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  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/version.h>
  25 #include <linux/errno.h>
  26 #include <linux/slab.h>
  27 #include <linux/init.h>
  28 #include <linux/config.h>
  29 #include <linux/timer.h>
  30 #include <linux/sched.h>
  31 #include <linux/list.h>
  32 #include <linux/file.h>
  33 #include <linux/mount.h>
  34 #include <linux/buffer_head.h>
  35 #include <linux/raid/md.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 /* debug macros */
  39
  40 #define DEBUG 0
  41
  42 #if DEBUG
  43 /* these are for debugging purposes only! */
  44
  45 /* define one and only one of these */
  46 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  47 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  48 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  49 #define INJECT_FAULTS_4 0 /* undef */
  50 #define INJECT_FAULTS_5 0 /* undef */
  51 #define INJECT_FAULTS_6 0
  52
  53 /* if these are defined, the driver will fail! debug only */
  54 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  55 #define INJECT_FATAL_FAULT_2 0 /* undef */
  56 #define INJECT_FATAL_FAULT_3 0 /* undef */
  57 #endif
  58
  59 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  60 #define DPRINTK(x...) do { } while(0)
  61
  62 #ifndef PRINTK
  63 #  if DEBUG > 0
  64 #    define PRINTK(x...) printk(KERN_DEBUG x)
  65 #  else
  66 #    define PRINTK(x...)
  67 #  endif
  68 #endif
  69
  70 static inline char * bmname(struct bitmap *bitmap)
  71 {
  72         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  73 }
  74
  75
  76 /*
  77  * test if the bitmap is active
  78  */
  79 int bitmap_active(struct bitmap *bitmap)
  80 {
  81         unsigned long flags;
  82         int res = 0;
  83
  84         if (!bitmap)
  85                 return res;
  86         spin_lock_irqsave(&bitmap->lock, flags);
  87         res = bitmap->flags & BITMAP_ACTIVE;
  88         spin_unlock_irqrestore(&bitmap->lock, flags);
  89         return res;
  90 }
  91
  92 #define WRITE_POOL_SIZE 256
  93 /* mempool for queueing pending writes on the bitmap file */
  94 static void *write_pool_alloc(unsigned int gfp_flags, void *data)
  95 {
  96         return kmalloc(sizeof(struct page_list), gfp_flags);
  97 }
  98
  99 static void write_pool_free(void *ptr, void *data)
 100 {
 101         kfree(ptr);
 102 }
 103
 104 /*
 105  * just a placeholder - calls kmalloc for bitmap pages
 106  */
 107 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 108 {
 109         unsigned char *page;
 110
 111 #ifdef INJECT_FAULTS_1
 112         page = NULL;
 113 #else
 114         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 115 #endif
 116         if (!page)
 117                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 118         else
 119                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 120                         bmname(bitmap), page);
 121         return page;
 122 }
 123
 124 /*
 125  * for now just a placeholder -- just calls kfree for bitmap pages
 126  */
 127 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 128 {
 129         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 130         kfree(page);
 131 }
 132
 133 /*
 134  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 135  *
 136  * 1) check to see if this page is allocated, if it's not then try to alloc
 137  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 138  *    page pointer directly as a counter
 139  *
 140  * if we find our page, we increment the page's refcount so that it stays
 141  * allocated while we're using it
 142  */
 143 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 144 {
 145         unsigned char *mappage;
 146
 147         if (page >= bitmap->pages) {
 148                 printk(KERN_ALERT
 149                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 150                         bmname(bitmap), page, bitmap->pages-1);
 151                 return -EINVAL;
 152         }
 153
 154
 155         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 156                 return 0;
 157
 158         if (bitmap->bp[page].map) /* page is already allocated, just return */
 159                 return 0;
 160
 161         if (!create)
 162                 return -ENOENT;
 163
 164         spin_unlock_irq(&bitmap->lock);
 165
 166         /* this page has not been allocated yet */
 167
 168         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 169                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 170                         bmname(bitmap));
 171                 /* failed - set the hijacked flag so that we can use the
 172                  * pointer as a counter */
 173                 spin_lock_irq(&bitmap->lock);
 174                 if (!bitmap->bp[page].map)
 175                         bitmap->bp[page].hijacked = 1;
 176                 goto out;
 177         }
 178
 179         /* got a page */
 180
 181         spin_lock_irq(&bitmap->lock);
 182
 183         /* recheck the page */
 184
 185         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 186                 /* somebody beat us to getting the page */
 187                 bitmap_free_page(bitmap, mappage);
 188                 return 0;
 189         }
 190
 191         /* no page was in place and we have one, so install it */
 192
 193         memset(mappage, 0, PAGE_SIZE);
 194         bitmap->bp[page].map = mappage;
 195         bitmap->missing_pages--;
 196 out:
 197         return 0;
 198 }
 199
 200
 201 /* if page is completely empty, put it back on the free list, or dealloc it */
 202 /* if page was hijacked, unmark the flag so it might get alloced next time */
 203 /* Note: lock should be held when calling this */
 204 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 205 {
 206         char *ptr;
 207
 208         if (bitmap->bp[page].count) /* page is still busy */
 209                 return;
 210
 211         /* page is no longer in use, it can be released */
 212
 213         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 214                 bitmap->bp[page].hijacked = 0;
 215                 bitmap->bp[page].map = NULL;
 216                 return;
 217         }
 218
 219         /* normal case, free the page */
 220
 221 #if 0
 222 /* actually ... let's not.  We will probably need the page again exactly when
 223  * memory is tight and we are flusing to disk
 224  */
 225         return;
 226 #else
 227         ptr = bitmap->bp[page].map;
 228         bitmap->bp[page].map = NULL;
 229         bitmap->missing_pages++;
 230         bitmap_free_page(bitmap, ptr);
 231         return;
 232 #endif
 233 }
 234
 235
 236 /*
 237  * bitmap file handling - read and write the bitmap file and its superblock
 238  */
 239
 240 /* copy the pathname of a file to a buffer */
 241 char *file_path(struct file *file, char *buf, int count)
 242 {
 243         struct dentry *d;
 244         struct vfsmount *v;
 245
 246         if (!buf)
 247                 return NULL;
 248
 249         d = file->f_dentry;
 250         v = file->f_vfsmnt;
 251
 252         buf = d_path(d, v, buf, count);
 253
 254         return IS_ERR(buf) ? NULL : buf;
 255 }
 256
 257 /*
 258  * basic page I/O operations
 259  */
 260
 261 /* IO operations when bitmap is stored near all superblocks */
