drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/version.h>
  25 #include <linux/errno.h>
  26 #include <linux/slab.h>
  27 #include <linux/init.h>
  28 #include <linux/config.h>
  29 #include <linux/timer.h>
  30 #include <linux/sched.h>
  31 #include <linux/list.h>
  32 #include <linux/file.h>
  33 #include <linux/mount.h>
  34 #include <linux/buffer_head.h>
  35 #include <linux/raid/md.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 /* debug macros */
  39
  40 #define DEBUG 0
  41
  42 #if DEBUG
  43 /* these are for debugging purposes only! */
  44
  45 /* define one and only one of these */
  46 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  47 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  48 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  49 #define INJECT_FAULTS_4 0 /* undef */
  50 #define INJECT_FAULTS_5 0 /* undef */
  51 #define INJECT_FAULTS_6 0
  52
  53 /* if these are defined, the driver will fail! debug only */
  54 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  55 #define INJECT_FATAL_FAULT_2 0 /* undef */
  56 #define INJECT_FATAL_FAULT_3 0 /* undef */
  57 #endif
  58
  59 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  60 #define DPRINTK(x...) do { } while(0)
  61
  62 #ifndef PRINTK
  63 #  if DEBUG > 0
  64 #    define PRINTK(x...) printk(KERN_DEBUG x)
  65 #  else
  66 #    define PRINTK(x...)
  67 #  endif
  68 #endif
  69
  70 static inline char * bmname(struct bitmap *bitmap)
  71 {
  72         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  73 }
  74
  75
  76 /*
  77  * test if the bitmap is active
  78  */
  79 int bitmap_active(struct bitmap *bitmap)
  80 {
  81         unsigned long flags;
  82         int res = 0;
  83
  84         if (!bitmap)
  85                 return res;
  86         spin_lock_irqsave(&bitmap->lock, flags);
  87         res = bitmap->flags & BITMAP_ACTIVE;
  88         spin_unlock_irqrestore(&bitmap->lock, flags);
  89         return res;
  90 }
  91
  92 #define WRITE_POOL_SIZE 256
  93 /* mempool for queueing pending writes on the bitmap file */
  94 static void *write_pool_alloc(unsigned int gfp_flags, void *data)
  95 {
  96         return kmalloc(sizeof(struct page_list), gfp_flags);
  97 }
  98
  99 static void write_pool_free(void *ptr, void *data)
 100 {
 101         kfree(ptr);
 102 }
 103
 104 /*
 105  * just a placeholder - calls kmalloc for bitmap pages
 106  */
 107 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 108 {
 109         unsigned char *page;
 110
 111 #ifdef INJECT_FAULTS_1
 112         page = NULL;
 113 #else
 114         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 115 #endif
 116         if (!page)
 117                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 118         else
 119                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 120                         bmname(bitmap), page);
 121         return page;
 122 }
 123
 124 /*
 125  * for now just a placeholder -- just calls kfree for bitmap pages
 126  */
 127 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 128 {
 129         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 130         kfree(page);
 131 }
 132
 133 /*
 134  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 135  *
 136  * 1) check to see if this page is allocated, if it's not then try to alloc
 137  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 138  *    page pointer directly as a counter
 139  *
 140  * if we find our page, we increment the page's refcount so that it stays
 141  * allocated while we're using it
 142  */
 143 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 144 {
 145         unsigned char *mappage;
 146
 147         if (page >= bitmap->pages) {
 148                 printk(KERN_ALERT
 149                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 150                         bmname(bitmap), page, bitmap->pages-1);
 151                 return -EINVAL;
 152         }
 153
 154
 155         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 156                 return 0;
 157
 158         if (bitmap->bp[page].map) /* page is already allocated, just return */
 159                 return 0;
 160
 161         if (!create)
 162                 return -ENOENT;
 163
 164         spin_unlock_irq(&bitmap->lock);
 165
 166         /* this page has not been allocated yet */
 167
 168         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 169                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 170                         bmname(bitmap));
 171                 /* failed - set the hijacked flag so that we can use the
 172                  * pointer as a counter */
 173                 spin_lock_irq(&bitmap->lock);
 174                 if (!bitmap->bp[page].map)
 175                         bitmap->bp[page].hijacked = 1;
 176                 goto out;
 177         }
 178
 179         /* got a page */
 180
 181         spin_lock_irq(&bitmap->lock);
 182
 183         /* recheck the page */
 184
 185         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 186                 /* somebody beat us to getting the page */
 187                 bitmap_free_page(bitmap, mappage);
 188                 return 0;
 189         }
 190
 191         /* no page was in place and we have one, so install it */
 192
 193         memset(mappage, 0, PAGE_SIZE);
 194         bitmap->bp[page].map = mappage;
 195         bitmap->missing_pages--;
 196 out:
 197         return 0;
 198 }
 199
 200
 201 /* if page is completely empty, put it back on the free list, or dealloc it */
 202 /* if page was hijacked, unmark the flag so it might get alloced next time */
 203 /* Note: lock should be held when calling this */
 204 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 205 {
 206         char *ptr;
 207
 208         if (bitmap->bp[page].count) /* page is still busy */
 209                 return;
 210
 211         /* page is no longer in use, it can be released */
 212
 213         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 214                 bitmap->bp[page].hijacked = 0;
 215                 bitmap->bp[page].map = NULL;
 216                 return;
 217         }
 218
 219         /* normal case, free the page */
 220
 221 #if 0
 222 /* actually ... let's not.  We will probably need the page again exactly when
 223  * memory is tight and we are flusing to disk
 224  */
 225         return;
 226 #else
 227         ptr = bitmap->bp[page].map;
 228         bitmap->bp[page].map = NULL;
 229         bitmap->missing_pages++;
 230         bitmap_free_page(bitmap, ptr);
 231         return;
 232 #endif
 233 }
 234
 235
 236 /*
 237  * bitmap file handling - read and write the bitmap file and its superblock
 238  */
 239
 240 /* copy the pathname of a file to a buffer */
 241 char *file_path(struct file *file, char *buf, int count)
 242 {
 243         struct dentry *d;
 244         struct vfsmount *v;
 245
 246         if (!buf)
 247                 return NULL;
 248
 249         d = file->f_dentry;
 250         v = file->f_vfsmnt;
 251
 252         buf = d_path(d, v, buf, count);
 253
 254         return IS_ERR(buf) ? NULL : buf;
 255 }
 256
 257 /*
 258  * basic page I/O operations
 259  */
 260
 261 /* IO operations when bitmap is stored near all superblocks */
 262 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 263 {
 264         /* choose a good rdev and read the page from there */
