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 "md.h"
  30 #include "bitmap.h"
  31
  32 /* debug macros */
  33
  34 #define DEBUG 0
  35
  36 #if DEBUG
  37 /* these are for debugging purposes only! */
  38
  39 /* define one and only one of these */
  40 #define INJECT_FAULTS_1 0 /* cause bitmap_alloc_page to fail always */
  41 #define INJECT_FAULTS_2 0 /* cause bitmap file to be kicked when first bit set*/
  42 #define INJECT_FAULTS_3 0 /* treat bitmap file as kicked at init time */
  43 #define INJECT_FAULTS_4 0 /* undef */
  44 #define INJECT_FAULTS_5 0 /* undef */
  45 #define INJECT_FAULTS_6 0
  46
  47 /* if these are defined, the driver will fail! debug only */
  48 #define INJECT_FATAL_FAULT_1 0 /* fail kmalloc, causing bitmap_create to fail */
  49 #define INJECT_FATAL_FAULT_2 0 /* undef */
  50 #define INJECT_FATAL_FAULT_3 0 /* undef */
  51 #endif
  52
  53 #ifndef PRINTK
  54 #  if DEBUG > 0
  55 #    define PRINTK(x...) printk(KERN_DEBUG x)
  56 #  else
  57 #    define PRINTK(x...)
  58 #  endif
  59 #endif
  60
  61 static inline char *bmname(struct bitmap *bitmap)
  62 {
  63         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
  64 }
  65
  66 /*
  67  * just a placeholder - calls kmalloc for bitmap pages
  68  */
  69 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
  70 {
  71         unsigned char *page;
  72
  73 #ifdef INJECT_FAULTS_1
  74         page = NULL;
  75 #else
  76         page = kzalloc(PAGE_SIZE, GFP_NOIO);
  77 #endif
  78         if (!page)
  79                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
  80         else
  81                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
  82                         bmname(bitmap), page);
  83         return page;
  84 }
  85
  86 /*
  87  * for now just a placeholder -- just calls kfree for bitmap pages
  88  */
  89 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
  90 {
  91         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
  92         kfree(page);
  93 }
  94
  95 /*
  96  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
  97  *
  98  * 1) check to see if this page is allocated, if it's not then try to alloc
  99  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
 100  *    page pointer directly as a counter
 101  *
 102  * if we find our page, we increment the page's refcount so that it stays
 103  * allocated while we're using it
 104  */
 105 static int bitmap_checkpage(struct bitmap *bitmap,
 106                             unsigned long page, int create)
 107 __releases(bitmap->lock)
 108 __acquires(bitmap->lock)
 109 {
 110         unsigned char *mappage;
 111
 112         if (page >= bitmap->pages) {
 113                 /* This can happen if bitmap_start_sync goes beyond
 114                  * End-of-device while looking for a whole page.
 115                  * It is harmless.
 116                  */
 117                 return -EINVAL;
 118         }
 119
 120         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
 121                 return 0;
 122
 123         if (bitmap->bp[page].map) /* page is already allocated, just return */
 124                 return 0;
 125
 126         if (!create)
 127                 return -ENOENT;
 128
 129         /* this page has not been allocated yet */
 130
 131         spin_unlock_irq(&bitmap->lock);
 132         mappage = bitmap_alloc_page(bitmap);
 133         spin_lock_irq(&bitmap->lock);
 134
 135         if (mappage == NULL) {
 136                 PRINTK("%s: bitmap map page allocation failed, hijacking\n",
 137                         bmname(bitmap));
 138                 /* failed - set the hijacked flag so that we can use the
 139                  * pointer as a counter */
 140                 if (!bitmap->bp[page].map)
 141                         bitmap->bp[page].hijacked = 1;
 142         } else if (bitmap->bp[page].map ||
 143                    bitmap->bp[page].hijacked) {
 144                 /* somebody beat us to getting the page */
 145                 bitmap_free_page(bitmap, mappage);
 146                 return 0;
 147         } else {
 148
 149                 /* no page was in place and we have one, so install it */
 150
 151                 bitmap->bp[page].map = mappage;
 152                 bitmap->missing_pages--;
 153         }
 154         return 0;
 155 }
 156
 157 /* if page is completely empty, put it back on the free list, or dealloc it */
 158 /* if page was hijacked, unmark the flag so it might get alloced next time */
 159 /* Note: lock should be held when calling this */
 160 static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
 161 {
 162         char *ptr;
 163
 164         if (bitmap->bp[page].count) /* page is still busy */
 165                 return;
 166
 167         /* page is no longer in use, it can be released */
 168
 169         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
 170                 bitmap->bp[page].hijacked = 0;
 171                 bitmap->bp[page].map = NULL;
 172         } else {
 173                 /* normal case, free the page */
 174                 ptr = bitmap->bp[page].map;
 175                 bitmap->bp[page].map = NULL;
 176                 bitmap->missing_pages++;
 177                 bitmap_free_page(bitmap, ptr);
 178         }
 179 }
 180
 181 /*
 182  * bitmap file handling - read and write the bitmap file and its superblock
 183  */
 184
 185 /*
 186  * basic page I/O operations
 187  */
 188
 189 /* IO operations when bitmap is stored near all superblocks */
 190 static struct page *read_sb_page(mddev_t *mddev, loff_t offset,
 191                                  struct page *page,
 192                                  unsigned long index, int size)
 193 {
 194         /* choose a good rdev and read the page from there */
 195
 196         mdk_rdev_t *rdev;
 197         sector_t target;
 198         int did_alloc = 0;
 199
 200         if (!page) {
 201                 page = alloc_page(GFP_KERNEL);
 202                 if (!page)
 203                         return ERR_PTR(-ENOMEM);
 204                 did_alloc = 1;
 205         }
 206
 207         list_for_each_entry(rdev, &mddev->disks, same_set) {
 208                 if (! test_bit(In_sync, &rdev->flags)
 209                     || test_bit(Faulty, &rdev->flags))
 210                         continue;
 211
 212                 target = rdev->sb_start + offset + index * (PAGE_SIZE/512);
 213
 214                 if (sync_page_io(rdev->bdev, target,
 215                                  roundup(size, bdev_logical_block_size(rdev->bdev)),
 216                                  page, READ)) {
 217                         page->index = index;
 218                         attach_page_buffers(page, NULL); /* so that free_buffer will
 219                                                           * quietly no-op */
 220                         return page;
 221                 }
 222         }
 223         if (did_alloc)
 224                 put_page(page);
 225         return ERR_PTR(-EIO);
 226
 227 }
 228
 229 static mdk_rdev_t *next_active_rdev(mdk_rdev_t *rdev, mddev_t *mddev)
 230 {
 231         /* Iterate the disks of an mddev, using rcu to protect access to the
 232          * linked list, and raising the refcount of devices we return to ensure
 233          * they don't disappear while in use.
 234          * As devices are only added or removed when raid_disk is < 0 and
 235          * nr_pending is 0 and In_sync is clear, the entries we return will
 236          * still be in the same position on the list when we re-enter
 237          * list_for_each_continue_rcu.