 262 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 263 {
 264         /* choose a good rdev and read the page from there */
 265
 266         mdk_rdev_t *rdev;
 267         struct list_head *tmp;
 268         struct page *page = alloc_page(GFP_KERNEL);
 269         sector_t target;
 270
 271         if (!page)
 272                 return ERR_PTR(-ENOMEM);
 273         do {
 274                 ITERATE_RDEV(mddev, rdev, tmp)
 275                         if (rdev->in_sync && !rdev->faulty)
 276                                 goto found;
 277                 return ERR_PTR(-EIO);
 278
 279         found:
 280                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 281
 282         } while (!sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ));
 283
 284         page->index = index;
 285         return page;
 286 }
 287
 288 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 289 {
 290         mdk_rdev_t *rdev;
 291         struct list_head *tmp;
 292
 293         ITERATE_RDEV(mddev, rdev, tmp)
 294                 if (rdev->in_sync && !rdev->faulty)
 295                         md_super_write(mddev, rdev,
 296                                        (rdev->sb_offset<<1) + offset
 297                                        + page->index * (PAGE_SIZE/512),
 298                                        PAGE_SIZE,
 299                                        page);
 300
 301         if (wait)
 302                 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 303         return 0;
 304 }
 305
 306 /*
 307  * write out a page to a file
 308  */
 309 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 310 {
 311         int ret = -ENOMEM;
 312
 313         if (bitmap->file == NULL)
 314                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 315
 316         if (wait)
 317                 lock_page(page);
 318         else {
 319                 if (TestSetPageLocked(page))
 320                         return -EAGAIN; /* already locked */
 321                 if (PageWriteback(page)) {
 322                         unlock_page(page);
 323                         return -EAGAIN;
 324                 }
 325         }
 326
 327         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 328         if (!ret)
 329                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 330                         PAGE_SIZE);
 331         if (ret) {
 332                 unlock_page(page);
 333                 return ret;
 334         }
 335
 336         set_page_dirty(page); /* force it to be written out */
 337
 338         if (!wait) {
 339                 /* add to list to be waited for by daemon */
 340                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 341                 item->page = page;
 342                 page_cache_get(page);
 343                 spin_lock(&bitmap->write_lock);
 344                 list_add(&item->list, &bitmap->complete_pages);
 345                 spin_unlock(&bitmap->write_lock);
 346                 md_wakeup_thread(bitmap->writeback_daemon);
 347         }
 348         return write_one_page(page, wait);
 349 }
 350
 351 /* read a page from a file, pinning it into cache, and return bytes_read */
 352 static struct page *read_page(struct file *file, unsigned long index,
 353                                         unsigned long *bytes_read)
 354 {
 355         struct inode *inode = file->f_mapping->host;
 356         struct page *page = NULL;
 357         loff_t isize = i_size_read(inode);
 358         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 359
 360         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 361                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 362
 363         page = read_cache_page(inode->i_mapping, index,
 364                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 365         if (IS_ERR(page))
 366                 goto out;
 367         wait_on_page_locked(page);
 368         if (!PageUptodate(page) || PageError(page)) {
 369                 page_cache_release(page);
 370                 page = ERR_PTR(-EIO);
 371                 goto out;
 372         }
 373
 374         if (index > end_index) /* we have read beyond EOF */
 375                 *bytes_read = 0;
 376         else if (index == end_index) /* possible short read */
 377                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 378         else
 379                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 380 out:
 381         if (IS_ERR(page))
 382                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 383                         (int)PAGE_CACHE_SIZE,
 384                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 385                         PTR_ERR(page));
 386         return page;
 387 }
 388
 389 /*
 390  * bitmap file superblock operations
 391  */
 392
 393 /* update the event counter and sync the superblock to disk */
 394 int bitmap_update_sb(struct bitmap *bitmap)
 395 {
 396         bitmap_super_t *sb;
 397         unsigned long flags;
 398
 399         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 400                 return 0;
 401         spin_lock_irqsave(&bitmap->lock, flags);
 402         if (!bitmap->sb_page) { /* no superblock */
 403                 spin_unlock_irqrestore(&bitmap->lock, flags);
 404                 return 0;
 405         }
 406         spin_unlock_irqrestore(&bitmap->lock, flags);
 407         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 408         sb->events = cpu_to_le64(bitmap->mddev->events);
 409         if (!bitmap->mddev->degraded)
 410                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 411         kunmap(bitmap->sb_page);
 412         return write_page(bitmap, bitmap->sb_page, 1);
 413 }
 414
 415 /* print out the bitmap file superblock */
 416 void bitmap_print_sb(struct bitmap *bitmap)
 417 {
 418         bitmap_super_t *sb;
 419
 420         if (!bitmap || !bitmap->sb_page)
 421                 return;
 422         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 423         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 424         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 425         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 426         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 427                                         *(__u32 *)(sb->uuid+0),
 428                                         *(__u32 *)(sb->uuid+4),
 429                                         *(__u32 *)(sb->uuid+8),
 430                                         *(__u32 *)(sb->uuid+12));
 431         printk(KERN_DEBUG "        events: %llu\n",
 432                         (unsigned long long) le64_to_cpu(sb->events));
 433         printk(KERN_DEBUG "events cleared: %llu\n",
 434                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 435         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 436         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 437         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 438         printk(KERN_DEBUG "     sync size: %llu KB\n",
 439                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 440         kunmap(bitmap->sb_page);
 441 }
 442
 443 /* read the superblock from the bitmap file and initialize some bitmap fields */
