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_zeroes_bios = stripes;
 161
 162         sc->chunk_size = chunk_size;
 163         if (chunk_size & (chunk_size - 1))
 164                 sc->chunk_size_shift = -1;
 165         else
 166                 sc->chunk_size_shift = __ffs(chunk_size);
 167
 168         /*
 169          * Get the stripe destinations.
 170          */
 171         for (i = 0; i < stripes; i++) {
 172                 argv += 2;
 173
 174                 r = get_stripe(ti, sc, i, argv);
 175                 if (r < 0) {
 176                         ti->error = "Couldn't parse stripe destination";
 177                         while (i--)
 178                                 dm_put_device(ti, sc->stripe[i].dev);
 179                         kfree(sc);
 180                         return r;
 181                 }
 182                 atomic_set(&(sc->stripe[i].error_count), 0);
 183         }
 184
 185         ti->private = sc;
 186
 187         return 0;
 188 }
 189
 190 static void stripe_dtr(struct dm_target *ti)
 191 {
 192         unsigned int i;
 193         struct stripe_c *sc = (struct stripe_c *) ti->private;
 194
 195         for (i = 0; i < sc->stripes; i++)
 196                 dm_put_device(ti, sc->stripe[i].dev);
 197
 198         flush_work(&sc->trigger_event);
 199         kfree(sc);
 200 }
 201
 202 static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
 203                               uint32_t *stripe, sector_t *result)
 204 {
 205         sector_t chunk = dm_target_offset(sc->ti, sector);
 206         sector_t chunk_offset;
 207
 208         if (sc->chunk_size_shift < 0)
 209                 chunk_offset = sector_div(chunk, sc->chunk_size);
 210         else {
 211                 chunk_offset = chunk & (sc->chunk_size - 1);
 212                 chunk >>= sc->chunk_size_shift;
 213         }
 214
 215         if (sc->stripes_shift < 0)
 216                 *stripe = sector_div(chunk, sc->stripes);
 217         else {
 218                 *stripe = chunk & (sc->stripes - 1);
 219                 chunk >>= sc->stripes_shift;
 220         }
 221
 222         if (sc->chunk_size_shift < 0)
 223                 chunk *= sc->chunk_size;
 224         else
 225                 chunk <<= sc->chunk_size_shift;
 226
 227         *result = chunk + chunk_offset;
 228 }
 229
 230 static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
 231                                     uint32_t target_stripe, sector_t *result)
 232 {
 233         uint32_t stripe;
 234
 235         stripe_map_sector(sc, sector, &stripe, result);
 236         if (stripe == target_stripe)
 237                 return;
 238
 239         /* round down */
 240         sector = *result;
 241         if (sc->chunk_size_shift < 0)
 242                 *result -= sector_div(sector, sc->chunk_size);
 243         else
 244                 *result = sector & ~(sector_t)(sc->chunk_size - 1);
 245
 246         if (target_stripe < stripe)
 247                 *result += sc->chunk_size;              /* next chunk */
 248 }
 249
 250 static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
 251                             uint32_t target_stripe)
 252 {
 253         sector_t begin, end;
 254
 255         stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
 256                                 target_stripe, &begin);
 257         stripe_map_range_sector(sc, bio_end_sector(bio),
 258                                 target_stripe, &end);
 259         if (begin < end) {
 260                 bio_set_dev(bio, sc->stripe[target_stripe].dev->bdev);
 261                 bio->bi_iter.bi_sector = begin +
 262                         sc->stripe[target_stripe].physical_start;
 263                 bio->bi_iter.bi_size = to_bytes(end - begin);
 264                 return DM_MAPIO_REMAPPED;
 265         } else {
