Add strong madvise() hint for cache pruning
[fio.git] / filesetup.c
index 7a9a2afed29b34d345e1628156c2d3ac6848163c..f8f81d2b2a54e9a7692cdb41289ac8c5ff472c2c 100644 (file)
@@ -8,7 +8,20 @@
 #include <sys/types.h>
 
 #include "fio.h"
+#include "smalloc.h"
+#include "filehash.h"
 
+static int root_warn;
+
+static inline void clear_error(struct thread_data *td)
+{
+       td->error = 0;
+       td->verror[0] = '\0';
+}
+
+/*
+ * Leaves f->fd open on success, caller must close
+ */
 static int extend_file(struct thread_data *td, struct fio_file *f)
 {
        int r, new_layout = 0, unlink_file = 0, flags;
@@ -16,18 +29,25 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        unsigned int bs;
        char *b;
 
+       if (read_only) {
+               log_err("fio: refusing extend of file due to read-only\n");
+               return 0;
+       }
+
        /*
         * check if we need to lay the file out complete again. fio
         * 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_write(td) && td->io_ops->flags & FIO_NOEXTEND))
                new_layout = 1;
        if (td_write(td) && !td->o.overwrite)
                unlink_file = 1;
 
-       if ((unlink_file || new_layout) && (f->flags & FIO_FILE_EXISTS)) {
-               if (unlink(f->file_name) < 0) {
+       if (unlink_file || new_layout) {
+               dprint(FD_FILE, "layout unlink %s\n", f->file_name);
+               if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
                        td_verror(td, errno, "unlink");
                        return 1;
                }
@@ -37,23 +57,37 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        if (new_layout)
                flags |= O_TRUNC;
 
+       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) {
                td_verror(td, errno, "open");
                return 1;
        }
 
-       if (ftruncate(f->fd, f->real_file_size) == -1) {
-               td_verror(td, errno, "ftruncate");
-               goto err;
-       }
-
        if (!new_layout)
                goto done;
 
-       if (posix_fallocate(f->fd, 0, f->real_file_size) < 0) {
-               td_verror(td, errno, "posix_fallocate");
-               goto err;
+       /*
+        * The size will be -1ULL when fill_device is used, so don't truncate
+        * or fallocate this file, just write it
+        */
+       if (!td->o.fill_device) {
+               dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
+                                                       f->real_file_size);
+               if (ftruncate(f->fd, f->real_file_size) == -1) {
+                       td_verror(td, errno, "ftruncate");
+                       goto err;
+               }
+
+#ifdef FIO_HAVE_FALLOCATE
+               dprint(FD_FILE, "fallocate file %s, size %llu\n", f->file_name,
+                                                       f->real_file_size);
+               r = posix_fallocate(f->fd, 0, f->real_file_size);
+               if (r < 0) {
+                       log_err("fio: posix_fallocate fails: %s\n",
+                                       strerror(-r));
+               }
+#endif
        }
 
        b = malloc(td->o.max_bs[DDIR_WRITE]);
@@ -67,28 +101,47 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
 
                r = write(f->fd, b, bs);
 
-               if (r == (int) bs) {
-                       left -= bs;
+               if (r > 0) {
+                       left -= r;
                        continue;
                } else {
-                       if (r < 0)
+                       if (r < 0) {
+                               int __e = errno;
+
+                               if (__e == ENOSPC) {
+                                       if (td->o.fill_device)
+                                               break;
+                                       log_info("fio: ENOSPC on laying out "
+                                                "file, stopping\n");
+                                       break;
+                               }
                                td_verror(td, errno, "write");
-                       else
+                       else
                                td_verror(td, EIO, "write");
 
                        break;
                }
        }
 
-       if (td->terminate)
+       if (td->terminate) {
+               dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
                unlink(f->file_name);
-       else if (td->o.create_fsync)
-               fsync(f->fd);
+       } else if (td->o.create_fsync) {
+               if (fsync(f->fd) < 0) {
+                       td_verror(td, errno, "fsync");
+                       goto err;
+               }
+       }
+       if (td->o.fill_device && !td_write(td)) {
+               fio_file_clear_size_known(f);
+               if (td_io_get_file_size(td, f))
+                       goto err;
+               if (f->io_size > f->real_file_size)
+                       f->io_size = f->real_file_size;
+       }
 
        free(b);
 done:
-       close(f->fd);
-       f->fd = -1;
        return 0;
 err:
        close(f->fd);
@@ -96,13 +149,66 @@ err:
        return 1;
 }
 
