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/errno.h>
  25 #include <linux/slab.h>
  26 #include <linux/init.h>
  27 #include <linux/config.h>
  28 #include <linux/timer.h>
  29 #include <linux/sched.h>
  30 #include <linux/list.h>
  31 #include <linux/file.h>
  32 #include <linux/mount.h>
  33 #include <linux/buffer_head.h>
  34 #include <linux/raid/md.h>
  35 #include <linux/raid/bitmap.h>
  36
  37 /* debug macros */
  38
  39 #define DEBUG 0
  40
  41 #if DEBUG
  42 /* these are for debugging purposes only! */
  43
  44 /* define one and only one of these */
  45 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  46 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  47 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  48 #define INJECT_FAULTS_4 0 /* undef */
  49 #define INJECT_FAULTS_5 0 /* undef */
  50 #define INJECT_FAULTS_6 0
  51
  52 /* if these are defined, the driver will fail! debug only */
  53 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  54 #define INJECT_FATAL_FAULT_2 0 /* undef */
  55 #define INJECT_FATAL_FAULT_3 0 /* undef */
  56 #endif
  57
  58 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  59 #define DPRINTK(x...) do { } while(0)
  60
  61 #ifndef PRINTK
  62 #  if DEBUG > 0
  63 #    define PRINTK(x...) printk(KERN_DEBUG x)
  64 #  else
  65 #    define PRINTK(x...)
  66 #  endif
  67 #endif
  68
  69 static inline char * bmname(struct bitmap *bitmap)
  70 {
  71         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  72 }
  73
  74
  75 /*
  76  * test if the bitmap is active
  77  */
  78 int bitmap_active(struct bitmap *bitmap)
  79 {
  80         unsigned long flags;
  81         int res = 0;
  82
  83         if (!bitmap)
  84                 return res;
  85         spin_lock_irqsave(&bitmap->lock, flags);
  86         res = bitmap->flags & BITMAP_ACTIVE;
  87         spin_unlock_irqrestore(&bitmap->lock, flags);
  88         return res;
  89 }
  90
  91 #define WRITE_POOL_SIZE 256
  92 /* mempool for queueing pending writes on the bitmap file */
  93 static void *write_pool_alloc(gfp_t gfp_flags, void *data)
  94 {
  95         return kmalloc(sizeof(struct page_list), gfp_flags);
  96 }
  97
  98 static void write_pool_free(void *ptr, void *data)
  99 {
 100         kfree(ptr);
 101 }
 102
 103 /*
 104  * just a placeholder - calls kmalloc for bitmap pages
 105  */
 106 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 107 {
 108         unsigned char *page;
 109
 110 #ifdef INJECT_FAULTS_1
 111         page = NULL;
 112 #else
 113         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 114 #endif
 115         if (!page)
 116                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 117         else
 118                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 119                         bmname(bitmap), page);
 120         return page;
 121 }
 122
 123 /*
 124  * for now just a placeholder -- just calls kfree for bitmap pages
 125  */
 126 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 127 {
 128         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 129         kfree(page);
 130 }
 131
 132 /*
 133  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 134  *
 135  * 1) check to see if this page is allocated, if it's not then try to alloc
 136  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 137  *    page pointer directly as a counter
 138  *
 139  * if we find our page, we increment the page's refcount so that it stays
 140  * allocated while we're using it
 141  */
 142 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 143 {
 144         unsigned char *mappage;
 145
 146         if (page >= bitmap->pages) {
 147                 printk(KERN_ALERT
 148                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 149                         bmname(bitmap), page, bitmap->pages-1);
 150                 return -EINVAL;
 151         }
 152
 153
 154         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 155                 return 0;
 156
 157         if (bitmap->bp[page].map) /* page is already allocated, just return */
 158                 return 0;
 159
 160         if (!create)
 161                 return -ENOENT;
 162
 163         spin_unlock_irq(&bitmap->lock);
 164
 165         /* this page has not been allocated yet */
 166
 167         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 168                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 169                         bmname(bitmap));
 170                 /* failed - set the hijacked flag so that we can use the
 171                  * pointer as a counter */
 172                 spin_lock_irq(&bitmap->lock);
 173                 if (!bitmap->bp[page].map)
 174                         bitmap->bp[page].hijacked = 1;
 175                 goto out;
 176         }
 177
 178         /* got a page */
 179
 180         spin_lock_irq(&bitmap->lock);
 181
 182         /* recheck the page */
 183
 184         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 185                 /* somebody beat us to getting the page */
 186                 bitmap_free_page(bitmap, mappage);
 187                 return 0;
 188         }
 189
 190         /* no page was in place and we have one, so install it */
 191
 192         memset(mappage, 0, PAGE_SIZE);
 193         bitmap->bp[page].map = mappage;
 194         bitmap->missing_pages--;
 195 out:
 196         return 0;
 197 }
 198
 199
 200 /* if page is completely empty, put it back on the free list, or dealloc it */
 201 /* if page was hijacked, unmark the flag so it might get alloced next time */
 202 /* Note: lock should be held when calling this */
 203 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 204 {
 205         char *ptr;
 206
 207         if (bitmap->bp[page].count) /* page is still busy */
 208                 return;
 209
 210         /* page is no longer in use, it can be released */
 211
 212         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 213                 bitmap->bp[page].hijacked = 0;
 214                 bitmap->bp[page].map = NULL;
 215                 return;
 216         }
 217
 218         /* normal case, free the page */
 219
 220 #if 0
 221 /* actually ... let's not.  We will probably need the page again exactly when
 222  * memory is tight and we are flusing to disk
 223  */
 224         return;
 225 #else
 226         ptr = bitmap->bp[page].map;
 227         bitmap->bp[page].map = NULL;
 228         bitmap->missing_pages++;
 229         bitmap_free_page(bitmap, ptr);
 230         return;
 231 #endif
 232 }
 233
 234
 235 /*
 236  * bitmap file handling - read and write the bitmap file and its superblock
 237  */
 238
 239 /* copy the pathname of a file to a buffer */
 240 char *file_path(struct file *file, char *buf, int count)
 241 {
 242         struct dentry *d;
 243         struct vfsmount *v;
 244
 245         if (!buf)
 246                 return NULL;
 247
