kernel/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          * Note that if entered with 'rdev == NULL' to start at the
 182          * beginning, we temporarily assign 'rdev' to an address which
 183          * isn't really an rdev, but which can be used by
 184          * list_for_each_entry_continue_rcu() to find the first entry.
 185          */
 186         rcu_read_lock();
 187         if (rdev == NULL)
 188                 /* start at the beginning */
 189                 rdev = list_entry(&mddev->disks, struct md_rdev, same_set);
 190         else {
 191                 /* release the previous rdev and start from there. */
 192                 rdev_dec_pending(rdev, mddev);
 193         }
 194         list_for_each_entry_continue_rcu(rdev, &mddev->disks, same_set) {
 195                 if (rdev->raid_disk >= 0 &&
 196                     !test_bit(Faulty, &rdev->flags)) {
 197                         /* this is a usable devices */
 198                         atomic_inc(&rdev->nr_pending);
 199                         rcu_read_unlock();
 200                         return rdev;
 201                 }
 202         }
 203         rcu_read_unlock();
 204         return NULL;
 205 }
 206
 207 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 208 {
 209         struct md_rdev *rdev = NULL;
 210         struct block_device *bdev;
 211         struct mddev *mddev = bitmap->mddev;
 212         struct bitmap_storage *store = &bitmap->storage;
 213         int node_offset = 0;
 214
 215         if (mddev_is_clustered(bitmap->mddev))
 216                 node_offset = bitmap->cluster_slot * store->file_pages;
 217
 218         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 219                 int size = PAGE_SIZE;
 220                 loff_t offset = mddev->bitmap_info.offset;
 221
 222                 bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
 223
 224                 if (page->index == store->file_pages-1) {
 225                         int last_page_size = store->bytes & (PAGE_SIZE-1);
 226                         if (last_page_size == 0)
 227                                 last_page_size = PAGE_SIZE;
 228                         size = roundup(last_page_size,
 229                                        bdev_logical_block_size(bdev));
 230                 }
 231                 /* Just make sure we aren't corrupting data or
 232                  * metadata
 233                  */
 234                 if (mddev->external) {
 235                         /* Bitmap could be anywhere. */
 236                         if (rdev->sb_start + offset + (page->index
 237                                                        * (PAGE_SIZE/512))
 238                             > rdev->data_offset
 239                             &&
 240                             rdev->sb_start + offset
 241                             < (rdev->data_offset + mddev->dev_sectors
 242                              + (PAGE_SIZE/512)))
 243                                 goto bad_alignment;
 244                 } else if (offset < 0) {
 245                         /* DATA  BITMAP METADATA  */
 246                         if (offset
 247                             + (long)(page->index * (PAGE_SIZE/512))
 248                             + size/512 > 0)
 249                                 /* bitmap runs in to metadata */
 250                                 goto bad_alignment;
 251                         if (rdev->data_offset + mddev->dev_sectors
 252                             > rdev->sb_start + offset)
 253                                 /* data runs in to bitmap */
 254                                 goto bad_alignment;
 255                 } else if (rdev->sb_start < rdev->data_offset) {
 256                         /* METADATA BITMAP DATA */
 257                         if (rdev->sb_start
 258                             + offset
 259                             + page->index*(PAGE_SIZE/512) + size/512
 260                             > rdev->data_offset)
 261                                 /* bitmap runs in to data */
 262                                 goto bad_alignment;
 263                 } else {
 264                         /* DATA METADATA BITMAP - no problems */
 265                 }
 266                 md_super_write(mddev, rdev,
 267                                rdev->sb_start + offset
 268                                + page->index * (PAGE_SIZE/512),
 269                                size,
 270                                page);
 271         }
 272
 273         if (wait)
 274                 md_super_wait(mddev);
 275         return 0;
 276
 277  bad_alignment:
 278         return -EINVAL;
 279 }
 280
 281 static void bitmap_file_kick(struct bitmap *bitmap);
 282 /*
 283  * write out a page to a file
 284  */
 285 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 286 {
 287         struct buffer_head *bh;
 288
 289         if (bitmap->storage.file == NULL) {
 290                 switch (write_sb_page(bitmap, page, wait)) {
 291                 case -EINVAL:
 292                         set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 293                 }
 294         } else {
 295
 296                 bh = page_buffers(page);
 297
 298                 while (bh && bh->b_blocknr) {
 299                         atomic_inc(&bitmap->pending_writes);
 300                         set_buffer_locked(bh);
 301                         set_buffer_mapped(bh);
 302                         submit_bh(WRITE | REQ_SYNC, bh);
 303                         bh = bh->b_this_page;
 304                 }
 305
 306                 if (wait)
 307                         wait_event(bitmap->write_wait,
 308                                    atomic_read(&bitmap->pending_writes)==0);
 309         }
 310         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 311                 bitmap_file_kick(bitmap);
 312 }
 313
 314 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 315 {
 316         struct bitmap *bitmap = bh->b_private;
 317
 318         if (!uptodate)
 319                 set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 320         if (atomic_dec_and_test(&bitmap->pending_writes))
 321                 wake_up(&bitmap->write_wait);
 322 }
 323
 324 /* copied from buffer.c */
 325 static void
 326 __clear_page_buffers(struct page *page)
 327 {
 328         ClearPagePrivate(page);
 329         set_page_private(page, 0);
 330         page_cache_release(page);
 331 }
 332 static void free_buffers(struct page *page)
 333 {
 334         struct buffer_head *bh;
 335
 336         if (!PagePrivate(page))
 337                 return;
 338
 339         bh = page_buffers(page);
 340         while (bh) {
 341                 struct buffer_head *next = bh->b_this_page;
 342                 free_buffer_head(bh);
 343                 bh = next;
 344         }
 345         __clear_page_buffers(page);
 346         put_page(page);
 347 }
 348
 349 /* read a page from a file.
 350  * We both read the page, and attach buffers to the page to record the
 351  * address of each block (using bmap).  These addresses will be used
 352  * to write the block later, completely bypassing the filesystem.
 353  * This usage is similar to how swap files are handled, and allows us
 354  * to write to a file with no concerns of memory allocation failing.
 355  */
 356 static int read_page(struct file *file, unsigned long index,
 357                      struct bitmap *bitmap,
 358                      unsigned long count,
 359                      struct page *page)
 360 {
 361         int ret = 0;
 362         struct inode *inode = file_inode(file);
 363         struct buffer_head *bh;
 364         sector_t block;
 365
 366         pr_debug("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 367                  (unsigned long long)index << PAGE_SHIFT);
 368
 369         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 370         if (!bh) {
 371                 ret = -ENOMEM;
 372                 goto out;
 373         }
 374         attach_page_buffers(page, bh);
 375         block = index << (PAGE_SHIFT - inode->i_blkbits);
 376         while (bh) {
 377                 if (count == 0)
 378                         bh->b_blocknr = 0;
 379                 else {
 380                         bh->b_blocknr = bmap(inode, block);
 381                         if (bh->b_blocknr == 0) {
 382                                 /* Cannot use this file! */
 383                                 ret = -EINVAL;
 384                                 goto out;
 385                         }
 386                         bh->b_bdev = inode->i_sb->s_bdev;
 387                         if (count < (1<<inode->i_blkbits))
 388                                 count = 0;
 389                         else
 390                                 count -= (1<<inode->i_blkbits);
 391
 392                         bh->b_end_io = end_bitmap_write;
 393                         bh->b_private = bitmap;
 394                         atomic_inc(&bitmap->pending_writes);
 395                         set_buffer_locked(bh);
 396                         set_buffer_mapped(bh);
 397                         submit_bh(READ, bh);
 398                 }
 399                 block++;
 400                 bh = bh->b_this_page;
 401         }
 402         page->index = index;
 403
 404         wait_event(bitmap->write_wait,
 405                    atomic_read(&bitmap->pending_writes)==0);
 406         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 407                 ret = -EIO;
 408 out:
 409         if (ret)
 410                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %d\n",
 411                         (int)PAGE_SIZE,
 412                         (unsigned long long)index << PAGE_SHIFT,
 413                         ret);
 414         return ret;
 415 }
 416
 417 /*
 418  * bitmap file superblock operations
 419  */
 420
 421 /* update the event counter and sync the superblock to disk */
 422 void bitmap_update_sb(struct bitmap *bitmap)
 423 {
 424         bitmap_super_t *sb;
 425
 426         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 427                 return;
 428         if (bitmap->mddev->bitmap_info.external)
 429                 return;
 430         if (!bitmap->storage.sb_page) /* no superblock */
 431                 return;
 432         sb = kmap_atomic(bitmap->storage.sb_page);
 433         sb->events = cpu_to_le64(bitmap->mddev->events);
 434         if (bitmap->mddev->events < bitmap->events_cleared)
 435                 /* rocking back to read-only */
 436                 bitmap->events_cleared = bitmap->mddev->events;
 437         sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 438         sb->state = cpu_to_le32(bitmap->flags);
 439         /* Just in case these have been changed via sysfs: */
 440         sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 441         sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
 442         /* This might have been changed by a reshape */
 443         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 444         sb->chunksize = cpu_to_le32(bitmap->mddev->bitmap_info.chunksize);
 445         sb->nodes = cpu_to_le32(bitmap->mddev->bitmap_info.nodes);
 446         sb->sectors_reserved = cpu_to_le32(bitmap->mddev->
 447                                            bitmap_info.space);
 448         kunmap_atomic(sb);
 449         write_page(bitmap, bitmap->storage.sb_page, 1);
 450 }
 451
 452 /* print out the bitmap file superblock */
 453 void bitmap_print_sb(struct bitmap *bitmap)
 454 {
 455         bitmap_super_t *sb;
 456
 457         if (!bitmap || !bitmap->storage.sb_page)
 458                 return;
 459         sb = kmap_atomic(bitmap->storage.sb_page);
 460         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 461         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 462         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 463         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 464                                         *(__u32 *)(sb->uuid+0),
 465                                         *(__u32 *)(sb->uuid+4),
 466                                         *(__u32 *)(sb->uuid+8),
 467                                         *(__u32 *)(sb->uuid+12));
 468         printk(KERN_DEBUG "        events: %llu\n",
 469                         (unsigned long long) le64_to_cpu(sb->events));
 470         printk(KERN_DEBUG "events cleared: %llu\n",
 471                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 472         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 473         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 474         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 475         printk(KERN_DEBUG "     sync size: %llu KB\n",
 476                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 477         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 478         kunmap_atomic(sb);
 479 }
 480
 481 /*
 482  * bitmap_new_disk_sb
 483  * @bitmap
 484  *
 485  * This function is somewhat the reverse of bitmap_read_sb.  bitmap_read_sb
 486  * reads and verifies the on-disk bitmap superblock and populates bitmap_info.
