drivers/md/dm-stripe.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2001-2003 Sistina Software (UK) Limited.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm.h"
   9 #include <linux/device-mapper.h>
  10
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/blkdev.h>
  14 #include <linux/bio.h>
  15 #include <linux/dax.h>
  16 #include <linux/slab.h>
  17 #include <linux/log2.h>
  18
  19 #define DM_MSG_PREFIX "striped"
  20 #define DM_IO_ERROR_THRESHOLD 15
  21
  22 struct stripe {
  23         struct dm_dev *dev;
  24         sector_t physical_start;
  25
  26         atomic_t error_count;
  27 };
  28
  29 struct stripe_c {
  30         uint32_t stripes;
  31         int stripes_shift;
  32
  33         /* The size of this target / num. stripes */
  34         sector_t stripe_width;
  35
  36         uint32_t chunk_size;
  37         int chunk_size_shift;
  38
  39         /* Needed for handling events */
  40         struct dm_target *ti;
  41
  42         /* Work struct used for triggering events*/
  43         struct work_struct trigger_event;
  44
  45         struct stripe stripe[];
  46 };
  47
  48 /*
  49  * An event is triggered whenever a drive
  50  * drops out of a stripe volume.
  51  */
  52 static void trigger_event(struct work_struct *work)
  53 {
  54         struct stripe_c *sc = container_of(work, struct stripe_c,
  55                                            trigger_event);
  56         dm_table_event(sc->ti->table);
  57 }
  58
  59 /*
  60  * Parse a single <dev> <sector> pair
  61  */
  62 static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
  63                       unsigned int stripe, char **argv)
  64 {
  65         unsigned long long start;
  66         char dummy;
  67         int ret;
  68
  69         if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
  70                 return -EINVAL;
  71
  72         ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
  73                             &sc->stripe[stripe].dev);
  74         if (ret)
  75                 return ret;
  76
  77         sc->stripe[stripe].physical_start = start;
  78
  79         return 0;
  80 }
  81
  82 /*
  83  * Construct a striped mapping.
  84  * <number of stripes> <chunk size> [<dev_path> <offset>]+
  85  */
  86 static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
  87 {
  88         struct stripe_c *sc;
  89         sector_t width, tmp_len;
  90         uint32_t stripes;
  91         uint32_t chunk_size;
  92         int r;
  93         unsigned int i;
  94
  95         if (argc < 2) {
  96                 ti->error = "Not enough arguments";
  97                 return -EINVAL;
  98         }
  99
 100         if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
 101                 ti->error = "Invalid stripe count";
 102                 return -EINVAL;
 103         }
 104
 105         if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
 106                 ti->error = "Invalid chunk_size";
 107                 return -EINVAL;
 108         }
 109
 110         width = ti->len;
 111         if (sector_div(width, stripes)) {
 112                 ti->error = "Target length not divisible by 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 chunk size";
 119                 return -EINVAL;
 120         }
 121
 122         /*
 123          * Do we have enough arguments for that many stripes ?
 124          */
 125         if (argc != (2 + 2 * stripes)) {
 126                 ti->error = "Not enough destinations specified";
 127                 return -EINVAL;
 128         }
 129
 130         sc = kmalloc(struct_size(sc, stripe, stripes), GFP_KERNEL);
 131         if (!sc) {
 132                 ti->error = "Memory allocation for striped context failed";
 133                 return -ENOMEM;
 134         }
 135
 136         INIT_WORK(&sc->trigger_event, trigger_event);
 137
 138         /* Set pointer to dm target; used in trigger_event */
 139         sc->ti = ti;
 140         sc->stripes = stripes;
 141         sc->stripe_width = width;
 142
 143         if (stripes & (stripes - 1))
 144                 sc->stripes_shift = -1;
 145         else
 146                 sc->stripes_shift = __ffs(stripes);
 147
 148         r = dm_set_target_max_io_len(ti, chunk_size);
 149         if (r) {
 150                 kfree(sc);
 151                 return r;
 152         }
 153
 154         ti->num_flush_bios = stripes;
 155         ti->num_discard_bios = stripes;
 156         ti->num_secure_erase_bios = stripes;
 157         ti->num_write_zeroes_bios = stripes;
 158
 159         sc->chunk_size = chunk_size;
 160         if (chunk_size & (chunk_size - 1))
 161                 sc->chunk_size_shift = -1;
 162         else
 163                 sc->chunk_size_shift = __ffs(chunk_size);
 164
 165         /*
 166          * Get the stripe destinations.
