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[];
  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 /*
  59  * Parse a single <dev> <sector> pair
  60  */
  61 static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
  62                       unsigned int stripe, char **argv)
  63 {
  64         unsigned long long start;
  65         char dummy;
  66         int ret;
  67
  68         if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
  69                 return -EINVAL;
  70
  71         ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
  72                             &sc->stripe[stripe].dev);
  73         if (ret)
  74                 return ret;
  75
  76         sc->stripe[stripe].physical_start = start;
  77
  78         return 0;
  79 }
  80
  81 /*
  82  * Construct a striped mapping.
  83  * <number of stripes> <chunk size> [<dev_path> <offset>]+
  84  */
  85 static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
  86 {
  87         struct stripe_c *sc;
  88         sector_t width, tmp_len;
  89         uint32_t stripes;
  90         uint32_t chunk_size;
  91         int r;
  92         unsigned int i;
  93
  94         if (argc < 2) {
  95                 ti->error = "Not enough arguments";
  96                 return -EINVAL;
  97         }
  98
  99         if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
 100                 ti->error = "Invalid stripe count";
 101                 return -EINVAL;
 102         }
 103
 104         if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
 105                 ti->error = "Invalid chunk_size";
 106                 return -EINVAL;
 107         }
 108
 109         width = ti->len;
 110         if (sector_div(width, stripes)) {
 111                 ti->error = "Target length not divisible by "
 112                     "number of stripes";
 113                 return -EINVAL;
 114         }
 115
 116         tmp_len = width;
 117         if (sector_div(tmp_len, chunk_size)) {
 118                 ti->error = "Target length not divisible by "
 119                     "chunk size";
 120                 return -EINVAL;
 121         }
 122
 123         /*
 124          * Do we have enough arguments for that many stripes ?
 125          */
 126         if (argc != (2 + 2 * stripes)) {
 127                 ti->error = "Not enough destinations "
 128                         "specified";
 129                 return -EINVAL;
 130         }
 131
 132         sc = kmalloc(struct_size(sc, stripe, stripes), GFP_KERNEL);
 133         if (!sc) {
 134                 ti->error = "Memory allocation for striped context "
 135                     "failed";
 136                 return -ENOMEM;
 137         }
 138
 139         INIT_WORK(&sc->trigger_event, trigger_event);
 140
 141         /* Set pointer to dm target; used in trigger_event */
 142         sc->ti = ti;
 143         sc->stripes = stripes;
 144         sc->stripe_width = width;
 145
 146         if (stripes & (stripes - 1))
 147                 sc->stripes_shift = -1;
 148         else
 149                 sc->stripes_shift = __ffs(stripes);
 150
 151         r = dm_set_target_max_io_len(ti, chunk_size);
 152         if (r) {
 153                 kfree(sc);
 154                 return r;
 155         }
 156
 157         ti->num_flush_bios = stripes;
 158         ti->num_discard_bios = stripes;
 159         ti->num_secure_erase_bios = stripes;
 160         ti->num_write_same_bios = stripes;
 161         ti->num_write_zeroes_bios = stripes;
 162
 163         sc->chunk_size = chunk_size;
 164         if (chunk_size & (chunk_size - 1))
 165                 sc->chunk_size_shift = -1;
 166         else
 167                 sc->chunk_size_shift = __ffs(chunk_size);
 168
 169         /*
 170          * Get the stripe destinations.
