[linux-2.6-block.git] / drivers / md / dm-ps-service-time.c

/*
 * Copyright (C) 2007-2009 NEC Corporation.  All Rights Reserved.
 *
 * Module Author: Kiyoshi Ueda
 *
 * This file is released under the GPL.
 *
 * Throughput oriented path selector.
 */

#include "dm.h"
#include "dm-path-selector.h"

#include <linux/slab.h>
#include <linux/module.h>

#define DM_MSG_PREFIX	"multipath service-time"
#define ST_MIN_IO	1
#define ST_MAX_RELATIVE_THROUGHPUT	100
#define ST_MAX_RELATIVE_THROUGHPUT_SHIFT	7
#define ST_MAX_INFLIGHT_SIZE	((size_t)-1 >> ST_MAX_RELATIVE_THROUGHPUT_SHIFT)
#define ST_VERSION	"0.3.0"

struct selector {
	struct list_head valid_paths;
	struct list_head failed_paths;
	spinlock_t lock;
};

struct path_info {
	struct list_head list;
	struct dm_path *path;
	unsigned repeat_count;
	unsigned relative_throughput;
	atomic_t in_flight_size;	/* Total size of in-flight I/Os */
};

static struct selector *alloc_selector(void)
{
	struct selector *s = kmalloc(sizeof(*s), GFP_KERNEL);

	if (s) {
		INIT_LIST_HEAD(&s->valid_paths);
		INIT_LIST_HEAD(&s->failed_paths);
		spin_lock_init(&s->lock);
	}

	return s;
}

static int st_create(struct path_selector *ps, unsigned argc, char **argv)
{
	struct selector *s = alloc_selector();

	if (!s)
		return -ENOMEM;

	ps->context = s;
	return 0;
}

static void free_paths(struct list_head *paths)
{
	struct path_info *pi, *next;

	list_for_each_entry_safe(pi, next, paths, list) {
		list_del(&pi->list);
		kfree(pi);
	}
}

static void st_destroy(struct path_selector *ps)
{
	struct selector *s = ps->context;

	free_paths(&s->valid_paths);
	free_paths(&s->failed_paths);
	kfree(s);
	ps->context = NULL;
}

static int st_status(struct path_selector *ps, struct dm_path *path,
		     status_type_t type, char *result, unsigned maxlen)
{
	unsigned sz = 0;
	struct path_info *pi;

	if (!path)
		DMEMIT("0 ");
	else {
		pi = path->pscontext;

		switch (type) {
		case STATUSTYPE_INFO:
			DMEMIT("%d %u ", atomic_read(&pi->in_flight_size),
			       pi->relative_throughput);
			break;
		case STATUSTYPE_TABLE:
			DMEMIT("%u %u ", pi->repeat_count,
			       pi->relative_throughput);
			break;
		case STATUSTYPE_IMA:
			result[0] = '\0';
			break;
		}
	}

	return sz;
}

static int st_add_path(struct path_selector *ps, struct dm_path *path,
		       int argc, char **argv, char **error)
{
	struct selector *s = ps->context;
	struct path_info *pi;
	unsigned repeat_count = ST_MIN_IO;
	unsigned relative_throughput = 1;
	char dummy;
	unsigned long flags;

	/*
	 * Arguments: [<repeat_count> [<relative_throughput>]]
	 * 	<repeat_count>: The number of I/Os before switching path.
	 * 			If not given, default (ST_MIN_IO) is used.
	 * 	<relative_throughput>: The relative throughput value of
	 *			the path among all paths in the path-group.
	 * 			The valid range: 0-<ST_MAX_RELATIVE_THROUGHPUT>
	 *			If not given, minimum value '1' is used.
	 *			If '0' is given, the path isn't selected while
	 * 			other paths having a positive value are
	 * 			available.
	 */
	if (argc > 2) {
		*error = "service-time ps: incorrect number of arguments";
		return -EINVAL;
	}

	if (argc && (sscanf(argv[0], "%u%c", &repeat_count, &dummy) != 1)) {
		*error = "service-time ps: invalid repeat count";
		return -EINVAL;
	}

	if (repeat_count > 1) {
		DMWARN_LIMIT("repeat_count > 1 is deprecated, using 1 instead");
		repeat_count = 1;
	}

	if ((argc == 2) &&
	    (sscanf(argv[1], "%u%c", &relative_throughput, &dummy) != 1 ||
	     relative_throughput > ST_MAX_RELATIVE_THROUGHPUT)) {
		*error = "service-time ps: invalid relative_throughput value";
		return -EINVAL;
	}

