[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 <fcntl.h>
#include <assert.h>

#include "fio.h"
#include "diskutil.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, engine_lib);
	if (!ops)
		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;
}

/*
 * For cleaning up an ioengine which never made it to init().
 */
void free_ioengine(struct thread_data *td)
{
	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);

	if (td->eo && td->io_ops->options) {
		options_free(td->io_ops->options, td->eo);
		free(td->eo);
		td->eo = NULL;
	}

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

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

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;
	}

	free_ioengine(td);
}

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;

	/*
	 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	 * server side gets a message from the client
	 * side that the task is finished, and
	 * td->done is set to 1 after td_io_commit(). In this case,
	 * there is no need to reap complete event in server side.
	 */
	if (td->done)
		return 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) {
		/*
 		 * Reflect that our submitted requests were retrieved with
		 * whatever OS async calls are in the underlying engine.
		 */
		td->io_u_in_flight -= r;
		io_u_mark_complete(td, r);
	} else
		td_verror(td, r, "get_events");

	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(fio_file_open(io_u->file));

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

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

	if (ddir_rw(io_u->ddir))
		td->io_issues[io_u->ddir]++;

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

	unlock_file(td, io_u->file);

	/*
	 * Add warning for O_DIRECT so that users have an easier time
	 * spotting potentially bad alignment. If this triggers for the first
	 * IO, then it's likely an alignment problem or because the host fs
	 * does not support O_DIRECT
	 */
	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
	    td->o.odirect) {

		log_info("fio: first direct IO errored. File system may not "
			 "support direct IO, or iomem_align= is bad.\n");
	}

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

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

		if (ddir_rw(io_u->ddir)) {
			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) {
		if (fio_fill_issue_time(td))
			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));
	}

	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");
		}
		if (!td->error)
			td->error = ret;
	}

	return ret;
}

int td_io_commit(struct thread_data *td)
{
	int ret;

	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);	

	if (td->io_ops->commit) {
		ret = td->io_ops->commit(td);
		if (ret)
			td_verror(td, -ret, "io commit");
	}
	
	/*
	 * Reflect that events were submitted as async IO requests.
	 */
	td->io_u_in_flight += td->io_u_queued;
	td->io_u_queued = 0;

	return 0;
}

int td_io_open_file(struct thread_data *td, struct fio_file *f)
{
	assert(!fio_file_open(f));
	assert(f->fd == -1);

	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);
		}

		assert(f->fd == -1);
		assert(!fio_file_open(f));
		return 1;
	}

	fio_file_reset(f);
	fio_file_set_open(f);
	fio_file_clear_closing(f);
	disk_util_inc(f->du);

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

#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");
			log_err("fio: the file system does not seem to support direct IO\n");
			goto err;
		}
	}
#endif

done:
	log_file(td, f, FIO_LOG_OPEN_FILE);
	return 0;
err:
	disk_util_dec(f->du);
	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 (!fio_file_closing(f))
		log_file(td, f, FIO_LOG_CLOSE_FILE);

	/*
	 * mark as closing, do real close when last io on it has completed
	 */
	fio_file_set_closing(f);

	disk_util_dec(f->du);
	unlock_file_all(td, f);

	return put_file(td, f);
}

int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
{
	if (!td->io_ops->get_file_size)
		return 0;

	return td->io_ops->get_file_size(td, f);
}

static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
{
	off64_t offset, nbytes;

	offset = f->first_write;
	nbytes = f->last_write - f->first_write;

	if (!nbytes)
		return 0;

	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
}

int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
{
	int ret;

	if (io_u->ddir == DDIR_SYNC) {
		ret = fsync(io_u->file->fd);
	} else if (io_u->ddir == DDIR_DATASYNC) {
#ifdef FIO_HAVE_FDATASYNC
		ret = fdatasync(io_u->file->fd);
#else
		ret = io_u->xfer_buflen;
		io_u->error = EINVAL;
#endif
	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
		ret = do_sync_file_range(td, io_u->file);
	else {
		ret = io_u->xfer_buflen;
		io_u->error = EINVAL;
	}

	if (ret < 0)
		io_u->error = errno;

	return ret;
}

int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
{
#ifndef FIO_HAVE_TRIM
	io_u->error = EINVAL;
	return 0;
#else
	struct fio_file *f = io_u->file;
	int ret;