 248         d = file->f_dentry;
 249         v = file->f_vfsmnt;
 250
 251         buf = d_path(d, v, buf, count);
 252
 253         return IS_ERR(buf) ? NULL : buf;
 254 }
 255
 256 /*
 257  * basic page I/O operations
 258  */
 259
 260 /* IO operations when bitmap is stored near all superblocks */
 261 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 262 {
 263         /* choose a good rdev and read the page from there */
 264
 265         mdk_rdev_t *rdev;
 266         struct list_head *tmp;
 267         struct page *page = alloc_page(GFP_KERNEL);
 268         sector_t target;
 269
 270         if (!page)
 271                 return ERR_PTR(-ENOMEM);
 272
 273         ITERATE_RDEV(mddev, rdev, tmp) {
 274                 if (! test_bit(In_sync, &rdev->flags)
 275                     || test_bit(Faulty, &rdev->flags))
 276                         continue;
 277
 278                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 279
 280                 if (sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ)) {
 281                         page->index = index;
 282                         return page;
 283                 }
 284         }
 285         return ERR_PTR(-EIO);
 286
 287 }
 288
 289 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 290 {
 291         mdk_rdev_t *rdev;
 292         struct list_head *tmp;
 293
 294         ITERATE_RDEV(mddev, rdev, tmp)
 295                 if (test_bit(In_sync, &rdev->flags)
 296                     && !test_bit(Faulty, &rdev->flags))
 297                         md_super_write(mddev, rdev,
 298                                        (rdev->sb_offset<<1) + offset
 299                                        + page->index * (PAGE_SIZE/512),
 300                                        PAGE_SIZE,
 301                                        page);
 302
 303         if (wait)
 304                 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 305         return 0;
 306 }
 307
 308 /*
 309  * write out a page to a file
 310  */
 311 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 312 {
 313         int ret = -ENOMEM;
 314
 315         if (bitmap->file == NULL)
 316                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 317
 318         if (wait)
 319                 lock_page(page);
 320         else {
 321                 if (TestSetPageLocked(page))
 322                         return -EAGAIN; /* already locked */
 323                 if (PageWriteback(page)) {
 324                         unlock_page(page);
 325                         return -EAGAIN;
 326                 }
 327         }
 328
 329         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 330         if (!ret)
 331                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 332                         PAGE_SIZE);
 333         if (ret) {
 334                 unlock_page(page);
 335                 return ret;
 336         }
 337
 338         set_page_dirty(page); /* force it to be written out */
 339
 340         if (!wait) {
 341                 /* add to list to be waited for by daemon */
 342                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 343                 item->page = page;
 344                 page_cache_get(page);
 345                 spin_lock(&bitmap->write_lock);
 346                 list_add(&item->list, &bitmap->complete_pages);
 347                 spin_unlock(&bitmap->write_lock);
 348                 md_wakeup_thread(bitmap->writeback_daemon);
 349         }
 350         return write_one_page(page, wait);
 351 }
 352
 353 /* read a page from a file, pinning it into cache, and return bytes_read */
 354 static struct page *read_page(struct file *file, unsigned long index,
 355                                         unsigned long *bytes_read)
 356 {
 357         struct inode *inode = file->f_mapping->host;
 358         struct page *page = NULL;
 359         loff_t isize = i_size_read(inode);
 360         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 361
 362         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 363                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 364
 365         page = read_cache_page(inode->i_mapping, index,
 366                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 367         if (IS_ERR(page))
 368                 goto out;
 369         wait_on_page_locked(page);
 370         if (!PageUptodate(page) || PageError(page)) {
 371                 page_cache_release(page);
 372                 page = ERR_PTR(-EIO);
 373                 goto out;
 374         }
 375
 376         if (index > end_index) /* we have read beyond EOF */
 377                 *bytes_read = 0;
 378         else if (index == end_index) /* possible short read */
 379                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 380         else
 381                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 382 out:
 383         if (IS_ERR(page))
 384                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 385                         (int)PAGE_CACHE_SIZE,
 386                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 387                         PTR_ERR(page));
 388         return page;
 389 }
 390
 391 /*
 392  * bitmap file superblock operations
 393  */
 394
 395 /* update the event counter and sync the superblock to disk */
 396 int bitmap_update_sb(struct bitmap *bitmap)
 397 {
 398         bitmap_super_t *sb;
 399         unsigned long flags;
 400
 401         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 402                 return 0;
 403         spin_lock_irqsave(&bitmap->lock, flags);
 404         if (!bitmap->sb_page) { /* no superblock */
 405                 spin_unlock_irqrestore(&bitmap->lock, flags);
 406                 return 0;
 407         }
 408         spin_unlock_irqrestore(&bitmap->lock, flags);
 409         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 410         sb->events = cpu_to_le64(bitmap->mddev->events);
 411         if (!bitmap->mddev->degraded)
 412                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 413         kunmap(bitmap->sb_page);
 414         return write_page(bitmap, bitmap->sb_page, 1);
 415 }
 416
 417 /* print out the bitmap file superblock */
 418 void bitmap_print_sb(struct bitmap *bitmap)
 419 {
 420         bitmap_super_t *sb;
 421
 422         if (!bitmap || !bitmap->sb_page)
 423                 return;
 424         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 425         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 426         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 427         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 428         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 429                                         *(__u32 *)(sb->uuid+0),
 430                                         *(__u32 *)(sb->uuid+4),
 431                                         *(__u32 *)(sb->uuid+8),
