drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/version.h>
  25 #include <linux/errno.h>
  26 #include <linux/slab.h>
  27 #include <linux/init.h>
  28 #include <linux/config.h>
  29 #include <linux/timer.h>
  30 #include <linux/sched.h>
  31 #include <linux/list.h>
  32 #include <linux/file.h>
  33 #include <linux/mount.h>
  34 #include <linux/buffer_head.h>
  35 #include <linux/raid/md.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 /* debug macros */
  39
  40 #define DEBUG 0
  41
  42 #if DEBUG
  43 /* these are for debugging purposes only! */
  44
  45 /* define one and only one of these */
  46 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  47 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  48 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  49 #define INJECT_FAULTS_4 0 /* undef */
  50 #define INJECT_FAULTS_5 0 /* undef */
  51 #define INJECT_FAULTS_6 0
  52
  53 /* if these are defined, the driver will fail! debug only */
  54 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  55 #define INJECT_FATAL_FAULT_2 0 /* undef */
  56 #define INJECT_FATAL_FAULT_3 0 /* undef */
  57 #endif
  58
  59 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  60 #define DPRINTK(x...) do { } while(0)
  61
  62 #ifndef PRINTK
  63 #  if DEBUG > 0
  64 #    define PRINTK(x...) printk(KERN_DEBUG x)
  65 #  else
  66 #    define PRINTK(x...)
  67 #  endif
  68 #endif
  69
  70 static inline char * bmname(struct bitmap *bitmap)
  71 {
  72         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  73 }
  74
  75
  76 /*
  77  * test if the bitmap is active
  78  */
  79 int bitmap_active(struct bitmap *bitmap)
  80 {
  81         unsigned long flags;
  82         int res = 0;
  83
  84         if (!bitmap)
  85                 return res;
  86         spin_lock_irqsave(&bitmap->lock, flags);
  87         res = bitmap->flags & BITMAP_ACTIVE;
  88         spin_unlock_irqrestore(&bitmap->lock, flags);
  89         return res;
  90 }
  91
  92 #define WRITE_POOL_SIZE 256
  93 /* mempool for queueing pending writes on the bitmap file */
  94 static void *write_pool_alloc(unsigned int gfp_flags, void *data)
  95 {
  96         return kmalloc(sizeof(struct page_list), gfp_flags);
  97 }
  98
  99 static void write_pool_free(void *ptr, void *data)
 100 {
 101         kfree(ptr);
 102 }
 103
 104 /*
 105  * just a placeholder - calls kmalloc for bitmap pages
 106  */
 107 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 108 {
 109         unsigned char *page;
 110
 111 #ifdef INJECT_FAULTS_1
 112         page = NULL;
 113 #else
 114         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 115 #endif
 116         if (!page)
 117                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 118         else
 119                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 120                         bmname(bitmap), page);
 121         return page;
 122 }
 123
 124 /*
 125  * for now just a placeholder -- just calls kfree for bitmap pages
 126  */
 127 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 128 {
 129         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 130         kfree(page);
 131 }
 132
 133 /*
 134  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 135  *
 136  * 1) check to see if this page is allocated, if it's not then try to alloc
 137  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 138  *    page pointer directly as a counter
 139  *
 140  * if we find our page, we increment the page's refcount so that it stays
 141  * allocated while we're using it
 142  */
 143 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 144 {
 145         unsigned char *mappage;
 146
 147         if (page >= bitmap->pages) {
 148                 printk(KERN_ALERT
 149                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 150                         bmname(bitmap), page, bitmap->pages-1);
 151                 return -EINVAL;
 152         }
 153
 154
 155         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 156                 return 0;
 157
 158         if (bitmap->bp[page].map) /* page is already allocated, just return */
 159                 return 0;
 160
 161         if (!create)
 162                 return -ENOENT;
 163
 164         spin_unlock_irq(&bitmap->lock);
 165
 166         /* this page has not been allocated yet */
 167
 168         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 169                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 170                         bmname(bitmap));
 171                 /* failed - set the hijacked flag so that we can use the
 172                  * pointer as a counter */
 173                 spin_lock_irq(&bitmap->lock);
 174                 if (!bitmap->bp[page].map)
 175                         bitmap->bp[page].hijacked = 1;
 176                 goto out;
 177         }
 178
 179         /* got a page */
 180
 181         spin_lock_irq(&bitmap->lock);
 182
 183         /* recheck the page */
 184
 185         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 186                 /* somebody beat us to getting the page */
 187                 bitmap_free_page(bitmap, mappage);
 188                 return 0;
 189         }
 190
 191         /* no page was in place and we have one, so install it */
 192
 193         memset(mappage, 0, PAGE_SIZE);
 194         bitmap->bp[page].map = mappage;
 195         bitmap->missing_pages--;
 196 out:
 197         return 0;
 198 }
 199
 200
 201 /* if page is completely empty, put it back on the free list, or dealloc it */
 202 /* if page was hijacked, unmark the flag so it might get alloced next time */
 203 /* Note: lock should be held when calling this */
 204 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 205 {
 206         char *ptr;
 207
 208         if (bitmap->bp[page].count) /* page is still busy */
 209                 return;
 210
 211         /* page is no longer in use, it can be released */
 212
 213         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 214                 bitmap->bp[page].hijacked = 0;
 215                 bitmap->bp[page].map = NULL;
 216                 return;
 217         }
 218
 219         /* normal case, free the page */
 220
 221 #if 0
 222 /* actually ... let's not.  We will probably need the page again exactly when
 223  * memory is tight and we are flusing to disk
 224  */
 225         return;
 226 #else
 227         ptr = bitmap->bp[page].map;
