NFSv4: Eliminate nfs4_path_walk()
[linux-2.6.git] / fs / ext3 / balloc.c
1 /*
2  *  linux/fs/ext3/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/slab.h>
18 #include <linux/jbd.h>
19 #include <linux/ext3_fs.h>
20 #include <linux/ext3_jbd.h>
21 #include <linux/quotaops.h>
22 #include <linux/buffer_head.h>
23
24 /*
25  * balloc.c contains the blocks allocation and deallocation routines
26  */
27
28 /*
29  * The free blocks are managed by bitmaps.  A file system contains several
30  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
31  * block for inodes, N blocks for the inode table and data blocks.
32  *
33  * The file system contains group descriptors which are located after the
34  * super block.  Each descriptor contains the number of the bitmap block and
35  * the free blocks count in the block.  The descriptors are loaded in memory
36  * when a file system is mounted (see ext3_fill_super).
37  */
38
39
40 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
41
42 /**
43  * ext3_get_group_desc() -- load group descriptor from disk
44  * @sb:                 super block
45  * @block_group:        given block group
46  * @bh:                 pointer to the buffer head to store the block
47  *                      group descriptor
48  */
49 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
50                                              unsigned int block_group,
51                                              struct buffer_head ** bh)
52 {
53         unsigned long group_desc;
54         unsigned long offset;
55         struct ext3_group_desc * desc;
56         struct ext3_sb_info *sbi = EXT3_SB(sb);
57
58         if (block_group >= sbi->s_groups_count) {
59                 ext3_error (sb, "ext3_get_group_desc",
60                             "block_group >= groups_count - "
61                             "block_group = %d, groups_count = %lu",
62                             block_group, sbi->s_groups_count);
63
64                 return NULL;
65         }
66         smp_rmb();
67
68         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
69         offset = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
70         if (!sbi->s_group_desc[group_desc]) {
71                 ext3_error (sb, "ext3_get_group_desc",
72                             "Group descriptor not loaded - "
73                             "block_group = %d, group_desc = %lu, desc = %lu",
74                              block_group, group_desc, offset);
75                 return NULL;
76         }
77
78         desc = (struct ext3_group_desc *) sbi->s_group_desc[group_desc]->b_data;
79         if (bh)
80                 *bh = sbi->s_group_desc[group_desc];
81         return desc + offset;
82 }
83
84 static int ext3_valid_block_bitmap(struct super_block *sb,
85                                         struct ext3_group_desc *desc,
86                                         unsigned int block_group,
87                                         struct buffer_head *bh)
88 {
89         ext3_grpblk_t offset;
90         ext3_grpblk_t next_zero_bit;
91         ext3_fsblk_t bitmap_blk;
92         ext3_fsblk_t group_first_block;
93
94         group_first_block = ext3_group_first_block_no(sb, block_group);
95
96         /* check whether block bitmap block number is set */
97         bitmap_blk = le32_to_cpu(desc->bg_block_bitmap);
98         offset = bitmap_blk - group_first_block;
99         if (!ext3_test_bit(offset, bh->b_data))
100                 /* bad block bitmap */
101                 goto err_out;
102
103         /* check whether the inode bitmap block number is set */
104         bitmap_blk = le32_to_cpu(desc->bg_inode_bitmap);
105         offset = bitmap_blk - group_first_block;
106         if (!ext3_test_bit(offset, bh->b_data))
107                 /* bad block bitmap */
108                 goto err_out;
109
110         /* check whether the inode table block number is set */
111         bitmap_blk = le32_to_cpu(desc->bg_inode_table);
112         offset = bitmap_blk - group_first_block;
113         next_zero_bit = ext3_find_next_zero_bit(bh->b_data,
114                                 offset + EXT3_SB(sb)->s_itb_per_group,
115                                 offset);
116         if (next_zero_bit >= offset + EXT3_SB(sb)->s_itb_per_group)
117                 /* good bitmap for inode tables */
118                 return 1;
119
120 err_out:
121         ext3_error(sb, __func__,
122                         "Invalid block bitmap - "
123                         "block_group = %d, block = %lu",
124                         block_group, bitmap_blk);
125         return 0;
126 }
127
128 /**
129  * read_block_bitmap()
130  * @sb:                 super block
131  * @block_group:        given block group
132  *
133  * Read the bitmap for a given block_group,and validate the
134  * bits for block/inode/inode tables are set in the bitmaps
135  *
136  * Return buffer_head on success or NULL in case of failure.
137  */
138 static struct buffer_head *
139 read_block_bitmap(struct super_block *sb, unsigned int block_group)
140 {
141         struct ext3_group_desc * desc;
142         struct buffer_head * bh = NULL;
143         ext3_fsblk_t bitmap_blk;
144
145         desc = ext3_get_group_desc(sb, block_group, NULL);
146         if (!desc)
147                 return NULL;
148         bitmap_blk = le32_to_cpu(desc->bg_block_bitmap);
149         bh = sb_getblk(sb, bitmap_blk);
150         if (unlikely(!bh)) {
151                 ext3_error(sb, __func__,
152                             "Cannot read block bitmap - "
153                             "block_group = %d, block_bitmap = %u",
154                             block_group, le32_to_cpu(desc->bg_block_bitmap));
155                 return NULL;
156         }
157         if (likely(bh_uptodate_or_lock(bh)))
158                 return bh;
159
160         if (bh_submit_read(bh) < 0) {
161                 brelse(bh);
162                 ext3_error(sb, __func__,
163                             "Cannot read block bitmap - "
164                             "block_group = %d, block_bitmap = %u",
165                             block_group, le32_to_cpu(desc->bg_block_bitmap));
166                 return NULL;
167         }
168         ext3_valid_block_bitmap(sb, desc, block_group, bh);
169         /*
170          * file system mounted not to panic on error, continue with corrupt
171          * bitmap
172          */
173         return bh;
174 }
175 /*
176  * The reservation window structure operations
177  * --------------------------------------------
178  * Operations include:
179  * dump, find, add, remove, is_empty, find_next_reservable_window, etc.
180  *
181  * We use a red-black tree to represent per-filesystem reservation
182  * windows.
183  *
184  */
185
186 /**
187  * __rsv_window_dump() -- Dump the filesystem block allocation reservation map
188  * @rb_root:            root of per-filesystem reservation rb tree
189  * @verbose:            verbose mode
190  * @fn:                 function which wishes to dump the reservation map
191  *
192  * If verbose is turned on, it will print the whole block reservation
193  * windows(start, end). Otherwise, it will only print out the "bad" windows,
194  * those windows that overlap with their immediate neighbors.
195  */
196 #if 1
197 static void __rsv_window_dump(struct rb_root *root, int verbose,
198                               const char *fn)
199 {
200         struct rb_node *n;
201         struct ext3_reserve_window_node *rsv, *prev;
202         int bad;
203
204 restart:
205         n = rb_first(root);
206         bad = 0;
207         prev = NULL;
208
209         printk("Block Allocation Reservation Windows Map (%s):\n", fn);
210         while (n) {
211                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
212                 if (verbose)
213                         printk("reservation window 0x%p "
214                                "start:  %lu, end:  %lu\n",
215                                rsv, rsv->rsv_start, rsv->rsv_end);
216                 if (rsv->rsv_start && rsv->rsv_start >= rsv->rsv_end) {
217                         printk("Bad reservation %p (start >= end)\n",
218                                rsv);
219                         bad = 1;
220                 }
221                 if (prev && prev->rsv_end >= rsv->rsv_start) {
222                         printk("Bad reservation %p (prev->end >= start)\n",
223                                rsv);
224                         bad = 1;
225                 }
226                 if (bad) {
227                         if (!verbose) {
228                                 printk("Restarting reservation walk in verbose mode\n");
229                                 verbose = 1;
230                                 goto restart;
231                         }
232                 }
233                 n = rb_next(n);
234                 prev = rsv;
235         }
236         printk("Window map complete.\n");
237         BUG_ON(bad);
238 }
239 #define rsv_window_dump(root, verbose) \
240         __rsv_window_dump((root), (verbose), __func__)
241 #else
242 #define rsv_window_dump(root, verbose) do {} while (0)
243 #endif
244
245 /**
246  * goal_in_my_reservation()
247  * @rsv:                inode's reservation window
248  * @grp_goal:           given goal block relative to the allocation block group
249  * @group:              the current allocation block group
250  * @sb:                 filesystem super block
251  *
252  * Test if the given goal block (group relative) is within the file's
253  * own block reservation window range.