 444 static int bitmap_read_sb(struct bitmap *bitmap)
 445 {
 446         char *reason = NULL;
 447         bitmap_super_t *sb;
 448         unsigned long chunksize, daemon_sleep;
 449         unsigned long bytes_read;
 450         unsigned long long events;
 451         int err = -EINVAL;
 452
 453         /* page 0 is the superblock, read it... */
 454         if (bitmap->file)
 455                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 456         else {
 457                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 458                 bytes_read = PAGE_SIZE;
 459         }
 460         if (IS_ERR(bitmap->sb_page)) {
 461                 err = PTR_ERR(bitmap->sb_page);
 462                 bitmap->sb_page = NULL;
 463                 return err;
 464         }
 465
 466         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 467
 468         if (bytes_read < sizeof(*sb)) { /* short read */
 469                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 470                         bmname(bitmap));
 471                 err = -ENOSPC;
 472                 goto out;
 473         }
 474
 475         chunksize = le32_to_cpu(sb->chunksize);
 476         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 477
 478         /* verify that the bitmap-specific fields are valid */
 479         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 480                 reason = "bad magic";
 481         else if (sb->version != cpu_to_le32(BITMAP_MAJOR))
 482                 reason = "unrecognized superblock version";
 483         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 484                 reason = "bitmap chunksize out of range (512B - 4MB)";
 485         else if ((1 << ffz(~chunksize)) != chunksize)
 486                 reason = "bitmap chunksize not a power of 2";
 487         else if (daemon_sleep < 1 || daemon_sleep > 15)
 488                 reason = "daemon sleep period out of range";
 489         if (reason) {
 490                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 491                         bmname(bitmap), reason);
 492                 goto out;
 493         }
 494
 495         /* keep the array size field of the bitmap superblock up to date */
 496         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 497
 498         if (!bitmap->mddev->persistent)
 499                 goto success;
 500
 501         /*
 502          * if we have a persistent array superblock, compare the
 503          * bitmap's UUID and event counter to the mddev's
 504          */
 505         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 506                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 507                         bmname(bitmap));
 508                 goto out;
 509         }
 510         events = le64_to_cpu(sb->events);
 511         if (events < bitmap->mddev->events) {
 512                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 513                         "-- forcing full recovery\n", bmname(bitmap), events,
 514                         (unsigned long long) bitmap->mddev->events);
 515                 sb->state |= BITMAP_STALE;
 516         }
 517 success:
 518         /* assign fields using values from superblock */
 519         bitmap->chunksize = chunksize;
 520         bitmap->daemon_sleep = daemon_sleep;
 521         bitmap->flags |= sb->state;
 522         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 523         if (sb->state & BITMAP_STALE)
 524                 bitmap->events_cleared = bitmap->mddev->events;
 525         err = 0;
 526 out:
 527         kunmap(bitmap->sb_page);
 528         if (err)
 529                 bitmap_print_sb(bitmap);
 530         return err;
 531 }
 532
 533 enum bitmap_mask_op {
 534         MASK_SET,
 535         MASK_UNSET
 536 };
 537
 538 /* record the state of the bitmap in the superblock */
 539 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 540                                 enum bitmap_mask_op op)
 541 {
 542         bitmap_super_t *sb;
 543         unsigned long flags;
 544
 545         spin_lock_irqsave(&bitmap->lock, flags);
 546         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 547                 spin_unlock_irqrestore(&bitmap->lock, flags);
 548                 return;
 549         }
 550         page_cache_get(bitmap->sb_page);
 551         spin_unlock_irqrestore(&bitmap->lock, flags);
 552         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 553         switch (op) {
 554                 case MASK_SET: sb->state |= bits;
 555                                 break;
 556                 case MASK_UNSET: sb->state &= ~bits;
 557                                 break;
 558                 default: BUG();
 559         }
 560         kunmap(bitmap->sb_page);
 561         page_cache_release(bitmap->sb_page);
 562 }
 563
 564 /*
 565  * general bitmap file operations
 566  */
 567
 568 /* calculate the index of the page that contains this bit */
 569 static inline unsigned long file_page_index(unsigned long chunk)
 570 {
 571         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 572 }
 573
 574 /* calculate the (bit) offset of this bit within a page */
 575 static inline unsigned long file_page_offset(unsigned long chunk)
 576 {
 577         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 578 }
 579
 580 /*
 581  * return a pointer to the page in the filemap that contains the given bit
 582  *
 583  * this lookup is complicated by the fact that the bitmap sb might be exactly
 584  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 585  * 0 or page 1
 586  */
 587 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 588                                         unsigned long chunk)
 589 {
 590         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 591 }
 592
 593
 594 static void bitmap_file_unmap(struct bitmap *bitmap)
 595 {
 596         struct page **map, *sb_page;
 597         unsigned long *attr;
 598         int pages;
 599         unsigned long flags;
 600
 601         spin_lock_irqsave(&bitmap->lock, flags);
 602         map = bitmap->filemap;
 603         bitmap->filemap = NULL;
 604         attr = bitmap->filemap_attr;
 605         bitmap->filemap_attr = NULL;
 606         pages = bitmap->file_pages;
 607         bitmap->file_pages = 0;
 608         sb_page = bitmap->sb_page;
 609         bitmap->sb_page = NULL;
 610         spin_unlock_irqrestore(&bitmap->lock, flags);
 611
 612         while (pages--)
 613                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 614                         page_cache_release(map[pages]);
 615         kfree(map);
 616         kfree(attr);
 617
 618         if (sb_page)
 619                 page_cache_release(sb_page);
 620 }
 621
 622 static void bitmap_stop_daemons(struct bitmap *bitmap);
 623
 624 /* dequeue the next item in a page list -- don't call from irq context */
 625 static struct page_list *dequeue_page(struct bitmap *bitmap)
 626 {
 627         struct page_list *item = NULL;
 628         struct list_head *head = &bitmap->complete_pages;
 629
 630         spin_lock(&bitmap->write_lock);
 631         if (list_empty(head))
 632                 goto out;
 633         item = list_entry(head->prev, struct page_list, list);
 634         list_del(head->prev);
 635 out:
 636         spin_unlock(&bitmap->write_lock);
