X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=ioengines.c;h=6e6e3dea019e2b54a127a1dfea9e6e2b90dde48a;hp=e0b0f22e96d994b74f00153dd64509dd7b2a930d;hb=f25f4ef6d123173d7669553ec712419f95c1a3ea;hpb=c6404a44d139fb494202f9d968f7affe8c3d986f

diff --git a/ioengines.c b/ioengines.c
index e0b0f22e..6e6e3dea 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -14,6 +14,7 @@
 #include <unistd.h>
 #include <string.h>
 #include <dlfcn.h>
+#include <fcntl.h>
 #include <assert.h>
 
 #include "fio.h"
@@ -21,39 +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)
@@ -103,24 +96,41 @@ static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
 	 * Unlike the included modules, external engines should have a
 	 * non-static ioengine structure that we can reference.
 	 */
-	ops = dlsym(dlhandle, "ioengine");
+	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)
 {
-	struct ioengine_ops *ops, *ret;
-	char engine[16];
+	struct ioengine_ops *ops;
+	char engine[64];
 
 	dprint(FD_IO, "load ioengine %s\n", name);
 
+	engine[sizeof(engine) - 1] = '\0';
 	strncpy(engine, name, sizeof(engine) - 1);
 
 	/*
@@ -144,11 +154,26 @@ 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 ops;
+}
 
-	return ret;
+/*
+ * For cleaning up an ioengine which never made it to init().
+ */
+void free_ioengine(struct thread_data *td)
+{
+	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
+
+	if (td->eo && td->io_ops->options) {
+		options_free(td->io_ops->options, td->eo);
+		free(td->eo);
+		td->eo = NULL;
+	}
+
+	if (td->io_ops_dlhandle)
+		dlclose(td->io_ops_dlhandle);
+
+	td->io_ops = NULL;
 }
 
 void close_ioengine(struct thread_data *td)
@@ -157,14 +182,10 @@ 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;
 	}
 
-	if (td->io_ops->dlhandle)
-		dlclose(td->io_ops->dlhandle);
-
-	free(td->io_ops);
-	td->io_ops = NULL;
+	free_ioengine(td);
 }
 
 int td_io_prep(struct thread_data *td, struct io_u *io_u)
@@ -187,10 +208,20 @@ 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;
 
+	/*
+	 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
+	 * server side gets a message from the client
+	 * side that the task is finished, and
+	 * td->done is set to 1 after td_io_commit(). In this case,
+	 * there is no need to reap complete event in server side.
+	 */
+	if (td->done)
+		return 0;
+
 	if (min > 0 && td->io_ops->commit) {
 		r = td->io_ops->commit(td);
 		if (r < 0)
@@ -205,9 +236,14 @@ int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
 	if (max && td->io_ops->getevents)
 		r = td->io_ops->getevents(td, min, max, t);
 out:
-	if (r >= 0)
+	if (r >= 0) {
+		/*
+		 * Reflect that our submitted requests were retrieved with
+		 * whatever OS async calls are in the underlying engine.
+		 */
+		td->io_u_in_flight -= r;
 		io_u_mark_complete(td, r);
-	else
+	} else
 		td_verror(td, r, "get_events");
 
 	dprint(FD_IO, "getevents: %d\n", r);
@@ -216,20 +252,27 @@ 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));
 
+	/*
+	 * If using a write iolog, store this entry.
+	 */
+	log_io_u(td, 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);
 
@@ -238,16 +281,33 @@ 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(io_u->ddir))
-		td->io_issues[io_u->ddir]++;
+	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
@@ -256,11 +316,13 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 	 */
 	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
 	    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. Try "
+			 "setting direct=0.\n");
 	}
 
-	if (!td->io_ops->commit) {
+	if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
 		io_u_mark_submit(td, 1);
 		io_u_mark_complete(td, 1);
 	}
@@ -269,15 +331,14 @@ int td_io_queue(struct thread_data *td, struct io_u *io_u)
 		if (ddir_rw(io_u->ddir)) {
 			io_u_mark_depth(td, 1);
 			td->ts.total_io_u[io_u->ddir]++;
-		} else if (io_u->ddir == DDIR_TRIM)
-			td->ts.total_io_u[2]++;
+		}
 	} 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);
@@ -286,7 +347,7 @@ 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);
 
