git.oblomov.eu Git - linux-2.6/blob - 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  * wait if count gets too high, wake when it drops to half.
  17  */
  18
  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/raid/md.h>
  30 #include <linux/raid/bitmap.h>
  31
  32 /* debug macros */
  33
  34 #define DEBUG 0
  35
  36 #if DEBUG
  37 /* these are for debugging purposes only! */
  38
  39 /* define one and only one of these */
  40 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  41 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  42 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  43 #define INJECT_FAULTS_4 0 /* undef */
  44 #define INJECT_FAULTS_5 0 /* undef */
  45 #define INJECT_FAULTS_6 0
  46
  47 /* if these are defined, the driver will fail! debug only */
  48 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  49 #define INJECT_FATAL_FAULT_2 0 /* undef */
  50 #define INJECT_FATAL_FAULT_3 0 /* undef */
  51 #endif
  52
  53 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  54 #define DPRINTK(x...) do { } while(0)
  55
  56 #ifndef PRINTK
  57 #  if DEBUG > 0
  58 #    define PRINTK(x...) printk(KERN_DEBUG x)
  59 #  else
  60 #    define PRINTK(x...)
  61 #  endif
  62 #endif
  63
  64 static inline char * bmname(struct bitmap *bitmap)
  65 {
  66         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  67 }
  68
  69
  70 /*
  71  * just a placeholder - calls kmalloc for bitmap pages
  72  */
  73 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
  74 {
  75         unsigned char *page;
  76
  77 #ifdef INJECT_FAULTS_1
  78         page = NULL;
  79 #else
  80         page = kmalloc(PAGE_SIZE, GFP_NOIO);
  81 #endif
  82         if (!page)
  83                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
  84         else
  85                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
  86                         bmname(bitmap), page);
  87         return page;
  88 }
  89
  90 /*
  91  * for now just a placeholder -- just calls kfree for bitmap pages
  92  */
  93 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
  94 {
  95         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
  96         kfree(page);
  97 }
  98
  99 /*
 100  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 101  *
 102  * 1) check to see if this page is allocated, if it's not then try to alloc
 103  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 104  *    page pointer directly as a counter
 105  *
 106  * if we find our page, we increment the page's refcount so that it stays
 107  * allocated while we're using it
 108  */
 109 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 110 {
 111         unsigned char *mappage;
 112
 113         if (page >= bitmap->pages) {
 114                 printk(KERN_ALERT
 115                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 116                         bmname(bitmap), page, bitmap->pages-1);
 117                 return -EINVAL;
 118         }
 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         spin_unlock_irq(&bitmap->lock);
 131
 132         /* this page has not been allocated yet */
 133
 134         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 135                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 136                         bmname(bitmap));
 137                 /* failed - set the hijacked flag so that we can use the
 138                  * pointer as a counter */
 139                 spin_lock_irq(&bitmap->lock);
 140                 if (!bitmap->bp[page].map)
 141                         bitmap->bp[page].hijacked = 1;
 142                 goto out;
 143         }
 144
 145         /* got a page */
 146
 147         spin_lock_irq(&bitmap->lock);
 148
 149         /* recheck the page */
 150
 151         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 152                 /* somebody beat us to getting the page */
 153                 bitmap_free_page(bitmap, mappage);
 154                 return 0;
 155         }
 156
 157         /* no page was in place and we have one, so install it */
 158
 159         memset(mappage, 0, PAGE_SIZE);
 160         bitmap->bp[page].map = mappage;
 161         bitmap->missing_pages--;
 162 out:
 163         return 0;
 164 }
 165
 166
 167 /* if page is completely empty, put it back on the free list, or dealloc it */
 168 /* if page was hijacked, unmark the flag so it might get alloced next time */
 169 /* Note: lock should be held when calling this */
 170 static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 171 {
 172         char *ptr;
 173
 174         if (bitmap->bp[page].count) /* page is still busy */
 175                 return;
 176
 177         /* page is no longer in use, it can be released */
 178
 179         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 180                 bitmap->bp[page].hijacked = 0;
 181                 bitmap->bp[page].map = NULL;
 182                 return;
 183         }
 184
 185         /* normal case, free the page */
 186
 187 #if 0
 188 /* actually ... let's not.  We will probably need the page again exactly when
 189  * memory is tight and we are flusing to disk
 190  */
 191         return;
 192 #else
 193         ptr = bitmap->bp[page].map;
 194         bitmap->bp[page].map = NULL;
 195         bitmap->missing_pages++;
 196         bitmap_free_page(bitmap, ptr);
 197         return;
 198 #endif
 199 }
 200
 201
 202 /*
 203  * bitmap file handling - read and write the bitmap file and its superblock
 204  */
 205
 206 /*
 207  * basic page I/O operations
 208  */
 209
 210 /* IO operations when bitmap is stored near all superblocks */
 211 static struct page *read_sb_page(mddev_t *mddev, long offset,
 212                                  struct page *page,
 213                                  unsigned long index, int size)
 214 {
 215         /* choose a good rdev and read the page from there */
 216
 217         mdk_rdev_t *rdev;
 218         struct list_head *tmp;
 219         sector_t target;
 220
 221         if (!page)
 222                 page = alloc_page(GFP_KERNEL);
 223         if (!page)
 224                 return ERR_PTR(-ENOMEM);
 225
 226         rdev_for_each(rdev, tmp, mddev) {
 227                 if (! test_bit(In_sync, &rdev->flags)
 228                     || test_bit(Faulty, &rdev->flags))
 229                         continue;
 230
 231                 target = rdev->sb_start + offset + index * (PAGE_SIZE/512);
 232
 233                 if (sync_page_io(rdev->bdev, target,
 234                                  roundup(size, bdev_hardsect_size(rdev->bdev)),
 235                                  page, READ)) {
 236                         page->index = index;
 237                         attach_page_buffers(page, NULL); /* so that free_buffer will
 238                                                           * quietly no-op */
 239                         return page;
 240                 }
 241         }
 242         return ERR_PTR(-EIO);
 243
 244 }
 245
 246 static mdk_rdev_t *next_active_rdev(mdk_rdev_t *rdev, mddev_t *mddev)
 247 {
 248         /* Iterate the disks of an mddev, using rcu to protect access to the
 249          * linked list, and raising the refcount of devices we return to ensure
 250          * they don't disappear while in use.
 251          * As devices are only added or removed when raid_disk is < 0 and
 252          * nr_pending is 0 and In_sync is clear, the entries we return will
 253          * still be in the same position on the list when we re-enter
 254          * list_for_each_continue_rcu.