 265
 266         mdk_rdev_t *rdev;
 267         struct list_head *tmp;
 268         struct page *page = alloc_page(GFP_KERNEL);
 269         sector_t target;
 270
 271         if (!page)
 272                 return ERR_PTR(-ENOMEM);
 273
 274         ITERATE_RDEV(mddev, rdev, tmp) {
 275                 if (! rdev->in_sync || rdev->faulty)
 276                         continue;
 277
 278                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 279
 280                 if (sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ)) {
 281                         page->index = index;
 282                         return page;
 283                 }
 284         }
 285         return ERR_PTR(-EIO);
 286
 287 }
 288
 289 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 290 {
 291         mdk_rdev_t *rdev;
 292         struct list_head *tmp;
 293
 294         ITERATE_RDEV(mddev, rdev, tmp)
 295                 if (rdev->in_sync && !rdev->faulty)
 296                         md_super_write(mddev, rdev,
 297                                        (rdev->sb_offset<<1) + offset
 298                                        + page->index * (PAGE_SIZE/512),
 299                                        PAGE_SIZE,
 300                                        page);
 301
 302         if (wait)
 303                 wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 304         return 0;
 305 }
 306
 307 /*
 308  * write out a page to a file
 309  */
 310 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 311 {
 312         int ret = -ENOMEM;
 313
 314         if (bitmap->file == NULL)
 315                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 316
 317         if (wait)
 318                 lock_page(page);
 319         else {
 320                 if (TestSetPageLocked(page))
 321                         return -EAGAIN; /* already locked */
 322                 if (PageWriteback(page)) {
 323                         unlock_page(page);
 324                         return -EAGAIN;
 325                 }
 326         }
 327
 328         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 329         if (!ret)
 330                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 331                         PAGE_SIZE);
 332         if (ret) {
 333                 unlock_page(page);
 334                 return ret;
 335         }
 336
 337         set_page_dirty(page); /* force it to be written out */
 338
 339         if (!wait) {
 340                 /* add to list to be waited for by daemon */
 341                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 342                 item->page = page;
 343                 page_cache_get(page);
 344                 spin_lock(&bitmap->write_lock);
 345                 list_add(&item->list, &bitmap->complete_pages);
 346                 spin_unlock(&bitmap->write_lock);
 347                 md_wakeup_thread(bitmap->writeback_daemon);
 348         }
 349         return write_one_page(page, wait);
 350 }
 351
 352 /* read a page from a file, pinning it into cache, and return bytes_read */
 353 static struct page *read_page(struct file *file, unsigned long index,
 354                                         unsigned long *bytes_read)
 355 {
 356         struct inode *inode = file->f_mapping->host;
 357         struct page *page = NULL;
 358         loff_t isize = i_size_read(inode);
 359         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 360
 361         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 362                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 363
 364         page = read_cache_page(inode->i_mapping, index,
 365                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 366         if (IS_ERR(page))
 367                 goto out;
 368         wait_on_page_locked(page);
 369         if (!PageUptodate(page) || PageError(page)) {
 370                 page_cache_release(page);
 371                 page = ERR_PTR(-EIO);
 372                 goto out;
 373         }
 374
 375         if (index > end_index) /* we have read beyond EOF */
 376                 *bytes_read = 0;
 377         else if (index == end_index) /* possible short read */
 378                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 379         else
 380                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 381 out:
 382         if (IS_ERR(page))
 383                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 384                         (int)PAGE_CACHE_SIZE,
 385                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 386                         PTR_ERR(page));
 387         return page;
 388 }
 389
 390 /*
 391  * bitmap file superblock operations
 392  */
 393
 394 /* update the event counter and sync the superblock to disk */
 395 int bitmap_update_sb(struct bitmap *bitmap)
 396 {
 397         bitmap_super_t *sb;
 398         unsigned long flags;
 399
 400         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 401                 return 0;
 402         spin_lock_irqsave(&bitmap->lock, flags);
 403         if (!bitmap->sb_page) { /* no superblock */
 404                 spin_unlock_irqrestore(&bitmap->lock, flags);
 405                 return 0;
 406         }
 407         spin_unlock_irqrestore(&bitmap->lock, flags);
 408         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 409         sb->events = cpu_to_le64(bitmap->mddev->events);
 410         if (!bitmap->mddev->degraded)
 411                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 412         kunmap(bitmap->sb_page);
 413         return write_page(bitmap, bitmap->sb_page, 1);
 414 }
 415
 416 /* print out the bitmap file superblock */
 417 void bitmap_print_sb(struct bitmap *bitmap)
 418 {
 419         bitmap_super_t *sb;
 420
 421         if (!bitmap || !bitmap->sb_page)
 422                 return;
 423         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 424         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 425         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 426         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 427         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 428                                         *(__u32 *)(sb->uuid+0),
 429                                         *(__u32 *)(sb->uuid+4),
 430                                         *(__u32 *)(sb->uuid+8),
 431                                         *(__u32 *)(sb->uuid+12));
 432         printk(KERN_DEBUG "        events: %llu\n",
 433                         (unsigned long long) le64_to_cpu(sb->events));
 434         printk(KERN_DEBUG "events cleared: %llu\n",
 435                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 436         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 437         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 438         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 439         printk(KERN_DEBUG "     sync size: %llu KB\n",
 440                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 441         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 442         kunmap(bitmap->sb_page);
 443 }
 444
 445 /* read the superblock from the bitmap file and initialize some bitmap fields */
 446 static int bitmap_read_sb(struct bitmap *bitmap)
 447 {
 448         char *reason = NULL;
