7666117738c75ca4eef600d6a1c754708667fa9e
[linux-2.6.git] / 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  * wait if count gets too high, wake when it drops to half.
17  */
18
19 #include <linux/module.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/timer.h>
24 #include <linux/sched.h>
25 #include <linux/list.h>
26 #include <linux/file.h>
27 #include <linux/mount.h>
28 #include <linux/buffer_head.h>
29 #include <linux/raid/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 //#define DPRINTK PRINTK /* set this NULL to avoid verbose debug output */
54 #define DPRINTK(x...) do { } while(0)
55
56 #ifndef PRINTK
57 #  if DEBUG > 0
58 #    define PRINTK(x...) printk(KERN_DEBUG x)
59 #  else
60 #    define PRINTK(x...)
61 #  endif
62 #endif
63
64 static inline char * bmname(struct bitmap *bitmap)
65 {
66         return bitmap->mddev ? mdname(bitmap->mddev) : "mdX";
67 }
68
69
70 /*
71  * just a placeholder - calls kmalloc for bitmap pages
72  */
73 static unsigned char *bitmap_alloc_page(struct bitmap *bitmap)
74 {
75         unsigned char *page;
76
77 #ifdef INJECT_FAULTS_1
78         page = NULL;
79 #else
80         page = kmalloc(PAGE_SIZE, GFP_NOIO);
81 #endif
82         if (!page)
83                 printk("%s: bitmap_alloc_page FAILED\n", bmname(bitmap));
84         else
85                 PRINTK("%s: bitmap_alloc_page: allocated page at %p\n",
86                         bmname(bitmap), page);
87         return page;
88 }
89
90 /*
91  * for now just a placeholder -- just calls kfree for bitmap pages
92  */
93 static void bitmap_free_page(struct bitmap *bitmap, unsigned char *page)
94 {
95         PRINTK("%s: bitmap_free_page: free page %p\n", bmname(bitmap), page);
96         kfree(page);
97 }
98
99 /*
100  * check a page and, if necessary, allocate it (or hijack it if the alloc fails)
101  *
102  * 1) check to see if this page is allocated, if it's not then try to alloc
103  * 2) if the alloc fails, set the page's hijacked flag so we'll use the
104  *    page pointer directly as a counter
105  *
106  * if we find our page, we increment the page's refcount so that it stays
107  * allocated while we're using it
108  */
109 static int bitmap_checkpage(struct bitmap *bitmap, unsigned long page, int create)
110 {
111         unsigned char *mappage;
112
113         if (page >= bitmap->pages) {
114                 /* This can happen if bitmap_start_sync goes beyond
115                  * End-of-device while looking for a whole page.
116                  * It is harmless.
117                  */
118                 return -EINVAL;
119         }
120
121
122         if (bitmap->bp[page].hijacked) /* it's hijacked, don't try to alloc */
123                 return 0;
124
125         if (bitmap->bp[page].map) /* page is already allocated, just return */
126                 return 0;
127
128         if (!create)
129                 return -ENOENT;
130
131         spin_unlock_irq(&bitmap->lock);
132
133         /* this page has not been allocated yet */
134
135         if ((mappage = bitmap_alloc_page(bitmap)) == 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                 spin_lock_irq(&bitmap->lock);
141                 if (!bitmap->bp[page].map)
142                         bitmap->bp[page].hijacked = 1;
143                 goto out;
144         }
145
146         /* got a page */
147
148         spin_lock_irq(&bitmap->lock);
149
150         /* recheck the page */
151
152         if (bitmap->bp[page].map || bitmap->bp[page].hijacked) {
153                 /* somebody beat us to getting the page */
154                 bitmap_free_page(bitmap, mappage);
155                 return 0;
156         }
157
158         /* no page was in place and we have one, so install it */
159
160         memset(mappage, 0, PAGE_SIZE);
161         bitmap->bp[page].map = mappage;
162         bitmap->missing_pages--;
163 out:
164         return 0;
165 }
166
167
168 /* if page is completely empty, put it back on the free list, or dealloc it */
169 /* if page was hijacked, unmark the flag so it might get alloced next time */
170 /* Note: lock should be held when calling this */
171 static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
172 {
173         char *ptr;
174
175         if (bitmap->bp[page].count) /* page is still busy */
176                 return;
177
178         /* page is no longer in use, it can be released */
179
180         if (bitmap->bp[page].hijacked) { /* page was hijacked, undo this now */
181                 bitmap->bp[page].hijacked = 0;
182                 bitmap->bp[page].map = NULL;
183                 return;
184         }
185
186         /* normal case, free the page */
187
188 #if 0
189 /* actually ... let's not.  We will probably need the page again exactly when
190  * memory is tight and we are flusing to disk
191  */
192         return;
193 #else
194         ptr = bitmap->bp[page].map;
195         bitmap->bp[page].map = NULL;
196         bitmap->missing_pages++;
197         bitmap_free_page(bitmap, ptr);
198         return;
199 #endif
200 }
201
202
203 /*
204  * bitmap file handling - read and write the bitmap file and its superblock
205  */
206
207 /*
208  * basic page I/O operations
209  */
210
211 /* IO operations when bitmap is stored near all superblocks */
212 static struct page *read_sb_page(mddev_t *mddev, long offset,
213                                  struct page *page,
214                                  unsigned long index, int size)
215 {
216         /* choose a good rdev and read the page from there */
217
218         mdk_rdev_t *rdev;
219         sector_t target;
220
221         if (!page)
222                 page = alloc_page(GFP_KERNEL);
223         if (!page)
224                 return ERR_PTR(-ENOMEM);
225
226         list_for_each_entry(rdev, &mddev->disks, same_set) {
227                 if (! test_bit(In_sync, &rdev->flags)
228                     || test_bit(Faulty, &rdev->flags))
229                         continue;
230
231                 target = rdev->sb_start + offset + index * (PAGE_SIZE/512);
232
233                 if (sync_page_io(rdev->bdev, target,
234                                  roundup(size, bdev_hardsect_size(rdev->bdev)),
235                                  page, READ)) {
236                         page->index = index;
237                         attach_page_buffers(page, NULL); /* so that free_buffer will
238                                                           * quietly no-op */
239                         return page;
240                 }
241         }
242         return ERR_PTR(-EIO);
243
244 }
245
246 static mdk_rdev_t *next_active_rdev(mdk_rdev_t *rdev, mddev_t *mddev)
247 {
248         /* Iterate the disks of an mddev, using rcu to protect access to the
249          * linked list, and raising the refcount of devices we return to ensure
250          * they don't disappear while in use.
251          * As devices are only added or removed when raid_disk is < 0 and
252          * nr_pending is 0 and In_sync is clear, the entries we return will
253          * still be in the same position on the list when we re-enter
254          * list_for_each_continue_rcu.
