[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");

	/*
	 * For some external engines (like C++ ones) it is not that trivial
	 * to provide a non-static ionengine structure that we can reference.
	 * Instead we call a method which allocates the required ioengine
	 * structure.
	 */
	if (!ops) {
		get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");

		if (get_ioengine)
			get_ioengine(&ops);
	}

	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,
		    const 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));

	/*
	 * If using a write iolog, store this entry.
	 */
	log_io_u(td, io_u);

	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(acct_ddir(io_u))) {
		td->io_issues[acct_ddir(io_u)]++;
		td->io_issue_bytes[acct_ddir(io_u)] += io_u->xfer_buflen;
	}

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

	unlock_file(td, io_u->file);

	if (ret == FIO_Q_BUSY && ddir_rw(acct_ddir(io_u))) {
		td->io_issues[acct_ddir(io_u)]--;
		td->io_issue_bytes[acct_ddir(io_u)] -= io_u->xfer_buflen;
	}

	/*
	 * If an error was seen and the io engine didn't propagate it
	 * back to 'td', do so.
	 */
	if (io_u->error && !td->error)
		td_verror(td, io_u->error, "td_io_queue");

	/*
	 * 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 || io_u->ddir == DDIR_TRIM) {
		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;
	}

	if (!ret && (td->io_ops->flags & FIO_NOIO))
		td->flags |= TD_F_NOIO;

	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(td, 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);

	if (td->o.file_lock_mode != FILE_LOCK_NONE)
		unlock_file_all(td, f);

	return put_file(td, f);
}

int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
{
	if (td->io_ops->unlink_file)
		return td->io_ops->unlink_file(td, f);
	else
		return unlink(f->file_name);
}

