[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 LIST_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);
	list_del(&ops->list);
	INIT_LIST_HEAD(&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);
}

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

	list_for_each(entry, &engine_list) {
		ops = list_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);

	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);

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

		dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
		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;
	}

	r = 0;
	if (td->io_ops->getevents)
		r = td->io_ops->getevents(td, min, max, t);
out:
	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);
		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);

	if (ret != FIO_Q_BUSY)
		io_u_mark_depth(td, io_u);

	if (ret == FIO_Q_QUEUED) {
		int r;

		td->io_u_queued++;
		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);
		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)
		return 0;

	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;
	}

	if (f->filetype == FIO_TYPE_PIPE) {
		if (td_random(td)) {
			log_err("fio: can't seek on pipes (no random io)\n");
			goto err;
		}
	}

	f->last_free_lookup = 0;
	f->last_completed_pos = 0;
	f->last_pos = f->file_offset;
	f->flags |= FIO_FILE_OPEN;
	f->flags &= ~FIO_FILE_CLOSING;

	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(long));

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

void 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;

	put_file(td, f);
}
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>
0c6e7517	17	#include <assert.h>
8c16d840	18
ebac4655	19	#include "fio.h"
ebac4655	20
5f350952 JA	21	static LIST_HEAD(engine_list);
5f350952 JA	22
8c16d840 JA	23	static int check_engine_ops(struct ioengine_ops *ops)
8c16d840 JA	24	{
5f350952 JA	25	if (ops->version != FIO_IOOPS_VERSION) {
	26	log_err("bad ioops version %d (want %d)\n", ops->version, FIO_IOOPS_VERSION);
	27	return 1;
	28	}
	29
36167d82 JA	30	if (!ops->queue) {
	31	log_err("%s: no queue handler\n", ops->name);
	32	return 1;
	33	}
	34
	35	/*
	36	* sync engines only need a ->queue()
	37	*/
	38	if (ops->flags & FIO_SYNCIO)
	39	return 0;
	40
8c16d840	41	if (!ops->event) {
36167d82	42	log_err("%s: no event handler\n", ops->name);
8c16d840 JA	43	return 1;
	44	}
	45	if (!ops->getevents) {
36167d82	46	log_err("%s: no getevents handler\n", ops->name);
8c16d840 JA	47	return 1;
	48	}
	49	if (!ops->queue) {
36167d82	50	log_err("%s: no queue handler\n", ops->name);
8c16d840 JA	51	return 1;
	52	}
	53
	54	return 0;
	55	}
	56
5f350952	57	void unregister_ioengine(struct ioengine_ops *ops)
ebac4655	58	{
ee56ad50	59	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
5f350952 JA	60	list_del(&ops->list);
	61	INIT_LIST_HEAD(&ops->list);
	62	}
	63
b2fdda43	64	void register_ioengine(struct ioengine_ops *ops)
5f350952	65	{
ee56ad50	66	dprint(FD_IO, "ioengine %s registered\n", ops->name);
5f350952 JA	67	INIT_LIST_HEAD(&ops->list);
5f350952 JA	68	list_add_tail(&ops->list, &engine_list);
5f350952 JA	69	}
	70
	71	static struct ioengine_ops find_ioengine(const char name)
	72	{
	73	struct ioengine_ops *ops;
	74	struct list_head *entry;
ebac4655	75
5f350952 JA	76	list_for_each(entry, &engine_list) {
5f350952 JA	77	ops = list_entry(entry, struct ioengine_ops, list);
bc5b77a8	78	if (!strcmp(name, ops->name))
5f350952 JA	79	return ops;
	80	}
	81
	82	return NULL;
	83	}
	84
	85	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	86	const char *engine_lib)
	87	{
	88	struct ioengine_ops *ops;
	89	void *dlhandle;
	90
ee56ad50 JA	91	dprint(FD_IO, "dload engine %s\n", engine_lib);
ee56ad50 JA	92
2866c82d JA	93	dlerror();