 432                                         *(__u32 *)(sb->uuid+12));
 433         printk(KERN_DEBUG "        events: %llu\n",
 434                         (unsigned long long) le64_to_cpu(sb->events));
 435         printk(KERN_DEBUG "events cleared: %llu\n",
 436                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 437         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 438         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 439         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 440         printk(KERN_DEBUG "     sync size: %llu KB\n",
 441                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 442         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 443         kunmap(bitmap->sb_page);
 444 }
 445
 446 /* read the superblock from the bitmap file and initialize some bitmap fields */
 447 static int bitmap_read_sb(struct bitmap *bitmap)
 448 {
 449         char *reason = NULL;
 450         bitmap_super_t *sb;
 451         unsigned long chunksize, daemon_sleep, write_behind;
 452         unsigned long bytes_read;
 453         unsigned long long events;
 454         int err = -EINVAL;
 455
 456         /* page 0 is the superblock, read it... */
 457         if (bitmap->file)
 458                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 459         else {
 460                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 461                 bytes_read = PAGE_SIZE;
 462         }
 463         if (IS_ERR(bitmap->sb_page)) {
 464                 err = PTR_ERR(bitmap->sb_page);
 465                 bitmap->sb_page = NULL;
 466                 return err;
 467         }
 468
 469         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 470
 471         if (bytes_read < sizeof(*sb)) { /* short read */
 472                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 473                         bmname(bitmap));
 474                 err = -ENOSPC;
 475                 goto out;
 476         }
 477
 478         chunksize = le32_to_cpu(sb->chunksize);
 479         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 480         write_behind = le32_to_cpu(sb->write_behind);
 481
 482         /* verify that the bitmap-specific fields are valid */
 483         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 484                 reason = "bad magic";
 485         else if (sb->version != cpu_to_le32(BITMAP_MAJOR))
 486                 reason = "unrecognized superblock version";
 487         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 488                 reason = "bitmap chunksize out of range (512B - 4MB)";
 489         else if ((1 << ffz(~chunksize)) != chunksize)
 490                 reason = "bitmap chunksize not a power of 2";
 491         else if (daemon_sleep < 1 || daemon_sleep > 15)
 492                 reason = "daemon sleep period out of range (1-15s)";
 493         else if (write_behind > COUNTER_MAX)
 494                 reason = "write-behind limit out of range (0 - 16383)";
 495         if (reason) {
 496                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 497                         bmname(bitmap), reason);
 498                 goto out;
 499         }
 500
 501         /* keep the array size field of the bitmap superblock up to date */
 502         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 503
 504         if (!bitmap->mddev->persistent)
 505                 goto success;
 506
 507         /*
 508          * if we have a persistent array superblock, compare the
 509          * bitmap's UUID and event counter to the mddev's
 510          */
 511         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 512                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 513                         bmname(bitmap));
 514                 goto out;
 515         }
 516         events = le64_to_cpu(sb->events);
 517         if (events < bitmap->mddev->events) {
 518                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 519                         "-- forcing full recovery\n", bmname(bitmap), events,
 520                         (unsigned long long) bitmap->mddev->events);
 521                 sb->state |= BITMAP_STALE;
 522         }
 523 success:
 524         /* assign fields using values from superblock */
 525         bitmap->chunksize = chunksize;
 526         bitmap->daemon_sleep = daemon_sleep;
 527         bitmap->daemon_lastrun = jiffies;
 528         bitmap->max_write_behind = write_behind;
 529         bitmap->flags |= sb->state;
 530         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 531         if (sb->state & BITMAP_STALE)
 532                 bitmap->events_cleared = bitmap->mddev->events;
 533         err = 0;
 534 out:
 535         kunmap(bitmap->sb_page);
 536         if (err)
 537                 bitmap_print_sb(bitmap);
 538         return err;
 539 }
 540
 541 enum bitmap_mask_op {
 542         MASK_SET,
 543         MASK_UNSET
 544 };
 545
 546 /* record the state of the bitmap in the superblock */
 547 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 548                                 enum bitmap_mask_op op)
 549 {
 550         bitmap_super_t *sb;
 551         unsigned long flags;
 552
 553         spin_lock_irqsave(&bitmap->lock, flags);
 554         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 555                 spin_unlock_irqrestore(&bitmap->lock, flags);
 556                 return;
 557         }
 558         page_cache_get(bitmap->sb_page);
 559         spin_unlock_irqrestore(&bitmap->lock, flags);
 560         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 561         switch (op) {
 562                 case MASK_SET: sb->state |= bits;
 563                                 break;
 564                 case MASK_UNSET: sb->state &= ~bits;
 565                                 break;
 566                 default: BUG();
 567         }
 568         kunmap(bitmap->sb_page);
 569         page_cache_release(bitmap->sb_page);
 570 }
 571
 572 /*
 573  * general bitmap file operations
 574  */
 575
 576 /* calculate the index of the page that contains this bit */
 577 static inline unsigned long file_page_index(unsigned long chunk)
 578 {
 579         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 580 }
 581
 582 /* calculate the (bit) offset of this bit within a page */
 583 static inline unsigned long file_page_offset(unsigned long chunk)
 584 {
 585         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 586 }
 587
 588 /*
 589  * return a pointer to the page in the filemap that contains the given bit
 590  *
 591  * this lookup is complicated by the fact that the bitmap sb might be exactly
 592  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 593  * 0 or page 1
 594  */
 595 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 596                                         unsigned long chunk)
 597 {
 598         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 599 }