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(const 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(const 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 CONFIG_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(const 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
	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>
	15	#include <string.h>
	16	#include <dlfcn.h>
	17	#include <fcntl.h>
	18	#include <assert.h>
	19
	20	#include "fio.h"
	21	#include "diskutil.h"
	22
	23	static FLIST_HEAD(engine_list);
	24
	25	static int check_engine_ops(struct ioengine_ops *ops)
	26	{
	27	if (ops->version != FIO_IOOPS_VERSION) {
	28	log_err("bad ioops version %d (want %d)\n", ops->version,
	29	FIO_IOOPS_VERSION);
	30	return 1;
	31	}
	32
	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;
	43
	44	if (!ops->event) {
	45	log_err("%s: no event handler\n", ops->name);
	46	return 1;
	47	}
	48	if (!ops->getevents) {
	49	log_err("%s: no getevents handler\n", ops->name);
	50	return 1;
	51	}
	52	if (!ops->queue) {
	53	log_err("%s: no queue handler\n", ops->name);
	54	return 1;
	55	}
	56
	57	return 0;
	58	}
	59
	60	void unregister_ioengine(struct ioengine_ops *ops)
	61	{
	62	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
	63	flist_del(&ops->list);
	64	INIT_FLIST_HEAD(&ops->list);
	65	}
	66
	67	void register_ioengine(struct ioengine_ops *ops)
	68	{
	69	dprint(FD_IO, "ioengine %s registered\n", ops->name);
	70	INIT_FLIST_HEAD(&ops->list);
	71	flist_add_tail(&ops->list, &engine_list);
	72	}
	73
	74	static struct ioengine_ops find_ioengine(const char name)
	75	{
	76	struct ioengine_ops *ops;
	77	struct flist_head *entry;
	78
	79	flist_for_each(entry, &engine_list) {
	80	ops = flist_entry(entry, struct ioengine_ops, list);
	81	if (!strcmp(name, ops->name))
	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
	94	dprint(FD_IO, "dload engine %s\n", engine_lib);
	95
	96	dlerror();
	97	dlhandle = dlopen(engine_lib, RTLD_LAZY);
	98	if (!dlhandle) {
	99	td_vmsg(td, -1, dlerror(), "dlopen");
	100	return NULL;
	101	}
	102
	103	/*
	104	* Unlike the included modules, external engines should have a
	105	* non-static ioengine structure that we can reference.
	106	*/
	107	ops = dlsym(dlhandle, engine_lib);
	108	if (!ops)
	109	ops = dlsym(dlhandle, "ioengine");
	110
	111	/*
	112	* For some external engines (like C++ ones) it is not that trivial
	113	* to provide a non-static ionengine structure that we can reference.
	114	* Instead we call a method which allocates the required ioengine
	115	* structure.
	116	*/
	117	if (!ops) {
	118	get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
	119
	120	if (get_ioengine)
	121	get_ioengine(&ops);
	122	}
	123
	124	if (!ops) {
	125	td_vmsg(td, -1, dlerror(), "dlsym");
	126	dlclose(dlhandle);
	127	return NULL;
	128	}
	129
	130	ops->dlhandle = dlhandle;
	131	return ops;
	132	}
	133
	134	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	135	{
	136	struct ioengine_ops ops, ret;
	137	char engine[16];
	138
	139	dprint(FD_IO, "load ioengine %s\n", name);
	140
	141	strncpy(engine, name, sizeof(engine) - 1);
	142
	143	/*
	144	* linux libaio has alias names, so convert to what we want
	145	*/
	146	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3))
	147	strcpy(engine, "libaio");
	148
	149	ops = find_ioengine(engine);
	150	if (!ops)
	151	ops = dlopen_ioengine(td, name);
	152
	153	if (!ops) {
	154	log_err("fio: engine %s not loadable\n", name);
	155	return NULL;
	156	}
	157
	158	/*
	159	* Check that the required methods are there.
	160	*/
	161	if (check_engine_ops(ops))
	162	return NULL;
	163
	164	ret = malloc(sizeof(*ret));
	165	memcpy(ret, ops, sizeof(*ret));
	166	ret->data = NULL;
	167
