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 "md.h"
  30 #include "bitmap.h"
  31
  32 #include <linux/dm-dirty-log.h>
  33 /* debug macros */
  34
  35 #define DEBUG 0
  36
  37 #if DEBUG
  38 /* these are for debugging purposes only! */
  39
  40 /* define one and only one of these */
  41 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  42 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  43 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  44 #define INJECT_FAULTS_4 0 /* undef */
  45 #define INJECT_FAULTS_5 0 /* undef */
  46 #define INJECT_FAULTS_6 0
  47
  48 /* if these are defined, the driver will fail! debug only */
  49 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  50 #define INJECT_FATAL_FAULT_2 0 /* undef */
  51 #define INJECT_FATAL_FAULT_3 0 /* undef */
  52 #endif
  53
  54 #ifndef PRINTK
  55 #  if DEBUG > 0
  56 #    define PRINTK(x...) printk(KERN_DEBUG x)
  57 #  else
  58 #    define PRINTK(x...)
  59 #  endif
  60 #endif
  61
  62 static inline char *bmname(struct bitmap *bitmap)
  63 {
  64         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  65 }
  66
  67 /*
  68  * just a placeholder - calls kmalloc for bitmap pages
  69  */
  70 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
  71 {
  72         unsigned char *page;
  73
  74 #ifdef INJECT_FAULTS_1
  75         page = NULL;
  76 #else
  77         page = kzalloc(PAGE_SIZE, GFP_NOIO);
  78 #endif
  79         if (!page)
  80                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
  81         else
  82                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
  83                         bmname(bitmap), page);
  84         return page;
  85 }
  86
  87 /*
  88  * for now just a placeholder -- just calls kfree for bitmap pages
  89  */
  90 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
  91 {
  92         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
  93         kfree(page);
  94 }
  95
  96 /*
  97  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
  98  *
  99  * 1) check to see if this page is allocated, if it's not then try to alloc
 100  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 101  *    page pointer directly as a counter
 102  *
 103  * if we find our page, we increment the page's refcount so that it stays
 104  * allocated while we're using it
 105  */
 106 static int bitmap_checkpage(struct bitmap *bitmap,
 107                             unsigned long page, int create)
 108 __releases(bitmap->lock)
 109 __acquires(bitmap->lock)
 110 {
 111         unsigned char *mappage;
 112
 113         if (page >= bitmap->pages) {
 114                 /* This can happen if bitmap_start_sync goes beyond
 115                  * End-of-device while looking for a whole page.
 116                  * It is harmless.
 117                  */
 118                 return -EINVAL;
 119         }
 120
 121         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 122                 return 0;
 123
 124         if (bitmap->bp[page].map) /* page is already allocated, just return */
 125                 return 0;
 126
 127         if (!create)
 128                 return -ENOENT;
 129
 130         /* this page has not been allocated yet */
 131
 132         spin_unlock_irq(&bitmap->lock);
 133         mappage = bitmap_alloc_page(bitmap);
 134         spin_lock_irq(&bitmap->lock);
 135
 136         if (mappage == NULL) {
 137                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 138                         bmname(bitmap));
 139                 /* failed - set the hijacked flag so that we can use the
 140                  * pointer as a counter */
 141                 if (!bitmap->bp[page].map)
 142                         bitmap->bp[page].hijacked = 1;
 143         } else if (bitmap->bp[page].map ||
 144                    bitmap->bp[page].hijacked) {
 145                 /* somebody beat us to getting the page */
 146                 bitmap_free_page(bitmap, mappage);
 147                 return 0;
 148         } else {
 149
 150                 /* no page was in place and we have one, so install it */
 151
 152                 bitmap->bp[page].map = mappage;
 153                 bitmap->missing_pages--;
 154         }
 155         return 0;
 156 }
 157
 158 /* if page is completely empty, put it back on the free list, or dealloc it */
 159 /* if page was hijacked, unmark the flag so it might get alloced next time */
 160 /* Note: lock should be held when calling this */
 161 static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 162 {
 163         char *ptr;
 164
 165         if (bitmap->bp[page].count) /* page is still busy */
 166                 return;
 167
 168         /* page is no longer in use, it can be released */
 169
 170         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 171                 bitmap->bp[page].hijacked = 0;
 172                 bitmap->bp[page].map = NULL;
 173         } else {
 174                 /* normal case, free the page */
 175                 ptr = bitmap->bp[page].map;
 176                 bitmap->bp[page].map = NULL;
 177                 bitmap->missing_pages++;
 178                 bitmap_free_page(bitmap, ptr);
 179         }
 180 }
 181
 182 /*
 183  * bitmap file handling - read and write the bitmap file and its superblock
 184  */
 185
 186 /*
 187  * basic page I/O operations
 188  */
 189
 190 /* IO operations when bitmap is stored near all superblocks */
 191 static struct page *read_sb_page(mddev_t *mddev, loff_t offset,
 192                                  struct page *page,
 193                                  unsigned long index, int size)
 194 {
 195         /* choose a good rdev and read the page from there */
 196
 197         mdk_rdev_t *rdev;
 198         sector_t target;
 199         int did_alloc = 0;
 200
 201         if (!page) {
 202                 page = alloc_page(GFP_KERNEL);
 203                 if (!page)
 204                         return ERR_PTR(-ENOMEM);
 205                 did_alloc = 1;
 206         }
 207
 208         list_for_each_entry(rdev, &mddev->disks, same_set) {
 209                 if (! test_bit(In_sync, &rdev->flags)
 210                     || test_bit(Faulty, &rdev->flags))
 211                         continue;
 212
 213                 target = offset + index * (PAGE_SIZE/512);
 214
 215                 if (sync_page_io(rdev, target,
 216                                  roundup(size, bdev_logical_block_size(rdev->bdev)),
 217                                  page, READ, true)) {
 218                         page->index = index;
 219                         attach_page_buffers(page, NULL); /* so that free_buffer will
 220                                                           * quietly no-op */
 221                         return page;
 222                 }
 223         }
 224         if (did_alloc)
 225                 put_page(page);
 226         return ERR_PTR(-EIO);
 227
 228 }
 229
 230 static mdk_rdev_t *next_active_rdev(mdk_rdev_t *rdev, mddev_t *mddev)
 231 {
 232         /* Iterate the disks of an mddev, using rcu to protect access to the
 233          * linked list, and raising the refcount of devices we return to ensure
 234          * they don't disappear while in use.
 235          * As devices are only added or removed when raid_disk is < 0 and
 236          * nr_pending is 0 and In_sync is clear, the entries we return will
 237          * still be in the same position on the list when we re-enter
 238          * list_for_each_continue_rcu.
 239          */
 240         struct list_head *pos;
 241         rcu_read_lock();
 242         if (rdev == NULL)
 243                 /* start at the beginning */
 244                 pos = &mddev->disks;
 245         else {
 246                 /* release the previous rdev and start from there. */
 247                 rdev_dec_pending(rdev, mddev);
 248                 pos = &rdev->same_set;
 249         }
 250         list_for_each_continue_rcu(pos, &mddev->disks) {
 251                 rdev = list_entry(pos, mdk_rdev_t, same_set);
 252                 if (rdev->raid_disk >= 0 &&
 253                     !test_bit(Faulty, &rdev->flags)) {
 254                         /* this is a usable devices */
 255                         atomic_inc(&rdev->nr_pending);
 256                         rcu_read_unlock();
 257                         return rdev;
 258                 }
 259         }
 260         rcu_read_unlock();
 261         return NULL;
 262 }
 263
 264 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 265 {
 266         mdk_rdev_t *rdev = NULL;
 267         struct block_device *bdev;
 268         mddev_t *mddev = bitmap->mddev;
 269
 270         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 271                 int size = PAGE_SIZE;
 272                 loff_t offset = mddev->bitmap_info.offset;
 273
 274                 bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
 275
 276                 if (page->index == bitmap->file_pages-1)
 277                         size = roundup(bitmap->last_page_size,
 278                                        bdev_logical_block_size(bdev));
 279                 /* Just make sure we aren't corrupting data or
 280                  * metadata
 281                  */
 282                 if (mddev->external) {
 283                         /* Bitmap could be anywhere. */
 284                         if (rdev->sb_start + offset + (page->index
 285                                                        * (PAGE_SIZE/512))
 286                             > rdev->data_offset
 287                             &&
 288                             rdev->sb_start + offset
 289                             < (rdev->data_offset + mddev->dev_sectors
 290                              + (PAGE_SIZE/512)))
 291                                 goto bad_alignment;
 292                 } else if (offset < 0) {
 293                         /* DATA  BITMAP METADATA  */
 294                         if (offset
 295                             + (long)(page->index * (PAGE_SIZE/512))
 296                             + size/512 > 0)
 297                                 /* bitmap runs in to metadata */
 298                                 goto bad_alignment;
 299                         if (rdev->data_offset + mddev->dev_sectors
 300                             > rdev->sb_start + offset)
 301                                 /* data runs in to bitmap */
 302                                 goto bad_alignment;
 303                 } else if (rdev->sb_start < rdev->data_offset) {
 304                         /* METADATA BITMAP DATA */
 305                         if (rdev->sb_start
 306                             + offset
 307                             + page->index*(PAGE_SIZE/512) + size/512
 308                             > rdev->data_offset)
 309                                 /* bitmap runs in to data */
 310                                 goto bad_alignment;
 311                 } else {
 312                         /* DATA METADATA BITMAP - no problems */
 313                 }
 314                 md_super_write(mddev, rdev,
 315                                rdev->sb_start + offset
 316                                + page->index * (PAGE_SIZE/512),
 317                                size,
 318                                page);
 319         }
 320
 321         if (wait)
 322                 md_super_wait(mddev);
 323         return 0;
 324
 325  bad_alignment:
 326         return -EINVAL;
 327 }
 328
 329 static void bitmap_file_kick(struct bitmap *bitmap);
 330 /*
 331  * write out a page to a file
 332  */
 333 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 334 {
 335         struct buffer_head *bh;
 336
 337         if (bitmap->file == NULL) {
 338                 switch (write_sb_page(bitmap, page, wait)) {
 339                 case -EINVAL:
 340                         bitmap->flags |= BITMAP_WRITE_ERROR;
 341                 }
 342         } else {
 343
 344                 bh = page_buffers(page);
 345
 346                 while (bh && bh->b_blocknr) {
 347                         atomic_inc(&bitmap->pending_writes);
 348                         set_buffer_locked(bh);
 349                         set_buffer_mapped(bh);
 350                         submit_bh(WRITE | REQ_SYNC, bh);
 351                         bh = bh->b_this_page;
 352                 }
 353
 354                 if (wait)
 355                         wait_event(bitmap->write_wait,
 356                                    atomic_read(&bitmap->pending_writes)==0);
 357         }
 358         if (bitmap->flags & BITMAP_WRITE_ERROR)
 359                 bitmap_file_kick(bitmap);
 360 }
 361
 362 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 363 {
 364         struct bitmap *bitmap = bh->b_private;
 365         unsigned long flags;
 366
 367         if (!uptodate) {
 368                 spin_lock_irqsave(&bitmap->lock, flags);
 369                 bitmap->flags |= BITMAP_WRITE_ERROR;
 370                 spin_unlock_irqrestore(&bitmap->lock, flags);
 371         }
 372         if (atomic_dec_and_test(&bitmap->pending_writes))
 373                 wake_up(&bitmap->write_wait);
 374 }
 375
 376 /* copied from buffer.c */
 377 static void
 378 __clear_page_buffers(struct page *page)
 379 {
 380         ClearPagePrivate(page);
 381         set_page_private(page, 0);
 382         page_cache_release(page);
 383 }
 384 static void free_buffers(struct page *page)
 385 {
 386         struct buffer_head *bh = page_buffers(page);
 387
 388         while (bh) {
 389                 struct buffer_head *next = bh->b_this_page;
 390                 free_buffer_head(bh);
 391                 bh = next;
 392         }
 393         __clear_page_buffers(page);
 394         put_page(page);
 395 }
 396
 397 /* read a page from a file.