 637         return item;
 638 }
 639
 640 static void drain_write_queues(struct bitmap *bitmap)
 641 {
 642         struct page_list *item;
 643
 644         while ((item = dequeue_page(bitmap))) {
 645                 /* don't bother to wait */
 646                 page_cache_release(item->page);
 647                 mempool_free(item, bitmap->write_pool);
 648         }
 649
 650         wake_up(&bitmap->write_wait);
 651 }
 652
 653 static void bitmap_file_put(struct bitmap *bitmap)
 654 {
 655         struct file *file;
 656         struct inode *inode;
 657         unsigned long flags;
 658
 659         spin_lock_irqsave(&bitmap->lock, flags);
 660         file = bitmap->file;
 661         bitmap->file = NULL;
 662         spin_unlock_irqrestore(&bitmap->lock, flags);
 663
 664         bitmap_stop_daemons(bitmap);
 665
 666         drain_write_queues(bitmap);
 667
 668         bitmap_file_unmap(bitmap);
 669
 670         if (file) {
 671                 inode = file->f_mapping->host;
 672                 spin_lock(&inode->i_lock);
 673                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 674                 spin_unlock(&inode->i_lock);
 675                 fput(file);
 676         }
 677 }
 678
 679
 680 /*
 681  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 682  * then it is no longer reliable, so we stop using it and we mark the file
 683  * as failed in the superblock
 684  */
 685 static void bitmap_file_kick(struct bitmap *bitmap)
 686 {
 687         char *path, *ptr = NULL;
 688
 689         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 690         bitmap_update_sb(bitmap);
 691
 692         if (bitmap->file) {
 693                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 694                 if (path)
 695                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 696
 697                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 698                        bmname(bitmap), ptr ? ptr : "");
 699
 700                 kfree(path);
 701         }
 702
 703         bitmap_file_put(bitmap);
 704
 705         return;
 706 }
 707
 708 enum bitmap_page_attr {
 709         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 710         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 711         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 712 };
 713
 714 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 715                                 enum bitmap_page_attr attr)
 716 {
 717         bitmap->filemap_attr[page->index] |= attr;
 718 }
 719
 720 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 721                                 enum bitmap_page_attr attr)
 722 {
 723         bitmap->filemap_attr[page->index] &= ~attr;
 724 }
 725
 726 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 727 {
 728         return bitmap->filemap_attr[page->index];
 729 }
 730
 731 /*
 732  * bitmap_file_set_bit -- called before performing a write to the md device
 733  * to set (and eventually sync) a particular bit in the bitmap file
 734  *
 735  * we set the bit immediately, then we record the page number so that
 736  * when an unplug occurs, we can flush the dirty pages out to disk
 737  */
 738 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 739 {
 740         unsigned long bit;
 741         struct page *page;
 742         void *kaddr;
 743         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 744
 745         if (!bitmap->filemap) {
 746                 return;
 747         }
 748
 749         page = filemap_get_page(bitmap, chunk);
 750         bit = file_page_offset(chunk);
 751
 752
 753         /* make sure the page stays cached until it gets written out */
 754         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 755                 page_cache_get(page);
 756
 757         /* set the bit */
 758         kaddr = kmap_atomic(page, KM_USER0);
 759         set_bit(bit, kaddr);
 760         kunmap_atomic(kaddr, KM_USER0);
 761         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 762
 763         /* record page number so it gets flushed to disk when unplug occurs */
 764         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 765
 766 }
 767
 768 /* this gets called when the md device is ready to unplug its underlying
 769  * (slave) device queues -- before we let any writes go down, we need to
 770  * sync the dirty pages of the bitmap file to disk */
 771 int bitmap_unplug(struct bitmap *bitmap)
 772 {
 773         unsigned long i, attr, flags;
 774         struct page *page;
 775         int wait = 0;
 776         int err;
 777
 778         if (!bitmap)
 779                 return 0;
 780
 781         /* look at each page to see if there are any set bits that need to be
 782          * flushed out to disk */
 783         for (i = 0; i < bitmap->file_pages; i++) {
 784                 spin_lock_irqsave(&bitmap->lock, flags);
 785                 if (!bitmap->filemap) {
 786                         spin_unlock_irqrestore(&bitmap->lock, flags);
 787                         return 0;
 788                 }
 789                 page = bitmap->filemap[i];
 790                 attr = get_page_attr(bitmap, page);
 791                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 792                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 793                 if ((attr & BITMAP_PAGE_DIRTY))
 794                         wait = 1;
 795                 spin_unlock_irqrestore(&bitmap->lock, flags);
 796
 797                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 798                         err = write_page(bitmap, page, 0);
 799                         if (err == -EAGAIN) {
 800                                 if (attr & BITMAP_PAGE_DIRTY)
 801                                         err = write_page(bitmap, page, 1);
 802                                 else
 803                                         err = 0;
 804                         }
 805                         if (err)
 806                                 return 1;
 807                 }
 808         }
 809         if (wait) { /* if any writes were performed, we need to wait on them */
 810                 if (bitmap->file) {
 811                         spin_lock_irq(&bitmap->write_lock);
 812                         wait_event_lock_irq(bitmap->write_wait,
 813                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 814                                             wake_up_process(bitmap->writeback_daemon->tsk));
 815                         spin_unlock_irq(&bitmap->write_lock);
 816                 } else
 817                         wait_event(bitmap->mddev->sb_wait,
 818                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 819         }
 820         return 0;
 821 }
 822
 823 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 824 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 825  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 826  * memory mapping of the bitmap file
 827  * Special cases:
 828  *   if there's no bitmap file, or if the bitmap file had been
 829  *   previously kicked from the array, we mark all the bits as
 830  *   1's in order to cause a full resync.
 831  *
 832  * We ignore all bits for sectors that end earlier than 'start'.
 833  * This is used when reading an out-of-date bitmap...