 171          */
 172         for (i = 0; i < stripes; i++) {
 173                 argv += 2;
 174
 175                 r = get_stripe(ti, sc, i, argv);
 176                 if (r < 0) {
 177                         ti->error = "Couldn't parse stripe destination";
 178                         while (i--)
 179                                 dm_put_device(ti, sc->stripe[i].dev);
 180                         kfree(sc);
 181                         return r;
 182                 }
 183                 atomic_set(&(sc->stripe[i].error_count), 0);
 184         }
 185
 186         ti->private = sc;
 187
 188         return 0;
 189 }
 190
 191 static void stripe_dtr(struct dm_target *ti)
 192 {
 193         unsigned int i;
 194         struct stripe_c *sc = (struct stripe_c *) ti->private;
 195
 196         for (i = 0; i < sc->stripes; i++)
 197                 dm_put_device(ti, sc->stripe[i].dev);
 198
 199         flush_work(&sc->trigger_event);
 200         kfree(sc);
 201 }
 202
 203 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 204                               uint32_t *stripe, sector_t *result)
 205 {
 206         sector_t chunk = dm_target_offset(sc->ti, sector);
 207         sector_t chunk_offset;
 208
 209         if (sc->chunk_size_shift < 0)
 210                 chunk_offset = sector_div(chunk, sc->chunk_size);
 211         else {
 212                 chunk_offset = chunk & (sc->chunk_size - 1);
 213                 chunk >>= sc->chunk_size_shift;
 214         }
 215
 216         if (sc->stripes_shift < 0)
 217                 *stripe = sector_div(chunk, sc->stripes);
 218         else {
 219                 *stripe = chunk & (sc->stripes - 1);
 220                 chunk >>= sc->stripes_shift;
 221         }
 222
 223         if (sc->chunk_size_shift < 0)
 224                 chunk *= sc->chunk_size;
 225         else
 226                 chunk <<= sc->chunk_size_shift;
 227
 228         *result = chunk + chunk_offset;
 229 }
 230
 231 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 232                                     uint32_t target_stripe, sector_t *result)
 233 {
 234         uint32_t stripe;
 235
 236         stripe_map_sector(sc, sector, &stripe, result);
 237         if (stripe == target_stripe)
 238                 return;
 239
 240         /* round down */
 241         sector = *result;
 242         if (sc->chunk_size_shift < 0)
 243                 *result -= sector_div(sector, sc->chunk_size);
 244         else
 245                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 246
 247         if (target_stripe < stripe)
 248                 *result += sc->chunk_size;              /* next chunk */
 249 }
 250
 251 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 252                             uint32_t target_stripe)
 253 {
 254         sector_t begin, end;
 255
 256         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 257                                 target_stripe, &begin);
 258         stripe_map_range_sector(sc, bio_end_sector(bio),
 259                                 target_stripe, &end);
 260         if (begin < end) {
 261                 bio_set_dev(bio, sc->stripe[target_stripe].dev->bdev);
 262                 bio->bi_iter.bi_sector = begin +
 263                         sc->stripe[target_stripe].physical_start;
 264                 bio->bi_iter.bi_size = to_bytes(end - begin);
 265                 return DM_MAPIO_REMAPPED;
 266         } else {
 267                 /* The range doesn't map to the target stripe */
 268                 bio_endio(bio);
 269                 return DM_MAPIO_SUBMITTED;
 270         }
 271 }
 272
 273 static int stripe_map(struct dm_target *ti, struct bio *bio)
 274 {
 275         struct stripe_c *sc = ti->private;
 276         uint32_t stripe;
 277         unsigned target_bio_nr;
 278
 279         if (bio->bi_opf & REQ_PREFLUSH) {
 280                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 281                 BUG_ON(target_bio_nr >= sc->stripes);
 282                 bio_set_dev(bio, sc->stripe[target_bio_nr].dev->bdev);
 283                 return DM_MAPIO_REMAPPED;
 284         }
 285         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 286             unlikely(bio_op(bio) == REQ_OP_SECURE_ERASE) ||
 287             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES) ||
 288             unlikely(bio_op(bio) == REQ_OP_WRITE_SAME)) {
 289                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 290                 BUG_ON(target_bio_nr >= sc->stripes);
