[fio.git] / ioengines.c

/*
 * The io parts of the fio tool, includes workers for sync and mmap'ed
 * io, as well as both posix and linux libaio support.
 *
 * sync io is implemented on top of aio.
 *
 * This is not really specific to fio, if the get_io_u/put_io_u and
 * structures was pulled into this as well it would be a perfectly
 * 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"

static FLIST_HEAD(engine_list);

static int 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;
	}

	if (!ops->queue) {
		log_err("%s: no queue handler\n", ops->name);
		return 1;
	}

	/*
	 * sync engines only need a ->queue()
	 */
	if (ops->flags & FIO_SYNCIO)
		return 0;

	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;
	}

	return 0;
}

void unregister_ioengine(struct ioengine_ops *ops)
{
	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
	flist_del(&ops->list);
	INIT_FLIST_HEAD(&ops->list);
}

void register_ioengine(struct ioengine_ops *ops)
{
	dprint(FD_IO, "ioengine %s registered\n", ops->name);
	INIT_FLIST_HEAD(&ops->list);
	flist_add_tail(&ops->list, &engine_list);
}

static struct ioengine_ops *find_ioengine(const char *name)
{
	struct ioengine_ops *ops;
	struct flist_head *entry;

	flist_for_each(entry, &engine_list) {
		ops = flist_entry(entry, struct ioengine_ops, list);
		if (!strcmp(name, ops->name))
			return ops;
	}

	return NULL;
}

static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
					    const char *engine_lib)
{
	struct ioengine_ops *ops;
	void *dlhandle;

	dprint(FD_IO, "dload engine %s\n", engine_lib);

	dlerror();
	dlhandle = dlopen(engine_lib, RTLD_LAZY);
	if (!dlhandle) {
		td_vmsg(td, -1, dlerror(), "dlopen");
		return NULL;
	}

	/*
	 * Unlike the included modules, external engines should have a
	 * non-static ioengine structure that we can reference.
	 */
	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;
	return ops;
}

struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
{
	struct ioengine_ops *ops, *ret;
	char engine[16];

	dprint(FD_IO, "load ioengine %s\n", name);

	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))
		strcpy(engine, "libaio");

	ops = find_ioengine(engine);
	if (!ops)
		ops = dlopen_ioengine(td, name);

	if (!ops) {
		log_err("fio: engine %s not loadable\n", name);
		return NULL;
	}

	/*
	 * Check that the required methods are there.
	 */
	if (check_engine_ops(ops))
		return NULL;

	ret = malloc(sizeof(*ret));
	memcpy(ret, ops, sizeof(*ret));
	ret->data = NULL;

	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);

	free(td->io_ops);
	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) {
		td->io_ops->cleanup(td);
		td->io_ops->data = NULL;
	}

	free_ioengine(td);
}

int td_io_prep(struct thread_data *td, struct io_u *io_u)
{
	dprint_io_u(io_u, "prep");
	fio_ro_check(td, io_u);

	lock_file(td, io_u->file, io_u->ddir);

	if (td->io_ops->prep) {
		int ret = td->io_ops->prep(td, io_u);

		dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
		if (ret)
			unlock_file(td, io_u->file);
		return ret;
	}

	return 0;
}

int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
		    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 (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)
{
	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_set(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 (fio_fill_issue_time(td))
			fio_gettime(&io_u->issue_time, NULL);

		/*
		 * only used for iolog
		 */
		if (td->o.read_iolog_file)
			memcpy(&td->last_issue, &io_u->issue_time,
					sizeof(struct timeval));
	}

