drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/errno.h>
  25 #include <linux/slab.h>
  26 #include <linux/init.h>
  27 #include <linux/config.h>
  28 #include <linux/timer.h>
  29 #include <linux/sched.h>
  30 #include <linux/list.h>
  31 #include <linux/file.h>
  32 #include <linux/mount.h>
  33 #include <linux/buffer_head.h>
  34 #include <linux/raid/md.h>
  35 #include <linux/raid/bitmap.h>
  36
  37 /* debug macros */
  38
  39 #define DEBUG 0
  40
  41 #if DEBUG
  42 /* these are for debugging purposes only! */
  43
  44 /* define one and only one of these */
  45 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  46 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  47 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  48 #define INJECT_FAULTS_4 0 /* undef */
  49 #define INJECT_FAULTS_5 0 /* undef */
  50 #define INJECT_FAULTS_6 0
  51
  52 /* if these are defined, the driver will fail! debug only */
  53 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  54 #define INJECT_FATAL_FAULT_2 0 /* undef */
  55 #define INJECT_FATAL_FAULT_3 0 /* undef */
  56 #endif
  57
  58 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  59 #define DPRINTK(x...) do { } while(0)
  60
  61 #ifndef PRINTK
  62 #  if DEBUG > 0
  63 #    define PRINTK(x...) printk(KERN_DEBUG x)
  64 #  else
  65 #    define PRINTK(x...)
  66 #  endif
  67 #endif
  68
  69 static inline char * bmname(struct bitmap *bitmap)
  70 {
  71         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  72 }
  73
  74
  75 /*
  76  * test if the bitmap is active
  77  */
  78 int bitmap_active(struct bitmap *bitmap)
  79 {
  80         unsigned long flags;
  81         int res = 0;
  82
  83         if (!bitmap)
  84                 return res;
  85         spin_lock_irqsave(&bitmap->lock, flags);
  86         res = bitmap->flags & BITMAP_ACTIVE;
  87         spin_unlock_irqrestore(&bitmap->lock, flags);
  88         return res;
  89 }
  90
  91 #define WRITE_POOL_SIZE 256
  92
  93 /*
  94  * just a placeholder - calls kmalloc for bitmap pages
  95  */
  96 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
  97 {
  98         unsigned char *page;
  99
 100 #ifdef INJECT_FAULTS_1
 101         page = NULL;
 102 #else
 103         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 104 #endif
 105         if (!page)
 106                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 107         else
 108                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 109                         bmname(bitmap), page);
 110         return page;
 111 }
 112
 113 /*
 114  * for now just a placeholder -- just calls kfree for bitmap pages
 115  */
 116 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 117 {
 118         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 119         kfree(page);
 120 }
 121
 122 /*
 123  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 124  *
 125  * 1) check to see if this page is allocated, if it's not then try to alloc
 126  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 127  *    page pointer directly as a counter
 128  *
 129  * if we find our page, we increment the page's refcount so that it stays
 130  * allocated while we're using it
 131  */
 132 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 133 {
 134         unsigned char *mappage;
 135
 136         if (page >= bitmap->pages) {
 137                 printk(KERN_ALERT
 138                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 139                         bmname(bitmap), page, bitmap->pages-1);
 140                 return -EINVAL;
 141         }
 142
 143
 144         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 145                 return 0;
 146
 147         if (bitmap->bp[page].map) /* page is already allocated, just return */
 148                 return 0;
 149
 150         if (!create)
 151                 return -ENOENT;
 152
 153         spin_unlock_irq(&bitmap->lock);
 154
 155         /* this page has not been allocated yet */
 156
 157         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 158                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 159                         bmname(bitmap));
 160                 /* failed - set the hijacked flag so that we can use the
 161                  * pointer as a counter */
 162                 spin_lock_irq(&bitmap->lock);
 163                 if (!bitmap->bp[page].map)
 164                         bitmap->bp[page].hijacked = 1;
 165                 goto out;
 166         }
 167
 168         /* got a page */
 169
 170         spin_lock_irq(&bitmap->lock);
 171
 172         /* recheck the page */
 173
 174         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 175                 /* somebody beat us to getting the page */
 176                 bitmap_free_page(bitmap, mappage);
 177                 return 0;
 178         }
 179
 180         /* no page was in place and we have one, so install it */
 181
 182         memset(mappage, 0, PAGE_SIZE);
 183         bitmap->bp[page].map = mappage;
 184         bitmap->missing_pages--;
 185 out:
 186         return 0;
 187 }
 188
 189
 190 /* if page is completely empty, put it back on the free list, or dealloc it */
 191 /* if page was hijacked, unmark the flag so it might get alloced next time */
 192 /* Note: lock should be held when calling this */
 193 static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 194 {
 195         char *ptr;
 196
 197         if (bitmap->bp[page].count) /* page is still busy */
 198                 return;
 199
 200         /* page is no longer in use, it can be released */
 201
 202         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 203                 bitmap->bp[page].hijacked = 0;
 204                 bitmap->bp[page].map = NULL;
 205                 return;
 206         }
 207
 208         /* normal case, free the page */
 209
 210 #if 0
 211 /* actually ... let's not.  We will probably need the page again exactly when
 212  * memory is tight and we are flusing to disk
 213  */
 214         return;
 215 #else
 216         ptr = bitmap->bp[page].map;
 217         bitmap->bp[page].map = NULL;
 218         bitmap->missing_pages++;
 219         bitmap_free_page(bitmap, ptr);
 220         return;
 221 #endif
 222 }
 223
 224
 225 /*
 226  * bitmap file handling - read and write the bitmap file and its superblock
 227  */
 228
 229 /* copy the pathname of a file to a buffer */
 230 char *file_path(struct file *file, char *buf, int count)
 231 {
 232         struct dentry *d;
 233         struct vfsmount *v;
 234
 235         if (!buf)
 236                 return NULL;
 237
 238         d = file->f_dentry;
 239         v = file->f_vfsmnt;
 240
 241         buf = d_path(d, v, buf, count);