	/* allocate the path */
	pi = kmalloc(sizeof(*pi), GFP_KERNEL);
	if (!pi) {
		*error = "service-time ps: Error allocating path context";
		return -ENOMEM;
	}

	pi->path = path;
	pi->repeat_count = repeat_count;
	pi->relative_throughput = relative_throughput;
	atomic_set(&pi->in_flight_size, 0);

	path->pscontext = pi;

	spin_lock_irqsave(&s->lock, flags);
	list_add_tail(&pi->list, &s->valid_paths);
	spin_unlock_irqrestore(&s->lock, flags);

	return 0;
}

static void st_fail_path(struct path_selector *ps, struct dm_path *path)
{
	struct selector *s = ps->context;
	struct path_info *pi = path->pscontext;
	unsigned long flags;

	spin_lock_irqsave(&s->lock, flags);
	list_move(&pi->list, &s->failed_paths);
	spin_unlock_irqrestore(&s->lock, flags);
}

static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
{
	struct selector *s = ps->context;
	struct path_info *pi = path->pscontext;
	unsigned long flags;

	spin_lock_irqsave(&s->lock, flags);
	list_move_tail(&pi->list, &s->valid_paths);
	spin_unlock_irqrestore(&s->lock, flags);

	return 0;
}

/*
 * Compare the estimated service time of 2 paths, pi1 and pi2,
 * for the incoming I/O.
 *
 * Returns:
 * < 0 : pi1 is better
 * 0   : no difference between pi1 and pi2
 * > 0 : pi2 is better
 *
 * Description:
 * Basically, the service time is estimated by:
 *     ('pi->in-flight-size' + 'incoming') / 'pi->relative_throughput'
 * To reduce the calculation, some optimizations are made.
 * (See comments inline)
 */
static int st_compare_load(struct path_info *pi1, struct path_info *pi2,
			   size_t incoming)
{
	size_t sz1, sz2, st1, st2;

	sz1 = atomic_read(&pi1->in_flight_size);
	sz2 = atomic_read(&pi2->in_flight_size);

	/*
	 * Case 1: Both have same throughput value. Choose less loaded path.
	 */
	if (pi1->relative_throughput == pi2->relative_throughput)
		return sz1 - sz2;

	/*
	 * Case 2a: Both have same load. Choose higher throughput path.
	 * Case 2b: One path has no throughput value. Choose the other one.
	 */
	if (sz1 == sz2 ||
	    !pi1->relative_throughput || !pi2->relative_throughput)
		return pi2->relative_throughput - pi1->relative_throughput;

	/*
	 * Case 3: Calculate service time. Choose faster path.
	 *         Service time using pi1:
	 *             st1 = (sz1 + incoming) / pi1->relative_throughput
	 *         Service time using pi2:
	 *             st2 = (sz2 + incoming) / pi2->relative_throughput
	 *
	 *         To avoid the division, transform the expression to use
	 *         multiplication.
	 *         Because ->relative_throughput > 0 here, if st1 < st2,
	 *         the expressions below are the same meaning:
	 *             (sz1 + incoming) / pi1->relative_throughput <
	 *                 (sz2 + incoming) / pi2->relative_throughput
	 *             (sz1 + incoming) * pi2->relative_throughput <
	 *                 (sz2 + incoming) * pi1->relative_throughput
	 *         So use the later one.
	 */
	sz1 += incoming;
	sz2 += incoming;
	if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE ||
		     sz2 >= ST_MAX_INFLIGHT_SIZE)) {
		/*
		 * Size may be too big for multiplying pi->relative_throughput
		 * and overflow.
		 * To avoid the overflow and mis-selection, shift down both.
		 */
		sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
		sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
	}
	st1 = sz1 * pi2->relative_throughput;
	st2 = sz2 * pi1->relative_throughput;
	if (st1 != st2)
		return st1 - st2;