	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
	if (!ret)
		return io_u->xfer_buflen;;

	io_u->error = ret;
	return 0;
#endif
}

int fio_show_ioengine_help(const char *engine)
{
	struct flist_head *entry;
	struct thread_data td;
	char *sep;
	int ret = 1;

	if (!engine || !*engine) {
		log_info("Available IO engines:\n");
		flist_for_each(entry, &engine_list) {
			td.io_ops = flist_entry(entry, struct ioengine_ops,
						list);
			log_info("\t%s\n", td.io_ops->name);
		}
		return 0;
	}
	sep = strchr(engine, ',');
	if (sep) {
		*sep = 0;
		sep++;
	}

	memset(&td, 0, sizeof(td));

	td.io_ops = load_ioengine(&td, engine);
	if (!td.io_ops) {
		log_info("IO engine %s not found\n", engine);
		return 1;
	}

	if (td.io_ops->options)
		ret = show_cmd_help(td.io_ops->options, sep);
	else
		log_info("IO engine %s has no options\n", td.io_ops->name);

	free_ioengine(&td);

	return ret;
}
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>
ecc314ba	17	#include <fcntl.h>
0c6e7517	18	#include <assert.h>
8c16d840	19
ebac4655	20	#include "fio.h"
7c9b1bce	21	#include "diskutil.h"
ebac4655	22
01743ee1	23	static FLIST_HEAD(engine_list);
5f350952	24
8c16d840 JA	25	static int check_engine_ops(struct ioengine_ops *ops)
8c16d840 JA	26	{
5f350952	27	if (ops->version != FIO_IOOPS_VERSION) {
5ec10eaa JA	28	log_err("bad ioops version %d (want %d)\n", ops->version,
5ec10eaa JA	29	FIO_IOOPS_VERSION);
5f350952 JA	30	return 1;
	31	}
	32
36167d82 JA	33	if (!ops->queue) {
	34	log_err("%s: no queue handler\n", ops->name);
	35	return 1;
	36	}
	37
	38	/*
	39	* sync engines only need a ->queue()
	40	*/
	41	if (ops->flags & FIO_SYNCIO)
	42	return 0;
5ec10eaa	43
8c16d840	44	if (!ops->event) {
36167d82	45	log_err("%s: no event handler\n", ops->name);
8c16d840 JA	46	return 1;
	47	}
	48	if (!ops->getevents) {
36167d82	49	log_err("%s: no getevents handler\n", ops->name);
8c16d840 JA	50	return 1;
	51	}
	52	if (!ops->queue) {
36167d82	53	log_err("%s: no queue handler\n", ops->name);
8c16d840 JA	54	return 1;
8c16d840 JA	55	}
5ec10eaa	56
8c16d840 JA	57	return 0;
	58	}
	59
5f350952	60	void unregister_ioengine(struct ioengine_ops *ops)
ebac4655	61	{
ee56ad50	62	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1 JA	63	flist_del(&ops->list);
01743ee1 JA	64	INIT_FLIST_HEAD(&ops->list);
5f350952 JA	65	}
5f350952 JA	66
b2fdda43	67	void register_ioengine(struct ioengine_ops *ops)
5f350952	68	{
ee56ad50	69	dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1 JA	70	INIT_FLIST_HEAD(&ops->list);
01743ee1 JA	71	flist_add_tail(&ops->list, &engine_list);
5f350952 JA	72	}
	73
	74	static struct ioengine_ops find_ioengine(const char name)
	75	{
	76	struct ioengine_ops *ops;
01743ee1	77	struct flist_head *entry;
ebac4655	78
01743ee1 JA	79	flist_for_each(entry, &engine_list) {
01743ee1 JA	80	ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8	81	if (!strcmp(name, ops->name))
5f350952 JA	82	return ops;
	83	}
	84
	85	return NULL;
	86	}
	87
	88	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	89	const char *engine_lib)
	90	{
	91	struct ioengine_ops *ops;
	92	void *dlhandle;
	93
ee56ad50 JA	94	dprint(FD_IO, "dload engine %s\n", engine_lib);
ee56ad50 JA	95
2866c82d JA	96	dlerror();
2866c82d JA	97	dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8	98	if (!dlhandle) {
e1161c32	99	td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8 JA	100	return NULL;
