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/dax.h>
  15 #include <linux/slab.h>
  16 #include <linux/log2.h>
  17
  18 #define DM_MSG_PREFIX "striped"
  19 #define DM_IO_ERROR_THRESHOLD 15
  20
  21 struct stripe {
  22         struct dm_dev *dev;
  23         sector_t physical_start;
  24
  25         atomic_t error_count;
  26 };
  27
  28 struct stripe_c {
  29         uint32_t stripes;
  30         int stripes_shift;
  31
  32         /* The size of this target / num. stripes */
  33         sector_t stripe_width;
  34
  35         uint32_t chunk_size;
  36         int chunk_size_shift;
  37
  38         /* Needed for handling events */
  39         struct dm_target *ti;
  40
  41         /* Work struct used for triggering events*/
  42         struct work_struct trigger_event;
  43
  44         struct stripe stripe[0];
  45 };
  46
  47 /*
  48  * An event is triggered whenever a drive
  49  * drops out of a stripe volume.
  50  */
  51 static void trigger_event(struct work_struct *work)
  52 {
  53         struct stripe_c *sc = container_of(work, struct stripe_c,
  54                                            trigger_event);
  55         dm_table_event(sc->ti->table);
  56 }
  57
  58 static inline struct stripe_c *alloc_context(unsigned int stripes)
  59 {
  60         size_t len;
  61
  62         if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
  63                              stripes))
  64                 return NULL;
  65
  66         len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
  67
  68         return kmalloc(len, GFP_KERNEL);
  69 }
  70
  71 /*
  72  * Parse a single <dev> <sector> pair
  73  */
  74 static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
  75                       unsigned int stripe, char **argv)
  76 {
  77         unsigned long long start;
  78         char dummy;
  79         int ret;
  80
  81         if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
  82                 return -EINVAL;
  83
  84         ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
  85                             &sc->stripe[stripe].dev);
  86         if (ret)
  87                 return ret;
  88
  89         sc->stripe[stripe].physical_start = start;
  90
  91         return 0;
  92 }
  93
  94 /*
  95  * Construct a striped mapping.
  96  * <number of stripes> <chunk size> [<dev_path> <offset>]+
  97  */
  98 static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
  99 {
 100         struct stripe_c *sc;
 101         sector_t width, tmp_len;
 102         uint32_t stripes;
 103         uint32_t chunk_size;
 104         int r;
 105         unsigned int i;
 106
 107         if (argc < 2) {
 108                 ti->error = "Not enough arguments";
 109                 return -EINVAL;
 110         }
 111
 112         if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
 113                 ti->error = "Invalid stripe count";
 114                 return -EINVAL;
 115         }
 116
 117         if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
 118                 ti->error = "Invalid chunk_size";
 119                 return -EINVAL;
 120         }
 121
 122         width = ti->len;
 123         if (sector_div(width, stripes)) {
 124                 ti->error = "Target length not divisible by "
 125                     "number of stripes";
 126                 return -EINVAL;
 127         }
 128
 129         tmp_len = width;
 130         if (sector_div(tmp_len, chunk_size)) {
 131                 ti->error = "Target length not divisible by "
 132                     "chunk size";
 133                 return -EINVAL;
 134         }
 135
 136         /*
 137          * Do we have enough arguments for that many stripes ?
 138          */
 139         if (argc != (2 + 2 * stripes)) {
 140                 ti->error = "Not enough destinations "
 141                         "specified";
 142                 return -EINVAL;
 143         }
 144
 145         sc = alloc_context(stripes);
 146         if (!sc) {
 147                 ti->error = "Memory allocation for striped context "
 148                     "failed";
 149                 return -ENOMEM;
 150         }
 151
 152         INIT_WORK(&sc->trigger_event, trigger_event);
 153
 154         /* Set pointer to dm target; used in trigger_event */
 155         sc->ti = ti;
 156         sc->stripes = stripes;
 157         sc->stripe_width = width;
 158
 159         if (stripes & (stripes - 1))
 160                 sc->stripes_shift = -1;
 161         else
 162                 sc->stripes_shift = __ffs(stripes);
 163
 164         r = dm_set_target_max_io_len(ti, chunk_size);
 165         if (r) {
 166                 kfree(sc);
 167                 return r;
 168         }
 169
 170         ti->num_flush_bios = stripes;
 171         ti->num_discard_bios = stripes;
 172         ti->num_write_same_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_SAME)) {
 298                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 299                 BUG_ON(target_bio_nr >= sc->stripes);
 300                 return stripe_map_range(sc, bio, target_bio_nr);
 301         }
 302
 303         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 304                           &stripe, &bio->bi_iter.bi_sector);
 305
 306         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 307         bio->bi_bdev = sc->stripe[stripe].dev->bdev;
 308
 309         return DM_MAPIO_REMAPPED;
 310 }
 311
 312 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 313                 long nr_pages, void **kaddr, pfn_t *pfn)
