Fix issue with openfiles= and file sizing
[fio.git] / filesetup.c
index 3054d9dab4803271de2bbee3369ac9ef66d1133d..bae60929df32f7f80f1dfa4ba00ab3ffbb9514b0 100644 (file)
@@ -79,7 +79,8 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                        break;
                case FIO_FALLOCATE_POSIX:
                        dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
-                                f->file_name, f->real_file_size);
+                                f->file_name,
+                                (unsigned long long) f->real_file_size);
 
                        r = posix_fallocate(f->fd, 0, f->real_file_size);
                        if (r > 0) {
@@ -91,8 +92,8 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                case FIO_FALLOCATE_KEEP_SIZE:
                        dprint(FD_FILE,
                                "fallocate(FALLOC_FL_KEEP_SIZE) "
-                               "file %s size %llu\n",
-                               f->file_name, f->real_file_size);
+                               "file %s size %llu\n", f->file_name,
+                               (unsigned long long) f->real_file_size);
 
                        r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
                                        f->real_file_size);
@@ -118,15 +119,16 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
         */
        if (!td->o.fill_device) {
                dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
-                                                       f->real_file_size);
+                                       (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) {
@@ -134,6 +136,8 @@ 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) {
@@ -354,7 +358,8 @@ 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;
        }
 
@@ -385,7 +390,8 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        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);
+               if (f->filetype == FIO_TYPE_BD)
+                       (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 #endif
        } else if (f->filetype == FIO_TYPE_FILE) {
                ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
@@ -440,6 +446,7 @@ int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
                f->shadow_fd = -1;
        }
 
+       f->engine_data = 0;
        return ret;
 }
 
@@ -455,9 +462,6 @@ 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);
@@ -517,6 +521,13 @@ 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)
@@ -582,7 +593,7 @@ 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)
@@ -721,13 +732,20 @@ 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->runstate;
+       td_set_runstate(td, TD_SETTING_UP);
+
+       if (o->read_iolog_file)
                goto done;
 
        /*
@@ -741,7 +759,7 @@ 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
@@ -755,17 +773,32 @@ 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) {
+               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;
        }
 
        /*
@@ -778,27 +811,33 @@ int setup_files(struct thread_data *td)
        for_each_file(td, f, i) {
                f->file_offset = get_start_offset(td);
 
-               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;
                        }
@@ -808,15 +847,15 @@ 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
@@ -825,8 +864,8 @@ int setup_files(struct thread_data *td)
                }
        }
 
-       if (!td->o.size || td->o.size > total_size)
-               td->o.size = total_size;
+       if (!o->size || o->size > total_size)
+               o->size = total_size;
 
        /*
         * See if we need to extend some files
@@ -835,7 +874,7 @@ 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) {
@@ -846,7 +885,7 @@ 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;
@@ -867,25 +906,28 @@ 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)
+               td->total_io_size = o->size * o->loops;
 
 done:
-       if (td->o.create_only)
+       if (o->create_only)
                td->done = 1;
 
+       td_set_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_set_runstate(td, old_state);
        return 1;
 }
 
@@ -958,14 +1000,14 @@ int init_random_map(struct thread_data *td)
        for_each_file(td, f, i) {
                uint64_t file_size = min(f->real_file_size, f->io_size);
 
-               blocks = (file_size + td->o.rw_min_bs - 1) /
-                               (unsigned long long) td->o.rw_min_bs;
+               blocks = file_size / (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, seed & 0xF))
                                continue;
                } else if (!td->o.norandommap) {
                        f->io_axmap = axmap_new(blocks);
@@ -1027,8 +1069,11 @@ void close_and_free_files(struct thread_data *td)
 
        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;
 }
 
@@ -1085,6 +1130,14 @@ 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;
@@ -1112,7 +1165,7 @@ 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);
@@ -1187,57 +1240,36 @@ 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)
                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)
@@ -1310,6 +1342,9 @@ 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;
 
@@ -1331,6 +1366,11 @@ 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;
        }
 }
@@ -1369,3 +1409,15 @@ void fio_file_reset(struct thread_data *td, struct fio_file *f)
        if (td->o.random_generator == FIO_RAND_GEN_LFSR)
                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;
+}