drivers/md/bitmap.c

   1 /*
   2  * bitmap.c two-level bitmap (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
   3  *
   4  * bitmap_create  - sets up the bitmap structure
   5  * bitmap_destroy - destroys the bitmap structure
   6  *
   7  * additions, Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.:
   8  * - added disk storage for bitmap
   9  * - changes to allow various bitmap chunk sizes
  10  * - added bitmap daemon (to asynchronously clear bitmap bits from disk)
  11  */
  12
  13 /*
  14  * Still to do:
  15  *
  16  * flush after percent set rather than just time based. (maybe both).
  17  * wait if count gets too high, wake when it drops to half.
  18  * allow bitmap to be mirrored with superblock (before or after...)
  19  * allow hot-add to re-instate a current device.
  20  * allow hot-add of bitmap after quiessing device
  21  */
  22
  23 #include <linux/module.h>
  24 #include <linux/errno.h>
  25 #include <linux/slab.h>
  26 #include <linux/init.h>
  27 #include <linux/config.h>
  28 #include <linux/timer.h>
  29 #include <linux/sched.h>
  30 #include <linux/list.h>
  31 #include <linux/file.h>
  32 #include <linux/mount.h>
  33 #include <linux/buffer_head.h>
  34 #include <linux/raid/md.h>
  35 #include <linux/raid/bitmap.h>
  36
  37 /* debug macros */
  38
  39 #define DEBUG 0
  40
  41 #if DEBUG
  42 /* these are for debugging purposes only! */
  43
  44 /* define one and only one of these */
  45 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  46 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  47 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  48 #define INJECT_FAULTS_4 0 /* undef */
  49 #define INJECT_FAULTS_5 0 /* undef */
  50 #define INJECT_FAULTS_6 0
  51
  52 /* if these are defined, the driver will fail! debug only */
  53 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  54 #define INJECT_FATAL_FAULT_2 0 /* undef */
  55 #define INJECT_FATAL_FAULT_3 0 /* undef */
  56 #endif
  57
  58 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
  59 #define DPRINTK(x...) do { } while(0)
  60
  61 #ifndef PRINTK
  62 #  if DEBUG > 0
  63 #    define PRINTK(x...) printk(KERN_DEBUG x)
  64 #  else
  65 #    define PRINTK(x...)
  66 #  endif
  67 #endif
  68
  69 static inline char * bmname(struct bitmap *bitmap)
  70 {
  71         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  72 }
  73
  74
  75 /*
  76  * test if the bitmap is active
  77  */
  78 int bitmap_active(struct bitmap *bitmap)
  79 {
  80         unsigned long flags;
  81         int res = 0;
  82
  83         if (!bitmap)
  84                 return res;
  85         spin_lock_irqsave(&bitmap->lock, flags);
  86         res = bitmap->flags & BITMAP_ACTIVE;
  87         spin_unlock_irqrestore(&bitmap->lock, flags);
  88         return res;
  89 }
  90
  91 #define WRITE_POOL_SIZE 256
  92 /* mempool for queueing pending writes on the bitmap file */
  93 static void *write_pool_alloc(gfp_t gfp_flags, void *data)
  94 {
  95         return kmalloc(sizeof(struct page_list), gfp_flags);
  96 }
  97
  98 static void write_pool_free(void *ptr, void *data)
  99 {
 100         kfree(ptr);
 101 }
 102
 103 /*
 104  * just a placeholder - calls kmalloc for bitmap pages
 105  */
 106 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
 107 {
 108         unsigned char *page;
 109
 110 #ifdef INJECT_FAULTS_1
 111         page = NULL;
 112 #else
 113         page = kmalloc(PAGE_SIZE, GFP_NOIO);
 114 #endif
 115         if (!page)
 116                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
 117         else
 118                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
 119                         bmname(bitmap), page);
 120         return page;
 121 }
 122
 123 /*
 124  * for now just a placeholder -- just calls kfree for bitmap pages
 125  */
 126 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
 127 {
 128         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
 129         kfree(page);
 130 }
 131
 132 /*
 133  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
 134  *
 135  * 1) check to see if this page is allocated, if it's not then try to alloc
 136  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 137  *    page pointer directly as a counter
 138  *
 139  * if we find our page, we increment the page's refcount so that it stays
 140  * allocated while we're using it
 141  */
 142 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
 143 {
 144         unsigned char *mappage;
 145
 146         if (page >= bitmap->pages) {
 147                 printk(KERN_ALERT
 148                         "%s: invalid bitmap page request: %lu (> %lu)\n",
 149                         bmname(bitmap), page, bitmap->pages-1);
 150                 return -EINVAL;
 151         }
 152
 153
 154         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 155                 return 0;
 156
 157         if (bitmap->bp[page].map) /* page is already allocated, just return */
 158                 return 0;
 159
 160         if (!create)
 161                 return -ENOENT;
 162
 163         spin_unlock_irq(&bitmap->lock);
 164
 165         /* this page has not been allocated yet */
 166
 167         if ((mappage = bitmap_alloc_page(bitmap)) == NULL) {
 168                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 169                         bmname(bitmap));
 170                 /* failed - set the hijacked flag so that we can use the
 171                  * pointer as a counter */
 172                 spin_lock_irq(&bitmap->lock);
 173                 if (!bitmap->bp[page].map)
 174                         bitmap->bp[page].hijacked = 1;
 175                 goto out;
 176         }
 177
 178         /* got a page */
 179
 180         spin_lock_irq(&bitmap->lock);
 181
 182         /* recheck the page */
 183
 184         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
 185                 /* somebody beat us to getting the page */
 186                 bitmap_free_page(bitmap, mappage);
 187                 return 0;
 188         }
 189
 190         /* no page was in place and we have one, so install it */
 191
 192         memset(mappage, 0, PAGE_SIZE);
 193         bitmap->bp[page].map = mappage;
 194         bitmap->missing_pages--;
 195 out:
 196         return 0;
 197 }
 198
 199
 200 /* if page is completely empty, put it back on the free list, or dealloc it */
 201 /* if page was hijacked, unmark the flag so it might get alloced next time */
 202 /* Note: lock should be held when calling this */
 203 static inline void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 204 {
 205         char *ptr;
 206
 207         if (bitmap->bp[page].count) /* page is still busy */
 208                 return;
 209
 210         /* page is no longer in use, it can be released */
 211
 212         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 213                 bitmap->bp[page].hijacked = 0;
 214                 bitmap->bp[page].map = NULL;
 215                 return;
 216         }
 217
 218         /* normal case, free the page */
 219
 220 #if 0
 221 /* actually ... let's not.  We will probably need the page again exactly when
 222  * memory is tight and we are flusing to disk
 223  */
 224         return;
 225 #else
 226         ptr = bitmap->bp[page].map;
 227         bitmap->bp[page].map = NULL;
 228         bitmap->missing_pages++;
 229         bitmap_free_page(bitmap, ptr);
 230         return;
 231 #endif
 232 }
 233
 234
 235 /*
 236  * bitmap file handling - read and write the bitmap file and its superblock
 237  */
 238
 239 /* copy the pathname of a file to a buffer */
 240 char *file_path(struct file *file, char *buf, int count)
 241 {
 242         struct dentry *d;
 243         struct vfsmount *v;
 244
