Only set O_RDWR for open() when readonly isn't set
[fio.git] / filesetup.c
index 4e2a36cd9debb6968302cd54bd117b967a6b78da..5c02f1c46408fb0e2f781f1ba15118e710935a4e 100644 (file)
@@ -96,8 +96,12 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
 
        if (td->terminate)
                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;
+               }
+       }
 
        free(b);
 done:
@@ -112,11 +116,13 @@ err:
 
 static unsigned long long get_rand_file_size(struct thread_data *td)
 {
-       unsigned long long ret;
+       unsigned long long ret, size_d;
        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 - td->o.file_size_low) * (r / (RAND_MAX + 1.0)));
+       size_d = td->o.file_size_high - td->o.file_size_low;
+       ret = (unsigned long long) ((double) size_d * (r / (RAND_MAX + 1.0)));
+       ret += td->o.file_size_low;
        ret -= (ret % td->o.rw_min_bs);
        return ret;
 }
@@ -167,7 +173,8 @@ 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;
        }
 
@@ -189,13 +196,15 @@ int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
         */
        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) {
+       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) {
                ret = blockdev_invalidate_cache(f->fd);
                if (ret < 0 && errno == EACCES && geteuid()) {
                        if (!root_warn) {
-                               log_err("fio: only root may flush block devices. Cache flush bypassed!\n");
+                               log_err("fio: only root may flush block "
+                                       "devices. Cache flush bypassed!\n");
                                root_warn = 1;
                        }
                        ret = 0;
@@ -226,9 +235,33 @@ int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
        return ret;
 }
 
-int generic_open_file(struct thread_data *td, struct fio_file *f)
+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;
@@ -258,25 +291,16 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
 
 open_again:
        if (td_write(td)) {
-               assert(!read_only);
-
-               flags |= O_RDWR;
+               if (!read_only)
+                       flags |= O_RDWR;
 
                if (f->filetype == FIO_TYPE_FILE)
                        flags |= O_CREAT;
 
                if (is_std)
                        f->fd = dup(STDOUT_FILENO);
-               else {
-                       __f = lookup_file_hash(f->file_name);
-                       if (__f) {
-                               f->sem = __f->sem;
-                               f->fd = dup(__f->fd);
-                               f->references++;
-                               from_hash = 1;
-                       } else
-                               f->fd = open(f->file_name, flags, 0600);
-               }
+               else
+                       from_hash = file_lookup_open(f, flags);
        } else {
                if (f->filetype == FIO_TYPE_CHAR && !read_only)
                        flags |= O_RDWR;
@@ -285,16 +309,8 @@ open_again:
 
                if (is_std)
                        f->fd = dup(STDIN_FILENO);
-               else {
-                       __f = lookup_file_hash(f->file_name);
-                       if (__f) {
-                               f->sem = __f->sem;
-                               f->fd = dup(__f->fd);
-                               f->references++;
-                               from_hash = 1;
-                       } else
-                               f->fd = open(f->file_name, flags);
-               }
+               else
+                       from_hash = file_lookup_open(f, flags);
        }
 
        if (f->fd == -1) {
@@ -344,7 +360,8 @@ int open_files(struct thread_data *td)
                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);
+                               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);
@@ -375,6 +392,9 @@ static int get_file_sizes(struct thread_data *td)
        int err = 0;
 
        for_each_file(td, f, i) {
+               dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
+                                                               f->file_name);
+
                if (td->io_ops->open_file(td, f)) {
                        if (td->error != ENOENT) {
                                log_err("%s\n", td->verror);
@@ -461,16 +481,19 @@ int setup_files(struct thread_data *td)
                                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) 
+                       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 - f->file_offset;
-                       else
-                               f->io_size = get_rand_file_size(td) - f->file_offset;
+                       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->file_offset;
 
@@ -485,7 +508,7 @@ int setup_files(struct thread_data *td)
                        need_extend++;
                        extend_size += (f->io_size + f->file_offset);
                        f->flags |= FIO_FILE_EXTEND;
-               }       
+               }
        }
 
        if (!td->o.size || td->o.size > total_size)
@@ -541,11 +564,15 @@ int init_random_map(struct thread_data *td)
                return 0;
 
        for_each_file(td, f, i) {
-               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;
+               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(long));
                if (!f->file_map) {
-                       log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
+                       log_err("fio: failed allocating random map. If running"
+                               " a large number of jobs, try the 'norandommap'"
+                               " option\n");
                        return 1;
                }
                f->num_maps = num_maps;
@@ -588,6 +615,7 @@ void close_and_free_files(struct thread_data *td)
 
        td->o.filename = NULL;
        free(td->files);
+       td->files_index = 0;
        td->files = NULL;
        td->o.nr_files = 0;
 }
@@ -623,8 +651,9 @@ int add_file(struct thread_data *td, const char *fname)
        f = smalloc(sizeof(*f));
        f->fd = -1;
 
-       td->files = realloc(td->files, (cur_files + 1) * sizeof(f));
+       dprint(FD_FILE, "resize file array to %d files\n", cur_files + 1);
 
+       td->files = realloc(td->files, (cur_files + 1) * sizeof(f));
        td->files[cur_files] = f;
 
        /*
@@ -641,13 +670,27 @@ int add_file(struct thread_data *td, const char *fname)
 
        get_file_type(f);
 
-       if (td->o.lockfile)
-               f->sem = fio_sem_init(1);
+       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;
 }
 
@@ -660,7 +703,7 @@ void get_file(struct fio_file *f)
 
 int put_file(struct thread_data *td, struct fio_file *f)
 {
-       int ret = 0;
+       int f_ret = 0, ret = 0;
 
        dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
 
@@ -672,43 +715,77 @@ int put_file(struct thread_data *td, struct fio_file *f)
                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)
                ret = td->io_ops->close_file(td, f);
 
+       if (!ret)
+               ret = !f_ret;
+
        td->nr_open_files--;
        f->flags &= ~FIO_FILE_OPEN;
        return ret;
 }
 
-void lock_file(struct thread_data *td, struct fio_file *f)
+void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
 {
-       if (f && f->sem) {
-               if (f->sem_owner == td && f->sem_batch--)
-                       return;
+       if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
+               return;
 
-               fio_sem_down(f->sem);
-               f->sem_owner = td;
-               f->sem_batch = td->o.lockfile_batch;
-       }
+       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 fio_file *f)
+void unlock_file(struct thread_data *td, struct fio_file *f)
 {
-       if (f && f->sem) {
-               int sem_val;
+       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 (f->sem_batch)
-                       return;
+               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;
 
-               sem_getvalue(&f->sem->sem, &sem_val);
-               if (!sem_val)
-                       f->sem_owner = NULL;
-               fio_sem_up(f->sem);
+               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)
 {
        struct dirent *dir;
@@ -748,7 +825,8 @@ 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;
        }
 
@@ -770,16 +848,15 @@ 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) {
+       for_each_file(org, f, i) {
                struct fio_file *__f;
 
                __f = smalloc(sizeof(*__f));