[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 <malloc.h>
#include <asm/unistd.h>

#include "../fio.h"
#include "../indirect.h"
#include "../syslet.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 syslet_ring *ring;
	void *stack;
};

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

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

static void fio_syslet_add_events(struct thread_data *td, unsigned int nr)
{
	struct syslet_data *sd = td->io_ops->data;
	unsigned int i;

	for (i = 0; i < nr; i++) {
		unsigned int idx = (i + sd->ring->user_tail) % td->o.iodepth;
		struct syslet_completion *comp = &sd->ring->comp[idx];
		struct io_u *io_u = (struct io_u *) (long) comp->caller_data;
		long ret;

		ret = comp->status;
		if (ret <= 0) {
			io_u->resid = io_u->xfer_buflen;
			io_u->error = -ret;
		} else {
			io_u->resid = io_u->xfer_buflen - ret;
			io_u->error = 0;
		}

		fio_syslet_add_event(td, io_u);
	}
}

static void fio_syslet_wait_for_events(struct thread_data *td)
{
	struct syslet_data *sd = td->io_ops->data;
	struct syslet_ring *ring = sd->ring;
	unsigned int events;

	events = 0;
	do {
		unsigned int kh = ring->kernel_head;
		int ret;

		/*
		 * first reap events that are already completed
		 */
		if (ring->user_tail != kh) {
			unsigned int nr = kh - ring->user_tail;

			fio_syslet_add_events(td, nr);
			events += nr;
			ring->user_tail = kh;
			continue;
		}

		/*
		 * block waiting for at least one event
		 */
		ret = syscall(__NR_syslet_ring_wait, ring, ring->user_tail);
		assert(!ret);
	} while (!events);
}

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

		fio_syslet_wait_for_events(td);
	} 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 fio_syslet_prep_sync(struct fio_file *f,
				 struct indirect_registers *regs)
{
	FILL_IN(*regs, __NR_fsync, (long) f->fd);
}

static void fio_syslet_prep_rw(struct io_u *io_u, struct fio_file *f,
			       struct indirect_registers *regs)
{
	long nr;

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

	FILL_IN(*regs, nr, (long) f->fd, (long) io_u->xfer_buf,
		(long) io_u->xfer_buflen, (long) io_u->offset);
}

static void fio_syslet_prep(struct io_u *io_u, struct indirect_registers *regs)
{
	struct fio_file *f = io_u->file;

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

static void ret_func(void)
{
	syscall(__NR_exit);
}

static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
{
	struct syslet_data *sd = td->io_ops->data;
	union indirect_params params;
	struct indirect_registers regs;
	int ret;

	fio_ro_check(td, io_u);

	memset(&params, 0, sizeof(params));
	fill_syslet_args(&params.syslet, sd->ring, (long)io_u, ret_func, sd->stack);

	fio_syslet_prep(io_u, &regs);

	ret = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
	if (ret == (int) io_u->xfer_buflen) {
		/*
		 * completed sync, account. this also catches fsync().
		 */
		return FIO_Q_COMPLETED;
	} else if (ret < 0) {
		/*
		 * queued for async execution
		 */
		if (errno == ESYSLETPENDING)
			return FIO_Q_QUEUED;
	}

	io_u->error = errno;
	td_verror(td, io_u->error, "xfer");
	return FIO_Q_COMPLETED;
}

static int check_syslet_support(struct syslet_data *sd)
{
	union indirect_params params;
	struct indirect_registers regs;
	pid_t pid, my_pid = getpid();

	memset(&params, 0, sizeof(params));
	fill_syslet_args(&params.syslet, sd->ring, 0, ret_func, sd->stack);

	FILL_IN(regs, __NR_getpid);

	pid = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
	if (pid == my_pid)
		return 0;

	return 1;
}

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

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

static int fio_syslet_init(struct thread_data *td)
{
	struct syslet_data *sd;
	void *ring = NULL, *stack = NULL;
	unsigned int ring_size, ring_nr;

	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);

