Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes...
[linux-2.6] / fs / ext4 / balloc.c
1 /*
2  *  linux/fs/ext4/balloc.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  Enhanced block allocation by Stephen Tweedie (sct@redhat.com), 1993
10  *  Big-endian to little-endian byte-swapping/bitmaps by
11  *        David S. Miller (davem@caip.rutgers.edu), 1995
12  */
13
14 #include <linux/time.h>
15 #include <linux/capability.h>
16 #include <linux/fs.h>
17 #include <linux/jbd2.h>
18 #include <linux/quotaops.h>
19 #include <linux/buffer_head.h>
20 #include "ext4.h"
21 #include "ext4_jbd2.h"
22 #include "group.h"
23
24 /*
25  * balloc.c contains the blocks allocation and deallocation routines
26  */
27
28 /*
29  * Calculate the block group number and offset, given a block number
30  */
31 void ext4_get_group_no_and_offset(struct super_block *sb, ext4_fsblk_t blocknr,
32                 ext4_group_t *blockgrpp, ext4_grpblk_t *offsetp)
33 {
34         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
35         ext4_grpblk_t offset;
36
37         blocknr = blocknr - le32_to_cpu(es->s_first_data_block);
38         offset = do_div(blocknr, EXT4_BLOCKS_PER_GROUP(sb));
39         if (offsetp)
40                 *offsetp = offset;
41         if (blockgrpp)
42                 *blockgrpp = blocknr;
43
44 }
45
46 /* Initializes an uninitialized block bitmap if given, and returns the
47  * number of blocks free in the group. */
48 unsigned ext4_init_block_bitmap(struct super_block *sb, struct buffer_head *bh,
49                  ext4_group_t block_group, struct ext4_group_desc *gdp)
50 {
51         int bit, bit_max;
52         unsigned free_blocks, group_blocks;
53         struct ext4_sb_info *sbi = EXT4_SB(sb);
54
55         if (bh) {
56                 J_ASSERT_BH(bh, buffer_locked(bh));
57
58                 /* If checksum is bad mark all blocks used to prevent allocation
59                  * essentially implementing a per-group read-only flag. */
60                 if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) {
61                         ext4_error(sb, __func__,
62                                   "Checksum bad for group %lu\n", block_group);
63                         gdp->bg_free_blocks_count = 0;
64                         gdp->bg_free_inodes_count = 0;
65                         gdp->bg_itable_unused = 0;
66                         memset(bh->b_data, 0xff, sb->s_blocksize);
67                         return 0;
68                 }
69                 memset(bh->b_data, 0, sb->s_blocksize);
70         }
71
72         /* Check for superblock and gdt backups in this group */
73         bit_max = ext4_bg_has_super(sb, block_group);
74
75         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_META_BG) ||
76             block_group < le32_to_cpu(sbi->s_es->s_first_meta_bg) *
77                           sbi->s_desc_per_block) {
78                 if (bit_max) {
79                         bit_max += ext4_bg_num_gdb(sb, block_group);
80                         bit_max +=
81                                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks);
82                 }
83         } else { /* For META_BG_BLOCK_GROUPS */
84                 int group_rel = (block_group -
85                                  le32_to_cpu(sbi->s_es->s_first_meta_bg)) %
86                                 EXT4_DESC_PER_BLOCK(sb);
87                 if (group_rel == 0 || group_rel == 1 ||
88                     (group_rel == EXT4_DESC_PER_BLOCK(sb) - 1))
89                         bit_max += 1;
90         }
91
92         if (block_group == sbi->s_groups_count - 1) {
93                 /*
94                  * Even though mke2fs always initialize first and last group
95                  * if some other tool enabled the EXT4_BG_BLOCK_UNINIT we need
96                  * to make sure we calculate the right free blocks
97                  */
98                 group_blocks = ext4_blocks_count(sbi->s_es) -
99                         le32_to_cpu(sbi->s_es->s_first_data_block) -
100                         (EXT4_BLOCKS_PER_GROUP(sb) * (sbi->s_groups_count -1));
101         } else {
102                 group_blocks = EXT4_BLOCKS_PER_GROUP(sb);
103         }
104
105         free_blocks = group_blocks - bit_max;
106
107         if (bh) {
108                 ext4_fsblk_t start;
109
110                 for (bit = 0; bit < bit_max; bit++)
111                         ext4_set_bit(bit, bh->b_data);
112
113                 start = ext4_group_first_block_no(sb, block_group);
114
115                 /* Set bits for block and inode bitmaps, and inode table */
116                 ext4_set_bit(ext4_block_bitmap(sb, gdp) - start, bh->b_data);
117                 ext4_set_bit(ext4_inode_bitmap(sb, gdp) - start, bh->b_data);
118                 for (bit = (ext4_inode_table(sb, gdp) - start),
119                      bit_max = bit + sbi->s_itb_per_group; bit < bit_max; bit++)
120                         ext4_set_bit(bit, bh->b_data);
121
122                 /*
123                  * Also if the number of blocks within the group is
124                  * less than the blocksize * 8 ( which is the size
125                  * of bitmap ), set rest of the block bitmap to 1
126                  */
127                 mark_bitmap_end(group_blocks, sb->s_blocksize * 8, bh->b_data);
128         }
129
130         return free_blocks - sbi->s_itb_per_group - 2;
131 }
132
133
134 /*
135  * The free blocks are managed by bitmaps.  A file system contains several
136  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
137  * block for inodes, N blocks for the inode table and data blocks.
138  *
139  * The file system contains group descriptors which are located after the
140  * super block.  Each descriptor contains the number of the bitmap block and
141  * the free blocks count in the block.  The descriptors are loaded in memory
142  * when a file system is mounted (see ext4_fill_super).
143  */
144
145
146 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
147
148 /**
149  * ext4_get_group_desc() -- load group descriptor from disk
150  * @sb:                 super block
151  * @block_group:        given block group
152  * @bh:                 pointer to the buffer head to store the block
153  *                      group descriptor
154  */
155 struct ext4_group_desc * ext4_get_group_desc(struct super_block * sb,
156                                              ext4_group_t block_group,
157                                              struct buffer_head ** bh)
158 {
159         unsigned long group_desc;
160         unsigned long offset;
161         struct ext4_group_desc * desc;
162         struct ext4_sb_info *sbi = EXT4_SB(sb);
163
164         if (block_group >= sbi->s_groups_count) {
165                 ext4_error (sb, "ext4_get_group_desc",
166                             "block_group >= groups_count - "
167                             "block_group = %lu, groups_count = %lu",
168                             block_group, sbi->s_groups_count);
169
170                 return NULL;
171         }
172         smp_rmb();
173
174         group_desc = block_group >> EXT4_DESC_PER_BLOCK_BITS(sb);
175         offset = block_group & (EXT4_DESC_PER_BLOCK(sb) - 1);
176         if (!sbi->s_group_desc[group_desc]) {
177                 ext4_error (sb, "ext4_get_group_desc",
178                             "Group descriptor not loaded - "
179                             "block_group = %lu, group_desc = %lu, desc = %lu",
180                              block_group, group_desc, offset);
181                 return NULL;
182         }
183
184         desc = (struct ext4_group_desc *)(
185                 (__u8 *)sbi->s_group_desc[group_desc]->b_data +
186                 offset * EXT4_DESC_SIZE(sb));
187         if (bh)
188                 *bh = sbi->s_group_desc[group_desc];
189         return desc;
190 }
191
192 static int ext4_valid_block_bitmap(struct super_block *sb,
193                                         struct ext4_group_desc *desc,
194                                         unsigned int block_group,
195                                         struct buffer_head *bh)
196 {
197         ext4_grpblk_t offset;
198         ext4_grpblk_t next_zero_bit;
199         ext4_fsblk_t bitmap_blk;
200         ext4_fsblk_t group_first_block;
201
202         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FLEX_BG)) {
203                 /* with FLEX_BG, the inode/block bitmaps and itable
204                  * blocks may not be in the group at all
205                  * so the bitmap validation will be skipped for those groups
206                  * or it has to also read the block group where the bitmaps
207                  * are located to verify they are set.
208                  */
209                 return 1;
210         }
211         group_first_block = ext4_group_first_block_no(sb, block_group);
212
213         /* check whether block bitmap block number is set */
214         bitmap_blk = ext4_block_bitmap(sb, desc);
215         offset = bitmap_blk - group_first_block;
216         if (!ext4_test_bit(offset, bh->b_data))
217                 /* bad block bitmap */
218                 goto err_out;
219
220         /* check whether the inode bitmap block number is set */
221         bitmap_blk = ext4_inode_bitmap(sb, desc);
222         offset = bitmap_blk - group_first_block;
223         if (!ext4_test_bit(offset, bh->b_data))
224                 /* bad block bitmap */
225                 goto err_out;
226
227         /* check whether the inode table block number is set */
228         bitmap_blk = ext4_inode_table(sb, desc);
229         offset = bitmap_blk - group_first_block;
230         next_zero_bit = ext4_find_next_zero_bit(bh->b_data,
231                                 offset + EXT4_SB(sb)->s_itb_per_group,
232                                 offset);
233         if (next_zero_bit >= offset + EXT4_SB(sb)->s_itb_per_group)
234                 /* good bitmap for inode tables */
235                 return 1;
236
237 err_out:
238         ext4_error(sb, __func__,
239                         "Invalid block bitmap - "
240                         "block_group = %d, block = %llu",
241                         block_group, bitmap_blk);
242         return 0;
243 }
244 /**
245  * read_block_bitmap()
246  * @sb:                 super block
247  * @block_group:        given block group
248  *
249  * Read the bitmap for a given block_group,and validate the
250  * bits for block/inode/inode tables are set in the bitmaps
251  *
252  * Return buffer_head on success or NULL in case of failure.
