git.oblomov.eu Git - linux-2.6/blob - drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/version.h>
  25 #include <linux/errno.h>
  26 #include <linux/slab.h>
  27 #include <linux/init.h>
  28 #include <linux/config.h>
  29 #include <linux/timer.h>
  30 #include <linux/sched.h>
  31 #include <linux/list.h>
  32 #include <linux/file.h>
  33 #include <linux/mount.h>
  34 #include <linux/buffer_head.h>
  35 #include <linux/raid/md.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 /* debug macros */
  39
  40 #define DEBUG 0
  41
  42 #if DEBUG
  43 /* these are for debugging purposes only! */
  44
  45 /* define one and only one of these */
  46 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  47 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  48 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  49 #define INJECT_FAULTS_4 0 /* undef */
  50 #define INJECT_FAULTS_5 0 /* undef */
  51 #define INJECT_FAULTS_6 0
  52
  53 /* if these are defined, the driver will fail! debug only */
  54 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  55 #define INJECT_FATAL_FAULT_2 0 /* undef */
  56 #define INJECT_FATAL_FAULT_3 0 /* undef */
  57 #endif
  58
  59 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  60 #define DPRINTK(x...) do { } while(0)
  61
  62 #ifndef PRINTK
  63 #  if DEBUG > 0
  64 #    define PRINTK(x...) printk(KERN_DEBUG x)
  65 #  else
  66 #    define PRINTK(x...)
  67 #  endif
  68 #endif
  69
  70 static inline char * bmname(struct bitmap *bitmap)
  71 {
  72         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  73 }
  74
  75
  76 /*
  77  * test if the bitmap is active
  78  */
  79 int bitmap_active(struct bitmap *bitmap)
  80 {
  81         unsigned long flags;
  82         int res = 0;
  83
  84         if (!bitmap)
  85                 return res;
  86         spin_lock_irqsave(&bitmap->lock, flags);
  87         res = bitmap->flags & BITMAP_ACTIVE;
  88         spin_unlock_irqrestore(&bitmap->lock, flags);
  89         return res;
  90 }
  91
  92 #define WRITE_POOL_SIZE 256
  93 /* mempool for queueing pending writes on the bitmap file */
  94 static void *write_pool_alloc(unsigned int gfp_flags, void *data)
  95 {
  96         return kmalloc(sizeof(struct page_list), gfp_flags);
  97 }
  98
  99 static void write_pool_free(void *ptr, void *data)
 100 {
 101         kfree(ptr);
 102 }
 103
 104 /*
 105  * just a placeholder - calls kmalloc for bitmap pages
 106  */
 107 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 108 {
 109         unsigned char *page;
 110
 111 #if INJECT_FAULTS_1
 112         page = NULL;
 113 #else
 114         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 115 #endif
 116         if (!page)
 117                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 118         else
 119                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 120                         bmname(bitmap), page);
 121         return page;
 122 }
 123
 124 /*
 125  * for now just a placeholder -- just calls kfree for bitmap pages
 126  */
 127 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 128 {
 129         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 130         kfree(page);
 131 }
 132
 133 /*
 134  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 135  *
 136  * 1) check to see if this page is allocated, if it's not then try to alloc
 137  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 138  *    page pointer directly as a counter
 139  *
 140  * if we find our page, we increment the page's refcount so that it stays
 141  * allocated while we're using it
 142  */
 143 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 144 {
 145         unsigned char *mappage;
 146
 147         if (page >= bitmap->pages) {
 148                 printk(KERN_ALERT
 149                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 150                         bmname(bitmap), page, bitmap->pages-1);
 151                 return -EINVAL;
 152         }
 153
 154
 155         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 156                 return 0;
 157
 158         if (bitmap->bp[page].map) /* page is already allocated, just return */
 159                 return 0;
 160
 161         if (!create)
 162                 return -ENOENT;
 163
 164         spin_unlock_irq(&bitmap->lock);
 165
 166         /* this page has not been allocated yet */
 167
 168         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 169                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 170                         bmname(bitmap));
 171                 /* failed - set the hijacked flag so that we can use the
 172                  * pointer as a counter */
 173                 spin_lock_irq(&bitmap->lock);
 174                 if (!bitmap->bp[page].map)
 175                         bitmap->bp[page].hijacked = 1;
 176                 goto out;
 177         }
 178
 179         /* got a page */
 180
 181         spin_lock_irq(&bitmap->lock);
 182
 183         /* recheck the page */
 184
 185         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 186                 /* somebody beat us to getting the page */
 187                 bitmap_free_page(bitmap, mappage);
 188                 return 0;
 189         }
 190
 191         /* no page was in place and we have one, so install it */
 192
 193         memset(mappage, 0, PAGE_SIZE);
 194         bitmap->bp[page].map = mappage;
 195         bitmap->missing_pages--;
 196 out:
 197         return 0;
 198 }
 199
 200
 201 /* if page is completely empty, put it back on the free list, or dealloc it */
 202 /* if page was hijacked, unmark the flag so it might get alloced next time */
 203 /* Note: lock should be held when calling this */
 204 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 205 {
 206         char *ptr;
 207
 208         if (bitmap->bp[page].count) /* page is still busy */
 209                 return;
 210
 211         /* page is no longer in use, it can be released */
 212
 213         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 214                 bitmap->bp[page].hijacked = 0;
 215                 bitmap->bp[page].map = NULL;
 216                 return;
 217         }
 218
 219         /* normal case, free the page */
 220
 221 #if 0
 222 /* actually ... let's not.  We will probably need the page again exactly when
 223  * memory is tight and we are flusing to disk
 224  */
 225         return;
 226 #else
 227         ptr = bitmap->bp[page].map;
 228         bitmap->bp[page].map = NULL;
 229         bitmap->missing_pages++;
 230         bitmap_free_page(bitmap, ptr);
 231         return;
 232 #endif
 233 }
 234
 235
 236 /*
 237  * bitmap file handling - read and write the bitmap file and its superblock
 238  */
 239
 240 /* copy the pathname of a file to a buffer */
 241 char *file_path(struct file *file, char *buf, int count)
 242 {
 243         struct dentry *d;
 244         struct vfsmount *v;
 245
 246         if (!buf)
 247                 return NULL;
 248
 249         d = file->f_dentry;
 250         v = file->f_vfsmnt;
 251
 252         buf = d_path(d, v, buf, count);
 253
 254         return IS_ERR(buf) ? NULL : buf;
 255 }
 256
 257 /*
 258  * basic page I/O operations
 259  */
 260