	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;
	}

	/*
	 * 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
	 * 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. Try "
			 "setting direct=0.\n");
	}

	if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
		io_u_mark_submit(td, 1);
		io_u_mark_complete(td, 1);
	}

	if (ret == FIO_Q_COMPLETED) {
		if (ddir_rw(io_u->ddir)) {
			io_u_mark_depth(td, 1);
			td->ts.total_io_u[io_u->ddir]++;
		}
	} else if (ret == FIO_Q_QUEUED) {
		int r;

		if (ddir_rw(io_u->ddir)) {
			td->io_u_queued++;
			td->ts.total_io_u[io_u->ddir]++;
		}

		if (td->io_u_queued >= td->o.iodepth_batch) {
			r = td_io_commit(td);
			if (r < 0)
				return r;
		}
	}

	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
		if (fio_fill_issue_time(td))
			fio_gettime(&io_u->issue_time, NULL);

		/*
		 * only used for iolog
		 */
		if (td->o.read_iolog_file)
			memcpy(&td->last_issue, &io_u->issue_time,
					sizeof(struct timeval));
	}

	return ret;
}

int td_io_init(struct thread_data *td)
{
	int ret = 0;

	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 (!td->error)
			td->error = ret;
	}

	if (!ret && (td->io_ops->flags & FIO_NOIO))
		td->flags |= TD_F_NOIO;

	return ret;
}

int td_io_commit(struct thread_data *td)
{
	int ret;

	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);

	if (!td->cur_depth || !td->io_u_queued)
		return 0;

	io_u_mark_depth(td, td->io_u_queued);

	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;

	return 0;
}

int td_io_open_file(struct thread_data *td, struct fio_file *f)
{
	assert(!fio_file_open(f));
	assert(f->fd == -1);

	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");
		if (td->error == EMFILE) {
			log_err("fio: try reducing/setting openfiles (failed"
				" at %u of %u)\n", td->nr_open_files,
							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");
			goto err;
		}
	}

	if (td->io_ops->flags & 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)) {
		int flags;

		if (td_random(td))
			flags = POSIX_FADV_RANDOM;
		else
			flags = POSIX_FADV_SEQUENTIAL;

		if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
			td_verror(td, errno, "fadvise");
			goto err;
		}
	}
#ifdef FIO_HAVE_STREAMID
	if (td->o.fadvise_stream &&
	    (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
		off_t stream = td->o.fadvise_stream;

		if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
			td_verror(td, errno, "fadvise streamid");
			goto err;
		}
	}
#endif

#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);

		if (ret) {
			td_verror(td, ret, "fio_set_odirect");
			log_err("fio: the file system does not seem to support direct IO\n");
			goto err;
		}
	}
#endif

done:
	log_file(td, f, FIO_LOG_OPEN_FILE);
	return 0;
err:
	disk_util_dec(f->du);
	if (td->io_ops->close_file)
		td->io_ops->close_file(td, f);
	return 1;
}

int td_io_close_file(struct thread_data *td, struct fio_file *f)
{
	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
	 */
	fio_file_set_closing(f);

	disk_util_dec(f->du);

	if (td->o.file_lock_mode != FILE_LOCK_NONE)
		unlock_file_all(td, 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
		return unlink(f->file_name);
}

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);
}

static int do_sync_file_range(const 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)
		return 0;

	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
}

int do_io_u_sync(const 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(const 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;

	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
	if (!ret)
		return io_u->xfer_buflen;

	io_u->error = ret;
	return 0;
#endif
}

int fio_show_ioengine_help(const char *engine)
{
	struct flist_head *entry;
	struct thread_data td;
	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);
		}
		return 0;
	}
	sep = strchr(engine, ',');
	if (sep) {
		*sep = 0;
		sep++;
	}

	memset(&td, 0, sizeof(td));

	td.io_ops = load_ioengine(&td, engine);
	if (!td.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);
	else
		log_info("IO engine %s has no options\n", td.io_ops->name);

	free_ioengine(&td);

