dm-service-time.c 8.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362
  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. #include "dm.h"
  11. #include "dm-path-selector.h"
  12. #include <linux/slab.h>
  13. #include <linux/module.h>
  14. #define DM_MSG_PREFIX "multipath service-time"
  15. #define ST_MIN_IO 1
  16. #define ST_MAX_RELATIVE_THROUGHPUT 100
  17. #define ST_MAX_RELATIVE_THROUGHPUT_SHIFT 7
  18. #define ST_MAX_INFLIGHT_SIZE ((size_t)-1 >> ST_MAX_RELATIVE_THROUGHPUT_SHIFT)
  19. #define ST_VERSION "0.3.0"
  20. struct selector {
  21. struct list_head valid_paths;
  22. struct list_head failed_paths;
  23. spinlock_t lock;
  24. };
  25. struct path_info {
  26. struct list_head list;
  27. struct dm_path *path;
  28. unsigned repeat_count;
  29. unsigned relative_throughput;
  30. atomic_t in_flight_size; /* Total size of in-flight I/Os */
  31. };
  32. static struct selector *alloc_selector(void)
  33. {
  34. struct selector *s = kmalloc(sizeof(*s), GFP_KERNEL);
  35. if (s) {
  36. INIT_LIST_HEAD(&s->valid_paths);
  37. INIT_LIST_HEAD(&s->failed_paths);
  38. spin_lock_init(&s->lock);
  39. }
  40. return s;
  41. }
  42. static int st_create(struct path_selector *ps, unsigned argc, char **argv)
  43. {
  44. struct selector *s = alloc_selector();
  45. if (!s)
  46. return -ENOMEM;
  47. ps->context = s;
  48. return 0;
  49. }
  50. static void free_paths(struct list_head *paths)
  51. {
  52. struct path_info *pi, *next;
  53. list_for_each_entry_safe(pi, next, paths, list) {
  54. list_del(&pi->list);
  55. kfree(pi);
  56. }
  57. }
  58. static void st_destroy(struct path_selector *ps)
  59. {
  60. struct selector *s = ps->context;
  61. free_paths(&s->valid_paths);
  62. free_paths(&s->failed_paths);
  63. kfree(s);
  64. ps->context = NULL;
  65. }
  66. static int st_status(struct path_selector *ps, struct dm_path *path,
  67. status_type_t type, char *result, unsigned maxlen)
  68. {
  69. unsigned sz = 0;
  70. struct path_info *pi;
  71. if (!path)
  72. DMEMIT("0 ");
  73. else {
  74. pi = path->pscontext;
  75. switch (type) {
  76. case STATUSTYPE_INFO:
  77. DMEMIT("%d %u ", atomic_read(&pi->in_flight_size),
  78. pi->relative_throughput);
  79. break;
  80. case STATUSTYPE_TABLE:
  81. DMEMIT("%u %u ", pi->repeat_count,
  82. pi->relative_throughput);
  83. break;
  84. }
  85. }
  86. return sz;
  87. }
  88. static int st_add_path(struct path_selector *ps, struct dm_path *path,
  89. int argc, char **argv, char **error)
  90. {
  91. struct selector *s = ps->context;
  92. struct path_info *pi;
  93. unsigned repeat_count = ST_MIN_IO;
  94. unsigned relative_throughput = 1;
  95. char dummy;
  96. unsigned long flags;
  97. /*
  98. * Arguments: [<repeat_count> [<relative_throughput>]]
  99. * <repeat_count>: The number of I/Os before switching path.
  100. * If not given, default (ST_MIN_IO) is used.
  101. * <relative_throughput>: The relative throughput value of
  102. * the path among all paths in the path-group.
  103. * The valid range: 0-<ST_MAX_RELATIVE_THROUGHPUT>
  104. * If not given, minimum value '1' is used.
  105. * If '0' is given, the path isn't selected while
  106. * other paths having a positive value are
  107. * available.
