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         ti->num_write_zeroes_bios = stripes;
 174
 175         sc->chunk_size = chunk_size;
 176         if (chunk_size & (chunk_size - 1))
 177                 sc->chunk_size_shift = -1;
 178         else
 179                 sc->chunk_size_shift = __ffs(chunk_size);
 180
 181         /*
 182          * Get the stripe destinations.
 183          */
 184         for (i = 0; i < stripes; i++) {
 185                 argv += 2;
 186
 187                 r = get_stripe(ti, sc, i, argv);
 188                 if (r < 0) {
 189                         ti->error = "Couldn't parse stripe destination";
 190                         while (i--)
 191                                 dm_put_device(ti, sc->stripe[i].dev);
 192                         kfree(sc);
 193                         return r;
 194                 }
 195                 atomic_set(&(sc->stripe[i].error_count), 0);
 196         }
 197
 198         ti->private = sc;
 199
 200         return 0;
 201 }
 202
 203 static void stripe_dtr(struct dm_target *ti)
 204 {
 205         unsigned int i;
 206         struct stripe_c *sc = (struct stripe_c *) ti->private;
 207
 208         for (i = 0; i < sc->stripes; i++)
 209                 dm_put_device(ti, sc->stripe[i].dev);
 210
 211         flush_work(&sc->trigger_event);
 212         kfree(sc);
 213 }
 214
 215 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 216                               uint32_t *stripe, sector_t *result)
 217 {
 218         sector_t chunk = dm_target_offset(sc->ti, sector);
 219         sector_t chunk_offset;
 220
 221         if (sc->chunk_size_shift < 0)
 222                 chunk_offset = sector_div(chunk, sc->chunk_size);
 223         else {
 224                 chunk_offset = chunk & (sc->chunk_size - 1);
 225                 chunk >>= sc->chunk_size_shift;
 226         }
 227
 228         if (sc->stripes_shift < 0)
 229                 *stripe = sector_div(chunk, sc->stripes);
 230         else {
 231                 *stripe = chunk & (sc->stripes - 1);
 232                 chunk >>= sc->stripes_shift;
 233         }
 234
 235         if (sc->chunk_size_shift < 0)
 236                 chunk *= sc->chunk_size;
 237         else
 238                 chunk <<= sc->chunk_size_shift;
 239
 240         *result = chunk + chunk_offset;
 241 }
 242
 243 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 244                                     uint32_t target_stripe, sector_t *result)
 245 {
 246         uint32_t stripe;
 247
 248         stripe_map_sector(sc, sector, &stripe, result);
 249         if (stripe == target_stripe)
 250                 return;
 251
 252         /* round down */
 253         sector = *result;
 254         if (sc->chunk_size_shift < 0)
 255                 *result -= sector_div(sector, sc->chunk_size);
 256         else
 257                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 258
 259         if (target_stripe < stripe)
 260                 *result += sc->chunk_size;              /* next chunk */
 261 }
 262
 263 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 264                             uint32_t target_stripe)
 265 {
 266         sector_t begin, end;
 267
 268         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 269                                 target_stripe, &begin);
 270         stripe_map_range_sector(sc, bio_end_sector(bio),
 271                                 target_stripe, &end);
 272         if (begin < end) {
 273                 bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
 274                 bio->bi_iter.bi_sector = begin +
 275                         sc->stripe[target_stripe].physical_start;
 276                 bio->bi_iter.bi_size = to_bytes(end - begin);
 277                 return DM_MAPIO_REMAPPED;
 278         } else {
 279                 /* The range doesn't map to the target stripe */
 280                 bio_endio(bio);
 281                 return DM_MAPIO_SUBMITTED;
 282         }
 283 }
 284
 285 static int stripe_map(struct dm_target *ti, struct bio *bio)
 286 {
 287         struct stripe_c *sc = ti->private;
 288         uint32_t stripe;
 289         unsigned target_bio_nr;
 290