	return ret;
}
Commit	Line	Data
ebac4655 JA	1	/*
	2	* The io parts of the fio tool, includes workers for sync and mmap'ed
	3	* io, as well as both posix and linux libaio support.
	4	*
	5	* sync io is implemented on top of aio.
	6	*
	7	* This is not really specific to fio, if the get_io_u/put_io_u and
	8	* structures was pulled into this as well it would be a perfectly
	9	* generic io engine that could be used for other projects.
	10	*
	11	*/
	12	#include <stdio.h>
	13	#include <stdlib.h>
	14	#include <unistd.h>
5c4e1dbc	15	#include <string.h>
2866c82d	16	#include <dlfcn.h>
ecc314ba	17	#include <fcntl.h>
0c6e7517	18	#include <assert.h>
8c16d840	19
ebac4655	20	#include "fio.h"
7c9b1bce	21	#include "diskutil.h"
ebac4655	22
01743ee1	23	static FLIST_HEAD(engine_list);
5f350952	24
8c16d840 JA	25	static int check_engine_ops(struct ioengine_ops *ops)
8c16d840 JA	26	{
5f350952	27	if (ops->version != FIO_IOOPS_VERSION) {
5ec10eaa JA	28	log_err("bad ioops version %d (want %d)\n", ops->version,
5ec10eaa JA	29	FIO_IOOPS_VERSION);
5f350952 JA	30	return 1;
	31	}
	32
36167d82 JA	33	if (!ops->queue) {
	34	log_err("%s: no queue handler\n", ops->name);
	35	return 1;
	36	}
	37
	38	/*
	39	* sync engines only need a ->queue()
	40	*/
	41	if (ops->flags & FIO_SYNCIO)
	42	return 0;
5ec10eaa	43
8c16d840	44	if (!ops->event) {
36167d82	45	log_err("%s: no event handler\n", ops->name);
8c16d840 JA	46	return 1;
	47	}
	48	if (!ops->getevents) {
36167d82	49	log_err("%s: no getevents handler\n", ops->name);
8c16d840 JA	50	return 1;
	51	}
	52	if (!ops->queue) {
36167d82	53	log_err("%s: no queue handler\n", ops->name);
8c16d840 JA	54	return 1;
8c16d840 JA	55	}
5ec10eaa	56
8c16d840 JA	57	return 0;
	58	}
	59
5f350952	60	void unregister_ioengine(struct ioengine_ops *ops)
ebac4655	61	{
ee56ad50	62	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1 JA	63	flist_del(&ops->list);
01743ee1 JA	64	INIT_FLIST_HEAD(&ops->list);
5f350952 JA	65	}
5f350952 JA	66
b2fdda43	67	void register_ioengine(struct ioengine_ops *ops)
5f350952	68	{
ee56ad50	69	dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1 JA	70	INIT_FLIST_HEAD(&ops->list);
01743ee1 JA	71	flist_add_tail(&ops->list, &engine_list);
5f350952 JA	72	}
	73
	74	static struct ioengine_ops find_ioengine(const char name)
	75	{
	76	struct ioengine_ops *ops;
01743ee1	77	struct flist_head *entry;
ebac4655	78
01743ee1 JA	79	flist_for_each(entry, &engine_list) {
01743ee1 JA	80	ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8	81	if (!strcmp(name, ops->name))
5f350952 JA	82	return ops;
	83	}
	84
	85	return NULL;
	86	}
	87
	88	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	89	const char *engine_lib)
	90	{
	91	struct ioengine_ops *ops;
	92	void *dlhandle;
	93
ee56ad50 JA	94	dprint(FD_IO, "dload engine %s\n", engine_lib);
ee56ad50 JA	95
2866c82d JA	96	dlerror();
2866c82d JA	97	dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8	98	if (!dlhandle) {
e1161c32	99	td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8 JA	100	return NULL;
d4dbaaa8 JA	101	}
8756e4d4	102
da51c050 JA	103	/*
	104	* Unlike the included modules, external engines should have a