 167          */
 168         for (i = 0; i < stripes; i++) {
 169                 argv += 2;
 170
 171                 r = get_stripe(ti, sc, i, argv);
 172                 if (r < 0) {
 173                         ti->error = "Couldn't parse stripe destination";
 174                         while (i--)
 175                                 dm_put_device(ti, sc->stripe[i].dev);
 176                         kfree(sc);
 177                         return r;
 178                 }
 179                 atomic_set(&(sc->stripe[i].error_count), 0);
 180         }
 181
 182         ti->private = sc;
 183
 184         return 0;
 185 }
 186
 187 static void stripe_dtr(struct dm_target *ti)
 188 {
 189         unsigned int i;
 190         struct stripe_c *sc = (struct stripe_c *) ti->private;
 191
 192         for (i = 0; i < sc->stripes; i++)
 193                 dm_put_device(ti, sc->stripe[i].dev);
 194
 195         flush_work(&sc->trigger_event);
 196         kfree(sc);
 197 }
 198
 199 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 200                               uint32_t *stripe, sector_t *result)
 201 {
 202         sector_t chunk = dm_target_offset(sc->ti, sector);
 203         sector_t chunk_offset;
 204
 205         if (sc->chunk_size_shift < 0)
 206                 chunk_offset = sector_div(chunk, sc->chunk_size);
 207         else {
 208                 chunk_offset = chunk & (sc->chunk_size - 1);
 209                 chunk >>= sc->chunk_size_shift;
 210         }
 211
 212         if (sc->stripes_shift < 0)
 213                 *stripe = sector_div(chunk, sc->stripes);
 214         else {
 215                 *stripe = chunk & (sc->stripes - 1);
 216                 chunk >>= sc->stripes_shift;
 217         }
 218
 219         if (sc->chunk_size_shift < 0)
 220                 chunk *= sc->chunk_size;
 221         else
 222                 chunk <<= sc->chunk_size_shift;
 223
 224         *result = chunk + chunk_offset;
 225 }
 226
 227 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 228                                     uint32_t target_stripe, sector_t *result)
 229 {
 230         uint32_t stripe;
 231
 232         stripe_map_sector(sc, sector, &stripe, result);
 233         if (stripe == target_stripe)
 234                 return;
 235
 236         /* round down */
 237         sector = *result;
 238         if (sc->chunk_size_shift < 0)
 239                 *result -= sector_div(sector, sc->chunk_size);
 240         else
 241                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 242
 243         if (target_stripe < stripe)
 244                 *result += sc->chunk_size;              /* next chunk */
 245 }
 246
 247 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 248                             uint32_t target_stripe)
 249 {
 250         sector_t begin, end;
 251
 252         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 253                                 target_stripe, &begin);
 254         stripe_map_range_sector(sc, bio_end_sector(bio),
 255                                 target_stripe, &end);
 256         if (begin < end) {
 257                 bio_set_dev(bio, sc->stripe[target_stripe].dev->bdev);
 258                 bio->bi_iter.bi_sector = begin +
 259                         sc->stripe[target_stripe].physical_start;
 260                 bio->bi_iter.bi_size = to_bytes(end - begin);
 261                 return DM_MAPIO_REMAPPED;
 262         } else {
 263                 /* The range doesn't map to the target stripe */
 264                 bio_endio(bio);
 265                 return DM_MAPIO_SUBMITTED;
 266         }
 267 }
 268
 269 static int stripe_map(struct dm_target *ti, struct bio *bio)
 270 {
 271         struct stripe_c *sc = ti->private;
 272         uint32_t stripe;
 273         unsigned int target_bio_nr;
 274
 275         if (bio->bi_opf & REQ_PREFLUSH) {
 276                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 277                 BUG_ON(target_bio_nr >= sc->stripes);
 278                 bio_set_dev(bio, sc->stripe[target_bio_nr].dev->bdev);
