[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 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 true;
	}

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

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

	if (!ops->event || !ops->getevents) {
		log_err("%s: no event/getevents handler\n", ops->name);
		return true;
	}

	return false;
}

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;
	}

	td->io_ops_dlhandle = dlhandle;
	return ops;
}

struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
{
	struct ioengine_ops *ops;
	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;

	return ops;
}

/*
 * 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)
{
	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;

		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);
			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
1cc954ba	25	static bool check_engine_ops(struct ioengine_ops *ops)
8c16d840	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);
1cc954ba	30	return true;
5f350952 JA	31	}
5f350952 JA	32
36167d82 JA	33	if (!ops->queue) {
36167d82 JA	34	log_err("%s: no queue handler\n", ops->name);
1cc954ba	35	return true;
36167d82 JA	36	}
	37
	38	/*
	39	* sync engines only need a ->queue()
	40	*/
	41	if (ops->flags & FIO_SYNCIO)
1cc954ba	42	return false;
5ec10eaa	43
1cc954ba JA	44	if (!ops->event \|\| !ops->getevents) {
	45	log_err("%s: no event/getevents handler\n", ops->name);
	46	return true;
8c16d840	47	}
5ec10eaa	48
1cc954ba	49	return false;
8c16d840 JA	50	}
8c16d840 JA	51
5f350952	52	void unregister_ioengine(struct ioengine_ops *ops)
ebac4655	53	{
ee56ad50	54	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1 JA	55	flist_del(&ops->list);
01743ee1 JA	56	INIT_FLIST_HEAD(&ops->list);
5f350952 JA	57	}
5f350952 JA	58
b2fdda43	59	void register_ioengine(struct ioengine_ops *ops)
5f350952	60	{
ee56ad50	61	dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1 JA	62	INIT_FLIST_HEAD(&ops->list);
01743ee1 JA	63	flist_add_tail(&ops->list, &engine_list);
5f350952 JA	64	}
	65
	66	static struct ioengine_ops find_ioengine(const char name)
	67	{
	68	struct ioengine_ops *ops;
01743ee1	69	struct flist_head *entry;
ebac4655	70
01743ee1 JA	71	flist_for_each(entry, &engine_list) {
01743ee1 JA	72	ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8	73	if (!strcmp(name, ops->name))
5f350952 JA	74	return ops;
	75	}
	76
	77	return NULL;
	78	}
	79
	80	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	81	const char *engine_lib)
	82	{
	83	struct ioengine_ops *ops;
	84	void *dlhandle;
	85
ee56ad50 JA	86	dprint(FD_IO, "dload engine %s\n", engine_lib);
ee56ad50 JA	87
2866c82d JA	88	dlerror();
2866c82d JA	89	dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8	90	if (!dlhandle) {
e1161c32	91	td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8 JA	92	return NULL;
d4dbaaa8 JA	93	}
8756e4d4	94
da51c050 JA	95	/*
	96	* Unlike the included modules, external engines should have a
	97	* non-static ioengine structure that we can reference.
	98	*/
0abea0b7 DM	99	ops = dlsym(dlhandle, engine_lib);
	100	if (!ops)
	101	ops = dlsym(dlhandle, "ioengine");
a8075704 DG	102
	103	/*
	104	* For some external engines (like C++ ones) it is not that trivial
	105	* to provide a non-static ionengine structure that we can reference.
	106	* Instead we call a method which allocates the required ioengine
	107	* structure.
	108	*/
	109	if (!ops) {
	110	get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
	111
	112	if (get_ioengine)
	113	get_ioengine(&ops);
	114	}
	115
d4dbaaa8	116	if (!ops) {
e1161c32	117	td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8 JA	118	dlclose(dlhandle);
	119	return NULL;
