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  */
  11
  12 /*
  13  * Still to do:
  14  *
  15  * flush after percent set rather than just time based. (maybe both).
  16  */
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/errno.h>
  21 #include <linux/slab.h>
  22 #include <linux/init.h>
  23 #include <linux/timer.h>
  24 #include <linux/sched.h>
  25 #include <linux/list.h>
  26 #include <linux/file.h>
  27 #include <linux/mount.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/seq_file.h>
  30 #include "md.h"
  31 #include "bitmap.h"
  32
  33 static inline char *bmname(struct bitmap *bitmap)
  34 {
  35         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  36 }
  37
  38 /*
  39  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
  40  *
  41  * 1) check to see if this page is allocated, if it's not then try to alloc
  42  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
  43  *    page pointer directly as a counter
  44  *
  45  * if we find our page, we increment the page's refcount so that it stays
  46  * allocated while we're using it
  47  */
  48 static int bitmap_checkpage(struct bitmap_counts *bitmap,
  49                             unsigned long page, int create)
  50 __releases(bitmap->lock)
  51 __acquires(bitmap->lock)
  52 {
  53         unsigned char *mappage;
  54
  55         if (page >= bitmap->pages) {
  56                 /* This can happen if bitmap_start_sync goes beyond
  57                  * End-of-device while looking for a whole page.
  58                  * It is harmless.
  59                  */
  60                 return -EINVAL;
  61         }
  62
  63         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
  64                 return 0;
  65
  66         if (bitmap->bp[page].map) /* page is already allocated, just return */
  67                 return 0;
  68
  69         if (!create)
  70                 return -ENOENT;
  71
  72         /* this page has not been allocated yet */
  73
  74         spin_unlock_irq(&bitmap->lock);
  75         /* It is possible that this is being called inside a
  76          * prepare_to_wait/finish_wait loop from raid5c:make_request().
  77          * In general it is not permitted to sleep in that context as it
  78          * can cause the loop to spin freely.
  79          * That doesn't apply here as we can only reach this point
  80          * once with any loop.
  81          * When this function completes, either bp[page].map or
  82          * bp[page].hijacked.  In either case, this function will
  83          * abort before getting to this point again.  So there is
  84          * no risk of a free-spin, and so it is safe to assert
  85          * that sleeping here is allowed.
  86          */
  87         sched_annotate_sleep();
  88         mappage = kzalloc(PAGE_SIZE, GFP_NOIO);
  89         spin_lock_irq(&bitmap->lock);
  90
  91         if (mappage == NULL) {
  92                 pr_debug("md/bitmap: map page allocation failed, hijacking\n");
  93                 /* failed - set the hijacked flag so that we can use the
  94                  * pointer as a counter */
  95                 if (!bitmap->bp[page].map)
  96                         bitmap->bp[page].hijacked = 1;
  97         } else if (bitmap->bp[page].map ||
  98                    bitmap->bp[page].hijacked) {
  99                 /* somebody beat us to getting the page */
 100                 kfree(mappage);
 101                 return 0;
 102         } else {
 103
 104                 /* no page was in place and we have one, so install it */
 105
 106                 bitmap->bp[page].map = mappage;
 107                 bitmap->missing_pages--;
 108         }
 109         return 0;
 110 }
 111
 112 /* if page is completely empty, put it back on the free list, or dealloc it */
 113 /* if page was hijacked, unmark the flag so it might get alloced next time */
 114 /* Note: lock should be held when calling this */
 115 static void bitmap_checkfree(struct bitmap_counts *bitmap, unsigned long page)
 116 {
 117         char *ptr;
 118
 119         if (bitmap->bp[page].count) /* page is still busy */
 120                 return;
 121
 122         /* page is no longer in use, it can be released */
 123
 124         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 125                 bitmap->bp[page].hijacked = 0;
 126                 bitmap->bp[page].map = NULL;
 127         } else {
 128                 /* normal case, free the page */
 129                 ptr = bitmap->bp[page].map;
 130                 bitmap->bp[page].map = NULL;
 131                 bitmap->missing_pages++;
 132                 kfree(ptr);
 133         }
 134 }
 135
 136 /*
 137  * bitmap file handling - read and write the bitmap file and its superblock
 138  */
 139
 140 /*
 141  * basic page I/O operations
 142  */
 143
 144 /* IO operations when bitmap is stored near all superblocks */
 145 static int read_sb_page(struct mddev *mddev, loff_t offset,
 146                         struct page *page,
 147                         unsigned long index, int size)
 148 {
 149         /* choose a good rdev and read the page from there */
 150
 151         struct md_rdev *rdev;
 152         sector_t target;
 153
 154         rdev_for_each(rdev, mddev) {
 155                 if (! test_bit(In_sync, &rdev->flags)
 156                     || test_bit(Faulty, &rdev->flags))
 157                         continue;
 158
 159                 target = offset + index * (PAGE_SIZE/512);
 160
 161                 if (sync_page_io(rdev, target,
 162                                  roundup(size, bdev_logical_block_size(rdev->bdev)),
 163                                  page, READ, true)) {
 164                         page->index = index;
 165                         return 0;
 166                 }
 167         }
 168         return -EIO;
 169 }
 170
 171 static struct md_rdev *next_active_rdev(struct md_rdev *rdev, struct mddev *mddev)
 172 {
 173         /* Iterate the disks of an mddev, using rcu to protect access to the
 174          * linked list, and raising the refcount of devices we return to ensure
 175          * they don't disappear while in use.
 176          * As devices are only added or removed when raid_disk is < 0 and
 177          * nr_pending is 0 and In_sync is clear, the entries we return will
 178          * still be in the same position on the list when we re-enter
 179          * list_for_each_entry_continue_rcu.
 180          */
 181         rcu_read_lock();
 182         if (rdev == NULL)
 183                 /* start at the beginning */
 184                 rdev = list_entry_rcu(&mddev->disks, struct md_rdev, same_set);
 185         else {
 186                 /* release the previous rdev and start from there. */
 187                 rdev_dec_pending(rdev, mddev);
 188         }
 189         list_for_each_entry_continue_rcu(rdev, &mddev->disks, same_set) {
 190                 if (rdev->raid_disk >= 0 &&
 191                     !test_bit(Faulty, &rdev->flags)) {
 192                         /* this is a usable devices */
 193                         atomic_inc(&rdev->nr_pending);
 194                         rcu_read_unlock();
 195                         return rdev;
 196                 }
 197         }
 198         rcu_read_unlock();
 199         return NULL;
 200 }
 201
 202 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 203 {
 204         struct md_rdev *rdev = NULL;
 205         struct block_device *bdev;
 206         struct mddev *mddev = bitmap->mddev;
 207         struct bitmap_storage *store = &bitmap->storage;
 208         int node_offset = 0;
 209
 210         if (mddev_is_clustered(bitmap->mddev))
 211                 node_offset = bitmap->cluster_slot * store->file_pages;
 212
 213         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 214                 int size = PAGE_SIZE;
 215                 loff_t offset = mddev->bitmap_info.offset;
 216
 217                 bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
 218
 219                 if (page->index == store->file_pages-1) {
 220                         int last_page_size = store->bytes & (PAGE_SIZE-1);
 221                         if (last_page_size == 0)
 222                                 last_page_size = PAGE_SIZE;
 223                         size = roundup(last_page_size,
 224                                        bdev_logical_block_size(bdev));
 225                 }
 226                 /* Just make sure we aren't corrupting data or
 227                  * metadata
 228                  */
 229                 if (mddev->external) {
 230                         /* Bitmap could be anywhere. */
 231                         if (rdev->sb_start + offset + (page->index
 232                                                        * (PAGE_SIZE/512))
 233                             > rdev->data_offset
 234                             &&
 235                             rdev->sb_start + offset
 236                             < (rdev->data_offset + mddev->dev_sectors
 237                              + (PAGE_SIZE/512)))
 238                                 goto bad_alignment;
 239                 } else if (offset < 0) {
 240                         /* DATA  BITMAP METADATA  */
 241                         if (offset
 242                             + (long)(page->index * (PAGE_SIZE/512))
 243                             + size/512 > 0)
 244                                 /* bitmap runs in to metadata */
 245                                 goto bad_alignment;
 246                         if (rdev->data_offset + mddev->dev_sectors
 247                             > rdev->sb_start + offset)
 248                                 /* data runs in to bitmap */
 249                                 goto bad_alignment;
 250                 } else if (rdev->sb_start < rdev->data_offset) {
 251                         /* METADATA BITMAP DATA */
 252                         if (rdev->sb_start
 253                             + offset
 254                             + page->index*(PAGE_SIZE/512) + size/512
 255                             > rdev->data_offset)
 256                                 /* bitmap runs in to data */
 257                                 goto bad_alignment;
 258                 } else {
 259                         /* DATA METADATA BITMAP - no problems */
 260                 }
 261                 md_super_write(mddev, rdev,
 262                                rdev->sb_start + offset
 263                                + page->index * (PAGE_SIZE/512),
 264                                size,
 265                                page);
 266         }
 267
 268         if (wait)
 269                 md_super_wait(mddev);
 270         return 0;
 271
 272  bad_alignment:
 273         return -EINVAL;
 274 }
 275
 276 static void bitmap_file_kick(struct bitmap *bitmap);
 277 /*
 278  * write out a page to a file
 279  */
 280 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 281 {
 282         struct buffer_head *bh;
 283
 284         if (bitmap->storage.file == NULL) {
 285                 switch (write_sb_page(bitmap, page, wait)) {
 286                 case -EINVAL:
 287                         set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 288                 }
 289         } else {
 290
 291                 bh = page_buffers(page);
 292
 293                 while (bh && bh->b_blocknr) {
 294                         atomic_inc(&bitmap->pending_writes);
 295                         set_buffer_locked(bh);
 296                         set_buffer_mapped(bh);
 297                         submit_bh(WRITE | REQ_SYNC, bh);
 298                         bh = bh->b_this_page;
 299                 }
 300
 301                 if (wait)
 302                         wait_event(bitmap->write_wait,
 303                                    atomic_read(&bitmap->pending_writes)==0);
 304         }
 305         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 306                 bitmap_file_kick(bitmap);
 307 }
 308
 309 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 310 {
 311         struct bitmap *bitmap = bh->b_private;
 312
 313         if (!uptodate)
 314                 set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 315         if (atomic_dec_and_test(&bitmap->pending_writes))
 316                 wake_up(&bitmap->write_wait);
 317 }
 318
 319 /* copied from buffer.c */
 320 static void
 321 __clear_page_buffers(struct page *page)
 322 {
 323         ClearPagePrivate(page);
 324         set_page_private(page, 0);
 325         page_cache_release(page);
 326 }
 327 static void free_buffers(struct page *page)
 328 {
 329         struct buffer_head *bh;
 330
 331         if (!PagePrivate(page))
 332                 return;
 333
 334         bh = page_buffers(page);
 335         while (bh) {
 336                 struct buffer_head *next = bh->b_this_page;
 337                 free_buffer_head(bh);
 338                 bh = next;
 339         }
 340         __clear_page_buffers(page);
 341         put_page(page);
 342 }
 343
 344 /* read a page from a file.