 245         if (!buf)
 246                 return NULL;
 247
 248         d = file->f_dentry;
 249         v = file->f_vfsmnt;
 250
 251         buf = d_path(d, v, buf, count);
 252
 253         return IS_ERR(buf) ? NULL : buf;
 254 }
 255
 256 /*
 257  * basic page I/O operations
 258  */
 259
 260 /* IO operations when bitmap is stored near all superblocks */
 261 static struct page *read_sb_page(mddev_t *mddev, long offset, unsigned long index)
 262 {
 263         /* choose a good rdev and read the page from there */
 264
 265         mdk_rdev_t *rdev;
 266         struct list_head *tmp;
 267         struct page *page = alloc_page(GFP_KERNEL);
 268         sector_t target;
 269
 270         if (!page)
 271                 return ERR_PTR(-ENOMEM);
 272
 273         ITERATE_RDEV(mddev, rdev, tmp) {
 274                 if (! test_bit(In_sync, &rdev->flags)
 275                     || test_bit(Faulty, &rdev->flags))
 276                         continue;
 277
 278                 target = (rdev->sb_offset << 1) + offset + index * (PAGE_SIZE/512);
 279
 280                 if (sync_page_io(rdev->bdev, target, PAGE_SIZE, page, READ)) {
 281                         page->index = index;
 282                         return page;
 283                 }
 284         }
 285         return ERR_PTR(-EIO);
 286
 287 }
 288
 289 static int write_sb_page(mddev_t *mddev, long offset, struct page *page, int wait)
 290 {
 291         mdk_rdev_t *rdev;
 292         struct list_head *tmp;
 293
 294         ITERATE_RDEV(mddev, rdev, tmp)
 295                 if (test_bit(In_sync, &rdev->flags)
 296                     && !test_bit(Faulty, &rdev->flags))
 297                         md_super_write(mddev, rdev,
 298                                        (rdev->sb_offset<<1) + offset
 299                                        + page->index * (PAGE_SIZE/512),
 300                                        PAGE_SIZE,
 301                                        page);
 302
 303         if (wait)
 304                 md_super_wait(mddev);
 305         return 0;
 306 }
 307
 308 /*
 309  * write out a page to a file
 310  */
 311 static int write_page(struct bitmap *bitmap, struct page *page, int wait)
 312 {
 313         int ret = -ENOMEM;
 314
 315         if (bitmap->file == NULL)
 316                 return write_sb_page(bitmap->mddev, bitmap->offset, page, wait);
 317
 318         if (wait)
 319                 lock_page(page);
 320         else {
 321                 if (TestSetPageLocked(page))
 322                         return -EAGAIN; /* already locked */
 323                 if (PageWriteback(page)) {
 324                         unlock_page(page);
 325                         return -EAGAIN;
 326                 }
 327         }
 328
 329         ret = page->mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 330         if (!ret)
 331                 ret = page->mapping->a_ops->commit_write(NULL, page, 0,
 332                         PAGE_SIZE);
 333         if (ret) {
 334                 unlock_page(page);
 335                 return ret;
 336         }
 337
 338         set_page_dirty(page); /* force it to be written out */
 339
 340         if (!wait) {
 341                 /* add to list to be waited for by daemon */
 342                 struct page_list *item = mempool_alloc(bitmap->write_pool, GFP_NOIO);
 343                 item->page = page;
 344                 page_cache_get(page);
 345                 spin_lock(&bitmap->write_lock);
 346                 list_add(&item->list, &bitmap->complete_pages);
 347                 spin_unlock(&bitmap->write_lock);
 348                 md_wakeup_thread(bitmap->writeback_daemon);
 349         }
 350         return write_one_page(page, wait);
 351 }
 352
 353 /* read a page from a file, pinning it into cache, and return bytes_read */
 354 static struct page *read_page(struct file *file, unsigned long index,
 355                                         unsigned long *bytes_read)
 356 {
 357         struct inode *inode = file->f_mapping->host;
 358         struct page *page = NULL;
 359         loff_t isize = i_size_read(inode);
 360         unsigned long end_index = isize >> PAGE_CACHE_SHIFT;
 361
 362         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_CACHE_SIZE,
 363                         (unsigned long long)index << PAGE_CACHE_SHIFT);
 364
 365         page = read_cache_page(inode->i_mapping, index,
 366                         (filler_t *)inode->i_mapping->a_ops->readpage, file);
 367         if (IS_ERR(page))
 368                 goto out;
 369         wait_on_page_locked(page);
 370         if (!PageUptodate(page) || PageError(page)) {
 371                 page_cache_release(page);
 372                 page = ERR_PTR(-EIO);
 373                 goto out;
 374         }
 375
 376         if (index > end_index) /* we have read beyond EOF */
 377                 *bytes_read = 0;
 378         else if (index == end_index) /* possible short read */
 379                 *bytes_read = isize & ~PAGE_CACHE_MASK;
 380         else
 381                 *bytes_read = PAGE_CACHE_SIZE; /* got a full page */
 382 out:
 383         if (IS_ERR(page))
 384                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
 385                         (int)PAGE_CACHE_SIZE,
 386                         (unsigned long long)index << PAGE_CACHE_SHIFT,
 387                         PTR_ERR(page));
 388         return page;
 389 }
 390
 391 /*
 392  * bitmap file superblock operations
 393  */
 394
 395 /* update the event counter and sync the superblock to disk */
 396 int bitmap_update_sb(struct bitmap *bitmap)
 397 {
 398         bitmap_super_t *sb;
 399         unsigned long flags;
 400
 401         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 402                 return 0;
 403         spin_lock_irqsave(&bitmap->lock, flags);
 404         if (!bitmap->sb_page) { /* no superblock */
 405                 spin_unlock_irqrestore(&bitmap->lock, flags);
 406                 return 0;
 407         }
 408         spin_unlock_irqrestore(&bitmap->lock, flags);
 409         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 410         sb->events = cpu_to_le64(bitmap->mddev->events);
 411         if (!bitmap->mddev->degraded)
 412                 sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 413         kunmap(bitmap->sb_page);
 414         return write_page(bitmap, bitmap->sb_page, 1);
 415 }
 416
 417 /* print out the bitmap file superblock */
 418 void bitmap_print_sb(struct bitmap *bitmap)
 419 {
 420         bitmap_super_t *sb;
 421
 422         if (!bitmap || !bitmap->sb_page)
 423                 return;
 424         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 425         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 426         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 427         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 428         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 429                                         *(__u32 *)(sb->uuid+0),
 430                                         *(__u32 *)(sb->uuid+4),
 431                                         *(__u32 *)(sb->uuid+8),
 432                                         *(__u32 *)(sb->uuid+12));
 433         printk(KERN_DEBUG "        events: %llu\n",
 434                         (unsigned long long) le64_to_cpu(sb->events));
 435         printk(KERN_DEBUG "events cleared: %llu\n",
 436                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 437         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 438         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 439         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 440         printk(KERN_DEBUG "     sync size: %llu KB\n",