2866c82d JA	94	dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8	95	if (!dlhandle) {
e1161c32	96	td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8 JA	97	return NULL;
d4dbaaa8 JA	98	}
8756e4d4	99
da51c050 JA	100	/*
	101	* Unlike the included modules, external engines should have a
	102	* non-static ioengine structure that we can reference.
	103	*/
2866c82d	104	ops = dlsym(dlhandle, "ioengine");
d4dbaaa8	105	if (!ops) {
e1161c32	106	td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8 JA	107	dlclose(dlhandle);
	108	return NULL;
	109	}
8756e4d4	110
5f350952 JA	111	ops->dlhandle = dlhandle;
	112	return ops;
	113	}
	114
	115	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	116	{
	117	struct ioengine_ops ops, ret;
	118	char engine[16];
	119
ee56ad50 JA	120	dprint(FD_IO, "load ioengine %s\n", name);
ee56ad50 JA	121
5f350952 JA	122	strncpy(engine, name, sizeof(engine) - 1);
	123
	124	/*
	125	* linux libaio has alias names, so convert to what we want
	126	*/
	127	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	128	strcpy(engine, "libaio");
	129
	130	ops = find_ioengine(engine);
	131	if (!ops)
	132	ops = dlopen_ioengine(td, name);
	133
	134	if (!ops) {
	135	log_err("fio: engine %s not loadable\n", name);
b902ceb5 JA	136	return NULL;
	137	}
	138
8c16d840 JA	139	/*
	140	* Check that the required methods are there.
	141	*/
5f350952	142	if (check_engine_ops(ops))
8c16d840	143	return NULL;
8c16d840	144
84585003 JA	145	ret = malloc(sizeof(*ret));
	146	memcpy(ret, ops, sizeof(*ret));
	147	ret->data = NULL;
84585003 JA	148
84585003 JA	149	return ret;
8756e4d4 JA	150	}
8756e4d4 JA	151
2866c82d	152	void close_ioengine(struct thread_data *td)
8756e4d4	153	{
ee56ad50 JA	154	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
ee56ad50 JA	155
2866c82d JA	156	if (td->io_ops->cleanup)
2866c82d JA	157	td->io_ops->cleanup(td);
b990b5c0	158
5f350952 JA	159	if (td->io_ops->dlhandle)
	160	dlclose(td->io_ops->dlhandle);
	161
84585003 JA	162	free(td->io_ops);
84585003 JA	163	td->io_ops = NULL;
b990b5c0	164	}
10ba535a JA	165
	166	int td_io_prep(struct thread_data td, struct io_u io_u)
	167	{
ee56ad50	168	dprint_io_u(io_u, "prep");
7101d9c2 JA	169	fio_ro_check(td, io_u);
7101d9c2 JA	170
2ba1c290 JA	171	if (td->io_ops->prep) {
	172	int ret = td->io_ops->prep(td, io_u);
	173
	174	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
	175	return ret;
	176	}
10ba535a JA	177
	178	return 0;
	179	}
	180
e7d2e616	181	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
10ba535a JA	182	struct timespec *t)
10ba535a JA	183	{
ee56ad50	184	int r = 0;
face81b2	185
ee56ad50 JA	186	if (min > 0 && td->io_ops->commit) {
ee56ad50 JA	187	r = td->io_ops->commit(td);
face81b2	188	if (r < 0)
ee56ad50	189	goto out;
face81b2	190	}
36167d82	191
ee56ad50 JA	192	r = 0;
	193	if (td->io_ops->getevents)
	194	r = td->io_ops->getevents(td, min, max, t);
	195	out:
	196	dprint(FD_IO, "getevents: %d\n", r);
	197	return r;
10ba535a JA	198	}
	199
	200	int td_io_queue(struct thread_data td, struct io_u io_u)
	201	{
7e77dd02 JA	202	int ret;
7e77dd02 JA	203
ee56ad50	204	dprint_io_u(io_u, "queue");
7101d9c2 JA	205	fio_ro_check(td, io_u);
7101d9c2 JA	206
0c6e7517 JA	207	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
	208	io_u->flags \|= IO_U_F_FLIGHT;
	209
3d7b485f JA	210	assert(io_u->file->flags & FIO_FILE_OPEN);
3d7b485f JA	211
11786802 JA	212	io_u->error = 0;
	213	io_u->resid = 0;
	214
433afcb4	215	if (td->io_ops->flags & FIO_SYNCIO) {
5aeb77df	216	fio_gettime(&io_u->issue_time, NULL);
a61eddec	217	memcpy(&td->last_issue, &io_u->issue_time, sizeof(struct timeval));
10ba535a	218
433afcb4 JA	219	/*
	220	* for a sync engine, set the timeout upfront
	221	*/
	222	if (mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
	223	io_u_set_timeout(td);
	224	}
