X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=ioengines.c;h=aa4ccd2755c96f843d69333448795b0b16061723;hp=984c01ac7e6442a99868ee03388a684c654318e8;hb=01bf5128d0581e267383f280c6a1dcd26517240f;hpb=391ba6ac169ad7deaf3bb3840046581f45d755e1

diff --git a/ioengines.c b/ioengines.c
index 984c01ac..aa4ccd27 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -9,73 +9,65 @@
  * generic io engine that could be used for other projects.
  *
  */
-#include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <string.h>
 #include <dlfcn.h>
+#include <fcntl.h>
 #include <assert.h>
 
 #include "fio.h"
+#include "diskutil.h"
+#include "zbd.h"
 
-static LIST_HEAD(engine_list);
+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)
 {
 	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
-	list_del(&ops->list);
-	INIT_LIST_HEAD(&ops->list);
+	flist_del_init(&ops->list);
 }
 
 void register_ioengine(struct ioengine_ops *ops)
 {
 	dprint(FD_IO, "ioengine %s registered\n", ops->name);
-	INIT_LIST_HEAD(&ops->list);
-	list_add_tail(&ops->list, &engine_list);
+	flist_add_tail(&ops->list, &engine_list);
 }
 
 static struct ioengine_ops *find_ioengine(const char *name)
 {
 	struct ioengine_ops *ops;
-	struct list_head *entry;
+	struct flist_head *entry;
 
-	list_for_each(entry, &engine_list) {
-		ops = list_entry(entry, struct ioengine_ops, list);
+	flist_for_each(entry, &engine_list) {
+		ops = flist_entry(entry, struct ioengine_ops, list);
 		if (!strcmp(name, ops->name))
 			return ops;
 	}
@@ -102,36 +94,79 @@ 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)
+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);
 
 	/*
 	 * linux libaio has alias names, so convert to what we want
 	 */
-	if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
+	if (!strncmp(engine, "linuxaio", 8)) {
+		dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
 		strcpy(engine, "libaio");
+	}
+
+	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;
 
-	ops = find_ioengine(engine);
+	/*
+	 * 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;
@@ -143,25 +178,40 @@ 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_dlhandle = NULL;
+	}
+
+	td->io_ops = NULL;
 }
 
 void close_ioengine(struct thread_data *td)
 {
 	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
 
-	if (td->io_ops->cleanup)
+	if (td->io_ops->cleanup) {
 		td->io_ops->cleanup(td);
+		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)
@@ -174,7 +224,8 @@ int td_io_prep(struct thread_data *td, struct io_u *io_u)
 	if (td->io_ops->prep) {
 		int ret = td->io_ops->prep(td, io_u);
 
-		dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
+		dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
+
 		if (ret)
 			unlock_file(td, io_u->file);
 		return ret;
@@ -184,86 +235,174 @@ 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)
 			goto out;
 	}
+	if (max > td->cur_depth)
+		max = td->cur_depth;
+	if (min > max)
+		max = min;
 
 	r = 0;
-	if (td->io_ops->getevents)
+	if (max && td->io_ops->getevents)
 		r = td->io_ops->getevents(td, min, max, t);
 out:
+	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
+		td_verror(td, r, "get_events");
+
 	dprint(FD_IO, "getevents: %d\n", r);
 	return r;
 }
 
-int td_io_queue(struct thread_data *td, struct io_u *io_u)
+enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
 {
-	int ret;
+	const enum fio_ddir ddir = acct_ddir(io_u);
+	unsigned long long buflen = io_u->xfer_buflen;
+	enum fio_q_status 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);
+
+	/*
+	 * If overlap checking was enabled in offload mode we
+	 * can release this lock that was acquired when we
+	 * started the overlap check because the IO_U_F_FLIGHT
+	 * flag is now set
+	 */
+	if (td_offload_overlap(td))
+		pthread_mutex_unlock(&overlap_check);
 