 255          */
 256         struct list_head *pos;
 257         rcu_read_lock();
 258         if (rdev == NULL)
 259                 /* start at the beginning */
 260                 pos = &mddev->disks;
 261         else {
 262                 /* release the previous rdev and start from there. */
 263                 rdev_dec_pending(rdev, mddev);
 264                 pos = &rdev->same_set;
 265         }
 266         list_for_each_continue_rcu(pos, &mddev->disks) {
 267                 rdev = list_entry(pos, mdk_rdev_t, same_set);
 268                 if (rdev->raid_disk >= 0 &&
 269                     test_bit(In_sync, &rdev->flags) &&
 270                     !test_bit(Faulty, &rdev->flags)) {
 271                         /* this is a usable devices */
 272                         atomic_inc(&rdev->nr_pending);
 273                         rcu_read_unlock();
 274                         return rdev;
 275                 }
 276         }
 277         rcu_read_unlock();
 278         return NULL;
 279 }
 280
 281 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 282 {
 283         mdk_rdev_t *rdev = NULL;
 284         mddev_t *mddev = bitmap->mddev;
 285
 286         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 287                         int size = PAGE_SIZE;
 288                         if (page->index == bitmap->file_pages-1)
 289                                 size = roundup(bitmap->last_page_size,
 290                                                bdev_hardsect_size(rdev->bdev));
 291                         /* Just make sure we aren't corrupting data or
 292                          * metadata
 293                          */
 294                         if (bitmap->offset < 0) {
 295                                 /* DATA  BITMAP METADATA  */
 296                                 if (bitmap->offset
 297                                     + (long)(page->index * (PAGE_SIZE/512))
 298                                     + size/512 > 0)
 299                                         /* bitmap runs in to metadata */
 300                                         goto bad_alignment;
 301                                 if (rdev->data_offset + mddev->size*2
 302                                     > rdev->sb_start + bitmap->offset)
 303                                         /* data runs in to bitmap */
 304                                         goto bad_alignment;
 305                         } else if (rdev->sb_start < rdev->data_offset) {
 306                                 /* METADATA BITMAP DATA */
 307                                 if (rdev->sb_start
 308                                     + bitmap->offset
 309                                     + page->index*(PAGE_SIZE/512) + size/512
 310                                     > rdev->data_offset)
 311                                         /* bitmap runs in to data */
 312                                         goto bad_alignment;
 313                         } else {
 314                                 /* DATA METADATA BITMAP - no problems */
 315                         }
 316                         md_super_write(mddev, rdev,
 317                                        rdev->sb_start + bitmap->offset
 318                                        + page->index * (PAGE_SIZE/512),
 319                                        size,
 320                                        page);
 321         }
 322
 323         if (wait)
 324                 md_super_wait(mddev);
 325         return 0;
 326
 327  bad_alignment:
 328         rcu_read_unlock();
 329         return -EINVAL;
 330 }
 331
 332 static void bitmap_file_kick(struct bitmap *bitmap);
 333 /*
 334  * write out a page to a file
 335  */
 336 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 337 {
 338         struct buffer_head *bh;
 339
 340         if (bitmap->file == NULL) {
 341                 switch (write_sb_page(bitmap, page, wait)) {
 342                 case -EINVAL:
 343                         bitmap->flags |= BITMAP_WRITE_ERROR;
 344                 }
 345         } else {
 346
 347                 bh = page_buffers(page);
 348
 349                 while (bh && bh->b_blocknr) {
 350                         atomic_inc(&bitmap->pending_writes);
 351                         set_buffer_locked(bh);
 352                         set_buffer_mapped(bh);
 353                         submit_bh(WRITE, bh);
 354                         bh = bh->b_this_page;
 355                 }
 356
 357                 if (wait) {
 358                         wait_event(bitmap->write_wait,
 359                                    atomic_read(&bitmap->pending_writes)==0);
 360                 }
 361         }
 362         if (bitmap->flags & BITMAP_WRITE_ERROR)
 363                 bitmap_file_kick(bitmap);
 364 }
 365
 366 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 367 {
 368         struct bitmap *bitmap = bh->b_private;
 369         unsigned long flags;
 370
 371         if (!uptodate) {
 372                 spin_lock_irqsave(&bitmap->lock, flags);
 373                 bitmap->flags |= BITMAP_WRITE_ERROR;
 374                 spin_unlock_irqrestore(&bitmap->lock, flags);
 375         }
 376         if (atomic_dec_and_test(&bitmap->pending_writes))
 377                 wake_up(&bitmap->write_wait);
 378 }
 379
 380 /* copied from buffer.c */
 381 static void
 382 __clear_page_buffers(struct page *page)
 383 {
 384         ClearPagePrivate(page);
 385         set_page_private(page, 0);
 386         page_cache_release(page);
 387 }
 388 static void free_buffers(struct page *page)
 389 {
 390         struct buffer_head *bh = page_buffers(page);
 391
 392         while (bh) {
 393                 struct buffer_head *next = bh->b_this_page;
 394                 free_buffer_head(bh);
 395                 bh = next;
 396         }
 397         __clear_page_buffers(page);
 398         put_page(page);
 399 }
 400
 401 /* read a page from a file.
 402  * We both read the page, and attach buffers to the page to record the
 403  * address of each block (using bmap).  These addresses will be used
 404  * to write the block later, completely bypassing the filesystem.
 405  * This usage is similar to how swap files are handled, and allows us
 406  * to write to a file with no concerns of memory allocation failing.