 228         bitmap->bp[page].map = NULL;
 229         bitmap->missing_pages++;
 230         bitmap_free_page(bitmap, ptr);
 231         return;
 232 #endif
 233 }
 234
 235
 236 /*
 237  * bitmap file handling - read and write the bitmap file and its superblock
 238  */
 239
 240 /* copy the pathname of a file to a buffer */
 241 char *file_path(struct file *file, char *buf, int count)
 242 {
 243         struct dentry *d;
 244         struct vfsmount *v;
 245
 246         if (!buf)
 247                 return NULL;
 248
 249         d = file->f_dentry;
 250         v = file->f_vfsmnt;
 251
 252         buf = d_path(d, v, buf, count);
 253
 254         return IS_ERR(buf) ? NULL : buf;
 255 }
 256
 257 /*
 258  * basic page I/O operations
 259  */
 260
 261 /* IO operations when bitmap is stored near all superblocks */
 262 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 263 {
 264         /* choose a good rdev and read the page from there */
 265
 266         mdk_rdev_t *rdev;
 267         struct list_head *tmp;
 268         struct page *page = alloc_page(GFP_KERNEL);
 269         sector_t target;
 270
 271         if (!page)
 272                 return ERR_PTR(-ENOMEM);
 273         do {
 274                 ITERATE_RDEV(mddev, rdev, tmp)
 275                         if (rdev->in_sync && !rdev->faulty)
 276                                 goto found;
 277                 return ERR_PTR(-EIO);
 278
 279         found:
 280                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 281
 282         } while (!sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ));
 283
 284         page->index = index;
 285         return page;
 286 }
 287
 288 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 289 {
 290         mdk_rdev_t *rdev;
 291         struct list_head *tmp;
 292
 293         ITERATE_RDEV(mddev, rdev, tmp)
 294                 if (rdev->in_sync && !rdev->faulty)
 295                         md_super_write(mddev, rdev,
 296                                        (rdev->sb_offset<<1) + offset
 297                                        + page->index * (PAGE_SIZE/512),
 298                                        PAGE_SIZE,
 299                                        page);
 300
 301         if (wait)
 302                 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 303         return 0;
 304 }
 305
 306 /*
 307  * write out a page to a file
 308  */
 309 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 310 {
 311         int ret = -ENOMEM;
 312
 313         if (bitmap->file == NULL)
 314                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 315
 316         if (wait)
 317                 lock_page(page);
 318         else {
 319                 if (TestSetPageLocked(page))
 320                         return -EAGAIN; /* already locked */
 321                 if (PageWriteback(page)) {
 322                         unlock_page(page);
 323                         return -EAGAIN;
 324                 }
 325         }
 326
 327         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 328         if (!ret)
 329                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 330                         PAGE_SIZE);
 331         if (ret) {
 332                 unlock_page(page);
 333                 return ret;
 334         }
 335
 336         set_page_dirty(page); /* force it to be written out */
 337
 338         if (!wait) {
 339                 /* add to list to be waited for by daemon */
 340                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 341                 item->page = page;
 342                 page_cache_get(page);
 343                 spin_lock(&bitmap->write_lock);
 344                 list_add(&item->list, &bitmap->complete_pages);
 345                 spin_unlock(&bitmap->write_lock);
 346                 md_wakeup_thread(bitmap->writeback_daemon);
 347         }
 348         return write_one_page(page, wait);
 349 }
 350
 351 /* read a page from a file, pinning it into cache, and return bytes_read */
 352 static struct page *read_page(struct file *file, unsigned long index,
 353                                         unsigned long *bytes_read)
 354 {
 355         struct inode *inode = file->f_mapping->host;
 356         struct page *page = NULL;
 357         loff_t isize = i_size_read(inode);
 358         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 359
 360         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 361                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 362
 363         page = read_cache_page(inode->i_mapping, index,
 364                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 365         if (IS_ERR(page))
 366                 goto out;
 367         wait_on_page_locked(page);
 368         if (!PageUptodate(page) || PageError(page)) {
 369                 page_cache_release(page);
 370                 page = ERR_PTR(-EIO);
 371                 goto out;
 372         }
 373
 374         if (index > end_index) /* we have read beyond EOF */
 375                 *bytes_read = 0;
 376         else if (index == end_index) /* possible short read */
 377                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 378         else
 379                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 380 out:
 381         if (IS_ERR(page))
 382                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 383                         (int)PAGE_CACHE_SIZE,
 384                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 385                         PTR_ERR(page));
 386         return page;
 387 }
 388
 389 /*
 390  * bitmap file superblock operations
 391  */
 392
 393 /* update the event counter and sync the superblock to disk */
 394 int bitmap_update_sb(struct bitmap *bitmap)
 395 {
 396         bitmap_super_t *sb;
 397         unsigned long flags;
 398
 399         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 400                 return 0;
 401         spin_lock_irqsave(&bitmap->lock, flags);
 402         if (!bitmap->sb_page) { /* no superblock */
 403                 spin_unlock_irqrestore(&bitmap->lock, flags);
 404                 return 0;
 405         }
 406         spin_unlock_irqrestore(&bitmap->lock, flags);
 407         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 408         sb->events = cpu_to_le64(bitmap->mddev->events);
 409         if (!bitmap->mddev->degraded)
 410                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 411         kunmap(bitmap->sb_page);