 345  * We both read the page, and attach buffers to the page to record the
 346  * address of each block (using bmap).  These addresses will be used
 347  * to write the block later, completely bypassing the filesystem.
 348  * This usage is similar to how swap files are handled, and allows us
 349  * to write to a file with no concerns of memory allocation failing.
 350  */
 351 static int read_page(struct file *file, unsigned long index,
 352                      struct bitmap *bitmap,
 353                      unsigned long count,
 354                      struct page *page)
 355 {
 356         int ret = 0;
 357         struct inode *inode = file_inode(file);
 358         struct buffer_head *bh;
 359         sector_t block;
 360
 361         pr_debug("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 362                  (unsigned long long)index << PAGE_SHIFT);
 363
 364         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 365         if (!bh) {
 366                 ret = -ENOMEM;
 367                 goto out;
 368         }
 369         attach_page_buffers(page, bh);
 370         block = index << (PAGE_SHIFT - inode->i_blkbits);
 371         while (bh) {
 372                 if (count == 0)
 373                         bh->b_blocknr = 0;
 374                 else {
 375                         bh->b_blocknr = bmap(inode, block);
 376                         if (bh->b_blocknr == 0) {
 377                                 /* Cannot use this file! */
 378                                 ret = -EINVAL;
 379                                 goto out;
 380                         }
 381                         bh->b_bdev = inode->i_sb->s_bdev;
 382                         if (count < (1<<inode->i_blkbits))
 383                                 count = 0;
 384                         else
 385                                 count -= (1<<inode->i_blkbits);
 386
 387                         bh->b_end_io = end_bitmap_write;
 388                         bh->b_private = bitmap;
 389                         atomic_inc(&bitmap->pending_writes);
 390                         set_buffer_locked(bh);
 391                         set_buffer_mapped(bh);
 392                         submit_bh(READ, bh);
 393                 }
 394                 block++;
 395                 bh = bh->b_this_page;
 396         }
 397         page->index = index;
 398
 399         wait_event(bitmap->write_wait,
 400                    atomic_read(&bitmap->pending_writes)==0);
 401         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 402                 ret = -EIO;
 403 out:
 404         if (ret)
 405                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %d\n",
 406                         (int)PAGE_SIZE,
 407                         (unsigned long long)index << PAGE_SHIFT,
 408                         ret);
 409         return ret;
 410 }
 411
 412 /*
 413  * bitmap file superblock operations
 414  */
 415
 416 /* update the event counter and sync the superblock to disk */
 417 void bitmap_update_sb(struct bitmap *bitmap)
 418 {
 419         bitmap_super_t *sb;
 420
 421         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 422                 return;
 423         if (bitmap->mddev->bitmap_info.external)
 424                 return;
 425         if (!bitmap->storage.sb_page) /* no superblock */
 426                 return;
 427         sb = kmap_atomic(bitmap->storage.sb_page);
 428         sb->events = cpu_to_le64(bitmap->mddev->events);
 429         if (bitmap->mddev->events < bitmap->events_cleared)
 430                 /* rocking back to read-only */
 431                 bitmap->events_cleared = bitmap->mddev->events;
 432         sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 433         sb->state = cpu_to_le32(bitmap->flags);
 434         /* Just in case these have been changed via sysfs: */
 435         sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 436         sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
 437         /* This might have been changed by a reshape */
 438         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 439         sb->chunksize = cpu_to_le32(bitmap->mddev->bitmap_info.chunksize);
 440         sb->nodes = cpu_to_le32(bitmap->mddev->bitmap_info.nodes);
 441         sb->sectors_reserved = cpu_to_le32(bitmap->mddev->
 442                                            bitmap_info.space);
 443         kunmap_atomic(sb);
 444         write_page(bitmap, bitmap->storage.sb_page, 1);
 445 }
 446
 447 /* print out the bitmap file superblock */
 448 void bitmap_print_sb(struct bitmap *bitmap)
 449 {
 450         bitmap_super_t *sb;
 451
 452         if (!bitmap || !bitmap->storage.sb_page)
 453                 return;
 454         sb = kmap_atomic(bitmap->storage.sb_page);
 455         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 456         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 457         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 458         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 459                                         *(__u32 *)(sb->uuid+0),
 460                                         *(__u32 *)(sb->uuid+4),
 461                                         *(__u32 *)(sb->uuid+8),
 462                                         *(__u32 *)(sb->uuid+12));
 463         printk(KERN_DEBUG "        events: %llu\n",
 464                         (unsigned long long) le64_to_cpu(sb->events));
 465         printk(KERN_DEBUG "events cleared: %llu\n",
 466                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 467         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 468         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 469         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 470         printk(KERN_DEBUG "     sync size: %llu KB\n",
 471                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 472         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 473         kunmap_atomic(sb);
 474 }
 475
 476 /*
 477  * bitmap_new_disk_sb
 478  * @bitmap
 479  *
 480  * This function is somewhat the reverse of bitmap_read_sb.  bitmap_read_sb
 481  * reads and verifies the on-disk bitmap superblock and populates bitmap_info.
 482  * This function verifies 'bitmap_info' and populates the on-disk bitmap
 483  * structure, which is to be written to disk.
 484  *
 485  * Returns: 0 on success, -Exxx on error
 486  */
 487 static int bitmap_new_disk_sb(struct bitmap *bitmap)
 488 {
 489         bitmap_super_t *sb;
 490         unsigned long chunksize, daemon_sleep, write_behind;
 491
 492         bitmap->storage.sb_page = alloc_page(GFP_KERNEL);
 493         if (bitmap->storage.sb_page == NULL)
 494                 return -ENOMEM;
 495         bitmap->storage.sb_page->index = 0;
 496
 497         sb = kmap_atomic(bitmap->storage.sb_page);
 498
 499         sb->magic = cpu_to_le32(BITMAP_MAGIC);
 500         sb->version = cpu_to_le32(BITMAP_MAJOR_HI);
 501
 502         chunksize = bitmap->mddev->bitmap_info.chunksize;
 503         BUG_ON(!chunksize);
 504         if (!is_power_of_2(chunksize)) {
 505                 kunmap_atomic(sb);
 506                 printk(KERN_ERR "bitmap chunksize not a power of 2\n");
 507                 return -EINVAL;
 508         }
 509         sb->chunksize = cpu_to_le32(chunksize);
 510
 511         daemon_sleep = bitmap->mddev->bitmap_info.daemon_sleep;
 512         if (!daemon_sleep ||
 513             (daemon_sleep < 1) || (daemon_sleep > MAX_SCHEDULE_TIMEOUT)) {
 514                 printk(KERN_INFO "Choosing daemon_sleep default (5 sec)\n");
 515                 daemon_sleep = 5 * HZ;
 516         }
 517         sb->daemon_sleep = cpu_to_le32(daemon_sleep);
 518         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 519
 520         /*
 521          * FIXME: write_behind for RAID1.  If not specified, what
 522          * is a good choice?  We choose COUNTER_MAX / 2 arbitrarily.
 523          */
 524         write_behind = bitmap->mddev->bitmap_info.max_write_behind;
 525         if (write_behind > COUNTER_MAX)
 526                 write_behind = COUNTER_MAX / 2;
 527         sb->write_behind = cpu_to_le32(write_behind);
 528         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 529
 530         /* keep the array size field of the bitmap superblock up to date */
 531         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 532
 533         memcpy(sb->uuid, bitmap->mddev->uuid, 16);
 534
 535         set_bit(BITMAP_STALE, &bitmap->flags);
 536         sb->state = cpu_to_le32(bitmap->flags);
 537         bitmap->events_cleared = bitmap->mddev->events;
 538         sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 539
 540         kunmap_atomic(sb);
 541
 542         return 0;
 543 }
 544
 545 /* read the superblock from the bitmap file and initialize some bitmap fields */
 546 static int bitmap_read_sb(struct bitmap *bitmap)
 547 {
 548         char *reason = NULL;
 549         bitmap_super_t *sb;
 550         unsigned long chunksize, daemon_sleep, write_behind;
 551         unsigned long long events;
 552         int nodes = 0;
 553         unsigned long sectors_reserved = 0;
 554         int err = -EINVAL;
 555         struct page *sb_page;
 556
 557         if (!bitmap->storage.file && !bitmap->mddev->bitmap_info.offset) {
 558                 chunksize = 128 * 1024 * 1024;
 559                 daemon_sleep = 5 * HZ;
 560                 write_behind = 0;
 561                 set_bit(BITMAP_STALE, &bitmap->flags);
 562                 err = 0;
 563                 goto out_no_sb;
 564         }
 565         /* page 0 is the superblock, read it... */
 566         sb_page = alloc_page(GFP_KERNEL);
 567         if (!sb_page)
 568                 return -ENOMEM;
 569         bitmap->storage.sb_page = sb_page;
 570
 571 re_read:
 572         /* If cluster_slot is set, the cluster is setup */
 573         if (bitmap->cluster_slot >= 0) {
 574                 sector_t bm_blocks = bitmap->mddev->resync_max_sectors;
 575
 576                 sector_div(bm_blocks,
 577                            bitmap->mddev->bitmap_info.chunksize >> 9);
 578                 bm_blocks = bm_blocks << 3;
 579                 bm_blocks = DIV_ROUND_UP_SECTOR_T(bm_blocks, 4096);
 580                 bitmap->mddev->bitmap_info.offset += bitmap->cluster_slot * (bm_blocks << 3);
 581                 pr_info("%s:%d bm slot: %d offset: %llu\n", __func__, __LINE__,
 582                         bitmap->cluster_slot, (unsigned long long)bitmap->mddev->bitmap_info.offset);
 583         }
 584
 585         if (bitmap->storage.file) {
 586                 loff_t isize = i_size_read(bitmap->storage.file->f_mapping->host);
 587                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 588
 589                 err = read_page(bitmap->storage.file, 0,