 441                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 442         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 443         kunmap(bitmap->sb_page);
 444 }
 445
 446 /* read the superblock from the bitmap file and initialize some bitmap fields */
 447 static int bitmap_read_sb(struct bitmap *bitmap)
 448 {
 449         char *reason = NULL;
 450         bitmap_super_t *sb;
 451         unsigned long chunksize, daemon_sleep, write_behind;
 452         unsigned long bytes_read;
 453         unsigned long long events;
 454         int err = -EINVAL;
 455
 456         /* page 0 is the superblock, read it... */
 457         if (bitmap->file)
 458                 bitmap->sb_page = read_page(bitmap->file, 0, &bytes_read);
 459         else {
 460                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset, 0);
 461                 bytes_read = PAGE_SIZE;
 462         }
 463         if (IS_ERR(bitmap->sb_page)) {
 464                 err = PTR_ERR(bitmap->sb_page);
 465                 bitmap->sb_page = NULL;
 466                 return err;
 467         }
 468
 469         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 470
 471         if (bytes_read < sizeof(*sb)) { /* short read */
 472                 printk(KERN_INFO "%s: bitmap file superblock truncated\n",
 473                         bmname(bitmap));
 474                 err = -ENOSPC;
 475                 goto out;
 476         }
 477
 478         chunksize = le32_to_cpu(sb->chunksize);
 479         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
 480         write_behind = le32_to_cpu(sb->write_behind);
 481
 482         /* verify that the bitmap-specific fields are valid */
 483         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 484                 reason = "bad magic";
 485         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 486                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 487                 reason = "unrecognized superblock version";
 488         else if (chunksize < 512 || chunksize > (1024 * 1024 * 4))
 489                 reason = "bitmap chunksize out of range (512B - 4MB)";
 490         else if ((1 << ffz(~chunksize)) != chunksize)
 491                 reason = "bitmap chunksize not a power of 2";
 492         else if (daemon_sleep < 1 || daemon_sleep > 15)
 493                 reason = "daemon sleep period out of range (1-15s)";
 494         else if (write_behind > COUNTER_MAX)
 495                 reason = "write-behind limit out of range (0 - 16383)";
 496         if (reason) {
 497                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 498                         bmname(bitmap), reason);
 499                 goto out;
 500         }
 501
 502         /* keep the array size field of the bitmap superblock up to date */
 503         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 504
 505         if (!bitmap->mddev->persistent)
 506                 goto success;
 507
 508         /*
 509          * if we have a persistent array superblock, compare the
 510          * bitmap's UUID and event counter to the mddev's
 511          */
 512         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 513                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 514                         bmname(bitmap));
 515                 goto out;
 516         }
 517         events = le64_to_cpu(sb->events);
 518         if (events < bitmap->mddev->events) {
 519                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 520                         "-- forcing full recovery\n", bmname(bitmap), events,
 521                         (unsigned long long) bitmap->mddev->events);
 522                 sb->state |= BITMAP_STALE;
 523         }
 524 success:
 525         /* assign fields using values from superblock */
 526         bitmap->chunksize = chunksize;
 527         bitmap->daemon_sleep = daemon_sleep;
 528         bitmap->daemon_lastrun = jiffies;
 529         bitmap->max_write_behind = write_behind;
 530         bitmap->flags |= sb->state;
 531         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 532                 bitmap->flags |= BITMAP_HOSTENDIAN;
 533         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 534         if (sb->state & BITMAP_STALE)
 535                 bitmap->events_cleared = bitmap->mddev->events;
 536         err = 0;
 537 out:
 538         kunmap(bitmap->sb_page);
 539         if (err)
 540                 bitmap_print_sb(bitmap);
 541         return err;
 542 }
 543
 544 enum bitmap_mask_op {
 545         MASK_SET,
 546         MASK_UNSET
 547 };
 548
 549 /* record the state of the bitmap in the superblock */
 550 static void bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 551                                 enum bitmap_mask_op op)
 552 {
 553         bitmap_super_t *sb;
 554         unsigned long flags;
 555
 556         spin_lock_irqsave(&bitmap->lock, flags);
 557         if (!bitmap || !bitmap->sb_page) { /* can't set the state */
 558                 spin_unlock_irqrestore(&bitmap->lock, flags);
 559                 return;
 560         }
 561         page_cache_get(bitmap->sb_page);
 562         spin_unlock_irqrestore(&bitmap->lock, flags);
 563         sb = (bitmap_super_t *)kmap(bitmap->sb_page);
 564         switch (op) {
 565                 case MASK_SET: sb->state |= bits;
 566                                 break;
 567                 case MASK_UNSET: sb->state &= ~bits;
 568                                 break;
 569                 default: BUG();
 570         }
 571         kunmap(bitmap->sb_page);
 572         page_cache_release(bitmap->sb_page);
 573 }
 574
 575 /*
 576  * general bitmap file operations
 577  */
 578
 579 /* calculate the index of the page that contains this bit */
 580 static inline unsigned long file_page_index(unsigned long chunk)
 581 {
 582         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
 583 }
 584
 585 /* calculate the (bit) offset of this bit within a page */
 586 static inline unsigned long file_page_offset(unsigned long chunk)
 587 {
 588         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
 589 }
 590
 591 /*
 592  * return a pointer to the page in the filemap that contains the given bit
 593  *
 594  * this lookup is complicated by the fact that the bitmap sb might be exactly
 595  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 596  * 0 or page 1
 597  */
 598 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 599                                         unsigned long chunk)
 600 {
 601         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
 602 }
 603
 604
 605 static void bitmap_file_unmap(struct bitmap *bitmap)
 606 {
 607         struct page **map, *sb_page;
 608         unsigned long *attr;
 609         int pages;
 610         unsigned long flags;
 611
 612         spin_lock_irqsave(&bitmap->lock, flags);
 613         map = bitmap->filemap;
 614         bitmap->filemap = NULL;
 615         attr = bitmap->filemap_attr;
 616         bitmap->filemap_attr = NULL;
 617         pages = bitmap->file_pages;
 618         bitmap->file_pages = 0;
 619         sb_page = bitmap->sb_page;
 620         bitmap->sb_page = NULL;
 621         spin_unlock_irqrestore(&bitmap->lock, flags);
 622
 623         while (pages--)
 624                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
 625                         page_cache_release(map[pages]);
 626         kfree(map);
 627         kfree(attr);
 628
 629         if (sb_page)
 630                 page_cache_release(sb_page);
 631 }
 632
 633 static void bitmap_stop_daemon(struct bitmap *bitmap);
 634
 635 /* dequeue the next item in a page list -- don't call from irq context */
 636 static struct page_list *dequeue_page(struct bitmap *bitmap)
 637 {