 412         return write_page(bitmap, bitmap->sb_page, 1);
 413 }
 414
 415 /* print out the bitmap file superblock */
 416 void bitmap_print_sb(struct bitmap *bitmap)
 417 {
 418         bitmap_super_t *sb;
 419
 420         if (!bitmap || !bitmap->sb_page)
 421                 return;
 422         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 423         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 424         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 425         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 426         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 427                                         *(__u32 *)(sb->uuid+0),
 428                                         *(__u32 *)(sb->uuid+4),
 429                                         *(__u32 *)(sb->uuid+8),
 430                                         *(__u32 *)(sb->uuid+12));
 431         printk(KERN_DEBUG "        events: %llu\n",
 432                         (unsigned long long) le64_to_cpu(sb->events));
 433         printk(KERN_DEBUG "events cleared: %llu\n",
 434                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 435         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 436         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 437         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 438         printk(KERN_DEBUG "     sync size: %llu KB\n",
 439                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 440         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->max_write_behind = write_behind;
 526         bitmap->flags |= sb->state;
 527         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 528         if (sb->state & BITMAP_STALE)
 529                 bitmap->events_cleared = bitmap->mddev->events;
 530         err = 0;
 531 out:
 532         kunmap(bitmap->sb_page);
 533         if (err)
 534                 bitmap_print_sb(bitmap);
 535         return err;
 536 }
 537
 538 enum bitmap_mask_op {
 539         MASK_SET,
 540         MASK_UNSET
 541 };
 542
 543 /* record the state of the bitmap in the superblock */
 544 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 545                                 enum bitmap_mask_op op)
 546 {
 547         bitmap_super_t *sb;
 548         unsigned long flags;
 549
 550         spin_lock_irqsave(&bitmap->lock, flags);
 551         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 552                 spin_unlock_irqrestore(&bitmap->lock, flags);
 553                 return;
 554         }
 555         page_cache_get(bitmap->sb_page);
 556         spin_unlock_irqrestore(&bitmap->lock, flags);
 557         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 558         switch (op) {
 559                 case MASK_SET: sb->state |= bits;
 560                                 break;
 561                 case MASK_UNSET: sb->state &= ~bits;
 562                                 break;
 563                 default: BUG();
 564         }
 565         kunmap(bitmap->sb_page);
 566         page_cache_release(bitmap->sb_page);
 567 }
 568
 569 /*
 570  * general bitmap file operations
 571  */
 572
 573 /* calculate the index of the page that contains this bit */
 574 static inline unsigned long file_page_index(unsigned long chunk)
 575 {
 576         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 577 }
 578
 579 /* calculate the (bit) offset of this bit within a page */
 580 static inline unsigned long file_page_offset(unsigned long chunk)
 581 {
 582         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 583 }
 584
 585 /*
 586  * return a pointer to the page in the filemap that contains the given bit
 587  *
 588  * this lookup is complicated by the fact that the bitmap sb might be exactly
 589  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 590  * 0 or page 1
 591  */
 592 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 593                                         unsigned long chunk)
 594 {
 595         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 596 }
 597
 598
 599 static void bitmap_file_unmap(struct bitmap *bitmap)
 600 {
 601         struct page **map, *sb_page;
 602         unsigned long *attr;
 603         int pages;
 604         unsigned long flags;
 605
 606         spin_lock_irqsave(&bitmap->lock, flags);
 607         map = bitmap->filemap;
 608         bitmap->filemap = NULL;
 609         attr = bitmap->filemap_attr;
 610         bitmap->filemap_attr = NULL;
 611         pages = bitmap->file_pages;
 612         bitmap->file_pages = 0;
 613         sb_page = bitmap->sb_page;
 614         bitmap->sb_page = NULL;
 615         spin_unlock_irqrestore(&bitmap->lock, flags);
 616
 617         while (pages--)
 618                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 619                         page_cache_release(map[pages]);
 620         kfree(map);
 621         kfree(attr);
 622
 623         if (sb_page)
 624                 page_cache_release(sb_page);
 625 }
 626
 627 static void bitmap_stop_daemons(struct bitmap *bitmap);
 628
 629 /* dequeue the next item in a page list -- don't call from irq context */
 630 static struct page_list *dequeue_page(struct bitmap *bitmap)
 631 {
 632         struct page_list *item = NULL;
 633         struct list_head *head = &bitmap->complete_pages;
 634
 635         spin_lock(&bitmap->write_lock);
 636         if (list_empty(head))
 637                 goto out;
 638         item = list_entry(head->prev, struct page_list, list);
 639         list_del(head->prev);
 640 out:
 641         spin_unlock(&bitmap->write_lock);
 642         return item;
 643 }
 644
 645 static void drain_write_queues(struct bitmap *bitmap)
 646 {
 647         struct page_list *item;
 648
 649         while ((item = dequeue_page(bitmap))) {
 650                 /* don't bother to wait */
 651                 page_cache_release(item->page);
 652                 mempool_free(item, bitmap->write_pool);
 653         }
 654
 655         wake_up(&bitmap->write_wait);
 656 }
 657
 658 static void bitmap_file_put(struct bitmap *bitmap)
 659 {
 660         struct file *file;
 661         struct inode *inode;
 662         unsigned long flags;
 663
 664         spin_lock_irqsave(&bitmap->lock, flags);
 665         file = bitmap->file;
 666         bitmap->file = NULL;
 667         spin_unlock_irqrestore(&bitmap->lock, flags);
 668
 669         bitmap_stop_daemons(bitmap);
 670