	105	* non-static ioengine structure that we can reference.
	106	*/
0abea0b7 DM	107	ops = dlsym(dlhandle, engine_lib);
	108	if (!ops)
	109	ops = dlsym(dlhandle, "ioengine");
a8075704 DG	110
	111	/*
	112	* For some external engines (like C++ ones) it is not that trivial
	113	* to provide a non-static ionengine structure that we can reference.
	114	* Instead we call a method which allocates the required ioengine
	115	* structure.
	116	*/
	117	if (!ops) {
	118	get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
	119
	120	if (get_ioengine)
	121	get_ioengine(&ops);
	122	}
	123
d4dbaaa8	124	if (!ops) {
e1161c32	125	td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8 JA	126	dlclose(dlhandle);
	127	return NULL;
	128	}
8756e4d4	129
5f350952 JA	130	ops->dlhandle = dlhandle;
	131	return ops;
	132	}
	133
	134	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	135	{
	136	struct ioengine_ops ops, ret;
	137	char engine[16];
	138
ee56ad50 JA	139	dprint(FD_IO, "load ioengine %s\n", name);
ee56ad50 JA	140
5f350952 JA	141	strncpy(engine, name, sizeof(engine) - 1);
	142
	143	/*
	144	* linux libaio has alias names, so convert to what we want
	145	*/
	146	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	147	strcpy(engine, "libaio");
	148
	149	ops = find_ioengine(engine);
	150	if (!ops)
	151	ops = dlopen_ioengine(td, name);
	152
	153	if (!ops) {
	154	log_err("fio: engine %s not loadable\n", name);
b902ceb5 JA	155	return NULL;
	156	}
	157
8c16d840 JA	158	/*
	159	* Check that the required methods are there.
	160	*/
5f350952	161	if (check_engine_ops(ops))
8c16d840	162	return NULL;
8c16d840	163
84585003 JA	164	ret = malloc(sizeof(*ret));
	165	memcpy(ret, ops, sizeof(*ret));
	166	ret->data = NULL;
84585003 JA	167
84585003 JA	168	return ret;
8756e4d4 JA	169	}
8756e4d4 JA	170
de890a1e SL	171	/*
	172	* For cleaning up an ioengine which never made it to init().
	173	*/
	174	void free_ioengine(struct thread_data *td)
8756e4d4	175	{
de890a1e	176	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50	177
de890a1e SL	178	if (td->eo && td->io_ops->options) {
	179	options_free(td->io_ops->options, td->eo);
	180	free(td->eo);
	181	td->eo = NULL;
2992b059	182	}
b990b5c0	183
5f350952 JA	184	if (td->io_ops->dlhandle)
	185	dlclose(td->io_ops->dlhandle);
	186
84585003 JA	187	free(td->io_ops);
84585003 JA	188	td->io_ops = NULL;
b990b5c0	189	}
10ba535a	190
de890a1e SL	191	void close_ioengine(struct thread_data *td)
	192	{
	193	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	194
	195	if (td->io_ops->cleanup) {
	196	td->io_ops->cleanup(td);
	197	td->io_ops->data = NULL;
	198	}
	199
	200	free_ioengine(td);
	201	}
	202
10ba535a JA	203	int td_io_prep(struct thread_data td, struct io_u io_u)
10ba535a JA	204	{
ee56ad50	205	dprint_io_u(io_u, "prep");
7101d9c2 JA	206	fio_ro_check(td, io_u);
7101d9c2 JA	207
4d4e80f2	208	lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9	209
2ba1c290 JA	210	if (td->io_ops->prep) {
	211	int ret = td->io_ops->prep(td, io_u);
	212
	213	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9	214	if (ret)