 291         if (bio->bi_opf & REQ_PREFLUSH) {
 292                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 293                 BUG_ON(target_bio_nr >= sc->stripes);
 294                 bio->bi_bdev = sc->stripe[target_bio_nr].dev->bdev;
 295                 return DM_MAPIO_REMAPPED;
 296         }
 297         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 298             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES) ||
 299             unlikely(bio_op(bio) == REQ_OP_WRITE_SAME)) {
 300                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 301                 BUG_ON(target_bio_nr >= sc->stripes);
 302                 return stripe_map_range(sc, bio, target_bio_nr);
 303         }
 304
 305         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 306                           &stripe, &bio->bi_iter.bi_sector);
 307
 308         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 309         bio->bi_bdev = sc->stripe[stripe].dev->bdev;
 310
 311         return DM_MAPIO_REMAPPED;
 312 }
 313
 314 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 315                 long nr_pages, void **kaddr, pfn_t *pfn)
 316 {
 317         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 318         struct stripe_c *sc = ti->private;
 319         struct dax_device *dax_dev;
 320         struct block_device *bdev;
 321         uint32_t stripe;
 322         long ret;
 323
 324         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 325         dev_sector += sc->stripe[stripe].physical_start;
 326         dax_dev = sc->stripe[stripe].dev->dax_dev;
 327         bdev = sc->stripe[stripe].dev->bdev;
 328
 329         ret = bdev_dax_pgoff(bdev, dev_sector, nr_pages * PAGE_SIZE, &pgoff);
 330         if (ret)
 331                 return ret;
 332         return dax_direct_access(dax_dev, pgoff, nr_pages, kaddr, pfn);
 333 }
 334
 335 static size_t stripe_dax_copy_from_iter(struct dm_target *ti, pgoff_t pgoff,
 336                 void *addr, size_t bytes, struct iov_iter *i)
 337 {
 338         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 339         struct stripe_c *sc = ti->private;
 340         struct dax_device *dax_dev;
 341         struct block_device *bdev;
 342         uint32_t stripe;
 343
 344         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 345         dev_sector += sc->stripe[stripe].physical_start;
 346         dax_dev = sc->stripe[stripe].dev->dax_dev;
 347         bdev = sc->stripe[stripe].dev->bdev;
 348
 349         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 350                 return 0;
 351         return dax_copy_from_iter(dax_dev, pgoff, addr, bytes, i);
 352 }
 353
 354 static void stripe_dax_flush(struct dm_target *ti, pgoff_t pgoff, void *addr,
 355                 size_t size)
 356 {
 357         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 358         struct stripe_c *sc = ti->private;
 359         struct dax_device *dax_dev;
 360         struct block_device *bdev;
 361         uint32_t stripe;
 362
 363         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 364         dev_sector += sc->stripe[stripe].physical_start;
 365         dax_dev = sc->stripe[stripe].dev->dax_dev;
 366         bdev = sc->stripe[stripe].dev->bdev;
 367
 368         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(size, PAGE_SIZE), &pgoff))
 369                 return;
 370         dax_flush(dax_dev, pgoff, addr, size);
 371 }
 372
 373 /*
 374  * Stripe status:
 375  *
 376  * INFO
 377  * #stripes [stripe_name <stripe_name>] [group word count]
 378  * [error count 'A|D' <error count 'A|D'>]
 379  *
 380  * TABLE
 381  * #stripes [stripe chunk size]
 382  * [stripe_name physical_start <stripe_name physical_start>]
 383  *
 384  */
 385
 386 static void stripe_status(struct dm_target *ti, status_type_t type,
 387                           unsigned status_flags, char *result, unsigned maxlen)
 388 {
 389         struct stripe_c *sc = (struct stripe_c *) ti->private;
 390         char buffer[sc->stripes + 1];
 391         unsigned int sz = 0;
 392         unsigned int i;
 393
 394         switch (type) {
 395         case STATUSTYPE_INFO:
 396                 DMEMIT("%d ", sc->stripes);