 671         drain_write_queues(bitmap);
 672
 673         bitmap_file_unmap(bitmap);
 674
 675         if (file) {
 676                 inode = file->f_mapping->host;
 677                 spin_lock(&inode->i_lock);
 678                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 679                 spin_unlock(&inode->i_lock);
 680                 fput(file);
 681         }
 682 }
 683
 684
 685 /*
 686  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 687  * then it is no longer reliable, so we stop using it and we mark the file
 688  * as failed in the superblock
 689  */
 690 static void bitmap_file_kick(struct bitmap *bitmap)
 691 {
 692         char *path, *ptr = NULL;
 693
 694         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 695         bitmap_update_sb(bitmap);
 696
 697         if (bitmap->file) {
 698                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 699                 if (path)
 700                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 701
 702                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 703                        bmname(bitmap), ptr ? ptr : "");
 704
 705                 kfree(path);
 706         }
 707
 708         bitmap_file_put(bitmap);
 709
 710         return;
 711 }
 712
 713 enum bitmap_page_attr {
 714         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 715         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 716         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 717 };
 718
 719 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 720                                 enum bitmap_page_attr attr)
 721 {
 722         bitmap->filemap_attr[page->index] |= attr;
 723 }
 724
 725 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 726                                 enum bitmap_page_attr attr)
 727 {
 728         bitmap->filemap_attr[page->index] &= ~attr;
 729 }
 730
 731 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 732 {
 733         return bitmap->filemap_attr[page->index];
 734 }
 735
 736 /*
 737  * bitmap_file_set_bit -- called before performing a write to the md device
 738  * to set (and eventually sync) a particular bit in the bitmap file
 739  *
 740  * we set the bit immediately, then we record the page number so that
 741  * when an unplug occurs, we can flush the dirty pages out to disk
 742  */
 743 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 744 {
 745         unsigned long bit;
 746         struct page *page;
 747         void *kaddr;
 748         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 749
 750         if (!bitmap->filemap) {
 751                 return;
 752         }
 753
 754         page = filemap_get_page(bitmap, chunk);
 755         bit = file_page_offset(chunk);
 756
 757
 758         /* make sure the page stays cached until it gets written out */
 759         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 760                 page_cache_get(page);
 761
 762         /* set the bit */
 763         kaddr = kmap_atomic(page, KM_USER0);
 764         set_bit(bit, kaddr);
 765         kunmap_atomic(kaddr, KM_USER0);
 766         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 767
 768         /* record page number so it gets flushed to disk when unplug occurs */
 769         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 770
 771 }
 772
 773 /* this gets called when the md device is ready to unplug its underlying
 774  * (slave) device queues -- before we let any writes go down, we need to
 775  * sync the dirty pages of the bitmap file to disk */
 776 int bitmap_unplug(struct bitmap *bitmap)
 777 {
 778         unsigned long i, attr, flags;
 779         struct page *page;
 780         int wait = 0;
 781         int err;
 782
 783         if (!bitmap)
 784                 return 0;
 785
 786         /* look at each page to see if there are any set bits that need to be
 787          * flushed out to disk */
 788         for (i = 0; i < bitmap->file_pages; i++) {
 789                 spin_lock_irqsave(&bitmap->lock, flags);
 790                 if (!bitmap->filemap) {
 791                         spin_unlock_irqrestore(&bitmap->lock, flags);
 792                         return 0;
 793                 }
 794                 page = bitmap->filemap[i];
 795                 attr = get_page_attr(bitmap, page);
 796                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 797                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 798                 if ((attr & BITMAP_PAGE_DIRTY))
 799                         wait = 1;
 800                 spin_unlock_irqrestore(&bitmap->lock, flags);
 801
 802                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 803                         err = write_page(bitmap, page, 0);
 804                         if (err == -EAGAIN) {
 805                                 if (attr & BITMAP_PAGE_DIRTY)
 806                                         err = write_page(bitmap, page, 1);
 807                                 else
 808                                         err = 0;
 809                         }
 810                         if (err)
 811                                 return 1;
 812                 }
 813         }
 814         if (wait) { /* if any writes were performed, we need to wait on them */
 815                 if (bitmap->file) {
 816                         spin_lock_irq(&bitmap->write_lock);
 817                         wait_event_lock_irq(bitmap->write_wait,
 818                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 819                                             wake_up_process(bitmap->writeback_daemon->tsk));
 820                         spin_unlock_irq(&bitmap->write_lock);
 821                 } else
 822                         wait_event(bitmap->mddev->sb_wait,
 823                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 824         }
 825         return 0;
 826 }
 827
 828 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 829 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 830  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 831  * memory mapping of the bitmap file
 832  * Special cases:
 833  *   if there's no bitmap file, or if the bitmap file had been
 834  *   previously kicked from the array, we mark all the bits as
 835  *   1's in order to cause a full resync.
 836  *
 837  * We ignore all bits for sectors that end earlier than 'start'.
 838  * This is used when reading an out-of-date bitmap...