 398  * We both read the page, and attach buffers to the page to record the
 399  * address of each block (using bmap).  These addresses will be used
 400  * to write the block later, completely bypassing the filesystem.
 401  * This usage is similar to how swap files are handled, and allows us
 402  * to write to a file with no concerns of memory allocation failing.
 403  */
 404 static struct page *read_page(struct file *file, unsigned long index,
 405                               struct bitmap *bitmap,
 406                               unsigned long count)
 407 {
 408         struct page *page = NULL;
 409         struct inode *inode = file->f_path.dentry->d_inode;
 410         struct buffer_head *bh;
 411         sector_t block;
 412
 413         PRINTK("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 414                         (unsigned long long)index << PAGE_SHIFT);
 415
 416         page = alloc_page(GFP_KERNEL);
 417         if (!page)
 418                 page = ERR_PTR(-ENOMEM);
 419         if (IS_ERR(page))
 420                 goto out;
 421
 422         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 423         if (!bh) {
 424                 put_page(page);
 425                 page = ERR_PTR(-ENOMEM);
 426                 goto out;
 427         }
 428         attach_page_buffers(page, bh);
 429         block = index << (PAGE_SHIFT - inode->i_blkbits);
 430         while (bh) {
 431                 if (count == 0)
 432                         bh->b_blocknr = 0;
 433                 else {
 434                         bh->b_blocknr = bmap(inode, block);
 435                         if (bh->b_blocknr == 0) {
 436                                 /* Cannot use this file! */
 437                                 free_buffers(page);
 438                                 page = ERR_PTR(-EINVAL);
 439                                 goto out;
 440                         }
 441                         bh->b_bdev = inode->i_sb->s_bdev;
 442                         if (count < (1<<inode->i_blkbits))
 443                                 count = 0;
 444                         else
 445                                 count -= (1<<inode->i_blkbits);
 446
 447                         bh->b_end_io = end_bitmap_write;
 448                         bh->b_private = bitmap;
 449                         atomic_inc(&bitmap->pending_writes);
 450                         set_buffer_locked(bh);
 451                         set_buffer_mapped(bh);
 452                         submit_bh(READ, bh);
 453                 }
 454                 block++;
 455                 bh = bh->b_this_page;
 456         }
 457         page->index = index;
 458
 459         wait_event(bitmap->write_wait,
 460                    atomic_read(&bitmap->pending_writes)==0);
 461         if (bitmap->flags & BITMAP_WRITE_ERROR) {
 462                 free_buffers(page);
 463                 page = ERR_PTR(-EIO);
 464         }
 465 out:
 466         if (IS_ERR(page))
 467                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %ld\n",
 468                         (int)PAGE_SIZE,
 469                         (unsigned long long)index << PAGE_SHIFT,
 470                         PTR_ERR(page));
 471         return page;
 472 }
 473
 474 /*
 475  * bitmap file superblock operations
 476  */
 477
 478 /* update the event counter and sync the superblock to disk */
 479 void bitmap_update_sb(struct bitmap *bitmap)
 480 {
 481         bitmap_super_t *sb;
 482         unsigned long flags;
 483
 484         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 485                 return;
 486         if (bitmap->mddev->bitmap_info.external)
 487                 return;
 488         spin_lock_irqsave(&bitmap->lock, flags);
 489         if (!bitmap->sb_page) { /* no superblock */
 490                 spin_unlock_irqrestore(&bitmap->lock, flags);
 491                 return;
 492         }
 493         spin_unlock_irqrestore(&bitmap->lock, flags);
 494         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 495         sb->events = cpu_to_le64(bitmap->mddev->events);
 496         if (bitmap->mddev->events < bitmap->events_cleared)
 497                 /* rocking back to read-only */
 498                 bitmap->events_cleared = bitmap->mddev->events;
 499         sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 500         sb->state = cpu_to_le32(bitmap->flags);
 501         /* Just in case these have been changed via sysfs: */
 502         sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 503         sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
 504         kunmap_atomic(sb, KM_USER0);
 505         write_page(bitmap, bitmap->sb_page, 1);
 506 }
 507
 508 /* print out the bitmap file superblock */
 509 void bitmap_print_sb(struct bitmap *bitmap)
 510 {
 511         bitmap_super_t *sb;
 512
 513         if (!bitmap || !bitmap->sb_page)
 514                 return;
 515         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 516         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 517         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 518         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 519         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 520                                         *(__u32 *)(sb->uuid+0),
 521                                         *(__u32 *)(sb->uuid+4),
 522                                         *(__u32 *)(sb->uuid+8),
 523                                         *(__u32 *)(sb->uuid+12));
 524         printk(KERN_DEBUG "        events: %llu\n",
 525                         (unsigned long long) le64_to_cpu(sb->events));
 526         printk(KERN_DEBUG "events cleared: %llu\n",
 527                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 528         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 529         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 530         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 531         printk(KERN_DEBUG "     sync size: %llu KB\n",
 532                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 533         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 534         kunmap_atomic(sb, KM_USER0);
 535 }
 536
 537 /*
 538  * bitmap_new_disk_sb
 539  * @bitmap
 540  *
 541  * This function is somewhat the reverse of bitmap_read_sb.  bitmap_read_sb
 542  * reads and verifies the on-disk bitmap superblock and populates bitmap_info.
 543  * This function verifies 'bitmap_info' and populates the on-disk bitmap
 544  * structure, which is to be written to disk.
 545  *
 546  * Returns: 0 on success, -Exxx on error
 547  */
 548 static int bitmap_new_disk_sb(struct bitmap *bitmap)
 549 {
 550         bitmap_super_t *sb;
 551         unsigned long chunksize, daemon_sleep, write_behind;
 552         int err = -EINVAL;
 553
 554         bitmap->sb_page = alloc_page(GFP_KERNEL);
 555         if (IS_ERR(bitmap->sb_page)) {
 556                 err = PTR_ERR(bitmap->sb_page);
 557                 bitmap->sb_page = NULL;
 558                 return err;
 559         }
 560         bitmap->sb_page->index = 0;
 561
 562         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 563
 564         sb->magic = cpu_to_le32(BITMAP_MAGIC);
 565         sb->version = cpu_to_le32(BITMAP_MAJOR_HI);
 566
 567         chunksize = bitmap->mddev->bitmap_info.chunksize;
 568         BUG_ON(!chunksize);
 569         if (!is_power_of_2(chunksize)) {
 570                 kunmap_atomic(sb, KM_USER0);
 571                 printk(KERN_ERR "bitmap chunksize not a power of 2\n");
 572                 return -EINVAL;
 573         }
 574         sb->chunksize = cpu_to_le32(chunksize);
 575
 576         daemon_sleep = bitmap->mddev->bitmap_info.daemon_sleep;
 577         if (!daemon_sleep ||
 578             (daemon_sleep < 1) || (daemon_sleep > MAX_SCHEDULE_TIMEOUT)) {
 579                 printk(KERN_INFO "Choosing daemon_sleep default (5 sec)\n");
 580                 daemon_sleep = 5 * HZ;
 581         }
 582         sb->daemon_sleep = cpu_to_le32(daemon_sleep);
 583         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 584
 585         /*
 586          * FIXME: write_behind for RAID1.  If not specified, what
 587          * is a good choice?  We choose COUNTER_MAX / 2 arbitrarily.