	168	return ret;
	169	}
	170
	171	/*
	172	* For cleaning up an ioengine which never made it to init().
	173	*/
	174	void free_ioengine(struct thread_data *td)
	175	{
	176	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
	177
	178	if (td->eo && td->io_ops->options) {
	179	options_free(td->io_ops->options, td->eo);
	180	free(td->eo);
	181	td->eo = NULL;
	182	}
	183
	184	if (td->io_ops->dlhandle)
	185	dlclose(td->io_ops->dlhandle);
	186
	187	free(td->io_ops);
	188	td->io_ops = NULL;
	189	}
	190
	191	void close_ioengine(struct thread_data *td)
	192	{
	193	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	194
	195	if (td->io_ops->cleanup) {
	196	td->io_ops->cleanup(td);
	197	td->io_ops->data = NULL;
	198	}
	199
	200	free_ioengine(td);
	201	}
	202
	203	int td_io_prep(struct thread_data td, struct io_u io_u)
	204	{
	205	dprint_io_u(io_u, "prep");
	206	fio_ro_check(td, io_u);
	207
	208	lock_file(td, io_u->file, io_u->ddir);
	209
	210	if (td->io_ops->prep) {
	211	int ret = td->io_ops->prep(td, io_u);
	212
	213	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
	214	if (ret)
	215	unlock_file(td, io_u->file);
	216	return ret;
	217	}
	218
	219	return 0;
	220	}
	221
	222	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
	223	const struct timespec *t)
	224	{
	225	int r = 0;
	226
	227	/*
	228	* For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	229	* server side gets a message from the client
	230	* side that the task is finished, and
	231	* td->done is set to 1 after td_io_commit(). In this case,
	232	* there is no need to reap complete event in server side.
	233	*/
	234	if (td->done)
	235	return 0;
	236
	237	if (min > 0 && td->io_ops->commit) {
	238	r = td->io_ops->commit(td);
	239	if (r < 0)
	240	goto out;
	241	}
	242	if (max > td->cur_depth)
	243	max = td->cur_depth;
	244	if (min > max)
	245	max = min;
	246
	247	r = 0;
	248	if (max && td->io_ops->getevents)
	249	r = td->io_ops->getevents(td, min, max, t);
	250	out:
	251	if (r >= 0) {
	252	/*
	253	* Reflect that our submitted requests were retrieved with
	254	* whatever OS async calls are in the underlying engine.
	255	*/
	256	td->io_u_in_flight -= r;
	257	io_u_mark_complete(td, r);
	258	} else
	259	td_verror(td, r, "get_events");
	260
	261	dprint(FD_IO, "getevents: %d\n", r);
	262	return r;
	263	}
	264
	265	int td_io_queue(struct thread_data td, struct io_u io_u)
	266	{
	267	int ret;
	268
	269	dprint_io_u(io_u, "queue");
	270	fio_ro_check(td, io_u);
	271
	272	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
	273	io_u->flags \|= IO_U_F_FLIGHT;
	274
	275	assert(fio_file_open(io_u->file));
	276
	277	/*
	278	* If using a write iolog, store this entry.
	279	*/
	280	log_io_u(td, io_u);
	281
	282	io_u->error = 0;
	283	io_u->resid = 0;
	284
	285	if (td->io_ops->flags & FIO_SYNCIO) {
	286	if (fio_fill_issue_time(td))
	287	fio_gettime(&io_u->issue_time, NULL);
	288
	289	/*
	290	* only used for iolog
	291	*/
	292	if (td->o.read_iolog_file)
	293	memcpy(&td->last_issue, &io_u->issue_time,
	294	sizeof(struct timeval));
	295	}
	296
	297	if (ddir_rw(acct_ddir(io_u))) {
	298	td->io_issues[acct_ddir(io_u)]++;
	299	td->io_issue_bytes[acct_ddir(io_u)] += io_u->xfer_buflen;
	300	}
	301
	302	ret = td->io_ops->queue(td, io_u);
	303
	304	unlock_file(td, io_u->file);
	305
	306	if (ret == FIO_Q_BUSY && ddir_rw(acct_ddir(io_u))) {
	307	td->io_issues[acct_ddir(io_u)]--;
	308	td->io_issue_bytes[acct_ddir(io_u)] -= io_u->xfer_buflen;
	309	}
	310
	311	/*
	312	* If an error was seen and the io engine didn't propagate it
	313	* back to 'td', do so.
	314	*/
	315	if (io_u->error && !td->error)
	316	td_verror(td, io_u->error, "td_io_queue");
	317
	318	/*
	319	* Add warning for O_DIRECT so that users have an easier time
