drivers/md/bitmap.c

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