	/*
	 * Case 4: Service time is equal. Choose higher throughput path.
	 */
	return pi2->relative_throughput - pi1->relative_throughput;
}

static struct dm_path *st_select_path(struct path_selector *ps, size_t nr_bytes)
{
	struct selector *s = ps->context;
	struct path_info *pi = NULL, *best = NULL;
	struct dm_path *ret = NULL;
	unsigned long flags;

	spin_lock_irqsave(&s->lock, flags);
	if (list_empty(&s->valid_paths))
		goto out;

	list_for_each_entry(pi, &s->valid_paths, list)
		if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
			best = pi;

	if (!best)
		goto out;

	/* Move most recently used to least preferred to evenly balance. */
	list_move_tail(&best->list, &s->valid_paths);

	ret = best->path;
out:
	spin_unlock_irqrestore(&s->lock, flags);
	return ret;
}

static int st_start_io(struct path_selector *ps, struct dm_path *path,
		       size_t nr_bytes)
{
	struct path_info *pi = path->pscontext;

	atomic_add(nr_bytes, &pi->in_flight_size);

	return 0;
}

static int st_end_io(struct path_selector *ps, struct dm_path *path,
		     size_t nr_bytes, u64 start_time)
{
	struct path_info *pi = path->pscontext;

	atomic_sub(nr_bytes, &pi->in_flight_size);

	return 0;
}

static struct path_selector_type st_ps = {
	.name		= "service-time",
	.module		= THIS_MODULE,
	.table_args	= 2,
	.info_args	= 2,
	.create		= st_create,
	.destroy	= st_destroy,
	.status		= st_status,
	.add_path	= st_add_path,
	.fail_path	= st_fail_path,
	.reinstate_path	= st_reinstate_path,
	.select_path	= st_select_path,
	.start_io	= st_start_io,
	.end_io		= st_end_io,
};

static int __init dm_st_init(void)
{
	int r = dm_register_path_selector(&st_ps);

	if (r < 0)
		DMERR("register failed %d", r);

	DMINFO("version " ST_VERSION " loaded");

	return r;
}

static void __exit dm_st_exit(void)
{
	int r = dm_unregister_path_selector(&st_ps);

	if (r < 0)
		DMERR("unregister failed %d", r);
}

module_init(dm_st_init);
module_exit(dm_st_exit);

MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
MODULE_LICENSE("GPL");
Commit	Line	Data
f392ba88 KU	1	/*
	2	* Copyright (C) 2007-2009 NEC Corporation. All Rights Reserved.
	3	*
	4	* Module Author: Kiyoshi Ueda
	5	*
	6	* This file is released under the GPL.
	7	*
	8	* Throughput oriented path selector.
	9	*/
	10
	11	#include "dm.h"
	12	#include "dm-path-selector.h"
	13
5a0e3ad6	14	#include <linux/slab.h>
056075c7	15	#include <linux/module.h>
5a0e3ad6	16
f392ba88 KU	17	#define DM_MSG_PREFIX "multipath service-time"
	18	#define ST_MIN_IO 1
	19	#define ST_MAX_RELATIVE_THROUGHPUT 100
	20	#define ST_MAX_RELATIVE_THROUGHPUT_SHIFT 7
	21	#define ST_MAX_INFLIGHT_SIZE ((size_t)-1 >> ST_MAX_RELATIVE_THROUGHPUT_SHIFT)
21136f89	22	#define ST_VERSION "0.3.0"
f392ba88 KU	23
	24	struct selector {
	25	struct list_head valid_paths;
	26	struct list_head failed_paths;
9659f811	27	spinlock_t lock;
f392ba88 KU	28	};
	29
	30	struct path_info {
	31	struct list_head list;
	32	struct dm_path *path;
	33	unsigned repeat_count;
	34	unsigned relative_throughput;
	35	atomic_t in_flight_size; /* Total size of in-flight I/Os */
	36	};
	37
	38	static struct selector *alloc_selector(void)
	39	{
	40	struct selector s = kmalloc(sizeof(s), GFP_KERNEL);
	41
	42	if (s) {
	43	INIT_LIST_HEAD(&s->valid_paths);
	44	INIT_LIST_HEAD(&s->failed_paths);
9659f811	45	spin_lock_init(&s->lock);
f392ba88 KU	46	}
	47
	48	return s;
	49	}
	50
	51	static int st_create(struct path_selector ps, unsigned argc, char *argv)
	52	{
	53	struct selector *s = alloc_selector();
	54
	55	if (!s)
	56	return -ENOMEM;
	57
	58	ps->context = s;
	59	return 0;
	60	}
	61
	62	static void free_paths(struct list_head *paths)
	63	{
	64	struct path_info pi, next;
	65
	66	list_for_each_entry_safe(pi, next, paths, list) {
	67	list_del(&pi->list);
	68	kfree(pi);
	69	}
	70	}
	71
	72	static void st_destroy(struct path_selector *ps)
	73	{
	74	struct selector *s = ps->context;
	75
	76	free_paths(&s->valid_paths);
	77	free_paths(&s->failed_paths);
	78	kfree(s);
	79	ps->context = NULL;
	80	}
	81
	82	static int st_status(struct path_selector ps, struct dm_path path,
	83	status_type_t type, char *result, unsigned maxlen)
	84	{
	85	unsigned sz = 0;
	86	struct path_info *pi;
	87
	88	if (!path)
	89	DMEMIT("0 ");
	90	else {
	91	pi = path->pscontext;
	92
	93	switch (type) {
	94	case STATUSTYPE_INFO:
	95	DMEMIT("%d %u ", atomic_read(&pi->in_flight_size),
	96	pi->relative_throughput);
	97	break;
	98	case STATUSTYPE_TABLE:
	99	DMEMIT("%u %u ", pi->repeat_count,
	100	pi->relative_throughput);
	101	break;
8ec45662 TS	102	case STATUSTYPE_IMA:
	103	result[0] = '\0';
	104	break;
f392ba88 KU	105	}
	106	}
	107
	108	return sz;
	109	}
	110
	111	static int st_add_path(struct path_selector ps, struct dm_path path,
	112	int argc, char argv, char error)
	113	{
	114	struct selector *s = ps->context;
	115	struct path_info *pi;
	116	unsigned repeat_count = ST_MIN_IO;
	117	unsigned relative_throughput = 1;
31998ef1	118	char dummy;
9659f811	119	unsigned long flags;
f392ba88 KU	120
	121	/*
	122	* Arguments: [<repeat_count> [<relative_throughput>]]
	123	* <repeat_count>: The number of I/Os before switching path.
	124	* If not given, default (ST_MIN_IO) is used.
	125	* <relative_throughput>: The relative throughput value of
	126	* the path among all paths in the path-group.
	127	* The valid range: 0-<ST_MAX_RELATIVE_THROUGHPUT>
	128	* If not given, minimum value '1' is used.
	129	* If '0' is given, the path isn't selected while
	130	* other paths having a positive value are
	131	* available.
	132	*/
	133	if (argc > 2) {