 638         struct page_list *item = NULL;
 639         struct list_head *head = &bitmap->complete_pages;
 640
 641         spin_lock(&bitmap->write_lock);
 642         if (list_empty(head))
 643                 goto out;
 644         item = list_entry(head->prev, struct page_list, list);
 645         list_del(head->prev);
 646 out:
 647         spin_unlock(&bitmap->write_lock);
 648         return item;
 649 }
 650
 651 static void drain_write_queues(struct bitmap *bitmap)
 652 {
 653         struct page_list *item;
 654
 655         while ((item = dequeue_page(bitmap))) {
 656                 /* don't bother to wait */
 657                 page_cache_release(item->page);
 658                 mempool_free(item, bitmap->write_pool);
 659         }
 660
 661         wake_up(&bitmap->write_wait);
 662 }
 663
 664 static void bitmap_file_put(struct bitmap *bitmap)
 665 {
 666         struct file *file;
 667         struct inode *inode;
 668         unsigned long flags;
 669
 670         spin_lock_irqsave(&bitmap->lock, flags);
 671         file = bitmap->file;
 672         bitmap->file = NULL;
 673         spin_unlock_irqrestore(&bitmap->lock, flags);
 674
 675         bitmap_stop_daemon(bitmap);
 676
 677         drain_write_queues(bitmap);
 678
 679         bitmap_file_unmap(bitmap);
 680
 681         if (file) {
 682                 inode = file->f_mapping->host;
 683                 spin_lock(&inode->i_lock);
 684                 atomic_set(&inode->i_writecount, 1); /* allow writes again */
 685                 spin_unlock(&inode->i_lock);
 686                 fput(file);
 687         }
 688 }
 689
 690
 691 /*
 692  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 693  * then it is no longer reliable, so we stop using it and we mark the file
 694  * as failed in the superblock
 695  */
 696 static void bitmap_file_kick(struct bitmap *bitmap)
 697 {
 698         char *path, *ptr = NULL;
 699
 700         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET);
 701         bitmap_update_sb(bitmap);
 702
 703         if (bitmap->file) {
 704                 path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 705                 if (path)
 706                         ptr = file_path(bitmap->file, path, PAGE_SIZE);
 707
 708                 printk(KERN_ALERT "%s: kicking failed bitmap file %s from array!\n",
 709                        bmname(bitmap), ptr ? ptr : "");
 710
 711                 kfree(path);
 712         }
 713
 714         bitmap_file_put(bitmap);
 715
 716         return;
 717 }
 718
 719 enum bitmap_page_attr {
 720         BITMAP_PAGE_DIRTY = 1, // there are set bits that need to be synced
 721         BITMAP_PAGE_CLEAN = 2, // there are bits that might need to be cleared
 722         BITMAP_PAGE_NEEDWRITE=4, // there are cleared bits that need to be synced
 723 };
 724
 725 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 726                                 enum bitmap_page_attr attr)
 727 {
 728         bitmap->filemap_attr[page->index] |= attr;
 729 }
 730
 731 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 732                                 enum bitmap_page_attr attr)
 733 {
 734         bitmap->filemap_attr[page->index] &= ~attr;
 735 }
 736
 737 static inline unsigned long get_page_attr(struct bitmap *bitmap, struct page *page)
 738 {
 739         return bitmap->filemap_attr[page->index];
 740 }
 741
 742 /*
 743  * bitmap_file_set_bit -- called before performing a write to the md device
 744  * to set (and eventually sync) a particular bit in the bitmap file
 745  *
 746  * we set the bit immediately, then we record the page number so that
 747  * when an unplug occurs, we can flush the dirty pages out to disk
 748  */
 749 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 750 {
 751         unsigned long bit;
 752         struct page *page;
 753         void *kaddr;
 754         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 755
 756         if (!bitmap->filemap) {
 757                 return;
 758         }
 759
 760         page = filemap_get_page(bitmap, chunk);
 761         bit = file_page_offset(chunk);
 762
 763
 764         /* make sure the page stays cached until it gets written out */
 765         if (! (get_page_attr(bitmap, page) & BITMAP_PAGE_DIRTY))
 766                 page_cache_get(page);
 767
 768         /* set the bit */
 769         kaddr = kmap_atomic(page, KM_USER0);
 770         if (bitmap->flags & BITMAP_HOSTENDIAN)
 771                 set_bit(bit, kaddr);
 772         else
 773                 ext2_set_bit(bit, kaddr);
 774         kunmap_atomic(kaddr, KM_USER0);
 775         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 776
 777         /* record page number so it gets flushed to disk when unplug occurs */
 778         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 779
 780 }
 781
 782 /* this gets called when the md device is ready to unplug its underlying
 783  * (slave) device queues -- before we let any writes go down, we need to
 784  * sync the dirty pages of the bitmap file to disk */
 785 int bitmap_unplug(struct bitmap *bitmap)
 786 {
 787         unsigned long i, attr, flags;
 788         struct page *page;
 789         int wait = 0;
 790         int err;
 791
 792         if (!bitmap)
 793                 return 0;
 794
 795         /* look at each page to see if there are any set bits that need to be
 796          * flushed out to disk */
 797         for (i = 0; i < bitmap->file_pages; i++) {
 798                 spin_lock_irqsave(&bitmap->lock, flags);
 799                 if (!bitmap->filemap) {
 800                         spin_unlock_irqrestore(&bitmap->lock, flags);
 801                         return 0;
 802                 }
 803                 page = bitmap->filemap[i];
 804                 attr = get_page_attr(bitmap, page);
 805                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 806                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 807                 if ((attr & BITMAP_PAGE_DIRTY))
 808                         wait = 1;
 809                 spin_unlock_irqrestore(&bitmap->lock, flags);
 810
 811                 if (attr & (BITMAP_PAGE_DIRTY | BITMAP_PAGE_NEEDWRITE)) {
 812                         err = write_page(bitmap, page, 0);
 813                         if (err == -EAGAIN) {
 814                                 if (attr & BITMAP_PAGE_DIRTY)
 815                                         err = write_page(bitmap, page, 1);
 816                                 else
 817                                         err = 0;
 818                         }
 819                         if (err)
 820                                 return 1;
 821                 }
 822         }
 823         if (wait) { /* if any writes were performed, we need to wait on them */
 824                 if (bitmap->file) {
 825                         spin_lock_irq(&bitmap->write_lock);
 826                         wait_event_lock_irq(bitmap->write_wait,
 827                                             list_empty(&bitmap->complete_pages), bitmap->write_lock,
 828                                             wake_up_process(bitmap->writeback_daemon->tsk));
 829                         spin_unlock_irq(&bitmap->write_lock);
 830                 } else
 831                         md_super_wait(bitmap->mddev);
 832         }
 833         return 0;
 834 }
 835
 836 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 837 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 838  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 839  * memory mapping of the bitmap file
 840  * Special cases:
 841  *   if there's no bitmap file, or if the bitmap file had been
 842  *   previously kicked from the array, we mark all the bits as
 843  *   1's in order to cause a full resync.