 590                                 bitmap, bytes, sb_page);
 591         } else {
 592                 err = read_sb_page(bitmap->mddev,
 593                                    bitmap->mddev->bitmap_info.offset,
 594                                    sb_page,
 595                                    0, sizeof(bitmap_super_t));
 596         }
 597         if (err)
 598                 return err;
 599
 600         err = -EINVAL;
 601         sb = kmap_atomic(sb_page);
 602
 603         chunksize = le32_to_cpu(sb->chunksize);
 604         daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 605         write_behind = le32_to_cpu(sb->write_behind);
 606         sectors_reserved = le32_to_cpu(sb->sectors_reserved);
 607         nodes = le32_to_cpu(sb->nodes);
 608         strlcpy(bitmap->mddev->bitmap_info.cluster_name, sb->cluster_name, 64);
 609
 610         /* verify that the bitmap-specific fields are valid */
 611         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 612                 reason = "bad magic";
 613         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 614                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 615                 reason = "unrecognized superblock version";
 616         else if (chunksize < 512)
 617                 reason = "bitmap chunksize too small";
 618         else if (!is_power_of_2(chunksize))
 619                 reason = "bitmap chunksize not a power of 2";
 620         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT)
 621                 reason = "daemon sleep period out of range";
 622         else if (write_behind > COUNTER_MAX)
 623                 reason = "write-behind limit out of range (0 - 16383)";
 624         if (reason) {
 625                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 626                         bmname(bitmap), reason);
 627                 goto out;
 628         }
 629
 630         /* keep the array size field of the bitmap superblock up to date */
 631         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 632
 633         if (bitmap->mddev->persistent) {
 634                 /*
 635                  * We have a persistent array superblock, so compare the
 636                  * bitmap's UUID and event counter to the mddev's
 637                  */
 638                 if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 639                         printk(KERN_INFO
 640                                "%s: bitmap superblock UUID mismatch\n",
 641                                bmname(bitmap));
 642                         goto out;
 643                 }
 644                 events = le64_to_cpu(sb->events);
 645                 if (!nodes && (events < bitmap->mddev->events)) {
 646                         printk(KERN_INFO
 647                                "%s: bitmap file is out of date (%llu < %llu) "
 648                                "-- forcing full recovery\n",
 649                                bmname(bitmap), events,
 650                                (unsigned long long) bitmap->mddev->events);
 651                         set_bit(BITMAP_STALE, &bitmap->flags);
 652                 }
 653         }
 654
 655         /* assign fields using values from superblock */
 656         bitmap->flags |= le32_to_cpu(sb->state);
 657         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 658                 set_bit(BITMAP_HOSTENDIAN, &bitmap->flags);
 659         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 660         strlcpy(bitmap->mddev->bitmap_info.cluster_name, sb->cluster_name, 64);
 661         err = 0;
 662
 663 out:
 664         kunmap_atomic(sb);
 665         /* Assiging chunksize is required for "re_read" */
 666         bitmap->mddev->bitmap_info.chunksize = chunksize;
 667         if (nodes && (bitmap->cluster_slot < 0)) {
 668                 err = md_setup_cluster(bitmap->mddev, nodes);
 669                 if (err) {
 670                         pr_err("%s: Could not setup cluster service (%d)\n",
 671                                         bmname(bitmap), err);
 672                         goto out_no_sb;
 673                 }
 674                 bitmap->cluster_slot = md_cluster_ops->slot_number(bitmap->mddev);
 675                 pr_info("%s:%d bm slot: %d offset: %llu\n", __func__, __LINE__,
 676                         bitmap->cluster_slot,
 677                         (unsigned long long)bitmap->mddev->bitmap_info.offset);
 678                 goto re_read;
 679         }
 680
 681
 682 out_no_sb:
 683         if (test_bit(BITMAP_STALE, &bitmap->flags))
 684                 bitmap->events_cleared = bitmap->mddev->events;
 685         bitmap->mddev->bitmap_info.chunksize = chunksize;
 686         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 687         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 688         bitmap->mddev->bitmap_info.nodes = nodes;
 689         if (bitmap->mddev->bitmap_info.space == 0 ||
 690             bitmap->mddev->bitmap_info.space > sectors_reserved)
 691                 bitmap->mddev->bitmap_info.space = sectors_reserved;
 692         if (err) {
 693                 bitmap_print_sb(bitmap);
 694                 if (bitmap->cluster_slot < 0)
 695                         md_cluster_stop(bitmap->mddev);
 696         }
 697         return err;
 698 }
 699
 700 /*
 701  * general bitmap file operations
 702  */
 703
 704 /*
 705  * on-disk bitmap:
 706  *
 707  * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
 708  * file a page at a time. There's a superblock at the start of the file.
 709  */
 710 /* calculate the index of the page that contains this bit */
 711 static inline unsigned long file_page_index(struct bitmap_storage *store,
 712                                             unsigned long chunk)
 713 {
 714         if (store->sb_page)
 715                 chunk += sizeof(bitmap_super_t) << 3;
 716         return chunk >> PAGE_BIT_SHIFT;
 717 }
 718
 719 /* calculate the (bit) offset of this bit within a page */
 720 static inline unsigned long file_page_offset(struct bitmap_storage *store,
 721                                              unsigned long chunk)
 722 {
 723         if (store->sb_page)
 724                 chunk += sizeof(bitmap_super_t) << 3;
 725         return chunk & (PAGE_BITS - 1);
 726 }
 727
 728 /*
 729  * return a pointer to the page in the filemap that contains the given bit
 730  *
 731  */
 732 static inline struct page *filemap_get_page(struct bitmap_storage *store,
 733                                             unsigned long chunk)
 734 {
 735         if (file_page_index(store, chunk) >= store->file_pages)
 736                 return NULL;
 737         return store->filemap[file_page_index(store, chunk)];
 738 }
 739
 740 static int bitmap_storage_alloc(struct bitmap_storage *store,
 741                                 unsigned long chunks, int with_super,
 742                                 int slot_number)
 743 {
 744         int pnum, offset = 0;
 745         unsigned long num_pages;
 746         unsigned long bytes;
 747
 748         bytes = DIV_ROUND_UP(chunks, 8);
 749         if (with_super)
 750                 bytes += sizeof(bitmap_super_t);
 751
 752         num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
 753         offset = slot_number * (num_pages - 1);
 754
 755         store->filemap = kmalloc(sizeof(struct page *)
 756                                  * num_pages, GFP_KERNEL);
 757         if (!store->filemap)
 758                 return -ENOMEM;
 759
 760         if (with_super && !store->sb_page) {
 761                 store->sb_page = alloc_page(GFP_KERNEL|__GFP_ZERO);
 762                 if (store->sb_page == NULL)
 763                         return -ENOMEM;
 764         }
 765
 766         pnum = 0;
 767         if (store->sb_page) {
 768                 store->filemap[0] = store->sb_page;
 769                 pnum = 1;
 770                 store->sb_page->index = offset;
 771         }
 772
 773         for ( ; pnum < num_pages; pnum++) {
 774                 store->filemap[pnum] = alloc_page(GFP_KERNEL|__GFP_ZERO);
 775                 if (!store->filemap[pnum]) {
 776                         store->file_pages = pnum;
 777                         return -ENOMEM;
 778                 }
 779                 store->filemap[pnum]->index = pnum + offset;
 780         }
 781         store->file_pages = pnum;
 782
 783         /* We need 4 bits per page, rounded up to a multiple
 784          * of sizeof(unsigned long) */
 785         store->filemap_attr = kzalloc(
 786                 roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
 787                 GFP_KERNEL);
 788         if (!store->filemap_attr)
 789                 return -ENOMEM;
 790
 791         store->bytes = bytes;
 792
 793         return 0;
 794 }
 795
 796 static void bitmap_file_unmap(struct bitmap_storage *store)
 797 {
 798         struct page **map, *sb_page;
 799         int pages;
 800         struct file *file;
 801
 802         file = store->file;
 803         map = store->filemap;
 804         pages = store->file_pages;
 805         sb_page = store->sb_page;
 806
 807         while (pages--)
 808                 if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 809                         free_buffers(map[pages]);
 810         kfree(map);
 811         kfree(store->filemap_attr);
 812
 813         if (sb_page)
 814                 free_buffers(sb_page);
 815
 816         if (file) {
 817                 struct inode *inode = file_inode(file);
 818                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 819                 fput(file);
 820         }
 821 }
 822
 823 /*
 824  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 825  * then it is no longer reliable, so we stop using it and we mark the file
 826  * as failed in the superblock
 827  */
 828 static void bitmap_file_kick(struct bitmap *bitmap)
 829 {
 830         char *path, *ptr = NULL;
 831
 832         if (!test_and_set_bit(BITMAP_STALE, &bitmap->flags)) {
 833                 bitmap_update_sb(bitmap);
 834
 835                 if (bitmap->storage.file) {
 836                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 837                         if (path)
 838                                 ptr = d_path(&bitmap->storage.file->f_path,
 839                                              path, PAGE_SIZE);
 840
 841                         printk(KERN_ALERT
 842                               "%s: kicking failed bitmap file %s from array!\n",
 843                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 844
 845                         kfree(path);
 846                 } else
 847                         printk(KERN_ALERT
 848                                "%s: disabling internal bitmap due to errors\n",
 849                                bmname(bitmap));