	225
755200a3 JA	226	if (io_u->ddir != DDIR_SYNC)
	227	td->io_issues[io_u->ddir]++;
	228
7e77dd02	229	ret = td->io_ops->queue(td, io_u);
5aeb77df	230
163f849e JA	231	if (ret != FIO_Q_BUSY)
	232	io_u_mark_depth(td, io_u);
	233
eb7c8ae2 JA	234	if (ret == FIO_Q_QUEUED) {
	235	int r;
	236
cb5ab512	237	td->io_u_queued++;
2dc1bbeb	238	if (td->io_u_queued > td->o.iodepth_batch) {
eb7c8ae2 JA	239	r = td_io_commit(td);
	240	if (r < 0)
	241	return r;
	242	}
	243	}
cb5ab512	244
433afcb4	245	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
5aeb77df	246	fio_gettime(&io_u->issue_time, NULL);
a61eddec	247	memcpy(&td->last_issue, &io_u->issue_time, sizeof(struct timeval));
5aeb77df	248
433afcb4 JA	249	/*
	250	* async engine, set the timeout here
	251	*/
	252	if (ret == FIO_Q_QUEUED &&
	253	mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
	254	io_u_set_timeout(td);
	255	}
	256
7e77dd02	257	return ret;
10ba535a	258	}
8c16d840 JA	259
	260	int td_io_init(struct thread_data *td)
	261	{
eeb12160	262	int ret = 0;
8c16d840	263
eeb12160 JA	264	if (td->io_ops->init) {
	265	ret = td->io_ops->init(td);
	266	if (ret && td->o.iodepth > 1)
9de6bebc	267	log_err("fio: io engine init failed. Perhaps try reducing io depth?\n");
eeb12160 JA	268	}
	269
	270	return ret;
8c16d840	271	}
755200a3 JA	272
	273	int td_io_commit(struct thread_data *td)
	274	{
ee56ad50 JA	275	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
ee56ad50 JA	276
e1161c32 JA	277	if (!td->cur_depth)
e1161c32 JA	278	return 0;
cb5ab512 JA	279
cb5ab512 JA	280	td->io_u_queued = 0;
755200a3 JA	281	if (td->io_ops->commit)
	282	return td->io_ops->commit(td);
	283
	284	return 0;
	285	}
b5af8293 JA	286
	287	int td_io_open_file(struct thread_data td, struct fio_file f)
	288	{
413d6693 JA	289	if (td->io_ops->open_file(td, f)) {
	290	if (td->error == EINVAL && td->o.odirect)
	291	log_err("fio: destination does not support O_DIRECT\n");
	292	if (td->error == EMFILE)
	293	log_err("fio: try reducing/setting openfiles (failed at %u of %u)\n", td->nr_open_files, td->o.nr_files);
	294
	295	return 1;
	296	}
	297
66159828 JA	298	if (f->filetype == FIO_TYPE_PIPE) {
	299	if (td_random(td)) {
	300	log_err("fio: can't seek on pipes (no random io)\n");
	301	goto err;
	302	}
	303	}
	304
413d6693 JA	305	f->last_free_lookup = 0;
413d6693 JA	306	f->last_completed_pos = 0;
bcdedd0a	307	f->last_pos = f->file_offset;
413d6693 JA	308	f->flags \|= FIO_FILE_OPEN;
	309	f->flags &= ~FIO_FILE_CLOSING;
	310
	311	if (td->io_ops->flags & FIO_DISKLESSIO)
	312	goto done;
	313
	314	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	315	goto err;
	316
66159828 JA	317	if (td->o.fadvise_hint &&
	318	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
	319
413d6693 JA	320	int flags;
	321
	322	if (td_random(td))
	323	flags = POSIX_FADV_RANDOM;
	324	else
	325	flags = POSIX_FADV_SEQUENTIAL;
	326
	327	if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
	328	td_verror(td, errno, "fadvise");
	329	goto err;
	330	}
7bb48f84	331	}
a978ba68	332
413d6693 JA	333	if (f->file_map)
413d6693 JA	334	memset(f->file_map, 0, f->num_maps * sizeof(long));
a978ba68	335
413d6693	336	done:
f29b25a3	337	log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693 JA	338	td->nr_open_files++;
	339	get_file(f);
	340	return 0;
	341	err:
b284075a JA	342	if (td->io_ops->close_file)
b284075a JA	343	td->io_ops->close_file(td, f);
7bb48f84	344	return 1;
b5af8293 JA	345	}
	346
	347	void td_io_close_file(struct thread_data td, struct fio_file f)
	348	{
f29b25a3 JA	349	if (!(f->flags & FIO_FILE_CLOSING))
	350	log_file(td, f, FIO_LOG_CLOSE_FILE);
	351
0ad920e7 JA	352	/*
	353	* mark as closing, do real close when last io on it has completed
	354	*/
	355	f->flags \|= FIO_FILE_CLOSING;
	356
	357	put_file(td, f);
b5af8293	358	}