4d4e80f2	215	unlock_file(td, io_u->file);
2ba1c290 JA	216	return ret;
2ba1c290 JA	217	}
10ba535a JA	218
	219	return 0;
	220	}
	221
e7d2e616	222	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece	223	const struct timespec *t)
10ba535a	224	{
ee56ad50	225	int r = 0;
face81b2	226
a05d62b2 YR	227	/*
	228	* For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	229	* server side gets a message from the client
	230	* side that the task is finished, and
	231	* td->done is set to 1 after td_io_commit(). In this case,
	232	* there is no need to reap complete event in server side.
	233	*/
	234	if (td->done)
	235	return 0;
	236
ee56ad50 JA	237	if (min > 0 && td->io_ops->commit) {
ee56ad50 JA	238	r = td->io_ops->commit(td);
face81b2	239	if (r < 0)
ee56ad50	240	goto out;
face81b2	241	}
4950421a JA	242	if (max > td->cur_depth)
	243	max = td->cur_depth;
	244	if (min > max)
	245	max = min;
36167d82	246
ee56ad50	247	r = 0;
4950421a	248	if (max && td->io_ops->getevents)
ee56ad50 JA	249	r = td->io_ops->getevents(td, min, max, t);
ee56ad50 JA	250	out:
422f9e4b RM	251	if (r >= 0) {
422f9e4b RM	252	/*
3fd9efbc	253	* Reflect that our submitted requests were retrieved with
422f9e4b RM	254	* whatever OS async calls are in the underlying engine.
	255	*/
	256	td->io_u_in_flight -= r;
838bc709	257	io_u_mark_complete(td, r);
422f9e4b	258	} else
7c639b14	259	td_verror(td, r, "get_events");
f3e11d05	260
ee56ad50 JA	261	dprint(FD_IO, "getevents: %d\n", r);
ee56ad50 JA	262	return r;
10ba535a JA	263	}
	264
	265	int td_io_queue(struct thread_data td, struct io_u io_u)
	266	{
a9da8ab2 JA	267	const enum fio_ddir ddir = acct_ddir(io_u);
a9da8ab2 JA	268	unsigned long buflen = io_u->xfer_buflen;
7e77dd02 JA	269	int ret;
7e77dd02 JA	270
ee56ad50	271	dprint_io_u(io_u, "queue");
7101d9c2 JA	272	fio_ro_check(td, io_u);
7101d9c2 JA	273
0c6e7517	274	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
a9da8ab2	275	io_u_set(io_u, IO_U_F_FLIGHT);
0c6e7517	276
d6aed795	277	assert(fio_file_open(io_u->file));
3d7b485f	278
bcd5abfa JA	279	/*
	280	* If using a write iolog, store this entry.
	281	*/
	282	log_io_u(td, io_u);
	283
11786802 JA	284	io_u->error = 0;
	285	io_u->resid = 0;
	286
433afcb4	287	if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841	288	if (fio_fill_issue_time(td))
9520ebb9	289	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	290
	291	/*
	292	* only used for iolog
	293	*/
	294	if (td->o.read_iolog_file)
	295	memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa	296	sizeof(struct timeval));
433afcb4 JA	297	}
433afcb4 JA	298
a9da8ab2 JA	299	if (ddir_rw(ddir)) {
	300	td->io_issues[ddir]++;
	301	td->io_issue_bytes[ddir] += buflen;
50a8ce86	302	td->rate_io_issue_bytes[ddir] += buflen;
74d6277f	303	}
755200a3	304
7e77dd02	305	ret = td->io_ops->queue(td, io_u);
5aeb77df	306
4d4e80f2	307	unlock_file(td, io_u->file);
b2bd2bd9	308
a9da8ab2 JA	309	if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
	310	td->io_issues[ddir]--;
	311	td->io_issue_bytes[ddir] -= buflen;
50a8ce86	312	td->rate_io_issue_bytes[ddir] -= buflen;
cd8a19e6 JA	313	}
cd8a19e6 JA	314
39a43d34 JA	315	/*
