drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  */
  11
  12 /*
  13  * Still to do:
  14  *
  15  * flush after percent set rather than just time based. (maybe both).
  16  */
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/errno.h>
  21 #include <linux/slab.h>
  22 #include <linux/init.h>
  23 #include <linux/timer.h>
  24 #include <linux/sched.h>
  25 #include <linux/list.h>
  26 #include <linux/file.h>
  27 #include <linux/mount.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/seq_file.h>
  30 #include "md.h"
  31 #include "bitmap.h"
  32
  33 static inline char *bmname(struct bitmap *bitmap)
  34 {
  35         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  36 }
  37
  38 /*
  39  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
  40  *
  41  * 1) check to see if this page is allocated, if it's not then try to alloc
  42  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
  43  *    page pointer directly as a counter
  44  *
  45  * if we find our page, we increment the page's refcount so that it stays
  46  * allocated while we're using it
  47  */
  48 static int bitmap_checkpage(struct bitmap_counts *bitmap,
  49                             unsigned long page, int create)
  50 __releases(bitmap->lock)
  51 __acquires(bitmap->lock)
  52 {
  53         unsigned char *mappage;
  54
  55         if (page >= bitmap->pages) {
  56                 /* This can happen if bitmap_start_sync goes beyond
  57                  * End-of-device while looking for a whole page.
  58                  * It is harmless.
  59                  */
  60                 return -EINVAL;
  61         }
  62
  63         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
  64                 return 0;
  65
  66         if (bitmap->bp[page].map) /* page is already allocated, just return */
  67                 return 0;
  68
  69         if (!create)
  70                 return -ENOENT;
  71
  72         /* this page has not been allocated yet */
  73
  74         spin_unlock_irq(&bitmap->lock);
  75         mappage = kzalloc(PAGE_SIZE, GFP_NOIO);
  76         spin_lock_irq(&bitmap->lock);
  77
  78         if (mappage == NULL) {
  79                 pr_debug("md/bitmap: map page allocation failed, hijacking\n");
  80                 /* failed - set the hijacked flag so that we can use the
  81                  * pointer as a counter */
  82                 if (!bitmap->bp[page].map)
  83                         bitmap->bp[page].hijacked = 1;
  84         } else if (bitmap->bp[page].map ||
  85                    bitmap->bp[page].hijacked) {
  86                 /* somebody beat us to getting the page */
  87                 kfree(mappage);
  88                 return 0;
  89         } else {
  90
  91                 /* no page was in place and we have one, so install it */
  92
  93                 bitmap->bp[page].map = mappage;
  94                 bitmap->missing_pages--;
  95         }
  96         return 0;
  97 }
  98
  99 /* if page is completely empty, put it back on the free list, or dealloc it */
 100 /* if page was hijacked, unmark the flag so it might get alloced next time */
 101 /* Note: lock should be held when calling this */
 102 static void bitmap_checkfree(struct bitmap_counts *bitmap, unsigned long page)
 103 {
 104         char *ptr;
 105
 106         if (bitmap->bp[page].count) /* page is still busy */
 107                 return;
 108
 109         /* page is no longer in use, it can be released */
 110
 111         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 112                 bitmap->bp[page].hijacked = 0;
 113                 bitmap->bp[page].map = NULL;
 114         } else {
 115                 /* normal case, free the page */
 116                 ptr = bitmap->bp[page].map;
 117                 bitmap->bp[page].map = NULL;
 118                 bitmap->missing_pages++;
 119                 kfree(ptr);
 120         }
 121 }
 122
 123 /*
 124  * bitmap file handling - read and write the bitmap file and its superblock
 125  */
 126
 127 /*
 128  * basic page I/O operations
 129  */
 130
 131 /* IO operations when bitmap is stored near all superblocks */
 132 static int read_sb_page(struct mddev *mddev, loff_t offset,
 133                         struct page *page,
 134                         unsigned long index, int size)
 135 {
 136         /* choose a good rdev and read the page from there */
 137
 138         struct md_rdev *rdev;
 139         sector_t target;
 140
 141         rdev_for_each(rdev, mddev) {
 142                 if (! test_bit(In_sync, &rdev->flags)
 143                     || test_bit(Faulty, &rdev->flags))
 144                         continue;
 145
 146                 target = offset + index * (PAGE_SIZE/512);
 147
 148                 if (sync_page_io(rdev, target,
 149                                  roundup(size, bdev_logical_block_size(rdev->bdev)),
 150                                  page, READ, true)) {
 151                         page->index = index;
 152                         return 0;
 153                 }
 154         }
 155         return -EIO;
 156 }
 157
 158 static struct md_rdev *next_active_rdev(struct md_rdev *rdev, struct mddev *mddev)
 159 {
 160         /* Iterate the disks of an mddev, using rcu to protect access to the
 161          * linked list, and raising the refcount of devices we return to ensure
 162          * they don't disappear while in use.
 163          * As devices are only added or removed when raid_disk is < 0 and
 164          * nr_pending is 0 and In_sync is clear, the entries we return will
 165          * still be in the same position on the list when we re-enter
 166          * list_for_each_entry_continue_rcu.
 167          */
 168         rcu_read_lock();
 169         if (rdev == NULL)
 170                 /* start at the beginning */
 171                 rdev = list_entry_rcu(&mddev->disks, struct md_rdev, same_set);
 172         else {
 173                 /* release the previous rdev and start from there. */
 174                 rdev_dec_pending(rdev, mddev);
 175         }
 176         list_for_each_entry_continue_rcu(rdev, &mddev->disks, same_set) {
 177                 if (rdev->raid_disk >= 0 &&
 178                     !test_bit(Faulty, &rdev->flags)) {
 179                         /* this is a usable devices */
 180                         atomic_inc(&rdev->nr_pending);
 181                         rcu_read_unlock();
 182                         return rdev;
 183                 }
 184         }
 185         rcu_read_unlock();
 186         return NULL;
 187 }
 188
 189 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 190 {
 191         struct md_rdev *rdev = NULL;
 192         struct block_device *bdev;
 193         struct mddev *mddev = bitmap->mddev;
 194         struct bitmap_storage *store = &bitmap->storage;
 195
 196         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 197                 int size = PAGE_SIZE;
 198                 loff_t offset = mddev->bitmap_info.offset;
 199
 200                 bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
 201
 202                 if (page->index == store->file_pages-1) {
 203                         int last_page_size = store->bytes & (PAGE_SIZE-1);
 204                         if (last_page_size == 0)
 205                                 last_page_size = PAGE_SIZE;
 206                         size = roundup(last_page_size,
 207                                        bdev_logical_block_size(bdev));
 208                 }
 209                 /* Just make sure we aren't corrupting data or
 210                  * metadata
 211                  */
 212                 if (mddev->external) {
 213                         /* Bitmap could be anywhere. */
 214                         if (rdev->sb_start + offset + (page->index
 215                                                        * (PAGE_SIZE/512))
 216                             > rdev->data_offset
 217                             &&
 218                             rdev->sb_start + offset
 219                             < (rdev->data_offset + mddev->dev_sectors
 220                              + (PAGE_SIZE/512)))
 221                                 goto bad_alignment;
 222                 } else if (offset < 0) {
 223                         /* DATA  BITMAP METADATA  */
 224                         if (offset
 225                             + (long)(page->index * (PAGE_SIZE/512))
 226                             + size/512 > 0)
 227                                 /* bitmap runs in to metadata */
 228                                 goto bad_alignment;
 229                         if (rdev->data_offset + mddev->dev_sectors
 230                             > rdev->sb_start + offset)
 231                                 /* data runs in to bitmap */
 232                                 goto bad_alignment;
 233                 } else if (rdev->sb_start < rdev->data_offset) {
 234                         /* METADATA BITMAP DATA */
 235                         if (rdev->sb_start
 236                             + offset
 237                             + page->index*(PAGE_SIZE/512) + size/512
 238                             > rdev->data_offset)
 239                                 /* bitmap runs in to data */
 240                                 goto bad_alignment;
 241                 } else {
 242                         /* DATA METADATA BITMAP - no problems */
 243                 }
 244                 md_super_write(mddev, rdev,
 245                                rdev->sb_start + offset
 246                                + page->index * (PAGE_SIZE/512),
 247                                size,
 248                                page);
 249         }
 250
 251         if (wait)
 252                 md_super_wait(mddev);
 253         return 0;
 254
 255  bad_alignment:
 256         return -EINVAL;
 257 }
 258
 259 static void bitmap_file_kick(struct bitmap *bitmap);
 260 /*
 261  * write out a page to a file
 262  */
 263 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 264 {
 265         struct buffer_head *bh;
 266
 267         if (bitmap->storage.file == NULL) {
 268                 switch (write_sb_page(bitmap, page, wait)) {
 269                 case -EINVAL:
 270                         set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 271                 }
 272         } else {
 273
 274                 bh = page_buffers(page);
 275
 276                 while (bh && bh->b_blocknr) {
 277                         atomic_inc(&bitmap->pending_writes);
 278                         set_buffer_locked(bh);
 279                         set_buffer_mapped(bh);
 280                         submit_bh(WRITE | REQ_SYNC, bh);
 281                         bh = bh->b_this_page;
 282                 }
 283
 284                 if (wait)
 285                         wait_event(bitmap->write_wait,
 286                                    atomic_read(&bitmap->pending_writes)==0);
 287         }
 288         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 289                 bitmap_file_kick(bitmap);
 290 }
 291
 292 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 293 {
 294         struct bitmap *bitmap = bh->b_private;
 295
 296         if (!uptodate)
 297                 set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 298         if (atomic_dec_and_test(&bitmap->pending_writes))
 299                 wake_up(&bitmap->write_wait);
 300 }
 301
 302 /* copied from buffer.c */
 303 static void
 304 __clear_page_buffers(struct page *page)
 305 {
 306         ClearPagePrivate(page);
 307         set_page_private(page, 0);
 308         page_cache_release(page);
 309 }
 310 static void free_buffers(struct page *page)
 311 {
 312         struct buffer_head *bh;
 313
 314         if (!PagePrivate(page))
 315                 return;
 316
 317         bh = page_buffers(page);
 318         while (bh) {
 319                 struct buffer_head *next = bh->b_this_page;
 320                 free_buffer_head(bh);
 321                 bh = next;
 322         }
 323         __clear_page_buffers(page);
 324         put_page(page);
 325 }
 326
 327 /* read a page from a file.