	134	*error = "service-time ps: incorrect number of arguments";
	135	return -EINVAL;
	136	}
	137
31998ef1	138	if (argc && (sscanf(argv[0], "%u%c", &repeat_count, &dummy) != 1)) {
f392ba88 KU	139	*error = "service-time ps: invalid repeat count";
	140	return -EINVAL;
	141	}
	142
21136f89 MS	143	if (repeat_count > 1) {
	144	DMWARN_LIMIT("repeat_count > 1 is deprecated, using 1 instead");
	145	repeat_count = 1;
	146	}
	147
f392ba88	148	if ((argc == 2) &&
31998ef1	149	(sscanf(argv[1], "%u%c", &relative_throughput, &dummy) != 1 \|\|
f392ba88 KU	150	relative_throughput > ST_MAX_RELATIVE_THROUGHPUT)) {
	151	*error = "service-time ps: invalid relative_throughput value";
	152	return -EINVAL;
	153	}
	154
	155	/* allocate the path */
	156	pi = kmalloc(sizeof(*pi), GFP_KERNEL);
	157	if (!pi) {
	158	*error = "service-time ps: Error allocating path context";
	159	return -ENOMEM;
	160	}
	161
	162	pi->path = path;
	163	pi->repeat_count = repeat_count;
	164	pi->relative_throughput = relative_throughput;
	165	atomic_set(&pi->in_flight_size, 0);
	166
	167	path->pscontext = pi;
	168
9659f811	169	spin_lock_irqsave(&s->lock, flags);
f392ba88	170	list_add_tail(&pi->list, &s->valid_paths);
9659f811	171	spin_unlock_irqrestore(&s->lock, flags);
f392ba88 KU	172
	173	return 0;
	174	}
	175
	176	static void st_fail_path(struct path_selector ps, struct dm_path path)
	177	{
	178	struct selector *s = ps->context;
	179	struct path_info *pi = path->pscontext;
9659f811	180	unsigned long flags;
f392ba88	181
9659f811	182	spin_lock_irqsave(&s->lock, flags);
f392ba88	183	list_move(&pi->list, &s->failed_paths);
9659f811	184	spin_unlock_irqrestore(&s->lock, flags);
f392ba88 KU	185	}
	186
	187	static int st_reinstate_path(struct path_selector ps, struct dm_path path)
	188	{
	189	struct selector *s = ps->context;
	190	struct path_info *pi = path->pscontext;
9659f811	191	unsigned long flags;
f392ba88	192
9659f811	193	spin_lock_irqsave(&s->lock, flags);
f392ba88	194	list_move_tail(&pi->list, &s->valid_paths);
9659f811	195	spin_unlock_irqrestore(&s->lock, flags);
f392ba88 KU	196
	197	return 0;
	198	}
	199
	200	/*
	201	* Compare the estimated service time of 2 paths, pi1 and pi2,
	202	* for the incoming I/O.
	203	*
	204	* Returns:
	205	* < 0 : pi1 is better
	206	* 0 : no difference between pi1 and pi2
	207	* > 0 : pi2 is better
	208	*
	209	* Description:
	210	* Basically, the service time is estimated by:
	211	* ('pi->in-flight-size' + 'incoming') / 'pi->relative_throughput'
	212	* To reduce the calculation, some optimizations are made.
	213	* (See comments inline)
	214	*/
	215	static int st_compare_load(struct path_info pi1, struct path_info pi2,
	216	size_t incoming)
	217	{
	218	size_t sz1, sz2, st1, st2;
	219
	220	sz1 = atomic_read(&pi1->in_flight_size);
	221	sz2 = atomic_read(&pi2->in_flight_size);
	222
	223	/*
	224	* Case 1: Both have same throughput value. Choose less loaded path.
	225	*/
	226	if (pi1->relative_throughput == pi2->relative_throughput)
	227	return sz1 - sz2;
	228
	229	/*
	230	* Case 2a: Both have same load. Choose higher throughput path.
	231	* Case 2b: One path has no throughput value. Choose the other one.
	232	*/
	233	if (sz1 == sz2 \|\|
	234	!pi1->relative_throughput \|\| !pi2->relative_throughput)
	235	return pi2->relative_throughput - pi1->relative_throughput;
	236
	237	/*
	238	* Case 3: Calculate service time. Choose faster path.
	239	* Service time using pi1:
	240	* st1 = (sz1 + incoming) / pi1->relative_throughput
	241	* Service time using pi2:
	242	* st2 = (sz2 + incoming) / pi2->relative_throughput
	243	*
	244	* To avoid the division, transform the expression to use