d4dbaaa8 JA	101	}
8756e4d4	102
da51c050 JA	103	/*
	104	* Unlike the included modules, external engines should have a
	105	* non-static ioengine structure that we can reference.
	106	*/
0abea0b7 DM	107	ops = dlsym(dlhandle, engine_lib);
	108	if (!ops)
	109	ops = dlsym(dlhandle, "ioengine");
d4dbaaa8	110	if (!ops) {
e1161c32	111	td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8 JA	112	dlclose(dlhandle);
	113	return NULL;
	114	}
8756e4d4	115
5f350952 JA	116	ops->dlhandle = dlhandle;
	117	return ops;
	118	}
	119
	120	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	121	{
	122	struct ioengine_ops ops, ret;
	123	char engine[16];
	124
ee56ad50 JA	125	dprint(FD_IO, "load ioengine %s\n", name);
ee56ad50 JA	126
5f350952 JA	127	strncpy(engine, name, sizeof(engine) - 1);
	128
	129	/*
	130	* linux libaio has alias names, so convert to what we want
	131	*/
	132	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	133	strcpy(engine, "libaio");
	134
	135	ops = find_ioengine(engine);
	136	if (!ops)
	137	ops = dlopen_ioengine(td, name);
	138
	139	if (!ops) {
	140	log_err("fio: engine %s not loadable\n", name);
b902ceb5 JA	141	return NULL;
	142	}
	143
8c16d840 JA	144	/*
	145	* Check that the required methods are there.
	146	*/
5f350952	147	if (check_engine_ops(ops))
8c16d840	148	return NULL;
8c16d840	149
84585003 JA	150	ret = malloc(sizeof(*ret));
	151	memcpy(ret, ops, sizeof(*ret));
	152	ret->data = NULL;
84585003 JA	153
84585003 JA	154	return ret;
8756e4d4 JA	155	}
8756e4d4 JA	156
de890a1e SL	157	/*
	158	* For cleaning up an ioengine which never made it to init().
	159	*/
	160	void free_ioengine(struct thread_data *td)
8756e4d4	161	{
de890a1e	162	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50	163
de890a1e SL	164	if (td->eo && td->io_ops->options) {
	165	options_free(td->io_ops->options, td->eo);
	166	free(td->eo);
	167	td->eo = NULL;
2992b059	168	}
b990b5c0	169
5f350952 JA	170	if (td->io_ops->dlhandle)
	171	dlclose(td->io_ops->dlhandle);
	172
84585003 JA	173	free(td->io_ops);
84585003 JA	174	td->io_ops = NULL;
b990b5c0	175	}
10ba535a	176
de890a1e SL	177	void close_ioengine(struct thread_data *td)
	178	{
	179	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	180
	181	if (td->io_ops->cleanup) {
	182	td->io_ops->cleanup(td);
	183	td->io_ops->data = NULL;
	184	}
	185
	186	free_ioengine(td);
	187	}
	188
10ba535a JA	189	int td_io_prep(struct thread_data td, struct io_u io_u)
10ba535a JA	190	{
ee56ad50	191	dprint_io_u(io_u, "prep");
7101d9c2 JA	192	fio_ro_check(td, io_u);
7101d9c2 JA	193
4d4e80f2	194	lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9	195
2ba1c290 JA	196	if (td->io_ops->prep) {
	197	int ret = td->io_ops->prep(td, io_u);
	198
	199	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9	200	if (ret)
4d4e80f2	201	unlock_file(td, io_u->file);
2ba1c290 JA	202	return ret;
2ba1c290 JA	203	}
10ba535a JA	204
	205	return 0;
	206	}
	207
e7d2e616	208	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
10ba535a JA	209	struct timespec *t)
10ba535a JA	210	{
ee56ad50	211	int r = 0;
face81b2	212
a05d62b2 YR	213	/*
	214	* For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	215	* server side gets a message from the client
	216	* side that the task is finished, and
	217	* td->done is set to 1 after td_io_commit(). In this case,
	218	* there is no need to reap complete event in server side.
	219	*/