 328  * We both read the page, and attach buffers to the page to record the
 329  * address of each block (using bmap).  These addresses will be used
 330  * to write the block later, completely bypassing the filesystem.
 331  * This usage is similar to how swap files are handled, and allows us
 332  * to write to a file with no concerns of memory allocation failing.
 333  */
 334 static int read_page(struct file *file, unsigned long index,
 335                      struct bitmap *bitmap,
 336                      unsigned long count,
 337                      struct page *page)
 338 {
 339         int ret = 0;
 340         struct inode *inode = file->f_path.dentry->d_inode;
 341         struct buffer_head *bh;
 342         sector_t block;
 343
 344         pr_debug("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 345                  (unsigned long long)index << PAGE_SHIFT);
 346
 347         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 348         if (!bh) {
 349                 ret = -ENOMEM;
 350                 goto out;
 351         }
 352         attach_page_buffers(page, bh);
 353         block = index << (PAGE_SHIFT - inode->i_blkbits);
 354         while (bh) {
 355                 if (count == 0)
 356                         bh->b_blocknr = 0;
 357                 else {
 358                         bh->b_blocknr = bmap(inode, block);
 359                         if (bh->b_blocknr == 0) {
 360                                 /* Cannot use this file! */
 361                                 ret = -EINVAL;
 362                                 goto out;
 363                         }
 364                         bh->b_bdev = inode->i_sb->s_bdev;
 365                         if (count < (1<<inode->i_blkbits))
 366                                 count = 0;
 367                         else
 368                                 count -= (1<<inode->i_blkbits);
 369
 370                         bh->b_end_io = end_bitmap_write;
 371                         bh->b_private = bitmap;
 372                         atomic_inc(&bitmap->pending_writes);
 373                         set_buffer_locked(bh);
 374                         set_buffer_mapped(bh);
 375                         submit_bh(READ, bh);
 376                 }
 377                 block++;
 378                 bh = bh->b_this_page;
 379         }
 380         page->index = index;
 381
 382         wait_event(bitmap->write_wait,
 383                    atomic_read(&bitmap->pending_writes)==0);
 384         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 385                 ret = -EIO;
 386 out:
 387         if (ret)
 388                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %d\n",
 389                         (int)PAGE_SIZE,
 390                         (unsigned long long)index << PAGE_SHIFT,
 391                         ret);
 392         return ret;
 393 }
 394
 395 /*
 396  * bitmap file superblock operations
 397  */
 398
 399 /* update the event counter and sync the superblock to disk */
 400 void bitmap_update_sb(struct bitmap *bitmap)
 401 {
 402         bitmap_super_t *sb;
 403
 404         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 405                 return;
 406         if (bitmap->mddev->bitmap_info.external)
 407                 return;
 408         if (!bitmap->storage.sb_page) /* no superblock */
 409                 return;
 410         sb = kmap_atomic(bitmap->storage.sb_page);
 411         sb->events = cpu_to_le64(bitmap->mddev->events);
 412         if (bitmap->mddev->events < bitmap->events_cleared)
 413                 /* rocking back to read-only */
 414                 bitmap->events_cleared = bitmap->mddev->events;
 415         sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 416         sb->state = cpu_to_le32(bitmap->flags);
 417         /* Just in case these have been changed via sysfs: */
 418         sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 419         sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
 420         /* This might have been changed by a reshape */
 421         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 422         sb->chunksize = cpu_to_le32(bitmap->mddev->bitmap_info.chunksize);
 423         sb->sectors_reserved = cpu_to_le32(bitmap->mddev->
 424                                            bitmap_info.space);
 425         kunmap_atomic(sb);
 426         write_page(bitmap, bitmap->storage.sb_page, 1);
 427 }
 428
 429 /* print out the bitmap file superblock */
 430 void bitmap_print_sb(struct bitmap *bitmap)
 431 {
 432         bitmap_super_t *sb;
 433
 434         if (!bitmap || !bitmap->storage.sb_page)
 435                 return;
 436         sb = kmap_atomic(bitmap->storage.sb_page);
 437         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 438         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 439         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 440         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 441                                         *(__u32 *)(sb->uuid+0),
 442                                         *(__u32 *)(sb->uuid+4),
 443                                         *(__u32 *)(sb->uuid+8),
 444                                         *(__u32 *)(sb->uuid+12));
 445         printk(KERN_DEBUG "        events: %llu\n",
 446                         (unsigned long long) le64_to_cpu(sb->events));
 447         printk(KERN_DEBUG "events cleared: %llu\n",
 448                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 449         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 450         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 451         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 452         printk(KERN_DEBUG "     sync size: %llu KB\n",
 453                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 454         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 455         kunmap_atomic(sb);
 456 }
 457
 458 /*
 459  * bitmap_new_disk_sb
 460  * @bitmap
 461  *
 462  * This function is somewhat the reverse of bitmap_read_sb.  bitmap_read_sb
 463  * reads and verifies the on-disk bitmap superblock and populates bitmap_info.
 464  * This function verifies 'bitmap_info' and populates the on-disk bitmap
 465  * structure, which is to be written to disk.
 466  *
 467  * Returns: 0 on success, -Exxx on error
 468  */
 469 static int bitmap_new_disk_sb(struct bitmap *bitmap)
 470 {
 471         bitmap_super_t *sb;
 472         unsigned long chunksize, daemon_sleep, write_behind;
 473         int err = -EINVAL;
 474
 475         bitmap->storage.sb_page = alloc_page(GFP_KERNEL);
 476         if (IS_ERR(bitmap->storage.sb_page)) {
 477                 err = PTR_ERR(bitmap->storage.sb_page);
 478                 bitmap->storage.sb_page = NULL;
 479                 return err;
 480         }
 481         bitmap->storage.sb_page->index = 0;
 482
 483         sb = kmap_atomic(bitmap->storage.sb_page);
 484
 485         sb->magic = cpu_to_le32(BITMAP_MAGIC);
 486         sb->version = cpu_to_le32(BITMAP_MAJOR_HI);
 487
 488         chunksize = bitmap->mddev->bitmap_info.chunksize;
 489         BUG_ON(!chunksize);
 490         if (!is_power_of_2(chunksize)) {
 491                 kunmap_atomic(sb);
 492                 printk(KERN_ERR "bitmap chunksize not a power of 2\n");
 493                 return -EINVAL;
 494         }
 495         sb->chunksize = cpu_to_le32(chunksize);
 496
 497         daemon_sleep = bitmap->mddev->bitmap_info.daemon_sleep;
 498         if (!daemon_sleep ||
 499             (daemon_sleep < 1) || (daemon_sleep > MAX_SCHEDULE_TIMEOUT)) {
 500                 printk(KERN_INFO "Choosing daemon_sleep default (5 sec)\n");
 501                 daemon_sleep = 5 * HZ;
 502         }
 503         sb->daemon_sleep = cpu_to_le32(daemon_sleep);
 504         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 505
 506         /*
 507          * FIXME: write_behind for RAID1.  If not specified, what
 508          * is a good choice?  We choose COUNTER_MAX / 2 arbitrarily.
 509          */
 510         write_behind = bitmap->mddev->bitmap_info.max_write_behind;
 511         if (write_behind > COUNTER_MAX)
 512                 write_behind = COUNTER_MAX / 2;
 513         sb->write_behind = cpu_to_le32(write_behind);
 514         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 515
 516         /* keep the array size field of the bitmap superblock up to date */
 517         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 518
 519         memcpy(sb->uuid, bitmap->mddev->uuid, 16);
 520
 521         set_bit(BITMAP_STALE, &bitmap->flags);
 522         sb->state = cpu_to_le32(bitmap->flags);
 523         bitmap->events_cleared = bitmap->mddev->events;
 524         sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 525
 526         kunmap_atomic(sb);
 527
 528         return 0;
 529 }
 530
 531 /* read the superblock from the bitmap file and initialize some bitmap fields */
 532 static int bitmap_read_sb(struct bitmap *bitmap)
 533 {
 534         char *reason = NULL;
 535         bitmap_super_t *sb;
 536         unsigned long chunksize, daemon_sleep, write_behind;
 537         unsigned long long events;
 538         unsigned long sectors_reserved = 0;
 539         int err = -EINVAL;
 540         struct page *sb_page;
 541
 542         if (!bitmap->storage.file && !bitmap->mddev->bitmap_info.offset) {
 543                 chunksize = 128 * 1024 * 1024;
 544                 daemon_sleep = 5 * HZ;
 545                 write_behind = 0;
 546                 set_bit(BITMAP_STALE, &bitmap->flags);
 547                 err = 0;
 548                 goto out_no_sb;
 549         }
 550         /* page 0 is the superblock, read it... */
 551         sb_page = alloc_page(GFP_KERNEL);
 552         if (!sb_page)
 553                 return -ENOMEM;
 554         bitmap->storage.sb_page = sb_page;
 555
 556         if (bitmap->storage.file) {
 557                 loff_t isize = i_size_read(bitmap->storage.file->f_mapping->host);
 558                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 559
 560                 err = read_page(bitmap->storage.file, 0,
 561                                 bitmap, bytes, sb_page);
 562         } else {
 563                 err = read_sb_page(bitmap->mddev,
 564                                    bitmap->mddev->bitmap_info.offset,
 565                                    sb_page,
 566                                    0, sizeof(bitmap_super_t));
 567         }
 568         if (err)
 569                 return err;
 570
 571         sb = kmap_atomic(sb_page);
 572
 573         chunksize = le32_to_cpu(sb->chunksize);
 574         daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 575         write_behind = le32_to_cpu(sb->write_behind);
 576         sectors_reserved = le32_to_cpu(sb->sectors_reserved);
 577
 578         /* verify that the bitmap-specific fields are valid */
 579         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 580                 reason = "bad magic";
 581         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 582                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 583                 reason = "unrecognized superblock version";
 584         else if (chunksize < 512)
 585                 reason = "bitmap chunksize too small";
 586         else if (!is_power_of_2(chunksize))
 587                 reason = "bitmap chunksize not a power of 2";
 588         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT)
 589                 reason = "daemon sleep period out of range";
 590         else if (write_behind > COUNTER_MAX)
 591                 reason = "write-behind limit out of range (0 - 16383)";
 592         if (reason) {
 593                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 594                         bmname(bitmap), reason);
 595                 goto out;
 596         }
 597
 598         /* keep the array size field of the bitmap superblock up to date */
 599         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 600
 601         if (bitmap->mddev->persistent) {
 602                 /*
 603                  * We have a persistent array superblock, so compare the
 604                  * bitmap's UUID and event counter to the mddev's
 605                  */
 606                 if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 607                         printk(KERN_INFO
 608                                "%s: bitmap superblock UUID mismatch\n",
 609                                bmname(bitmap));
 610                         goto out;
 611                 }
 612                 events = le64_to_cpu(sb->events);
 613                 if (events < bitmap->mddev->events) {
 614                         printk(KERN_INFO
 615                                "%s: bitmap file is out of date (%llu < %llu) "
 616                                "-- forcing full recovery\n",
 617                                bmname(bitmap), events,
 618                                (unsigned long long) bitmap->mddev->events);
 619                         set_bit(BITMAP_STALE, &bitmap->flags);
 620                 }
 621         }
 622
 623         /* assign fields using values from superblock */
 624         bitmap->flags |= le32_to_cpu(sb->state);
 625         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 626                 set_bit(BITMAP_HOSTENDIAN, &bitmap->flags);
 627         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 628         err = 0;
 629 out:
 630         kunmap_atomic(sb);
 631 out_no_sb:
 632         if (test_bit(BITMAP_STALE, &bitmap->flags))
 633                 bitmap->events_cleared = bitmap->mddev->events;
 634         bitmap->mddev->bitmap_info.chunksize = chunksize;
 635         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 636         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 637         if (bitmap->mddev->bitmap_info.space == 0 ||
 638             bitmap->mddev->bitmap_info.space > sectors_reserved)
 639                 bitmap->mddev->bitmap_info.space = sectors_reserved;
 640         if (err)
 641                 bitmap_print_sb(bitmap);
 642         return err;
 643 }
 644
 645 /*
 646  * general bitmap file operations
 647  */
 648
 649 /*
 650  * on-disk bitmap:
 651  *
 652  * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
 653  * file a page at a time. There's a superblock at the start of the file.