 242
 243         return IS_ERR(buf) ? NULL : buf;
 244 }
 245
 246 /*
 247  * basic page I/O operations
 248  */
 249
 250 /* IO operations when bitmap is stored near all superblocks */
 251 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 252 {
 253         /* choose a good rdev and read the page from there */
 254
 255         mdk_rdev_t *rdev;
 256         struct list_head *tmp;
 257         struct page *page = alloc_page(GFP_KERNEL);
 258         sector_t target;
 259
 260         if (!page)
 261                 return ERR_PTR(-ENOMEM);
 262
 263         ITERATE_RDEV(mddev, rdev, tmp) {
 264                 if (! test_bit(In_sync, &rdev->flags)
 265                     || test_bit(Faulty, &rdev->flags))
 266                         continue;
 267
 268                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 269
 270                 if (sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ)) {
 271                         page->index = index;
 272                         return page;
 273                 }
 274         }
 275         return ERR_PTR(-EIO);
 276
 277 }
 278
 279 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 280 {
 281         mdk_rdev_t *rdev;
 282         struct list_head *tmp;
 283
 284         ITERATE_RDEV(mddev, rdev, tmp)
 285                 if (test_bit(In_sync, &rdev->flags)
 286                     && !test_bit(Faulty, &rdev->flags))
 287                         md_super_write(mddev, rdev,
 288                                        (rdev->sb_offset<<1) + offset
 289                                        + page->index * (PAGE_SIZE/512),
 290                                        PAGE_SIZE,
 291                                        page);
 292
 293         if (wait)
 294                 md_super_wait(mddev);
 295         return 0;
 296 }
 297
 298 /*
 299  * write out a page to a file
 300  */
 301 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 302 {
 303         int ret = -ENOMEM;
 304
 305         if (bitmap->file == NULL)
 306                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 307
 308         flush_dcache_page(page); /* make sure visible to anyone reading the file */
 309
 310         if (wait)
 311                 lock_page(page);
 312         else {
 313                 if (TestSetPageLocked(page))
 314                         return -EAGAIN; /* already locked */
 315                 if (PageWriteback(page)) {
 316                         unlock_page(page);
 317                         return -EAGAIN;
 318                 }
 319         }
 320
 321         ret = page->mapping->a_ops->prepare_write(bitmap->file, page, 0, PAGE_SIZE);
 322         if (!ret)
 323                 ret = page->mapping->a_ops->commit_write(bitmap->file, page, 0,
 324                         PAGE_SIZE);
 325         if (ret) {
 326                 unlock_page(page);
 327                 return ret;
 328         }
 329
 330         set_page_dirty(page); /* force it to be written out */
 331
 332         if (!wait) {
 333                 /* add to list to be waited for by daemon */
 334                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 335                 item->page = page;
 336                 get_page(page);
 337                 spin_lock(&bitmap->write_lock);
 338                 list_add(&item->list, &bitmap->complete_pages);
 339                 spin_unlock(&bitmap->write_lock);
 340                 md_wakeup_thread(bitmap->writeback_daemon);
 341         }
 342         return write_one_page(page, wait);
 343 }
 344
 345 /* read a page from a file, pinning it into cache, and return bytes_read */
 346 static struct page *read_page(struct file *file, unsigned long index,
 347                                         unsigned long *bytes_read)
 348 {
 349         struct inode *inode = file->f_mapping->host;
 350         struct page *page = NULL;
 351         loff_t isize = i_size_read(inode);
 352         unsigned long end_index = isize >> PAGE_SHIFT;
 353
 354         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_SIZE,
 355                         (unsigned long long)index << PAGE_SHIFT);
 356
 357         page = read_cache_page(inode->i_mapping, index,
 358                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 359         if (IS_ERR(page))
 360                 goto out;
 361         wait_on_page_locked(page);
 362         if (!PageUptodate(page) || PageError(page)) {
 363                 put_page(page);
 364                 page = ERR_PTR(-EIO);
 365                 goto out;
 366         }
 367
 368         if (index > end_index) /* we have read beyond EOF */
 369                 *bytes_read = 0;
 370         else if (index == end_index) /* possible short read */
 371                 *bytes_read = isize & ~PAGE_MASK;
 372         else
 373                 *bytes_read = PAGE_SIZE; /* got a full page */
 374 out:
 375         if (IS_ERR(page))
 376                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 377                         (int)PAGE_SIZE,
 378                         (unsigned long long)index << PAGE_SHIFT,
 379                         PTR_ERR(page));
 380         return page;
 381 }
 382
 383 /*
 384  * bitmap file superblock operations
 385  */
 386
 387 /* update the event counter and sync the superblock to disk */
 388 int bitmap_update_sb(struct bitmap *bitmap)
 389 {
 390         bitmap_super_t *sb;
 391         unsigned long flags;
 392
 393         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 394                 return 0;
 395         spin_lock_irqsave(&bitmap->lock, flags);
 396         if (!bitmap->sb_page) { /* no superblock */
 397                 spin_unlock_irqrestore(&bitmap->lock, flags);
 398                 return 0;
 399         }
 400         spin_unlock_irqrestore(&bitmap->lock, flags);
 401         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 402         sb->events = cpu_to_le64(bitmap->mddev->events);
 403         if (!bitmap->mddev->degraded)
 404                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 405         kunmap_atomic(sb, KM_USER0);
 406         return write_page(bitmap, bitmap->sb_page, 1);
 407 }
 408
 409 /* print out the bitmap file superblock */
 410 void bitmap_print_sb(struct bitmap *bitmap)
 411 {
 412         bitmap_super_t *sb;
 413
 414         if (!bitmap || !bitmap->sb_page)
 415                 return;
 416         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 417         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 418         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 419         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 420         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 421                                         *(__u32 *)(sb->uuid+0),
