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
14#include <linux/slab.h>
15#include <linux/module.h>
16
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)
22#define ST_VERSION "0.3.0"
23
24struct selector {
25 struct list_head valid_paths;
26 struct list_head failed_paths;
27 spinlock_t lock;
28};
29
30struct 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
38static 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);
45 spin_lock_init(&s->lock);
46 }
47
48 return s;
49}
50
51static 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
62static 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
72static 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
82static 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;
102 }
103 }
104
105 return sz;
106}
107
108static int st_add_path(struct path_selector *ps, struct dm_path *path,
109 int argc, char **argv, char **error)
110{
111 struct selector *s = ps->context;
112 struct path_info *pi;
113 unsigned repeat_count = ST_MIN_IO;
114 unsigned relative_throughput = 1;
115 char dummy;
116 unsigned long flags;
117
118 /*
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
128 * available.
129 */
130 if (argc > 2) {
131 *error = "service-time ps: incorrect number of arguments";
132 return -EINVAL;
133 }
134
135 if (argc && (sscanf(argv[0], "%u%c", &repeat_count, &dummy) != 1)) {
136 *error = "service-time ps: invalid repeat count";
137 return -EINVAL;
138 }
139
140 if (repeat_count > 1) {
141 DMWARN_LIMIT("repeat_count > 1 is deprecated, using 1 instead");
142 repeat_count = 1;
143 }
144
145 if ((argc == 2) &&
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";
149 return -EINVAL;
150 }
151
152 /* allocate the path */
153 pi = kmalloc(sizeof(*pi), GFP_KERNEL);
154 if (!pi) {
155 *error = "service-time ps: Error allocating path context";
156 return -ENOMEM;
157 }
158
159 pi->path = path;
160 pi->repeat_count = repeat_count;
161 pi->relative_throughput = relative_throughput;
162 atomic_set(&pi->in_flight_size, 0);
163
164 path->pscontext = pi;
165
166 spin_lock_irqsave(&s->lock, flags);
167 list_add_tail(&pi->list, &s->valid_paths);
168 spin_unlock_irqrestore(&s->lock, flags);
169
170 return 0;
171}
172
173static void st_fail_path(struct path_selector *ps, struct dm_path *path)
174{
175 struct selector *s = ps->context;
176 struct path_info *pi = path->pscontext;
177 unsigned long flags;
178
179 spin_lock_irqsave(&s->lock, flags);
180 list_move(&pi->list, &s->failed_paths);
181 spin_unlock_irqrestore(&s->lock, flags);
182}
183
184static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
185{
186 struct selector *s = ps->context;
187 struct path_info *pi = path->pscontext;
188 unsigned long flags;
189
190 spin_lock_irqsave(&s->lock, flags);
191 list_move_tail(&pi->list, &s->valid_paths);
192 spin_unlock_irqrestore(&s->lock, flags);
193
194 return 0;
195}
196
197/*
198 * Compare the estimated service time of 2 paths, pi1 and pi2,
199 * for the incoming I/O.
200 *
201 * Returns:
202 * < 0 : pi1 is better
203 * 0 : no difference between pi1 and pi2
204 * > 0 : pi2 is better
205 *
206 * Description:
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)
211 */
212static int st_compare_load(struct path_info *pi1, struct path_info *pi2,
213 size_t incoming)
214{
215 size_t sz1, sz2, st1, st2;
216
217 sz1 = atomic_read(&pi1->in_flight_size);
218 sz2 = atomic_read(&pi2->in_flight_size);
219
220 /*
221 * Case 1: Both have same throughput value. Choose less loaded path.
222 */
223 if (pi1->relative_throughput == pi2->relative_throughput)
224 return sz1 - sz2;
225
226 /*
227 * Case 2a: Both have same load. Choose higher throughput path.
228 * Case 2b: One path has no throughput value. Choose the other one.
229 */
230 if (sz1 == sz2 ||
231 !pi1->relative_throughput || !pi2->relative_throughput)
232 return pi2->relative_throughput - pi1->relative_throughput;
233
234 /*
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
240 *
241 * To avoid the division, transform the expression to use
242 * multiplication.
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.
250 */
251 sz1 += incoming;
252 sz2 += incoming;
253 if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE ||
254 sz2 >= ST_MAX_INFLIGHT_SIZE)) {
255 /*
256 * Size may be too big for multiplying pi->relative_throughput
257 * and overflow.
258 * To avoid the overflow and mis-selection, shift down both.
259 */
260 sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
261 sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
262 }
263 st1 = sz1 * pi2->relative_throughput;
264 st2 = sz2 * pi1->relative_throughput;
265 if (st1 != st2)
266 return st1 - st2;
267
268 /*
269 * Case 4: Service time is equal. Choose higher throughput path.
270 */
271 return pi2->relative_throughput - pi1->relative_throughput;
272}
273
274static struct dm_path *st_select_path(struct path_selector *ps, size_t nr_bytes)
275{
276 struct selector *s = ps->context;
277 struct path_info *pi = NULL, *best = NULL;
278 struct dm_path *ret = NULL;
279 unsigned long flags;
280
281 spin_lock_irqsave(&s->lock, flags);
282 if (list_empty(&s->valid_paths))
283 goto out;
284
285 list_for_each_entry(pi, &s->valid_paths, list)
286 if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
287 best = pi;
288
289 if (!best)
290 goto out;
291
292 /* Move most recently used to least preferred to evenly balance. */
293 list_move_tail(&best->list, &s->valid_paths);
294
295 ret = best->path;
296out:
297 spin_unlock_irqrestore(&s->lock, flags);
298 return ret;
299}
300
301static int st_start_io(struct path_selector *ps, struct dm_path *path,
302 size_t nr_bytes)
303{
304 struct path_info *pi = path->pscontext;
305
306 atomic_add(nr_bytes, &pi->in_flight_size);
307
308 return 0;
309}
310
311static int st_end_io(struct path_selector *ps, struct dm_path *path,
312 size_t nr_bytes)
313{
314 struct path_info *pi = path->pscontext;
315
316 atomic_sub(nr_bytes, &pi->in_flight_size);
317
318 return 0;
319}
320
321static struct path_selector_type st_ps = {
322 .name = "service-time",
323 .module = THIS_MODULE,
324 .table_args = 2,
325 .info_args = 2,
326 .create = st_create,
327 .destroy = st_destroy,
328 .status = st_status,
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,
334 .end_io = st_end_io,
335};
336
337static int __init dm_st_init(void)
338{
339 int r = dm_register_path_selector(&st_ps);
340
341 if (r < 0)
342 DMERR("register failed %d", r);
343
344 DMINFO("version " ST_VERSION " loaded");
345
346 return r;
347}
348
349static void __exit dm_st_exit(void)
350{
351 int r = dm_unregister_path_selector(&st_ps);
352
353 if (r < 0)
354 DMERR("unregister failed %d", r);
355}
356
357module_init(dm_st_init);
358module_exit(dm_st_exit);
359
360MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
361MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
362MODULE_LICENSE("GPL");
363