dm mpath: add service time load balancer

This patch adds a service time oriented dynamic load balancer,
dm-service-time, which selects the path with the shortest estimated
service time for the incoming I/O.
The service time is estimated by dividing the in-flight I/O size
by a performance value of each path.

The performance value can be given as a table argument at the table
loading time.  If no performance value is given, all paths are
considered equal.

Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
diff --git a/Documentation/device-mapper/dm-service-time.txt b/Documentation/device-mapper/dm-service-time.txt
new file mode 100644
index 0000000..7d00668
--- /dev/null
+++ b/Documentation/device-mapper/dm-service-time.txt
@@ -0,0 +1,91 @@
+dm-service-time
+===============
+
+dm-service-time is a path selector module for device-mapper targets,
+which selects a path with the shortest estimated service time for
+the incoming I/O.
+
+The service time for each path is estimated by dividing the total size
+of in-flight I/Os on a path with the performance value of the path.
+The performance value is a relative throughput value among all paths
+in a path-group, and it can be specified as a table argument.
+
+The path selector name is 'service-time'.
+
+Table parameters for each path: [<repeat_count> [<relative_throughput>]]
+	<repeat_count>: The number of I/Os to dispatch using the selected
+			path before switching to the next path.
+			If not given, internal default is used.  To check
+			the default value, see the activated table.
+	<relative_throughput>: The relative throughput value of the path
+			among all paths in the path-group.
+			The valid range is 0-100.
+			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.
+
+Status for each path: <status> <fail-count> <in-flight-size> \
+		      <relative_throughput>
+	<status>: 'A' if the path is active, 'F' if the path is failed.
+	<fail-count>: The number of path failures.
+	<in-flight-size>: The size of in-flight I/Os on the path.
+	<relative_throughput>: The relative throughput value of the path
+			among all paths in the path-group.
+
+
+Algorithm
+=========
+
+dm-service-time adds the I/O size to 'in-flight-size' when the I/O is
+dispatched and substracts when completed.
+Basically, dm-service-time selects a path having minimum service time
+which is calculated by:
+
+	('in-flight-size' + 'size-of-incoming-io') / 'relative_throughput'
+
+However, some optimizations below are used to reduce the calculation
+as much as possible.
+
+	1. If the paths have the same 'relative_throughput', skip
+	   the division and just compare the 'in-flight-size'.
+
+	2. If the paths have the same 'in-flight-size', skip the division
+	   and just compare the 'relative_throughput'.
+
+	3. If some paths have non-zero 'relative_throughput' and others
+	   have zero 'relative_throughput', ignore those paths with zero
+	   'relative_throughput'.
+
+If such optimizations can't be applied, calculate service time, and
+compare service time.
+If calculated service time is equal, the path having maximum
+'relative_throughput' may be better.  So compare 'relative_throughput'
+then.
+
+
+Examples
+========
+In case that 2 paths (sda and sdb) are used with repeat_count == 128
+and sda has an average throughput 1GB/s and sdb has 4GB/s,
+'relative_throughput' value may be '1' for sda and '4' for sdb.
+
+# echo "0 10 multipath 0 0 1 1 service-time 0 2 2 8:0 128 1 8:16 128 4" \
+  dmsetup create test
+#
+# dmsetup table
+test: 0 10 multipath 0 0 1 1 service-time 0 2 2 8:0 128 1 8:16 128 4
+#
+# dmsetup status
+test: 0 10 multipath 2 0 0 0 1 1 E 0 2 2 8:0 A 0 0 1 8:16 A 0 0 4
+
+
+Or '2' for sda and '8' for sdb would be also true.
+
+# echo "0 10 multipath 0 0 1 1 service-time 0 2 2 8:0 128 2 8:16 128 8" \
+  dmsetup create test
+#
+# dmsetup table
+test: 0 10 multipath 0 0 1 1 service-time 0 2 2 8:0 128 2 8:16 128 8
+#
+# dmsetup status
+test: 0 10 multipath 2 0 0 0 1 1 E 0 2 2 8:0 A 0 0 2 8:16 A 0 0 8
diff --git a/drivers/md/Kconfig b/drivers/md/Kconfig
index 3b311d2..09f93fa 100644
--- a/drivers/md/Kconfig
+++ b/drivers/md/Kconfig
@@ -258,6 +258,16 @@
 
 	  If unsure, say N.
 
+config DM_MULTIPATH_ST
+	tristate "I/O Path Selector based on the service time"
+	depends on DM_MULTIPATH
+	---help---
+	  This path selector is a dynamic load balancer which selects
+	  the path expected to complete the incoming I/O in the shortest
+	  time.
+
+	  If unsure, say N.
+
 config DM_DELAY
 	tristate "I/O delaying target (EXPERIMENTAL)"
 	depends on BLK_DEV_DM && EXPERIMENTAL
diff --git a/drivers/md/Makefile b/drivers/md/Makefile
index ff9f545..dade52f 100644
--- a/drivers/md/Makefile
+++ b/drivers/md/Makefile
@@ -37,6 +37,7 @@
 obj-$(CONFIG_DM_DELAY)		+= dm-delay.o
 obj-$(CONFIG_DM_MULTIPATH)	+= dm-multipath.o dm-round-robin.o
 obj-$(CONFIG_DM_MULTIPATH_QL)	+= dm-queue-length.o
+obj-$(CONFIG_DM_MULTIPATH_ST)	+= dm-service-time.o
 obj-$(CONFIG_DM_SNAPSHOT)	+= dm-snapshot.o
 obj-$(CONFIG_DM_MIRROR)		+= dm-mirror.o dm-log.o dm-region-hash.o
 obj-$(CONFIG_DM_ZERO)		+= dm-zero.o
diff --git a/drivers/md/dm-service-time.c b/drivers/md/dm-service-time.c
new file mode 100644
index 0000000..cfa668f
--- /dev/null
+++ b/drivers/md/dm-service-time.c
@@ -0,0 +1,339 @@
+/*
+ * 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"
+
+#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	"0.2.0"
+
+struct selector {
+	struct list_head valid_paths;
+	struct list_head failed_paths;
+};
+
+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);
+	}
+
+	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;
+		}
+	}
+
+	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;
+
+	/*
+	 * 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", &repeat_count) != 1)) {
+		*error = "service-time ps: invalid repeat count";
+		return -EINVAL;
+	}
+
+	if ((argc == 2) &&
+	    (sscanf(argv[1], "%u", &relative_throughput) != 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;
+
+	list_add_tail(&pi->list, &s->valid_paths);
+
+	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;
+
+	list_move(&pi->list, &s->failed_paths);
+}
+
+static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
+{
+	struct selector *s = ps->context;
+	struct path_info *pi = path->pscontext;
+
+	list_move_tail(&pi->list, &s->valid_paths);
+
+	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,
+				      unsigned *repeat_count, size_t nr_bytes)
+{
+	struct selector *s = ps->context;
+	struct path_info *pi = NULL, *best = NULL;
+
+	if (list_empty(&s->valid_paths))
+		return NULL;
+
+	/* Change preferred (first in list) path to evenly balance. */
+	list_move_tail(s->valid_paths.next, &s->valid_paths);
+
+	list_for_each_entry(pi, &s->valid_paths, list)
+		if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
+			best = pi;
+
+	if (!best)
+		return NULL;
+
+	*repeat_count = best->repeat_count;
+
+	return best->path;
+}
+
+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)
+{
+	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");