 487  * This function verifies 'bitmap_info' and populates the on-disk bitmap
 488  * structure, which is to be written to disk.
 489  *
 490  * Returns: 0 on success, -Exxx on error
 491  */
 492 static int bitmap_new_disk_sb(struct bitmap *bitmap)
 493 {
 494         bitmap_super_t *sb;
 495         unsigned long chunksize, daemon_sleep, write_behind;
 496
 497         bitmap->storage.sb_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 498         if (bitmap->storage.sb_page == NULL)
 499                 return -ENOMEM;
 500         bitmap->storage.sb_page->index = 0;
 501
 502         sb = kmap_atomic(bitmap->storage.sb_page);
 503
 504         sb->magic = cpu_to_le32(BITMAP_MAGIC);
 505         sb->version = cpu_to_le32(BITMAP_MAJOR_HI);
 506
 507         chunksize = bitmap->mddev->bitmap_info.chunksize;
 508         BUG_ON(!chunksize);
 509         if (!is_power_of_2(chunksize)) {
 510                 kunmap_atomic(sb);
 511                 printk(KERN_ERR "bitmap chunksize not a power of 2\n");
 512                 return -EINVAL;
 513         }
 514         sb->chunksize = cpu_to_le32(chunksize);
 515
 516         daemon_sleep = bitmap->mddev->bitmap_info.daemon_sleep;
 517         if (!daemon_sleep ||
 518             (daemon_sleep < 1) || (daemon_sleep > MAX_SCHEDULE_TIMEOUT)) {
 519                 printk(KERN_INFO "Choosing daemon_sleep default (5 sec)\n");
 520                 daemon_sleep = 5 * HZ;
 521         }
 522         sb->daemon_sleep = cpu_to_le32(daemon_sleep);
 523         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 524
 525         /*
 526          * FIXME: write_behind for RAID1.  If not specified, what
 527          * is a good choice?  We choose COUNTER_MAX / 2 arbitrarily.
 528          */
 529         write_behind = bitmap->mddev->bitmap_info.max_write_behind;
 530         if (write_behind > COUNTER_MAX)
 531                 write_behind = COUNTER_MAX / 2;
 532         sb->write_behind = cpu_to_le32(write_behind);
 533         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 534
 535         /* keep the array size field of the bitmap superblock up to date */
 536         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 537
 538         memcpy(sb->uuid, bitmap->mddev->uuid, 16);
 539
 540         set_bit(BITMAP_STALE, &bitmap->flags);
 541         sb->state = cpu_to_le32(bitmap->flags);
 542         bitmap->events_cleared = bitmap->mddev->events;
 543         sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 544         bitmap->mddev->bitmap_info.nodes = 0;
 545
 546         kunmap_atomic(sb);
 547
 548         return 0;
 549 }
 550
 551 /* read the superblock from the bitmap file and initialize some bitmap fields */
 552 static int bitmap_read_sb(struct bitmap *bitmap)
 553 {
 554         char *reason = NULL;
 555         bitmap_super_t *sb;
 556         unsigned long chunksize, daemon_sleep, write_behind;
 557         unsigned long long events;
 558         int nodes = 0;
 559         unsigned long sectors_reserved = 0;
 560         int err = -EINVAL;
 561         struct page *sb_page;
 562
 563         if (!bitmap->storage.file && !bitmap->mddev->bitmap_info.offset) {
 564                 chunksize = 128 * 1024 * 1024;
 565                 daemon_sleep = 5 * HZ;
 566                 write_behind = 0;
 567                 set_bit(BITMAP_STALE, &bitmap->flags);
 568                 err = 0;
 569                 goto out_no_sb;
 570         }
 571         /* page 0 is the superblock, read it... */
 572         sb_page = alloc_page(GFP_KERNEL);
 573         if (!sb_page)
 574                 return -ENOMEM;
 575         bitmap->storage.sb_page = sb_page;
 576
 577 re_read:
 578         /* If cluster_slot is set, the cluster is setup */
 579         if (bitmap->cluster_slot >= 0) {
 580                 sector_t bm_blocks = bitmap->mddev->resync_max_sectors;
 581
 582                 sector_div(bm_blocks,
 583                            bitmap->mddev->bitmap_info.chunksize >> 9);
 584                 /* bits to bytes */
 585                 bm_blocks = ((bm_blocks+7) >> 3) + sizeof(bitmap_super_t);
 586                 /* to 4k blocks */
 587                 bm_blocks = DIV_ROUND_UP_SECTOR_T(bm_blocks, 4096);
 588                 bitmap->mddev->bitmap_info.offset += bitmap->cluster_slot * (bm_blocks << 3);
 589                 pr_info("%s:%d bm slot: %d offset: %llu\n", __func__, __LINE__,
 590                         bitmap->cluster_slot, (unsigned long long)bitmap->mddev->bitmap_info.offset);
 591         }
 592
 593         if (bitmap->storage.file) {
 594                 loff_t isize = i_size_read(bitmap->storage.file->f_mapping->host);
 595                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 596
 597                 err = read_page(bitmap->storage.file, 0,
 598                                 bitmap, bytes, sb_page);
 599         } else {
 600                 err = read_sb_page(bitmap->mddev,
 601                                    bitmap->mddev->bitmap_info.offset,
 602                                    sb_page,
 603                                    0, sizeof(bitmap_super_t));
 604         }
 605         if (err)
 606                 return err;
 607
 608         err = -EINVAL;
 609         sb = kmap_atomic(sb_page);
 610
 611         chunksize = le32_to_cpu(sb->chunksize);
 612         daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 613         write_behind = le32_to_cpu(sb->write_behind);
 614         sectors_reserved = le32_to_cpu(sb->sectors_reserved);
 615         /* XXX: This is a hack to ensure that we don't use clustering
 616          *  in case:
 617          *      - dm-raid is in use and
 618          *      - the nodes written in bitmap_sb is erroneous.
 619          */
 620         if (!bitmap->mddev->sync_super) {
 621                 nodes = le32_to_cpu(sb->nodes);
 622                 strlcpy(bitmap->mddev->bitmap_info.cluster_name,
 623                                 sb->cluster_name, 64);
 624         }
 625
 626         /* verify that the bitmap-specific fields are valid */
 627         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 628                 reason = "bad magic";
 629         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 630                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 631                 reason = "unrecognized superblock version";
 632         else if (chunksize < 512)
 633                 reason = "bitmap chunksize too small";
 634         else if (!is_power_of_2(chunksize))
 635                 reason = "bitmap chunksize not a power of 2";
 636         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT)
 637                 reason = "daemon sleep period out of range";
 638         else if (write_behind > COUNTER_MAX)
 639                 reason = "write-behind limit out of range (0 - 16383)";
 640         if (reason) {
 641                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 642                         bmname(bitmap), reason);
 643                 goto out;
 644         }
 645
 646         /* keep the array size field of the bitmap superblock up to date */
 647         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 648
 649         if (bitmap->mddev->persistent) {
 650                 /*
 651                  * We have a persistent array superblock, so compare the
 652                  * bitmap's UUID and event counter to the mddev's
 653                  */
 654                 if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 655                         printk(KERN_INFO
 656                                "%s: bitmap superblock UUID mismatch\n",
 657                                bmname(bitmap));
 658                         goto out;
 659                 }
 660                 events = le64_to_cpu(sb->events);
 661                 if (!nodes && (events < bitmap->mddev->events)) {
 662                         printk(KERN_INFO
 663                                "%s: bitmap file is out of date (%llu < %llu) "
 664                                "-- forcing full recovery\n",
 665                                bmname(bitmap), events,
 666                                (unsigned long long) bitmap->mddev->events);
 667                         set_bit(BITMAP_STALE, &bitmap->flags);
 668                 }
 669         }
 670
 671         /* assign fields using values from superblock */
 672         bitmap->flags |= le32_to_cpu(sb->state);
 673         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 674                 set_bit(BITMAP_HOSTENDIAN, &bitmap->flags);
 675         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 676         strlcpy(bitmap->mddev->bitmap_info.cluster_name, sb->cluster_name, 64);
 677         err = 0;
 678
 679 out:
 680         kunmap_atomic(sb);
 681         /* Assiging chunksize is required for "re_read" */
 682         bitmap->mddev->bitmap_info.chunksize = chunksize;
 683         if (nodes && (bitmap->cluster_slot < 0)) {
 684                 err = md_setup_cluster(bitmap->mddev, nodes);
 685                 if (err) {
 686                         pr_err("%s: Could not setup cluster service (%d)\n",
 687                                         bmname(bitmap), err);
 688                         goto out_no_sb;
 689                 }
 690                 bitmap->cluster_slot = md_cluster_ops->slot_number(bitmap->mddev);
 691                 goto re_read;
 692         }
 693
 694
 695 out_no_sb:
 696         if (test_bit(BITMAP_STALE, &bitmap->flags))
 697                 bitmap->events_cleared = bitmap->mddev->events;
 698         bitmap->mddev->bitmap_info.chunksize = chunksize;
 699         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 700         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 701         bitmap->mddev->bitmap_info.nodes = nodes;
 702         if (bitmap->mddev->bitmap_info.space == 0 ||
 703             bitmap->mddev->bitmap_info.space > sectors_reserved)
 704                 bitmap->mddev->bitmap_info.space = sectors_reserved;
 705         if (err) {
 706                 bitmap_print_sb(bitmap);
 707                 if (bitmap->cluster_slot < 0)
 708                         md_cluster_stop(bitmap->mddev);
 709         }
 710         return err;
 711 }
 712
 713 /*
 714  * general bitmap file operations
 715  */
 716
 717 /*
 718  * on-disk bitmap:
 719  *
 720  * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
 721  * file a page at a time. There's a superblock at the start of the file.