 407  */
 408 static struct page *read_page(struct file *file, unsigned long index,
 409                               struct bitmap *bitmap,
 410                               unsigned long count)
 411 {
 412         struct page *page = NULL;
 413         struct inode *inode = file->f_path.dentry->d_inode;
 414         struct buffer_head *bh;
 415         sector_t block;
 416
 417         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_SIZE,
 418                         (unsigned long long)index << PAGE_SHIFT);
 419
 420         page = alloc_page(GFP_KERNEL);
 421         if (!page)
 422                 page = ERR_PTR(-ENOMEM);
 423         if (IS_ERR(page))
 424                 goto out;
 425
 426         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 427         if (!bh) {
 428                 put_page(page);
 429                 page = ERR_PTR(-ENOMEM);
 430                 goto out;
 431         }
 432         attach_page_buffers(page, bh);
 433         block = index << (PAGE_SHIFT - inode->i_blkbits);
 434         while (bh) {
 435                 if (count == 0)
 436                         bh->b_blocknr = 0;
 437                 else {
 438                         bh->b_blocknr = bmap(inode, block);
 439                         if (bh->b_blocknr == 0) {
 440                                 /* Cannot use this file! */
 441                                 free_buffers(page);
 442                                 page = ERR_PTR(-EINVAL);
 443                                 goto out;
 444                         }
 445                         bh->b_bdev = inode->i_sb->s_bdev;
 446                         if (count < (1<<inode->i_blkbits))
 447                                 count = 0;
 448                         else
 449                                 count -= (1<<inode->i_blkbits);
 450
 451                         bh->b_end_io = end_bitmap_write;
 452                         bh->b_private = bitmap;
 453                         atomic_inc(&bitmap->pending_writes);
 454                         set_buffer_locked(bh);
 455                         set_buffer_mapped(bh);
 456                         submit_bh(READ, bh);
 457                 }
 458                 block++;
 459                 bh = bh->b_this_page;
 460         }
 461         page->index = index;
 462
 463         wait_event(bitmap->write_wait,
 464                    atomic_read(&bitmap->pending_writes)==0);
 465         if (bitmap->flags & BITMAP_WRITE_ERROR) {
 466                 free_buffers(page);
 467                 page = ERR_PTR(-EIO);
 468         }
 469 out:
 470         if (IS_ERR(page))
 471                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 472                         (int)PAGE_SIZE,
 473                         (unsigned long long)index << PAGE_SHIFT,
 474                         PTR_ERR(page));
 475         return page;
 476 }
 477
 478 /*
 479  * bitmap file superblock operations
 480  */
 481
 482 /* update the event counter and sync the superblock to disk */
 483 void bitmap_update_sb(struct bitmap *bitmap)
 484 {
 485         bitmap_super_t *sb;
 486         unsigned long flags;
 487
 488         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 489                 return;
 490         spin_lock_irqsave(&bitmap->lock, flags);
 491         if (!bitmap->sb_page) { /* no superblock */
 492                 spin_unlock_irqrestore(&bitmap->lock, flags);
 493                 return;
 494         }
 495         spin_unlock_irqrestore(&bitmap->lock, flags);
 496         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 497         sb->events = cpu_to_le64(bitmap->mddev->events);
 498         if (bitmap->mddev->events < bitmap->events_cleared) {
 499                 /* rocking back to read-only */
 500                 bitmap->events_cleared = bitmap->mddev->events;
 501                 sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 502         }
 503         kunmap_atomic(sb, KM_USER0);
 504         write_page(bitmap, bitmap->sb_page, 1);
 505 }
 506
 507 /* print out the bitmap file superblock */
 508 void bitmap_print_sb(struct bitmap *bitmap)
 509 {
 510         bitmap_super_t *sb;
 511
 512         if (!bitmap || !bitmap->sb_page)
 513                 return;
 514         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 515         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 516         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 517         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 518         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 519                                         *(__u32 *)(sb->uuid+0),
 520                                         *(__u32 *)(sb->uuid+4),
 521                                         *(__u32 *)(sb->uuid+8),
 522                                         *(__u32 *)(sb->uuid+12));
 523         printk(KERN_DEBUG "        events: %llu\n",
 524                         (unsigned long long) le64_to_cpu(sb->events));
 525         printk(KERN_DEBUG "events cleared: %llu\n",
 526                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 527         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 528         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 529         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 530         printk(KERN_DEBUG "     sync size: %llu KB\n",
 531                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 532         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 533         kunmap_atomic(sb, KM_USER0);
 534 }
 535
 536 /* read the superblock from the bitmap file and initialize some bitmap fields */
 537 static int bitmap_read_sb(struct bitmap *bitmap)
 538 {
 539         char *reason = NULL;
 540         bitmap_super_t *sb;
 541         unsigned long chunksize, daemon_sleep, write_behind;
 542         unsigned long long events;
 543         int err = -EINVAL;
 544
 545         /* page 0 is the superblock, read it... */
 546         if (bitmap->file) {
 547                 loff_t isize = i_size_read(bitmap->file->f_mapping->host);
 548                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 549
 550                 bitmap->sb_page = read_page(bitmap->file, 0, bitmap, bytes);
 551         } else {
 552                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset,
 553                                                NULL,
 554                                                0, sizeof(bitmap_super_t));
 555         }
 556         if (IS_ERR(bitmap->sb_page)) {
 557                 err = PTR_ERR(bitmap->sb_page);
 558                 bitmap->sb_page = NULL;
 559                 return err;
 560         }
 561
 562         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 563
 564         chunksize = le32_to_cpu(sb->chunksize);
 565         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 566         write_behind = le32_to_cpu(sb->write_behind);
 567
 568         /* verify that the bitmap-specific fields are valid */
 569         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 570                 reason = "bad magic";
 571         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 572                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 573                 reason = "unrecognized superblock version";
 574         else if (chunksize < PAGE_SIZE)
 575                 reason = "bitmap chunksize too small";
 576         else if ((1 << ffz(~chunksize)) != chunksize)
 577                 reason = "bitmap chunksize not a power of 2";
 578         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT / HZ)
 579                 reason = "daemon sleep period out of range";
 580         else if (write_behind > COUNTER_MAX)
 581                 reason = "write-behind limit out of range (0 - 16383)";
 582         if (reason) {
 583                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 584                         bmname(bitmap), reason);
 585                 goto out;
 586         }
 587
 588         /* keep the array size field of the bitmap superblock up to date */
 589         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 590
 591         if (!bitmap->mddev->persistent)
 592                 goto success;
 593
 594         /*
 595          * if we have a persistent array superblock, compare the
 596          * bitmap's UUID and event counter to the mddev's
 597          */
 598         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 599                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 600                         bmname(bitmap));
 601                 goto out;
 602         }
 603         events = le64_to_cpu(sb->events);
 604         if (events < bitmap->mddev->events) {
 605                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 606                         "-- forcing full recovery\n", bmname(bitmap), events,
 607                         (unsigned long long) bitmap->mddev->events);
 608                 sb->state |= cpu_to_le32(BITMAP_STALE);
 609         }
 610 success:
 611         /* assign fields using values from superblock */
 612         bitmap->chunksize = chunksize;
 613         bitmap->daemon_sleep = daemon_sleep;
 614         bitmap->daemon_lastrun = jiffies;
 615         bitmap->max_write_behind = write_behind;
 616         bitmap->flags |= le32_to_cpu(sb->state);
 617         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 618                 bitmap->flags |= BITMAP_HOSTENDIAN;
 619         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 620         if (sb->state & cpu_to_le32(BITMAP_STALE))
 621                 bitmap->events_cleared = bitmap->mddev->events;
 622         err = 0;
 623 out:
 624         kunmap_atomic(sb, KM_USER0);
 625         if (err)
 626                 bitmap_print_sb(bitmap);
 627         return err;
 628 }
 629