 279                 return DM_MAPIO_REMAPPED;
 280         }
 281         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 282             unlikely(bio_op(bio) == REQ_OP_SECURE_ERASE) ||
 283             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES)) {
 284                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 285                 BUG_ON(target_bio_nr >= sc->stripes);
 286                 return stripe_map_range(sc, bio, target_bio_nr);
 287         }
 288
 289         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 290                           &stripe, &bio->bi_iter.bi_sector);
 291
 292         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 293         bio_set_dev(bio, sc->stripe[stripe].dev->bdev);
 294
 295         return DM_MAPIO_REMAPPED;
 296 }
 297
 298 #if IS_ENABLED(CONFIG_FS_DAX)
 299 static struct dax_device *stripe_dax_pgoff(struct dm_target *ti, pgoff_t *pgoff)
 300 {
 301         struct stripe_c *sc = ti->private;
 302         struct block_device *bdev;
 303         sector_t dev_sector;
 304         uint32_t stripe;
 305
 306         stripe_map_sector(sc, *pgoff * PAGE_SECTORS, &stripe, &dev_sector);
 307         dev_sector += sc->stripe[stripe].physical_start;
 308         bdev = sc->stripe[stripe].dev->bdev;
 309
 310         *pgoff = (get_start_sect(bdev) + dev_sector) >> PAGE_SECTORS_SHIFT;
 311         return sc->stripe[stripe].dev->dax_dev;
 312 }
 313
 314 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 315                 long nr_pages, enum dax_access_mode mode, void **kaddr,
 316                 pfn_t *pfn)
 317 {
 318         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 319
 320         return dax_direct_access(dax_dev, pgoff, nr_pages, mode, kaddr, pfn);
 321 }
 322
 323 static int stripe_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
 324                                       size_t nr_pages)
 325 {
 326         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 327
 328         return dax_zero_page_range(dax_dev, pgoff, nr_pages);
 329 }
 330
 331 static size_t stripe_dax_recovery_write(struct dm_target *ti, pgoff_t pgoff,
 332                 void *addr, size_t bytes, struct iov_iter *i)
 333 {
 334         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 335
 336         return dax_recovery_write(dax_dev, pgoff, addr, bytes, i);
 337 }
 338
 339 #else
 340 #define stripe_dax_direct_access NULL
 341 #define stripe_dax_zero_page_range NULL
 342 #define stripe_dax_recovery_write NULL
 343 #endif
 344
 345 /*
 346  * Stripe status:
 347  *
 348  * INFO
 349  * #stripes [stripe_name <stripe_name>] [group word count]
 350  * [error count 'A|D' <error count 'A|D'>]
 351  *
 352  * TABLE
 353  * #stripes [stripe chunk size]
 354  * [stripe_name physical_start <stripe_name physical_start>]
 355  *
 356  */
 357
 358 static void stripe_status(struct dm_target *ti, status_type_t type,
 359                           unsigned int status_flags, char *result, unsigned int maxlen)
 360 {
 361         struct stripe_c *sc = (struct stripe_c *) ti->private;
 362         unsigned int sz = 0;
 363         unsigned int i;
 364
 365         switch (type) {
 366         case STATUSTYPE_INFO:
 367                 DMEMIT("%d ", sc->stripes);
 368                 for (i = 0; i < sc->stripes; i++)
 369                         DMEMIT("%s ", sc->stripe[i].dev->name);
 370
 371                 DMEMIT("1 ");
 372                 for (i = 0; i < sc->stripes; i++)
 373                         DMEMIT("%c", atomic_read(&(sc->stripe[i].error_count)) ?  'D' : 'A');
 374                 break;
 375
 376         case STATUSTYPE_TABLE:
 377                 DMEMIT("%d %llu", sc->stripes,
 378                         (unsigned long long)sc->chunk_size);
 379                 for (i = 0; i < sc->stripes; i++)
 380                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 381                             (unsigned long long)sc->stripe[i].physical_start);
 382                 break;
 383
 384         case STATUSTYPE_IMA:
 385                 DMEMIT_TARGET_NAME_VERSION(ti->type);