253  */
254 struct buffer_head *
255 read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
256 {
257         struct ext4_group_desc * desc;
258         struct buffer_head * bh = NULL;
259         ext4_fsblk_t bitmap_blk;
260
261         desc = ext4_get_group_desc(sb, block_group, NULL);
262         if (!desc)
263                 return NULL;
264         bitmap_blk = ext4_block_bitmap(sb, desc);
265         bh = sb_getblk(sb, bitmap_blk);
266         if (unlikely(!bh)) {
267                 ext4_error(sb, __func__,
268                             "Cannot read block bitmap - "
269                             "block_group = %d, block_bitmap = %llu",
270                             (int)block_group, (unsigned long long)bitmap_blk);
271                 return NULL;
272         }
273         if (bh_uptodate_or_lock(bh))
274                 return bh;
275
276         if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
277                 ext4_init_block_bitmap(sb, bh, block_group, desc);
278                 set_buffer_uptodate(bh);
279                 unlock_buffer(bh);
280                 return bh;
281         }
282         if (bh_submit_read(bh) < 0) {
283                 put_bh(bh);
284                 ext4_error(sb, __func__,
285                             "Cannot read block bitmap - "
286                             "block_group = %d, block_bitmap = %llu",
287                             (int)block_group, (unsigned long long)bitmap_blk);
288                 return NULL;
289         }
290         if (!ext4_valid_block_bitmap(sb, desc, block_group, bh)) {
291                 put_bh(bh);
292                 return NULL;
293         }
294
295         return bh;
296 }
297 /*
298  * The reservation window structure operations
299  * --------------------------------------------
300  * Operations include:
301  * dump, find, add, remove, is_empty, find_next_reservable_window, etc.
302  *
303  * We use a red-black tree to represent per-filesystem reservation
304  * windows.
305  *
306  */
307
308 /**
309  * __rsv_window_dump() -- Dump the filesystem block allocation reservation map
310  * @rb_root:            root of per-filesystem reservation rb tree
311  * @verbose:            verbose mode
312  * @fn:                 function which wishes to dump the reservation map
313  *
314  * If verbose is turned on, it will print the whole block reservation
315  * windows(start, end). Otherwise, it will only print out the "bad" windows,
316  * those windows that overlap with their immediate neighbors.
317  */
318 #if 1
319 static void __rsv_window_dump(struct rb_root *root, int verbose,
320                               const char *fn)
321 {
322         struct rb_node *n;
323         struct ext4_reserve_window_node *rsv, *prev;
324         int bad;
325
326 restart:
327         n = rb_first(root);
328         bad = 0;
329         prev = NULL;
330
331         printk("Block Allocation Reservation Windows Map (%s):\n", fn);
332         while (n) {
333                 rsv = rb_entry(n, struct ext4_reserve_window_node, rsv_node);
334                 if (verbose)
335                         printk("reservation window 0x%p "
336                                "start:  %llu, end:  %llu\n",
337                                rsv, rsv->rsv_start, rsv->rsv_end);
338                 if (rsv->rsv_start && rsv->rsv_start >= rsv->rsv_end) {
339                         printk("Bad reservation %p (start >= end)\n",
340                                rsv);
341                         bad = 1;
342                 }
343                 if (prev && prev->rsv_end >= rsv->rsv_start) {
344                         printk("Bad reservation %p (prev->end >= start)\n",
345                                rsv);
346                         bad = 1;
347                 }
348                 if (bad) {
349                         if (!verbose) {
350                                 printk("Restarting reservation walk in verbose mode\n");
351                                 verbose = 1;
352                                 goto restart;
353                         }
354                 }
355                 n = rb_next(n);
356                 prev = rsv;
357         }
358         printk("Window map complete.\n");
359         if (bad)
360                 BUG();
361 }
362 #define rsv_window_dump(root, verbose) \
363         __rsv_window_dump((root), (verbose), __func__)
364 #else
365 #define rsv_window_dump(root, verbose) do {} while (0)
366 #endif
367
368 /**
369  * goal_in_my_reservation()
370  * @rsv:                inode's reservation window
371  * @grp_goal:           given goal block relative to the allocation block group
372  * @group:              the current allocation block group
373  * @sb:                 filesystem super block
374  *
375  * Test if the given goal block (group relative) is within the file's
376  * own block reservation window range.
377  *
378  * If the reservation window is outside the goal allocation group, return 0;
379  * grp_goal (given goal block) could be -1, which means no specific
380  * goal block. In this case, always return 1.
381  * If the goal block is within the reservation window, return 1;
382  * otherwise, return 0;
383  */
384 static int
385 goal_in_my_reservation(struct ext4_reserve_window *rsv, ext4_grpblk_t grp_goal,
386                         ext4_group_t group, struct super_block *sb)
387 {
388         ext4_fsblk_t group_first_block, group_last_block;
389
390         group_first_block = ext4_group_first_block_no(sb, group);
391         group_last_block = group_first_block + (EXT4_BLOCKS_PER_GROUP(sb) - 1);
392
393         if ((rsv->_rsv_start > group_last_block) ||
394             (rsv->_rsv_end < group_first_block))
395                 return 0;
396         if ((grp_goal >= 0) && ((grp_goal + group_first_block < rsv->_rsv_start)
397                 || (grp_goal + group_first_block > rsv->_rsv_end)))
398                 return 0;
399         return 1;
400 }
401
402 /**
403  * search_reserve_window()
404  * @rb_root:            root of reservation tree
405  * @goal:               target allocation block
406  *
407  * Find the reserved window which includes the goal, or the previous one
408  * if the goal is not in any window.
409  * Returns NULL if there are no windows or if all windows start after the goal.
410  */
411 static struct ext4_reserve_window_node *
412 search_reserve_window(struct rb_root *root, ext4_fsblk_t goal)
413 {
414         struct rb_node *n = root->rb_node;
415         struct ext4_reserve_window_node *rsv;
416
417         if (!n)
418                 return NULL;
419
420         do {
421                 rsv = rb_entry(n, struct ext4_reserve_window_node, rsv_node);
422
423                 if (goal < rsv->rsv_start)
424                         n = n->rb_left;
425                 else if (goal > rsv->rsv_end)
426                         n = n->rb_right;
427                 else
428                         return rsv;
429         } while (n);
430         /*
431          * We've fallen off the end of the tree: the goal wasn't inside
432          * any particular node.  OK, the previous node must be to one
433          * side of the interval containing the goal.  If it's the RHS,
434          * we need to back up one.
435          */
436         if (rsv->rsv_start > goal) {
437                 n = rb_prev(&rsv->rsv_node);
438                 rsv = rb_entry(n, struct ext4_reserve_window_node, rsv_node);
439         }
440         return rsv;
441 }
442
443 /**
444  * ext4_rsv_window_add() -- Insert a window to the block reservation rb tree.
445  * @sb:                 super block
446  * @rsv:                reservation window to add
447  *
448  * Must be called with rsv_lock hold.
449  */
450 void ext4_rsv_window_add(struct super_block *sb,
451                     struct ext4_reserve_window_node *rsv)
452 {
453         struct rb_root *root = &EXT4_SB(sb)->s_rsv_window_root;
454         struct rb_node *node = &rsv->rsv_node;
455         ext4_fsblk_t start = rsv->rsv_start;
456
457         struct rb_node ** p = &root->rb_node;
458         struct rb_node * parent = NULL;
459         struct ext4_reserve_window_node *this;
460
461         while (*p)
462         {
463                 parent = *p;
464                 this = rb_entry(parent, struct ext4_reserve_window_node, rsv_node);
465
466                 if (start < this->rsv_start)
467                         p = &(*p)->rb_left;
468                 else if (start > this->rsv_end)
469                         p = &(*p)->rb_right;
470                 else {
471                         rsv_window_dump(root, 1);
472                         BUG();
473                 }
474         }
475
476         rb_link_node(node, parent, p);
477         rb_insert_color(node, root);
478 }
479
480 /**
481  * ext4_rsv_window_remove() -- unlink a window from the reservation rb tree
482  * @sb:                 super block
483  * @rsv:                reservation window to remove
484  *
485  * Mark the block reservation window as not allocated, and unlink it
486  * from the filesystem reservation window rb tree. Must be called with
487  * rsv_lock hold.
488  */
489 static void rsv_window_remove(struct super_block *sb,
490                               struct ext4_reserve_window_node *rsv)
491 {
492         rsv->rsv_start = EXT4_RESERVE_WINDOW_NOT_ALLOCATED;
493         rsv->rsv_end = EXT4_RESERVE_WINDOW_NOT_ALLOCATED;
494         rsv->rsv_alloc_hit = 0;
495         rb_erase(&rsv->rsv_node, &EXT4_SB(sb)->s_rsv_window_root);
496 }
497
498 /*
499  * rsv_is_empty() -- Check if the reservation window is allocated.
500  * @rsv:                given reservation window to check
501  *
502  * returns 1 if the end block is EXT4_RESERVE_WINDOW_NOT_ALLOCATED.
