[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 bool 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 true;
	}

	if (!ops->queue) {
		log_err("%s: no queue handler\n", ops->name);
		return true;
	}

	/*
	 * sync engines only need a ->queue()
	 */
	if (ops->flags & FIO_SYNCIO)
		return false;

	if (!ops->event || !ops->getevents) {
		log_err("%s: no event/getevents handler\n", ops->name);
		return true;
	}

	return false;
}

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;
	}

	td->io_ops_dlhandle = dlhandle;
	return ops;
}

struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
{
	struct ioengine_ops *ops;
	char engine[64];

	dprint(FD_IO, "load ioengine %s\n", name);

	engine[sizeof(engine) - 1] = '\0';
	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)) {
		dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
		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;

	return ops;
}

/*
 * 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);

	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)
{
	const enum fio_ddir ddir = acct_ddir(io_u);
	unsigned long buflen = io_u->xfer_buflen;
	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_set(td, io_u, 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_ioengine_flagged(td, 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(ddir)) {
		td->io_issues[ddir]++;
		td->io_issue_bytes[ddir] += buflen;
		td->rate_io_issue_bytes[ddir] += buflen;
	}

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

	unlock_file(td, io_u->file);

	if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
		td->io_issues[ddir]--;
		td->io_issue_bytes[ddir] -= buflen;
		td->rate_io_issue_bytes[ddir] -= buflen;
		io_u_clear(td, io_u, IO_U_F_FLIGHT);
	}

	/*
	 * 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. Try "
			 "setting direct=0.\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;

		td->io_u_queued++;

		if (ddir_rw(io_u->ddir))
			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_ioengine_flagged(td, 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));
	}

	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)
			log_err("fio: io engine %s init failed.%s\n",
				td->io_ops->name,
				td->o.iodepth > 1 ?
				" Perhaps try reducing io depth?" : "");
		else
			td->io_ops_init = 1;
		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);
	assert(td->io_ops->open_file);

	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_ioengine_flagged(td, FIO_DISKLESSIO))
		goto done;

	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
		goto err;

	if (td->o.fadvise_hint != F_ADV_NONE &&
	    (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
		int flags;

		if (td->o.fadvise_hint == F_ADV_TYPE) {
			if (td_random(td))
				flags = POSIX_FADV_RANDOM;
			else
				flags = POSIX_FADV_SEQUENTIAL;
		} else if (td->o.fadvise_hint == F_ADV_RANDOM)
			flags = POSIX_FADV_RANDOM;
		else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
			flags = POSIX_FADV_SEQUENTIAL;
		else {
			log_err("fio: unknown fadvise type %d\n",
							td->o.fadvise_hint);
			flags = POSIX_FADV_NORMAL;
		}

		if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
			td_verror(td, errno, "fadvise");
			goto err;
		}
	}
#ifdef FIO_HAVE_STREAMID
	if (td->o.fadvise_stream &&
	    (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
		off_t stream = td->o.fadvise_stream;

		if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
			td_verror(td, errno, "fadvise streamid");
			goto err;
		}
	}
#endif

#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");
			if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
				log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
			} else {
				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 {
		int ret;

		ret = unlink(f->file_name);
		if (ret < 0)
			return errno;