 850         }
 851 }
 852
 853 enum bitmap_page_attr {
 854         BITMAP_PAGE_DIRTY = 0,     /* there are set bits that need to be synced */
 855         BITMAP_PAGE_PENDING = 1,   /* there are bits that are being cleaned.
 856                                     * i.e. counter is 1 or 2. */
 857         BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 858 };
 859
 860 static inline void set_page_attr(struct bitmap *bitmap, int pnum,
 861                                  enum bitmap_page_attr attr)
 862 {
 863         set_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 864 }
 865
 866 static inline void clear_page_attr(struct bitmap *bitmap, int pnum,
 867                                    enum bitmap_page_attr attr)
 868 {
 869         clear_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 870 }
 871
 872 static inline int test_page_attr(struct bitmap *bitmap, int pnum,
 873                                  enum bitmap_page_attr attr)
 874 {
 875         return test_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 876 }
 877
 878 static inline int test_and_clear_page_attr(struct bitmap *bitmap, int pnum,
 879                                            enum bitmap_page_attr attr)
 880 {
 881         return test_and_clear_bit((pnum<<2) + attr,
 882                                   bitmap->storage.filemap_attr);
 883 }
 884 /*
 885  * bitmap_file_set_bit -- called before performing a write to the md device
 886  * to set (and eventually sync) a particular bit in the bitmap file
 887  *
 888  * we set the bit immediately, then we record the page number so that
 889  * when an unplug occurs, we can flush the dirty pages out to disk
 890  */
 891 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 892 {
 893         unsigned long bit;
 894         struct page *page;
 895         void *kaddr;
 896         unsigned long chunk = block >> bitmap->counts.chunkshift;
 897
 898         page = filemap_get_page(&bitmap->storage, chunk);
 899         if (!page)
 900                 return;
 901         bit = file_page_offset(&bitmap->storage, chunk);
 902
 903         /* set the bit */
 904         kaddr = kmap_atomic(page);
 905         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 906                 set_bit(bit, kaddr);
 907         else
 908                 set_bit_le(bit, kaddr);
 909         kunmap_atomic(kaddr);
 910         pr_debug("set file bit %lu page %lu\n", bit, page->index);
 911         /* record page number so it gets flushed to disk when unplug occurs */
 912         set_page_attr(bitmap, page->index, BITMAP_PAGE_DIRTY);
 913 }
 914
 915 static void bitmap_file_clear_bit(struct bitmap *bitmap, sector_t block)
 916 {
 917         unsigned long bit;
 918         struct page *page;
 919         void *paddr;
 920         unsigned long chunk = block >> bitmap->counts.chunkshift;
 921
 922         page = filemap_get_page(&bitmap->storage, chunk);
 923         if (!page)
 924                 return;
 925         bit = file_page_offset(&bitmap->storage, chunk);
 926         paddr = kmap_atomic(page);
 927         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 928                 clear_bit(bit, paddr);
 929         else
 930                 clear_bit_le(bit, paddr);
 931         kunmap_atomic(paddr);
 932         if (!test_page_attr(bitmap, page->index, BITMAP_PAGE_NEEDWRITE)) {
 933                 set_page_attr(bitmap, page->index, BITMAP_PAGE_PENDING);
 934                 bitmap->allclean = 0;
 935         }
 936 }
 937
 938 static int bitmap_file_test_bit(struct bitmap *bitmap, sector_t block)
 939 {
 940         unsigned long bit;
 941         struct page *page;
 942         void *paddr;
 943         unsigned long chunk = block >> bitmap->counts.chunkshift;
 944         int set = 0;
 945
 946         page = filemap_get_page(&bitmap->storage, chunk);
 947         if (!page)
 948                 return -EINVAL;
 949         bit = file_page_offset(&bitmap->storage, chunk);
 950         paddr = kmap_atomic(page);
 951         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 952                 set = test_bit(bit, paddr);
 953         else
 954                 set = test_bit_le(bit, paddr);
 955         kunmap_atomic(paddr);
 956         return set;
 957 }
 958
 959
 960 /* this gets called when the md device is ready to unplug its underlying
 961  * (slave) device queues -- before we let any writes go down, we need to
 962  * sync the dirty pages of the bitmap file to disk */
 963 void bitmap_unplug(struct bitmap *bitmap)
 964 {
 965         unsigned long i;
 966         int dirty, need_write;
 967
 968         if (!bitmap || !bitmap->storage.filemap ||
 969             test_bit(BITMAP_STALE, &bitmap->flags))
 970                 return;
 971
 972         /* look at each page to see if there are any set bits that need to be
 973          * flushed out to disk */
 974         for (i = 0; i < bitmap->storage.file_pages; i++) {
 975                 if (!bitmap->storage.filemap)
 976                         return;
 977                 dirty = test_and_clear_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
 978                 need_write = test_and_clear_page_attr(bitmap, i,
 979                                                       BITMAP_PAGE_NEEDWRITE);
 980                 if (dirty || need_write) {
 981                         clear_page_attr(bitmap, i, BITMAP_PAGE_PENDING);
 982                         write_page(bitmap, bitmap->storage.filemap[i], 0);
 983                 }
 984         }
 985         if (bitmap->storage.file)
 986                 wait_event(bitmap->write_wait,
 987                            atomic_read(&bitmap->pending_writes)==0);
 988         else
 989                 md_super_wait(bitmap->mddev);
 990
 991         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 992                 bitmap_file_kick(bitmap);
 993 }
 994 EXPORT_SYMBOL(bitmap_unplug);
 995
 996 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 997 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 998  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 999  * memory mapping of the bitmap file
1000  * Special cases:
1001  *   if there's no bitmap file, or if the bitmap file had been
1002  *   previously kicked from the array, we mark all the bits as
1003  *   1's in order to cause a full resync.
1004  *
1005  * We ignore all bits for sectors that end earlier than 'start'.
1006  * This is used when reading an out-of-date bitmap...
1007  */
1008 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
1009 {
1010         unsigned long i, chunks, index, oldindex, bit, node_offset = 0;
1011         struct page *page = NULL;
1012         unsigned long bit_cnt = 0;
1013         struct file *file;
1014         unsigned long offset;
1015         int outofdate;
1016         int ret = -ENOSPC;
1017         void *paddr;
1018         struct bitmap_storage *store = &bitmap->storage;
1019
1020         chunks = bitmap->counts.chunks;
1021         file = store->file;
1022
1023         if (!file && !bitmap->mddev->bitmap_info.offset) {
1024                 /* No permanent bitmap - fill with '1s'. */
1025                 store->filemap = NULL;
1026                 store->file_pages = 0;
1027                 for (i = 0; i < chunks ; i++) {
1028                         /* if the disk bit is set, set the memory bit */
1029                         int needed = ((sector_t)(i+1) << (bitmap->counts.chunkshift)
1030                                       >= start);
1031                         bitmap_set_memory_bits(bitmap,
1032                                                (sector_t)i << bitmap->counts.chunkshift,
1033                                                needed);
1034                 }
1035                 return 0;
1036         }
1037
1038         outofdate = test_bit(BITMAP_STALE, &bitmap->flags);
1039         if (outofdate)
1040                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
1041                         "recovery\n", bmname(bitmap));
1042
1043         if (file && i_size_read(file->f_mapping->host) < store->bytes) {
1044                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
1045                        bmname(bitmap),
1046                        (unsigned long) i_size_read(file->f_mapping->host),
1047                        store->bytes);
1048                 goto err;
1049         }
1050
1051         oldindex = ~0L;
1052         offset = 0;
1053         if (!bitmap->mddev->bitmap_info.external)
1054                 offset = sizeof(bitmap_super_t);
1055
1056         if (mddev_is_clustered(bitmap->mddev))
1057                 node_offset = bitmap->cluster_slot * (DIV_ROUND_UP(store->bytes, PAGE_SIZE));
1058
1059         for (i = 0; i < chunks; i++) {
1060                 int b;
1061                 index = file_page_index(&bitmap->storage, i);
1062                 bit = file_page_offset(&bitmap->storage, i);
1063                 if (index != oldindex) { /* this is a new page, read it in */
1064                         int count;
1065                         /* unmap the old page, we're done with it */
1066                         if (index == store->file_pages-1)
1067                                 count = store->bytes - index * PAGE_SIZE;
1068                         else
1069                                 count = PAGE_SIZE;
1070                         page = store->filemap[index];
1071                         if (file)
1072                                 ret = read_page(file, index, bitmap,
1073                                                 count, page);
1074                         else
1075                                 ret = read_sb_page(
1076                                         bitmap->mddev,
1077                                         bitmap->mddev->bitmap_info.offset,
1078                                         page,
1079                                         index + node_offset, count);
1080
1081                         if (ret)
1082                                 goto err;
1083
1084                         oldindex = index;
1085
1086                         if (outofdate) {
1087                                 /*
1088                                  * if bitmap is out of date, dirty the
1089                                  * whole page and write it out
1090                                  */
1091                                 paddr = kmap_atomic(page);
1092                                 memset(paddr + offset, 0xff,
1093                                        PAGE_SIZE - offset);
1094                                 kunmap_atomic(paddr);
1095                                 write_page(bitmap, page, 1);
1096
1097                                 ret = -EIO;
1098                                 if (test_bit(BITMAP_WRITE_ERROR,
1099                                              &bitmap->flags))
1100                                         goto err;
1101                         }
1102                 }
1103                 paddr = kmap_atomic(page);
1104                 if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
1105                         b = test_bit(bit, paddr);
1106                 else
1107                         b = test_bit_le(bit, paddr);
1108                 kunmap_atomic(paddr);
1109                 if (b) {
1110                         /* if the disk bit is set, set the memory bit */
1111                         int needed = ((sector_t)(i+1) << bitmap->counts.chunkshift
1112                                       >= start);
1113                         bitmap_set_memory_bits(bitmap,
1114                                                (sector_t)i << bitmap->counts.chunkshift,
1115                                                needed);
1116                         bit_cnt++;
1117                 }
1118                 offset = 0;
1119         }
1120
1121         printk(KERN_INFO "%s: bitmap initialized from disk: "
1122                "read %lu pages, set %lu of %lu bits\n",
1123                bmname(bitmap), store->file_pages,
1124                bit_cnt, chunks);
1125
1126         return 0;
1127
1128  err:
1129         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1130                bmname(bitmap), ret);
1131         return ret;
1132 }
1133
1134 void bitmap_write_all(struct bitmap *bitmap)
1135 {
1136         /* We don't actually write all bitmap blocks here,
1137          * just flag them as needing to be written
1138          */
1139         int i;
1140
1141         if (!bitmap || !bitmap->storage.filemap)
1142                 return;
1143         if (bitmap->storage.file)
1144                 /* Only one copy, so nothing needed */
1145                 return;
1146
1147         for (i = 0; i < bitmap->storage.file_pages; i++)
1148                 set_page_attr(bitmap, i,
1149                               BITMAP_PAGE_NEEDWRITE);
1150         bitmap->allclean = 0;
1151 }
1152
1153 static void bitmap_count_page(struct bitmap_counts *bitmap,
1154                               sector_t offset, int inc)
1155 {
1156         sector_t chunk = offset >> bitmap->chunkshift;
1157         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1158         bitmap->bp[page].count += inc;
1159         bitmap_checkfree(bitmap, page);
1160 }
1161
1162 static void bitmap_set_pending(struct bitmap_counts *bitmap, sector_t offset)
1163 {
1164         sector_t chunk = offset >> bitmap->chunkshift;
1165         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1166         struct bitmap_page *bp = &bitmap->bp[page];
1167
1168         if (!bp->pending)
1169                 bp->pending = 1;
1170 }
1171
1172 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1173                                             sector_t offset, sector_t *blocks,
1174                                             int create);
1175
1176 /*
1177  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1178  *                      out to disk
1179  */
1180
1181 void bitmap_daemon_work(struct mddev *mddev)
1182 {
1183         struct bitmap *bitmap;
1184         unsigned long j;
1185         unsigned long nextpage;
1186         sector_t blocks;
1187         struct bitmap_counts *counts;
1188
1189         /* Use a mutex to guard daemon_work against
1190          * bitmap_destroy.
1191          */
1192         mutex_lock(&mddev->bitmap_info.mutex);
1193         bitmap = mddev->bitmap;
1194         if (bitmap == NULL) {
1195                 mutex_unlock(&mddev->bitmap_info.mutex);
1196                 return;
1197         }
1198         if (time_before(jiffies, bitmap->daemon_lastrun
1199                         + mddev->bitmap_info.daemon_sleep))
1200                 goto done;
1201
1202         bitmap->daemon_lastrun = jiffies;
1203         if (bitmap->allclean) {
1204                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1205                 goto done;
1206         }
1207         bitmap->allclean = 1;
1208
1209         /* Any file-page which is PENDING now needs to be written.
1210          * So set NEEDWRITE now, then after we make any last-minute changes
1211          * we will write it.
1212          */
1213         for (j = 0; j < bitmap->storage.file_pages; j++)
1214                 if (test_and_clear_page_attr(bitmap, j,
1215                                              BITMAP_PAGE_PENDING))
1216                         set_page_attr(bitmap, j,
1217                                       BITMAP_PAGE_NEEDWRITE);
1218
1219         if (bitmap->need_sync &&
1220             mddev->bitmap_info.external == 0) {
1221                 /* Arrange for superblock update as well as
1222                  * other changes */
1223                 bitmap_super_t *sb;
1224                 bitmap->need_sync = 0;
1225                 if (bitmap->storage.filemap) {
1226                         sb = kmap_atomic(bitmap->storage.sb_page);
1227                         sb->events_cleared =
1228                                 cpu_to_le64(bitmap->events_cleared);
1229                         kunmap_atomic(sb);
1230                         set_page_attr(bitmap, 0,
1231                                       BITMAP_PAGE_NEEDWRITE);
1232                 }
1233         }
1234         /* Now look at the bitmap counters and if any are '2' or '1',
1235          * decrement and handle accordingly.