 839  */
 840 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 841 {
 842         unsigned long i, chunks, index, oldindex, bit;
 843         struct page *page = NULL, *oldpage = NULL;
 844         unsigned long num_pages, bit_cnt = 0;
 845         struct file *file;
 846         unsigned long bytes, offset, dummy;
 847         int outofdate;
 848         int ret = -ENOSPC;
 849
 850         chunks = bitmap->chunks;
 851         file = bitmap->file;
 852
 853         BUG_ON(!file && !bitmap->offset);
 854
 855 #ifdef INJECT_FAULTS_3
 856         outofdate = 1;
 857 #else
 858         outofdate = bitmap->flags & BITMAP_STALE;
 859 #endif
 860         if (outofdate)
 861                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 862                         "recovery\n", bmname(bitmap));
 863
 864         bytes = (chunks + 7) / 8;
 865
 866         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 867
 868         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 869                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 870                         bmname(bitmap),
 871                         (unsigned long) i_size_read(file->f_mapping->host),
 872                         bytes + sizeof(bitmap_super_t));
 873                 goto out;
 874         }
 875
 876         ret = -ENOMEM;
 877
 878         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 879         if (!bitmap->filemap)
 880                 goto out;
 881
 882         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 883         if (!bitmap->filemap_attr)
 884                 goto out;
 885
 886         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 887
 888         oldindex = ~0L;
 889
 890         for (i = 0; i < chunks; i++) {
 891                 index = file_page_index(i);
 892                 bit = file_page_offset(i);
 893                 if (index != oldindex) { /* this is a new page, read it in */
 894                         /* unmap the old page, we're done with it */
 895                         if (oldpage != NULL)
 896                                 kunmap(oldpage);
 897                         if (index == 0) {
 898                                 /*
 899                                  * if we're here then the superblock page
 900                                  * contains some bits (PAGE_SIZE != sizeof sb)
 901                                  * we've already read it in, so just use it
 902                                  */
 903                                 page = bitmap->sb_page;
 904                                 offset = sizeof(bitmap_super_t);
 905                         } else if (file) {
 906                                 page = read_page(file, index, &dummy);
 907                                 offset = 0;
 908                         } else {
 909                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 910                                 offset = 0;
 911                         }
 912                         if (IS_ERR(page)) { /* read error */
 913                                 ret = PTR_ERR(page);
 914                                 goto out;
 915                         }
 916
 917                         oldindex = index;
 918                         oldpage = page;
 919                         kmap(page);
 920
 921                         if (outofdate) {
 922                                 /*
 923                                  * if bitmap is out of date, dirty the
 924                                  * whole page and write it out
 925                                  */
 926                                 memset(page_address(page) + offset, 0xff,
 927                                        PAGE_SIZE - offset);
 928                                 ret = write_page(bitmap, page, 1);
 929                                 if (ret) {
 930                                         kunmap(page);
 931                                         /* release, page not in filemap yet */
 932                                         page_cache_release(page);
 933                                         goto out;
 934                                 }
 935                         }
 936
 937                         bitmap->filemap[bitmap->file_pages++] = page;
 938                 }
 939                 if (test_bit(bit, page_address(page))) {
 940                         /* if the disk bit is set, set the memory bit */
 941                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 942                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 943                                 );
 944                         bit_cnt++;
 945                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 946                 }
 947         }
 948
 949         /* everything went OK */
 950         ret = 0;
 951         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 952
 953         if (page) /* unmap the last page */
 954                 kunmap(page);
 955
 956         if (bit_cnt) { /* Kick recovery if any bits were set */
 957                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 958                 md_wakeup_thread(bitmap->mddev->thread);
 959         }
 960
 961 out:
 962         printk(KERN_INFO "%s: bitmap initialized from disk: "
 963                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 964                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 965
 966         return ret;
 967 }
 968
 969 void bitmap_write_all(struct bitmap *bitmap)
 970 {
 971         /* We don't actually write all bitmap blocks here,
 972          * just flag them as needing to be written
 973          */
 974
 975         unsigned long chunks = bitmap->chunks;
 976         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 977         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 978         while (num_pages--)
 979                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 980 }
 981
 982
 983 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 984 {
 985         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 986         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 987         bitmap->bp[page].count += inc;
 988 /*
 989         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 990                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 991 */
 992         bitmap_checkfree(bitmap, page);
 993 }
 994 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 995                                             sector_t offset, int *blocks,
 996                                             int create);
 997
 998 /*
 999  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1000  *                      out to disk
1001  */
1002
1003 int bitmap_daemon_work(struct bitmap *bitmap)
1004 {
1005         unsigned long j;
1006         unsigned long flags;
1007         struct page *page = NULL, *lastpage = NULL;
1008         int err = 0;
1009         int blocks;
1010         int attr;
1011
1012         if (bitmap == NULL)
1013                 return 0;
1014         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1015                 return 0;
1016         bitmap->daemon_lastrun = jiffies;
1017
1018         for (j = 0; j < bitmap->chunks; j++) {
1019                 bitmap_counter_t *bmc;
1020                 spin_lock_irqsave(&bitmap->lock, flags);
1021                 if (!bitmap->filemap) {
1022                         /* error or shutdown */
1023                         spin_unlock_irqrestore(&bitmap->lock, flags);
1024                         break;
1025                 }
1026
1027                 page = filemap_get_page(bitmap, j);
1028
1029                 if (page != lastpage) {
1030                         /* skip this page unless it's marked as needing cleaning */
1031                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1032                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1033                                         page_cache_get(page);
1034                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1035                                 }
1036                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1037                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1038                                         switch (write_page(bitmap, page, 0)) {
1039                                         case -EAGAIN:
1040                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1041                                                 break;
1042                                         case 0:
1043                                                 break;
1044                                         default:
1045                                                 bitmap_file_kick(bitmap);
1046                                         }
1047                                         page_cache_release(page);