 291                 return stripe_map_range(sc, bio, target_bio_nr);
 292         }
 293
 294         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 295                           &stripe, &bio->bi_iter.bi_sector);
 296
 297         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 298         bio_set_dev(bio, sc->stripe[stripe].dev->bdev);
 299
 300         return DM_MAPIO_REMAPPED;
 301 }
 302
 303 #if IS_ENABLED(CONFIG_DAX_DRIVER)
 304 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 305                 long nr_pages, void **kaddr, pfn_t *pfn)
 306 {
 307         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 308         struct stripe_c *sc = ti->private;
 309         struct dax_device *dax_dev;
 310         struct block_device *bdev;
 311         uint32_t stripe;
 312         long ret;
 313
 314         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 315         dev_sector += sc->stripe[stripe].physical_start;
 316         dax_dev = sc->stripe[stripe].dev->dax_dev;
 317         bdev = sc->stripe[stripe].dev->bdev;
 318
 319         ret = bdev_dax_pgoff(bdev, dev_sector, nr_pages * PAGE_SIZE, &pgoff);
 320         if (ret)
 321                 return ret;
 322         return dax_direct_access(dax_dev, pgoff, nr_pages, kaddr, pfn);
 323 }
 324
 325 static size_t stripe_dax_copy_from_iter(struct dm_target *ti, pgoff_t pgoff,
 326                 void *addr, size_t bytes, struct iov_iter *i)
 327 {
 328         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 329         struct stripe_c *sc = ti->private;
 330         struct dax_device *dax_dev;
 331         struct block_device *bdev;
 332         uint32_t stripe;
 333
 334         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 335         dev_sector += sc->stripe[stripe].physical_start;
 336         dax_dev = sc->stripe[stripe].dev->dax_dev;
 337         bdev = sc->stripe[stripe].dev->bdev;
 338
 339         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 340                 return 0;
 341         return dax_copy_from_iter(dax_dev, pgoff, addr, bytes, i);
 342 }
 343
 344 static size_t stripe_dax_copy_to_iter(struct dm_target *ti, pgoff_t pgoff,
 345                 void *addr, size_t bytes, struct iov_iter *i)
 346 {
 347         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 348         struct stripe_c *sc = ti->private;
 349         struct dax_device *dax_dev;
 350         struct block_device *bdev;
 351         uint32_t stripe;
 352
 353         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 354         dev_sector += sc->stripe[stripe].physical_start;
 355         dax_dev = sc->stripe[stripe].dev->dax_dev;
 356         bdev = sc->stripe[stripe].dev->bdev;
 357
 358         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 359                 return 0;
 360         return dax_copy_to_iter(dax_dev, pgoff, addr, bytes, i);
 361 }
 362
 363 static int stripe_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
 364                                       size_t nr_pages)
 365 {
 366         int ret;
 367         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 368         struct stripe_c *sc = ti->private;
 369         struct dax_device *dax_dev;
 370         struct block_device *bdev;
 371         uint32_t stripe;
 372
 373         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 374         dev_sector += sc->stripe[stripe].physical_start;
 375         dax_dev = sc->stripe[stripe].dev->dax_dev;
 376         bdev = sc->stripe[stripe].dev->bdev;
 377
 378         ret = bdev_dax_pgoff(bdev, dev_sector, nr_pages << PAGE_SHIFT, &pgoff);
 379         if (ret)
 380                 return ret;
 381         return dax_zero_page_range(dax_dev, pgoff, nr_pages);
 382 }
 383
 384 #else
 385 #define stripe_dax_direct_access NULL
 386 #define stripe_dax_copy_from_iter NULL
 387 #define stripe_dax_copy_to_iter NULL
 388 #define stripe_dax_zero_page_range NULL
 389 #endif
 390
 391 /*
 392  * Stripe status:
 393  *
 394  * INFO
 395  * #stripes [stripe_name <stripe_name>] [group word count]
 396  * [error count 'A|D' <error count 'A|D'>]
 397  *
 398  * TABLE
 399  * #stripes [stripe chunk size]
 400  * [stripe_name physical_start <stripe_name physical_start>]
 401  *
 402  */
 403
 404 static void stripe_status(struct dm_target *ti, status_type_t type,
 405                           unsigned status_flags, char *result, unsigned maxlen)
 406 {
 407         struct stripe_c *sc = (struct stripe_c *) ti->private;
 408         unsigned int sz = 0;
 409         unsigned int i;
 410
 411         switch (type) {
 412         case STATUSTYPE_INFO:
 413                 DMEMIT("%d ", sc->stripes);
 414                 for (i = 0; i < sc->stripes; i++)  {
 415                         DMEMIT("%s ", sc->stripe[i].dev->name);
 416                 }
 417                 DMEMIT("1 ");
 418                 for (i = 0; i < sc->stripes; i++) {
 419                         DMEMIT("%c", atomic_read(&(sc->stripe[i].error_count)) ?
 420                                'D' : 'A');
 421                 }
 422                 break;
 423
 424         case STATUSTYPE_TABLE:
 425                 DMEMIT("%d %llu", sc->stripes,
 426                         (unsigned long long)sc->chunk_size);
 427                 for (i = 0; i < sc->stripes; i++)
 428                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 429                             (unsigned long long)sc->stripe[i].physical_start);
 430                 break;
 431
 432         case STATUSTYPE_IMA:
 433                 DMEMIT_TARGET_NAME_VERSION(ti->type);
 434                 DMEMIT(",stripes=%d,chunk_size=%llu", sc->stripes,
 435                        (unsigned long long)sc->chunk_size);
 436
 437                 for (i = 0; i < sc->stripes; i++) {
 438                         DMEMIT(",stripe_%d_device_name=%s", i, sc->stripe[i].dev->name);
 439                         DMEMIT(",stripe_%d_physical_start=%llu", i,
 440                                (unsigned long long)sc->stripe[i].physical_start);
 441                         DMEMIT(",stripe_%d_status=%c", i,
 442                                atomic_read(&(sc->stripe[i].error_count)) ? 'D' : 'A');
 443                 }
 444                 DMEMIT(";");
 445                 break;
 446         }
 447 }
 448
 449 static int stripe_end_io(struct dm_target *ti, struct bio *bio,
 450                 blk_status_t *error)
 451 {
 452         unsigned i;
 453         char major_minor[16];
 454         struct stripe_c *sc = ti->private;
 455
 456         if (!*error)
 457                 return DM_ENDIO_DONE; /* I/O complete */
 458
 459         if (bio->bi_opf & REQ_RAHEAD)
 460                 return DM_ENDIO_DONE;
 461
 462         if (*error == BLK_STS_NOTSUPP)
 463                 return DM_ENDIO_DONE;
 464
 465         memset(major_minor, 0, sizeof(major_minor));
 466         sprintf(major_minor, "%d:%d", MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)));
 467
 468         /*
 469          * Test to see which stripe drive triggered the event
 470          * and increment error count for all stripes on that device.
 471          * If the error count for a given device exceeds the threshold
 472          * value we will no longer trigger any further events.
 473          */
 474         for (i = 0; i < sc->stripes; i++)
 475                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 476                         atomic_inc(&(sc->stripe[i].error_count));
 477                         if (atomic_read(&(sc->stripe[i].error_count)) <
 478                             DM_IO_ERROR_THRESHOLD)
 479                                 schedule_work(&sc->trigger_event);
 480                 }
 481
 482         return DM_ENDIO_DONE;
 483 }
 484
 485 static int stripe_iterate_devices(struct dm_target *ti,
 486                                   iterate_devices_callout_fn fn, void *data)
 487 {
 488         struct stripe_c *sc = ti->private;
 489         int ret = 0;
 490         unsigned i = 0;
 491
 492         do {
 493                 ret = fn(ti, sc->stripe[i].dev,
 494                          sc->stripe[i].physical_start,
 495                          sc->stripe_width, data);
 496         } while (!ret && ++i < sc->stripes);
 497
 498         return ret;
 499 }
 500
 501 static void stripe_io_hints(struct dm_target *ti,
 502                             struct queue_limits *limits)
 503 {
 504         struct stripe_c *sc = ti->private;
 505         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 506
 507         blk_limits_io_min(limits, chunk_size);
 508         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 509 }
 510
 511 static struct target_type stripe_target = {
 512         .name   = "striped",
 513         .version = {1, 6, 0},
 514         .features = DM_TARGET_PASSES_INTEGRITY | DM_TARGET_NOWAIT,
 515         .module = THIS_MODULE,
 516         .ctr    = stripe_ctr,
 517         .dtr    = stripe_dtr,
 518         .map    = stripe_map,
 519         .end_io = stripe_end_io,
 520         .status = stripe_status,
 521         .iterate_devices = stripe_iterate_devices,
 522         .io_hints = stripe_io_hints,
 523         .direct_access = stripe_dax_direct_access,
 524         .dax_copy_from_iter = stripe_dax_copy_from_iter,
 525         .dax_copy_to_iter = stripe_dax_copy_to_iter,
 526         .dax_zero_page_range = stripe_dax_zero_page_range,
 527 };
 528
 529 int __init dm_stripe_init(void)
 530 {
 531         int r;
 532
 533         r = dm_register_target(&stripe_target);
 534         if (r < 0)
 535                 DMWARN("target registration failed");
 536
 537         return r;
 538 }
 539
 540 void dm_stripe_exit(void)
 541 {
 542         dm_unregister_target(&stripe_target);
 543 }