 449         bitmap_super_t *sb;
 450         unsigned long chunksize, daemon_sleep, write_behind;
 451         unsigned long bytes_read;
 452         unsigned long long events;
 453         int err = -EINVAL;
 454
 455         /* page 0 is the superblock, read it... */
 456         if (bitmap->file)
 457                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 458         else {
 459                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 460                 bytes_read = PAGE_SIZE;
 461         }
 462         if (IS_ERR(bitmap->sb_page)) {
 463                 err = PTR_ERR(bitmap->sb_page);
 464                 bitmap->sb_page = NULL;
 465                 return err;
 466         }
 467
 468         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 469
 470         if (bytes_read < sizeof(*sb)) { /* short read */
 471                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 472                         bmname(bitmap));
 473                 err = -ENOSPC;
 474                 goto out;
 475         }
 476
 477         chunksize = le32_to_cpu(sb->chunksize);
 478         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 479         write_behind = le32_to_cpu(sb->write_behind);
 480
 481         /* verify that the bitmap-specific fields are valid */
 482         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 483                 reason = "bad magic";
 484         else if (sb->version != cpu_to_le32(BITMAP_MAJOR))
 485                 reason = "unrecognized superblock version";
 486         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 487                 reason = "bitmap chunksize out of range (512B - 4MB)";
 488         else if ((1 << ffz(~chunksize)) != chunksize)
 489                 reason = "bitmap chunksize not a power of 2";
 490         else if (daemon_sleep < 1 || daemon_sleep > 15)
 491                 reason = "daemon sleep period out of range (1-15s)";
 492         else if (write_behind > COUNTER_MAX)
 493                 reason = "write-behind limit out of range (0 - 16383)";
 494         if (reason) {
 495                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 496                         bmname(bitmap), reason);
 497                 goto out;
 498         }
 499
 500         /* keep the array size field of the bitmap superblock up to date */
 501         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 502
 503         if (!bitmap->mddev->persistent)
 504                 goto success;
 505
 506         /*
 507          * if we have a persistent array superblock, compare the
 508          * bitmap's UUID and event counter to the mddev's
 509          */
 510         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 511                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 512                         bmname(bitmap));
 513                 goto out;
 514         }
 515         events = le64_to_cpu(sb->events);
 516         if (events < bitmap->mddev->events) {
 517                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 518                         "-- forcing full recovery\n", bmname(bitmap), events,
 519                         (unsigned long long) bitmap->mddev->events);
 520                 sb->state |= BITMAP_STALE;
 521         }
 522 success:
 523         /* assign fields using values from superblock */
 524         bitmap->chunksize = chunksize;
 525         bitmap->daemon_sleep = daemon_sleep;
 526         bitmap->daemon_lastrun = jiffies;
 527         bitmap->max_write_behind = write_behind;
 528         bitmap->flags |= sb->state;
 529         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 530         if (sb->state & BITMAP_STALE)
 531                 bitmap->events_cleared = bitmap->mddev->events;
 532         err = 0;
 533 out:
 534         kunmap(bitmap->sb_page);
 535         if (err)
 536                 bitmap_print_sb(bitmap);
 537         return err;
 538 }
 539
 540 enum bitmap_mask_op {
 541         MASK_SET,
 542         MASK_UNSET
 543 };
 544
 545 /* record the state of the bitmap in the superblock */
 546 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 547                                 enum bitmap_mask_op op)
 548 {
 549         bitmap_super_t *sb;
 550         unsigned long flags;
 551
 552         spin_lock_irqsave(&bitmap->lock, flags);
 553         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 554                 spin_unlock_irqrestore(&bitmap->lock, flags);
 555                 return;
 556         }
 557         page_cache_get(bitmap->sb_page);
 558         spin_unlock_irqrestore(&bitmap->lock, flags);
 559         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 560         switch (op) {
 561                 case MASK_SET: sb->state |= bits;
 562                                 break;
 563                 case MASK_UNSET: sb->state &= ~bits;
 564                                 break;
 565                 default: BUG();
 566         }
 567         kunmap(bitmap->sb_page);
 568         page_cache_release(bitmap->sb_page);
 569 }
 570
 571 /*
 572  * general bitmap file operations
 573  */
 574
 575 /* calculate the index of the page that contains this bit */
 576 static inline unsigned long file_page_index(unsigned long chunk)
 577 {
 578         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 579 }
 580
 581 /* calculate the (bit) offset of this bit within a page */
 582 static inline unsigned long file_page_offset(unsigned long chunk)
 583 {
 584         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 585 }
 586
 587 /*
 588  * return a pointer to the page in the filemap that contains the given bit
 589  *
 590  * this lookup is complicated by the fact that the bitmap sb might be exactly
 591  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 592  * 0 or page 1
 593  */
 594 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 595                                         unsigned long chunk)
 596 {
 597         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 598 }
 599
 600
 601 static void bitmap_file_unmap(struct bitmap *bitmap)
 602 {
 603         struct page **map, *sb_page;
 604         unsigned long *attr;
 605         int pages;
 606         unsigned long flags;
 607
 608         spin_lock_irqsave(&bitmap->lock, flags);
 609         map = bitmap->filemap;
 610         bitmap->filemap = NULL;
 611         attr = bitmap->filemap_attr;
 612         bitmap->filemap_attr = NULL;
 613         pages = bitmap->file_pages;
 614         bitmap->file_pages = 0;
 615         sb_page = bitmap->sb_page;
 616         bitmap->sb_page = NULL;
 617         spin_unlock_irqrestore(&bitmap->lock, flags);
 618
 619         while (pages--)
 620                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 621                         page_cache_release(map[pages]);
 622         kfree(map);
 623         kfree(attr);
 624
 625         if (sb_page)
 626                 page_cache_release(sb_page);
 627 }
 628
 629 static void bitmap_stop_daemon(struct bitmap *bitmap);
 630
 631 /* dequeue the next item in a page list -- don't call from irq context */
 632 static struct page_list *dequeue_page(struct bitmap *bitmap)
 633 {
 634         struct page_list *item = NULL;
 635         struct list_head *head = &bitmap->complete_pages;
 636
 637         spin_lock(&bitmap->write_lock);
 638         if (list_empty(head))