 238          */
 239         struct list_head *pos;
 240         rcu_read_lock();
 241         if (rdev == NULL)
 242                 /* start at the beginning */
 243                 pos = &mddev->disks;
 244         else {
 245                 /* release the previous rdev and start from there. */
 246                 rdev_dec_pending(rdev, mddev);
 247                 pos = &rdev->same_set;
 248         }
 249         list_for_each_continue_rcu(pos, &mddev->disks) {
 250                 rdev = list_entry(pos, mdk_rdev_t, same_set);
 251                 if (rdev->raid_disk >= 0 &&
 252                     !test_bit(Faulty, &rdev->flags)) {
 253                         /* this is a usable devices */
 254                         atomic_inc(&rdev->nr_pending);
 255                         rcu_read_unlock();
 256                         return rdev;
 257                 }
 258         }
 259         rcu_read_unlock();
 260         return NULL;
 261 }
 262
 263 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 264 {
 265         mdk_rdev_t *rdev = NULL;
 266         mddev_t *mddev = bitmap->mddev;
 267
 268         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 269                 int size = PAGE_SIZE;
 270                 loff_t offset = mddev->bitmap_info.offset;
 271                 if (page->index == bitmap->file_pages-1)
 272                         size = roundup(bitmap->last_page_size,
 273                                        bdev_logical_block_size(rdev->bdev));
 274                 /* Just make sure we aren't corrupting data or
 275                  * metadata
 276                  */
 277                 if (mddev->external) {
 278                         /* Bitmap could be anywhere. */
 279                         if (rdev->sb_start + offset + (page->index
 280                                                        * (PAGE_SIZE/512))
 281                             > rdev->data_offset
 282                             &&
 283                             rdev->sb_start + offset
 284                             < (rdev->data_offset + mddev->dev_sectors
 285                              + (PAGE_SIZE/512)))
 286                                 goto bad_alignment;
 287                 } else if (offset < 0) {
 288                         /* DATA  BITMAP METADATA  */
 289                         if (offset
 290                             + (long)(page->index * (PAGE_SIZE/512))
 291                             + size/512 > 0)
 292                                 /* bitmap runs in to metadata */
 293                                 goto bad_alignment;
 294                         if (rdev->data_offset + mddev->dev_sectors
 295                             > rdev->sb_start + offset)
 296                                 /* data runs in to bitmap */
 297                                 goto bad_alignment;
 298                 } else if (rdev->sb_start < rdev->data_offset) {
 299                         /* METADATA BITMAP DATA */
 300                         if (rdev->sb_start
 301                             + offset
 302                             + page->index*(PAGE_SIZE/512) + size/512
 303                             > rdev->data_offset)
 304                                 /* bitmap runs in to data */
 305                                 goto bad_alignment;
 306                 } else {
 307                         /* DATA METADATA BITMAP - no problems */
 308                 }
 309                 md_super_write(mddev, rdev,
 310                                rdev->sb_start + offset
 311                                + page->index * (PAGE_SIZE/512),
 312                                size,
 313                                page);
 314         }
 315
 316         if (wait)
 317                 md_super_wait(mddev);
 318         return 0;
 319
 320  bad_alignment:
 321         return -EINVAL;
 322 }
 323
 324 static void bitmap_file_kick(struct bitmap *bitmap);
 325 /*
 326  * write out a page to a file
 327  */
 328 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
 329 {
 330         struct buffer_head *bh;
 331
 332         if (bitmap->file == NULL) {
 333                 switch (write_sb_page(bitmap, page, wait)) {
 334                 case -EINVAL:
 335                         bitmap->flags |= BITMAP_WRITE_ERROR;
 336                 }
 337         } else {
 338
 339                 bh = page_buffers(page);
 340
 341                 while (bh && bh->b_blocknr) {
 342                         atomic_inc(&bitmap->pending_writes);
 343                         set_buffer_locked(bh);
 344                         set_buffer_mapped(bh);
 345                         submit_bh(WRITE, bh);
 346                         bh = bh->b_this_page;
 347                 }
 348
 349                 if (wait)
 350                         wait_event(bitmap->write_wait,
 351                                    atomic_read(&bitmap->pending_writes)==0);
 352         }
 353         if (bitmap->flags & BITMAP_WRITE_ERROR)
 354                 bitmap_file_kick(bitmap);
 355 }
 356
 357 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 358 {
 359         struct bitmap *bitmap = bh->b_private;
 360         unsigned long flags;
 361
 362         if (!uptodate) {
 363                 spin_lock_irqsave(&bitmap->lock, flags);
 364                 bitmap->flags |= BITMAP_WRITE_ERROR;
 365                 spin_unlock_irqrestore(&bitmap->lock, flags);
 366         }
 367         if (atomic_dec_and_test(&bitmap->pending_writes))
 368                 wake_up(&bitmap->write_wait);
 369 }
 370
 371 /* copied from buffer.c */
 372 static void
 373 __clear_page_buffers(struct page *page)
 374 {
 375         ClearPagePrivate(page);
 376         set_page_private(page, 0);
 377         page_cache_release(page);
 378 }
 379 static void free_buffers(struct page *page)
 380 {
 381         struct buffer_head *bh = page_buffers(page);
 382
 383         while (bh) {
 384                 struct buffer_head *next = bh->b_this_page;
 385                 free_buffer_head(bh);
 386                 bh = next;
 387         }
 388         __clear_page_buffers(page);
 389         put_page(page);
 390 }
 391
 392 /* read a page from a file.
 393  * We both read the page, and attach buffers to the page to record the
 394  * address of each block (using bmap).  These addresses will be used
 395  * to write the block later, completely bypassing the filesystem.
 396  * This usage is similar to how swap files are handled, and allows us
 397  * to write to a file with no concerns of memory allocation failing.
 398  */
 399 static struct page *read_page(struct file *file, unsigned long index,
 400                               struct bitmap *bitmap,
 401                               unsigned long count)
 402 {
 403         struct page *page = NULL;
 404         struct inode *inode = file->f_path.dentry->d_inode;
 405         struct buffer_head *bh;
 406         sector_t block;
 407
 408         PRINTK("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 409                         (unsigned long long)index << PAGE_SHIFT);
 410
 411         page = alloc_page(GFP_KERNEL);
 412         if (!page)
 413                 page = ERR_PTR(-ENOMEM);
 414         if (IS_ERR(page))
 415                 goto out;
 416
 417         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 418         if (!bh) {
 419                 put_page(page);
 420                 page = ERR_PTR(-ENOMEM);
 421                 goto out;
 422         }
 423         attach_page_buffers(page, bh);
 424         block = index << (PAGE_SHIFT - inode->i_blkbits);
 425         while (bh) {
 426                 if (count == 0)
 427                         bh->b_blocknr = 0;
 428                 else {
 429                         bh->b_blocknr = bmap(inode, block);
 430                         if (bh->b_blocknr == 0) {
 431                                 /* Cannot use this file! */
 432                                 free_buffers(page);
 433                                 page = ERR_PTR(-EINVAL);
 434                                 goto out;
 435                         }
 436                         bh->b_bdev = inode->i_sb->s_bdev;
 437                         if (count < (1<<inode->i_blkbits))
 438                                 count = 0;
 439                         else
 440                                 count -= (1<<inode->i_blkbits);
 441
 442                         bh->b_end_io = end_bitmap_write;
 443                         bh->b_private = bitmap;
 444                         atomic_inc(&bitmap->pending_writes);
 445                         set_buffer_locked(bh);
 446                         set_buffer_mapped(bh);
 447                         submit_bh(READ, bh);
 448                 }
 449                 block++;
 450                 bh = bh->b_this_page;
 451         }
 452         page->index = index;
 453
 454         wait_event(bitmap->write_wait,
 455                    atomic_read(&bitmap->pending_writes)==0);
 456         if (bitmap->flags & BITMAP_WRITE_ERROR) {
 457                 free_buffers(page);
 458                 page = ERR_PTR(-EIO);
 459         }
 460 out:
 461         if (IS_ERR(page))
 462                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %ld\n",
 463                         (int)PAGE_SIZE,
 464                         (unsigned long long)index << PAGE_SHIFT,
 465                         PTR_ERR(page));
 466         return page;
 467 }
 468
 469 /*
 470  * bitmap file superblock operations
 471  */
 472
 473 /* update the event counter and sync the superblock to disk */
 474 void bitmap_update_sb(struct bitmap *bitmap)
 475 {
 476         bitmap_super_t *sb;
 477         unsigned long flags;
 478
 479         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
 480                 return;
 481         if (bitmap->mddev->bitmap_info.external)
 482                 return;
 483         spin_lock_irqsave(&bitmap->lock, flags);
 484         if (!bitmap->sb_page) { /* no superblock */
 485                 spin_unlock_irqrestore(&bitmap->lock, flags);
 486                 return;
 487         }
 488         spin_unlock_irqrestore(&bitmap->lock, flags);
 489         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 490         sb->events = cpu_to_le64(bitmap->mddev->events);
 491         if (bitmap->mddev->events < bitmap->events_cleared) {
 492                 /* rocking back to read-only */
 493                 bitmap->events_cleared = bitmap->mddev->events;
 494                 sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
 495         }
 496         /* Just in case these have been changed via sysfs: */
 497         sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 498         sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
 499         kunmap_atomic(sb, KM_USER0);
 500         write_page(bitmap, bitmap->sb_page, 1);
 501 }
 502
 503 /* print out the bitmap file superblock */
 504 void bitmap_print_sb(struct bitmap *bitmap)
 505 {
 506         bitmap_super_t *sb;
 507
 508         if (!bitmap || !bitmap->sb_page)
 509                 return;
 510         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 511         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 512         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 513         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
 514         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
 515                                         *(__u32 *)(sb->uuid+0),
 516                                         *(__u32 *)(sb->uuid+4),
 517                                         *(__u32 *)(sb->uuid+8),
 518                                         *(__u32 *)(sb->uuid+12));
 519         printk(KERN_DEBUG "        events: %llu\n",
 520                         (unsigned long long) le64_to_cpu(sb->events));
 521         printk(KERN_DEBUG "events cleared: %llu\n",