 654  */
 655 /* calculate the index of the page that contains this bit */
 656 static inline unsigned long file_page_index(struct bitmap_storage *store,
 657                                             unsigned long chunk)
 658 {
 659         if (store->sb_page)
 660                 chunk += sizeof(bitmap_super_t) << 3;
 661         return chunk >> PAGE_BIT_SHIFT;
 662 }
 663
 664 /* calculate the (bit) offset of this bit within a page */
 665 static inline unsigned long file_page_offset(struct bitmap_storage *store,
 666                                              unsigned long chunk)
 667 {
 668         if (store->sb_page)
 669                 chunk += sizeof(bitmap_super_t) << 3;
 670         return chunk & (PAGE_BITS - 1);
 671 }
 672
 673 /*
 674  * return a pointer to the page in the filemap that contains the given bit
 675  *
 676  * this lookup is complicated by the fact that the bitmap sb might be exactly
 677  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 678  * 0 or page 1
 679  */
 680 static inline struct page *filemap_get_page(struct bitmap_storage *store,
 681                                             unsigned long chunk)
 682 {
 683         if (file_page_index(store, chunk) >= store->file_pages)
 684                 return NULL;
 685         return store->filemap[file_page_index(store, chunk)
 686                               - file_page_index(store, 0)];
 687 }
 688
 689 static int bitmap_storage_alloc(struct bitmap_storage *store,
 690                                 unsigned long chunks, int with_super)
 691 {
 692         int pnum;
 693         unsigned long num_pages;
 694         unsigned long bytes;
 695
 696         bytes = DIV_ROUND_UP(chunks, 8);
 697         if (with_super)
 698                 bytes += sizeof(bitmap_super_t);
 699
 700         num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
 701
 702         store->filemap = kmalloc(sizeof(struct page *)
 703                                  * num_pages, GFP_KERNEL);
 704         if (!store->filemap)
 705                 return -ENOMEM;
 706
 707         if (with_super && !store->sb_page) {
 708                 store->sb_page = alloc_page(GFP_KERNEL|__GFP_ZERO);
 709                 if (store->sb_page == NULL)
 710                         return -ENOMEM;
 711                 store->sb_page->index = 0;
 712         }
 713         pnum = 0;
 714         if (store->sb_page) {
 715                 store->filemap[0] = store->sb_page;
 716                 pnum = 1;
 717         }
 718         for ( ; pnum < num_pages; pnum++) {
 719                 store->filemap[pnum] = alloc_page(GFP_KERNEL|__GFP_ZERO);
 720                 if (!store->filemap[pnum]) {
 721                         store->file_pages = pnum;
 722                         return -ENOMEM;
 723                 }
 724                 store->filemap[pnum]->index = pnum;
 725         }
 726         store->file_pages = pnum;
 727
 728         /* We need 4 bits per page, rounded up to a multiple
 729          * of sizeof(unsigned long) */
 730         store->filemap_attr = kzalloc(
 731                 roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
 732                 GFP_KERNEL);
 733         if (!store->filemap_attr)
 734                 return -ENOMEM;
 735
 736         store->bytes = bytes;
 737
 738         return 0;
 739 }
 740
 741 static void bitmap_file_unmap(struct bitmap_storage *store)
 742 {
 743         struct page **map, *sb_page;
 744         int pages;
 745         struct file *file;
 746
 747         file = store->file;
 748         map = store->filemap;
 749         pages = store->file_pages;
 750         sb_page = store->sb_page;
 751
 752         while (pages--)
 753                 if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 754                         free_buffers(map[pages]);
 755         kfree(map);
 756         kfree(store->filemap_attr);
 757
 758         if (sb_page)
 759                 free_buffers(sb_page);
 760
 761         if (file) {
 762                 struct inode *inode = file->f_path.dentry->d_inode;
 763                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 764                 fput(file);
 765         }
 766 }
 767
 768 /*
 769  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 770  * then it is no longer reliable, so we stop using it and we mark the file
 771  * as failed in the superblock
 772  */
 773 static void bitmap_file_kick(struct bitmap *bitmap)
 774 {
 775         char *path, *ptr = NULL;
 776
 777         if (!test_and_set_bit(BITMAP_STALE, &bitmap->flags)) {
 778                 bitmap_update_sb(bitmap);
 779
 780                 if (bitmap->storage.file) {
 781                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 782                         if (path)
 783                                 ptr = d_path(&bitmap->storage.file->f_path,
 784                                              path, PAGE_SIZE);
 785
 786                         printk(KERN_ALERT
 787                               "%s: kicking failed bitmap file %s from array!\n",
 788                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 789
 790                         kfree(path);
 791                 } else
 792                         printk(KERN_ALERT
 793                                "%s: disabling internal bitmap due to errors\n",
 794                                bmname(bitmap));
 795         }
 796 }
 797
 798 enum bitmap_page_attr {
 799         BITMAP_PAGE_DIRTY = 0,     /* there are set bits that need to be synced */
 800         BITMAP_PAGE_PENDING = 1,   /* there are bits that are being cleaned.
 801                                     * i.e. counter is 1 or 2. */
 802         BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 803 };
 804
 805 static inline void set_page_attr(struct bitmap *bitmap, int pnum,
 806                                  enum bitmap_page_attr attr)
 807 {
 808         set_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 809 }
 810
 811 static inline void clear_page_attr(struct bitmap *bitmap, int pnum,
 812                                    enum bitmap_page_attr attr)
 813 {
 814         clear_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 815 }
 816
 817 static inline int test_page_attr(struct bitmap *bitmap, int pnum,
 818                                  enum bitmap_page_attr attr)
 819 {
 820         return test_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 821 }
 822
 823 static inline int test_and_clear_page_attr(struct bitmap *bitmap, int pnum,
 824                                            enum bitmap_page_attr attr)
 825 {
 826         return test_and_clear_bit((pnum<<2) + attr,
 827                                   bitmap->storage.filemap_attr);
 828 }
 829 /*
 830  * bitmap_file_set_bit -- called before performing a write to the md device
 831  * to set (and eventually sync) a particular bit in the bitmap file
 832  *
 833  * we set the bit immediately, then we record the page number so that
 834  * when an unplug occurs, we can flush the dirty pages out to disk
 835  */
 836 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 837 {
 838         unsigned long bit;
 839         struct page *page;
 840         void *kaddr;
 841         unsigned long chunk = block >> bitmap->counts.chunkshift;
 842
 843         page = filemap_get_page(&bitmap->storage, chunk);
 844         if (!page)
 845                 return;
 846         bit = file_page_offset(&bitmap->storage, chunk);
 847
 848         /* set the bit */
 849         kaddr = kmap_atomic(page);
 850         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 851                 set_bit(bit, kaddr);
 852         else
 853                 test_and_set_bit_le(bit, kaddr);
 854         kunmap_atomic(kaddr);
 855         pr_debug("set file bit %lu page %lu\n", bit, page->index);
 856         /* record page number so it gets flushed to disk when unplug occurs */
 857         set_page_attr(bitmap, page->index, BITMAP_PAGE_DIRTY);
 858 }
 859
 860 static void bitmap_file_clear_bit(struct bitmap *bitmap, sector_t block)
 861 {
 862         unsigned long bit;
 863         struct page *page;
 864         void *paddr;
 865         unsigned long chunk = block >> bitmap->counts.chunkshift;
 866
 867         page = filemap_get_page(&bitmap->storage, chunk);
 868         if (!page)
 869                 return;
 870         bit = file_page_offset(&bitmap->storage, chunk);
 871         paddr = kmap_atomic(page);
 872         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 873                 clear_bit(bit, paddr);
 874         else
 875                 test_and_clear_bit_le(bit, paddr);
 876         kunmap_atomic(paddr);
 877         if (!test_page_attr(bitmap, page->index, BITMAP_PAGE_NEEDWRITE)) {
 878                 set_page_attr(bitmap, page->index, BITMAP_PAGE_PENDING);
 879                 bitmap->allclean = 0;
 880         }
 881 }
 882
 883 /* this gets called when the md device is ready to unplug its underlying
 884  * (slave) device queues -- before we let any writes go down, we need to
 885  * sync the dirty pages of the bitmap file to disk */
 886 void bitmap_unplug(struct bitmap *bitmap)
 887 {
 888         unsigned long i;
 889         int dirty, need_write;
 890         int wait = 0;
 891
 892         if (!bitmap || !bitmap->storage.filemap ||
 893             test_bit(BITMAP_STALE, &bitmap->flags))
 894                 return;
 895
 896         /* look at each page to see if there are any set bits that need to be
 897          * flushed out to disk */
 898         for (i = 0; i < bitmap->storage.file_pages; i++) {
 899                 if (!bitmap->storage.filemap)
 900                         return;
 901                 dirty = test_and_clear_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
 902                 need_write = test_and_clear_page_attr(bitmap, i,
 903                                                       BITMAP_PAGE_NEEDWRITE);
 904                 if (dirty || need_write) {
 905                         clear_page_attr(bitmap, i, BITMAP_PAGE_PENDING);
 906                         write_page(bitmap, bitmap->storage.filemap[i], 0);
 907                 }
 908                 if (dirty)
 909                         wait = 1;
 910         }
 911         if (wait) { /* if any writes were performed, we need to wait on them */
 912                 if (bitmap->storage.file)
 913                         wait_event(bitmap->write_wait,
 914                                    atomic_read(&bitmap->pending_writes)==0);
 915                 else
 916                         md_super_wait(bitmap->mddev);
 917         }
 918         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 919                 bitmap_file_kick(bitmap);
 920 }
 921 EXPORT_SYMBOL(bitmap_unplug);
 922
 923 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 924 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 925  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 926  * memory mapping of the bitmap file
 927  * Special cases:
 928  *   if there's no bitmap file, or if the bitmap file had been
 929  *   previously kicked from the array, we mark all the bits as
 930  *   1's in order to cause a full resync.