503  */
504 static inline int rsv_is_empty(struct ext4_reserve_window *rsv)
505 {
506         /* a valid reservation end block could not be 0 */
507         return rsv->_rsv_end == EXT4_RESERVE_WINDOW_NOT_ALLOCATED;
508 }
509
510 /**
511  * ext4_init_block_alloc_info()
512  * @inode:              file inode structure
513  *
514  * Allocate and initialize the  reservation window structure, and
515  * link the window to the ext4 inode structure at last
516  *
517  * The reservation window structure is only dynamically allocated
518  * and linked to ext4 inode the first time the open file
519  * needs a new block. So, before every ext4_new_block(s) call, for
520  * regular files, we should check whether the reservation window
521  * structure exists or not. In the latter case, this function is called.
522  * Fail to do so will result in block reservation being turned off for that
523  * open file.
524  *
525  * This function is called from ext4_get_blocks_handle(), also called
526  * when setting the reservation window size through ioctl before the file
527  * is open for write (needs block allocation).
528  *
529  * Needs down_write(i_data_sem) protection prior to call this function.
530  */
531 void ext4_init_block_alloc_info(struct inode *inode)
532 {
533         struct ext4_inode_info *ei = EXT4_I(inode);
534         struct ext4_block_alloc_info *block_i = ei->i_block_alloc_info;
535         struct super_block *sb = inode->i_sb;
536
537         block_i = kmalloc(sizeof(*block_i), GFP_NOFS);
538         if (block_i) {
539                 struct ext4_reserve_window_node *rsv = &block_i->rsv_window_node;
540
541                 rsv->rsv_start = EXT4_RESERVE_WINDOW_NOT_ALLOCATED;
542                 rsv->rsv_end = EXT4_RESERVE_WINDOW_NOT_ALLOCATED;
543
544                 /*
545                  * if filesystem is mounted with NORESERVATION, the goal
546                  * reservation window size is set to zero to indicate
547                  * block reservation is off
548                  */
549                 if (!test_opt(sb, RESERVATION))
550                         rsv->rsv_goal_size = 0;
551                 else
552                         rsv->rsv_goal_size = EXT4_DEFAULT_RESERVE_BLOCKS;
553                 rsv->rsv_alloc_hit = 0;
554                 block_i->last_alloc_logical_block = 0;
555                 block_i->last_alloc_physical_block = 0;
556         }
557         ei->i_block_alloc_info = block_i;
558 }
559
560 /**
561  * ext4_discard_reservation()
562  * @inode:              inode
563  *
564  * Discard(free) block reservation window on last file close, or truncate
565  * or at last iput().
566  *
567  * It is being called in three cases:
568  *      ext4_release_file(): last writer close the file
569  *      ext4_clear_inode(): last iput(), when nobody link to this file.
570  *      ext4_truncate(): when the block indirect map is about to change.
571  *
572  */
573 void ext4_discard_reservation(struct inode *inode)
574 {
575         struct ext4_inode_info *ei = EXT4_I(inode);
576         struct ext4_block_alloc_info *block_i = ei->i_block_alloc_info;
577         struct ext4_reserve_window_node *rsv;
578         spinlock_t *rsv_lock = &EXT4_SB(inode->i_sb)->s_rsv_window_lock;
579
580         ext4_mb_discard_inode_preallocations(inode);
581
582         if (!block_i)
583                 return;
584
585         rsv = &block_i->rsv_window_node;
586         if (!rsv_is_empty(&rsv->rsv_window)) {
587                 spin_lock(rsv_lock);
588                 if (!rsv_is_empty(&rsv->rsv_window))
589                         rsv_window_remove(inode->i_sb, rsv);
590                 spin_unlock(rsv_lock);
591         }
592 }
593
594 /**
595  * ext4_free_blocks_sb() -- Free given blocks and update quota
596  * @handle:                     handle to this transaction
597  * @sb:                         super block
598  * @block:                      start physcial block to free
599  * @count:                      number of blocks to free
600  * @pdquot_freed_blocks:        pointer to quota
601  */
602 void ext4_free_blocks_sb(handle_t *handle, struct super_block *sb,
603                          ext4_fsblk_t block, unsigned long count,
604                          unsigned long *pdquot_freed_blocks)
605 {
606         struct buffer_head *bitmap_bh = NULL;
607         struct buffer_head *gd_bh;
608         ext4_group_t block_group;
609         ext4_grpblk_t bit;
610         unsigned long i;
611         unsigned long overflow;
612         struct ext4_group_desc * desc;
613         struct ext4_super_block * es;
614         struct ext4_sb_info *sbi;
615         int err = 0, ret;
616         ext4_grpblk_t group_freed;
617
618         *pdquot_freed_blocks = 0;
619         sbi = EXT4_SB(sb);
620         es = sbi->s_es;
621         if (block < le32_to_cpu(es->s_first_data_block) ||
622             block + count < block ||
623             block + count > ext4_blocks_count(es)) {
624                 ext4_error (sb, "ext4_free_blocks",
625                             "Freeing blocks not in datazone - "
626                             "block = %llu, count = %lu", block, count);
627                 goto error_return;
628         }
629
630         ext4_debug ("freeing block(s) %llu-%llu\n", block, block + count - 1);
631
632 do_more:
633         overflow = 0;
634         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
635         /*
636          * Check to see if we are freeing blocks across a group
637          * boundary.
638          */
639         if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
640                 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
641                 count -= overflow;
642         }
643         brelse(bitmap_bh);
644         bitmap_bh = read_block_bitmap(sb, block_group);
645         if (!bitmap_bh)
646                 goto error_return;
647         desc = ext4_get_group_desc (sb, block_group, &gd_bh);
648         if (!desc)
649                 goto error_return;
650
651         if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
652             in_range(ext4_inode_bitmap(sb, desc), block, count) ||
653             in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
654             in_range(block + count - 1, ext4_inode_table(sb, desc),
655                      sbi->s_itb_per_group)) {
656                 ext4_error (sb, "ext4_free_blocks",
657                             "Freeing blocks in system zones - "
658                             "Block = %llu, count = %lu",
659                             block, count);
660                 goto error_return;
661         }
662
663         /*
664          * We are about to start releasing blocks in the bitmap,
665          * so we need undo access.
666          */
667         /* @@@ check errors */
668         BUFFER_TRACE(bitmap_bh, "getting undo access");
669         err = ext4_journal_get_undo_access(handle, bitmap_bh);
670         if (err)
671                 goto error_return;
672
673         /*
674          * We are about to modify some metadata.  Call the journal APIs
675          * to unshare ->b_data if a currently-committing transaction is
676          * using it
677          */
678         BUFFER_TRACE(gd_bh, "get_write_access");
679         err = ext4_journal_get_write_access(handle, gd_bh);
680         if (err)
681                 goto error_return;
682
683         jbd_lock_bh_state(bitmap_bh);
684
685         for (i = 0, group_freed = 0; i < count; i++) {
686                 /*
687                  * An HJ special.  This is expensive...
688                  */
689 #ifdef CONFIG_JBD2_DEBUG
690                 jbd_unlock_bh_state(bitmap_bh);
691                 {
692                         struct buffer_head *debug_bh;
693                         debug_bh = sb_find_get_block(sb, block + i);
694                         if (debug_bh) {
695                                 BUFFER_TRACE(debug_bh, "Deleted!");
696                                 if (!bh2jh(bitmap_bh)->b_committed_data)
697                                         BUFFER_TRACE(debug_bh,
698                                                 "No commited data in bitmap");
699                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
700                                 __brelse(debug_bh);
701                         }
702                 }
703                 jbd_lock_bh_state(bitmap_bh);
704 #endif
705                 if (need_resched()) {
706                         jbd_unlock_bh_state(bitmap_bh);
707                         cond_resched();
708                         jbd_lock_bh_state(bitmap_bh);
709                 }
710                 /* @@@ This prevents newly-allocated data from being
711                  * freed and then reallocated within the same
712                  * transaction.
713                  *
714                  * Ideally we would want to allow that to happen, but to
715                  * do so requires making jbd2_journal_forget() capable of
716                  * revoking the queued write of a data block, which
717                  * implies blocking on the journal lock.  *forget()
718                  * cannot block due to truncate races.
719                  *
720                  * Eventually we can fix this by making jbd2_journal_forget()
721                  * return a status indicating whether or not it was able
722                  * to revoke the buffer.  On successful revoke, it is
723                  * safe not to set the allocation bit in the committed
724                  * bitmap, because we know that there is no outstanding
725                  * activity on the buffer any more and so it is safe to
726                  * reallocate it.
727                  */
728                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
729                 J_ASSERT_BH(bitmap_bh,
730                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
731                 ext4_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
732                                 bh2jh(bitmap_bh)->b_committed_data);
733
734                 /*
735                  * We clear the bit in the bitmap after setting the committed
736                  * data bit, because this is the reverse order to that which
737                  * the allocator uses.
