drivers/md/dm-stripe.c

   1 /*
   2  * Copyright (C) 2001-2003 Sistina Software (UK) Limited.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm.h"
   8 #include <linux/device-mapper.h>
   9
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/bio.h>
  14 #include <linux/slab.h>
  15 #include <linux/log2.h>
  16
  17 #define DM_MSG_PREFIX "striped"
  18 #define DM_IO_ERROR_THRESHOLD 15
  19
  20 struct stripe {
  21         struct dm_dev *dev;
  22         sector_t physical_start;
  23
  24         atomic_t error_count;
  25 };
  26
  27 struct stripe_c {
  28         uint32_t stripes;
  29         int stripes_shift;
  30
  31         /* The size of this target / num. stripes */
  32         sector_t stripe_width;
  33
  34         uint32_t chunk_size;
  35         int chunk_size_shift;
  36
  37         /* Needed for handling events */
  38         struct dm_target *ti;
  39
  40         /* Work struct used for triggering events*/
  41         struct work_struct trigger_event;
  42
  43         struct stripe stripe[0];
  44 };
  45
  46 /*
  47  * An event is triggered whenever a drive
  48  * drops out of a stripe volume.
  49  */
  50 static void trigger_event(struct work_struct *work)
  51 {
  52         struct stripe_c *sc = container_of(work, struct stripe_c,
  53                                            trigger_event);
  54         dm_table_event(sc->ti->table);
  55 }
  56
  57 static inline struct stripe_c *alloc_context(unsigned int stripes)
  58 {
  59         size_t len;
  60
  61         if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
  62                              stripes))
  63                 return NULL;
  64
  65         len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
  66
  67         return kmalloc(len, GFP_KERNEL);
  68 }
  69
  70 /*
  71  * Parse a single <dev> <sector> pair
  72  */
  73 static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
  74                       unsigned int stripe, char **argv)
  75 {
  76         unsigned long long start;
  77         char dummy;
  78         int ret;
  79
  80         if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
  81                 return -EINVAL;
  82
  83         ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
  84                             &sc->stripe[stripe].dev);
  85         if (ret)
  86                 return ret;
  87
  88         sc->stripe[stripe].physical_start = start;
  89
  90         return 0;
  91 }
  92
  93 /*
  94  * Construct a striped mapping.
  95  * <number of stripes> <chunk size> [<dev_path> <offset>]+
  96  */
  97 static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
  98 {
  99         struct stripe_c *sc;
 100         sector_t width, tmp_len;
 101         uint32_t stripes;
 102         uint32_t chunk_size;
 103         int r;
 104         unsigned int i;
 105
 106         if (argc < 2) {
 107                 ti->error = "Not enough arguments";
 108                 return -EINVAL;
 109         }
 110
 111         if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
 112                 ti->error = "Invalid stripe count";
 113                 return -EINVAL;
 114         }
 115
 116         if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
 117                 ti->error = "Invalid chunk_size";
 118                 return -EINVAL;
 119         }
 120
 121         width = ti->len;
 122         if (sector_div(width, stripes)) {
 123                 ti->error = "Target length not divisible by "
 124                     "number of stripes";
 125                 return -EINVAL;
 126         }
 127
 128         tmp_len = width;
 129         if (sector_div(tmp_len, chunk_size)) {
 130                 ti->error = "Target length not divisible by "
 131                     "chunk size";
 132                 return -EINVAL;
 133         }
 134
 135         /*
 136          * Do we have enough arguments for that many stripes ?
 137          */
 138         if (argc != (2 + 2 * stripes)) {
 139                 ti->error = "Not enough destinations "
 140                         "specified";
 141                 return -EINVAL;
 142         }
 143
 144         sc = alloc_context(stripes);
 145         if (!sc) {
 146                 ti->error = "Memory allocation for striped context "
 147                     "failed";
 148                 return -ENOMEM;
 149         }
 150
 151         INIT_WORK(&sc->trigger_event, trigger_event);
 152
 153         /* Set pointer to dm target; used in trigger_event */
 154         sc->ti = ti;
 155         sc->stripes = stripes;
 156         sc->stripe_width = width;
 157
 158         if (stripes & (stripes - 1))
 159                 sc->stripes_shift = -1;
 160         else
 161                 sc->stripes_shift = __ffs(stripes);
 162
 163         r = dm_set_target_max_io_len(ti, chunk_size);
 164         if (r) {
 165                 kfree(sc);
 166                 return r;
 167         }
 168
 169         ti->num_flush_bios = stripes;
 170         ti->num_discard_bios = stripes;
 171         ti->num_write_same_bios = stripes;
 172         ti->num_write_zeroes_bios = stripes;
 173
 174         sc->chunk_size = chunk_size;
 175         if (chunk_size & (chunk_size - 1))
 176                 sc->chunk_size_shift = -1;
 177         else
 178                 sc->chunk_size_shift = __ffs(chunk_size);
 179
 180         /*
 181          * Get the stripe destinations.
