = 0;
 
+#ifdef CONFIG_ESX
+       return 0;
+#endif
+
        if (len == -1ULL)
                len = f->io_size;
        if (off == -1ULL)
@@ -380,17 +398,11 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
                                                                len);
 
-       /*
-        * FIXME: add blockdev flushing too
-        */
-       if (f->mmap_ptr) {
-               ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
-#ifdef FIO_MADV_FREE
-               (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
-#endif
-       } else if (f->filetype == FIO_TYPE_FILE) {
+       if (td->io_ops->invalidate)
+               ret = td->io_ops->invalidate(td, f);
+       else if (f->filetype == FIO_TYPE_FILE)
                ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
-       else if (f->filetype == FIO_TYPE_BD) {
+       else if (f->filetype == FIO_TYPE_BD) {
                ret = blockdev_invalidate_cache(f);
                if (ret < 0 && errno == EACCES && geteuid()) {
                        if (!root_warn) {
@@ -403,15 +415,18 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
                ret = 0;
 
-       if (ret < 0) {
-               td_verror(td, errno, "invalidate_cache");
-               return 1;
-       } else if (ret > 0) {
-               td_verror(td, ret, "invalidate_cache");
-               return 1;
+       /*
+        * Cache flushing isn't a fatal condition, and we know it will
+        * happen on some platforms where we don't have the proper
+        * function to flush eg block device caches. So just warn and
+        * continue on our way.
+        */
+       if (ret) {
+               log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
+               ret = 0;
        }
 
-       return ret;
+       return 0;
 
 }
 
@@ -441,6 +456,7 @@ int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
                f->shadow_fd = -1;
        }
 
+       f->engine_data = 0;
        return ret;
 }
 
@@ -456,15 +472,16 @@ int file_lookup_open(struct fio_file *f, int flags)
                 * racy, need the __f->lock locked
                 */
                f->lock = __f->lock;
-               f->lock_owner = __f->lock_owner;
-               f->lock_batch = __f->lock_batch;
-               f->lock_ddir = __f->lock_ddir;
                from_hash = 1;
        } else {
                dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
                from_hash = 0;
        }
 
+#ifdef WIN32
+       flags |= _O_BINARY;
+#endif
+
        f->fd = open(f->file_name, flags, 0600);
        return from_hash;
 }
@@ -518,6 +535,13 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                goto skip_flags;
        if (td->o.odirect)
                flags |= OS_O_DIRECT;
+       if (td->o.oatomic) {
+               if (!FIO_O_ATOMIC) {
+                       td_verror(td, EINVAL, "OS does not support atomic IO");
+                       return 1;
+               }
+               flags |= OS_O_DIRECT | FIO_O_ATOMIC;
+       }
        if (td->o.sync_io)
                flags |= O_SYNC;
        if (td->o.create_on_open)
@@ -572,6 +596,7 @@ open_again:
                }
 
                td_verror(td, __e, buf);
+               return 1;
        }
 
        if (!from_hash && f->fd != -1) {
@@ -583,7 +608,7 @@ open_again:
                         * work-around a "feature" on Linux, where a close of
                         * an fd that has been opened for write will trigger
                         * udev to call blkid to check partitions, fs id, etc.
-                        * That polutes the device cache, which can slow down
+                        * That pollutes the device cache, which can slow down
                         * unbuffered accesses.
                         */
                        if (f->shadow_fd == -1)
@@ -624,6 +649,7 @@ static int get_file_sizes(struct thread_data *td)
                        if (td->error != ENOENT) {
                                log_err("%s\n", td->verror);
                                err = 1;
+                               break;
                        }
                        clear_error(td);
                }
@@ -665,7 +691,8 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
                } else if (f->filetype != FIO_TYPE_FILE)
                        continue;
 
-               strcpy(buf, f->file_name);
+               buf[255] = '\0';
+               strncpy(buf, f->file_name, 255);
 
                if (stat(buf, &sb) < 0) {
                        if (errno != ENOENT)
@@ -687,8 +714,8 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
                if (fm)
                        continue;
 
-               fm = malloc(sizeof(*fm));
-               strcpy(fm->__base, buf);
+               fm = calloc(1, sizeof(*fm));
+               strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
                fm->base = basename(fm->__base);
                fm->key = sb.st_dev;
                flist_add(&fm->list, &list);
@@ -710,10 +737,15 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
        return ret;
 }
 
