[fio.git] / engines / syslet-rw.c

/*
 * syslet engine
 *
 * IO engine that does regular pread(2)/pwrite(2) to transfer data, but
 * with syslets to make the execution async.
 *
 */
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <assert.h>

#include "../fio.h"
#include "../os.h"

#ifdef FIO_HAVE_SYSLET

struct syslet_data {
	struct io_u **events;
	unsigned int nr_events;
	
	struct async_head_user ahu;
	struct syslet_uatom **ring;

	struct syslet_uatom *head, *tail;
};

static void fio_syslet_complete_atom(struct thread_data *td,
				     struct syslet_uatom *atom)
{
	struct syslet_data *sd = td->io_ops->data;
	struct syslet_uatom *last;
	struct io_u *io_u;

	/*
	 * complete from the beginning of the sequence up to (and
	 * including) this atom
	 */
	last = atom;
	io_u = atom->private;
	atom = io_u->req.head;

	/*
	 * now complete in right order
	 */
	do {
		long ret;

		io_u = atom->private;
		ret = *atom->ret_ptr;
		if (ret >= 0)
			io_u->resid = io_u->xfer_buflen - ret;
		else if (ret < 0)
			io_u->error = ret;

		assert(sd->nr_events < td->iodepth);
		sd->events[sd->nr_events++] = io_u;

		if (atom == last)
			break;

		atom = atom->next;
	} while (1);

	assert(!last->next);
}

/*
 * Inspect the ring to see if we have completed events
 */
static void fio_syslet_complete(struct thread_data *td)
{
	struct syslet_data *sd = td->io_ops->data;

	do {
		struct syslet_uatom *atom;

		atom = sd->ring[sd->ahu.user_ring_idx];
		if (!atom)
			break;

		sd->ring[sd->ahu.user_ring_idx] = NULL;
		if (++sd->ahu.user_ring_idx == td->iodepth)
			sd->ahu.user_ring_idx = 0;

		fio_syslet_complete_atom(td, atom);
	} while (1);
}

static int fio_syslet_getevents(struct thread_data *td, int min,
				int fio_unused max,
				struct timespec fio_unused *t)
{
	struct syslet_data *sd = td->io_ops->data;
	long ret;

	do {
		fio_syslet_complete(td);

		/*
		 * do we have enough immediate completions?
		 */
		if (sd->nr_events >= (unsigned int) min)
			break;

		/*
		 * OK, we need to wait for some events...
		 */
		ret = async_wait(1, sd->ahu.user_ring_idx, &sd->ahu);
		if (ret < 0)
			return -errno;
	} while (1);

	ret = sd->nr_events;
	sd->nr_events = 0;
	return ret;
}

static struct io_u *fio_syslet_event(struct thread_data *td, int event)
{
	struct syslet_data *sd = td->io_ops->data;

	return sd->events[event];
}

static void init_atom(struct syslet_uatom *atom, int nr, void *arg0,
		      void *arg1, void *arg2, void *arg3, void *ret_ptr,
		      unsigned long flags, void *priv)
{
	atom->flags = flags;
	atom->nr = nr;
	atom->ret_ptr = ret_ptr;
	atom->next = NULL;
	atom->arg_ptr[0] = arg0;
	atom->arg_ptr[1] = arg1;
	atom->arg_ptr[2] = arg2;
	atom->arg_ptr[3] = arg3;
	atom->arg_ptr[4] = atom->arg_ptr[5] = NULL;
	atom->private = priv;
}

/*
 * Use seek atom for sync
 */
static void fio_syslet_prep_sync(struct io_u *io_u, struct fio_file *f)
{
	init_atom(&io_u->req.atom, __NR_fsync, &f->fd, NULL, NULL, NULL,
		  &io_u->req.ret, 0, io_u);
}

static void fio_syslet_prep_rw(struct io_u *io_u, struct fio_file *f)
{
	int nr;

	/*
	 * prepare rw
	 */
	if (io_u->ddir == DDIR_READ)
		nr = __NR_pread64;
	else
		nr = __NR_pwrite64;

	init_atom(&io_u->req.atom, nr, &f->fd, &io_u->xfer_buf,
		  &io_u->xfer_buflen, &io_u->offset, &io_u->req.ret, 0, io_u);
}

static int fio_syslet_prep(struct thread_data fio_unused *td, struct io_u *io_u)
{
	struct fio_file *f = io_u->file;

	if (io_u->ddir == DDIR_SYNC)
		fio_syslet_prep_sync(io_u, f);
	else
		fio_syslet_prep_rw(io_u, f);

	return 0;
}

static void cachemiss_thread_start(void)
{
	while (1)
		async_thread(NULL, NULL);
}