	316	* If an error was seen and the io engine didn't propagate it
	317	* back to 'td', do so.
	318	*/
	319	if (io_u->error && !td->error)
	320	td_verror(td, io_u->error, "td_io_queue");
	321
cb211682 JA	322	/*
	323	* Add warning for O_DIRECT so that users have an easier time
	324	* spotting potentially bad alignment. If this triggers for the first
	325	* IO, then it's likely an alignment problem or because the host fs
	326	* does not support O_DIRECT
	327	*/
ff58fced	328	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682	329	td->o.odirect) {
214ac7e0	330
cb211682	331	log_info("fio: first direct IO errored. File system may not "
44cc9c8f JA	332	"support direct IO, or iomem_align= is bad. Try "
44cc9c8f JA	333	"setting direct=0.\n");
cb211682 JA	334	}
cb211682 JA	335
77b451cc	336	if (!td->io_ops->commit \|\| io_u->ddir == DDIR_TRIM) {
838bc709 JA	337	io_u_mark_submit(td, 1);
	338	io_u_mark_complete(td, 1);
	339	}
	340
d8005759	341	if (ret == FIO_Q_COMPLETED) {
ff58fced	342	if (ddir_rw(io_u->ddir)) {
d8005759 JA	343	io_u_mark_depth(td, 1);
d8005759 JA	344	td->ts.total_io_u[io_u->ddir]++;
6eaf09d6	345	}
d8005759	346	} else if (ret == FIO_Q_QUEUED) {
eb7c8ae2 JA	347	int r;
eb7c8ae2 JA	348
ff58fced	349	if (ddir_rw(io_u->ddir)) {
d8005759 JA	350	td->io_u_queued++;
	351	td->ts.total_io_u[io_u->ddir]++;
	352	}
	353
	354	if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2 JA	355	r = td_io_commit(td);
	356	if (r < 0)
	357	return r;
	358	}
	359	}
cb5ab512	360
433afcb4	361	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841	362	if (fio_fill_issue_time(td))
9520ebb9	363	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	364
	365	/*
	366	* only used for iolog
	367	*/
	368	if (td->o.read_iolog_file)
	369	memcpy(&td->last_issue, &io_u->issue_time,
	370	sizeof(struct timeval));
433afcb4 JA	371	}
433afcb4 JA	372
7e77dd02	373	return ret;
10ba535a	374	}
8c16d840 JA	375
	376	int td_io_init(struct thread_data *td)
	377	{
eeb12160	378	int ret = 0;
8c16d840	379
eeb12160 JA	380	if (td->io_ops->init) {
eeb12160 JA	381	ret = td->io_ops->init(td);
5ec10eaa JA	382	if (ret && td->o.iodepth > 1) {
	383	log_err("fio: io engine init failed. Perhaps try"
	384	" reducing io depth?\n");
	385	}
7c973896 JA	386	if (!td->error)
7c973896 JA	387	td->error = ret;
eeb12160 JA	388	}
eeb12160 JA	389
046395d7 JA	390	if (!ret && (td->io_ops->flags & FIO_NOIO))
	391	td->flags \|= TD_F_NOIO;
	392
eeb12160	393	return ret;
8c16d840	394	}
755200a3 JA	395
	396	int td_io_commit(struct thread_data *td)
	397	{
f3e11d05 JA	398	int ret;
f3e11d05 JA	399
ee56ad50 JA	400	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
ee56ad50 JA	401
d8005759	402	if (!td->cur_depth \|\| !td->io_u_queued)
e1161c32	403	return 0;
cb5ab512	404
3fd9efbc	405	io_u_mark_depth(td, td->io_u_queued);
d8005759	406
f3e11d05 JA	407	if (td->io_ops->commit) {
	408	ret = td->io_ops->commit(td);
	409	if (ret)
	410	td_verror(td, -ret, "io commit");
	411	}
3fd9efbc	412
422f9e4b RM	413	/*
	414	* Reflect that events were submitted as async IO requests.
	415	*/
	416	td->io_u_in_flight += td->io_u_queued;