 630 enum bitmap_mask_op {
 631         MASK_SET,
 632         MASK_UNSET
 633 };
 634
 635 /* record the state of the bitmap in the superblock.  Return the old value */
 636 static int bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 637                              enum bitmap_mask_op op)
 638 {
 639         bitmap_super_t *sb;
 640         unsigned long flags;
 641         int old;
 642
 643         spin_lock_irqsave(&bitmap->lock, flags);
 644         if (!bitmap->sb_page) { /* can't set the state */
 645                 spin_unlock_irqrestore(&bitmap->lock, flags);
 646                 return 0;
 647         }
 648         spin_unlock_irqrestore(&bitmap->lock, flags);
 649         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 650         old = le32_to_cpu(sb->state) & bits;
 651         switch (op) {
 652                 case MASK_SET: sb->state |= cpu_to_le32(bits);
 653                                 break;
 654                 case MASK_UNSET: sb->state &= cpu_to_le32(~bits);
 655                                 break;
 656                 default: BUG();
 657         }
 658         kunmap_atomic(sb, KM_USER0);
 659         return old;
 660 }
 661
 662 /*
 663  * general bitmap file operations
 664  */
 665
 666 /* calculate the index of the page that contains this bit */
 667 static inline unsigned long file_page_index(unsigned long chunk)
 668 {
 669         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 670 }
 671
 672 /* calculate the (bit) offset of this bit within a page */
 673 static inline unsigned long file_page_offset(unsigned long chunk)
 674 {
 675         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 676 }
 677
 678 /*
 679  * return a pointer to the page in the filemap that contains the given bit
 680  *
 681  * this lookup is complicated by the fact that the bitmap sb might be exactly
 682  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 683  * 0 or page 1
 684  */
 685 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 686                                         unsigned long chunk)
 687 {
 688         if (file_page_index(chunk) >= bitmap->file_pages) return NULL;
 689         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 690 }
 691
 692
 693 static void bitmap_file_unmap(struct bitmap *bitmap)
 694 {
 695         struct page **map, *sb_page;
 696         unsigned long *attr;
 697         int pages;
 698         unsigned long flags;
 699
 700         spin_lock_irqsave(&bitmap->lock, flags);
 701         map = bitmap->filemap;
 702         bitmap->filemap = NULL;
 703         attr = bitmap->filemap_attr;
 704         bitmap->filemap_attr = NULL;
 705         pages = bitmap->file_pages;
 706         bitmap->file_pages = 0;
 707         sb_page = bitmap->sb_page;
 708         bitmap->sb_page = NULL;
 709         spin_unlock_irqrestore(&bitmap->lock, flags);
 710
 711         while (pages--)
 712                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 713                         free_buffers(map[pages]);
 714         kfree(map);
 715         kfree(attr);
 716
 717         if (sb_page)
 718                 free_buffers(sb_page);
 719 }
 720
 721 static void bitmap_file_put(struct bitmap *bitmap)
 722 {
 723         struct file *file;
 724         unsigned long flags;
 725
 726         spin_lock_irqsave(&bitmap->lock, flags);
 727         file = bitmap->file;
 728         bitmap->file = NULL;
 729         spin_unlock_irqrestore(&bitmap->lock, flags);
 730
 731         if (file)
 732                 wait_event(bitmap->write_wait,
 733                            atomic_read(&bitmap->pending_writes)==0);
 734         bitmap_file_unmap(bitmap);
 735
 736         if (file) {
 737                 struct inode *inode = file->f_path.dentry->d_inode;
 738                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 739                 fput(file);
 740         }
 741 }
 742
 743
 744 /*
 745  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 746  * then it is no longer reliable, so we stop using it and we mark the file
 747  * as failed in the superblock
 748  */
 749 static void bitmap_file_kick(struct bitmap *bitmap)
 750 {
 751         char *path, *ptr = NULL;
 752
 753         if (bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET) == 0) {
 754                 bitmap_update_sb(bitmap);
 755
 756                 if (bitmap->file) {
 757                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 758                         if (path)
 759                                 ptr = d_path(&bitmap->file->f_path, path,
 760                                              PAGE_SIZE);
 761
 762
 763                         printk(KERN_ALERT
 764                               "%s: kicking failed bitmap file %s from array!\n",
 765                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 766
 767                         kfree(path);
 768                 } else
 769                         printk(KERN_ALERT
 770                                "%s: disabling internal bitmap due to errors\n",
 771                                bmname(bitmap));
 772         }
 773
 774         bitmap_file_put(bitmap);
 775
 776         return;
 777 }
 778
 779 enum bitmap_page_attr {
 780         BITMAP_PAGE_DIRTY = 0, // there are set bits that need to be synced
 781         BITMAP_PAGE_CLEAN = 1, // there are bits that might need to be cleared
 782         BITMAP_PAGE_NEEDWRITE=2, // there are cleared bits that need to be synced
 783 };
 784
 785 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 786                                 enum bitmap_page_attr attr)
 787 {
 788         __set_bit((page->index<<2) + attr, bitmap->filemap_attr);
 789 }
 790
 791 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 792                                 enum bitmap_page_attr attr)
 793 {
 794         __clear_bit((page->index<<2) + attr, bitmap->filemap_attr);
 795 }
 796
 797 static inline unsigned long test_page_attr(struct bitmap *bitmap, struct page *page,
 798                                            enum bitmap_page_attr attr)
 799 {
 800         return test_bit((page->index<<2) + attr, bitmap->filemap_attr);
 801 }
 802
 803 /*
 804  * bitmap_file_set_bit -- called before performing a write to the md device
 805  * to set (and eventually sync) a particular bit in the bitmap file
 806  *
 807  * we set the bit immediately, then we record the page number so that
 808  * when an unplug occurs, we can flush the dirty pages out to disk
 809  */
 810 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 811 {
 812         unsigned long bit;
 813         struct page *page;
 814         void *kaddr;
 815         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 816
 817         if (!bitmap->filemap) {
 818                 return;
 819         }
 820
 821         page = filemap_get_page(bitmap, chunk);
 822         if (!page) return;
 823         bit = file_page_offset(chunk);
 824
 825         /* set the bit */
 826         kaddr = kmap_atomic(page, KM_USER0);
 827         if (bitmap->flags & BITMAP_HOSTENDIAN)
 828                 set_bit(bit, kaddr);
 829         else
 830                 ext2_set_bit(bit, kaddr);
 831         kunmap_atomic(kaddr, KM_USER0);
 832         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 833
 834         /* record page number so it gets flushed to disk when unplug occurs */
 835         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 836
 837 }
 838
 839 /* this gets called when the md device is ready to unplug its underlying
 840  * (slave) device queues -- before we let any writes go down, we need to
 841  * sync the dirty pages of the bitmap file to disk */
 842 void bitmap_unplug(struct bitmap *bitmap)
 843 {
 844         unsigned long i, flags;
 845         int dirty, need_write;
 846         struct page *page;
 847         int wait = 0;
 848
 849         if (!bitmap)
 850                 return;
 851
 852         /* look at each page to see if there are any set bits that need to be
 853          * flushed out to disk */
 854         for (i = 0; i < bitmap->file_pages; i++) {
 855                 spin_lock_irqsave(&bitmap->lock, flags);
 856                 if (!bitmap->filemap) {
 857                         spin_unlock_irqrestore(&bitmap->lock, flags);
 858                         return;
 859                 }
 860                 page = bitmap->filemap[i];
 861                 dirty = test_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 862                 need_write = test_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 863                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 864                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 865                 if (dirty)
 866                         wait = 1;
 867                 spin_unlock_irqrestore(&bitmap->lock, flags);
 868
 869                 if (dirty | need_write)
 870                         write_page(bitmap, page, 0);
 871         }
 872         if (wait) { /* if any writes were performed, we need to wait on them */
 873                 if (bitmap->file)
 874                         wait_event(bitmap->write_wait,
 875                                    atomic_read(&bitmap->pending_writes)==0);
 876                 else
 877                         md_super_wait(bitmap->mddev);
 878         }
 879         if (bitmap->flags & BITMAP_WRITE_ERROR)
 880                 bitmap_file_kick(bitmap);
 881 }
 882
 883 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 884 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 885  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 886  * memory mapping of the bitmap file
 887  * Special cases:
 888  *   if there's no bitmap file, or if the bitmap file had been
 889  *   previously kicked from the array, we mark all the bits as
 890  *   1's in order to cause a full resync.