+static int pre_read_file(struct thread_data *td, struct fio_file *f)
+{
+       int r, did_open = 0, old_runstate;
+       unsigned long long left;
+       unsigned int bs;
+       char *b;
+
+       if (td->io_ops->flags & FIO_PIPEIO)
+               return 0;
+
+       if (!fio_file_open(f)) {
+               if (td->io_ops->open_file(td, f)) {
+                       log_err("fio: cannot pre-read, failed to open file\n");
+                       return 1;
+               }
+               did_open = 1;
+       }
+
+       old_runstate = td->runstate;
+       td_set_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);
+       left = f->io_size;
+
+       while (left && !td->terminate) {
+               if (bs > left)
+                       bs = left;
+
+               r = read(f->fd, b, bs);
+
+               if (r == (int) bs) {
+                       left -= bs;
+                       continue;
+               } else {
+                       td_verror(td, EIO, "pre_read");
+                       break;
+               }
+       }
+
+       td_set_runstate(td, old_runstate);
+
+       if (did_open)
+               td->io_ops->close_file(td, f);
+       free(b);
+       return 0;
+}
+
 static unsigned long long get_rand_file_size(struct thread_data *td)
 {
-       unsigned long long ret;
+       unsigned long long ret, sized;
        long r;
 
        r = os_random_long(&td->file_size_state);
-       ret = td->o.file_size_low + (unsigned long long) ((double) td->o.file_size_high * (r / (RAND_MAX + 1.0)));
+       sized = td->o.file_size_high - td->o.file_size_low;
+       ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
+       ret += td->o.file_size_low;
        ret -= (ret % td->o.rw_min_bs);
        return ret;
 }
@@ -111,7 +217,7 @@ static int file_size(struct thread_data *td, struct fio_file *f)
 {
        struct stat st;
 
-       if (fstat(f->fd, &st) == -1) {
+       if (stat(f->file_name, &st) == -1) {
                td_verror(td, errno, "fstat");
                return 1;
        }
@@ -125,21 +231,36 @@ static int bdev_size(struct thread_data *td, struct fio_file *f)
        unsigned long long bytes;
        int r;
 
+       if (td->io_ops->open_file(td, f)) {
+               log_err("fio: failed opening blockdev %s for size check\n",
+                       f->file_name);
+               return 1;
+       }
+
        r = blockdev_size(f->fd, &bytes);
        if (r) {
                td_verror(td, r, "blockdev_size");
-               return 1;
+               goto err;
+       }
+
+       if (!bytes) {
+               log_err("%s: zero sized block device?\n", f->file_name);
+               goto err;
        }
 
        f->real_file_size = bytes;
+       td->io_ops->close_file(td, f);
        return 0;
+err:
+       td->io_ops->close_file(td, f);
+       return 1;
 }
 
 static int get_file_size(struct thread_data *td, struct fio_file *f)
 {
        int ret = 0;
 
-       if (f->flags & FIO_SIZE_KNOWN)
+       if (fio_file_size_known(f))
                return 0;
 
        if (f->filetype == FIO_TYPE_FILE)
@@ -153,32 +274,50 @@ static int get_file_size(struct thread_data *td, struct fio_file *f)
                return ret;
 
        if (f->file_offset > f->real_file_size) {
-               log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name, f->file_offset, f->real_file_size);
+               log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name,
+                                       f->file_offset, f->real_file_size);
                return 1;
        }
 
-       f->flags |= FIO_SIZE_KNOWN;
+       fio_file_set_size_known(f);
        return 0;
 }
 
-int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
+static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
+                                  unsigned long long off,
+                                  unsigned long long len)
 {
        int ret = 0;
 
-       if (td->o.odirect)
+       if (len == -1ULL)
+               len = f->io_size;
+       if (off == -1ULL)
+               off = f->file_offset;
+
+       if (len == -1ULL || off == -1ULL)
                return 0;
 