 844  *
 845  * We ignore all bits for sectors that end earlier than 'start'.
 846  * This is used when reading an out-of-date bitmap...
 847  */
 848 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 849 {
 850         unsigned long i, chunks, index, oldindex, bit;
 851         struct page *page = NULL, *oldpage = NULL;
 852         unsigned long num_pages, bit_cnt = 0;
 853         struct file *file;
 854         unsigned long bytes, offset, dummy;
 855         int outofdate;
 856         int ret = -ENOSPC;
 857
 858         chunks = bitmap->chunks;
 859         file = bitmap->file;
 860
 861         BUG_ON(!file && !bitmap->offset);
 862
 863 #ifdef INJECT_FAULTS_3
 864         outofdate = 1;
 865 #else
 866         outofdate = bitmap->flags & BITMAP_STALE;
 867 #endif
 868         if (outofdate)
 869                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 870                         "recovery\n", bmname(bitmap));
 871
 872         bytes = (chunks + 7) / 8;
 873
 874         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
 875
 876         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
 877                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 878                         bmname(bitmap),
 879                         (unsigned long) i_size_read(file->f_mapping->host),
 880                         bytes + sizeof(bitmap_super_t));
 881                 goto out;
 882         }
 883
 884         ret = -ENOMEM;
 885
 886         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 887         if (!bitmap->filemap)
 888                 goto out;
 889
 890         bitmap->filemap_attr = kmalloc(sizeof(long) * num_pages, GFP_KERNEL);
 891         if (!bitmap->filemap_attr)
 892                 goto out;
 893
 894         memset(bitmap->filemap_attr, 0, sizeof(long) * num_pages);
 895
 896         oldindex = ~0L;
 897
 898         for (i = 0; i < chunks; i++) {
 899                 int b;
 900                 index = file_page_index(i);
 901                 bit = file_page_offset(i);
 902                 if (index != oldindex) { /* this is a new page, read it in */
 903                         /* unmap the old page, we're done with it */
 904                         if (oldpage != NULL)
 905                                 kunmap(oldpage);
 906                         if (index == 0) {
 907                                 /*
 908                                  * if we're here then the superblock page
 909                                  * contains some bits (PAGE_SIZE != sizeof sb)
 910                                  * we've already read it in, so just use it
 911                                  */
 912                                 page = bitmap->sb_page;
 913                                 offset = sizeof(bitmap_super_t);
 914                         } else if (file) {
 915                                 page = read_page(file, index, &dummy);
 916                                 offset = 0;
 917                         } else {
 918                                 page = read_sb_page(bitmap->mddev, bitmap->offset, index);
 919                                 offset = 0;
 920                         }
 921                         if (IS_ERR(page)) { /* read error */
 922                                 ret = PTR_ERR(page);
 923                                 goto out;
 924                         }
 925
 926                         oldindex = index;
 927                         oldpage = page;
 928                         kmap(page);
 929
 930                         if (outofdate) {
 931                                 /*
 932                                  * if bitmap is out of date, dirty the
 933                                  * whole page and write it out
 934                                  */
 935                                 memset(page_address(page) + offset, 0xff,
 936                                        PAGE_SIZE - offset);
 937                                 ret = write_page(bitmap, page, 1);
 938                                 if (ret) {
 939                                         kunmap(page);
 940                                         /* release, page not in filemap yet */
 941                                         page_cache_release(page);
 942                                         goto out;
 943                                 }
 944                         }
 945
 946                         bitmap->filemap[bitmap->file_pages++] = page;
 947                 }
 948                 if (bitmap->flags & BITMAP_HOSTENDIAN)
 949                         b = test_bit(bit, page_address(page));
 950                 else
 951                         b = ext2_test_bit(bit, page_address(page));
 952                 if (b) {
 953                         /* if the disk bit is set, set the memory bit */
 954                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
 955                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
 956                                 );
 957                         bit_cnt++;
 958                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
 959                 }
 960         }
 961
 962         /* everything went OK */
 963         ret = 0;
 964         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
 965
 966         if (page) /* unmap the last page */
 967                 kunmap(page);
 968
 969         if (bit_cnt) { /* Kick recovery if any bits were set */
 970                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 971                 md_wakeup_thread(bitmap->mddev->thread);
 972         }
 973
 974 out:
 975         printk(KERN_INFO "%s: bitmap initialized from disk: "
 976                 "read %lu/%lu pages, set %lu bits, status: %d\n",
 977                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, ret);
 978
 979         return ret;
 980 }
 981
 982 void bitmap_write_all(struct bitmap *bitmap)
 983 {
 984         /* We don't actually write all bitmap blocks here,
 985          * just flag them as needing to be written
 986          */
 987
 988         unsigned long chunks = bitmap->chunks;
 989         unsigned long bytes = (chunks+7)/8 + sizeof(bitmap_super_t);
 990         unsigned long num_pages = (bytes + PAGE_SIZE-1) / PAGE_SIZE;
 991         while (num_pages--)
 992                 bitmap->filemap_attr[num_pages] |= BITMAP_PAGE_NEEDWRITE;
 993 }
 994
 995
 996 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
 997 {
 998         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
 999         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1000         bitmap->bp[page].count += inc;
1001 /*
1002         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
1003                               (unsigned long long)offset, inc, bitmap->bp[page].count);
1004 */
1005         bitmap_checkfree(bitmap, page);
1006 }
1007 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1008                                             sector_t offset, int *blocks,
1009                                             int create);
1010
1011 /*
1012  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1013  *                      out to disk
1014  */
1015
1016 int bitmap_daemon_work(struct bitmap *bitmap)
1017 {
1018         unsigned long j;
1019         unsigned long flags;
1020         struct page *page = NULL, *lastpage = NULL;
1021         int err = 0;
1022         int blocks;
1023         int attr;
1024
1025         if (bitmap == NULL)
1026                 return 0;
1027         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1028                 return 0;
1029         bitmap->daemon_lastrun = jiffies;
1030
1031         for (j = 0; j < bitmap->chunks; j++) {
1032                 bitmap_counter_t *bmc;
1033                 spin_lock_irqsave(&bitmap->lock, flags);
1034                 if (!bitmap->filemap) {
1035                         /* error or shutdown */
1036                         spin_unlock_irqrestore(&bitmap->lock, flags);
1037                         break;
1038                 }
1039
1040                 page = filemap_get_page(bitmap, j);
1041
1042                 if (page != lastpage) {
1043                         /* skip this page unless it's marked as needing cleaning */
1044                         if (!((attr=get_page_attr(bitmap, page)) & BITMAP_PAGE_CLEAN)) {
1045                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1046                                         page_cache_get(page);
1047                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1048                                 }
1049                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1050                                 if (attr & BITMAP_PAGE_NEEDWRITE) {
1051                                         switch (write_page(bitmap, page, 0)) {
1052                                         case -EAGAIN:
1053                                                 set_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1054                                                 break;