 639                 goto out;
 640         item = list_entry(head->prev, struct page_list, list);
 641         list_del(head->prev);
 642 out:
 643         spin_unlock(&bitmap->write_lock);
 644         return item;
 645 }
 646
 647 static void drain_write_queues(struct bitmap *bitmap)
 648 {
 649         struct page_list *item;
 650
 651         while ((item = dequeue_page(bitmap))) {
 652                 /* don't bother to wait */
 653                 page_cache_release(item->page);
 654                 mempool_free(item, bitmap->write_pool);
 655         }
 656
 657         wake_up(&bitmap->write_wait);
 658 }
 659
 660 static void bitmap_file_put(struct bitmap *bitmap)
 661 {
 662         struct file *file;
 663         struct inode *inode;
 664         unsigned long flags;
 665
 666         spin_lock_irqsave(&bitmap->lock, flags);
 667         file = bitmap->file;
 668         bitmap->file = NULL;
 669         spin_unlock_irqrestore(&bitmap->lock, flags);
 670
 671         bitmap_stop_daemon(bitmap);
 672
 673         drain_write_queues(bitmap);
 674
 675         bitmap_file_unmap(bitmap);
 676
 677         if (file) {
 678                 inode = file->f_mapping->host;
 679                 spin_lock(&inode->i_lock);
 680                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 681                 spin_unlock(&inode->i_lock);
 682                 fput(file);
 683         }
 684 }
 685
 686
 687 /*
 688  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 689  * then it is no longer reliable, so we stop using it and we mark the file
 690  * as failed in the superblock
 691  */
 692 static void bitmap_file_kick(struct bitmap *bitmap)
 693 {
 694         char *path, *ptr = NULL;
 695
 696         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 697         bitmap_update_sb(bitmap);
 698
 699         if (bitmap->file) {
 700                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 701                 if (path)
 702                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 703
 704                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 705                        bmname(bitmap), ptr ? ptr : "");
 706
 707                 kfree(path);
 708         }
 709
 710         bitmap_file_put(bitmap);
 711
 712         return;
 713 }
 714
 715 enum bitmap_page_attr {
 716         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 717         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 718         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 719 };
 720
 721 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 722                                 enum bitmap_page_attr attr)
 723 {
 724         bitmap->filemap_attr[page->index] |= attr;
 725 }
 726
 727 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 728                                 enum bitmap_page_attr attr)
 729 {
 730         bitmap->filemap_attr[page->index] &= ~attr;
 731 }
 732
 733 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 734 {
 735         return bitmap->filemap_attr[page->index];
 736 }
 737
 738 /*
 739  * bitmap_file_set_bit -- called before performing a write to the md device
 740  * to set (and eventually sync) a particular bit in the bitmap file
 741  *
 742  * we set the bit immediately, then we record the page number so that
 743  * when an unplug occurs, we can flush the dirty pages out to disk
 744  */
 745 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 746 {
 747         unsigned long bit;
 748         struct page *page;
 749         void *kaddr;
 750         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 751
 752         if (!bitmap->filemap) {
 753                 return;
 754         }
 755
 756         page = filemap_get_page(bitmap, chunk);
 757         bit = file_page_offset(chunk);
 758
 759
 760         /* make sure the page stays cached until it gets written out */
 761         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 762                 page_cache_get(page);
 763
 764         /* set the bit */
 765         kaddr = kmap_atomic(page, KM_USER0);
 766         set_bit(bit, kaddr);
 767         kunmap_atomic(kaddr, KM_USER0);
 768         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 769
 770         /* record page number so it gets flushed to disk when unplug occurs */
 771         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 772
 773 }
 774
 775 /* this gets called when the md device is ready to unplug its underlying
 776  * (slave) device queues -- before we let any writes go down, we need to
 777  * sync the dirty pages of the bitmap file to disk */
 778 int bitmap_unplug(struct bitmap *bitmap)
 779 {
 780         unsigned long i, attr, flags;
 781         struct page *page;
 782         int wait = 0;
 783         int err;
 784
 785         if (!bitmap)
 786                 return 0;
 787
 788         /* look at each page to see if there are any set bits that need to be
 789          * flushed out to disk */
 790         for (i = 0; i < bitmap->file_pages; i++) {
 791                 spin_lock_irqsave(&bitmap->lock, flags);
 792                 if (!bitmap->filemap) {
 793                         spin_unlock_irqrestore(&bitmap->lock, flags);
 794                         return 0;
 795                 }
 796                 page = bitmap->filemap[i];
 797                 attr = get_page_attr(bitmap, page);
 798                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 799                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 800                 if ((attr & BITMAP_PAGE_DIRTY))
 801                         wait = 1;
 802                 spin_unlock_irqrestore(&bitmap->lock, flags);
 803
 804                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 805                         err = write_page(bitmap, page, 0);
 806                         if (err == -EAGAIN) {
 807                                 if (attr & BITMAP_PAGE_DIRTY)
 808                                         err = write_page(bitmap, page, 1);
 809                                 else
 810                                         err = 0;
 811                         }
 812                         if (err)
 813                                 return 1;
 814                 }
 815         }
 816         if (wait) { /* if any writes were performed, we need to wait on them */
 817                 if (bitmap->file) {
 818                         spin_lock_irq(&bitmap->write_lock);
 819                         wait_event_lock_irq(bitmap->write_wait,
 820                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 821                                             wake_up_process(bitmap->writeback_daemon->tsk));
 822                         spin_unlock_irq(&bitmap->write_lock);
 823                 } else
 824                         wait_event(bitmap->mddev->sb_wait,
 825                                    atomic_read(&bitmap->mddev->pending_writes)==0);
 826         }
 827         return 0;
 828 }
 829
 830 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 831 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 832  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 833  * memory mapping of the bitmap file
 834  * Special cases:
 835  *   if there's no bitmap file, or if the bitmap file had been
 836  *   previously kicked from the array, we mark all the bits as
 837  *   1's in order to cause a full resync.