 422                                         *(__u32 *)(sb->uuid+4),
 423                                         *(__u32 *)(sb->uuid+8),
 424                                         *(__u32 *)(sb->uuid+12));
 425         printk(KERN_DEBUG "        events: %llu\n",
 426                         (unsigned long long) le64_to_cpu(sb->events));
 427         printk(KERN_DEBUG "events cleared: %llu\n",
 428                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 429         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 430         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 431         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 432         printk(KERN_DEBUG "     sync size: %llu KB\n",
 433                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 434         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 435         kunmap_atomic(sb, KM_USER0);
 436 }
 437
 438 /* read the superblock from the bitmap file and initialize some bitmap fields */
 439 static int bitmap_read_sb(struct bitmap *bitmap)
 440 {
 441         char *reason = NULL;
 442         bitmap_super_t *sb;
 443         unsigned long chunksize, daemon_sleep, write_behind;
 444         unsigned long bytes_read;
 445         unsigned long long events;
 446         int err = -EINVAL;
 447
 448         /* page 0 is the superblock, read it... */
 449         if (bitmap->file)
 450                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 451         else {
 452                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 453                 bytes_read = PAGE_SIZE;
 454         }
 455         if (IS_ERR(bitmap->sb_page)) {
 456                 err = PTR_ERR(bitmap->sb_page);
 457                 bitmap->sb_page = NULL;
 458                 return err;
 459         }
 460
 461         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 462
 463         if (bytes_read < sizeof(*sb)) { /* short read */
 464                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 465                         bmname(bitmap));
 466                 err = -ENOSPC;
 467                 goto out;
 468         }
 469
 470         chunksize = le32_to_cpu(sb->chunksize);
 471         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 472         write_behind = le32_to_cpu(sb->write_behind);
 473
 474         /* verify that the bitmap-specific fields are valid */
 475         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 476                 reason = "bad magic";
 477         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 478                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 479                 reason = "unrecognized superblock version";
 480         else if (chunksize < PAGE_SIZE)
 481                 reason = "bitmap chunksize too small";
 482         else if ((1 << ffz(~chunksize)) != chunksize)
 483                 reason = "bitmap chunksize not a power of 2";
 484         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT / HZ)
 485                 reason = "daemon sleep period out of range";
 486         else if (write_behind > COUNTER_MAX)
 487                 reason = "write-behind limit out of range (0 - 16383)";
 488         if (reason) {
 489                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 490                         bmname(bitmap), reason);
 491                 goto out;
 492         }
 493
 494         /* keep the array size field of the bitmap superblock up to date */
 495         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 496
 497         if (!bitmap->mddev->persistent)
 498                 goto success;
 499
 500         /*
 501          * if we have a persistent array superblock, compare the
 502          * bitmap's UUID and event counter to the mddev's
 503          */
 504         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 505                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 506                         bmname(bitmap));
 507                 goto out;
 508         }
 509         events = le64_to_cpu(sb->events);
 510         if (events < bitmap->mddev->events) {
 511                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 512                         "-- forcing full recovery\n", bmname(bitmap), events,
 513                         (unsigned long long) bitmap->mddev->events);
 514                 sb->state |= BITMAP_STALE;
 515         }
 516 success:
 517         /* assign fields using values from superblock */
 518         bitmap->chunksize = chunksize;
 519         bitmap->daemon_sleep = daemon_sleep;
 520         bitmap->daemon_lastrun = jiffies;
 521         bitmap->max_write_behind = write_behind;
 522         bitmap->flags |= sb->state;
 523         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 524                 bitmap->flags |= BITMAP_HOSTENDIAN;
 525         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 526         if (sb->state & BITMAP_STALE)
 527                 bitmap->events_cleared = bitmap->mddev->events;
 528         err = 0;
 529 out:
 530         kunmap_atomic(sb, KM_USER0);
 531         if (err)
 532                 bitmap_print_sb(bitmap);
 533         return err;
 534 }
 535
 536 enum bitmap_mask_op {
 537         MASK_SET,
 538         MASK_UNSET
 539 };
 540
 541 /* record the state of the bitmap in the superblock */
 542 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 543                                 enum bitmap_mask_op op)
 544 {
 545         bitmap_super_t *sb;
 546         unsigned long flags;
 547
 548         spin_lock_irqsave(&bitmap->lock, flags);
 549         if (!bitmap->sb_page) { /* can't set the state */
 550                 spin_unlock_irqrestore(&bitmap->lock, flags);
 551                 return;
 552         }
 553         get_page(bitmap->sb_page);
 554         spin_unlock_irqrestore(&bitmap->lock, flags);
 555         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
 556         switch (op) {
 557                 case MASK_SET: sb->state |= bits;
 558                                 break;
 559                 case MASK_UNSET: sb->state &= ~bits;
 560                                 break;
 561                 default: BUG();
 562         }
 563         kunmap_atomic(sb, KM_USER0);
 564         put_page(bitmap->sb_page);
 565 }
 566
 567 /*
 568  * general bitmap file operations
 569  */
 570
 571 /* calculate the index of the page that contains this bit */
 572 static inline unsigned long file_page_index(unsigned long chunk)
 573 {
 574         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 575 }
 576
 577 /* calculate the (bit) offset of this bit within a page */
 578 static inline unsigned long file_page_offset(unsigned long chunk)
 579 {