1048                                 }
1049                                 continue;
1050                         }
1051
1052                         /* grab the new page, sync and release the old */
1053                         page_cache_get(page);
1054                         if (lastpage != NULL) {
1055                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1056                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1057                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1058                                         err = write_page(bitmap, lastpage, 0);
1059                                         if (err == -EAGAIN) {
1060                                                 err = 0;
1061                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1062                                         }
1063                                 } else {
1064                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1065                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1066                                 }
1067                                 kunmap(lastpage);
1068                                 page_cache_release(lastpage);
1069                                 if (err)
1070                                         bitmap_file_kick(bitmap);
1071                         } else
1072                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1073                         lastpage = page;
1074                         kmap(page);
1075 /*
1076                         printk("bitmap clean at page %lu\n", j);
1077 */
1078                         spin_lock_irqsave(&bitmap->lock, flags);
1079                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1080                 }
1081                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1082                                         &blocks, 0);
1083                 if (bmc) {
1084 /*
1085   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1086 */
1087                         if (*bmc == 2) {
1088                                 *bmc=1; /* maybe clear the bit next time */
1089                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1090                         } else if (*bmc == 1) {
1091                                 /* we can clear the bit */
1092                                 *bmc = 0;
1093                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1094                                                   -1);
1095
1096                                 /* clear the bit */
1097                                 clear_bit(file_page_offset(j), page_address(page));
1098                         }
1099                 }
1100                 spin_unlock_irqrestore(&bitmap->lock, flags);
1101         }
1102
1103         /* now sync the final page */
1104         if (lastpage != NULL) {
1105                 kunmap(lastpage);
1106                 spin_lock_irqsave(&bitmap->lock, flags);
1107                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1108                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1109                         spin_unlock_irqrestore(&bitmap->lock, flags);
1110                         err = write_page(bitmap, lastpage, 0);
1111                         if (err == -EAGAIN) {
1112                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1113                                 err = 0;
1114                         }
1115                 } else {
1116                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1117                         spin_unlock_irqrestore(&bitmap->lock, flags);
1118                 }
1119
1120                 page_cache_release(lastpage);
1121         }
1122
1123         return err;
1124 }
1125
1126 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1127 {
1128         mdk_thread_t *dmn;
1129         unsigned long flags;
1130
1131         /* if no one is waiting on us, we'll free the md thread struct
1132          * and exit, otherwise we let the waiter clean things up */
1133         spin_lock_irqsave(&bitmap->lock, flags);
1134         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1135                 *daemon = NULL;
1136                 spin_unlock_irqrestore(&bitmap->lock, flags);
1137                 kfree(dmn);
1138                 complete_and_exit(NULL, 0); /* do_exit not exported */
1139         }
1140         spin_unlock_irqrestore(&bitmap->lock, flags);
1141 }
1142
1143 static void bitmap_writeback_daemon(mddev_t *mddev)
1144 {
1145         struct bitmap *bitmap = mddev->bitmap;
1146         struct page *page;
1147         struct page_list *item;
1148         int err = 0;
1149
1150         if (signal_pending(current)) {
1151                 printk(KERN_INFO
1152                        "%s: bitmap writeback daemon got signal, exiting...\n",
1153                        bmname(bitmap));
1154                 err = -EINTR;
1155                 goto out;
1156         }
1157
1158         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1159         /* wait on bitmap page writebacks */
1160         while ((item = dequeue_page(bitmap))) {
1161                 page = item->page;
1162                 mempool_free(item, bitmap->write_pool);
1163                 PRINTK("wait on page writeback: %p\n", page);
1164                 wait_on_page_writeback(page);
1165                 PRINTK("finished page writeback: %p\n", page);
1166
1167                 err = PageError(page);
1168                 page_cache_release(page);
1169                 if (err) {
1170                         printk(KERN_WARNING "%s: bitmap file writeback "
1171                                "failed (page %lu): %d\n",
1172                                bmname(bitmap), page->index, err);
1173                         bitmap_file_kick(bitmap);
1174                         goto out;
1175                 }
1176         }
1177  out:
1178         wake_up(&bitmap->write_wait);
1179         if (err) {
1180                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1181                        bmname(bitmap), err);
1182                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1183         }
1184 }
1185
1186 static int bitmap_start_daemon(struct bitmap *bitmap, mdk_thread_t **ptr,
1187                                 void (*func)(mddev_t *), char *name)
1188 {
1189         mdk_thread_t *daemon;
1190         unsigned long flags;
1191         char namebuf[32];
1192
1193         spin_lock_irqsave(&bitmap->lock, flags);
1194         *ptr = NULL;
1195
1196         if (!bitmap->file) /* no need for daemon if there's no backing file */
1197                 goto out_unlock;
1198
1199         spin_unlock_irqrestore(&bitmap->lock, flags);
1200
1201 #ifdef INJECT_FATAL_FAULT_2
1202         daemon = NULL;
1203 #else
1204         sprintf(namebuf, "%%s_%s", name);
1205         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1206 #endif
1207         if (!daemon) {
1208                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1209                         bmname(bitmap));
1210                 return -ECHILD;
1211         }
1212
1213         spin_lock_irqsave(&bitmap->lock, flags);
1214         *ptr = daemon;
1215
1216         md_wakeup_thread(daemon); /* start it running */
1217
1218         PRINTK("%s: %s daemon (pid %d) started...\n",
1219                 bmname(bitmap), name, daemon->tsk->pid);
1220 out_unlock:
1221         spin_unlock_irqrestore(&bitmap->lock, flags);
1222         return 0;
1223 }
1224
1225 static int bitmap_start_daemons(struct bitmap *bitmap)
1226 {
1227         int err = bitmap_start_daemon(bitmap, &bitmap->writeback_daemon,
1228                                         bitmap_writeback_daemon, "bitmap_wb");
1229         return err;
1230 }
1231
1232 static void bitmap_stop_daemon(struct bitmap *bitmap, mdk_thread_t **ptr)
1233 {
1234         mdk_thread_t *daemon;
1235         unsigned long flags;
1236
1237         spin_lock_irqsave(&bitmap->lock, flags);
1238         daemon = *ptr;
1239         *ptr = NULL;
1240         spin_unlock_irqrestore(&bitmap->lock, flags);
1241         if (daemon)
1242                 md_unregister_thread(daemon); /* destroy the thread */
1243 }
1244
1245 static void bitmap_stop_daemons(struct bitmap *bitmap)
1246 {
1247         /* the daemons can't stop themselves... they'll just exit instead... */
1248         if (bitmap->writeback_daemon &&
1249             current->pid != bitmap->writeback_daemon->tsk->pid)
1250                 bitmap_stop_daemon(bitmap, &bitmap->writeback_daemon);
1251 }
1252
1253 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1254                                             sector_t offset, int *blocks,
1255                                             int create)
1256 {
1257         /* If 'create', we might release the lock and reclaim it.