-uint64_t get_start_offset(struct thread_data *td)
+uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 {
+       struct thread_options *o = &td->o;
+
+       if (o->file_append && f->filetype == FIO_TYPE_FILE)
+               return f->real_file_size;
+
        return td->o.start_offset +
-               (td->thread_number - 1) * td->o.offset_increment;
+               td->subjob_number * td->o.offset_increment;
 }
 
 /*
@@ -722,13 +754,19 @@ uint64_t get_start_offset(struct thread_data *td)
 int setup_files(struct thread_data *td)
 {
        unsigned long long total_size, extend_size;
+       struct thread_options *o = &td->o;
        struct fio_file *f;
-       unsigned int i;
+       unsigned int i, nr_fs_extra = 0;
        int err = 0, need_extend;
+       int old_state;
+       const unsigned int bs = td_min_bs(td);
+       uint64_t fs = 0;
 
        dprint(FD_FILE, "setup files\n");
 
-       if (td->o.read_iolog_file)
+       old_state = td_bump_runstate(td, TD_SETTING_UP);
+
+       if (o->read_iolog_file)
                goto done;
 
        /*
@@ -742,7 +780,7 @@ int setup_files(struct thread_data *td)
                err = get_file_sizes(td);
 
        if (err)
-               return err;
+               goto err_out;
 
        /*
         * check sizes. if the files/devices do not exist and the size
@@ -756,17 +794,32 @@ int setup_files(struct thread_data *td)
                        total_size += f->real_file_size;
        }
 
-       if (td->o.fill_device)
+       if (o->fill_device)
                td->fill_device_size = get_fs_free_counts(td);
 
        /*
         * device/file sizes are zero and no size given, punt
         */
-       if ((!total_size || total_size == -1ULL) && !td->o.size &&
-           !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
-               log_err("%s: you need to specify size=\n", td->o.name);
+       if ((!total_size || total_size == -1ULL) && !o->size &&
+           !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
+           !(o->nr_files && (o->file_size_low || o->file_size_high))) {
+               log_err("%s: you need to specify size=\n", o->name);
                td_verror(td, EINVAL, "total_file_size");
-               return 1;
+               goto err_out;
+       }
+
+       /*
+        * Calculate per-file size and potential extra size for the
+        * first files, if needed.
+        */
+       if (!o->file_size_low && o->nr_files) {
+               uint64_t all_fs;
+
+               fs = o->size / o->nr_files;
+               all_fs = fs * o->nr_files;
+
+               if (all_fs < o->size)
+                       nr_fs_extra = (o->size - all_fs) / bs;
        }
 
        /*
@@ -777,29 +830,35 @@ int setup_files(struct thread_data *td)
        extend_size = total_size = 0;
        need_extend = 0;
        for_each_file(td, f, i) {
-               f->file_offset = get_start_offset(td);
+               f->file_offset = get_start_offset(td, f);
 
-               if (!td->o.file_size_low) {
+               if (!o->file_size_low) {
                        /*
                         * no file size range given, file size is equal to
-                        * total size divided by number of files. if that is
-                        * zero, set it to the real file size.
+                        * total size divided by number of files. If that is
+                        * zero, set it to the real file size. If the size
+                        * doesn't divide nicely with the min blocksize,
+                        * make the first files bigger.
                         */
-                       f->io_size = td->o.size / td->o.nr_files;
+                       f->io_size = fs;
+                       if (nr_fs_extra) {
+                               nr_fs_extra--;
+                               f->io_size += bs;
+                       }
+
                        if (!f->io_size)
                                f->io_size = f->real_file_size - f->file_offset;
-               } else if (f->real_file_size < td->o.file_size_low ||
-                          f->real_file_size > td->o.file_size_high) {
-                       if (f->file_offset > td->o.file_size_low)
+               } else if (f->real_file_size < o->file_size_low ||
+                          f->real_file_size > o->file_size_high) {
+                       if (f->file_offset > o->file_size_low)
                                goto err_offset;
                        /*
                         * file size given. if it's fixed, use that. if it's a
                         * range, generate a random size in-between.
                         */
-                       if (td->o.file_size_low == td->o.file_size_high) {
-                               f->io_size = td->o.file_size_low
-                                               - f->file_offset;
-                       } else {
+                       if (o->file_size_low == o->file_size_high)
+                               f->io_size = o->file_size_low - f->file_offset;
+                       else {
                                f->io_size = get_rand_file_size(td)
                                                - f->file_offset;
                        }
@@ -809,15 +868,15 @@ int setup_files(struct thread_data *td)
                if (f->io_size == -1ULL)
                        total_size = -1ULL;
                else {
-                        if (td->o.size_percent)
-                                f->io_size = (f->io_size * td->o.size_percent) / 100;
+                        if (o->size_percent)
+                                f->io_size = (f->io_size * o->size_percent) / 100;
                        total_size += f->io_size;
                }
 