+       dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
+                                                               len);
+
        /*
         * FIXME: add blockdev flushing too
         */
-       if (f->mmap)
-               ret = madvise(f->mmap, f->io_size, MADV_DONTNEED);
-       else if (f->filetype == FIO_TYPE_FILE)
-               ret = fadvise(f->fd, f->file_offset, f->io_size, POSIX_FADV_DONTNEED);
-       else if (f->filetype == FIO_TYPE_BD) {
+       if (f->mmap_ptr) {
+               ret = madvise(f->mmap_ptr, f->mmap_sz, MADV_DONTNEED);
+#ifdef FIO_MADV_FREE
+               (void) madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
+#endif
+       } else if (f->filetype == FIO_TYPE_FILE) {
+               ret = fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
+       } else if (f->filetype == FIO_TYPE_BD) {
                ret = blockdev_invalidate_cache(f->fd);
                if (ret < 0 && errno == EACCES && geteuid()) {
-                       log_err("fio: only root may flush block devices. Cache flush bypassed!\n");
+                       if (!root_warn) {
+                               log_err("fio: only root may flush block "
+                                       "devices. Cache flush bypassed!\n");
+                               root_warn = 1;
+                       }
                        ret = 0;
                }
        } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
@@ -187,21 +326,70 @@ int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
        if (ret < 0) {
                td_verror(td, errno, "invalidate_cache");
                return 1;
+       } else if (ret > 0) {
+               td_verror(td, ret, "invalidate_cache");
+               return 1;
        }
 
        return ret;
+
 }
 
-void generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
+int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 {
-       close(f->fd);
+       if (!fio_file_open(f))
+               return 0;
+
+       return __file_invalidate_cache(td, f, -1ULL, -1ULL);
+}
+
+int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
+{
+       int ret = 0;
+
+       dprint(FD_FILE, "fd close %s\n", f->file_name);
+
+       remove_file_hash(f);
+
+       if (close(f->fd) < 0)
+               ret = errno;
+
        f->fd = -1;
+       return ret;
+}
+
+static int file_lookup_open(struct fio_file *f, int flags)
+{
+       struct fio_file *__f;
+       int from_hash;
+
+       __f = lookup_file_hash(f->file_name);
+       if (__f) {
+               dprint(FD_FILE, "found file in hash %s\n", f->file_name);
+               /*
+                * 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;
+       }
+
+       f->fd = open(f->file_name, flags, 0600);
+       return from_hash;
 }
 
 int generic_open_file(struct thread_data *td, struct fio_file *f)
 {
        int is_std = 0;
        int flags = 0;
+       int from_hash = 0;
+
+       dprint(FD_FILE, "fd open %s\n", f->file_name);
 
        if (!strcmp(f->file_name, "-")) {
                if (td_rw(td)) {
@@ -221,9 +409,15 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                flags |= OS_O_DIRECT;
        if (td->o.sync_io)
                flags |= O_SYNC;
+       if (f->filetype != FIO_TYPE_FILE)
+               flags |= FIO_O_NOATIME;
+       if (td->o.create_on_open)
+               flags |= O_CREAT;
 
+open_again:
        if (td_write(td)) {
-               flags |= O_RDWR;
+               if (!read_only)
+                       flags |= O_RDWR;
 
                if (f->filetype == FIO_TYPE_FILE)
                        flags |= O_CREAT;
@@ -231,9 +425,9 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                if (is_std)
                        f->fd = dup(STDOUT_FILENO);
                else
-                       f->fd = open(f->file_name, flags, 0600);
+                       from_hash = file_lookup_open(f, flags);
        } else {
-               if (f->filetype == FIO_TYPE_CHAR)
+               if (f->filetype == FIO_TYPE_CHAR && !read_only)
                        flags |= O_RDWR;
                else
                        flags |= O_RDONLY;
@@ -241,56 +435,41 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                if (is_std)
                        f->fd = dup(STDIN_FILENO);
                else
-                       f->fd = open(f->file_name, flags);
+                       from_hash = file_lookup_open(f, flags);
        }
 
        if (f->fd == -1) {
                char buf[FIO_VERROR_SIZE];
                int __e = errno;
 
+               if (__e == EPERM && (flags & FIO_O_NOATIME)) {
+                       flags &= ~FIO_O_NOATIME;
+                       goto open_again;
+               }
+
                snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
 
                td_verror(td, __e, buf);
        }
 
-       if (get_file_size(td, f))
-               goto err;
-
-       return 0;
-err:
-       close(f->fd);
-       return 1;
-}
+       if (!from_hash && f->fd != -1) {
+               if (add_file_hash(f)) {
+                       int ret;
 
-int open_files(struct thread_data *td)
-{
-       struct fio_file *f;
-       unsigned int i;
-       int err = 0;
-
-       for_each_file(td, f, i) {
-               err = td_io_open_file(td, f);
-               if (err) {
-                       if (td->error == EMFILE) {
-                               log_err("fio: limited open files to: %d\n", td->nr_open_files);
-                               td->o.open_files = td->nr_open_files;
-                               err = 0;
-                               clear_error(td);
-                       }
-                       break;
+                       /*
+                        * OK to ignore, we haven't done anything with it
+                        */
+                       ret = generic_close_file(td, f);
+                       goto open_again;
                }
-
-               if (td->o.open_files == td->nr_open_files)
-                       break;
        }
 