1258          * The lock must have been taken with interrupts enabled.
1259          * If !create, we don't release the lock.
1260          */
1261         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1262         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1263         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1264         sector_t csize;
1265
1266         if (bitmap_checkpage(bitmap, page, create) < 0) {
1267                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1268                 *blocks = csize - (offset & (csize- 1));
1269                 return NULL;
1270         }
1271         /* now locked ... */
1272
1273         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1274                 /* should we use the first or second counter field
1275                  * of the hijacked pointer? */
1276                 int hi = (pageoff > PAGE_COUNTER_MASK);
1277                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1278                                           PAGE_COUNTER_SHIFT - 1);
1279                 *blocks = csize - (offset & (csize- 1));
1280                 return  &((bitmap_counter_t *)
1281                           &bitmap->bp[page].map)[hi];
1282         } else { /* page is allocated */
1283                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1284                 *blocks = csize - (offset & (csize- 1));
1285                 return (bitmap_counter_t *)
1286                         &(bitmap->bp[page].map[pageoff]);
1287         }
1288 }
1289
1290 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1291 {
1292         if (!bitmap) return 0;
1293
1294         if (behind) {
1295                 atomic_inc(&bitmap->behind_writes);
1296                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1297                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1298         }
1299
1300         while (sectors) {
1301                 int blocks;
1302                 bitmap_counter_t *bmc;
1303
1304                 spin_lock_irq(&bitmap->lock);
1305                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1306                 if (!bmc) {
1307                         spin_unlock_irq(&bitmap->lock);
1308                         return 0;
1309                 }
1310
1311                 switch(*bmc) {
1312                 case 0:
1313                         bitmap_file_set_bit(bitmap, offset);
1314                         bitmap_count_page(bitmap,offset, 1);
1315                         blk_plug_device(bitmap->mddev->queue);
1316                         /* fall through */
1317                 case 1:
1318                         *bmc = 2;
1319                 }
1320                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1321                 (*bmc)++;
1322
1323                 spin_unlock_irq(&bitmap->lock);
1324
1325                 offset += blocks;
1326                 if (sectors > blocks)
1327                         sectors -= blocks;
1328                 else sectors = 0;
1329         }
1330         return 0;
1331 }
1332
1333 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1334                      int success, int behind)
1335 {
1336         if (!bitmap) return;
1337         if (behind) {
1338                 atomic_dec(&bitmap->behind_writes);
1339                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1340                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1341         }
1342
1343         while (sectors) {
1344                 int blocks;
1345                 unsigned long flags;
1346                 bitmap_counter_t *bmc;
1347
1348                 spin_lock_irqsave(&bitmap->lock, flags);
1349                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1350                 if (!bmc) {
1351                         spin_unlock_irqrestore(&bitmap->lock, flags);
1352                         return;
1353                 }
1354
1355                 if (!success && ! (*bmc & NEEDED_MASK))
1356                         *bmc |= NEEDED_MASK;
1357
1358                 (*bmc)--;
1359                 if (*bmc <= 2) {
1360                         set_page_attr(bitmap,
1361                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1362                                       BITMAP_PAGE_CLEAN);
1363                 }
1364                 spin_unlock_irqrestore(&bitmap->lock, flags);
1365                 offset += blocks;
1366                 if (sectors > blocks)
1367                         sectors -= blocks;
1368                 else sectors = 0;
1369         }
1370 }
1371
1372 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1373                         int degraded)
1374 {
1375         bitmap_counter_t *bmc;
1376         int rv;
1377         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1378                 *blocks = 1024;
1379                 return 1; /* always resync if no bitmap */
1380         }
1381         spin_lock_irq(&bitmap->lock);
1382         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1383         rv = 0;
1384         if (bmc) {
1385                 /* locked */
1386                 if (RESYNC(*bmc))
1387                         rv = 1;
1388                 else if (NEEDED(*bmc)) {
1389                         rv = 1;
1390                         if (!degraded) { /* don't set/clear bits if degraded */
1391                                 *bmc |= RESYNC_MASK;
1392                                 *bmc &= ~NEEDED_MASK;
1393                         }
1394                 }
1395         }
1396         spin_unlock_irq(&bitmap->lock);
1397         return rv;
1398 }
1399
1400 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1401 {
1402         bitmap_counter_t *bmc;
1403         unsigned long flags;
1404 /*
1405         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1406 */      if (bitmap == NULL) {
1407                 *blocks = 1024;
1408                 return;
1409         }
1410         spin_lock_irqsave(&bitmap->lock, flags);
1411         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1412         if (bmc == NULL)
1413                 goto unlock;
1414         /* locked */
1415 /*
1416         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1417 */
1418         if (RESYNC(*bmc)) {
1419                 *bmc &= ~RESYNC_MASK;
1420
1421                 if (!NEEDED(*bmc) && aborted)
1422                         *bmc |= NEEDED_MASK;
1423                 else {
1424                         if (*bmc <= 2) {
1425                                 set_page_attr(bitmap,
1426                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1427                                               BITMAP_PAGE_CLEAN);
1428                         }
1429                 }
1430         }
1431  unlock:
1432         spin_unlock_irqrestore(&bitmap->lock, flags);
1433 }