		return 0;
	}
}

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);
}

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

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

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

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

	if (io_ops->options)
		ret = show_cmd_help(io_ops->options, sep);
	else
		log_info("IO engine %s has no options\n", 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
1cc954ba	25	static bool check_engine_ops(struct ioengine_ops *ops)
8c16d840	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);
1cc954ba	30	return true;
5f350952 JA	31	}
5f350952 JA	32
36167d82 JA	33	if (!ops->queue) {
36167d82 JA	34	log_err("%s: no queue handler\n", ops->name);
1cc954ba	35	return true;
36167d82 JA	36	}
	37
	38	/*
	39	* sync engines only need a ->queue()
	40	*/
	41	if (ops->flags & FIO_SYNCIO)
1cc954ba	42	return false;
5ec10eaa	43
1cc954ba JA	44	if (!ops->event \|\| !ops->getevents) {
	45	log_err("%s: no event/getevents handler\n", ops->name);
	46	return true;
8c16d840	47	}
5ec10eaa	48
1cc954ba	49	return false;
8c16d840 JA	50	}
8c16d840 JA	51
5f350952	52	void unregister_ioengine(struct ioengine_ops *ops)
ebac4655	53	{
ee56ad50	54	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1 JA	55	flist_del(&ops->list);
01743ee1 JA	56	INIT_FLIST_HEAD(&ops->list);
5f350952 JA	57	}
5f350952 JA	58
b2fdda43	59	void register_ioengine(struct ioengine_ops *ops)
5f350952	60	{
ee56ad50	61	dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1 JA	62	INIT_FLIST_HEAD(&ops->list);
01743ee1 JA	63	flist_add_tail(&ops->list, &engine_list);
5f350952 JA	64	}
	65
	66	static struct ioengine_ops find_ioengine(const char name)
	67	{
	68	struct ioengine_ops *ops;
01743ee1	69	struct flist_head *entry;
ebac4655	70
01743ee1 JA	71	flist_for_each(entry, &engine_list) {
01743ee1 JA	72	ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8	73	if (!strcmp(name, ops->name))
5f350952 JA	74	return ops;
	75	}
	76
	77	return NULL;
	78	}
	79
	80	static struct ioengine_ops dlopen_ioengine(struct thread_data td,
	81	const char *engine_lib)
	82	{
	83	struct ioengine_ops *ops;
	84	void *dlhandle;
	85
ee56ad50 JA	86	dprint(FD_IO, "dload engine %s\n", engine_lib);
ee56ad50 JA	87
2866c82d JA	88	dlerror();
2866c82d JA	89	dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8	90	if (!dlhandle) {
e1161c32	91	td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8 JA	92	return NULL;
d4dbaaa8 JA	93	}
8756e4d4	94
da51c050 JA	95	/*
	96	* Unlike the included modules, external engines should have a
	97	* non-static ioengine structure that we can reference.
	98	*/
0abea0b7 DM	99	ops = dlsym(dlhandle, engine_lib);
	100	if (!ops)
	101	ops = dlsym(dlhandle, "ioengine");
a8075704 DG	102
	103	/*
	104	* For some external engines (like C++ ones) it is not that trivial
	105	* to provide a non-static ionengine structure that we can reference.
	106	* Instead we call a method which allocates the required ioengine
	107	* structure.
	108	*/
	109	if (!ops) {
	110	get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
	111
	112	if (get_ioengine)
	113	get_ioengine(&ops);
	114	}
	115
d4dbaaa8	116	if (!ops) {
e1161c32	117	td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8 JA	118	dlclose(dlhandle);
	119	return NULL;
	120	}
8756e4d4	121
565e784d	122	td->io_ops_dlhandle = dlhandle;
5f350952 JA	123	return ops;
	124	}
	125
	126	struct ioengine_ops load_ioengine(struct thread_data td, const char *name)
	127	{
565e784d	128	struct ioengine_ops *ops;
