X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=ioengines.c;h=9638d8043753627f60ffc223b8308ca6f61d57af;hp=f4eae04db6df5d072103191f05b187fcb6f19e80;hb=78d4a262a0758d47631bfd44672bbd5f9eff9b42;hpb=9db01ef976006c002b05fa4e4ec589eb029aac5b

diff --git a/ioengines.c b/ioengines.c
index f4eae04d..9638d804 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -22,39 +22,31 @@
 
 static FLIST_HEAD(engine_list);
 
-static int check_engine_ops(struct ioengine_ops *ops)
+static bool check_engine_ops(struct ioengine_ops *ops)
 {
 	if (ops->version != FIO_IOOPS_VERSION) {
 		log_err("bad ioops version %d (want %d)\n", ops->version,
 							FIO_IOOPS_VERSION);
-		return 1;
+		return true;
 	}
 
 	if (!ops->queue) {
 		log_err("%s: no queue handler\n", ops->name);
-		return 1;
+		return true;
 	}
 
 	/*
 	 * sync engines only need a ->queue()
 	 */
 	if (ops->flags & FIO_SYNCIO)
-		return 0;
+		return false;
 
-	if (!ops->event) {
-		log_err("%s: no event handler\n", ops->name);
-		return 1;
-	}
-	if (!ops->getevents) {
-		log_err("%s: no getevents handler\n", ops->name);
-		return 1;
-	}
-	if (!ops->queue) {
-		log_err("%s: no queue handler\n", ops->name);
-		return 1;
+	if (!ops->event || !ops->getevents) {
+		log_err("%s: no event/getevents handler\n", ops->name);
+		return true;
 	}
 
-	return 0;
+	return false;
 }
 
 void unregister_ioengine(struct ioengine_ops *ops)
@@ -107,23 +99,35 @@ static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
 	ops = dlsym(dlhandle, engine_lib);
 	if (!ops)
 		ops = dlsym(dlhandle, "ioengine");
+
+	/*
+	 * For some external engines (like C++ ones) it is not that trivial
+	 * to provide a non-static ionengine structure that we can reference.
+	 * Instead we call a method which allocates the required ioengine
+	 * structure.
+	 */
+	if (!ops) {
+		get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
+
+		if (get_ioengine)
+			get_ioengine(&ops);
+	}
+
 	if (!ops) {
 		td_vmsg(td, -1, dlerror(), "dlsym");
 		dlclose(dlhandle);
 		return NULL;
 	}
 
-	ops->dlhandle = dlhandle;
+	td->io_ops_dlhandle = dlhandle;
 	return ops;
 }
 
-struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
+static struct ioengine_ops *__load_ioengine(const char *name)
 {
-	struct ioengine_ops *ops, *ret;
-	char engine[16];
-
-	dprint(FD_IO, "load ioengine %s\n", name);
+	char engine[64];
 
+	engine[sizeof(engine) - 1] = '\0';
 	strncpy(engine, name, sizeof(engine) - 1);
 
 	/*
@@ -132,10 +136,37 @@ struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
 	if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
 		strcpy(engine, "libaio");
 
-	ops = find_ioengine(engine);
+	dprint(FD_IO, "load ioengine %s\n", engine);
+	return find_ioengine(engine);
+}
+
+struct ioengine_ops *load_ioengine(struct thread_data *td)
+{
+	struct ioengine_ops *ops = NULL;
+	const char *name;
+
+	/*
+	 * Use ->ioengine_so_path if an external ioengine path is specified.
+	 * In this case, ->ioengine is "external" which also means the prefix
+	 * for external ioengines "external:" is properly used.
+	 */
+	name = td->o.ioengine_so_path ?: td->o.ioengine;
+
+	/*
+	 * Try to load ->ioengine first, and if failed try to dlopen(3) either
+	 * ->ioengine or ->ioengine_so_path.  This is redundant for an external
+	 * ioengine with prefix, and also leaves the possibility of unexpected
+	 * behavior (e.g. if the "external" ioengine exists), but we do this
+	 * so as not to break job files not using the prefix.
+	 */
+	ops = __load_ioengine(td->o.ioengine);
 	if (!ops)
 		ops = dlopen_ioengine(td, name);
 
