Pileus Git - ~andy/linux/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/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 (! rdev->in_sync || rdev->faulty)
 275                         continue;
 276
 277                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 278
 279                 if (sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ)) {
 280                         page->index = index;
 281                         return page;
 282                 }
 283         }
 284         return ERR_PTR(-EIO);
 285
 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         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 441         kunmap(bitmap->sb_page);
 442 }
 443
 444 /* read the superblock from the bitmap file and initialize some bitmap fields */
 445 static int bitmap_read_sb(struct bitmap *bitmap)
 446 {
 447         char *reason = NULL;
 448         bitmap_super_t *sb;
 449         unsigned long chunksize, daemon_sleep, write_behind;
 450         unsigned long bytes_read;
 451         unsigned long long events;
 452         int err = -EINVAL;
 453
 454         /* page 0 is the superblock, read it... */
 455         if (bitmap->file)
 456                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 457         else {
 458                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 459                 bytes_read = PAGE_SIZE;
 460         }
 461         if (IS_ERR(bitmap->sb_page)) {
 462                 err = PTR_ERR(bitmap->sb_page);
 463                 bitmap->sb_page = NULL;
 464                 return err;
 465         }
 466
 467         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 468
 469         if (bytes_read < sizeof(*sb)) { /* short read */
 470                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 471                         bmname(bitmap));
 472                 err = -ENOSPC;
 473                 goto out;
 474         }
 475
 476         chunksize = le32_to_cpu(sb->chunksize);
 477         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 478         write_behind = le32_to_cpu(sb->write_behind);
 479
 480         /* verify that the bitmap-specific fields are valid */
 481         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 482                 reason = "bad magic";
 483         else if (sb->version != cpu_to_le32(BITMAP_MAJOR))
 484                 reason = "unrecognized superblock version";
 485         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 486                 reason = "bitmap chunksize out of range (512B - 4MB)";
 487         else if ((1 << ffz(~chunksize)) != chunksize)
 488                 reason = "bitmap chunksize not a power of 2";
 489         else if (daemon_sleep < 1 || daemon_sleep > 15)
 490                 reason = "daemon sleep period out of range (1-15s)";
 491         else if (write_behind > COUNTER_MAX)
 492                 reason = "write-behind limit out of range (0 - 16383)";
 493         if (reason) {
 494                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 495                         bmname(bitmap), reason);
 496                 goto out;
 497         }
 498
 499         /* keep the array size field of the bitmap superblock up to date */
 500         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 501
 502         if (!bitmap->mddev->persistent)
 503                 goto success;
 504
 505         /*
 506          * if we have a persistent array superblock, compare the
 507          * bitmap's UUID and event counter to the mddev's
 508          */
 509         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 510                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 511                         bmname(bitmap));
 512                 goto out;
 513         }
 514         events = le64_to_cpu(sb->events);
 515         if (events < bitmap->mddev->events) {
 516                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 517                         "-- forcing full recovery\n", bmname(bitmap), events,
 518                         (unsigned long long) bitmap->mddev->events);
 519                 sb->state |= BITMAP_STALE;
 520         }
 521 success:
 522         /* assign fields using values from superblock */
 523         bitmap->chunksize = chunksize;
 524         bitmap->daemon_sleep = daemon_sleep;
 525         bitmap->daemon_lastrun = jiffies;
 526         bitmap->max_write_behind = write_behind;
 527         bitmap->flags |= sb->state;
 528         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 529         if (sb->state & BITMAP_STALE)
 530                 bitmap->events_cleared = bitmap->mddev->events;
 531         err = 0;
 532 out:
 533         kunmap(bitmap->sb_page);
 534         if (err)
 535                 bitmap_print_sb(bitmap);
 536         return err;
 537 }
 538
 539 enum bitmap_mask_op {
 540         MASK_SET,
 541         MASK_UNSET
 542 };
 543
 544 /* record the state of the bitmap in the superblock */
 545 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 546                                 enum bitmap_mask_op op)
 547 {
 548         bitmap_super_t *sb;
 549         unsigned long flags;
 550
 551         spin_lock_irqsave(&bitmap->lock, flags);