 588          */
 589         write_behind = bitmap->mddev->bitmap_info.max_write_behind;
 590         if (write_behind > COUNTER_MAX)
 591                 write_behind = COUNTER_MAX / 2;
 592         sb->write_behind = cpu_to_le32(write_behind);
 593         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 594
 595         /* keep the array size field of the bitmap superblock up to date */
 596         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 597
 598         memcpy(sb->uuid, bitmap->mddev->uuid, 16);
 599
 600         bitmap->flags |= BITMAP_STALE;
 601         sb->state |= cpu_to_le32(BITMAP_STALE);
 602         bitmap->events_cleared = bitmap->mddev->events;
 603         sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 604
 605         bitmap->flags |= BITMAP_HOSTENDIAN;
 606         sb->version = cpu_to_le32(BITMAP_MAJOR_HOSTENDIAN);
 607
 608         kunmap_atomic(sb, KM_USER0);
 609
 610         return 0;
 611 }
 612
 613 /* read the superblock from the bitmap file and initialize some bitmap fields */
 614 static int bitmap_read_sb(struct bitmap *bitmap)
 615 {
 616         char *reason = NULL;
 617         bitmap_super_t *sb;
 618         unsigned long chunksize, daemon_sleep, write_behind;
 619         unsigned long long events;
 620         int err = -EINVAL;
 621
 622         /* page 0 is the superblock, read it... */
 623         if (bitmap->file) {
 624                 loff_t isize = i_size_read(bitmap->file->f_mapping->host);
 625                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 626
 627                 bitmap->sb_page = read_page(bitmap->file, 0, bitmap, bytes);
 628         } else {
 629                 bitmap->sb_page = read_sb_page(bitmap->mddev,
 630                                                bitmap->mddev->bitmap_info.offset,
 631                                                NULL,
 632                                                0, sizeof(bitmap_super_t));
 633         }
 634         if (IS_ERR(bitmap->sb_page)) {
 635                 err = PTR_ERR(bitmap->sb_page);
 636                 bitmap->sb_page = NULL;
 637                 return err;
 638         }
 639
 640         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 641
 642         chunksize = le32_to_cpu(sb->chunksize);
 643         daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 644         write_behind = le32_to_cpu(sb->write_behind);
 645
 646         /* verify that the bitmap-specific fields are valid */
 647         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 648                 reason = "bad magic";
 649         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 650                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 651                 reason = "unrecognized superblock version";
 652         else if (chunksize < 512)
 653                 reason = "bitmap chunksize too small";
 654         else if (!is_power_of_2(chunksize))
 655                 reason = "bitmap chunksize not a power of 2";
 656         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT)
 657                 reason = "daemon sleep period out of range";
 658         else if (write_behind > COUNTER_MAX)
 659                 reason = "write-behind limit out of range (0 - 16383)";
 660         if (reason) {
 661                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 662                         bmname(bitmap), reason);
 663                 goto out;
 664         }
 665
 666         /* keep the array size field of the bitmap superblock up to date */
 667         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 668
 669         if (!bitmap->mddev->persistent)
 670                 goto success;
 671
 672         /*
 673          * if we have a persistent array superblock, compare the
 674          * bitmap's UUID and event counter to the mddev's
 675          */
 676         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 677                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 678                         bmname(bitmap));
 679                 goto out;
 680         }
 681         events = le64_to_cpu(sb->events);
 682         if (events < bitmap->mddev->events) {
 683                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 684                         "-- forcing full recovery\n", bmname(bitmap), events,
 685                         (unsigned long long) bitmap->mddev->events);
 686                 sb->state |= cpu_to_le32(BITMAP_STALE);
 687         }
 688 success:
 689         /* assign fields using values from superblock */
 690         bitmap->mddev->bitmap_info.chunksize = chunksize;
 691         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 692         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 693         bitmap->flags |= le32_to_cpu(sb->state);
 694         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 695                 bitmap->flags |= BITMAP_HOSTENDIAN;
 696         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 697         if (bitmap->flags & BITMAP_STALE)
 698                 bitmap->events_cleared = bitmap->mddev->events;
 699         err = 0;
 700 out:
 701         kunmap_atomic(sb, KM_USER0);
 702         if (err)
 703                 bitmap_print_sb(bitmap);
 704         return err;
 705 }
 706
 707 enum bitmap_mask_op {
 708         MASK_SET,
 709         MASK_UNSET
 710 };
 711
 712 /* record the state of the bitmap in the superblock.  Return the old value */
 713 static int bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 714                              enum bitmap_mask_op op)
 715 {
 716         bitmap_super_t *sb;
 717         unsigned long flags;
 718         int old;
 719
 720         spin_lock_irqsave(&bitmap->lock, flags);
 721         if (!bitmap->sb_page) { /* can't set the state */
 722                 spin_unlock_irqrestore(&bitmap->lock, flags);
 723                 return 0;
 724         }
 725         spin_unlock_irqrestore(&bitmap->lock, flags);
 726         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 727         old = le32_to_cpu(sb->state) & bits;
 728         switch (op) {
 729         case MASK_SET:
 730                 sb->state |= cpu_to_le32(bits);
 731                 bitmap->flags |= bits;
 732                 break;
 733         case MASK_UNSET:
 734                 sb->state &= cpu_to_le32(~bits);
 735                 bitmap->flags &= ~bits;
 736                 break;
 737         default:
 738                 BUG();
 739         }
 740         kunmap_atomic(sb, KM_USER0);
 741         return old;
 742 }
 743
 744 /*
 745  * general bitmap file operations
 746  */
 747
 748 /*
 749  * on-disk bitmap:
 750  *
 751  * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
 752  * file a page at a time. There's a superblock at the start of the file.
 753  */
 754 /* calculate the index of the page that contains this bit */
 755 static inline unsigned long file_page_index(struct bitmap *bitmap, unsigned long chunk)
 756 {
 757         if (!bitmap->mddev->bitmap_info.external)
 758                 chunk += sizeof(bitmap_super_t) << 3;
 759         return chunk >> PAGE_BIT_SHIFT;
 760 }
 761
 762 /* calculate the (bit) offset of this bit within a page */
 763 static inline unsigned long file_page_offset(struct bitmap *bitmap, unsigned long chunk)
 764 {
 765         if (!bitmap->mddev->bitmap_info.external)
 766                 chunk += sizeof(bitmap_super_t) << 3;
 767         return chunk & (PAGE_BITS - 1);
 768 }
 769
 770 /*
 771  * return a pointer to the page in the filemap that contains the given bit
 772  *
 773  * this lookup is complicated by the fact that the bitmap sb might be exactly
 774  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 775  * 0 or page 1
 776  */
 777 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 778                                         unsigned long chunk)
 779 {
 780         if (bitmap->filemap == NULL)
 781                 return NULL;
 782         if (file_page_index(bitmap, chunk) >= bitmap->file_pages)
 783                 return NULL;
 784         return bitmap->filemap[file_page_index(bitmap, chunk)
 785                                - file_page_index(bitmap, 0)];
 786 }
 787
 788 static void bitmap_file_unmap(struct bitmap *bitmap)
 789 {
 790         struct page **map, *sb_page;
 791         unsigned long *attr;
 792         int pages;
 793         unsigned long flags;
 794
 795         spin_lock_irqsave(&bitmap->lock, flags);
 796         map = bitmap->filemap;
 797         bitmap->filemap = NULL;
 798         attr = bitmap->filemap_attr;
 799         bitmap->filemap_attr = NULL;
 800         pages = bitmap->file_pages;
 801         bitmap->file_pages = 0;
 802         sb_page = bitmap->sb_page;
 803         bitmap->sb_page = NULL;
 804         spin_unlock_irqrestore(&bitmap->lock, flags);
 805
 806         while (pages--)
 807                 if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 808                         free_buffers(map[pages]);
 809         kfree(map);
 810         kfree(attr);
 811
 812         if (sb_page)
 813                 free_buffers(sb_page);
 814 }
 815
 816 static void bitmap_file_put(struct bitmap *bitmap)
 817 {
 818         struct file *file;
 819         unsigned long flags;
 820
 821         spin_lock_irqsave(&bitmap->lock, flags);
 822         file = bitmap->file;
 823         bitmap->file = NULL;
 824         spin_unlock_irqrestore(&bitmap->lock, flags);
 825
 826         if (file)
 827                 wait_event(bitmap->write_wait,
 828                            atomic_read(&bitmap->pending_writes)==0);
 829         bitmap_file_unmap(bitmap);
 830
 831         if (file) {
 832                 struct inode *inode = file->f_path.dentry->d_inode;
 833                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 834                 fput(file);
 835         }
 836 }
 837
 838 /*
 839  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 840  * then it is no longer reliable, so we stop using it and we mark the file
 841  * as failed in the superblock
 842  */
 843 static void bitmap_file_kick(struct bitmap *bitmap)
 844 {
 845         char *path, *ptr = NULL;
 846
 847         if (bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET) == 0) {
 848                 bitmap_update_sb(bitmap);
 849
 850                 if (bitmap->file) {
 851                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 852                         if (path)
 853                                 ptr = d_path(&bitmap->file->f_path, path,
 854                                              PAGE_SIZE);
 855
 856                         printk(KERN_ALERT
 857                               "%s: kicking failed bitmap file %s from array!\n",
 858                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 859
 860                         kfree(path);
 861                 } else
 862                         printk(KERN_ALERT
 863                                "%s: disabling internal bitmap due to errors\n",
 864                                bmname(bitmap));
 865         }
 866
 867         bitmap_file_put(bitmap);
 868
 869         return;
 870 }
 871
 872 enum bitmap_page_attr {
 873         BITMAP_PAGE_DIRTY = 0,     /* there are set bits that need to be synced */
 874         BITMAP_PAGE_CLEAN = 1,     /* there are bits that might need to be cleared */
 875         BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 876 };
 877
 878 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 879                                 enum bitmap_page_attr attr)
 880 {
 881         if (page)
 882                 __set_bit((page->index<<2) + attr, bitmap->filemap_attr);
 883         else
 884                 __set_bit(attr, &bitmap->logattrs);
 885 }
 886
 887 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 888                                 enum bitmap_page_attr attr)
 889 {
 890         if (page)
 891                 __clear_bit((page->index<<2) + attr, bitmap->filemap_attr);
 892         else
 893                 __clear_bit(attr, &bitmap->logattrs);
 894 }
 895
 896 static inline unsigned long test_page_attr(struct bitmap *bitmap, struct page *page,
 897                                            enum bitmap_page_attr attr)
 898 {
 899         if (page)
 900                 return test_bit((page->index<<2) + attr, bitmap->filemap_attr);
 901         else
 902                 return test_bit(attr, &bitmap->logattrs);
 903 }
 904
 905 /*
 906  * bitmap_file_set_bit -- called before performing a write to the md device
 907  * to set (and eventually sync) a particular bit in the bitmap file
 908  *
 909  * we set the bit immediately, then we record the page number so that
 910  * when an unplug occurs, we can flush the dirty pages out to disk
 911  */
 912 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 913 {