#define THREAD_STACK_SIZE (16384)

static unsigned long thread_stack_alloc()
{
	return (unsigned long) malloc(THREAD_STACK_SIZE) + THREAD_STACK_SIZE;
}

static void fio_syslet_queued(struct thread_data *td, struct syslet_data *sd)
{
	struct syslet_uatom *atom;
	struct timeval now;

	fio_gettime(&now, NULL);

	atom = sd->head;
	while (atom) {
		struct io_u *io_u = atom->private;

		memcpy(&io_u->issue_time, &now, sizeof(now));
		io_u_queued(td, io_u);
		atom = atom->next;
	}
}

static int fio_syslet_commit(struct thread_data *td)
{
	struct syslet_data *sd = td->io_ops->data;
	struct syslet_uatom *done;

	if (!sd->head)
		return 0;

	assert(!sd->tail->next);

	if (!sd->ahu.new_thread_stack)
		sd->ahu.new_thread_stack = thread_stack_alloc();

	fio_syslet_queued(td, sd);

	/*
	 * On sync completion, the atom is returned. So on NULL return
	 * it's queued asynchronously.
	 */
	done = async_exec(sd->head, &sd->ahu);

	sd->head = sd->tail = NULL;

	if (done)
		fio_syslet_complete_atom(td, done);

	return 0;
}

static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
{
	struct syslet_data *sd = td->io_ops->data;

	if (sd->tail) {
		sd->tail->next = &io_u->req.atom;
		sd->tail = &io_u->req.atom;
	} else
		sd->head = sd->tail = &io_u->req.atom;

	io_u->req.head = sd->head;
	return FIO_Q_QUEUED;
}

static int async_head_init(struct syslet_data *sd, unsigned int depth)
{
	unsigned long ring_size;

	memset(&sd->ahu, 0, sizeof(struct async_head_user));

	ring_size = sizeof(struct syslet_uatom *) * depth;
	sd->ring = malloc(ring_size);
	memset(sd->ring, 0, ring_size);

	sd->ahu.user_ring_idx = 0;
	sd->ahu.completion_ring = sd->ring;
	sd->ahu.ring_size_bytes = ring_size;
	sd->ahu.head_stack = thread_stack_alloc();
	sd->ahu.head_eip = (unsigned long) cachemiss_thread_start;
	sd->ahu.new_thread_eip = (unsigned long) cachemiss_thread_start;

	return 0;
}

static void async_head_exit(struct syslet_data *sd)
{
	free(sd->ring);
}

static void fio_syslet_cleanup(struct thread_data *td)
{
	struct syslet_data *sd = td->io_ops->data;

	if (sd) {
		async_head_exit(sd);
		free(sd->events);
		free(sd);
		td->io_ops->data = NULL;
	}
}

static int fio_syslet_init(struct thread_data *td)
{
	struct syslet_data *sd;


	sd = malloc(sizeof(*sd));
	memset(sd, 0, sizeof(*sd));
	sd->events = malloc(sizeof(struct io_u *) * td->iodepth);
	memset(sd->events, 0, sizeof(struct io_u *) * td->iodepth);

	/*
	 * This will handily fail for kernels where syslet isn't available
	 */
	if (async_head_init(sd, td->iodepth)) {
		free(sd->events);
		free(sd);
		return 1;
	}

	td->io_ops->data = sd;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name		= "syslet-rw",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_syslet_init,
	.prep		= fio_syslet_prep,
	.queue		= fio_syslet_queue,
	.commit		= fio_syslet_commit,
	.getevents	= fio_syslet_getevents,
	.event		= fio_syslet_event,
	.cleanup	= fio_syslet_cleanup,
	.open_file	= generic_open_file,
	.close_file	= generic_close_file,
};

#else /* FIO_HAVE_SYSLET */

/*
 * When we have a proper configure system in place, we simply wont build
 * and install this io engine. For now install a crippled version that
 * just complains and fails to load.
 */
static int fio_syslet_init(struct thread_data fio_unused *td)
{
	fprintf(stderr, "fio: syslet not available\n");
	return 1;
}

static struct ioengine_ops ioengine = {
	.name		= "syslet-rw",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_syslet_init,
};