 386                 DMEMIT(",stripes=%d,chunk_size=%llu", sc->stripes,
 387                        (unsigned long long)sc->chunk_size);
 388
 389                 for (i = 0; i < sc->stripes; i++) {
 390                         DMEMIT(",stripe_%d_device_name=%s", i, sc->stripe[i].dev->name);
 391                         DMEMIT(",stripe_%d_physical_start=%llu", i,
 392                                (unsigned long long)sc->stripe[i].physical_start);
 393                         DMEMIT(",stripe_%d_status=%c", i,
 394                                atomic_read(&(sc->stripe[i].error_count)) ? 'D' : 'A');
 395                 }
 396                 DMEMIT(";");
 397                 break;
 398         }
 399 }
 400
 401 static int stripe_end_io(struct dm_target *ti, struct bio *bio,
 402                 blk_status_t *error)
 403 {
 404         unsigned int i;
 405         char major_minor[16];
 406         struct stripe_c *sc = ti->private;
 407
 408         if (!*error)
 409                 return DM_ENDIO_DONE; /* I/O complete */
 410
 411         if (bio->bi_opf & REQ_RAHEAD)
 412                 return DM_ENDIO_DONE;
 413
 414         if (*error == BLK_STS_NOTSUPP)
 415                 return DM_ENDIO_DONE;
 416
 417         memset(major_minor, 0, sizeof(major_minor));
 418         sprintf(major_minor, "%d:%d", MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)));
 419
 420         /*
 421          * Test to see which stripe drive triggered the event
 422          * and increment error count for all stripes on that device.
 423          * If the error count for a given device exceeds the threshold
 424          * value we will no longer trigger any further events.
 425          */
 426         for (i = 0; i < sc->stripes; i++)
 427                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 428                         atomic_inc(&(sc->stripe[i].error_count));
 429                         if (atomic_read(&(sc->stripe[i].error_count)) <
 430                             DM_IO_ERROR_THRESHOLD)
 431                                 schedule_work(&sc->trigger_event);
 432                 }
 433
 434         return DM_ENDIO_DONE;
 435 }
 436
 437 static int stripe_iterate_devices(struct dm_target *ti,
 438                                   iterate_devices_callout_fn fn, void *data)
 439 {
 440         struct stripe_c *sc = ti->private;
 441         int ret = 0;
 442         unsigned int i = 0;
 443
 444         do {
 445                 ret = fn(ti, sc->stripe[i].dev,
 446                          sc->stripe[i].physical_start,
 447                          sc->stripe_width, data);
 448         } while (!ret && ++i < sc->stripes);
 449
 450         return ret;
 451 }
 452
 453 static void stripe_io_hints(struct dm_target *ti,
 454                             struct queue_limits *limits)
 455 {
 456         struct stripe_c *sc = ti->private;
 457         unsigned int chunk_size = sc->chunk_size << SECTOR_SHIFT;
 458
 459         blk_limits_io_min(limits, chunk_size);
 460         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 461 }
 462
 463 static struct target_type stripe_target = {
 464         .name   = "striped",
 465         .version = {1, 6, 0},
 466         .features = DM_TARGET_PASSES_INTEGRITY | DM_TARGET_NOWAIT,
 467         .module = THIS_MODULE,
 468         .ctr    = stripe_ctr,
 469         .dtr    = stripe_dtr,
 470         .map    = stripe_map,
 471         .end_io = stripe_end_io,
 472         .status = stripe_status,
 473         .iterate_devices = stripe_iterate_devices,
 474         .io_hints = stripe_io_hints,
 475         .direct_access = stripe_dax_direct_access,
 476         .dax_zero_page_range = stripe_dax_zero_page_range,
 477         .dax_recovery_write = stripe_dax_recovery_write,
 478 };
 479
 480 int __init dm_stripe_init(void)
 481 {
 482         int r;
 483
 484         r = dm_register_target(&stripe_target);
 485         if (r < 0)
 486                 DMWARN("target registration failed");
 487
 488         return r;
 489 }
 490
 491 void dm_stripe_exit(void)
 492 {
 493         dm_unregister_target(&stripe_target);
 494 }