 600
 601
 602 static void bitmap_file_unmap(struct bitmap *bitmap)
 603 {
 604         struct page **map, *sb_page;
 605         unsigned long *attr;
 606         int pages;
 607         unsigned long flags;
 608
 609         spin_lock_irqsave(&bitmap->lock, flags);
 610         map = bitmap->filemap;
 611         bitmap->filemap = NULL;
 612         attr = bitmap->filemap_attr;
 613         bitmap->filemap_attr = NULL;
 614         pages = bitmap->file_pages;
 615         bitmap->file_pages = 0;
 616         sb_page = bitmap->sb_page;
 617         bitmap->sb_page = NULL;
 618         spin_unlock_irqrestore(&bitmap->lock, flags);
 619
 620         while (pages--)
 621                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 622                         page_cache_release(map[pages]);
 623         kfree(map);
 624         kfree(attr);
 625
 626         if (sb_page)
 627                 page_cache_release(sb_page);
 628 }
 629
 630 static void bitmap_stop_daemon(struct bitmap *bitmap);
 631
 632 /* dequeue the next item in a page list -- don't call from irq context */
 633 static struct page_list *dequeue_page(struct bitmap *bitmap)
 634 {
 635         struct page_list *item = NULL;
 636         struct list_head *head = &bitmap->complete_pages;
 637
 638         spin_lock(&bitmap->write_lock);
 639         if (list_empty(head))
 640                 goto out;
 641         item = list_entry(head->prev, struct page_list, list);
 642         list_del(head->prev);
 643 out:
 644         spin_unlock(&bitmap->write_lock);
 645         return item;
 646 }
 647
 648 static void drain_write_queues(struct bitmap *bitmap)
 649 {
 650         struct page_list *item;
 651
 652         while ((item = dequeue_page(bitmap))) {
 653                 /* don't bother to wait */
 654                 page_cache_release(item->page);
 655                 mempool_free(item, bitmap->write_pool);
 656         }
 657
 658         wake_up(&bitmap->write_wait);
 659 }
 660
 661 static void bitmap_file_put(struct bitmap *bitmap)
 662 {
 663         struct file *file;
 664         struct inode *inode;
 665         unsigned long flags;
 666
 667         spin_lock_irqsave(&bitmap->lock, flags);
 668         file = bitmap->file;
 669         bitmap->file = NULL;
 670         spin_unlock_irqrestore(&bitmap->lock, flags);
 671
 672         bitmap_stop_daemon(bitmap);
 673
 674         drain_write_queues(bitmap);
 675
 676         bitmap_file_unmap(bitmap);
 677
 678         if (file) {
 679                 inode = file->f_mapping->host;
 680                 spin_lock(&inode->i_lock);
 681                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 682                 spin_unlock(&inode->i_lock);
 683                 fput(file);
 684         }
 685 }
 686
 687
 688 /*
 689  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 690  * then it is no longer reliable, so we stop using it and we mark the file
 691  * as failed in the superblock
 692  */
 693 static void bitmap_file_kick(struct bitmap *bitmap)
 694 {
 695         char *path, *ptr = NULL;
 696
 697         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 698         bitmap_update_sb(bitmap);
 699
 700         if (bitmap->file) {
 701                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 702                 if (path)
 703                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 704
 705                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 706                        bmname(bitmap), ptr ? ptr : "");
 707
 708                 kfree(path);
 709         }
 710
 711         bitmap_file_put(bitmap);
 712
 713         return;
 714 }
 715
 716 enum bitmap_page_attr {
 717         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 718         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 719         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 720 };
 721
 722 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 723                                 enum bitmap_page_attr attr)
 724 {
 725         bitmap->filemap_attr[page->index] |= attr;
 726 }
 727
 728 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 729                                 enum bitmap_page_attr attr)
 730 {
 731         bitmap->filemap_attr[page->index] &= ~attr;
 732 }
 733
 734 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 735 {
 736         return bitmap->filemap_attr[page->index];
 737 }
 738
 739 /*
 740  * bitmap_file_set_bit -- called before performing a write to the md device
 741  * to set (and eventually sync) a particular bit in the bitmap file
 742  *
 743  * we set the bit immediately, then we record the page number so that
 744  * when an unplug occurs, we can flush the dirty pages out to disk
 745  */
 746 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 747 {
 748         unsigned long bit;
 749         struct page *page;
 750         void *kaddr;
 751         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 752
 753         if (!bitmap->filemap) {
 754                 return;
 755         }
 756
 757         page = filemap_get_page(bitmap, chunk);
 758         bit = file_page_offset(chunk);
 759
 760
 761         /* make sure the page stays cached until it gets written out */
 762         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 763                 page_cache_get(page);
 764
 765         /* set the bit */
 766         kaddr = kmap_atomic(page, KM_USER0);
 767         set_bit(bit, kaddr);
 768         kunmap_atomic(kaddr, KM_USER0);
 769         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 770
 771         /* record page number so it gets flushed to disk when unplug occurs */
 772         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 773
 774 }
 775
 776 /* this gets called when the md device is ready to unplug its underlying
 777  * (slave) device queues -- before we let any writes go down, we need to
 778  * sync the dirty pages of the bitmap file to disk */
 779 int bitmap_unplug(struct bitmap *bitmap)
 780 {
 781         unsigned long i, attr, flags;
 782         struct page *page;
 783         int wait = 0;
 784         int err;
 785
 786         if (!bitmap)
 787                 return 0;
 788
 789         /* look at each page to see if there are any set bits that need to be
 790          * flushed out to disk */
 791         for (i = 0; i < bitmap->file_pages; i++) {
 792                 spin_lock_irqsave(&bitmap->lock, flags);
 793                 if (!bitmap->filemap) {
 794                         spin_unlock_irqrestore(&bitmap->lock, flags);
 795                         return 0;
 796                 }
 797                 page = bitmap->filemap[i];
 798                 attr = get_page_attr(bitmap, page);
 799                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 800                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 801                 if ((attr & BITMAP_PAGE_DIRTY))
 802                         wait = 1;