254  *
255  * If the reservation window is outside the goal allocation group, return 0;
256  * grp_goal (given goal block) could be -1, which means no specific
257  * goal block. In this case, always return 1.
258  * If the goal block is within the reservation window, return 1;
259  * otherwise, return 0;
260  */
261 static int
262 goal_in_my_reservation(struct ext3_reserve_window *rsv, ext3_grpblk_t grp_goal,
263                         unsigned int group, struct super_block * sb)
264 {
265         ext3_fsblk_t group_first_block, group_last_block;
266
267         group_first_block = ext3_group_first_block_no(sb, group);
268         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
269
270         if ((rsv->_rsv_start > group_last_block) ||
271             (rsv->_rsv_end < group_first_block))
272                 return 0;
273         if ((grp_goal >= 0) && ((grp_goal + group_first_block < rsv->_rsv_start)
274                 || (grp_goal + group_first_block > rsv->_rsv_end)))
275                 return 0;
276         return 1;
277 }
278
279 /**
280  * search_reserve_window()
281  * @rb_root:            root of reservation tree
282  * @goal:               target allocation block
283  *
284  * Find the reserved window which includes the goal, or the previous one
285  * if the goal is not in any window.
286  * Returns NULL if there are no windows or if all windows start after the goal.
287  */
288 static struct ext3_reserve_window_node *
289 search_reserve_window(struct rb_root *root, ext3_fsblk_t goal)
290 {
291         struct rb_node *n = root->rb_node;
292         struct ext3_reserve_window_node *rsv;
293
294         if (!n)
295                 return NULL;
296
297         do {
298                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
299
300                 if (goal < rsv->rsv_start)
301                         n = n->rb_left;
302                 else if (goal > rsv->rsv_end)
303                         n = n->rb_right;
304                 else
305                         return rsv;
306         } while (n);
307         /*
308          * We've fallen off the end of the tree: the goal wasn't inside
309          * any particular node.  OK, the previous node must be to one
310          * side of the interval containing the goal.  If it's the RHS,
311          * we need to back up one.
312          */
313         if (rsv->rsv_start > goal) {
314                 n = rb_prev(&rsv->rsv_node);
315                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
316         }
317         return rsv;
318 }
319
320 /**
321  * ext3_rsv_window_add() -- Insert a window to the block reservation rb tree.
322  * @sb:                 super block
323  * @rsv:                reservation window to add
324  *
325  * Must be called with rsv_lock hold.
326  */
327 void ext3_rsv_window_add(struct super_block *sb,
328                     struct ext3_reserve_window_node *rsv)
329 {
330         struct rb_root *root = &EXT3_SB(sb)->s_rsv_window_root;
331         struct rb_node *node = &rsv->rsv_node;
332         ext3_fsblk_t start = rsv->rsv_start;
333
334         struct rb_node ** p = &root->rb_node;
335         struct rb_node * parent = NULL;
336         struct ext3_reserve_window_node *this;
337
338         while (*p)
339         {
340                 parent = *p;
341                 this = rb_entry(parent, struct ext3_reserve_window_node, rsv_node);
342
343                 if (start < this->rsv_start)
344                         p = &(*p)->rb_left;
345                 else if (start > this->rsv_end)
346                         p = &(*p)->rb_right;
347                 else {
348                         rsv_window_dump(root, 1);
349                         BUG();
350                 }
351         }
352
353         rb_link_node(node, parent, p);
354         rb_insert_color(node, root);
355 }
356
357 /**
358  * ext3_rsv_window_remove() -- unlink a window from the reservation rb tree
359  * @sb:                 super block
360  * @rsv:                reservation window to remove
361  *
362  * Mark the block reservation window as not allocated, and unlink it
363  * from the filesystem reservation window rb tree. Must be called with
364  * rsv_lock hold.
365  */
366 static void rsv_window_remove(struct super_block *sb,
367                               struct ext3_reserve_window_node *rsv)
368 {
369         rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
370         rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
371         rsv->rsv_alloc_hit = 0;
372         rb_erase(&rsv->rsv_node, &EXT3_SB(sb)->s_rsv_window_root);
373 }
374
375 /*
376  * rsv_is_empty() -- Check if the reservation window is allocated.
377  * @rsv:                given reservation window to check
378  *
379  * returns 1 if the end block is EXT3_RESERVE_WINDOW_NOT_ALLOCATED.
380  */
381 static inline int rsv_is_empty(struct ext3_reserve_window *rsv)
382 {
383         /* a valid reservation end block could not be 0 */
384         return rsv->_rsv_end == EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
385 }
386
387 /**
388  * ext3_init_block_alloc_info()
389  * @inode:              file inode structure
390  *
391  * Allocate and initialize the  reservation window structure, and
392  * link the window to the ext3 inode structure at last
393  *
394  * The reservation window structure is only dynamically allocated
395  * and linked to ext3 inode the first time the open file
396  * needs a new block. So, before every ext3_new_block(s) call, for
397  * regular files, we should check whether the reservation window
398  * structure exists or not. In the latter case, this function is called.
399  * Fail to do so will result in block reservation being turned off for that
400  * open file.
401  *
402  * This function is called from ext3_get_blocks_handle(), also called
403  * when setting the reservation window size through ioctl before the file
404  * is open for write (needs block allocation).
405  *
406  * Needs truncate_mutex protection prior to call this function.
407  */
408 void ext3_init_block_alloc_info(struct inode *inode)
409 {
410         struct ext3_inode_info *ei = EXT3_I(inode);
411         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
412         struct super_block *sb = inode->i_sb;
413
414         block_i = kmalloc(sizeof(*block_i), GFP_NOFS);
415         if (block_i) {
416                 struct ext3_reserve_window_node *rsv = &block_i->rsv_window_node;
417
418                 rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
419                 rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
420
421                 /*
422                  * if filesystem is mounted with NORESERVATION, the goal
423                  * reservation window size is set to zero to indicate
424                  * block reservation is off
425                  */
426                 if (!test_opt(sb, RESERVATION))
427                         rsv->rsv_goal_size = 0;
428                 else
429                         rsv->rsv_goal_size = EXT3_DEFAULT_RESERVE_BLOCKS;
430                 rsv->rsv_alloc_hit = 0;
431                 block_i->last_alloc_logical_block = 0;
432                 block_i->last_alloc_physical_block = 0;
433         }
434         ei->i_block_alloc_info = block_i;
435 }
436
437 /**
438  * ext3_discard_reservation()
439  * @inode:              inode
440  *
441  * Discard(free) block reservation window on last file close, or truncate
442  * or at last iput().
443  *
444  * It is being called in three cases:
445  *      ext3_release_file(): last writer close the file
446  *      ext3_clear_inode(): last iput(), when nobody link to this file.
447  *      ext3_truncate(): when the block indirect map is about to change.