	417	td->io_u_queued = 0;
755200a3 JA	418
	419	return 0;
	420	}
b5af8293 JA	421
	422	int td_io_open_file(struct thread_data td, struct fio_file f)
	423	{
22a57ba8 JA	424	assert(!fio_file_open(f));
	425	assert(f->fd == -1);
	426
413d6693 JA	427	if (td->io_ops->open_file(td, f)) {
	428	if (td->error == EINVAL && td->o.odirect)
	429	log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa JA	430	if (td->error == EMFILE) {
	431	log_err("fio: try reducing/setting openfiles (failed"
	432	" at %u of %u)\n", td->nr_open_files,
	433	td->o.nr_files);
	434	}
413d6693	435
22a57ba8 JA	436	assert(f->fd == -1);
22a57ba8 JA	437	assert(!fio_file_open(f));
413d6693 JA	438	return 1;
	439	}
	440
33c48814	441	fio_file_reset(td, f);
d6aed795 JA	442	fio_file_set_open(f);
d6aed795 JA	443	fio_file_clear_closing(f);
c97bd0fa	444	disk_util_inc(f->du);
d5707a35 JA	445
	446	td->nr_open_files++;
	447	get_file(f);
	448
66159828 JA	449	if (f->filetype == FIO_TYPE_PIPE) {
	450	if (td_random(td)) {
	451	log_err("fio: can't seek on pipes (no random io)\n");
	452	goto err;
	453	}
	454	}
	455
413d6693 JA	456	if (td->io_ops->flags & FIO_DISKLESSIO)
	457	goto done;
	458
	459	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	460	goto err;
	461
66159828 JA	462	if (td->o.fadvise_hint &&
66159828 JA	463	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
413d6693 JA	464	int flags;
	465
	466	if (td_random(td))
	467	flags = POSIX_FADV_RANDOM;
	468	else
	469	flags = POSIX_FADV_SEQUENTIAL;
	470
ecc314ba	471	if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693 JA	472	td_verror(td, errno, "fadvise");
	473	goto err;
	474	}
7bb48f84	475	}
37659335 JA	476	#ifdef FIO_HAVE_STREAMID
	477	if (td->o.fadvise_stream &&
	478	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
	479	off_t stream = td->o.fadvise_stream;
	480
	481	if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
	482	td_verror(td, errno, "fadvise streamid");
	483	goto err;
	484	}
	485	}
	486	#endif
a978ba68	487
e116f2b9 JA	488	#ifdef FIO_OS_DIRECTIO
	489	/*
	490	* Some OS's have a distinct call to mark the file non-buffered,
	491	* instead of using O_DIRECT (Solaris)
	492	*/
	493	if (td->o.odirect) {
	494	int ret = fio_set_odirect(f->fd);
	495
	496	if (ret) {
	497	td_verror(td, ret, "fio_set_odirect");
78e51c72	498	log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9 JA	499	goto err;
	500	}
	501	}
	502	#endif
	503
413d6693	504	done:
f29b25a3	505	log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693 JA	506	return 0;
413d6693 JA	507	err:
c97bd0fa	508	disk_util_dec(f->du);
b284075a JA	509	if (td->io_ops->close_file)
b284075a JA	510	td->io_ops->close_file(td, f);
7bb48f84	511	return 1;
b5af8293 JA	512	}
b5af8293 JA	513
6977bcd0	514	int td_io_close_file(struct thread_data td, struct fio_file f)
b5af8293	515	{
d6aed795	516	if (!fio_file_closing(f))
f29b25a3 JA	517	log_file(td, f, FIO_LOG_CLOSE_FILE);
f29b25a3 JA	518
0ad920e7 JA	519	/*
	520	* mark as closing, do real close when last io on it has completed
	521	*/
d6aed795	522	fio_file_set_closing(f);
0ad920e7	523
c97bd0fa	524	disk_util_dec(f->du);
cba5243b JA	525