 722  */
 723 /* calculate the index of the page that contains this bit */
 724 static inline unsigned long file_page_index(struct bitmap_storage *store,
 725                                             unsigned long chunk)
 726 {
 727         if (store->sb_page)
 728                 chunk += sizeof(bitmap_super_t) << 3;
 729         return chunk >> PAGE_BIT_SHIFT;
 730 }
 731
 732 /* calculate the (bit) offset of this bit within a page */
 733 static inline unsigned long file_page_offset(struct bitmap_storage *store,
 734                                              unsigned long chunk)
 735 {
 736         if (store->sb_page)
 737                 chunk += sizeof(bitmap_super_t) << 3;
 738         return chunk & (PAGE_BITS - 1);
 739 }
 740
 741 /*
 742  * return a pointer to the page in the filemap that contains the given bit
 743  *
 744  */
 745 static inline struct page *filemap_get_page(struct bitmap_storage *store,
 746                                             unsigned long chunk)
 747 {
 748         if (file_page_index(store, chunk) >= store->file_pages)
 749                 return NULL;
 750         return store->filemap[file_page_index(store, chunk)];
 751 }
 752
 753 static int bitmap_storage_alloc(struct bitmap_storage *store,
 754                                 unsigned long chunks, int with_super,
 755                                 int slot_number)
 756 {
 757         int pnum, offset = 0;
 758         unsigned long num_pages;
 759         unsigned long bytes;
 760
 761         bytes = DIV_ROUND_UP(chunks, 8);
 762         if (with_super)
 763                 bytes += sizeof(bitmap_super_t);
 764
 765         num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
 766         offset = slot_number * (num_pages - 1);
 767
 768         store->filemap = kmalloc(sizeof(struct page *)
 769                                  * num_pages, GFP_KERNEL);
 770         if (!store->filemap)
 771                 return -ENOMEM;
 772
 773         if (with_super && !store->sb_page) {
 774                 store->sb_page = alloc_page(GFP_KERNEL|__GFP_ZERO);
 775                 if (store->sb_page == NULL)
 776                         return -ENOMEM;
 777         }
 778
 779         pnum = 0;
 780         if (store->sb_page) {
 781                 store->filemap[0] = store->sb_page;
 782                 pnum = 1;
 783                 store->sb_page->index = offset;
 784         }
 785
 786         for ( ; pnum < num_pages; pnum++) {
 787                 store->filemap[pnum] = alloc_page(GFP_KERNEL|__GFP_ZERO);
 788                 if (!store->filemap[pnum]) {
 789                         store->file_pages = pnum;
 790                         return -ENOMEM;
 791                 }
 792                 store->filemap[pnum]->index = pnum + offset;
 793         }
 794         store->file_pages = pnum;
 795
 796         /* We need 4 bits per page, rounded up to a multiple
 797          * of sizeof(unsigned long) */
 798         store->filemap_attr = kzalloc(
 799                 roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
 800                 GFP_KERNEL);
 801         if (!store->filemap_attr)
 802                 return -ENOMEM;
 803
 804         store->bytes = bytes;
 805
 806         return 0;
 807 }
 808
 809 static void bitmap_file_unmap(struct bitmap_storage *store)
 810 {
 811         struct page **map, *sb_page;
 812         int pages;
 813         struct file *file;
 814
 815         file = store->file;
 816         map = store->filemap;
 817         pages = store->file_pages;
 818         sb_page = store->sb_page;
 819
 820         while (pages--)
 821                 if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 822                         free_buffers(map[pages]);
 823         kfree(map);
 824         kfree(store->filemap_attr);
 825
 826         if (sb_page)
 827                 free_buffers(sb_page);
 828
 829         if (file) {
 830                 struct inode *inode = file_inode(file);
 831                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 832                 fput(file);
 833         }
 834 }
 835
 836 /*
 837  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 838  * then it is no longer reliable, so we stop using it and we mark the file
 839  * as failed in the superblock
 840  */
 841 static void bitmap_file_kick(struct bitmap *bitmap)
 842 {
 843         char *path, *ptr = NULL;
 844
 845         if (!test_and_set_bit(BITMAP_STALE, &bitmap->flags)) {
 846                 bitmap_update_sb(bitmap);
 847
 848                 if (bitmap->storage.file) {
 849                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 850                         if (path)
 851                                 ptr = d_path(&bitmap->storage.file->f_path,
 852                                              path, PAGE_SIZE);
 853
 854                         printk(KERN_ALERT
 855                               "%s: kicking failed bitmap file %s from array!\n",
 856                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 857
 858                         kfree(path);
 859                 } else
 860                         printk(KERN_ALERT
 861                                "%s: disabling internal bitmap due to errors\n",
 862                                bmname(bitmap));
 863         }
 864 }
 865
 866 enum bitmap_page_attr {
 867         BITMAP_PAGE_DIRTY = 0,     /* there are set bits that need to be synced */
 868         BITMAP_PAGE_PENDING = 1,   /* there are bits that are being cleaned.
 869                                     * i.e. counter is 1 or 2. */
 870         BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 871 };
 872
 873 static inline void set_page_attr(struct bitmap *bitmap, int pnum,
 874                                  enum bitmap_page_attr attr)
 875 {
 876         set_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 877 }
 878
 879 static inline void clear_page_attr(struct bitmap *bitmap, int pnum,
 880                                    enum bitmap_page_attr attr)
 881 {
 882         clear_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 883 }
 884
 885 static inline int test_page_attr(struct bitmap *bitmap, int pnum,
 886                                  enum bitmap_page_attr attr)
 887 {
 888         return test_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 889 }
 890
 891 static inline int test_and_clear_page_attr(struct bitmap *bitmap, int pnum,
 892                                            enum bitmap_page_attr attr)
 893 {
 894         return test_and_clear_bit((pnum<<2) + attr,
 895                                   bitmap->storage.filemap_attr);
 896 }
 897 /*
 898  * bitmap_file_set_bit -- called before performing a write to the md device
 899  * to set (and eventually sync) a particular bit in the bitmap file
 900  *
 901  * we set the bit immediately, then we record the page number so that
 902  * when an unplug occurs, we can flush the dirty pages out to disk
 903  */
 904 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 905 {
 906         unsigned long bit;
 907         struct page *page;
 908         void *kaddr;
 909         unsigned long chunk = block >> bitmap->counts.chunkshift;
 910
 911         page = filemap_get_page(&bitmap->storage, chunk);
 912         if (!page)
 913                 return;
 914         bit = file_page_offset(&bitmap->storage, chunk);
 915
 916         /* set the bit */
 917         kaddr = kmap_atomic(page);
 918         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 919                 set_bit(bit, kaddr);
 920         else
 921                 set_bit_le(bit, kaddr);
 922         kunmap_atomic(kaddr);
 923         pr_debug("set file bit %lu page %lu\n", bit, page->index);
 924         /* record page number so it gets flushed to disk when unplug occurs */
 925         set_page_attr(bitmap, page->index, BITMAP_PAGE_DIRTY);
 926 }
 927
 928 static void bitmap_file_clear_bit(struct bitmap *bitmap, sector_t block)
 929 {
 930         unsigned long bit;
 931         struct page *page;
 932         void *paddr;
 933         unsigned long chunk = block >> bitmap->counts.chunkshift;
 934
 935         page = filemap_get_page(&bitmap->storage, chunk);
 936         if (!page)
 937                 return;
 938         bit = file_page_offset(&bitmap->storage, chunk);
 939         paddr = kmap_atomic(page);
 940         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 941                 clear_bit(bit, paddr);
 942         else
 943                 clear_bit_le(bit, paddr);
 944         kunmap_atomic(paddr);
 945         if (!test_page_attr(bitmap, page->index, BITMAP_PAGE_NEEDWRITE)) {
 946                 set_page_attr(bitmap, page->index, BITMAP_PAGE_PENDING);
 947                 bitmap->allclean = 0;
 948         }
 949 }
 950
 951 static int bitmap_file_test_bit(struct bitmap *bitmap, sector_t block)
 952 {
 953         unsigned long bit;
 954         struct page *page;
 955         void *paddr;
 956         unsigned long chunk = block >> bitmap->counts.chunkshift;
 957         int set = 0;
 958
 959         page = filemap_get_page(&bitmap->storage, chunk);
 960         if (!page)
 961                 return -EINVAL;
 962         bit = file_page_offset(&bitmap->storage, chunk);
 963         paddr = kmap_atomic(page);
 964         if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 965                 set = test_bit(bit, paddr);
 966         else
 967                 set = test_bit_le(bit, paddr);
 968         kunmap_atomic(paddr);
 969         return set;
 970 }
 971
 972
 973 /* this gets called when the md device is ready to unplug its underlying
 974  * (slave) device queues -- before we let any writes go down, we need to
 975  * sync the dirty pages of the bitmap file to disk */
 976 void bitmap_unplug(struct bitmap *bitmap)
 977 {
 978         unsigned long i;
 979         int dirty, need_write;
 980
 981         if (!bitmap || !bitmap->storage.filemap ||
 982             test_bit(BITMAP_STALE, &bitmap->flags))
 983                 return;
 984
 985         /* look at each page to see if there are any set bits that need to be
 986          * flushed out to disk */
 987         for (i = 0; i < bitmap->storage.file_pages; i++) {
 988                 if (!bitmap->storage.filemap)
 989                         return;
 990                 dirty = test_and_clear_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
 991                 need_write = test_and_clear_page_attr(bitmap, i,
 992                                                       BITMAP_PAGE_NEEDWRITE);
 993                 if (dirty || need_write) {
 994                         clear_page_attr(bitmap, i, BITMAP_PAGE_PENDING);
 995                         write_page(bitmap, bitmap->storage.filemap[i], 0);
 996                 }
 997         }
 998         if (bitmap->storage.file)
 999                 wait_event(bitmap->write_wait,
1000                            atomic_read(&bitmap->pending_writes)==0);
1001         else
1002                 md_super_wait(bitmap->mddev);
1003
1004         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
1005                 bitmap_file_kick(bitmap);
1006 }
1007 EXPORT_SYMBOL(bitmap_unplug);
1008
1009 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
1010 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
1011  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
1012  * memory mapping of the bitmap file
1013  * Special cases:
1014  *   if there's no bitmap file, or if the bitmap file had been
1015  *   previously kicked from the array, we mark all the bits as
1016  *   1's in order to cause a full resync.
1017  *
1018  * We ignore all bits for sectors that end earlier than 'start'.
1019  * This is used when reading an out-of-date bitmap...