	120	}
8756e4d4	121
565e784d	122	td->io_ops_dlhandle = dlhandle;
5f350952 JA	123	return ops;
	124	}
	125
	126	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	127	{
565e784d	128	struct ioengine_ops *ops;
5f350952 JA	129	char engine[16];
5f350952 JA	130
ee56ad50 JA	131	dprint(FD_IO, "load ioengine %s\n", name);
ee56ad50 JA	132
5f350952 JA	133	strncpy(engine, name, sizeof(engine) - 1);
	134
	135	/*
	136	* linux libaio has alias names, so convert to what we want
	137	*/
	138	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	139	strcpy(engine, "libaio");
	140
	141	ops = find_ioengine(engine);
	142	if (!ops)
	143	ops = dlopen_ioengine(td, name);
	144
	145	if (!ops) {
	146	log_err("fio: engine %s not loadable\n", name);
b902ceb5 JA	147	return NULL;
	148	}
	149
8c16d840 JA	150	/*
	151	* Check that the required methods are there.
	152	*/
5f350952	153	if (check_engine_ops(ops))
8c16d840	154	return NULL;
8c16d840	155
565e784d	156	return ops;
8756e4d4 JA	157	}
8756e4d4 JA	158
de890a1e SL	159	/*
	160	* For cleaning up an ioengine which never made it to init().
	161	*/
	162	void free_ioengine(struct thread_data *td)
8756e4d4	163	{
de890a1e	164	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50	165
de890a1e SL	166	if (td->eo && td->io_ops->options) {
	167	options_free(td->io_ops->options, td->eo);
	168	free(td->eo);
	169	td->eo = NULL;
2992b059	170	}
b990b5c0	171
565e784d TK	172	if (td->io_ops_dlhandle)
565e784d TK	173	dlclose(td->io_ops_dlhandle);
5f350952	174
84585003	175	td->io_ops = NULL;
b990b5c0	176	}
10ba535a	177
de890a1e SL	178	void close_ioengine(struct thread_data *td)
	179	{
	180	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	181
	182	if (td->io_ops->cleanup) {
	183	td->io_ops->cleanup(td);
565e784d	184	td->io_ops_data = NULL;
de890a1e SL	185	}
	186
	187	free_ioengine(td);
	188	}
	189
10ba535a JA	190	int td_io_prep(struct thread_data td, struct io_u io_u)
10ba535a JA	191	{
ee56ad50	192	dprint_io_u(io_u, "prep");
7101d9c2 JA	193	fio_ro_check(td, io_u);
7101d9c2 JA	194
4d4e80f2	195	lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9	196
2ba1c290 JA	197	if (td->io_ops->prep) {
	198	int ret = td->io_ops->prep(td, io_u);
	199
	200	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9	201	if (ret)
4d4e80f2	202	unlock_file(td, io_u->file);
2ba1c290 JA	203	return ret;
2ba1c290 JA	204	}
10ba535a JA	205
	206	return 0;
	207	}
	208
e7d2e616	209	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece	210	const struct timespec *t)
10ba535a	211	{
ee56ad50	212	int r = 0;
face81b2	213
a05d62b2 YR	214	/*
	215	* For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	216	* server side gets a message from the client
	217	* side that the task is finished, and
	218	* td->done is set to 1 after td_io_commit(). In this case,
	219	* there is no need to reap complete event in server side.
	220	*/
	221	if (td->done)
	222	return 0;
	223
ee56ad50 JA	224	if (min > 0 && td->io_ops->commit) {
ee56ad50 JA	225	r = td->io_ops->commit(td);
face81b2	226	if (r < 0)
ee56ad50	227	goto out;
face81b2	228	}
4950421a JA	229	if (max > td->cur_depth)
	230	max = td->cur_depth;
	231	if (min > max)
	232	max = min;
36167d82	233
ee56ad50	234	r = 0;
4950421a	235	if (max && td->io_ops->getevents)
ee56ad50 JA	236	r = td->io_ops->getevents(td, min, max, t);
ee56ad50 JA	237	out:
422f9e4b RM	238	if (r >= 0) {
422f9e4b RM	239	/*
3fd9efbc	240	* Reflect that our submitted requests were retrieved with
422f9e4b RM	241	* whatever OS async calls are in the underlying engine.
	242	*/
	243	td->io_u_in_flight -= r;
838bc709	244	io_u_mark_complete(td, r);
422f9e4b	245	} else
7c639b14	246	td_verror(td, r, "get_events");
f3e11d05	247