 522                         (unsigned long long) le64_to_cpu(sb->events_cleared));
 523         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
 524         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
 525         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
 526         printk(KERN_DEBUG "     sync size: %llu KB\n",
 527                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
 528         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
 529         kunmap_atomic(sb, KM_USER0);
 530 }
 531
 532 /* read the superblock from the bitmap file and initialize some bitmap fields */
 533 static int bitmap_read_sb(struct bitmap *bitmap)
 534 {
 535         char *reason = NULL;
 536         bitmap_super_t *sb;
 537         unsigned long chunksize, daemon_sleep, write_behind;
 538         unsigned long long events;
 539         int err = -EINVAL;
 540
 541         /* page 0 is the superblock, read it... */
 542         if (bitmap->file) {
 543                 loff_t isize = i_size_read(bitmap->file->f_mapping->host);
 544                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 545
 546                 bitmap->sb_page = read_page(bitmap->file, 0, bitmap, bytes);
 547         } else {
 548                 bitmap->sb_page = read_sb_page(bitmap->mddev,
 549                                                bitmap->mddev->bitmap_info.offset,
 550                                                NULL,
 551                                                0, sizeof(bitmap_super_t));
 552         }
 553         if (IS_ERR(bitmap->sb_page)) {
 554                 err = PTR_ERR(bitmap->sb_page);
 555                 bitmap->sb_page = NULL;
 556                 return err;
 557         }
 558
 559         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 560
 561         chunksize = le32_to_cpu(sb->chunksize);
 562         daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 563         write_behind = le32_to_cpu(sb->write_behind);
 564
 565         /* verify that the bitmap-specific fields are valid */
 566         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
 567                 reason = "bad magic";
 568         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
 569                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
 570                 reason = "unrecognized superblock version";
 571         else if (chunksize < 512)
 572                 reason = "bitmap chunksize too small";
 573         else if ((1 << ffz(~chunksize)) != chunksize)
 574                 reason = "bitmap chunksize not a power of 2";
 575         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT)
 576                 reason = "daemon sleep period out of range";
 577         else if (write_behind > COUNTER_MAX)
 578                 reason = "write-behind limit out of range (0 - 16383)";
 579         if (reason) {
 580                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
 581                         bmname(bitmap), reason);
 582                 goto out;
 583         }
 584
 585         /* keep the array size field of the bitmap superblock up to date */
 586         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
 587
 588         if (!bitmap->mddev->persistent)
 589                 goto success;
 590
 591         /*
 592          * if we have a persistent array superblock, compare the
 593          * bitmap's UUID and event counter to the mddev's
 594          */
 595         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
 596                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
 597                         bmname(bitmap));
 598                 goto out;
 599         }
 600         events = le64_to_cpu(sb->events);
 601         if (events < bitmap->mddev->events) {
 602                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
 603                         "-- forcing full recovery\n", bmname(bitmap), events,
 604                         (unsigned long long) bitmap->mddev->events);
 605                 sb->state |= cpu_to_le32(BITMAP_STALE);
 606         }
 607 success:
 608         /* assign fields using values from superblock */
 609         bitmap->mddev->bitmap_info.chunksize = chunksize;
 610         bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
 611         bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 612         bitmap->flags |= le32_to_cpu(sb->state);
 613         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
 614                 bitmap->flags |= BITMAP_HOSTENDIAN;
 615         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
 616         if (sb->state & cpu_to_le32(BITMAP_STALE))
 617                 bitmap->events_cleared = bitmap->mddev->events;
 618         err = 0;
 619 out:
 620         kunmap_atomic(sb, KM_USER0);
 621         if (err)
 622                 bitmap_print_sb(bitmap);
 623         return err;
 624 }
 625
 626 enum bitmap_mask_op {
 627         MASK_SET,
 628         MASK_UNSET
 629 };
 630
 631 /* record the state of the bitmap in the superblock.  Return the old value */
 632 static int bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
 633                              enum bitmap_mask_op op)
 634 {
 635         bitmap_super_t *sb;
 636         unsigned long flags;
 637         int old;
 638
 639         spin_lock_irqsave(&bitmap->lock, flags);
 640         if (!bitmap->sb_page) { /* can't set the state */
 641                 spin_unlock_irqrestore(&bitmap->lock, flags);
 642                 return 0;
 643         }
 644         spin_unlock_irqrestore(&bitmap->lock, flags);
 645         sb = kmap_atomic(bitmap->sb_page, KM_USER0);
 646         old = le32_to_cpu(sb->state) & bits;
 647         switch (op) {
 648         case MASK_SET:
 649                 sb->state |= cpu_to_le32(bits);
 650                 break;
 651         case MASK_UNSET:
 652                 sb->state &= cpu_to_le32(~bits);
 653                 break;
 654         default:
 655                 BUG();
 656         }
 657         kunmap_atomic(sb, KM_USER0);
 658         return old;
 659 }
 660
 661 /*
 662  * general bitmap file operations
 663  */
 664
 665 /*
 666  * on-disk bitmap:
 667  *
 668  * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
 669  * file a page at a time. There's a superblock at the start of the file.
 670  */
 671 /* calculate the index of the page that contains this bit */
 672 static inline unsigned long file_page_index(struct bitmap *bitmap, unsigned long chunk)
 673 {
 674         if (!bitmap->mddev->bitmap_info.external)
 675                 chunk += sizeof(bitmap_super_t) << 3;
 676         return chunk >> PAGE_BIT_SHIFT;
 677 }
 678
 679 /* calculate the (bit) offset of this bit within a page */
 680 static inline unsigned long file_page_offset(struct bitmap *bitmap, unsigned long chunk)
 681 {
 682         if (!bitmap->mddev->bitmap_info.external)
 683                 chunk += sizeof(bitmap_super_t) << 3;
 684         return chunk & (PAGE_BITS - 1);
 685 }
 686
 687 /*
 688  * return a pointer to the page in the filemap that contains the given bit
 689  *
 690  * this lookup is complicated by the fact that the bitmap sb might be exactly
 691  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
 692  * 0 or page 1
 693  */
 694 static inline struct page *filemap_get_page(struct bitmap *bitmap,
 695                                         unsigned long chunk)
 696 {
 697         if (file_page_index(bitmap, chunk) >= bitmap->file_pages)
 698                 return NULL;
 699         return bitmap->filemap[file_page_index(bitmap, chunk)
 700                                - file_page_index(bitmap, 0)];
 701 }
 702
 703 static void bitmap_file_unmap(struct bitmap *bitmap)
 704 {
 705         struct page **map, *sb_page;
 706         unsigned long *attr;
 707         int pages;
 708         unsigned long flags;
 709
 710         spin_lock_irqsave(&bitmap->lock, flags);
 711         map = bitmap->filemap;
 712         bitmap->filemap = NULL;
 713         attr = bitmap->filemap_attr;
 714         bitmap->filemap_attr = NULL;
 715         pages = bitmap->file_pages;
 716         bitmap->file_pages = 0;
 717         sb_page = bitmap->sb_page;
 718         bitmap->sb_page = NULL;
 719         spin_unlock_irqrestore(&bitmap->lock, flags);
 720
 721         while (pages--)
 722                 if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 723                         free_buffers(map[pages]);
 724         kfree(map);
 725         kfree(attr);
 726
 727         if (sb_page)
 728                 free_buffers(sb_page);
 729 }
 730
 731 static void bitmap_file_put(struct bitmap *bitmap)
 732 {
 733         struct file *file;
 734         unsigned long flags;
 735
 736         spin_lock_irqsave(&bitmap->lock, flags);
 737         file = bitmap->file;
 738         bitmap->file = NULL;
 739         spin_unlock_irqrestore(&bitmap->lock, flags);
 740
 741         if (file)
 742                 wait_event(bitmap->write_wait,
 743                            atomic_read(&bitmap->pending_writes)==0);
 744         bitmap_file_unmap(bitmap);
 745
 746         if (file) {
 747                 struct inode *inode = file->f_path.dentry->d_inode;
 748                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
 749                 fput(file);
 750         }
 751 }
 752
 753 /*
 754  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
 755  * then it is no longer reliable, so we stop using it and we mark the file
 756  * as failed in the superblock
 757  */
 758 static void bitmap_file_kick(struct bitmap *bitmap)
 759 {
 760         char *path, *ptr = NULL;
 761
 762         if (bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET) == 0) {
 763                 bitmap_update_sb(bitmap);
 764
 765                 if (bitmap->file) {
 766                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 767                         if (path)
 768                                 ptr = d_path(&bitmap->file->f_path, path,
 769                                              PAGE_SIZE);
 770
 771                         printk(KERN_ALERT
 772                               "%s: kicking failed bitmap file %s from array!\n",
 773                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
 774
 775                         kfree(path);
 776                 } else
 777                         printk(KERN_ALERT
 778                                "%s: disabling internal bitmap due to errors\n",
 779                                bmname(bitmap));
 780         }
 781
 782         bitmap_file_put(bitmap);
 783
 784         return;
 785 }
 786
 787 enum bitmap_page_attr {
 788         BITMAP_PAGE_DIRTY = 0,     /* there are set bits that need to be synced */
 789         BITMAP_PAGE_CLEAN = 1,     /* there are bits that might need to be cleared */
 790         BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 791 };
 792
 793 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
 794                                 enum bitmap_page_attr attr)
 795 {
 796         __set_bit((page->index<<2) + attr, bitmap->filemap_attr);
 797 }
 798
 799 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
 800                                 enum bitmap_page_attr attr)
 801 {
 802         __clear_bit((page->index<<2) + attr, bitmap->filemap_attr);
 803 }
 804
 805 static inline unsigned long test_page_attr(struct bitmap *bitmap, struct page *page,
 806                                            enum bitmap_page_attr attr)
 807 {
 808         return test_bit((page->index<<2) + attr, bitmap->filemap_attr);
 809 }
 810
 811 /*
 812  * bitmap_file_set_bit -- called before performing a write to the md device
 813  * to set (and eventually sync) a particular bit in the bitmap file
 814  *
 815  * we set the bit immediately, then we record the page number so that
 816  * when an unplug occurs, we can flush the dirty pages out to disk
 817  */