 552         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 553                 spin_unlock_irqrestore(&bitmap->lock, flags);
 554                 return;
 555         }
 556         page_cache_get(bitmap->sb_page);
 557         spin_unlock_irqrestore(&bitmap->lock, flags);
 558         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 559         switch (op) {
 560                 case MASK_SET: sb->state |= bits;
 561                                 break;
 562                 case MASK_UNSET: sb->state &= ~bits;
 563                                 break;
 564                 default: BUG();
 565         }
 566         kunmap(bitmap->sb_page);
 567         page_cache_release(bitmap->sb_page);
 568 }
 569
 570 /*
 571  * general bitmap file operations
 572  */
 573
 574 /* calculate the index of the page that contains this bit */
 575 static inline unsigned long file_page_index(unsigned long chunk)
 576 {
 577         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 578 }
 579
 580 /* calculate the (bit) offset of this bit within a page */
 581 static inline unsigned long file_page_offset(unsigned long chunk)
 582 {
 583         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 584 }
 585
 586 /*
 587  * return a pointer to the page in the filemap that contains the given bit
 588  *
 589  * this lookup is complicated by the fact that the bitmap sb might be exactly
 590  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 591  * 0 or page 1
 592  */
 593 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 594                                         unsigned long chunk)
 595 {
 596         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 597 }
 598
 599
 600 static void bitmap_file_unmap(struct bitmap *bitmap)
 601 {
 602         struct page **map, *sb_page;
 603         unsigned long *attr;
 604         int pages;
 605         unsigned long flags;
 606
 607         spin_lock_irqsave(&bitmap->lock, flags);
 608         map = bitmap->filemap;
 609         bitmap->filemap = NULL;
 610         attr = bitmap->filemap_attr;
 611         bitmap->filemap_attr = NULL;
 612         pages = bitmap->file_pages;
 613         bitmap->file_pages = 0;
 614         sb_page = bitmap->sb_page;
 615         bitmap->sb_page = NULL;
 616         spin_unlock_irqrestore(&bitmap->lock, flags);
 617
 618         while (pages--)
 619                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 620                         page_cache_release(map[pages]);
 621         kfree(map);
 622         kfree(attr);
 623
 624         if (sb_page)
 625                 page_cache_release(sb_page);
 626 }
 627
 628 static void bitmap_stop_daemon(struct bitmap *bitmap);
 629
 630 /* dequeue the next item in a page list -- don't call from irq context */
 631 static struct page_list *dequeue_page(struct bitmap *bitmap)
 632 {
 633         struct page_list *item = NULL;
 634         struct list_head *head = &bitmap->complete_pages;
 635
 636         spin_lock(&bitmap->write_lock);
 637         if (list_empty(head))
 638                 goto out;
 639         item = list_entry(head->prev, struct page_list, list);
 640         list_del(head->prev);
 641 out:
 642         spin_unlock(&bitmap->write_lock);
 643         return item;
 644 }
 645
 646 static void drain_write_queues(struct bitmap *bitmap)
 647 {
 648         struct page_list *item;
 649
 650         while ((item = dequeue_page(bitmap))) {
 651                 /* don't bother to wait */
 652                 page_cache_release(item->page);
 653                 mempool_free(item, bitmap->write_pool);
 654         }
 655
 656         wake_up(&bitmap->write_wait);
 657 }
 658
 659 static void bitmap_file_put(struct bitmap *bitmap)
 660 {
 661         struct file *file;
 662         struct inode *inode;
 663         unsigned long flags;
 664
 665         spin_lock_irqsave(&bitmap->lock, flags);
 666         file = bitmap->file;
 667         bitmap->file = NULL;
 668         spin_unlock_irqrestore(&bitmap->lock, flags);
 669
 670         bitmap_stop_daemon(bitmap);
 671
 672         drain_write_queues(bitmap);
 673
 674         bitmap_file_unmap(bitmap);
 675
 676         if (file) {
 677                 inode = file->f_mapping->host;
 678                 spin_lock(&inode->i_lock);
 679                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 680                 spin_unlock(&inode->i_lock);
 681                 fput(file);
 682         }
 683 }
 684
 685
 686 /*
 687  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 688  * then it is no longer reliable, so we stop using it and we mark the file
 689  * as failed in the superblock
 690  */
 691 static void bitmap_file_kick(struct bitmap *bitmap)
 692 {
 693         char *path, *ptr = NULL;
 694
 695         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 696         bitmap_update_sb(bitmap);
 697
 698         if (bitmap->file) {
 699                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 700                 if (path)
 701                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 702
 703                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 704                        bmname(bitmap), ptr ? ptr : "");
 705