 803                 spin_unlock_irqrestore(&bitmap->lock, flags);
 804
 805                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 806                         err = write_page(bitmap, page, 0);
 807                         if (err == -EAGAIN) {
 808                                 if (attr & BITMAP_PAGE_DIRTY)
 809                                         err = write_page(bitmap, page, 1);
 810                                 else
 811                                         err = 0;
 812                         }
 813                         if (err)
 814                                 return 1;
 815                 }
 816         }
 817         if (wait) { /* if any writes were performed, we need to wait on them */
 818                 if (bitmap->file) {
 819                         spin_lock_irq(&bitmap->write_lock);
 820                         wait_event_lock_irq(bitmap->write_wait,
 821                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 822                                             wake_up_process(bitmap->writeback_daemon->tsk));
 823                         spin_unlock_irq(&bitmap->write_lock);
 824                 } else
 825                         wait_event(bitmap->mddev->sb_wait,
 826                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 827         }
 828         return 0;
 829 }
 830
 831 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 832 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 833  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 834  * memory mapping of the bitmap file
 835  * Special cases:
 836  *   if there's no bitmap file, or if the bitmap file had been
 837  *   previously kicked from the array, we mark all the bits as
 838  *   1's in order to cause a full resync.
 839  *
 840  * We ignore all bits for sectors that end earlier than 'start'.
 841  * This is used when reading an out-of-date bitmap...
 842  */
 843 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 844 {
 845         unsigned long i, chunks, index, oldindex, bit;
 846         struct page *page = NULL, *oldpage = NULL;
 847         unsigned long num_pages, bit_cnt = 0;
 848         struct file *file;
 849         unsigned long bytes, offset, dummy;
 850         int outofdate;
 851         int ret = -ENOSPC;
 852
 853         chunks = bitmap->chunks;
 854         file = bitmap->file;
 855
 856         BUG_ON(!file && !bitmap->offset);
 857
 858 #ifdef INJECT_FAULTS_3
 859         outofdate = 1;
 860 #else
 861         outofdate = bitmap->flags & BITMAP_STALE;
 862 #endif
 863         if (outofdate)
 864                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 865                         "recovery\n", bmname(bitmap));
 866
 867         bytes = (chunks + 7) / 8;
 868
 869         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 870
 871         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 872                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 873                         bmname(bitmap),
 874                         (unsigned long) i_size_read(file->f_mapping->host),
 875                         bytes + sizeof(bitmap_super_t));
 876                 goto out;
 877         }
 878
 879         ret = -ENOMEM;
 880
 881         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 882         if (!bitmap->filemap)
 883                 goto out;
 884
 885         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 886         if (!bitmap->filemap_attr)
 887                 goto out;
 888
 889         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 890
 891         oldindex = ~0L;
 892
 893         for (i = 0; i < chunks; i++) {
 894                 index = file_page_index(i);
 895                 bit = file_page_offset(i);
 896                 if (index != oldindex) { /* this is a new page, read it in */
 897                         /* unmap the old page, we're done with it */
 898                         if (oldpage != NULL)
 899                                 kunmap(oldpage);
 900                         if (index == 0) {
 901                                 /*
 902                                  * if we're here then the superblock page
 903                                  * contains some bits (PAGE_SIZE != sizeof sb)
 904                                  * we've already read it in, so just use it
 905                                  */
 906                                 page = bitmap->sb_page;
 907                                 offset = sizeof(bitmap_super_t);
 908                         } else if (file) {
 909                                 page = read_page(file, index, &dummy);
 910                                 offset = 0;
 911                         } else {
 912                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 913                                 offset = 0;
 914                         }
 915                         if (IS_ERR(page)) { /* read error */
 916                                 ret = PTR_ERR(page);
 917                                 goto out;
 918                         }
 919
 920                         oldindex = index;
 921                         oldpage = page;
 922                         kmap(page);
 923
 924                         if (outofdate) {
 925                                 /*
 926                                  * if bitmap is out of date, dirty the
 927                                  * whole page and write it out
 928                                  */
 929                                 memset(page_address(page) + offset, 0xff,
 930                                        PAGE_SIZE - offset);
 931                                 ret = write_page(bitmap, page, 1);
 932                                 if (ret) {
 933                                         kunmap(page);
 934                                         /* release, page not in filemap yet */
 935                                         page_cache_release(page);
 936                                         goto out;
 937                                 }
 938                         }
 939
 940                         bitmap->filemap[bitmap->file_pages++] = page;
 941                 }
 942                 if (test_bit(bit, page_address(page))) {
 943                         /* if the disk bit is set, set the memory bit */
 944                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 945                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 946                                 );
 947                         bit_cnt++;
 948                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 949                 }
 950         }
 951
 952         /* everything went OK */
 953         ret = 0;
 954         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 955
 956         if (page) /* unmap the last page */
 957                 kunmap(page);
 958
 959         if (bit_cnt) { /* Kick recovery if any bits were set */