1020  */
1021 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
1022 {
1023         unsigned long i, chunks, index, oldindex, bit, node_offset = 0;
1024         struct page *page = NULL;
1025         unsigned long bit_cnt = 0;
1026         struct file *file;
1027         unsigned long offset;
1028         int outofdate;
1029         int ret = -ENOSPC;
1030         void *paddr;
1031         struct bitmap_storage *store = &bitmap->storage;
1032
1033         chunks = bitmap->counts.chunks;
1034         file = store->file;
1035
1036         if (!file && !bitmap->mddev->bitmap_info.offset) {
1037                 /* No permanent bitmap - fill with '1s'. */
1038                 store->filemap = NULL;
1039                 store->file_pages = 0;
1040                 for (i = 0; i < chunks ; i++) {
1041                         /* if the disk bit is set, set the memory bit */
1042                         int needed = ((sector_t)(i+1) << (bitmap->counts.chunkshift)
1043                                       >= start);
1044                         bitmap_set_memory_bits(bitmap,
1045                                                (sector_t)i << bitmap->counts.chunkshift,
1046                                                needed);
1047                 }
1048                 return 0;
1049         }
1050
1051         outofdate = test_bit(BITMAP_STALE, &bitmap->flags);
1052         if (outofdate)
1053                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
1054                         "recovery\n", bmname(bitmap));
1055
1056         if (file && i_size_read(file->f_mapping->host) < store->bytes) {
1057                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
1058                        bmname(bitmap),
1059                        (unsigned long) i_size_read(file->f_mapping->host),
1060                        store->bytes);
1061                 goto err;
1062         }
1063
1064         oldindex = ~0L;
1065         offset = 0;
1066         if (!bitmap->mddev->bitmap_info.external)
1067                 offset = sizeof(bitmap_super_t);
1068
1069         if (mddev_is_clustered(bitmap->mddev))
1070                 node_offset = bitmap->cluster_slot * (DIV_ROUND_UP(store->bytes, PAGE_SIZE));
1071
1072         for (i = 0; i < chunks; i++) {
1073                 int b;
1074                 index = file_page_index(&bitmap->storage, i);
1075                 bit = file_page_offset(&bitmap->storage, i);
1076                 if (index != oldindex) { /* this is a new page, read it in */
1077                         int count;
1078                         /* unmap the old page, we're done with it */
1079                         if (index == store->file_pages-1)
1080                                 count = store->bytes - index * PAGE_SIZE;
1081                         else
1082                                 count = PAGE_SIZE;
1083                         page = store->filemap[index];
1084                         if (file)
1085                                 ret = read_page(file, index, bitmap,
1086                                                 count, page);
1087                         else
1088                                 ret = read_sb_page(
1089                                         bitmap->mddev,
1090                                         bitmap->mddev->bitmap_info.offset,
1091                                         page,
1092                                         index + node_offset, count);
1093
1094                         if (ret)
1095                                 goto err;
1096
1097                         oldindex = index;
1098
1099                         if (outofdate) {
1100                                 /*
1101                                  * if bitmap is out of date, dirty the
1102                                  * whole page and write it out
1103                                  */
1104                                 paddr = kmap_atomic(page);
1105                                 memset(paddr + offset, 0xff,
1106                                        PAGE_SIZE - offset);
1107                                 kunmap_atomic(paddr);
1108                                 write_page(bitmap, page, 1);
1109
1110                                 ret = -EIO;
1111                                 if (test_bit(BITMAP_WRITE_ERROR,
1112                                              &bitmap->flags))
1113                                         goto err;
1114                         }
1115                 }
1116                 paddr = kmap_atomic(page);
1117                 if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
1118                         b = test_bit(bit, paddr);
1119                 else
1120                         b = test_bit_le(bit, paddr);
1121                 kunmap_atomic(paddr);
1122                 if (b) {
1123                         /* if the disk bit is set, set the memory bit */
1124                         int needed = ((sector_t)(i+1) << bitmap->counts.chunkshift
1125                                       >= start);
1126                         bitmap_set_memory_bits(bitmap,
1127                                                (sector_t)i << bitmap->counts.chunkshift,
1128                                                needed);
1129                         bit_cnt++;
1130                 }
1131                 offset = 0;
1132         }
1133
1134         printk(KERN_INFO "%s: bitmap initialized from disk: "
1135                "read %lu pages, set %lu of %lu bits\n",
1136                bmname(bitmap), store->file_pages,
1137                bit_cnt, chunks);
1138
1139         return 0;
1140
1141  err:
1142         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1143                bmname(bitmap), ret);
1144         return ret;
1145 }
1146
1147 void bitmap_write_all(struct bitmap *bitmap)
1148 {
1149         /* We don't actually write all bitmap blocks here,
1150          * just flag them as needing to be written
1151          */
1152         int i;
1153
1154         if (!bitmap || !bitmap->storage.filemap)
1155                 return;
1156         if (bitmap->storage.file)
1157                 /* Only one copy, so nothing needed */
1158                 return;
1159
1160         for (i = 0; i < bitmap->storage.file_pages; i++)
1161                 set_page_attr(bitmap, i,
1162                               BITMAP_PAGE_NEEDWRITE);
1163         bitmap->allclean = 0;
1164 }
1165
1166 static void bitmap_count_page(struct bitmap_counts *bitmap,
1167                               sector_t offset, int inc)
1168 {
1169         sector_t chunk = offset >> bitmap->chunkshift;
1170         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1171         bitmap->bp[page].count += inc;
1172         bitmap_checkfree(bitmap, page);
1173 }
1174
1175 static void bitmap_set_pending(struct bitmap_counts *bitmap, sector_t offset)
1176 {
1177         sector_t chunk = offset >> bitmap->chunkshift;
1178         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1179         struct bitmap_page *bp = &bitmap->bp[page];
1180
1181         if (!bp->pending)
1182                 bp->pending = 1;
1183 }
1184
1185 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1186                                             sector_t offset, sector_t *blocks,
1187                                             int create);
1188
1189 /*
1190  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1191  *                      out to disk
1192  */
1193
1194 void bitmap_daemon_work(struct mddev *mddev)
1195 {
1196         struct bitmap *bitmap;
1197         unsigned long j;
1198         unsigned long nextpage;
1199         sector_t blocks;
1200         struct bitmap_counts *counts;
1201
1202         /* Use a mutex to guard daemon_work against
1203          * bitmap_destroy.
1204          */
1205         mutex_lock(&mddev->bitmap_info.mutex);
1206         bitmap = mddev->bitmap;
1207         if (bitmap == NULL) {
1208                 mutex_unlock(&mddev->bitmap_info.mutex);
1209                 return;
1210         }
1211         if (time_before(jiffies, bitmap->daemon_lastrun
1212                         + mddev->bitmap_info.daemon_sleep))
1213                 goto done;
1214
1215         bitmap->daemon_lastrun = jiffies;
1216         if (bitmap->allclean) {
1217                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1218                 goto done;
1219         }
1220         bitmap->allclean = 1;
1221
1222         /* Any file-page which is PENDING now needs to be written.
1223          * So set NEEDWRITE now, then after we make any last-minute changes
1224          * we will write it.
1225          */
1226         for (j = 0; j < bitmap->storage.file_pages; j++)
1227                 if (test_and_clear_page_attr(bitmap, j,
1228                                              BITMAP_PAGE_PENDING))
1229                         set_page_attr(bitmap, j,
1230                                       BITMAP_PAGE_NEEDWRITE);
1231
1232         if (bitmap->need_sync &&
1233             mddev->bitmap_info.external == 0) {
1234                 /* Arrange for superblock update as well as
1235                  * other changes */
1236                 bitmap_super_t *sb;
1237                 bitmap->need_sync = 0;
1238                 if (bitmap->storage.filemap) {
1239                         sb = kmap_atomic(bitmap->storage.sb_page);
1240                         sb->events_cleared =
1241                                 cpu_to_le64(bitmap->events_cleared);
1242                         kunmap_atomic(sb);
1243                         set_page_attr(bitmap, 0,
1244                                       BITMAP_PAGE_NEEDWRITE);
1245                 }
1246         }
1247         /* Now look at the bitmap counters and if any are '2' or '1',
1248          * decrement and handle accordingly.
1249          */
1250         counts = &bitmap->counts;
1251         spin_lock_irq(&counts->lock);
1252         nextpage = 0;
1253         for (j = 0; j < counts->chunks; j++) {
1254                 bitmap_counter_t *bmc;
1255                 sector_t  block = (sector_t)j << counts->chunkshift;
1256
1257                 if (j == nextpage) {
1258                         nextpage += PAGE_COUNTER_RATIO;
1259                         if (!counts->bp[j >> PAGE_COUNTER_SHIFT].pending) {
1260                                 j |= PAGE_COUNTER_MASK;
1261                                 continue;
1262                         }
1263                         counts->bp[j >> PAGE_COUNTER_SHIFT].pending = 0;
1264                 }
1265                 bmc = bitmap_get_counter(counts,
1266                                          block,
1267                                          &blocks, 0);
1268
1269                 if (!bmc) {
1270                         j |= PAGE_COUNTER_MASK;
1271                         continue;
1272                 }
1273                 if (*bmc == 1 && !bitmap->need_sync) {
1274                         /* We can clear the bit */
1275                         *bmc = 0;
1276                         bitmap_count_page(counts, block, -1);
1277                         bitmap_file_clear_bit(bitmap, block);
1278                 } else if (*bmc && *bmc <= 2) {
1279                         *bmc = 1;
1280                         bitmap_set_pending(counts, block);
1281                         bitmap->allclean = 0;
1282                 }
1283         }
1284         spin_unlock_irq(&counts->lock);
1285
1286         /* Now start writeout on any page in NEEDWRITE that isn't DIRTY.
1287          * DIRTY pages need to be written by bitmap_unplug so it can wait
1288          * for them.
1289          * If we find any DIRTY page we stop there and let bitmap_unplug
1290          * handle all the rest.  This is important in the case where
1291          * the first blocking holds the superblock and it has been updated.
1292          * We mustn't write any other blocks before the superblock.
1293          */
1294         for (j = 0;
1295              j < bitmap->storage.file_pages
1296                      && !test_bit(BITMAP_STALE, &bitmap->flags);
1297              j++) {
1298                 if (test_page_attr(bitmap, j,
1299                                    BITMAP_PAGE_DIRTY))
1300                         /* bitmap_unplug will handle the rest */
1301                         break;
1302                 if (test_and_clear_page_attr(bitmap, j,
1303                                              BITMAP_PAGE_NEEDWRITE)) {
1304                         write_page(bitmap, bitmap->storage.filemap[j], 0);
1305                 }
1306         }
1307
1308  done:
1309         if (bitmap->allclean == 0)
1310                 mddev->thread->timeout =
1311                         mddev->bitmap_info.daemon_sleep;
1312         mutex_unlock(&mddev->bitmap_info.mutex);
1313 }
1314
1315 static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
1316                                             sector_t offset, sector_t *blocks,
1317                                             int create)
1318 __releases(bitmap->lock)
1319 __acquires(bitmap->lock)
1320 {
1321         /* If 'create', we might release the lock and reclaim it.