 261 /* IO operations when bitmap is stored near all superblocks */
 262 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 263 {
 264         /* choose a good rdev and read the page from there */
 265
 266         mdk_rdev_t *rdev;
 267         struct list_head *tmp;
 268         struct page *page = alloc_page(GFP_KERNEL);
 269         sector_t target;
 270
 271         if (!page)
 272                 return ERR_PTR(-ENOMEM);
 273         do {
 274                 ITERATE_RDEV(mddev, rdev, tmp)
 275                         if (rdev->in_sync && !rdev->faulty)
 276                                 goto found;
 277                 return ERR_PTR(-EIO);
 278
 279         found:
 280                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 281
 282         } while (!sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ));
 283
 284         page->index = index;
 285         return page;
 286 }
 287
 288 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 289 {
 290         mdk_rdev_t *rdev;
 291         struct list_head *tmp;
 292
 293         ITERATE_RDEV(mddev, rdev, tmp)
 294                 if (rdev->in_sync && !rdev->faulty)
 295                         md_super_write(mddev, rdev,
 296                                        (rdev->sb_offset<<1) + offset
 297                                        + page->index * (PAGE_SIZE/512),
 298                                        PAGE_SIZE,
 299                                        page);
 300
 301         if (wait)
 302                 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 303         return 0;
 304 }
 305
 306 /*
 307  * write out a page to a file
 308  */
 309 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 310 {
 311         int ret = -ENOMEM;
 312
 313         if (bitmap->file == NULL)
 314                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 315
 316         if (wait)
 317                 lock_page(page);
 318         else {
 319                 if (TestSetPageLocked(page))
 320                         return -EAGAIN; /* already locked */
 321                 if (PageWriteback(page)) {
 322                         unlock_page(page);
 323                         return -EAGAIN;
 324                 }
 325         }
 326
 327         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 328         if (!ret)
 329                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 330                         PAGE_SIZE);
 331         if (ret) {
 332                 unlock_page(page);
 333                 return ret;
 334         }
 335
 336         set_page_dirty(page); /* force it to be written out */
 337
 338         if (!wait) {
 339                 /* add to list to be waited for by daemon */
 340                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 341                 item->page = page;
 342                 page_cache_get(page);
 343                 spin_lock(&bitmap->write_lock);
 344                 list_add(&item->list, &bitmap->complete_pages);
 345                 spin_unlock(&bitmap->write_lock);
 346                 md_wakeup_thread(bitmap->writeback_daemon);
 347         }
 348         return write_one_page(page, wait);
 349 }
 350
 351 /* read a page from a file, pinning it into cache, and return bytes_read */
 352 static struct page *read_page(struct file *file, unsigned long index,
 353                                         unsigned long *bytes_read)
 354 {
 355         struct inode *inode = file->f_mapping->host;
 356         struct page *page = NULL;
 357         loff_t isize = i_size_read(inode);
 358         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 359
 360         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 361                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 362
 363         page = read_cache_page(inode->i_mapping, index,
 364                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 365         if (IS_ERR(page))
 366                 goto out;
 367         wait_on_page_locked(page);
 368         if (!PageUptodate(page) || PageError(page)) {
 369                 page_cache_release(page);
 370                 page = ERR_PTR(-EIO);
 371                 goto out;
 372         }
 373
 374         if (index > end_index) /* we have read beyond EOF */
 375                 *bytes_read = 0;
 376         else if (index == end_index) /* possible short read */
 377                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 378         else
 379                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 380 out:
 381         if (IS_ERR(page))
 382                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 383                         (int)PAGE_CACHE_SIZE,
 384                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 385                         PTR_ERR(page));
 386         return page;
 387 }
 388
 389 /*
 390  * bitmap file superblock operations
 391  */
 392
 393 /* update the event counter and sync the superblock to disk */
 394 int bitmap_update_sb(struct bitmap *bitmap)
 395 {
 396         bitmap_super_t *sb;
 397         unsigned long flags;
 398
 399         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 400                 return 0;
 401         spin_lock_irqsave(&bitmap->lock, flags);
 402         if (!bitmap->sb_page) { /* no superblock */
 403                 spin_unlock_irqrestore(&bitmap->lock, flags);
 404                 return 0;
 405         }
 406         spin_unlock_irqrestore(&bitmap->lock, flags);
 407         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 408         sb->events = cpu_to_le64(bitmap->mddev->events);
 409         if (!bitmap->mddev->degraded)
 410                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 411         kunmap(bitmap->sb_page);
 412         return write_page(bitmap, bitmap->sb_page, 1);
 413 }
 414
 415 /* print out the bitmap file superblock */
 416 void bitmap_print_sb(struct bitmap *bitmap)
 417 {
 418         bitmap_super_t *sb;
 419
 420         if (!bitmap || !bitmap->sb_page)
 421                 return;
 422         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 423         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 424         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 425         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 426         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 427                                         *(__u32 *)(sb->uuid+0),
 428                                         *(__u32 *)(sb->uuid+4),
 429                                         *(__u32 *)(sb->uuid+8),
 430                                         *(__u32 *)(sb->uuid+12));
 431         printk(KERN_DEBUG "        events: %llu\n",
 432                         (unsigned long long) le64_to_cpu(sb->events));
 433         printk(KERN_DEBUG "events cleared: %llu\n",
 434                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 435         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 436         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 437         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 438         printk(KERN_DEBUG "     sync size: %llu KB\n",
 439                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 440         kunmap(bitmap->sb_page);
 441 }
 442