-       if (!err)
-               return 0;
-
-       for_each_file(td, f, i)
-               td_io_close_file(td, f);
+       return 0;
+}
 
-       return err;
+int generic_get_file_size(struct thread_data *td, struct fio_file *f)
+{
+       return get_file_size(td, f);
 }
 
 /*
@@ -303,15 +482,15 @@ static int get_file_sizes(struct thread_data *td)
        int err = 0;
 
        for_each_file(td, f, i) {
-               if (td->io_ops->open_file(td, f)) {
+               dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
+                                                               f->file_name);
+
+               if (td_io_get_file_size(td, f)) {
                        if (td->error != ENOENT) {
                                log_err("%s\n", td->verror);
                                err = 1;
                        }
                        clear_error(td);
-               } else {
-                       if (td->io_ops->close_file)
-                               td->io_ops->close_file(td, f);
                }
 
                if (f->real_file_size == -1ULL && td->o.size)
@@ -331,6 +510,11 @@ int setup_files(struct thread_data *td)
        unsigned int i;
        int err = 0, need_extend;
 
+       dprint(FD_FILE, "setup files\n");
+
+       if (td->o.read_iolog_file)
+               return 0;
+
        /*
         * if ioengine defines a setup() method, it's responsible for
         * opening the files and setting f->real_file_size to indicate
@@ -359,7 +543,8 @@ int setup_files(struct thread_data *td)
        /*
         * device/file sizes are zero and no size given, punt
         */
-       if ((!total_size || total_size == -1ULL) && !td->o.size) {
+       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);
                td_verror(td, EINVAL, "total_file_size");
                return 1;
@@ -373,6 +558,8 @@ int setup_files(struct thread_data *td)
        extend_size = total_size = 0;
        need_extend = 0;
        for_each_file(td, f, i) {
+               f->file_offset = td->o.start_offset;
+
                if (!td->o.file_size_low) {
                        /*
                         * no file size range given, file size is equal to
@@ -381,19 +568,24 @@ int setup_files(struct thread_data *td)
                         */
                        f->io_size = td->o.size / td->o.nr_files;
                        if (!f->io_size)
-                               f->io_size = f->real_file_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)
+                               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;
-                       else
-                               f->io_size = get_rand_file_size(td);
+                       if (td->o.file_size_low == td->o.file_size_high) {
+                               f->io_size = td->o.file_size_low
+                                               - f->file_offset;
+                       } else {
+                               f->io_size = get_rand_file_size(td)
+                                               - f->file_offset;
+                       }
                } else
-                       f->io_size = f->real_file_size;
+                       f->io_size = f->real_file_size - f->file_offset;
 
                if (f->io_size == -1ULL)
                        total_size = -1ULL;
@@ -401,15 +593,18 @@ int setup_files(struct thread_data *td)
                        total_size += f->io_size;
 
                if (f->filetype == FIO_TYPE_FILE &&
-                   f->io_size > f->real_file_size &&
+                   (f->io_size + f->file_offset) > f->real_file_size &&
                    !(td->io_ops->flags & FIO_DISKLESSIO)) {
-                       need_extend++;
-                       extend_size += f->io_size;
-                       f->flags |= FIO_FILE_EXTEND;
+                       if (!td->o.create_on_open) {
+                               need_extend++;
+                               extend_size += (f->io_size + f->file_offset);
+                       } else
+                               f->real_file_size = f->io_size + f->file_offset;
+                       fio_file_set_extend(f);
                }
        }
 
-       if (!td->o.size)
+       if (!td->o.size || td->o.size > total_size)
                td->o.size = total_size;
 
        /*
@@ -417,19 +612,34 @@ int setup_files(struct thread_data *td)
         */
        if (need_extend) {
                temp_stall_ts = 1;
-               log_info("%s: Laying out IO file(s) (%u files / %LuMiB)\n",
-                       td->o.name, need_extend, extend_size >> 20);
+               if (!terse_output)
+                       log_info("%s: Laying out IO file(s) (%u file(s) /"
+                                " %LuMB)\n", td->o.name, need_extend,
+                                       extend_size >> 20);
 
                for_each_file(td, f, i) {
-                       if (!(f->flags & FIO_FILE_EXTEND))
+                       unsigned long long old_len = -1ULL, extend_len = -1ULL;
+
+                       if (!fio_file_extend(f))
                                continue;
 