 397                 for (i = 0; i < sc->stripes; i++)  {
 398                         DMEMIT("%s ", sc->stripe[i].dev->name);
 399                         buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
 400                                 'D' : 'A';
 401                 }
 402                 buffer[i] = '\0';
 403                 DMEMIT("1 %s", buffer);
 404                 break;
 405
 406         case STATUSTYPE_TABLE:
 407                 DMEMIT("%d %llu", sc->stripes,
 408                         (unsigned long long)sc->chunk_size);
 409                 for (i = 0; i < sc->stripes; i++)
 410                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 411                             (unsigned long long)sc->stripe[i].physical_start);
 412                 break;
 413         }
 414 }
 415
 416 static int stripe_end_io(struct dm_target *ti, struct bio *bio,
 417                 blk_status_t *error)
 418 {
 419         unsigned i;
 420         char major_minor[16];
 421         struct stripe_c *sc = ti->private;
 422
 423         if (!*error)
 424                 return DM_ENDIO_DONE; /* I/O complete */
 425
 426         if (bio->bi_opf & REQ_RAHEAD)
 427                 return DM_ENDIO_DONE;
 428
 429         if (*error == BLK_STS_NOTSUPP)
 430                 return DM_ENDIO_DONE;
 431
 432         memset(major_minor, 0, sizeof(major_minor));
 433         sprintf(major_minor, "%d:%d",
 434                 MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
 435                 MINOR(disk_devt(bio->bi_bdev->bd_disk)));
 436
 437         /*
 438          * Test to see which stripe drive triggered the event
 439          * and increment error count for all stripes on that device.
 440          * If the error count for a given device exceeds the threshold
 441          * value we will no longer trigger any further events.
 442          */
 443         for (i = 0; i < sc->stripes; i++)
 444                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 445                         atomic_inc(&(sc->stripe[i].error_count));
 446                         if (atomic_read(&(sc->stripe[i].error_count)) <
 447                             DM_IO_ERROR_THRESHOLD)
 448                                 schedule_work(&sc->trigger_event);
 449                 }
 450
 451         return DM_ENDIO_DONE;
 452 }
 453
 454 static int stripe_iterate_devices(struct dm_target *ti,
 455                                   iterate_devices_callout_fn fn, void *data)
 456 {
 457         struct stripe_c *sc = ti->private;
 458         int ret = 0;
 459         unsigned i = 0;
 460
 461         do {
 462                 ret = fn(ti, sc->stripe[i].dev,
 463                          sc->stripe[i].physical_start,
 464                          sc->stripe_width, data);
 465         } while (!ret && ++i < sc->stripes);
 466
 467         return ret;
 468 }
 469
 470 static void stripe_io_hints(struct dm_target *ti,
 471                             struct queue_limits *limits)
 472 {
 473         struct stripe_c *sc = ti->private;
 474         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 475
 476         blk_limits_io_min(limits, chunk_size);
 477         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 478 }
 479
 480 static struct target_type stripe_target = {
 481         .name   = "striped",
 482         .version = {1, 6, 0},
 483         .features = DM_TARGET_PASSES_INTEGRITY,
 484         .module = THIS_MODULE,
 485         .ctr    = stripe_ctr,
 486         .dtr    = stripe_dtr,
 487         .map    = stripe_map,
 488         .end_io = stripe_end_io,
 489         .status = stripe_status,
 490         .iterate_devices = stripe_iterate_devices,
 491         .io_hints = stripe_io_hints,
 492         .direct_access = stripe_dax_direct_access,
 493         .dax_copy_from_iter = stripe_dax_copy_from_iter,
 494         .dax_flush = stripe_dax_flush,
 495 };
 496
 497 int __init dm_stripe_init(void)
 498 {
 499         int r;
 500
 501         r = dm_register_target(&stripe_target);
 502         if (r < 0)
 503                 DMWARN("target registration failed");
 504
 505         return r;
 506 }
 507
 508 void dm_stripe_exit(void)
 509 {
 510         dm_unregister_target(&stripe_target);
 511 }