255          */
256         struct list_head *pos;
257         rcu_read_lock();
258         if (rdev == NULL)
259                 /* start at the beginning */
260                 pos = &mddev->disks;
261         else {
262                 /* release the previous rdev and start from there. */
263                 rdev_dec_pending(rdev, mddev);
264                 pos = &rdev->same_set;
265         }
266         list_for_each_continue_rcu(pos, &mddev->disks) {
267                 rdev = list_entry(pos, mdk_rdev_t, same_set);
268                 if (rdev->raid_disk >= 0 &&
269                     !test_bit(Faulty, &rdev->flags)) {
270                         /* this is a usable devices */
271                         atomic_inc(&rdev->nr_pending);
272                         rcu_read_unlock();
273                         return rdev;
274                 }
275         }
276         rcu_read_unlock();
277         return NULL;
278 }
279
280 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
281 {
282         mdk_rdev_t *rdev = NULL;
283         mddev_t *mddev = bitmap->mddev;
284
285         while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
286                         int size = PAGE_SIZE;
287                         if (page->index == bitmap->file_pages-1)
288                                 size = roundup(bitmap->last_page_size,
289                                                bdev_hardsect_size(rdev->bdev));
290                         /* Just make sure we aren't corrupting data or
291                          * metadata
292                          */
293                         if (bitmap->offset < 0) {
294                                 /* DATA  BITMAP METADATA  */
295                                 if (bitmap->offset
296                                     + (long)(page->index * (PAGE_SIZE/512))
297                                     + size/512 > 0)
298                                         /* bitmap runs in to metadata */
299                                         goto bad_alignment;
300                                 if (rdev->data_offset + mddev->size*2
301                                     > rdev->sb_start + bitmap->offset)
302                                         /* data runs in to bitmap */
303                                         goto bad_alignment;
304                         } else if (rdev->sb_start < rdev->data_offset) {
305                                 /* METADATA BITMAP DATA */
306                                 if (rdev->sb_start
307                                     + bitmap->offset
308                                     + page->index*(PAGE_SIZE/512) + size/512
309                                     > rdev->data_offset)
310                                         /* bitmap runs in to data */
311                                         goto bad_alignment;
312                         } else {
313                                 /* DATA METADATA BITMAP - no problems */
314                         }
315                         md_super_write(mddev, rdev,
316                                        rdev->sb_start + bitmap->offset
317                                        + page->index * (PAGE_SIZE/512),
318                                        size,
319                                        page);
320         }
321
322         if (wait)
323                 md_super_wait(mddev);
324         return 0;
325
326  bad_alignment:
327         rcu_read_unlock();
328         return -EINVAL;
329 }
330
331 static void bitmap_file_kick(struct bitmap *bitmap);
332 /*
333  * write out a page to a file
334  */
335 static void write_page(struct bitmap *bitmap, struct page *page, int wait)
336 {
337         struct buffer_head *bh;
338
339         if (bitmap->file == NULL) {
340                 switch (write_sb_page(bitmap, page, wait)) {
341                 case -EINVAL:
342                         bitmap->flags |= BITMAP_WRITE_ERROR;
343                 }
344         } else {
345
346                 bh = page_buffers(page);
347
348                 while (bh && bh->b_blocknr) {
349                         atomic_inc(&bitmap->pending_writes);
350                         set_buffer_locked(bh);
351                         set_buffer_mapped(bh);
352                         submit_bh(WRITE, bh);
353                         bh = bh->b_this_page;
354                 }
355
356                 if (wait) {
357                         wait_event(bitmap->write_wait,
358                                    atomic_read(&bitmap->pending_writes)==0);
359                 }
360         }
361         if (bitmap->flags & BITMAP_WRITE_ERROR)
362                 bitmap_file_kick(bitmap);
363 }
364
365 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
366 {
367         struct bitmap *bitmap = bh->b_private;
368         unsigned long flags;
369
370         if (!uptodate) {
371                 spin_lock_irqsave(&bitmap->lock, flags);
372                 bitmap->flags |= BITMAP_WRITE_ERROR;
373                 spin_unlock_irqrestore(&bitmap->lock, flags);
374         }
375         if (atomic_dec_and_test(&bitmap->pending_writes))
376                 wake_up(&bitmap->write_wait);
377 }
378
379 /* copied from buffer.c */
380 static void
381 __clear_page_buffers(struct page *page)
382 {
383         ClearPagePrivate(page);
384         set_page_private(page, 0);
385         page_cache_release(page);
386 }
387 static void free_buffers(struct page *page)
388 {
389         struct buffer_head *bh = page_buffers(page);
390
391         while (bh) {
392                 struct buffer_head *next = bh->b_this_page;
393                 free_buffer_head(bh);
394                 bh = next;
395         }
396         __clear_page_buffers(page);
397         put_page(page);
398 }
399
400 /* read a page from a file.
401  * We both read the page, and attach buffers to the page to record the
402  * address of each block (using bmap).  These addresses will be used
403  * to write the block later, completely bypassing the filesystem.
404  * This usage is similar to how swap files are handled, and allows us
405  * to write to a file with no concerns of memory allocation failing.
406  */
407 static struct page *read_page(struct file *file, unsigned long index,
408                               struct bitmap *bitmap,
409                               unsigned long count)
410 {
411         struct page *page = NULL;
412         struct inode *inode = file->f_path.dentry->d_inode;
413         struct buffer_head *bh;
414         sector_t block;
415
416         PRINTK("read bitmap file (%dB @ %Lu)\n", (int)PAGE_SIZE,
417                         (unsigned long long)index << PAGE_SHIFT);
418
419         page = alloc_page(GFP_KERNEL);
420         if (!page)
421                 page = ERR_PTR(-ENOMEM);
422         if (IS_ERR(page))
423                 goto out;
424
425         bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
426         if (!bh) {
427                 put_page(page);
428                 page = ERR_PTR(-ENOMEM);
429                 goto out;
430         }
431         attach_page_buffers(page, bh);
432         block = index << (PAGE_SHIFT - inode->i_blkbits);
433         while (bh) {
434                 if (count == 0)
435                         bh->b_blocknr = 0;
436                 else {
437                         bh->b_blocknr = bmap(inode, block);
438                         if (bh->b_blocknr == 0) {
439                                 /* Cannot use this file! */
440                                 free_buffers(page);
441                                 page = ERR_PTR(-EINVAL);
442                                 goto out;
443                         }
444                         bh->b_bdev = inode->i_sb->s_bdev;
445                         if (count < (1<<inode->i_blkbits))
446                                 count = 0;
447                         else
448                                 count -= (1<<inode->i_blkbits);
449
450                         bh->b_end_io = end_bitmap_write;
451                         bh->b_private = bitmap;
452                         atomic_inc(&bitmap->pending_writes);
453                         set_buffer_locked(bh);
454                         set_buffer_mapped(bh);
455                         submit_bh(READ, bh);
456                 }
457                 block++;
458                 bh = bh->b_this_page;
459         }
460         page->index = index;
461
462         wait_event(bitmap->write_wait,
463                    atomic_read(&bitmap->pending_writes)==0);
464         if (bitmap->flags & BITMAP_WRITE_ERROR) {
465                 free_buffers(page);