 580         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 581 }
 582
 583 /*
 584  * return a pointer to the page in the filemap that contains the given bit
 585  *
 586  * this lookup is complicated by the fact that the bitmap sb might be exactly
 587  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 588  * 0 or page 1
 589  */
 590 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 591                                         unsigned long chunk)
 592 {
 593         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 594 }
 595
 596
 597 static void bitmap_file_unmap(struct bitmap *bitmap)
 598 {
 599         struct page **map, *sb_page;
 600         unsigned long *attr;
 601         int pages;
 602         unsigned long flags;
 603
 604         spin_lock_irqsave(&bitmap->lock, flags);
 605         map = bitmap->filemap;
 606         bitmap->filemap = NULL;
 607         attr = bitmap->filemap_attr;
 608         bitmap->filemap_attr = NULL;
 609         pages = bitmap->file_pages;
 610         bitmap->file_pages = 0;
 611         sb_page = bitmap->sb_page;
 612         bitmap->sb_page = NULL;
 613         spin_unlock_irqrestore(&bitmap->lock, flags);
 614
 615         while (pages--)
 616                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 617                         put_page(map[pages]);
 618         kfree(map);
 619         kfree(attr);
 620
 621         safe_put_page(sb_page);
 622 }
 623
 624 static void bitmap_stop_daemon(struct bitmap *bitmap);
 625
 626 /* dequeue the next item in a page list -- don't call from irq context */
 627 static struct page_list *dequeue_page(struct bitmap *bitmap)
 628 {
 629         struct page_list *item = NULL;
 630         struct list_head *head = &bitmap->complete_pages;
 631
 632         spin_lock(&bitmap->write_lock);
 633         if (list_empty(head))
 634                 goto out;
 635         item = list_entry(head->prev, struct page_list, list);
 636         list_del(head->prev);
 637 out:
 638         spin_unlock(&bitmap->write_lock);
 639         return item;
 640 }
 641
 642 static void drain_write_queues(struct bitmap *bitmap)
 643 {
 644         struct page_list *item;
 645
 646         while ((item = dequeue_page(bitmap))) {
 647                 /* don't bother to wait */
 648                 put_page(item->page);
 649                 mempool_free(item, bitmap->write_pool);
 650         }
 651
 652         wake_up(&bitmap->write_wait);
 653 }
 654
 655 static void bitmap_file_put(struct bitmap *bitmap)
 656 {
 657         struct file *file;
 658         struct inode *inode;
 659         unsigned long flags;
 660
 661         spin_lock_irqsave(&bitmap->lock, flags);
 662         file = bitmap->file;
 663         bitmap->file = NULL;
 664         spin_unlock_irqrestore(&bitmap->lock, flags);
 665
 666         bitmap_stop_daemon(bitmap);
 667
 668         drain_write_queues(bitmap);
 669
 670         bitmap_file_unmap(bitmap);
 671
 672         if (file) {
 673                 inode = file->f_mapping->host;
 674                 spin_lock(&inode->i_lock);
 675                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 676                 spin_unlock(&inode->i_lock);
 677                 fput(file);
 678         }
 679 }
 680
 681
 682 /*
 683  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 684  * then it is no longer reliable, so we stop using it and we mark the file
 685  * as failed in the superblock
 686  */
 687 static void bitmap_file_kick(struct bitmap *bitmap)
 688 {
 689         char *path, *ptr = NULL;
 690
 691         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 692         bitmap_update_sb(bitmap);
 693
 694         if (bitmap->file) {
 695                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 696                 if (path)
 697                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 698
 699                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 700                        bmname(bitmap), ptr ? ptr : "");
 701
 702                 kfree(path);
 703         }
 704
 705         bitmap_file_put(bitmap);
 706
 707         return;
 708 }
 709
 710 enum bitmap_page_attr {
 711         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 712         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 713         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 714 };
 715
 716 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 717                                 enum bitmap_page_attr attr)
 718 {
 719         bitmap->filemap_attr[page->index] |= attr;
 720 }
 721
 722 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 723                                 enum bitmap_page_attr attr)
 724 {
 725         bitmap->filemap_attr[page->index] &= ~attr;
 726 }
 727
 728 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 729 {
 730         return bitmap->filemap_attr[page->index];
 731 }
 732
 733 /*
 734  * bitmap_file_set_bit -- called before performing a write to the md device
 735  * to set (and eventually sync) a particular bit in the bitmap file
 736  *
 737  * we set the bit immediately, then we record the page number so that
 738  * when an unplug occurs, we can flush the dirty pages out to disk
 739  */
 740 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 741 {
 742         unsigned long bit;
 743         struct page *page;
 744         void *kaddr;
 745         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 746
 747         if (!bitmap->filemap) {
 748                 return;
 749         }
 750
 751         page = filemap_get_page(bitmap, chunk);
 752         bit = file_page_offset(chunk);
 753
 754
 755         /* make sure the page stays cached until it gets written out */
 756         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 757                 get_page(page);
 758
 759         /* set the bit */
 760         kaddr = kmap_atomic(page, KM_USER0);
 761         if (bitmap->flags & BITMAP_HOSTENDIAN)
 762                 set_bit(bit, kaddr);
 763         else
 764                 ext2_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                         md_super_wait(bitmap->mddev);
 823         }
 824         return 0;
 825 }
 826
 827 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 828 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 829  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 830  * memory mapping of the bitmap file
 831  * Special cases:
 832  *   if there's no bitmap file, or if the bitmap file had been
 833  *   previously kicked from the array, we mark all the bits as
 834  *   1's in order to cause a full resync.
 835  *
 836  * We ignore all bits for sectors that end earlier than 'start'.
 837  * This is used when reading an out-of-date bitmap...