@@ -295,7 +356,7 @@ 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;
@@ -307,10 +368,15 @@ 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;
 	}
 
 	return ret;
@@ -326,7 +392,6 @@ int td_io_commit(struct thread_data *td)
 		return 0;
 
 	io_u_mark_depth(td, td->io_u_queued);
-	td->io_u_queued = 0;
 
 	if (td->io_ops->commit) {
 		ret = td->io_ops->commit(td);
@@ -334,6 +399,12 @@ int td_io_commit(struct thread_data *td)
 			td_verror(td, -ret, "io commit");
 	}
 
+	/*
+	 * Reflect that events were submitted as async IO requests.
+	 */
+	td->io_u_in_flight += td->io_u_queued;
+	td->io_u_queued = 0;
+
 	return 0;
 }
 
@@ -356,7 +427,7 @@ int td_io_open_file(struct thread_data *td, struct fio_file *f)
 		return 1;
 	}
 
-	fio_file_reset(f);
+	fio_file_reset(td, f);
 	fio_file_set_open(f);
 	fio_file_clear_closing(f);
 	disk_util_inc(f->du);
@@ -371,26 +442,58 @@ 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 (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
+		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;
+
+		/*
+		 * 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 (fcntl(f->fd, cmd, &hint) < 0) {
+			td_verror(td, errno, "fcntl write hint");
+			goto err;
+		}
+	}
+#endif
 
 #ifdef FIO_OS_DIRECTIO
 	/*
@@ -402,6 +505,12 @@ int td_io_open_file(struct thread_data *td, struct fio_file *f)
 
 		if (ret) {
 			td_verror(td, ret, "fio_set_odirect");
+			if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
+				log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
+			} else {
+				log_err("fio: the file system does not seem to support direct IO\n");
+			}
+
 			goto err;
 		}
 	}
@@ -428,72 +537,72 @@ 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);
+
+	if (td->o.file_lock_mode != FILE_LOCK_NONE)
+		unlock_file_all(td, f);
 
 	return put_file(td, f);
 }
 
-int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
+int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
 {
-	if (!td->io_ops->get_file_size)
-		return 0;
+	if (td->io_ops->unlink_file)
+		return td->io_ops->unlink_file(td, f);
+	else {
+		int ret;
 
-	return td->io_ops->get_file_size(td, f);
+		ret = unlink(f->file_name);
+		if (ret < 0)
+			return errno;
+
+		return 0;
+	}
 }
 
-static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
+int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
 {
-	off64_t offset, nbytes;
-
-	offset = f->first_write;
-	nbytes = f->last_write - f->first_write;
-
-	if (!nbytes)
+	if (!td->io_ops->get_file_size)
 		return 0;
 
-	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
+	return td->io_ops->get_file_size(td, f);
 }
 
-int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
+int fio_show_ioengine_help(const char *engine)
 {
-	int ret;
-
-	if (io_u->ddir == DDIR_SYNC) {
-		ret = fsync(io_u->file->fd);
-	} else if (io_u->ddir == DDIR_DATASYNC) {
-#ifdef FIO_HAVE_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;
+	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) {
+			io_ops = flist_entry(entry, struct ioengine_ops, list);
+			log_info("\t%s\n", io_ops->name);
+		}
+		return 0;
+	}
+	sep = strchr(engine, ',');
+	if (sep) {
+		*sep = 0;
+		sep++;
 	}
 
-	if (ret < 0)
-		io_u->error = errno;
+	memset(&td, 0, sizeof(td));
 
-	return ret;
-}
+	io_ops = load_ioengine(&td, engine);
+	if (!io_ops) {
+		log_info("IO engine %s not found\n", engine);
+		return 1;
+	}
 
-int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
-{
-#ifndef FIO_HAVE_TRIM
-	io_u->error = EINVAL;
-	return 0;
-#else
-	struct fio_file *f = io_u->file;
-	int ret;
+	if (io_ops->options)
+		ret = show_cmd_help(io_ops->options, sep);
+	else
+		log_info("IO engine %s has no options\n", io_ops->name);
 
-	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
-	if (!ret)
-		return io_u->xfer_buflen;;
+	free_ioengine(&td);
 
-	io_u->error = ret;
-	return 0;
-#endif
+	return ret;
 }