[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 <asm/unistd.h>

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

#ifdef FIO_HAVE_SYSLET

#ifdef __NR_pread64
#define __NR_fio_pread	__NR_pread64
#define __NR_fio_pwrite	__NR_pwrite64
#else
#define __NR_fio_pread	__NR_pread
#define __NR_fio_pwrite	__NR_pwrite
#endif

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->o.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->o.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_fio_pread;
	else
		nr = __NR_fio_pwrite;

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

	/*
	 * This will handily fail for kernels where syslet isn't available
	 */
	if (async_head_init(sd, td->o.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>
b8846354	13	#include <asm/unistd.h>
a4f4fdd7 JA	14
	15	#include "../fio.h"
	16	#include "../os.h"
	17
	18	#ifdef FIO_HAVE_SYSLET
	19
1760e679 JA	20	#ifdef __NR_pread64
	21	#define __NR_fio_pread __NR_pread64
	22	#define __NR_fio_pwrite __NR_pwrite64
	23	#else
	24	#define __NR_fio_pread __NR_pread
	25	#define __NR_fio_pwrite __NR_pwrite
	26	#endif
	27
a4f4fdd7 JA	28	struct syslet_data {
	29	struct io_u **events;
	30	unsigned int nr_events;
	31
bf0dc8fa	32	struct async_head_user ahu;
a4f4fdd7	33	struct syslet_uatom **ring;
9ff9de69 JA	34
9ff9de69 JA	35	struct syslet_uatom head, tail;
a4f4fdd7 JA	36	};
a4f4fdd7 JA	37
9ff9de69 JA	38	static void fio_syslet_complete_atom(struct thread_data *td,
	39	struct syslet_uatom *atom)
	40	{
	41	struct syslet_data *sd = td->io_ops->data;
5b38ee84	42	struct syslet_uatom *last;
9ff9de69	43	struct io_u *io_u;
9ff9de69 JA	44
9ff9de69 JA	45	/*
5b38ee84 JA	46	* complete from the beginning of the sequence up to (and
5b38ee84 JA	47	* including) this atom
9ff9de69	48	*/
5b38ee84 JA	49	last = atom;
	50	io_u = atom->private;
	51	atom = io_u->req.head;
9ff9de69 JA	52
	53	/*
	54	* now complete in right order
	55	*/
5b38ee84	56	do {
9ff9de69 JA	57	long ret;
9ff9de69 JA	58
9ff9de69 JA	59	io_u = atom->private;
9ff9de69 JA	60	ret = *atom->ret_ptr;
e2e67912	61	if (ret >= 0)
9ff9de69 JA	62	io_u->resid = io_u->xfer_buflen - ret;
	63	else if (ret < 0)
	64	io_u->error = ret;
	65
2dc1bbeb	66	assert(sd->nr_events < td->o.iodepth);
9ff9de69	67	sd->events[sd->nr_events++] = io_u;
9ff9de69	68
5b38ee84 JA	69	if (atom == last)
5b38ee84 JA	70	break;
9ff9de69	71
5b38ee84 JA	72	atom = atom->next;
	73	} while (1);
	74
	75	assert(!last->next);
9ff9de69 JA	76	}
9ff9de69 JA	77
a4f4fdd7 JA	78	/*
	79	* Inspect the ring to see if we have completed events
	80	*/
	81	static void fio_syslet_complete(struct thread_data *td)
	82	{
	83	struct syslet_data *sd = td->io_ops->data;
	84
	85	do {
	86	struct syslet_uatom *atom;
a4f4fdd7	87
bf0dc8fa	88	atom = sd->ring[sd->ahu.user_ring_idx];
a4f4fdd7 JA	89	if (!atom)
	90	break;
	91
bf0dc8fa	92	sd->ring[sd->ahu.user_ring_idx] = NULL;
2dc1bbeb	93	if (++sd->ahu.user_ring_idx == td->o.iodepth)
bf0dc8fa	94	sd->ahu.user_ring_idx = 0;
a4f4fdd7	95
9ff9de69	96	fio_syslet_complete_atom(td, atom);
a4f4fdd7 JA	97	} while (1);
	98	}
	99
	100	static int fio_syslet_getevents(struct thread_data *td, int min,
	101	int fio_unused max,
	102	struct timespec fio_unused *t)
	103	{
	104	struct syslet_data *sd = td->io_ops->data;
a4f4fdd7 JA	105	long ret;
	106
	107	do {
	108	fio_syslet_complete(td);
	109
	110	/*
	111	* do we have enough immediate completions?
	112	*/
	113	if (sd->nr_events >= (unsigned int) min)
	114	break;
	115
	116	/*
	117	* OK, we need to wait for some events...
	118	*/
9ff9de69	119	ret = async_wait(1, sd->ahu.user_ring_idx, &sd->ahu);