 838  */
 839 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 840 {
 841         unsigned long i, chunks, index, oldindex, bit;
 842         struct page *page = NULL, *oldpage = NULL;
 843         unsigned long num_pages, bit_cnt = 0;
 844         struct file *file;
 845         unsigned long bytes, offset, dummy;
 846         int outofdate;
 847         int ret = -ENOSPC;
 848         void *paddr;
 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 = kzalloc(sizeof(long) * num_pages, GFP_KERNEL);
 883         if (!bitmap->filemap_attr)
 884                 goto out;
 885
 886         oldindex = ~0L;
 887
 888         for (i = 0; i < chunks; i++) {
 889                 int b;
 890                 index = file_page_index(i);
 891                 bit = file_page_offset(i);
 892                 if (index != oldindex) { /* this is a new page, read it in */
 893                         /* unmap the old page, we're done with it */
 894                         if (index == 0) {
 895                                 /*
 896                                  * if we're here then the superblock page
 897                                  * contains some bits (PAGE_SIZE != sizeof sb)
 898                                  * we've already read it in, so just use it
 899                                  */
 900                                 page = bitmap->sb_page;
 901                                 offset = sizeof(bitmap_super_t);
 902                         } else if (file) {
 903                                 page = read_page(file, index, &dummy);
 904                                 offset = 0;
 905                         } else {
 906                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 907                                 offset = 0;
 908                         }
 909                         if (IS_ERR(page)) { /* read error */
 910                                 ret = PTR_ERR(page);
 911                                 goto out;
 912                         }
 913
 914                         oldindex = index;
 915                         oldpage = page;
 916
 917                         if (outofdate) {
 918                                 /*
 919                                  * if bitmap is out of date, dirty the
 920                                  * whole page and write it out
 921                                  */
 922                                 paddr = kmap_atomic(page, KM_USER0);
 923                                 memset(paddr + offset, 0xff,
 924                                        PAGE_SIZE - offset);
 925                                 kunmap_atomic(paddr, KM_USER0);
 926                                 ret = write_page(bitmap, page, 1);
 927                                 if (ret) {
 928                                         /* release, page not in filemap yet */
 929                                         put_page(page);
 930                                         goto out;
 931                                 }
 932                         }
 933
 934                         bitmap->filemap[bitmap->file_pages++] = page;
 935                 }
 936                 paddr = kmap_atomic(page, KM_USER0);
 937                 if (bitmap->flags & BITMAP_HOSTENDIAN)
 938                         b = test_bit(bit, paddr);
 939                 else
 940                         b = ext2_test_bit(bit, paddr);
 941                 kunmap_atomic(paddr, KM_USER0);
 942                 if (b) {
 943                         /* if the disk bit is set, set the memory bit */
 944                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 945                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 946                                 );
 947                         bit_cnt++;
 948                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 949                 }
 950         }
 951
 952         /* everything went OK */
 953         ret = 0;
 954         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 955
 956         if (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         void *paddr;
1012
1013         if (bitmap == NULL)
1014                 return 0;
1015         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1016                 return 0;
1017         bitmap->daemon_lastrun = jiffies;
1018
1019         for (j = 0; j < bitmap->chunks; j++) {
1020                 bitmap_counter_t *bmc;
1021                 spin_lock_irqsave(&bitmap->lock, flags);
1022                 if (!bitmap->filemap) {
1023                         /* error or shutdown */
1024                         spin_unlock_irqrestore(&bitmap->lock, flags);
1025                         break;
1026                 }
1027
1028                 page = filemap_get_page(bitmap, j);
1029
1030                 if (page != lastpage) {
1031                         /* skip this page unless it's marked as needing cleaning */
1032                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1033                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1034                                         get_page(page);
1035                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1036                                 }
1037                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1038                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1039                                         switch (write_page(bitmap, page, 0)) {
1040                                         case -EAGAIN:
1041                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1042                                                 break;
1043                                         case 0:
1044                                                 break;
1045                                         default:
1046                                                 bitmap_file_kick(bitmap);
1047                                         }
1048                                         put_page(page);
1049                                 }
1050                                 continue;
1051                         }
1052
1053                         /* grab the new page, sync and release the old */
1054                         get_page(page);
1055                         if (lastpage != NULL) {
1056                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1057                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1058                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1059                                         err = write_page(bitmap, lastpage, 0);
1060                                         if (err == -EAGAIN) {
1061                                                 err = 0;
1062                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1063                                         }
1064                                 } else {
1065                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1066                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1067                                 }
1068                                 put_page(lastpage);
1069                                 if (err)
1070                                         bitmap_file_kick(bitmap);