1055                                         case 0:
1056                                                 break;
1057                                         default:
1058                                                 bitmap_file_kick(bitmap);
1059                                         }
1060                                         page_cache_release(page);
1061                                 }
1062                                 continue;
1063                         }
1064
1065                         /* grab the new page, sync and release the old */
1066                         page_cache_get(page);
1067                         if (lastpage != NULL) {
1068                                 if (get_page_attr(bitmap, lastpage) & BITMAP_PAGE_NEEDWRITE) {
1069                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1070                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1071                                         err = write_page(bitmap, lastpage, 0);
1072                                         if (err == -EAGAIN) {
1073                                                 err = 0;
1074                                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1075                                         }
1076                                 } else {
1077                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1078                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1079                                 }
1080                                 kunmap(lastpage);
1081                                 page_cache_release(lastpage);
1082                                 if (err)
1083                                         bitmap_file_kick(bitmap);
1084                         } else
1085                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1086                         lastpage = page;
1087                         kmap(page);
1088 /*
1089                         printk("bitmap clean at page %lu\n", j);
1090 */
1091                         spin_lock_irqsave(&bitmap->lock, flags);
1092                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1093                 }
1094                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1095                                         &blocks, 0);
1096                 if (bmc) {
1097 /*
1098   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1099 */
1100                         if (*bmc == 2) {
1101                                 *bmc=1; /* maybe clear the bit next time */
1102                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1103                         } else if (*bmc == 1) {
1104                                 /* we can clear the bit */
1105                                 *bmc = 0;
1106                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1107                                                   -1);
1108
1109                                 /* clear the bit */
1110                                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1111                                         clear_bit(file_page_offset(j), page_address(page));
1112                                 else
1113                                         ext2_clear_bit(file_page_offset(j), page_address(page));
1114                         }
1115                 }
1116                 spin_unlock_irqrestore(&bitmap->lock, flags);
1117         }
1118
1119         /* now sync the final page */
1120         if (lastpage != NULL) {
1121                 kunmap(lastpage);
1122                 spin_lock_irqsave(&bitmap->lock, flags);
1123                 if (get_page_attr(bitmap, lastpage) &BITMAP_PAGE_NEEDWRITE) {
1124                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1125                         spin_unlock_irqrestore(&bitmap->lock, flags);
1126                         err = write_page(bitmap, lastpage, 0);
1127                         if (err == -EAGAIN) {
1128                                 set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1129                                 err = 0;
1130                         }
1131                 } else {
1132                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1133                         spin_unlock_irqrestore(&bitmap->lock, flags);
1134                 }
1135
1136                 page_cache_release(lastpage);
1137         }
1138
1139         return err;
1140 }
1141
1142 static void daemon_exit(struct bitmap *bitmap, mdk_thread_t **daemon)
1143 {
1144         mdk_thread_t *dmn;
1145         unsigned long flags;
1146
1147         /* if no one is waiting on us, we'll free the md thread struct
1148          * and exit, otherwise we let the waiter clean things up */
1149         spin_lock_irqsave(&bitmap->lock, flags);
1150         if ((dmn = *daemon)) { /* no one is waiting, cleanup and exit */
1151                 *daemon = NULL;
1152                 spin_unlock_irqrestore(&bitmap->lock, flags);
1153                 kfree(dmn);
1154                 complete_and_exit(NULL, 0); /* do_exit not exported */
1155         }
1156         spin_unlock_irqrestore(&bitmap->lock, flags);
1157 }
1158
1159 static void bitmap_writeback_daemon(mddev_t *mddev)
1160 {
1161         struct bitmap *bitmap = mddev->bitmap;
1162         struct page *page;
1163         struct page_list *item;
1164         int err = 0;
1165
1166         if (signal_pending(current)) {
1167                 printk(KERN_INFO
1168                        "%s: bitmap writeback daemon got signal, exiting...\n",
1169                        bmname(bitmap));
1170                 err = -EINTR;
1171                 goto out;
1172         }
1173         if (bitmap == NULL)
1174                 /* about to be stopped. */
1175                 return;
1176
1177         PRINTK("%s: bitmap writeback daemon woke up...\n", bmname(bitmap));
1178         /* wait on bitmap page writebacks */
1179         while ((item = dequeue_page(bitmap))) {
1180                 page = item->page;
1181                 mempool_free(item, bitmap->write_pool);
1182                 PRINTK("wait on page writeback: %p\n", page);
1183                 wait_on_page_writeback(page);
1184                 PRINTK("finished page writeback: %p\n", page);
1185
1186                 err = PageError(page);
1187                 page_cache_release(page);
1188                 if (err) {
1189                         printk(KERN_WARNING "%s: bitmap file writeback "
1190                                "failed (page %lu): %d\n",
1191                                bmname(bitmap), page->index, err);
1192                         bitmap_file_kick(bitmap);
1193                         goto out;
1194                 }
1195         }
1196  out:
1197         wake_up(&bitmap->write_wait);
1198         if (err) {
1199                 printk(KERN_INFO "%s: bitmap writeback daemon exiting (%d)\n",
1200                        bmname(bitmap), err);
1201                 daemon_exit(bitmap, &bitmap->writeback_daemon);
1202         }
1203 }
1204
1205 static mdk_thread_t *bitmap_start_daemon(struct bitmap *bitmap,
1206                                 void (*func)(mddev_t *), char *name)
1207 {
1208         mdk_thread_t *daemon;
1209         char namebuf[32];
1210
1211 #ifdef INJECT_FATAL_FAULT_2
1212         daemon = NULL;
1213 #else
1214         sprintf(namebuf, "%%s_%s", name);
1215         daemon = md_register_thread(func, bitmap->mddev, namebuf);
1216 #endif
1217         if (!daemon) {
1218                 printk(KERN_ERR "%s: failed to start bitmap daemon\n",
1219                         bmname(bitmap));
1220                 return ERR_PTR(-ECHILD);
1221         }
1222
1223         md_wakeup_thread(daemon); /* start it running */
1224
1225         PRINTK("%s: %s daemon (pid %d) started...\n",
1226                 bmname(bitmap), name, daemon->tsk->pid);
1227
1228         return daemon;
1229 }
1230
1231 static void bitmap_stop_daemon(struct bitmap *bitmap)
1232 {
1233         /* the daemon can't stop itself... it'll just exit instead... */
1234         if (bitmap->writeback_daemon && ! IS_ERR(bitmap->writeback_daemon) &&
1235             current->pid != bitmap->writeback_daemon->tsk->pid) {
1236                 mdk_thread_t *daemon;
1237                 unsigned long flags;
1238
1239                 spin_lock_irqsave(&bitmap->lock, flags);
1240                 daemon = bitmap->writeback_daemon;
1241                 bitmap->writeback_daemon = NULL;
1242                 spin_unlock_irqrestore(&bitmap->lock, flags);
1243                 if (daemon && ! IS_ERR(daemon))
1244                         md_unregister_thread(daemon); /* destroy the thread */
1245         }
1246 }
1247
1248 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1249                                             sector_t offset, int *blocks,
1250                                             int create)
1251 {
1252         /* If 'create', we might release the lock and reclaim it.