	320	* spotting potentially bad alignment. If this triggers for the first
	321	* IO, then it's likely an alignment problem or because the host fs
	322	* does not support O_DIRECT
	323	*/
	324	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
	325	td->o.odirect) {
	326
	327	log_info("fio: first direct IO errored. File system may not "
	328	"support direct IO, or iomem_align= is bad.\n");
	329	}
	330
	331	if (!td->io_ops->commit \|\| io_u->ddir == DDIR_TRIM) {
	332	io_u_mark_submit(td, 1);
	333	io_u_mark_complete(td, 1);
	334	}
	335
	336	if (ret == FIO_Q_COMPLETED) {
	337	if (ddir_rw(io_u->ddir)) {
	338	io_u_mark_depth(td, 1);
	339	td->ts.total_io_u[io_u->ddir]++;
	340	}
	341	} else if (ret == FIO_Q_QUEUED) {
	342	int r;
	343
	344	if (ddir_rw(io_u->ddir)) {
	345	td->io_u_queued++;
	346	td->ts.total_io_u[io_u->ddir]++;
	347	}
	348
	349	if (td->io_u_queued >= td->o.iodepth_batch) {
	350	r = td_io_commit(td);
	351	if (r < 0)
	352	return r;
	353	}
	354	}
	355
	356	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
	357	if (fio_fill_issue_time(td))
	358	fio_gettime(&io_u->issue_time, NULL);
	359
	360	/*
	361	* only used for iolog
	362	*/
	363	if (td->o.read_iolog_file)
	364	memcpy(&td->last_issue, &io_u->issue_time,
	365	sizeof(struct timeval));
	366	}
	367
	368	return ret;
	369	}
	370
	371	int td_io_init(struct thread_data *td)
	372	{
	373	int ret = 0;
	374
	375	if (td->io_ops->init) {
	376	ret = td->io_ops->init(td);
	377	if (ret && td->o.iodepth > 1) {
	378	log_err("fio: io engine init failed. Perhaps try"
	379	" reducing io depth?\n");
	380	}
	381	if (!td->error)
	382	td->error = ret;
	383	}
	384
	385	if (!ret && (td->io_ops->flags & FIO_NOIO))
	386	td->flags \|= TD_F_NOIO;
	387
	388	return ret;
	389	}
	390
	391	int td_io_commit(struct thread_data *td)
	392	{
	393	int ret;
	394
	395	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
	396
	397	if (!td->cur_depth \|\| !td->io_u_queued)
	398	return 0;
	399
	400	io_u_mark_depth(td, td->io_u_queued);
	401
	402	if (td->io_ops->commit) {
	403	ret = td->io_ops->commit(td);
	404	if (ret)
	405	td_verror(td, -ret, "io commit");
	406	}
	407
	408	/*
	409	* Reflect that events were submitted as async IO requests.
	410	*/
	411	td->io_u_in_flight += td->io_u_queued;
	412	td->io_u_queued = 0;
	413
	414	return 0;
	415	}
	416
	417	int td_io_open_file(struct thread_data td, struct fio_file f)
	418	{
	419	assert(!fio_file_open(f));
	420	assert(f->fd == -1);
	421
	422	if (td->io_ops->open_file(td, f)) {
	423	if (td->error == EINVAL && td->o.odirect)
	424	log_err("fio: destination does not support O_DIRECT\n");
	425	if (td->error == EMFILE) {
	426	log_err("fio: try reducing/setting openfiles (failed"
	427	" at %u of %u)\n", td->nr_open_files,
	428	td->o.nr_files);
	429	}
	430
	431	assert(f->fd == -1);
	432	assert(!fio_file_open(f));
	433	return 1;
	434	}
	435
	436	fio_file_reset(td, f);
	437	fio_file_set_open(f);
	438	fio_file_clear_closing(f);
	439	disk_util_inc(f->du);
	440
	441	td->nr_open_files++;
	442	get_file(f);
	443
	444	if (f->filetype == FIO_TYPE_PIPE) {
	445	if (td_random(td)) {
	446	log_err("fio: can't seek on pipes (no random io)\n");
	447	goto err;
	448	}
	449	}
	450
	451	if (td->io_ops->flags & FIO_DISKLESSIO)
	452	goto done;
	453
	454	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	455	goto err;
	456
	457	if (td->o.fadvise_hint &&
	458	(f->filetype == FIO_TYPE_BD \|\| f->filetype == FIO_TYPE_FILE)) {
	459	int flags;
	460
	461	if (td_random(td))
	462	flags = POSIX_FADV_RANDOM;
	463	else
	464	flags = POSIX_FADV_SEQUENTIAL;
	465
	466	if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
	467	td_verror(td, errno, "fadvise");
	468	goto err;