1071                         } else
1072                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1073                         lastpage = page;
1074 /*
1075                         printk("bitmap clean at page %lu\n", j);
1076 */
1077                         spin_lock_irqsave(&bitmap->lock, flags);
1078                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1079                 }
1080                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1081                                         &blocks, 0);
1082                 if (bmc) {
1083 /*
1084   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1085 */
1086                         if (*bmc == 2) {
1087                                 *bmc=1; /* maybe clear the bit next time */
1088                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1089                         } else if (*bmc == 1) {
1090                                 /* we can clear the bit */
1091                                 *bmc = 0;
1092                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1093                                                   -1);
1094
1095                                 /* clear the bit */
1096                                 paddr = kmap_atomic(page, KM_USER0);
1097                                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1098                                         clear_bit(file_page_offset(j), paddr);
1099                                 else
1100                                         ext2_clear_bit(file_page_offset(j), paddr);
1101                                 kunmap_atomic(paddr, KM_USER0);
1102                         }
1103                 }
1104                 spin_unlock_irqrestore(&bitmap->lock, flags);
1105         }
1106
1107         /* now sync the final page */
1108         if (lastpage != NULL) {
1109                 spin_lock_irqsave(&bitmap->lock, flags);
1110                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1111                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1112                         spin_unlock_irqrestore(&bitmap->lock, flags);
1113                         err = write_page(bitmap, lastpage, 0);
1114                         if (err == -EAGAIN) {
1115                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1116                                 err = 0;
1117                         }
1118                 } else {
1119                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1120                         spin_unlock_irqrestore(&bitmap->lock, flags);
1121                 }
1122
1123                 put_page(lastpage);
1124         }
1125
1126         return err;
1127 }
1128
1129 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1130 {
1131         mdk_thread_t *dmn;
1132         unsigned long flags;
1133
1134         /* if no one is waiting on us, we'll free the md thread struct
1135          * and exit, otherwise we let the waiter clean things up */
1136         spin_lock_irqsave(&bitmap->lock, flags);
1137         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1138                 *daemon = NULL;
1139                 spin_unlock_irqrestore(&bitmap->lock, flags);
1140                 kfree(dmn);
1141                 complete_and_exit(NULL, 0); /* do_exit not exported */
1142         }
1143         spin_unlock_irqrestore(&bitmap->lock, flags);
1144 }
1145
1146 static void bitmap_writeback_daemon(mddev_t *mddev)
1147 {
1148         struct bitmap *bitmap = mddev->bitmap;
1149         struct page *page;
1150         struct page_list *item;
1151         int err = 0;
1152
1153         if (signal_pending(current)) {
1154                 printk(KERN_INFO
1155                        "%s: bitmap writeback daemon got signal, exiting...\n",
1156                        bmname(bitmap));
1157                 err = -EINTR;
1158                 goto out;
1159         }
1160         if (bitmap == NULL)
1161                 /* about to be stopped. */
1162                 return;
1163
1164         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1165         /* wait on bitmap page writebacks */
1166         while ((item = dequeue_page(bitmap))) {
1167                 page = item->page;
1168                 mempool_free(item, bitmap->write_pool);
1169                 PRINTK("wait on page writeback: %p\n", page);
1170                 wait_on_page_writeback(page);
1171                 PRINTK("finished page writeback: %p\n", page);
1172
1173                 err = PageError(page);
1174                 put_page(page);
1175                 if (err) {
1176                         printk(KERN_WARNING "%s: bitmap file writeback "
1177                                "failed (page %lu): %d\n",
1178                                bmname(bitmap), page->index, err);
1179                         bitmap_file_kick(bitmap);
1180                         goto out;
1181                 }
1182         }
1183  out:
1184         wake_up(&bitmap->write_wait);
1185         if (err) {
1186                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1187                        bmname(bitmap), err);
1188                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1189         }
1190 }
1191
1192 static mdk_thread_t *bitmap_start_daemon(struct bitmap *bitmap,
1193                                 void (*func)(mddev_t *), char *name)
1194 {
1195         mdk_thread_t *daemon;
1196         char namebuf[32];
1197
1198 #ifdef INJECT_FATAL_FAULT_2
1199         daemon = NULL;
1200 #else
1201         sprintf(namebuf, "%%s_%s", name);
1202         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1203 #endif
1204         if (!daemon) {
1205                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1206                         bmname(bitmap));
1207                 return ERR_PTR(-ECHILD);
1208         }
1209
1210         md_wakeup_thread(daemon); /* start it running */
1211
1212         PRINTK("%s: %s daemon (pid %d) started...\n",
1213                 bmname(bitmap), name, daemon->tsk->pid);
1214
1215         return daemon;
1216 }
1217
1218 static void bitmap_stop_daemon(struct bitmap *bitmap)
1219 {
1220         /* the daemon can't stop itself... it'll just exit instead... */
1221         if (bitmap->writeback_daemon && ! IS_ERR(bitmap->writeback_daemon) &&
1222             current->pid != bitmap->writeback_daemon->tsk->pid) {
1223                 mdk_thread_t *daemon;
1224                 unsigned long flags;
1225
1226                 spin_lock_irqsave(&bitmap->lock, flags);
1227                 daemon = bitmap->writeback_daemon;
1228                 bitmap->writeback_daemon = NULL;
1229                 spin_unlock_irqrestore(&bitmap->lock, flags);
1230                 if (daemon && ! IS_ERR(daemon))
1231                         md_unregister_thread(daemon); /* destroy the thread */
1232         }
1233 }
1234
1235 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1236                                             sector_t offset, int *blocks,
1237                                             int create)
1238 {
1239         /* If 'create', we might release the lock and reclaim it.
1240          * The lock must have been taken with interrupts enabled.
1241          * If !create, we don't release the lock.