 834  */
 835 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 836 {
 837         unsigned long i, chunks, index, oldindex, bit;
 838         struct page *page = NULL, *oldpage = NULL;
 839         unsigned long num_pages, bit_cnt = 0;
 840         struct file *file;
 841         unsigned long bytes, offset, dummy;
 842         int outofdate;
 843         int ret = -ENOSPC;
 844
 845         chunks = bitmap->chunks;
 846         file = bitmap->file;
 847
 848         BUG_ON(!file && !bitmap->offset);
 849
 850 #ifdef INJECT_FAULTS_3
 851         outofdate = 1;
 852 #else
 853         outofdate = bitmap->flags & BITMAP_STALE;
 854 #endif
 855         if (outofdate)
 856                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 857                         "recovery\n", bmname(bitmap));
 858
 859         bytes = (chunks + 7) / 8;
 860
 861         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 862
 863         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 864                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 865                         bmname(bitmap),
 866                         (unsigned long) i_size_read(file->f_mapping->host),
 867                         bytes + sizeof(bitmap_super_t));
 868                 goto out;
 869         }
 870
 871         ret = -ENOMEM;
 872
 873         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 874         if (!bitmap->filemap)
 875                 goto out;
 876
 877         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 878         if (!bitmap->filemap_attr)
 879                 goto out;
 880
 881         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 882
 883         oldindex = ~0L;
 884
 885         for (i = 0; i < chunks; i++) {
 886                 index = file_page_index(i);
 887                 bit = file_page_offset(i);
 888                 if (index != oldindex) { /* this is a new page, read it in */
 889                         /* unmap the old page, we're done with it */
 890                         if (oldpage != NULL)
 891                                 kunmap(oldpage);
 892                         if (index == 0) {
 893                                 /*
 894                                  * if we're here then the superblock page
 895                                  * contains some bits (PAGE_SIZE != sizeof sb)
 896                                  * we've already read it in, so just use it
 897                                  */
 898                                 page = bitmap->sb_page;
 899                                 offset = sizeof(bitmap_super_t);
 900                         } else if (file) {
 901                                 page = read_page(file, index, &dummy);
 902                                 offset = 0;
 903                         } else {
 904                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 905                                 offset = 0;
 906                         }
 907                         if (IS_ERR(page)) { /* read error */
 908                                 ret = PTR_ERR(page);
 909                                 goto out;
 910                         }
 911
 912                         oldindex = index;
 913                         oldpage = page;
 914                         kmap(page);
 915
 916                         if (outofdate) {
 917                                 /*
 918                                  * if bitmap is out of date, dirty the
 919                                  * whole page and write it out
 920                                  */
 921                                 memset(page_address(page) + offset, 0xff,
 922                                        PAGE_SIZE - offset);
 923                                 ret = write_page(bitmap, page, 1);
 924                                 if (ret) {
 925                                         kunmap(page);
 926                                         /* release, page not in filemap yet */
 927                                         page_cache_release(page);
 928                                         goto out;
 929                                 }
 930                         }
 931
 932                         bitmap->filemap[bitmap->file_pages++] = page;
 933                 }
 934                 if (test_bit(bit, page_address(page))) {
 935                         /* if the disk bit is set, set the memory bit */
 936                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 937                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 938                                 );
 939                         bit_cnt++;
 940                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 941                 }
 942         }
 943
 944         /* everything went OK */
 945         ret = 0;
 946         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 947
 948         if (page) /* unmap the last page */
 949                 kunmap(page);
 950
 951         if (bit_cnt) { /* Kick recovery if any bits were set */
 952                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 953                 md_wakeup_thread(bitmap->mddev->thread);
 954         }
 955
 956 out:
 957         printk(KERN_INFO "%s: bitmap initialized from disk: "
 958                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 959                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 960
 961         return ret;
 962 }
 963
 964 void bitmap_write_all(struct bitmap *bitmap)
 965 {
 966         /* We don't actually write all bitmap blocks here,
 967          * just flag them as needing to be written
 968          */
 969
 970         unsigned long chunks = bitmap->chunks;
 971         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 972         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 973         while (num_pages--)
 974                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 975 }
 976
 977
 978 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 979 {
 980         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 981         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 982         bitmap->bp[page].count += inc;
 983 /*
 984         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 985                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 986 */
 987         bitmap_checkfree(bitmap, page);
 988 }
 989 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 990                                             sector_t offset, int *blocks,
 991                                             int create);
 992
 993 /*
 994  * bitmap daemon -- periodically wakes up to clean bits and flush pages
 995  *                      out to disk
 996  */
 997
 998 int bitmap_daemon_work(struct bitmap *bitmap)
 999 {
1000         unsigned long j;
1001         unsigned long flags;
1002         struct page *page = NULL, *lastpage = NULL;
1003         int err = 0;
1004         int blocks;
1005         int attr;
1006
1007         if (bitmap == NULL)
1008                 return 0;
1009         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1010                 return 0;
1011         bitmap->daemon_lastrun = jiffies;
1012
1013         for (j = 0; j < bitmap->chunks; j++) {
1014                 bitmap_counter_t *bmc;
1015                 spin_lock_irqsave(&bitmap->lock, flags);
1016                 if (!bitmap->filemap) {
1017                         /* error or shutdown */
1018                         spin_unlock_irqrestore(&bitmap->lock, flags);
1019                         break;
1020                 }
1021
1022                 page = filemap_get_page(bitmap, j);
1023
1024                 if (page != lastpage) {
1025                         /* skip this page unless it's marked as needing cleaning */
1026                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1027                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1028                                         page_cache_get(page);
1029                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1030                                 }
1031                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1032                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1033                                         switch (write_page(bitmap, page, 0)) {
1034                                         case -EAGAIN:
1035                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1036                                                 break;
1037                                         case 0:
1038                                                 break;
1039                                         default:
1040                                                 bitmap_file_kick(bitmap);
1041                                         }