+	/*
+	 * If ops is NULL, we failed to load ->ioengine, and also failed to
+	 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
+	 */
 	if (!ops) {
 		log_err("fio: engine %s not loadable\n", name);
 		return NULL;
@@ -147,11 +178,7 @@ struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
 	if (check_engine_ops(ops))
 		return NULL;
 
-	ret = malloc(sizeof(*ret));
-	memcpy(ret, ops, sizeof(*ret));
-	ret->data = NULL;
-
-	return ret;
+	return ops;
 }
 
 /*
@@ -167,10 +194,9 @@ void free_ioengine(struct thread_data *td)
 		td->eo = NULL;
 	}
 
-	if (td->io_ops->dlhandle)
-		dlclose(td->io_ops->dlhandle);
+	if (td->io_ops_dlhandle)
+		dlclose(td->io_ops_dlhandle);
 
-	free(td->io_ops);
 	td->io_ops = NULL;
 }
 
@@ -180,7 +206,7 @@ void close_ioengine(struct thread_data *td)
 
 	if (td->io_ops->cleanup) {
 		td->io_ops->cleanup(td);
-		td->io_ops->data = NULL;
+		td->io_ops_data = NULL;
 	}
 
 	free_ioengine(td);
@@ -206,7 +232,7 @@ int td_io_prep(struct thread_data *td, struct io_u *io_u)
 }
 
 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
-		    struct timespec *t)
+		    const struct timespec *t)
 {
 	int r = 0;
 
@@ -250,13 +276,15 @@ out:
 
 int td_io_queue(struct thread_data *td, struct io_u *io_u)
 {
+	const enum fio_ddir ddir = acct_ddir(io_u);
+	unsigned long buflen = io_u->xfer_buflen;
 	int ret;
 
 	dprint_io_u(io_u, "queue");
 	fio_ro_check(td, io_u);
 
 	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
-	io_u->flags |= IO_U_F_FLIGHT;
+	io_u_set(td, io_u, IO_U_F_FLIGHT);
 
 	assert(fio_file_open(io_u->file));
 
@@ -268,7 +296,7 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 	io_u->error = 0;
 	io_u->resid = 0;
 
-	if (td->io_ops->flags & FIO_SYNCIO) {
+	if (td_ioengine_flagged(td, FIO_SYNCIO)) {
 		if (fio_fill_issue_time(td))
 			fio_gettime(&io_u->issue_time, NULL);
 
@@ -277,16 +305,33 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 		 */
 		if (td->o.read_iolog_file)
 			memcpy(&td->last_issue, &io_u->issue_time,
-					sizeof(struct timeval));
+					sizeof(io_u->issue_time));
 	}
 
-	if (ddir_rw(acct_ddir(io_u)))
-		td->io_issues[acct_ddir(io_u)]++;
+	if (ddir_rw(ddir)) {
+		td->io_issues[ddir]++;
+		td->io_issue_bytes[ddir] += buflen;
+		td->rate_io_issue_bytes[ddir] += buflen;
+	}
 
 	ret = td->io_ops->queue(td, io_u);
 
 	unlock_file(td, io_u->file);
 
+	if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
+		td->io_issues[ddir]--;
+		td->io_issue_bytes[ddir] -= buflen;
+		td->rate_io_issue_bytes[ddir] -= buflen;
+		io_u_clear(td, io_u, IO_U_F_FLIGHT);
+	}
+
+	/*
+	 * If an error was seen and the io engine didn't propagate it
+	 * back to 'td', do so.
+	 */
+	if (io_u->error && !td->error)
+		td_verror(td, io_u->error, "td_io_queue");
+
 	/*
 	 * Add warning for O_DIRECT so that users have an easier time
 	 * spotting potentially bad alignment. If this triggers for the first
@@ -297,10 +342,11 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 	    td->o.odirect) {
 
 		log_info("fio: first direct IO errored. File system may not "
-			 "support direct IO, or iomem_align= is bad.\n");
+			 "support direct IO, or iomem_align= is bad, or "
+			 "invalid block size. Try setting direct=0.\n");
 	}
 