 960                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 961                 md_wakeup_thread(bitmap->mddev->thread);
 962         }
 963
 964 out:
 965         printk(KERN_INFO "%s: bitmap initialized from disk: "
 966                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 967                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 968
 969         return ret;
 970 }
 971
 972 void bitmap_write_all(struct bitmap *bitmap)
 973 {
 974         /* We don't actually write all bitmap blocks here,
 975          * just flag them as needing to be written
 976          */
 977
 978         unsigned long chunks = bitmap->chunks;
 979         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 980         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 981         while (num_pages--)
 982                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 983 }
 984
 985
 986 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 987 {
 988         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 989         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 990         bitmap->bp[page].count += inc;
 991 /*
 992         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 993                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 994 */
 995         bitmap_checkfree(bitmap, page);
 996 }
 997 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 998                                             sector_t offset, int *blocks,
 999                                             int create);
1000
1001 /*
1002  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1003  *                      out to disk
1004  */
1005
1006 int bitmap_daemon_work(struct bitmap *bitmap)
1007 {
1008         unsigned long j;
1009         unsigned long flags;
1010         struct page *page = NULL, *lastpage = NULL;
1011         int err = 0;
1012         int blocks;
1013         int attr;
1014
1015         if (bitmap == NULL)
1016                 return 0;
1017         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1018                 return 0;
1019         bitmap->daemon_lastrun = jiffies;
1020
1021         for (j = 0; j < bitmap->chunks; j++) {
1022                 bitmap_counter_t *bmc;
1023                 spin_lock_irqsave(&bitmap->lock, flags);
1024                 if (!bitmap->filemap) {
1025                         /* error or shutdown */
1026                         spin_unlock_irqrestore(&bitmap->lock, flags);
1027                         break;
1028                 }
1029
1030                 page = filemap_get_page(bitmap, j);
1031
1032                 if (page != lastpage) {
1033                         /* skip this page unless it's marked as needing cleaning */
1034                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1035                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1036                                         page_cache_get(page);
1037                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1038                                 }
1039                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1040                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1041                                         switch (write_page(bitmap, page, 0)) {
1042                                         case -EAGAIN:
1043                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1044                                                 break;
1045                                         case 0:
1046                                                 break;
1047                                         default:
1048                                                 bitmap_file_kick(bitmap);
1049                                         }
1050                                         page_cache_release(page);
1051                                 }
1052                                 continue;
1053                         }
1054
1055                         /* grab the new page, sync and release the old */
1056                         page_cache_get(page);
1057                         if (lastpage != NULL) {
1058                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1059                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1060                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1061                                         err = write_page(bitmap, lastpage, 0);
1062                                         if (err == -EAGAIN) {
1063                                                 err = 0;
1064                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1065                                         }
1066                                 } else {
1067                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1068                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1069                                 }
1070                                 kunmap(lastpage);
1071                                 page_cache_release(lastpage);
1072                                 if (err)
1073                                         bitmap_file_kick(bitmap);
1074                         } else
1075                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1076                         lastpage = page;
1077                         kmap(page);
1078 /*
1079                         printk("bitmap clean at page %lu\n", j);
1080 */
1081                         spin_lock_irqsave(&bitmap->lock, flags);
1082                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1083                 }
1084                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1085                                         &blocks, 0);
1086                 if (bmc) {
1087 /*
1088   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1089 */
1090                         if (*bmc == 2) {
1091                                 *bmc=1; /* maybe clear the bit next time */
1092                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1093                         } else if (*bmc == 1) {
1094                                 /* we can clear the bit */
1095                                 *bmc = 0;
1096                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1097                                                   -1);
1098
1099                                 /* clear the bit */
1100                                 clear_bit(file_page_offset(j), page_address(page));
1101                         }
1102                 }
1103                 spin_unlock_irqrestore(&bitmap->lock, flags);
1104         }
1105
1106         /* now sync the final page */
1107         if (lastpage != NULL) {
1108                 kunmap(lastpage);
1109                 spin_lock_irqsave(&bitmap->lock, flags);
1110                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1111                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1112                         spin_unlock_irqrestore(&bitmap->lock, flags);