 182          */
 183         for (i = 0; i < stripes; i++) {
 184                 argv += 2;
 185
 186                 r = get_stripe(ti, sc, i, argv);
 187                 if (r < 0) {
 188                         ti->error = "Couldn't parse stripe destination";
 189                         while (i--)
 190                                 dm_put_device(ti, sc->stripe[i].dev);
 191                         kfree(sc);
 192                         return r;
 193                 }
 194                 atomic_set(&(sc->stripe[i].error_count), 0);
 195         }
 196
 197         ti->private = sc;
 198
 199         return 0;
 200 }
 201
 202 static void stripe_dtr(struct dm_target *ti)
 203 {
 204         unsigned int i;
 205         struct stripe_c *sc = (struct stripe_c *) ti->private;
 206
 207         for (i = 0; i < sc->stripes; i++)
 208                 dm_put_device(ti, sc->stripe[i].dev);
 209
 210         flush_work(&sc->trigger_event);
 211         kfree(sc);
 212 }
 213
 214 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 215                               uint32_t *stripe, sector_t *result)
 216 {
 217         sector_t chunk = dm_target_offset(sc->ti, sector);
 218         sector_t chunk_offset;
 219
 220         if (sc->chunk_size_shift < 0)
 221                 chunk_offset = sector_div(chunk, sc->chunk_size);
 222         else {
 223                 chunk_offset = chunk & (sc->chunk_size - 1);
 224                 chunk >>= sc->chunk_size_shift;
 225         }
 226
 227         if (sc->stripes_shift < 0)
 228                 *stripe = sector_div(chunk, sc->stripes);
 229         else {
 230                 *stripe = chunk & (sc->stripes - 1);
 231                 chunk >>= sc->stripes_shift;
 232         }
 233
 234         if (sc->chunk_size_shift < 0)
 235                 chunk *= sc->chunk_size;
 236         else
 237                 chunk <<= sc->chunk_size_shift;
 238
 239         *result = chunk + chunk_offset;
 240 }
 241
 242 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 243                                     uint32_t target_stripe, sector_t *result)
 244 {
 245         uint32_t stripe;
 246
 247         stripe_map_sector(sc, sector, &stripe, result);
 248         if (stripe == target_stripe)
 249                 return;
 250
 251         /* round down */
 252         sector = *result;
 253         if (sc->chunk_size_shift < 0)
 254                 *result -= sector_div(sector, sc->chunk_size);
 255         else
 256                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 257
 258         if (target_stripe < stripe)
 259                 *result += sc->chunk_size;              /* next chunk */
 260 }
 261
 262 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 263                             uint32_t target_stripe)
 264 {
 265         sector_t begin, end;
 266
 267         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 268                                 target_stripe, &begin);
 269         stripe_map_range_sector(sc, bio_end_sector(bio),
 270                                 target_stripe, &end);
 271         if (begin < end) {
 272                 bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
 273                 bio->bi_iter.bi_sector = begin +
 274                         sc->stripe[target_stripe].physical_start;
 275                 bio->bi_iter.bi_size = to_bytes(end - begin);
 276                 return DM_MAPIO_REMAPPED;
 277         } else {
 278                 /* The range doesn't map to the target stripe */
 279                 bio_endio(bio);
 280                 return DM_MAPIO_SUBMITTED;
 281         }
 282 }
 283
 284 static int stripe_map(struct dm_target *ti, struct bio *bio)
 285 {
 286         struct stripe_c *sc = ti->private;
 287         uint32_t stripe;
 288         unsigned target_bio_nr;
 289
 290         if (bio->bi_opf & REQ_PREFLUSH) {
 291                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 292                 BUG_ON(target_bio_nr >= sc->stripes);
 293                 bio->bi_bdev = sc->stripe[target_bio_nr].dev->bdev;
 294                 return DM_MAPIO_REMAPPED;
 295         }
 296         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 297             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES) ||
 298             unlikely(bio_op(bio) == REQ_OP_WRITE_SAME)) {
 299                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 300                 BUG_ON(target_bio_nr >= sc->stripes);
 301                 return stripe_map_range(sc, bio, target_bio_nr);
 302         }
 303
 304         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 305                           &stripe, &bio->bi_iter.bi_sector);
 306
 307         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 308         bio->bi_bdev = sc->stripe[stripe].dev->bdev;
 309
 310         return DM_MAPIO_REMAPPED;
 311 }
 312
 313 static long stripe_direct_access(struct dm_target *ti, sector_t sector,
 314                                  void **kaddr, pfn_t *pfn, long size)
 315 {
 316         struct stripe_c *sc = ti->private;
 317         uint32_t stripe;
 318         struct block_device *bdev;
 319         struct blk_dax_ctl dax = {