466                 page = ERR_PTR(-EIO);
467         }
468 out:
469         if (IS_ERR(page))
470                 printk(KERN_ALERT "md: bitmap read error: (%dB @ %Lu): %ld\n",
471                         (int)PAGE_SIZE,
472                         (unsigned long long)index << PAGE_SHIFT,
473                         PTR_ERR(page));
474         return page;
475 }
476
477 /*
478  * bitmap file superblock operations
479  */
480
481 /* update the event counter and sync the superblock to disk */
482 void bitmap_update_sb(struct bitmap *bitmap)
483 {
484         bitmap_super_t *sb;
485         unsigned long flags;
486
487         if (!bitmap || !bitmap->mddev) /* no bitmap for this array */
488                 return;
489         spin_lock_irqsave(&bitmap->lock, flags);
490         if (!bitmap->sb_page) { /* no superblock */
491                 spin_unlock_irqrestore(&bitmap->lock, flags);
492                 return;
493         }
494         spin_unlock_irqrestore(&bitmap->lock, flags);
495         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
496         sb->events = cpu_to_le64(bitmap->mddev->events);
497         if (bitmap->mddev->events < bitmap->events_cleared) {
498                 /* rocking back to read-only */
499                 bitmap->events_cleared = bitmap->mddev->events;
500                 sb->events_cleared = cpu_to_le64(bitmap->events_cleared);
501         }
502         kunmap_atomic(sb, KM_USER0);
503         write_page(bitmap, bitmap->sb_page, 1);
504 }
505
506 /* print out the bitmap file superblock */
507 void bitmap_print_sb(struct bitmap *bitmap)
508 {
509         bitmap_super_t *sb;
510
511         if (!bitmap || !bitmap->sb_page)
512                 return;
513         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
514         printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
515         printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
516         printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
517         printk(KERN_DEBUG "          uuid: %08x.%08x.%08x.%08x\n",
518                                         *(__u32 *)(sb->uuid+0),
519                                         *(__u32 *)(sb->uuid+4),
520                                         *(__u32 *)(sb->uuid+8),
521                                         *(__u32 *)(sb->uuid+12));
522         printk(KERN_DEBUG "        events: %llu\n",
523                         (unsigned long long) le64_to_cpu(sb->events));
524         printk(KERN_DEBUG "events cleared: %llu\n",
525                         (unsigned long long) le64_to_cpu(sb->events_cleared));
526         printk(KERN_DEBUG "         state: %08x\n", le32_to_cpu(sb->state));
527         printk(KERN_DEBUG "     chunksize: %d B\n", le32_to_cpu(sb->chunksize));
528         printk(KERN_DEBUG "  daemon sleep: %ds\n", le32_to_cpu(sb->daemon_sleep));
529         printk(KERN_DEBUG "     sync size: %llu KB\n",
530                         (unsigned long long)le64_to_cpu(sb->sync_size)/2);
531         printk(KERN_DEBUG "max write behind: %d\n", le32_to_cpu(sb->write_behind));
532         kunmap_atomic(sb, KM_USER0);
533 }
534
535 /* read the superblock from the bitmap file and initialize some bitmap fields */
536 static int bitmap_read_sb(struct bitmap *bitmap)
537 {
538         char *reason = NULL;
539         bitmap_super_t *sb;
540         unsigned long chunksize, daemon_sleep, write_behind;
541         unsigned long long events;
542         int err = -EINVAL;
543
544         /* page 0 is the superblock, read it... */
545         if (bitmap->file) {
546                 loff_t isize = i_size_read(bitmap->file->f_mapping->host);
547                 int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
548
549                 bitmap->sb_page = read_page(bitmap->file, 0, bitmap, bytes);
550         } else {
551                 bitmap->sb_page = read_sb_page(bitmap->mddev, bitmap->offset,
552                                                NULL,
553                                                0, sizeof(bitmap_super_t));
554         }
555         if (IS_ERR(bitmap->sb_page)) {
556                 err = PTR_ERR(bitmap->sb_page);
557                 bitmap->sb_page = NULL;
558                 return err;
559         }
560
561         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
562
563         chunksize = le32_to_cpu(sb->chunksize);
564         daemon_sleep = le32_to_cpu(sb->daemon_sleep);
565         write_behind = le32_to_cpu(sb->write_behind);
566
567         /* verify that the bitmap-specific fields are valid */
568         if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
569                 reason = "bad magic";
570         else if (le32_to_cpu(sb->version) < BITMAP_MAJOR_LO ||
571                  le32_to_cpu(sb->version) > BITMAP_MAJOR_HI)
572                 reason = "unrecognized superblock version";
573         else if (chunksize < 512)
574                 reason = "bitmap chunksize too small";
575         else if ((1 << ffz(~chunksize)) != chunksize)
576                 reason = "bitmap chunksize not a power of 2";
577         else if (daemon_sleep < 1 || daemon_sleep > MAX_SCHEDULE_TIMEOUT / HZ)
578                 reason = "daemon sleep period out of range";
579         else if (write_behind > COUNTER_MAX)
580                 reason = "write-behind limit out of range (0 - 16383)";
581         if (reason) {
582                 printk(KERN_INFO "%s: invalid bitmap file superblock: %s\n",
583                         bmname(bitmap), reason);
584                 goto out;
585         }
586
587         /* keep the array size field of the bitmap superblock up to date */
588         sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
589
590         if (!bitmap->mddev->persistent)
591                 goto success;
592
593         /*
594          * if we have a persistent array superblock, compare the
595          * bitmap's UUID and event counter to the mddev's
596          */
597         if (memcmp(sb->uuid, bitmap->mddev->uuid, 16)) {
598                 printk(KERN_INFO "%s: bitmap superblock UUID mismatch\n",
599                         bmname(bitmap));
600                 goto out;
601         }
602         events = le64_to_cpu(sb->events);
603         if (events < bitmap->mddev->events) {
604                 printk(KERN_INFO "%s: bitmap file is out of date (%llu < %llu) "
605                         "-- forcing full recovery\n", bmname(bitmap), events,
606                         (unsigned long long) bitmap->mddev->events);
607                 sb->state |= cpu_to_le32(BITMAP_STALE);
608         }
609 success:
610         /* assign fields using values from superblock */
611         bitmap->chunksize = chunksize;
612         bitmap->daemon_sleep = daemon_sleep;
613         bitmap->daemon_lastrun = jiffies;
614         bitmap->max_write_behind = write_behind;
615         bitmap->flags |= le32_to_cpu(sb->state);
616         if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
617                 bitmap->flags |= BITMAP_HOSTENDIAN;
618         bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
619         if (sb->state & cpu_to_le32(BITMAP_STALE))
620                 bitmap->events_cleared = bitmap->mddev->events;
621         err = 0;
622 out:
623         kunmap_atomic(sb, KM_USER0);
624         if (err)
625                 bitmap_print_sb(bitmap);
626         return err;
627 }
628
629 enum bitmap_mask_op {
630         MASK_SET,
631         MASK_UNSET
632 };
633
634 /* record the state of the bitmap in the superblock.  Return the old value */
635 static int bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
636                              enum bitmap_mask_op op)
637 {
638         bitmap_super_t *sb;
639         unsigned long flags;
640         int old;
641
642         spin_lock_irqsave(&bitmap->lock, flags);
643         if (!bitmap->sb_page) { /* can't set the state */
644                 spin_unlock_irqrestore(&bitmap->lock, flags);
645                 return 0;
646         }
647         spin_unlock_irqrestore(&bitmap->lock, flags);
648         sb = (bitmap_super_t *)kmap_atomic(bitmap->sb_page, KM_USER0);
649         old = le32_to_cpu(sb->state) & bits;
650         switch (op) {
651                 case MASK_SET: sb->state |= cpu_to_le32(bits);
652                                 break;
653                 case MASK_UNSET: sb->state &= cpu_to_le32(~bits);
654                                 break;
655                 default: BUG();
656         }
657         kunmap_atomic(sb, KM_USER0);
658         return old;
659 }
660
661 /*
662  * general bitmap file operations
