2 * Copyright (C) 2007-2009 NEC Corporation. All Rights Reserved.
4 * Module Author: Kiyoshi Ueda
6 * This file is released under the GPL.
8 * Throughput oriented path selector.
12 #include "dm-path-selector.h"
14 #include <linux/slab.h>
15 #include <linux/module.h>
17 #define DM_MSG_PREFIX "multipath service-time"
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)
22 #define ST_VERSION "0.3.0"
25 struct list_head valid_paths;
26 struct list_head failed_paths;
31 struct list_head list;
33 unsigned repeat_count;
34 unsigned relative_throughput;
35 atomic_t in_flight_size; /* Total size of in-flight I/Os */
38 static struct selector *alloc_selector(void)
40 struct selector *s = kmalloc(sizeof(*s), GFP_KERNEL);
43 INIT_LIST_HEAD(&s->valid_paths);
44 INIT_LIST_HEAD(&s->failed_paths);
45 spin_lock_init(&s->lock);
51 static int st_create(struct path_selector *ps, unsigned argc, char **argv)
53 struct selector *s = alloc_selector();
62 static void free_paths(struct list_head *paths)
64 struct path_info *pi, *next;
66 list_for_each_entry_safe(pi, next, paths, list) {
72 static void st_destroy(struct path_selector *ps)
74 struct selector *s = ps->context;
76 free_paths(&s->valid_paths);
77 free_paths(&s->failed_paths);
82 static int st_status(struct path_selector *ps, struct dm_path *path,
83 status_type_t type, char *result, unsigned maxlen)
95 DMEMIT("%d %u ", atomic_read(&pi->in_flight_size),
96 pi->relative_throughput);
98 case STATUSTYPE_TABLE:
99 DMEMIT("%u %u ", pi->repeat_count,
100 pi->relative_throughput);
108 static int st_add_path(struct path_selector *ps, struct dm_path *path,
109 int argc, char **argv, char **error)
111 struct selector *s = ps->context;
112 struct path_info *pi;
113 unsigned repeat_count = ST_MIN_IO;
114 unsigned relative_throughput = 1;
119 * Arguments: [<repeat_count> [<relative_throughput>]]
120 * <repeat_count>: The number of I/Os before switching path.
121 * If not given, default (ST_MIN_IO) is used.
122 * <relative_throughput>: The relative throughput value of
123 * the path among all paths in the path-group.
124 * The valid range: 0-<ST_MAX_RELATIVE_THROUGHPUT>
125 * If not given, minimum value '1' is used.
126 * If '0' is given, the path isn't selected while
127 * other paths having a positive value are
131 *error = "service-time ps: incorrect number of arguments";
135 if (argc && (sscanf(argv[0], "%u%c", &repeat_count, &dummy) != 1)) {
136 *error = "service-time ps: invalid repeat count";
140 if (repeat_count > 1) {
141 DMWARN_LIMIT("repeat_count > 1 is deprecated, using 1 instead");
146 (sscanf(argv[1], "%u%c", &relative_throughput, &dummy) != 1 ||
147 relative_throughput > ST_MAX_RELATIVE_THROUGHPUT)) {
148 *error = "service-time ps: invalid relative_throughput value";
152 /* allocate the path */
153 pi = kmalloc(sizeof(*pi), GFP_KERNEL);
155 *error = "service-time ps: Error allocating path context";
160 pi->repeat_count = repeat_count;
161 pi->relative_throughput = relative_throughput;
162 atomic_set(&pi->in_flight_size, 0);
164 path->pscontext = pi;
166 spin_lock_irqsave(&s->lock, flags);
167 list_add_tail(&pi->list, &s->valid_paths);
168 spin_unlock_irqrestore(&s->lock, flags);
173 static void st_fail_path(struct path_selector *ps, struct dm_path *path)
175 struct selector *s = ps->context;
176 struct path_info *pi = path->pscontext;
179 spin_lock_irqsave(&s->lock, flags);
180 list_move(&pi->list, &s->failed_paths);
181 spin_unlock_irqrestore(&s->lock, flags);
184 static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
186 struct selector *s = ps->context;
187 struct path_info *pi = path->pscontext;
190 spin_lock_irqsave(&s->lock, flags);
191 list_move_tail(&pi->list, &s->valid_paths);
192 spin_unlock_irqrestore(&s->lock, flags);
198 * Compare the estimated service time of 2 paths, pi1 and pi2,
199 * for the incoming I/O.