aa2b823c	129	char engine[64];
5f350952	130
ee56ad50 JA	131	dprint(FD_IO, "load ioengine %s\n", name);
ee56ad50 JA	132
087d0ed0	133	engine[sizeof(engine) - 1] = '\0';
5f350952 JA	134	strncpy(engine, name, sizeof(engine) - 1);
	135
	136	/*
	137	* linux libaio has alias names, so convert to what we want
	138	*/
966fcbd4	139	if (!strncmp(engine, "linuxaio", 8) \|\| !strncmp(engine, "aio", 3)) {
966fcbd4	140	dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
5f350952	141	strcpy(engine, "libaio");
966fcbd4	142	}
5f350952 JA	143
	144	ops = find_ioengine(engine);
	145	if (!ops)
	146	ops = dlopen_ioengine(td, name);
	147
	148	if (!ops) {
	149	log_err("fio: engine %s not loadable\n", name);
b902ceb5 JA	150	return NULL;
	151	}
	152
8c16d840 JA	153	/*
	154	* Check that the required methods are there.
	155	*/
5f350952	156	if (check_engine_ops(ops))
8c16d840	157	return NULL;
8c16d840	158
565e784d	159	return ops;
8756e4d4 JA	160	}
8756e4d4 JA	161
de890a1e SL	162	/*
	163	* For cleaning up an ioengine which never made it to init().
	164	*/
	165	void free_ioengine(struct thread_data *td)
8756e4d4	166	{
de890a1e	167	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50	168
de890a1e SL	169	if (td->eo && td->io_ops->options) {
	170	options_free(td->io_ops->options, td->eo);
	171	free(td->eo);
	172	td->eo = NULL;
2992b059	173	}
b990b5c0	174
565e784d TK	175	if (td->io_ops_dlhandle)
565e784d TK	176	dlclose(td->io_ops_dlhandle);
5f350952	177
84585003	178	td->io_ops = NULL;
b990b5c0	179	}
10ba535a	180
de890a1e SL	181	void close_ioengine(struct thread_data *td)
	182	{
	183	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
	184
	185	if (td->io_ops->cleanup) {
	186	td->io_ops->cleanup(td);
565e784d	187	td->io_ops_data = NULL;
de890a1e SL	188	}
	189
	190	free_ioengine(td);
	191	}
	192
10ba535a JA	193	int td_io_prep(struct thread_data td, struct io_u io_u)
10ba535a JA	194	{
ee56ad50	195	dprint_io_u(io_u, "prep");
7101d9c2 JA	196	fio_ro_check(td, io_u);
7101d9c2 JA	197
4d4e80f2	198	lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9	199
2ba1c290 JA	200	if (td->io_ops->prep) {
	201	int ret = td->io_ops->prep(td, io_u);
	202
	203	dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9	204	if (ret)
4d4e80f2	205	unlock_file(td, io_u->file);
2ba1c290 JA	206	return ret;
2ba1c290 JA	207	}
10ba535a JA	208
	209	return 0;
	210	}
	211
e7d2e616	212	int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece	213	const struct timespec *t)
10ba535a	214	{
ee56ad50	215	int r = 0;
face81b2	216
a05d62b2 YR	217	/*
	218	* For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
	219	* server side gets a message from the client
	220	* side that the task is finished, and
	221	* td->done is set to 1 after td_io_commit(). In this case,
	222	* there is no need to reap complete event in server side.
	223	*/
	224	if (td->done)
	225	return 0;
	226
ee56ad50 JA	227	if (min > 0 && td->io_ops->commit) {
ee56ad50 JA	228	r = td->io_ops->commit(td);
face81b2	229	if (r < 0)
ee56ad50	230	goto out;
face81b2	231	}
4950421a JA	232	if (max > td->cur_depth)
	233	max = td->cur_depth;
	234	if (min > max)
	235	max = min;
36167d82	236
ee56ad50	237	r = 0;
4950421a	238	if (max && td->io_ops->getevents)
ee56ad50 JA	239	r = td->io_ops->getevents(td, min, max, t);
ee56ad50 JA	240	out:
422f9e4b RM	241	if (r >= 0) {
422f9e4b RM	242	/*
3fd9efbc	243	* Reflect that our submitted requests were retrieved with
422f9e4b RM	244	* whatever OS async calls are in the underlying engine.
	245	*/
	246	td->io_u_in_flight -= r;