 931  *
 932  * We ignore all bits for sectors that end earlier than 'start'.
 933  * This is used when reading an out-of-date bitmap...
 934  */
 935 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 936 {
 937         unsigned long i, chunks, index, oldindex, bit;
 938         struct page *page = NULL;
 939         unsigned long bit_cnt = 0;
 940         struct file *file;
 941         unsigned long offset;
 942         int outofdate;
 943         int ret = -ENOSPC;
 944         void *paddr;
 945         struct bitmap_storage *store = &bitmap->storage;
 946
 947         chunks = bitmap->counts.chunks;
 948         file = store->file;
 949
 950         if (!file && !bitmap->mddev->bitmap_info.offset) {
 951                 /* No permanent bitmap - fill with '1s'. */
 952                 store->filemap = NULL;
 953                 store->file_pages = 0;
 954                 for (i = 0; i < chunks ; i++) {
 955                         /* if the disk bit is set, set the memory bit */
 956                         int needed = ((sector_t)(i+1) << (bitmap->counts.chunkshift)
 957                                       >= start);
 958                         bitmap_set_memory_bits(bitmap,
 959                                                (sector_t)i << bitmap->counts.chunkshift,
 960                                                needed);
 961                 }
 962                 return 0;
 963         }
 964
 965         outofdate = test_bit(BITMAP_STALE, &bitmap->flags);
 966         if (outofdate)
 967                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 968                         "recovery\n", bmname(bitmap));
 969
 970         if (file && i_size_read(file->f_mapping->host) < store->bytes) {
 971                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 972                        bmname(bitmap),
 973                        (unsigned long) i_size_read(file->f_mapping->host),
 974                        store->bytes);
 975                 goto err;
 976         }
 977
 978         oldindex = ~0L;
 979         offset = 0;
 980         if (!bitmap->mddev->bitmap_info.external)
 981                 offset = sizeof(bitmap_super_t);
 982
 983         for (i = 0; i < chunks; i++) {
 984                 int b;
 985                 index = file_page_index(&bitmap->storage, i);
 986                 bit = file_page_offset(&bitmap->storage, i);
 987                 if (index != oldindex) { /* this is a new page, read it in */
 988                         int count;
 989                         /* unmap the old page, we're done with it */
 990                         if (index == store->file_pages-1)
 991                                 count = store->bytes - index * PAGE_SIZE;
 992                         else
 993                                 count = PAGE_SIZE;
 994                         page = store->filemap[index];
 995                         if (file)
 996                                 ret = read_page(file, index, bitmap,
 997                                                 count, page);
 998                         else
 999                                 ret = read_sb_page(
1000                                         bitmap->mddev,
1001                                         bitmap->mddev->bitmap_info.offset,
1002                                         page,
1003                                         index, count);
1004
1005                         if (ret)
1006                                 goto err;
1007
1008                         oldindex = index;
1009
1010                         if (outofdate) {
1011                                 /*
1012                                  * if bitmap is out of date, dirty the
1013                                  * whole page and write it out
1014                                  */
1015                                 paddr = kmap_atomic(page);
1016                                 memset(paddr + offset, 0xff,
1017                                        PAGE_SIZE - offset);
1018                                 kunmap_atomic(paddr);
1019                                 write_page(bitmap, page, 1);
1020
1021                                 ret = -EIO;
1022                                 if (test_bit(BITMAP_WRITE_ERROR,
1023                                              &bitmap->flags))
1024                                         goto err;
1025                         }
1026                 }
1027                 paddr = kmap_atomic(page);
1028                 if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
1029                         b = test_bit(bit, paddr);
1030                 else
1031                         b = test_bit_le(bit, paddr);
1032                 kunmap_atomic(paddr);
1033                 if (b) {
1034                         /* if the disk bit is set, set the memory bit */
1035                         int needed = ((sector_t)(i+1) << bitmap->counts.chunkshift
1036                                       >= start);
1037                         bitmap_set_memory_bits(bitmap,
1038                                                (sector_t)i << bitmap->counts.chunkshift,
1039                                                needed);
1040                         bit_cnt++;
1041                 }
1042                 offset = 0;
1043         }
1044
1045         printk(KERN_INFO "%s: bitmap initialized from disk: "
1046                "read %lu pages, set %lu of %lu bits\n",
1047                bmname(bitmap), store->file_pages,
1048                bit_cnt, chunks);
1049
1050         return 0;
1051
1052  err:
1053         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1054                bmname(bitmap), ret);
1055         return ret;
1056 }
1057
1058 void bitmap_write_all(struct bitmap *bitmap)
1059 {
1060         /* We don't actually write all bitmap blocks here,
1061          * just flag them as needing to be written
1062          */
1063         int i;
1064
1065         if (!bitmap || !bitmap->storage.filemap)
1066                 return;
1067         if (bitmap->storage.file)
1068                 /* Only one copy, so nothing needed */
1069                 return;
1070
1071         for (i = 0; i < bitmap->storage.file_pages; i++)
1072                 set_page_attr(bitmap, i,
1073                               BITMAP_PAGE_NEEDWRITE);
1074         bitmap->allclean = 0;
1075 }
1076
1077 static void bitmap_count_page(struct bitmap_counts *bitmap,
1078                               sector_t offset, int inc)
1079 {
1080         sector_t chunk = offset >> bitmap->chunkshift;
1081         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1082         bitmap->bp[page].count += inc;
1083         bitmap_checkfree(bitmap, page);
1084 }
1085
1086 static void bitmap_set_pending(struct bitmap_counts *bitmap, sector_t offset)
1087 {
1088         sector_t chunk = offset >> bitmap->chunkshift;
1089         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1090         struct bitmap_page *bp = &bitmap->bp[page];
1091
1092         if (!bp->pending)
1093                 bp->pending = 1;
1094 }
1095
1096 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1097                                             sector_t offset, sector_t *blocks,
1098                                             int create);
1099
1100 /*
1101  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1102  *                      out to disk
1103  */
1104
1105 void bitmap_daemon_work(struct mddev *mddev)
1106 {
1107         struct bitmap *bitmap;
1108         unsigned long j;
1109         unsigned long nextpage;
1110         sector_t blocks;
1111         struct bitmap_counts *counts;
1112
1113         /* Use a mutex to guard daemon_work against
1114          * bitmap_destroy.
1115          */
1116         mutex_lock(&mddev->bitmap_info.mutex);
1117         bitmap = mddev->bitmap;
1118         if (bitmap == NULL) {
1119                 mutex_unlock(&mddev->bitmap_info.mutex);
1120                 return;
1121         }
1122         if (time_before(jiffies, bitmap->daemon_lastrun
1123                         + mddev->bitmap_info.daemon_sleep))
1124                 goto done;
1125
1126         bitmap->daemon_lastrun = jiffies;
1127         if (bitmap->allclean) {
1128                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1129                 goto done;
1130         }
1131         bitmap->allclean = 1;
1132
1133         /* Any file-page which is PENDING now needs to be written.
1134          * So set NEEDWRITE now, then after we make any last-minute changes
1135          * we will write it.
1136          */
1137         for (j = 0; j < bitmap->storage.file_pages; j++)
1138                 if (test_and_clear_page_attr(bitmap, j,
1139                                              BITMAP_PAGE_PENDING))
1140                         set_page_attr(bitmap, j,
1141                                       BITMAP_PAGE_NEEDWRITE);
1142
1143         if (bitmap->need_sync &&
1144             mddev->bitmap_info.external == 0) {
1145                 /* Arrange for superblock update as well as
1146                  * other changes */
1147                 bitmap_super_t *sb;
1148                 bitmap->need_sync = 0;
1149                 if (bitmap->storage.filemap) {
1150                         sb = kmap_atomic(bitmap->storage.sb_page);
1151                         sb->events_cleared =
1152                                 cpu_to_le64(bitmap->events_cleared);
1153                         kunmap_atomic(sb);
1154                         set_page_attr(bitmap, 0,
1155                                       BITMAP_PAGE_NEEDWRITE);
1156                 }
1157         }
1158         /* Now look at the bitmap counters and if any are '2' or '1',
1159          * decrement and handle accordingly.