1236          */
1237         counts = &bitmap->counts;
1238         spin_lock_irq(&counts->lock);
1239         nextpage = 0;
1240         for (j = 0; j < counts->chunks; j++) {
1241                 bitmap_counter_t *bmc;
1242                 sector_t  block = (sector_t)j << counts->chunkshift;
1243
1244                 if (j == nextpage) {
1245                         nextpage += PAGE_COUNTER_RATIO;
1246                         if (!counts->bp[j >> PAGE_COUNTER_SHIFT].pending) {
1247                                 j |= PAGE_COUNTER_MASK;
1248                                 continue;
1249                         }
1250                         counts->bp[j >> PAGE_COUNTER_SHIFT].pending = 0;
1251                 }
1252                 bmc = bitmap_get_counter(counts,
1253                                          block,
1254                                          &blocks, 0);
1255
1256                 if (!bmc) {
1257                         j |= PAGE_COUNTER_MASK;
1258                         continue;
1259                 }
1260                 if (*bmc == 1 && !bitmap->need_sync) {
1261                         /* We can clear the bit */
1262                         *bmc = 0;
1263                         bitmap_count_page(counts, block, -1);
1264                         bitmap_file_clear_bit(bitmap, block);
1265                 } else if (*bmc && *bmc <= 2) {
1266                         *bmc = 1;
1267                         bitmap_set_pending(counts, block);
1268                         bitmap->allclean = 0;
1269                 }
1270         }
1271         spin_unlock_irq(&counts->lock);
1272
1273         /* Now start writeout on any page in NEEDWRITE that isn't DIRTY.
1274          * DIRTY pages need to be written by bitmap_unplug so it can wait
1275          * for them.
1276          * If we find any DIRTY page we stop there and let bitmap_unplug
1277          * handle all the rest.  This is important in the case where
1278          * the first blocking holds the superblock and it has been updated.
1279          * We mustn't write any other blocks before the superblock.
1280          */
1281         for (j = 0;
1282              j < bitmap->storage.file_pages
1283                      && !test_bit(BITMAP_STALE, &bitmap->flags);
1284              j++) {
1285                 if (test_page_attr(bitmap, j,
1286                                    BITMAP_PAGE_DIRTY))
1287                         /* bitmap_unplug will handle the rest */
1288                         break;
1289                 if (test_and_clear_page_attr(bitmap, j,
1290                                              BITMAP_PAGE_NEEDWRITE)) {
1291                         write_page(bitmap, bitmap->storage.filemap[j], 0);
1292                 }
1293         }
1294
1295  done:
1296         if (bitmap->allclean == 0)
1297                 mddev->thread->timeout =
1298                         mddev->bitmap_info.daemon_sleep;
1299         mutex_unlock(&mddev->bitmap_info.mutex);
1300 }
1301
1302 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1303                                             sector_t offset, sector_t *blocks,
1304                                             int create)
1305 __releases(bitmap->lock)
1306 __acquires(bitmap->lock)
1307 {
1308         /* If 'create', we might release the lock and reclaim it.
1309          * The lock must have been taken with interrupts enabled.
1310          * If !create, we don't release the lock.
1311          */
1312         sector_t chunk = offset >> bitmap->chunkshift;
1313         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1314         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1315         sector_t csize;
1316         int err;
1317
1318         err = bitmap_checkpage(bitmap, page, create);
1319
1320         if (bitmap->bp[page].hijacked ||
1321             bitmap->bp[page].map == NULL)
1322                 csize = ((sector_t)1) << (bitmap->chunkshift +
1323                                           PAGE_COUNTER_SHIFT - 1);
1324         else
1325                 csize = ((sector_t)1) << bitmap->chunkshift;
1326         *blocks = csize - (offset & (csize - 1));
1327
1328         if (err < 0)
1329                 return NULL;
1330
1331         /* now locked ... */
1332
1333         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1334                 /* should we use the first or second counter field
1335                  * of the hijacked pointer? */
1336                 int hi = (pageoff > PAGE_COUNTER_MASK);
1337                 return  &((bitmap_counter_t *)
1338                           &bitmap->bp[page].map)[hi];
1339         } else /* page is allocated */
1340                 return (bitmap_counter_t *)
1341                         &(bitmap->bp[page].map[pageoff]);
1342 }
1343
1344 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1345 {
1346         if (!bitmap)
1347                 return 0;
1348
1349         if (behind) {
1350                 int bw;
1351                 atomic_inc(&bitmap->behind_writes);
1352                 bw = atomic_read(&bitmap->behind_writes);
1353                 if (bw > bitmap->behind_writes_used)
1354                         bitmap->behind_writes_used = bw;
1355
1356                 pr_debug("inc write-behind count %d/%lu\n",
1357                          bw, bitmap->mddev->bitmap_info.max_write_behind);
1358         }
1359
1360         while (sectors) {
1361                 sector_t blocks;
1362                 bitmap_counter_t *bmc;
1363
1364                 spin_lock_irq(&bitmap->counts.lock);
1365                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 1);
1366                 if (!bmc) {
1367                         spin_unlock_irq(&bitmap->counts.lock);
1368                         return 0;
1369                 }
1370
1371                 if (unlikely(COUNTER(*bmc) == COUNTER_MAX)) {
1372                         DEFINE_WAIT(__wait);
1373                         /* note that it is safe to do the prepare_to_wait
1374                          * after the test as long as we do it before dropping
1375                          * the spinlock.
1376                          */
1377                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1378                                         TASK_UNINTERRUPTIBLE);
1379                         spin_unlock_irq(&bitmap->counts.lock);
1380                         schedule();
1381                         finish_wait(&bitmap->overflow_wait, &__wait);
1382                         continue;
1383                 }
1384
1385                 switch (*bmc) {
1386                 case 0:
1387                         bitmap_file_set_bit(bitmap, offset);
1388                         bitmap_count_page(&bitmap->counts, offset, 1);
1389                         /* fall through */
1390                 case 1:
1391                         *bmc = 2;
1392                 }
1393
1394                 (*bmc)++;
1395
1396                 spin_unlock_irq(&bitmap->counts.lock);
1397
1398                 offset += blocks;
1399                 if (sectors > blocks)
1400                         sectors -= blocks;
1401                 else
1402                         sectors = 0;
1403         }
1404         return 0;
1405 }
1406 EXPORT_SYMBOL(bitmap_startwrite);
1407
1408 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1409                      int success, int behind)
1410 {
1411         if (!bitmap)
1412                 return;
1413         if (behind) {
1414                 if (atomic_dec_and_test(&bitmap->behind_writes))
1415                         wake_up(&bitmap->behind_wait);
1416                 pr_debug("dec write-behind count %d/%lu\n",
1417                          atomic_read(&bitmap->behind_writes),
1418                          bitmap->mddev->bitmap_info.max_write_behind);
1419         }
1420
1421         while (sectors) {
1422                 sector_t blocks;
1423                 unsigned long flags;
1424                 bitmap_counter_t *bmc;
1425
1426                 spin_lock_irqsave(&bitmap->counts.lock, flags);
1427                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 0);
1428                 if (!bmc) {
1429                         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1430                         return;
1431                 }
1432
1433                 if (success && !bitmap->mddev->degraded &&
1434                     bitmap->events_cleared < bitmap->mddev->events) {
1435                         bitmap->events_cleared = bitmap->mddev->events;
1436                         bitmap->need_sync = 1;
1437                         sysfs_notify_dirent_safe(bitmap->sysfs_can_clear);
1438                 }
1439
1440                 if (!success && !NEEDED(*bmc))
1441                         *bmc |= NEEDED_MASK;
1442
1443                 if (COUNTER(*bmc) == COUNTER_MAX)
1444                         wake_up(&bitmap->overflow_wait);
1445
1446                 (*bmc)--;
1447                 if (*bmc <= 2) {
1448                         bitmap_set_pending(&bitmap->counts, offset);
1449                         bitmap->allclean = 0;
1450                 }
1451                 spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1452                 offset += blocks;
1453                 if (sectors > blocks)
1454                         sectors -= blocks;
1455                 else
1456                         sectors = 0;
1457         }
1458 }
1459 EXPORT_SYMBOL(bitmap_endwrite);
1460
1461 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1462                                int degraded)
1463 {
1464         bitmap_counter_t *bmc;
1465         int rv;
1466         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1467                 *blocks = 1024;
1468                 return 1; /* always resync if no bitmap */
1469         }
1470         spin_lock_irq(&bitmap->counts.lock);
1471         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1472         rv = 0;
1473         if (bmc) {
1474                 /* locked */
1475                 if (RESYNC(*bmc))
1476                         rv = 1;
1477                 else if (NEEDED(*bmc)) {
1478                         rv = 1;
1479                         if (!degraded) { /* don't set/clear bits if degraded */
1480                                 *bmc |= RESYNC_MASK;
1481                                 *bmc &= ~NEEDED_MASK;
1482                         }
1483                 }
1484         }
1485         spin_unlock_irq(&bitmap->counts.lock);
1486         return rv;
1487 }
1488
1489 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1490                       int degraded)
1491 {
1492         /* bitmap_start_sync must always report on multiples of whole
1493          * pages, otherwise resync (which is very PAGE_SIZE based) will
1494          * get confused.
1495          * So call __bitmap_start_sync repeatedly (if needed) until
1496          * At least PAGE_SIZE>>9 blocks are covered.
1497          * Return the 'or' of the result.