 320                 .size = size,
 321         };
 322         long ret;
 323
 324         stripe_map_sector(sc, sector, &stripe, &dax.sector);
 325
 326         dax.sector += sc->stripe[stripe].physical_start;
 327         bdev = sc->stripe[stripe].dev->bdev;
 328
 329         ret = bdev_direct_access(bdev, &dax);
 330         *kaddr = dax.addr;
 331         *pfn = dax.pfn;
 332
 333         return ret;
 334 }
 335
 336 /*
 337  * Stripe status:
 338  *
 339  * INFO
 340  * #stripes [stripe_name <stripe_name>] [group word count]
 341  * [error count 'A|D' <error count 'A|D'>]
 342  *
 343  * TABLE
 344  * #stripes [stripe chunk size]
 345  * [stripe_name physical_start <stripe_name physical_start>]
 346  *
 347  */
 348
 349 static void stripe_status(struct dm_target *ti, status_type_t type,
 350                           unsigned status_flags, char *result, unsigned maxlen)
 351 {
 352         struct stripe_c *sc = (struct stripe_c *) ti->private;
 353         char buffer[sc->stripes + 1];
 354         unsigned int sz = 0;
 355         unsigned int i;
 356
 357         switch (type) {
 358         case STATUSTYPE_INFO:
 359                 DMEMIT("%d ", sc->stripes);
 360                 for (i = 0; i < sc->stripes; i++)  {
 361                         DMEMIT("%s ", sc->stripe[i].dev->name);
 362                         buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
 363                                 'D' : 'A';
 364                 }
 365                 buffer[i] = '\0';
 366                 DMEMIT("1 %s", buffer);
 367                 break;
 368
 369         case STATUSTYPE_TABLE:
 370                 DMEMIT("%d %llu", sc->stripes,
 371                         (unsigned long long)sc->chunk_size);
 372                 for (i = 0; i < sc->stripes; i++)
 373                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 374                             (unsigned long long)sc->stripe[i].physical_start);
 375                 break;
 376         }
 377 }
 378
 379 static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error)
 380 {
 381         unsigned i;
 382         char major_minor[16];
 383         struct stripe_c *sc = ti->private;
 384
 385         if (!error)
 386                 return 0; /* I/O complete */
 387
 388         if ((error == -EWOULDBLOCK) && (bio->bi_opf & REQ_RAHEAD))
 389                 return error;
 390
 391         if (error == -EOPNOTSUPP)
 392                 return error;
 393
 394         memset(major_minor, 0, sizeof(major_minor));
 395         sprintf(major_minor, "%d:%d",
 396                 MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
 397                 MINOR(disk_devt(bio->bi_bdev->bd_disk)));
 398
 399         /*
 400          * Test to see which stripe drive triggered the event
 401          * and increment error count for all stripes on that device.
 402          * If the error count for a given device exceeds the threshold
 403          * value we will no longer trigger any further events.
 404          */
 405         for (i = 0; i < sc->stripes; i++)
 406                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 407                         atomic_inc(&(sc->stripe[i].error_count));
 408                         if (atomic_read(&(sc->stripe[i].error_count)) <
 409                             DM_IO_ERROR_THRESHOLD)
 410                                 schedule_work(&sc->trigger_event);
 411                 }
 412
 413         return error;
 414 }
 415
 416 static int stripe_iterate_devices(struct dm_target *ti,
 417                                   iterate_devices_callout_fn fn, void *data)
 418 {
 419         struct stripe_c *sc = ti->private;
 420         int ret = 0;
 421         unsigned i = 0;
 422
 423         do {
 424                 ret = fn(ti, sc->stripe[i].dev,
 425                          sc->stripe[i].physical_start,
 426                          sc->stripe_width, data);
 427         } while (!ret && ++i < sc->stripes);
 428
 429         return ret;
 430 }
 431
 432 static void stripe_io_hints(struct dm_target *ti,
 433                             struct queue_limits *limits)
 434 {
 435         struct stripe_c *sc = ti->private;
 436         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 437
 438         blk_limits_io_min(limits, chunk_size);
 439         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 440 }
 441
 442 static struct target_type stripe_target = {
 443         .name   = "striped",
 444         .version = {1, 6, 0},
 445         .module = THIS_MODULE,
 446         .ctr    = stripe_ctr,
 447         .dtr    = stripe_dtr,
 448         .map    = stripe_map,
 449         .end_io = stripe_end_io,
 450         .status = stripe_status,
 451         .iterate_devices = stripe_iterate_devices,
 452         .io_hints = stripe_io_hints,
 453         .direct_access = stripe_direct_access,
 454 };
 455
 456 int __init dm_stripe_init(void)
 457 {
 458         int r;
 459
 460         r = dm_register_target(&stripe_target);
 461         if (r < 0)
 462                 DMWARN("target registration failed");
 463
 464         return r;
 465 }
 466
 467 void dm_stripe_exit(void)
 468 {
 469         dm_unregister_target(&stripe_target);
 470 }