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

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