	/*
	 * The ring needs to be a power-of-2, so round it up if we have to
	 */
	ring_nr = td->o.iodepth;
	if (ring_nr & (ring_nr - 1)) {
		int bits = 1;

		while (ring_nr >>= 1)
			bits++;

		ring_nr = 1 << bits;
	}

	ring_size = sizeof(struct syslet_ring) +
			ring_nr * sizeof(struct syslet_completion);
	if (posix_memalign(&ring, sizeof(uint64_t), ring_size))
		goto err_mem;
	if (posix_memalign(&stack, page_size, page_size))
		goto err_mem;

	sd->ring = ring;
	sd->stack = stack;

	memset(sd->ring, 0, ring_size);
	sd->ring->elements = ring_nr;

	if (!check_syslet_support(sd)) {
		td->io_ops->data = sd;
		return 0;
	}

	log_err("fio: syslets do not appear to work\n");
err_mem:
	free(sd->events);
	if (ring)
		free(ring);
	if (stack)
		free(stack);
	free(sd);
	return 1;
}

static struct ioengine_ops ioengine = {
	.name		= "syslet-rw",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_syslet_init,
	.queue		= fio_syslet_queue,
	.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>
12cbb469	13	#include <malloc.h>
b8846354	14	#include <asm/unistd.h>
a4f4fdd7 JA	15
a4f4fdd7 JA	16	#include "../fio.h"
12cbb469 JA	17	#include "../indirect.h"
12cbb469 JA	18	#include "../syslet.h"
a4f4fdd7 JA	19
	20	#ifdef FIO_HAVE_SYSLET
	21
1760e679 JA	22	#ifdef __NR_pread64
	23	#define __NR_fio_pread __NR_pread64
	24	#define __NR_fio_pwrite __NR_pwrite64
	25	#else
	26	#define __NR_fio_pread __NR_pread
	27	#define __NR_fio_pwrite __NR_pwrite
	28	#endif
	29
a4f4fdd7 JA	30	struct syslet_data {
	31	struct io_u **events;
	32	unsigned int nr_events;
	33
12cbb469 JA	34	struct syslet_ring *ring;
12cbb469 JA	35	void *stack;
a4f4fdd7 JA	36	};
a4f4fdd7 JA	37
c6effc3e	38	static void fio_syslet_add_event(struct thread_data td, struct io_u io_u)
9ff9de69 JA	39	{
9ff9de69 JA	40	struct syslet_data *sd = td->io_ops->data;
9ff9de69	41
12cbb469 JA	42	assert(sd->nr_events < td->o.iodepth);
	43	sd->events[sd->nr_events++] = io_u;
	44	}
9ff9de69	45
c6effc3e	46	static void fio_syslet_add_events(struct thread_data *td, unsigned int nr)
12cbb469 JA	47	{
	48	struct syslet_data *sd = td->io_ops->data;
	49	unsigned int i;
9ff9de69	50
12cbb469 JA	51	for (i = 0; i < nr; i++) {
	52	unsigned int idx = (i + sd->ring->user_tail) % td->o.iodepth;
	53	struct syslet_completion *comp = &sd->ring->comp[idx];
	54	struct io_u io_u = (struct io_u ) (long) comp->caller_data;
c6effc3e JA	55	long ret;
	56
	57	ret = comp->status;
	58	if (ret <= 0) {
	59	io_u->resid = io_u->xfer_buflen;
	60	io_u->error = -ret;
	61	} else {
	62	io_u->resid = io_u->xfer_buflen - ret;
	63	io_u->error = 0;
	64	}
5b38ee84	65
c6effc3e	66	fio_syslet_add_event(td, io_u);
12cbb469	67	}
9ff9de69 JA	68	}
9ff9de69 JA	69
12cbb469	70	static void fio_syslet_wait_for_events(struct thread_data *td)
a4f4fdd7 JA	71	{
a4f4fdd7 JA	72	struct syslet_data *sd = td->io_ops->data;
12cbb469 JA	73	struct syslet_ring *ring = sd->ring;
12cbb469 JA	74	unsigned int events;
a4f4fdd7	75
12cbb469	76	events = 0;
a4f4fdd7	77	do {
12cbb469 JA	78	unsigned int kh = ring->kernel_head;
12cbb469 JA	79	int ret;
a4f4fdd7	80
12cbb469 JA	81	/*
	82	* first reap events that are already completed
	83	*/
	84	if (ring->user_tail != kh) {
	85	unsigned int nr = kh - ring->user_tail;
a4f4fdd7	86
c6effc3e	87	fio_syslet_add_events(td, nr);
12cbb469 JA	88	events += nr;
	89	ring->user_tail = kh;
	90	continue;
	91	}
a4f4fdd7	92
12cbb469 JA	93	/*