	526	if (td->o.file_lock_mode != FILE_LOCK_NONE)
	527	unlock_file_all(td, f);
29c1349f	528
6977bcd0	529	return put_file(td, f);
b5af8293	530	}
df9c26b1	531
38ef9c90 CF	532	int td_io_unlink_file(struct thread_data td, struct fio_file f)
	533	{
	534	if (td->io_ops->unlink_file)
	535	return td->io_ops->unlink_file(td, f);
	536	else
	537	return unlink(f->file_name);
	538	}
	539
df9c26b1 JA	540	int td_io_get_file_size(struct thread_data td, struct fio_file f)
	541	{
	542	if (!td->io_ops->get_file_size)
	543	return 0;
	544
	545	return td->io_ops->get_file_size(td, f);
	546	}
44f29692	547
effd6ff0 JA	548	static int do_sync_file_range(const struct thread_data *td,
effd6ff0 JA	549	struct fio_file *f)
44f29692 JA	550	{
	551	off64_t offset, nbytes;
	552
	553	offset = f->first_write;
	554	nbytes = f->last_write - f->first_write;
	555
3843deb3 JA	556	if (!nbytes)
3843deb3 JA	557	return 0;
44f29692	558
3843deb3	559	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692	560	}
0a28ecda	561
effd6ff0	562	int do_io_u_sync(const struct thread_data td, struct io_u io_u)
0a28ecda JA	563	{
	564	int ret;
	565
	566	if (io_u->ddir == DDIR_SYNC) {
	567	ret = fsync(io_u->file->fd);
	568	} else if (io_u->ddir == DDIR_DATASYNC) {
67bf9823	569	#ifdef CONFIG_FDATASYNC
0a28ecda JA	570	ret = fdatasync(io_u->file->fd);
	571	#else
	572	ret = io_u->xfer_buflen;
	573	io_u->error = EINVAL;
	574	#endif
	575	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
	576	ret = do_sync_file_range(td, io_u->file);
	577	else {
	578	ret = io_u->xfer_buflen;
	579	io_u->error = EINVAL;
	580	}
	581
cb849a79 JA	582	if (ret < 0)
	583	io_u->error = errno;
	584
0a28ecda JA	585	return ret;
0a28ecda JA	586	}
a5f3027c	587
effd6ff0	588	int do_io_u_trim(const struct thread_data td, struct io_u io_u)
a5f3027c JA	589	{
	590	#ifndef FIO_HAVE_TRIM
	591	io_u->error = EINVAL;
ff58fced	592	return 0;
a5f3027c JA	593	#else
	594	struct fio_file *f = io_u->file;
	595	int ret;
	596
c6404a44	597	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c	598	if (!ret)
3fd9efbc	599	return io_u->xfer_buflen;
a5f3027c	600
ff58fced JA	601	io_u->error = ret;
ff58fced JA	602	return 0;
a5f3027c JA	603	#endif
a5f3027c JA	604	}
de890a1e SL	605
	606	int fio_show_ioengine_help(const char *engine)
	607	{
	608	struct flist_head *entry;
	609	struct thread_data td;
	610	char *sep;
	611	int ret = 1;
	612
	613	if (!engine \|\| !*engine) {
	614	log_info("Available IO engines:\n");
	615	flist_for_each(entry, &engine_list) {
	616	td.io_ops = flist_entry(entry, struct ioengine_ops,
	617	list);
	618	log_info("\t%s\n", td.io_ops->name);
	619	}
	620	return 0;
	621	}
	622	sep = strchr(engine, ',');
	623	if (sep) {
	624	*sep = 0;
	625	sep++;
	626	}
	627
	628	memset(&td, 0, sizeof(td));
	629
	630	td.io_ops = load_ioengine(&td, engine);
	631	if (!td.io_ops) {
	632	log_info("IO engine %s not found\n", engine);
	633	return 1;
	634	}
	635
	636	if (td.io_ops->options)
	637	ret = show_cmd_help(td.io_ops->options, sep);
	638	else
	639	log_info("IO engine %s has no options\n", td.io_ops->name);
	640
	641	free_ioengine(&td);
	642
	643	return ret;
	644	}