ee56ad50 JA	248	dprint(FD_IO, "getevents: %d\n", r);
ee56ad50 JA	249	return r;
10ba535a JA	250	}
	251
	252	int td_io_queue(struct thread_data td, struct io_u io_u)
	253	{
a9da8ab2 JA	254	const enum fio_ddir ddir = acct_ddir(io_u);
a9da8ab2 JA	255	unsigned long buflen = io_u->xfer_buflen;
7e77dd02 JA	256	int ret;
7e77dd02 JA	257
ee56ad50	258	dprint_io_u(io_u, "queue");
7101d9c2 JA	259	fio_ro_check(td, io_u);
7101d9c2 JA	260
0c6e7517	261	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
a9da8ab2	262	io_u_set(io_u, IO_U_F_FLIGHT);
0c6e7517	263
d6aed795	264	assert(fio_file_open(io_u->file));
3d7b485f	265
bcd5abfa JA	266	/*
	267	* If using a write iolog, store this entry.
	268	*/
	269	log_io_u(td, io_u);
	270
11786802 JA	271	io_u->error = 0;
	272	io_u->resid = 0;
	273
433afcb4	274	if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841	275	if (fio_fill_issue_time(td))
9520ebb9	276	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	277
	278	/*
	279	* only used for iolog
	280	*/
	281	if (td->o.read_iolog_file)
	282	memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa	283	sizeof(struct timeval));
433afcb4 JA	284	}
433afcb4 JA	285
a9da8ab2 JA	286	if (ddir_rw(ddir)) {
	287	td->io_issues[ddir]++;
	288	td->io_issue_bytes[ddir] += buflen;
50a8ce86	289	td->rate_io_issue_bytes[ddir] += buflen;
74d6277f	290	}
755200a3	291
7e77dd02	292	ret = td->io_ops->queue(td, io_u);
5aeb77df	293
4d4e80f2	294	unlock_file(td, io_u->file);
b2bd2bd9	295
a9da8ab2 JA	296	if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
	297	td->io_issues[ddir]--;
	298	td->io_issue_bytes[ddir] -= buflen;
50a8ce86	299	td->rate_io_issue_bytes[ddir] -= buflen;
cd8a19e6 JA	300	}
cd8a19e6 JA	301
39a43d34 JA	302	/*
	303	* If an error was seen and the io engine didn't propagate it
	304	* back to 'td', do so.
	305	*/
	306	if (io_u->error && !td->error)
	307	td_verror(td, io_u->error, "td_io_queue");
	308
cb211682 JA	309	/*
	310	* Add warning for O_DIRECT so that users have an easier time
	311	* spotting potentially bad alignment. If this triggers for the first
	312	* IO, then it's likely an alignment problem or because the host fs
	313	* does not support O_DIRECT
	314	*/
ff58fced	315	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682	316	td->o.odirect) {
214ac7e0	317
cb211682	318	log_info("fio: first direct IO errored. File system may not "
44cc9c8f JA	319	"support direct IO, or iomem_align= is bad. Try "
44cc9c8f JA	320	"setting direct=0.\n");
cb211682 JA	321	}
cb211682 JA	322
77b451cc	323	if (!td->io_ops->commit \|\| io_u->ddir == DDIR_TRIM) {
838bc709 JA	324	io_u_mark_submit(td, 1);
	325	io_u_mark_complete(td, 1);
	326	}
	327
d8005759	328	if (ret == FIO_Q_COMPLETED) {
ff58fced	329	if (ddir_rw(io_u->ddir)) {
d8005759 JA	330	io_u_mark_depth(td, 1);
d8005759 JA	331	td->ts.total_io_u[io_u->ddir]++;
6eaf09d6	332	}
d8005759	333	} else if (ret == FIO_Q_QUEUED) {
eb7c8ae2 JA	334	int r;
eb7c8ae2 JA	335
e6727cbd JA	336	td->io_u_queued++;
	337
	338	if (ddir_rw(io_u->ddir))
d8005759	339	td->ts.total_io_u[io_u->ddir]++;
d8005759 JA	340
d8005759 JA	341	if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2 JA	342	r = td_io_commit(td);
	343	if (r < 0)
	344	return r;
	345	}
	346	}
cb5ab512	347
433afcb4	348	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841	349	if (fio_fill_issue_time(td))
9520ebb9	350	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	351
	352	/*
	353	* only used for iolog
	354	*/
	355	if (td->o.read_iolog_file)
	356	memcpy(&td->last_issue, &io_u->issue_time,
	357	sizeof(struct timeval));
433afcb4 JA	358	}
433afcb4 JA	359
7e77dd02	360	return ret;
10ba535a	361	}
8c16d840 JA	362
	363	int td_io_init(struct thread_data *td)
