SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/md/bitmap.c

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