1160          */
1161         counts = &bitmap->counts;
1162         spin_lock_irq(&counts->lock);
1163         nextpage = 0;
1164         for (j = 0; j < counts->chunks; j++) {
1165                 bitmap_counter_t *bmc;
1166                 sector_t  block = (sector_t)j << counts->chunkshift;
1167
1168                 if (j == nextpage) {
1169                         nextpage += PAGE_COUNTER_RATIO;
1170                         if (!counts->bp[j >> PAGE_COUNTER_SHIFT].pending) {
1171                                 j |= PAGE_COUNTER_MASK;
1172                                 continue;
1173                         }
1174                         counts->bp[j >> PAGE_COUNTER_SHIFT].pending = 0;
1175                 }
1176                 bmc = bitmap_get_counter(counts,
1177                                          block,
1178                                          &blocks, 0);
1179
1180                 if (!bmc) {
1181                         j |= PAGE_COUNTER_MASK;
1182                         continue;
1183                 }
1184                 if (*bmc == 1 && !bitmap->need_sync) {
1185                         /* We can clear the bit */
1186                         *bmc = 0;
1187                         bitmap_count_page(counts, block, -1);
1188                         bitmap_file_clear_bit(bitmap, block);
1189                 } else if (*bmc && *bmc <= 2) {
1190                         *bmc = 1;
1191                         bitmap_set_pending(counts, block);
1192                         bitmap->allclean = 0;
1193                 }
1194         }
1195         spin_unlock_irq(&counts->lock);
1196
1197         /* Now start writeout on any page in NEEDWRITE that isn't DIRTY.
1198          * DIRTY pages need to be written by bitmap_unplug so it can wait
1199          * for them.
1200          * If we find any DIRTY page we stop there and let bitmap_unplug
1201          * handle all the rest.  This is important in the case where
1202          * the first blocking holds the superblock and it has been updated.
1203          * We mustn't write any other blocks before the superblock.
1204          */
1205         for (j = 0;
1206              j < bitmap->storage.file_pages
1207                      && !test_bit(BITMAP_STALE, &bitmap->flags);
1208              j++) {
1209
1210                 if (test_page_attr(bitmap, j,
1211                                    BITMAP_PAGE_DIRTY))
1212                         /* bitmap_unplug will handle the rest */
1213                         break;
1214                 if (test_and_clear_page_attr(bitmap, j,
1215                                              BITMAP_PAGE_NEEDWRITE)) {
1216                         write_page(bitmap, bitmap->storage.filemap[j], 0);
1217                 }
1218         }
1219
1220  done:
1221         if (bitmap->allclean == 0)
1222                 mddev->thread->timeout =
1223                         mddev->bitmap_info.daemon_sleep;
1224         mutex_unlock(&mddev->bitmap_info.mutex);
1225 }
1226
1227 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1228                                             sector_t offset, sector_t *blocks,
1229                                             int create)
1230 __releases(bitmap->lock)
1231 __acquires(bitmap->lock)
1232 {
1233         /* If 'create', we might release the lock and reclaim it.
1234          * The lock must have been taken with interrupts enabled.
1235          * If !create, we don't release the lock.
1236          */
1237         sector_t chunk = offset >> bitmap->chunkshift;
1238         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1239         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1240         sector_t csize;
1241         int err;
1242
1243         err = bitmap_checkpage(bitmap, page, create);
1244
1245         if (bitmap->bp[page].hijacked ||
1246             bitmap->bp[page].map == NULL)
1247                 csize = ((sector_t)1) << (bitmap->chunkshift +
1248                                           PAGE_COUNTER_SHIFT - 1);
1249         else
1250                 csize = ((sector_t)1) << bitmap->chunkshift;
1251         *blocks = csize - (offset & (csize - 1));
1252
1253         if (err < 0)
1254                 return NULL;
1255
1256         /* now locked ... */
1257
1258         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1259                 /* should we use the first or second counter field
1260                  * of the hijacked pointer? */
1261                 int hi = (pageoff > PAGE_COUNTER_MASK);
1262                 return  &((bitmap_counter_t *)
1263                           &bitmap->bp[page].map)[hi];
1264         } else /* page is allocated */
1265                 return (bitmap_counter_t *)
1266                         &(bitmap->bp[page].map[pageoff]);
1267 }
1268
1269 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1270 {
1271         if (!bitmap)
1272                 return 0;
1273
1274         if (behind) {
1275                 int bw;
1276                 atomic_inc(&bitmap->behind_writes);
1277                 bw = atomic_read(&bitmap->behind_writes);
1278                 if (bw > bitmap->behind_writes_used)
1279                         bitmap->behind_writes_used = bw;
1280
1281                 pr_debug("inc write-behind count %d/%lu\n",
1282                          bw, bitmap->mddev->bitmap_info.max_write_behind);
1283         }
1284
1285         while (sectors) {
1286                 sector_t blocks;
1287                 bitmap_counter_t *bmc;
1288
1289                 spin_lock_irq(&bitmap->counts.lock);
1290                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 1);
1291                 if (!bmc) {
1292                         spin_unlock_irq(&bitmap->counts.lock);
1293                         return 0;
1294                 }
1295
1296                 if (unlikely(COUNTER(*bmc) == COUNTER_MAX)) {
1297                         DEFINE_WAIT(__wait);
1298                         /* note that it is safe to do the prepare_to_wait
1299                          * after the test as long as we do it before dropping
1300                          * the spinlock.
1301                          */
1302                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1303                                         TASK_UNINTERRUPTIBLE);
1304                         spin_unlock_irq(&bitmap->counts.lock);
1305                         schedule();
1306                         finish_wait(&bitmap->overflow_wait, &__wait);
1307                         continue;
1308                 }
1309
1310                 switch (*bmc) {
1311                 case 0:
1312                         bitmap_file_set_bit(bitmap, offset);
1313                         bitmap_count_page(&bitmap->counts, offset, 1);
1314                         /* fall through */
1315                 case 1:
1316                         *bmc = 2;
1317                 }
1318
1319                 (*bmc)++;
1320
1321                 spin_unlock_irq(&bitmap->counts.lock);
1322
1323                 offset += blocks;
1324                 if (sectors > blocks)
1325                         sectors -= blocks;
1326                 else
1327                         sectors = 0;
1328         }
1329         return 0;
1330 }
1331 EXPORT_SYMBOL(bitmap_startwrite);
1332
1333 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1334                      int success, int behind)
1335 {
1336         if (!bitmap)
1337                 return;
1338         if (behind) {
1339                 if (atomic_dec_and_test(&bitmap->behind_writes))
1340                         wake_up(&bitmap->behind_wait);
1341                 pr_debug("dec write-behind count %d/%lu\n",
1342                          atomic_read(&bitmap->behind_writes),
1343                          bitmap->mddev->bitmap_info.max_write_behind);
1344         }
1345
1346         while (sectors) {
1347                 sector_t blocks;
1348                 unsigned long flags;
1349                 bitmap_counter_t *bmc;
1350
1351                 spin_lock_irqsave(&bitmap->counts.lock, flags);
1352                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 0);
1353                 if (!bmc) {
1354                         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1355                         return;
1356                 }
1357
1358                 if (success && !bitmap->mddev->degraded &&
1359                     bitmap->events_cleared < bitmap->mddev->events) {
1360                         bitmap->events_cleared = bitmap->mddev->events;
1361                         bitmap->need_sync = 1;
1362                         sysfs_notify_dirent_safe(bitmap->sysfs_can_clear);
1363                 }
1364
1365                 if (!success && !NEEDED(*bmc))
1366                         *bmc |= NEEDED_MASK;
1367
1368                 if (COUNTER(*bmc) == COUNTER_MAX)
1369                         wake_up(&bitmap->overflow_wait);
1370
1371                 (*bmc)--;
1372                 if (*bmc <= 2) {
1373                         bitmap_set_pending(&bitmap->counts, offset);
1374                         bitmap->allclean = 0;
1375                 }
1376                 spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1377                 offset += blocks;
1378                 if (sectors > blocks)
1379                         sectors -= blocks;
1380                 else
1381                         sectors = 0;
1382         }
1383 }
1384 EXPORT_SYMBOL(bitmap_endwrite);
1385
1386 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1387                                int degraded)
1388 {
1389         bitmap_counter_t *bmc;
1390         int rv;
1391         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1392                 *blocks = 1024;
1393                 return 1; /* always resync if no bitmap */
1394         }
1395         spin_lock_irq(&bitmap->counts.lock);
1396         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1397         rv = 0;
1398         if (bmc) {
1399                 /* locked */
1400                 if (RESYNC(*bmc))
1401                         rv = 1;
1402                 else if (NEEDED(*bmc)) {
1403                         rv = 1;
1404                         if (!degraded) { /* don't set/clear bits if degraded */
1405                                 *bmc |= RESYNC_MASK;
1406                                 *bmc &= ~NEEDED_MASK;
1407                         }
1408                 }
1409         }
1410         spin_unlock_irq(&bitmap->counts.lock);
1411         return rv;
1412 }
1413
1414 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1415                       int degraded)
1416 {
1417         /* bitmap_start_sync must always report on multiples of whole
1418          * pages, otherwise resync (which is very PAGE_SIZE based) will
1419          * get confused.
1420          * So call __bitmap_start_sync repeatedly (if needed) until
1421          * At least PAGE_SIZE>>9 blocks are covered.
1422          * Return the 'or' of the result.