                        assert(f->filetype == FIO_TYPE_FILE);
-                       f->flags &= ~FIO_FILE_EXTEND;
-                       f->real_file_size = f->io_size;
+                       fio_file_clear_extend(f);
+                       if (!td->o.fill_device) {
+                               old_len = f->real_file_size;
+                               extend_len = f->io_size + f->file_offset - old_len;
+                       }
+                       f->real_file_size = (f->io_size + f->file_offset);
                        err = extend_file(td, f);
                        if (err)
                                break;
+
+                       err = __file_invalidate_cache(td, f, old_len,
+                                                               extend_len);
+                       close(f->fd);
+                       f->fd = -1;
+                       if (err)
+                               break;
                }
                temp_stall_ts = 0;
        }
@@ -440,29 +650,62 @@ int setup_files(struct thread_data *td)
        if (!td->o.zone_size)
                td->o.zone_size = td->o.size;
 
-       td->total_io_size = td->o.size * td->o.loops;
+       /*
+        * 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;
        return 0;
+err_offset:
+       log_err("%s: you need to specify valid offset=\n", td->o.name);
+       return 1;
+}
+
+int pre_read_files(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       dprint(FD_FILE, "pre_read files\n");
+
+       for_each_file(td, f, i) {
+               pre_read_file(td, f);
+       }
+
+       return 1;
 }
 
 int init_random_map(struct thread_data *td)
 {
-       int num_maps, blocks;
+       unsigned long long blocks, num_maps;
        struct fio_file *f;
        unsigned int i;
 
-       if (td->o.norandommap)
+       if (td->o.norandommap || !td_random(td))
                return 0;
 
        for_each_file(td, f, i) {
-               blocks = (f->real_file_size + td->o.rw_min_bs - 1) / td->o.rw_min_bs;
-               num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
-               f->file_map = malloc(num_maps * sizeof(long));
-               if (!f->file_map) {
-                       log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
+               blocks = (f->real_file_size + td->o.rw_min_bs - 1) /
+                               (unsigned long long) td->o.rw_min_bs;
+               num_maps = (blocks + BLOCKS_PER_MAP - 1) /
+                               (unsigned long long) BLOCKS_PER_MAP;
+               f->file_map = smalloc(num_maps * sizeof(int));
+               if (f->file_map) {
+                       f->num_maps = num_maps;
+                       continue;
+               }
+               if (!td->o.softrandommap) {
+                       log_err("fio: failed allocating random map. If running"
+                               " a large number of jobs, try the 'norandommap'"
+                               " option or set 'softrandommap'. Or give"
+                               " a larger --alloc-size to fio.\n");
                        return 1;
                }
-               f->num_maps = num_maps;
-               memset(f->file_map, 0, num_maps * sizeof(long));
+
+               log_info("fio: file %s failed allocating random map. Running "
+                        "job without.\n", f->file_name);
+               f->num_maps = 0;
        }
 
        return 0;
@@ -474,22 +717,42 @@ void close_files(struct thread_data *td)
        unsigned int i;
 
        for_each_file(td, f, i) {
-               if (td->o.unlink && f->filetype == FIO_TYPE_FILE)
+               if (fio_file_open(f))
+                       td_io_close_file(td, f);
+       }
+}
+
+void close_and_free_files(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       dprint(FD_FILE, "close files\n");
+
+       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_close_file(td, f);
+               if (fio_file_open(f))
+                       td_io_close_file(td, f);
 
-               free(f->file_name);
+               remove_file_hash(f);
+
+               sfree(f->file_name);
                f->file_name = NULL;
 
                if (f->file_map) {
-                       free(f->file_map);
+                       sfree(f->file_map);
                        f->file_map = NULL;
                }
+               sfree(f);
        }
 
        td->o.filename = NULL;
        free(td->files);
+       td->files_index = 0;
        td->files = NULL;
        td->o.nr_files = 0;
 }
@@ -520,12 +783,26 @@ int add_file(struct thread_data *td, const char *fname)
        struct fio_file *f;
        int len = 0;
 
-       td->files = realloc(td->files, (cur_files + 1) * sizeof(*f));
+       dprint(FD_FILE, "add file %s\n", fname);
 
-       f = &td->files[cur_files];
-       memset(f, 0, sizeof(*f));
+       f = smalloc(sizeof(*f));
+       if (!f) {
+               log_err("fio: smalloc OOM\n");
+               assert(0);
+       }
+               
        f->fd = -1;
 