	94	* block waiting for at least one event
	95	*/
	96	ret = syscall(__NR_syslet_ring_wait, ring, ring->user_tail);
	97	assert(!ret);
	98	} while (!events);
a4f4fdd7 JA	99	}
	100
	101	static int fio_syslet_getevents(struct thread_data *td, int min,
	102	int fio_unused max,
	103	struct timespec fio_unused *t)
	104	{
	105	struct syslet_data *sd = td->io_ops->data;
a4f4fdd7 JA	106	long ret;
	107
	108	do {
a4f4fdd7 JA	109	/*
	110	* do we have enough immediate completions?
	111	*/
	112	if (sd->nr_events >= (unsigned int) min)
	113	break;
	114
12cbb469	115	fio_syslet_wait_for_events(td);
a4f4fdd7 JA	116	} while (1);
	117
	118	ret = sd->nr_events;
	119	sd->nr_events = 0;
	120	return ret;
	121	}
	122
	123	static struct io_u fio_syslet_event(struct thread_data td, int event)
	124	{
	125	struct syslet_data *sd = td->io_ops->data;
	126
	127	return sd->events[event];
	128	}
	129
12cbb469 JA	130	static void fio_syslet_prep_sync(struct fio_file *f,
12cbb469 JA	131	struct indirect_registers *regs)
a4f4fdd7	132	{
12cbb469	133	FILL_IN(*regs, __NR_fsync, (long) f->fd);
a4f4fdd7 JA	134	}
a4f4fdd7 JA	135
12cbb469 JA	136	static void fio_syslet_prep_rw(struct io_u io_u, struct fio_file f,
12cbb469 JA	137	struct indirect_registers *regs)
a4f4fdd7	138	{
12cbb469	139	long nr;
a4f4fdd7	140
a4f4fdd7 JA	141	/*
	142	* prepare rw
	143	*/
	144	if (io_u->ddir == DDIR_READ)
1760e679	145	nr = __NR_fio_pread;
a4f4fdd7	146	else
1760e679	147	nr = __NR_fio_pwrite;
a4f4fdd7	148
12cbb469 JA	149	FILL_IN(*regs, nr, (long) f->fd, (long) io_u->xfer_buf,
12cbb469 JA	150	(long) io_u->xfer_buflen, (long) io_u->offset);
a4f4fdd7 JA	151	}
a4f4fdd7 JA	152
12cbb469	153	static void fio_syslet_prep(struct io_u io_u, struct indirect_registers regs)
a4f4fdd7 JA	154	{
	155	struct fio_file *f = io_u->file;
	156
	157	if (io_u->ddir == DDIR_SYNC)
12cbb469	158	fio_syslet_prep_sync(f, regs);
a4f4fdd7	159	else
12cbb469	160	fio_syslet_prep_rw(io_u, f, regs);
a0a930ef JA	161	}
a0a930ef JA	162
12cbb469	163	static void ret_func(void)
a4f4fdd7	164	{
12cbb469	165	syscall(__NR_exit);
9ff9de69 JA	166	}
	167
	168	static int fio_syslet_queue(struct thread_data td, struct io_u io_u)
	169	{
	170	struct syslet_data *sd = td->io_ops->data;
12cbb469 JA	171	union indirect_params params;
	172	struct indirect_registers regs;
	173	int ret;
bf0dc8fa	174
7101d9c2 JA	175	fio_ro_check(td, io_u);
7101d9c2 JA	176
12cbb469 JA	177	memset(&params, 0, sizeof(params));
12cbb469 JA	178	fill_syslet_args(&params.syslet, sd->ring, (long)io_u, ret_func, sd->stack);
a4f4fdd7	179
12cbb469	180	fio_syslet_prep(io_u, &regs);
2ca50be4	181
12cbb469 JA	182	ret = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
	183	if (ret == (int) io_u->xfer_buflen) {
	184	/*
	185	* completed sync, account. this also catches fsync().
	186	*/
	187	return FIO_Q_COMPLETED;
	188	} else if (ret < 0) {
	189	/*
	190	* queued for async execution
	191	*/
	192	if (errno == ESYSLETPENDING)
	193	return FIO_Q_QUEUED;
	194	}
a4f4fdd7	195
12cbb469 JA	196	io_u->error = errno;
	197	td_verror(td, io_u->error, "xfer");
	198	return FIO_Q_COMPLETED;
a4f4fdd7 JA	199	}
a4f4fdd7 JA	200
76f58b92 JA	201	static int check_syslet_support(struct syslet_data *sd)
76f58b92 JA	202	{
12cbb469 JA	203	union indirect_params params;
	204	struct indirect_registers regs;
	205	pid_t pid, my_pid = getpid();
76f58b92	206