-	assert(io_u->file->flags & FIO_FILE_OPEN);
+	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) {
-		fio_gettime(&io_u->issue_time, NULL);
-		memcpy(&td->last_issue, &io_u->issue_time,
-					sizeof(struct timeval));
+	if (td_ioengine_flagged(td, FIO_SYNCIO) ||
+		(td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) && 
+		io_u->ddir == DDIR_TRIM)) {
+		if (fio_fill_issue_time(td))
+			fio_gettime(&io_u->issue_time, NULL);
 
 		/*
-		 * for a sync engine, set the timeout upfront
+		 * only used for iolog
 		 */
-		if (mtime_since(&td->timeout_end, &io_u->issue_time)
-		    < IO_U_TIMEOUT)
-			io_u_set_timeout(td);
+		if (td->o.read_iolog_file)
+			memcpy(&td->last_issue, &io_u->issue_time,
+					sizeof(io_u->issue_time));
 	}
 
-	if (io_u->ddir != DDIR_SYNC)
-		td->io_issues[io_u->ddir]++;
+	if (ddir_rw(ddir)) {
+		if (!(io_u->flags & IO_U_F_VER_LIST)) {
+			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);
+	zbd_queue_io_u(io_u, ret);
 
 	unlock_file(td, io_u->file);
 
-	if (ret != FIO_Q_BUSY)
-		io_u_mark_depth(td, io_u, 1);
+	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 (ret == FIO_Q_QUEUED) {
-		int r;
+	/*
+	 * 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");
 
-		td->io_u_queued++;
-		if (td->io_u_queued > td->o.iodepth_batch) {
-			r = td_io_commit(td);
-			if (r < 0)
-				return r;
+	/*
+	 * Add warning for O_DIRECT so that users have an easier time
+	 * spotting potentially bad alignment. If this triggers for the first
+	 * IO, then it's likely an alignment problem or because the host fs
+	 * does not support O_DIRECT
+	 */
+	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, or "
+			 "invalid block size. Try setting direct=0.\n");
+	}
+
+	if (zbd_unaligned_write(io_u->error) &&
+	    td->io_issues[io_u->ddir & 1] == 1 &&
+	    td->o.zone_mode != ZONE_MODE_ZBD) {
+		log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
+			 io_u->file->file_name);
+	}
+
+	if (!td->io_ops->commit) {
+		io_u_mark_submit(td, 1);
+		io_u_mark_complete(td, 1);
+		zbd_put_io_u(io_u);
+	}
+
+	if (ret == FIO_Q_COMPLETED) {
+		if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir)) {
+			io_u_mark_depth(td, 1);
+			td->ts.total_io_u[io_u->ddir]++;
 		}
+	} else if (ret == FIO_Q_QUEUED) {
+		td->io_u_queued++;
+
+		if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir))
+			td->ts.total_io_u[io_u->ddir]++;
+
+		if (td->io_u_queued >= td->o.iodepth_batch)
+			td_io_commit(td);
 	}
 
-	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
-		fio_gettime(&io_u->issue_time, NULL);
-		memcpy(&td->last_issue, &io_u->issue_time,
-				sizeof(struct timeval));
+	if (!td_ioengine_flagged(td, FIO_SYNCIO) &&
+		(!td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) ||
+		 io_u->ddir != DDIR_TRIM)) {
+		if (fio_fill_issue_time(td))
+			fio_gettime(&io_u->issue_time, NULL);
 
 		/*
-		 * async engine, set the timeout here
+		 * only used for iolog
 		 */
-		if (ret == FIO_Q_QUEUED &&
-		    (mtime_since(&td->timeout_end, &io_u->issue_time)
-			< IO_U_TIMEOUT)) {
-			io_u_set_timeout(td);
-		}
+		if (td->o.read_iolog_file)
+			memcpy(&td->last_issue, &io_u->issue_time,
+					sizeof(io_u->issue_time));
 	}
 
 	return ret;
@@ -275,31 +414,58 @@ 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;
 }
 
-int td_io_commit(struct thread_data *td)
+void td_io_commit(struct thread_data *td)
 {
+	int ret;
+
 	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
 