1242          */
1243         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1244         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1245         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1246         sector_t csize;
1247
1248         if (bitmap_checkpage(bitmap, page, create) < 0) {
1249                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1250                 *blocks = csize - (offset & (csize- 1));
1251                 return NULL;
1252         }
1253         /* now locked ... */
1254
1255         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1256                 /* should we use the first or second counter field
1257                  * of the hijacked pointer? */
1258                 int hi = (pageoff > PAGE_COUNTER_MASK);
1259                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1260                                           PAGE_COUNTER_SHIFT - 1);
1261                 *blocks = csize - (offset & (csize- 1));
1262                 return  &((bitmap_counter_t *)
1263                           &bitmap->bp[page].map)[hi];
1264         } else { /* page is allocated */
1265                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1266                 *blocks = csize - (offset & (csize- 1));
1267                 return (bitmap_counter_t *)
1268                         &(bitmap->bp[page].map[pageoff]);
1269         }
1270 }
1271
1272 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1273 {
1274         if (!bitmap) return 0;
1275
1276         if (behind) {
1277                 atomic_inc(&bitmap->behind_writes);
1278                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1279                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1280         }
1281
1282         while (sectors) {
1283                 int blocks;
1284                 bitmap_counter_t *bmc;
1285
1286                 spin_lock_irq(&bitmap->lock);
1287                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1288                 if (!bmc) {
1289                         spin_unlock_irq(&bitmap->lock);
1290                         return 0;
1291                 }
1292
1293                 switch(*bmc) {
1294                 case 0:
1295                         bitmap_file_set_bit(bitmap, offset);
1296                         bitmap_count_page(bitmap,offset, 1);
1297                         blk_plug_device(bitmap->mddev->queue);
1298                         /* fall through */
1299                 case 1:
1300                         *bmc = 2;
1301                 }
1302                 BUG_ON((*bmc & COUNTER_MAX) == COUNTER_MAX);
1303                 (*bmc)++;
1304
1305                 spin_unlock_irq(&bitmap->lock);
1306
1307                 offset += blocks;
1308                 if (sectors > blocks)
1309                         sectors -= blocks;
1310                 else sectors = 0;
1311         }
1312         return 0;
1313 }
1314
1315 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1316                      int success, int behind)
1317 {
1318         if (!bitmap) return;
1319         if (behind) {
1320                 atomic_dec(&bitmap->behind_writes);
1321                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1322                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1323         }
1324
1325         while (sectors) {
1326                 int blocks;
1327                 unsigned long flags;
1328                 bitmap_counter_t *bmc;
1329
1330                 spin_lock_irqsave(&bitmap->lock, flags);
1331                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1332                 if (!bmc) {
1333                         spin_unlock_irqrestore(&bitmap->lock, flags);
1334                         return;
1335                 }
1336
1337                 if (!success && ! (*bmc & NEEDED_MASK))
1338                         *bmc |= NEEDED_MASK;
1339
1340                 (*bmc)--;
1341                 if (*bmc <= 2) {
1342                         set_page_attr(bitmap,
1343                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1344                                       BITMAP_PAGE_CLEAN);
1345                 }
1346                 spin_unlock_irqrestore(&bitmap->lock, flags);
1347                 offset += blocks;
1348                 if (sectors > blocks)
1349                         sectors -= blocks;
1350                 else sectors = 0;
1351         }
1352 }
1353
1354 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1355                         int degraded)
1356 {
1357         bitmap_counter_t *bmc;
1358         int rv;
1359         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1360                 *blocks = 1024;
1361                 return 1; /* always resync if no bitmap */
1362         }
1363         spin_lock_irq(&bitmap->lock);
1364         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1365         rv = 0;
1366         if (bmc) {
1367                 /* locked */
1368                 if (RESYNC(*bmc))
1369                         rv = 1;
1370                 else if (NEEDED(*bmc)) {
1371                         rv = 1;
1372                         if (!degraded) { /* don't set/clear bits if degraded */
1373                                 *bmc |= RESYNC_MASK;
1374                                 *bmc &= ~NEEDED_MASK;
1375                         }
1376                 }
1377         }
1378         spin_unlock_irq(&bitmap->lock);
1379         return rv;
1380 }
1381
1382 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1383 {
1384         bitmap_counter_t *bmc;
1385         unsigned long flags;
1386 /*
1387         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1388 */      if (bitmap == NULL) {
1389                 *blocks = 1024;
1390                 return;
1391         }
1392         spin_lock_irqsave(&bitmap->lock, flags);
1393         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1394         if (bmc == NULL)
1395                 goto unlock;
1396         /* locked */
1397 /*
1398         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1399 */
1400         if (RESYNC(*bmc)) {
1401                 *bmc &= ~RESYNC_MASK;
1402
1403                 if (!NEEDED(*bmc) && aborted)
1404                         *bmc |= NEEDED_MASK;
1405                 else {
1406                         if (*bmc <= 2) {
1407                                 set_page_attr(bitmap,
1408                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1409                                               BITMAP_PAGE_CLEAN);
1410                         }
1411                 }
1412         }
1413  unlock:
1414         spin_unlock_irqrestore(&bitmap->lock, flags);
1415 }
1416
1417 void bitmap_close_sync(struct bitmap *bitmap)
1418 {
1419         /* Sync has finished, and any bitmap chunks that weren't synced
1420          * properly have been aborted.  It remains to us to clear the
1421          * RESYNC bit wherever it is still on
1422          */
1423         sector_t sector = 0;
1424         int blocks;
1425         if (!bitmap) return;
1426         while (sector < bitmap->mddev->resync_max_sectors) {
1427                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1428 /*