 838  *
 839  * We ignore all bits for sectors that end earlier than 'start'.
 840  * This is used when reading an out-of-date bitmap...
 841  */
 842 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 843 {
 844         unsigned long i, chunks, index, oldindex, bit;
 845         struct page *page = NULL, *oldpage = NULL;
 846         unsigned long num_pages, bit_cnt = 0;
 847         struct file *file;
 848         unsigned long bytes, offset, dummy;
 849         int outofdate;
 850         int ret = -ENOSPC;
 851
 852         chunks = bitmap->chunks;
 853         file = bitmap->file;
 854
 855         BUG_ON(!file && !bitmap->offset);
 856
 857 #ifdef INJECT_FAULTS_3
 858         outofdate = 1;
 859 #else
 860         outofdate = bitmap->flags & BITMAP_STALE;
 861 #endif
 862         if (outofdate)
 863                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 864                         "recovery\n", bmname(bitmap));
 865
 866         bytes = (chunks + 7) / 8;
 867
 868         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 869
 870         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 871                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 872                         bmname(bitmap),
 873                         (unsigned long) i_size_read(file->f_mapping->host),
 874                         bytes + sizeof(bitmap_super_t));
 875                 goto out;
 876         }
 877
 878         ret = -ENOMEM;
 879
 880         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 881         if (!bitmap->filemap)
 882                 goto out;
 883
 884         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 885         if (!bitmap->filemap_attr)
 886                 goto out;
 887
 888         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 889
 890         oldindex = ~0L;
 891
 892         for (i = 0; i < chunks; i++) {
 893                 index = file_page_index(i);
 894                 bit = file_page_offset(i);
 895                 if (index != oldindex) { /* this is a new page, read it in */
 896                         /* unmap the old page, we're done with it */
 897                         if (oldpage != NULL)
 898                                 kunmap(oldpage);
 899                         if (index == 0) {
 900                                 /*
 901                                  * if we're here then the superblock page
 902                                  * contains some bits (PAGE_SIZE != sizeof sb)
 903                                  * we've already read it in, so just use it
 904                                  */
 905                                 page = bitmap->sb_page;
 906                                 offset = sizeof(bitmap_super_t);
 907                         } else if (file) {
 908                                 page = read_page(file, index, &dummy);
 909                                 offset = 0;
 910                         } else {
 911                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 912                                 offset = 0;
 913                         }
 914                         if (IS_ERR(page)) { /* read error */
 915                                 ret = PTR_ERR(page);
 916                                 goto out;
 917                         }
 918
 919                         oldindex = index;
 920                         oldpage = page;
 921                         kmap(page);
 922
 923                         if (outofdate) {
 924                                 /*
 925                                  * if bitmap is out of date, dirty the
 926                                  * whole page and write it out
 927                                  */
 928                                 memset(page_address(page) + offset, 0xff,
 929                                        PAGE_SIZE - offset);
 930                                 ret = write_page(bitmap, page, 1);
 931                                 if (ret) {
 932                                         kunmap(page);
 933                                         /* release, page not in filemap yet */
 934                                         page_cache_release(page);
 935                                         goto out;
 936                                 }
 937                         }
 938
 939                         bitmap->filemap[bitmap->file_pages++] = page;
 940                 }
 941                 if (test_bit(bit, page_address(page))) {
 942                         /* if the disk bit is set, set the memory bit */
 943                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 944                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 945                                 );
 946                         bit_cnt++;
 947                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 948                 }
 949         }
 950
 951         /* everything went OK */
 952         ret = 0;
 953         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 954
 955         if (page) /* unmap the last page */
 956                 kunmap(page);
 957
 958         if (bit_cnt) { /* Kick recovery if any bits were set */
 959                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 960                 md_wakeup_thread(bitmap->mddev->thread);
 961         }
 962
 963 out:
 964         printk(KERN_INFO "%s: bitmap initialized from disk: "
 965                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 966                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 967
 968         return ret;
 969 }
 970
 971 void bitmap_write_all(struct bitmap *bitmap)
 972 {
 973         /* We don't actually write all bitmap blocks here,
 974          * just flag them as needing to be written
 975          */
 976
 977         unsigned long chunks = bitmap->chunks;
 978         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 979         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 980         while (num_pages--)
 981                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 982 }
 983
 984
 985 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 986 {
 987         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 988         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 989         bitmap->bp[page].count += inc;
 990 /*
 991         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
 992                               (unsigned long long)offset, inc, bitmap->bp[page].count);
 993 */
 994         bitmap_checkfree(bitmap, page);
 995 }
 996 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
 997                                             sector_t offset, int *blocks,
 998                                             int create);
 999
1000 /*
1001  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1002  *                      out to disk
1003  */
1004
1005 int bitmap_daemon_work(struct bitmap *bitmap)
1006 {
1007         unsigned long j;
1008         unsigned long flags;
1009         struct page *page = NULL, *lastpage = NULL;
1010         int err = 0;
1011         int blocks;
1012         int attr;
1013
1014         if (bitmap == NULL)
1015                 return 0;
1016         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1017                 return 0;
1018         bitmap->daemon_lastrun = jiffies;
1019
1020         for (j = 0; j < bitmap->chunks; j++) {
1021                 bitmap_counter_t *bmc;
1022                 spin_lock_irqsave(&bitmap->lock, flags);
1023                 if (!bitmap->filemap) {
1024                         /* error or shutdown */
1025                         spin_unlock_irqrestore(&bitmap->lock, flags);
1026                         break;
1027                 }
1028
1029                 page = filemap_get_page(bitmap, j);
1030
1031                 if (page != lastpage) {
1032                         /* skip this page unless it's marked as needing cleaning */
1033                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1034                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1035                                         page_cache_get(page);
1036                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1037                                 }
1038                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1039                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1040                                         switch (write_page(bitmap, page, 0)) {
1041                                         case -EAGAIN:
1042                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1043                                                 break;