 706                 kfree(path);
 707         }
 708
 709         bitmap_file_put(bitmap);
 710
 711         return;
 712 }
 713
 714 enum bitmap_page_attr {
 715         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 716         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 717         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 718 };
 719
 720 static inline void set_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 void clear_page_attr(struct bitmap *bitmap, struct page *page,
 727                                 enum bitmap_page_attr attr)
 728 {
 729         bitmap->filemap_attr[page->index] &= ~attr;
 730 }
 731
 732 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 733 {
 734         return bitmap->filemap_attr[page->index];
 735 }
 736
 737 /*
 738  * bitmap_file_set_bit -- called before performing a write to the md device
 739  * to set (and eventually sync) a particular bit in the bitmap file
 740  *
 741  * we set the bit immediately, then we record the page number so that
 742  * when an unplug occurs, we can flush the dirty pages out to disk
 743  */
 744 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 745 {
 746         unsigned long bit;
 747         struct page *page;
 748         void *kaddr;
 749         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 750
 751         if (!bitmap->filemap) {
 752                 return;
 753         }
 754
 755         page = filemap_get_page(bitmap, chunk);
 756         bit = file_page_offset(chunk);
 757
 758
 759         /* make sure the page stays cached until it gets written out */
 760         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 761                 page_cache_get(page);
 762
 763         /* set the bit */
 764         kaddr = kmap_atomic(page, KM_USER0);
 765         set_bit(bit, kaddr);
 766         kunmap_atomic(kaddr, KM_USER0);
 767         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 768
 769         /* record page number so it gets flushed to disk when unplug occurs */
 770         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 771
 772 }
 773
 774 /* this gets called when the md device is ready to unplug its underlying
 775  * (slave) device queues -- before we let any writes go down, we need to
 776  * sync the dirty pages of the bitmap file to disk */
 777 int bitmap_unplug(struct bitmap *bitmap)
 778 {
 779         unsigned long i, attr, flags;
 780         struct page *page;
 781         int wait = 0;
 782         int err;
 783
 784         if (!bitmap)
 785                 return 0;
 786
 787         /* look at each page to see if there are any set bits that need to be
 788          * flushed out to disk */
 789         for (i = 0; i < bitmap->file_pages; i++) {
 790                 spin_lock_irqsave(&bitmap->lock, flags);
 791                 if (!bitmap->filemap) {
 792                         spin_unlock_irqrestore(&bitmap->lock, flags);
 793                         return 0;
 794                 }
 795                 page = bitmap->filemap[i];
 796                 attr = get_page_attr(bitmap, page);
 797                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 798                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 799                 if ((attr & BITMAP_PAGE_DIRTY))
 800                         wait = 1;
 801                 spin_unlock_irqrestore(&bitmap->lock, flags);
 802
 803                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 804                         err = write_page(bitmap, page, 0);
 805                         if (err == -EAGAIN) {
 806                                 if (attr & BITMAP_PAGE_DIRTY)
 807                                         err = write_page(bitmap, page, 1);
 808                                 else
 809                                         err = 0;
 810                         }
 811                         if (err)
 812                                 return 1;
 813                 }
 814         }
 815         if (wait) { /* if any writes were performed, we need to wait on them */
 816                 if (bitmap->file) {
 817                         spin_lock_irq(&bitmap->write_lock);
 818                         wait_event_lock_irq(bitmap->write_wait,
 819                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 820                                             wake_up_process(bitmap->writeback_daemon->tsk));
 821                         spin_unlock_irq(&bitmap->write_lock);
 822                 } else
 823                         wait_event(bitmap->mddev->sb_wait,
 824                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 825         }
 826         return 0;
 827 }
 828
 829 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 830 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 831  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 832  * memory mapping of the bitmap file
 833  * Special cases:
 834  *   if there's no bitmap file, or if the bitmap file had been
 835  *   previously kicked from the array, we mark all the bits as
 836  *   1's in order to cause a full resync.
 837  *
 838  * We ignore all bits for sectors that end earlier than 'start'.
 839  * This is used when reading an out-of-date bitmap...