 914         unsigned long bit;
 915         struct page *page = NULL;
 916         void *kaddr;
 917         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 918
 919         if (!bitmap->filemap) {
 920                 struct dm_dirty_log *log = bitmap->mddev->bitmap_info.log;
 921                 if (log)
 922                         log->type->mark_region(log, chunk);
 923         } else {
 924
 925                 page = filemap_get_page(bitmap, chunk);
 926                 if (!page)
 927                         return;
 928                 bit = file_page_offset(bitmap, chunk);
 929
 930                 /* set the bit */
 931                 kaddr = kmap_atomic(page, KM_USER0);
 932                 if (bitmap->flags & BITMAP_HOSTENDIAN)
 933                         set_bit(bit, kaddr);
 934                 else
 935                         __set_bit_le(bit, kaddr);
 936                 kunmap_atomic(kaddr, KM_USER0);
 937                 PRINTK("set file bit %lu page %lu\n", bit, page->index);
 938         }
 939         /* record page number so it gets flushed to disk when unplug occurs */
 940         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 941 }
 942
 943 /* this gets called when the md device is ready to unplug its underlying
 944  * (slave) device queues -- before we let any writes go down, we need to
 945  * sync the dirty pages of the bitmap file to disk */
 946 void bitmap_unplug(struct bitmap *bitmap)
 947 {
 948         unsigned long i, flags;
 949         int dirty, need_write;
 950         struct page *page;
 951         int wait = 0;
 952
 953         if (!bitmap)
 954                 return;
 955         if (!bitmap->filemap) {
 956                 /* Must be using a dirty_log */
 957                 struct dm_dirty_log *log = bitmap->mddev->bitmap_info.log;
 958                 dirty = test_and_clear_bit(BITMAP_PAGE_DIRTY, &bitmap->logattrs);
 959                 need_write = test_and_clear_bit(BITMAP_PAGE_NEEDWRITE, &bitmap->logattrs);
 960                 if (dirty || need_write)
 961                         if (log->type->flush(log))
 962                                 bitmap->flags |= BITMAP_WRITE_ERROR;
 963                 goto out;
 964         }
 965
 966         /* look at each page to see if there are any set bits that need to be
 967          * flushed out to disk */
 968         for (i = 0; i < bitmap->file_pages; i++) {
 969                 spin_lock_irqsave(&bitmap->lock, flags);
 970                 if (!bitmap->filemap) {
 971                         spin_unlock_irqrestore(&bitmap->lock, flags);
 972                         return;
 973                 }
 974                 page = bitmap->filemap[i];
 975                 dirty = test_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 976                 need_write = test_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 977                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 978                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 979                 if (dirty)
 980                         wait = 1;
 981                 spin_unlock_irqrestore(&bitmap->lock, flags);
 982
 983                 if (dirty || need_write)
 984                         write_page(bitmap, page, 0);
 985         }
 986         if (wait) { /* if any writes were performed, we need to wait on them */
 987                 if (bitmap->file)
 988                         wait_event(bitmap->write_wait,
 989                                    atomic_read(&bitmap->pending_writes)==0);
 990                 else
 991                         md_super_wait(bitmap->mddev);
 992         }
 993 out:
 994         if (bitmap->flags & BITMAP_WRITE_ERROR)
 995                 bitmap_file_kick(bitmap);
 996 }
 997 EXPORT_SYMBOL(bitmap_unplug);
 998
 999 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
1000 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
1001  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
1002  * memory mapping of the bitmap file
1003  * Special cases:
1004  *   if there's no bitmap file, or if the bitmap file had been
1005  *   previously kicked from the array, we mark all the bits as
1006  *   1's in order to cause a full resync.
1007  *
1008  * We ignore all bits for sectors that end earlier than 'start'.
1009  * This is used when reading an out-of-date bitmap...
1010  */
1011 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
1012 {
1013         unsigned long i, chunks, index, oldindex, bit;
1014         struct page *page = NULL, *oldpage = NULL;
1015         unsigned long num_pages, bit_cnt = 0;
1016         struct file *file;
1017         unsigned long bytes, offset;
1018         int outofdate;
1019         int ret = -ENOSPC;
1020         void *paddr;
1021
1022         chunks = bitmap->chunks;
1023         file = bitmap->file;
1024
1025         BUG_ON(!file && !bitmap->mddev->bitmap_info.offset);
1026
1027 #ifdef INJECT_FAULTS_3
1028         outofdate = 1;
1029 #else
1030         outofdate = bitmap->flags & BITMAP_STALE;
1031 #endif
1032         if (outofdate)
1033                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
1034                         "recovery\n", bmname(bitmap));
1035
1036         bytes = DIV_ROUND_UP(bitmap->chunks, 8);
1037         if (!bitmap->mddev->bitmap_info.external)
1038                 bytes += sizeof(bitmap_super_t);
1039
1040         num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
1041
1042         if (file && i_size_read(file->f_mapping->host) < bytes) {
1043                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
1044                         bmname(bitmap),
1045                         (unsigned long) i_size_read(file->f_mapping->host),
1046                         bytes);
1047                 goto err;
1048         }
1049
1050         ret = -ENOMEM;
1051
1052         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
1053         if (!bitmap->filemap)
1054                 goto err;
1055
1056         /* We need 4 bits per page, rounded up to a multiple of sizeof(unsigned long) */
1057         bitmap->filemap_attr = kzalloc(
1058                 roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
1059                 GFP_KERNEL);
1060         if (!bitmap->filemap_attr)
1061                 goto err;
1062
1063         oldindex = ~0L;
1064
1065         for (i = 0; i < chunks; i++) {
1066                 int b;
1067                 index = file_page_index(bitmap, i);
1068                 bit = file_page_offset(bitmap, i);
1069                 if (index != oldindex) { /* this is a new page, read it in */
1070                         int count;
1071                         /* unmap the old page, we're done with it */
1072                         if (index == num_pages-1)
1073                                 count = bytes - index * PAGE_SIZE;
1074                         else
1075                                 count = PAGE_SIZE;
1076                         if (index == 0 && bitmap->sb_page) {
1077                                 /*
1078                                  * if we're here then the superblock page
1079                                  * contains some bits (PAGE_SIZE != sizeof sb)
1080                                  * we've already read it in, so just use it
1081                                  */
1082                                 page = bitmap->sb_page;
1083                                 offset = sizeof(bitmap_super_t);
1084                                 if (!file)
1085                                         page = read_sb_page(
1086                                                 bitmap->mddev,
1087                                                 bitmap->mddev->bitmap_info.offset,
1088                                                 page,
1089                                                 index, count);
1090                         } else if (file) {
1091                                 page = read_page(file, index, bitmap, count);
1092                                 offset = 0;
1093                         } else {
1094                                 page = read_sb_page(bitmap->mddev,
1095                                                     bitmap->mddev->bitmap_info.offset,
1096                                                     NULL,
1097                                                     index, count);
1098                                 offset = 0;
1099                         }
1100                         if (IS_ERR(page)) { /* read error */
1101                                 ret = PTR_ERR(page);
1102                                 goto err;
1103                         }
1104
1105                         oldindex = index;
1106                         oldpage = page;
1107
1108                         bitmap->filemap[bitmap->file_pages++] = page;
1109                         bitmap->last_page_size = count;
1110
1111                         if (outofdate) {
1112                                 /*
1113                                  * if bitmap is out of date, dirty the
1114                                  * whole page and write it out
1115                                  */
1116                                 paddr = kmap_atomic(page, KM_USER0);
1117                                 memset(paddr + offset, 0xff,
1118                                        PAGE_SIZE - offset);
1119                                 kunmap_atomic(paddr, KM_USER0);
1120                                 write_page(bitmap, page, 1);
1121
1122                                 ret = -EIO;
1123                                 if (bitmap->flags & BITMAP_WRITE_ERROR)
1124                                         goto err;
1125                         }
1126                 }
1127                 paddr = kmap_atomic(page, KM_USER0);
1128                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1129                         b = test_bit(bit, paddr);
1130                 else
1131                         b = test_bit_le(bit, paddr);
1132                 kunmap_atomic(paddr, KM_USER0);
1133                 if (b) {
1134                         /* if the disk bit is set, set the memory bit */
1135                         int needed = ((sector_t)(i+1) << (CHUNK_BLOCK_SHIFT(bitmap))
1136                                       >= start);
1137                         bitmap_set_memory_bits(bitmap,
1138                                                (sector_t)i << CHUNK_BLOCK_SHIFT(bitmap),
1139                                                needed);
1140                         bit_cnt++;
1141                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1142                 }
1143         }
1144
1145         /* everything went OK */
1146         ret = 0;
1147         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
1148
1149         if (bit_cnt) { /* Kick recovery if any bits were set */
1150                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1151                 md_wakeup_thread(bitmap->mddev->thread);
1152         }
1153
1154         printk(KERN_INFO "%s: bitmap initialized from disk: "
1155                "read %lu/%lu pages, set %lu of %lu bits\n",
1156                bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, chunks);
1157
1158         return 0;
1159
1160  err:
1161         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1162                bmname(bitmap), ret);
1163         return ret;
1164 }
1165
1166 void bitmap_write_all(struct bitmap *bitmap)
1167 {
1168         /* We don't actually write all bitmap blocks here,
1169          * just flag them as needing to be written
1170          */
1171         int i;
1172
1173         for (i = 0; i < bitmap->file_pages; i++)
1174                 set_page_attr(bitmap, bitmap->filemap[i],
1175                               BITMAP_PAGE_NEEDWRITE);
1176 }
1177
1178 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
1179 {
1180         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1181         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1182         bitmap->bp[page].count += inc;
1183         bitmap_checkfree(bitmap, page);
1184 }
1185 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1186                                             sector_t offset, sector_t *blocks,
1187                                             int create);
1188
1189 /*
1190  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1191  *                      out to disk
1192  */
1193
1194 void bitmap_daemon_work(mddev_t *mddev)
1195 {
1196         struct bitmap *bitmap;
1197         unsigned long j;
1198         unsigned long flags;
1199         struct page *page = NULL, *lastpage = NULL;
1200         sector_t blocks;
1201         void *paddr;
1202         struct dm_dirty_log *log = mddev->bitmap_info.log;
1203
1204         /* Use a mutex to guard daemon_work against
1205          * bitmap_destroy.