738                  */
739                 BUFFER_TRACE(bitmap_bh, "clear bit");
740                 if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
741                                                 bit + i, bitmap_bh->b_data)) {
742                         jbd_unlock_bh_state(bitmap_bh);
743                         ext4_error(sb, __func__,
744                                    "bit already cleared for block %llu",
745                                    (ext4_fsblk_t)(block + i));
746                         jbd_lock_bh_state(bitmap_bh);
747                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
748                 } else {
749                         group_freed++;
750                 }
751         }
752         jbd_unlock_bh_state(bitmap_bh);
753
754         spin_lock(sb_bgl_lock(sbi, block_group));
755         le16_add_cpu(&desc->bg_free_blocks_count, group_freed);
756         desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
757         spin_unlock(sb_bgl_lock(sbi, block_group));
758         percpu_counter_add(&sbi->s_freeblocks_counter, count);
759
760         /* We dirtied the bitmap block */
761         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
762         err = ext4_journal_dirty_metadata(handle, bitmap_bh);
763
764         /* And the group descriptor block */
765         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
766         ret = ext4_journal_dirty_metadata(handle, gd_bh);
767         if (!err) err = ret;
768         *pdquot_freed_blocks += group_freed;
769
770         if (overflow && !err) {
771                 block += count;
772                 count = overflow;
773                 goto do_more;
774         }
775         sb->s_dirt = 1;
776 error_return:
777         brelse(bitmap_bh);
778         ext4_std_error(sb, err);
779         return;
780 }
781
782 /**
783  * ext4_free_blocks() -- Free given blocks and update quota
784  * @handle:             handle for this transaction
785  * @inode:              inode
786  * @block:              start physical block to free
787  * @count:              number of blocks to count
788  * @metadata:           Are these metadata blocks
789  */
790 void ext4_free_blocks(handle_t *handle, struct inode *inode,
791                         ext4_fsblk_t block, unsigned long count,
792                         int metadata)
793 {
794         struct super_block * sb;
795         unsigned long dquot_freed_blocks;
796
797         /* this isn't the right place to decide whether block is metadata
798          * inode.c/extents.c knows better, but for safety ... */
799         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
800                         ext4_should_journal_data(inode))
801                 metadata = 1;
802
803         sb = inode->i_sb;
804
805         if (!test_opt(sb, MBALLOC) || !EXT4_SB(sb)->s_group_info)
806                 ext4_free_blocks_sb(handle, sb, block, count,
807                                                 &dquot_freed_blocks);
808         else
809                 ext4_mb_free_blocks(handle, inode, block, count,
810                                                 metadata, &dquot_freed_blocks);
811         if (dquot_freed_blocks)
812                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
813         return;
814 }
815
816 /**
817  * ext4_test_allocatable()
818  * @nr:                 given allocation block group
819  * @bh:                 bufferhead contains the bitmap of the given block group
820  *
821  * For ext4 allocations, we must not reuse any blocks which are
822  * allocated in the bitmap buffer's "last committed data" copy.  This
823  * prevents deletes from freeing up the page for reuse until we have
824  * committed the delete transaction.
825  *
826  * If we didn't do this, then deleting something and reallocating it as
827  * data would allow the old block to be overwritten before the
828  * transaction committed (because we force data to disk before commit).
829  * This would lead to corruption if we crashed between overwriting the
830  * data and committing the delete.
831  *
832  * @@@ We may want to make this allocation behaviour conditional on
833  * data-writes at some point, and disable it for metadata allocations or
834  * sync-data inodes.
835  */
836 static int ext4_test_allocatable(ext4_grpblk_t nr, struct buffer_head *bh)
837 {
838         int ret;
839         struct journal_head *jh = bh2jh(bh);
840
841         if (ext4_test_bit(nr, bh->b_data))
842                 return 0;
843
844         jbd_lock_bh_state(bh);
845         if (!jh->b_committed_data)
846                 ret = 1;
847         else
848                 ret = !ext4_test_bit(nr, jh->b_committed_data);
849         jbd_unlock_bh_state(bh);
850         return ret;
851 }
852
853 /**
854  * bitmap_search_next_usable_block()
855  * @start:              the starting block (group relative) of the search
856  * @bh:                 bufferhead contains the block group bitmap
857  * @maxblocks:          the ending block (group relative) of the reservation
858  *
859  * The bitmap search --- search forward alternately through the actual
860  * bitmap on disk and the last-committed copy in journal, until we find a
861  * bit free in both bitmaps.
862  */
863 static ext4_grpblk_t
864 bitmap_search_next_usable_block(ext4_grpblk_t start, struct buffer_head *bh,
865                                         ext4_grpblk_t maxblocks)
866 {
867         ext4_grpblk_t next;
868         struct journal_head *jh = bh2jh(bh);
869
870         while (start < maxblocks) {
871                 next = ext4_find_next_zero_bit(bh->b_data, maxblocks, start);
872                 if (next >= maxblocks)
873                         return -1;
874                 if (ext4_test_allocatable(next, bh))
875                         return next;
876                 jbd_lock_bh_state(bh);
877                 if (jh->b_committed_data)
878                         start = ext4_find_next_zero_bit(jh->b_committed_data,
879                                                         maxblocks, next);
880                 jbd_unlock_bh_state(bh);
881         }
882         return -1;
883 }
884
885 /**
886  * find_next_usable_block()
887  * @start:              the starting block (group relative) to find next
888  *                      allocatable block in bitmap.
889  * @bh:                 bufferhead contains the block group bitmap
890  * @maxblocks:          the ending block (group relative) for the search
891  *
892  * Find an allocatable block in a bitmap.  We honor both the bitmap and
893  * its last-committed copy (if that exists), and perform the "most
894  * appropriate allocation" algorithm of looking for a free block near
895  * the initial goal; then for a free byte somewhere in the bitmap; then
896  * for any free bit in the bitmap.
897  */
898 static ext4_grpblk_t
899 find_next_usable_block(ext4_grpblk_t start, struct buffer_head *bh,
900                         ext4_grpblk_t maxblocks)
901 {
902         ext4_grpblk_t here, next;
903         char *p, *r;
904
905         if (start > 0) {
906                 /*
907                  * The goal was occupied; search forward for a free
908                  * block within the next XX blocks.
909                  *
910                  * end_goal is more or less random, but it has to be
911                  * less than EXT4_BLOCKS_PER_GROUP. Aligning up to the
912                  * next 64-bit boundary is simple..
913                  */
914                 ext4_grpblk_t end_goal = (start + 63) & ~63;
915                 if (end_goal > maxblocks)
916                         end_goal = maxblocks;
917                 here = ext4_find_next_zero_bit(bh->b_data, end_goal, start);
918                 if (here < end_goal && ext4_test_allocatable(here, bh))
919                         return here;
920                 ext4_debug("Bit not found near goal\n");
921         }
922
923         here = start;
924         if (here < 0)
925                 here = 0;
926
927         p = ((char *)bh->b_data) + (here >> 3);
928         r = memscan(p, 0, ((maxblocks + 7) >> 3) - (here >> 3));
929         next = (r - ((char *)bh->b_data)) << 3;
930
931         if (next < maxblocks && next >= start && ext4_test_allocatable(next, bh))
932                 return next;
933
934         /*
935          * The bitmap search --- search forward alternately through the actual
936          * bitmap and the last-committed copy until we find a bit free in
937          * both
938          */
939         here = bitmap_search_next_usable_block(here, bh, maxblocks);
940         return here;
941 }
942
943 /**
944  * claim_block()
945  * @block:              the free block (group relative) to allocate
946  * @bh:                 the bufferhead containts the block group bitmap
947  *
948  * We think we can allocate this block in this bitmap.  Try to set the bit.
949  * If that succeeds then check that nobody has allocated and then freed the
950  * block since we saw that is was not marked in b_committed_data.  If it _was_
951  * allocated and freed then clear the bit in the bitmap again and return
952  * zero (failure).
953  */
954 static inline int
955 claim_block(spinlock_t *lock, ext4_grpblk_t block, struct buffer_head *bh)
956 {
957         struct journal_head *jh = bh2jh(bh);
958         int ret;
959
960         if (ext4_set_bit_atomic(lock, block, bh->b_data))
961                 return 0;
962         jbd_lock_bh_state(bh);
963         if (jh->b_committed_data && ext4_test_bit(block,jh->b_committed_data)) {
964                 ext4_clear_bit_atomic(lock, block, bh->b_data);
965                 ret = 0;
966         } else {
967                 ret = 1;
968         }
969         jbd_unlock_bh_state(bh);
970         return ret;
971 }
972
973 /**
974  * ext4_try_to_allocate()
975  * @sb:                 superblock
976  * @handle:             handle to this transaction
977  * @group:              given allocation block group
978  * @bitmap_bh:          bufferhead holds the block bitmap
979  * @grp_goal:           given target block within the group
980  * @count:              target number of blocks to allocate
981  * @my_rsv:             reservation window
982  *
983  * Attempt to allocate blocks within a give range. Set the range of allocation
984  * first, then find the first free bit(s) from the bitmap (within the range),
985  * and at last, allocate the blocks by claiming the found free bit as allocated.
986  *
987  * To set the range of this allocation:
988  *      if there is a reservation window, only try to allocate block(s) from the
989  *      file's own reservation window;
990  *      Otherwise, the allocation range starts from the give goal block, ends at
991  *      the block group's last block.
992  *
993  * If we failed to allocate the desired block then we may end up crossing to a
994  * new bitmap.  In that case we must release write access to the old one via
995  * ext4_journal_release_buffer(), else we'll run out of credits.