                if (f->filetype == FIO_TYPE_FILE &&
                    (f->io_size + f->file_offset) > f->real_file_size &&
                    !(td->io_ops->flags & FIO_DISKLESSIO)) {
-                       if (!td->o.create_on_open) {
+                       if (!o->create_on_open) {
                                need_extend++;
                                extend_size += (f->io_size + f->file_offset);
                        } else
@@ -826,8 +885,13 @@ int setup_files(struct thread_data *td)
                }
        }
 
-       if (!td->o.size || td->o.size > total_size)
-               td->o.size = total_size;
+       if (!o->size || (total_size && o->size > total_size))
+               o->size = total_size;
+
+       if (o->size < td_min_bs(td)) {
+               log_err("fio: blocksize too large for data set\n");
+               goto err_out;
+       }
 
        /*
         * See if we need to extend some files
@@ -836,7 +900,7 @@ int setup_files(struct thread_data *td)
                temp_stall_ts = 1;
                if (output_format == FIO_OUTPUT_NORMAL)
                        log_info("%s: Laying out IO file(s) (%u file(s) /"
-                                " %lluMB)\n", td->o.name, need_extend,
+                                " %lluMB)\n", o->name, need_extend,
                                        extend_size >> 20);
 
                for_each_file(td, f, i) {
@@ -847,7 +911,7 @@ int setup_files(struct thread_data *td)
 
                        assert(f->filetype == FIO_TYPE_FILE);
                        fio_file_clear_extend(f);
-                       if (!td->o.fill_device) {
+                       if (!o->fill_device) {
                                old_len = f->real_file_size;
                                extend_len = f->io_size + f->file_offset -
                                                old_len;
@@ -859,7 +923,13 @@ int setup_files(struct thread_data *td)
 
                        err = __file_invalidate_cache(td, f, old_len,
                                                                extend_len);
-                       close(f->fd);
+
+                       /*
+                        * Shut up static checker
+                        */
+                       if (f->fd != -1)
+                               close(f->fd);
+
                        f->fd = -1;
                        if (err)
                                break;
@@ -868,25 +938,32 @@ int setup_files(struct thread_data *td)
        }
 
        if (err)
-               return err;
+               goto err_out;
 
-       if (!td->o.zone_size)
-               td->o.zone_size = td->o.size;
+       if (!o->zone_size)
+               o->zone_size = o->size;
 
        /*
         * iolog already set the total io size, if we read back
         * stored entries.
         */
-       if (!td->o.read_iolog_file)
-               td->total_io_size = td->o.size * td->o.loops;
+       if (!o->read_iolog_file) {
+               if (o->io_limit)
+                       td->total_io_size = o->io_limit * o->loops;
+               else
+                       td->total_io_size = o->size * o->loops;
+       }
 
 done:
-       if (td->o.create_only)
+       if (o->create_only)
                td->done = 1;
 
+       td_restore_runstate(td, old_state);
        return 0;
 err_offset:
-       log_err("%s: you need to specify valid offset=\n", td->o.name);
+       log_err("%s: you need to specify valid offset=\n", o->name);
+err_out:
+       td_restore_runstate(td, old_state);
        return 1;
 }
 
@@ -908,12 +985,12 @@ static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 {
        unsigned int range_size, seed;
        unsigned long nranges;
-       uint64_t file_size;
+       uint64_t fsize;
 
        range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
-       file_size = min(f->real_file_size, f->io_size);
+       fsize = min(f->real_file_size, f->io_size);
 
-       nranges = (file_size + range_size - 1) / range_size;
+       nranges = (fsize + range_size - 1) / range_size;
 
        seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
        if (!td->o.rand_repeatable)
@@ -921,8 +998,10 @@ static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 
        if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
                zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
-       else
+       else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
                pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
+       else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+               gauss_init(&f->gauss, nranges, td->o.gauss_dev, seed);
 
        return 1;
 }
@@ -936,11 +1015,12 @@ static int init_rand_distribution(struct thread_data *td)
        if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
                return 0;
 