1206          */
1207         mutex_lock(&mddev->bitmap_info.mutex);
1208         bitmap = mddev->bitmap;
1209         if (bitmap == NULL) {
1210                 mutex_unlock(&mddev->bitmap_info.mutex);
1211                 return;
1212         }
1213         if (time_before(jiffies, bitmap->daemon_lastrun
1214                         + bitmap->mddev->bitmap_info.daemon_sleep))
1215                 goto done;
1216
1217         bitmap->daemon_lastrun = jiffies;
1218         if (bitmap->allclean) {
1219                 bitmap->mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1220                 goto done;
1221         }
1222         bitmap->allclean = 1;
1223
1224         spin_lock_irqsave(&bitmap->lock, flags);
1225         for (j = 0; j < bitmap->chunks; j++) {
1226                 bitmap_counter_t *bmc;
1227                 if (!bitmap->filemap) {
1228                         if (!log)
1229                                 /* error or shutdown */
1230                                 break;
1231                 } else
1232                         page = filemap_get_page(bitmap, j);
1233
1234                 if (page != lastpage) {
1235                         /* skip this page unless it's marked as needing cleaning */
1236                         if (!test_page_attr(bitmap, page, BITMAP_PAGE_CLEAN)) {
1237                                 int need_write = test_page_attr(bitmap, page,
1238                                                                 BITMAP_PAGE_NEEDWRITE);
1239                                 if (need_write)
1240                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1241
1242                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1243                                 if (need_write) {
1244                                         write_page(bitmap, page, 0);
1245                                         bitmap->allclean = 0;
1246                                 }
1247                                 spin_lock_irqsave(&bitmap->lock, flags);
1248                                 j |= (PAGE_BITS - 1);
1249                                 continue;
1250                         }
1251
1252                         /* grab the new page, sync and release the old */
1253                         if (lastpage != NULL) {
1254                                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1255                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1256                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1257                                         write_page(bitmap, lastpage, 0);
1258                                 } else {
1259                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1260                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1261                                 }
1262                         } else
1263                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1264                         lastpage = page;
1265
1266                         /* We are possibly going to clear some bits, so make
1267                          * sure that events_cleared is up-to-date.
1268                          */
1269                         if (bitmap->need_sync &&
1270                             bitmap->mddev->bitmap_info.external == 0) {
1271                                 bitmap_super_t *sb;
1272                                 bitmap->need_sync = 0;
1273                                 sb = kmap_atomic(bitmap->sb_page, KM_USER0);
1274                                 sb->events_cleared =
1275                                         cpu_to_le64(bitmap->events_cleared);
1276                                 kunmap_atomic(sb, KM_USER0);
1277                                 write_page(bitmap, bitmap->sb_page, 1);
1278                         }
1279                         spin_lock_irqsave(&bitmap->lock, flags);
1280                         if (!bitmap->need_sync)
1281                                 clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1282                 }
1283                 bmc = bitmap_get_counter(bitmap,
1284                                          (sector_t)j << CHUNK_BLOCK_SHIFT(bitmap),
1285                                          &blocks, 0);
1286                 if (bmc) {
1287                         if (*bmc)
1288                                 bitmap->allclean = 0;
1289
1290                         if (*bmc == 2) {
1291                                 *bmc = 1; /* maybe clear the bit next time */
1292                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1293                         } else if (*bmc == 1 && !bitmap->need_sync) {
1294                                 /* we can clear the bit */
1295                                 *bmc = 0;
1296                                 bitmap_count_page(bitmap,
1297                                                   (sector_t)j << CHUNK_BLOCK_SHIFT(bitmap),
1298                                                   -1);
1299
1300                                 /* clear the bit */
1301                                 if (page) {
1302                                         paddr = kmap_atomic(page, KM_USER0);
1303                                         if (bitmap->flags & BITMAP_HOSTENDIAN)
1304                                                 clear_bit(file_page_offset(bitmap, j),
1305                                                           paddr);
1306                                         else
1307                                                 __clear_bit_le(
1308                                                         file_page_offset(bitmap,
1309                                                                          j),
1310                                                         paddr);
1311                                         kunmap_atomic(paddr, KM_USER0);
1312                                 } else
1313                                         log->type->clear_region(log, j);
1314                         }
1315                 } else
1316                         j |= PAGE_COUNTER_MASK;
1317         }
1318         spin_unlock_irqrestore(&bitmap->lock, flags);
1319
1320         /* now sync the final page */
1321         if (lastpage != NULL || log != NULL) {
1322                 spin_lock_irqsave(&bitmap->lock, flags);
1323                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1324                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1325                         spin_unlock_irqrestore(&bitmap->lock, flags);
1326                         if (lastpage)
1327                                 write_page(bitmap, lastpage, 0);
1328                         else
1329                                 if (log->type->flush(log))
1330                                         bitmap->flags |= BITMAP_WRITE_ERROR;
1331                 } else {
1332                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1333                         spin_unlock_irqrestore(&bitmap->lock, flags);
1334                 }
1335         }
1336
1337  done:
1338         if (bitmap->allclean == 0)
1339                 bitmap->mddev->thread->timeout =
1340                         bitmap->mddev->bitmap_info.daemon_sleep;
1341         mutex_unlock(&mddev->bitmap_info.mutex);
1342 }
1343
1344 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1345                                             sector_t offset, sector_t *blocks,
1346                                             int create)
1347 __releases(bitmap->lock)
1348 __acquires(bitmap->lock)
1349 {
1350         /* If 'create', we might release the lock and reclaim it.
1351          * The lock must have been taken with interrupts enabled.
1352          * If !create, we don't release the lock.
1353          */
1354         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1355         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1356         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1357         sector_t csize;
1358         int err;
1359
1360         err = bitmap_checkpage(bitmap, page, create);
1361
1362         if (bitmap->bp[page].hijacked ||
1363             bitmap->bp[page].map == NULL)
1364                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1365                                           PAGE_COUNTER_SHIFT - 1);
1366         else
1367                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1368         *blocks = csize - (offset & (csize - 1));
1369
1370         if (err < 0)
1371                 return NULL;
1372
1373         /* now locked ... */
1374
1375         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1376                 /* should we use the first or second counter field
1377                  * of the hijacked pointer? */
1378                 int hi = (pageoff > PAGE_COUNTER_MASK);
1379                 return  &((bitmap_counter_t *)
1380                           &bitmap->bp[page].map)[hi];
1381         } else /* page is allocated */
1382                 return (bitmap_counter_t *)
1383                         &(bitmap->bp[page].map[pageoff]);
1384 }
1385
1386 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1387 {
1388         if (!bitmap)
1389                 return 0;
1390
1391         if (behind) {
1392                 int bw;
1393                 atomic_inc(&bitmap->behind_writes);
1394                 bw = atomic_read(&bitmap->behind_writes);
1395                 if (bw > bitmap->behind_writes_used)
1396                         bitmap->behind_writes_used = bw;
1397
1398                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1399                        bw, bitmap->max_write_behind);
1400         }
1401
1402         while (sectors) {
1403                 sector_t blocks;
1404                 bitmap_counter_t *bmc;
1405
1406                 spin_lock_irq(&bitmap->lock);
1407                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1408                 if (!bmc) {
1409                         spin_unlock_irq(&bitmap->lock);
1410                         return 0;
1411                 }
1412
1413                 if (unlikely(COUNTER(*bmc) == COUNTER_MAX)) {
1414                         DEFINE_WAIT(__wait);
1415                         /* note that it is safe to do the prepare_to_wait
1416                          * after the test as long as we do it before dropping
1417                          * the spinlock.