 818 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
 819 {
 820         unsigned long bit;
 821         struct page *page;
 822         void *kaddr;
 823         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
 824
 825         if (!bitmap->filemap)
 826                 return;
 827
 828         page = filemap_get_page(bitmap, chunk);
 829         if (!page)
 830                 return;
 831         bit = file_page_offset(bitmap, chunk);
 832
 833         /* set the bit */
 834         kaddr = kmap_atomic(page, KM_USER0);
 835         if (bitmap->flags & BITMAP_HOSTENDIAN)
 836                 set_bit(bit, kaddr);
 837         else
 838                 ext2_set_bit(bit, kaddr);
 839         kunmap_atomic(kaddr, KM_USER0);
 840         PRINTK("set file bit %lu page %lu\n", bit, page->index);
 841
 842         /* record page number so it gets flushed to disk when unplug occurs */
 843         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 844 }
 845
 846 /* this gets called when the md device is ready to unplug its underlying
 847  * (slave) device queues -- before we let any writes go down, we need to
 848  * sync the dirty pages of the bitmap file to disk */
 849 void bitmap_unplug(struct bitmap *bitmap)
 850 {
 851         unsigned long i, flags;
 852         int dirty, need_write;
 853         struct page *page;
 854         int wait = 0;
 855
 856         if (!bitmap)
 857                 return;
 858
 859         /* look at each page to see if there are any set bits that need to be
 860          * flushed out to disk */
 861         for (i = 0; i < bitmap->file_pages; i++) {
 862                 spin_lock_irqsave(&bitmap->lock, flags);
 863                 if (!bitmap->filemap) {
 864                         spin_unlock_irqrestore(&bitmap->lock, flags);
 865                         return;
 866                 }
 867                 page = bitmap->filemap[i];
 868                 dirty = test_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 869                 need_write = test_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 870                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
 871                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 872                 if (dirty)
 873                         wait = 1;
 874                 spin_unlock_irqrestore(&bitmap->lock, flags);
 875
 876                 if (dirty || need_write)
 877                         write_page(bitmap, page, 0);
 878         }
 879         if (wait) { /* if any writes were performed, we need to wait on them */
 880                 if (bitmap->file)
 881                         wait_event(bitmap->write_wait,
 882                                    atomic_read(&bitmap->pending_writes)==0);
 883                 else
 884                         md_super_wait(bitmap->mddev);
 885         }
 886         if (bitmap->flags & BITMAP_WRITE_ERROR)
 887                 bitmap_file_kick(bitmap);
 888 }
 889 EXPORT_SYMBOL(bitmap_unplug);
 890
 891 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
 892 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
 893  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
 894  * memory mapping of the bitmap file
 895  * Special cases:
 896  *   if there's no bitmap file, or if the bitmap file had been
 897  *   previously kicked from the array, we mark all the bits as
 898  *   1's in order to cause a full resync.
 899  *
 900  * We ignore all bits for sectors that end earlier than 'start'.
 901  * This is used when reading an out-of-date bitmap...
 902  */
 903 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 904 {
 905         unsigned long i, chunks, index, oldindex, bit;
 906         struct page *page = NULL, *oldpage = NULL;
 907         unsigned long num_pages, bit_cnt = 0;
 908         struct file *file;
 909         unsigned long bytes, offset;
 910         int outofdate;
 911         int ret = -ENOSPC;
 912         void *paddr;
 913
 914         chunks = bitmap->chunks;
 915         file = bitmap->file;
 916
 917         BUG_ON(!file && !bitmap->mddev->bitmap_info.offset);
 918
 919 #ifdef INJECT_FAULTS_3
 920         outofdate = 1;
 921 #else
 922         outofdate = bitmap->flags & BITMAP_STALE;
 923 #endif
 924         if (outofdate)
 925                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 926                         "recovery\n", bmname(bitmap));
 927
 928         bytes = (chunks + 7) / 8;
 929         if (!bitmap->mddev->bitmap_info.external)
 930                 bytes += sizeof(bitmap_super_t);
 931
 932         num_pages = (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
 933
 934         if (file && i_size_read(file->f_mapping->host) < bytes) {
 935                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
 936                         bmname(bitmap),
 937                         (unsigned long) i_size_read(file->f_mapping->host),
 938                         bytes);
 939                 goto err;
 940         }
 941
 942         ret = -ENOMEM;
 943
 944         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
 945         if (!bitmap->filemap)
 946                 goto err;
 947
 948         /* We need 4 bits per page, rounded up to a multiple of sizeof(unsigned long) */
 949         bitmap->filemap_attr = kzalloc(
 950                 roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
 951                 GFP_KERNEL);
 952         if (!bitmap->filemap_attr)
 953                 goto err;
 954
 955         oldindex = ~0L;
 956
 957         for (i = 0; i < chunks; i++) {
 958                 int b;
 959                 index = file_page_index(bitmap, i);
 960                 bit = file_page_offset(bitmap, i);
 961                 if (index != oldindex) { /* this is a new page, read it in */
 962                         int count;
 963                         /* unmap the old page, we're done with it */
 964                         if (index == num_pages-1)
 965                                 count = bytes - index * PAGE_SIZE;
 966                         else
 967                                 count = PAGE_SIZE;
 968                         if (index == 0 && bitmap->sb_page) {
 969                                 /*
 970                                  * if we're here then the superblock page
 971                                  * contains some bits (PAGE_SIZE != sizeof sb)
 972                                  * we've already read it in, so just use it
 973                                  */
 974                                 page = bitmap->sb_page;
 975                                 offset = sizeof(bitmap_super_t);
 976                                 if (!file)
 977                                         read_sb_page(bitmap->mddev,
 978                                                      bitmap->mddev->bitmap_info.offset,
 979                                                      page,
 980                                                      index, count);
 981                         } else if (file) {
 982                                 page = read_page(file, index, bitmap, count);
 983                                 offset = 0;
 984                         } else {
 985                                 page = read_sb_page(bitmap->mddev,
 986                                                     bitmap->mddev->bitmap_info.offset,
 987                                                     NULL,
 988                                                     index, count);
 989                                 offset = 0;
 990                         }
 991                         if (IS_ERR(page)) { /* read error */
 992                                 ret = PTR_ERR(page);
 993                                 goto err;
 994                         }
 995
 996                         oldindex = index;
 997                         oldpage = page;
 998
 999                         bitmap->filemap[bitmap->file_pages++] = page;
1000                         bitmap->last_page_size = count;
1001
1002                         if (outofdate) {
1003                                 /*
1004                                  * if bitmap is out of date, dirty the
1005                                  * whole page and write it out
1006                                  */
1007                                 paddr = kmap_atomic(page, KM_USER0);
1008                                 memset(paddr + offset, 0xff,
1009                                        PAGE_SIZE - offset);
1010                                 kunmap_atomic(paddr, KM_USER0);
1011                                 write_page(bitmap, page, 1);
1012
1013                                 ret = -EIO;
1014                                 if (bitmap->flags & BITMAP_WRITE_ERROR)
1015                                         goto err;
1016                         }
1017                 }
1018                 paddr = kmap_atomic(page, KM_USER0);
1019                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1020                         b = test_bit(bit, paddr);
1021                 else
1022                         b = ext2_test_bit(bit, paddr);
1023                 kunmap_atomic(paddr, KM_USER0);
1024                 if (b) {
1025                         /* if the disk bit is set, set the memory bit */
1026                         int needed = ((sector_t)(i+1) << (CHUNK_BLOCK_SHIFT(bitmap))
1027                                       >= start);
1028                         bitmap_set_memory_bits(bitmap,
1029                                                (sector_t)i << CHUNK_BLOCK_SHIFT(bitmap),
1030                                                needed);
1031                         bit_cnt++;
1032                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1033                 }
1034         }
1035
1036         /* everything went OK */
1037         ret = 0;
1038         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
1039
1040         if (bit_cnt) { /* Kick recovery if any bits were set */
1041                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1042                 md_wakeup_thread(bitmap->mddev->thread);
1043         }
1044
1045         printk(KERN_INFO "%s: bitmap initialized from disk: "
1046                 "read %lu/%lu pages, set %lu bits\n",
1047                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt);
1048
1049         return 0;
1050
1051  err:
1052         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1053                bmname(bitmap), ret);
1054         return ret;
1055 }
1056
1057 void bitmap_write_all(struct bitmap *bitmap)
1058 {
1059         /* We don't actually write all bitmap blocks here,
1060          * just flag them as needing to be written
1061          */
1062         int i;
1063
1064         for (i = 0; i < bitmap->file_pages; i++)
1065                 set_page_attr(bitmap, bitmap->filemap[i],
1066                               BITMAP_PAGE_NEEDWRITE);
1067 }
1068
1069 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
1070 {
1071         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1072         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1073         bitmap->bp[page].count += inc;
1074         bitmap_checkfree(bitmap, page);
1075 }
1076 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1077                                             sector_t offset, int *blocks,
1078                                             int create);
1079
1080 /*
1081  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1082  *                      out to disk
1083  */
1084
1085 void bitmap_daemon_work(mddev_t *mddev)
1086 {
1087         struct bitmap *bitmap;
1088         unsigned long j;
1089         unsigned long flags;
1090         struct page *page = NULL, *lastpage = NULL;
1091         int blocks;
1092         void *paddr;
1093
1094         /* Use a mutex to guard daemon_work against
1095          * bitmap_destroy.