 840  */
 841 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 842 {
 843         unsigned long i, chunks, index, oldindex, bit;
 844         struct page *page = NULL, *oldpage = NULL;
 845         unsigned long num_pages, bit_cnt = 0;
 846         struct file *file;
 847         unsigned long bytes, offset, dummy;
 848         int outofdate;
 849         int ret = -ENOSPC;
 850
 851         chunks = bitmap->chunks;
 852         file = bitmap->file;
 853
 854         BUG_ON(!file && !bitmap->offset);
 855
 856 #ifdef INJECT_FAULTS_3
 857         outofdate = 1;
 858 #else
 859         outofdate = bitmap->flags & BITMAP_STALE;
 860 #endif
 861         if (outofdate)
 862                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 863                         "recovery\n", bmname(bitmap));
 864
 865         bytes = (chunks + 7) / 8;
 866
 867         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 868
 869         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 870                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 871                         bmname(bitmap),
 872                         (unsigned long) i_size_read(file->f_mapping->host),
 873                         bytes + sizeof(bitmap_super_t));
 874                 goto out;
 875         }
 876
 877         ret = -ENOMEM;
 878
 879         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 880         if (!bitmap->filemap)
 881                 goto out;
 882
 883         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 884         if (!bitmap->filemap_attr)
 885                 goto out;
 886
 887         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 888
 889         oldindex = ~0L;
 890
 891         for (i = 0; i < chunks; i++) {
 892                 index = file_page_index(i);
 893                 bit = file_page_offset(i);
 894                 if (index != oldindex) { /* this is a new page, read it in */
 895                         /* unmap the old page, we're done with it */
 896                         if (oldpage != NULL)
 897                                 kunmap(oldpage);
 898                         if (index == 0) {
 899                                 /*
 900                                  * if we're here then the superblock page
 901                                  * contains some bits (PAGE_SIZE != sizeof sb)
 902                                  * we've already read it in, so just use it
 903                                  */
 904                                 page = bitmap->sb_page;
 905                                 offset = sizeof(bitmap_super_t);
 906                         } else if (file) {
 907                                 page = read_page(file, index, &dummy);
 908                                 offset = 0;
 909                         } else {
 910                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 911                                 offset = 0;
 912                         }
 913                         if (IS_ERR(page)) { /* read error */
 914                                 ret = PTR_ERR(page);
 915                                 goto out;
 916                         }
 917
 918                         oldindex = index;
 919                         oldpage = page;
 920                         kmap(page);
 921
 922                         if (outofdate) {
 923                                 /*
 924                                  * if bitmap is out of date, dirty the
 925                                  * whole page and write it out
 926                                  */
 927                                 memset(page_address(page) + offset, 0xff,
 928                                        PAGE_SIZE - offset);
 929                                 ret = write_page(bitmap, page, 1);
 930                                 if (ret) {
 931                                         kunmap(page);
 932                                         /* release, page not in filemap yet */
 933                                         page_cache_release(page);
 934                                         goto out;
 935                                 }
 936                         }
 937
 938                         bitmap->filemap[bitmap->file_pages++] = page;
 939                 }
 940                 if (test_bit(bit, page_address(page))) {
 941                         /* if the disk bit is set, set the memory bit */
 942                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 943                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 944                                 );
 945                         bit_cnt++;
 946                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 947                 }
 948         }
 949
 950         /* everything went OK */
 951         ret = 0;
 952         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 953
 954         if (page) /* unmap the last page */
 955                 kunmap(page);
 956
 957         if (bit_cnt) { /* Kick recovery if any bits were set */
 958                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 959                 md_wakeup_thread(bitmap->mddev->thread);
 960         }
 961
 962 out:
 963         printk(KERN_INFO "%s: bitmap initialized from disk: "
 964                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 965                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 966
 967         return ret;
 968 }
 969
 970 void bitmap_write_all(struct bitmap *bitmap)
 971 {
 972         /* We don't actually write all bitmap blocks here,
 973          * just flag them as needing to be written
 974          */
 975
 976         unsigned long chunks = bitmap->chunks;
 977         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 978         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 979         while (num_pages--)
 980                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 981 }
 982
 983
 984 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 985 {
 986         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 987         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 988         bitmap->bp[page].count += inc;
 989 /*
 990         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 991                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 992 */
 993         bitmap_checkfree(bitmap, page);
 994 }
 995 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 996                                             sector_t offset, int *blocks,
 997                                             int create);
 998
 999 /*
1000  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1001  *                      out to disk
1002  */
1003
1004 int bitmap_daemon_work(struct bitmap *bitmap)
1005 {
1006         unsigned long j;
1007         unsigned long flags;
1008         struct page *page = NULL, *lastpage = NULL;
1009         int err = 0;
1010         int blocks;
1011         int attr;
1012
1013         if (bitmap == NULL)
1014                 return 0;
1015         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1016                 return 0;
1017         bitmap->daemon_lastrun = jiffies;
1018
1019         for (j = 0; j < bitmap->chunks; j++) {
1020                 bitmap_counter_t *bmc;
1021                 spin_lock_irqsave(&bitmap->lock, flags);
1022                 if (!bitmap->filemap) {
1023                         /* error or shutdown */
1024                         spin_unlock_irqrestore(&bitmap->lock, flags);
1025                         break;
1026                 }
1027
1028                 page = filemap_get_page(bitmap, j);
1029
1030                 if (page != lastpage) {
1031                         /* skip this page unless it's marked as needing cleaning */
1032                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1033                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1034                                         page_cache_get(page);
1035                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1036                                 }
1037                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1038                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1039                                         switch (write_page(bitmap, page, 0)) {
1040                                         case -EAGAIN:
1041                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1042                                                 break;
1043                                         case 0:
1044                                                 break;
1045                                         default:
1046                                                 bitmap_file_kick(bitmap);
1047                                         }
1048                                         page_cache_release(page);
1049                                 }
1050                                 continue;
1051                         }
1052
1053                         /* grab the new page, sync and release the old */
1054                         page_cache_get(page);
1055                         if (lastpage != NULL) {
1056                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1057                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1058                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1059                                         err = write_page(bitmap, lastpage, 0);
1060                                         if (err == -EAGAIN) {
1061                                                 err = 0;
1062                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1063                                         }
1064                                 } else {
1065                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1066                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1067                                 }
1068                                 kunmap(lastpage);
1069                                 page_cache_release(lastpage);
1070                                 if (err)
1071                                         bitmap_file_kick(bitmap);
1072                         } else
1073                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1074                         lastpage = page;
1075                         kmap(page);
1076 /*
1077                         printk("bitmap clean at page %lu\n", j);
1078 */
1079                         spin_lock_irqsave(&bitmap->lock, flags);
1080                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1081                 }
1082                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1083                                         &blocks, 0);
1084                 if (bmc) {
1085 /*
1086   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1087 */
1088                         if (*bmc == 2) {
1089                                 *bmc=1; /* maybe clear the bit next time */
1090                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1091                         } else if (*bmc == 1) {
1092                                 /* we can clear the bit */
1093                                 *bmc = 0;
1094                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1095                                                   -1);
1096
1097                                 /* clear the bit */
1098                                 clear_bit(file_page_offset(j), page_address(page));
1099                         }
1100                 }
1101                 spin_unlock_irqrestore(&bitmap->lock, flags);
1102         }
1103
1104         /* now sync the final page */
1105         if (lastpage != NULL) {
1106                 kunmap(lastpage);
1107                 spin_lock_irqsave(&bitmap->lock, flags);
1108                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1109                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1110                         spin_unlock_irqrestore(&bitmap->lock, flags);
1111                         err = write_page(bitmap, lastpage, 0);
1112                         if (err == -EAGAIN) {
1113                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1114                                 err = 0;
1115                         }
1116                 } else {
1117                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1118                         spin_unlock_irqrestore(&bitmap->lock, flags);
1119                 }
1120
1121                 page_cache_release(lastpage);
1122         }
1123
1124         return err;
1125 }
1126
1127 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1128 {
1129         mdk_thread_t *dmn;
1130         unsigned long flags;
1131
1132         /* if no one is waiting on us, we'll free the md thread struct
1133          * and exit, otherwise we let the waiter clean things up */
1134         spin_lock_irqsave(&bitmap->lock, flags);
1135         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1136                 *daemon = NULL;
1137                 spin_unlock_irqrestore(&bitmap->lock, flags);
1138                 kfree(dmn);
1139                 complete_and_exit(NULL, 0); /* do_exit not exported */
1140         }
1141         spin_unlock_irqrestore(&bitmap->lock, flags);
1142 }
1143
1144 static void bitmap_writeback_daemon(mddev_t *mddev)
1145 {
1146         struct bitmap *bitmap = mddev->bitmap;
1147         struct page *page;
1148         struct page_list *item;
1149         int err = 0;
1150
1151         if (signal_pending(current)) {
1152                 printk(KERN_INFO
1153                        "%s: bitmap writeback daemon got signal, exiting...\n",
1154                        bmname(bitmap));
1155                 err = -EINTR;
1156                 goto out;
1157         }
1158         if (bitmap == NULL)
1159                 /* about to be stopped. */
1160                 return;
1161
1162         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1163         /* wait on bitmap page writebacks */
1164         while ((item = dequeue_page(bitmap))) {
1165                 page = item->page;
1166                 mempool_free(item, bitmap->write_pool);
1167                 PRINTK("wait on page writeback: %p\n", page);
1168                 wait_on_page_writeback(page);
1169                 PRINTK("finished page writeback: %p\n", page);
1170
1171                 err = PageError(page);
1172                 page_cache_release(page);
1173                 if (err) {
1174                         printk(KERN_WARNING "%s: bitmap file writeback "
1175                                "failed (page %lu): %d\n",
1176                                bmname(bitmap), page->index, err);
1177                         bitmap_file_kick(bitmap);
1178                         goto out;
1179                 }
1180         }
1181  out:
1182         wake_up(&bitmap->write_wait);
1183         if (err) {
1184                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1185                        bmname(bitmap), err);
1186                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1187         }
1188 }
1189
1190 static mdk_thread_t *bitmap_start_daemon(struct bitmap *bitmap,
1191                                 void (*func)(mddev_t *), char *name)
1192 {
1193         mdk_thread_t *daemon;
1194         char namebuf[32];
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 ERR_PTR(-ECHILD);
1206         }
1207
1208         md_wakeup_thread(daemon); /* start it running */
1209
1210         PRINTK("%s: %s daemon (pid %d) started...\n",
1211                 bmname(bitmap), name, daemon->tsk->pid);
1212
1213         return daemon;
1214 }
1215
1216 static void bitmap_stop_daemon(struct bitmap *bitmap)
1217 {
1218         /* the daemon can't stop itself... it'll just exit instead... */
1219         if (bitmap->writeback_daemon && ! IS_ERR(bitmap->writeback_daemon) &&
1220             current->pid != bitmap->writeback_daemon->tsk->pid) {
1221                 mdk_thread_t *daemon;
1222                 unsigned long flags;
1223
1224                 spin_lock_irqsave(&bitmap->lock, flags);
1225                 daemon = bitmap->writeback_daemon;
1226                 bitmap->writeback_daemon = NULL;
1227                 spin_unlock_irqrestore(&bitmap->lock, flags);
1228                 if (daemon && ! IS_ERR(daemon))
1229                         md_unregister_thread(daemon); /* destroy the thread */
1230         }
1231 }
1232
1233 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1234                                             sector_t offset, int *blocks,
1235                                             int create)
1236 {
1237         /* If 'create', we might release the lock and reclaim it.