996  */
997 static ext4_grpblk_t
998 ext4_try_to_allocate(struct super_block *sb, handle_t *handle,
999                         ext4_group_t group, struct buffer_head *bitmap_bh,
1000                         ext4_grpblk_t grp_goal, unsigned long *count,
1001                         struct ext4_reserve_window *my_rsv)
1002 {
1003         ext4_fsblk_t group_first_block;
1004         ext4_grpblk_t start, end;
1005         unsigned long num = 0;
1006
1007         /* we do allocation within the reservation window if we have a window */
1008         if (my_rsv) {
1009                 group_first_block = ext4_group_first_block_no(sb, group);
1010                 if (my_rsv->_rsv_start >= group_first_block)
1011                         start = my_rsv->_rsv_start - group_first_block;
1012                 else
1013                         /* reservation window cross group boundary */
1014                         start = 0;
1015                 end = my_rsv->_rsv_end - group_first_block + 1;
1016                 if (end > EXT4_BLOCKS_PER_GROUP(sb))
1017                         /* reservation window crosses group boundary */
1018                         end = EXT4_BLOCKS_PER_GROUP(sb);
1019                 if ((start <= grp_goal) && (grp_goal < end))
1020                         start = grp_goal;
1021                 else
1022                         grp_goal = -1;
1023         } else {
1024                 if (grp_goal > 0)
1025                         start = grp_goal;
1026                 else
1027                         start = 0;
1028                 end = EXT4_BLOCKS_PER_GROUP(sb);
1029         }
1030
1031         BUG_ON(start > EXT4_BLOCKS_PER_GROUP(sb));
1032
1033 repeat:
1034         if (grp_goal < 0 || !ext4_test_allocatable(grp_goal, bitmap_bh)) {
1035                 grp_goal = find_next_usable_block(start, bitmap_bh, end);
1036                 if (grp_goal < 0)
1037                         goto fail_access;
1038                 if (!my_rsv) {
1039                         int i;
1040
1041                         for (i = 0; i < 7 && grp_goal > start &&
1042                                         ext4_test_allocatable(grp_goal - 1,
1043                                                                 bitmap_bh);
1044                                         i++, grp_goal--)
1045                                 ;
1046                 }
1047         }
1048         start = grp_goal;
1049
1050         if (!claim_block(sb_bgl_lock(EXT4_SB(sb), group),
1051                 grp_goal, bitmap_bh)) {
1052                 /*
1053                  * The block was allocated by another thread, or it was
1054                  * allocated and then freed by another thread
1055                  */
1056                 start++;
1057                 grp_goal++;
1058                 if (start >= end)
1059                         goto fail_access;
1060                 goto repeat;
1061         }
1062         num++;
1063         grp_goal++;
1064         while (num < *count && grp_goal < end
1065                 && ext4_test_allocatable(grp_goal, bitmap_bh)
1066                 && claim_block(sb_bgl_lock(EXT4_SB(sb), group),
1067                                 grp_goal, bitmap_bh)) {
1068                 num++;
1069                 grp_goal++;
1070         }
1071         *count = num;
1072         return grp_goal - num;
1073 fail_access:
1074         *count = num;
1075         return -1;
1076 }
1077
1078 /**
1079  *      find_next_reservable_window():
1080  *              find a reservable space within the given range.
1081  *              It does not allocate the reservation window for now:
1082  *              alloc_new_reservation() will do the work later.
1083  *
1084  *      @search_head: the head of the searching list;
1085  *              This is not necessarily the list head of the whole filesystem
1086  *
1087  *              We have both head and start_block to assist the search
1088  *              for the reservable space. The list starts from head,
1089  *              but we will shift to the place where start_block is,
1090  *              then start from there, when looking for a reservable space.
1091  *
1092  *      @size: the target new reservation window size
1093  *
1094  *      @group_first_block: the first block we consider to start
1095  *                      the real search from
1096  *
1097  *      @last_block:
1098  *              the maximum block number that our goal reservable space
1099  *              could start from. This is normally the last block in this
1100  *              group. The search will end when we found the start of next
1101  *              possible reservable space is out of this boundary.
1102  *              This could handle the cross boundary reservation window
1103  *              request.
1104  *
1105  *      basically we search from the given range, rather than the whole
1106  *      reservation double linked list, (start_block, last_block)
1107  *      to find a free region that is of my size and has not
1108  *      been reserved.
1109  *
1110  */
1111 static int find_next_reservable_window(
1112                                 struct ext4_reserve_window_node *search_head,
1113                                 struct ext4_reserve_window_node *my_rsv,
1114                                 struct super_block * sb,
1115                                 ext4_fsblk_t start_block,
1116                                 ext4_fsblk_t last_block)
1117 {
1118         struct rb_node *next;
1119         struct ext4_reserve_window_node *rsv, *prev;
1120         ext4_fsblk_t cur;
1121         int size = my_rsv->rsv_goal_size;
1122
1123         /* TODO: make the start of the reservation window byte-aligned */
1124         /* cur = *start_block & ~7;*/
1125         cur = start_block;
1126         rsv = search_head;
1127         if (!rsv)
1128                 return -1;
1129
1130         while (1) {
1131                 if (cur <= rsv->rsv_end)
1132                         cur = rsv->rsv_end + 1;
1133
1134                 /* TODO?
1135                  * in the case we could not find a reservable space
1136                  * that is what is expected, during the re-search, we could
1137                  * remember what's the largest reservable space we could have
1138                  * and return that one.
1139                  *
1140                  * For now it will fail if we could not find the reservable
1141                  * space with expected-size (or more)...
1142                  */
1143                 if (cur > last_block)
1144                         return -1;              /* fail */
1145
1146                 prev = rsv;
1147                 next = rb_next(&rsv->rsv_node);
1148                 rsv = rb_entry(next,struct ext4_reserve_window_node,rsv_node);
1149
1150                 /*
1151                  * Reached the last reservation, we can just append to the
1152                  * previous one.
1153                  */
1154                 if (!next)
1155                         break;
1156
1157                 if (cur + size <= rsv->rsv_start) {
1158                         /*
1159                          * Found a reserveable space big enough.  We could
1160                          * have a reservation across the group boundary here
1161                          */
1162                         break;
1163                 }
1164         }
1165         /*
1166          * we come here either :
1167          * when we reach the end of the whole list,
1168          * and there is empty reservable space after last entry in the list.
1169          * append it to the end of the list.
1170          *
1171          * or we found one reservable space in the middle of the list,
1172          * return the reservation window that we could append to.
1173          * succeed.
1174          */
1175
1176         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
1177                 rsv_window_remove(sb, my_rsv);
1178
1179         /*
1180          * Let's book the whole avaliable window for now.  We will check the
1181          * disk bitmap later and then, if there are free blocks then we adjust
1182          * the window size if it's larger than requested.
1183          * Otherwise, we will remove this node from the tree next time
1184          * call find_next_reservable_window.
1185          */
1186         my_rsv->rsv_start = cur;
1187         my_rsv->rsv_end = cur + size - 1;
1188         my_rsv->rsv_alloc_hit = 0;
1189
1190         if (prev != my_rsv)
1191                 ext4_rsv_window_add(sb, my_rsv);
1192
1193         return 0;
1194 }
1195
1196 /**
1197  *      alloc_new_reservation()--allocate a new reservation window
1198  *
1199  *              To make a new reservation, we search part of the filesystem
1200  *              reservation list (the list that inside the group). We try to
1201  *              allocate a new reservation window near the allocation goal,
1202  *              or the beginning of the group, if there is no goal.
1203  *
1204  *              We first find a reservable space after the goal, then from
1205  *              there, we check the bitmap for the first free block after
1206  *              it. If there is no free block until the end of group, then the
1207  *              whole group is full, we failed. Otherwise, check if the free
1208  *              block is inside the expected reservable space, if so, we
1209  *              succeed.
1210  *              If the first free block is outside the reservable space, then
1211  *              start from the first free block, we search for next available
1212  *              space, and go on.
1213  *
1214  *      on succeed, a new reservation will be found and inserted into the list
1215  *      It contains at least one free block, and it does not overlap with other
1216  *      reservation windows.
1217  *
1218  *      failed: we failed to find a reservation window in this group
1219  *
1220  *      @rsv: the reservation
1221  *
1222  *      @grp_goal: The goal (group-relative).  It is where the search for a
1223  *              free reservable space should start from.
1224  *              if we have a grp_goal(grp_goal >0 ), then start from there,
1225  *              no grp_goal(grp_goal = -1), we start from the first block
1226  *              of the group.
1227  *
1228  *      @sb: the super block
1229  *      @group: the group we are trying to allocate in
1230  *      @bitmap_bh: the block group block bitmap
1231  *
1232  */
1233 static int alloc_new_reservation(struct ext4_reserve_window_node *my_rsv,
1234                 ext4_grpblk_t grp_goal, struct super_block *sb,
1235                 ext4_group_t group, struct buffer_head *bitmap_bh)
1236 {
1237         struct ext4_reserve_window_node *search_head;
1238         ext4_fsblk_t group_first_block, group_end_block, start_block;
1239         ext4_grpblk_t first_free_block;
1240         struct rb_root *fs_rsv_root = &EXT4_SB(sb)->s_rsv_window_root;
1241         unsigned long size;
1242         int ret;
1243         spinlock_t *rsv_lock = &EXT4_SB(sb)->s_rsv_window_lock;
1244
1245         group_first_block = ext4_group_first_block_no(sb, group);
1246         group_end_block = group_first_block + (EXT4_BLOCKS_PER_GROUP(sb) - 1);
1247
1248         if (grp_goal < 0)
1249                 start_block = group_first_block;
1250         else
1251                 start_block = grp_goal + group_first_block;
1252
1253         size = my_rsv->rsv_goal_size;
1254
1255         if (!rsv_is_empty(&my_rsv->rsv_window)) {
1256                 /*
1257                  * if the old reservation is cross group boundary
1258                  * and if the goal is inside the old reservation window,
1259                  * we will come here when we just failed to allocate from
1260                  * the first part of the window. We still have another part
1261                  * that belongs to the next group. In this case, there is no
1262                  * point to discard our window and try to allocate a new one
1263                  * in this group(which will fail). we should
1264                  * keep the reservation window, just simply move on.
1265                  *
1266                  * Maybe we could shift the start block of the reservation
1267                  * window to the first block of next group.