1096          */
1097         mutex_lock(&mddev->bitmap_info.mutex);
1098         bitmap = mddev->bitmap;
1099         if (bitmap == NULL) {
1100                 mutex_unlock(&mddev->bitmap_info.mutex);
1101                 return;
1102         }
1103         if (time_before(jiffies, bitmap->daemon_lastrun
1104                         + bitmap->mddev->bitmap_info.daemon_sleep))
1105                 goto done;
1106
1107         bitmap->daemon_lastrun = jiffies;
1108         if (bitmap->allclean) {
1109                 bitmap->mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1110                 goto done;
1111         }
1112         bitmap->allclean = 1;
1113
1114         spin_lock_irqsave(&bitmap->lock, flags);
1115         for (j = 0; j < bitmap->chunks; j++) {
1116                 bitmap_counter_t *bmc;
1117                 if (!bitmap->filemap)
1118                         /* error or shutdown */
1119                         break;
1120
1121                 page = filemap_get_page(bitmap, j);
1122
1123                 if (page != lastpage) {
1124                         /* skip this page unless it's marked as needing cleaning */
1125                         if (!test_page_attr(bitmap, page, BITMAP_PAGE_CLEAN)) {
1126                                 int need_write = test_page_attr(bitmap, page,
1127                                                                 BITMAP_PAGE_NEEDWRITE);
1128                                 if (need_write)
1129                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1130
1131                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1132                                 if (need_write) {
1133                                         write_page(bitmap, page, 0);
1134                                         bitmap->allclean = 0;
1135                                 }
1136                                 spin_lock_irqsave(&bitmap->lock, flags);
1137                                 j |= (PAGE_BITS - 1);
1138                                 continue;
1139                         }
1140
1141                         /* grab the new page, sync and release the old */
1142                         if (lastpage != NULL) {
1143                                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1144                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1145                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1146                                         write_page(bitmap, lastpage, 0);
1147                                 } else {
1148                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1149                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1150                                 }
1151                         } else
1152                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1153                         lastpage = page;
1154
1155                         /* We are possibly going to clear some bits, so make
1156                          * sure that events_cleared is up-to-date.
1157                          */
1158                         if (bitmap->need_sync &&
1159                             bitmap->mddev->bitmap_info.external == 0) {
1160                                 bitmap_super_t *sb;
1161                                 bitmap->need_sync = 0;
1162                                 sb = kmap_atomic(bitmap->sb_page, KM_USER0);
1163                                 sb->events_cleared =
1164                                         cpu_to_le64(bitmap->events_cleared);
1165                                 kunmap_atomic(sb, KM_USER0);
1166                                 write_page(bitmap, bitmap->sb_page, 1);
1167                         }
1168                         spin_lock_irqsave(&bitmap->lock, flags);
1169                         if (!bitmap->need_sync)
1170                                 clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1171                 }
1172                 bmc = bitmap_get_counter(bitmap,
1173                                          (sector_t)j << CHUNK_BLOCK_SHIFT(bitmap),
1174                                          &blocks, 0);
1175                 if (bmc) {
1176                         if (*bmc)
1177                                 bitmap->allclean = 0;
1178
1179                         if (*bmc == 2) {
1180                                 *bmc = 1; /* maybe clear the bit next time */
1181                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1182                         } else if (*bmc == 1 && !bitmap->need_sync) {
1183                                 /* we can clear the bit */
1184                                 *bmc = 0;
1185                                 bitmap_count_page(bitmap,
1186                                                   (sector_t)j << CHUNK_BLOCK_SHIFT(bitmap),
1187                                                   -1);
1188
1189                                 /* clear the bit */
1190                                 paddr = kmap_atomic(page, KM_USER0);
1191                                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1192                                         clear_bit(file_page_offset(bitmap, j),
1193                                                   paddr);
1194                                 else
1195                                         ext2_clear_bit(file_page_offset(bitmap, j),
1196                                                        paddr);
1197                                 kunmap_atomic(paddr, KM_USER0);
1198                         }
1199                 } else
1200                         j |= PAGE_COUNTER_MASK;
1201         }
1202         spin_unlock_irqrestore(&bitmap->lock, flags);
1203
1204         /* now sync the final page */
1205         if (lastpage != NULL) {
1206                 spin_lock_irqsave(&bitmap->lock, flags);
1207                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1208                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1209                         spin_unlock_irqrestore(&bitmap->lock, flags);
1210                         write_page(bitmap, lastpage, 0);
1211                 } else {
1212                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1213                         spin_unlock_irqrestore(&bitmap->lock, flags);
1214                 }
1215         }
1216
1217  done:
1218         if (bitmap->allclean == 0)
1219                 bitmap->mddev->thread->timeout =
1220                         bitmap->mddev->bitmap_info.daemon_sleep;
1221         mutex_unlock(&mddev->bitmap_info.mutex);
1222 }
1223
1224 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1225                                             sector_t offset, int *blocks,
1226                                             int create)
1227 __releases(bitmap->lock)
1228 __acquires(bitmap->lock)
1229 {
1230         /* If 'create', we might release the lock and reclaim it.
1231          * The lock must have been taken with interrupts enabled.
1232          * If !create, we don't release the lock.
1233          */
1234         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1235         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1236         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1237         sector_t csize;
1238
1239         if (bitmap_checkpage(bitmap, page, create) < 0) {
1240                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1241                 *blocks = csize - (offset & (csize - 1));
1242                 return NULL;
1243         }
1244         /* now locked ... */
1245
1246         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1247                 /* should we use the first or second counter field
1248                  * of the hijacked pointer? */
1249                 int hi = (pageoff > PAGE_COUNTER_MASK);
1250                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1251                                           PAGE_COUNTER_SHIFT - 1);
1252                 *blocks = csize - (offset & (csize - 1));
1253                 return  &((bitmap_counter_t *)
1254                           &bitmap->bp[page].map)[hi];
1255         } else { /* page is allocated */
1256                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1257                 *blocks = csize - (offset & (csize - 1));
1258                 return (bitmap_counter_t *)
1259                         &(bitmap->bp[page].map[pageoff]);
1260         }
1261 }
1262
1263 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1264 {
1265         if (!bitmap)
1266                 return 0;
1267
1268         if (behind) {
1269                 int bw;
1270                 atomic_inc(&bitmap->behind_writes);
1271                 bw = atomic_read(&bitmap->behind_writes);
1272                 if (bw > bitmap->behind_writes_used)
1273                         bitmap->behind_writes_used = bw;
1274
1275                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1276                        bw, bitmap->max_write_behind);
1277         }
1278
1279         while (sectors) {
1280                 int blocks;
1281                 bitmap_counter_t *bmc;
1282
1283                 spin_lock_irq(&bitmap->lock);
1284                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1285                 if (!bmc) {
1286                         spin_unlock_irq(&bitmap->lock);
1287                         return 0;
1288                 }
1289
1290                 if (unlikely((*bmc & COUNTER_MAX) == COUNTER_MAX)) {
1291                         DEFINE_WAIT(__wait);
1292                         /* note that it is safe to do the prepare_to_wait
1293                          * after the test as long as we do it before dropping
1294                          * the spinlock.