-       state = td->runstate;
-       td_set_runstate(td, TD_SETTING_UP);
+       state = td_bump_runstate(td, TD_SETTING_UP);
+
        for_each_file(td, f, i)
                __init_rand_distribution(td, f);
-       td_set_runstate(td, state);
+
+       td_restore_runstate(td, state);
 
        return 1;
 }
@@ -957,21 +1037,25 @@ int init_random_map(struct thread_data *td)
                return 0;
 
        for_each_file(td, f, i) {
-               uint64_t file_size = min(f->real_file_size, f->io_size);
+               uint64_t fsize = min(f->real_file_size, f->io_size);
 
-               blocks = file_size / (unsigned long long) td->o.rw_min_bs;
+               blocks = fsize / (unsigned long long) td->o.rw_min_bs;
 
                if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
                        unsigned long seed;
 
                        seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
-                       
-                       if (!lfsr_init(&f->lfsr, blocks, seed))
+
+                       if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
+                               fio_file_set_lfsr(f);
                                continue;
+                       }
                } else if (!td->o.norandommap) {
                        f->io_axmap = axmap_new(blocks);
-                       if (f->io_axmap)
+                       if (f->io_axmap) {
+                               fio_file_set_axmap(f);
                                continue;
+                       }
                } else if (td->o.norandommap)
                        continue;
 
@@ -1011,7 +1095,7 @@ void close_and_free_files(struct thread_data *td)
        for_each_file(td, f, i) {
                if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
                        dprint(FD_FILE, "free unlink %s\n", f->file_name);
-                       unlink(f->file_name);
+                       td_io_unlink_file(td, f);
                }
 
                if (fio_file_open(f))
@@ -1019,17 +1103,27 @@ void close_and_free_files(struct thread_data *td)
 
                remove_file_hash(f);
 
+               if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
+                       dprint(FD_FILE, "free unlink %s\n", f->file_name);
+                       td_io_unlink_file(td, f);
+               }
+
                sfree(f->file_name);
                f->file_name = NULL;
-               axmap_free(f->io_axmap);
-               f->io_axmap = NULL;
+               if (fio_file_axmap(f)) {
+                       axmap_free(f->io_axmap);
+                       f->io_axmap = NULL;
+               }
                sfree(f);
        }
 
        td->o.filename = NULL;
        free(td->files);
+       free(td->file_locks);
        td->files_index = 0;
        td->files = NULL;
+       td->file_locks = NULL;
+       td->o.file_lock_mode = FILE_LOCK_NONE;
        td->o.nr_files = 0;
 }
 
@@ -1057,24 +1151,110 @@ static void get_file_type(struct fio_file *f)
        }
 }
 
-int add_file(struct thread_data *td, const char *fname)
+static int __is_already_allocated(const char *fname)
 {
-       int cur_files = td->files_index;
-       char file_name[PATH_MAX];
-       struct fio_file *f;
-       int len = 0;
+       struct flist_head *entry;
+       char *filename;
 
-       dprint(FD_FILE, "add file %s\n", fname);
+       if (flist_empty(&filename_list))
+               return 0;
+
+       flist_for_each(entry, &filename_list) {
+               filename = flist_entry(entry, struct file_name, list)->filename;
+
+               if (strcmp(filename, fname) == 0)
+                       return 1;
+       }
+
+       return 0;
+}
+
+static int is_already_allocated(const char *fname)
+{
+       int ret;
+
+       fio_file_hash_lock();
+       ret = __is_already_allocated(fname);
+       fio_file_hash_unlock();
+       return ret;
+}
+
+static void set_already_allocated(const char *fname)
+{
+       struct file_name *fn;
+
+       fn = malloc(sizeof(struct file_name));
+       fn->filename = strdup(fname);
+
+       fio_file_hash_lock();
+       if (!__is_already_allocated(fname)) {
+               flist_add_tail(&fn->list, &filename_list);
+               fn = NULL;
+       }
+       fio_file_hash_unlock();
+
+       if (fn) {
+               free(fn->filename);
+               free(fn);
+       }
+}
+
+
+static void free_already_allocated(void)
+{
+       struct flist_head *entry, *tmp;
+       struct file_name *fn;
+
+       if (flist_empty(&filename_list))
+               return;
+
+       fio_file_hash_lock();
+       flist_for_each_safe(entry, tmp, &filename_list) {
+               fn = flist_entry(entry, struct file_name, list);
+               free(fn->filename);
+               flist_del(&fn->list);
+               free(fn);
+       }
+
+       fio_file_hash_unlock();
+}
+
+static struct fio_file *alloc_new_file(struct thread_data *td)
+{
+       struct fio_file *f;
 