1268                  */
1269
1270                 if ((my_rsv->rsv_start <= group_end_block) &&
1271                                 (my_rsv->rsv_end > group_end_block) &&
1272                                 (start_block >= my_rsv->rsv_start))
1273                         return -1;
1274
1275                 if ((my_rsv->rsv_alloc_hit >
1276                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
1277                         /*
1278                          * if the previously allocation hit ratio is
1279                          * greater than 1/2, then we double the size of
1280                          * the reservation window the next time,
1281                          * otherwise we keep the same size window
1282                          */
1283                         size = size * 2;
1284                         if (size > EXT4_MAX_RESERVE_BLOCKS)
1285                                 size = EXT4_MAX_RESERVE_BLOCKS;
1286                         my_rsv->rsv_goal_size= size;
1287                 }
1288         }
1289
1290         spin_lock(rsv_lock);
1291         /*
1292          * shift the search start to the window near the goal block
1293          */
1294         search_head = search_reserve_window(fs_rsv_root, start_block);
1295
1296         /*
1297          * find_next_reservable_window() simply finds a reservable window
1298          * inside the given range(start_block, group_end_block).
1299          *
1300          * To make sure the reservation window has a free bit inside it, we
1301          * need to check the bitmap after we found a reservable window.
1302          */
1303 retry:
1304         ret = find_next_reservable_window(search_head, my_rsv, sb,
1305                                                 start_block, group_end_block);
1306
1307         if (ret == -1) {
1308                 if (!rsv_is_empty(&my_rsv->rsv_window))
1309                         rsv_window_remove(sb, my_rsv);
1310                 spin_unlock(rsv_lock);
1311                 return -1;
1312         }
1313
1314         /*
1315          * On success, find_next_reservable_window() returns the
1316          * reservation window where there is a reservable space after it.
1317          * Before we reserve this reservable space, we need
1318          * to make sure there is at least a free block inside this region.
1319          *
1320          * searching the first free bit on the block bitmap and copy of
1321          * last committed bitmap alternatively, until we found a allocatable
1322          * block. Search start from the start block of the reservable space
1323          * we just found.
1324          */
1325         spin_unlock(rsv_lock);
1326         first_free_block = bitmap_search_next_usable_block(
1327                         my_rsv->rsv_start - group_first_block,
1328                         bitmap_bh, group_end_block - group_first_block + 1);
1329
1330         if (first_free_block < 0) {
1331                 /*
1332                  * no free block left on the bitmap, no point
1333                  * to reserve the space. return failed.
1334                  */
1335                 spin_lock(rsv_lock);
1336                 if (!rsv_is_empty(&my_rsv->rsv_window))
1337                         rsv_window_remove(sb, my_rsv);
1338                 spin_unlock(rsv_lock);
1339                 return -1;              /* failed */
1340         }
1341
1342         start_block = first_free_block + group_first_block;
1343         /*
1344          * check if the first free block is within the
1345          * free space we just reserved
1346          */
1347         if (start_block >= my_rsv->rsv_start && start_block <= my_rsv->rsv_end)
1348                 return 0;               /* success */
1349         /*
1350          * if the first free bit we found is out of the reservable space
1351          * continue search for next reservable space,
1352          * start from where the free block is,
1353          * we also shift the list head to where we stopped last time
1354          */
1355         search_head = my_rsv;
1356         spin_lock(rsv_lock);
1357         goto retry;
1358 }
1359
1360 /**
1361  * try_to_extend_reservation()
1362  * @my_rsv:             given reservation window
1363  * @sb:                 super block
1364  * @size:               the delta to extend
1365  *
1366  * Attempt to expand the reservation window large enough to have
1367  * required number of free blocks
1368  *
1369  * Since ext4_try_to_allocate() will always allocate blocks within
1370  * the reservation window range, if the window size is too small,
1371  * multiple blocks allocation has to stop at the end of the reservation
1372  * window. To make this more efficient, given the total number of
1373  * blocks needed and the current size of the window, we try to
1374  * expand the reservation window size if necessary on a best-effort
1375  * basis before ext4_new_blocks() tries to allocate blocks,
1376  */
1377 static void try_to_extend_reservation(struct ext4_reserve_window_node *my_rsv,
1378                         struct super_block *sb, int size)
1379 {
1380         struct ext4_reserve_window_node *next_rsv;
1381         struct rb_node *next;
1382         spinlock_t *rsv_lock = &EXT4_SB(sb)->s_rsv_window_lock;
1383
1384         if (!spin_trylock(rsv_lock))
1385                 return;
1386
1387         next = rb_next(&my_rsv->rsv_node);
1388
1389         if (!next)
1390                 my_rsv->rsv_end += size;
1391         else {
1392                 next_rsv = rb_entry(next, struct ext4_reserve_window_node, rsv_node);
1393
1394                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1395                         my_rsv->rsv_end += size;
1396                 else
1397                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1398         }
1399         spin_unlock(rsv_lock);
1400 }
1401
1402 /**
1403  * ext4_try_to_allocate_with_rsv()
1404  * @sb:                 superblock
1405  * @handle:             handle to this transaction
1406  * @group:              given allocation block group
1407  * @bitmap_bh:          bufferhead holds the block bitmap
1408  * @grp_goal:           given target block within the group
1409  * @count:              target number of blocks to allocate
1410  * @my_rsv:             reservation window
1411  * @errp:               pointer to store the error code
1412  *
1413  * This is the main function used to allocate a new block and its reservation
1414  * window.
1415  *
1416  * Each time when a new block allocation is need, first try to allocate from
1417  * its own reservation.  If it does not have a reservation window, instead of
1418  * looking for a free bit on bitmap first, then look up the reservation list to
1419  * see if it is inside somebody else's reservation window, we try to allocate a
1420  * reservation window for it starting from the goal first. Then do the block
1421  * allocation within the reservation window.
1422  *
1423  * This will avoid keeping on searching the reservation list again and
1424  * again when somebody is looking for a free block (without
1425  * reservation), and there are lots of free blocks, but they are all
1426  * being reserved.
1427  *
1428  * We use a red-black tree for the per-filesystem reservation list.
1429  *
1430  */
1431 static ext4_grpblk_t
1432 ext4_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1433                         ext4_group_t group, struct buffer_head *bitmap_bh,
1434                         ext4_grpblk_t grp_goal,
1435                         struct ext4_reserve_window_node * my_rsv,
1436                         unsigned long *count, int *errp)
1437 {
1438         ext4_fsblk_t group_first_block, group_last_block;
1439         ext4_grpblk_t ret = 0;
1440         int fatal;
1441         unsigned long num = *count;
1442
1443         *errp = 0;
1444
1445         /*
1446          * Make sure we use undo access for the bitmap, because it is critical
1447          * that we do the frozen_data COW on bitmap buffers in all cases even
1448          * if the buffer is in BJ_Forget state in the committing transaction.
1449          */
1450         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1451         fatal = ext4_journal_get_undo_access(handle, bitmap_bh);
1452         if (fatal) {
1453                 *errp = fatal;
1454                 return -1;
1455         }
1456
1457         /*
1458          * we don't deal with reservation when
1459          * filesystem is mounted without reservation
1460          * or the file is not a regular file
1461          * or last attempt to allocate a block with reservation turned on failed
1462          */
1463         if (my_rsv == NULL ) {
1464                 ret = ext4_try_to_allocate(sb, handle, group, bitmap_bh,
1465                                                 grp_goal, count, NULL);
1466                 goto out;
1467         }
1468         /*
1469          * grp_goal is a group relative block number (if there is a goal)
1470          * 0 <= grp_goal < EXT4_BLOCKS_PER_GROUP(sb)
1471          * first block is a filesystem wide block number
1472          * first block is the block number of the first block in this group
1473          */
1474         group_first_block = ext4_group_first_block_no(sb, group);
1475         group_last_block = group_first_block + (EXT4_BLOCKS_PER_GROUP(sb) - 1);
1476
1477         /*
1478          * Basically we will allocate a new block from inode's reservation
1479          * window.
1480          *
1481          * We need to allocate a new reservation window, if:
1482          * a) inode does not have a reservation window; or
1483          * b) last attempt to allocate a block from existing reservation
1484          *    failed; or
1485          * c) we come here with a goal and with a reservation window
1486          *
1487          * We do not need to allocate a new reservation window if we come here
1488          * at the beginning with a goal and the goal is inside the window, or
1489          * we don't have a goal but already have a reservation window.
1490          * then we could go to allocate from the reservation window directly.
1491          */
1492         while (1) {
1493                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1494                         !goal_in_my_reservation(&my_rsv->rsv_window,
1495                                                 grp_goal, group, sb)) {
1496                         if (my_rsv->rsv_goal_size < *count)
1497                                 my_rsv->rsv_goal_size = *count;
1498                         ret = alloc_new_reservation(my_rsv, grp_goal, sb,
1499                                                         group, bitmap_bh);
1500                         if (ret < 0)
1501                                 break;                  /* failed */
1502
1503                         if (!goal_in_my_reservation(&my_rsv->rsv_window,
1504                                                         grp_goal, group, sb))
1505                                 grp_goal = -1;
1506                 } else if (grp_goal >= 0) {
1507                         int curr = my_rsv->rsv_end -
1508                                         (grp_goal + group_first_block) + 1;
1509
1510                         if (curr < *count)
1511                                 try_to_extend_reservation(my_rsv, sb,
1512                                                         *count - curr);
1513                 }
1514
1515                 if ((my_rsv->rsv_start > group_last_block) ||
1516                                 (my_rsv->rsv_end < group_first_block)) {
1517                         rsv_window_dump(&EXT4_SB(sb)->s_rsv_window_root, 1);
1518                         BUG();
1519                 }
1520                 ret = ext4_try_to_allocate(sb, handle, group, bitmap_bh,
1521                                            grp_goal, &num, &my_rsv->rsv_window);
1522                 if (ret >= 0) {
1523                         my_rsv->rsv_alloc_hit += num;
1524                         *count = num;
1525                         break;                          /* succeed */
1526                 }
1527                 num = *count;
1528         }
1529 out:
1530         if (ret >= 0) {
1531                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1532                                         "bitmap block");
1533                 fatal = ext4_journal_dirty_metadata(handle, bitmap_bh);
1534                 if (fatal) {
1535                         *errp = fatal;
1536                         return -1;
1537                 }
1538                 return ret;
1539         }
1540
1541         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1542         ext4_journal_release_buffer(handle, bitmap_bh);
1543         return ret;
1544 }
1545
1546 /**
1547  * ext4_has_free_blocks()
1548  * @sbi:                in-core super block structure.