1434
1435 void bitmap_close_sync(struct bitmap *bitmap)
1436 {
1437         /* Sync has finished, and any bitmap chunks that weren't synced
1438          * properly have been aborted.  It remains to us to clear the
1439          * RESYNC bit wherever it is still on
1440          */
1441         sector_t sector = 0;
1442         int blocks;
1443         if (!bitmap) return;
1444         while (sector < bitmap->mddev->resync_max_sectors) {
1445                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1446 /*
1447                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1448                                          (unsigned long long)sector, blocks);
1449 */              sector += blocks;
1450         }
1451 }
1452
1453 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1454 {
1455         /* For each chunk covered by any of these sectors, set the
1456          * counter to 1 and set resync_needed.  They should all
1457          * be 0 at this point
1458          */
1459
1460         int secs;
1461         bitmap_counter_t *bmc;
1462         spin_lock_irq(&bitmap->lock);
1463         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1464         if (!bmc) {
1465                 spin_unlock_irq(&bitmap->lock);
1466                 return;
1467         }
1468         if (! *bmc) {
1469                 struct page *page;
1470                 *bmc = 1 | (needed?NEEDED_MASK:0);
1471                 bitmap_count_page(bitmap, offset, 1);
1472                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1473                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1474         }
1475         spin_unlock_irq(&bitmap->lock);
1476
1477 }
1478
1479 /*
1480  * flush out any pending updates
1481  */
1482 void bitmap_flush(mddev_t *mddev)
1483 {
1484         struct bitmap *bitmap = mddev->bitmap;
1485         int sleep;
1486
1487         if (!bitmap) /* there was no bitmap */
1488                 return;
1489
1490         /* run the daemon_work three time to ensure everything is flushed
1491          * that can be
1492          */
1493         sleep = bitmap->daemon_sleep;
1494         bitmap->daemon_sleep = 0;
1495         bitmap_daemon_work(bitmap);
1496         bitmap_daemon_work(bitmap);
1497         bitmap_daemon_work(bitmap);
1498         bitmap->daemon_sleep = sleep;
1499         bitmap_update_sb(bitmap);
1500 }
1501
1502 /*
1503  * free memory that was allocated
1504  */
1505 void bitmap_destroy(mddev_t *mddev)
1506 {
1507         unsigned long k, pages;
1508         struct bitmap_page *bp;
1509         struct bitmap *bitmap = mddev->bitmap;
1510
1511         if (!bitmap) /* there was no bitmap */
1512                 return;
1513
1514         mddev->bitmap = NULL; /* disconnect from the md device */
1515
1516         /* release the bitmap file and kill the daemon */
1517         bitmap_file_put(bitmap);
1518
1519         bp = bitmap->bp;
1520         pages = bitmap->pages;
1521
1522         /* free all allocated memory */
1523
1524         mempool_destroy(bitmap->write_pool);
1525
1526         if (bp) /* deallocate the page memory */
1527                 for (k = 0; k < pages; k++)
1528                         if (bp[k].map && !bp[k].hijacked)
1529                                 kfree(bp[k].map);
1530         kfree(bp);
1531         kfree(bitmap);
1532 }
1533
1534 /*
1535  * initialize the bitmap structure
1536  * if this returns an error, bitmap_destroy must be called to do clean up
1537  */
1538 int bitmap_create(mddev_t *mddev)
1539 {
1540         struct bitmap *bitmap;
1541         unsigned long blocks = mddev->resync_max_sectors;
1542         unsigned long chunks;
1543         unsigned long pages;
1544         struct file *file = mddev->bitmap_file;
1545         int err;
1546         sector_t start;
1547
1548         BUG_ON(sizeof(bitmap_super_t) != 256);
1549
1550         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1551                 return 0;
1552
1553         BUG_ON(file && mddev->bitmap_offset);
1554
1555         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1556         if (!bitmap)
1557                 return -ENOMEM;
1558
1559         memset(bitmap, 0, sizeof(*bitmap));
1560
1561         spin_lock_init(&bitmap->lock);
1562         bitmap->mddev = mddev;
1563         mddev->bitmap = bitmap;
1564
1565         spin_lock_init(&bitmap->write_lock);
1566         INIT_LIST_HEAD(&bitmap->complete_pages);
1567         init_waitqueue_head(&bitmap->write_wait);
1568         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1569                                 write_pool_free, NULL);
1570         if (!bitmap->write_pool)
1571                 return -ENOMEM;
1572
1573         bitmap->file = file;
1574         bitmap->offset = mddev->bitmap_offset;
1575         if (file) get_file(file);
1576         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1577         err = bitmap_read_sb(bitmap);
1578         if (err)
1579                 return err;
1580
1581         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1582                                         sizeof(bitmap->chunksize));
1583
1584         /* now that chunksize and chunkshift are set, we can use these macros */
1585         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1586                         CHUNK_BLOCK_RATIO(bitmap);
1587         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1588
1589         BUG_ON(!pages);
1590
1591         bitmap->chunks = chunks;
1592         bitmap->pages = pages;
1593         bitmap->missing_pages = pages;
1594         bitmap->counter_bits = COUNTER_BITS;
1595
1596         bitmap->syncchunk = ~0UL;
1597
1598 #ifdef INJECT_FATAL_FAULT_1
1599         bitmap->bp = NULL;
1600 #else
1601         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1602 #endif
1603         if (!bitmap->bp)
1604                 return -ENOMEM;
1605         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1606
1607         bitmap->flags |= BITMAP_ACTIVE;
1608
1609         /* now that we have some pages available, initialize the in-memory
1610          * bitmap from the on-disk bitmap */
1611         start = 0;
1612         if (mddev->degraded == 0
1613             || bitmap->events_cleared == mddev->events)
1614                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1615                 start = mddev->recovery_cp;
1616         err = bitmap_init_from_disk(bitmap, start);
1617
1618         if (err)
1619                 return err;
1620
1621         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1622                 pages, bmname(bitmap));
1623
1624         /* kick off the bitmap daemons */
1625         err = bitmap_start_daemons(bitmap);
1626         if (err)
1627                 return err;
1628         return bitmap_update_sb(bitmap);
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);