838bc709	247	io_u_mark_complete(td, r);
422f9e4b	248	} else
7c639b14	249	td_verror(td, r, "get_events");
f3e11d05	250
ee56ad50 JA	251	dprint(FD_IO, "getevents: %d\n", r);
ee56ad50 JA	252	return r;
10ba535a JA	253	}
	254
	255	int td_io_queue(struct thread_data td, struct io_u io_u)
	256	{
a9da8ab2 JA	257	const enum fio_ddir ddir = acct_ddir(io_u);
a9da8ab2 JA	258	unsigned long buflen = io_u->xfer_buflen;
7e77dd02 JA	259	int ret;
7e77dd02 JA	260
ee56ad50	261	dprint_io_u(io_u, "queue");
7101d9c2 JA	262	fio_ro_check(td, io_u);
7101d9c2 JA	263
0c6e7517	264	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
1651e431	265	io_u_set(td, io_u, IO_U_F_FLIGHT);
0c6e7517	266
d6aed795	267	assert(fio_file_open(io_u->file));
3d7b485f	268
bcd5abfa JA	269	/*
	270	* If using a write iolog, store this entry.
	271	*/
	272	log_io_u(td, io_u);
	273
11786802 JA	274	io_u->error = 0;
	275	io_u->resid = 0;
	276
9b87f09b	277	if (td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841	278	if (fio_fill_issue_time(td))
9520ebb9	279	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	280
	281	/*
	282	* only used for iolog
	283	*/
	284	if (td->o.read_iolog_file)
	285	memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa	286	sizeof(struct timeval));
433afcb4 JA	287	}
433afcb4 JA	288
a9da8ab2 JA	289	if (ddir_rw(ddir)) {
	290	td->io_issues[ddir]++;
	291	td->io_issue_bytes[ddir] += buflen;
50a8ce86	292	td->rate_io_issue_bytes[ddir] += buflen;
74d6277f	293	}
755200a3	294
7e77dd02	295	ret = td->io_ops->queue(td, io_u);
5aeb77df	296
4d4e80f2	297	unlock_file(td, io_u->file);
b2bd2bd9	298
a9da8ab2 JA	299	if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
	300	td->io_issues[ddir]--;
	301	td->io_issue_bytes[ddir] -= buflen;
50a8ce86	302	td->rate_io_issue_bytes[ddir] -= buflen;
871467d9	303	io_u_clear(td, io_u, IO_U_F_FLIGHT);
cd8a19e6 JA	304	}
cd8a19e6 JA	305
39a43d34 JA	306	/*
	307	* If an error was seen and the io engine didn't propagate it
	308	* back to 'td', do so.
	309	*/
	310	if (io_u->error && !td->error)
	311	td_verror(td, io_u->error, "td_io_queue");
	312
cb211682 JA	313	/*
	314	* Add warning for O_DIRECT so that users have an easier time
	315	* spotting potentially bad alignment. If this triggers for the first
	316	* IO, then it's likely an alignment problem or because the host fs
	317	* does not support O_DIRECT
	318	*/
ff58fced	319	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682	320	td->o.odirect) {
214ac7e0	321
cb211682	322	log_info("fio: first direct IO errored. File system may not "
44cc9c8f JA	323	"support direct IO, or iomem_align= is bad. Try "
44cc9c8f JA	324	"setting direct=0.\n");
cb211682 JA	325	}
cb211682 JA	326
77b451cc	327	if (!td->io_ops->commit \|\| io_u->ddir == DDIR_TRIM) {
838bc709 JA	328	io_u_mark_submit(td, 1);
	329	io_u_mark_complete(td, 1);
	330	}
	331
d8005759	332	if (ret == FIO_Q_COMPLETED) {
ff58fced	333	if (ddir_rw(io_u->ddir)) {
d8005759 JA	334	io_u_mark_depth(td, 1);
d8005759 JA	335	td->ts.total_io_u[io_u->ddir]++;
6eaf09d6	336	}
d8005759	337	} else if (ret == FIO_Q_QUEUED) {
eb7c8ae2 JA	338	int r;
eb7c8ae2 JA	339
e6727cbd JA	340	td->io_u_queued++;
	341
	342	if (ddir_rw(io_u->ddir))
d8005759	343	td->ts.total_io_u[io_u->ddir]++;
d8005759 JA	344
d8005759 JA	345	if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2 JA	346	r = td_io_commit(td);
	347	if (r < 0)
	348	return r;
	349	}
	350	}
cb5ab512	351
9b87f09b	352	if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841	353	if (fio_fill_issue_time(td))
