[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 <assert.h>

#include "fio.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, "ioengine");
	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;
}

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;
	}

	if (td->io_ops->dlhandle)
		dlclose(td->io_ops->dlhandle);

	free(td->io_ops);
	td->io_ops = NULL;
}

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,
		    struct timespec *t)
{
	int r = 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)
		io_u_mark_complete(td, r);
	dprint(FD_IO, "getevents: %d\n", r);
	return r;
}

int td_io_queue(struct thread_data *td, struct io_u *io_u)
{
	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;

	assert(io_u->file->flags & FIO_FILE_OPEN);

	io_u->error = 0;
	io_u->resid = 0;

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

		/*
		 * for a sync engine, set the timeout upfront
		 */
		if (mtime_since(&td->timeout_end, &io_u->issue_time)
		    < IO_U_TIMEOUT)
			io_u_set_timeout(td);
	}

	if (io_u->ddir != DDIR_SYNC)
		td->io_issues[io_u->ddir]++;

	ret = td->io_ops->queue(td, io_u);

	unlock_file(td, io_u->file);

	if (!td->io_ops->commit) {
		io_u_mark_submit(td, 1);
		io_u_mark_complete(td, 1);
	}

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

		if (io_u->ddir != DDIR_SYNC) {
			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) {
		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));

		/*
		 * async engine, set the timeout here
		 */
		if (ret == FIO_Q_QUEUED &&
		    (mtime_since(&td->timeout_end, &io_u->issue_time)
			< IO_U_TIMEOUT)) {
			io_u_set_timeout(td);
		}
	}

	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");
		}
	}

	return ret;
}

int td_io_commit(struct thread_data *td)
{
	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);
	td->io_u_queued = 0;

	if (td->io_ops->commit)
		return td->io_ops->commit(td);

	return 0;
}

int td_io_open_file(struct thread_data *td, struct fio_file *f)
{
	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);
		}

		return 1;
	}

	fio_file_reset(f);
	f->flags |= FIO_FILE_OPEN;
	f->flags &= ~FIO_FILE_CLOSING;

	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 (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
			td_verror(td, errno, "fadvise");
			goto err;
		}
	}

	if (f->file_map)
		memset(f->file_map, 0, f->num_maps * sizeof(int));

#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");
			goto err;
		}
	}
#endif

done:
	log_file(td, f, FIO_LOG_OPEN_FILE);
	return 0;
err:
	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 (!(f->flags & FIO_FILE_CLOSING))
		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);