1549  *
1550  * Check if filesystem has at least 1 free block available for allocation.
1551  */
1552 static int ext4_has_free_blocks(struct ext4_sb_info *sbi)
1553 {
1554         ext4_fsblk_t free_blocks, root_blocks;
1555
1556         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1557         root_blocks = ext4_r_blocks_count(sbi->s_es);
1558         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
1559                 sbi->s_resuid != current->fsuid &&
1560                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
1561                 return 0;
1562         }
1563         return 1;
1564 }
1565
1566 /**
1567  * ext4_should_retry_alloc()
1568  * @sb:                 super block
1569  * @retries             number of attemps has been made
1570  *
1571  * ext4_should_retry_alloc() is called when ENOSPC is returned, and if
1572  * it is profitable to retry the operation, this function will wait
1573  * for the current or commiting transaction to complete, and then
1574  * return TRUE.
1575  *
1576  * if the total number of retries exceed three times, return FALSE.
1577  */
1578 int ext4_should_retry_alloc(struct super_block *sb, int *retries)
1579 {
1580         if (!ext4_has_free_blocks(EXT4_SB(sb)) || (*retries)++ > 3)
1581                 return 0;
1582
1583         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1584
1585         return jbd2_journal_force_commit_nested(EXT4_SB(sb)->s_journal);
1586 }
1587
1588 /**
1589  * ext4_new_blocks_old() -- core block(s) allocation function
1590  * @handle:             handle to this transaction
1591  * @inode:              file inode
1592  * @goal:               given target block(filesystem wide)
1593  * @count:              target number of blocks to allocate
1594  * @errp:               error code
1595  *
1596  * ext4_new_blocks uses a goal block to assist allocation.  It tries to
1597  * allocate block(s) from the block group contains the goal block first. If that
1598  * fails, it will try to allocate block(s) from other block groups without
1599  * any specific goal block.
1600  *
1601  */
1602 ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
1603                         ext4_fsblk_t goal, unsigned long *count, int *errp)
1604 {
1605         struct buffer_head *bitmap_bh = NULL;
1606         struct buffer_head *gdp_bh;
1607         ext4_group_t group_no;
1608         ext4_group_t goal_group;
1609         ext4_grpblk_t grp_target_blk;   /* blockgroup relative goal block */
1610         ext4_grpblk_t grp_alloc_blk;    /* blockgroup-relative allocated block*/
1611         ext4_fsblk_t ret_block;         /* filesyetem-wide allocated block */
1612         ext4_group_t bgi;                       /* blockgroup iteration index */
1613         int fatal = 0, err;
1614         int performed_allocation = 0;
1615         ext4_grpblk_t free_blocks;      /* number of free blocks in a group */
1616         struct super_block *sb;
1617         struct ext4_group_desc *gdp;
1618         struct ext4_super_block *es;
1619         struct ext4_sb_info *sbi;
1620         struct ext4_reserve_window_node *my_rsv = NULL;
1621         struct ext4_block_alloc_info *block_i;
1622         unsigned short windowsz = 0;
1623         ext4_group_t ngroups;
1624         unsigned long num = *count;
1625
1626         *errp = -ENOSPC;
1627         sb = inode->i_sb;
1628         if (!sb) {
1629                 printk("ext4_new_block: nonexistent device");
1630                 return 0;
1631         }
1632
1633         /*
1634          * Check quota for allocation of this block.
1635          */
1636         if (DQUOT_ALLOC_BLOCK(inode, num)) {
1637                 *errp = -EDQUOT;
1638                 return 0;
1639         }
1640
1641         sbi = EXT4_SB(sb);
1642         es = EXT4_SB(sb)->s_es;
1643         ext4_debug("goal=%llu.\n", goal);
1644         /*
1645          * Allocate a block from reservation only when
1646          * filesystem is mounted with reservation(default,-o reservation), and
1647          * it's a regular file, and
1648          * the desired window size is greater than 0 (One could use ioctl
1649          * command EXT4_IOC_SETRSVSZ to set the window size to 0 to turn off
1650          * reservation on that particular file)
1651          */
1652         block_i = EXT4_I(inode)->i_block_alloc_info;
1653         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1654                 my_rsv = &block_i->rsv_window_node;
1655
1656         if (!ext4_has_free_blocks(sbi)) {
1657                 *errp = -ENOSPC;
1658                 goto out;
1659         }
1660
1661         /*
1662          * First, test whether the goal block is free.
1663          */
1664         if (goal < le32_to_cpu(es->s_first_data_block) ||
1665             goal >= ext4_blocks_count(es))
1666                 goal = le32_to_cpu(es->s_first_data_block);
1667         ext4_get_group_no_and_offset(sb, goal, &group_no, &grp_target_blk);
1668         goal_group = group_no;
1669 retry_alloc:
1670         gdp = ext4_get_group_desc(sb, group_no, &gdp_bh);
1671         if (!gdp)
1672                 goto io_error;
1673
1674         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1675         /*
1676          * if there is not enough free blocks to make a new resevation
1677          * turn off reservation for this allocation
1678          */
1679         if (my_rsv && (free_blocks < windowsz)
1680                 && (rsv_is_empty(&my_rsv->rsv_window)))
1681                 my_rsv = NULL;
1682
1683         if (free_blocks > 0) {
1684                 bitmap_bh = read_block_bitmap(sb, group_no);
1685                 if (!bitmap_bh)
1686                         goto io_error;
1687                 grp_alloc_blk = ext4_try_to_allocate_with_rsv(sb, handle,
1688                                         group_no, bitmap_bh, grp_target_blk,
1689                                         my_rsv, &num, &fatal);
1690                 if (fatal)
1691                         goto out;
1692                 if (grp_alloc_blk >= 0)
1693                         goto allocated;
1694         }
1695
1696         ngroups = EXT4_SB(sb)->s_groups_count;
1697         smp_rmb();
1698
1699         /*
1700          * Now search the rest of the groups.  We assume that
1701          * group_no and gdp correctly point to the last group visited.
1702          */
1703         for (bgi = 0; bgi < ngroups; bgi++) {
1704                 group_no++;
1705                 if (group_no >= ngroups)
1706                         group_no = 0;
1707                 gdp = ext4_get_group_desc(sb, group_no, &gdp_bh);
1708                 if (!gdp)
1709                         goto io_error;
1710                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1711                 /*
1712                  * skip this group if the number of
1713                  * free blocks is less than half of the reservation
1714                  * window size.
1715                  */
1716                 if (free_blocks <= (windowsz/2))
1717                         continue;
1718
1719                 brelse(bitmap_bh);
1720                 bitmap_bh = read_block_bitmap(sb, group_no);
1721                 if (!bitmap_bh)
1722                         goto io_error;
1723                 /*
1724                  * try to allocate block(s) from this group, without a goal(-1).
1725                  */
1726                 grp_alloc_blk = ext4_try_to_allocate_with_rsv(sb, handle,
1727                                         group_no, bitmap_bh, -1, my_rsv,
1728                                         &num, &fatal);
1729                 if (fatal)
1730                         goto out;
1731                 if (grp_alloc_blk >= 0)
1732                         goto allocated;
1733         }
1734         /*
1735          * We may end up a bogus ealier ENOSPC error due to
1736          * filesystem is "full" of reservations, but
1737          * there maybe indeed free blocks avaliable on disk
1738          * In this case, we just forget about the reservations
1739          * just do block allocation as without reservations.