	220	if (td->done)
	221	return 0;
	222
ee56ad50 JA	223	if (min > 0 && td->io_ops->commit) {
ee56ad50 JA	224	r = td->io_ops->commit(td);
face81b2	225	if (r < 0)
ee56ad50	226	goto out;
face81b2	227	}
4950421a JA	228	if (max > td->cur_depth)
	229	max = td->cur_depth;
	230	if (min > max)
	231	max = min;
36167d82	232
ee56ad50	233	r = 0;
4950421a	234	if (max && td->io_ops->getevents)
ee56ad50 JA	235	r = td->io_ops->getevents(td, min, max, t);
ee56ad50 JA	236	out:
422f9e4b RM	237	if (r >= 0) {
	238	/*
	239	* Reflect that our submitted requests were retrieved with
	240	* whatever OS async calls are in the underlying engine.
	241	*/
	242	td->io_u_in_flight -= r;
838bc709	243	io_u_mark_complete(td, r);
422f9e4b	244	} else
7c639b14	245	td_verror(td, r, "get_events");
f3e11d05	246
ee56ad50 JA	247	dprint(FD_IO, "getevents: %d\n", r);
ee56ad50 JA	248	return r;
10ba535a JA	249	}
	250
	251	int td_io_queue(struct thread_data td, struct io_u io_u)
	252	{
7e77dd02 JA	253	int ret;
7e77dd02 JA	254
ee56ad50	255	dprint_io_u(io_u, "queue");
7101d9c2 JA	256	fio_ro_check(td, io_u);
7101d9c2 JA	257
0c6e7517 JA	258	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
	259	io_u->flags \|= IO_U_F_FLIGHT;
	260
d6aed795	261	assert(fio_file_open(io_u->file));
3d7b485f	262
11786802 JA	263	io_u->error = 0;
	264	io_u->resid = 0;
	265
433afcb4	266	if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841	267	if (fio_fill_issue_time(td))
9520ebb9	268	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	269
	270	/*
	271	* only used for iolog
	272	*/
	273	if (td->o.read_iolog_file)
	274	memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa	275	sizeof(struct timeval));
433afcb4 JA	276	}
433afcb4 JA	277
ff58fced	278	if (ddir_rw(io_u->ddir))
755200a3 JA	279	td->io_issues[io_u->ddir]++;
755200a3 JA	280
7e77dd02	281	ret = td->io_ops->queue(td, io_u);
5aeb77df	282
4d4e80f2	283	unlock_file(td, io_u->file);
b2bd2bd9	284
cb211682 JA	285	/*
	286	* Add warning for O_DIRECT so that users have an easier time
	287	* spotting potentially bad alignment. If this triggers for the first
	288	* IO, then it's likely an alignment problem or because the host fs
	289	* does not support O_DIRECT
	290	*/
ff58fced	291	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682	292	td->o.odirect) {
214ac7e0	293
cb211682 JA	294	log_info("fio: first direct IO errored. File system may not "
	295	"support direct IO, or iomem_align= is bad.\n");
	296	}
	297
6eaf09d6	298	if (!td->io_ops->commit \|\| ddir_trim(io_u->ddir)) {
838bc709 JA	299	io_u_mark_submit(td, 1);
	300	io_u_mark_complete(td, 1);
	301	}
	302
d8005759	303	if (ret == FIO_Q_COMPLETED) {
ff58fced	304	if (ddir_rw(io_u->ddir)) {
d8005759 JA	305	io_u_mark_depth(td, 1);
d8005759 JA	306	td->ts.total_io_u[io_u->ddir]++;
6eaf09d6	307	}
d8005759	308	} else if (ret == FIO_Q_QUEUED) {
eb7c8ae2 JA	309	int r;
eb7c8ae2 JA	310
ff58fced	311	if (ddir_rw(io_u->ddir)) {
d8005759 JA	312	td->io_u_queued++;
	313	td->ts.total_io_u[io_u->ddir]++;
	314	}
	315
	316	if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2 JA	317	r = td_io_commit(td);
	318	if (r < 0)
	319	return r;
	320	}
	321	}
cb5ab512	322
433afcb4	323	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841	324	if (fio_fill_issue_time(td))
9520ebb9	325	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	326
	327	/*
	328	* only used for iolog
	329	*/