	return put_file(td, f);
}
Commit	Line	Data
	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>
	15	#include <string.h>
	16	#include <dlfcn.h>
	17	#include <assert.h>
	18
	19	#include "fio.h"
	20
	21	static FLIST_HEAD(engine_list);
	22
	23	static int check_engine_ops(struct ioengine_ops *ops)
	24	{
	25	if (ops->version != FIO_IOOPS_VERSION) {
	26	log_err("bad ioops version %d (want %d)\n", ops->version,
	27	FIO_IOOPS_VERSION);
	28	return 1;
	29	}
	30
	31	if (!ops->queue) {
	32	log_err("%s: no queue handler\n", ops->name);
	33	return 1;
	34	}
	35
	36	/*
	37	* sync engines only need a ->queue()
	38	*/
	39	if (ops->flags & FIO_SYNCIO)
	40	return 0;
	41
	42	if (!ops->event) {
	43	log_err("%s: no event handler\n", ops->name);
	44	return 1;
	45	}
	46	if (!ops->getevents) {
	47	log_err("%s: no getevents handler\n", ops->name);
	48	return 1;
	49	}
	50	if (!ops->queue) {
	51	log_err("%s: no queue handler\n", ops->name);
	52	return 1;
	53	}
	54
	55	return 0;
	56	}
	57
	58	void unregister_ioengine(struct ioengine_ops *ops)
	59	{
	60	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
	61	flist_del(&ops->list);
	62	INIT_FLIST_HEAD(&ops->list);
	63	}
	64
	65	void register_ioengine(struct ioengine_ops *ops)
	66	{
	67	dprint(FD_IO, "ioengine %s registered\n", ops->name);
	68	INIT_FLIST_HEAD(&ops->list);
	69	flist_add_tail(&ops->list, &engine_list);
	70	}
	71
	72	static struct ioengine_ops find_ioengine(const char name)
	73	{
	74	struct ioengine_ops *ops;
	75	struct flist_head *entry;
	76
	77	flist_for_each(entry, &engine_list) {
	78	ops = flist_entry(entry, struct ioengine_ops, list);
	79	if (!strcmp(name, ops->name))
	80	return ops;
	81	}
	82
	83	return NULL;
	84	}
	85
	86	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	87	const char *engine_lib)
	88	{
	89	struct ioengine_ops *ops;
	90	void *dlhandle;
	91
	92	dprint(FD_IO, "dload engine %s\n", engine_lib);
	93
	94	dlerror();
	95	dlhandle = dlopen(engine_lib, RTLD_LAZY);
	96	if (!dlhandle) {
	97	td_vmsg(td, -1, dlerror(), "dlopen");
	98	return NULL;
	99	}
	100
	101	/*
	102	* Unlike the included modules, external engines should have a
	103	* non-static ioengine structure that we can reference.
	104	*/
	105	ops = dlsym(dlhandle, "ioengine");
	106	if (!ops) {
	107	td_vmsg(td, -1, dlerror(), "dlsym");
	108	dlclose(dlhandle);
	109	return NULL;
	110	}
	111
	112	ops->dlhandle = dlhandle;
	113	return ops;
	114	}
	115
	116	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	117	{
	118	struct ioengine_ops ops, ret;
	119	char engine[16];
	120
	121	dprint(FD_IO, "load ioengine %s\n", name);
	122
	123	strncpy(engine, name, sizeof(engine) - 1);
	124
	125	/*
	126	* linux libaio has alias names, so convert to what we want
	127	*/
	128	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	129	strcpy(engine, "libaio");
	130
	131	ops = find_ioengine(engine);
	132	if (!ops)
	133	ops = dlopen_ioengine(td, name);
	134
	135	if (!ops) {
	136	log_err("fio: engine %s not loadable\n", name);
	137	return NULL;
	138	}
	139
	140	/*
	141	* Check that the required methods are there.
	142	*/
	143	if (check_engine_ops(ops))
	144	return NULL;
	145
	146	ret = malloc(sizeof(*ret));
	147	memcpy(ret, ops, sizeof(*ret));
	148	ret->data = NULL;
	149
	150	return ret;
	151	}
	152
	153	void close_ioengine(struct thread_data *td)
	154	{
	155	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	156
	157	if (td->io_ops->cleanup) {
	158	td->io_ops->cleanup(td);
	159	td->io_ops->data = NULL;
	160	}
	161
	162	if (td->io_ops->dlhandle)
	163	dlclose(td->io_ops->dlhandle);
	164
	165	free(td->io_ops);
	166	td->io_ops = NULL;
	167	}
	168
	169	int td_io_prep(struct thread_data td, struct io_u io_u)
	170	{
	171	dprint_io_u(io_u, "prep");
	172	fio_ro_check(td, io_u);
	173
	174	lock_file(td, io_u->file, io_u->ddir);
	175
	176	if (td->io_ops->prep) {
	177	int ret = td->io_ops->prep(td, io_u);
	178
	179	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
	180	if (ret)
	181	unlock_file(td, io_u->file);
	182	return ret;
	183	}
	184
	185	return 0;
	186	}
	187
	188	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
	189	struct timespec *t)
	190	{
	191	int r = 0;
	192
	193	if (min > 0 && td->io_ops->commit) {
	194	r = td->io_ops->commit(td);
	195	if (r < 0)
	196	goto out;
	197	}
	198	if (max > td->cur_depth)
	199	max = td->cur_depth;
	200	if (min > max)
	201	max = min;
	202
	203	r = 0;
	204	if (max && td->io_ops->getevents)
	205	r = td->io_ops->getevents(td, min, max, t);
	206	out:
	207	if (r >= 0)
	208	io_u_mark_complete(td, r);
	209	dprint(FD_IO, "getevents: %d\n", r);
	210	return r;
	211	}
	212
	213	int td_io_queue(struct thread_data td, struct io_u io_u)
	214	{
	215	int ret;
	216
	217	dprint_io_u(io_u, "queue");