a4f4fdd7	120	if (ret < 0)
e49499f8	121	return -errno;
a4f4fdd7 JA	122	} while (1);
	123
	124	ret = sd->nr_events;
	125	sd->nr_events = 0;
	126	return ret;
	127	}
	128
	129	static struct io_u fio_syslet_event(struct thread_data td, int event)
	130	{
	131	struct syslet_data *sd = td->io_ops->data;
	132
	133	return sd->events[event];
	134	}
	135
	136	static void init_atom(struct syslet_uatom atom, int nr, void arg0,
a2e1b08a JA	137	void arg1, void arg2, void arg3, void ret_ptr,
a2e1b08a JA	138	unsigned long flags, void *priv)
a4f4fdd7 JA	139	{
	140	atom->flags = flags;
	141	atom->nr = nr;
	142	atom->ret_ptr = ret_ptr;
a2e1b08a	143	atom->next = NULL;
a4f4fdd7 JA	144	atom->arg_ptr[0] = arg0;
	145	atom->arg_ptr[1] = arg1;
	146	atom->arg_ptr[2] = arg2;
a2e1b08a JA	147	atom->arg_ptr[3] = arg3;
a2e1b08a JA	148	atom->arg_ptr[4] = atom->arg_ptr[5] = NULL;
a4f4fdd7 JA	149	atom->private = priv;
	150	}
	151
	152	/*
	153	* Use seek atom for sync
	154	*/
	155	static void fio_syslet_prep_sync(struct io_u io_u, struct fio_file f)
	156	{
a2e1b08a	157	init_atom(&io_u->req.atom, __NR_fsync, &f->fd, NULL, NULL, NULL,
7d44a745	158	&io_u->req.ret, 0, io_u);
a4f4fdd7 JA	159	}
	160
	161	static void fio_syslet_prep_rw(struct io_u io_u, struct fio_file f)
	162	{
	163	int nr;
	164
a4f4fdd7 JA	165	/*
	166	* prepare rw
	167	*/
	168	if (io_u->ddir == DDIR_READ)
1760e679	169	nr = __NR_fio_pread;
a4f4fdd7	170	else
1760e679	171	nr = __NR_fio_pwrite;
a4f4fdd7	172
a2e1b08a	173	init_atom(&io_u->req.atom, nr, &f->fd, &io_u->xfer_buf,
7d44a745	174	&io_u->xfer_buflen, &io_u->offset, &io_u->req.ret, 0, io_u);
a4f4fdd7 JA	175	}
	176
	177	static int fio_syslet_prep(struct thread_data fio_unused td, struct io_u io_u)
	178	{
	179	struct fio_file *f = io_u->file;
	180
	181	if (io_u->ddir == DDIR_SYNC)
	182	fio_syslet_prep_sync(io_u, f);
	183	else
	184	fio_syslet_prep_rw(io_u, f);
	185
	186	return 0;
	187	}
	188
bf0dc8fa IM	189	static void cachemiss_thread_start(void)
	190	{
	191	while (1)
7756b0d0	192	async_thread(NULL, NULL);
bf0dc8fa IM	193	}
	194
	195	#define THREAD_STACK_SIZE (16384)
	196
	197	static unsigned long thread_stack_alloc()
	198	{
5b38ee84	199	return (unsigned long) malloc(THREAD_STACK_SIZE) + THREAD_STACK_SIZE;
bf0dc8fa IM	200	}
bf0dc8fa IM	201
a0a930ef JA	202	static void fio_syslet_queued(struct thread_data td, struct syslet_data sd)
	203	{
	204	struct syslet_uatom *atom;
	205	struct timeval now;
	206
	207	fio_gettime(&now, NULL);
	208
	209	atom = sd->head;
	210	while (atom) {
	211	struct io_u *io_u = atom->private;
	212
	213	memcpy(&io_u->issue_time, &now, sizeof(now));
	214	io_u_queued(td, io_u);
	215	atom = atom->next;
	216	}
	217	}
	218
9ff9de69	219	static int fio_syslet_commit(struct thread_data *td)
a4f4fdd7 JA	220	{
a4f4fdd7 JA	221	struct syslet_data *sd = td->io_ops->data;
bf0dc8fa	222	struct syslet_uatom *done;
9ff9de69 JA	223
	224	if (!sd->head)
	225	return 0;
a4f4fdd7	226
5b38ee84 JA	227	assert(!sd->tail->next);
5b38ee84 JA	228
bf0dc8fa IM	229	if (!sd->ahu.new_thread_stack)
	230	sd->ahu.new_thread_stack = thread_stack_alloc();
	231
a0a930ef JA	232	fio_syslet_queued(td, sd);
a0a930ef JA	233
7d44a745 JA	234	/*
	235	* On sync completion, the atom is returned. So on NULL return
	236	* it's queued asynchronously.
	237	*/
9ff9de69	238	done = async_exec(sd->head, &sd->ahu);
bf0dc8fa	239
9ff9de69	240	sd->head = sd->tail = NULL;
a4f4fdd7	241
9ff9de69 JA	242	if (done)
9ff9de69 JA	243	fio_syslet_complete_atom(td, done);
a4f4fdd7	244
9ff9de69 JA	245	return 0;
	246	}
	247