1423          */
1424         int rv = 0;
1425         sector_t blocks1;
1426
1427         *blocks = 0;
1428         while (*blocks < (PAGE_SIZE>>9)) {
1429                 rv |= __bitmap_start_sync(bitmap, offset,
1430                                           &blocks1, degraded);
1431                 offset += blocks1;
1432                 *blocks += blocks1;
1433         }
1434         return rv;
1435 }
1436 EXPORT_SYMBOL(bitmap_start_sync);
1437
1438 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks, int aborted)
1439 {
1440         bitmap_counter_t *bmc;
1441         unsigned long flags;
1442
1443         if (bitmap == NULL) {
1444                 *blocks = 1024;
1445                 return;
1446         }
1447         spin_lock_irqsave(&bitmap->counts.lock, flags);
1448         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1449         if (bmc == NULL)
1450                 goto unlock;
1451         /* locked */
1452         if (RESYNC(*bmc)) {
1453                 *bmc &= ~RESYNC_MASK;
1454
1455                 if (!NEEDED(*bmc) && aborted)
1456                         *bmc |= NEEDED_MASK;
1457                 else {
1458                         if (*bmc <= 2) {
1459                                 bitmap_set_pending(&bitmap->counts, offset);
1460                                 bitmap->allclean = 0;
1461                         }
1462                 }
1463         }
1464  unlock:
1465         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1466 }
1467 EXPORT_SYMBOL(bitmap_end_sync);
1468
1469 void bitmap_close_sync(struct bitmap *bitmap)
1470 {
1471         /* Sync has finished, and any bitmap chunks that weren't synced
1472          * properly have been aborted.  It remains to us to clear the
1473          * RESYNC bit wherever it is still on
1474          */
1475         sector_t sector = 0;
1476         sector_t blocks;
1477         if (!bitmap)
1478                 return;
1479         while (sector < bitmap->mddev->resync_max_sectors) {
1480                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1481                 sector += blocks;
1482         }
1483 }
1484 EXPORT_SYMBOL(bitmap_close_sync);
1485
1486 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1487 {
1488         sector_t s = 0;
1489         sector_t blocks;
1490
1491         if (!bitmap)
1492                 return;
1493         if (sector == 0) {
1494                 bitmap->last_end_sync = jiffies;
1495                 return;
1496         }
1497         if (time_before(jiffies, (bitmap->last_end_sync
1498                                   + bitmap->mddev->bitmap_info.daemon_sleep)))
1499                 return;
1500         wait_event(bitmap->mddev->recovery_wait,
1501                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1502
1503         bitmap->mddev->curr_resync_completed = sector;
1504         set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
1505         sector &= ~((1ULL << bitmap->counts.chunkshift) - 1);
1506         s = 0;
1507         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1508                 bitmap_end_sync(bitmap, s, &blocks, 0);
1509                 s += blocks;
1510         }
1511         bitmap->last_end_sync = jiffies;
1512         sysfs_notify(&bitmap->mddev->kobj, NULL, "sync_completed");
1513 }
1514 EXPORT_SYMBOL(bitmap_cond_end_sync);
1515
1516 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1517 {
1518         /* For each chunk covered by any of these sectors, set the
1519          * counter to 2 and possibly set resync_needed.  They should all
1520          * be 0 at this point
1521          */
1522
1523         sector_t secs;
1524         bitmap_counter_t *bmc;
1525         spin_lock_irq(&bitmap->counts.lock);
1526         bmc = bitmap_get_counter(&bitmap->counts, offset, &secs, 1);
1527         if (!bmc) {
1528                 spin_unlock_irq(&bitmap->counts.lock);
1529                 return;
1530         }
1531         if (!*bmc) {
1532                 *bmc = 2 | (needed ? NEEDED_MASK : 0);
1533                 bitmap_count_page(&bitmap->counts, offset, 1);
1534                 bitmap_set_pending(&bitmap->counts, offset);
1535                 bitmap->allclean = 0;
1536         }
1537         spin_unlock_irq(&bitmap->counts.lock);
1538 }
1539
1540 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1541 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1542 {
1543         unsigned long chunk;
1544
1545         for (chunk = s; chunk <= e; chunk++) {
1546                 sector_t sec = (sector_t)chunk << bitmap->counts.chunkshift;
1547                 bitmap_set_memory_bits(bitmap, sec, 1);
1548                 bitmap_file_set_bit(bitmap, sec);
1549                 if (sec < bitmap->mddev->recovery_cp)
1550                         /* We are asserting that the array is dirty,
1551                          * so move the recovery_cp address back so
1552                          * that it is obvious that it is dirty
1553                          */
1554                         bitmap->mddev->recovery_cp = sec;
1555         }
1556 }
1557
1558 /*
1559  * flush out any pending updates
1560  */
1561 void bitmap_flush(struct mddev *mddev)
1562 {
1563         struct bitmap *bitmap = mddev->bitmap;
1564         long sleep;
1565
1566         if (!bitmap) /* there was no bitmap */
1567                 return;
1568
1569         /* run the daemon_work three time to ensure everything is flushed
1570          * that can be
1571          */
1572         sleep = mddev->bitmap_info.daemon_sleep * 2;
1573         bitmap->daemon_lastrun -= sleep;
1574         bitmap_daemon_work(mddev);
1575         bitmap->daemon_lastrun -= sleep;
1576         bitmap_daemon_work(mddev);
1577         bitmap->daemon_lastrun -= sleep;
1578         bitmap_daemon_work(mddev);
1579         bitmap_update_sb(bitmap);
1580 }
1581
1582 /*
1583  * free memory that was allocated
1584  */
1585 static void bitmap_free(struct bitmap *bitmap)
1586 {
1587         unsigned long k, pages;
1588         struct bitmap_page *bp;
1589
1590         if (!bitmap) /* there was no bitmap */
1591                 return;
1592
1593         /* Shouldn't be needed - but just in case.... */
1594         wait_event(bitmap->write_wait,
1595                    atomic_read(&bitmap->pending_writes) == 0);
1596
1597         /* release the bitmap file  */
1598         bitmap_file_unmap(&bitmap->storage);
1599
1600         bp = bitmap->counts.bp;
1601         pages = bitmap->counts.pages;
1602
1603         /* free all allocated memory */
1604
1605         if (bp) /* deallocate the page memory */
1606                 for (k = 0; k < pages; k++)
1607                         if (bp[k].map && !bp[k].hijacked)
1608                                 kfree(bp[k].map);
1609         kfree(bp);
1610         kfree(bitmap);
1611 }
1612
1613 void bitmap_destroy(struct mddev *mddev)
1614 {
1615         struct bitmap *bitmap = mddev->bitmap;
1616
1617         if (!bitmap) /* there was no bitmap */
1618                 return;
1619
1620         mutex_lock(&mddev->bitmap_info.mutex);
1621         mddev->bitmap = NULL; /* disconnect from the md device */
1622         mutex_unlock(&mddev->bitmap_info.mutex);
1623         if (mddev->thread)
1624                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1625
1626         if (bitmap->sysfs_can_clear)
1627                 sysfs_put(bitmap->sysfs_can_clear);
1628
1629         bitmap_free(bitmap);
1630 }
1631
1632 /*
1633  * initialize the bitmap structure
1634  * if this returns an error, bitmap_destroy must be called to do clean up
1635  */
1636 int bitmap_create(struct mddev *mddev)
1637 {
1638         struct bitmap *bitmap;
1639         sector_t blocks = mddev->resync_max_sectors;
1640         struct file *file = mddev->bitmap_info.file;
1641         int err;
1642         struct sysfs_dirent *bm = NULL;
1643
1644         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1645
1646         BUG_ON(file && mddev->bitmap_info.offset);
1647
1648         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1649         if (!bitmap)
1650                 return -ENOMEM;
1651
1652         spin_lock_init(&bitmap->counts.lock);
1653         atomic_set(&bitmap->pending_writes, 0);
1654         init_waitqueue_head(&bitmap->write_wait);
1655         init_waitqueue_head(&bitmap->overflow_wait);
1656         init_waitqueue_head(&bitmap->behind_wait);
1657
1658         bitmap->mddev = mddev;
1659
1660         if (mddev->kobj.sd)
1661                 bm = sysfs_get_dirent(mddev->kobj.sd, NULL, "bitmap");
1662         if (bm) {
1663                 bitmap->sysfs_can_clear = sysfs_get_dirent(bm, NULL, "can_clear");
1664                 sysfs_put(bm);
1665         } else
1666                 bitmap->sysfs_can_clear = NULL;
1667
1668         bitmap->storage.file = file;
1669         if (file) {
1670                 get_file(file);
1671                 /* As future accesses to this file will use bmap,
1672                  * and bypass the page cache, we must sync the file
1673                  * first.
1674                  */
1675                 vfs_fsync(file, 1);
1676         }
1677         /* read superblock from bitmap file (this sets mddev->bitmap_info.chunksize) */
1678         if (!mddev->bitmap_info.external) {
1679                 /*
1680                  * If 'MD_ARRAY_FIRST_USE' is set, then device-mapper is
1681                  * instructing us to create a new on-disk bitmap instance.
1682                  */
1683                 if (test_and_clear_bit(MD_ARRAY_FIRST_USE, &mddev->flags))
1684                         err = bitmap_new_disk_sb(bitmap);
1685                 else
1686                         err = bitmap_read_sb(bitmap);
1687         } else {
1688                 err = 0;
1689                 if (mddev->bitmap_info.chunksize == 0 ||
1690                     mddev->bitmap_info.daemon_sleep == 0)
1691                         /* chunksize and time_base need to be
1692                          * set first. */
1693                         err = -EINVAL;
1694         }
1695         if (err)
1696                 goto error;
1697
1698         bitmap->daemon_lastrun = jiffies;
1699         err = bitmap_resize(bitmap, blocks, mddev->bitmap_info.chunksize, 1);
1700         if (err)
1701                 goto error;
1702
1703         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1704                bitmap->counts.pages, bmname(bitmap));
1705
1706         mddev->bitmap = bitmap;
1707         return test_bit(BITMAP_WRITE_ERROR, &bitmap->flags) ? -EIO : 0;
1708
1709  error:
1710         bitmap_free(bitmap);
1711         return err;
1712 }
1713
1714 int bitmap_load(struct mddev *mddev)
1715 {
1716         int err = 0;
1717         sector_t start = 0;
1718         sector_t sector = 0;
1719         struct bitmap *bitmap = mddev->bitmap;
1720
1721         if (!bitmap)
1722                 goto out;
1723
1724         /* Clear out old bitmap info first:  Either there is none, or we
1725          * are resuming after someone else has possibly changed things,
1726          * so we should forget old cached info.
1727          * All chunks should be clean, but some might need_sync.