1418                          */
1419                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1420                                         TASK_UNINTERRUPTIBLE);
1421                         spin_unlock_irq(&bitmap->lock);
1422                         io_schedule();
1423                         finish_wait(&bitmap->overflow_wait, &__wait);
1424                         continue;
1425                 }
1426
1427                 switch (*bmc) {
1428                 case 0:
1429                         bitmap_file_set_bit(bitmap, offset);
1430                         bitmap_count_page(bitmap, offset, 1);
1431                         /* fall through */
1432                 case 1:
1433                         *bmc = 2;
1434                 }
1435
1436                 (*bmc)++;
1437
1438                 spin_unlock_irq(&bitmap->lock);
1439
1440                 offset += blocks;
1441                 if (sectors > blocks)
1442                         sectors -= blocks;
1443                 else
1444                         sectors = 0;
1445         }
1446         bitmap->allclean = 0;
1447         return 0;
1448 }
1449 EXPORT_SYMBOL(bitmap_startwrite);
1450
1451 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1452                      int success, int behind)
1453 {
1454         if (!bitmap)
1455                 return;
1456         if (behind) {
1457                 if (atomic_dec_and_test(&bitmap->behind_writes))
1458                         wake_up(&bitmap->behind_wait);
1459                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1460                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1461         }
1462         if (bitmap->mddev->degraded)
1463                 /* Never clear bits or update events_cleared when degraded */
1464                 success = 0;
1465
1466         while (sectors) {
1467                 sector_t blocks;
1468                 unsigned long flags;
1469                 bitmap_counter_t *bmc;
1470
1471                 spin_lock_irqsave(&bitmap->lock, flags);
1472                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1473                 if (!bmc) {
1474                         spin_unlock_irqrestore(&bitmap->lock, flags);
1475                         return;
1476                 }
1477
1478                 if (success &&
1479                     bitmap->events_cleared < bitmap->mddev->events) {
1480                         bitmap->events_cleared = bitmap->mddev->events;
1481                         bitmap->need_sync = 1;
1482                         sysfs_notify_dirent_safe(bitmap->sysfs_can_clear);
1483                 }
1484
1485                 if (!success && !NEEDED(*bmc))
1486                         *bmc |= NEEDED_MASK;
1487
1488                 if (COUNTER(*bmc) == COUNTER_MAX)
1489                         wake_up(&bitmap->overflow_wait);
1490
1491                 (*bmc)--;
1492                 if (*bmc <= 2)
1493                         set_page_attr(bitmap,
1494                                       filemap_get_page(
1495                                               bitmap,
1496                                               offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1497                                       BITMAP_PAGE_CLEAN);
1498
1499                 spin_unlock_irqrestore(&bitmap->lock, flags);
1500                 offset += blocks;
1501                 if (sectors > blocks)
1502                         sectors -= blocks;
1503                 else
1504                         sectors = 0;
1505         }
1506 }
1507 EXPORT_SYMBOL(bitmap_endwrite);
1508
1509 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1510                                int degraded)
1511 {
1512         bitmap_counter_t *bmc;
1513         int rv;
1514         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1515                 *blocks = 1024;
1516                 return 1; /* always resync if no bitmap */
1517         }
1518         spin_lock_irq(&bitmap->lock);
1519         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1520         rv = 0;
1521         if (bmc) {
1522                 /* locked */
1523                 if (RESYNC(*bmc))
1524                         rv = 1;
1525                 else if (NEEDED(*bmc)) {
1526                         rv = 1;
1527                         if (!degraded) { /* don't set/clear bits if degraded */
1528                                 *bmc |= RESYNC_MASK;
1529                                 *bmc &= ~NEEDED_MASK;
1530                         }
1531                 }
1532         }
1533         spin_unlock_irq(&bitmap->lock);
1534         bitmap->allclean = 0;
1535         return rv;
1536 }
1537
1538 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1539                       int degraded)
1540 {
1541         /* bitmap_start_sync must always report on multiples of whole
1542          * pages, otherwise resync (which is very PAGE_SIZE based) will
1543          * get confused.
1544          * So call __bitmap_start_sync repeatedly (if needed) until
1545          * At least PAGE_SIZE>>9 blocks are covered.
1546          * Return the 'or' of the result.
1547          */
1548         int rv = 0;
1549         sector_t blocks1;
1550
1551         *blocks = 0;
1552         while (*blocks < (PAGE_SIZE>>9)) {
1553                 rv |= __bitmap_start_sync(bitmap, offset,
1554                                           &blocks1, degraded);
1555                 offset += blocks1;
1556                 *blocks += blocks1;
1557         }
1558         return rv;
1559 }
1560 EXPORT_SYMBOL(bitmap_start_sync);
1561
1562 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks, int aborted)
1563 {
1564         bitmap_counter_t *bmc;
1565         unsigned long flags;
1566
1567         if (bitmap == NULL) {
1568                 *blocks = 1024;
1569                 return;
1570         }
1571         spin_lock_irqsave(&bitmap->lock, flags);
1572         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1573         if (bmc == NULL)
1574                 goto unlock;
1575         /* locked */
1576         if (RESYNC(*bmc)) {
1577                 *bmc &= ~RESYNC_MASK;
1578
1579                 if (!NEEDED(*bmc) && aborted)
1580                         *bmc |= NEEDED_MASK;
1581                 else {
1582                         if (*bmc <= 2)
1583                                 set_page_attr(bitmap,
1584                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1585                                               BITMAP_PAGE_CLEAN);
1586                 }
1587         }
1588  unlock:
1589         spin_unlock_irqrestore(&bitmap->lock, flags);
1590         bitmap->allclean = 0;
1591 }
1592 EXPORT_SYMBOL(bitmap_end_sync);
1593
1594 void bitmap_close_sync(struct bitmap *bitmap)
1595 {
1596         /* Sync has finished, and any bitmap chunks that weren't synced
1597          * properly have been aborted.  It remains to us to clear the
1598          * RESYNC bit wherever it is still on
1599          */
1600         sector_t sector = 0;
1601         sector_t blocks;
1602         if (!bitmap)
1603                 return;
1604         while (sector < bitmap->mddev->resync_max_sectors) {
1605                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1606                 sector += blocks;
1607         }
1608 }
1609 EXPORT_SYMBOL(bitmap_close_sync);
1610
1611 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1612 {
1613         sector_t s = 0;
1614         sector_t blocks;
1615
1616         if (!bitmap)
1617                 return;
1618         if (sector == 0) {
1619                 bitmap->last_end_sync = jiffies;
1620                 return;
1621         }
1622         if (time_before(jiffies, (bitmap->last_end_sync
1623                                   + bitmap->mddev->bitmap_info.daemon_sleep)))
1624                 return;
1625         wait_event(bitmap->mddev->recovery_wait,
1626                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1627
1628         bitmap->mddev->curr_resync_completed = sector;
1629         set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
1630         sector &= ~((1ULL << CHUNK_BLOCK_SHIFT(bitmap)) - 1);
1631         s = 0;
1632         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1633                 bitmap_end_sync(bitmap, s, &blocks, 0);
1634                 s += blocks;
1635         }
1636         bitmap->last_end_sync = jiffies;
1637         sysfs_notify(&bitmap->mddev->kobj, NULL, "sync_completed");
1638 }
1639 EXPORT_SYMBOL(bitmap_cond_end_sync);
1640
1641 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1642 {
1643         /* For each chunk covered by any of these sectors, set the
1644          * counter to 1 and set resync_needed.  They should all
1645          * be 0 at this point
1646          */
1647
1648         sector_t secs;
1649         bitmap_counter_t *bmc;
1650         spin_lock_irq(&bitmap->lock);
1651         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1652         if (!bmc) {
1653                 spin_unlock_irq(&bitmap->lock);
1654                 return;
1655         }
1656         if (!*bmc) {
1657                 struct page *page;
1658                 *bmc = 1 | (needed ? NEEDED_MASK : 0);
1659                 bitmap_count_page(bitmap, offset, 1);
1660                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1661                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1662         }
1663         spin_unlock_irq(&bitmap->lock);
1664         bitmap->allclean = 0;
1665 }
1666
1667 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1668 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1669 {
1670         unsigned long chunk;
1671
1672         for (chunk = s; chunk <= e; chunk++) {
1673                 sector_t sec = (sector_t)chunk << CHUNK_BLOCK_SHIFT(bitmap);
1674                 bitmap_set_memory_bits(bitmap, sec, 1);
1675                 bitmap_file_set_bit(bitmap, sec);
1676                 if (sec < bitmap->mddev->recovery_cp)
1677                         /* We are asserting that the array is dirty,
1678                          * so move the recovery_cp address back so
1679                          * that it is obvious that it is dirty
1680                          */
1681                         bitmap->mddev->recovery_cp = sec;
1682         }
1683 }
1684
1685 /*
1686  * flush out any pending updates
1687  */
1688 void bitmap_flush(mddev_t *mddev)
1689 {
1690         struct bitmap *bitmap = mddev->bitmap;
1691         long sleep;
1692
1693         if (!bitmap) /* there was no bitmap */
1694                 return;
1695
1696         /* run the daemon_work three time to ensure everything is flushed
1697          * that can be
1698          */
1699         sleep = mddev->bitmap_info.daemon_sleep * 2;
1700         bitmap->daemon_lastrun -= sleep;
1701         bitmap_daemon_work(mddev);
1702         bitmap->daemon_lastrun -= sleep;
1703         bitmap_daemon_work(mddev);
1704         bitmap->daemon_lastrun -= sleep;
1705         bitmap_daemon_work(mddev);
1706         bitmap_update_sb(bitmap);
1707 }
1708
1709 /*
1710  * free memory that was allocated
1711  */
1712 static void bitmap_free(struct bitmap *bitmap)
1713 {