663  */
664
665 /* calculate the index of the page that contains this bit */
666 static inline unsigned long file_page_index(unsigned long chunk)
667 {
668         return CHUNK_BIT_OFFSET(chunk) >> PAGE_BIT_SHIFT;
669 }
670
671 /* calculate the (bit) offset of this bit within a page */
672 static inline unsigned long file_page_offset(unsigned long chunk)
673 {
674         return CHUNK_BIT_OFFSET(chunk) & (PAGE_BITS - 1);
675 }
676
677 /*
678  * return a pointer to the page in the filemap that contains the given bit
679  *
680  * this lookup is complicated by the fact that the bitmap sb might be exactly
681  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
682  * 0 or page 1
683  */
684 static inline struct page *filemap_get_page(struct bitmap *bitmap,
685                                         unsigned long chunk)
686 {
687         if (file_page_index(chunk) >= bitmap->file_pages) return NULL;
688         return bitmap->filemap[file_page_index(chunk) - file_page_index(0)];
689 }
690
691
692 static void bitmap_file_unmap(struct bitmap *bitmap)
693 {
694         struct page **map, *sb_page;
695         unsigned long *attr;
696         int pages;
697         unsigned long flags;
698
699         spin_lock_irqsave(&bitmap->lock, flags);
700         map = bitmap->filemap;
701         bitmap->filemap = NULL;
702         attr = bitmap->filemap_attr;
703         bitmap->filemap_attr = NULL;
704         pages = bitmap->file_pages;
705         bitmap->file_pages = 0;
706         sb_page = bitmap->sb_page;
707         bitmap->sb_page = NULL;
708         spin_unlock_irqrestore(&bitmap->lock, flags);
709
710         while (pages--)
711                 if (map[pages]->index != 0) /* 0 is sb_page, release it below */
712                         free_buffers(map[pages]);
713         kfree(map);
714         kfree(attr);
715
716         if (sb_page)
717                 free_buffers(sb_page);
718 }
719
720 static void bitmap_file_put(struct bitmap *bitmap)
721 {
722         struct file *file;
723         unsigned long flags;
724
725         spin_lock_irqsave(&bitmap->lock, flags);
726         file = bitmap->file;
727         bitmap->file = NULL;
728         spin_unlock_irqrestore(&bitmap->lock, flags);
729
730         if (file)
731                 wait_event(bitmap->write_wait,
732                            atomic_read(&bitmap->pending_writes)==0);
733         bitmap_file_unmap(bitmap);
734
735         if (file) {
736                 struct inode *inode = file->f_path.dentry->d_inode;
737                 invalidate_mapping_pages(inode->i_mapping, 0, -1);
738                 fput(file);
739         }
740 }
741
742
743 /*
744  * bitmap_file_kick - if an error occurs while manipulating the bitmap file
745  * then it is no longer reliable, so we stop using it and we mark the file
746  * as failed in the superblock
747  */
748 static void bitmap_file_kick(struct bitmap *bitmap)
749 {
750         char *path, *ptr = NULL;
751
752         if (bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET) == 0) {
753                 bitmap_update_sb(bitmap);
754
755                 if (bitmap->file) {
756                         path = kmalloc(PAGE_SIZE, GFP_KERNEL);
757                         if (path)
758                                 ptr = d_path(&bitmap->file->f_path, path,
759                                              PAGE_SIZE);
760
761
762                         printk(KERN_ALERT
763                               "%s: kicking failed bitmap file %s from array!\n",
764                               bmname(bitmap), IS_ERR(ptr) ? "" : ptr);
765
766                         kfree(path);
767                 } else
768                         printk(KERN_ALERT
769                                "%s: disabling internal bitmap due to errors\n",
770                                bmname(bitmap));
771         }
772
773         bitmap_file_put(bitmap);
774
775         return;
776 }
777
778 enum bitmap_page_attr {
779         BITMAP_PAGE_DIRTY = 0, // there are set bits that need to be synced
780         BITMAP_PAGE_CLEAN = 1, // there are bits that might need to be cleared
781         BITMAP_PAGE_NEEDWRITE=2, // there are cleared bits that need to be synced
782 };
783
784 static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
785                                 enum bitmap_page_attr attr)
786 {
787         __set_bit((page->index<<2) + attr, bitmap->filemap_attr);
788 }
789
790 static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
791                                 enum bitmap_page_attr attr)
792 {
793         __clear_bit((page->index<<2) + attr, bitmap->filemap_attr);
794 }
795
796 static inline unsigned long test_page_attr(struct bitmap *bitmap, struct page *page,
797                                            enum bitmap_page_attr attr)
798 {
799         return test_bit((page->index<<2) + attr, bitmap->filemap_attr);
800 }
801
802 /*
803  * bitmap_file_set_bit -- called before performing a write to the md device
804  * to set (and eventually sync) a particular bit in the bitmap file
805  *
806  * we set the bit immediately, then we record the page number so that
807  * when an unplug occurs, we can flush the dirty pages out to disk
808  */
809 static void bitmap_file_set_bit(struct bitmap *bitmap, sector_t block)
810 {
811         unsigned long bit;
812         struct page *page;
813         void *kaddr;
814         unsigned long chunk = block >> CHUNK_BLOCK_SHIFT(bitmap);
815
816         if (!bitmap->filemap) {
817                 return;
818         }
819
820         page = filemap_get_page(bitmap, chunk);
821         if (!page) return;
822         bit = file_page_offset(chunk);
823
824         /* set the bit */
825         kaddr = kmap_atomic(page, KM_USER0);
826         if (bitmap->flags & BITMAP_HOSTENDIAN)
827                 set_bit(bit, kaddr);
828         else
829                 ext2_set_bit(bit, kaddr);
830         kunmap_atomic(kaddr, KM_USER0);
831         PRINTK("set file bit %lu page %lu\n", bit, page->index);
832
833         /* record page number so it gets flushed to disk when unplug occurs */
834         set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
835
836 }
837
838 /* this gets called when the md device is ready to unplug its underlying
839  * (slave) device queues -- before we let any writes go down, we need to
840  * sync the dirty pages of the bitmap file to disk */
841 void bitmap_unplug(struct bitmap *bitmap)
842 {
843         unsigned long i, flags;
844         int dirty, need_write;
845         struct page *page;
846         int wait = 0;
847
848         if (!bitmap)
849                 return;
850
851         /* look at each page to see if there are any set bits that need to be
852          * flushed out to disk */
853         for (i = 0; i < bitmap->file_pages; i++) {
854                 spin_lock_irqsave(&bitmap->lock, flags);
855                 if (!bitmap->filemap) {
856                         spin_unlock_irqrestore(&bitmap->lock, flags);
857                         return;
858                 }
859                 page = bitmap->filemap[i];
860                 dirty = test_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
861                 need_write = test_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
862                 clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
863                 clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
864                 if (dirty)
865                         wait = 1;
866                 spin_unlock_irqrestore(&bitmap->lock, flags);
867
868                 if (dirty | need_write)
869                         write_page(bitmap, page, 0);
870         }
871         if (wait) { /* if any writes were performed, we need to wait on them */
872                 if (bitmap->file)
873                         wait_event(bitmap->write_wait,
874                                    atomic_read(&bitmap->pending_writes)==0);
875                 else
876                         md_super_wait(bitmap->mddev);
877         }
878         if (bitmap->flags & BITMAP_WRITE_ERROR)
879                 bitmap_file_kick(bitmap);
880 }
881
882 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed);
883 /* * bitmap_init_from_disk -- called at bitmap_create time to initialize
884  * the in-memory bitmap from the on-disk bitmap -- also, sets up the
885  * memory mapping of the bitmap file
886  * Special cases:
887  *   if there's no bitmap file, or if the bitmap file had been
888  *   previously kicked from the array, we mark all the bits as
889  *   1's in order to cause a full resync.