#endif /* FIO_HAVE_SYSLET */

static void fio_init fio_syslet_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_syslet_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
a4f4fdd7	1	/*
da751ca9 JA	2	* syslet engine
	3	*
	4	* IO engine that does regular pread(2)/pwrite(2) to transfer data, but
	5	* with syslets to make the execution async.
a4f4fdd7 JA	6	*
	7	*/
	8	#include <stdio.h>
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
	12	#include <assert.h>
	13
	14	#include "../fio.h"
	15	#include "../os.h"
	16
	17	#ifdef FIO_HAVE_SYSLET
	18
	19	struct syslet_data {
	20	struct io_u **events;
	21	unsigned int nr_events;
	22
bf0dc8fa	23	struct async_head_user ahu;
a4f4fdd7	24	struct syslet_uatom **ring;
9ff9de69 JA	25
9ff9de69 JA	26	struct syslet_uatom head, tail;
a4f4fdd7 JA	27	};
a4f4fdd7 JA	28
9ff9de69 JA	29	static void fio_syslet_complete_atom(struct thread_data *td,
	30	struct syslet_uatom *atom)
	31	{
	32	struct syslet_data *sd = td->io_ops->data;
5b38ee84	33	struct syslet_uatom *last;
9ff9de69	34	struct io_u *io_u;
9ff9de69 JA	35
9ff9de69 JA	36	/*
5b38ee84 JA	37	* complete from the beginning of the sequence up to (and
5b38ee84 JA	38	* including) this atom
9ff9de69	39	*/
5b38ee84 JA	40	last = atom;
	41	io_u = atom->private;
	42	atom = io_u->req.head;
9ff9de69 JA	43
	44	/*
	45	* now complete in right order
	46	*/
5b38ee84	47	do {
9ff9de69 JA	48	long ret;
9ff9de69 JA	49
9ff9de69 JA	50	io_u = atom->private;
9ff9de69 JA	51	ret = *atom->ret_ptr;
e2e67912	52	if (ret >= 0)
9ff9de69 JA	53	io_u->resid = io_u->xfer_buflen - ret;
	54	else if (ret < 0)
	55	io_u->error = ret;
	56
	57	assert(sd->nr_events < td->iodepth);
	58	sd->events[sd->nr_events++] = io_u;
9ff9de69	59
5b38ee84 JA	60	if (atom == last)
5b38ee84 JA	61	break;
9ff9de69	62
5b38ee84 JA	63	atom = atom->next;
	64	} while (1);
	65
	66	assert(!last->next);
9ff9de69 JA	67	}
9ff9de69 JA	68
a4f4fdd7 JA	69	/*
	70	* Inspect the ring to see if we have completed events
	71	*/
	72	static void fio_syslet_complete(struct thread_data *td)
	73	{
	74	struct syslet_data *sd = td->io_ops->data;
	75
	76	do {
	77	struct syslet_uatom *atom;
a4f4fdd7	78
bf0dc8fa	79	atom = sd->ring[sd->ahu.user_ring_idx];
a4f4fdd7 JA	80	if (!atom)
	81	break;
	82
bf0dc8fa IM	83	sd->ring[sd->ahu.user_ring_idx] = NULL;
	84	if (++sd->ahu.user_ring_idx == td->iodepth)
	85	sd->ahu.user_ring_idx = 0;
a4f4fdd7	86
9ff9de69	87	fio_syslet_complete_atom(td, atom);
a4f4fdd7 JA	88	} while (1);
	89	}
	90
	91	static int fio_syslet_getevents(struct thread_data *td, int min,
	92	int fio_unused max,
	93	struct timespec fio_unused *t)
	94	{
	95	struct syslet_data *sd = td->io_ops->data;
a4f4fdd7 JA	96	long ret;
	97
	98	do {
	99	fio_syslet_complete(td);
	100
	101	/*
	102	* do we have enough immediate completions?
	103	*/
	104	if (sd->nr_events >= (unsigned int) min)
	105	break;
	106
	107	/*
	108	* OK, we need to wait for some events...
	109	*/
9ff9de69	110	ret = async_wait(1, sd->ahu.user_ring_idx, &sd->ahu);
a4f4fdd7	111	if (ret < 0)
e49499f8	112	return -errno;
a4f4fdd7 JA	113	} while (1);
	114
	115	ret = sd->nr_events;
	116	sd->nr_events = 0;
	117	return ret;
	118	}
	119
	120	static struct io_u fio_syslet_event(struct thread_data td, int event)
	121	{
	122	struct syslet_data *sd = td->io_ops->data;
	123
	124	return sd->events[event];
	125	}
	126