        f = smalloc(sizeof(*f));
        if (!f) {
                log_err("fio: smalloc OOM\n");
                assert(0);
+               return NULL;
        }
 
        f->fd = -1;
        f->shadow_fd = -1;
        fio_file_reset(td, f);
+       return f;
+}
+
+int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
+{
+       int cur_files = td->files_index;
+       char file_name[PATH_MAX];
+       struct fio_file *f;
+       int len = 0;
+
+       dprint(FD_FILE, "add file %s\n", fname);
+
+       if (td->o.directory)
+               len = set_name_idx(file_name, td->o.directory, numjob);
+
+       sprintf(file_name + len, "%s", fname);
+
+       /* clean cloned siblings using existing files */
+       if (numjob && is_already_allocated(file_name))
+               return 0;
+
+       f = alloc_new_file(td);
 
        if (td->files_size <= td->files_index) {
                unsigned int new_size = td->o.nr_files + 1;
@@ -1086,6 +1266,14 @@ int add_file(struct thread_data *td, const char *fname)
                        log_err("fio: realloc OOM\n");
                        assert(0);
                }
+               if (td->o.file_lock_mode != FILE_LOCK_NONE) {
+                       td->file_locks = realloc(td->file_locks, new_size);
+                       if (!td->file_locks) {
+                               log_err("fio: realloc OOM\n");
+                               assert(0);
+                       }
+                       td->file_locks[cur_files] = FILE_LOCK_NONE;
+               }
                td->files_size = new_size;
        }
        td->files[cur_files] = f;
@@ -1097,10 +1285,6 @@ int add_file(struct thread_data *td, const char *fname)
        if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
                f->real_file_size = -1ULL;
 
-       if (td->o.directory)
-               len = sprintf(file_name, "%s/", td->o.directory);
-
-       sprintf(file_name + len, "%s", fname);
        f->file_name = smalloc_strdup(file_name);
        if (!f->file_name) {
                log_err("fio: smalloc OOM\n");
@@ -1113,7 +1297,7 @@ int add_file(struct thread_data *td, const char *fname)
        case FILE_LOCK_NONE:
                break;
        case FILE_LOCK_READWRITE:
-               f->lock = fio_mutex_rw_init();
+               f->rwlock = fio_rwlock_init();
                break;
        case FILE_LOCK_EXCLUSIVE:
                f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
@@ -1127,6 +1311,11 @@ int add_file(struct thread_data *td, const char *fname)
        if (f->filetype == FIO_TYPE_FILE)
                td->nr_normal_files++;
 
+       set_already_allocated(file_name);
+
+       if (inc)
+               td->o.nr_files++;
+
        dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
                                                        cur_files);
 
@@ -1143,7 +1332,7 @@ int add_file_exclusive(struct thread_data *td, const char *fname)
                        return i;
        }
 
-       return add_file(td, fname);
+       return add_file(td, fname, 0, 1);
 }
 
 void get_file(struct fio_file *f)
@@ -1168,8 +1357,11 @@ int put_file(struct thread_data *td, struct fio_file *f)
        if (--f->references)
                return 0;
 
-       if (should_fsync(td) && td->o.fsync_on_close)
+       if (should_fsync(td) && td->o.fsync_on_close) {
                f_ret = fsync(f->fd);
+               if (f_ret < 0)
+                       f_ret = errno;
+       }
 
        if (td->io_ops->close_file)
                ret = td->io_ops->close_file(td, f);
@@ -1188,57 +1380,36 @@ void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
        if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
                return;
 
-       if (f->lock_owner == td && f->lock_batch--)
-               return;
-
        if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
                if (ddir == DDIR_READ)
-                       fio_mutex_down_read(f->lock);
+                       fio_rwlock_read(f->rwlock);
                else
-                       fio_mutex_down_write(f->lock);
+                       fio_rwlock_write(f->rwlock);
        } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
                fio_mutex_down(f->lock);
 