1740          */
1741         if (my_rsv) {
1742                 my_rsv = NULL;
1743                 windowsz = 0;
1744                 group_no = goal_group;
1745                 goto retry_alloc;
1746         }
1747         /* No space left on the device */
1748         *errp = -ENOSPC;
1749         goto out;
1750
1751 allocated:
1752
1753         ext4_debug("using block group %lu(%d)\n",
1754                         group_no, gdp->bg_free_blocks_count);
1755
1756         BUFFER_TRACE(gdp_bh, "get_write_access");
1757         fatal = ext4_journal_get_write_access(handle, gdp_bh);
1758         if (fatal)
1759                 goto out;
1760
1761         ret_block = grp_alloc_blk + ext4_group_first_block_no(sb, group_no);
1762
1763         if (in_range(ext4_block_bitmap(sb, gdp), ret_block, num) ||
1764             in_range(ext4_inode_bitmap(sb, gdp), ret_block, num) ||
1765             in_range(ret_block, ext4_inode_table(sb, gdp),
1766                      EXT4_SB(sb)->s_itb_per_group) ||
1767             in_range(ret_block + num - 1, ext4_inode_table(sb, gdp),
1768                      EXT4_SB(sb)->s_itb_per_group)) {
1769                 ext4_error(sb, "ext4_new_block",
1770                             "Allocating block in system zone - "
1771                             "blocks from %llu, length %lu",
1772                              ret_block, num);
1773                 goto out;
1774         }
1775
1776         performed_allocation = 1;
1777
1778 #ifdef CONFIG_JBD2_DEBUG
1779         {
1780                 struct buffer_head *debug_bh;
1781
1782                 /* Record bitmap buffer state in the newly allocated block */
1783                 debug_bh = sb_find_get_block(sb, ret_block);
1784                 if (debug_bh) {
1785                         BUFFER_TRACE(debug_bh, "state when allocated");
1786                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1787                         brelse(debug_bh);
1788                 }
1789         }
1790         jbd_lock_bh_state(bitmap_bh);
1791         spin_lock(sb_bgl_lock(sbi, group_no));
1792         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1793                 int i;
1794
1795                 for (i = 0; i < num; i++) {
1796                         if (ext4_test_bit(grp_alloc_blk+i,
1797                                         bh2jh(bitmap_bh)->b_committed_data)) {
1798                                 printk("%s: block was unexpectedly set in "
1799                                         "b_committed_data\n", __func__);
1800                         }
1801                 }
1802         }
1803         ext4_debug("found bit %d\n", grp_alloc_blk);
1804         spin_unlock(sb_bgl_lock(sbi, group_no));
1805         jbd_unlock_bh_state(bitmap_bh);
1806 #endif
1807
1808         if (ret_block + num - 1 >= ext4_blocks_count(es)) {
1809                 ext4_error(sb, "ext4_new_block",
1810                             "block(%llu) >= blocks count(%llu) - "
1811                             "block_group = %lu, es == %p ", ret_block,
1812                         ext4_blocks_count(es), group_no, es);
1813                 goto out;
1814         }
1815
1816         /*
1817          * It is up to the caller to add the new buffer to a journal
1818          * list of some description.  We don't know in advance whether
1819          * the caller wants to use it as metadata or data.
1820          */
1821         spin_lock(sb_bgl_lock(sbi, group_no));
1822         if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1823                 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
1824         le16_add_cpu(&gdp->bg_free_blocks_count, -num);
1825         gdp->bg_checksum = ext4_group_desc_csum(sbi, group_no, gdp);
1826         spin_unlock(sb_bgl_lock(sbi, group_no));
1827         percpu_counter_sub(&sbi->s_freeblocks_counter, num);
1828
1829         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1830         err = ext4_journal_dirty_metadata(handle, gdp_bh);
1831         if (!fatal)
1832                 fatal = err;
1833
1834         sb->s_dirt = 1;
1835         if (fatal)
1836                 goto out;
1837
1838         *errp = 0;
1839         brelse(bitmap_bh);
1840         DQUOT_FREE_BLOCK(inode, *count-num);
1841         *count = num;
1842         return ret_block;
1843
1844 io_error:
1845         *errp = -EIO;
1846 out:
1847         if (fatal) {
1848                 *errp = fatal;
1849                 ext4_std_error(sb, fatal);
1850         }
1851         /*
1852          * Undo the block allocation
1853          */
1854         if (!performed_allocation)
1855                 DQUOT_FREE_BLOCK(inode, *count);
1856         brelse(bitmap_bh);
1857         return 0;
1858 }
1859
1860 ext4_fsblk_t ext4_new_block(handle_t *handle, struct inode *inode,
1861                 ext4_fsblk_t goal, int *errp)
1862 {
1863         struct ext4_allocation_request ar;
1864         ext4_fsblk_t ret;
1865
1866         if (!test_opt(inode->i_sb, MBALLOC)) {
1867                 unsigned long count = 1;
1868                 ret = ext4_new_blocks_old(handle, inode, goal, &count, errp);
1869                 return ret;
1870         }
1871
1872         memset(&ar, 0, sizeof(ar));
1873         ar.inode = inode;
1874         ar.goal = goal;
1875         ar.len = 1;
1876         ret = ext4_mb_new_blocks(handle, &ar, errp);
1877         return ret;
1878 }
1879
1880 ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
1881                 ext4_fsblk_t goal, unsigned long *count, int *errp)
1882 {
1883         struct ext4_allocation_request ar;
1884         ext4_fsblk_t ret;
1885
1886         if (!test_opt(inode->i_sb, MBALLOC)) {
1887                 ret = ext4_new_blocks_old(handle, inode, goal, count, errp);
1888                 return ret;
1889         }
1890
1891         memset(&ar, 0, sizeof(ar));
1892         ar.inode = inode;
1893         ar.goal = goal;
1894         ar.len = *count;
1895         ret = ext4_mb_new_blocks(handle, &ar, errp);
1896         *count = ar.len;
1897         return ret;
1898 }
1899
1900
1901 /**
1902  * ext4_count_free_blocks() -- count filesystem free blocks
1903  * @sb:         superblock
1904  *
1905  * Adds up the number of free blocks from each block group.
1906  */
1907 ext4_fsblk_t ext4_count_free_blocks(struct super_block *sb)
1908 {
1909         ext4_fsblk_t desc_count;
1910         struct ext4_group_desc *gdp;
1911         ext4_group_t i;
1912         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
1913 #ifdef EXT4FS_DEBUG
1914         struct ext4_super_block *es;
1915         ext4_fsblk_t bitmap_count;
1916         unsigned long x;
1917         struct buffer_head *bitmap_bh = NULL;
1918
1919         es = EXT4_SB(sb)->s_es;
1920         desc_count = 0;
1921         bitmap_count = 0;
1922         gdp = NULL;
1923
1924         smp_rmb();
1925         for (i = 0; i < ngroups; i++) {
1926                 gdp = ext4_get_group_desc(sb, i, NULL);
1927                 if (!gdp)
1928                         continue;
1929                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1930                 brelse(bitmap_bh);
1931                 bitmap_bh = read_block_bitmap(sb, i);
1932                 if (bitmap_bh == NULL)
1933                         continue;
1934
1935                 x = ext4_count_free(bitmap_bh, sb->s_blocksize);
1936                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
1937                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1938                 bitmap_count += x;
1939         }
1940         brelse(bitmap_bh);
1941         printk("ext4_count_free_blocks: stored = %llu"
1942                 ", computed = %llu, %llu\n",
1943                 ext4_free_blocks_count(es),
1944                 desc_count, bitmap_count);
1945         return bitmap_count;
1946 #else
1947         desc_count = 0;
1948         smp_rmb();
1949         for (i = 0; i < ngroups; i++) {
1950                 gdp = ext4_get_group_desc(sb, i, NULL);
1951                 if (!gdp)
1952                         continue;
1953                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1954         }
1955
1956         return desc_count;
1957 #endif
1958 }
1959
1960 static inline int test_root(ext4_group_t a, int b)
1961 {
1962         int num = b;
1963
1964         while (a > num)
1965                 num *= b;
1966         return num == a;
1967 }
1968
1969 static int ext4_group_sparse(ext4_group_t group)
1970 {
1971         if (group <= 1)
1972                 return 1;
1973         if (!(group & 1))
1974                 return 0;
1975         return (test_root(group, 7) || test_root(group, 5) ||
1976                 test_root(group, 3));
1977 }
1978
1979 /**
1980  *      ext4_bg_has_super - number of blocks used by the superblock in group
1981  *      @sb: superblock for filesystem
1982  *      @group: group number to check
1983  *
1984  *      Return the number of blocks used by the superblock (primary or backup)
1985  *      in this group.  Currently this will be only 0 or 1.
1986  */
1987 int ext4_bg_has_super(struct super_block *sb, ext4_group_t group)
1988 {
1989         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
1990                                 EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1991                         !ext4_group_sparse(group))
1992                 return 0;
1993         return 1;
1994 }
1995
1996 static unsigned long ext4_bg_num_gdb_meta(struct super_block *sb,
1997                                         ext4_group_t group)
1998 {
1999         unsigned long metagroup = group / EXT4_DESC_PER_BLOCK(sb);
2000         ext4_group_t first = metagroup * EXT4_DESC_PER_BLOCK(sb);
2001         ext4_group_t last = first + EXT4_DESC_PER_BLOCK(sb) - 1;
2002
2003         if (group == first || group == first + 1 || group == last)
2004                 return 1;
2005         return 0;
2006 }
2007
2008 static unsigned long ext4_bg_num_gdb_nometa(struct super_block *sb,
2009                                         ext4_group_t group)
2010 {
2011         return ext4_bg_has_super(sb, group) ? EXT4_SB(sb)->s_gdb_count : 0;
2012 }
2013
2014 /**
2015  *      ext4_bg_num_gdb - number of blocks used by the group table in group
2016  *      @sb: superblock for filesystem
2017  *      @group: group number to check
2018  *
2019  *      Return the number of blocks used by the group descriptor table
2020  *      (primary or backup) in this group.  In the future there may be a
2021  *      different number of descriptor blocks in each group.
2022  */
2023 unsigned long ext4_bg_num_gdb(struct super_block *sb, ext4_group_t group)
2024 {
2025         unsigned long first_meta_bg =
2026                         le32_to_cpu(EXT4_SB(sb)->s_es->s_first_meta_bg);
2027         unsigned long metagroup = group / EXT4_DESC_PER_BLOCK(sb);
2028
2029         if (!EXT4_HAS_INCOMPAT_FEATURE(sb,EXT4_FEATURE_INCOMPAT_META_BG) ||
2030                         metagroup < first_meta_bg)
2031                 return ext4_bg_num_gdb_nometa(sb,group);
2032
2033         return ext4_bg_num_gdb_meta(sb,group);
2034
2035 }