1498          */
1499         int rv = 0;
1500         sector_t blocks1;
1501
1502         *blocks = 0;
1503         while (*blocks < (PAGE_SIZE>>9)) {
1504                 rv |= __bitmap_start_sync(bitmap, offset,
1505                                           &blocks1, degraded);
1506                 offset += blocks1;
1507                 *blocks += blocks1;
1508         }
1509         return rv;
1510 }
1511 EXPORT_SYMBOL(bitmap_start_sync);
1512
1513 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks, int aborted)
1514 {
1515         bitmap_counter_t *bmc;
1516         unsigned long flags;
1517
1518         if (bitmap == NULL) {
1519                 *blocks = 1024;
1520                 return;
1521         }
1522         spin_lock_irqsave(&bitmap->counts.lock, flags);
1523         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1524         if (bmc == NULL)
1525                 goto unlock;
1526         /* locked */
1527         if (RESYNC(*bmc)) {
1528                 *bmc &= ~RESYNC_MASK;
1529
1530                 if (!NEEDED(*bmc) && aborted)
1531                         *bmc |= NEEDED_MASK;
1532                 else {
1533                         if (*bmc <= 2) {
1534                                 bitmap_set_pending(&bitmap->counts, offset);
1535                                 bitmap->allclean = 0;
1536                         }
1537                 }
1538         }
1539  unlock:
1540         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1541 }
1542 EXPORT_SYMBOL(bitmap_end_sync);
1543
1544 void bitmap_close_sync(struct bitmap *bitmap)
1545 {
1546         /* Sync has finished, and any bitmap chunks that weren't synced
1547          * properly have been aborted.  It remains to us to clear the
1548          * RESYNC bit wherever it is still on
1549          */
1550         sector_t sector = 0;
1551         sector_t blocks;
1552         if (!bitmap)
1553                 return;
1554         while (sector < bitmap->mddev->resync_max_sectors) {
1555                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1556                 sector += blocks;
1557         }
1558 }
1559 EXPORT_SYMBOL(bitmap_close_sync);
1560
1561 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1562 {
1563         sector_t s = 0;
1564         sector_t blocks;
1565
1566         if (!bitmap)
1567                 return;
1568         if (sector == 0) {
1569                 bitmap->last_end_sync = jiffies;
1570                 return;
1571         }
1572         if (time_before(jiffies, (bitmap->last_end_sync
1573                                   + bitmap->mddev->bitmap_info.daemon_sleep)))
1574                 return;
1575         wait_event(bitmap->mddev->recovery_wait,
1576                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1577
1578         bitmap->mddev->curr_resync_completed = sector;
1579         set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
1580         sector &= ~((1ULL << bitmap->counts.chunkshift) - 1);
1581         s = 0;
1582         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1583                 bitmap_end_sync(bitmap, s, &blocks, 0);
1584                 s += blocks;
1585         }
1586         bitmap->last_end_sync = jiffies;
1587         sysfs_notify(&bitmap->mddev->kobj, NULL, "sync_completed");
1588 }
1589 EXPORT_SYMBOL(bitmap_cond_end_sync);
1590
1591 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1592 {
1593         /* For each chunk covered by any of these sectors, set the
1594          * counter to 2 and possibly set resync_needed.  They should all
1595          * be 0 at this point
1596          */
1597
1598         sector_t secs;
1599         bitmap_counter_t *bmc;
1600         spin_lock_irq(&bitmap->counts.lock);
1601         bmc = bitmap_get_counter(&bitmap->counts, offset, &secs, 1);
1602         if (!bmc) {
1603                 spin_unlock_irq(&bitmap->counts.lock);
1604                 return;
1605         }
1606         if (!*bmc) {
1607                 *bmc = 2;
1608                 bitmap_count_page(&bitmap->counts, offset, 1);
1609                 bitmap_set_pending(&bitmap->counts, offset);
1610                 bitmap->allclean = 0;
1611         }
1612         if (needed)
1613                 *bmc |= NEEDED_MASK;
1614         spin_unlock_irq(&bitmap->counts.lock);
1615 }
1616
1617 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1618 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1619 {
1620         unsigned long chunk;
1621
1622         for (chunk = s; chunk <= e; chunk++) {
1623                 sector_t sec = (sector_t)chunk << bitmap->counts.chunkshift;
1624                 bitmap_set_memory_bits(bitmap, sec, 1);
1625                 bitmap_file_set_bit(bitmap, sec);
1626                 if (sec < bitmap->mddev->recovery_cp)
1627                         /* We are asserting that the array is dirty,
1628                          * so move the recovery_cp address back so
1629                          * that it is obvious that it is dirty
1630                          */
1631                         bitmap->mddev->recovery_cp = sec;
1632         }
1633 }
1634
1635 /*
1636  * flush out any pending updates
1637  */
1638 void bitmap_flush(struct mddev *mddev)
1639 {
1640         struct bitmap *bitmap = mddev->bitmap;
1641         long sleep;
1642
1643         if (!bitmap) /* there was no bitmap */
1644                 return;
1645
1646         /* run the daemon_work three time to ensure everything is flushed
1647          * that can be
1648          */
1649         sleep = mddev->bitmap_info.daemon_sleep * 2;
1650         bitmap->daemon_lastrun -= sleep;
1651         bitmap_daemon_work(mddev);
1652         bitmap->daemon_lastrun -= sleep;
1653         bitmap_daemon_work(mddev);
1654         bitmap->daemon_lastrun -= sleep;
1655         bitmap_daemon_work(mddev);
1656         bitmap_update_sb(bitmap);
1657 }
1658
1659 /*
1660  * free memory that was allocated
1661  */
1662 static void bitmap_free(struct bitmap *bitmap)
1663 {
1664         unsigned long k, pages;
1665         struct bitmap_page *bp;
1666
1667         if (!bitmap) /* there was no bitmap */
1668                 return;
1669
1670         if (mddev_is_clustered(bitmap->mddev) && bitmap->mddev->cluster_info &&
1671                 bitmap->cluster_slot == md_cluster_ops->slot_number(bitmap->mddev))
1672                 md_cluster_stop(bitmap->mddev);
1673
1674         /* Shouldn't be needed - but just in case.... */
1675         wait_event(bitmap->write_wait,
1676                    atomic_read(&bitmap->pending_writes) == 0);
1677
1678         /* release the bitmap file  */
1679         bitmap_file_unmap(&bitmap->storage);
1680
1681         bp = bitmap->counts.bp;
1682         pages = bitmap->counts.pages;
1683
1684         /* free all allocated memory */
1685
1686         if (bp) /* deallocate the page memory */
1687                 for (k = 0; k < pages; k++)
1688                         if (bp[k].map && !bp[k].hijacked)
1689                                 kfree(bp[k].map);
1690         kfree(bp);
1691         kfree(bitmap);
1692 }
1693
1694 void bitmap_destroy(struct mddev *mddev)
1695 {
1696         struct bitmap *bitmap = mddev->bitmap;
1697
1698         if (!bitmap) /* there was no bitmap */
1699                 return;
1700
1701         mutex_lock(&mddev->bitmap_info.mutex);
1702         spin_lock(&mddev->lock);
1703         mddev->bitmap = NULL; /* disconnect from the md device */
1704         spin_unlock(&mddev->lock);
1705         mutex_unlock(&mddev->bitmap_info.mutex);
1706         if (mddev->thread)
1707                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1708
1709         if (bitmap->sysfs_can_clear)
1710                 sysfs_put(bitmap->sysfs_can_clear);
1711
1712         bitmap_free(bitmap);
1713 }
1714
1715 /*
1716  * initialize the bitmap structure
1717  * if this returns an error, bitmap_destroy must be called to do clean up
1718  */
1719 struct bitmap *bitmap_create(struct mddev *mddev, int slot)
1720 {
1721         struct bitmap *bitmap;
1722         sector_t blocks = mddev->resync_max_sectors;
1723         struct file *file = mddev->bitmap_info.file;
1724         int err;
1725         struct kernfs_node *bm = NULL;
1726
1727         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1728
1729         BUG_ON(file && mddev->bitmap_info.offset);
1730
1731         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1732         if (!bitmap)
1733                 return ERR_PTR(-ENOMEM);
1734
1735         spin_lock_init(&bitmap->counts.lock);
1736         atomic_set(&bitmap->pending_writes, 0);
1737         init_waitqueue_head(&bitmap->write_wait);
1738         init_waitqueue_head(&bitmap->overflow_wait);
1739         init_waitqueue_head(&bitmap->behind_wait);
1740
1741         bitmap->mddev = mddev;
1742         bitmap->cluster_slot = slot;
1743
1744         if (mddev->kobj.sd)
1745                 bm = sysfs_get_dirent(mddev->kobj.sd, "bitmap");
1746         if (bm) {
1747                 bitmap->sysfs_can_clear = sysfs_get_dirent(bm, "can_clear");
1748                 sysfs_put(bm);
1749         } else
1750                 bitmap->sysfs_can_clear = NULL;
1751
1752         bitmap->storage.file = file;
1753         if (file) {
1754                 get_file(file);
1755                 /* As future accesses to this file will use bmap,
1756                  * and bypass the page cache, we must sync the file
1757                  * first.
1758                  */
1759                 vfs_fsync(file, 1);
1760         }
1761         /* read superblock from bitmap file (this sets mddev->bitmap_info.chunksize) */
1762         if (!mddev->bitmap_info.external) {
1763                 /*
1764                  * If 'MD_ARRAY_FIRST_USE' is set, then device-mapper is
1765                  * instructing us to create a new on-disk bitmap instance.
1766                  */
1767                 if (test_and_clear_bit(MD_ARRAY_FIRST_USE, &mddev->flags))
1768                         err = bitmap_new_disk_sb(bitmap);
1769                 else
1770                         err = bitmap_read_sb(bitmap);
1771         } else {
1772                 err = 0;
1773                 if (mddev->bitmap_info.chunksize == 0 ||
1774                     mddev->bitmap_info.daemon_sleep == 0)
1775                         /* chunksize and time_base need to be
1776                          * set first. */
1777                         err = -EINVAL;
1778         }
1779         if (err)
1780                 goto error;
1781
1782         bitmap->daemon_lastrun = jiffies;
1783         err = bitmap_resize(bitmap, blocks, mddev->bitmap_info.chunksize, 1);
1784         if (err)
1785                 goto error;
1786
1787         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1788                bitmap->counts.pages, bmname(bitmap));
1789
1790         err = test_bit(BITMAP_WRITE_ERROR, &bitmap->flags) ? -EIO : 0;
1791         if (err)
1792                 goto error;
1793
1794         return bitmap;
1795  error:
1796         bitmap_free(bitmap);
1797         return ERR_PTR(err);
1798 }
1799
1800 int bitmap_load(struct mddev *mddev)
1801 {
1802         int err = 0;
1803         sector_t start = 0;
1804         sector_t sector = 0;
1805         struct bitmap *bitmap = mddev->bitmap;
1806
1807         if (!bitmap)
1808                 goto out;
1809
1810         /* Clear out old bitmap info first:  Either there is none, or we
1811          * are resuming after someone else has possibly changed things,
1812          * so we should forget old cached info.
1813          * All chunks should be clean, but some might need_sync.