	127	static void init_atom(struct syslet_uatom atom, int nr, void arg0,
a2e1b08a JA	128	void arg1, void arg2, void arg3, void ret_ptr,
a2e1b08a JA	129	unsigned long flags, void *priv)
a4f4fdd7 JA	130	{
	131	atom->flags = flags;
	132	atom->nr = nr;
	133	atom->ret_ptr = ret_ptr;
a2e1b08a	134	atom->next = NULL;
a4f4fdd7 JA	135	atom->arg_ptr[0] = arg0;
	136	atom->arg_ptr[1] = arg1;
	137	atom->arg_ptr[2] = arg2;
a2e1b08a JA	138	atom->arg_ptr[3] = arg3;
a2e1b08a JA	139	atom->arg_ptr[4] = atom->arg_ptr[5] = NULL;
a4f4fdd7 JA	140	atom->private = priv;
	141	}
	142
	143	/*
	144	* Use seek atom for sync
	145	*/
	146	static void fio_syslet_prep_sync(struct io_u io_u, struct fio_file f)
	147	{
a2e1b08a	148	init_atom(&io_u->req.atom, __NR_fsync, &f->fd, NULL, NULL, NULL,
7d44a745	149	&io_u->req.ret, 0, io_u);
a4f4fdd7 JA	150	}
	151
	152	static void fio_syslet_prep_rw(struct io_u io_u, struct fio_file f)
	153	{
	154	int nr;
	155
a4f4fdd7 JA	156	/*
	157	* prepare rw
	158	*/
	159	if (io_u->ddir == DDIR_READ)
a2e1b08a	160	nr = __NR_pread64;
a4f4fdd7	161	else
a2e1b08a	162	nr = __NR_pwrite64;
a4f4fdd7	163
a2e1b08a	164	init_atom(&io_u->req.atom, nr, &f->fd, &io_u->xfer_buf,
7d44a745	165	&io_u->xfer_buflen, &io_u->offset, &io_u->req.ret, 0, io_u);
a4f4fdd7 JA	166	}
	167
	168	static int fio_syslet_prep(struct thread_data fio_unused td, struct io_u io_u)
	169	{
	170	struct fio_file *f = io_u->file;
	171
	172	if (io_u->ddir == DDIR_SYNC)
	173	fio_syslet_prep_sync(io_u, f);
	174	else
	175	fio_syslet_prep_rw(io_u, f);
	176
	177	return 0;
	178	}
	179
bf0dc8fa IM	180	static void cachemiss_thread_start(void)
	181	{
	182	while (1)
7756b0d0	183	async_thread(NULL, NULL);
bf0dc8fa IM	184	}
	185
	186	#define THREAD_STACK_SIZE (16384)
	187
	188	static unsigned long thread_stack_alloc()
	189	{
5b38ee84	190	return (unsigned long) malloc(THREAD_STACK_SIZE) + THREAD_STACK_SIZE;
bf0dc8fa IM	191	}
bf0dc8fa IM	192
a0a930ef JA	193	static void fio_syslet_queued(struct thread_data td, struct syslet_data sd)
	194	{
	195	struct syslet_uatom *atom;
	196	struct timeval now;
	197
	198	fio_gettime(&now, NULL);
	199
	200	atom = sd->head;
	201	while (atom) {
	202	struct io_u *io_u = atom->private;
	203
	204	memcpy(&io_u->issue_time, &now, sizeof(now));
	205	io_u_queued(td, io_u);
	206	atom = atom->next;
	207	}
	208	}
	209
9ff9de69	210	static int fio_syslet_commit(struct thread_data *td)
a4f4fdd7 JA	211	{
a4f4fdd7 JA	212	struct syslet_data *sd = td->io_ops->data;
bf0dc8fa	213	struct syslet_uatom *done;
9ff9de69 JA	214
	215	if (!sd->head)
	216	return 0;
a4f4fdd7	217
5b38ee84 JA	218	assert(!sd->tail->next);
5b38ee84 JA	219
bf0dc8fa IM	220	if (!sd->ahu.new_thread_stack)
	221	sd->ahu.new_thread_stack = thread_stack_alloc();
	222
a0a930ef JA	223	fio_syslet_queued(td, sd);
a0a930ef JA	224
7d44a745 JA	225	/*
	226	* On sync completion, the atom is returned. So on NULL return
	227	* it's queued asynchronously.
	228	*/
9ff9de69	229	done = async_exec(sd->head, &sd->ahu);
bf0dc8fa	230
9ff9de69	231	sd->head = sd->tail = NULL;
a4f4fdd7	232
9ff9de69 JA	233	if (done)
9ff9de69 JA	234	fio_syslet_complete_atom(td, done);
a4f4fdd7	235
9ff9de69 JA	236	return 0;
	237	}
	238
	239	static int fio_syslet_queue(struct thread_data td, struct io_u io_u)
	240	{
	241	struct syslet_data *sd = td->io_ops->data;
bf0dc8fa	242
9ff9de69 JA	243	if (sd->tail) {