1322          * The lock must have been taken with interrupts enabled.
1323          * If !create, we don't release the lock.
1324          */
1325         sector_t chunk = offset >> bitmap->chunkshift;
1326         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1327         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1328         sector_t csize;
1329         int err;
1330
1331         err = bitmap_checkpage(bitmap, page, create);
1332
1333         if (bitmap->bp[page].hijacked ||
1334             bitmap->bp[page].map == NULL)
1335                 csize = ((sector_t)1) << (bitmap->chunkshift +
1336                                           PAGE_COUNTER_SHIFT - 1);
1337         else
1338                 csize = ((sector_t)1) << bitmap->chunkshift;
1339         *blocks = csize - (offset & (csize - 1));
1340
1341         if (err < 0)
1342                 return NULL;
1343
1344         /* now locked ... */
1345
1346         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1347                 /* should we use the first or second counter field
1348                  * of the hijacked pointer? */
1349                 int hi = (pageoff > PAGE_COUNTER_MASK);
1350                 return  &((bitmap_counter_t *)
1351                           &bitmap->bp[page].map)[hi];
1352         } else /* page is allocated */
1353                 return (bitmap_counter_t *)
1354                         &(bitmap->bp[page].map[pageoff]);
1355 }
1356
1357 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1358 {
1359         if (!bitmap)
1360                 return 0;
1361
1362         if (behind) {
1363                 int bw;
1364                 atomic_inc(&bitmap->behind_writes);
1365                 bw = atomic_read(&bitmap->behind_writes);
1366                 if (bw > bitmap->behind_writes_used)
1367                         bitmap->behind_writes_used = bw;
1368
1369                 pr_debug("inc write-behind count %d/%lu\n",
1370                          bw, bitmap->mddev->bitmap_info.max_write_behind);
1371         }
1372
1373         while (sectors) {
1374                 sector_t blocks;
1375                 bitmap_counter_t *bmc;
1376
1377                 spin_lock_irq(&bitmap->counts.lock);
1378                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 1);
1379                 if (!bmc) {
1380                         spin_unlock_irq(&bitmap->counts.lock);
1381                         return 0;
1382                 }
1383
1384                 if (unlikely(COUNTER(*bmc) == COUNTER_MAX)) {
1385                         DEFINE_WAIT(__wait);
1386                         /* note that it is safe to do the prepare_to_wait
1387                          * after the test as long as we do it before dropping
1388                          * the spinlock.
1389                          */
1390                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1391                                         TASK_UNINTERRUPTIBLE);
1392                         spin_unlock_irq(&bitmap->counts.lock);
1393                         schedule();
1394                         finish_wait(&bitmap->overflow_wait, &__wait);
1395                         continue;
1396                 }
1397
1398                 switch (*bmc) {
1399                 case 0:
1400                         bitmap_file_set_bit(bitmap, offset);
1401                         bitmap_count_page(&bitmap->counts, offset, 1);
1402                         /* fall through */
1403                 case 1:
1404                         *bmc = 2;
1405                 }
1406
1407                 (*bmc)++;
1408
1409                 spin_unlock_irq(&bitmap->counts.lock);
1410
1411                 offset += blocks;
1412                 if (sectors > blocks)
1413                         sectors -= blocks;
1414                 else
1415                         sectors = 0;
1416         }
1417         return 0;
1418 }
1419 EXPORT_SYMBOL(bitmap_startwrite);
1420
1421 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1422                      int success, int behind)
1423 {
1424         if (!bitmap)
1425                 return;
1426         if (behind) {
1427                 if (atomic_dec_and_test(&bitmap->behind_writes))
1428                         wake_up(&bitmap->behind_wait);
1429                 pr_debug("dec write-behind count %d/%lu\n",
1430                          atomic_read(&bitmap->behind_writes),
1431                          bitmap->mddev->bitmap_info.max_write_behind);
1432         }
1433
1434         while (sectors) {
1435                 sector_t blocks;
1436                 unsigned long flags;
1437                 bitmap_counter_t *bmc;
1438
1439                 spin_lock_irqsave(&bitmap->counts.lock, flags);
1440                 bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 0);
1441                 if (!bmc) {
1442                         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1443                         return;
1444                 }
1445
1446                 if (success && !bitmap->mddev->degraded &&
1447                     bitmap->events_cleared < bitmap->mddev->events) {
1448                         bitmap->events_cleared = bitmap->mddev->events;
1449                         bitmap->need_sync = 1;
1450                         sysfs_notify_dirent_safe(bitmap->sysfs_can_clear);
1451                 }
1452
1453                 if (!success && !NEEDED(*bmc))
1454                         *bmc |= NEEDED_MASK;
1455
1456                 if (COUNTER(*bmc) == COUNTER_MAX)
1457                         wake_up(&bitmap->overflow_wait);
1458
1459                 (*bmc)--;
1460                 if (*bmc <= 2) {
1461                         bitmap_set_pending(&bitmap->counts, offset);
1462                         bitmap->allclean = 0;
1463                 }
1464                 spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1465                 offset += blocks;
1466                 if (sectors > blocks)
1467                         sectors -= blocks;
1468                 else
1469                         sectors = 0;
1470         }
1471 }
1472 EXPORT_SYMBOL(bitmap_endwrite);
1473
1474 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1475                                int degraded)
1476 {
1477         bitmap_counter_t *bmc;
1478         int rv;
1479         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1480                 *blocks = 1024;
1481                 return 1; /* always resync if no bitmap */
1482         }
1483         spin_lock_irq(&bitmap->counts.lock);
1484         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1485         rv = 0;
1486         if (bmc) {
1487                 /* locked */
1488                 if (RESYNC(*bmc))
1489                         rv = 1;
1490                 else if (NEEDED(*bmc)) {
1491                         rv = 1;
1492                         if (!degraded) { /* don't set/clear bits if degraded */
1493                                 *bmc |= RESYNC_MASK;
1494                                 *bmc &= ~NEEDED_MASK;
1495                         }
1496                 }
1497         }
1498         spin_unlock_irq(&bitmap->counts.lock);
1499         return rv;
1500 }
1501
1502 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks,
1503                       int degraded)
1504 {
1505         /* bitmap_start_sync must always report on multiples of whole
1506          * pages, otherwise resync (which is very PAGE_SIZE based) will
1507          * get confused.
1508          * So call __bitmap_start_sync repeatedly (if needed) until
1509          * At least PAGE_SIZE>>9 blocks are covered.
1510          * Return the 'or' of the result.
1511          */
1512         int rv = 0;
1513         sector_t blocks1;
1514
1515         *blocks = 0;
1516         while (*blocks < (PAGE_SIZE>>9)) {
1517                 rv |= __bitmap_start_sync(bitmap, offset,
1518                                           &blocks1, degraded);
1519                 offset += blocks1;
1520                 *blocks += blocks1;
1521         }
1522         return rv;
1523 }
1524 EXPORT_SYMBOL(bitmap_start_sync);
1525
1526 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, sector_t *blocks, int aborted)
1527 {
1528         bitmap_counter_t *bmc;
1529         unsigned long flags;
1530
1531         if (bitmap == NULL) {
1532                 *blocks = 1024;
1533                 return;
1534         }
1535         spin_lock_irqsave(&bitmap->counts.lock, flags);
1536         bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
1537         if (bmc == NULL)
1538                 goto unlock;
1539         /* locked */
1540         if (RESYNC(*bmc)) {
1541                 *bmc &= ~RESYNC_MASK;
1542
1543                 if (!NEEDED(*bmc) && aborted)
1544                         *bmc |= NEEDED_MASK;
1545                 else {
1546                         if (*bmc <= 2) {
1547                                 bitmap_set_pending(&bitmap->counts, offset);
1548                                 bitmap->allclean = 0;
1549                         }
1550                 }
1551         }
1552  unlock:
1553         spin_unlock_irqrestore(&bitmap->counts.lock, flags);
1554 }
1555 EXPORT_SYMBOL(bitmap_end_sync);
1556
1557 void bitmap_close_sync(struct bitmap *bitmap)
1558 {
1559         /* Sync has finished, and any bitmap chunks that weren't synced
1560          * properly have been aborted.  It remains to us to clear the
1561          * RESYNC bit wherever it is still on
1562          */
1563         sector_t sector = 0;
1564         sector_t blocks;
1565         if (!bitmap)
1566                 return;
1567         while (sector < bitmap->mddev->resync_max_sectors) {
1568                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1569                 sector += blocks;
1570         }
1571 }
1572 EXPORT_SYMBOL(bitmap_close_sync);
1573
1574 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1575 {
1576         sector_t s = 0;
1577         sector_t blocks;
1578
1579         if (!bitmap)
1580                 return;
1581         if (sector == 0) {
1582                 bitmap->last_end_sync = jiffies;
1583                 return;
1584         }
1585         if (time_before(jiffies, (bitmap->last_end_sync
1586                                   + bitmap->mddev->bitmap_info.daemon_sleep)))
1587                 return;
1588         wait_event(bitmap->mddev->recovery_wait,
1589                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1590
1591         bitmap->mddev->curr_resync_completed = sector;
1592         set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
1593         sector &= ~((1ULL << bitmap->counts.chunkshift) - 1);
1594         s = 0;
1595         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1596                 bitmap_end_sync(bitmap, s, &blocks, 0);
1597                 s += blocks;
1598         }
1599         bitmap->last_end_sync = jiffies;
1600         sysfs_notify(&bitmap->mddev->kobj, NULL, "sync_completed");
1601 }
1602 EXPORT_SYMBOL(bitmap_cond_end_sync);
1603