 266                 /* The range doesn't map to the target stripe */
 267                 bio_endio(bio);
 268                 return DM_MAPIO_SUBMITTED;
 269         }
 270 }
 271
 272 static int stripe_map(struct dm_target *ti, struct bio *bio)
 273 {
 274         struct stripe_c *sc = ti->private;
 275         uint32_t stripe;
 276         unsigned target_bio_nr;
 277
 278         if (bio->bi_opf & REQ_PREFLUSH) {
 279                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 280                 BUG_ON(target_bio_nr >= sc->stripes);
 281                 bio_set_dev(bio, sc->stripe[target_bio_nr].dev->bdev);
 282                 return DM_MAPIO_REMAPPED;
 283         }
 284         if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
 285             unlikely(bio_op(bio) == REQ_OP_SECURE_ERASE) ||
 286             unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES)) {
 287                 target_bio_nr = dm_bio_get_target_bio_nr(bio);
 288                 BUG_ON(target_bio_nr >= sc->stripes);
 289                 return stripe_map_range(sc, bio, target_bio_nr);
 290         }
 291
 292         stripe_map_sector(sc, bio->bi_iter.bi_sector,
 293                           &stripe, &bio->bi_iter.bi_sector);
 294
 295         bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
 296         bio_set_dev(bio, sc->stripe[stripe].dev->bdev);
 297
 298         return DM_MAPIO_REMAPPED;
 299 }
 300
 301 #if IS_ENABLED(CONFIG_FS_DAX)
 302 static struct dax_device *stripe_dax_pgoff(struct dm_target *ti, pgoff_t *pgoff)
 303 {
 304         struct stripe_c *sc = ti->private;
 305         struct block_device *bdev;
 306         sector_t dev_sector;
 307         uint32_t stripe;
 308
 309         stripe_map_sector(sc, *pgoff * PAGE_SECTORS, &stripe, &dev_sector);
 310         dev_sector += sc->stripe[stripe].physical_start;
 311         bdev = sc->stripe[stripe].dev->bdev;
 312
 313         *pgoff = (get_start_sect(bdev) + dev_sector) >> PAGE_SECTORS_SHIFT;
 314         return sc->stripe[stripe].dev->dax_dev;
 315 }
 316
 317 static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 318                 long nr_pages, enum dax_access_mode mode, void **kaddr,
 319                 pfn_t *pfn)
 320 {
 321         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 322
 323         return dax_direct_access(dax_dev, pgoff, nr_pages, mode, kaddr, pfn);
 324 }
 325
 326 static int stripe_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
 327                                       size_t nr_pages)
 328 {
 329         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 330
 331         return dax_zero_page_range(dax_dev, pgoff, nr_pages);
 332 }
 333
 334 static size_t stripe_dax_recovery_write(struct dm_target *ti, pgoff_t pgoff,
 335                 void *addr, size_t bytes, struct iov_iter *i)
 336 {
 337         struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
 338
 339         return dax_recovery_write(dax_dev, pgoff, addr, bytes, i);
 340 }
 341
 342 #else
 343 #define stripe_dax_direct_access NULL
 344 #define stripe_dax_zero_page_range NULL
 345 #define stripe_dax_recovery_write NULL
 346 #endif
 347
 348 /*
 349  * Stripe status:
 350  *
 351  * INFO
 352  * #stripes [stripe_name <stripe_name>] [group word count]
 353  * [error count 'A|D' <error count 'A|D'>]
 354  *
 355  * TABLE
 356  * #stripes [stripe chunk size]
 357  * [stripe_name physical_start <stripe_name physical_start>]
 358  *
 359  */
 360
 361 static void stripe_status(struct dm_target *ti, status_type_t type,
 362                           unsigned status_flags, char *result, unsigned maxlen)
 363 {
 364         struct stripe_c *sc = (struct stripe_c *) ti->private;
 365         unsigned int sz = 0;
 366         unsigned int i;
 367
 368         switch (type) {
 369         case STATUSTYPE_INFO:
 370                 DMEMIT("%d ", sc->stripes);
 371                 for (i = 0; i < sc->stripes; i++)  {
 372                         DMEMIT("%s ", sc->stripe[i].dev->name);
 373                 }
 374                 DMEMIT("1 ");
 375                 for (i = 0; i < sc->stripes; i++) {
 376                         DMEMIT("%c", atomic_read(&(sc->stripe[i].error_count)) ?
 377                                'D' : 'A');
 378                 }
 379                 break;
 380
 381         case STATUSTYPE_TABLE:
 382                 DMEMIT("%d %llu", sc->stripes,
 383                         (unsigned long long)sc->chunk_size);
 384                 for (i = 0; i < sc->stripes; i++)
 385                         DMEMIT(" %s %llu", sc->stripe[i].dev->name,
 386                             (unsigned long long)sc->stripe[i].physical_start);
 387                 break;
 388
 389         case STATUSTYPE_IMA:
 390                 DMEMIT_TARGET_NAME_VERSION(ti->type);
 391                 DMEMIT(",stripes=%d,chunk_size=%llu", sc->stripes,
 392                        (unsigned long long)sc->chunk_size);
 393
 394                 for (i = 0; i < sc->stripes; i++) {
 395                         DMEMIT(",stripe_%d_device_name=%s", i, sc->stripe[i].dev->name);
 396                         DMEMIT(",stripe_%d_physical_start=%llu", i,
 397                                (unsigned long long)sc->stripe[i].physical_start);
 398                         DMEMIT(",stripe_%d_status=%c", i,
 399                                atomic_read(&(sc->stripe[i].error_count)) ? 'D' : 'A');
 400                 }
 401                 DMEMIT(";");
 402                 break;
 403         }
 404 }
 405
 406 static int stripe_end_io(struct dm_target *ti, struct bio *bio,
 407                 blk_status_t *error)
 408 {
 409         unsigned i;
 410         char major_minor[16];
 411         struct stripe_c *sc = ti->private;
 412
 413         if (!*error)
 414                 return DM_ENDIO_DONE; /* I/O complete */
 415
 416         if (bio->bi_opf & REQ_RAHEAD)
 417                 return DM_ENDIO_DONE;
 418
 419         if (*error == BLK_STS_NOTSUPP)
 420                 return DM_ENDIO_DONE;
 421
 422         memset(major_minor, 0, sizeof(major_minor));
 423         sprintf(major_minor, "%d:%d", MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)));
 424
 425         /*
 426          * Test to see which stripe drive triggered the event
 427          * and increment error count for all stripes on that device.
 428          * If the error count for a given device exceeds the threshold
 429          * value we will no longer trigger any further events.
 430          */
 431         for (i = 0; i < sc->stripes; i++)
 432                 if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
 433                         atomic_inc(&(sc->stripe[i].error_count));
 434                         if (atomic_read(&(sc->stripe[i].error_count)) <
 435                             DM_IO_ERROR_THRESHOLD)
 436                                 schedule_work(&sc->trigger_event);
 437                 }
 438
 439         return DM_ENDIO_DONE;
 440 }
 441
 442 static int stripe_iterate_devices(struct dm_target *ti,
 443                                   iterate_devices_callout_fn fn, void *data)
 444 {
 445         struct stripe_c *sc = ti->private;
 446         int ret = 0;
 447         unsigned i = 0;
 448
 449         do {
 450                 ret = fn(ti, sc->stripe[i].dev,
 451                          sc->stripe[i].physical_start,
 452                          sc->stripe_width, data);
 453         } while (!ret && ++i < sc->stripes);
 454
 455         return ret;
 456 }
 457
 458 static void stripe_io_hints(struct dm_target *ti,
 459                             struct queue_limits *limits)
 460 {
 461         struct stripe_c *sc = ti->private;
 462         unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
 463
 464         blk_limits_io_min(limits, chunk_size);
 465         blk_limits_io_opt(limits, chunk_size * sc->stripes);
 466 }
 467
 468 static struct target_type stripe_target = {
 469         .name   = "striped",
 470         .version = {1, 6, 0},
 471         .features = DM_TARGET_PASSES_INTEGRITY | DM_TARGET_NOWAIT,
 472         .module = THIS_MODULE,
 473         .ctr    = stripe_ctr,
 474         .dtr    = stripe_dtr,
 475         .map    = stripe_map,
 476         .end_io = stripe_end_io,
 477         .status = stripe_status,
 478         .iterate_devices = stripe_iterate_devices,
 479         .io_hints = stripe_io_hints,
 480         .direct_access = stripe_dax_direct_access,
 481         .dax_zero_page_range = stripe_dax_zero_page_range,
 482         .dax_recovery_write = stripe_dax_recovery_write,
 483 };
 484
 485 int __init dm_stripe_init(void)
 486 {
 487         int r;
 488
 489         r = dm_register_target(&stripe_target);
 490         if (r < 0)
 491                 DMWARN("target registration failed");
 492
 493         return r;
 494 }
 495
 496 void dm_stripe_exit(void)
 497 {
 498         dm_unregister_target(&stripe_target);
 499 }