1044                                         case 0:
1045                                                 break;
1046                                         default:
1047                                                 bitmap_file_kick(bitmap);
1048                                         }
1049                                         page_cache_release(page);
1050                                 }
1051                                 continue;
1052                         }
1053
1054                         /* grab the new page, sync and release the old */
1055                         page_cache_get(page);
1056                         if (lastpage != NULL) {
1057                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1058                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1059                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1060                                         err = write_page(bitmap, lastpage, 0);
1061                                         if (err == -EAGAIN) {
1062                                                 err = 0;
1063                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1064                                         }
1065                                 } else {
1066                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1067                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1068                                 }
1069                                 kunmap(lastpage);
1070                                 page_cache_release(lastpage);
1071                                 if (err)
1072                                         bitmap_file_kick(bitmap);
1073                         } else
1074                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1075                         lastpage = page;
1076                         kmap(page);
1077 /*
1078                         printk("bitmap clean at page %lu\n", j);
1079 */
1080                         spin_lock_irqsave(&bitmap->lock, flags);
1081                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1082                 }
1083                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1084                                         &blocks, 0);
1085                 if (bmc) {
1086 /*
1087   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1088 */
1089                         if (*bmc == 2) {
1090                                 *bmc=1; /* maybe clear the bit next time */
1091                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1092                         } else if (*bmc == 1) {
1093                                 /* we can clear the bit */
1094                                 *bmc = 0;
1095                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1096                                                   -1);
1097
1098                                 /* clear the bit */
1099                                 clear_bit(file_page_offset(j), page_address(page));
1100                         }
1101                 }
1102                 spin_unlock_irqrestore(&bitmap->lock, flags);
1103         }
1104
1105         /* now sync the final page */
1106         if (lastpage != NULL) {
1107                 kunmap(lastpage);
1108                 spin_lock_irqsave(&bitmap->lock, flags);
1109                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1110                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1111                         spin_unlock_irqrestore(&bitmap->lock, flags);
1112                         err = write_page(bitmap, lastpage, 0);
1113                         if (err == -EAGAIN) {
1114                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1115                                 err = 0;
1116                         }
1117                 } else {
1118                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1119                         spin_unlock_irqrestore(&bitmap->lock, flags);
1120                 }
1121
1122                 page_cache_release(lastpage);
1123         }
1124
1125         return err;
1126 }
1127
1128 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1129 {
1130         mdk_thread_t *dmn;
1131         unsigned long flags;
1132
1133         /* if no one is waiting on us, we'll free the md thread struct
1134          * and exit, otherwise we let the waiter clean things up */
1135         spin_lock_irqsave(&bitmap->lock, flags);
1136         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1137                 *daemon = NULL;
1138                 spin_unlock_irqrestore(&bitmap->lock, flags);
1139                 kfree(dmn);
1140                 complete_and_exit(NULL, 0); /* do_exit not exported */
1141         }
1142         spin_unlock_irqrestore(&bitmap->lock, flags);
1143 }
1144
1145 static void bitmap_writeback_daemon(mddev_t *mddev)
1146 {
1147         struct bitmap *bitmap = mddev->bitmap;
1148         struct page *page;
1149         struct page_list *item;
1150         int err = 0;
1151
1152         if (signal_pending(current)) {
1153                 printk(KERN_INFO
1154                        "%s: bitmap writeback daemon got signal, exiting...\n",
1155                        bmname(bitmap));
1156                 err = -EINTR;
1157                 goto out;
1158         }
1159         if (bitmap == NULL)
1160                 /* about to be stopped. */
1161                 return;
1162
1163         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1164         /* wait on bitmap page writebacks */
1165         while ((item = dequeue_page(bitmap))) {
1166                 page = item->page;
1167                 mempool_free(item, bitmap->write_pool);
1168                 PRINTK("wait on page writeback: %p\n", page);
1169                 wait_on_page_writeback(page);
1170                 PRINTK("finished page writeback: %p\n", page);
1171
1172                 err = PageError(page);
1173                 page_cache_release(page);
1174                 if (err) {
1175                         printk(KERN_WARNING "%s: bitmap file writeback "
1176                                "failed (page %lu): %d\n",
1177                                bmname(bitmap), page->index, err);
1178                         bitmap_file_kick(bitmap);
1179                         goto out;
1180                 }
1181         }
1182  out:
1183         wake_up(&bitmap->write_wait);
1184         if (err) {
1185                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1186                        bmname(bitmap), err);
1187                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1188         }
1189 }
1190
1191 static mdk_thread_t *bitmap_start_daemon(struct bitmap *bitmap,
1192                                 void (*func)(mddev_t *), char *name)
1193 {
1194         mdk_thread_t *daemon;
1195         char namebuf[32];
1196
1197 #ifdef INJECT_FATAL_FAULT_2
1198         daemon = NULL;
1199 #else
1200         sprintf(namebuf, "%%s_%s", name);
1201         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1202 #endif
1203         if (!daemon) {
1204                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1205                         bmname(bitmap));
1206                 return ERR_PTR(-ECHILD);
1207         }
1208
1209         md_wakeup_thread(daemon); /* start it running */
1210
1211         PRINTK("%s: %s daemon (pid %d) started...\n",
1212                 bmname(bitmap), name, daemon->tsk->pid);
1213
1214         return daemon;
1215 }
1216
1217 static void bitmap_stop_daemon(struct bitmap *bitmap)
1218 {
1219         /* the daemon can't stop itself... it'll just exit instead... */
1220         if (bitmap->writeback_daemon && ! IS_ERR(bitmap->writeback_daemon) &&
1221             current->pid != bitmap->writeback_daemon->tsk->pid) {
1222                 mdk_thread_t *daemon;
1223                 unsigned long flags;
1224
1225                 spin_lock_irqsave(&bitmap->lock, flags);
1226                 daemon = bitmap->writeback_daemon;
1227                 bitmap->writeback_daemon = NULL;
1228                 spin_unlock_irqrestore(&bitmap->lock, flags);
1229                 if (daemon && ! IS_ERR(daemon))
1230                         md_unregister_thread(daemon); /* destroy the thread */
1231         }
1232 }
1233
1234 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1235                                             sector_t offset, int *blocks,
1236                                             int create)
1237 {
1238         /* If 'create', we might release the lock and reclaim it.