1429                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1430                                          (unsigned long long)sector, blocks);
1431 */              sector += blocks;
1432         }
1433 }
1434
1435 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1436 {
1437         /* For each chunk covered by any of these sectors, set the
1438          * counter to 1 and set resync_needed.  They should all
1439          * be 0 at this point
1440          */
1441
1442         int secs;
1443         bitmap_counter_t *bmc;
1444         spin_lock_irq(&bitmap->lock);
1445         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1446         if (!bmc) {
1447                 spin_unlock_irq(&bitmap->lock);
1448                 return;
1449         }
1450         if (! *bmc) {
1451                 struct page *page;
1452                 *bmc = 1 | (needed?NEEDED_MASK:0);
1453                 bitmap_count_page(bitmap, offset, 1);
1454                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1455                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1456         }
1457         spin_unlock_irq(&bitmap->lock);
1458
1459 }
1460
1461 /*
1462  * flush out any pending updates
1463  */
1464 void bitmap_flush(mddev_t *mddev)
1465 {
1466         struct bitmap *bitmap = mddev->bitmap;
1467         int sleep;
1468
1469         if (!bitmap) /* there was no bitmap */
1470                 return;
1471
1472         /* run the daemon_work three time to ensure everything is flushed
1473          * that can be
1474          */
1475         sleep = bitmap->daemon_sleep;
1476         bitmap->daemon_sleep = 0;
1477         bitmap_daemon_work(bitmap);
1478         bitmap_daemon_work(bitmap);
1479         bitmap_daemon_work(bitmap);
1480         bitmap->daemon_sleep = sleep;
1481         bitmap_update_sb(bitmap);
1482 }
1483
1484 /*
1485  * free memory that was allocated
1486  */
1487 static void bitmap_free(struct bitmap *bitmap)
1488 {
1489         unsigned long k, pages;
1490         struct bitmap_page *bp;
1491
1492         if (!bitmap) /* there was no bitmap */
1493                 return;
1494
1495         /* release the bitmap file and kill the daemon */
1496         bitmap_file_put(bitmap);
1497
1498         bp = bitmap->bp;
1499         pages = bitmap->pages;
1500
1501         /* free all allocated memory */
1502
1503         mempool_destroy(bitmap->write_pool);
1504
1505         if (bp) /* deallocate the page memory */
1506                 for (k = 0; k < pages; k++)
1507                         if (bp[k].map && !bp[k].hijacked)
1508                                 kfree(bp[k].map);
1509         kfree(bp);
1510         kfree(bitmap);
1511 }
1512 void bitmap_destroy(mddev_t *mddev)
1513 {
1514         struct bitmap *bitmap = mddev->bitmap;
1515
1516         if (!bitmap) /* there was no bitmap */
1517                 return;
1518
1519         mddev->bitmap = NULL; /* disconnect from the md device */
1520         if (mddev->thread)
1521                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1522
1523         bitmap_free(bitmap);
1524 }
1525
1526 /*
1527  * initialize the bitmap structure
1528  * if this returns an error, bitmap_destroy must be called to do clean up
1529  */
1530 int bitmap_create(mddev_t *mddev)
1531 {
1532         struct bitmap *bitmap;
1533         unsigned long blocks = mddev->resync_max_sectors;
1534         unsigned long chunks;
1535         unsigned long pages;
1536         struct file *file = mddev->bitmap_file;
1537         int err;
1538         sector_t start;
1539
1540         BUG_ON(sizeof(bitmap_super_t) != 256);
1541
1542         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1543                 return 0;
1544
1545         BUG_ON(file && mddev->bitmap_offset);
1546
1547         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1548         if (!bitmap)
1549                 return -ENOMEM;
1550
1551         spin_lock_init(&bitmap->lock);
1552         bitmap->mddev = mddev;
1553
1554         spin_lock_init(&bitmap->write_lock);
1555         INIT_LIST_HEAD(&bitmap->complete_pages);
1556         init_waitqueue_head(&bitmap->write_wait);
1557         bitmap->write_pool = mempool_create_kmalloc_pool(WRITE_POOL_SIZE,
1558                                                 sizeof(struct page_list));
1559         err = -ENOMEM;
1560         if (!bitmap->write_pool)
1561                 goto error;
1562
1563         bitmap->file = file;
1564         bitmap->offset = mddev->bitmap_offset;
1565         if (file) get_file(file);
1566         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1567         err = bitmap_read_sb(bitmap);
1568         if (err)
1569                 goto error;
1570
1571         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1572                                         sizeof(bitmap->chunksize));
1573
1574         /* now that chunksize and chunkshift are set, we can use these macros */
1575         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1576                         CHUNK_BLOCK_RATIO(bitmap);
1577         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1578
1579         BUG_ON(!pages);
1580
1581         bitmap->chunks = chunks;
1582         bitmap->pages = pages;
1583         bitmap->missing_pages = pages;
1584         bitmap->counter_bits = COUNTER_BITS;
1585
1586         bitmap->syncchunk = ~0UL;
1587
1588 #ifdef INJECT_FATAL_FAULT_1
1589         bitmap->bp = NULL;
1590 #else
1591         bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1592 #endif
1593         err = -ENOMEM;
1594         if (!bitmap->bp)
1595                 goto error;
1596
1597         bitmap->flags |= BITMAP_ACTIVE;
1598
1599         /* now that we have some pages available, initialize the in-memory
1600          * bitmap from the on-disk bitmap */
1601         start = 0;
1602         if (mddev->degraded == 0
1603             || bitmap->events_cleared == mddev->events)
1604                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1605                 start = mddev->recovery_cp;
1606         err = bitmap_init_from_disk(bitmap, start);
1607
1608         if (err)
1609                 goto error;
1610
1611         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1612                 pages, bmname(bitmap));
1613
1614         mddev->bitmap = bitmap;
1615
1616         if (file)
1617                 /* kick off the bitmap writeback daemon */
1618                 bitmap->writeback_daemon =
1619                         bitmap_start_daemon(bitmap,
1620                                             bitmap_writeback_daemon,
1621                                             "bitmap_wb");
1622
1623         if (IS_ERR(bitmap->writeback_daemon))
1624                 return PTR_ERR(bitmap->writeback_daemon);
1625         mddev->thread->timeout = bitmap->daemon_sleep * HZ;
1626
1627         return bitmap_update_sb(bitmap);
1628
1629  error:
1630         bitmap_free(bitmap);
1631         return err;
1632 }
1633
1634 /* the bitmap API -- for raid personalities */
1635 EXPORT_SYMBOL(bitmap_startwrite);
1636 EXPORT_SYMBOL(bitmap_endwrite);
1637 EXPORT_SYMBOL(bitmap_start_sync);
1638 EXPORT_SYMBOL(bitmap_end_sync);
1639 EXPORT_SYMBOL(bitmap_unplug);
1640 EXPORT_SYMBOL(bitmap_close_sync);
1641 EXPORT_SYMBOL(bitmap_daemon_work);