1042                                         page_cache_release(page);
1043                                 }
1044                                 continue;
1045                         }
1046
1047                         /* grab the new page, sync and release the old */
1048                         page_cache_get(page);
1049                         if (lastpage != NULL) {
1050                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1051                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1052                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1053                                         err = write_page(bitmap, lastpage, 0);
1054                                         if (err == -EAGAIN) {
1055                                                 err = 0;
1056                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1057                                         }
1058                                 } else {
1059                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1060                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1061                                 }
1062                                 kunmap(lastpage);
1063                                 page_cache_release(lastpage);
1064                                 if (err)
1065                                         bitmap_file_kick(bitmap);
1066                         } else
1067                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1068                         lastpage = page;
1069                         kmap(page);
1070 /*
1071                         printk("bitmap clean at page %lu\n", j);
1072 */
1073                         spin_lock_irqsave(&bitmap->lock, flags);
1074                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1075                 }
1076                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1077                                         &blocks, 0);
1078                 if (bmc) {
1079 /*
1080   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1081 */
1082                         if (*bmc == 2) {
1083                                 *bmc=1; /* maybe clear the bit next time */
1084                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1085                         } else if (*bmc == 1) {
1086                                 /* we can clear the bit */
1087                                 *bmc = 0;
1088                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1089                                                   -1);
1090
1091                                 /* clear the bit */
1092                                 clear_bit(file_page_offset(j), page_address(page));
1093                         }
1094                 }
1095                 spin_unlock_irqrestore(&bitmap->lock, flags);
1096         }
1097
1098         /* now sync the final page */
1099         if (lastpage != NULL) {
1100                 kunmap(lastpage);
1101                 spin_lock_irqsave(&bitmap->lock, flags);
1102                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1103                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1104                         spin_unlock_irqrestore(&bitmap->lock, flags);
1105                         err = write_page(bitmap, lastpage, 0);
1106                         if (err == -EAGAIN) {
1107                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1108                                 err = 0;
1109                         }
1110                 } else {
1111                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1112                         spin_unlock_irqrestore(&bitmap->lock, flags);
1113                 }
1114
1115                 page_cache_release(lastpage);
1116         }
1117
1118         return err;
1119 }
1120
1121 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1122 {
1123         mdk_thread_t *dmn;
1124         unsigned long flags;
1125
1126         /* if no one is waiting on us, we'll free the md thread struct
1127          * and exit, otherwise we let the waiter clean things up */
1128         spin_lock_irqsave(&bitmap->lock, flags);
1129         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1130                 *daemon = NULL;
1131                 spin_unlock_irqrestore(&bitmap->lock, flags);
1132                 kfree(dmn);
1133                 complete_and_exit(NULL, 0); /* do_exit not exported */
1134         }
1135         spin_unlock_irqrestore(&bitmap->lock, flags);
1136 }
1137
1138 static void bitmap_writeback_daemon(mddev_t *mddev)
1139 {
1140         struct bitmap *bitmap = mddev->bitmap;
1141         struct page *page;
1142         struct page_list *item;
1143         int err = 0;
1144
1145         if (signal_pending(current)) {
1146                 printk(KERN_INFO
1147                        "%s: bitmap writeback daemon got signal, exiting...\n",
1148                        bmname(bitmap));
1149                 err = -EINTR;
1150                 goto out;
1151         }
1152
1153         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1154         /* wait on bitmap page writebacks */
1155         while ((item = dequeue_page(bitmap))) {
1156                 page = item->page;
1157                 mempool_free(item, bitmap->write_pool);
1158                 PRINTK("wait on page writeback: %p\n", page);
1159                 wait_on_page_writeback(page);
1160                 PRINTK("finished page writeback: %p\n", page);
1161
1162                 err = PageError(page);
1163                 page_cache_release(page);
1164                 if (err) {
1165                         printk(KERN_WARNING "%s: bitmap file writeback "
1166                                "failed (page %lu): %d\n",
1167                                bmname(bitmap), page->index, err);
1168                         bitmap_file_kick(bitmap);
1169                         goto out;
1170                 }
1171         }
1172  out:
1173         wake_up(&bitmap->write_wait);
1174         if (err) {
1175                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1176                        bmname(bitmap), err);
1177                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1178         }
1179 }
1180
1181 static int bitmap_start_daemon(struct bitmap *bitmap, mdk_thread_t **ptr,
1182                                 void (*func)(mddev_t *), char *name)
1183 {
1184         mdk_thread_t *daemon;
1185         unsigned long flags;
1186         char namebuf[32];
1187
1188         spin_lock_irqsave(&bitmap->lock, flags);
1189         *ptr = NULL;
1190
1191         if (!bitmap->file) /* no need for daemon if there's no backing file */
1192                 goto out_unlock;
1193
1194         spin_unlock_irqrestore(&bitmap->lock, flags);
1195
1196 #ifdef INJECT_FATAL_FAULT_2
1197         daemon = NULL;
1198 #else
1199         sprintf(namebuf, "%%s_%s", name);
1200         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1201 #endif
1202         if (!daemon) {
1203                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1204                         bmname(bitmap));
1205                 return -ECHILD;
1206         }
1207
1208         spin_lock_irqsave(&bitmap->lock, flags);
1209         *ptr = daemon;
1210
1211         md_wakeup_thread(daemon); /* start it running */
1212
1213         PRINTK("%s: %s daemon (pid %d) started...\n",
1214                 bmname(bitmap), name, daemon->tsk->pid);
1215 out_unlock:
1216         spin_unlock_irqrestore(&bitmap->lock, flags);
1217         return 0;
1218 }
1219
1220 static int bitmap_start_daemons(struct bitmap *bitmap)
1221 {
1222         int err = bitmap_start_daemon(bitmap, &bitmap->writeback_daemon,
1223                                         bitmap_writeback_daemon, "bitmap_wb");
1224         return err;
1225 }
1226
1227 static void bitmap_stop_daemon(struct bitmap *bitmap, mdk_thread_t **ptr)
1228 {
1229         mdk_thread_t *daemon;
1230         unsigned long flags;
1231
1232         spin_lock_irqsave(&bitmap->lock, flags);
1233         daemon = *ptr;
1234         *ptr = NULL;
1235         spin_unlock_irqrestore(&bitmap->lock, flags);
1236         if (daemon)
1237                 md_unregister_thread(daemon); /* destroy the thread */
1238 }
1239
1240 static void bitmap_stop_daemons(struct bitmap *bitmap)
1241 {
1242         /* the daemons can't stop themselves... they'll just exit instead... */
1243         if (bitmap->writeback_daemon &&
1244             current->pid != bitmap->writeback_daemon->tsk->pid)
1245                 bitmap_stop_daemon(bitmap, &bitmap->writeback_daemon);
1246 }
1247
1248 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1249                                             sector_t offset, int *blocks,
1250                                             int create)
1251 {
1252         /* If 'create', we might release the lock and reclaim it.