12cbb469 JA	207	memset(&params, 0, sizeof(params));
12cbb469 JA	208	fill_syslet_args(&params.syslet, sd->ring, 0, ret_func, sd->stack);
76f58b92	209
12cbb469 JA	210	FILL_IN(regs, __NR_getpid);
	211
	212	pid = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
	213	if (pid == my_pid)
	214	return 0;
	215
	216	return 1;
76f58b92 JA	217	}
76f58b92 JA	218
a4f4fdd7 JA	219	static void fio_syslet_cleanup(struct thread_data *td)
	220	{
	221	struct syslet_data *sd = td->io_ops->data;
	222
	223	if (sd) {
a4f4fdd7	224	free(sd->events);
12cbb469 JA	225	free(sd->ring);
12cbb469 JA	226	free(sd->stack);
a4f4fdd7 JA	227	free(sd);
	228	td->io_ops->data = NULL;
	229	}
	230	}
	231
	232	static int fio_syslet_init(struct thread_data *td)
	233	{
	234	struct syslet_data *sd;
e81f7340	235	void ring = NULL, stack = NULL;
05f6a172	236	unsigned int ring_size, ring_nr;
a4f4fdd7 JA	237
	238	sd = malloc(sizeof(*sd));
	239	memset(sd, 0, sizeof(*sd));
12cbb469	240
2dc1bbeb JA	241	sd->events = malloc(sizeof(struct io_u ) td->o.iodepth);
2dc1bbeb JA	242	memset(sd->events, 0, sizeof(struct io_u ) td->o.iodepth);
05f6a172 JA	243
	244	/*
	245	* The ring needs to be a power-of-2, so round it up if we have to
	246	*/
	247	ring_nr = td->o.iodepth;
	248	if (ring_nr & (ring_nr - 1)) {
	249	int bits = 1;
	250
	251	while (ring_nr >>= 1)
	252	bits++;
	253
	254	ring_nr = 1 << bits;
	255	}
	256
	257	ring_size = sizeof(struct syslet_ring) +
	258	ring_nr * sizeof(struct syslet_completion);
	259	if (posix_memalign(&ring, sizeof(uint64_t), ring_size))
e81f7340	260	goto err_mem;
12cbb469	261	if (posix_memalign(&stack, page_size, page_size))
e81f7340	262	goto err_mem;
12cbb469 JA	263
	264	sd->ring = ring;
	265	sd->stack = stack;
	266
05f6a172 JA	267	memset(sd->ring, 0, ring_size);
05f6a172 JA	268	sd->ring->elements = ring_nr;
db64e9bc	269
e81f7340 JA	270	if (!check_syslet_support(sd)) {
	271	td->io_ops->data = sd;
	272	return 0;
76f58b92 JA	273	}
76f58b92 JA	274
e81f7340 JA	275	log_err("fio: syslets do not appear to work\n");
	276	err_mem:
	277	free(sd->events);
	278	if (ring)
	279	free(ring);
	280	if (stack)
	281	free(stack);
	282	free(sd);
	283	return 1;
a4f4fdd7 JA	284	}
	285
	286	static struct ioengine_ops ioengine = {
	287	.name = "syslet-rw",
	288	.version = FIO_IOOPS_VERSION,
	289	.init = fio_syslet_init,
a4f4fdd7 JA	290	.queue = fio_syslet_queue,
	291	.getevents = fio_syslet_getevents,
	292	.event = fio_syslet_event,
	293	.cleanup = fio_syslet_cleanup,
b5af8293 JA	294	.open_file = generic_open_file,
b5af8293 JA	295	.close_file = generic_close_file,
a4f4fdd7 JA	296	};
	297
	298	#else /* FIO_HAVE_SYSLET */
	299
	300	/*
	301	* When we have a proper configure system in place, we simply wont build
	302	* and install this io engine. For now install a crippled version that
	303	* just complains and fails to load.
	304	*/
	305	static int fio_syslet_init(struct thread_data fio_unused *td)
	306	{
	307	fprintf(stderr, "fio: syslet not available\n");
	308	return 1;
	309	}
	310
	311	static struct ioengine_ops ioengine = {
	312	.name = "syslet-rw",
	313	.version = FIO_IOOPS_VERSION,
	314	.init = fio_syslet_init,
	315	};
	316
	317	#endif /* FIO_HAVE_SYSLET */
	318
	319	static void fio_init fio_syslet_register(void)
	320	{
	321	register_ioengine(&ioengine);
	322	}
	323
	324	static void fio_exit fio_syslet_unregister(void)
	325	{
	326	unregister_ioengine(&ioengine);
	327	}