 891  *
 892  * We ignore all bits for sectors that end earlier than 'start'.
 893  * This is used when reading an out-of-date bitmap...
 894  */
 895 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 896 {
 897         unsigned long i, chunks, index, oldindex, bit;
 898         struct page *page = NULL, *oldpage = NULL;
 899         unsigned long num_pages, bit_cnt = 0;
 900         struct file *file;
 901         unsigned long bytes, offset;
 902         int outofdate;
 903         int ret = -ENOSPC;
 904         void *paddr;
 905
 906         chunks = bitmap->chunks;
 907         file = bitmap->file;
 908
 909         BUG_ON(!file && !bitmap->offset);
 910
 911 #ifdef INJECT_FAULTS_3
 912         outofdate = 1;
 913 #else
 914         outofdate = bitmap->flags & BITMAP_STALE;
 915 #endif
 916         if (outofdate)
 917                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 918                         "recovery\n", bmname(bitmap));
 919
 920         bytes = (chunks + 7) / 8;
 921
 922         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 923
 924         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 925                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 926                         bmname(bitmap),
 927                         (unsigned long) i_size_read(file->f_mapping->host),
 928                         bytes + sizeof(bitmap_super_t));
 929                 goto err;
 930         }
 931
 932         ret = -ENOMEM;
 933
 934         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 935         if (!bitmap->filemap)
 936                 goto err;
 937
 938         /* We need 4 bits per page, rounded up to a multiple of sizeof(unsigned long) */
 939         bitmap->filemap_attr = kzalloc(
 940                 roundup( DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
 941                 GFP_KERNEL);
 942         if (!bitmap->filemap_attr)
 943                 goto err;
 944
 945         oldindex = ~0L;
 946
 947         for (i = 0; i < chunks; i++) {
 948                 int b;
 949                 index = file_page_index(i);
 950                 bit = file_page_offset(i);
 951                 if (index != oldindex) { /* this is a new page, read it in */
 952                         int count;
 953                         /* unmap the old page, we're done with it */
 954                         if (index == num_pages-1)
 955                                 count = bytes + sizeof(bitmap_super_t)
 956                                         - index * PAGE_SIZE;
 957                         else
 958                                 count = PAGE_SIZE;
 959                         if (index == 0) {
 960                                 /*
 961                                  * if we're here then the superblock page
 962                                  * contains some bits (PAGE_SIZE != sizeof sb)
 963                                  * we've already read it in, so just use it
 964                                  */
 965                                 page = bitmap->sb_page;
 966                                 offset = sizeof(bitmap_super_t);
 967                                 read_sb_page(bitmap->mddev, bitmap->offset,
 968                                              page,
 969                                              index, count);
 970                         } else if (file) {
 971                                 page = read_page(file, index, bitmap, count);
 972                                 offset = 0;
 973                         } else {
 974                                 page = read_sb_page(bitmap->mddev, bitmap->offset,
 975                                                     NULL,
 976                                                     index, count);
 977                                 offset = 0;
 978                         }
 979                         if (IS_ERR(page)) { /* read error */
 980                                 ret = PTR_ERR(page);
 981                                 goto err;
 982                         }
 983
 984                         oldindex = index;
 985                         oldpage = page;
 986
 987                         if (outofdate) {
 988                                 /*
 989                                  * if bitmap is out of date, dirty the
 990                                  * whole page and write it out
 991                                  */
 992                                 paddr = kmap_atomic(page, KM_USER0);
 993                                 memset(paddr + offset, 0xff,
 994                                        PAGE_SIZE - offset);
 995                                 kunmap_atomic(paddr, KM_USER0);
 996                                 write_page(bitmap, page, 1);
 997
 998                                 ret = -EIO;
 999                                 if (bitmap->flags & BITMAP_WRITE_ERROR) {
1000                                         /* release, page not in filemap yet */
1001                                         put_page(page);
1002                                         goto err;
1003                                 }
1004                         }
1005
1006                         bitmap->filemap[bitmap->file_pages++] = page;
1007                         bitmap->last_page_size = count;
1008                 }
1009                 paddr = kmap_atomic(page, KM_USER0);
1010                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1011                         b = test_bit(bit, paddr);
1012                 else
1013                         b = ext2_test_bit(bit, paddr);
1014                 kunmap_atomic(paddr, KM_USER0);
1015                 if (b) {
1016                         /* if the disk bit is set, set the memory bit */
1017                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
1018                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
1019                                 );
1020                         bit_cnt++;
1021                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1022                 }
1023         }
1024
1025         /* everything went OK */
1026         ret = 0;
1027         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
1028
1029         if (bit_cnt) { /* Kick recovery if any bits were set */
1030                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1031                 md_wakeup_thread(bitmap->mddev->thread);
1032         }
1033
1034         printk(KERN_INFO "%s: bitmap initialized from disk: "
1035                 "read %lu/%lu pages, set %lu bits\n",
1036                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt);
1037
1038         return 0;
1039
1040  err:
1041         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1042                bmname(bitmap), ret);
1043         return ret;
1044 }
1045
1046 void bitmap_write_all(struct bitmap *bitmap)
1047 {
1048         /* We don't actually write all bitmap blocks here,
1049          * just flag them as needing to be written
1050          */
1051         int i;
1052
1053         for (i=0; i < bitmap->file_pages; i++)
1054                 set_page_attr(bitmap, bitmap->filemap[i],
1055                               BITMAP_PAGE_NEEDWRITE);
1056 }
1057
1058
1059 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
1060 {
1061         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1062         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1063         bitmap->bp[page].count += inc;
1064 /*
1065         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
1066                               (unsigned long long)offset, inc, bitmap->bp[page].count);
1067 */
1068         bitmap_checkfree(bitmap, page);
1069 }
1070 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1071                                             sector_t offset, int *blocks,
1072                                             int create);
1073
1074 /*
1075  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1076  *                      out to disk
1077  */
1078
1079 void bitmap_daemon_work(struct bitmap *bitmap)
1080 {
1081         unsigned long j;
1082         unsigned long flags;
1083         struct page *page = NULL, *lastpage = NULL;
1084         int blocks;
1085         void *paddr;
1086
1087         if (bitmap == NULL)
1088                 return;
1089         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1090                 goto done;
1091
1092         bitmap->daemon_lastrun = jiffies;