 314 {
 315         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 316         struct stripe_c *sc = ti->private;
 317         struct dax_device *dax_dev;
 318         struct block_device *bdev;
 319         uint32_t stripe;
 320         long ret;
 321
 322         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 323         dev_sector += sc->stripe[stripe].physical_start;
 324         dax_dev = sc->stripe[stripe].dev->dax_dev;
 325         bdev = sc->stripe[stripe].dev->bdev;
 326
 327         ret = bdev_dax_pgoff(bdev, dev_sector, nr_pages * PAGE_SIZE, &pgoff);
 328         if (ret)
 329                 return ret;
 330         return dax_direct_access(dax_dev, pgoff, nr_pages, kaddr, pfn);
 331 }
 332
 333 /*
 334  * Stripe status:
 335  *
 336  * INFO
 337  * #stripes [stripe_name <stripe_name>] [group word count]
 338  * [error count 'A|D' <error count 'A|D'>]
 339  *
 340  * TABLE
 341  * #stripes [stripe chunk size]
 342  * [stripe_name physical_start <stripe_name physical_start>]
 343  *
 344  */
 345
 346 static void stripe_status(struct dm_target *ti, status_type_t type,
 347                           unsigned status_flags, char *result, unsigned maxlen)
 348 {
 349         struct stripe_c *sc = (struct stripe_c *) ti->private;
 350         char buffer[sc->stripes + 1];
 351         unsigned int sz = 0;
 352         unsigned int i;
 353
 354         switch (type) {
 355         case STATUSTYPE_INFO:
 356                 DMEMIT("%d ", sc->stripes);
 357                 for (i = 0; i < sc->stripes; i++)  {
 358                         DMEMIT("%s ", sc->stripe[i].dev->name);
 359                         buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
 360                                 'D' : 'A';
 361                 }
 362                 buffer[i] = '\0';
 363                 DMEMIT("1 %s", buffer);
 364                 break;
 365
 366         case STATUSTYPE_TABLE:
 367                 DMEMIT("%d %llu", sc->stripes,
 368                         (unsigned long long)sc->chunk_size);
 369                 for (i = 0; i < sc->stripes; i++)
 370                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 371                             (unsigned long long)sc->stripe[i].physical_start);
 372                 break;
 373         }
 374 }
 375
 376 static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error)
 377 {
 378         unsigned i;
 379         char major_minor[16];
 380         struct stripe_c *sc = ti->private;
 381
 382         if (!error)
 383                 return 0; /* I/O complete */
 384
 385         if ((error == -EWOULDBLOCK) && (bio->bi_opf & REQ_RAHEAD))
 386                 return error;
 387
 388         if (error == -EOPNOTSUPP)
 389                 return error;
 390
 391         memset(major_minor, 0, sizeof(major_minor));
 392         sprintf(major_minor, "%d:%d",
 393                 MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
 394                 MINOR(disk_devt(bio->bi_bdev->bd_disk)));
 395
 396         /*
 397          * Test to see which stripe drive triggered the event
 398          * and increment error count for all stripes on that device.
 399          * If the error count for a given device exceeds the threshold
 400          * value we will no longer trigger any further events.
 401          */
 402         for (i = 0; i < sc->stripes; i++)
 403                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 404                         atomic_inc(&(sc->stripe[i].error_count));
 405                         if (atomic_read(&(sc->stripe[i].error_count)) <
 406                             DM_IO_ERROR_THRESHOLD)
 407                                 schedule_work(&sc->trigger_event);
 408                 }
 409
 410         return error;
 411 }
 412
 413 static int stripe_iterate_devices(struct dm_target *ti,
 414                                   iterate_devices_callout_fn fn, void *data)
 415 {
 416         struct stripe_c *sc = ti->private;
 417         int ret = 0;
 418         unsigned i = 0;
 419
 420         do {
 421                 ret = fn(ti, sc->stripe[i].dev,
 422                          sc->stripe[i].physical_start,
 423                          sc->stripe_width, data);
 424         } while (!ret && ++i < sc->stripes);
 425
 426         return ret;
 427 }
 428
 429 static void stripe_io_hints(struct dm_target *ti,
 430                             struct queue_limits *limits)
 431 {
 432         struct stripe_c *sc = ti->private;
 433         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 434
 435         blk_limits_io_min(limits, chunk_size);
 436         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 437 }
 438
 439 static struct target_type stripe_target = {
 440         .name   = "striped",
 441         .version = {1, 6, 0},
 442         .module = THIS_MODULE,
 443         .ctr    = stripe_ctr,
 444         .dtr    = stripe_dtr,
 445         .map    = stripe_map,
 446         .end_io = stripe_end_io,
 447         .status = stripe_status,
 448         .iterate_devices = stripe_iterate_devices,
 449         .io_hints = stripe_io_hints,
 450         .direct_access = stripe_dax_direct_access,
 451 };
 452
 453 int __init dm_stripe_init(void)
 454 {
 455         int r;
 456
 457         r = dm_register_target(&stripe_target);
 458         if (r < 0)
 459                 DMWARN("target registration failed");
 460
 461         return r;
 462 }
 463
 464 void dm_stripe_exit(void)
 465 {
 466         dm_unregister_target(&stripe_target);
 467 }