	469	}
	470	}
	471
	472	#ifdef FIO_OS_DIRECTIO
	473	/*
	474	* Some OS's have a distinct call to mark the file non-buffered,
	475	* instead of using O_DIRECT (Solaris)
	476	*/
	477	if (td->o.odirect) {
	478	int ret = fio_set_odirect(f->fd);
	479
	480	if (ret) {
	481	td_verror(td, ret, "fio_set_odirect");
	482	log_err("fio: the file system does not seem to support direct IO\n");
	483	goto err;
	484	}
	485	}
	486	#endif
	487
	488	done:
	489	log_file(td, f, FIO_LOG_OPEN_FILE);
	490	return 0;
	491	err:
	492	disk_util_dec(f->du);
	493	if (td->io_ops->close_file)
	494	td->io_ops->close_file(td, f);
	495	return 1;
	496	}
	497
	498	int td_io_close_file(struct thread_data td, struct fio_file f)
	499	{
	500	if (!fio_file_closing(f))
	501	log_file(td, f, FIO_LOG_CLOSE_FILE);
	502
	503	/*
	504	* mark as closing, do real close when last io on it has completed
	505	*/
	506	fio_file_set_closing(f);
	507
	508	disk_util_dec(f->du);
	509
	510	if (td->o.file_lock_mode != FILE_LOCK_NONE)
	511	unlock_file_all(td, f);
	512
	513	return put_file(td, f);
	514	}
	515
	516	int td_io_unlink_file(struct thread_data td, struct fio_file f)
	517	{
	518	if (td->io_ops->unlink_file)
	519	return td->io_ops->unlink_file(td, f);
	520	else
	521	return unlink(f->file_name);
	522	}
	523
	524	int td_io_get_file_size(struct thread_data td, struct fio_file f)
	525	{
	526	if (!td->io_ops->get_file_size)
	527	return 0;
	528
	529	return td->io_ops->get_file_size(td, f);
	530	}
	531
	532	static int do_sync_file_range(const struct thread_data *td,
	533	struct fio_file *f)
	534	{
	535	off64_t offset, nbytes;
	536
	537	offset = f->first_write;
	538	nbytes = f->last_write - f->first_write;
	539
	540	if (!nbytes)
	541	return 0;
	542
	543	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
	544	}
	545
	546	int do_io_u_sync(const struct thread_data td, struct io_u io_u)
	547	{
	548	int ret;
	549
	550	if (io_u->ddir == DDIR_SYNC) {
	551	ret = fsync(io_u->file->fd);
	552	} else if (io_u->ddir == DDIR_DATASYNC) {
	553	#ifdef CONFIG_FDATASYNC
	554	ret = fdatasync(io_u->file->fd);
	555	#else
	556	ret = io_u->xfer_buflen;
	557	io_u->error = EINVAL;
	558	#endif
	559	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
	560	ret = do_sync_file_range(td, io_u->file);
	561	else {
	562	ret = io_u->xfer_buflen;
	563	io_u->error = EINVAL;
	564	}
	565
	566	if (ret < 0)
	567	io_u->error = errno;
	568
	569	return ret;
	570	}
	571
	572	int do_io_u_trim(const struct thread_data td, struct io_u io_u)
	573	{
	574	#ifndef FIO_HAVE_TRIM
	575	io_u->error = EINVAL;
	576	return 0;
	577	#else
	578	struct fio_file *f = io_u->file;
	579	int ret;
	580
	581	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
	582	if (!ret)
	583	return io_u->xfer_buflen;
	584
	585	io_u->error = ret;
	586	return 0;
	587	#endif
	588	}
	589
	590	int fio_show_ioengine_help(const char *engine)
	591	{
	592	struct flist_head *entry;
	593	struct thread_data td;
	594	char *sep;
	595	int ret = 1;
	596
	597	if (!engine \|\| !*engine) {
	598	log_info("Available IO engines:\n");
	599	flist_for_each(entry, &engine_list) {
	600	td.io_ops = flist_entry(entry, struct ioengine_ops,
	601	list);
	602	log_info("\t%s\n", td.io_ops->name);
	603	}
	604	return 0;
	605	}
	606	sep = strchr(engine, ',');
	607	if (sep) {
	608	*sep = 0;
	609	sep++;
	610	}
	611
	612	memset(&td, 0, sizeof(td));
	613
	614	td.io_ops = load_ioengine(&td, engine);
	615	if (!td.io_ops) {
	616	log_info("IO engine %s not found\n", engine);
	617	return 1;
	618	}
	619
	620	if (td.io_ops->options)
	621	ret = show_cmd_help(td.io_ops->options, sep);
	622	else
	623	log_info("IO engine %s has no options\n", td.io_ops->name);
	624
	625	free_ioengine(&td);
	626
	627	return ret;
	628	}