1093         if (bitmap->allclean) {
1094                 bitmap->mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1095                 return;
1096         }
1097         bitmap->allclean = 1;
1098
1099         for (j = 0; j < bitmap->chunks; j++) {
1100                 bitmap_counter_t *bmc;
1101                 spin_lock_irqsave(&bitmap->lock, flags);
1102                 if (!bitmap->filemap) {
1103                         /* error or shutdown */
1104                         spin_unlock_irqrestore(&bitmap->lock, flags);
1105                         break;
1106                 }
1107
1108                 page = filemap_get_page(bitmap, j);
1109
1110                 if (page != lastpage) {
1111                         /* skip this page unless it's marked as needing cleaning */
1112                         if (!test_page_attr(bitmap, page, BITMAP_PAGE_CLEAN)) {
1113                                 int need_write = test_page_attr(bitmap, page,
1114                                                                 BITMAP_PAGE_NEEDWRITE);
1115                                 if (need_write)
1116                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1117
1118                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1119                                 if (need_write) {
1120                                         write_page(bitmap, page, 0);
1121                                         bitmap->allclean = 0;
1122                                 }
1123                                 continue;
1124                         }
1125
1126                         /* grab the new page, sync and release the old */
1127                         if (lastpage != NULL) {
1128                                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1129                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1130                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1131                                         write_page(bitmap, lastpage, 0);
1132                                 } else {
1133                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1134                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1135                                 }
1136                         } else
1137                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1138                         lastpage = page;
1139
1140                         /* We are possibly going to clear some bits, so make
1141                          * sure that events_cleared is up-to-date.
1142                          */
1143                         if (bitmap->need_sync) {
1144                                 bitmap_super_t *sb;
1145                                 bitmap->need_sync = 0;
1146                                 sb = kmap_atomic(bitmap->sb_page, KM_USER0);
1147                                 sb->events_cleared =
1148                                         cpu_to_le64(bitmap->events_cleared);
1149                                 kunmap_atomic(sb, KM_USER0);
1150                                 write_page(bitmap, bitmap->sb_page, 1);
1151                         }
1152                         spin_lock_irqsave(&bitmap->lock, flags);
1153                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1154                 }
1155                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1156                                         &blocks, 0);
1157                 if (bmc) {
1158 /*
1159   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1160 */
1161                         if (*bmc)
1162                                 bitmap->allclean = 0;
1163
1164                         if (*bmc == 2) {
1165                                 *bmc=1; /* maybe clear the bit next time */
1166                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1167                         } else if (*bmc == 1) {
1168                                 /* we can clear the bit */
1169                                 *bmc = 0;
1170                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1171                                                   -1);
1172
1173                                 /* clear the bit */
1174                                 paddr = kmap_atomic(page, KM_USER0);
1175                                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1176                                         clear_bit(file_page_offset(j), paddr);
1177                                 else
1178                                         ext2_clear_bit(file_page_offset(j), paddr);
1179                                 kunmap_atomic(paddr, KM_USER0);
1180                         }
1181                 }
1182                 spin_unlock_irqrestore(&bitmap->lock, flags);
1183         }
1184
1185         /* now sync the final page */
1186         if (lastpage != NULL) {
1187                 spin_lock_irqsave(&bitmap->lock, flags);
1188                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1189                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1190                         spin_unlock_irqrestore(&bitmap->lock, flags);
1191                         write_page(bitmap, lastpage, 0);
1192                 } else {
1193                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1194                         spin_unlock_irqrestore(&bitmap->lock, flags);
1195                 }
1196         }
1197
1198  done:
1199         if (bitmap->allclean == 0)
1200                 bitmap->mddev->thread->timeout = bitmap->daemon_sleep * HZ;
1201 }
1202
1203 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1204                                             sector_t offset, int *blocks,
1205                                             int create)
1206 {
1207         /* If 'create', we might release the lock and reclaim it.
1208          * The lock must have been taken with interrupts enabled.
1209          * If !create, we don't release the lock.
1210          */
1211         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1212         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1213         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1214         sector_t csize;
1215
1216         if (bitmap_checkpage(bitmap, page, create) < 0) {
1217                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1218                 *blocks = csize - (offset & (csize- 1));
1219                 return NULL;
1220         }
1221         /* now locked ... */
1222
1223         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1224                 /* should we use the first or second counter field
1225                  * of the hijacked pointer? */
1226                 int hi = (pageoff > PAGE_COUNTER_MASK);
1227                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1228                                           PAGE_COUNTER_SHIFT - 1);
1229                 *blocks = csize - (offset & (csize- 1));
1230                 return  &((bitmap_counter_t *)
1231                           &bitmap->bp[page].map)[hi];
1232         } else { /* page is allocated */
1233                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1234                 *blocks = csize - (offset & (csize- 1));
1235                 return (bitmap_counter_t *)
1236                         &(bitmap->bp[page].map[pageoff]);
1237         }
1238 }
1239
1240 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1241 {
1242         if (!bitmap) return 0;
1243
1244         if (behind) {
1245                 atomic_inc(&bitmap->behind_writes);
1246                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1247                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1248         }
1249
1250         while (sectors) {
1251                 int blocks;
1252                 bitmap_counter_t *bmc;
1253
1254                 spin_lock_irq(&bitmap->lock);
1255                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1256                 if (!bmc) {
1257                         spin_unlock_irq(&bitmap->lock);
1258                         return 0;
1259                 }
1260
1261                 if (unlikely((*bmc & COUNTER_MAX) == COUNTER_MAX)) {
1262                         DEFINE_WAIT(__wait);
1263                         /* note that it is safe to do the prepare_to_wait
1264                          * after the test as long as we do it before dropping
1265                          * the spinlock.