-	if (!td->io_ops->commit || ddir_trim(io_u->ddir)) {
+	if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
 		io_u_mark_submit(td, 1);
 		io_u_mark_complete(td, 1);
 	}
@@ -313,10 +359,10 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 	} else if (ret == FIO_Q_QUEUED) {
 		int r;
 
-		if (ddir_rw(io_u->ddir)) {
-			td->io_u_queued++;
+		td->io_u_queued++;
+
+		if (ddir_rw(io_u->ddir))
 			td->ts.total_io_u[io_u->ddir]++;
-		}
 
 		if (td->io_u_queued >= td->o.iodepth_batch) {
 			r = td_io_commit(td);
@@ -325,7 +371,7 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 		}
 	}
 
-	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
+	if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
 		if (fio_fill_issue_time(td))
 			fio_gettime(&io_u->issue_time, NULL);
 
@@ -334,7 +380,7 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 		 */
 		if (td->o.read_iolog_file)
 			memcpy(&td->last_issue, &io_u->issue_time,
-					sizeof(struct timeval));
+					sizeof(io_u->issue_time));
 	}
 
 	return ret;
@@ -346,10 +392,13 @@ int td_io_init(struct thread_data *td)
 
 	if (td->io_ops->init) {
 		ret = td->io_ops->init(td);
-		if (ret && td->o.iodepth > 1) {
-			log_err("fio: io engine init failed. Perhaps try"
-				" reducing io depth?\n");
-		}
+		if (ret)
+			log_err("fio: io engine %s init failed.%s\n",
+				td->io_ops->name,
+				td->o.iodepth > 1 ?
+				" Perhaps try reducing io depth?" : "");
+		else
+			td->io_ops_init = 1;
 		if (!td->error)
 			td->error = ret;
 	}
@@ -417,43 +466,62 @@ int td_io_open_file(struct thread_data *td, struct fio_file *f)
 		}
 	}
 
-	if (td->io_ops->flags & FIO_DISKLESSIO)
+	if (td_ioengine_flagged(td, FIO_DISKLESSIO))
 		goto done;
 
 	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
 		goto err;
 
-	if (td->o.fadvise_hint &&
-	    (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
+	if (td->o.fadvise_hint != F_ADV_NONE &&
+	    (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
 		int flags;
 
-		if (td_random(td))
+		if (td->o.fadvise_hint == F_ADV_TYPE) {
+			if (td_random(td))
+				flags = POSIX_FADV_RANDOM;
+			else
+				flags = POSIX_FADV_SEQUENTIAL;
+		} else if (td->o.fadvise_hint == F_ADV_RANDOM)
 			flags = POSIX_FADV_RANDOM;
-		else
+		else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
 			flags = POSIX_FADV_SEQUENTIAL;
+		else {
+			log_err("fio: unknown fadvise type %d\n",
+							td->o.fadvise_hint);
+			flags = POSIX_FADV_NORMAL;
+		}
 
 		if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
 			td_verror(td, errno, "fadvise");
 			goto err;
 		}
 	}
+#ifdef FIO_HAVE_WRITE_HINT
+	if (fio_option_is_set(&td->o, write_hint) &&
+	    (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
+		uint64_t hint = td->o.write_hint;
+		int cmd;
 
-#ifdef FIO_OS_DIRECTIO
-	/*
-	 * Some OS's have a distinct call to mark the file non-buffered,
-	 * instead of using O_DIRECT (Solaris)
-	 */
-	if (td->o.odirect) {
-		int ret = fio_set_odirect(f->fd);
+		/*
+		 * For direct IO, we just need/want to set the hint on
+		 * the file descriptor. For buffered IO, we need to set
+		 * it on the inode.
+		 */
+		if (td->o.odirect)
+			cmd = F_SET_FILE_RW_HINT;
+		else
+			cmd = F_SET_RW_HINT;
 