1253          * The lock must have been taken with interrupts enabled.
1254          * If !create, we don't release the lock.
1255          */
1256         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1257         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1258         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1259         sector_t csize;
1260
1261         if (bitmap_checkpage(bitmap, page, create) < 0) {
1262                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1263                 *blocks = csize - (offset & (csize- 1));
1264                 return NULL;
1265         }
1266         /* now locked ... */
1267
1268         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1269                 /* should we use the first or second counter field
1270                  * of the hijacked pointer? */
1271                 int hi = (pageoff > PAGE_COUNTER_MASK);
1272                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1273                                           PAGE_COUNTER_SHIFT - 1);
1274                 *blocks = csize - (offset & (csize- 1));
1275                 return  &((bitmap_counter_t *)
1276                           &bitmap->bp[page].map)[hi];
1277         } else { /* page is allocated */
1278                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1279                 *blocks = csize - (offset & (csize- 1));
1280                 return (bitmap_counter_t *)
1281                         &(bitmap->bp[page].map[pageoff]);
1282         }
1283 }
1284
1285 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors)
1286 {
1287         if (!bitmap) return 0;
1288         while (sectors) {
1289                 int blocks;
1290                 bitmap_counter_t *bmc;
1291
1292                 spin_lock_irq(&bitmap->lock);
1293                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1294                 if (!bmc) {
1295                         spin_unlock_irq(&bitmap->lock);
1296                         return 0;
1297                 }
1298
1299                 switch(*bmc) {
1300                 case 0:
1301                         bitmap_file_set_bit(bitmap, offset);
1302                         bitmap_count_page(bitmap,offset, 1);
1303                         blk_plug_device(bitmap->mddev->queue);
1304                         /* fall through */
1305                 case 1:
1306                         *bmc = 2;
1307                 }
1308                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1309                 (*bmc)++;
1310
1311                 spin_unlock_irq(&bitmap->lock);
1312
1313                 offset += blocks;
1314                 if (sectors > blocks)
1315                         sectors -= blocks;
1316                 else sectors = 0;
1317         }
1318         return 0;
1319 }
1320
1321 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1322                      int success)
1323 {
1324         if (!bitmap) return;
1325         while (sectors) {
1326                 int blocks;
1327                 unsigned long flags;
1328                 bitmap_counter_t *bmc;
1329
1330                 spin_lock_irqsave(&bitmap->lock, flags);
1331                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1332                 if (!bmc) {
1333                         spin_unlock_irqrestore(&bitmap->lock, flags);
1334                         return;
1335                 }
1336
1337                 if (!success && ! (*bmc & NEEDED_MASK))
1338                         *bmc |= NEEDED_MASK;
1339
1340                 (*bmc)--;
1341                 if (*bmc <= 2) {
1342                         set_page_attr(bitmap,
1343                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1344                                       BITMAP_PAGE_CLEAN);
1345                 }
1346                 spin_unlock_irqrestore(&bitmap->lock, flags);
1347                 offset += blocks;
1348                 if (sectors > blocks)
1349                         sectors -= blocks;
1350                 else sectors = 0;
1351         }
1352 }
1353
1354 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1355                         int degraded)
1356 {
1357         bitmap_counter_t *bmc;
1358         int rv;
1359         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1360                 *blocks = 1024;
1361                 return 1; /* always resync if no bitmap */
1362         }
1363         spin_lock_irq(&bitmap->lock);
1364         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1365         rv = 0;
1366         if (bmc) {
1367                 /* locked */
1368                 if (RESYNC(*bmc))
1369                         rv = 1;
1370                 else if (NEEDED(*bmc)) {
1371                         rv = 1;
1372                         if (!degraded) { /* don't set/clear bits if degraded */
1373                                 *bmc |= RESYNC_MASK;
1374                                 *bmc &= ~NEEDED_MASK;
1375                         }
1376                 }
1377         }
1378         spin_unlock_irq(&bitmap->lock);
1379         return rv;
1380 }
1381
1382 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1383 {
1384         bitmap_counter_t *bmc;
1385         unsigned long flags;
1386 /*
1387         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1388 */      if (bitmap == NULL) {
1389                 *blocks = 1024;
1390                 return;
1391         }
1392         spin_lock_irqsave(&bitmap->lock, flags);
1393         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1394         if (bmc == NULL)
1395                 goto unlock;
1396         /* locked */
1397 /*
1398         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1399 */
1400         if (RESYNC(*bmc)) {
1401                 *bmc &= ~RESYNC_MASK;
1402
1403                 if (!NEEDED(*bmc) && aborted)
1404                         *bmc |= NEEDED_MASK;
1405                 else {
1406                         if (*bmc <= 2) {
1407                                 set_page_attr(bitmap,
1408                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1409                                               BITMAP_PAGE_CLEAN);
1410                         }
1411                 }
1412         }
1413  unlock:
1414         spin_unlock_irqrestore(&bitmap->lock, flags);
1415 }
1416
1417 void bitmap_close_sync(struct bitmap *bitmap)
1418 {
1419         /* Sync has finished, and any bitmap chunks that weren't synced
1420          * properly have been aborted.  It remains to us to clear the
1421          * RESYNC bit wherever it is still on
1422          */
1423         sector_t sector = 0;
1424         int blocks;
1425         if (!bitmap) return;