1238          * The lock must have been taken with interrupts enabled.
1239          * If !create, we don't release the lock.
1240          */
1241         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1242         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1243         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1244         sector_t csize;
1245
1246         if (bitmap_checkpage(bitmap, page, create) < 0) {
1247                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1248                 *blocks = csize - (offset & (csize- 1));
1249                 return NULL;
1250         }
1251         /* now locked ... */
1252
1253         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1254                 /* should we use the first or second counter field
1255                  * of the hijacked pointer? */
1256                 int hi = (pageoff > PAGE_COUNTER_MASK);
1257                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1258                                           PAGE_COUNTER_SHIFT - 1);
1259                 *blocks = csize - (offset & (csize- 1));
1260                 return  &((bitmap_counter_t *)
1261                           &bitmap->bp[page].map)[hi];
1262         } else { /* page is allocated */
1263                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1264                 *blocks = csize - (offset & (csize- 1));
1265                 return (bitmap_counter_t *)
1266                         &(bitmap->bp[page].map[pageoff]);
1267         }
1268 }
1269
1270 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1271 {
1272         if (!bitmap) return 0;
1273
1274         if (behind) {
1275                 atomic_inc(&bitmap->behind_writes);
1276                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1277                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1278         }
1279
1280         while (sectors) {
1281                 int blocks;
1282                 bitmap_counter_t *bmc;
1283
1284                 spin_lock_irq(&bitmap->lock);
1285                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1286                 if (!bmc) {
1287                         spin_unlock_irq(&bitmap->lock);
1288                         return 0;
1289                 }
1290
1291                 switch(*bmc) {
1292                 case 0:
1293                         bitmap_file_set_bit(bitmap, offset);
1294                         bitmap_count_page(bitmap,offset, 1);
1295                         blk_plug_device(bitmap->mddev->queue);
1296                         /* fall through */
1297                 case 1:
1298                         *bmc = 2;
1299                 }
1300                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1301                 (*bmc)++;
1302
1303                 spin_unlock_irq(&bitmap->lock);
1304
1305                 offset += blocks;
1306                 if (sectors > blocks)
1307                         sectors -= blocks;
1308                 else sectors = 0;
1309         }
1310         return 0;
1311 }
1312
1313 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1314                      int success, int behind)
1315 {
1316         if (!bitmap) return;
1317         if (behind) {
1318                 atomic_dec(&bitmap->behind_writes);
1319                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1320                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1321         }
1322
1323         while (sectors) {
1324                 int blocks;
1325                 unsigned long flags;
1326                 bitmap_counter_t *bmc;
1327
1328                 spin_lock_irqsave(&bitmap->lock, flags);
1329                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1330                 if (!bmc) {
1331                         spin_unlock_irqrestore(&bitmap->lock, flags);
1332                         return;
1333                 }
1334
1335                 if (!success && ! (*bmc & NEEDED_MASK))
1336                         *bmc |= NEEDED_MASK;
1337
1338                 (*bmc)--;
1339                 if (*bmc <= 2) {
1340                         set_page_attr(bitmap,
1341                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1342                                       BITMAP_PAGE_CLEAN);
1343                 }
1344                 spin_unlock_irqrestore(&bitmap->lock, flags);
1345                 offset += blocks;
1346                 if (sectors > blocks)
1347                         sectors -= blocks;
1348                 else sectors = 0;
1349         }
1350 }
1351
1352 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1353                         int degraded)
1354 {
1355         bitmap_counter_t *bmc;
1356         int rv;
1357         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1358                 *blocks = 1024;