	244	sd->tail->next = &io_u->req.atom;
	245	sd->tail = &io_u->req.atom;
	246	} else
	247	sd->head = sd->tail = &io_u->req.atom;
a4f4fdd7	248
5b38ee84	249	io_u->req.head = sd->head;
9ff9de69	250	return FIO_Q_QUEUED;
a4f4fdd7 JA	251	}
a4f4fdd7 JA	252
db64e9bc	253	static int async_head_init(struct syslet_data *sd, unsigned int depth)
a4f4fdd7	254	{
a4f4fdd7 JA	255	unsigned long ring_size;
a4f4fdd7 JA	256
bf0dc8fa	257	memset(&sd->ahu, 0, sizeof(struct async_head_user));
2ca50be4	258
a4f4fdd7 JA	259	ring_size = sizeof(struct syslet_uatom ) depth;
	260	sd->ring = malloc(ring_size);
	261	memset(sd->ring, 0, ring_size);
	262
bf0dc8fa IM	263	sd->ahu.user_ring_idx = 0;
	264	sd->ahu.completion_ring = sd->ring;
	265	sd->ahu.ring_size_bytes = ring_size;
	266	sd->ahu.head_stack = thread_stack_alloc();
5b38ee84 JA	267	sd->ahu.head_eip = (unsigned long) cachemiss_thread_start;
5b38ee84 JA	268	sd->ahu.new_thread_eip = (unsigned long) cachemiss_thread_start;
db64e9bc JA	269
db64e9bc JA	270	return 0;
a4f4fdd7 JA	271	}
a4f4fdd7 JA	272
2ca50be4	273	static void async_head_exit(struct syslet_data *sd)
a4f4fdd7	274	{
7f059a76	275	free(sd->ring);
a4f4fdd7 JA	276	}
	277
	278	static void fio_syslet_cleanup(struct thread_data *td)
	279	{
	280	struct syslet_data *sd = td->io_ops->data;
	281
	282	if (sd) {
2ca50be4	283	async_head_exit(sd);
a4f4fdd7 JA	284	free(sd->events);
	285	free(sd);
	286	td->io_ops->data = NULL;
	287	}
	288	}
	289
	290	static int fio_syslet_init(struct thread_data *td)
	291	{
	292	struct syslet_data *sd;
	293
db64e9bc	294
a4f4fdd7 JA	295	sd = malloc(sizeof(*sd));
	296	memset(sd, 0, sizeof(*sd));
	297	sd->events = malloc(sizeof(struct io_u ) td->iodepth);
	298	memset(sd->events, 0, sizeof(struct io_u ) td->iodepth);
db64e9bc JA	299
	300	/*
	301	* This will handily fail for kernels where syslet isn't available
	302	*/
	303	if (async_head_init(sd, td->iodepth)) {
	304	free(sd->events);
	305	free(sd);
	306	return 1;
	307	}
	308
a4f4fdd7	309	td->io_ops->data = sd;
a4f4fdd7 JA	310	return 0;
	311	}
	312
	313	static struct ioengine_ops ioengine = {
	314	.name = "syslet-rw",
	315	.version = FIO_IOOPS_VERSION,
	316	.init = fio_syslet_init,
	317	.prep = fio_syslet_prep,
	318	.queue = fio_syslet_queue,
9ff9de69	319	.commit = fio_syslet_commit,
a4f4fdd7 JA	320	.getevents = fio_syslet_getevents,
	321	.event = fio_syslet_event,
	322	.cleanup = fio_syslet_cleanup,
b5af8293 JA	323	.open_file = generic_open_file,
b5af8293 JA	324	.close_file = generic_close_file,
a4f4fdd7 JA	325	};
	326
	327	#else /* FIO_HAVE_SYSLET */
	328
	329	/*
	330	* When we have a proper configure system in place, we simply wont build
	331	* and install this io engine. For now install a crippled version that
	332	* just complains and fails to load.
	333	*/
	334	static int fio_syslet_init(struct thread_data fio_unused *td)
	335	{
	336	fprintf(stderr, "fio: syslet not available\n");
	337	return 1;
	338	}
	339
	340	static struct ioengine_ops ioengine = {
	341	.name = "syslet-rw",
	342	.version = FIO_IOOPS_VERSION,
	343	.init = fio_syslet_init,
	344	};
	345
	346	#endif /* FIO_HAVE_SYSLET */
	347
	348	static void fio_init fio_syslet_register(void)
	349	{
	350	register_ioengine(&ioengine);
	351	}
	352
	353	static void fio_exit fio_syslet_unregister(void)
	354	{
	355	unregister_ioengine(&ioengine);
	356	}