9520ebb9	354	fio_gettime(&io_u->issue_time, NULL);
d0c15328 JA	355
	356	/*
	357	* only used for iolog
	358	*/
	359	if (td->o.read_iolog_file)
	360	memcpy(&td->last_issue, &io_u->issue_time,
	361	sizeof(struct timeval));
433afcb4 JA	362	}
433afcb4 JA	363
7e77dd02	364	return ret;
10ba535a	365	}
8c16d840 JA	366
	367	int td_io_init(struct thread_data *td)
	368	{
eeb12160	369	int ret = 0;
8c16d840	370
eeb12160 JA	371	if (td->io_ops->init) {
eeb12160 JA	372	ret = td->io_ops->init(td);
356ef1a1 TK	373	if (ret)
	374	log_err("fio: io engine %s init failed.%s\n",
	375	td->io_ops->name,
	376	td->o.iodepth > 1 ?
	377	" Perhaps try reducing io depth?" : "");
	378	else
	379	td->io_ops_init = 1;
7c973896 JA	380	if (!td->error)
7c973896 JA	381	td->error = ret;
eeb12160 JA	382	}
	383
	384	return ret;
8c16d840	385	}
755200a3 JA	386
	387	int td_io_commit(struct thread_data *td)
	388	{
f3e11d05 JA	389	int ret;
f3e11d05 JA	390
ee56ad50 JA	391	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
ee56ad50 JA	392
d8005759	393	if (!td->cur_depth \|\| !td->io_u_queued)
e1161c32	394	return 0;
cb5ab512	395
3fd9efbc	396	io_u_mark_depth(td, td->io_u_queued);
d8005759	397
f3e11d05 JA	398	if (td->io_ops->commit) {
	399	ret = td->io_ops->commit(td);
	400	if (ret)
	401	td_verror(td, -ret, "io commit");
	402	}
3fd9efbc	403
422f9e4b RM	404	/*
	405	* Reflect that events were submitted as async IO requests.
	406	*/
	407	td->io_u_in_flight += td->io_u_queued;
	408	td->io_u_queued = 0;
755200a3 JA	409
	410	return 0;
	411	}
b5af8293 JA	412
	413	int td_io_open_file(struct thread_data td, struct fio_file f)
	414	{
22a57ba8 JA	415	assert(!fio_file_open(f));
22a57ba8 JA	416	assert(f->fd == -1);
966fcbd4	417	assert(td->io_ops->open_file);
22a57ba8	418
413d6693 JA	419	if (td->io_ops->open_file(td, f)) {
	420	if (td->error == EINVAL && td->o.odirect)
	421	log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa JA	422	if (td->error == EMFILE) {
	423	log_err("fio: try reducing/setting openfiles (failed"
	424	" at %u of %u)\n", td->nr_open_files,
	425	td->o.nr_files);
	426	}
413d6693	427
22a57ba8 JA	428	assert(f->fd == -1);
22a57ba8 JA	429	assert(!fio_file_open(f));
413d6693 JA	430	return 1;
	431	}
	432
33c48814	433	fio_file_reset(td, f);
d6aed795 JA	434	fio_file_set_open(f);
d6aed795 JA	435	fio_file_clear_closing(f);
c97bd0fa	436	disk_util_inc(f->du);
d5707a35 JA	437
	438	td->nr_open_files++;
	439	get_file(f);
	440
66159828 JA	441	if (f->filetype == FIO_TYPE_PIPE) {
	442	if (td_random(td)) {
	443	log_err("fio: can't seek on pipes (no random io)\n");
	444	goto err;
	445	}
	446	}
	447
9b87f09b	448	if (td_ioengine_flagged(td, FIO_DISKLESSIO))
413d6693 JA	449	goto done;
	450
	451	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
	452	goto err;
	453
ecb2083d	454	if (td->o.fadvise_hint != F_ADV_NONE &&
686fbd31	455	(f->filetype == FIO_TYPE_BLOCK \|\| f->filetype == FIO_TYPE_FILE)) {
413d6693 JA	456	int flags;
413d6693 JA	457
ecb2083d JA	458	if (td->o.fadvise_hint == F_ADV_TYPE) {
	459	if (td_random(td))
	460	flags = POSIX_FADV_RANDOM;
	461	else
	462	flags = POSIX_FADV_SEQUENTIAL;
	463	} else if (td->o.fadvise_hint == F_ADV_RANDOM)
413d6693	464	flags = POSIX_FADV_RANDOM;
ecb2083d	465	else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
413d6693	466	flags = POSIX_FADV_SEQUENTIAL;
ecb2083d JA	467	else {
	468	log_err("fio: unknown fadvise type %d\n",
	469	td->o.fadvise_hint);
	470	flags = POSIX_FADV_NORMAL;
	471	}
413d6693	472
ecc314ba	473	if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693 JA	474	td_verror(td, errno, "fadvise");