448  *
449  */
450 void ext3_discard_reservation(struct inode *inode)
451 {
452         struct ext3_inode_info *ei = EXT3_I(inode);
453         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
454         struct ext3_reserve_window_node *rsv;
455         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
456
457         if (!block_i)
458                 return;
459
460         rsv = &block_i->rsv_window_node;
461         if (!rsv_is_empty(&rsv->rsv_window)) {
462                 spin_lock(rsv_lock);
463                 if (!rsv_is_empty(&rsv->rsv_window))
464                         rsv_window_remove(inode->i_sb, rsv);
465                 spin_unlock(rsv_lock);
466         }
467 }
468
469 /**
470  * ext3_free_blocks_sb() -- Free given blocks and update quota
471  * @handle:                     handle to this transaction
472  * @sb:                         super block
473  * @block:                      start physcial block to free
474  * @count:                      number of blocks to free
475  * @pdquot_freed_blocks:        pointer to quota
476  */
477 void ext3_free_blocks_sb(handle_t *handle, struct super_block *sb,
478                          ext3_fsblk_t block, unsigned long count,
479                          unsigned long *pdquot_freed_blocks)
480 {
481         struct buffer_head *bitmap_bh = NULL;
482         struct buffer_head *gd_bh;
483         unsigned long block_group;
484         ext3_grpblk_t bit;
485         unsigned long i;
486         unsigned long overflow;
487         struct ext3_group_desc * desc;
488         struct ext3_super_block * es;
489         struct ext3_sb_info *sbi;
490         int err = 0, ret;
491         ext3_grpblk_t group_freed;
492
493         *pdquot_freed_blocks = 0;
494         sbi = EXT3_SB(sb);
495         es = sbi->s_es;
496         if (block < le32_to_cpu(es->s_first_data_block) ||
497             block + count < block ||
498             block + count > le32_to_cpu(es->s_blocks_count)) {
499                 ext3_error (sb, "ext3_free_blocks",
500                             "Freeing blocks not in datazone - "
501                             "block = "E3FSBLK", count = %lu", block, count);
502                 goto error_return;
503         }
504
505         ext3_debug ("freeing block(s) %lu-%lu\n", block, block + count - 1);
506
507 do_more:
508         overflow = 0;
509         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
510                       EXT3_BLOCKS_PER_GROUP(sb);
511         bit = (block - le32_to_cpu(es->s_first_data_block)) %
512                       EXT3_BLOCKS_PER_GROUP(sb);
513         /*
514          * Check to see if we are freeing blocks across a group
515          * boundary.
516          */
517         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
518                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
519                 count -= overflow;
520         }
521         brelse(bitmap_bh);
522         bitmap_bh = read_block_bitmap(sb, block_group);
523         if (!bitmap_bh)
524                 goto error_return;
525         desc = ext3_get_group_desc (sb, block_group, &gd_bh);
526         if (!desc)
527                 goto error_return;
528
529         if (in_range (le32_to_cpu(desc->bg_block_bitmap), block, count) ||
530             in_range (le32_to_cpu(desc->bg_inode_bitmap), block, count) ||
531             in_range (block, le32_to_cpu(desc->bg_inode_table),
532                       sbi->s_itb_per_group) ||
533             in_range (block + count - 1, le32_to_cpu(desc->bg_inode_table),
534                       sbi->s_itb_per_group)) {
535                 ext3_error (sb, "ext3_free_blocks",
536                             "Freeing blocks in system zones - "
537                             "Block = "E3FSBLK", count = %lu",
538                             block, count);
539                 goto error_return;
540         }
541
542         /*
543          * We are about to start releasing blocks in the bitmap,
544          * so we need undo access.
545          */
546         /* @@@ check errors */
547         BUFFER_TRACE(bitmap_bh, "getting undo access");
548         err = ext3_journal_get_undo_access(handle, bitmap_bh);
549         if (err)
550                 goto error_return;
551
552         /*
553          * We are about to modify some metadata.  Call the journal APIs
554          * to unshare ->b_data if a currently-committing transaction is
555          * using it
556          */
557         BUFFER_TRACE(gd_bh, "get_write_access");
558         err = ext3_journal_get_write_access(handle, gd_bh);
559         if (err)
560                 goto error_return;
561
562         jbd_lock_bh_state(bitmap_bh);
563
564         for (i = 0, group_freed = 0; i < count; i++) {
565                 /*
566                  * An HJ special.  This is expensive...
567                  */
568 #ifdef CONFIG_JBD_DEBUG
569                 jbd_unlock_bh_state(bitmap_bh);
570                 {
571                         struct buffer_head *debug_bh;
572                         debug_bh = sb_find_get_block(sb, block + i);
573                         if (debug_bh) {
574                                 BUFFER_TRACE(debug_bh, "Deleted!");
575                                 if (!bh2jh(bitmap_bh)->b_committed_data)
576                                         BUFFER_TRACE(debug_bh,
577                                                 "No commited data in bitmap");
578                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
579                                 __brelse(debug_bh);
580                         }
581                 }
582                 jbd_lock_bh_state(bitmap_bh);
583 #endif
584                 if (need_resched()) {
585                         jbd_unlock_bh_state(bitmap_bh);
586                         cond_resched();
587                         jbd_lock_bh_state(bitmap_bh);
588                 }
589                 /* @@@ This prevents newly-allocated data from being
590                  * freed and then reallocated within the same
591                  * transaction.
592                  *
593                  * Ideally we would want to allow that to happen, but to
594                  * do so requires making journal_forget() capable of
595                  * revoking the queued write of a data block, which
596                  * implies blocking on the journal lock.  *forget()
597                  * cannot block due to truncate races.
598                  *
599                  * Eventually we can fix this by making journal_forget()
600                  * return a status indicating whether or not it was able
601                  * to revoke the buffer.  On successful revoke, it is
602                  * safe not to set the allocation bit in the committed
603                  * bitmap, because we know that there is no outstanding
604                  * activity on the buffer any more and so it is safe to
605                  * reallocate it.
606                  */
607                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
608                 J_ASSERT_BH(bitmap_bh,
609                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
610                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
611                                 bh2jh(bitmap_bh)->b_committed_data);
612
613                 /*
614                  * We clear the bit in the bitmap after setting the committed
615                  * data bit, because this is the reverse order to that which
616                  * the allocator uses.
617                  */
618                 BUFFER_TRACE(bitmap_bh, "clear bit");
619                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
620                                                 bit + i, bitmap_bh->b_data)) {
621                         jbd_unlock_bh_state(bitmap_bh);
622                         ext3_error(sb, __func__,
623                                 "bit already cleared for block "E3FSBLK,
624                                  block + i);
625                         jbd_lock_bh_state(bitmap_bh);
626                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
627                 } else {
628                         group_freed++;
629                 }
630         }
631         jbd_unlock_bh_state(bitmap_bh);
632
633         spin_lock(sb_bgl_lock(sbi, block_group));
634         le16_add_cpu(&desc->bg_free_blocks_count, group_freed);
635         spin_unlock(sb_bgl_lock(sbi, block_group));
636         percpu_counter_add(&sbi->s_freeblocks_counter, count);
637
638         /* We dirtied the bitmap block */
639         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
640         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
641
642         /* And the group descriptor block */
643         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
644         ret = ext3_journal_dirty_metadata(handle, gd_bh);
645         if (!err) err = ret;
646         *pdquot_freed_blocks += group_freed;
647
648         if (overflow && !err) {
649                 block += count;
650                 count = overflow;
651                 goto do_more;
652         }
653
654 error_return:
655         brelse(bitmap_bh);
656         ext3_std_error(sb, err);
657         return;
658 }
659
660 /**
661  * ext3_free_blocks() -- Free given blocks and update quota
662  * @handle:             handle for this transaction
663  * @inode:              inode
664  * @block:              start physical block to free
665  * @count:              number of blocks to count
666  */
667 void ext3_free_blocks(handle_t *handle, struct inode *inode,
668                         ext3_fsblk_t block, unsigned long count)
669 {
670         struct super_block * sb;
671         unsigned long dquot_freed_blocks;
672
673         sb = inode->i_sb;
674         if (!sb) {
675                 printk ("ext3_free_blocks: nonexistent device");
676                 return;
677         }
678         ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
679         if (dquot_freed_blocks)
680                 dquot_free_block(inode, dquot_freed_blocks);
681         return;
682 }
683
684 /**
685  * ext3_test_allocatable()
686  * @nr:                 given allocation block group
687  * @bh:                 bufferhead contains the bitmap of the given block group
688  *
689  * For ext3 allocations, we must not reuse any blocks which are
690  * allocated in the bitmap buffer's "last committed data" copy.  This
691  * prevents deletes from freeing up the page for reuse until we have
692  * committed the delete transaction.