1253          * The lock must have been taken with interrupts enabled.
1254          * If !create, we don't release the lock.
1255          */
1256         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1257         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1258         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1259         sector_t csize;
1260
1261         if (bitmap_checkpage(bitmap, page, create) < 0) {
1262                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1263                 *blocks = csize - (offset & (csize- 1));
1264                 return NULL;
1265         }
1266         /* now locked ... */
1267
1268         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1269                 /* should we use the first or second counter field
1270                  * of the hijacked pointer? */
1271                 int hi = (pageoff > PAGE_COUNTER_MASK);
1272                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1273                                           PAGE_COUNTER_SHIFT - 1);
1274                 *blocks = csize - (offset & (csize- 1));
1275                 return  &((bitmap_counter_t *)
1276                           &bitmap->bp[page].map)[hi];
1277         } else { /* page is allocated */
1278                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1279                 *blocks = csize - (offset & (csize- 1));
1280                 return (bitmap_counter_t *)
1281                         &(bitmap->bp[page].map[pageoff]);
1282         }
1283 }
1284
1285 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1286 {
1287         if (!bitmap) return 0;
1288
1289         if (behind) {
1290                 atomic_inc(&bitmap->behind_writes);
1291                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1292                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1293         }
1294
1295         while (sectors) {
1296                 int blocks;
1297                 bitmap_counter_t *bmc;
1298
1299                 spin_lock_irq(&bitmap->lock);
1300                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1301                 if (!bmc) {
1302                         spin_unlock_irq(&bitmap->lock);
1303                         return 0;
1304                 }
1305
1306                 switch(*bmc) {
1307                 case 0:
1308                         bitmap_file_set_bit(bitmap, offset);
1309                         bitmap_count_page(bitmap,offset, 1);
1310                         blk_plug_device(bitmap->mddev->queue);
1311                         /* fall through */
1312                 case 1:
1313                         *bmc = 2;
1314                 }
1315                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX) BUG();
1316                 (*bmc)++;
1317
1318                 spin_unlock_irq(&bitmap->lock);
1319
1320                 offset += blocks;
1321                 if (sectors > blocks)
1322                         sectors -= blocks;
1323                 else sectors = 0;
1324         }
1325         return 0;
1326 }
1327
1328 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1329                      int success, int behind)
1330 {
1331         if (!bitmap) return;
1332         if (behind) {
1333                 atomic_dec(&bitmap->behind_writes);
1334                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1335                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1336         }
1337
1338         while (sectors) {
1339                 int blocks;
1340                 unsigned long flags;
1341                 bitmap_counter_t *bmc;
1342
1343                 spin_lock_irqsave(&bitmap->lock, flags);
1344                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1345                 if (!bmc) {
1346                         spin_unlock_irqrestore(&bitmap->lock, flags);
1347                         return;
1348                 }
1349
1350                 if (!success && ! (*bmc & NEEDED_MASK))
1351                         *bmc |= NEEDED_MASK;
1352
1353                 (*bmc)--;
1354                 if (*bmc <= 2) {
1355                         set_page_attr(bitmap,
1356                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1357                                       BITMAP_PAGE_CLEAN);
1358                 }
1359                 spin_unlock_irqrestore(&bitmap->lock, flags);
1360                 offset += blocks;
1361                 if (sectors > blocks)
1362                         sectors -= blocks;
1363                 else sectors = 0;
1364         }
1365 }
1366
1367 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1368                         int degraded)
1369 {
1370         bitmap_counter_t *bmc;
1371         int rv;
1372         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1373                 *blocks = 1024;
1374                 return 1; /* always resync if no bitmap */
1375         }
1376         spin_lock_irq(&bitmap->lock);
1377         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1378         rv = 0;
1379         if (bmc) {
1380                 /* locked */
1381                 if (RESYNC(*bmc))
1382                         rv = 1;
1383                 else if (NEEDED(*bmc)) {
1384                         rv = 1;
1385                         if (!degraded) { /* don't set/clear bits if degraded */
1386                                 *bmc |= RESYNC_MASK;
1387                                 *bmc &= ~NEEDED_MASK;
1388                         }
1389                 }
1390         }
1391         spin_unlock_irq(&bitmap->lock);
1392         return rv;
1393 }
1394
1395 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1396 {
1397         bitmap_counter_t *bmc;
1398         unsigned long flags;
1399 /*
1400         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1401 */      if (bitmap == NULL) {
1402                 *blocks = 1024;
1403                 return;
1404         }
1405         spin_lock_irqsave(&bitmap->lock, flags);
1406         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1407         if (bmc == NULL)
1408                 goto unlock;
1409         /* locked */
1410 /*
1411         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1412 */
1413         if (RESYNC(*bmc)) {
1414                 *bmc &= ~RESYNC_MASK;
1415
1416                 if (!NEEDED(*bmc) && aborted)
1417                         *bmc |= NEEDED_MASK;
1418                 else {
1419                         if (*bmc <= 2) {
1420                                 set_page_attr(bitmap,
1421                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1422                                               BITMAP_PAGE_CLEAN);
1423                         }
1424                 }
1425         }
1426  unlock:
1427         spin_unlock_irqrestore(&bitmap->lock, flags);
1428 }
1429
1430 void bitmap_close_sync(struct bitmap *bitmap)
1431 {
1432         /* Sync has finished, and any bitmap chunks that weren't synced
1433          * properly have been aborted.  It remains to us to clear the
1434          * RESYNC bit wherever it is still on
1435          */
1436         sector_t sector = 0;
1437         int blocks;
1438         if (!bitmap) return;
1439         while (sector < bitmap->mddev->resync_max_sectors) {