1113                         err = write_page(bitmap, lastpage, 0);
1114                         if (err == -EAGAIN) {
1115                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1116                                 err = 0;
1117                         }
1118                 } else {
1119                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1120                         spin_unlock_irqrestore(&bitmap->lock, flags);
1121                 }
1122
1123                 page_cache_release(lastpage);
1124         }
1125
1126         return err;
1127 }
1128
1129 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1130 {
1131         mdk_thread_t *dmn;
1132         unsigned long flags;
1133
1134         /* if no one is waiting on us, we'll free the md thread struct
1135          * and exit, otherwise we let the waiter clean things up */
1136         spin_lock_irqsave(&bitmap->lock, flags);
1137         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1138                 *daemon = NULL;
1139                 spin_unlock_irqrestore(&bitmap->lock, flags);
1140                 kfree(dmn);
1141                 complete_and_exit(NULL, 0); /* do_exit not exported */
1142         }
1143         spin_unlock_irqrestore(&bitmap->lock, flags);
1144 }
1145
1146 static void bitmap_writeback_daemon(mddev_t *mddev)
1147 {
1148         struct bitmap *bitmap = mddev->bitmap;
1149         struct page *page;
1150         struct page_list *item;
1151         int err = 0;
1152
1153         if (signal_pending(current)) {
1154                 printk(KERN_INFO
1155                        "%s: bitmap writeback daemon got signal, exiting...\n",
1156                        bmname(bitmap));
1157                 err = -EINTR;
1158                 goto out;
1159         }
1160         if (bitmap == NULL)
1161                 /* about to be stopped. */
1162                 return;
1163
1164         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1165         /* wait on bitmap page writebacks */
1166         while ((item = dequeue_page(bitmap))) {
1167                 page = item->page;
1168                 mempool_free(item, bitmap->write_pool);
1169                 PRINTK("wait on page writeback: %p\n", page);
1170                 wait_on_page_writeback(page);
1171                 PRINTK("finished page writeback: %p\n", page);
1172
1173                 err = PageError(page);
1174                 page_cache_release(page);
1175                 if (err) {
1176                         printk(KERN_WARNING "%s: bitmap file writeback "
1177                                "failed (page %lu): %d\n",
1178                                bmname(bitmap), page->index, err);
1179                         bitmap_file_kick(bitmap);
1180                         goto out;
1181                 }
1182         }
1183  out:
1184         wake_up(&bitmap->write_wait);
1185         if (err) {
1186                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1187                        bmname(bitmap), err);
1188                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1189         }
1190 }
1191
1192 static mdk_thread_t *bitmap_start_daemon(struct bitmap *bitmap,
1193                                 void (*func)(mddev_t *), char *name)
1194 {
1195         mdk_thread_t *daemon;
1196         char namebuf[32];
1197
1198 #ifdef INJECT_FATAL_FAULT_2
1199         daemon = NULL;
1200 #else
1201         sprintf(namebuf, "%%s_%s", name);
1202         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1203 #endif
1204         if (!daemon) {
1205                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1206                         bmname(bitmap));
1207                 return ERR_PTR(-ECHILD);
1208         }
1209
1210         md_wakeup_thread(daemon); /* start it running */
1211
1212         PRINTK("%s: %s daemon (pid %d) started...\n",
1213                 bmname(bitmap), name, daemon->tsk->pid);
1214
1215         return daemon;
1216 }
1217
1218 static void bitmap_stop_daemon(struct bitmap *bitmap)
1219 {
1220         /* the daemon can't stop itself... it'll just exit instead... */
1221         if (bitmap->writeback_daemon && ! IS_ERR(bitmap->writeback_daemon) &&
1222             current->pid != bitmap->writeback_daemon->tsk->pid) {
1223                 mdk_thread_t *daemon;
1224                 unsigned long flags;
1225
1226                 spin_lock_irqsave(&bitmap->lock, flags);
1227                 daemon = bitmap->writeback_daemon;
1228                 bitmap->writeback_daemon = NULL;
1229                 spin_unlock_irqrestore(&bitmap->lock, flags);
1230                 if (daemon && ! IS_ERR(daemon))
1231                         md_unregister_thread(daemon); /* destroy the thread */
1232         }
1233 }
1234
1235 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1236                                             sector_t offset, int *blocks,
1237                                             int create)
1238 {
1239         /* If 'create', we might release the lock and reclaim it.
1240          * The lock must have been taken with interrupts enabled.
1241          * If !create, we don't release the lock.
1242          */
1243         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1244         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1245         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1246         sector_t csize;
1247
1248         if (bitmap_checkpage(bitmap, page, create) < 0) {
1249                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1250                 *blocks = csize - (offset & (csize- 1));
1251                 return NULL;
1252         }
1253         /* now locked ... */
1254
1255         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1256                 /* should we use the first or second counter field
1257                  * of the hijacked pointer? */
1258                 int hi = (pageoff > PAGE_COUNTER_MASK);
1259                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1260                                           PAGE_COUNTER_SHIFT - 1);
1261                 *blocks = csize - (offset & (csize- 1));
1262                 return  &((bitmap_counter_t *)
1263                           &bitmap->bp[page].map)[hi];
1264         } else { /* page is allocated */
1265                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1266                 *blocks = csize - (offset & (csize- 1));
1267                 return (bitmap_counter_t *)
1268                         &(bitmap->bp[page].map[pageoff]);
1269         }
1270 }
1271
1272 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1273 {
1274         if (!bitmap) return 0;
1275
1276         if (behind) {
1277                 atomic_inc(&bitmap->behind_writes);
1278                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1279                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1280         }