890  *
891  * We ignore all bits for sectors that end earlier than 'start'.
892  * This is used when reading an out-of-date bitmap...
893  */
894 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
895 {
896         unsigned long i, chunks, index, oldindex, bit;
897         struct page *page = NULL, *oldpage = NULL;
898         unsigned long num_pages, bit_cnt = 0;
899         struct file *file;
900         unsigned long bytes, offset;
901         int outofdate;
902         int ret = -ENOSPC;
903         void *paddr;
904
905         chunks = bitmap->chunks;
906         file = bitmap->file;
907
908         BUG_ON(!file && !bitmap->offset);
909
910 #ifdef INJECT_FAULTS_3
911         outofdate = 1;
912 #else
913         outofdate = bitmap->flags & BITMAP_STALE;
914 #endif
915         if (outofdate)
916                 printk(KERN_INFO "%s: bitmap file is out of date, doing full "
917                         "recovery\n", bmname(bitmap));
918
919         bytes = (chunks + 7) / 8;
920
921         num_pages = (bytes + sizeof(bitmap_super_t) + PAGE_SIZE - 1) / PAGE_SIZE;
922
923         if (file && i_size_read(file->f_mapping->host) < bytes + sizeof(bitmap_super_t)) {
924                 printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
925                         bmname(bitmap),
926                         (unsigned long) i_size_read(file->f_mapping->host),
927                         bytes + sizeof(bitmap_super_t));
928                 goto err;
929         }
930
931         ret = -ENOMEM;
932
933         bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
934         if (!bitmap->filemap)
935                 goto err;
936
937         /* We need 4 bits per page, rounded up to a multiple of sizeof(unsigned long) */
938         bitmap->filemap_attr = kzalloc(
939                 roundup( DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
940                 GFP_KERNEL);
941         if (!bitmap->filemap_attr)
942                 goto err;
943
944         oldindex = ~0L;
945
946         for (i = 0; i < chunks; i++) {
947                 int b;
948                 index = file_page_index(i);
949                 bit = file_page_offset(i);
950                 if (index != oldindex) { /* this is a new page, read it in */
951                         int count;
952                         /* unmap the old page, we're done with it */
953                         if (index == num_pages-1)
954                                 count = bytes + sizeof(bitmap_super_t)
955                                         - index * PAGE_SIZE;
956                         else
957                                 count = PAGE_SIZE;
958                         if (index == 0) {
959                                 /*
960                                  * if we're here then the superblock page
961                                  * contains some bits (PAGE_SIZE != sizeof sb)
962                                  * we've already read it in, so just use it
963                                  */
964                                 page = bitmap->sb_page;
965                                 offset = sizeof(bitmap_super_t);
966                                 if (!file)
967                                         read_sb_page(bitmap->mddev,
968                                                      bitmap->offset,
969                                                      page,
970                                                      index, count);
971                         } else if (file) {
972                                 page = read_page(file, index, bitmap, count);
973                                 offset = 0;
974                         } else {
975                                 page = read_sb_page(bitmap->mddev, bitmap->offset,
976                                                     NULL,
977                                                     index, count);
978                                 offset = 0;
979                         }
980                         if (IS_ERR(page)) { /* read error */
981                                 ret = PTR_ERR(page);
982                                 goto err;
983                         }
984
985                         oldindex = index;
986                         oldpage = page;
987
988                         if (outofdate) {
989                                 /*
990                                  * if bitmap is out of date, dirty the
991                                  * whole page and write it out
992                                  */
993                                 paddr = kmap_atomic(page, KM_USER0);
994                                 memset(paddr + offset, 0xff,
995                                        PAGE_SIZE - offset);
996                                 kunmap_atomic(paddr, KM_USER0);
997                                 write_page(bitmap, page, 1);
998
999                                 ret = -EIO;
1000                                 if (bitmap->flags & BITMAP_WRITE_ERROR) {
1001                                         /* release, page not in filemap yet */
1002                                         put_page(page);
1003                                         goto err;
1004                                 }
1005                         }
1006
1007                         bitmap->filemap[bitmap->file_pages++] = page;
1008                         bitmap->last_page_size = count;
1009                 }
1010                 paddr = kmap_atomic(page, KM_USER0);
1011                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1012                         b = test_bit(bit, paddr);
1013                 else
1014                         b = ext2_test_bit(bit, paddr);
1015                 kunmap_atomic(paddr, KM_USER0);
1016                 if (b) {
1017                         /* if the disk bit is set, set the memory bit */
1018                         bitmap_set_memory_bits(bitmap, i << CHUNK_BLOCK_SHIFT(bitmap),
1019                                                ((i+1) << (CHUNK_BLOCK_SHIFT(bitmap)) >= start)
1020                                 );
1021                         bit_cnt++;
1022                         set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1023                 }
1024         }
1025
1026         /* everything went OK */
1027         ret = 0;
1028         bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
1029
1030         if (bit_cnt) { /* Kick recovery if any bits were set */
1031                 set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
1032                 md_wakeup_thread(bitmap->mddev->thread);
1033         }
1034
1035         printk(KERN_INFO "%s: bitmap initialized from disk: "
1036                 "read %lu/%lu pages, set %lu bits\n",
1037                 bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt);
1038
1039         return 0;
1040
1041  err:
1042         printk(KERN_INFO "%s: bitmap initialisation failed: %d\n",
1043                bmname(bitmap), ret);
1044         return ret;
1045 }
1046
1047 void bitmap_write_all(struct bitmap *bitmap)
1048 {
1049         /* We don't actually write all bitmap blocks here,
1050          * just flag them as needing to be written
1051          */
1052         int i;
1053
1054         for (i=0; i < bitmap->file_pages; i++)
1055                 set_page_attr(bitmap, bitmap->filemap[i],
1056                               BITMAP_PAGE_NEEDWRITE);
1057 }
1058
1059
1060 static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
1061 {
1062         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1063         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1064         bitmap->bp[page].count += inc;
1065 /*
1066         if (page == 0) printk("count page 0, offset %llu: %d gives %d\n",
1067                               (unsigned long long)offset, inc, bitmap->bp[page].count);
1068 */
1069         bitmap_checkfree(bitmap, page);
1070 }
1071 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1072                                             sector_t offset, int *blocks,
1073                                             int create);
1074
1075 /*
1076  * bitmap daemon -- periodically wakes up to clean bits and flush pages
1077  *                      out to disk
1078  */
1079
1080 void bitmap_daemon_work(struct bitmap *bitmap)
1081 {
1082         unsigned long j;
1083         unsigned long flags;
1084         struct page *page = NULL, *lastpage = NULL;
1085         int blocks;
1086         void *paddr;
1087
1088         if (bitmap == NULL)
1089                 return;
1090         if (time_before(jiffies, bitmap->daemon_lastrun + bitmap->daemon_sleep*HZ))
1091                 goto done;
1092
1093         bitmap->daemon_lastrun = jiffies;
1094         if (bitmap->allclean) {
1095                 bitmap->mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1096                 return;
1097         }
1098         bitmap->allclean = 1;
1099
1100         for (j = 0; j < bitmap->chunks; j++) {
1101                 bitmap_counter_t *bmc;
1102                 spin_lock_irqsave(&bitmap->lock, flags);
1103                 if (!bitmap->filemap) {
1104                         /* error or shutdown */
1105                         spin_unlock_irqrestore(&bitmap->lock, flags);
1106                         break;
1107                 }
1108
1109                 page = filemap_get_page(bitmap, j);
1110
1111                 if (page != lastpage) {
1112                         /* skip this page unless it's marked as needing cleaning */
1113                         if (!test_page_attr(bitmap, page, BITMAP_PAGE_CLEAN)) {
1114                                 int need_write = test_page_attr(bitmap, page,
1115                                                                 BITMAP_PAGE_NEEDWRITE);
1116                                 if (need_write)
1117                                         clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
1118
1119                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1120                                 if (need_write) {
1121                                         write_page(bitmap, page, 0);
1122                                         bitmap->allclean = 0;
1123                                 }
1124                                 continue;
1125                         }
1126
1127                         /* grab the new page, sync and release the old */
1128                         if (lastpage != NULL) {
1129                                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1130                                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1131                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1132                                         write_page(bitmap, lastpage, 0);
1133                                 } else {
1134                                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1135                                         spin_unlock_irqrestore(&bitmap->lock, flags);
1136                                 }
1137                         } else
1138                                 spin_unlock_irqrestore(&bitmap->lock, flags);
1139                         lastpage = page;
1140
1141                         /* We are possibly going to clear some bits, so make
1142                          * sure that events_cleared is up-to-date.