 443 /* read the superblock from the bitmap file and initialize some bitmap fields */
 444 static int bitmap_read_sb(struct bitmap *bitmap)
 445 {
 446         char *reason = NULL;
 447         bitmap_super_t *sb;
 448         unsigned long chunksize, daemon_sleep;
 449         unsigned long bytes_read;
 450         unsigned long long events;
 451         int err = -EINVAL;
 452
 453         /* page 0 is the superblock, read it... */
 454         if (bitmap->file)
 455                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 456         else {
 457                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 458                 bytes_read = PAGE_SIZE;
 459         }
 460         if (IS_ERR(bitmap->sb_page)) {
 461                 err = PTR_ERR(bitmap->sb_page);
 462                 bitmap->sb_page = NULL;
 463                 return err;
 464         }
 465
 466         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 467
 468         if (bytes_read < sizeof(*sb)) { /* short read */
 469                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 470                         bmname(bitmap));
 471                 err = -ENOSPC;
 472                 goto out;
 473         }
 474
 475         chunksize = le32_to_cpu(sb->chunksize);
 476         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 477
 478         /* verify that the bitmap-specific fields are valid */
 479         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 480                 reason = "bad magic";
 481         else if (sb->version != cpu_to_le32(BITMAP_MAJOR))
 482                 reason = "unrecognized superblock version";
 483         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 484                 reason = "bitmap chunksize out of range (512B - 4MB)";
 485         else if ((1 << ffz(~chunksize)) != chunksize)
 486                 reason = "bitmap chunksize not a power of 2";
 487         else if (daemon_sleep < 1 || daemon_sleep > 15)
 488                 reason = "daemon sleep period out of range";
 489         if (reason) {
 490                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 491                         bmname(bitmap), reason);
 492                 goto out;
 493         }
 494
 495         /* keep the array size field of the bitmap superblock up to date */
 496         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 497
 498         if (!bitmap->mddev->persistent)
 499                 goto success;
 500
 501         /*
 502          * if we have a persistent array superblock, compare the
 503          * bitmap's UUID and event counter to the mddev's
 504          */
 505         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 506                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 507                         bmname(bitmap));
 508                 goto out;
 509         }
 510         events = le64_to_cpu(sb->events);
 511         if (events < bitmap->mddev->events) {
 512                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 513                         "-- forcing full recovery\n", bmname(bitmap), events,
 514                         (unsigned long long) bitmap->mddev->events);
 515                 sb->state |= BITMAP_STALE;
 516         }
 517 success:
 518         /* assign fields using values from superblock */
 519         bitmap->chunksize = chunksize;
 520         bitmap->daemon_sleep = daemon_sleep;
 521         bitmap->flags |= sb->state;
 522         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 523         err = 0;
 524 out:
 525         kunmap(bitmap->sb_page);
 526         if (err)
 527                 bitmap_print_sb(bitmap);
 528         return err;
 529 }
 530
 531 enum bitmap_mask_op {
 532         MASK_SET,
 533         MASK_UNSET
 534 };
 535
 536 /* record the state of the bitmap in the superblock */
 537 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 538                                 enum bitmap_mask_op op)
 539 {
 540         bitmap_super_t *sb;
 541         unsigned long flags;
 542
 543         spin_lock_irqsave(&bitmap->lock, flags);
 544         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 545                 spin_unlock_irqrestore(&bitmap->lock, flags);
 546                 return;
 547         }
 548         page_cache_get(bitmap->sb_page);
 549         spin_unlock_irqrestore(&bitmap->lock, flags);
 550         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 551         switch (op) {
 552                 case MASK_SET: sb->state |= bits;
 553                                 break;
 554                 case MASK_UNSET: sb->state &= ~bits;
 555                                 break;
 556                 default: BUG();
 557         }
 558         kunmap(bitmap->sb_page);
 559         page_cache_release(bitmap->sb_page);
 560 }
 561
 562 /*
 563  * general bitmap file operations
 564  */
 565
 566 /* calculate the index of the page that contains this bit */
 567 static inline unsigned long file_page_index(unsigned long chunk)
 568 {
 569         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 570 }
 571
 572 /* calculate the (bit) offset of this bit within a page */
 573 static inline unsigned long file_page_offset(unsigned long chunk)
 574 {
 575         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 576 }
 577
 578 /*
 579  * return a pointer to the page in the filemap that contains the given bit
 580  *
 581  * this lookup is complicated by the fact that the bitmap sb might be exactly
 582  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 583  * 0 or page 1
 584  */
 585 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 586                                         unsigned long chunk)
 587 {
 588         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 589 }
 590
 591
 592 static void bitmap_file_unmap(struct bitmap *bitmap)
 593 {
 594         struct page **map, *sb_page;
 595         unsigned long *attr;
 596         int pages;
 597         unsigned long flags;
 598
 599         spin_lock_irqsave(&bitmap->lock, flags);
 600         map = bitmap->filemap;
 601         bitmap->filemap = NULL;
 602         attr = bitmap->filemap_attr;
 603         bitmap->filemap_attr = NULL;
 604         pages = bitmap->file_pages;
 605         bitmap->file_pages = 0;
 606         sb_page = bitmap->sb_page;
 607         bitmap->sb_page = NULL;
 608         spin_unlock_irqrestore(&bitmap->lock, flags);
 609
 610         while (pages--)
 611                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 612                         page_cache_release(map[pages]);
 613         kfree(map);
 614         kfree(attr);
 615
 616         if (sb_page)
 617                 page_cache_release(sb_page);
 618 }
 619
 620 static void bitmap_stop_daemons(struct bitmap *bitmap);
 621
 622 /* dequeue the next item in a page list -- don't call from irq context */
 623 static struct page_list *dequeue_page(struct bitmap *bitmap)
 624 {
 625         struct page_list *item = NULL;
 626         struct list_head *head = &bitmap->complete_pages;
 627
 628         spin_lock(&bitmap->write_lock);
 629         if (list_empty(head))
 630                 goto out;
 631         item = list_entry(head->prev, struct page_list, list);
 632         list_del(head->prev);
 633 out:
 634         spin_unlock(&bitmap->write_lock);
 635         return item;
 636 }
 637
 638 static void drain_write_queues(struct bitmap *bitmap)
 639 {
 640         struct page_list *item;
 641
 642         while ((item = dequeue_page(bitmap))) {
 643                 /* don't bother to wait */
 644                 page_cache_release(item->page);
 645                 mempool_free(item, bitmap->write_pool);
 646         }
 647
 648         wake_up(&bitmap->write_wait);
 649 }
 650
 651 static void bitmap_file_put(struct bitmap *bitmap)
 652 {
 653         struct file *file;
 654         struct inode *inode;
 655         unsigned long flags;
 656
 657         spin_lock_irqsave(&bitmap->lock, flags);
 658         file = bitmap->file;
 659         bitmap->file = NULL;
 660         spin_unlock_irqrestore(&bitmap->lock, flags);
 661
 662         bitmap_stop_daemons(bitmap);
 663
 664         drain_write_queues(bitmap);
 665
 666         bitmap_file_unmap(bitmap);
 667
 668         if (file) {
 669                 inode = file->f_mapping->host;
 670                 spin_lock(&inode->i_lock);
 671                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 672                 spin_unlock(&inode->i_lock);
 673                 fput(file);
 674         }
 675 }
 676
 677
 678 /*
 679  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 680  * then it is no longer reliable, so we stop using it and we mark the file
 681  * as failed in the superblock
 682  */
 683 static void bitmap_file_kick(struct bitmap *bitmap)
 684 {
 685         char *path, *ptr = NULL;
 686
 687         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 688         bitmap_update_sb(bitmap);
 689
 690         if (bitmap->file) {
 691                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 692                 if (path)
 693                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 694
 695                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 696                        bmname(bitmap), ptr ? ptr : "");
 697
 698                 kfree(path);
 699         }
 700
 701         bitmap_file_put(bitmap);
 702
 703         return;
 704 }
 705
 706 enum bitmap_page_attr {
 707         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 708         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 709         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 710 };
 711
 712 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 713                                 enum bitmap_page_attr attr)
 714 {
 715         bitmap->filemap_attr[page->index] |= attr;
 716 }
 717
 718 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 719                                 enum bitmap_page_attr attr)
 720 {
 721         bitmap->filemap_attr[page->index] &= ~attr;
 722 }
 723
 724 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 725 {
 726         return bitmap->filemap_attr[page->index];
 727 }
 728
 729 /*
 730  * bitmap_file_set_bit -- called before performing a write to the md device
 731  * to set (and eventually sync) a particular bit in the bitmap file
 732  *
 733  * we set the bit immediately, then we record the page number so that
 734  * when an unplug occurs, we can flush the dirty pages out to disk
 735  */
 736 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 737 {
 738         unsigned long bit;
 739         struct page *page;
 740         void *kaddr;
 741         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 742
 743         if (!bitmap->filemap) {
 744                 return;
 745         }
 746
 747         page = filemap_get_page(bitmap, chunk);
 748         bit = file_page_offset(chunk);
 749
 750
 751         /* make sure the page stays cached until it gets written out */
 752         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 753                 page_cache_get(page);
 754
 755         /* set the bit */
 756         kaddr = kmap_atomic(page, KM_USER0);
 757         set_bit(bit, kaddr);
 758         kunmap_atomic(kaddr, KM_USER0);
 759         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 760
 761         /* record page number so it gets flushed to disk when unplug occurs */
 762         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 763
 764 }
 765
 766 /* this gets called when the md device is ready to unplug its underlying
 767  * (slave) device queues -- before we let any writes go down, we need to
 768  * sync the dirty pages of the bitmap file to disk */
 769 int bitmap_unplug(struct bitmap *bitmap)
 770 {
 771         unsigned long i, attr, flags;
 772         struct page *page;
 773         int wait = 0;
 774         int err;
 775
 776         if (!bitmap)
 777                 return 0;
 778
 779         /* look at each page to see if there are any set bits that need to be
 780          * flushed out to disk */
 781         for (i = 0; i < bitmap->file_pages; i++) {
 782                 spin_lock_irqsave(&bitmap->lock, flags);
 783                 if (!bitmap->filemap) {
 784                         spin_unlock_irqrestore(&bitmap->lock, flags);
 785                         return 0;
 786                 }
 787                 page = bitmap->filemap[i];
 788                 attr = get_page_attr(bitmap, page);
 789                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 790                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 791                 if ((attr & BITMAP_PAGE_DIRTY))
 792                         wait = 1;
 793                 spin_unlock_irqrestore(&bitmap->lock, flags);
 794
 795                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 796                         err = write_page(bitmap, page, 0);
 797                         if (err == -EAGAIN) {
 798                                 if (attr & BITMAP_PAGE_DIRTY)
 799                                         err = write_page(bitmap, page, 1);
 800                                 else
 801                                         err = 0;
 802                         }
 803                         if (err)
 804                                 return 1;
 805                 }
 806         }
 807         if (wait) { /* if any writes were performed, we need to wait on them */
 808                 if (bitmap->file) {
 809                         spin_lock_irq(&bitmap->write_lock);
 810                         wait_event_lock_irq(bitmap->write_wait,
 811                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 812                                             wake_up_process(bitmap->writeback_daemon->tsk));
 813                         spin_unlock_irq(&bitmap->write_lock);
 814                 } else
 815                         wait_event(bitmap->mddev->sb_wait,
 816                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 817         }
 818         return 0;
 819 }
 820
 821 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset,
 822         unsigned long sectors, int in_sync);
 823 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 824  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 825  * memory mapping of the bitmap file
 826  * Special cases:
 827  *   if there's no bitmap file, or if the bitmap file had been
 828  *   previously kicked from the array, we mark all the bits as
 829  *   1's in order to cause a full resync.