1814          */
1815         while (sector < mddev->resync_max_sectors) {
1816                 sector_t blocks;
1817                 bitmap_start_sync(bitmap, sector, &blocks, 0);
1818                 sector += blocks;
1819         }
1820         bitmap_close_sync(bitmap);
1821
1822         if (mddev->degraded == 0
1823             || bitmap->events_cleared == mddev->events)
1824                 /* no need to keep dirty bits to optimise a
1825                  * re-add of a missing device */
1826                 start = mddev->recovery_cp;
1827
1828         mutex_lock(&mddev->bitmap_info.mutex);
1829         err = bitmap_init_from_disk(bitmap, start);
1830         mutex_unlock(&mddev->bitmap_info.mutex);
1831
1832         if (err)
1833                 goto out;
1834         clear_bit(BITMAP_STALE, &bitmap->flags);
1835
1836         /* Kick recovery in case any bits were set */
1837         set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1838
1839         mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
1840         md_wakeup_thread(mddev->thread);
1841
1842         bitmap_update_sb(bitmap);
1843
1844         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
1845                 err = -EIO;
1846 out:
1847         return err;
1848 }
1849 EXPORT_SYMBOL_GPL(bitmap_load);
1850
1851 /* Loads the bitmap associated with slot and copies the resync information
1852  * to our bitmap
1853  */
1854 int bitmap_copy_from_slot(struct mddev *mddev, int slot,
1855                 sector_t *low, sector_t *high)
1856 {
1857         int rv = 0, i, j;
1858         sector_t block, lo = 0, hi = 0;
1859         struct bitmap_counts *counts;
1860         struct bitmap *bitmap = bitmap_create(mddev, slot);
1861
1862         if (IS_ERR(bitmap))
1863                 return PTR_ERR(bitmap);
1864
1865         rv = bitmap_read_sb(bitmap);
1866         if (rv)
1867                 goto err;
1868
1869         rv = bitmap_init_from_disk(bitmap, 0);
1870         if (rv)
1871                 goto err;
1872
1873         counts = &bitmap->counts;
1874         for (j = 0; j < counts->chunks; j++) {
1875                 block = (sector_t)j << counts->chunkshift;
1876                 if (bitmap_file_test_bit(bitmap, block)) {
1877                         if (!lo)
1878                                 lo = block;
1879                         hi = block;
1880                         bitmap_file_clear_bit(bitmap, block);
1881                         bitmap_set_memory_bits(mddev->bitmap, block, 1);
1882                         bitmap_file_set_bit(mddev->bitmap, block);
1883                 }
1884         }
1885
1886         bitmap_update_sb(bitmap);
1887         /* Setting this for the ev_page should be enough.
1888          * And we do not require both write_all and PAGE_DIRT either
1889          */
1890         for (i = 0; i < bitmap->storage.file_pages; i++)
1891                 set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
1892         bitmap_write_all(bitmap);
1893         bitmap_unplug(bitmap);
1894         *low = lo;
1895         *high = hi;
1896 err:
1897         bitmap_free(bitmap);
1898         return rv;
1899 }
1900 EXPORT_SYMBOL_GPL(bitmap_copy_from_slot);
1901
1902
1903 void bitmap_status(struct seq_file *seq, struct bitmap *bitmap)
1904 {
1905         unsigned long chunk_kb;
1906         struct bitmap_counts *counts;
1907
1908         if (!bitmap)
1909                 return;
1910
1911         counts = &bitmap->counts;
1912
1913         chunk_kb = bitmap->mddev->bitmap_info.chunksize >> 10;
1914         seq_printf(seq, "bitmap: %lu/%lu pages [%luKB], "
1915                    "%lu%s chunk",
1916                    counts->pages - counts->missing_pages,
1917                    counts->pages,
1918                    (counts->pages - counts->missing_pages)
1919                    << (PAGE_SHIFT - 10),
1920                    chunk_kb ? chunk_kb : bitmap->mddev->bitmap_info.chunksize,
1921                    chunk_kb ? "KB" : "B");
1922         if (bitmap->storage.file) {
1923                 seq_printf(seq, ", file: ");
1924                 seq_path(seq, &bitmap->storage.file->f_path, " \t\n");
1925         }
1926
1927         seq_printf(seq, "\n");
1928 }
1929
1930 int bitmap_resize(struct bitmap *bitmap, sector_t blocks,
1931                   int chunksize, int init)
1932 {
1933         /* If chunk_size is 0, choose an appropriate chunk size.
1934          * Then possibly allocate new storage space.
1935          * Then quiesce, copy bits, replace bitmap, and re-start
1936          *
1937          * This function is called both to set up the initial bitmap
1938          * and to resize the bitmap while the array is active.
1939          * If this happens as a result of the array being resized,
1940          * chunksize will be zero, and we need to choose a suitable
1941          * chunksize, otherwise we use what we are given.
1942          */
1943         struct bitmap_storage store;
1944         struct bitmap_counts old_counts;
1945         unsigned long chunks;
1946         sector_t block;
1947         sector_t old_blocks, new_blocks;
1948         int chunkshift;
1949         int ret = 0;
1950         long pages;
1951         struct bitmap_page *new_bp;
1952
1953         if (chunksize == 0) {
1954                 /* If there is enough space, leave the chunk size unchanged,
1955                  * else increase by factor of two until there is enough space.
1956                  */
1957                 long bytes;
1958                 long space = bitmap->mddev->bitmap_info.space;
1959
1960                 if (space == 0) {
1961                         /* We don't know how much space there is, so limit
1962                          * to current size - in sectors.
1963                          */
1964                         bytes = DIV_ROUND_UP(bitmap->counts.chunks, 8);
1965                         if (!bitmap->mddev->bitmap_info.external)
1966                                 bytes += sizeof(bitmap_super_t);
1967                         space = DIV_ROUND_UP(bytes, 512);
1968                         bitmap->mddev->bitmap_info.space = space;
1969                 }
1970                 chunkshift = bitmap->counts.chunkshift;
1971                 chunkshift--;
1972                 do {
1973                         /* 'chunkshift' is shift from block size to chunk size */
1974                         chunkshift++;
1975                         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1976                         bytes = DIV_ROUND_UP(chunks, 8);
1977                         if (!bitmap->mddev->bitmap_info.external)
1978                                 bytes += sizeof(bitmap_super_t);
1979                 } while (bytes > (space << 9));
1980         } else
1981                 chunkshift = ffz(~chunksize) - BITMAP_BLOCK_SHIFT;
1982
1983         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1984         memset(&store, 0, sizeof(store));
1985         if (bitmap->mddev->bitmap_info.offset || bitmap->mddev->bitmap_info.file)
1986                 ret = bitmap_storage_alloc(&store, chunks,
1987                                            !bitmap->mddev->bitmap_info.external,
1988                                            bitmap->cluster_slot);
1989         if (ret)
1990                 goto err;
1991
1992         pages = DIV_ROUND_UP(chunks, PAGE_COUNTER_RATIO);
1993
1994         new_bp = kzalloc(pages * sizeof(*new_bp), GFP_KERNEL);
1995         ret = -ENOMEM;
1996         if (!new_bp) {
1997                 bitmap_file_unmap(&store);
1998                 goto err;
1999         }
2000
2001         if (!init)
2002                 bitmap->mddev->pers->quiesce(bitmap->mddev, 1);
2003
2004         store.file = bitmap->storage.file;
2005         bitmap->storage.file = NULL;
2006
2007         if (store.sb_page && bitmap->storage.sb_page)
2008                 memcpy(page_address(store.sb_page),
2009                        page_address(bitmap->storage.sb_page),
2010                        sizeof(bitmap_super_t));
2011         bitmap_file_unmap(&bitmap->storage);
2012         bitmap->storage = store;
2013
2014         old_counts = bitmap->counts;
2015         bitmap->counts.bp = new_bp;
2016         bitmap->counts.pages = pages;
2017         bitmap->counts.missing_pages = pages;
2018         bitmap->counts.chunkshift = chunkshift;
2019         bitmap->counts.chunks = chunks;
2020         bitmap->mddev->bitmap_info.chunksize = 1 << (chunkshift +
2021                                                      BITMAP_BLOCK_SHIFT);
2022
2023         blocks = min(old_counts.chunks << old_counts.chunkshift,
2024                      chunks << chunkshift);
2025
2026         spin_lock_irq(&bitmap->counts.lock);
2027         for (block = 0; block < blocks; ) {
2028                 bitmap_counter_t *bmc_old, *bmc_new;
2029                 int set;
2030
2031                 bmc_old = bitmap_get_counter(&old_counts, block,
2032                                              &old_blocks, 0);
2033                 set = bmc_old && NEEDED(*bmc_old);
2034
2035                 if (set) {
2036                         bmc_new = bitmap_get_counter(&bitmap->counts, block,
2037                                                      &new_blocks, 1);
2038                         if (*bmc_new == 0) {
2039                                 /* need to set on-disk bits too. */
2040                                 sector_t end = block + new_blocks;
2041                                 sector_t start = block >> chunkshift;
2042                                 start <<= chunkshift;
2043                                 while (start < end) {
2044                                         bitmap_file_set_bit(bitmap, block);
2045                                         start += 1 << chunkshift;
2046                                 }
2047                                 *bmc_new = 2;
2048                                 bitmap_count_page(&bitmap->counts,
2049                                                   block, 1);
2050                                 bitmap_set_pending(&bitmap->counts,
2051                                                    block);
2052                         }
2053                         *bmc_new |= NEEDED_MASK;
2054                         if (new_blocks < old_blocks)
2055                                 old_blocks = new_blocks;
2056                 }
2057                 block += old_blocks;
2058         }
2059
2060         if (!init) {
2061                 int i;
2062                 while (block < (chunks << chunkshift)) {
2063                         bitmap_counter_t *bmc;
2064                         bmc = bitmap_get_counter(&bitmap->counts, block,
2065                                                  &new_blocks, 1);
2066                         if (bmc) {
2067                                 /* new space.  It needs to be resynced, so
2068                                  * we set NEEDED_MASK.
2069                                  */
2070                                 if (*bmc == 0) {
2071                                         *bmc = NEEDED_MASK | 2;
2072                                         bitmap_count_page(&bitmap->counts,
2073                                                           block, 1);
2074                                         bitmap_set_pending(&bitmap->counts,
2075                                                            block);
2076                                 }
2077                         }
2078                         block += new_blocks;
2079                 }
2080                 for (i = 0; i < bitmap->storage.file_pages; i++)
2081                         set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
2082         }
2083         spin_unlock_irq(&bitmap->counts.lock);
2084
2085         if (!init) {
2086                 bitmap_unplug(bitmap);
2087                 bitmap->mddev->pers->quiesce(bitmap->mddev, 0);
2088         }
2089         ret = 0;
2090 err:
2091         return ret;
2092 }
2093 EXPORT_SYMBOL_GPL(bitmap_resize);
2094
2095 static ssize_t
2096 location_show(struct mddev *mddev, char *page)
2097 {
2098         ssize_t len;
2099         if (mddev->bitmap_info.file)
2100                 len = sprintf(page, "file");
2101         else if (mddev->bitmap_info.offset)
2102                 len = sprintf(page, "%+lld", (long long)mddev->bitmap_info.offset);
2103         else
2104                 len = sprintf(page, "none");
2105         len += sprintf(page+len, "\n");
2106         return len;
2107 }
2108
2109 static ssize_t
2110 location_store(struct mddev *mddev, const char *buf, size_t len)
2111 {
2112
2113         if (mddev->pers) {
2114                 if (!mddev->pers->quiesce)
2115                         return -EBUSY;
2116                 if (mddev->recovery || mddev->sync_thread)
2117                         return -EBUSY;
2118         }
2119
2120         if (mddev->bitmap || mddev->bitmap_info.file ||
2121             mddev->bitmap_info.offset) {
2122                 /* bitmap already configured.  Only option is to clear it */
2123                 if (strncmp(buf, "none", 4) != 0)
2124                         return -EBUSY;
2125                 if (mddev->pers) {
2126                         mddev->pers->quiesce(mddev, 1);
2127                         bitmap_destroy(mddev);
2128                         mddev->pers->quiesce(mddev, 0);
2129                 }
2130                 mddev->bitmap_info.offset = 0;
2131                 if (mddev->bitmap_info.file) {
2132                         struct file *f = mddev->bitmap_info.file;
2133                         mddev->bitmap_info.file = NULL;
2134                         fput(f);
2135                 }
2136         } else {
2137                 /* No bitmap, OK to set a location */
2138                 long long offset;
2139                 if (strncmp(buf, "none", 4) == 0)
2140                         /* nothing to be done */;
2141                 else if (strncmp(buf, "file:", 5) == 0) {
2142                         /* Not supported yet */
2143                         return -EINVAL;
2144                 } else {
2145                         int rv;
2146                         if (buf[0] == '+')
2147                                 rv = kstrtoll(buf+1, 10, &offset);
2148                         else
2149                                 rv = kstrtoll(buf, 10, &offset);
2150                         if (rv)
2151                                 return rv;
2152                         if (offset == 0)
2153                                 return -EINVAL;
2154                         if (mddev->bitmap_info.external == 0 &&
2155                             mddev->major_version == 0 &&
2156                             offset != mddev->bitmap_info.default_offset)
2157                                 return -EINVAL;
2158                         mddev->bitmap_info.offset = offset;
2159                         if (mddev->pers) {
2160                                 struct bitmap *bitmap;
2161                                 mddev->pers->quiesce(mddev, 1);
2162                                 bitmap = bitmap_create(mddev, -1);
2163                                 if (IS_ERR(bitmap))
2164                                         rv = PTR_ERR(bitmap);
2165                                 else {
2166                                         mddev->bitmap = bitmap;
2167                                         rv = bitmap_load(mddev);
2168                                         if (rv) {
2169                                                 bitmap_destroy(mddev);
2170                                                 mddev->bitmap_info.offset = 0;
2171                                         }
2172                                 }
2173                                 mddev->pers->quiesce(mddev, 0);
2174                                 if (rv)
2175                                         return rv;
2176                         }
2177                 }
2178         }
2179         if (!mddev->external) {
2180                 /* Ensure new bitmap info is stored in
2181                  * metadata promptly.