-		if (ret) {
-			td_verror(td, ret, "fio_set_odirect");
-			log_err("fio: the file system does not seem to support direct IO\n");
+		if (fcntl(f->fd, cmd, &hint) < 0) {
+			td_verror(td, errno, "fcntl write hint");
 			goto err;
 		}
 	}
 #endif
 
+	if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
+		goto err;
+
 done:
 	log_file(td, f, FIO_LOG_OPEN_FILE);
 	return 0;
@@ -475,89 +543,48 @@ int td_io_close_file(struct thread_data *td, struct fio_file *f)
 	fio_file_set_closing(f);
 
 	disk_util_dec(f->du);
-	unlock_file_all(td, f);
 
-	return put_file(td, f);
-}
+	if (td->o.file_lock_mode != FILE_LOCK_NONE)
+		unlock_file_all(td, f);
 
-int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
-{
-	if (!td->io_ops->get_file_size)
-		return 0;
-
-	return td->io_ops->get_file_size(td, f);
+	return put_file(td, f);
 }
 
-static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
+int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
 {
-	off64_t offset, nbytes;
+	if (td->io_ops->unlink_file)
+		return td->io_ops->unlink_file(td, f);
+	else {
+		int ret;
 
-	offset = f->first_write;
-	nbytes = f->last_write - f->first_write;
+		ret = unlink(f->file_name);
+		if (ret < 0)
+			return errno;
 
-	if (!nbytes)
 		return 0;
-
-	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
-}
-
-int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
-{
-	int ret;
-
-	if (io_u->ddir == DDIR_SYNC) {
-		ret = fsync(io_u->file->fd);
-	} else if (io_u->ddir == DDIR_DATASYNC) {
-#ifdef CONFIG_FDATASYNC
-		ret = fdatasync(io_u->file->fd);
-#else
-		ret = io_u->xfer_buflen;
-		io_u->error = EINVAL;
-#endif
-	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
-		ret = do_sync_file_range(td, io_u->file);
-	else {
-		ret = io_u->xfer_buflen;
-		io_u->error = EINVAL;
 	}
-
-	if (ret < 0)
-		io_u->error = errno;
-
-	return ret;
 }
 
-int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
+int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
 {
-#ifndef FIO_HAVE_TRIM
-	io_u->error = EINVAL;
-	return 0;
-#else
-	struct fio_file *f = io_u->file;
-	int ret;
-
-	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
-	if (!ret)
-		return io_u->xfer_buflen;
+	if (!td->io_ops->get_file_size)
+		return 0;
 
-	io_u->error = ret;
-	return 0;
-#endif
+	return td->io_ops->get_file_size(td, f);
 }
 
 int fio_show_ioengine_help(const char *engine)
 {
 	struct flist_head *entry;
-	struct thread_data td;
+	struct ioengine_ops *io_ops;
 	char *sep;
 	int ret = 1;
 
 	if (!engine || !*engine) {
 		log_info("Available IO engines:\n");
 		flist_for_each(entry, &engine_list) {
-			td.io_ops = flist_entry(entry, struct ioengine_ops,
-						list);
-			log_info("\t%s\n", td.io_ops->name);
+			io_ops = flist_entry(entry, struct ioengine_ops, list);
+			log_info("\t%s\n", io_ops->name);
 		}
 		return 0;
 	}
@@ -567,20 +594,16 @@ int fio_show_ioengine_help(const char *engine)
 		sep++;
 	}
 
-	memset(&td, 0, sizeof(td));
-
-	td.io_ops = load_ioengine(&td, engine);
-	if (!td.io_ops) {
+	io_ops = __load_ioengine(engine);
+	if (!io_ops) {
 		log_info("IO engine %s not found\n", engine);
 		return 1;
 	}
 
-	if (td.io_ops->options)
-		ret = show_cmd_help(td.io_ops->options, sep);
+	if (io_ops->options)
+		ret = show_cmd_help(io_ops->options, sep);
 	else
-		log_info("IO engine %s has no options\n", td.io_ops->name);
-
-	free_ioengine(&td);
+		log_info("IO engine %s has no options\n", io_ops->name);
 
 	return ret;
 }