 830  */
 831 static int bitmap_init_from_disk(struct bitmap *bitmap, int in_sync)
 832 {
 833         unsigned long i, chunks, index, oldindex, bit;
 834         struct page *page = NULL, *oldpage = NULL;
 835         unsigned long num_pages, bit_cnt = 0;
 836         struct file *file;
 837         unsigned long bytes, offset, dummy;
 838         int outofdate;
 839         int ret = -ENOSPC;
 840
 841         chunks = bitmap->chunks;
 842         file = bitmap->file;
 843
 844         BUG_ON(!file && !bitmap->offset);
 845
 846 #if INJECT_FAULTS_3
 847         outofdate = 1;
 848 #else
 849         outofdate = bitmap->flags & BITMAP_STALE;
 850 #endif
 851         if (outofdate)
 852                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 853                         "recovery\n", bmname(bitmap));
 854
 855         bytes = (chunks + 7) / 8;
 856
 857         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 858
 859         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 860                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 861                         bmname(bitmap),
 862                         (unsigned long) i_size_read(file->f_mapping->host),
 863                         bytes + sizeof(bitmap_super_t));
 864                 goto out;
 865         }
 866
 867         ret = -ENOMEM;
 868
 869         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 870         if (!bitmap->filemap)
 871                 goto out;
 872
 873         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 874         if (!bitmap->filemap_attr)
 875                 goto out;
 876
 877         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 878
 879         oldindex = ~0L;
 880
 881         for (i = 0; i < chunks; i++) {
 882                 index = file_page_index(i);
 883                 bit = file_page_offset(i);
 884                 if (index != oldindex) { /* this is a new page, read it in */
 885                         /* unmap the old page, we're done with it */
 886                         if (oldpage != NULL)
 887                                 kunmap(oldpage);
 888                         if (index == 0) {
 889                                 /*
 890                                  * if we're here then the superblock page
 891                                  * contains some bits (PAGE_SIZE != sizeof sb)
 892                                  * we've already read it in, so just use it
 893                                  */
 894                                 page = bitmap->sb_page;
 895                                 offset = sizeof(bitmap_super_t);
 896                         } else if (file) {
 897                                 page = read_page(file, index, &dummy);
 898                                 offset = 0;
 899                         } else {
 900                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 901                                 offset = 0;
 902                         }
 903                         if (IS_ERR(page)) { /* read error */
 904                                 ret = PTR_ERR(page);
 905                                 goto out;
 906                         }
 907
 908                         oldindex = index;
 909                         oldpage = page;
 910                         kmap(page);
 911
 912                         if (outofdate) {
 913                                 /*
 914                                  * if bitmap is out of date, dirty the
 915                                  * whole page and write it out
 916                                  */
 917                                 memset(page_address(page) + offset, 0xff,
 918                                         PAGE_SIZE - offset);
 919                                 ret = write_page(bitmap, page, 1);
 920                                 if (ret) {
 921                                         kunmap(page);
 922                                         /* release, page not in filemap yet */
 923                                         page_cache_release(page);
 924                                         goto out;
 925                                 }
 926                         }
 927
 928                         bitmap->filemap[bitmap->file_pages++] = page;
 929                 }
 930                 if (test_bit(bit, page_address(page))) {
 931                         /* if the disk bit is set, set the memory bit */
 932                         bitmap_set_memory_bits(bitmap,
 933                                         i << CHUNK_BLOCK_SHIFT(bitmap), 1, in_sync);
 934                         bit_cnt++;
 935                 }
 936         }
 937
 938         /* everything went OK */
 939         ret = 0;
 940         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 941
 942         if (page) /* unmap the last page */
 943                 kunmap(page);
 944
 945         if (bit_cnt) { /* Kick recovery if any bits were set */
 946                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 947                 md_wakeup_thread(bitmap->mddev->thread);
 948         }
 949
 950 out:
 951         printk(KERN_INFO "%s: bitmap initialized from disk: "
 952                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 953                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 954
 955         return ret;
 956 }
 957
 958 void bitmap_write_all(struct bitmap *bitmap)
 959 {
 960         /* We don't actually write all bitmap blocks here,
 961          * just flag them as needing to be written
 962          */
 963
 964         unsigned long chunks = bitmap->chunks;
 965         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 966         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 967         while (num_pages--)
 968                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 969 }
 970
 971
 972 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 973 {
 974         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 975         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 976         bitmap->bp[page].count += inc;
 977 /*
 978         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 979                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 980 */
 981         bitmap_checkfree(bitmap, page);
 982 }
 983 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 984                                             sector_t offset, int *blocks,
 985                                             int create);
 986
 987 /*
 988  * bitmap daemon -- periodically wakes up to clean bits and flush pages
 989  *                      out to disk
 990  */
 991
 992 int bitmap_daemon_work(struct bitmap *bitmap)
 993 {
 994         unsigned long j;
 995         unsigned long flags;
 996         struct page *page = NULL, *lastpage = NULL;
 997         int err = 0;
 998         int blocks;
 999         int attr;
1000
1001         if (bitmap == NULL)
1002                 return 0;
1003         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1004                 return 0;
1005         bitmap->daemon_lastrun = jiffies;
1006
1007         for (j = 0; j < bitmap->chunks; j++) {
1008                 bitmap_counter_t *bmc;
1009                 spin_lock_irqsave(&bitmap->lock, flags);
1010                 if (!bitmap->filemap) {
1011                         /* error or shutdown */
1012                         spin_unlock_irqrestore(&bitmap->lock, flags);
1013                         break;
1014                 }
1015
1016                 page = filemap_get_page(bitmap, j);
1017
1018                 if (page != lastpage) {
1019                         /* skip this page unless it's marked as needing cleaning */
1020                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1021                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1022                                         page_cache_get(page);
1023                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1024                                 }
1025                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1026                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1027                                         switch (write_page(bitmap, page, 0)) {
1028                                         case -EAGAIN:
1029                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1030                                                 break;
1031                                         case 0:
1032                                                 break;
1033                                         default:
1034                                                 bitmap_file_kick(bitmap);
1035                                         }
1036                                         page_cache_release(page);