2182                  */
2183                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2184                 md_wakeup_thread(mddev->thread);
2185         }
2186         return len;
2187 }
2188
2189 static struct md_sysfs_entry bitmap_location =
2190 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
2191
2192 /* 'bitmap/space' is the space available at 'location' for the
2193  * bitmap.  This allows the kernel to know when it is safe to
2194  * resize the bitmap to match a resized array.
2195  */
2196 static ssize_t
2197 space_show(struct mddev *mddev, char *page)
2198 {
2199         return sprintf(page, "%lu\n", mddev->bitmap_info.space);
2200 }
2201
2202 static ssize_t
2203 space_store(struct mddev *mddev, const char *buf, size_t len)
2204 {
2205         unsigned long sectors;
2206         int rv;
2207
2208         rv = kstrtoul(buf, 10, &sectors);
2209         if (rv)
2210                 return rv;
2211
2212         if (sectors == 0)
2213                 return -EINVAL;
2214
2215         if (mddev->bitmap &&
2216             sectors < (mddev->bitmap->storage.bytes + 511) >> 9)
2217                 return -EFBIG; /* Bitmap is too big for this small space */
2218
2219         /* could make sure it isn't too big, but that isn't really
2220          * needed - user-space should be careful.
2221          */
2222         mddev->bitmap_info.space = sectors;
2223         return len;
2224 }
2225
2226 static struct md_sysfs_entry bitmap_space =
2227 __ATTR(space, S_IRUGO|S_IWUSR, space_show, space_store);
2228
2229 static ssize_t
2230 timeout_show(struct mddev *mddev, char *page)
2231 {
2232         ssize_t len;
2233         unsigned long secs = mddev->bitmap_info.daemon_sleep / HZ;
2234         unsigned long jifs = mddev->bitmap_info.daemon_sleep % HZ;
2235
2236         len = sprintf(page, "%lu", secs);
2237         if (jifs)
2238                 len += sprintf(page+len, ".%03u", jiffies_to_msecs(jifs));
2239         len += sprintf(page+len, "\n");
2240         return len;
2241 }
2242
2243 static ssize_t
2244 timeout_store(struct mddev *mddev, const char *buf, size_t len)
2245 {
2246         /* timeout can be set at any time */
2247         unsigned long timeout;
2248         int rv = strict_strtoul_scaled(buf, &timeout, 4);
2249         if (rv)
2250                 return rv;
2251
2252         /* just to make sure we don't overflow... */
2253         if (timeout >= LONG_MAX / HZ)
2254                 return -EINVAL;
2255
2256         timeout = timeout * HZ / 10000;
2257
2258         if (timeout >= MAX_SCHEDULE_TIMEOUT)
2259                 timeout = MAX_SCHEDULE_TIMEOUT-1;
2260         if (timeout < 1)
2261                 timeout = 1;
2262         mddev->bitmap_info.daemon_sleep = timeout;
2263         if (mddev->thread) {
2264                 /* if thread->timeout is MAX_SCHEDULE_TIMEOUT, then
2265                  * the bitmap is all clean and we don't need to
2266                  * adjust the timeout right now
2267                  */
2268                 if (mddev->thread->timeout < MAX_SCHEDULE_TIMEOUT) {
2269                         mddev->thread->timeout = timeout;
2270                         md_wakeup_thread(mddev->thread);
2271                 }
2272         }
2273         return len;
2274 }
2275
2276 static struct md_sysfs_entry bitmap_timeout =
2277 __ATTR(time_base, S_IRUGO|S_IWUSR, timeout_show, timeout_store);
2278
2279 static ssize_t
2280 backlog_show(struct mddev *mddev, char *page)
2281 {
2282         return sprintf(page, "%lu\n", mddev->bitmap_info.max_write_behind);
2283 }
2284
2285 static ssize_t
2286 backlog_store(struct mddev *mddev, const char *buf, size_t len)
2287 {
2288         unsigned long backlog;
2289         int rv = kstrtoul(buf, 10, &backlog);
2290         if (rv)
2291                 return rv;
2292         if (backlog > COUNTER_MAX)
2293                 return -EINVAL;
2294         mddev->bitmap_info.max_write_behind = backlog;
2295         return len;
2296 }
2297
2298 static struct md_sysfs_entry bitmap_backlog =
2299 __ATTR(backlog, S_IRUGO|S_IWUSR, backlog_show, backlog_store);
2300
2301 static ssize_t
2302 chunksize_show(struct mddev *mddev, char *page)
2303 {
2304         return sprintf(page, "%lu\n", mddev->bitmap_info.chunksize);
2305 }
2306
2307 static ssize_t
2308 chunksize_store(struct mddev *mddev, const char *buf, size_t len)
2309 {
2310         /* Can only be changed when no bitmap is active */
2311         int rv;
2312         unsigned long csize;
2313         if (mddev->bitmap)
2314                 return -EBUSY;
2315         rv = kstrtoul(buf, 10, &csize);
2316         if (rv)
2317                 return rv;
2318         if (csize < 512 ||
2319             !is_power_of_2(csize))
2320                 return -EINVAL;
2321         mddev->bitmap_info.chunksize = csize;
2322         return len;
2323 }
2324
2325 static struct md_sysfs_entry bitmap_chunksize =
2326 __ATTR(chunksize, S_IRUGO|S_IWUSR, chunksize_show, chunksize_store);
2327
2328 static ssize_t metadata_show(struct mddev *mddev, char *page)
2329 {
2330         if (mddev_is_clustered(mddev))
2331                 return sprintf(page, "clustered\n");
2332         return sprintf(page, "%s\n", (mddev->bitmap_info.external
2333                                       ? "external" : "internal"));
2334 }
2335
2336 static ssize_t metadata_store(struct mddev *mddev, const char *buf, size_t len)
2337 {
2338         if (mddev->bitmap ||
2339             mddev->bitmap_info.file ||
2340             mddev->bitmap_info.offset)
2341                 return -EBUSY;
2342         if (strncmp(buf, "external", 8) == 0)
2343                 mddev->bitmap_info.external = 1;
2344         else if ((strncmp(buf, "internal", 8) == 0) ||
2345                         (strncmp(buf, "clustered", 9) == 0))
2346                 mddev->bitmap_info.external = 0;
2347         else
2348                 return -EINVAL;
2349         return len;
2350 }
2351
2352 static struct md_sysfs_entry bitmap_metadata =
2353 __ATTR(metadata, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
2354
2355 static ssize_t can_clear_show(struct mddev *mddev, char *page)
2356 {
2357         int len;
2358         spin_lock(&mddev->lock);
2359         if (mddev->bitmap)
2360                 len = sprintf(page, "%s\n", (mddev->bitmap->need_sync ?
2361                                              "false" : "true"));
2362         else
2363                 len = sprintf(page, "\n");
2364         spin_unlock(&mddev->lock);
2365         return len;
2366 }
2367
2368 static ssize_t can_clear_store(struct mddev *mddev, const char *buf, size_t len)
2369 {
2370         if (mddev->bitmap == NULL)
2371                 return -ENOENT;
2372         if (strncmp(buf, "false", 5) == 0)
2373                 mddev->bitmap->need_sync = 1;
2374         else if (strncmp(buf, "true", 4) == 0) {
2375                 if (mddev->degraded)
2376                         return -EBUSY;
2377                 mddev->bitmap->need_sync = 0;
2378         } else
2379                 return -EINVAL;
2380         return len;
2381 }
2382
2383 static struct md_sysfs_entry bitmap_can_clear =
2384 __ATTR(can_clear, S_IRUGO|S_IWUSR, can_clear_show, can_clear_store);
2385
2386 static ssize_t
2387 behind_writes_used_show(struct mddev *mddev, char *page)
2388 {
2389         ssize_t ret;
2390         spin_lock(&mddev->lock);
2391         if (mddev->bitmap == NULL)
2392                 ret = sprintf(page, "0\n");
2393         else
2394                 ret = sprintf(page, "%lu\n",
2395                               mddev->bitmap->behind_writes_used);
2396         spin_unlock(&mddev->lock);
2397         return ret;
2398 }
2399
2400 static ssize_t
2401 behind_writes_used_reset(struct mddev *mddev, const char *buf, size_t len)
2402 {
2403         if (mddev->bitmap)
2404                 mddev->bitmap->behind_writes_used = 0;
2405         return len;
2406 }
2407
2408 static struct md_sysfs_entry max_backlog_used =
2409 __ATTR(max_backlog_used, S_IRUGO | S_IWUSR,
2410        behind_writes_used_show, behind_writes_used_reset);
2411
2412 static struct attribute *md_bitmap_attrs[] = {
2413         &bitmap_location.attr,
2414         &bitmap_space.attr,
2415         &bitmap_timeout.attr,
2416         &bitmap_backlog.attr,
2417         &bitmap_chunksize.attr,
2418         &bitmap_metadata.attr,
2419         &bitmap_can_clear.attr,
2420         &max_backlog_used.attr,
2421         NULL
2422 };
2423 struct attribute_group md_bitmap_group = {
2424         .name = "bitmap",
2425         .attrs = md_bitmap_attrs,
2426 };
2427