drivers/md/dm-service-time.c

   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      "0.2.0"
  23
  24 struct selector {
  25         struct list_head valid_paths;
  26         struct list_head failed_paths;
  27 };
  28
  29 struct path_info {
  30         struct list_head list;
  31         struct dm_path *path;
  32         unsigned repeat_count;
  33         unsigned relative_throughput;
  34         atomic_t in_flight_size;        /* Total size of in-flight I/Os */
  35 };
  36
  37 static struct selector *alloc_selector(void)
  38 {
  39         struct selector *s = kmalloc(sizeof(*s), GFP_KERNEL);
  40
  41         if (s) {
  42                 INIT_LIST_HEAD(&s->valid_paths);
  43                 INIT_LIST_HEAD(&s->failed_paths);
  44         }
  45
  46         return s;
  47 }
  48
  49 static int st_create(struct path_selector *ps, unsigned argc, char **argv)
  50 {
  51         struct selector *s = alloc_selector();
  52
  53         if (!s)
  54                 return -ENOMEM;
  55
  56         ps->context = s;
  57         return 0;
  58 }
  59
  60 static void free_paths(struct list_head *paths)
  61 {
  62         struct path_info *pi, *next;
  63
  64         list_for_each_entry_safe(pi, next, paths, list) {
  65                 list_del(&pi->list);
  66                 kfree(pi);
  67         }
  68 }
  69
  70 static void st_destroy(struct path_selector *ps)
  71 {
  72         struct selector *s = ps->context;
  73
  74         free_paths(&s->valid_paths);
  75         free_paths(&s->failed_paths);
  76         kfree(s);
  77         ps->context = NULL;
  78 }
  79
  80 static int st_status(struct path_selector *ps, struct dm_path *path,
  81                      status_type_t type, char *result, unsigned maxlen)
  82 {
  83         unsigned sz = 0;
  84         struct path_info *pi;
  85
  86         if (!path)
  87                 DMEMIT("0 ");
  88         else {
  89                 pi = path->pscontext;
  90
  91                 switch (type) {
  92                 case STATUSTYPE_INFO:
  93                         DMEMIT("%d %u ", atomic_read(&pi->in_flight_size),
  94                                pi->relative_throughput);
  95                         break;
  96                 case STATUSTYPE_TABLE:
  97                         DMEMIT("%u %u ", pi->repeat_count,
  98                                pi->relative_throughput);
  99                         break;
 100                 }
 101         }
 102
 103         return sz;
 104 }
 105
 106 static int st_add_path(struct path_selector *ps, struct dm_path *path,
 107                        int argc, char **argv, char **error)
 108 {
 109         struct selector *s = ps->context;
 110         struct path_info *pi;
 111         unsigned repeat_count = ST_MIN_IO;
 112         unsigned relative_throughput = 1;
 113
 114         /*
 115          * Arguments: [<repeat_count> [<relative_throughput>]]
 116          *      <repeat_count>: The number of I/Os before switching path.
 117          *                      If not given, default (ST_MIN_IO) is used.
 118          *      <relative_throughput>: The relative throughput value of
 119          *                      the path among all paths in the path-group.
 120          *                      The valid range: 0-<ST_MAX_RELATIVE_THROUGHPUT>
 121          *                      If not given, minimum value '1' is used.
 122          *                      If '0' is given, the path isn't selected while
 123          *                      other paths having a positive value are
 124          *                      available.
 125          */
 126         if (argc > 2) {
 127                 *error = "service-time ps: incorrect number of arguments";
 128                 return -EINVAL;
 129         }
 130
 131         if (argc && (sscanf(argv[0], "%u", &repeat_count) != 1)) {
 132                 *error = "service-time ps: invalid repeat count";
 133                 return -EINVAL;
 134         }
 135
 136         if ((argc == 2) &&
 137             (sscanf(argv[1], "%u", &relative_throughput) != 1 ||
 138              relative_throughput > ST_MAX_RELATIVE_THROUGHPUT)) {
 139                 *error = "service-time ps: invalid relative_throughput value";
 140                 return -EINVAL;
 141         }
 142
 143         /* allocate the path */
 144         pi = kmalloc(sizeof(*pi), GFP_KERNEL);
 145         if (!pi) {
 146                 *error = "service-time ps: Error allocating path context";
 147                 return -ENOMEM;
 148         }
 149
 150         pi->path = path;
 151         pi->repeat_count = repeat_count;
 152         pi->relative_throughput = relative_throughput;
 153         atomic_set(&pi->in_flight_size, 0);
 154
 155         path->pscontext = pi;
 156
 157         list_add_tail(&pi->list, &s->valid_paths);
 158
 159         return 0;
 160 }
 161
 162 static void st_fail_path(struct path_selector *ps, struct dm_path *path)
 163 {
 164         struct selector *s = ps->context;
 165         struct path_info *pi = path->pscontext;
 166
 167         list_move(&pi->list, &s->failed_paths);
 168 }
 169
 170 static int st_reinstate_path(struct path_selector *ps, struct dm_path *path)
 171 {
 172         struct selector *s = ps->context;
 173         struct path_info *pi = path->pscontext;
 174
 175         list_move_tail(&pi->list, &s->valid_paths);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * Compare the estimated service time of 2 paths, pi1 and pi2,
 182  * for the incoming I/O.
 183  *
 184  * Returns:
 185  * < 0 : pi1 is better
 186  * 0   : no difference between pi1 and pi2
 187  * > 0 : pi2 is better
 188  *
 189  * Description:
 190  * Basically, the service time is estimated by:
 191  *     ('pi->in-flight-size' + 'incoming') / 'pi->relative_throughput'
 192  * To reduce the calculation, some optimizations are made.