1295                          */
1296                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1297                                         TASK_UNINTERRUPTIBLE);
1298                         spin_unlock_irq(&bitmap->lock);
1299                         blk_unplug(bitmap->mddev->queue);
1300                         schedule();
1301                         finish_wait(&bitmap->overflow_wait, &__wait);
1302                         continue;
1303                 }
1304
1305                 switch (*bmc) {
1306                 case 0:
1307                         bitmap_file_set_bit(bitmap, offset);
1308                         bitmap_count_page(bitmap, offset, 1);
1309                         blk_plug_device_unlocked(bitmap->mddev->queue);
1310                         /* fall through */
1311                 case 1:
1312                         *bmc = 2;
1313                 }
1314
1315                 (*bmc)++;
1316
1317                 spin_unlock_irq(&bitmap->lock);
1318
1319                 offset += blocks;
1320                 if (sectors > blocks)
1321                         sectors -= blocks;
1322                 else
1323                         sectors = 0;
1324         }
1325         bitmap->allclean = 0;
1326         return 0;
1327 }
1328 EXPORT_SYMBOL(bitmap_startwrite);
1329
1330 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1331                      int success, int behind)
1332 {
1333         if (!bitmap)
1334                 return;
1335         if (behind) {
1336                 if (atomic_dec_and_test(&bitmap->behind_writes))
1337                         wake_up(&bitmap->behind_wait);
1338                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1339                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1340         }
1341         if (bitmap->mddev->degraded)
1342                 /* Never clear bits or update events_cleared when degraded */
1343                 success = 0;
1344
1345         while (sectors) {
1346                 int blocks;
1347                 unsigned long flags;
1348                 bitmap_counter_t *bmc;
1349
1350                 spin_lock_irqsave(&bitmap->lock, flags);
1351                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1352                 if (!bmc) {
1353                         spin_unlock_irqrestore(&bitmap->lock, flags);
1354                         return;
1355                 }
1356
1357                 if (success &&
1358                     bitmap->events_cleared < bitmap->mddev->events) {
1359                         bitmap->events_cleared = bitmap->mddev->events;
1360                         bitmap->need_sync = 1;
1361                         sysfs_notify_dirent_safe(bitmap->sysfs_can_clear);
1362                 }
1363
1364                 if (!success && ! (*bmc & NEEDED_MASK))
1365                         *bmc |= NEEDED_MASK;
1366
1367                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX)
1368                         wake_up(&bitmap->overflow_wait);
1369
1370                 (*bmc)--;
1371                 if (*bmc <= 2)
1372                         set_page_attr(bitmap,
1373                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1374                                       BITMAP_PAGE_CLEAN);
1375
1376                 spin_unlock_irqrestore(&bitmap->lock, flags);
1377                 offset += blocks;
1378                 if (sectors > blocks)
1379                         sectors -= blocks;
1380                 else
1381                         sectors = 0;
1382         }
1383 }
1384 EXPORT_SYMBOL(bitmap_endwrite);
1385
1386 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1387                                int degraded)
1388 {
1389         bitmap_counter_t *bmc;
1390         int rv;
1391         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1392                 *blocks = 1024;
1393                 return 1; /* always resync if no bitmap */
1394         }
1395         spin_lock_irq(&bitmap->lock);
1396         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1397         rv = 0;
1398         if (bmc) {
1399                 /* locked */
1400                 if (RESYNC(*bmc))
1401                         rv = 1;
1402                 else if (NEEDED(*bmc)) {
1403                         rv = 1;
1404                         if (!degraded) { /* don't set/clear bits if degraded */
1405                                 *bmc |= RESYNC_MASK;
1406                                 *bmc &= ~NEEDED_MASK;
1407                         }
1408                 }
1409         }
1410         spin_unlock_irq(&bitmap->lock);
1411         bitmap->allclean = 0;
1412         return rv;
1413 }
1414
1415 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1416                       int degraded)
1417 {
1418         /* bitmap_start_sync must always report on multiples of whole
1419          * pages, otherwise resync (which is very PAGE_SIZE based) will
1420          * get confused.
1421          * So call __bitmap_start_sync repeatedly (if needed) until
1422          * At least PAGE_SIZE>>9 blocks are covered.
1423          * Return the 'or' of the result.
1424          */
1425         int rv = 0;
1426         int blocks1;
1427
1428         *blocks = 0;
1429         while (*blocks < (PAGE_SIZE>>9)) {
1430                 rv |= __bitmap_start_sync(bitmap, offset,
1431                                           &blocks1, degraded);
1432                 offset += blocks1;
1433                 *blocks += blocks1;
1434         }
1435         return rv;
1436 }
1437 EXPORT_SYMBOL(bitmap_start_sync);
1438
1439 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1440 {
1441         bitmap_counter_t *bmc;
1442         unsigned long flags;
1443
1444         if (bitmap == NULL) {
1445                 *blocks = 1024;
1446                 return;
1447         }
1448         spin_lock_irqsave(&bitmap->lock, flags);
1449         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1450         if (bmc == NULL)
1451                 goto unlock;
1452         /* locked */
1453         if (RESYNC(*bmc)) {
1454                 *bmc &= ~RESYNC_MASK;
1455
1456                 if (!NEEDED(*bmc) && aborted)
1457                         *bmc |= NEEDED_MASK;
1458                 else {
1459                         if (*bmc <= 2)
1460                                 set_page_attr(bitmap,
1461                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1462                                               BITMAP_PAGE_CLEAN);
1463                 }
1464         }
1465  unlock:
1466         spin_unlock_irqrestore(&bitmap->lock, flags);
1467         bitmap->allclean = 0;
1468 }
1469 EXPORT_SYMBOL(bitmap_end_sync);
1470
1471 void bitmap_close_sync(struct bitmap *bitmap)
1472 {
1473         /* Sync has finished, and any bitmap chunks that weren't synced
1474          * properly have been aborted.  It remains to us to clear the
1475          * RESYNC bit wherever it is still on
1476          */
1477         sector_t sector = 0;
1478         int blocks;
1479         if (!bitmap)
1480                 return;
1481         while (sector < bitmap->mddev->resync_max_sectors) {
1482                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1483                 sector += blocks;
1484         }
1485 }
1486 EXPORT_SYMBOL(bitmap_close_sync);
1487
1488 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1489 {
1490         sector_t s = 0;
1491         int blocks;
1492
1493         if (!bitmap)
1494                 return;
1495         if (sector == 0) {
1496                 bitmap->last_end_sync = jiffies;
1497                 return;
1498         }
1499         if (time_before(jiffies, (bitmap->last_end_sync
1500                                   + bitmap->mddev->bitmap_info.daemon_sleep)))
1501                 return;
1502         wait_event(bitmap->mddev->recovery_wait,
1503                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1504
1505         bitmap->mddev->curr_resync_completed = bitmap->mddev->curr_resync;
1506         if (bitmap->mddev->persistent)
1507                 set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
1508         sector &= ~((1ULL << CHUNK_BLOCK_SHIFT(bitmap)) - 1);
1509         s = 0;
1510         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1511                 bitmap_end_sync(bitmap, s, &blocks, 0);
1512                 s += blocks;
1513         }
1514         bitmap->last_end_sync = jiffies;
1515         sysfs_notify(&bitmap->mddev->kobj, NULL, "sync_completed");
1516 }
1517 EXPORT_SYMBOL(bitmap_cond_end_sync);
1518
1519 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1520 {
1521         /* For each chunk covered by any of these sectors, set the
1522          * counter to 1 and set resync_needed.  They should all
1523          * be 0 at this point
1524          */
1525
1526         int secs;
1527         bitmap_counter_t *bmc;
1528         spin_lock_irq(&bitmap->lock);
1529         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1530         if (!bmc) {
1531                 spin_unlock_irq(&bitmap->lock);
1532                 return;
1533         }
1534         if (!*bmc) {
1535                 struct page *page;
1536                 *bmc = 1 | (needed ? NEEDED_MASK : 0);
1537                 bitmap_count_page(bitmap, offset, 1);
1538                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1539                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1540         }
1541         spin_unlock_irq(&bitmap->lock);
1542         bitmap->allclean = 0;
1543 }
1544
1545 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1546 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1547 {
1548         unsigned long chunk;
1549
1550         for (chunk = s; chunk <= e; chunk++) {
1551                 sector_t sec = (sector_t)chunk << CHUNK_BLOCK_SHIFT(bitmap);
1552                 bitmap_set_memory_bits(bitmap, sec, 1);
1553                 bitmap_file_set_bit(bitmap, sec);
1554                 if (sec < bitmap->mddev->recovery_cp)
1555                         /* We are asserting that the array is dirty,
1556                          * so move the recovery_cp address back so
1557                          * that it is obvious that it is dirty
1558                          */