693  *
694  * If we didn't do this, then deleting something and reallocating it as
695  * data would allow the old block to be overwritten before the
696  * transaction committed (because we force data to disk before commit).
697  * This would lead to corruption if we crashed between overwriting the
698  * data and committing the delete.
699  *
700  * @@@ We may want to make this allocation behaviour conditional on
701  * data-writes at some point, and disable it for metadata allocations or
702  * sync-data inodes.
703  */
704 static int ext3_test_allocatable(ext3_grpblk_t nr, struct buffer_head *bh)
705 {
706         int ret;
707         struct journal_head *jh = bh2jh(bh);
708
709         if (ext3_test_bit(nr, bh->b_data))
710                 return 0;
711
712         jbd_lock_bh_state(bh);
713         if (!jh->b_committed_data)
714                 ret = 1;
715         else
716                 ret = !ext3_test_bit(nr, jh->b_committed_data);
717         jbd_unlock_bh_state(bh);
718         return ret;
719 }
720
721 /**
722  * bitmap_search_next_usable_block()
723  * @start:              the starting block (group relative) of the search
724  * @bh:                 bufferhead contains the block group bitmap
725  * @maxblocks:          the ending block (group relative) of the reservation
726  *
727  * The bitmap search --- search forward alternately through the actual
728  * bitmap on disk and the last-committed copy in journal, until we find a
729  * bit free in both bitmaps.
730  */
731 static ext3_grpblk_t
732 bitmap_search_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
733                                         ext3_grpblk_t maxblocks)
734 {
735         ext3_grpblk_t next;
736         struct journal_head *jh = bh2jh(bh);
737
738         while (start < maxblocks) {
739                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, start);
740                 if (next >= maxblocks)
741                         return -1;
742                 if (ext3_test_allocatable(next, bh))
743                         return next;
744                 jbd_lock_bh_state(bh);
745                 if (jh->b_committed_data)
746                         start = ext3_find_next_zero_bit(jh->b_committed_data,
747                                                         maxblocks, next);
748                 jbd_unlock_bh_state(bh);
749         }
750         return -1;
751 }
752
753 /**
754  * find_next_usable_block()
755  * @start:              the starting block (group relative) to find next
756  *                      allocatable block in bitmap.
757  * @bh:                 bufferhead contains the block group bitmap
758  * @maxblocks:          the ending block (group relative) for the search
759  *
760  * Find an allocatable block in a bitmap.  We honor both the bitmap and
761  * its last-committed copy (if that exists), and perform the "most
762  * appropriate allocation" algorithm of looking for a free block near
763  * the initial goal; then for a free byte somewhere in the bitmap; then
764  * for any free bit in the bitmap.
765  */
766 static ext3_grpblk_t
767 find_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
768                         ext3_grpblk_t maxblocks)
769 {
770         ext3_grpblk_t here, next;
771         char *p, *r;
772
773         if (start > 0) {
774                 /*
775                  * The goal was occupied; search forward for a free
776                  * block within the next XX blocks.
777                  *
778                  * end_goal is more or less random, but it has to be
779                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
780                  * next 64-bit boundary is simple..
781                  */
782                 ext3_grpblk_t end_goal = (start + 63) & ~63;
783                 if (end_goal > maxblocks)
784                         end_goal = maxblocks;
785                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
786                 if (here < end_goal && ext3_test_allocatable(here, bh))
787                         return here;
788                 ext3_debug("Bit not found near goal\n");
789         }
790
791         here = start;
792         if (here < 0)
793                 here = 0;
794
795         p = ((char *)bh->b_data) + (here >> 3);
796         r = memscan(p, 0, ((maxblocks + 7) >> 3) - (here >> 3));
797         next = (r - ((char *)bh->b_data)) << 3;
798
799         if (next < maxblocks && next >= start && ext3_test_allocatable(next, bh))
800                 return next;
801
802         /*
803          * The bitmap search --- search forward alternately through the actual
804          * bitmap and the last-committed copy until we find a bit free in
805          * both
806          */
807         here = bitmap_search_next_usable_block(here, bh, maxblocks);
808         return here;
809 }
810
811 /**
812  * claim_block()
813  * @block:              the free block (group relative) to allocate
814  * @bh:                 the bufferhead containts the block group bitmap
815  *
816  * We think we can allocate this block in this bitmap.  Try to set the bit.
817  * If that succeeds then check that nobody has allocated and then freed the
818  * block since we saw that is was not marked in b_committed_data.  If it _was_
819  * allocated and freed then clear the bit in the bitmap again and return
820  * zero (failure).
821  */
822 static inline int
823 claim_block(spinlock_t *lock, ext3_grpblk_t block, struct buffer_head *bh)
824 {
825         struct journal_head *jh = bh2jh(bh);
826         int ret;
827
828         if (ext3_set_bit_atomic(lock, block, bh->b_data))
829                 return 0;
830         jbd_lock_bh_state(bh);
831         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
832                 ext3_clear_bit_atomic(lock, block, bh->b_data);
833                 ret = 0;
834         } else {
835                 ret = 1;
836         }
837         jbd_unlock_bh_state(bh);
838         return ret;
839 }
840
841 /**
842  * ext3_try_to_allocate()
843  * @sb:                 superblock
844  * @handle:             handle to this transaction
845  * @group:              given allocation block group
846  * @bitmap_bh:          bufferhead holds the block bitmap
847  * @grp_goal:           given target block within the group
848  * @count:              target number of blocks to allocate
849  * @my_rsv:             reservation window
850  *
851  * Attempt to allocate blocks within a give range. Set the range of allocation
852  * first, then find the first free bit(s) from the bitmap (within the range),
853  * and at last, allocate the blocks by claiming the found free bit as allocated.
854  *
855  * To set the range of this allocation:
856  *      if there is a reservation window, only try to allocate block(s) from the
857  *      file's own reservation window;
858  *      Otherwise, the allocation range starts from the give goal block, ends at
859  *      the block group's last block.
860  *
861  * If we failed to allocate the desired block then we may end up crossing to a
862  * new bitmap.  In that case we must release write access to the old one via
863  * ext3_journal_release_buffer(), else we'll run out of credits.
864  */
865 static ext3_grpblk_t
866 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
867                         struct buffer_head *bitmap_bh, ext3_grpblk_t grp_goal,
868                         unsigned long *count, struct ext3_reserve_window *my_rsv)
869 {
870         ext3_fsblk_t group_first_block;
871         ext3_grpblk_t start, end;
872         unsigned long num = 0;
873
874         /* we do allocation within the reservation window if we have a window */
875         if (my_rsv) {
876                 group_first_block = ext3_group_first_block_no(sb, group);
877                 if (my_rsv->_rsv_start >= group_first_block)
878                         start = my_rsv->_rsv_start - group_first_block;
879                 else
880                         /* reservation window cross group boundary */
881                         start = 0;
882                 end = my_rsv->_rsv_end - group_first_block + 1;
883                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
884                         /* reservation window crosses group boundary */
885                         end = EXT3_BLOCKS_PER_GROUP(sb);
886                 if ((start <= grp_goal) && (grp_goal < end))
887                         start = grp_goal;
888                 else
889                         grp_goal = -1;
890         } else {
891                 if (grp_goal > 0)
892                         start = grp_goal;
893                 else
894                         start = 0;
895                 end = EXT3_BLOCKS_PER_GROUP(sb);
896         }
897
898         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
899
900 repeat:
901         if (grp_goal < 0 || !ext3_test_allocatable(grp_goal, bitmap_bh)) {
902                 grp_goal = find_next_usable_block(start, bitmap_bh, end);
903                 if (grp_goal < 0)
904                         goto fail_access;
905                 if (!my_rsv) {
906                         int i;
907
908                         for (i = 0; i < 7 && grp_goal > start &&
909                                         ext3_test_allocatable(grp_goal - 1,
910                                                                 bitmap_bh);
911                                         i++, grp_goal--)
912                                 ;
913                 }
914         }
915         start = grp_goal;
916
917         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group),
918                 grp_goal, bitmap_bh)) {
919                 /*
920                  * The block was allocated by another thread, or it was
921                  * allocated and then freed by another thread
922                  */
923                 start++;
924                 grp_goal++;
925                 if (start >= end)
926                         goto fail_access;
927                 goto repeat;
928         }
929         num++;
930         grp_goal++;
931         while (num < *count && grp_goal < end
932                 && ext3_test_allocatable(grp_goal, bitmap_bh)
933                 && claim_block(sb_bgl_lock(EXT3_SB(sb), group),
934                                 grp_goal, bitmap_bh)) {
935                 num++;
936                 grp_goal++;
937         }
938         *count = num;
939         return grp_goal - num;
940 fail_access:
941         *count = num;
942         return -1;
943 }
944
945 /**
946  *      find_next_reservable_window():
947  *              find a reservable space within the given range.