  108. */
  109. if (argc > 2) {
  110. *error = "service-time ps: incorrect number of arguments";
  111. return -EINVAL;
  112. }
  113. if (argc && (sscanf(argv[0], "%u%c", &repeat_count, &dummy) != 1)) {
  114. *error = "service-time ps: invalid repeat count";
  115. return -EINVAL;
  116. }
  117. if (repeat_count > 1) {
  118. DMWARN_LIMIT("repeat_count > 1 is deprecated, using 1 instead");
  119. repeat_count = 1;
  120. }
  121. if ((argc == 2) &&
  122. (sscanf(argv[1], "%u%c", &relative_throughput, &dummy) != 1 ||
  123. relative_throughput > ST_MAX_RELATIVE_THROUGHPUT)) {
  124. *error = "service-time ps: invalid relative_throughput value";
  125. return -EINVAL;
  126. }
  127. /* allocate the path */
  128. pi = kmalloc(sizeof(*pi), GFP_KERNEL);
  129. if (!pi) {
  130. *error = "service-time ps: Error allocating path context";
  131. return -ENOMEM;
  132. }
  133. pi->path = path;
  134. pi->repeat_count = repeat_count;
  135. pi->relative_throughput = relative_throughput;
  136. atomic_set(&pi->in_flight_size, 0);
  137. path->pscontext = pi;
  138. spin_lock_irqsave(&s->lock, flags);
  139. list_add_tail(&pi->list, &s->valid_paths);
  140. spin_unlock_irqrestore(&s->lock, flags);
  141. return 0;
  142. }
  143. static void st_fail_path(struct path_selector *ps, struct dm_path *path)
  144. {
  145. struct selector *s = ps->context;
  146. struct path_info *pi = path->pscontext;
  147. unsigned long flags;
  148. spin_lock_irqsave(&s->lock, flags);
  149. list_move(&pi->list, &s->failed_paths);
  150. spin_unlock_irqrestore(&s->lock, flags);
  151. }
  152. static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
  153. {
  154. struct selector *s = ps->context;
  155. struct path_info *pi = path->pscontext;
  156. unsigned long flags;
  157. spin_lock_irqsave(&s->lock, flags);
  158. list_move_tail(&pi->list, &s->valid_paths);
  159. spin_unlock_irqrestore(&s->lock, flags);
  160. return 0;
  161. }
  162. /*
  163. * Compare the estimated service time of 2 paths, pi1 and pi2,
  164. * for the incoming I/O.
  165. *
  166. * Returns:
  167. * < 0 : pi1 is better
  168. * 0 : no difference between pi1 and pi2
  169. * > 0 : pi2 is better
  170. *
  171. * Description:
  172. * Basically, the service time is estimated by:
  173. * ('pi->in-flight-size' + 'incoming') / 'pi->relative_throughput'
  174. * To reduce the calculation, some optimizations are made.
  175. * (See comments inline)
  176. */
  177. static int st_compare_load(struct path_info *pi1, struct path_info *pi2,
  178. size_t incoming)
  179. {
  180. size_t sz1, sz2, st1, st2;
  181. sz1 = atomic_read(&pi1->in_flight_size);
  182. sz2 = atomic_read(&pi2->in_flight_size);
  183. /*
  184. * Case 1: Both have same throughput value. Choose less loaded path.
  185. */
  186. if (pi1->relative_throughput == pi2->relative_throughput)
  187. return sz1 - sz2;
  188. /*
  189. * Case 2a: Both have same load. Choose higher throughput path.
  190. * Case 2b: One path has no throughput value. Choose the other one.
  191. */
  192. if (sz1 == sz2 ||
  193. !pi1->relative_throughput || !pi2->relative_throughput)
  194. return pi2->relative_throughput - pi1->relative_throughput;
  195. /*
  196. * Case 3: Calculate service time. Choose faster path.