202 * < 0 : pi1 is better
203 * 0 : no difference between pi1 and pi2
204 * > 0 : pi2 is better
207 * Basically, the service time is estimated by:
208 * ('pi->in-flight-size' + 'incoming') / 'pi->relative_throughput'
209 * To reduce the calculation, some optimizations are made.
210 * (See comments inline)
212 static int st_compare_load(struct path_info *pi1, struct path_info *pi2,
215 size_t sz1, sz2, st1, st2;
217 sz1 = atomic_read(&pi1->in_flight_size);
218 sz2 = atomic_read(&pi2->in_flight_size);
221 * Case 1: Both have same throughput value. Choose less loaded path.
223 if (pi1->relative_throughput == pi2->relative_throughput)
227 * Case 2a: Both have same load. Choose higher throughput path.
228 * Case 2b: One path has no throughput value. Choose the other one.
231 !pi1->relative_throughput || !pi2->relative_throughput)
232 return pi2->relative_throughput - pi1->relative_throughput;
235 * Case 3: Calculate service time. Choose faster path.
236 * Service time using pi1:
237 * st1 = (sz1 + incoming) / pi1->relative_throughput
238 * Service time using pi2:
239 * st2 = (sz2 + incoming) / pi2->relative_throughput
241 * To avoid the division, transform the expression to use
243 * Because ->relative_throughput > 0 here, if st1 < st2,
244 * the expressions below are the same meaning:
245 * (sz1 + incoming) / pi1->relative_throughput <
246 * (sz2 + incoming) / pi2->relative_throughput
247 * (sz1 + incoming) * pi2->relative_throughput <
248 * (sz2 + incoming) * pi1->relative_throughput
249 * So use the later one.
253 if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE ||
254 sz2 >= ST_MAX_INFLIGHT_SIZE)) {
256 * Size may be too big for multiplying pi->relative_throughput
258 * To avoid the overflow and mis-selection, shift down both.
260 sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
261 sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
263 st1 = sz1 * pi2->relative_throughput;
264 st2 = sz2 * pi1->relative_throughput;
269 * Case 4: Service time is equal. Choose higher throughput path.
271 return pi2->relative_throughput - pi1->relative_throughput;
274 static struct dm_path *st_select_path(struct path_selector *ps, size_t nr_bytes)
276 struct selector *s = ps->context;
277 struct path_info *pi = NULL, *best = NULL;
278 struct dm_path *ret = NULL;
281 spin_lock_irqsave(&s->lock, flags);
282 if (list_empty(&s->valid_paths))
285 /* Change preferred (first in list) path to evenly balance. */
286 list_move_tail(s->valid_paths.next, &s->valid_paths);
288 list_for_each_entry(pi, &s->valid_paths, list)
289 if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
297 spin_unlock_irqrestore(&s->lock, flags);
301 static int st_start_io(struct path_selector *ps, struct dm_path *path,
304 struct path_info *pi = path->pscontext;
306 atomic_add(nr_bytes, &pi->in_flight_size);
311 static int st_end_io(struct path_selector *ps, struct dm_path *path,
314 struct path_info *pi = path->pscontext;
316 atomic_sub(nr_bytes, &pi->in_flight_size);
321 static struct path_selector_type st_ps = {
322 .name = "service-time",
323 .module = THIS_MODULE,
327 .destroy = st_destroy,
329 .add_path = st_add_path,
330 .fail_path = st_fail_path,
331 .reinstate_path = st_reinstate_path,
332 .select_path = st_select_path,
333 .start_io = st_start_io,
337 static int __init dm_st_init(void)
339 int r = dm_register_path_selector(&st_ps);
342 DMERR("register failed %d", r);
344 DMINFO("version " ST_VERSION " loaded");
349 static void __exit dm_st_exit(void)
351 int r = dm_unregister_path_selector(&st_ps);
354 DMERR("unregister failed %d", r);
357 module_init(dm_st_init);
358 module_exit(dm_st_exit);
360 MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
361 MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
362 MODULE_LICENSE("GPL");