948  *              It does not allocate the reservation window for now:
949  *              alloc_new_reservation() will do the work later.
950  *
951  *      @search_head: the head of the searching list;
952  *              This is not necessarily the list head of the whole filesystem
953  *
954  *              We have both head and start_block to assist the search
955  *              for the reservable space. The list starts from head,
956  *              but we will shift to the place where start_block is,
957  *              then start from there, when looking for a reservable space.
958  *
959  *      @size: the target new reservation window size
960  *
961  *      @group_first_block: the first block we consider to start
962  *                      the real search from
963  *
964  *      @last_block:
965  *              the maximum block number that our goal reservable space
966  *              could start from. This is normally the last block in this
967  *              group. The search will end when we found the start of next
968  *              possible reservable space is out of this boundary.
969  *              This could handle the cross boundary reservation window
970  *              request.
971  *
972  *      basically we search from the given range, rather than the whole
973  *      reservation double linked list, (start_block, last_block)
974  *      to find a free region that is of my size and has not
975  *      been reserved.
976  *
977  */
978 static int find_next_reservable_window(
979                                 struct ext3_reserve_window_node *search_head,
980                                 struct ext3_reserve_window_node *my_rsv,
981                                 struct super_block * sb,
982                                 ext3_fsblk_t start_block,
983                                 ext3_fsblk_t last_block)
984 {
985         struct rb_node *next;
986         struct ext3_reserve_window_node *rsv, *prev;
987         ext3_fsblk_t cur;
988         int size = my_rsv->rsv_goal_size;
989
990         /* TODO: make the start of the reservation window byte-aligned */
991         /* cur = *start_block & ~7;*/
992         cur = start_block;
993         rsv = search_head;
994         if (!rsv)
995                 return -1;
996
997         while (1) {
998                 if (cur <= rsv->rsv_end)
999                         cur = rsv->rsv_end + 1;
1000
1001                 /* TODO?
1002                  * in the case we could not find a reservable space
1003                  * that is what is expected, during the re-search, we could
1004                  * remember what's the largest reservable space we could have
1005                  * and return that one.
1006                  *
1007                  * For now it will fail if we could not find the reservable
1008                  * space with expected-size (or more)...
1009                  */
1010                 if (cur > last_block)
1011                         return -1;              /* fail */
1012
1013                 prev = rsv;
1014                 next = rb_next(&rsv->rsv_node);
1015                 rsv = rb_entry(next,struct ext3_reserve_window_node,rsv_node);
1016
1017                 /*
1018                  * Reached the last reservation, we can just append to the
1019                  * previous one.
1020                  */
1021                 if (!next)
1022                         break;
1023
1024                 if (cur + size <= rsv->rsv_start) {
1025                         /*
1026                          * Found a reserveable space big enough.  We could
1027                          * have a reservation across the group boundary here
1028                          */
1029                         break;
1030                 }
1031         }
1032         /*
1033          * we come here either :
1034          * when we reach the end of the whole list,
1035          * and there is empty reservable space after last entry in the list.
1036          * append it to the end of the list.
1037          *
1038          * or we found one reservable space in the middle of the list,
1039          * return the reservation window that we could append to.
1040          * succeed.
1041          */
1042
1043         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
1044                 rsv_window_remove(sb, my_rsv);
1045
1046         /*
1047          * Let's book the whole avaliable window for now.  We will check the
1048          * disk bitmap later and then, if there are free blocks then we adjust
1049          * the window size if it's larger than requested.
1050          * Otherwise, we will remove this node from the tree next time
1051          * call find_next_reservable_window.
1052          */
1053         my_rsv->rsv_start = cur;
1054         my_rsv->rsv_end = cur + size - 1;
1055         my_rsv->rsv_alloc_hit = 0;
1056
1057         if (prev != my_rsv)
1058                 ext3_rsv_window_add(sb, my_rsv);
1059
1060         return 0;
1061 }
1062
1063 /**
1064  *      alloc_new_reservation()--allocate a new reservation window
1065  *
1066  *              To make a new reservation, we search part of the filesystem
1067  *              reservation list (the list that inside the group). We try to
1068  *              allocate a new reservation window near the allocation goal,
1069  *              or the beginning of the group, if there is no goal.
1070  *
1071  *              We first find a reservable space after the goal, then from
1072  *              there, we check the bitmap for the first free block after
1073  *              it. If there is no free block until the end of group, then the
1074  *              whole group is full, we failed. Otherwise, check if the free
1075  *              block is inside the expected reservable space, if so, we
1076  *              succeed.
1077  *              If the first free block is outside the reservable space, then
1078  *              start from the first free block, we search for next available
1079  *              space, and go on.
1080  *
1081  *      on succeed, a new reservation will be found and inserted into the list
1082  *      It contains at least one free block, and it does not overlap with other
1083  *      reservation windows.
1084  *
1085  *      failed: we failed to find a reservation window in this group
1086  *
1087  *      @rsv: the reservation
1088  *
1089  *      @grp_goal: The goal (group-relative).  It is where the search for a
1090  *              free reservable space should start from.
1091  *              if we have a grp_goal(grp_goal >0 ), then start from there,
1092  *              no grp_goal(grp_goal = -1), we start from the first block
1093  *              of the group.
1094  *
1095  *      @sb: the super block
1096  *      @group: the group we are trying to allocate in
1097  *      @bitmap_bh: the block group block bitmap
1098  *
1099  */
1100 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
1101                 ext3_grpblk_t grp_goal, struct super_block *sb,
1102                 unsigned int group, struct buffer_head *bitmap_bh)
1103 {
1104         struct ext3_reserve_window_node *search_head;
1105         ext3_fsblk_t group_first_block, group_end_block, start_block;
1106         ext3_grpblk_t first_free_block;
1107         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
1108         unsigned long size;
1109         int ret;
1110         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1111
1112         group_first_block = ext3_group_first_block_no(sb, group);
1113         group_end_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1114
1115         if (grp_goal < 0)
1116                 start_block = group_first_block;
1117         else
1118                 start_block = grp_goal + group_first_block;
1119
1120         size = my_rsv->rsv_goal_size;
1121
1122         if (!rsv_is_empty(&my_rsv->rsv_window)) {
1123                 /*
1124                  * if the old reservation is cross group boundary
1125                  * and if the goal is inside the old reservation window,
1126                  * we will come here when we just failed to allocate from
1127                  * the first part of the window. We still have another part
1128                  * that belongs to the next group. In this case, there is no
1129                  * point to discard our window and try to allocate a new one
1130                  * in this group(which will fail). we should
1131                  * keep the reservation window, just simply move on.