+       if (td->files_size <= td->files_index) {
+               int new_size = td->o.nr_files + 1;
+
+               dprint(FD_FILE, "resize file array to %d files\n", new_size);
+
+               td->files = realloc(td->files, new_size * sizeof(f));
+               td->files_size = new_size;
+       }
+       td->files[cur_files] = f;
+
        /*
         * init function, io engine may not be loaded yet
         */
@@ -536,40 +813,131 @@ int add_file(struct thread_data *td, const char *fname)
                len = sprintf(file_name, "%s/", td->o.directory);
 
        sprintf(file_name + len, "%s", fname);
-       f->file_name = strdup(file_name);
-
+       f->file_name = smalloc_strdup(file_name);
+       if (!f->file_name) {
+               log_err("fio: smalloc OOM\n");
+               assert(0);
+       }
+       
        get_file_type(f);
 
+       switch (td->o.file_lock_mode) {
+       case FILE_LOCK_NONE:
+               break;
+       case FILE_LOCK_READWRITE:
+               f->lock = fio_mutex_rw_init();
+               break;
+       case FILE_LOCK_EXCLUSIVE:
+               f->lock = fio_mutex_init(1);
+               break;
+       default:
+               log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
+               assert(0);
+       }
+
        td->files_index++;
        if (f->filetype == FIO_TYPE_FILE)
                td->nr_normal_files++;
 
+       dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
+                                                       cur_files);
+
        return cur_files;
 }
 
 void get_file(struct fio_file *f)
 {
-       assert(f->flags & FIO_FILE_OPEN);
+       dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
+       assert(fio_file_open(f));
        f->references++;
 }
 
-void put_file(struct thread_data *td, struct fio_file *f)
+int put_file(struct thread_data *td, struct fio_file *f)
 {
-       if (!(f->flags & FIO_FILE_OPEN))
-               return;
+       int f_ret = 0, ret = 0;
+
+       dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
+
+       if (!fio_file_open(f)) {
+               assert(f->fd == -1);
+               return 0;
+       }
 
        assert(f->references);
        if (--f->references)
-               return;
+               return 0;
 
        if (should_fsync(td) && td->o.fsync_on_close)
-               fsync(f->fd);
+               f_ret = fsync(f->fd);
 
        if (td->io_ops->close_file)
-               td->io_ops->close_file(td, f);
+               ret = td->io_ops->close_file(td, f);
+
+       if (!ret)
+               ret = f_ret;
 
        td->nr_open_files--;
-       f->flags &= ~FIO_FILE_OPEN;
+       fio_file_clear_open(f);
+       assert(f->fd == -1);
+       return ret;
+}
+
+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);
+               else
+                       fio_mutex_down_write(f->lock);
+       } 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;
+}
+
+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;
+
+               fio_mutex_up(f->lock);
+       }
+}
+
+void unlock_file_all(struct thread_data *td, struct fio_file *f)
+{
+       if (f->lock_owner != td)
+               return;
+
+       f->lock_batch = 0;
+       unlock_file(td, f);
 }
 
 static int recurse_dir(struct thread_data *td, const char *dirname)
@@ -611,7 +979,8 @@ static int recurse_dir(struct thread_data *td, const char *dirname)
                if (!S_ISDIR(sb.st_mode))
                        continue;
 
-               if ((ret = recurse_dir(td, full_path)) != 0)
+               ret = recurse_dir(td, full_path);
+               if (ret)
                        break;
        }
 
@@ -633,18 +1002,35 @@ void dup_files(struct thread_data *td, struct thread_data *org)
 {
        struct fio_file *f;
        unsigned int i;
-       size_t bytes;
+
+       dprint(FD_FILE, "dup files: %d\n", org->files_index);
 
        if (!org->files)
                return;
 
-       bytes = org->files_index * sizeof(*f);
-       td->files = malloc(bytes);
-       memcpy(td->files, org->files, bytes);
+       td->files = malloc(org->files_index * sizeof(f));
 
-       for_each_file(td, f, i) {
-               if (f->file_name)
-                       f->file_name = strdup(f->file_name);
+       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;
+       
+               if (f->file_name) {
+                       __f->file_name = smalloc_strdup(f->file_name);
+                       if (!__f->file_name) {
+                               log_err("fio: smalloc OOM\n");
+                               assert(0);
+                       }
+       
+                       __f->filetype = f->filetype;
+               }
+
+               td->files[i] = __f;
        }
 }