	364	{
eeb12160	365	int ret = 0;
8c16d840	366
eeb12160 JA	367	if (td->io_ops->init) {
eeb12160 JA	368	ret = td->io_ops->init(td);
5ec10eaa JA	369	if (ret && td->o.iodepth > 1) {
	370	log_err("fio: io engine init failed. Perhaps try"
	371	" reducing io depth?\n");
	372	}
7c973896 JA	373	if (!td->error)
7c973896 JA	374	td->error = ret;
eeb12160 JA	375	}
eeb12160 JA	376
046395d7 JA	377	if (!ret && (td->io_ops->flags & FIO_NOIO))
	378	td->flags \|= TD_F_NOIO;
	379
eeb12160	380	return ret;
8c16d840	381	}
755200a3 JA	382
	383	int td_io_commit(struct thread_data *td)
	384	{
f3e11d05 JA	385	int ret;
f3e11d05 JA	386
ee56ad50 JA	387	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
ee56ad50 JA	388
d8005759	389	if (!td->cur_depth \|\| !td->io_u_queued)
e1161c32	390	return 0;
cb5ab512	391
3fd9efbc	392	io_u_mark_depth(td, td->io_u_queued);
d8005759	393
f3e11d05 JA	394	if (td->io_ops->commit) {
	395	ret = td->io_ops->commit(td);
	396	if (ret)
	397	td_verror(td, -ret, "io commit");
	398	}
3fd9efbc	399
422f9e4b RM	400	/*
	401	* Reflect that events were submitted as async IO requests.
	402	*/
	403	td->io_u_in_flight += td->io_u_queued;
	404	td->io_u_queued = 0;
755200a3 JA	405
	406	return 0;
	407	}
b5af8293 JA	408
	409	int td_io_open_file(struct thread_data td, struct fio_file f)
	410	{
22a57ba8 JA	411	assert(!fio_file_open(f));
	412	assert(f->fd == -1);
	413
413d6693 JA	414	if (td->io_ops->open_file(td, f)) {
	415	if (td->error == EINVAL && td->o.odirect)
	416	log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa JA	417	if (td->error == EMFILE) {
	418	log_err("fio: try reducing/setting openfiles (failed"
	419	" at %u of %u)\n", td->nr_open_files,
	420	td->o.nr_files);
	421	}
413d6693	422
22a57ba8 JA	423	assert(f->fd == -1);
22a57ba8 JA	424	assert(!fio_file_open(f));
413d6693 JA	425	return 1;
	426	}
	427
33c48814	428	fio_file_reset(td, f);
d6aed795 JA	429	fio_file_set_open(f);
d6aed795 JA	430	fio_file_clear_closing(f);
c97bd0fa	431	disk_util_inc(f->du);
d5707a35 JA	432
	433	td->nr_open_files++;
	434	get_file(f);
	435
66159828 JA	436	if (f->filetype == FIO_TYPE_PIPE) {
	437	if (td_random(td)) {
	438	log_err("fio: can't seek on pipes (no random io)\n");
	439	goto err;
	440	}
	441	}
	442
413d6693 JA	443	if (td->io_ops->flags & FIO_DISKLESSIO)
	444	goto done;
	445
	446	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	447	goto err;
	448
66159828 JA	449	if (td->o.fadvise_hint &&
66159828 JA	450	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
413d6693 JA	451	int flags;
	452
	453	if (td_random(td))
	454	flags = POSIX_FADV_RANDOM;
	455	else
	456	flags = POSIX_FADV_SEQUENTIAL;
	457
ecc314ba	458	if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693 JA	459	td_verror(td, errno, "fadvise");
	460	goto err;
	461	}
7bb48f84	462	}
37659335 JA	463	#ifdef FIO_HAVE_STREAMID
	464	if (td->o.fadvise_stream &&
	465	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
	466	off_t stream = td->o.fadvise_stream;
	467
	468	if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
	469	td_verror(td, errno, "fadvise streamid");
	470	goto err;
	471	}
	472	}
	473	#endif
a978ba68	474
e116f2b9 JA	475	#ifdef FIO_OS_DIRECTIO
	476	/*
	477	* Some OS's have a distinct call to mark the file non-buffered,
	478	* instead of using O_DIRECT (Solaris)
	479	*/
	480	if (td->o.odirect) {
	481	int ret = fio_set_odirect(f->fd);
	482
	483	if (ret) {
	484	td_verror(td, ret, "fio_set_odirect");
78e51c72	485	log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9 JA	486	goto err;
	487	}
	488	}
	489	#endif
	490
413d6693	491	done:
f29b25a3	492	log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693 JA	493	return 0;
413d6693 JA	494	err:
c97bd0fa	495	disk_util_dec(f->du);
b284075a JA	496	if (td->io_ops->close_file)
b284075a JA	497	td->io_ops->close_file(td, f);
7bb48f84	498	return 1;
b5af8293 JA	499	}
b5af8293 JA	500
6977bcd0	501	int td_io_close_file(struct thread_data td, struct fio_file f)
b5af8293	502	{
d6aed795	503	if (!fio_file_closing(f))
f29b25a3 JA	504	log_file(td, f, FIO_LOG_CLOSE_FILE);
f29b25a3 JA	505
0ad920e7 JA	506	/*
	507	* mark as closing, do real close when last io on it has completed
	508	*/
d6aed795	509	fio_file_set_closing(f);
0ad920e7	510
c97bd0fa	511	disk_util_dec(f->du);
cba5243b JA	512
	513	if (td->o.file_lock_mode != FILE_LOCK_NONE)
	514	unlock_file_all(td, f);
29c1349f	515
6977bcd0	516	return put_file(td, f);
b5af8293	517	}
df9c26b1	518
38ef9c90 CF	519	int td_io_unlink_file(struct thread_data td, struct fio_file f)
	520	{
	521	if (td->io_ops->unlink_file)
	522	return td->io_ops->unlink_file(td, f);
	523	else
	524	return unlink(f->file_name);
	525	}
	526
df9c26b1 JA	527	int td_io_get_file_size(struct thread_data td, struct fio_file f)
	528	{
	529	if (!td->io_ops->get_file_size)
	530	return 0;
	531
	532	return td->io_ops->get_file_size(td, f);
	533	}
44f29692	534
effd6ff0 JA	535	static int do_sync_file_range(const struct thread_data *td,
effd6ff0 JA	536	struct fio_file *f)
44f29692 JA	537	{
	538	off64_t offset, nbytes;
	539
	540	offset = f->first_write;
	541	nbytes = f->last_write - f->first_write;
	542
3843deb3 JA	543	if (!nbytes)
3843deb3 JA	544	return 0;
44f29692	545
3843deb3	546	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692	547	}
0a28ecda	548
effd6ff0	549	int do_io_u_sync(const struct thread_data td, struct io_u io_u)
0a28ecda JA	550	{
	551	int ret;
	552
	553	if (io_u->ddir == DDIR_SYNC) {
	554	ret = fsync(io_u->file->fd);
	555	} else if (io_u->ddir == DDIR_DATASYNC) {
67bf9823	556	#ifdef CONFIG_FDATASYNC
0a28ecda JA	557	ret = fdatasync(io_u->file->fd);
	558	#else
	559	ret = io_u->xfer_buflen;
	560	io_u->error = EINVAL;
	561	#endif
	562	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
	563	ret = do_sync_file_range(td, io_u->file);
	564	else {
	565	ret = io_u->xfer_buflen;
	566	io_u->error = EINVAL;
	567	}
	568
cb849a79 JA	569	if (ret < 0)
	570	io_u->error = errno;
	571
0a28ecda JA	572	return ret;
0a28ecda JA	573	}
a5f3027c	574
effd6ff0	575	int do_io_u_trim(const struct thread_data td, struct io_u io_u)
a5f3027c JA	576	{
	577	#ifndef FIO_HAVE_TRIM
	578	io_u->error = EINVAL;
ff58fced	579	return 0;
a5f3027c JA	580	#else
	581	struct fio_file *f = io_u->file;
	582	int ret;
	583
c6404a44	584	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c	585	if (!ret)
3fd9efbc	586	return io_u->xfer_buflen;
a5f3027c	587
ff58fced JA	588	io_u->error = ret;
ff58fced JA	589	return 0;
a5f3027c JA	590	#endif
a5f3027c JA	591	}
de890a1e SL	592
	593	int fio_show_ioengine_help(const char *engine)
	594	{
	595	struct flist_head *entry;
	596	struct thread_data td;
	597	char *sep;
	598	int ret = 1;
	599
	600	if (!engine \|\| !*engine) {
	601	log_info("Available IO engines:\n");
	602	flist_for_each(entry, &engine_list) {
	603	td.io_ops = flist_entry(entry, struct ioengine_ops,
	604	list);
	605	log_info("\t%s\n", td.io_ops->name);
	606	}
	607	return 0;
	608	}
	609	sep = strchr(engine, ',');
	610	if (sep) {
	611	*sep = 0;
	612	sep++;
	613	}
	614
	615	memset(&td, 0, sizeof(td));
	616
	617	td.io_ops = load_ioengine(&td, engine);
	618	if (!td.io_ops) {
	619	log_info("IO engine %s not found\n", engine);
	620	return 1;
	621	}
	622
	623	if (td.io_ops->options)
	624	ret = show_cmd_help(td.io_ops->options, sep);
	625	else
	626	log_info("IO engine %s has no options\n", td.io_ops->name);
	627
	628	free_ioengine(&td);
	629
	630	return ret;
	631	}