1143                          */
1144                         if (bitmap->need_sync) {
1145                                 bitmap_super_t *sb;
1146                                 bitmap->need_sync = 0;
1147                                 sb = kmap_atomic(bitmap->sb_page, KM_USER0);
1148                                 sb->events_cleared =
1149                                         cpu_to_le64(bitmap->events_cleared);
1150                                 kunmap_atomic(sb, KM_USER0);
1151                                 write_page(bitmap, bitmap->sb_page, 1);
1152                         }
1153                         spin_lock_irqsave(&bitmap->lock, flags);
1154                         clear_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1155                 }
1156                 bmc = bitmap_get_counter(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1157                                         &blocks, 0);
1158                 if (bmc) {
1159 /*
1160   if (j < 100) printk("bitmap: j=%lu, *bmc = 0x%x\n", j, *bmc);
1161 */
1162                         if (*bmc)
1163                                 bitmap->allclean = 0;
1164
1165                         if (*bmc == 2) {
1166                                 *bmc=1; /* maybe clear the bit next time */
1167                                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1168                         } else if (*bmc == 1) {
1169                                 /* we can clear the bit */
1170                                 *bmc = 0;
1171                                 bitmap_count_page(bitmap, j << CHUNK_BLOCK_SHIFT(bitmap),
1172                                                   -1);
1173
1174                                 /* clear the bit */
1175                                 paddr = kmap_atomic(page, KM_USER0);
1176                                 if (bitmap->flags & BITMAP_HOSTENDIAN)
1177                                         clear_bit(file_page_offset(j), paddr);
1178                                 else
1179                                         ext2_clear_bit(file_page_offset(j), paddr);
1180                                 kunmap_atomic(paddr, KM_USER0);
1181                         }
1182                 }
1183                 spin_unlock_irqrestore(&bitmap->lock, flags);
1184         }
1185
1186         /* now sync the final page */
1187         if (lastpage != NULL) {
1188                 spin_lock_irqsave(&bitmap->lock, flags);
1189                 if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
1190                         clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1191                         spin_unlock_irqrestore(&bitmap->lock, flags);
1192                         write_page(bitmap, lastpage, 0);
1193                 } else {
1194                         set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
1195                         spin_unlock_irqrestore(&bitmap->lock, flags);
1196                 }
1197         }
1198
1199  done:
1200         if (bitmap->allclean == 0)
1201                 bitmap->mddev->thread->timeout = bitmap->daemon_sleep * HZ;
1202 }
1203
1204 static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
1205                                             sector_t offset, int *blocks,
1206                                             int create)
1207 {
1208         /* If 'create', we might release the lock and reclaim it.
1209          * The lock must have been taken with interrupts enabled.
1210          * If !create, we don't release the lock.
1211          */
1212         sector_t chunk = offset >> CHUNK_BLOCK_SHIFT(bitmap);
1213         unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
1214         unsigned long pageoff = (chunk & PAGE_COUNTER_MASK) << COUNTER_BYTE_SHIFT;
1215         sector_t csize;
1216
1217         if (bitmap_checkpage(bitmap, page, create) < 0) {
1218                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1219                 *blocks = csize - (offset & (csize- 1));
1220                 return NULL;
1221         }
1222         /* now locked ... */
1223
1224         if (bitmap->bp[page].hijacked) { /* hijacked pointer */
1225                 /* should we use the first or second counter field
1226                  * of the hijacked pointer? */
1227                 int hi = (pageoff > PAGE_COUNTER_MASK);
1228                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap) +
1229                                           PAGE_COUNTER_SHIFT - 1);
1230                 *blocks = csize - (offset & (csize- 1));
1231                 return  &((bitmap_counter_t *)
1232                           &bitmap->bp[page].map)[hi];
1233         } else { /* page is allocated */
1234                 csize = ((sector_t)1) << (CHUNK_BLOCK_SHIFT(bitmap));
1235                 *blocks = csize - (offset & (csize- 1));
1236                 return (bitmap_counter_t *)
1237                         &(bitmap->bp[page].map[pageoff]);
1238         }
1239 }
1240
1241 int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors, int behind)
1242 {
1243         if (!bitmap) return 0;
1244
1245         if (behind) {
1246                 atomic_inc(&bitmap->behind_writes);
1247                 PRINTK(KERN_DEBUG "inc write-behind count %d/%d\n",
1248                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1249         }
1250
1251         while (sectors) {
1252                 int blocks;
1253                 bitmap_counter_t *bmc;
1254
1255                 spin_lock_irq(&bitmap->lock);
1256                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
1257                 if (!bmc) {
1258                         spin_unlock_irq(&bitmap->lock);
1259                         return 0;
1260                 }
1261
1262                 if (unlikely((*bmc & COUNTER_MAX) == COUNTER_MAX)) {
1263                         DEFINE_WAIT(__wait);
1264                         /* note that it is safe to do the prepare_to_wait
1265                          * after the test as long as we do it before dropping
1266                          * the spinlock.