1266                          */
1267                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1268                                         TASK_UNINTERRUPTIBLE);
1269                         spin_unlock_irq(&bitmap->lock);
1270                         blk_unplug(bitmap->mddev->queue);
1271                         schedule();
1272                         finish_wait(&bitmap->overflow_wait, &__wait);
1273                         continue;
1274                 }
1275
1276                 switch(*bmc) {
1277                 case 0:
1278                         bitmap_file_set_bit(bitmap, offset);
1279                         bitmap_count_page(bitmap,offset, 1);
1280                         blk_plug_device_unlocked(bitmap->mddev->queue);
1281                         /* fall through */
1282                 case 1:
1283                         *bmc = 2;
1284                 }
1285
1286                 (*bmc)++;
1287
1288                 spin_unlock_irq(&bitmap->lock);
1289
1290                 offset += blocks;
1291                 if (sectors > blocks)
1292                         sectors -= blocks;
1293                 else sectors = 0;
1294         }
1295         bitmap->allclean = 0;
1296         return 0;
1297 }
1298
1299 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1300                      int success, int behind)
1301 {
1302         if (!bitmap) return;
1303         if (behind) {
1304                 atomic_dec(&bitmap->behind_writes);
1305                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1306                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1307         }
1308
1309         while (sectors) {
1310                 int blocks;
1311                 unsigned long flags;
1312                 bitmap_counter_t *bmc;
1313
1314                 spin_lock_irqsave(&bitmap->lock, flags);
1315                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1316                 if (!bmc) {
1317                         spin_unlock_irqrestore(&bitmap->lock, flags);
1318                         return;
1319                 }
1320
1321                 if (success &&
1322                     bitmap->events_cleared < bitmap->mddev->events) {
1323                         bitmap->events_cleared = bitmap->mddev->events;
1324                         bitmap->need_sync = 1;
1325                 }
1326
1327                 if (!success && ! (*bmc & NEEDED_MASK))
1328                         *bmc |= NEEDED_MASK;
1329
1330                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX)
1331                         wake_up(&bitmap->overflow_wait);
1332
1333                 (*bmc)--;
1334                 if (*bmc <= 2) {
1335                         set_page_attr(bitmap,
1336                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1337                                       BITMAP_PAGE_CLEAN);
1338                 }
1339                 spin_unlock_irqrestore(&bitmap->lock, flags);
1340                 offset += blocks;
1341                 if (sectors > blocks)
1342                         sectors -= blocks;
1343                 else sectors = 0;
1344         }
1345 }
1346
1347 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1348                         int degraded)
1349 {
1350         bitmap_counter_t *bmc;
1351         int rv;
1352         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1353                 *blocks = 1024;
1354                 return 1; /* always resync if no bitmap */
1355         }
1356         spin_lock_irq(&bitmap->lock);
1357         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1358         rv = 0;
1359         if (bmc) {
1360                 /* locked */
1361                 if (RESYNC(*bmc))
1362                         rv = 1;
1363                 else if (NEEDED(*bmc)) {
1364                         rv = 1;
1365                         if (!degraded) { /* don't set/clear bits if degraded */
1366                                 *bmc |= RESYNC_MASK;
1367                                 *bmc &= ~NEEDED_MASK;
1368                         }
1369                 }
1370         }
1371         spin_unlock_irq(&bitmap->lock);
1372         bitmap->allclean = 0;
1373         return rv;
1374 }
1375
1376 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1377 {
1378         bitmap_counter_t *bmc;
1379         unsigned long flags;
1380 /*
1381         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1382 */      if (bitmap == NULL) {
1383                 *blocks = 1024;
1384                 return;
1385         }
1386         spin_lock_irqsave(&bitmap->lock, flags);
1387         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1388         if (bmc == NULL)
1389                 goto unlock;
1390         /* locked */
1391 /*
1392         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1393 */
1394         if (RESYNC(*bmc)) {
1395                 *bmc &= ~RESYNC_MASK;
1396
1397                 if (!NEEDED(*bmc) && aborted)
1398                         *bmc |= NEEDED_MASK;
1399                 else {
1400                         if (*bmc <= 2) {
1401                                 set_page_attr(bitmap,
1402                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1403                                               BITMAP_PAGE_CLEAN);
1404                         }
1405                 }
1406         }
1407  unlock:
1408         spin_unlock_irqrestore(&bitmap->lock, flags);
1409         bitmap->allclean = 0;
1410 }
1411
1412 void bitmap_close_sync(struct bitmap *bitmap)
1413 {
1414         /* Sync has finished, and any bitmap chunks that weren't synced
1415          * properly have been aborted.  It remains to us to clear the
1416          * RESYNC bit wherever it is still on
1417          */
1418         sector_t sector = 0;
1419         int blocks;
1420         if (!bitmap)
1421                 return;
1422         while (sector < bitmap->mddev->resync_max_sectors) {
1423                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1424                 sector += blocks;
1425         }
1426 }
1427
1428 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1429 {
1430         sector_t s = 0;
1431         int blocks;
1432
1433         if (!bitmap)
1434                 return;
1435         if (sector == 0) {
1436                 bitmap->last_end_sync = jiffies;
1437                 return;
1438         }
1439         if (time_before(jiffies, (bitmap->last_end_sync
1440                                   + bitmap->daemon_sleep * HZ)))
1441                 return;
1442         wait_event(bitmap->mddev->recovery_wait,
1443                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1444
1445         sector &= ~((1ULL << CHUNK_BLOCK_SHIFT(bitmap)) - 1);
1446         s = 0;
1447         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1448                 bitmap_end_sync(bitmap, s, &blocks, 0);
1449                 s += blocks;
1450         }
1451         bitmap->last_end_sync = jiffies;
1452 }
1453