1426         while (sector < bitmap->mddev->resync_max_sectors) {
1427                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1428 /*
1429                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1430                                          (unsigned long long)sector, blocks);
1431 */              sector += blocks;
1432         }
1433 }
1434
1435 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1436 {
1437         /* For each chunk covered by any of these sectors, set the
1438          * counter to 1 and set resync_needed.  They should all
1439          * be 0 at this point
1440          */
1441
1442         int secs;
1443         bitmap_counter_t *bmc;
1444         spin_lock_irq(&bitmap->lock);
1445         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1446         if (!bmc) {
1447                 spin_unlock_irq(&bitmap->lock);
1448                 return;
1449         }
1450         if (! *bmc) {
1451                 struct page *page;
1452                 *bmc = 1 | (needed?NEEDED_MASK:0);
1453                 bitmap_count_page(bitmap, offset, 1);
1454                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1455                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1456         }
1457         spin_unlock_irq(&bitmap->lock);
1458
1459 }
1460
1461 /*
1462  * flush out any pending updates
1463  */
1464 void bitmap_flush(mddev_t *mddev)
1465 {
1466         struct bitmap *bitmap = mddev->bitmap;
1467         int sleep;
1468
1469         if (!bitmap) /* there was no bitmap */
1470                 return;
1471
1472         /* run the daemon_work three time to ensure everything is flushed
1473          * that can be
1474          */
1475         sleep = bitmap->daemon_sleep;
1476         bitmap->daemon_sleep = 0;
1477         bitmap_daemon_work(bitmap);
1478         bitmap_daemon_work(bitmap);
1479         bitmap_daemon_work(bitmap);
1480         bitmap->daemon_sleep = sleep;
1481         bitmap_update_sb(bitmap);
1482 }
1483
1484 /*
1485  * free memory that was allocated
1486  */
1487 void bitmap_destroy(mddev_t *mddev)
1488 {
1489         unsigned long k, pages;
1490         struct bitmap_page *bp;
1491         struct bitmap *bitmap = mddev->bitmap;
1492
1493         if (!bitmap) /* there was no bitmap */
1494                 return;
1495
1496         mddev->bitmap = NULL; /* disconnect from the md device */
1497
1498         /* release the bitmap file and kill the daemon */
1499         bitmap_file_put(bitmap);
1500
1501         bp = bitmap->bp;
1502         pages = bitmap->pages;
1503
1504         /* free all allocated memory */
1505
1506         mempool_destroy(bitmap->write_pool);
1507
1508         if (bp) /* deallocate the page memory */
1509                 for (k = 0; k < pages; k++)
1510                         if (bp[k].map && !bp[k].hijacked)
1511                                 kfree(bp[k].map);
1512         kfree(bp);
1513         kfree(bitmap);
1514 }
1515
1516 /*
1517  * initialize the bitmap structure
1518  * if this returns an error, bitmap_destroy must be called to do clean up
1519  */
1520 int bitmap_create(mddev_t *mddev)
1521 {
1522         struct bitmap *bitmap;
1523         unsigned long blocks = mddev->resync_max_sectors;
1524         unsigned long chunks;
1525         unsigned long pages;
1526         struct file *file = mddev->bitmap_file;
1527         int err;
1528         sector_t start;
1529
1530         BUG_ON(sizeof(bitmap_super_t) != 256);
1531
1532         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1533                 return 0;
1534
1535         BUG_ON(file && mddev->bitmap_offset);
1536
1537         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1538         if (!bitmap)
1539                 return -ENOMEM;
1540
1541         memset(bitmap, 0, sizeof(*bitmap));
1542
1543         spin_lock_init(&bitmap->lock);
1544         bitmap->mddev = mddev;
1545         mddev->bitmap = bitmap;
1546
1547         spin_lock_init(&bitmap->write_lock);
1548         INIT_LIST_HEAD(&bitmap->complete_pages);
1549         init_waitqueue_head(&bitmap->write_wait);
1550         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1551                                 write_pool_free, NULL);
1552         if (!bitmap->write_pool)
1553                 return -ENOMEM;
1554
1555         bitmap->file = file;
1556         bitmap->offset = mddev->bitmap_offset;
1557         if (file) get_file(file);
1558         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1559         err = bitmap_read_sb(bitmap);
1560         if (err)
1561                 return err;
1562
1563         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1564                                         sizeof(bitmap->chunksize));
1565
1566         /* now that chunksize and chunkshift are set, we can use these macros */
1567         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1568                         CHUNK_BLOCK_RATIO(bitmap);
1569         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1570
1571         BUG_ON(!pages);
1572
1573         bitmap->chunks = chunks;
1574         bitmap->pages = pages;
1575         bitmap->missing_pages = pages;
1576         bitmap->counter_bits = COUNTER_BITS;
1577
1578         bitmap->syncchunk = ~0UL;
1579
1580 #ifdef INJECT_FATAL_FAULT_1
1581         bitmap->bp = NULL;
1582 #else
1583         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1584 #endif
1585         if (!bitmap->bp)
1586                 return -ENOMEM;
1587         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1588
1589         bitmap->flags |= BITMAP_ACTIVE;
1590
1591         /* now that we have some pages available, initialize the in-memory
1592          * bitmap from the on-disk bitmap */
1593         start = 0;
1594         if (mddev->degraded == 0
1595             || bitmap->events_cleared == mddev->events)
1596                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1597                 start = mddev->recovery_cp;
1598         err = bitmap_init_from_disk(bitmap, start);
1599
1600         if (err)
1601                 return err;
1602
1603         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1604                 pages, bmname(bitmap));
1605
1606         /* kick off the bitmap daemons */
1607         err = bitmap_start_daemons(bitmap);
1608         if (err)
1609                 return err;
1610         return bitmap_update_sb(bitmap);
1611 }
1612
1613 /* the bitmap API -- for raid personalities */
1614 EXPORT_SYMBOL(bitmap_startwrite);
1615 EXPORT_SYMBOL(bitmap_endwrite);
1616 EXPORT_SYMBOL(bitmap_start_sync);
1617 EXPORT_SYMBOL(bitmap_end_sync);
1618 EXPORT_SYMBOL(bitmap_unplug);
1619 EXPORT_SYMBOL(bitmap_close_sync);
1620 EXPORT_SYMBOL(bitmap_daemon_work);