	330	if (td->o.read_iolog_file)
	331	memcpy(&td->last_issue, &io_u->issue_time,
	332	sizeof(struct timeval));
433afcb4 JA	333	}
433afcb4 JA	334
7e77dd02	335	return ret;
10ba535a	336	}
8c16d840 JA	337
	338	int td_io_init(struct thread_data *td)
	339	{
eeb12160	340	int ret = 0;
8c16d840	341
eeb12160 JA	342	if (td->io_ops->init) {
eeb12160 JA	343	ret = td->io_ops->init(td);
5ec10eaa JA	344	if (ret && td->o.iodepth > 1) {
	345	log_err("fio: io engine init failed. Perhaps try"
	346	" reducing io depth?\n");
	347	}
7c973896 JA	348	if (!td->error)
7c973896 JA	349	td->error = ret;
eeb12160 JA	350	}
	351
	352	return ret;
8c16d840	353	}
755200a3 JA	354
	355	int td_io_commit(struct thread_data *td)
	356	{
f3e11d05 JA	357	int ret;
f3e11d05 JA	358
ee56ad50 JA	359	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
ee56ad50 JA	360
d8005759	361	if (!td->cur_depth \|\| !td->io_u_queued)
e1161c32	362	return 0;
cb5ab512	363
422f9e4b	364	io_u_mark_depth(td, td->io_u_queued);
d8005759	365
f3e11d05 JA	366	if (td->io_ops->commit) {
	367	ret = td->io_ops->commit(td);
	368	if (ret)
	369	td_verror(td, -ret, "io commit");
	370	}
422f9e4b RM	371
	372	/*
	373	* Reflect that events were submitted as async IO requests.
	374	*/
	375	td->io_u_in_flight += td->io_u_queued;
	376	td->io_u_queued = 0;
755200a3 JA	377
	378	return 0;
	379	}
b5af8293 JA	380
	381	int td_io_open_file(struct thread_data td, struct fio_file f)
	382	{
22a57ba8 JA	383	assert(!fio_file_open(f));
	384	assert(f->fd == -1);
	385
413d6693 JA	386	if (td->io_ops->open_file(td, f)) {
	387	if (td->error == EINVAL && td->o.odirect)
	388	log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa JA	389	if (td->error == EMFILE) {
	390	log_err("fio: try reducing/setting openfiles (failed"
	391	" at %u of %u)\n", td->nr_open_files,
	392	td->o.nr_files);
	393	}
413d6693	394
22a57ba8 JA	395	assert(f->fd == -1);
22a57ba8 JA	396	assert(!fio_file_open(f));
413d6693 JA	397	return 1;
	398	}
	399
d5707a35	400	fio_file_reset(f);
d6aed795 JA	401	fio_file_set_open(f);
d6aed795 JA	402	fio_file_clear_closing(f);
c97bd0fa	403	disk_util_inc(f->du);
d5707a35 JA	404
	405	td->nr_open_files++;
	406	get_file(f);
	407
66159828 JA	408	if (f->filetype == FIO_TYPE_PIPE) {
	409	if (td_random(td)) {
	410	log_err("fio: can't seek on pipes (no random io)\n");
	411	goto err;
	412	}
	413	}
	414
413d6693 JA	415	if (td->io_ops->flags & FIO_DISKLESSIO)
	416	goto done;
	417
	418	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	419	goto err;
	420
66159828 JA	421	if (td->o.fadvise_hint &&
66159828 JA	422	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
413d6693 JA	423	int flags;
	424
	425	if (td_random(td))
	426	flags = POSIX_FADV_RANDOM;
	427	else
	428	flags = POSIX_FADV_SEQUENTIAL;
	429
ecc314ba	430	if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693 JA	431	td_verror(td, errno, "fadvise");
	432	goto err;
	433	}
7bb48f84	434	}
a978ba68	435
e116f2b9 JA	436	#ifdef FIO_OS_DIRECTIO
	437	/*
	438	* Some OS's have a distinct call to mark the file non-buffered,
	439	* instead of using O_DIRECT (Solaris)
	440	*/
	441	if (td->o.odirect) {
	442	int ret = fio_set_odirect(f->fd);
	443
	444	if (ret) {
	445	td_verror(td, ret, "fio_set_odirect");
78e51c72	446	log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9 JA	447	goto err;
	448	}
	449	}
	450	#endif
	451
413d6693	452	done:
f29b25a3	453	log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693 JA	454	return 0;