1604 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1605 {
1606         /* For each chunk covered by any of these sectors, set the
1607          * counter to 2 and possibly set resync_needed.  They should all
1608          * be 0 at this point
1609          */
1610
1611         sector_t secs;
1612         bitmap_counter_t *bmc;
1613         spin_lock_irq(&bitmap->counts.lock);
1614         bmc = bitmap_get_counter(&bitmap->counts, offset, &secs, 1);
1615         if (!bmc) {
1616                 spin_unlock_irq(&bitmap->counts.lock);
1617                 return;
1618         }
1619         if (!*bmc) {
1620                 *bmc = 2;
1621                 bitmap_count_page(&bitmap->counts, offset, 1);
1622                 bitmap_set_pending(&bitmap->counts, offset);
1623                 bitmap->allclean = 0;
1624         }
1625         if (needed)
1626                 *bmc |= NEEDED_MASK;
1627         spin_unlock_irq(&bitmap->counts.lock);
1628 }
1629
1630 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1631 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1632 {
1633         unsigned long chunk;
1634
1635         for (chunk = s; chunk <= e; chunk++) {
1636                 sector_t sec = (sector_t)chunk << bitmap->counts.chunkshift;
1637                 bitmap_set_memory_bits(bitmap, sec, 1);
1638                 bitmap_file_set_bit(bitmap, sec);
1639                 if (sec < bitmap->mddev->recovery_cp)
1640                         /* We are asserting that the array is dirty,
1641                          * so move the recovery_cp address back so
1642                          * that it is obvious that it is dirty
1643                          */
1644                         bitmap->mddev->recovery_cp = sec;
1645         }
1646 }
1647
1648 /*
1649  * flush out any pending updates
1650  */
1651 void bitmap_flush(struct mddev *mddev)
1652 {
1653         struct bitmap *bitmap = mddev->bitmap;
1654         long sleep;
1655
1656         if (!bitmap) /* there was no bitmap */
1657                 return;
1658
1659         /* run the daemon_work three time to ensure everything is flushed
1660          * that can be
1661          */
1662         sleep = mddev->bitmap_info.daemon_sleep * 2;
1663         bitmap->daemon_lastrun -= sleep;
1664         bitmap_daemon_work(mddev);
1665         bitmap->daemon_lastrun -= sleep;
1666         bitmap_daemon_work(mddev);
1667         bitmap->daemon_lastrun -= sleep;
1668         bitmap_daemon_work(mddev);
1669         bitmap_update_sb(bitmap);
1670 }
1671
1672 /*
1673  * free memory that was allocated
1674  */
1675 static void bitmap_free(struct bitmap *bitmap)
1676 {
1677         unsigned long k, pages;
1678         struct bitmap_page *bp;
1679
1680         if (!bitmap) /* there was no bitmap */
1681                 return;
1682
1683         if (mddev_is_clustered(bitmap->mddev) && bitmap->mddev->cluster_info &&
1684                 bitmap->cluster_slot == md_cluster_ops->slot_number(bitmap->mddev))
1685                 md_cluster_stop(bitmap->mddev);
1686
1687         /* Shouldn't be needed - but just in case.... */
1688         wait_event(bitmap->write_wait,
1689                    atomic_read(&bitmap->pending_writes) == 0);
1690
1691         /* release the bitmap file  */
1692         bitmap_file_unmap(&bitmap->storage);
1693
1694         bp = bitmap->counts.bp;
1695         pages = bitmap->counts.pages;
1696
1697         /* free all allocated memory */
1698
1699         if (bp) /* deallocate the page memory */
1700                 for (k = 0; k < pages; k++)
1701                         if (bp[k].map && !bp[k].hijacked)
1702                                 kfree(bp[k].map);
1703         kfree(bp);
1704         kfree(bitmap);
1705 }
1706
1707 void bitmap_destroy(struct mddev *mddev)
1708 {
1709         struct bitmap *bitmap = mddev->bitmap;
1710
1711         if (!bitmap) /* there was no bitmap */
1712                 return;
1713
1714         mutex_lock(&mddev->bitmap_info.mutex);
1715         spin_lock(&mddev->lock);
1716         mddev->bitmap = NULL; /* disconnect from the md device */
1717         spin_unlock(&mddev->lock);
1718         mutex_unlock(&mddev->bitmap_info.mutex);
1719         if (mddev->thread)
1720                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1721
1722         if (bitmap->sysfs_can_clear)
1723                 sysfs_put(bitmap->sysfs_can_clear);
1724
1725         bitmap_free(bitmap);
1726 }
1727
1728 /*
1729  * initialize the bitmap structure
1730  * if this returns an error, bitmap_destroy must be called to do clean up
1731  */
1732 struct bitmap *bitmap_create(struct mddev *mddev, int slot)
1733 {
1734         struct bitmap *bitmap;
1735         sector_t blocks = mddev->resync_max_sectors;
1736         struct file *file = mddev->bitmap_info.file;
1737         int err;
1738         struct kernfs_node *bm = NULL;
1739
1740         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1741
1742         BUG_ON(file && mddev->bitmap_info.offset);
1743
1744         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1745         if (!bitmap)
1746                 return ERR_PTR(-ENOMEM);
1747
1748         spin_lock_init(&bitmap->counts.lock);
1749         atomic_set(&bitmap->pending_writes, 0);
1750         init_waitqueue_head(&bitmap->write_wait);
1751         init_waitqueue_head(&bitmap->overflow_wait);
1752         init_waitqueue_head(&bitmap->behind_wait);
1753
1754         bitmap->mddev = mddev;
1755         bitmap->cluster_slot = slot;
1756
1757         if (mddev->kobj.sd)
1758                 bm = sysfs_get_dirent(mddev->kobj.sd, "bitmap");
1759         if (bm) {
1760                 bitmap->sysfs_can_clear = sysfs_get_dirent(bm, "can_clear");
1761                 sysfs_put(bm);
1762         } else
1763                 bitmap->sysfs_can_clear = NULL;
1764
1765         bitmap->storage.file = file;
1766         if (file) {
1767                 get_file(file);
1768                 /* As future accesses to this file will use bmap,
1769                  * and bypass the page cache, we must sync the file
1770                  * first.
1771                  */
1772                 vfs_fsync(file, 1);
1773         }
1774         /* read superblock from bitmap file (this sets mddev->bitmap_info.chunksize) */
1775         if (!mddev->bitmap_info.external) {
1776                 /*
1777                  * If 'MD_ARRAY_FIRST_USE' is set, then device-mapper is
1778                  * instructing us to create a new on-disk bitmap instance.
1779                  */
1780                 if (test_and_clear_bit(MD_ARRAY_FIRST_USE, &mddev->flags))
1781                         err = bitmap_new_disk_sb(bitmap);
1782                 else
1783                         err = bitmap_read_sb(bitmap);
1784         } else {
1785                 err = 0;
1786                 if (mddev->bitmap_info.chunksize == 0 ||
1787                     mddev->bitmap_info.daemon_sleep == 0)
1788                         /* chunksize and time_base need to be
1789                          * set first. */
1790                         err = -EINVAL;
1791         }
1792         if (err)
1793                 goto error;
1794
1795         bitmap->daemon_lastrun = jiffies;
1796         err = bitmap_resize(bitmap, blocks, mddev->bitmap_info.chunksize, 1);
1797         if (err)
1798                 goto error;
1799
1800         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1801                bitmap->counts.pages, bmname(bitmap));
1802
1803         err = test_bit(BITMAP_WRITE_ERROR, &bitmap->flags) ? -EIO : 0;
1804         if (err)
1805                 goto error;
1806
1807         return bitmap;
1808  error:
1809         bitmap_free(bitmap);
1810         return ERR_PTR(err);
1811 }
1812
1813 int bitmap_load(struct mddev *mddev)
1814 {
1815         int err = 0;
1816         sector_t start = 0;
1817         sector_t sector = 0;
1818         struct bitmap *bitmap = mddev->bitmap;
1819
1820         if (!bitmap)
1821                 goto out;
1822
1823         /* Clear out old bitmap info first:  Either there is none, or we
1824          * are resuming after someone else has possibly changed things,
1825          * so we should forget old cached info.
1826          * All chunks should be clean, but some might need_sync.
1827          */
1828         while (sector < mddev->resync_max_sectors) {
1829                 sector_t blocks;
1830                 bitmap_start_sync(bitmap, sector, &blocks, 0);
1831                 sector += blocks;
1832         }
1833         bitmap_close_sync(bitmap);
1834
1835         if (mddev->degraded == 0
1836             || bitmap->events_cleared == mddev->events)
1837                 /* no need to keep dirty bits to optimise a
1838                  * re-add of a missing device */
1839                 start = mddev->recovery_cp;
1840
1841         mutex_lock(&mddev->bitmap_info.mutex);
1842         err = bitmap_init_from_disk(bitmap, start);
1843         mutex_unlock(&mddev->bitmap_info.mutex);
1844
1845         if (err)
1846                 goto out;
1847         clear_bit(BITMAP_STALE, &bitmap->flags);
1848
1849         /* Kick recovery in case any bits were set */
1850         set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1851
1852         mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
1853         md_wakeup_thread(mddev->thread);
1854
1855         bitmap_update_sb(bitmap);
1856
1857         if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
1858                 err = -EIO;
1859 out:
1860         return err;
1861 }
1862 EXPORT_SYMBOL_GPL(bitmap_load);
1863
1864 /* Loads the bitmap associated with slot and copies the resync information
1865  * to our bitmap
1866  */
1867 int bitmap_copy_from_slot(struct mddev *mddev, int slot,
1868                 sector_t *low, sector_t *high, bool clear_bits)
1869 {
1870         int rv = 0, i, j;
1871         sector_t block, lo = 0, hi = 0;
1872         struct bitmap_counts *counts;
1873         struct bitmap *bitmap = bitmap_create(mddev, slot);
1874
1875         if (IS_ERR(bitmap))
1876                 return PTR_ERR(bitmap);
1877
1878         rv = bitmap_read_sb(bitmap);
1879         if (rv)
1880                 goto err;
1881
1882         rv = bitmap_init_from_disk(bitmap, 0);
1883         if (rv)
1884                 goto err;
1885
1886         counts = &bitmap->counts;
1887         for (j = 0; j < counts->chunks; j++) {
1888                 block = (sector_t)j << counts->chunkshift;
1889                 if (bitmap_file_test_bit(bitmap, block)) {
1890                         if (!lo)
1891                                 lo = block;
1892                         hi = block;
1893                         bitmap_file_clear_bit(bitmap, block);
1894                         bitmap_set_memory_bits(mddev->bitmap, block, 1);
1895                         bitmap_file_set_bit(mddev->bitmap, block);
1896                 }
1897         }
1898
1899         if (clear_bits) {
1900                 bitmap_update_sb(bitmap);
1901                 /* Setting this for the ev_page should be enough.