1728          */
1729         while (sector < mddev->resync_max_sectors) {
1730                 sector_t blocks;
1731                 bitmap_start_sync(bitmap, sector, &blocks, 0);
1732                 sector += blocks;
1733         }
1734         bitmap_close_sync(bitmap);
1735
1736         if (mddev->degraded == 0
1737             || bitmap->events_cleared == mddev->events)
1738                 /* no need to keep dirty bits to optimise a
1739                  * re-add of a missing device */
1740                 start = mddev->recovery_cp;
1741
1742         mutex_lock(&mddev->bitmap_info.mutex);
1743         err = bitmap_init_from_disk(bitmap, start);
1744         mutex_unlock(&mddev->bitmap_info.mutex);
1745
1746         if (err)
1747                 goto out;
1748         clear_bit(BITMAP_STALE, &bitmap->flags);
1749
1750         /* Kick recovery in case any bits were set */
1751         set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1752
1753         mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
1754         md_wakeup_thread(mddev->thread);
1755
1756         bitmap_update_sb(bitmap);
1757
1758         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
1759                 err = -EIO;
1760 out:
1761         return err;
1762 }
1763 EXPORT_SYMBOL_GPL(bitmap_load);
1764
1765 void bitmap_status(struct seq_file *seq, struct bitmap *bitmap)
1766 {
1767         unsigned long chunk_kb;
1768         struct bitmap_counts *counts;
1769
1770         if (!bitmap)
1771                 return;
1772
1773         counts = &bitmap->counts;
1774
1775         chunk_kb = bitmap->mddev->bitmap_info.chunksize >> 10;
1776         seq_printf(seq, "bitmap: %lu/%lu pages [%luKB], "
1777                    "%lu%s chunk",
1778                    counts->pages - counts->missing_pages,
1779                    counts->pages,
1780                    (counts->pages - counts->missing_pages)
1781                    << (PAGE_SHIFT - 10),
1782                    chunk_kb ? chunk_kb : bitmap->mddev->bitmap_info.chunksize,
1783                    chunk_kb ? "KB" : "B");
1784         if (bitmap->storage.file) {
1785                 seq_printf(seq, ", file: ");
1786                 seq_path(seq, &bitmap->storage.file->f_path, " \t\n");
1787         }
1788
1789         seq_printf(seq, "\n");
1790 }
1791
1792 int bitmap_resize(struct bitmap *bitmap, sector_t blocks,
1793                   int chunksize, int init)
1794 {
1795         /* If chunk_size is 0, choose an appropriate chunk size.
1796          * Then possibly allocate new storage space.
1797          * Then quiesce, copy bits, replace bitmap, and re-start
1798          *
1799          * This function is called both to set up the initial bitmap
1800          * and to resize the bitmap while the array is active.
1801          * If this happens as a result of the array being resized,
1802          * chunksize will be zero, and we need to choose a suitable
1803          * chunksize, otherwise we use what we are given.
1804          */
1805         struct bitmap_storage store;
1806         struct bitmap_counts old_counts;
1807         unsigned long chunks;
1808         sector_t block;
1809         sector_t old_blocks, new_blocks;
1810         int chunkshift;
1811         int ret = 0;
1812         long pages;
1813         struct bitmap_page *new_bp;
1814
1815         if (chunksize == 0) {
1816                 /* If there is enough space, leave the chunk size unchanged,
1817                  * else increase by factor of two until there is enough space.
1818                  */
1819                 long bytes;
1820                 long space = bitmap->mddev->bitmap_info.space;
1821
1822                 if (space == 0) {
1823                         /* We don't know how much space there is, so limit
1824                          * to current size - in sectors.
1825                          */
1826                         bytes = DIV_ROUND_UP(bitmap->counts.chunks, 8);
1827                         if (!bitmap->mddev->bitmap_info.external)
1828                                 bytes += sizeof(bitmap_super_t);
1829                         space = DIV_ROUND_UP(bytes, 512);
1830                         bitmap->mddev->bitmap_info.space = space;
1831                 }
1832                 chunkshift = bitmap->counts.chunkshift;
1833                 chunkshift--;
1834                 do {
1835                         /* 'chunkshift' is shift from block size to chunk size */
1836                         chunkshift++;
1837                         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1838                         bytes = DIV_ROUND_UP(chunks, 8);
1839                         if (!bitmap->mddev->bitmap_info.external)
1840                                 bytes += sizeof(bitmap_super_t);
1841                 } while (bytes > (space << 9));
1842         } else
1843                 chunkshift = ffz(~chunksize) - BITMAP_BLOCK_SHIFT;
1844
1845         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1846         memset(&store, 0, sizeof(store));
1847         if (bitmap->mddev->bitmap_info.offset || bitmap->mddev->bitmap_info.file)
1848                 ret = bitmap_storage_alloc(&store, chunks,
1849                                            !bitmap->mddev->bitmap_info.external);
1850         if (ret)
1851                 goto err;
1852
1853         pages = DIV_ROUND_UP(chunks, PAGE_COUNTER_RATIO);
1854
1855         new_bp = kzalloc(pages * sizeof(*new_bp), GFP_KERNEL);
1856         ret = -ENOMEM;
1857         if (!new_bp) {
1858                 bitmap_file_unmap(&store);
1859                 goto err;
1860         }
1861
1862         if (!init)
1863                 bitmap->mddev->pers->quiesce(bitmap->mddev, 1);
1864
1865         store.file = bitmap->storage.file;
1866         bitmap->storage.file = NULL;
1867
1868         if (store.sb_page && bitmap->storage.sb_page)
1869                 memcpy(page_address(store.sb_page),
1870                        page_address(bitmap->storage.sb_page),
1871                        sizeof(bitmap_super_t));
1872         bitmap_file_unmap(&bitmap->storage);
1873         bitmap->storage = store;
1874
1875         old_counts = bitmap->counts;
1876         bitmap->counts.bp = new_bp;
1877         bitmap->counts.pages = pages;
1878         bitmap->counts.missing_pages = pages;
1879         bitmap->counts.chunkshift = chunkshift;
1880         bitmap->counts.chunks = chunks;
1881         bitmap->mddev->bitmap_info.chunksize = 1 << (chunkshift +
1882                                                      BITMAP_BLOCK_SHIFT);
1883
1884         blocks = min(old_counts.chunks << old_counts.chunkshift,
1885                      chunks << chunkshift);
1886
1887         spin_lock_irq(&bitmap->counts.lock);
1888         for (block = 0; block < blocks; ) {
1889                 bitmap_counter_t *bmc_old, *bmc_new;
1890                 int set;
1891
1892                 bmc_old = bitmap_get_counter(&old_counts, block,
1893                                              &old_blocks, 0);
1894                 set = bmc_old && NEEDED(*bmc_old);
1895
1896                 if (set) {
1897                         bmc_new = bitmap_get_counter(&bitmap->counts, block,
1898                                                      &new_blocks, 1);
1899                         if (*bmc_new == 0) {
1900                                 /* need to set on-disk bits too. */
1901                                 sector_t end = block + new_blocks;
1902                                 sector_t start = block >> chunkshift;
1903                                 start <<= chunkshift;
1904                                 while (start < end) {
1905                                         bitmap_file_set_bit(bitmap, block);
1906                                         start += 1 << chunkshift;
1907                                 }
1908                                 *bmc_new = 2;
1909                                 bitmap_count_page(&bitmap->counts,
1910                                                   block, 1);
1911                                 bitmap_set_pending(&bitmap->counts,
1912                                                    block);
1913                         }
1914                         *bmc_new |= NEEDED_MASK;
1915                         if (new_blocks < old_blocks)
1916                                 old_blocks = new_blocks;
1917                 }
1918                 block += old_blocks;
1919         }
1920
1921         if (!init) {
1922                 int i;
1923                 while (block < (chunks << chunkshift)) {
1924                         bitmap_counter_t *bmc;
1925                         bmc = bitmap_get_counter(&bitmap->counts, block,
1926                                                  &new_blocks, 1);
1927                         if (bmc) {
1928                                 /* new space.  It needs to be resynced, so
1929                                  * we set NEEDED_MASK.
1930                                  */
1931                                 if (*bmc == 0) {
1932                                         *bmc = NEEDED_MASK | 2;
1933                                         bitmap_count_page(&bitmap->counts,
1934                                                           block, 1);
1935                                         bitmap_set_pending(&bitmap->counts,
1936                                                            block);
1937                                 }
1938                         }
1939                         block += new_blocks;
1940                 }
1941                 for (i = 0; i < bitmap->storage.file_pages; i++)
1942                         set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
1943         }
1944         spin_unlock_irq(&bitmap->counts.lock);
1945
1946         if (!init) {
1947                 bitmap_unplug(bitmap);
1948                 bitmap->mddev->pers->quiesce(bitmap->mddev, 0);
1949         }
1950         ret = 0;
1951 err:
1952         return ret;
1953 }
1954 EXPORT_SYMBOL_GPL(bitmap_resize);
1955
1956 static ssize_t
1957 location_show(struct mddev *mddev, char *page)
1958 {
1959         ssize_t len;
1960         if (mddev->bitmap_info.file)
1961                 len = sprintf(page, "file");
1962         else if (mddev->bitmap_info.offset)
1963                 len = sprintf(page, "%+lld", (long long)mddev->bitmap_info.offset);
1964         else
1965                 len = sprintf(page, "none");
1966         len += sprintf(page+len, "\n");
1967         return len;
1968 }
1969
1970 static ssize_t
1971 location_store(struct mddev *mddev, const char *buf, size_t len)
1972 {
1973
1974         if (mddev->pers) {
1975                 if (!mddev->pers->quiesce)
1976                         return -EBUSY;
1977                 if (mddev->recovery || mddev->sync_thread)
1978                         return -EBUSY;
1979         }
1980
1981         if (mddev->bitmap || mddev->bitmap_info.file ||
1982             mddev->bitmap_info.offset) {
1983                 /* bitmap already configured.  Only option is to clear it */
1984                 if (strncmp(buf, "none", 4) != 0)
1985                         return -EBUSY;
1986                 if (mddev->pers) {
1987                         mddev->pers->quiesce(mddev, 1);
1988                         bitmap_destroy(mddev);
1989                         mddev->pers->quiesce(mddev, 0);
1990                 }
1991                 mddev->bitmap_info.offset = 0;
1992                 if (mddev->bitmap_info.file) {
1993                         struct file *f = mddev->bitmap_info.file;
1994                         mddev->bitmap_info.file = NULL;
1995                         restore_bitmap_write_access(f);
1996                         fput(f);
1997                 }
1998         } else {
1999                 /* No bitmap, OK to set a location */
2000                 long long offset;
2001                 if (strncmp(buf, "none", 4) == 0)
2002                         /* nothing to be done */;
2003                 else if (strncmp(buf, "file:", 5) == 0) {
2004                         /* Not supported yet */
2005                         return -EINVAL;
2006                 } else {
2007                         int rv;
2008                         if (buf[0] == '+')
2009                                 rv = strict_strtoll(buf+1, 10, &offset);
2010                         else
2011                                 rv = strict_strtoll(buf, 10, &offset);
2012                         if (rv)
2013                                 return rv;
2014                         if (offset == 0)
2015                                 return -EINVAL;
2016                         if (mddev->bitmap_info.external == 0 &&
2017                             mddev->major_version == 0 &&
2018                             offset != mddev->bitmap_info.default_offset)
2019                                 return -EINVAL;
2020                         mddev->bitmap_info.offset = offset;
2021                         if (mddev->pers) {
2022                                 mddev->pers->quiesce(mddev, 1);
2023                                 rv = bitmap_create(mddev);
2024                                 if (!rv)
2025                                         rv = bitmap_load(mddev);
2026                                 if (rv) {
2027                                         bitmap_destroy(mddev);
2028                                         mddev->bitmap_info.offset = 0;
2029                                 }
2030                                 mddev->pers->quiesce(mddev, 0);
2031                                 if (rv)
2032                                         return rv;
2033                         }
2034                 }
2035         }
2036         if (!mddev->external) {
2037                 /* Ensure new bitmap info is stored in
2038                  * metadata promptly.