1267                          */
1268                         prepare_to_wait(&bitmap->overflow_wait, &__wait,
1269                                         TASK_UNINTERRUPTIBLE);
1270                         spin_unlock_irq(&bitmap->lock);
1271                         blk_unplug(bitmap->mddev->queue);
1272                         schedule();
1273                         finish_wait(&bitmap->overflow_wait, &__wait);
1274                         continue;
1275                 }
1276
1277                 switch(*bmc) {
1278                 case 0:
1279                         bitmap_file_set_bit(bitmap, offset);
1280                         bitmap_count_page(bitmap,offset, 1);
1281                         blk_plug_device_unlocked(bitmap->mddev->queue);
1282                         /* fall through */
1283                 case 1:
1284                         *bmc = 2;
1285                 }
1286
1287                 (*bmc)++;
1288
1289                 spin_unlock_irq(&bitmap->lock);
1290
1291                 offset += blocks;
1292                 if (sectors > blocks)
1293                         sectors -= blocks;
1294                 else sectors = 0;
1295         }
1296         bitmap->allclean = 0;
1297         return 0;
1298 }
1299
1300 void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
1301                      int success, int behind)
1302 {
1303         if (!bitmap) return;
1304         if (behind) {
1305                 atomic_dec(&bitmap->behind_writes);
1306                 PRINTK(KERN_DEBUG "dec write-behind count %d/%d\n",
1307                   atomic_read(&bitmap->behind_writes), bitmap->max_write_behind);
1308         }
1309         if (bitmap->mddev->degraded)
1310                 /* Never clear bits or update events_cleared when degraded */
1311                 success = 0;
1312
1313         while (sectors) {
1314                 int blocks;
1315                 unsigned long flags;
1316                 bitmap_counter_t *bmc;
1317
1318                 spin_lock_irqsave(&bitmap->lock, flags);
1319                 bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
1320                 if (!bmc) {
1321                         spin_unlock_irqrestore(&bitmap->lock, flags);
1322                         return;
1323                 }
1324
1325                 if (success &&
1326                     bitmap->events_cleared < bitmap->mddev->events) {
1327                         bitmap->events_cleared = bitmap->mddev->events;
1328                         bitmap->need_sync = 1;
1329                 }
1330
1331                 if (!success && ! (*bmc & NEEDED_MASK))
1332                         *bmc |= NEEDED_MASK;
1333
1334                 if ((*bmc & COUNTER_MAX) == COUNTER_MAX)
1335                         wake_up(&bitmap->overflow_wait);
1336
1337                 (*bmc)--;
1338                 if (*bmc <= 2) {
1339                         set_page_attr(bitmap,
1340                                       filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1341                                       BITMAP_PAGE_CLEAN);
1342                 }
1343                 spin_unlock_irqrestore(&bitmap->lock, flags);
1344                 offset += blocks;
1345                 if (sectors > blocks)
1346                         sectors -= blocks;
1347                 else sectors = 0;
1348         }
1349 }
1350
1351 static int __bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1352                                int degraded)
1353 {
1354         bitmap_counter_t *bmc;
1355         int rv;
1356         if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
1357                 *blocks = 1024;
1358                 return 1; /* always resync if no bitmap */
1359         }
1360         spin_lock_irq(&bitmap->lock);
1361         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1362         rv = 0;
1363         if (bmc) {
1364                 /* locked */
1365                 if (RESYNC(*bmc))
1366                         rv = 1;
1367                 else if (NEEDED(*bmc)) {
1368                         rv = 1;
1369                         if (!degraded) { /* don't set/clear bits if degraded */
1370                                 *bmc |= RESYNC_MASK;
1371                                 *bmc &= ~NEEDED_MASK;
1372                         }
1373                 }
1374         }
1375         spin_unlock_irq(&bitmap->lock);
1376         bitmap->allclean = 0;
1377         return rv;
1378 }
1379
1380 int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks,
1381                       int degraded)
1382 {
1383         /* bitmap_start_sync must always report on multiples of whole
1384          * pages, otherwise resync (which is very PAGE_SIZE based) will
1385          * get confused.
1386          * So call __bitmap_start_sync repeatedly (if needed) until
1387          * At least PAGE_SIZE>>9 blocks are covered.
1388          * Return the 'or' of the result.
1389          */
1390         int rv = 0;
1391         int blocks1;
1392
1393         *blocks = 0;
1394         while (*blocks < (PAGE_SIZE>>9)) {
1395                 rv |= __bitmap_start_sync(bitmap, offset,
1396                                           &blocks1, degraded);
1397                 offset += blocks1;
1398                 *blocks += blocks1;
1399         }
1400         return rv;
1401 }
1402
1403 void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted)
1404 {
1405         bitmap_counter_t *bmc;
1406         unsigned long flags;
1407 /*
1408         if (offset == 0) printk("bitmap_end_sync 0 (%d)\n", aborted);
1409 */      if (bitmap == NULL) {
1410                 *blocks = 1024;
1411                 return;
1412         }
1413         spin_lock_irqsave(&bitmap->lock, flags);
1414         bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
1415         if (bmc == NULL)
1416                 goto unlock;
1417         /* locked */
1418 /*
1419         if (offset == 0) printk("bitmap_end sync found 0x%x, blocks %d\n", *bmc, *blocks);
1420 */
1421         if (RESYNC(*bmc)) {
1422                 *bmc &= ~RESYNC_MASK;
1423
1424                 if (!NEEDED(*bmc) && aborted)
1425                         *bmc |= NEEDED_MASK;
1426                 else {
1427                         if (*bmc <= 2) {
1428                                 set_page_attr(bitmap,
1429                                               filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap)),
1430                                               BITMAP_PAGE_CLEAN);
1431                         }
1432                 }
1433         }
1434  unlock:
1435         spin_unlock_irqrestore(&bitmap->lock, flags);
1436         bitmap->allclean = 0;
1437 }
1438
1439 void bitmap_close_sync(struct bitmap *bitmap)
1440 {
1441         /* Sync has finished, and any bitmap chunks that weren't synced
1442          * properly have been aborted.  It remains to us to clear the
1443          * RESYNC bit wherever it is still on
1444          */
1445         sector_t sector = 0;
1446         int blocks;
1447         if (!bitmap)
1448                 return;
1449         while (sector < bitmap->mddev->resync_max_sectors) {
1450                 bitmap_end_sync(bitmap, sector, &blocks, 0);
1451                 sector += blocks;
1452         }
1453 }
1454
1455 void bitmap_cond_end_sync(struct bitmap *bitmap, sector_t sector)
1456 {
1457         sector_t s = 0;
1458         int blocks;
1459
1460         if (!bitmap)
1461                 return;
1462         if (sector == 0) {
1463                 bitmap->last_end_sync = jiffies;
1464                 return;
1465         }
1466         if (time_before(jiffies, (bitmap->last_end_sync
1467                                   + bitmap->daemon_sleep * HZ)))