1559                         bitmap->mddev->recovery_cp = sec;
1560         }
1561 }
1562
1563 /*
1564  * flush out any pending updates
1565  */
1566 void bitmap_flush(mddev_t *mddev)
1567 {
1568         struct bitmap *bitmap = mddev->bitmap;
1569         long sleep;
1570
1571         if (!bitmap) /* there was no bitmap */
1572                 return;
1573
1574         /* run the daemon_work three time to ensure everything is flushed
1575          * that can be
1576          */
1577         sleep = mddev->bitmap_info.daemon_sleep * 2;
1578         bitmap->daemon_lastrun -= sleep;
1579         bitmap_daemon_work(mddev);
1580         bitmap->daemon_lastrun -= sleep;
1581         bitmap_daemon_work(mddev);
1582         bitmap->daemon_lastrun -= sleep;
1583         bitmap_daemon_work(mddev);
1584         bitmap_update_sb(bitmap);
1585 }
1586
1587 /*
1588  * free memory that was allocated
1589  */
1590 static void bitmap_free(struct bitmap *bitmap)
1591 {
1592         unsigned long k, pages;
1593         struct bitmap_page *bp;
1594
1595         if (!bitmap) /* there was no bitmap */
1596                 return;
1597
1598         /* release the bitmap file and kill the daemon */
1599         bitmap_file_put(bitmap);
1600
1601         bp = bitmap->bp;
1602         pages = bitmap->pages;
1603
1604         /* free all allocated memory */
1605
1606         if (bp) /* deallocate the page memory */
1607                 for (k = 0; k < pages; k++)
1608                         if (bp[k].map && !bp[k].hijacked)
1609                                 kfree(bp[k].map);
1610         kfree(bp);
1611         kfree(bitmap);
1612 }
1613
1614 void bitmap_destroy(mddev_t *mddev)
1615 {
1616         struct bitmap *bitmap = mddev->bitmap;
1617
1618         if (!bitmap) /* there was no bitmap */
1619                 return;
1620
1621         mutex_lock(&mddev->bitmap_info.mutex);
1622         mddev->bitmap = NULL; /* disconnect from the md device */
1623         mutex_unlock(&mddev->bitmap_info.mutex);
1624         if (mddev->thread)
1625                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1626
1627         if (bitmap->sysfs_can_clear)
1628                 sysfs_put(bitmap->sysfs_can_clear);
1629
1630         bitmap_free(bitmap);
1631 }
1632
1633 /*
1634  * initialize the bitmap structure
1635  * if this returns an error, bitmap_destroy must be called to do clean up
1636  */
1637 int bitmap_create(mddev_t *mddev)
1638 {
1639         struct bitmap *bitmap;
1640         sector_t blocks = mddev->resync_max_sectors;
1641         unsigned long chunks;
1642         unsigned long pages;
1643         struct file *file = mddev->bitmap_info.file;
1644         int err;
1645         sector_t start;
1646         struct sysfs_dirent *bm = NULL;
1647
1648         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1649
1650         if (!file && !mddev->bitmap_info.offset) /* bitmap disabled, nothing to do */
1651                 return 0;
1652
1653         BUG_ON(file && mddev->bitmap_info.offset);
1654
1655         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1656         if (!bitmap)
1657                 return -ENOMEM;
1658
1659         spin_lock_init(&bitmap->lock);
1660         atomic_set(&bitmap->pending_writes, 0);
1661         init_waitqueue_head(&bitmap->write_wait);
1662         init_waitqueue_head(&bitmap->overflow_wait);
1663         init_waitqueue_head(&bitmap->behind_wait);
1664
1665         bitmap->mddev = mddev;
1666
1667         if (mddev->kobj.sd)
1668                 bm = sysfs_get_dirent(mddev->kobj.sd, NULL, "bitmap");
1669         if (bm) {
1670                 bitmap->sysfs_can_clear = sysfs_get_dirent(bm, NULL, "can_clear");
1671                 sysfs_put(bm);
1672         } else
1673                 bitmap->sysfs_can_clear = NULL;
1674
1675         bitmap->file = file;
1676         if (file) {
1677                 get_file(file);
1678                 /* As future accesses to this file will use bmap,
1679                  * and bypass the page cache, we must sync the file
1680                  * first.
1681                  */
1682                 vfs_fsync(file, 1);
1683         }
1684         /* read superblock from bitmap file (this sets mddev->bitmap_info.chunksize) */
1685         if (!mddev->bitmap_info.external)
1686                 err = bitmap_read_sb(bitmap);
1687         else {
1688                 err = 0;
1689                 if (mddev->bitmap_info.chunksize == 0 ||
1690                     mddev->bitmap_info.daemon_sleep == 0)
1691                         /* chunksize and time_base need to be
1692                          * set first. */
1693                         err = -EINVAL;
1694         }
1695         if (err)
1696                 goto error;
1697
1698         bitmap->daemon_lastrun = jiffies;
1699         bitmap->chunkshift = ffz(~mddev->bitmap_info.chunksize);
1700
1701         /* now that chunksize and chunkshift are set, we can use these macros */
1702         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) >>
1703                         CHUNK_BLOCK_SHIFT(bitmap);
1704         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1705
1706         BUG_ON(!pages);
1707
1708         bitmap->chunks = chunks;
1709         bitmap->pages = pages;
1710         bitmap->missing_pages = pages;
1711         bitmap->counter_bits = COUNTER_BITS;
1712
1713         bitmap->syncchunk = ~0UL;
1714
1715 #ifdef INJECT_FATAL_FAULT_1
1716         bitmap->bp = NULL;
1717 #else
1718         bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1719 #endif
1720         err = -ENOMEM;
1721         if (!bitmap->bp)
1722                 goto error;
1723
1724         /* now that we have some pages available, initialize the in-memory
1725          * bitmap from the on-disk bitmap */
1726         start = 0;
1727         if (mddev->degraded == 0
1728             || bitmap->events_cleared == mddev->events)
1729                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1730                 start = mddev->recovery_cp;
1731         err = bitmap_init_from_disk(bitmap, start);
1732
1733         if (err)
1734                 goto error;
1735
1736         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1737                 pages, bmname(bitmap));
1738
1739         mddev->bitmap = bitmap;
1740
1741         mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
1742         md_wakeup_thread(mddev->thread);
1743
1744         bitmap_update_sb(bitmap);
1745
1746         return (bitmap->flags & BITMAP_WRITE_ERROR) ? -EIO : 0;
1747
1748  error:
1749         bitmap_free(bitmap);
1750         return err;
1751 }
1752
1753 static ssize_t
1754 location_show(mddev_t *mddev, char *page)
1755 {
1756         ssize_t len;
1757         if (mddev->bitmap_info.file)
1758                 len = sprintf(page, "file");
1759         else if (mddev->bitmap_info.offset)
1760                 len = sprintf(page, "%+lld", (long long)mddev->bitmap_info.offset);
1761         else
1762                 len = sprintf(page, "none");
1763         len += sprintf(page+len, "\n");
1764         return len;
1765 }
1766
1767 static ssize_t
1768 location_store(mddev_t *mddev, const char *buf, size_t len)
1769 {
1770
1771         if (mddev->pers) {
1772                 if (!mddev->pers->quiesce)
1773                         return -EBUSY;
1774                 if (mddev->recovery || mddev->sync_thread)
1775                         return -EBUSY;
1776         }
1777
1778         if (mddev->bitmap || mddev->bitmap_info.file ||
1779             mddev->bitmap_info.offset) {
1780                 /* bitmap already configured.  Only option is to clear it */
1781                 if (strncmp(buf, "none", 4) != 0)
1782                         return -EBUSY;
1783                 if (mddev->pers) {
1784                         mddev->pers->quiesce(mddev, 1);
1785                         bitmap_destroy(mddev);
1786                         mddev->pers->quiesce(mddev, 0);
1787                 }
1788                 mddev->bitmap_info.offset = 0;
1789                 if (mddev->bitmap_info.file) {
1790                         struct file *f = mddev->bitmap_info.file;
1791                         mddev->bitmap_info.file = NULL;
1792                         restore_bitmap_write_access(f);
1793                         fput(f);
1794                 }
1795         } else {
1796                 /* No bitmap, OK to set a location */
1797                 long long offset;
1798                 if (strncmp(buf, "none", 4) == 0)
1799                         /* nothing to be done */;
1800                 else if (strncmp(buf, "file:", 5) == 0) {
1801                         /* Not supported yet */
1802                         return -EINVAL;
1803                 } else {
1804                         int rv;
1805                         if (buf[0] == '+')
1806                                 rv = strict_strtoll(buf+1, 10, &offset);
1807                         else
1808                                 rv = strict_strtoll(buf, 10, &offset);
1809                         if (rv)
1810                                 return rv;
1811                         if (offset == 0)
1812                                 return -EINVAL;
1813                         if (mddev->bitmap_info.external == 0 &&
1814                             mddev->major_version == 0 &&
1815                             offset != mddev->bitmap_info.default_offset)
1816                                 return -EINVAL;
1817                         mddev->bitmap_info.offset = offset;
1818                         if (mddev->pers) {
1819                                 mddev->pers->quiesce(mddev, 1);
1820                                 rv = bitmap_create(mddev);
1821                                 if (rv) {
1822                                         bitmap_destroy(mddev);
1823                                         mddev->bitmap_info.offset = 0;
1824                                 }
1825                                 mddev->pers->quiesce(mddev, 0);
1826                                 if (rv)
1827                                         return rv;
1828                         }
1829                 }
1830         }
1831         if (!mddev->external) {
1832                 /* Ensure new bitmap info is stored in
1833                  * metadata promptly.