1454 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1455 {
1456         /* For each chunk covered by any of these sectors, set the
1457          * counter to 1 and set resync_needed.  They should all
1458          * be 0 at this point
1459          */
1460
1461         int secs;
1462         bitmap_counter_t *bmc;
1463         spin_lock_irq(&bitmap->lock);
1464         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1465         if (!bmc) {
1466                 spin_unlock_irq(&bitmap->lock);
1467                 return;
1468         }
1469         if (! *bmc) {
1470                 struct page *page;
1471                 *bmc = 1 | (needed?NEEDED_MASK:0);
1472                 bitmap_count_page(bitmap, offset, 1);
1473                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1474                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1475         }
1476         spin_unlock_irq(&bitmap->lock);
1477         bitmap->allclean = 0;
1478 }
1479
1480 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1481 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1482 {
1483         unsigned long chunk;
1484
1485         for (chunk = s; chunk <= e; chunk++) {
1486                 sector_t sec = chunk << CHUNK_BLOCK_SHIFT(bitmap);
1487                 bitmap_set_memory_bits(bitmap, sec, 1);
1488                 bitmap_file_set_bit(bitmap, sec);
1489         }
1490 }
1491
1492 /*
1493  * flush out any pending updates
1494  */
1495 void bitmap_flush(mddev_t *mddev)
1496 {
1497         struct bitmap *bitmap = mddev->bitmap;
1498         int sleep;
1499
1500         if (!bitmap) /* there was no bitmap */
1501                 return;
1502
1503         /* run the daemon_work three time to ensure everything is flushed
1504          * that can be
1505          */
1506         sleep = bitmap->daemon_sleep;
1507         bitmap->daemon_sleep = 0;
1508         bitmap_daemon_work(bitmap);
1509         bitmap_daemon_work(bitmap);
1510         bitmap_daemon_work(bitmap);
1511         bitmap->daemon_sleep = sleep;
1512         bitmap_update_sb(bitmap);
1513 }
1514
1515 /*
1516  * free memory that was allocated
1517  */
1518 static void bitmap_free(struct bitmap *bitmap)
1519 {
1520         unsigned long k, pages;
1521         struct bitmap_page *bp;
1522
1523         if (!bitmap) /* there was no bitmap */
1524                 return;
1525
1526         /* release the bitmap file and kill the daemon */
1527         bitmap_file_put(bitmap);
1528
1529         bp = bitmap->bp;
1530         pages = bitmap->pages;
1531
1532         /* free all allocated memory */
1533
1534         if (bp) /* deallocate the page memory */
1535                 for (k = 0; k < pages; k++)
1536                         if (bp[k].map && !bp[k].hijacked)
1537                                 kfree(bp[k].map);
1538         kfree(bp);
1539         kfree(bitmap);
1540 }
1541 void bitmap_destroy(mddev_t *mddev)
1542 {
1543         struct bitmap *bitmap = mddev->bitmap;
1544
1545         if (!bitmap) /* there was no bitmap */
1546                 return;
1547
1548         mddev->bitmap = NULL; /* disconnect from the md device */
1549         if (mddev->thread)
1550                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1551
1552         bitmap_free(bitmap);
1553 }
1554
1555 /*
1556  * initialize the bitmap structure
1557  * if this returns an error, bitmap_destroy must be called to do clean up
1558  */
1559 int bitmap_create(mddev_t *mddev)
1560 {
1561         struct bitmap *bitmap;
1562         unsigned long blocks = mddev->resync_max_sectors;
1563         unsigned long chunks;
1564         unsigned long pages;
1565         struct file *file = mddev->bitmap_file;
1566         int err;
1567         sector_t start;
1568
1569         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1570
1571         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1572                 return 0;
1573
1574         BUG_ON(file && mddev->bitmap_offset);
1575
1576         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1577         if (!bitmap)
1578                 return -ENOMEM;
1579
1580         spin_lock_init(&bitmap->lock);
1581         atomic_set(&bitmap->pending_writes, 0);
1582         init_waitqueue_head(&bitmap->write_wait);
1583         init_waitqueue_head(&bitmap->overflow_wait);
1584
1585         bitmap->mddev = mddev;
1586
1587         bitmap->file = file;
1588         bitmap->offset = mddev->bitmap_offset;
1589         if (file) {
1590                 get_file(file);
1591                 do_sync_mapping_range(file->f_mapping, 0, LLONG_MAX,
1592                                       SYNC_FILE_RANGE_WAIT_BEFORE |
1593                                       SYNC_FILE_RANGE_WRITE |
1594                                       SYNC_FILE_RANGE_WAIT_AFTER);
1595         }
1596         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1597         err = bitmap_read_sb(bitmap);
1598         if (err)
1599                 goto error;
1600
1601         bitmap->chunkshift = ffz(~bitmap->chunksize);
1602
1603         /* now that chunksize and chunkshift are set, we can use these macros */
1604         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1605                         CHUNK_BLOCK_RATIO(bitmap);
1606         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1607
1608         BUG_ON(!pages);
1609
1610         bitmap->chunks = chunks;
1611         bitmap->pages = pages;
1612         bitmap->missing_pages = pages;
1613         bitmap->counter_bits = COUNTER_BITS;
1614
1615         bitmap->syncchunk = ~0UL;
1616
1617 #ifdef INJECT_FATAL_FAULT_1
1618         bitmap->bp = NULL;
1619 #else
1620         bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1621 #endif
1622         err = -ENOMEM;
1623         if (!bitmap->bp)
1624                 goto error;
1625
1626         /* now that we have some pages available, initialize the in-memory
1627          * bitmap from the on-disk bitmap */
1628         start = 0;
1629         if (mddev->degraded == 0
1630             || bitmap->events_cleared == mddev->events)
1631                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1632                 start = mddev->recovery_cp;
1633         err = bitmap_init_from_disk(bitmap, start);
1634
1635         if (err)
1636                 goto error;
1637
1638         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1639                 pages, bmname(bitmap));
1640
1641         mddev->bitmap = bitmap;
1642
1643         mddev->thread->timeout = bitmap->daemon_sleep * HZ;
1644
1645         bitmap_update_sb(bitmap);
1646
1647         return (bitmap->flags & BITMAP_WRITE_ERROR) ? -EIO : 0;
1648
1649  error:
1650         bitmap_free(bitmap);
1651         return err;
1652 }
1653
1654 /* the bitmap API -- for raid personalities */
1655 EXPORT_SYMBOL(bitmap_startwrite);
1656 EXPORT_SYMBOL(bitmap_endwrite);
1657 EXPORT_SYMBOL(bitmap_start_sync);
1658 EXPORT_SYMBOL(bitmap_end_sync);
1659 EXPORT_SYMBOL(bitmap_unplug);
1660 EXPORT_SYMBOL(bitmap_close_sync);
1661 EXPORT_SYMBOL(bitmap_cond_end_sync);