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_secure_erase_bios = stripes;
 173         ti->num_write_same_bios = stripes;
 174         ti->num_write_zeroes_bios = stripes;
 175
 176         sc->chunk_size = chunk_size;
 177         if (chunk_size & (chunk_size - 1))
 178                 sc->chunk_size_shift = -1;
 179         else
 180                 sc->chunk_size_shift = __ffs(chunk_size);
 181
 182         /*
 183          * Get the stripe destinations.
 184          */
 185         for (i = 0; i < stripes; i++) {
 186                 argv += 2;
 187
 188                 r = get_stripe(ti, sc, i, argv);
 189                 if (r < 0) {
 190                         ti->error = "Couldn't parse stripe destination";
 191                         while (i--)
 192                                 dm_put_device(ti, sc->stripe[i].dev);
 193                         kfree(sc);
 194                         return r;
 195                 }
 196                 atomic_set(&(sc->stripe[i].error_count), 0);
 197         }
 198
 199         ti->private = sc;
 200
 201         return 0;
 202 }
 203
 204 static void stripe_dtr(struct dm_target *ti)
 205 {
 206         unsigned int i;
 207         struct stripe_c *sc = (struct stripe_c *) ti->private;
 208
 209         for (i = 0; i < sc->stripes; i++)
 210                 dm_put_device(ti, sc->stripe[i].dev);
 211
 212         flush_work(&sc->trigger_event);
 213         kfree(sc);
 214 }
 215
 216 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 217                               uint32_t *stripe, sector_t *result)
 218 {
 219         sector_t chunk = dm_target_offset(sc->ti, sector);
 220         sector_t chunk_offset;
 221
 222         if (sc->chunk_size_shift < 0)
 223                 chunk_offset = sector_div(chunk, sc->chunk_size);
 224         else {
 225                 chunk_offset = chunk & (sc->chunk_size - 1);
 226                 chunk >>= sc->chunk_size_shift;
 227         }
 228
 229         if (sc->stripes_shift < 0)
 230                 *stripe = sector_div(chunk, sc->stripes);
 231         else {
 232                 *stripe = chunk & (sc->stripes - 1);
 233                 chunk >>= sc->stripes_shift;
 234         }
 235
 236         if (sc->chunk_size_shift < 0)
 237                 chunk *= sc->chunk_size;
 238         else
 239                 chunk <<= sc->chunk_size_shift;
 240
 241         *result = chunk + chunk_offset;
 242 }
 243
 244 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 245                                     uint32_t target_stripe, sector_t *result)
 246 {
 247         uint32_t stripe;
 248
 249         stripe_map_sector(sc, sector, &stripe, result);
 250         if (stripe == target_stripe)
 251                 return;
 252
 253         /* round down */
 254         sector = *result;
 255         if (sc->chunk_size_shift < 0)
 256                 *result -= sector_div(sector, sc->chunk_size);
 257         else
 258                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 259
 260         if (target_stripe < stripe)
 261                 *result += sc->chunk_size;              /* next chunk */
 262 }
 263
 264 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 265                             uint32_t target_stripe)
 266 {
 267         sector_t begin, end;
 268
 269         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 270                                 target_stripe, &begin);
 271         stripe_map_range_sector(sc, bio_end_sector(bio),
 272                                 target_stripe, &end);
 273         if (begin < end) {
 274                 bio_set_dev(bio, sc->stripe[target_stripe].dev->bdev);
 275                 bio->bi_iter.bi_sector = begin +
 276                         sc->stripe[target_stripe].physical_start;
 277                 bio->bi_iter.bi_size = to_bytes(end - begin);
 278                 return DM_MAPIO_REMAPPED;
 279         } else {
 280                 /* The range doesn't map to the target stripe */
 281                 bio_endio(bio);
 282                 return DM_MAPIO_SUBMITTED;
 283         }
 284 }
 285
 286 static int stripe_map(struct dm_target *ti, struct bio *bio)
 287 {
 288         struct stripe_c *sc = ti->private;
 289         uint32_t stripe;
 290         unsigned target_bio_nr;
 291
 292         if (bio->bi_opf & REQ_PREFLUSH) {