1834                  */
1835                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1836                 md_wakeup_thread(mddev->thread);
1837         }
1838         return len;
1839 }
1840
1841 static struct md_sysfs_entry bitmap_location =
1842 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
1843
1844 static ssize_t
1845 timeout_show(mddev_t *mddev, char *page)
1846 {
1847         ssize_t len;
1848         unsigned long secs = mddev->bitmap_info.daemon_sleep / HZ;
1849         unsigned long jifs = mddev->bitmap_info.daemon_sleep % HZ;
1850
1851         len = sprintf(page, "%lu", secs);
1852         if (jifs)
1853                 len += sprintf(page+len, ".%03u", jiffies_to_msecs(jifs));
1854         len += sprintf(page+len, "\n");
1855         return len;
1856 }
1857
1858 static ssize_t
1859 timeout_store(mddev_t *mddev, const char *buf, size_t len)
1860 {
1861         /* timeout can be set at any time */
1862         unsigned long timeout;
1863         int rv = strict_strtoul_scaled(buf, &timeout, 4);
1864         if (rv)
1865                 return rv;
1866
1867         /* just to make sure we don't overflow... */
1868         if (timeout >= LONG_MAX / HZ)
1869                 return -EINVAL;
1870
1871         timeout = timeout * HZ / 10000;
1872
1873         if (timeout >= MAX_SCHEDULE_TIMEOUT)
1874                 timeout = MAX_SCHEDULE_TIMEOUT-1;
1875         if (timeout < 1)
1876                 timeout = 1;
1877         mddev->bitmap_info.daemon_sleep = timeout;
1878         if (mddev->thread) {
1879                 /* if thread->timeout is MAX_SCHEDULE_TIMEOUT, then
1880                  * the bitmap is all clean and we don't need to
1881                  * adjust the timeout right now
1882                  */
1883                 if (mddev->thread->timeout < MAX_SCHEDULE_TIMEOUT) {
1884                         mddev->thread->timeout = timeout;
1885                         md_wakeup_thread(mddev->thread);
1886                 }
1887         }
1888         return len;
1889 }
1890
1891 static struct md_sysfs_entry bitmap_timeout =
1892 __ATTR(time_base, S_IRUGO|S_IWUSR, timeout_show, timeout_store);
1893
1894 static ssize_t
1895 backlog_show(mddev_t *mddev, char *page)
1896 {
1897         return sprintf(page, "%lu\n", mddev->bitmap_info.max_write_behind);
1898 }
1899
1900 static ssize_t
1901 backlog_store(mddev_t *mddev, const char *buf, size_t len)
1902 {
1903         unsigned long backlog;
1904         int rv = strict_strtoul(buf, 10, &backlog);
1905         if (rv)
1906                 return rv;
1907         if (backlog > COUNTER_MAX)
1908                 return -EINVAL;
1909         mddev->bitmap_info.max_write_behind = backlog;
1910         return len;
1911 }
1912
1913 static struct md_sysfs_entry bitmap_backlog =
1914 __ATTR(backlog, S_IRUGO|S_IWUSR, backlog_show, backlog_store);
1915
1916 static ssize_t
1917 chunksize_show(mddev_t *mddev, char *page)
1918 {
1919         return sprintf(page, "%lu\n", mddev->bitmap_info.chunksize);
1920 }
1921
1922 static ssize_t
1923 chunksize_store(mddev_t *mddev, const char *buf, size_t len)
1924 {
1925         /* Can only be changed when no bitmap is active */
1926         int rv;
1927         unsigned long csize;
1928         if (mddev->bitmap)
1929                 return -EBUSY;
1930         rv = strict_strtoul(buf, 10, &csize);
1931         if (rv)
1932                 return rv;
1933         if (csize < 512 ||
1934             !is_power_of_2(csize))
1935                 return -EINVAL;
1936         mddev->bitmap_info.chunksize = csize;
1937         return len;
1938 }
1939
1940 static struct md_sysfs_entry bitmap_chunksize =
1941 __ATTR(chunksize, S_IRUGO|S_IWUSR, chunksize_show, chunksize_store);
1942
1943 static ssize_t metadata_show(mddev_t *mddev, char *page)
1944 {
1945         return sprintf(page, "%s\n", (mddev->bitmap_info.external
1946                                       ? "external" : "internal"));
1947 }
1948
1949 static ssize_t metadata_store(mddev_t *mddev, const char *buf, size_t len)
1950 {
1951         if (mddev->bitmap ||
1952             mddev->bitmap_info.file ||
1953             mddev->bitmap_info.offset)
1954                 return -EBUSY;
1955         if (strncmp(buf, "external", 8) == 0)
1956                 mddev->bitmap_info.external = 1;
1957         else if (strncmp(buf, "internal", 8) == 0)
1958                 mddev->bitmap_info.external = 0;
1959         else
1960                 return -EINVAL;
1961         return len;
1962 }
1963
1964 static struct md_sysfs_entry bitmap_metadata =
1965 __ATTR(metadata, S_IRUGO|S_IWUSR, metadata_show, metadata_store);
1966
1967 static ssize_t can_clear_show(mddev_t *mddev, char *page)
1968 {
1969         int len;
1970         if (mddev->bitmap)
1971                 len = sprintf(page, "%s\n", (mddev->bitmap->need_sync ?
1972                                              "false" : "true"));
1973         else
1974                 len = sprintf(page, "\n");
1975         return len;
1976 }
1977
1978 static ssize_t can_clear_store(mddev_t *mddev, const char *buf, size_t len)
1979 {
1980         if (mddev->bitmap == NULL)
1981                 return -ENOENT;
1982         if (strncmp(buf, "false", 5) == 0)
1983                 mddev->bitmap->need_sync = 1;
1984         else if (strncmp(buf, "true", 4) == 0) {
1985                 if (mddev->degraded)
1986                         return -EBUSY;
1987                 mddev->bitmap->need_sync = 0;
1988         } else
1989                 return -EINVAL;
1990         return len;
1991 }
1992
1993 static struct md_sysfs_entry bitmap_can_clear =
1994 __ATTR(can_clear, S_IRUGO|S_IWUSR, can_clear_show, can_clear_store);
1995
1996 static ssize_t
1997 behind_writes_used_show(mddev_t *mddev, char *page)
1998 {
1999         if (mddev->bitmap == NULL)
2000                 return sprintf(page, "0\n");
2001         return sprintf(page, "%lu\n",
2002                        mddev->bitmap->behind_writes_used);
2003 }
2004
2005 static ssize_t
2006 behind_writes_used_reset(mddev_t *mddev, const char *buf, size_t len)
2007 {
2008         if (mddev->bitmap)
2009                 mddev->bitmap->behind_writes_used = 0;
2010         return len;
2011 }
2012
2013 static struct md_sysfs_entry max_backlog_used =
2014 __ATTR(max_backlog_used, S_IRUGO | S_IWUSR,
2015        behind_writes_used_show, behind_writes_used_reset);
2016
2017 static struct attribute *md_bitmap_attrs[] = {
2018         &bitmap_location.attr,
2019         &bitmap_timeout.attr,
2020         &bitmap_backlog.attr,
2021         &bitmap_chunksize.attr,
2022         &bitmap_metadata.attr,
2023         &bitmap_can_clear.attr,
2024         &max_backlog_used.attr,
2025         NULL
2026 };
2027 struct attribute_group md_bitmap_group = {
2028         .name = "bitmap",
2029         .attrs = md_bitmap_attrs,
2030 };
2031