1359                 return 1; /* always resync if no bitmap */
1360         }
1361         spin_lock_irq(&bitmap->lock);
1362         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1363         rv = 0;
1364         if (bmc) {
1365                 /* locked */
1366                 if (RESYNC(*bmc))
1367                         rv = 1;
1368                 else if (NEEDED(*bmc)) {
1369                         rv = 1;
1370                         if (!degraded) { /* don't set/clear bits if degraded */
1371                                 *bmc |= RESYNC_MASK;
1372                                 *bmc &= ~NEEDED_MASK;
1373                         }
1374                 }
1375         }
1376         spin_unlock_irq(&bitmap->lock);
1377         return rv;
1378 }
1379
1380 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1381 {
1382         bitmap_counter_t *bmc;
1383         unsigned long flags;
1384 /*
1385         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1386 */      if (bitmap == NULL) {
1387                 *blocks = 1024;
1388                 return;
1389         }
1390         spin_lock_irqsave(&bitmap->lock, flags);
1391         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1392         if (bmc == NULL)
1393                 goto unlock;
1394         /* locked */
1395 /*
1396         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1397 */
1398         if (RESYNC(*bmc)) {
1399                 *bmc &= ~RESYNC_MASK;
1400
1401                 if (!NEEDED(*bmc) && aborted)
1402                         *bmc |= NEEDED_MASK;
1403                 else {
1404                         if (*bmc <= 2) {
1405                                 set_page_attr(bitmap,
1406                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1407                                               BITMAP_PAGE_CLEAN);
1408                         }
1409                 }
1410         }
1411  unlock:
1412         spin_unlock_irqrestore(&bitmap->lock, flags);
1413 }
1414
1415 void bitmap_close_sync(struct bitmap *bitmap)
1416 {
1417         /* Sync has finished, and any bitmap chunks that weren't synced
1418          * properly have been aborted.  It remains to us to clear the
1419          * RESYNC bit wherever it is still on
1420          */
1421         sector_t sector = 0;
1422         int blocks;
1423         if (!bitmap) return;
1424         while (sector < bitmap->mddev->resync_max_sectors) {
1425                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1426 /*
1427                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1428                                          (unsigned long long)sector, blocks);
1429 */              sector += blocks;
1430         }
1431 }
1432
1433 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1434 {
1435         /* For each chunk covered by any of these sectors, set the
1436          * counter to 1 and set resync_needed.  They should all
1437          * be 0 at this point
1438          */
1439
1440         int secs;
1441         bitmap_counter_t *bmc;
1442         spin_lock_irq(&bitmap->lock);
1443         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1444         if (!bmc) {
1445                 spin_unlock_irq(&bitmap->lock);
1446                 return;
1447         }
1448         if (! *bmc) {
1449                 struct page *page;
1450                 *bmc = 1 | (needed?NEEDED_MASK:0);
1451                 bitmap_count_page(bitmap, offset, 1);
1452                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1453                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1454         }
1455         spin_unlock_irq(&bitmap->lock);
1456
1457 }
1458
1459 /*
1460  * flush out any pending updates
1461  */
1462 void bitmap_flush(mddev_t *mddev)
1463 {
1464         struct bitmap *bitmap = mddev->bitmap;
1465         int sleep;
1466
1467         if (!bitmap) /* there was no bitmap */
1468                 return;
1469
1470         /* run the daemon_work three time to ensure everything is flushed
1471          * that can be
1472          */
1473         sleep = bitmap->daemon_sleep;
1474         bitmap->daemon_sleep = 0;
1475         bitmap_daemon_work(bitmap);
1476         bitmap_daemon_work(bitmap);
1477         bitmap_daemon_work(bitmap);
1478         bitmap->daemon_sleep = sleep;
1479         bitmap_update_sb(bitmap);
1480 }
1481
1482 /*
1483  * free memory that was allocated
1484  */
1485 static void bitmap_free(struct bitmap *bitmap)
1486 {
1487         unsigned long k, pages;
1488         struct bitmap_page *bp;
1489
1490         if (!bitmap) /* there was no bitmap */
1491                 return;
1492
1493         /* release the bitmap file and kill the daemon */
1494         bitmap_file_put(bitmap);
1495
1496         bp = bitmap->bp;
1497         pages = bitmap->pages;
1498
1499         /* free all allocated memory */