	475	goto err;
	476	}
7bb48f84	477	}
37659335 JA	478	#ifdef FIO_HAVE_STREAMID
37659335 JA	479	if (td->o.fadvise_stream &&
686fbd31	480	(f->filetype == FIO_TYPE_BLOCK \|\| f->filetype == FIO_TYPE_FILE)) {
37659335 JA	481	off_t stream = td->o.fadvise_stream;
	482
	483	if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
	484	td_verror(td, errno, "fadvise streamid");
	485	goto err;
	486	}
	487	}
	488	#endif
a978ba68	489
e116f2b9 JA	490	#ifdef FIO_OS_DIRECTIO
	491	/*
	492	* Some OS's have a distinct call to mark the file non-buffered,
	493	* instead of using O_DIRECT (Solaris)
	494	*/
	495	if (td->o.odirect) {
	496	int ret = fio_set_odirect(f->fd);
	497
	498	if (ret) {
	499	td_verror(td, ret, "fio_set_odirect");
42f1ee68 SZ	500	if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
	501	log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
	502	} else {
	503	log_err("fio: the file system does not seem to support direct IO\n");
	504	}
	505
e116f2b9 JA	506	goto err;
	507	}
	508	}
	509	#endif
	510
413d6693	511	done:
f29b25a3	512	log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693 JA	513	return 0;
413d6693 JA	514	err:
c97bd0fa	515	disk_util_dec(f->du);
b284075a JA	516	if (td->io_ops->close_file)
b284075a JA	517	td->io_ops->close_file(td, f);
7bb48f84	518	return 1;
b5af8293 JA	519	}
b5af8293 JA	520
6977bcd0	521	int td_io_close_file(struct thread_data td, struct fio_file f)
b5af8293	522	{
d6aed795	523	if (!fio_file_closing(f))
f29b25a3 JA	524	log_file(td, f, FIO_LOG_CLOSE_FILE);
f29b25a3 JA	525
0ad920e7 JA	526	/*
	527	* mark as closing, do real close when last io on it has completed
	528	*/
d6aed795	529	fio_file_set_closing(f);
0ad920e7	530
c97bd0fa	531	disk_util_dec(f->du);
cba5243b JA	532
	533	if (td->o.file_lock_mode != FILE_LOCK_NONE)
	534	unlock_file_all(td, f);
29c1349f	535
6977bcd0	536	return put_file(td, f);
b5af8293	537	}
df9c26b1	538
38ef9c90 CF	539	int td_io_unlink_file(struct thread_data td, struct fio_file f)
	540	{
	541	if (td->io_ops->unlink_file)
	542	return td->io_ops->unlink_file(td, f);
2442c935 JA	543	else {
	544	int ret;
	545
	546	ret = unlink(f->file_name);
	547	if (ret < 0)
	548	return errno;
	549
	550	return 0;
	551	}
38ef9c90 CF	552	}
38ef9c90 CF	553
df9c26b1 JA	554	int td_io_get_file_size(struct thread_data td, struct fio_file f)
	555	{
	556	if (!td->io_ops->get_file_size)
	557	return 0;
	558
	559	return td->io_ops->get_file_size(td, f);
	560	}
44f29692	561
de890a1e SL	562	int fio_show_ioengine_help(const char *engine)
	563	{
	564	struct flist_head *entry;
	565	struct thread_data td;
755dcbbd	566	struct ioengine_ops *io_ops;
de890a1e SL	567	char *sep;
	568	int ret = 1;
	569
	570	if (!engine \|\| !*engine) {
	571	log_info("Available IO engines:\n");
	572	flist_for_each(entry, &engine_list) {
755dcbbd TK	573	io_ops = flist_entry(entry, struct ioengine_ops, list);
755dcbbd TK	574	log_info("\t%s\n", io_ops->name);
de890a1e SL	575	}
	576	return 0;
	577	}
	578	sep = strchr(engine, ',');
	579	if (sep) {
	580	*sep = 0;
	581	sep++;
	582	}
	583
	584	memset(&td, 0, sizeof(td));
	585
755dcbbd TK	586	io_ops = load_ioengine(&td, engine);
755dcbbd TK	587	if (!io_ops) {
de890a1e SL	588	log_info("IO engine %s not found\n", engine);
	589	return 1;
	590	}
	591
755dcbbd TK	592	if (io_ops->options)
755dcbbd TK	593	ret = show_cmd_help(io_ops->options, sep);
de890a1e	594	else
755dcbbd	595	log_info("IO engine %s has no options\n", io_ops->name);
de890a1e SL	596
	597	free_ioengine(&td);
	598
	599	return ret;
	600	}