413d6693 JA	455	err:
c97bd0fa	456	disk_util_dec(f->du);
b284075a JA	457	if (td->io_ops->close_file)
b284075a JA	458	td->io_ops->close_file(td, f);
7bb48f84	459	return 1;
b5af8293 JA	460	}
b5af8293 JA	461
6977bcd0	462	int td_io_close_file(struct thread_data td, struct fio_file f)
b5af8293	463	{
d6aed795	464	if (!fio_file_closing(f))
f29b25a3 JA	465	log_file(td, f, FIO_LOG_CLOSE_FILE);
f29b25a3 JA	466
0ad920e7 JA	467	/*
	468	* mark as closing, do real close when last io on it has completed
	469	*/
d6aed795	470	fio_file_set_closing(f);
0ad920e7	471
c97bd0fa	472	disk_util_dec(f->du);
4d4e80f2	473	unlock_file_all(td, f);
29c1349f	474
6977bcd0	475	return put_file(td, f);
b5af8293	476	}
df9c26b1 JA	477
	478	int td_io_get_file_size(struct thread_data td, struct fio_file f)
	479	{
	480	if (!td->io_ops->get_file_size)
	481	return 0;
	482
	483	return td->io_ops->get_file_size(td, f);
	484	}
44f29692	485
0a28ecda	486	static int do_sync_file_range(struct thread_data td, struct fio_file f)
44f29692 JA	487	{
	488	off64_t offset, nbytes;
	489
	490	offset = f->first_write;
	491	nbytes = f->last_write - f->first_write;
	492
3843deb3 JA	493	if (!nbytes)
3843deb3 JA	494	return 0;
44f29692	495
3843deb3	496	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692	497	}
0a28ecda JA	498
	499	int do_io_u_sync(struct thread_data td, struct io_u io_u)
	500	{
	501	int ret;
	502
	503	if (io_u->ddir == DDIR_SYNC) {
	504	ret = fsync(io_u->file->fd);
	505	} else if (io_u->ddir == DDIR_DATASYNC) {
	506	#ifdef FIO_HAVE_FDATASYNC
	507	ret = fdatasync(io_u->file->fd);
	508	#else
	509	ret = io_u->xfer_buflen;
	510	io_u->error = EINVAL;
	511	#endif
	512	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
	513	ret = do_sync_file_range(td, io_u->file);
	514	else {
	515	ret = io_u->xfer_buflen;
	516	io_u->error = EINVAL;
	517	}
	518
cb849a79 JA	519	if (ret < 0)
	520	io_u->error = errno;
	521
0a28ecda JA	522	return ret;
0a28ecda JA	523	}
a5f3027c JA	524
	525	int do_io_u_trim(struct thread_data td, struct io_u io_u)
	526	{
	527	#ifndef FIO_HAVE_TRIM
	528	io_u->error = EINVAL;
ff58fced	529	return 0;
a5f3027c JA	530	#else
	531	struct fio_file *f = io_u->file;
	532	int ret;
	533
c6404a44	534	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c	535	if (!ret)
ff58fced	536	return io_u->xfer_buflen;;
a5f3027c	537
ff58fced JA	538	io_u->error = ret;
ff58fced JA	539	return 0;
a5f3027c JA	540	#endif
a5f3027c JA	541	}
de890a1e SL	542
	543	int fio_show_ioengine_help(const char *engine)
	544	{
	545	struct flist_head *entry;
	546	struct thread_data td;
	547	char *sep;
	548	int ret = 1;
	549
	550	if (!engine \|\| !*engine) {
	551	log_info("Available IO engines:\n");
	552	flist_for_each(entry, &engine_list) {
	553	td.io_ops = flist_entry(entry, struct ioengine_ops,
	554	list);
	555	log_info("\t%s\n", td.io_ops->name);
	556	}
	557	return 0;
	558	}
	559	sep = strchr(engine, ',');
	560	if (sep) {
	561	*sep = 0;
	562	sep++;
	563	}
	564
	565	memset(&td, 0, sizeof(td));
	566
	567	td.io_ops = load_ioengine(&td, engine);
	568	if (!td.io_ops) {
	569	log_info("IO engine %s not found\n", engine);
	570	return 1;
	571	}
	572
	573	if (td.io_ops->options)
	574	ret = show_cmd_help(td.io_ops->options, sep);
	575	else
	576	log_info("IO engine %s has no options\n", td.io_ops->name);
	577
	578	free_ioengine(&td);
	579
	580	return ret;
	581	}