1468                 return;
1469         wait_event(bitmap->mddev->recovery_wait,
1470                    atomic_read(&bitmap->mddev->recovery_active) == 0);
1471
1472         sector &= ~((1ULL << CHUNK_BLOCK_SHIFT(bitmap)) - 1);
1473         s = 0;
1474         while (s < sector && s < bitmap->mddev->resync_max_sectors) {
1475                 bitmap_end_sync(bitmap, s, &blocks, 0);
1476                 s += blocks;
1477         }
1478         bitmap->last_end_sync = jiffies;
1479 }
1480
1481 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
1482 {
1483         /* For each chunk covered by any of these sectors, set the
1484          * counter to 1 and set resync_needed.  They should all
1485          * be 0 at this point
1486          */
1487
1488         int secs;
1489         bitmap_counter_t *bmc;
1490         spin_lock_irq(&bitmap->lock);
1491         bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
1492         if (!bmc) {
1493                 spin_unlock_irq(&bitmap->lock);
1494                 return;
1495         }
1496         if (! *bmc) {
1497                 struct page *page;
1498                 *bmc = 1 | (needed?NEEDED_MASK:0);
1499                 bitmap_count_page(bitmap, offset, 1);
1500                 page = filemap_get_page(bitmap, offset >> CHUNK_BLOCK_SHIFT(bitmap));
1501                 set_page_attr(bitmap, page, BITMAP_PAGE_CLEAN);
1502         }
1503         spin_unlock_irq(&bitmap->lock);
1504         bitmap->allclean = 0;
1505 }
1506
1507 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
1508 void bitmap_dirty_bits(struct bitmap *bitmap, unsigned long s, unsigned long e)
1509 {
1510         unsigned long chunk;
1511
1512         for (chunk = s; chunk <= e; chunk++) {
1513                 sector_t sec = chunk << CHUNK_BLOCK_SHIFT(bitmap);
1514                 bitmap_set_memory_bits(bitmap, sec, 1);
1515                 bitmap_file_set_bit(bitmap, sec);
1516         }
1517 }
1518
1519 /*
1520  * flush out any pending updates
1521  */
1522 void bitmap_flush(mddev_t *mddev)
1523 {
1524         struct bitmap *bitmap = mddev->bitmap;
1525         int sleep;
1526
1527         if (!bitmap) /* there was no bitmap */
1528                 return;
1529
1530         /* run the daemon_work three time to ensure everything is flushed
1531          * that can be
1532          */
1533         sleep = bitmap->daemon_sleep;
1534         bitmap->daemon_sleep = 0;
1535         bitmap_daemon_work(bitmap);
1536         bitmap_daemon_work(bitmap);
1537         bitmap_daemon_work(bitmap);
1538         bitmap->daemon_sleep = sleep;
1539         bitmap_update_sb(bitmap);
1540 }
1541
1542 /*
1543  * free memory that was allocated
1544  */
1545 static void bitmap_free(struct bitmap *bitmap)
1546 {
1547         unsigned long k, pages;
1548         struct bitmap_page *bp;
1549
1550         if (!bitmap) /* there was no bitmap */
1551                 return;
1552
1553         /* release the bitmap file and kill the daemon */
1554         bitmap_file_put(bitmap);
1555
1556         bp = bitmap->bp;
1557         pages = bitmap->pages;
1558
1559         /* free all allocated memory */
1560
1561         if (bp) /* deallocate the page memory */
1562                 for (k = 0; k < pages; k++)
1563                         if (bp[k].map && !bp[k].hijacked)
1564                                 kfree(bp[k].map);
1565         kfree(bp);
1566         kfree(bitmap);
1567 }
1568 void bitmap_destroy(mddev_t *mddev)
1569 {
1570         struct bitmap *bitmap = mddev->bitmap;
1571
1572         if (!bitmap) /* there was no bitmap */
1573                 return;
1574
1575         mddev->bitmap = NULL; /* disconnect from the md device */
1576         if (mddev->thread)
1577                 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
1578
1579         bitmap_free(bitmap);
1580 }
1581
1582 /*
1583  * initialize the bitmap structure
1584  * if this returns an error, bitmap_destroy must be called to do clean up
1585  */
1586 int bitmap_create(mddev_t *mddev)
1587 {
1588         struct bitmap *bitmap;
1589         unsigned long blocks = mddev->resync_max_sectors;
1590         unsigned long chunks;
1591         unsigned long pages;
1592         struct file *file = mddev->bitmap_file;
1593         int err;
1594         sector_t start;
1595
1596         BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
1597
1598         if (!file && !mddev->bitmap_offset) /* bitmap disabled, nothing to do */
1599                 return 0;
1600
1601         BUG_ON(file && mddev->bitmap_offset);
1602
1603         bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
1604         if (!bitmap)
1605                 return -ENOMEM;
1606
1607         spin_lock_init(&bitmap->lock);
1608         atomic_set(&bitmap->pending_writes, 0);
1609         init_waitqueue_head(&bitmap->write_wait);
1610         init_waitqueue_head(&bitmap->overflow_wait);
1611
1612         bitmap->mddev = mddev;
1613
1614         bitmap->file = file;
1615         bitmap->offset = mddev->bitmap_offset;
1616         if (file) {
1617                 get_file(file);
1618                 do_sync_mapping_range(file->f_mapping, 0, LLONG_MAX,
1619                                       SYNC_FILE_RANGE_WAIT_BEFORE |
1620                                       SYNC_FILE_RANGE_WRITE |
1621                                       SYNC_FILE_RANGE_WAIT_AFTER);
1622         }
1623         /* read superblock from bitmap file (this sets bitmap->chunksize) */
1624         err = bitmap_read_sb(bitmap);
1625         if (err)
1626                 goto error;
1627
1628         bitmap->chunkshift = ffz(~bitmap->chunksize);
1629
1630         /* now that chunksize and chunkshift are set, we can use these macros */
1631         chunks = (blocks + CHUNK_BLOCK_RATIO(bitmap) - 1) /
1632                         CHUNK_BLOCK_RATIO(bitmap);
1633         pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
1634
1635         BUG_ON(!pages);
1636
1637         bitmap->chunks = chunks;
1638         bitmap->pages = pages;
1639         bitmap->missing_pages = pages;
1640         bitmap->counter_bits = COUNTER_BITS;
1641
1642         bitmap->syncchunk = ~0UL;
1643
1644 #ifdef INJECT_FATAL_FAULT_1
1645         bitmap->bp = NULL;
1646 #else
1647         bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
1648 #endif
1649         err = -ENOMEM;
1650         if (!bitmap->bp)
1651                 goto error;
1652
1653         /* now that we have some pages available, initialize the in-memory
1654          * bitmap from the on-disk bitmap */
1655         start = 0;
1656         if (mddev->degraded == 0
1657             || bitmap->events_cleared == mddev->events)
1658                 /* no need to keep dirty bits to optimise a re-add of a missing device */
1659                 start = mddev->recovery_cp;
1660         err = bitmap_init_from_disk(bitmap, start);
1661
1662         if (err)
1663                 goto error;
1664
1665         printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
1666                 pages, bmname(bitmap));
1667
1668         mddev->bitmap = bitmap;
1669
1670         mddev->thread->timeout = bitmap->daemon_sleep * HZ;
1671
1672         bitmap_update_sb(bitmap);
1673
1674         return (bitmap->flags & BITMAP_WRITE_ERROR) ? -EIO : 0;
1675
1676  error:
1677         bitmap_free(bitmap);
1678         return err;
1679 }
1680
1681 /* the bitmap API -- for raid personalities */
1682 EXPORT_SYMBOL(bitmap_startwrite);
1683 EXPORT_SYMBOL(bitmap_endwrite);
1684 EXPORT_SYMBOL(bitmap_start_sync);
1685 EXPORT_SYMBOL(bitmap_end_sync);
1686 EXPORT_SYMBOL(bitmap_unplug);
1687 EXPORT_SYMBOL(bitmap_close_sync);
1688 EXPORT_SYMBOL(bitmap_cond_end_sync);