	218	fio_ro_check(td, io_u);
	219
	220	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
	221	io_u->flags \|= IO_U_F_FLIGHT;
	222
	223	assert(io_u->file->flags & FIO_FILE_OPEN);
	224
	225	io_u->error = 0;
	226	io_u->resid = 0;
	227
	228	if (td->io_ops->flags & FIO_SYNCIO) {
	229	fio_gettime(&io_u->issue_time, NULL);
	230
	231	/*
	232	* only used for iolog
	233	*/
	234	if (td->o.read_iolog_file)
	235	memcpy(&td->last_issue, &io_u->issue_time,
	236	sizeof(struct timeval));
	237
	238	/*
	239	* for a sync engine, set the timeout upfront
	240	*/
	241	if (mtime_since(&td->timeout_end, &io_u->issue_time)
	242	< IO_U_TIMEOUT)
	243	io_u_set_timeout(td);
	244	}
	245
	246	if (io_u->ddir != DDIR_SYNC)
	247	td->io_issues[io_u->ddir]++;
	248
	249	ret = td->io_ops->queue(td, io_u);
	250
	251	unlock_file(td, io_u->file);
	252
	253	if (!td->io_ops->commit) {
	254	io_u_mark_submit(td, 1);
	255	io_u_mark_complete(td, 1);
	256	}
	257
	258	if (ret == FIO_Q_COMPLETED) {
	259	if (io_u->ddir != DDIR_SYNC) {
	260	io_u_mark_depth(td, 1);
	261	td->ts.total_io_u[io_u->ddir]++;
	262	}
	263	} else if (ret == FIO_Q_QUEUED) {
	264	int r;
	265
	266	if (io_u->ddir != DDIR_SYNC) {
	267	td->io_u_queued++;
	268	td->ts.total_io_u[io_u->ddir]++;
	269	}
	270
	271	if (td->io_u_queued >= td->o.iodepth_batch) {
	272	r = td_io_commit(td);
	273	if (r < 0)
	274	return r;
	275	}
	276	}
	277
	278	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
	279	fio_gettime(&io_u->issue_time, NULL);
	280
	281	/*
	282	* only used for iolog
	283	*/
	284	if (td->o.read_iolog_file)
	285	memcpy(&td->last_issue, &io_u->issue_time,
	286	sizeof(struct timeval));
	287
	288	/*
	289	* async engine, set the timeout here
	290	*/
	291	if (ret == FIO_Q_QUEUED &&
	292	(mtime_since(&td->timeout_end, &io_u->issue_time)
	293	< IO_U_TIMEOUT)) {
	294	io_u_set_timeout(td);
	295	}
	296	}
	297
	298	return ret;
	299	}
	300
	301	int td_io_init(struct thread_data *td)
	302	{
	303	int ret = 0;
	304
	305	if (td->io_ops->init) {
	306	ret = td->io_ops->init(td);
	307	if (ret && td->o.iodepth > 1) {
	308	log_err("fio: io engine init failed. Perhaps try"
	309	" reducing io depth?\n");
	310	}
	311	}
	312
	313	return ret;
	314	}
	315
	316	int td_io_commit(struct thread_data *td)
	317	{
	318	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
	319
	320	if (!td->cur_depth \|\| !td->io_u_queued)
	321	return 0;
	322
	323	io_u_mark_depth(td, td->io_u_queued);
	324	td->io_u_queued = 0;
	325
	326	if (td->io_ops->commit)
	327	return td->io_ops->commit(td);
	328
	329	return 0;
	330	}
	331
	332	int td_io_open_file(struct thread_data td, struct fio_file f)
	333	{
	334	if (td->io_ops->open_file(td, f)) {
	335	if (td->error == EINVAL && td->o.odirect)
	336	log_err("fio: destination does not support O_DIRECT\n");
	337	if (td->error == EMFILE) {
	338	log_err("fio: try reducing/setting openfiles (failed"
	339	" at %u of %u)\n", td->nr_open_files,
	340	td->o.nr_files);
	341	}
	342
	343	return 1;
	344	}
	345
	346	fio_file_reset(f);
	347	f->flags \|= FIO_FILE_OPEN;
	348	f->flags &= ~FIO_FILE_CLOSING;
	349
	350	td->nr_open_files++;
	351	get_file(f);
	352
	353	if (f->filetype == FIO_TYPE_PIPE) {
	354	if (td_random(td)) {
	355	log_err("fio: can't seek on pipes (no random io)\n");
	356	goto err;
	357	}
	358	}
	359
	360	if (td->io_ops->flags & FIO_DISKLESSIO)
	361	goto done;
	362
	363	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	364	goto err;
	365
	366	if (td->o.fadvise_hint &&
	367	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
	368	int flags;
	369
	370	if (td_random(td))
	371	flags = POSIX_FADV_RANDOM;
	372	else
	373	flags = POSIX_FADV_SEQUENTIAL;
	374
	375	if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
	376	td_verror(td, errno, "fadvise");
	377	goto err;
	378	}
	379	}
	380
	381	if (f->file_map)
	382	memset(f->file_map, 0, f->num_maps * sizeof(int));
	383
	384	#ifdef FIO_OS_DIRECTIO
	385	/*
	386	* Some OS's have a distinct call to mark the file non-buffered,
	387	* instead of using O_DIRECT (Solaris)
	388	*/
	389	if (td->o.odirect) {
	390	int ret = fio_set_odirect(f->fd);
	391
	392	if (ret) {
	393	td_verror(td, ret, "fio_set_odirect");
	394	goto err;
	395	}
	396	}
	397	#endif
	398
	399	done:
	400	log_file(td, f, FIO_LOG_OPEN_FILE);
	401	return 0;
	402	err:
	403	if (td->io_ops->close_file)
	404	td->io_ops->close_file(td, f);
	405	return 1;
	406	}
	407
	408	int td_io_close_file(struct thread_data td, struct fio_file f)
	409	{
	410	if (!(f->flags & FIO_FILE_CLOSING))
	411	log_file(td, f, FIO_LOG_CLOSE_FILE);
	412
	413	/*
	414	* mark as closing, do real close when last io on it has completed
	415	*/
	416	f->flags \|= FIO_FILE_CLOSING;
	417
	418	unlock_file_all(td, f);
	419
	420	return put_file(td, f);
	421	}