1037                                 }
1038                                 continue;
1039                         }
1040
1041                         /* grab the new page, sync and release the old */
1042                         page_cache_get(page);
1043                         if (lastpage != NULL) {
1044                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1045                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1046                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1047                                         err = write_page(bitmap, lastpage, 0);
1048                                         if (err == -EAGAIN) {
1049                                                 err = 0;
1050                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1051                                         }
1052                                 } else {
1053                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1054                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1055                                 }
1056                                 kunmap(lastpage);
1057                                 page_cache_release(lastpage);
1058                                 if (err)
1059                                         bitmap_file_kick(bitmap);
1060                         } else
1061                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1062                         lastpage = page;
1063                         kmap(page);
1064 /*
1065                         printk("bitmap clean at page %lu\n", j);
1066 */
1067                         spin_lock_irqsave(&bitmap->lock, flags);
1068                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1069                 }
1070                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1071                                         &blocks, 0);
1072                 if (bmc) {
1073 /*
1074   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1075 */
1076                         if (*bmc == 2) {
1077                                 *bmc=1; /* maybe clear the bit next time */
1078                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1079                         } else if (*bmc == 1) {
1080                                 /* we can clear the bit */
1081                                 *bmc = 0;
1082                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1083                                                   -1);
1084
1085                                 /* clear the bit */
1086                                 clear_bit(file_page_offset(j), page_address(page));
1087                         }
1088                 }
1089                 spin_unlock_irqrestore(&bitmap->lock, flags);
1090         }
1091
1092         /* now sync the final page */
1093         if (lastpage != NULL) {
1094                 kunmap(lastpage);
1095                 spin_lock_irqsave(&bitmap->lock, flags);
1096                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1097                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1098                         spin_unlock_irqrestore(&bitmap->lock, flags);
1099                         err = write_page(bitmap, lastpage, 0);
1100                         if (err == -EAGAIN) {
1101                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1102                                 err = 0;
1103                         }
1104                 } else {
1105                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1106                         spin_unlock_irqrestore(&bitmap->lock, flags);
1107                 }
1108
1109                 page_cache_release(lastpage);
1110         }
1111
1112         return err;
1113 }
1114
1115 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1116 {
1117         mdk_thread_t *dmn;
1118         unsigned long flags;
1119
1120         /* if no one is waiting on us, we'll free the md thread struct
1121          * and exit, otherwise we let the waiter clean things up */
1122         spin_lock_irqsave(&bitmap->lock, flags);
1123         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1124                 *daemon = NULL;
1125                 spin_unlock_irqrestore(&bitmap->lock, flags);
1126                 kfree(dmn);
1127                 complete_and_exit(NULL, 0); /* do_exit not exported */
1128         }
1129         spin_unlock_irqrestore(&bitmap->lock, flags);
1130 }
1131
1132 static void bitmap_writeback_daemon(mddev_t *mddev)
1133 {
1134         struct bitmap *bitmap = mddev->bitmap;
1135         struct page *page;
1136         struct page_list *item;
1137         int err = 0;
1138
1139         if (signal_pending(current)) {
1140                 printk(KERN_INFO
1141                        "%s: bitmap writeback daemon got signal, exiting...\n",
1142                        bmname(bitmap));
1143                 err = -EINTR;
1144                 goto out;
1145         }
1146
1147         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1148         /* wait on bitmap page writebacks */
1149         while ((item = dequeue_page(bitmap))) {
1150                 page = item->page;
1151                 mempool_free(item, bitmap->write_pool);
1152                 PRINTK("wait on page writeback: %p\n", page);
1153                 wait_on_page_writeback(page);
1154                 PRINTK("finished page writeback: %p\n", page);
1155
1156                 err = PageError(page);
1157                 page_cache_release(page);
1158                 if (err) {
1159                         printk(KERN_WARNING "%s: bitmap file writeback "
1160                                "failed (page %lu): %d\n",
1161                                bmname(bitmap), page->index, err);
1162                         bitmap_file_kick(bitmap);
1163                         goto out;
1164                 }
1165         }
1166  out:
1167         wake_up(&bitmap->write_wait);
1168         if (err) {
1169                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1170                        bmname(bitmap), err);
1171                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1172         }
1173 }
1174
1175 static int bitmap_start_daemon(struct bitmap *bitmap, mdk_thread_t **ptr,
1176                                 void (*func)(mddev_t *), char *name)
1177 {
1178         mdk_thread_t *daemon;
1179         unsigned long flags;
1180         char namebuf[32];
1181
1182         spin_lock_irqsave(&bitmap->lock, flags);
1183         *ptr = NULL;
1184
1185         if (!bitmap->file) /* no need for daemon if there's no backing file */
1186                 goto out_unlock;
1187
1188         spin_unlock_irqrestore(&bitmap->lock, flags);
1189
1190 #if INJECT_FATAL_FAULT_2
1191         daemon = NULL;
1192 #else
1193         sprintf(namebuf, "%%s_%s", name);
1194         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1195 #endif
1196         if (!daemon) {
1197                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1198                         bmname(bitmap));
1199                 return -ECHILD;
1200         }
1201
1202         spin_lock_irqsave(&bitmap->lock, flags);
1203         *ptr = daemon;
1204
1205         md_wakeup_thread(daemon); /* start it running */
1206
1207         PRINTK("%s: %s daemon (pid %d) started...\n",
1208                 bmname(bitmap), name, daemon->tsk->pid);
1209 out_unlock:
1210         spin_unlock_irqrestore(&bitmap->lock, flags);
1211         return 0;
1212 }
1213
1214 static int bitmap_start_daemons(struct bitmap *bitmap)
1215 {
1216         int err = bitmap_start_daemon(bitmap, &bitmap->writeback_daemon,
1217                                         bitmap_writeback_daemon, "bitmap_wb");
1218         return err;
1219 }
1220
1221 static void bitmap_stop_daemon(struct bitmap *bitmap, mdk_thread_t **ptr)
1222 {
1223         mdk_thread_t *daemon;
1224         unsigned long flags;
1225
1226         spin_lock_irqsave(&bitmap->lock, flags);
1227         daemon = *ptr;
1228         *ptr = NULL;
1229         spin_unlock_irqrestore(&bitmap->lock, flags);
1230         if (daemon)
1231                 md_unregister_thread(daemon); /* destroy the thread */
1232 }
1233
1234 static void bitmap_stop_daemons(struct bitmap *bitmap)
1235 {
1236         /* the daemons can't stop themselves... they'll just exit instead... */
1237         if (bitmap->writeback_daemon &&
1238             current->pid != bitmap->writeback_daemon->tsk->pid)
1239                 bitmap_stop_daemon(bitmap, &bitmap->writeback_daemon);
1240 }
1241
1242 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1243                                             sector_t offset, int *blocks,
1244                                             int create)
1245 {
1246         /* If 'create', we might release the lock and reclaim it.