 193  * (See comments inline)
 194  */
 195 static int st_compare_load(struct path_info *pi1, struct path_info *pi2,
 196                            size_t incoming)
 197 {
 198         size_t sz1, sz2, st1, st2;
 199
 200         sz1 = atomic_read(&pi1->in_flight_size);
 201         sz2 = atomic_read(&pi2->in_flight_size);
 202
 203         /*
 204          * Case 1: Both have same throughput value. Choose less loaded path.
 205          */
 206         if (pi1->relative_throughput == pi2->relative_throughput)
 207                 return sz1 - sz2;
 208
 209         /*
 210          * Case 2a: Both have same load. Choose higher throughput path.
 211          * Case 2b: One path has no throughput value. Choose the other one.
 212          */
 213         if (sz1 == sz2 ||
 214             !pi1->relative_throughput || !pi2->relative_throughput)
 215                 return pi2->relative_throughput - pi1->relative_throughput;
 216
 217         /*
 218          * Case 3: Calculate service time. Choose faster path.
 219          *         Service time using pi1:
 220          *             st1 = (sz1 + incoming) / pi1->relative_throughput
 221          *         Service time using pi2:
 222          *             st2 = (sz2 + incoming) / pi2->relative_throughput
 223          *
 224          *         To avoid the division, transform the expression to use
 225          *         multiplication.
 226          *         Because ->relative_throughput > 0 here, if st1 < st2,
 227          *         the expressions below are the same meaning:
 228          *             (sz1 + incoming) / pi1->relative_throughput <
 229          *                 (sz2 + incoming) / pi2->relative_throughput
 230          *             (sz1 + incoming) * pi2->relative_throughput <
 231          *                 (sz2 + incoming) * pi1->relative_throughput
 232          *         So use the later one.
 233          */
 234         sz1 += incoming;
 235         sz2 += incoming;
 236         if (unlikely(sz1 >= ST_MAX_INFLIGHT_SIZE ||
 237                      sz2 >= ST_MAX_INFLIGHT_SIZE)) {
 238                 /*
 239                  * Size may be too big for multiplying pi->relative_throughput
 240                  * and overflow.
 241                  * To avoid the overflow and mis-selection, shift down both.
 242                  */
 243                 sz1 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
 244                 sz2 >>= ST_MAX_RELATIVE_THROUGHPUT_SHIFT;
 245         }
 246         st1 = sz1 * pi2->relative_throughput;
 247         st2 = sz2 * pi1->relative_throughput;
 248         if (st1 != st2)
 249                 return st1 - st2;
 250
 251         /*
 252          * Case 4: Service time is equal. Choose higher throughput path.
 253          */
 254         return pi2->relative_throughput - pi1->relative_throughput;
 255 }
 256
 257 static struct dm_path *st_select_path(struct path_selector *ps,
 258                                       unsigned *repeat_count, size_t nr_bytes)
 259 {
 260         struct selector *s = ps->context;
 261         struct path_info *pi = NULL, *best = NULL;
 262
 263         if (list_empty(&s->valid_paths))
 264                 return NULL;
 265
 266         /* Change preferred (first in list) path to evenly balance. */
 267         list_move_tail(s->valid_paths.next, &s->valid_paths);
 268
 269         list_for_each_entry(pi, &s->valid_paths, list)
 270                 if (!best || (st_compare_load(pi, best, nr_bytes) < 0))
 271                         best = pi;
 272
 273         if (!best)
 274                 return NULL;
 275
 276         *repeat_count = best->repeat_count;
 277
 278         return best->path;
 279 }
 280
 281 static int st_start_io(struct path_selector *ps, struct dm_path *path,
 282                        size_t nr_bytes)
 283 {
 284         struct path_info *pi = path->pscontext;
 285
 286         atomic_add(nr_bytes, &pi->in_flight_size);
 287
 288         return 0;
 289 }
 290
 291 static int st_end_io(struct path_selector *ps, struct dm_path *path,
 292                      size_t nr_bytes)
 293 {
 294         struct path_info *pi = path->pscontext;
 295
 296         atomic_sub(nr_bytes, &pi->in_flight_size);
 297
 298         return 0;
 299 }
 300
 301 static struct path_selector_type st_ps = {
 302         .name           = "service-time",
 303         .module         = THIS_MODULE,
 304         .table_args     = 2,
 305         .info_args      = 2,
 306         .create         = st_create,
 307         .destroy        = st_destroy,
 308         .status         = st_status,
 309         .add_path       = st_add_path,
 310         .fail_path      = st_fail_path,
 311         .reinstate_path = st_reinstate_path,
 312         .select_path    = st_select_path,
 313         .start_io       = st_start_io,
 314         .end_io         = st_end_io,
 315 };
 316
 317 static int __init dm_st_init(void)
 318 {
 319         int r = dm_register_path_selector(&st_ps);
 320
 321         if (r < 0)
 322                 DMERR("register failed %d", r);
 323
 324         DMINFO("version " ST_VERSION " loaded");
 325
 326         return r;
 327 }
 328
 329 static void __exit dm_st_exit(void)
 330 {
 331         int r = dm_unregister_path_selector(&st_ps);
 332
 333         if (r < 0)
 334                 DMERR("unregister failed %d", r);
 335 }
 336
 337 module_init(dm_st_init);
 338 module_exit(dm_st_exit);
 339
 340 MODULE_DESCRIPTION(DM_NAME " throughput oriented path selector");
 341 MODULE_AUTHOR("Kiyoshi Ueda <k-ueda@ct.jp.nec.com>");
 342 MODULE_LICENSE("GPL");