1902                  * And we do not require both write_all and PAGE_DIRT either
1903                  */
1904                 for (i = 0; i < bitmap->storage.file_pages; i++)
1905                         set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
1906                 bitmap_write_all(bitmap);
1907                 bitmap_unplug(bitmap);
1908         }
1909         *low = lo;
1910         *high = hi;
1911 err:
1912         bitmap_free(bitmap);
1913         return rv;
1914 }
1915 EXPORT_SYMBOL_GPL(bitmap_copy_from_slot);
1916
1917
1918 void bitmap_status(struct seq_file *seq, struct bitmap *bitmap)
1919 {
1920         unsigned long chunk_kb;
1921         struct bitmap_counts *counts;
1922
1923         if (!bitmap)
1924                 return;
1925
1926         counts = &bitmap->counts;
1927
1928         chunk_kb = bitmap->mddev->bitmap_info.chunksize >> 10;
1929         seq_printf(seq, "bitmap: %lu/%lu pages [%luKB], "
1930                    "%lu%s chunk",
1931                    counts->pages - counts->missing_pages,
1932                    counts->pages,
1933                    (counts->pages - counts->missing_pages)
1934                    << (PAGE_SHIFT - 10),
1935                    chunk_kb ? chunk_kb : bitmap->mddev->bitmap_info.chunksize,
1936                    chunk_kb ? "KB" : "B");
1937         if (bitmap->storage.file) {
1938                 seq_printf(seq, ", file: ");
1939                 seq_path(seq, &bitmap->storage.file->f_path, " \t\n");
1940         }
1941
1942         seq_printf(seq, "\n");
1943 }
1944
1945 int bitmap_resize(struct bitmap *bitmap, sector_t blocks,
1946                   int chunksize, int init)
1947 {
1948         /* If chunk_size is 0, choose an appropriate chunk size.
1949          * Then possibly allocate new storage space.
1950          * Then quiesce, copy bits, replace bitmap, and re-start
1951          *
1952          * This function is called both to set up the initial bitmap
1953          * and to resize the bitmap while the array is active.
1954          * If this happens as a result of the array being resized,
1955          * chunksize will be zero, and we need to choose a suitable
1956          * chunksize, otherwise we use what we are given.
1957          */
1958         struct bitmap_storage store;
1959         struct bitmap_counts old_counts;
1960         unsigned long chunks;
1961         sector_t block;
1962         sector_t old_blocks, new_blocks;
1963         int chunkshift;
1964         int ret = 0;
1965         long pages;
1966         struct bitmap_page *new_bp;
1967
1968         if (chunksize == 0) {
1969                 /* If there is enough space, leave the chunk size unchanged,
1970                  * else increase by factor of two until there is enough space.
1971                  */
1972                 long bytes;
1973                 long space = bitmap->mddev->bitmap_info.space;
1974
1975                 if (space == 0) {
1976                         /* We don't know how much space there is, so limit
1977                          * to current size - in sectors.
1978                          */
1979                         bytes = DIV_ROUND_UP(bitmap->counts.chunks, 8);
1980                         if (!bitmap->mddev->bitmap_info.external)
1981                                 bytes += sizeof(bitmap_super_t);
1982                         space = DIV_ROUND_UP(bytes, 512);
1983                         bitmap->mddev->bitmap_info.space = space;
1984                 }
1985                 chunkshift = bitmap->counts.chunkshift;
1986                 chunkshift--;
1987                 do {
1988                         /* 'chunkshift' is shift from block size to chunk size */
1989                         chunkshift++;
1990                         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1991                         bytes = DIV_ROUND_UP(chunks, 8);
1992                         if (!bitmap->mddev->bitmap_info.external)
1993                                 bytes += sizeof(bitmap_super_t);
1994                 } while (bytes > (space << 9));
1995         } else
1996                 chunkshift = ffz(~chunksize) - BITMAP_BLOCK_SHIFT;
1997
1998         chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
1999         memset(&store, 0, sizeof(store));
2000         if (bitmap->mddev->bitmap_info.offset || bitmap->mddev->bitmap_info.file)
2001                 ret = bitmap_storage_alloc(&store, chunks,
2002                                            !bitmap->mddev->bitmap_info.external,
2003                                            bitmap->cluster_slot);
2004         if (ret)
2005                 goto err;
2006
2007         pages = DIV_ROUND_UP(chunks, PAGE_COUNTER_RATIO);
2008
2009         new_bp = kzalloc(pages * sizeof(*new_bp), GFP_KERNEL);
2010         ret = -ENOMEM;
2011         if (!new_bp) {
2012                 bitmap_file_unmap(&store);
2013                 goto err;
2014         }
2015
2016         if (!init)
2017                 bitmap->mddev->pers->quiesce(bitmap->mddev, 1);
2018
2019         store.file = bitmap->storage.file;
2020         bitmap->storage.file = NULL;
2021
2022         if (store.sb_page && bitmap->storage.sb_page)
2023                 memcpy(page_address(store.sb_page),
2024                        page_address(bitmap->storage.sb_page),
2025                        sizeof(bitmap_super_t));
2026         bitmap_file_unmap(&bitmap->storage);
2027         bitmap->storage = store;
2028
2029         old_counts = bitmap->counts;
2030         bitmap->counts.bp = new_bp;
2031         bitmap->counts.pages = pages;
2032         bitmap->counts.missing_pages = pages;
2033         bitmap->counts.chunkshift = chunkshift;
2034         bitmap->counts.chunks = chunks;
2035         bitmap->mddev->bitmap_info.chunksize = 1 << (chunkshift +
2036                                                      BITMAP_BLOCK_SHIFT);
2037
2038         blocks = min(old_counts.chunks << old_counts.chunkshift,
2039                      chunks << chunkshift);
2040
2041         spin_lock_irq(&bitmap->counts.lock);
2042         for (block = 0; block < blocks; ) {
2043                 bitmap_counter_t *bmc_old, *bmc_new;
2044                 int set;
2045
2046                 bmc_old = bitmap_get_counter(&old_counts, block,
2047                                              &old_blocks, 0);
2048                 set = bmc_old && NEEDED(*bmc_old);
2049
2050                 if (set) {
2051                         bmc_new = bitmap_get_counter(&bitmap->counts, block,
2052                                                      &new_blocks, 1);
2053                         if (*bmc_new == 0) {
2054                                 /* need to set on-disk bits too. */
2055                                 sector_t end = block + new_blocks;
2056                                 sector_t start = block >> chunkshift;
2057                                 start <<= chunkshift;
2058                                 while (start < end) {
2059                                         bitmap_file_set_bit(bitmap, block);
2060                                         start += 1 << chunkshift;
2061                                 }
2062                                 *bmc_new = 2;
2063                                 bitmap_count_page(&bitmap->counts,
2064                                                   block, 1);
2065                                 bitmap_set_pending(&bitmap->counts,
2066                                                    block);
2067                         }
2068                         *bmc_new |= NEEDED_MASK;
2069                         if (new_blocks < old_blocks)
2070                                 old_blocks = new_blocks;
2071                 }
2072                 block += old_blocks;
2073         }
2074
2075         if (!init) {
2076                 int i;
2077                 while (block < (chunks << chunkshift)) {
2078                         bitmap_counter_t *bmc;
2079                         bmc = bitmap_get_counter(&bitmap->counts, block,
2080                                                  &new_blocks, 1);
2081                         if (bmc) {
2082                                 /* new space.  It needs to be resynced, so
2083                                  * we set NEEDED_MASK.
2084                                  */
2085                                 if (*bmc == 0) {
2086                                         *bmc = NEEDED_MASK | 2;
2087                                         bitmap_count_page(&bitmap->counts,
2088                                                           block, 1);
2089                                         bitmap_set_pending(&bitmap->counts,
2090                                                            block);
2091                                 }
2092                         }
2093                         block += new_blocks;
2094                 }
2095                 for (i = 0; i < bitmap->storage.file_pages; i++)
2096                         set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
2097         }
2098         spin_unlock_irq(&bitmap->counts.lock);
2099
2100         if (!init) {
2101                 bitmap_unplug(bitmap);
2102                 bitmap->mddev->pers->quiesce(bitmap->mddev, 0);
2103         }
2104         ret = 0;
2105 err:
2106         return ret;
2107 }
2108 EXPORT_SYMBOL_GPL(bitmap_resize);
2109
2110 static ssize_t
2111 location_show(struct mddev *mddev, char *page)
2112 {
2113         ssize_t len;
2114         if (mddev->bitmap_info.file)
2115                 len = sprintf(page, "file");
2116         else if (mddev->bitmap_info.offset)
2117                 len = sprintf(page, "%+lld", (long long)mddev->bitmap_info.offset);
2118         else
2119                 len = sprintf(page, "none");
2120         len += sprintf(page+len, "\n");
2121         return len;
2122 }
2123
2124 static ssize_t
2125 location_store(struct mddev *mddev, const char *buf, size_t len)
2126 {
2127
2128         if (mddev->pers) {
2129                 if (!mddev->pers->quiesce)
2130                         return -EBUSY;
2131                 if (mddev->recovery || mddev->sync_thread)
2132                         return -EBUSY;
2133         }
2134
2135         if (mddev->bitmap || mddev->bitmap_info.file ||
2136             mddev->bitmap_info.offset) {
2137                 /* bitmap already configured.  Only option is to clear it */
2138                 if (strncmp(buf, "none", 4) != 0)
2139                         return -EBUSY;
2140                 if (mddev->pers) {
2141                         mddev->pers->quiesce(mddev, 1);
2142                         bitmap_destroy(mddev);
2143                         mddev->pers->quiesce(mddev, 0);
2144                 }
2145                 mddev->bitmap_info.offset = 0;
2146                 if (mddev->bitmap_info.file) {
2147                         struct file *f = mddev->bitmap_info.file;
2148                         mddev->bitmap_info.file = NULL;
2149                         fput(f);
2150                 }
2151         } else {
2152                 /* No bitmap, OK to set a location */
2153                 long long offset;
2154                 if (strncmp(buf, "none", 4) == 0)
2155                         /* nothing to be done */;
2156                 else if (strncmp(buf, "file:", 5) == 0) {
2157                         /* Not supported yet */
2158                         return -EINVAL;
2159                 } else {
2160                         int rv;
2161                         if (buf[0] == '+')
2162                                 rv = kstrtoll(buf+1, 10, &offset);
2163                         else
2164                                 rv = kstrtoll(buf, 10, &offset);
2165                         if (rv)
2166                                 return rv;
2167                         if (offset == 0)
2168                                 return -EINVAL;
2169                         if (mddev->bitmap_info.external == 0 &&
2170                             mddev->major_version == 0 &&
2171                             offset != mddev->bitmap_info.default_offset)
2172                                 return -EINVAL;
2173                         mddev->bitmap_info.offset = offset;
2174                         if (mddev->pers) {
2175                                 struct bitmap *bitmap;
2176                                 mddev->pers->quiesce(mddev, 1);
2177                                 bitmap = bitmap_create(mddev, -1);
2178                                 if (IS_ERR(bitmap))
2179                                         rv = PTR_ERR(bitmap);
2180                                 else {
2181                                         mddev->bitmap = bitmap;
2182                                         rv = bitmap_load(mddev);
2183                                         if (rv) {
2184                                                 bitmap_destroy(mddev);
2185                                                 mddev->bitmap_info.offset = 0;
2186                                         }
2187                                 }
2188                                 mddev->pers->quiesce(mddev, 0);
2189                                 if (rv)
2190                                         return rv;
2191                         }
2192                 }
2193         }
2194         if (!mddev->external) {
2195                 /* Ensure new bitmap info is stored in
2196                  * metadata promptly.