1132                  *
1133                  * Maybe we could shift the start block of the reservation
1134                  * window to the first block of next group.
1135                  */
1136
1137                 if ((my_rsv->rsv_start <= group_end_block) &&
1138                                 (my_rsv->rsv_end > group_end_block) &&
1139                                 (start_block >= my_rsv->rsv_start))
1140                         return -1;
1141
1142                 if ((my_rsv->rsv_alloc_hit >
1143                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
1144                         /*
1145                          * if the previously allocation hit ratio is
1146                          * greater than 1/2, then we double the size of
1147                          * the reservation window the next time,
1148                          * otherwise we keep the same size window
1149                          */
1150                         size = size * 2;
1151                         if (size > EXT3_MAX_RESERVE_BLOCKS)
1152                                 size = EXT3_MAX_RESERVE_BLOCKS;
1153                         my_rsv->rsv_goal_size= size;
1154                 }
1155         }
1156
1157         spin_lock(rsv_lock);
1158         /*
1159          * shift the search start to the window near the goal block
1160          */
1161         search_head = search_reserve_window(fs_rsv_root, start_block);
1162
1163         /*
1164          * find_next_reservable_window() simply finds a reservable window
1165          * inside the given range(start_block, group_end_block).
1166          *
1167          * To make sure the reservation window has a free bit inside it, we
1168          * need to check the bitmap after we found a reservable window.
1169          */
1170 retry:
1171         ret = find_next_reservable_window(search_head, my_rsv, sb,
1172                                                 start_block, group_end_block);
1173
1174         if (ret == -1) {
1175                 if (!rsv_is_empty(&my_rsv->rsv_window))
1176                         rsv_window_remove(sb, my_rsv);
1177                 spin_unlock(rsv_lock);
1178                 return -1;
1179         }
1180
1181         /*
1182          * On success, find_next_reservable_window() returns the
1183          * reservation window where there is a reservable space after it.
1184          * Before we reserve this reservable space, we need
1185          * to make sure there is at least a free block inside this region.
1186          *
1187          * searching the first free bit on the block bitmap and copy of
1188          * last committed bitmap alternatively, until we found a allocatable
1189          * block. Search start from the start block of the reservable space
1190          * we just found.
1191          */
1192         spin_unlock(rsv_lock);
1193         first_free_block = bitmap_search_next_usable_block(
1194                         my_rsv->rsv_start - group_first_block,
1195                         bitmap_bh, group_end_block - group_first_block + 1);
1196
1197         if (first_free_block < 0) {
1198                 /*
1199                  * no free block left on the bitmap, no point
1200                  * to reserve the space. return failed.
1201                  */
1202                 spin_lock(rsv_lock);
1203                 if (!rsv_is_empty(&my_rsv->rsv_window))
1204                         rsv_window_remove(sb, my_rsv);
1205                 spin_unlock(rsv_lock);
1206                 return -1;              /* failed */
1207         }
1208
1209         start_block = first_free_block + group_first_block;
1210         /*
1211          * check if the first free block is within the
1212          * free space we just reserved
1213          */
1214         if (start_block >= my_rsv->rsv_start && start_block <= my_rsv->rsv_end)
1215                 return 0;               /* success */
1216         /*
1217          * if the first free bit we found is out of the reservable space
1218          * continue search for next reservable space,
1219          * start from where the free block is,
1220          * we also shift the list head to where we stopped last time
1221          */
1222         search_head = my_rsv;
1223         spin_lock(rsv_lock);
1224         goto retry;
1225 }
1226
1227 /**
1228  * try_to_extend_reservation()
1229  * @my_rsv:             given reservation window
1230  * @sb:                 super block
1231  * @size:               the delta to extend
1232  *
1233  * Attempt to expand the reservation window large enough to have
1234  * required number of free blocks
1235  *
1236  * Since ext3_try_to_allocate() will always allocate blocks within
1237  * the reservation window range, if the window size is too small,
1238  * multiple blocks allocation has to stop at the end of the reservation
1239  * window. To make this more efficient, given the total number of
1240  * blocks needed and the current size of the window, we try to
1241  * expand the reservation window size if necessary on a best-effort
1242  * basis before ext3_new_blocks() tries to allocate blocks,
1243  */
1244 static void try_to_extend_reservation(struct ext3_reserve_window_node *my_rsv,
1245                         struct super_block *sb, int size)
1246 {
1247         struct ext3_reserve_window_node *next_rsv;
1248         struct rb_node *next;
1249         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1250
1251         if (!spin_trylock(rsv_lock))
1252                 return;
1253
1254         next = rb_next(&my_rsv->rsv_node);
1255
1256         if (!next)
1257                 my_rsv->rsv_end += size;
1258         else {
1259                 next_rsv = rb_entry(next, struct ext3_reserve_window_node, rsv_node);
1260
1261                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1262                         my_rsv->rsv_end += size;
1263                 else
1264                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1265         }
1266         spin_unlock(rsv_lock);
1267 }
1268
1269 /**
1270  * ext3_try_to_allocate_with_rsv()
1271  * @sb:                 superblock
1272  * @handle:             handle to this transaction
1273  * @group:              given allocation block group
1274  * @bitmap_bh:          bufferhead holds the block bitmap
1275  * @grp_goal:           given target block within the group
1276  * @count:              target number of blocks to allocate
1277  * @my_rsv:             reservation window
1278  * @errp:               pointer to store the error code
1279  *
1280  * This is the main function used to allocate a new block and its reservation
1281  * window.
1282  *
1283  * Each time when a new block allocation is need, first try to allocate from
1284  * its own reservation.  If it does not have a reservation window, instead of
1285  * looking for a free bit on bitmap first, then look up the reservation list to
1286  * see if it is inside somebody else's reservation window, we try to allocate a
1287  * reservation window for it starting from the goal first. Then do the block
1288  * allocation within the reservation window.
1289  *
1290  * This will avoid keeping on searching the reservation list again and
1291  * again when somebody is looking for a free block (without
1292  * reservation), and there are lots of free blocks, but they are all
1293  * being reserved.
1294  *
1295  * We use a red-black tree for the per-filesystem reservation list.
1296  *
1297  */
1298 static ext3_grpblk_t
1299 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1300                         unsigned int group, struct buffer_head *bitmap_bh,
1301                         ext3_grpblk_t grp_goal,
1302                         struct ext3_reserve_window_node * my_rsv,
1303                         unsigned long *count, int *errp)
1304 {
1305         ext3_fsblk_t group_first_block, group_last_block;
1306         ext3_grpblk_t ret = 0;
1307         int fatal;
1308         unsigned long num = *count;
1309
1310         *errp = 0;
1311
1312         /*
1313          * Make sure we use undo access for the bitmap, because it is critical
1314          * that we do the frozen_data COW on bitmap buffers in all cases even
1315          * if the buffer is in BJ_Forget state in the committing transaction.
1316          */
1317         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1318         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1319         if (fatal) {
1320                 *errp = fatal;
1321                 return -1;
1322         }
1323
1324         /*
1325          * we don't deal with reservation when
1326          * filesystem is mounted without reservation
1327          * or the file is not a regular file
1328          * or last attempt to allocate a block with reservation turned on failed
1329          */
1330         if (my_rsv == NULL ) {
1331                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1332                                                 grp_goal, count, NULL);
1333                 goto out;
1334         }
1335         /*
1336          * grp_goal is a group relative block number (if there is a goal)
1337          * 0 <= grp_goal < EXT3_BLOCKS_PER_GROUP(sb)
1338          * first block is a filesystem wide block number
1339          * first block is the block number of the first block in this group
1340          */
1341         group_first_block = ext3_group_first_block_no(sb, group);
1342         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1343
1344         /*
1345          * Basically we will allocate a new block from inode's reservation
1346          * window.