  197. * Service time using pi1:
  198. * st1 = (sz1 + incoming) / pi1->relative_throughput
  199. * Service time using pi2:
  200. * st2 = (sz2 + incoming) / pi2->relative_throughput
  201. *
  202. * To avoid the division, transform the expression to use
  203. * multiplication.
  204. * Because ->relative_throughput > 0 here, if st1 < st2,
  205. * the expressions below are the same meaning:
  206. * (sz1 + incoming) / pi1->relative_throughput <
  207. * (sz2 + incoming) / pi2->relative_throughput
  208. * (sz1 + incoming) * pi2->relative_throughput <
  209. * (sz2 + incoming) * pi1->relative_throughput
  210. * So use the later one.
  211. */
  212. sz1 += incoming;
  213. sz2 += incoming;
  214. if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE ||
  215. sz2 >= ST_MAX_INFLIGHT_SIZE)) {
  216. /*
  217. * Size may be too big for multiplying pi->relative_throughput
  218. * and overflow.
  219. * To avoid the overflow and mis-selection, shift down both.
  220. */
  221. sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
  222. sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
  223. }
  224. st1 = sz1 * pi2->relative_throughput;
  225. st2 = sz2 * pi1->relative_throughput;
  226. if (st1 != st2)
  227. return st1 - st2;
  228. /*
  229. * Case 4: Service time is equal. Choose higher throughput path.
  230. */
  231. return pi2->relative_throughput - pi1->relative_throughput;
  232. }
  233. static struct dm_path *st_select_path(struct path_selector *ps, size_t nr_bytes)
  234. {
  235. struct selector *s = ps->context;
  236. struct path_info *pi = NULL, *best = NULL;
  237. struct dm_path *ret = NULL;
  238. unsigned long flags;
  239. spin_lock_irqsave(&s->lock, flags);
  240. if (list_empty(&s->valid_paths))
  241. goto out;
  242. list_for_each_entry(pi, &s->valid_paths, list)
  243. if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
  244. best = pi;
  245. if (!best)
  246. goto out;
  247. /* Move most recently used to least preferred to evenly balance. */
  248. list_move_tail(&best->list, &s->valid_paths);
  249. ret = best->path;
  250. out:
  251. spin_unlock_irqrestore(&s->lock, flags);
  252. return ret;
  253. }
  254. static int st_start_io(struct path_selector *ps, struct dm_path *path,
  255. size_t nr_bytes)
  256. {
  257. struct path_info *pi = path->pscontext;
  258. atomic_add(nr_bytes, &pi->in_flight_size);
  259. return 0;
  260. }
  261. static int st_end_io(struct path_selector *ps, struct dm_path *path,
  262. size_t nr_bytes, u64 start_time)
  263. {
  264. struct path_info *pi = path->pscontext;
  265. atomic_sub(nr_bytes, &pi->in_flight_size);
  266. return 0;
  267. }
  268. static struct path_selector_type st_ps = {
  269. .name = "service-time",
  270. .module = THIS_MODULE,
  271. .table_args = 2,
  272. .info_args = 2,
  273. .create = st_create,
  274. .destroy = st_destroy,
  275. .status = st_status,
  276. .add_path = st_add_path,
  277. .fail_path = st_fail_path,
  278. .reinstate_path = st_reinstate_path,
  279. .select_path = st_select_path,
  280. .start_io = st_start_io,
  281. .end_io = st_end_io,
  282. };
  283. static int __init dm_st_init(void)
  284. {
  285. int r = dm_register_path_selector(&st_ps);
  286. if (r < 0)
  287. DMERR("register failed %d", r);
  288. DMINFO("version " ST_VERSION " loaded");
  289. return r;
  290. }
  291. static void __exit dm_st_exit(void)
  292. {
  293. int r = dm_unregister_path_selector(&st_ps);
  294. if (r < 0)
  295. DMERR("unregister failed %d", r);
  296. }
  297. module_init(dm_st_init);
  298. module_exit(dm_st_exit);
  299. MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
  300. MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
  301. MODULE_LICENSE("GPL");