 293                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 294                 BUG_ON(target_bio_nr >= sc->stripes);
 295                 bio_set_dev(bio, sc->stripe[target_bio_nr].dev->bdev);
 296                 return DM_MAPIO_REMAPPED;
 297         }
 298         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 299             unlikely(bio_op(bio) == REQ_OP_SECURE_ERASE) ||
 300             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES) ||
 301             unlikely(bio_op(bio) == REQ_OP_WRITE_SAME)) {
 302                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 303                 BUG_ON(target_bio_nr >= sc->stripes);
 304                 return stripe_map_range(sc, bio, target_bio_nr);
 305         }
 306
 307         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 308                           &stripe, &bio->bi_iter.bi_sector);
 309
 310         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 311         bio_set_dev(bio, sc->stripe[stripe].dev->bdev);
 312
 313         return DM_MAPIO_REMAPPED;
 314 }
 315
 316 #if IS_ENABLED(CONFIG_DAX_DRIVER)
 317 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 318                 long nr_pages, void **kaddr, pfn_t *pfn)
 319 {
 320         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 321         struct stripe_c *sc = ti->private;
 322         struct dax_device *dax_dev;
 323         struct block_device *bdev;
 324         uint32_t stripe;
 325         long ret;
 326
 327         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 328         dev_sector += sc->stripe[stripe].physical_start;
 329         dax_dev = sc->stripe[stripe].dev->dax_dev;
 330         bdev = sc->stripe[stripe].dev->bdev;
 331
 332         ret = bdev_dax_pgoff(bdev, dev_sector, nr_pages * PAGE_SIZE, &pgoff);
 333         if (ret)
 334                 return ret;
 335         return dax_direct_access(dax_dev, pgoff, nr_pages, kaddr, pfn);
 336 }
 337
 338 static size_t stripe_dax_copy_from_iter(struct dm_target *ti, pgoff_t pgoff,
 339                 void *addr, size_t bytes, struct iov_iter *i)
 340 {
 341         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 342         struct stripe_c *sc = ti->private;
 343         struct dax_device *dax_dev;
 344         struct block_device *bdev;
 345         uint32_t stripe;
 346
 347         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 348         dev_sector += sc->stripe[stripe].physical_start;
 349         dax_dev = sc->stripe[stripe].dev->dax_dev;
 350         bdev = sc->stripe[stripe].dev->bdev;
 351
 352         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 353                 return 0;
 354         return dax_copy_from_iter(dax_dev, pgoff, addr, bytes, i);
 355 }
 356
 357 static size_t stripe_dax_copy_to_iter(struct dm_target *ti, pgoff_t pgoff,
 358                 void *addr, size_t bytes, struct iov_iter *i)
 359 {
 360         sector_t dev_sector, sector = pgoff * PAGE_SECTORS;
 361         struct stripe_c *sc = ti->private;
 362         struct dax_device *dax_dev;
 363         struct block_device *bdev;
 364         uint32_t stripe;
 365
 366         stripe_map_sector(sc, sector, &stripe, &dev_sector);
 367         dev_sector += sc->stripe[stripe].physical_start;
 368         dax_dev = sc->stripe[stripe].dev->dax_dev;
 369         bdev = sc->stripe[stripe].dev->bdev;
 370
 371         if (bdev_dax_pgoff(bdev, dev_sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 372                 return 0;
 373         return dax_copy_to_iter(dax_dev, pgoff, addr, bytes, i);
 374 }
 375
 376 #else
 377 #define stripe_dax_direct_access NULL
 378 #define stripe_dax_copy_from_iter NULL
 379 #define stripe_dax_copy_to_iter NULL
 380 #endif
 381
 382 /*
 383  * Stripe status:
 384  *
 385  * INFO
 386  * #stripes [stripe_name <stripe_name>] [group word count]
 387  * [error count 'A|D' <error count 'A|D'>]
 388  *
 389  * TABLE
 390  * #stripes [stripe chunk size]
 391  * [stripe_name physical_start <stripe_name physical_start>]
 392  *
 393  */
 394
 395 static void stripe_status(struct dm_target *ti, status_type_t type,
 396                           unsigned status_flags, char *result, unsigned maxlen)
 397 {
 398         struct stripe_c *sc = (struct stripe_c *) ti->private;
 399         unsigned int sz = 0;
 400         unsigned int i;
 401
 402         switch (type) {
 403         case STATUSTYPE_INFO:
 404                 DMEMIT("%d ", sc->stripes);
 405                 for (i = 0; i < sc->stripes; i++)  {
 406                         DMEMIT("%s ", sc->stripe[i].dev->name);
 407                 }
 408                 DMEMIT("1 ");
 409                 for (i = 0; i < sc->stripes; i++) {
 410                         DMEMIT("%c", atomic_read(&(sc->stripe[i].error_count)) ?
 411                                'D' : 'A');
 412                 }
 413                 break;
 414
 415         case STATUSTYPE_TABLE:
 416                 DMEMIT("%d %llu", sc->stripes,
 417                         (unsigned long long)sc->chunk_size);
 418                 for (i = 0; i < sc->stripes; i++)
 419                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 420                             (unsigned long long)sc->stripe[i].physical_start);
 421                 break;
 422         }
 423 }
 424
 425 static int stripe_end_io(struct dm_target *ti, struct bio *bio,
 426                 blk_status_t *error)
 427 {
 428         unsigned i;
 429         char major_minor[16];
 430         struct stripe_c *sc = ti->private;
 431
 432         if (!*error)
 433                 return DM_ENDIO_DONE; /* I/O complete */
 434
 435         if (bio->bi_opf & REQ_RAHEAD)
 436                 return DM_ENDIO_DONE;
 437
 438         if (*error == BLK_STS_NOTSUPP)
 439                 return DM_ENDIO_DONE;
 440
 441         memset(major_minor, 0, sizeof(major_minor));
 442         sprintf(major_minor, "%d:%d", MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)));
 443
 444         /*
 445          * Test to see which stripe drive triggered the event
 446          * and increment error count for all stripes on that device.
 447          * If the error count for a given device exceeds the threshold
 448          * value we will no longer trigger any further events.
 449          */
 450         for (i = 0; i < sc->stripes; i++)
 451                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 452                         atomic_inc(&(sc->stripe[i].error_count));
 453                         if (atomic_read(&(sc->stripe[i].error_count)) <
 454                             DM_IO_ERROR_THRESHOLD)
 455                                 schedule_work(&sc->trigger_event);
 456                 }
 457
 458         return DM_ENDIO_DONE;
 459 }
 460
 461 static int stripe_iterate_devices(struct dm_target *ti,
 462                                   iterate_devices_callout_fn fn, void *data)
 463 {
 464         struct stripe_c *sc = ti->private;
 465         int ret = 0;
 466         unsigned i = 0;
 467
 468         do {
 469                 ret = fn(ti, sc->stripe[i].dev,
 470                          sc->stripe[i].physical_start,
 471                          sc->stripe_width, data);
 472         } while (!ret && ++i < sc->stripes);
 473
 474         return ret;
 475 }
 476
 477 static void stripe_io_hints(struct dm_target *ti,
 478                             struct queue_limits *limits)
 479 {
 480         struct stripe_c *sc = ti->private;
 481         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 482
 483         blk_limits_io_min(limits, chunk_size);
 484         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 485 }
 486
 487 static struct target_type stripe_target = {
 488         .name   = "striped",
 489         .version = {1, 6, 0},
 490         .features = DM_TARGET_PASSES_INTEGRITY,
 491         .module = THIS_MODULE,
 492         .ctr    = stripe_ctr,
 493         .dtr    = stripe_dtr,
 494         .map    = stripe_map,
 495         .end_io = stripe_end_io,
 496         .status = stripe_status,
 497         .iterate_devices = stripe_iterate_devices,
 498         .io_hints = stripe_io_hints,
 499         .direct_access = stripe_dax_direct_access,
 500         .dax_copy_from_iter = stripe_dax_copy_from_iter,
 501         .dax_copy_to_iter = stripe_dax_copy_to_iter,
 502 };
 503
 504 int __init dm_stripe_init(void)
 505 {
 506         int r;
 507
 508         r = dm_register_target(&stripe_target);
 509         if (r < 0)
 510                 DMWARN("target registration failed");
 511
 512         return r;
 513 }
 514
 515 void dm_stripe_exit(void)
 516 {
 517         dm_unregister_target(&stripe_target);
 518 }