	248	static int fio_syslet_queue(struct thread_data td, struct io_u io_u)
	249	{
	250	struct syslet_data *sd = td->io_ops->data;
bf0dc8fa	251
9ff9de69 JA	252	if (sd->tail) {
	253	sd->tail->next = &io_u->req.atom;
	254	sd->tail = &io_u->req.atom;
	255	} else
	256	sd->head = sd->tail = &io_u->req.atom;
a4f4fdd7	257
5b38ee84	258	io_u->req.head = sd->head;
9ff9de69	259	return FIO_Q_QUEUED;
a4f4fdd7 JA	260	}
a4f4fdd7 JA	261
db64e9bc	262	static int async_head_init(struct syslet_data *sd, unsigned int depth)
a4f4fdd7	263	{
a4f4fdd7 JA	264	unsigned long ring_size;
a4f4fdd7 JA	265
bf0dc8fa	266	memset(&sd->ahu, 0, sizeof(struct async_head_user));
2ca50be4	267
a4f4fdd7 JA	268	ring_size = sizeof(struct syslet_uatom ) depth;
	269	sd->ring = malloc(ring_size);
	270	memset(sd->ring, 0, ring_size);
	271
bf0dc8fa IM	272	sd->ahu.user_ring_idx = 0;
	273	sd->ahu.completion_ring = sd->ring;
	274	sd->ahu.ring_size_bytes = ring_size;
	275	sd->ahu.head_stack = thread_stack_alloc();
5b38ee84 JA	276	sd->ahu.head_eip = (unsigned long) cachemiss_thread_start;
5b38ee84 JA	277	sd->ahu.new_thread_eip = (unsigned long) cachemiss_thread_start;
db64e9bc JA	278
db64e9bc JA	279	return 0;
a4f4fdd7 JA	280	}
a4f4fdd7 JA	281
2ca50be4	282	static void async_head_exit(struct syslet_data *sd)
a4f4fdd7	283	{
7f059a76	284	free(sd->ring);
a4f4fdd7 JA	285	}
	286
	287	static void fio_syslet_cleanup(struct thread_data *td)
	288	{
	289	struct syslet_data *sd = td->io_ops->data;
	290
	291	if (sd) {
2ca50be4	292	async_head_exit(sd);
a4f4fdd7 JA	293	free(sd->events);
	294	free(sd);
	295	td->io_ops->data = NULL;
	296	}
	297	}
	298
	299	static int fio_syslet_init(struct thread_data *td)
	300	{
	301	struct syslet_data *sd;
	302
db64e9bc	303
a4f4fdd7 JA	304	sd = malloc(sizeof(*sd));
a4f4fdd7 JA	305	memset(sd, 0, sizeof(*sd));
2dc1bbeb JA	306	sd->events = malloc(sizeof(struct io_u ) td->o.iodepth);
2dc1bbeb JA	307	memset(sd->events, 0, sizeof(struct io_u ) td->o.iodepth);
db64e9bc JA	308
	309	/*
	310	* This will handily fail for kernels where syslet isn't available
	311	*/
2dc1bbeb	312	if (async_head_init(sd, td->o.iodepth)) {
db64e9bc JA	313	free(sd->events);
	314	free(sd);
	315	return 1;
	316	}
	317
a4f4fdd7	318	td->io_ops->data = sd;
a4f4fdd7 JA	319	return 0;
	320	}
	321
	322	static struct ioengine_ops ioengine = {
	323	.name = "syslet-rw",
	324	.version = FIO_IOOPS_VERSION,
	325	.init = fio_syslet_init,
	326	.prep = fio_syslet_prep,
	327	.queue = fio_syslet_queue,
9ff9de69	328	.commit = fio_syslet_commit,
a4f4fdd7 JA	329	.getevents = fio_syslet_getevents,
	330	.event = fio_syslet_event,
	331	.cleanup = fio_syslet_cleanup,
b5af8293 JA	332	.open_file = generic_open_file,
b5af8293 JA	333	.close_file = generic_close_file,
a4f4fdd7 JA	334	};
	335
	336	#else /* FIO_HAVE_SYSLET */
	337
	338	/*
	339	* When we have a proper configure system in place, we simply wont build
	340	* and install this io engine. For now install a crippled version that
	341	* just complains and fails to load.
	342	*/
	343	static int fio_syslet_init(struct thread_data fio_unused *td)
	344	{
	345	fprintf(stderr, "fio: syslet not available\n");
	346	return 1;
	347	}
	348
	349	static struct ioengine_ops ioengine = {
	350	.name = "syslet-rw",
	351	.version = FIO_IOOPS_VERSION,
	352	.init = fio_syslet_init,
	353	};
	354
	355	#endif /* FIO_HAVE_SYSLET */
	356
	357	static void fio_init fio_syslet_register(void)
	358	{
	359	register_ioengine(&ioengine);
	360	}
	361
	362	static void fio_exit fio_syslet_unregister(void)
	363	{
	364	unregister_ioengine(&ioengine);
	365	}