1714         unsigned long k, pages;
1715         struct bitmap_page *bp;
1716
1717         if (!bitmap) /* there was no bitmap */
1718                 return;
1719
1720         /* release the bitmap file and kill the daemon */
1721         bitmap_file_put(bitmap);
1722
1723         bp = bitmap->bp;
1724         pages = bitmap->pages;
1725
1726         /* free all allocated memory */
1727
1728         if (bp) /* deallocate the page memory */
1729                 for (k = 0; k < pages; k++)
1730                         if (bp[k].map && !bp[k].hijacked)
1731                                 kfree(bp[k].map);
1732         kfree(bp);
1733         kfree(bitmap);
1734 }
1735
1736 void bitmap_destroy(mddev_t *mddev)
1737 {
1738         struct bitmap *bitmap = mddev->bitmap;
1739
1740         if (!bitmap) /* there was no bitmap */
1741                 return;
1742
1743         mutex_lock(&mddev->bitmap_info.mutex);
1744         mddev->bitmap = NULL; /* disconnect from the md device */
1745         mutex_unlock(&mddev->bitmap_info.mutex);
1746         if (mddev->thread)
1747                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1748
1749         if (bitmap->sysfs_can_clear)
1750                 sysfs_put(bitmap->sysfs_can_clear);
1751
1752         bitmap_free(bitmap);
1753 }
1754
1755 /*
1756  * initialize the bitmap structure
1757  * if this returns an error, bitmap_destroy must be called to do clean up
1758  */
1759 int bitmap_create(mddev_t *mddev)
1760 {
1761         struct bitmap *bitmap;
1762         sector_t blocks = mddev->resync_max_sectors;
1763         unsigned long chunks;
1764         unsigned long pages;
1765         struct file *file = mddev->bitmap_info.file;
1766         int err;
1767         struct sysfs_dirent *bm = NULL;
1768
1769         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1770
1771         if (!file
1772             && !mddev->bitmap_info.offset
1773             && !mddev->bitmap_info.log) /* bitmap disabled, nothing to do */
1774                 return 0;
1775
1776         BUG_ON(file && mddev->bitmap_info.offset);
1777         BUG_ON(mddev->bitmap_info.offset && mddev->bitmap_info.log);
1778
1779         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1780         if (!bitmap)
1781                 return -ENOMEM;
1782
1783         spin_lock_init(&bitmap->lock);
1784         atomic_set(&bitmap->pending_writes, 0);
1785         init_waitqueue_head(&bitmap->write_wait);
1786         init_waitqueue_head(&bitmap->overflow_wait);
1787         init_waitqueue_head(&bitmap->behind_wait);
1788
1789         bitmap->mddev = mddev;
1790
1791         if (mddev->kobj.sd)
1792                 bm = sysfs_get_dirent(mddev->kobj.sd, NULL, "bitmap");
1793         if (bm) {
1794                 bitmap->sysfs_can_clear = sysfs_get_dirent(bm, NULL, "can_clear");
1795                 sysfs_put(bm);
1796         } else
1797                 bitmap->sysfs_can_clear = NULL;
1798
1799         bitmap->file = file;
1800         if (file) {
1801                 get_file(file);
1802                 /* As future accesses to this file will use bmap,
1803                  * and bypass the page cache, we must sync the file
1804                  * first.
1805                  */
1806                 vfs_fsync(file, 1);
1807         }
1808         /* read superblock from bitmap file (this sets mddev->bitmap_info.chunksize) */
1809         if (!mddev->bitmap_info.external) {
1810                 /*
1811                  * If 'MD_ARRAY_FIRST_USE' is set, then device-mapper is
1812                  * instructing us to create a new on-disk bitmap instance.
1813                  */
1814                 if (test_and_clear_bit(MD_ARRAY_FIRST_USE, &mddev->flags))
1815                         err = bitmap_new_disk_sb(bitmap);
1816                 else
1817                         err = bitmap_read_sb(bitmap);
1818         } else {
1819                 err = 0;
1820                 if (mddev->bitmap_info.chunksize == 0 ||
1821                     mddev->bitmap_info.daemon_sleep == 0)
1822                         /* chunksize and time_base need to be
1823                          * set first. */
1824                         err = -EINVAL;
1825         }
1826         if (err)
1827                 goto error;
1828
1829         bitmap->daemon_lastrun = jiffies;
1830         bitmap->chunkshift = ffz(~mddev->bitmap_info.chunksize);
1831
1832         /* now that chunksize and chunkshift are set, we can use these macros */
1833         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) >>
1834                         CHUNK_BLOCK_SHIFT(bitmap);
1835         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1836
1837         BUG_ON(!pages);
1838
1839         bitmap->chunks = chunks;
1840         bitmap->pages = pages;
1841         bitmap->missing_pages = pages;
1842
1843 #ifdef INJECT_FATAL_FAULT_1
1844         bitmap->bp = NULL;
1845 #else
1846         bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1847 #endif
1848         err = -ENOMEM;
1849         if (!bitmap->bp)
1850                 goto error;
1851
1852         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1853                 pages, bmname(bitmap));
1854
1855         mddev->bitmap = bitmap;
1856
1857
1858         return (bitmap->flags & BITMAP_WRITE_ERROR) ? -EIO : 0;
1859
1860  error:
1861         bitmap_free(bitmap);
1862         return err;
1863 }
1864
1865 int bitmap_load(mddev_t *mddev)
1866 {
1867         int err = 0;
1868         sector_t sector = 0;
1869         struct bitmap *bitmap = mddev->bitmap;
1870
1871         if (!bitmap)
1872                 goto out;
1873
1874         /* Clear out old bitmap info first:  Either there is none, or we
1875          * are resuming after someone else has possibly changed things,
1876          * so we should forget old cached info.
1877          * All chunks should be clean, but some might need_sync.
1878          */
1879         while (sector < mddev->resync_max_sectors) {
1880                 sector_t blocks;
1881                 bitmap_start_sync(bitmap, sector, &blocks, 0);
1882                 sector += blocks;
1883         }
1884         bitmap_close_sync(bitmap);
1885
1886         if (mddev->bitmap_info.log) {
1887                 unsigned long i;
1888                 struct dm_dirty_log *log = mddev->bitmap_info.log;
1889                 for (i = 0; i < bitmap->chunks; i++)
1890                         if (!log->type->in_sync(log, i, 1))
1891                                 bitmap_set_memory_bits(bitmap,
1892                                                        (sector_t)i << CHUNK_BLOCK_SHIFT(bitmap),
1893                                                        1);
1894         } else {
1895                 sector_t start = 0;
1896                 if (mddev->degraded == 0
1897                     || bitmap->events_cleared == mddev->events)
1898                         /* no need to keep dirty bits to optimise a
1899                          * re-add of a missing device */
1900                         start = mddev->recovery_cp;
1901
1902                 err = bitmap_init_from_disk(bitmap, start);
1903         }
1904         if (err)
1905                 goto out;
1906
1907         mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
1908         md_wakeup_thread(mddev->thread);
1909
1910         bitmap_update_sb(bitmap);
1911
1912         if (bitmap->flags & BITMAP_WRITE_ERROR)
1913                 err = -EIO;
1914 out:
1915         return err;
1916 }
1917 EXPORT_SYMBOL_GPL(bitmap_load);
1918
1919 static ssize_t
1920 location_show(mddev_t *mddev, char *page)
1921 {
1922         ssize_t len;
1923         if (mddev->bitmap_info.file)
1924                 len = sprintf(page, "file");
1925         else if (mddev->bitmap_info.offset)
1926                 len = sprintf(page, "%+lld", (long long)mddev->bitmap_info.offset);
1927         else
1928                 len = sprintf(page, "none");
1929         len += sprintf(page+len, "\n");
1930         return len;
1931 }
1932
1933 static ssize_t
1934 location_store(mddev_t *mddev, const char *buf, size_t len)
1935 {
1936
1937         if (mddev->pers) {
1938                 if (!mddev->pers->quiesce)
1939                         return -EBUSY;
1940                 if (mddev->recovery || mddev->sync_thread)
1941                         return -EBUSY;
1942         }
1943
1944         if (mddev->bitmap || mddev->bitmap_info.file ||
1945             mddev->bitmap_info.offset) {
1946                 /* bitmap already configured.  Only option is to clear it */
1947                 if (strncmp(buf, "none", 4) != 0)
1948                         return -EBUSY;
1949                 if (mddev->pers) {
1950                         mddev->pers->quiesce(mddev, 1);
1951                         bitmap_destroy(mddev);
1952                         mddev->pers->quiesce(mddev, 0);
1953                 }
1954                 mddev->bitmap_info.offset = 0;
1955                 if (mddev->bitmap_info.file) {
1956                         struct file *f = mddev->bitmap_info.file;
1957                         mddev->bitmap_info.file = NULL;
1958                         restore_bitmap_write_access(f);
1959                         fput(f);
1960                 }
1961         } else {
1962                 /* No bitmap, OK to set a location */
1963                 long long offset;
1964                 if (strncmp(buf, "none", 4) == 0)
1965                         /* nothing to be done */;
1966                 else if (strncmp(buf, "file:", 5) == 0) {
1967                         /* Not supported yet */
1968                         return -EINVAL;
1969                 } else {
1970                         int rv;
1971                         if (buf[0] == '+')
1972                                 rv = strict_strtoll(buf+1, 10, &offset);
1973                         else
1974                                 rv = strict_strtoll(buf, 10, &offset);
1975                         if (rv)
1976                                 return rv;
1977                         if (offset == 0)
1978                                 return -EINVAL;
1979                         if (mddev->bitmap_info.external == 0 &&
1980                             mddev->major_version == 0 &&
1981                             offset != mddev->bitmap_info.default_offset)
1982                                 return -EINVAL;
1983                         mddev->bitmap_info.offset = offset;
1984                         if (mddev->pers) {
1985                                 mddev->pers->quiesce(mddev, 1);
1986                                 rv = bitmap_create(mddev);
1987                                 if (rv) {
1988                                         bitmap_destroy(mddev);
1989                                         mddev->bitmap_info.offset = 0;
1990                                 }
1991                                 mddev->pers->quiesce(mddev, 0);
1992                                 if (rv)
1993                                         return rv;
1994                         }
1995                 }
1996         }
1997         if (!mddev->external) {
1998                 /* Ensure new bitmap info is stored in
1999                  * metadata promptly.