	245	* multiplication.
	246	* Because ->relative_throughput > 0 here, if st1 < st2,
	247	* the expressions below are the same meaning:
	248	* (sz1 + incoming) / pi1->relative_throughput <
	249	* (sz2 + incoming) / pi2->relative_throughput
	250	* (sz1 + incoming) * pi2->relative_throughput <
	251	* (sz2 + incoming) * pi1->relative_throughput
	252	* So use the later one.
	253	*/
	254	sz1 += incoming;
	255	sz2 += incoming;
	256	if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE \|\|
	257	sz2 >= ST_MAX_INFLIGHT_SIZE)) {
	258	/*
	259	* Size may be too big for multiplying pi->relative_throughput
260	* and overflow.
261	* To avoid the overflow and mis-selection, shift down both.
262	*/
263	sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
264	sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
265	}
266	st1 = sz1 * pi2->relative_throughput;
267	st2 = sz2 * pi1->relative_throughput;
268	if (st1 != st2)
269	return st1 - st2;
270
271	/*
272	* Case 4: Service time is equal. Choose higher throughput path.
273	*/
274	return pi2->relative_throughput - pi1->relative_throughput;
275	}
276
90a4323c	277	static struct dm_path st_select_path(struct path_selector ps, size_t nr_bytes)
f392ba88 KU	278	{
	279	struct selector *s = ps->context;
	280	struct path_info pi = NULL, best = NULL;
9659f811 MS	281	struct dm_path *ret = NULL;
9659f811 MS	282	unsigned long flags;
f392ba88	283
9659f811	284	spin_lock_irqsave(&s->lock, flags);
f392ba88	285	if (list_empty(&s->valid_paths))
9659f811	286	goto out;
f392ba88	287
f392ba88 KU	288	list_for_each_entry(pi, &s->valid_paths, list)
	289	if (!best \|\| (st_compare_load(pi, best, nr_bytes) < 0))
	290	best = pi;
	291
	292	if (!best)
9659f811	293	goto out;
f392ba88	294
f2042605 KK	295	/* Move most recently used to least preferred to evenly balance. */
	296	list_move_tail(&best->list, &s->valid_paths);
	297
9659f811 MS	298	ret = best->path;
	299	out:
	300	spin_unlock_irqrestore(&s->lock, flags);
	301	return ret;
f392ba88 KU	302	}
	303
	304	static int st_start_io(struct path_selector ps, struct dm_path path,
	305	size_t nr_bytes)
	306	{
	307	struct path_info *pi = path->pscontext;
	308
	309	atomic_add(nr_bytes, &pi->in_flight_size);
	310
	311	return 0;
	312	}
	313
	314	static int st_end_io(struct path_selector ps, struct dm_path path,
087615bf	315	size_t nr_bytes, u64 start_time)
f392ba88 KU	316	{
	317	struct path_info *pi = path->pscontext;
	318
	319	atomic_sub(nr_bytes, &pi->in_flight_size);
	320
	321	return 0;
	322	}
	323
	324	static struct path_selector_type st_ps = {
	325	.name = "service-time",
	326	.module = THIS_MODULE,
	327	.table_args = 2,
	328	.info_args = 2,
	329	.create = st_create,
	330	.destroy = st_destroy,
	331	.status = st_status,
	332	.add_path = st_add_path,
	333	.fail_path = st_fail_path,
	334	.reinstate_path = st_reinstate_path,
	335	.select_path = st_select_path,
	336	.start_io = st_start_io,
	337	.end_io = st_end_io,
	338	};
	339
	340	static int __init dm_st_init(void)
	341	{
	342	int r = dm_register_path_selector(&st_ps);
	343
	344	if (r < 0)
	345	DMERR("register failed %d", r);
	346
	347	DMINFO("version " ST_VERSION " loaded");
	348
	349	return r;
	350	}
	351
	352	static void __exit dm_st_exit(void)
	353	{
	354	int r = dm_unregister_path_selector(&st_ps);
	355
	356	if (r < 0)
	357	DMERR("unregister failed %d", r);
	358	}
	359
	360	module_init(dm_st_init);
	361	module_exit(dm_st_exit);
	362
	363	MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
	364	MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
	365	MODULE_LICENSE("GPL");