-	if (!td->cur_depth)
-		return 0;
+	if (!td->cur_depth || !td->io_u_queued)
+		return;
 
-	td->io_u_queued = 0;
-	if (td->io_ops->commit)
-		return td->io_ops->commit(td);
+	io_u_mark_depth(td, td->io_u_queued);
 
-	return 0;
+	if (td->io_ops->commit) {
+		ret = td->io_ops->commit(td);
+		if (ret)
+			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;
 }
 
 int td_io_open_file(struct thread_data *td, struct fio_file *f)
 {
+	if (fio_file_closing(f)) {
+		/*
+		 * Open translates to undo closing.
+		 */
+		fio_file_clear_closing(f);
+		get_file(f);
+		return 0;
+	}
+	assert(!fio_file_open(f));
+	assert(f->fd == -1);
+	assert(td->io_ops->open_file);
+
 	if (td->io_ops->open_file(td, f)) {
 		if (td->error == EINVAL && td->o.odirect)
 			log_err("fio: destination does not support O_DIRECT\n");
@@ -309,9 +475,19 @@ int td_io_open_file(struct thread_data *td, struct fio_file *f)
 							td->o.nr_files);
 		}
 
+		assert(f->fd == -1);
+		assert(!fio_file_open(f));
 		return 1;
 	}
 
+	fio_file_reset(td, f);
+	fio_file_set_open(f);
+	fio_file_clear_closing(f);
+	disk_util_inc(f->du);
+
+	td->nr_open_files++;
+	get_file(f);
+
 	if (f->filetype == FIO_TYPE_PIPE) {
 		if (td_random(td)) {
 			log_err("fio: can't seek on pipes (no random io)\n");
@@ -319,41 +495,67 @@ int td_io_open_file(struct thread_data *td, struct fio_file *f)
 		}
 	}
 
-	f->last_free_lookup = 0;
-	f->last_pos = f->file_offset;
-	f->flags |= FIO_FILE_OPEN;
-	f->flags &= ~FIO_FILE_CLOSING;
-
-	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) {
+			if (!fio_did_warn(FIO_WARN_FADVISE))
+				log_err("fio: fadvise hint failed\n");
+		}
+	}
+#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;
 
-		if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
-			td_verror(td, errno, "fadvise");
+		/*
+		 * 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
 
-	if (f->file_map)
-		memset(f->file_map, 0, f->num_maps * sizeof(long));
+	if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
+		goto err;
 
 done:
 	log_file(td, f, FIO_LOG_OPEN_FILE);
-	td->nr_open_files++;
-	get_file(f);
 	return 0;
 err:
+	disk_util_dec(f->du);
 	if (td->io_ops->close_file)
 		td->io_ops->close_file(td, f);
 	return 1;
@@ -361,15 +563,76 @@ err:
 
 int td_io_close_file(struct thread_data *td, struct fio_file *f)
 {
-	if (!(f->flags & FIO_FILE_CLOSING))
+	if (!fio_file_closing(f))
 		log_file(td, f, FIO_LOG_CLOSE_FILE);
 
 	/*
 	 * mark as closing, do real close when last io on it has completed
 	 */
-	f->flags |= FIO_FILE_CLOSING;
-
-	unlock_file_all(td, f);
+	fio_file_set_closing(f);
 
 	return put_file(td, f);
 }
+
+int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
+{
+	if (td->io_ops->unlink_file)
+		return td->io_ops->unlink_file(td, f);
+	else {
+		int ret;
+
+		ret = unlink(f->file_name);
+		if (ret < 0)
+			return errno;
+
+		return 0;
+	}
+}
+
+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);
+}
+
+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) {
+			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++;
+	}
+
+	memset(&td, 0, sizeof(struct thread_data));
+	td.o.ioengine = (char *)engine;
+	io_ops = load_ioengine(&td);
+
+	if (!io_ops) {
+		log_info("IO engine %s not found\n", engine);
+		return 1;
+	}
+
+	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);
+
+	free_ioengine(&td);
+	return ret;
+}