1239          * The lock must have been taken with interrupts enabled.
1240          * If !create, we don't release the lock.
1241          */
1242         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1243         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1244         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1245         sector_t csize;
1246
1247         if (bitmap_checkpage(bitmap, page, create) < 0) {
1248                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1249                 *blocks = csize - (offset & (csize- 1));
1250                 return NULL;
1251         }
1252         /* now locked ... */
1253
1254         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1255                 /* should we use the first or second counter field
1256                  * of the hijacked pointer? */
1257                 int hi = (pageoff > PAGE_COUNTER_MASK);
1258                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1259                                           PAGE_COUNTER_SHIFT - 1);
1260                 *blocks = csize - (offset & (csize- 1));
1261                 return  &((bitmap_counter_t *)
1262                           &bitmap->bp[page].map)[hi];
1263         } else { /* page is allocated */
1264                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1265                 *blocks = csize - (offset & (csize- 1));
1266                 return (bitmap_counter_t *)
1267                         &(bitmap->bp[page].map[pageoff]);
1268         }
1269 }
1270
1271 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1272 {
1273         if (!bitmap) return 0;
1274
1275         if (behind) {
1276                 atomic_inc(&bitmap->behind_writes);
1277                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1278                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1279         }
1280
1281         while (sectors) {
1282                 int blocks;
1283                 bitmap_counter_t *bmc;
1284
1285                 spin_lock_irq(&bitmap->lock);
1286                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1287                 if (!bmc) {
1288                         spin_unlock_irq(&bitmap->lock);
1289                         return 0;
1290                 }
1291
1292                 switch(*bmc) {
1293                 case 0:
1294                         bitmap_file_set_bit(bitmap, offset);
1295                         bitmap_count_page(bitmap,offset, 1);
1296                         blk_plug_device(bitmap->mddev->queue);
1297                         /* fall through */
1298                 case 1:
1299                         *bmc = 2;
1300                 }
1301                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1302                 (*bmc)++;
1303
1304                 spin_unlock_irq(&bitmap->lock);
1305
1306                 offset += blocks;
1307                 if (sectors > blocks)
1308                         sectors -= blocks;
1309                 else sectors = 0;
1310         }
1311         return 0;
1312 }
1313
1314 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1315                      int success, int behind)
1316 {
1317         if (!bitmap) return;
1318         if (behind) {
1319                 atomic_dec(&bitmap->behind_writes);
1320                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1321                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1322         }
1323
1324         while (sectors) {
1325                 int blocks;
1326                 unsigned long flags;
1327                 bitmap_counter_t *bmc;
1328
1329                 spin_lock_irqsave(&bitmap->lock, flags);
1330                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1331                 if (!bmc) {
1332                         spin_unlock_irqrestore(&bitmap->lock, flags);
1333                         return;
1334                 }
1335
1336                 if (!success && ! (*bmc & NEEDED_MASK))
1337                         *bmc |= NEEDED_MASK;
1338
1339                 (*bmc)--;
1340                 if (*bmc <= 2) {
1341                         set_page_attr(bitmap,
1342                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1343                                       BITMAP_PAGE_CLEAN);
1344                 }
1345                 spin_unlock_irqrestore(&bitmap->lock, flags);
1346                 offset += blocks;
1347                 if (sectors > blocks)
1348                         sectors -= blocks;
1349                 else sectors = 0;
1350         }
1351 }
1352
1353 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1354                         int degraded)
1355 {
1356         bitmap_counter_t *bmc;
1357         int rv;
1358         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1359                 *blocks = 1024;
1360                 return 1; /* always resync if no bitmap */
1361         }
1362         spin_lock_irq(&bitmap->lock);
1363         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1364         rv = 0;
1365         if (bmc) {
1366                 /* locked */
1367                 if (RESYNC(*bmc))
1368                         rv = 1;
1369                 else if (NEEDED(*bmc)) {
1370                         rv = 1;
1371                         if (!degraded) { /* don't set/clear bits if degraded */
1372                                 *bmc |= RESYNC_MASK;
1373                                 *bmc &= ~NEEDED_MASK;
1374                         }
1375                 }
1376         }
1377         spin_unlock_irq(&bitmap->lock);
1378         return rv;
1379 }
1380
1381 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1382 {
1383         bitmap_counter_t *bmc;
1384         unsigned long flags;
1385 /*
1386         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1387 */      if (bitmap == NULL) {
1388                 *blocks = 1024;
1389                 return;
1390         }
1391         spin_lock_irqsave(&bitmap->lock, flags);
1392         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1393         if (bmc == NULL)
1394                 goto unlock;
1395         /* locked */
1396 /*
1397         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1398 */
1399         if (RESYNC(*bmc)) {
1400                 *bmc &= ~RESYNC_MASK;
1401
1402                 if (!NEEDED(*bmc) && aborted)
1403                         *bmc |= NEEDED_MASK;
1404                 else {
1405                         if (*bmc <= 2) {
1406                                 set_page_attr(bitmap,
1407                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1408                                               BITMAP_PAGE_CLEAN);
1409                         }