1347          *
1348          * We need to allocate a new reservation window, if:
1349          * a) inode does not have a reservation window; or
1350          * b) last attempt to allocate a block from existing reservation
1351          *    failed; or
1352          * c) we come here with a goal and with a reservation window
1353          *
1354          * We do not need to allocate a new reservation window if we come here
1355          * at the beginning with a goal and the goal is inside the window, or
1356          * we don't have a goal but already have a reservation window.
1357          * then we could go to allocate from the reservation window directly.
1358          */
1359         while (1) {
1360                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1361                         !goal_in_my_reservation(&my_rsv->rsv_window,
1362                                                 grp_goal, group, sb)) {
1363                         if (my_rsv->rsv_goal_size < *count)
1364                                 my_rsv->rsv_goal_size = *count;
1365                         ret = alloc_new_reservation(my_rsv, grp_goal, sb,
1366                                                         group, bitmap_bh);
1367                         if (ret < 0)
1368                                 break;                  /* failed */
1369
1370                         if (!goal_in_my_reservation(&my_rsv->rsv_window,
1371                                                         grp_goal, group, sb))
1372                                 grp_goal = -1;
1373                 } else if (grp_goal >= 0) {
1374                         int curr = my_rsv->rsv_end -
1375                                         (grp_goal + group_first_block) + 1;
1376
1377                         if (curr < *count)
1378                                 try_to_extend_reservation(my_rsv, sb,
1379                                                         *count - curr);
1380                 }
1381
1382                 if ((my_rsv->rsv_start > group_last_block) ||
1383                                 (my_rsv->rsv_end < group_first_block)) {
1384                         rsv_window_dump(&EXT3_SB(sb)->s_rsv_window_root, 1);
1385                         BUG();
1386                 }
1387                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1388                                            grp_goal, &num, &my_rsv->rsv_window);
1389                 if (ret >= 0) {
1390                         my_rsv->rsv_alloc_hit += num;
1391                         *count = num;
1392                         break;                          /* succeed */
1393                 }
1394                 num = *count;
1395         }
1396 out:
1397         if (ret >= 0) {
1398                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1399                                         "bitmap block");
1400                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1401                 if (fatal) {
1402                         *errp = fatal;
1403                         return -1;
1404                 }
1405                 return ret;
1406         }
1407
1408         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1409         ext3_journal_release_buffer(handle, bitmap_bh);
1410         return ret;
1411 }
1412
1413 /**
1414  * ext3_has_free_blocks()
1415  * @sbi:                in-core super block structure.
1416  *
1417  * Check if filesystem has at least 1 free block available for allocation.
1418  */
1419 static int ext3_has_free_blocks(struct ext3_sb_info *sbi)
1420 {
1421         ext3_fsblk_t free_blocks, root_blocks;
1422
1423         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1424         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1425         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
1426                 sbi->s_resuid != current_fsuid() &&
1427                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
1428                 return 0;
1429         }
1430         return 1;
1431 }
1432
1433 /**
1434  * ext3_should_retry_alloc()
1435  * @sb:                 super block
1436  * @retries             number of attemps has been made
1437  *
1438  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1439  * it is profitable to retry the operation, this function will wait
1440  * for the current or commiting transaction to complete, and then
1441  * return TRUE.
1442  *
1443  * if the total number of retries exceed three times, return FALSE.
1444  */
1445 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1446 {
1447         if (!ext3_has_free_blocks(EXT3_SB(sb)) || (*retries)++ > 3)
1448                 return 0;
1449
1450         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1451
1452         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1453 }
1454
1455 /**
1456  * ext3_new_blocks() -- core block(s) allocation function
1457  * @handle:             handle to this transaction
1458  * @inode:              file inode
1459  * @goal:               given target block(filesystem wide)
1460  * @count:              target number of blocks to allocate
1461  * @errp:               error code
1462  *
1463  * ext3_new_blocks uses a goal block to assist allocation.  It tries to
1464  * allocate block(s) from the block group contains the goal block first. If that
1465  * fails, it will try to allocate block(s) from other block groups without
1466  * any specific goal block.
1467  *
1468  */
1469 ext3_fsblk_t ext3_new_blocks(handle_t *handle, struct inode *inode,
1470                         ext3_fsblk_t goal, unsigned long *count, int *errp)
1471 {
1472         struct buffer_head *bitmap_bh = NULL;
1473         struct buffer_head *gdp_bh;
1474         int group_no;
1475         int goal_group;
1476         ext3_grpblk_t grp_target_blk;   /* blockgroup relative goal block */
1477         ext3_grpblk_t grp_alloc_blk;    /* blockgroup-relative allocated block*/
1478         ext3_fsblk_t ret_block;         /* filesyetem-wide allocated block */
1479         int bgi;                        /* blockgroup iteration index */
1480         int fatal = 0, err;
1481         int performed_allocation = 0;
1482         ext3_grpblk_t free_blocks;      /* number of free blocks in a group */
1483         struct super_block *sb;
1484         struct ext3_group_desc *gdp;
1485         struct ext3_super_block *es;
1486         struct ext3_sb_info *sbi;
1487         struct ext3_reserve_window_node *my_rsv = NULL;
1488         struct ext3_block_alloc_info *block_i;
1489         unsigned short windowsz = 0;
1490 #ifdef EXT3FS_DEBUG
1491         static int goal_hits, goal_attempts;
1492 #endif
1493         unsigned long ngroups;
1494         unsigned long num = *count;
1495
1496         *errp = -ENOSPC;
1497         sb = inode->i_sb;
1498         if (!sb) {
1499                 printk("ext3_new_block: nonexistent device");
1500                 return 0;
1501         }
1502
1503         /*
1504          * Check quota for allocation of this block.
1505          */
1506         err = dquot_alloc_block(inode, num);
1507         if (err) {
1508                 *errp = err;
1509                 return 0;
1510         }
1511
1512         sbi = EXT3_SB(sb);
1513         es = EXT3_SB(sb)->s_es;
1514         ext3_debug("goal=%lu.\n", goal);
1515         /*
1516          * Allocate a block from reservation only when
1517          * filesystem is mounted with reservation(default,-o reservation), and
1518          * it's a regular file, and
1519          * the desired window size is greater than 0 (One could use ioctl
1520          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1521          * reservation on that particular file)
1522          */
1523         block_i = EXT3_I(inode)->i_block_alloc_info;
1524         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1525                 my_rsv = &block_i->rsv_window_node;
1526
1527         if (!ext3_has_free_blocks(sbi)) {
1528                 *errp = -ENOSPC;
1529                 goto out;
1530         }
1531
1532         /*
1533          * First, test whether the goal block is free.
1534          */
1535         if (goal < le32_to_cpu(es->s_first_data_block) ||
1536             goal >= le32_to_cpu(es->s_blocks_count))
1537                 goal = le32_to_cpu(es->s_first_data_block);
1538         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1539                         EXT3_BLOCKS_PER_GROUP(sb);
1540         goal_group = group_no;
1541 retry_alloc:
1542         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1543         if (!gdp)
1544                 goto io_error;
1545
1546         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1547         /*
1548          * if there is not enough free blocks to make a new resevation
1549          * turn off reservation for this allocation
1550          */
1551         if (my_rsv && (free_blocks < windowsz)
1552                 && (free_blocks > 0)
1553                 && (rsv_is_empty(&my_rsv->rsv_window)))
1554                 my_rsv = NULL;
1555
1556         if (free_blocks > 0) {
1557                 grp_target_blk = ((goal - le32_to_cpu(es->s_first_data_block)) %
1558                                 EXT3_BLOCKS_PER_GROUP(sb));
1559                 bitmap_bh = read_block_bitmap(sb, group_no);
1560                 if (!bitmap_bh)
1561                         goto io_error;
1562                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1563                                         group_no, bitmap_bh, grp_target_blk,
1564                                         my_rsv, &num, &fatal);
1565                 if (fatal)
1566                         goto out;
1567                 if (grp_alloc_blk >= 0)
1568                         goto allocated;
1569         }
1570
1571         ngroups = EXT3_SB(sb)->s_groups_count;
1572         smp_rmb();
1573
1574         /*
1575          * Now search the rest of the groups.  We assume that
1576          * group_no and gdp correctly point to the last group visited.