-       f->lock_owner = td;
-       f->lock_batch = td->o.lockfile_batch;
-       f->lock_ddir = ddir;
+       td->file_locks[f->fileno] = td->o.file_lock_mode;
 }
 
 void unlock_file(struct thread_data *td, struct fio_file *f)
 {
        if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
                return;
-       if (f->lock_batch)
-               return;
-
-       if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
-               const int is_read = f->lock_ddir == DDIR_READ;
-               int val = fio_mutex_getval(f->lock);
-
-               if ((is_read && val == 1) || (!is_read && val == -1))
-                       f->lock_owner = NULL;
-
-               if (is_read)
-                       fio_mutex_up_read(f->lock);
-               else
-                       fio_mutex_up_write(f->lock);
-       } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
-               int val = fio_mutex_getval(f->lock);
-
-               if (val == 0)
-                       f->lock_owner = NULL;
 
+       if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
+               fio_rwlock_unlock(f->rwlock);
+       else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
                fio_mutex_up(f->lock);
-       }
+
+       td->file_locks[f->fileno] = FILE_LOCK_NONE;
 }
 
 void unlock_file_all(struct thread_data *td, struct fio_file *f)
 {
-       if (f->lock_owner != td)
+       if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
                return;
-
-       f->lock_batch = 0;
-       unlock_file(td, f);
+       if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
+               unlock_file(td, f);
 }
 
 static int recurse_dir(struct thread_data *td, const char *dirname)
@@ -1268,13 +1439,13 @@ static int recurse_dir(struct thread_data *td, const char *dirname)
                if (lstat(full_path, &sb) == -1) {
                        if (errno != ENOENT) {
                                td_verror(td, errno, "stat");
-                               return 1;
+                               ret = 1;
+                               break;
                        }
                }
 
                if (S_ISREG(sb.st_mode)) {
-                       add_file(td, full_path);
-                       td->o.nr_files++;
+                       add_file(td, full_path, 0, 1);
                        continue;
                }
                if (!S_ISDIR(sb.st_mode))
@@ -1311,16 +1482,13 @@ void dup_files(struct thread_data *td, struct thread_data *org)
 
        td->files = malloc(org->files_index * sizeof(f));
 
+       if (td->o.file_lock_mode != FILE_LOCK_NONE)
+               td->file_locks = malloc(org->files_index);
+
        for_each_file(org, f, i) {
                struct fio_file *__f;
 
-               __f = smalloc(sizeof(*__f));
-               if (!__f) {
-                       log_err("fio: smalloc OOM\n");
-                       assert(0);
-               }
-               __f->fd = -1;
-               fio_file_reset(td, __f);
+               __f = alloc_new_file(td);
 
                if (f->file_name) {
                        __f->file_name = smalloc_strdup(f->file_name);
@@ -1332,6 +1500,11 @@ void dup_files(struct thread_data *td, struct thread_data *org)
                        __f->filetype = f->filetype;
                }
 
+               if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
+                       __f->lock = f->lock;
+               else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
+                       __f->rwlock = f->rwlock;
+
                td->files[i] = __f;
        }
 }
@@ -1357,16 +1530,41 @@ int get_fileno(struct thread_data *td, const char *fname)
 void free_release_files(struct thread_data *td)
 {
        close_files(td);
+       td->o.nr_files = 0;
+       td->o.open_files = 0;
        td->files_index = 0;
        td->nr_normal_files = 0;
 }
 
 void fio_file_reset(struct thread_data *td, struct fio_file *f)
 {
-       f->last_pos = f->file_offset;
-       f->last_start = -1ULL;
-       if (f->io_axmap)
+       int i;
+
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               f->last_pos[i] = f->file_offset;
+               f->last_start[i] = -1ULL;
+       }
+
+       if (fio_file_axmap(f))
                axmap_reset(f->io_axmap);
-       if (td->o.random_generator == FIO_RAND_GEN_LFSR)
+       else if (fio_file_lfsr(f))
                lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
 }
+
+int fio_files_done(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       for_each_file(td, f, i)
+               if (!fio_file_done(f))
+                       return 0;
+
+       return 1;
+}
+
+/* free memory used in initialization phase only */
+void filesetup_mem_free(void)
+{
+       free_already_allocated();
+}