1247          * The lock must have been taken with interrupts enabled.
1248          * If !create, we don't release the lock.
1249          */
1250         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1251         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1252         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1253         sector_t csize;
1254
1255         if (bitmap_checkpage(bitmap, page, create) < 0) {
1256                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1257                 *blocks = csize - (offset & (csize- 1));
1258                 return NULL;
1259         }
1260         /* now locked ... */
1261
1262         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1263                 /* should we use the first or second counter field
1264                  * of the hijacked pointer? */
1265                 int hi = (pageoff > PAGE_COUNTER_MASK);
1266                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1267                                           PAGE_COUNTER_SHIFT - 1);
1268                 *blocks = csize - (offset & (csize- 1));
1269                 return  &((bitmap_counter_t *)
1270                           &bitmap->bp[page].map)[hi];
1271         } else { /* page is allocated */
1272                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1273                 *blocks = csize - (offset & (csize- 1));
1274                 return (bitmap_counter_t *)
1275                         &(bitmap->bp[page].map[pageoff]);
1276         }
1277 }
1278
1279 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors)
1280 {
1281         if (!bitmap) return 0;
1282         while (sectors) {
1283                 int blocks;
1284                 bitmap_counter_t *bmc;
1285
1286                 spin_lock_irq(&bitmap->lock);
1287                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1288                 if (!bmc) {
1289                         spin_unlock_irq(&bitmap->lock);
1290                         return 0;
1291                 }
1292
1293                 switch(*bmc) {
1294                 case 0:
1295                         bitmap_file_set_bit(bitmap, offset);
1296                         bitmap_count_page(bitmap,offset, 1);
1297                         blk_plug_device(bitmap->mddev->queue);
1298                         /* fall through */
1299                 case 1:
1300                         *bmc = 2;
1301                 }
1302                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1303                 (*bmc)++;
1304
1305                 spin_unlock_irq(&bitmap->lock);
1306
1307                 offset += blocks;
1308                 if (sectors > blocks)
1309                         sectors -= blocks;
1310                 else sectors = 0;
1311         }
1312         return 0;
1313 }
1314
1315 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1316                      int success)
1317 {
1318         if (!bitmap) return;
1319         while (sectors) {
1320                 int blocks;
1321                 unsigned long flags;
1322                 bitmap_counter_t *bmc;
1323
1324                 spin_lock_irqsave(&bitmap->lock, flags);
1325                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1326                 if (!bmc) {
1327                         spin_unlock_irqrestore(&bitmap->lock, flags);
1328                         return;
1329                 }
1330
1331                 if (!success && ! (*bmc & NEEDED_MASK))
1332                         *bmc |= NEEDED_MASK;
1333
1334                 (*bmc)--;
1335                 if (*bmc <= 2) {
1336                         set_page_attr(bitmap,
1337                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1338                                       BITMAP_PAGE_CLEAN);
1339                 }
1340                 spin_unlock_irqrestore(&bitmap->lock, flags);
1341                 offset += blocks;
1342                 if (sectors > blocks)
1343                         sectors -= blocks;
1344                 else sectors = 0;
1345         }
1346 }
1347
1348 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1349                         int degraded)
1350 {
1351         bitmap_counter_t *bmc;
1352         int rv;
1353         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1354                 *blocks = 1024;
1355                 return 1; /* always resync if no bitmap */
1356         }
1357         spin_lock_irq(&bitmap->lock);
1358         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1359         rv = 0;
1360         if (bmc) {
1361                 /* locked */
1362                 if (RESYNC(*bmc))
1363                         rv = 1;
1364                 else if (NEEDED(*bmc)) {
1365                         rv = 1;
1366                         if (!degraded) { /* don't set/clear bits if degraded */
1367                                 *bmc |= RESYNC_MASK;
1368                                 *bmc &= ~NEEDED_MASK;
1369                         }
1370                 }
1371         }
1372         spin_unlock_irq(&bitmap->lock);
1373         return rv;
1374 }
1375
1376 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1377 {
1378         bitmap_counter_t *bmc;
1379         unsigned long flags;
1380 /*
1381         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1382 */      if (bitmap == NULL) {
1383                 *blocks = 1024;
1384                 return;
1385         }
1386         spin_lock_irqsave(&bitmap->lock, flags);
1387         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1388         if (bmc == NULL)
1389                 goto unlock;
1390         /* locked */
1391 /*
1392         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1393 */
1394         if (RESYNC(*bmc)) {
1395                 *bmc &= ~RESYNC_MASK;
1396
1397                 if (!NEEDED(*bmc) && aborted)
1398                         *bmc |= NEEDED_MASK;
1399                 else {
1400                         if (*bmc <= 2) {
1401                                 set_page_attr(bitmap,
1402                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1403                                               BITMAP_PAGE_CLEAN);
1404                         }
1405                 }
1406         }
1407  unlock:
1408         spin_unlock_irqrestore(&bitmap->lock, flags);
1409 }
1410
1411 void bitmap_close_sync(struct bitmap *bitmap)
1412 {
1413         /* Sync has finished, and any bitmap chunks that weren't synced