2039                  */
2040                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2041                 md_wakeup_thread(mddev->thread);
2042         }
2043         return len;
2044 }
2045
2046 static struct md_sysfs_entry bitmap_location =
2047 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
2048
2049 /* 'bitmap/space' is the space available at 'location' for the
2050  * bitmap.  This allows the kernel to know when it is safe to
2051  * resize the bitmap to match a resized array.
2052  */
2053 static ssize_t
2054 space_show(struct mddev *mddev, char *page)
2055 {
2056         return sprintf(page, "%lu\n", mddev->bitmap_info.space);
2057 }
2058
2059 static ssize_t
2060 space_store(struct mddev *mddev, const char *buf, size_t len)
2061 {
2062         unsigned long sectors;
2063         int rv;
2064
2065         rv = kstrtoul(buf, 10, &sectors);
2066         if (rv)
2067                 return rv;
2068
2069         if (sectors == 0)
2070                 return -EINVAL;
2071
2072         if (mddev->bitmap &&
2073             sectors < (mddev->bitmap->storage.bytes + 511) >> 9)
2074                 return -EFBIG; /* Bitmap is too big for this small space */
2075
2076         /* could make sure it isn't too big, but that isn't really
2077          * needed - user-space should be careful.
2078          */
2079         mddev->bitmap_info.space = sectors;
2080         return len;
2081 }
2082
2083 static struct md_sysfs_entry bitmap_space =
2084 __ATTR(space, S_IRUGO|S_IWUSR, space_show, space_store);
2085
2086 static ssize_t
2087 timeout_show(struct mddev *mddev, char *page)
2088 {
2089         ssize_t len;
2090         unsigned long secs = mddev->bitmap_info.daemon_sleep / HZ;
2091         unsigned long jifs = mddev->bitmap_info.daemon_sleep % HZ;
2092
2093         len = sprintf(page, "%lu", secs);
2094         if (jifs)
2095                 len += sprintf(page+len, ".%03u", jiffies_to_msecs(jifs));
2096         len += sprintf(page+len, "\n");
2097         return len;
2098 }
2099
2100 static ssize_t
2101 timeout_store(struct mddev *mddev, const char *buf, size_t len)
2102 {
2103         /* timeout can be set at any time */
2104         unsigned long timeout;
2105         int rv = strict_strtoul_scaled(buf, &timeout, 4);
2106         if (rv)
2107                 return rv;
2108
2109         /* just to make sure we don't overflow... */
2110         if (timeout >= LONG_MAX / HZ)
2111                 return -EINVAL;
2112
2113         timeout = timeout * HZ / 10000;
2114
2115         if (timeout >= MAX_SCHEDULE_TIMEOUT)
2116                 timeout = MAX_SCHEDULE_TIMEOUT-1;
2117         if (timeout < 1)
2118                 timeout = 1;
2119         mddev->bitmap_info.daemon_sleep = timeout;
2120         if (mddev->thread) {
2121                 /* if thread->timeout is MAX_SCHEDULE_TIMEOUT, then
2122                  * the bitmap is all clean and we don't need to
2123                  * adjust the timeout right now
2124                  */
2125                 if (mddev->thread->timeout < MAX_SCHEDULE_TIMEOUT) {
2126                         mddev->thread->timeout = timeout;
2127                         md_wakeup_thread(mddev->thread);
2128                 }
2129         }
2130         return len;
2131 }
2132
2133 static struct md_sysfs_entry bitmap_timeout =
2134 __ATTR(time_base, S_IRUGO|S_IWUSR, timeout_show, timeout_store);
2135
2136 static ssize_t
2137 backlog_show(struct mddev *mddev, char *page)
2138 {
2139         return sprintf(page, "%lu\n", mddev->bitmap_info.max_write_behind);
2140 }
2141
2142 static ssize_t
2143 backlog_store(struct mddev *mddev, const char *buf, size_t len)
2144 {
2145         unsigned long backlog;
2146         int rv = strict_strtoul(buf, 10, &backlog);
2147         if (rv)
2148                 return rv;
2149         if (backlog > COUNTER_MAX)
2150                 return -EINVAL;
2151         mddev->bitmap_info.max_write_behind = backlog;
2152         return len;
2153 }
2154
2155 static struct md_sysfs_entry bitmap_backlog =
2156 __ATTR(backlog, S_IRUGO|S_IWUSR, backlog_show, backlog_store);
2157
2158 static ssize_t
2159 chunksize_show(struct mddev *mddev, char *page)
2160 {
2161         return sprintf(page, "%lu\n", mddev->bitmap_info.chunksize);
2162 }
2163
2164 static ssize_t
2165 chunksize_store(struct mddev *mddev, const char *buf, size_t len)
2166 {
2167         /* Can only be changed when no bitmap is active */
2168         int rv;
2169         unsigned long csize;
2170         if (mddev->bitmap)
2171                 return -EBUSY;
2172         rv = strict_strtoul(buf, 10, &csize);
2173         if (rv)
2174                 return rv;
2175         if (csize < 512 ||
2176             !is_power_of_2(csize))
2177                 return -EINVAL;
2178         mddev->bitmap_info.chunksize = csize;
2179         return len;
2180 }
2181
2182 static struct md_sysfs_entry bitmap_chunksize =
2183 __ATTR(chunksize, S_IRUGO|S_IWUSR, chunksize_show, chunksize_store);
2184
2185 static ssize_t metadata_show(struct mddev *mddev, char *page)
2186 {
2187         return sprintf(page, "%s\n", (mddev->bitmap_info.external
2188                                       ? "external" : "internal"));
2189 }
2190
2191 static ssize_t metadata_store(struct mddev *mddev, const char *buf, size_t len)
2192 {
2193         if (mddev->bitmap ||
2194             mddev->bitmap_info.file ||
2195             mddev->bitmap_info.offset)
2196                 return -EBUSY;
2197         if (strncmp(buf, "external", 8) == 0)
2198                 mddev->bitmap_info.external = 1;
2199         else if (strncmp(buf, "internal", 8) == 0)
2200                 mddev->bitmap_info.external = 0;
2201         else
2202                 return -EINVAL;
2203         return len;
2204 }
2205
2206 static struct md_sysfs_entry bitmap_metadata =
2207 __ATTR(metadata, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
2208
2209 static ssize_t can_clear_show(struct mddev *mddev, char *page)
2210 {
2211         int len;
2212         if (mddev->bitmap)
2213                 len = sprintf(page, "%s\n", (mddev->bitmap->need_sync ?
2214                                              "false" : "true"));
2215         else
2216                 len = sprintf(page, "\n");
2217         return len;
2218 }
2219
2220 static ssize_t can_clear_store(struct mddev *mddev, const char *buf, size_t len)
2221 {
2222         if (mddev->bitmap == NULL)
2223                 return -ENOENT;
2224         if (strncmp(buf, "false", 5) == 0)
2225                 mddev->bitmap->need_sync = 1;
2226         else if (strncmp(buf, "true", 4) == 0) {
2227                 if (mddev->degraded)
2228                         return -EBUSY;
2229                 mddev->bitmap->need_sync = 0;
2230         } else
2231                 return -EINVAL;
2232         return len;
2233 }
2234
2235 static struct md_sysfs_entry bitmap_can_clear =
2236 __ATTR(can_clear, S_IRUGO|S_IWUSR, can_clear_show, can_clear_store);
2237
2238 static ssize_t
2239 behind_writes_used_show(struct mddev *mddev, char *page)
2240 {
2241         if (mddev->bitmap == NULL)
2242                 return sprintf(page, "0\n");
2243         return sprintf(page, "%lu\n",
2244                        mddev->bitmap->behind_writes_used);
2245 }
2246
2247 static ssize_t
2248 behind_writes_used_reset(struct mddev *mddev, const char *buf, size_t len)
2249 {
2250         if (mddev->bitmap)
2251                 mddev->bitmap->behind_writes_used = 0;
2252         return len;
2253 }
2254
2255 static struct md_sysfs_entry max_backlog_used =
2256 __ATTR(max_backlog_used, S_IRUGO | S_IWUSR,
2257        behind_writes_used_show, behind_writes_used_reset);
2258
2259 static struct attribute *md_bitmap_attrs[] = {
2260         &bitmap_location.attr,
2261         &bitmap_space.attr,
2262         &bitmap_timeout.attr,
2263         &bitmap_backlog.attr,
2264         &bitmap_chunksize.attr,
2265         &bitmap_metadata.attr,
2266         &bitmap_can_clear.attr,
2267         &max_backlog_used.attr,
2268         NULL
2269 };
2270 struct attribute_group md_bitmap_group = {
2271         .name = "bitmap",
2272         .attrs = md_bitmap_attrs,
2273 };
2274