1440                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1441 /*
1442                 if (sector < 500) printk("bitmap_close_sync: sec %llu blks %d\n",
1443                                          (unsigned long long)sector, blocks);
1444 */              sector += blocks;
1445         }
1446 }
1447
1448 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1449 {
1450         /* For each chunk covered by any of these sectors, set the
1451          * counter to 1 and set resync_needed.  They should all
1452          * be 0 at this point
1453          */
1454
1455         int secs;
1456         bitmap_counter_t *bmc;
1457         spin_lock_irq(&bitmap->lock);
1458         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1459         if (!bmc) {
1460                 spin_unlock_irq(&bitmap->lock);
1461                 return;
1462         }
1463         if (! *bmc) {
1464                 struct page *page;
1465                 *bmc = 1 | (needed?NEEDED_MASK:0);
1466                 bitmap_count_page(bitmap, offset, 1);
1467                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1468                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1469         }
1470         spin_unlock_irq(&bitmap->lock);
1471
1472 }
1473
1474 /*
1475  * flush out any pending updates
1476  */
1477 void bitmap_flush(mddev_t *mddev)
1478 {
1479         struct bitmap *bitmap = mddev->bitmap;
1480         int sleep;
1481
1482         if (!bitmap) /* there was no bitmap */
1483                 return;
1484
1485         /* run the daemon_work three time to ensure everything is flushed
1486          * that can be
1487          */
1488         sleep = bitmap->daemon_sleep;
1489         bitmap->daemon_sleep = 0;
1490         bitmap_daemon_work(bitmap);
1491         bitmap_daemon_work(bitmap);
1492         bitmap_daemon_work(bitmap);
1493         bitmap->daemon_sleep = sleep;
1494         bitmap_update_sb(bitmap);
1495 }
1496
1497 /*
1498  * free memory that was allocated
1499  */
1500 static void bitmap_free(struct bitmap *bitmap)
1501 {
1502         unsigned long k, pages;
1503         struct bitmap_page *bp;
1504
1505         if (!bitmap) /* there was no bitmap */
1506                 return;
1507
1508         /* release the bitmap file and kill the daemon */
1509         bitmap_file_put(bitmap);
1510
1511         bp = bitmap->bp;
1512         pages = bitmap->pages;
1513
1514         /* free all allocated memory */
1515
1516         mempool_destroy(bitmap->write_pool);
1517
1518         if (bp) /* deallocate the page memory */
1519                 for (k = 0; k < pages; k++)
1520                         if (bp[k].map && !bp[k].hijacked)
1521                                 kfree(bp[k].map);
1522         kfree(bp);
1523         kfree(bitmap);
1524 }
1525 void bitmap_destroy(mddev_t *mddev)
1526 {
1527         struct bitmap *bitmap = mddev->bitmap;
1528
1529         if (!bitmap) /* there was no bitmap */
1530                 return;
1531
1532         mddev->bitmap = NULL; /* disconnect from the md device */
1533
1534         bitmap_free(bitmap);
1535 }
1536
1537 /*
1538  * initialize the bitmap structure
1539  * if this returns an error, bitmap_destroy must be called to do clean up
1540  */
1541 int bitmap_create(mddev_t *mddev)
1542 {
1543         struct bitmap *bitmap;
1544         unsigned long blocks = mddev->resync_max_sectors;
1545         unsigned long chunks;
1546         unsigned long pages;
1547         struct file *file = mddev->bitmap_file;
1548         int err;
1549         sector_t start;
1550
1551         BUG_ON(sizeof(bitmap_super_t) != 256);
1552
1553         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1554                 return 0;
1555
1556         BUG_ON(file && mddev->bitmap_offset);
1557
1558         bitmap = kmalloc(sizeof(*bitmap), GFP_KERNEL);
1559         if (!bitmap)
1560                 return -ENOMEM;
1561
1562         memset(bitmap, 0, sizeof(*bitmap));
1563
1564         spin_lock_init(&bitmap->lock);
1565         bitmap->mddev = mddev;
1566
1567         spin_lock_init(&bitmap->write_lock);
1568         INIT_LIST_HEAD(&bitmap->complete_pages);
1569         init_waitqueue_head(&bitmap->write_wait);
1570         bitmap->write_pool = mempool_create(WRITE_POOL_SIZE, write_pool_alloc,
1571                                 write_pool_free, NULL);
1572         err = -ENOMEM;
1573         if (!bitmap->write_pool)
1574                 goto error;
1575
1576         bitmap->file = file;
1577         bitmap->offset = mddev->bitmap_offset;
1578         if (file) get_file(file);
1579         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1580         err = bitmap_read_sb(bitmap);
1581         if (err)
1582                 goto error;
1583
1584         bitmap->chunkshift = find_first_bit(&bitmap->chunksize,
1585                                         sizeof(bitmap->chunksize));
1586
1587         /* now that chunksize and chunkshift are set, we can use these macros */
1588         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1589                         CHUNK_BLOCK_RATIO(bitmap);
1590         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1591
1592         BUG_ON(!pages);
1593
1594         bitmap->chunks = chunks;
1595         bitmap->pages = pages;
1596         bitmap->missing_pages = pages;
1597         bitmap->counter_bits = COUNTER_BITS;
1598
1599         bitmap->syncchunk = ~0UL;
1600
1601 #ifdef INJECT_FATAL_FAULT_1
1602         bitmap->bp = NULL;
1603 #else
1604         bitmap->bp = kmalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1605 #endif
1606         err = -ENOMEM;
1607         if (!bitmap->bp)
1608                 goto error;
1609         memset(bitmap->bp, 0, pages * sizeof(*bitmap->bp));
1610
1611         bitmap->flags |= BITMAP_ACTIVE;
1612
1613         /* now that we have some pages available, initialize the in-memory
1614          * bitmap from the on-disk bitmap */
1615         start = 0;
1616         if (mddev->degraded == 0
1617             || bitmap->events_cleared == mddev->events)
1618                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1619                 start = mddev->recovery_cp;
1620         err = bitmap_init_from_disk(bitmap, start);
1621
1622         if (err)
1623                 goto error;
1624
1625         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1626                 pages, bmname(bitmap));
1627
1628         mddev->bitmap = bitmap;
1629
1630         if (file)
1631                 /* kick off the bitmap writeback daemon */
1632                 bitmap->writeback_daemon =
1633                         bitmap_start_daemon(bitmap,
1634                                             bitmap_writeback_daemon,
1635                                             "bitmap_wb");
1636
1637         if (IS_ERR(bitmap->writeback_daemon))
1638                 return PTR_ERR(bitmap->writeback_daemon);
1639         return bitmap_update_sb(bitmap);
1640
1641  error:
1642         bitmap_free(bitmap);
1643         return err;
1644 }
1645
1646 /* the bitmap API -- for raid personalities */
1647 EXPORT_SYMBOL(bitmap_startwrite);
1648 EXPORT_SYMBOL(bitmap_endwrite);
1649 EXPORT_SYMBOL(bitmap_start_sync);
1650 EXPORT_SYMBOL(bitmap_end_sync);
1651 EXPORT_SYMBOL(bitmap_unplug);
1652 EXPORT_SYMBOL(bitmap_close_sync);
1653 EXPORT_SYMBOL(bitmap_daemon_work);