2197                  */
2198                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2199                 md_wakeup_thread(mddev->thread);
2200         }
2201         return len;
2202 }
2203
2204 static struct md_sysfs_entry bitmap_location =
2205 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
2206
2207 /* 'bitmap/space' is the space available at 'location' for the
2208  * bitmap.  This allows the kernel to know when it is safe to
2209  * resize the bitmap to match a resized array.
2210  */
2211 static ssize_t
2212 space_show(struct mddev *mddev, char *page)
2213 {
2214         return sprintf(page, "%lu\n", mddev->bitmap_info.space);
2215 }
2216
2217 static ssize_t
2218 space_store(struct mddev *mddev, const char *buf, size_t len)
2219 {
2220         unsigned long sectors;
2221         int rv;
2222
2223         rv = kstrtoul(buf, 10, &sectors);
2224         if (rv)
2225                 return rv;
2226
2227         if (sectors == 0)
2228                 return -EINVAL;
2229
2230         if (mddev->bitmap &&
2231             sectors < (mddev->bitmap->storage.bytes + 511) >> 9)
2232                 return -EFBIG; /* Bitmap is too big for this small space */
2233
2234         /* could make sure it isn't too big, but that isn't really
2235          * needed - user-space should be careful.
2236          */
2237         mddev->bitmap_info.space = sectors;
2238         return len;
2239 }
2240
2241 static struct md_sysfs_entry bitmap_space =
2242 __ATTR(space, S_IRUGO|S_IWUSR, space_show, space_store);
2243
2244 static ssize_t
2245 timeout_show(struct mddev *mddev, char *page)
2246 {
2247         ssize_t len;
2248         unsigned long secs = mddev->bitmap_info.daemon_sleep / HZ;
2249         unsigned long jifs = mddev->bitmap_info.daemon_sleep % HZ;
2250
2251         len = sprintf(page, "%lu", secs);
2252         if (jifs)
2253                 len += sprintf(page+len, ".%03u", jiffies_to_msecs(jifs));
2254         len += sprintf(page+len, "\n");
2255         return len;
2256 }
2257
2258 static ssize_t
2259 timeout_store(struct mddev *mddev, const char *buf, size_t len)
2260 {
2261         /* timeout can be set at any time */
2262         unsigned long timeout;
2263         int rv = strict_strtoul_scaled(buf, &timeout, 4);
2264         if (rv)
2265                 return rv;
2266
2267         /* just to make sure we don't overflow... */
2268         if (timeout >= LONG_MAX / HZ)
2269                 return -EINVAL;
2270
2271         timeout = timeout * HZ / 10000;
2272
2273         if (timeout >= MAX_SCHEDULE_TIMEOUT)
2274                 timeout = MAX_SCHEDULE_TIMEOUT-1;
2275         if (timeout < 1)
2276                 timeout = 1;
2277         mddev->bitmap_info.daemon_sleep = timeout;
2278         if (mddev->thread) {
2279                 /* if thread->timeout is MAX_SCHEDULE_TIMEOUT, then
2280                  * the bitmap is all clean and we don't need to
2281                  * adjust the timeout right now
2282                  */
2283                 if (mddev->thread->timeout < MAX_SCHEDULE_TIMEOUT) {
2284                         mddev->thread->timeout = timeout;
2285                         md_wakeup_thread(mddev->thread);
2286                 }
2287         }
2288         return len;
2289 }
2290
2291 static struct md_sysfs_entry bitmap_timeout =
2292 __ATTR(time_base, S_IRUGO|S_IWUSR, timeout_show, timeout_store);
2293
2294 static ssize_t
2295 backlog_show(struct mddev *mddev, char *page)
2296 {
2297         return sprintf(page, "%lu\n", mddev->bitmap_info.max_write_behind);
2298 }
2299
2300 static ssize_t
2301 backlog_store(struct mddev *mddev, const char *buf, size_t len)
2302 {
2303         unsigned long backlog;
2304         int rv = kstrtoul(buf, 10, &backlog);
2305         if (rv)
2306                 return rv;
2307         if (backlog > COUNTER_MAX)
2308                 return -EINVAL;
2309         mddev->bitmap_info.max_write_behind = backlog;
2310         return len;
2311 }
2312
2313 static struct md_sysfs_entry bitmap_backlog =
2314 __ATTR(backlog, S_IRUGO|S_IWUSR, backlog_show, backlog_store);
2315
2316 static ssize_t
2317 chunksize_show(struct mddev *mddev, char *page)
2318 {
2319         return sprintf(page, "%lu\n", mddev->bitmap_info.chunksize);
2320 }
2321
2322 static ssize_t
2323 chunksize_store(struct mddev *mddev, const char *buf, size_t len)
2324 {
2325         /* Can only be changed when no bitmap is active */
2326         int rv;
2327         unsigned long csize;
2328         if (mddev->bitmap)
2329                 return -EBUSY;
2330         rv = kstrtoul(buf, 10, &csize);
2331         if (rv)
2332                 return rv;
2333         if (csize < 512 ||
2334             !is_power_of_2(csize))
2335                 return -EINVAL;
2336         mddev->bitmap_info.chunksize = csize;
2337         return len;
2338 }
2339
2340 static struct md_sysfs_entry bitmap_chunksize =
2341 __ATTR(chunksize, S_IRUGO|S_IWUSR, chunksize_show, chunksize_store);
2342
2343 static ssize_t metadata_show(struct mddev *mddev, char *page)
2344 {
2345         if (mddev_is_clustered(mddev))
2346                 return sprintf(page, "clustered\n");
2347         return sprintf(page, "%s\n", (mddev->bitmap_info.external
2348                                       ? "external" : "internal"));
2349 }
2350
2351 static ssize_t metadata_store(struct mddev *mddev, const char *buf, size_t len)
2352 {
2353         if (mddev->bitmap ||
2354             mddev->bitmap_info.file ||
2355             mddev->bitmap_info.offset)
2356                 return -EBUSY;
2357         if (strncmp(buf, "external", 8) == 0)
2358                 mddev->bitmap_info.external = 1;
2359         else if ((strncmp(buf, "internal", 8) == 0) ||
2360                         (strncmp(buf, "clustered", 9) == 0))
2361                 mddev->bitmap_info.external = 0;
2362         else
2363                 return -EINVAL;
2364         return len;
2365 }
2366
2367 static struct md_sysfs_entry bitmap_metadata =
2368 __ATTR(metadata, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
2369
2370 static ssize_t can_clear_show(struct mddev *mddev, char *page)
2371 {
2372         int len;
2373         spin_lock(&mddev->lock);
2374         if (mddev->bitmap)
2375                 len = sprintf(page, "%s\n", (mddev->bitmap->need_sync ?
2376                                              "false" : "true"));
2377         else
2378                 len = sprintf(page, "\n");
2379         spin_unlock(&mddev->lock);
2380         return len;
2381 }
2382
2383 static ssize_t can_clear_store(struct mddev *mddev, const char *buf, size_t len)
2384 {
2385         if (mddev->bitmap == NULL)
2386                 return -ENOENT;
2387         if (strncmp(buf, "false", 5) == 0)
2388                 mddev->bitmap->need_sync = 1;
2389         else if (strncmp(buf, "true", 4) == 0) {
2390                 if (mddev->degraded)
2391                         return -EBUSY;
2392                 mddev->bitmap->need_sync = 0;
2393         } else
2394                 return -EINVAL;
2395         return len;
2396 }
2397
2398 static struct md_sysfs_entry bitmap_can_clear =
2399 __ATTR(can_clear, S_IRUGO|S_IWUSR, can_clear_show, can_clear_store);
2400
2401 static ssize_t
2402 behind_writes_used_show(struct mddev *mddev, char *page)
2403 {
2404         ssize_t ret;
2405         spin_lock(&mddev->lock);
2406         if (mddev->bitmap == NULL)
2407                 ret = sprintf(page, "0\n");
2408         else
2409                 ret = sprintf(page, "%lu\n",
2410                               mddev->bitmap->behind_writes_used);
2411         spin_unlock(&mddev->lock);
2412         return ret;
2413 }
2414
2415 static ssize_t
2416 behind_writes_used_reset(struct mddev *mddev, const char *buf, size_t len)
2417 {
2418         if (mddev->bitmap)
2419                 mddev->bitmap->behind_writes_used = 0;
2420         return len;
2421 }
2422
2423 static struct md_sysfs_entry max_backlog_used =
2424 __ATTR(max_backlog_used, S_IRUGO | S_IWUSR,
2425        behind_writes_used_show, behind_writes_used_reset);
2426
2427 static struct attribute *md_bitmap_attrs[] = {
2428         &bitmap_location.attr,
2429         &bitmap_space.attr,
2430         &bitmap_timeout.attr,
2431         &bitmap_backlog.attr,
2432         &bitmap_chunksize.attr,
2433         &bitmap_metadata.attr,
2434         &bitmap_can_clear.attr,
2435         &max_backlog_used.attr,
2436         NULL
2437 };
2438 struct attribute_group md_bitmap_group = {
2439         .name = "bitmap",
2440         .attrs = md_bitmap_attrs,
2441 };
2442