1500
1501         mempool_destroy(bitmap->write_pool);
1502
1503         if (bp) /* deallocate the page memory */
1504                 for (k = 0; k < pages; k++)
1505                         if (bp[k].map && !bp[k].hijacked)
1506                                 kfree(bp[k].map);
1507         kfree(bp);
1508         kfree(bitmap);
1509 }
1510 void bitmap_destroy(mddev_t *mddev)
1511 {
1512         struct bitmap *bitmap = mddev->bitmap;
1513
1514         if (!bitmap) /* there was no bitmap */
1515                 return;
1516
1517         mddev->bitmap = NULL; /* disconnect from the md device */
1518
1519         bitmap_free(bitmap);
1520 }
1521
1522 /*
1523  * initialize the bitmap structure
1524  * if this returns an error, bitmap_destroy must be called to do clean up
1525  */
1526 int bitmap_create(mddev_t *mddev)
1527 {
1528         struct bitmap *bitmap;
1529         unsigned long blocks = mddev->resync_max_sectors;
1530         unsigned long chunks;
1531         unsigned long pages;
1532         struct file *file = mddev->bitmap_file;
1533         int err;
1534         sector_t start;
1535
1536         BUG_ON(sizeof(bitmap_super_t) != 256);
1537
1538         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1539                 return 0;
1540
1541         BUG_ON(file && mddev->bitmap_offset);
1542
1543         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1544         if (!bitmap)
1545                 return -ENOMEM;
1546
1547         memset(bitmap, 0, sizeof(*bitmap));
1548
1549         spin_lock_init(&bitmap->lock);
1550         bitmap->mddev = mddev;
1551
1552         spin_lock_init(&bitmap->write_lock);
1553         INIT_LIST_HEAD(&bitmap->complete_pages);
1554         init_waitqueue_head(&bitmap->write_wait);
1555         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1556                                 write_pool_free, NULL);
1557         err = -ENOMEM;
1558         if (!bitmap->write_pool)
1559                 goto error;
1560
1561         bitmap->file = file;
1562         bitmap->offset = mddev->bitmap_offset;
1563         if (file) get_file(file);
1564         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1565         err = bitmap_read_sb(bitmap);
1566         if (err)
1567                 goto error;
1568
1569         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1570                                         sizeof(bitmap->chunksize));
1571
1572         /* now that chunksize and chunkshift are set, we can use these macros */
1573         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1574                         CHUNK_BLOCK_RATIO(bitmap);
1575         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1576
1577         BUG_ON(!pages);
1578
1579         bitmap->chunks = chunks;
1580         bitmap->pages = pages;
1581         bitmap->missing_pages = pages;
1582         bitmap->counter_bits = COUNTER_BITS;
1583
1584         bitmap->syncchunk = ~0UL;
1585
1586 #ifdef INJECT_FATAL_FAULT_1
1587         bitmap->bp = NULL;
1588 #else
1589         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1590 #endif
1591         err = -ENOMEM;
1592         if (!bitmap->bp)
1593                 goto error;
1594         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1595
1596         bitmap->flags |= BITMAP_ACTIVE;
1597
1598         /* now that we have some pages available, initialize the in-memory
1599          * bitmap from the on-disk bitmap */
1600         start = 0;
1601         if (mddev->degraded == 0
1602             || bitmap->events_cleared == mddev->events)
1603                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1604                 start = mddev->recovery_cp;
1605         err = bitmap_init_from_disk(bitmap, start);
1606
1607         if (err)
1608                 goto error;
1609
1610         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1611                 pages, bmname(bitmap));
1612
1613         mddev->bitmap = bitmap;
1614
1615         if (file)
1616                 /* kick off the bitmap writeback daemon */
1617                 bitmap->writeback_daemon =
1618                         bitmap_start_daemon(bitmap,
1619                                             bitmap_writeback_daemon,
1620                                             "bitmap_wb");
1621
1622         if (IS_ERR(bitmap->writeback_daemon))
1623                 return PTR_ERR(bitmap->writeback_daemon);
1624         return bitmap_update_sb(bitmap);
1625
1626  error:
1627         bitmap_free(bitmap);
1628         return err;
1629 }
1630
1631 /* the bitmap API -- for raid personalities */
1632 EXPORT_SYMBOL(bitmap_startwrite);
1633 EXPORT_SYMBOL(bitmap_endwrite);
1634 EXPORT_SYMBOL(bitmap_start_sync);
1635 EXPORT_SYMBOL(bitmap_end_sync);
1636 EXPORT_SYMBOL(bitmap_unplug);
1637 EXPORT_SYMBOL(bitmap_close_sync);
1638 EXPORT_SYMBOL(bitmap_daemon_work);