1410                 }
1411         }
1412  unlock:
1413         spin_unlock_irqrestore(&bitmap->lock, flags);
1414 }
1415
1416 void bitmap_close_sync(struct bitmap *bitmap)
1417 {
1418         /* Sync has finished, and any bitmap chunks that weren't synced
1419          * properly have been aborted.  It remains to us to clear the
1420          * RESYNC bit wherever it is still on
1421          */
1422         sector_t sector = 0;
1423         int blocks;
1424         if (!bitmap) return;
1425         while (sector < bitmap->mddev->resync_max_sectors) {
1426                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1427 /*
1428                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1429                                          (unsigned long long)sector, blocks);
1430 */              sector += blocks;
1431         }
1432 }
1433
1434 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1435 {
1436         /* For each chunk covered by any of these sectors, set the
1437          * counter to 1 and set resync_needed.  They should all
1438          * be 0 at this point
1439          */
1440
1441         int secs;
1442         bitmap_counter_t *bmc;
1443         spin_lock_irq(&bitmap->lock);
1444         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1445         if (!bmc) {
1446                 spin_unlock_irq(&bitmap->lock);
1447                 return;
1448         }
1449         if (! *bmc) {
1450                 struct page *page;
1451                 *bmc = 1 | (needed?NEEDED_MASK:0);
1452                 bitmap_count_page(bitmap, offset, 1);
1453                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1454                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1455         }
1456         spin_unlock_irq(&bitmap->lock);
1457
1458 }
1459
1460 /*
1461  * flush out any pending updates
1462  */
1463 void bitmap_flush(mddev_t *mddev)
1464 {
1465         struct bitmap *bitmap = mddev->bitmap;
1466         int sleep;
1467
1468         if (!bitmap) /* there was no bitmap */
1469                 return;
1470
1471         /* run the daemon_work three time to ensure everything is flushed
1472          * that can be
1473          */
1474         sleep = bitmap->daemon_sleep;
1475         bitmap->daemon_sleep = 0;
1476         bitmap_daemon_work(bitmap);
1477         bitmap_daemon_work(bitmap);
1478         bitmap_daemon_work(bitmap);
1479         bitmap->daemon_sleep = sleep;
1480         bitmap_update_sb(bitmap);
1481 }
1482
1483 /*
1484  * free memory that was allocated
1485  */
1486 static void bitmap_free(struct bitmap *bitmap)
1487 {
1488         unsigned long k, pages;
1489         struct bitmap_page *bp;
1490
1491         if (!bitmap) /* there was no bitmap */
1492                 return;
1493
1494         /* release the bitmap file and kill the daemon */
1495         bitmap_file_put(bitmap);
1496
1497         bp = bitmap->bp;
1498         pages = bitmap->pages;
1499
1500         /* free all allocated memory */
1501
1502         mempool_destroy(bitmap->write_pool);
1503
1504         if (bp) /* deallocate the page memory */
1505                 for (k = 0; k < pages; k++)
1506                         if (bp[k].map && !bp[k].hijacked)
1507                                 kfree(bp[k].map);
1508         kfree(bp);
1509         kfree(bitmap);
1510 }
1511 void bitmap_destroy(mddev_t *mddev)
1512 {
1513         struct bitmap *bitmap = mddev->bitmap;
1514
1515         if (!bitmap) /* there was no bitmap */
1516                 return;
1517
1518         mddev->bitmap = NULL; /* disconnect from the md device */
1519
1520         bitmap_free(bitmap);
1521 }
1522
1523 /*
1524  * initialize the bitmap structure
1525  * if this returns an error, bitmap_destroy must be called to do clean up
1526  */
1527 int bitmap_create(mddev_t *mddev)
1528 {
1529         struct bitmap *bitmap;
1530         unsigned long blocks = mddev->resync_max_sectors;
1531         unsigned long chunks;
1532         unsigned long pages;
1533         struct file *file = mddev->bitmap_file;
1534         int err;
1535         sector_t start;
1536
1537         BUG_ON(sizeof(bitmap_super_t) != 256);
1538
1539         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1540                 return 0;
1541
1542         BUG_ON(file && mddev->bitmap_offset);
1543
1544         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1545         if (!bitmap)
1546                 return -ENOMEM;
1547
1548         memset(bitmap, 0, sizeof(*bitmap));
1549
1550         spin_lock_init(&bitmap->lock);
1551         bitmap->mddev = mddev;
1552
1553         spin_lock_init(&bitmap->write_lock);
1554         INIT_LIST_HEAD(&bitmap->complete_pages);
1555         init_waitqueue_head(&bitmap->write_wait);
1556         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1557                                 write_pool_free, NULL);
1558         err = -ENOMEM;
1559         if (!bitmap->write_pool)
1560                 goto error;
1561
1562         bitmap->file = file;
1563         bitmap->offset = mddev->bitmap_offset;
1564         if (file) get_file(file);
1565         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1566         err = bitmap_read_sb(bitmap);
1567         if (err)
1568                 goto error;
1569
1570         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1571                                         sizeof(bitmap->chunksize));
1572
1573         /* now that chunksize and chunkshift are set, we can use these macros */
1574         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1575                         CHUNK_BLOCK_RATIO(bitmap);
1576         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1577
1578         BUG_ON(!pages);
1579
1580         bitmap->chunks = chunks;
1581         bitmap->pages = pages;
1582         bitmap->missing_pages = pages;
1583         bitmap->counter_bits = COUNTER_BITS;
1584
1585         bitmap->syncchunk = ~0UL;
1586
1587 #ifdef INJECT_FATAL_FAULT_1
1588         bitmap->bp = NULL;
1589 #else
1590         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1591 #endif
1592         err = -ENOMEM;
1593         if (!bitmap->bp)
1594                 goto error;
1595         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
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         return bitmap_update_sb(bitmap);
1626
1627  error:
1628         bitmap_free(bitmap);
1629         return err;
1630 }
1631
1632 /* the bitmap API -- for raid personalities */
1633 EXPORT_SYMBOL(bitmap_startwrite);
1634 EXPORT_SYMBOL(bitmap_endwrite);
1635 EXPORT_SYMBOL(bitmap_start_sync);
1636 EXPORT_SYMBOL(bitmap_end_sync);
1637 EXPORT_SYMBOL(bitmap_unplug);
1638 EXPORT_SYMBOL(bitmap_close_sync);
1639 EXPORT_SYMBOL(bitmap_daemon_work);