1414          * properly have been aborted.  It remains to us to clear the
1415          * RESYNC bit wherever it is still on
1416          */
1417         sector_t sector = 0;
1418         int blocks;
1419         if (!bitmap) return;
1420         while (sector < bitmap->mddev->resync_max_sectors) {
1421                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1422 /*
1423                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1424                                          (unsigned long long)sector, blocks);
1425 */              sector += blocks;
1426         }
1427 }
1428
1429 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset,
1430                                    unsigned long sectors, int in_sync)
1431 {
1432         /* For each chunk covered by any of these sectors, set the
1433          * counter to 1 and set resync_needed unless in_sync.  They should all
1434          * be 0 at this point
1435          */
1436         while (sectors) {
1437                 int secs;
1438                 bitmap_counter_t *bmc;
1439                 spin_lock_irq(&bitmap->lock);
1440                 bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1441                 if (!bmc) {
1442                         spin_unlock_irq(&bitmap->lock);
1443                         return;
1444                 }
1445                 if (! *bmc) {
1446                         struct page *page;
1447                         *bmc = 1 | (in_sync? 0 : NEEDED_MASK);
1448                         bitmap_count_page(bitmap, offset, 1);
1449                         page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1450                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1451                 }
1452                 spin_unlock_irq(&bitmap->lock);
1453                 if (sectors > secs)
1454                         sectors -= secs;
1455                 else
1456                         sectors = 0;
1457         }
1458 }
1459
1460 /*
1461  * free memory that was allocated
1462  */
1463 void bitmap_destroy(mddev_t *mddev)
1464 {
1465         unsigned long k, pages;
1466         struct bitmap_page *bp;
1467         struct bitmap *bitmap = mddev->bitmap;
1468
1469         if (!bitmap) /* there was no bitmap */
1470                 return;
1471
1472         mddev->bitmap = NULL; /* disconnect from the md device */
1473
1474         /* release the bitmap file and kill the daemon */
1475         bitmap_file_put(bitmap);
1476
1477         bp = bitmap->bp;
1478         pages = bitmap->pages;
1479
1480         /* free all allocated memory */
1481
1482         mempool_destroy(bitmap->write_pool);
1483
1484         if (bp) /* deallocate the page memory */
1485                 for (k = 0; k < pages; k++)
1486                         if (bp[k].map && !bp[k].hijacked)
1487                                 kfree(bp[k].map);
1488         kfree(bp);
1489         kfree(bitmap);
1490 }
1491
1492 /*
1493  * initialize the bitmap structure
1494  * if this returns an error, bitmap_destroy must be called to do clean up
1495  */
1496 int bitmap_create(mddev_t *mddev)
1497 {
1498         struct bitmap *bitmap;
1499         unsigned long blocks = mddev->resync_max_sectors;
1500         unsigned long chunks;
1501         unsigned long pages;
1502         struct file *file = mddev->bitmap_file;
1503         int err;
1504
1505         BUG_ON(sizeof(bitmap_super_t) != 256);
1506
1507         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1508                 return 0;
1509
1510         BUG_ON(file && mddev->bitmap_offset);
1511
1512         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1513         if (!bitmap)
1514                 return -ENOMEM;
1515
1516         memset(bitmap, 0, sizeof(*bitmap));
1517
1518         spin_lock_init(&bitmap->lock);
1519         bitmap->mddev = mddev;
1520         mddev->bitmap = bitmap;
1521
1522         spin_lock_init(&bitmap->write_lock);
1523         INIT_LIST_HEAD(&bitmap->complete_pages);
1524         init_waitqueue_head(&bitmap->write_wait);
1525         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1526                                 write_pool_free, NULL);
1527         if (!bitmap->write_pool)
1528                 return -ENOMEM;
1529
1530         bitmap->file = file;
1531         bitmap->offset = mddev->bitmap_offset;
1532         if (file) get_file(file);
1533         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1534         err = bitmap_read_sb(bitmap);
1535         if (err)
1536                 return err;
1537
1538         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1539                                         sizeof(bitmap->chunksize));
1540
1541         /* now that chunksize and chunkshift are set, we can use these macros */
1542         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1543                         CHUNK_BLOCK_RATIO(bitmap);
1544         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1545
1546         BUG_ON(!pages);
1547
1548         bitmap->chunks = chunks;
1549         bitmap->pages = pages;
1550         bitmap->missing_pages = pages;
1551         bitmap->counter_bits = COUNTER_BITS;
1552
1553         bitmap->syncchunk = ~0UL;
1554
1555 #if INJECT_FATAL_FAULT_1
1556         bitmap->bp = NULL;
1557 #else
1558         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1559 #endif
1560         if (!bitmap->bp)
1561                 return -ENOMEM;
1562         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1563
1564         bitmap->flags |= BITMAP_ACTIVE;
1565
1566         /* now that we have some pages available, initialize the in-memory
1567          * bitmap from the on-disk bitmap */
1568         err = bitmap_init_from_disk(bitmap, mddev->recovery_cp == MaxSector);
1569         if (err)
1570                 return err;
1571
1572         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1573                 pages, bmname(bitmap));
1574
1575         /* kick off the bitmap daemons */
1576         err = bitmap_start_daemons(bitmap);
1577         if (err)
1578                 return err;
1579         return bitmap_update_sb(bitmap);
1580 }
1581
1582 /* the bitmap API -- for raid personalities */
1583 EXPORT_SYMBOL(bitmap_startwrite);
1584 EXPORT_SYMBOL(bitmap_endwrite);
1585 EXPORT_SYMBOL(bitmap_start_sync);
1586 EXPORT_SYMBOL(bitmap_end_sync);
1587 EXPORT_SYMBOL(bitmap_unplug);
1588 EXPORT_SYMBOL(bitmap_close_sync);
1589 EXPORT_SYMBOL(bitmap_daemon_work);