1281
1282         while (sectors) {
1283                 int blocks;
1284                 bitmap_counter_t *bmc;
1285
1286                 spin_lock_irq(&bitmap->lock);
1287                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1288                 if (!bmc) {
1289                         spin_unlock_irq(&bitmap->lock);
1290                         return 0;
1291                 }
1292
1293                 switch(*bmc) {
1294                 case 0:
1295                         bitmap_file_set_bit(bitmap, offset);
1296                         bitmap_count_page(bitmap,offset, 1);
1297                         blk_plug_device(bitmap->mddev->queue);
1298                         /* fall through */
1299                 case 1:
1300                         *bmc = 2;
1301                 }
1302                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1303                 (*bmc)++;
1304
1305                 spin_unlock_irq(&bitmap->lock);
1306
1307                 offset += blocks;
1308                 if (sectors > blocks)
1309                         sectors -= blocks;
1310                 else sectors = 0;
1311         }
1312         return 0;
1313 }
1314
1315 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1316                      int success, int behind)
1317 {
1318         if (!bitmap) return;
1319         if (behind) {
1320                 atomic_dec(&bitmap->behind_writes);
1321                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1322                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1323         }
1324
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 static void bitmap_free(struct bitmap *bitmap)
1488 {
1489         unsigned long k, pages;
1490         struct bitmap_page *bp;
1491
1492         if (!bitmap) /* there was no bitmap */
1493                 return;
1494
1495         /* release the bitmap file and kill the daemon */
1496         bitmap_file_put(bitmap);
1497
1498         bp = bitmap->bp;
1499         pages = bitmap->pages;
1500
1501         /* free all allocated memory */
1502
1503         mempool_destroy(bitmap->write_pool);
1504
1505         if (bp) /* deallocate the page memory */
1506                 for (k = 0; k < pages; k++)
1507                         if (bp[k].map && !bp[k].hijacked)
1508                                 kfree(bp[k].map);
1509         kfree(bp);
1510         kfree(bitmap);
1511 }
1512 void bitmap_destroy(mddev_t *mddev)
1513 {
1514         struct bitmap *bitmap = mddev->bitmap;
1515
1516         if (!bitmap) /* there was no bitmap */
1517                 return;
1518
1519         mddev->bitmap = NULL; /* disconnect from the md device */
1520
1521         bitmap_free(bitmap);
1522 }
1523
1524 /*
1525  * initialize the bitmap structure
1526  * if this returns an error, bitmap_destroy must be called to do clean up
1527  */
1528 int bitmap_create(mddev_t *mddev)
1529 {
1530         struct bitmap *bitmap;
1531         unsigned long blocks = mddev->resync_max_sectors;
1532         unsigned long chunks;
1533         unsigned long pages;
1534         struct file *file = mddev->bitmap_file;
1535         int err;
1536         sector_t start;
1537
1538         BUG_ON(sizeof(bitmap_super_t) != 256);
1539
1540         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1541                 return 0;
1542
1543         BUG_ON(file && mddev->bitmap_offset);
1544
1545         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1546         if (!bitmap)
1547                 return -ENOMEM;
1548
1549         memset(bitmap, 0, sizeof(*bitmap));
1550
1551         spin_lock_init(&bitmap->lock);
1552         bitmap->mddev = mddev;
1553
1554         spin_lock_init(&bitmap->write_lock);
1555         INIT_LIST_HEAD(&bitmap->complete_pages);
1556         init_waitqueue_head(&bitmap->write_wait);
1557         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1558                                 write_pool_free, NULL);
1559         err = -ENOMEM;
1560         if (!bitmap->write_pool)
1561                 goto error;
1562
1563         bitmap->file = file;
1564         bitmap->offset = mddev->bitmap_offset;
1565         if (file) get_file(file);
1566         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1567         err = bitmap_read_sb(bitmap);
1568         if (err)
1569                 goto error;
1570
1571         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1572                                         sizeof(bitmap->chunksize));
1573
1574         /* now that chunksize and chunkshift are set, we can use these macros */
1575         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1576                         CHUNK_BLOCK_RATIO(bitmap);
1577         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1578
1579         BUG_ON(!pages);
1580
1581         bitmap->chunks = chunks;
1582         bitmap->pages = pages;
1583         bitmap->missing_pages = pages;
1584         bitmap->counter_bits = COUNTER_BITS;
1585
1586         bitmap->syncchunk = ~0UL;
1587
1588 #ifdef INJECT_FATAL_FAULT_1
1589         bitmap->bp = NULL;
1590 #else
1591         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1592 #endif
1593         err = -ENOMEM;
1594         if (!bitmap->bp)
1595                 goto error;
1596         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1597
1598         bitmap->flags |= BITMAP_ACTIVE;
1599
1600         /* now that we have some pages available, initialize the in-memory
1601          * bitmap from the on-disk bitmap */
1602         start = 0;
1603         if (mddev->degraded == 0
1604             || bitmap->events_cleared == mddev->events)
1605                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1606                 start = mddev->recovery_cp;
1607         err = bitmap_init_from_disk(bitmap, start);
1608
1609         if (err)
1610                 goto error;
1611
1612         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1613                 pages, bmname(bitmap));
1614
1615         mddev->bitmap = bitmap;
1616
1617         if (file)
1618                 /* kick off the bitmap writeback daemon */
1619                 bitmap->writeback_daemon =
1620                         bitmap_start_daemon(bitmap,
1621                                             bitmap_writeback_daemon,
1622                                             "bitmap_wb");
1623
1624         if (IS_ERR(bitmap->writeback_daemon))
1625                 return PTR_ERR(bitmap->writeback_daemon);
1626         return bitmap_update_sb(bitmap);
1627
1628  error:
1629         bitmap_free(bitmap);
1630         return err;
1631 }
1632
1633 /* the bitmap API -- for raid personalities */
1634 EXPORT_SYMBOL(bitmap_startwrite);
1635 EXPORT_SYMBOL(bitmap_endwrite);
1636 EXPORT_SYMBOL(bitmap_start_sync);
1637 EXPORT_SYMBOL(bitmap_end_sync);
1638 EXPORT_SYMBOL(bitmap_unplug);
1639 EXPORT_SYMBOL(bitmap_close_sync);
1640 EXPORT_SYMBOL(bitmap_daemon_work);