2000                  */
2001                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2002                 md_wakeup_thread(mddev->thread);
2003         }
2004         return len;
2005 }
2006
2007 static struct md_sysfs_entry bitmap_location =
2008 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
2009
2010 static ssize_t
2011 timeout_show(mddev_t *mddev, char *page)
2012 {
2013         ssize_t len;
2014         unsigned long secs = mddev->bitmap_info.daemon_sleep / HZ;
2015         unsigned long jifs = mddev->bitmap_info.daemon_sleep % HZ;
2016
2017         len = sprintf(page, "%lu", secs);
2018         if (jifs)
2019                 len += sprintf(page+len, ".%03u", jiffies_to_msecs(jifs));
2020         len += sprintf(page+len, "\n");
2021         return len;
2022 }
2023
2024 static ssize_t
2025 timeout_store(mddev_t *mddev, const char *buf, size_t len)
2026 {
2027         /* timeout can be set at any time */
2028         unsigned long timeout;
2029         int rv = strict_strtoul_scaled(buf, &timeout, 4);
2030         if (rv)
2031                 return rv;
2032
2033         /* just to make sure we don't overflow... */
2034         if (timeout >= LONG_MAX / HZ)
2035                 return -EINVAL;
2036
2037         timeout = timeout * HZ / 10000;
2038
2039         if (timeout >= MAX_SCHEDULE_TIMEOUT)
2040                 timeout = MAX_SCHEDULE_TIMEOUT-1;
2041         if (timeout < 1)
2042                 timeout = 1;
2043         mddev->bitmap_info.daemon_sleep = timeout;
2044         if (mddev->thread) {
2045                 /* if thread->timeout is MAX_SCHEDULE_TIMEOUT, then
2046                  * the bitmap is all clean and we don't need to
2047                  * adjust the timeout right now
2048                  */
2049                 if (mddev->thread->timeout < MAX_SCHEDULE_TIMEOUT) {
2050                         mddev->thread->timeout = timeout;
2051                         md_wakeup_thread(mddev->thread);
2052                 }
2053         }
2054         return len;
2055 }
2056
2057 static struct md_sysfs_entry bitmap_timeout =
2058 __ATTR(time_base, S_IRUGO|S_IWUSR, timeout_show, timeout_store);
2059
2060 static ssize_t
2061 backlog_show(mddev_t *mddev, char *page)
2062 {
2063         return sprintf(page, "%lu\n", mddev->bitmap_info.max_write_behind);
2064 }
2065
2066 static ssize_t
2067 backlog_store(mddev_t *mddev, const char *buf, size_t len)
2068 {
2069         unsigned long backlog;
2070         int rv = strict_strtoul(buf, 10, &backlog);
2071         if (rv)
2072                 return rv;
2073         if (backlog > COUNTER_MAX)
2074                 return -EINVAL;
2075         mddev->bitmap_info.max_write_behind = backlog;
2076         return len;
2077 }
2078
2079 static struct md_sysfs_entry bitmap_backlog =
2080 __ATTR(backlog, S_IRUGO|S_IWUSR, backlog_show, backlog_store);
2081
2082 static ssize_t
2083 chunksize_show(mddev_t *mddev, char *page)
2084 {
2085         return sprintf(page, "%lu\n", mddev->bitmap_info.chunksize);
2086 }
2087
2088 static ssize_t
2089 chunksize_store(mddev_t *mddev, const char *buf, size_t len)
2090 {
2091         /* Can only be changed when no bitmap is active */
2092         int rv;
2093         unsigned long csize;
2094         if (mddev->bitmap)
2095                 return -EBUSY;
2096         rv = strict_strtoul(buf, 10, &csize);
2097         if (rv)
2098                 return rv;
2099         if (csize < 512 ||
2100             !is_power_of_2(csize))
2101                 return -EINVAL;
2102         mddev->bitmap_info.chunksize = csize;
2103         return len;
2104 }
2105
2106 static struct md_sysfs_entry bitmap_chunksize =
2107 __ATTR(chunksize, S_IRUGO|S_IWUSR, chunksize_show, chunksize_store);
2108
2109 static ssize_t metadata_show(mddev_t *mddev, char *page)
2110 {
2111         return sprintf(page, "%s\n", (mddev->bitmap_info.external
2112                                       ? "external" : "internal"));
2113 }
2114
2115 static ssize_t metadata_store(mddev_t *mddev, const char *buf, size_t len)
2116 {
2117         if (mddev->bitmap ||
2118             mddev->bitmap_info.file ||
2119             mddev->bitmap_info.offset)
2120                 return -EBUSY;
2121         if (strncmp(buf, "external", 8) == 0)
2122                 mddev->bitmap_info.external = 1;
2123         else if (strncmp(buf, "internal", 8) == 0)
2124                 mddev->bitmap_info.external = 0;
2125         else
2126                 return -EINVAL;
2127         return len;
2128 }
2129
2130 static struct md_sysfs_entry bitmap_metadata =
2131 __ATTR(metadata, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
2132
2133 static ssize_t can_clear_show(mddev_t *mddev, char *page)
2134 {
2135         int len;
2136         if (mddev->bitmap)
2137                 len = sprintf(page, "%s\n", (mddev->bitmap->need_sync ?
2138                                              "false" : "true"));
2139         else
2140                 len = sprintf(page, "\n");
2141         return len;
2142 }
2143
2144 static ssize_t can_clear_store(mddev_t *mddev, const char *buf, size_t len)
2145 {
2146         if (mddev->bitmap == NULL)
2147                 return -ENOENT;
2148         if (strncmp(buf, "false", 5) == 0)
2149                 mddev->bitmap->need_sync = 1;
2150         else if (strncmp(buf, "true", 4) == 0) {
2151                 if (mddev->degraded)
2152                         return -EBUSY;
2153                 mddev->bitmap->need_sync = 0;
2154         } else
2155                 return -EINVAL;
2156         return len;
2157 }
2158
2159 static struct md_sysfs_entry bitmap_can_clear =
2160 __ATTR(can_clear, S_IRUGO|S_IWUSR, can_clear_show, can_clear_store);
2161
2162 static ssize_t
2163 behind_writes_used_show(mddev_t *mddev, char *page)
2164 {
2165         if (mddev->bitmap == NULL)
2166                 return sprintf(page, "0\n");
2167         return sprintf(page, "%lu\n",
2168                        mddev->bitmap->behind_writes_used);
2169 }
2170
2171 static ssize_t
2172 behind_writes_used_reset(mddev_t *mddev, const char *buf, size_t len)
2173 {
2174         if (mddev->bitmap)
2175                 mddev->bitmap->behind_writes_used = 0;
2176         return len;
2177 }
2178
2179 static struct md_sysfs_entry max_backlog_used =
2180 __ATTR(max_backlog_used, S_IRUGO | S_IWUSR,
2181        behind_writes_used_show, behind_writes_used_reset);
2182
2183 static struct attribute *md_bitmap_attrs[] = {
2184         &bitmap_location.attr,
2185         &bitmap_timeout.attr,
2186         &bitmap_backlog.attr,
2187         &bitmap_chunksize.attr,
2188         &bitmap_metadata.attr,
2189         &bitmap_can_clear.attr,
2190         &max_backlog_used.attr,
2191         NULL
2192 };
2193 struct attribute_group md_bitmap_group = {
2194         .name = "bitmap",
2195         .attrs = md_bitmap_attrs,
2196 };
2197