1577          */
1578         for (bgi = 0; bgi < ngroups; bgi++) {
1579                 group_no++;
1580                 if (group_no >= ngroups)
1581                         group_no = 0;
1582                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1583                 if (!gdp)
1584                         goto io_error;
1585                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1586                 /*
1587                  * skip this group if the number of
1588                  * free blocks is less than half of the reservation
1589                  * window size.
1590                  */
1591                 if (my_rsv && (free_blocks <= (windowsz/2)))
1592                         continue;
1593
1594                 brelse(bitmap_bh);
1595                 bitmap_bh = read_block_bitmap(sb, group_no);
1596                 if (!bitmap_bh)
1597                         goto io_error;
1598                 /*
1599                  * try to allocate block(s) from this group, without a goal(-1).
1600                  */
1601                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1602                                         group_no, bitmap_bh, -1, my_rsv,
1603                                         &num, &fatal);
1604                 if (fatal)
1605                         goto out;
1606                 if (grp_alloc_blk >= 0)
1607                         goto allocated;
1608         }
1609         /*
1610          * We may end up a bogus ealier ENOSPC error due to
1611          * filesystem is "full" of reservations, but
1612          * there maybe indeed free blocks avaliable on disk
1613          * In this case, we just forget about the reservations
1614          * just do block allocation as without reservations.
1615          */
1616         if (my_rsv) {
1617                 my_rsv = NULL;
1618                 windowsz = 0;
1619                 group_no = goal_group;
1620                 goto retry_alloc;
1621         }
1622         /* No space left on the device */
1623         *errp = -ENOSPC;
1624         goto out;
1625
1626 allocated:
1627
1628         ext3_debug("using block group %d(%d)\n",
1629                         group_no, gdp->bg_free_blocks_count);
1630
1631         BUFFER_TRACE(gdp_bh, "get_write_access");
1632         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1633         if (fatal)
1634                 goto out;
1635
1636         ret_block = grp_alloc_blk + ext3_group_first_block_no(sb, group_no);
1637
1638         if (in_range(le32_to_cpu(gdp->bg_block_bitmap), ret_block, num) ||
1639             in_range(le32_to_cpu(gdp->bg_inode_bitmap), ret_block, num) ||
1640             in_range(ret_block, le32_to_cpu(gdp->bg_inode_table),
1641                       EXT3_SB(sb)->s_itb_per_group) ||
1642             in_range(ret_block + num - 1, le32_to_cpu(gdp->bg_inode_table),
1643                       EXT3_SB(sb)->s_itb_per_group)) {
1644                 ext3_error(sb, "ext3_new_block",
1645                             "Allocating block in system zone - "
1646                             "blocks from "E3FSBLK", length %lu",
1647                              ret_block, num);
1648                 /*
1649                  * claim_block() marked the blocks we allocated as in use. So we
1650                  * may want to selectively mark some of the blocks as free.
1651                  */
1652                 goto retry_alloc;
1653         }
1654
1655         performed_allocation = 1;
1656
1657 #ifdef CONFIG_JBD_DEBUG
1658         {
1659                 struct buffer_head *debug_bh;
1660
1661                 /* Record bitmap buffer state in the newly allocated block */
1662                 debug_bh = sb_find_get_block(sb, ret_block);
1663                 if (debug_bh) {
1664                         BUFFER_TRACE(debug_bh, "state when allocated");
1665                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1666                         brelse(debug_bh);
1667                 }
1668         }
1669         jbd_lock_bh_state(bitmap_bh);
1670         spin_lock(sb_bgl_lock(sbi, group_no));
1671         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1672                 int i;
1673
1674                 for (i = 0; i < num; i++) {
1675                         if (ext3_test_bit(grp_alloc_blk+i,
1676                                         bh2jh(bitmap_bh)->b_committed_data)) {
1677                                 printk("%s: block was unexpectedly set in "
1678                                         "b_committed_data\n", __func__);
1679                         }
1680                 }
1681         }
1682         ext3_debug("found bit %d\n", grp_alloc_blk);
1683         spin_unlock(sb_bgl_lock(sbi, group_no));
1684         jbd_unlock_bh_state(bitmap_bh);
1685 #endif
1686
1687         if (ret_block + num - 1 >= le32_to_cpu(es->s_blocks_count)) {
1688                 ext3_error(sb, "ext3_new_block",
1689                             "block("E3FSBLK") >= blocks count(%d) - "
1690                             "block_group = %d, es == %p ", ret_block,
1691                         le32_to_cpu(es->s_blocks_count), group_no, es);
1692                 goto out;
1693         }
1694
1695         /*
1696          * It is up to the caller to add the new buffer to a journal
1697          * list of some description.  We don't know in advance whether
1698          * the caller wants to use it as metadata or data.
1699          */
1700         ext3_debug("allocating block %lu. Goal hits %d of %d.\n",
1701                         ret_block, goal_hits, goal_attempts);
1702
1703         spin_lock(sb_bgl_lock(sbi, group_no));
1704         le16_add_cpu(&gdp->bg_free_blocks_count, -num);
1705         spin_unlock(sb_bgl_lock(sbi, group_no));
1706         percpu_counter_sub(&sbi->s_freeblocks_counter, num);
1707
1708         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1709         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1710         if (!fatal)
1711                 fatal = err;
1712
1713         if (fatal)
1714                 goto out;
1715
1716         *errp = 0;
1717         brelse(bitmap_bh);
1718         dquot_free_block(inode, *count-num);
1719         *count = num;
1720         return ret_block;
1721
1722 io_error:
1723         *errp = -EIO;
1724 out:
1725         if (fatal) {
1726                 *errp = fatal;
1727                 ext3_std_error(sb, fatal);
1728         }
1729         /*
1730          * Undo the block allocation
1731          */
1732         if (!performed_allocation)
1733                 dquot_free_block(inode, *count);
1734         brelse(bitmap_bh);
1735         return 0;
1736 }
1737
1738 ext3_fsblk_t ext3_new_block(handle_t *handle, struct inode *inode,
1739                         ext3_fsblk_t goal, int *errp)
1740 {
1741         unsigned long count = 1;
1742
1743         return ext3_new_blocks(handle, inode, goal, &count, errp);
1744 }
1745
1746 /**
1747  * ext3_count_free_blocks() -- count filesystem free blocks
1748  * @sb:         superblock
1749  *
1750  * Adds up the number of free blocks from each block group.
1751  */
1752 ext3_fsblk_t ext3_count_free_blocks(struct super_block *sb)
1753 {
1754         ext3_fsblk_t desc_count;
1755         struct ext3_group_desc *gdp;
1756         int i;
1757         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1758 #ifdef EXT3FS_DEBUG
1759         struct ext3_super_block *es;
1760         ext3_fsblk_t bitmap_count;
1761         unsigned long x;
1762         struct buffer_head *bitmap_bh = NULL;
1763
1764         es = EXT3_SB(sb)->s_es;
1765         desc_count = 0;
1766         bitmap_count = 0;
1767         gdp = NULL;
1768
1769         smp_rmb();
1770         for (i = 0; i < ngroups; i++) {
1771                 gdp = ext3_get_group_desc(sb, i, NULL);
1772                 if (!gdp)
1773                         continue;
1774                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1775                 brelse(bitmap_bh);
1776                 bitmap_bh = read_block_bitmap(sb, i);
1777                 if (bitmap_bh == NULL)
1778                         continue;
1779
1780                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1781                 printk("group %d: stored = %d, counted = %lu\n",
1782                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1783                 bitmap_count += x;
1784         }
1785         brelse(bitmap_bh);
1786         printk("ext3_count_free_blocks: stored = "E3FSBLK
1787                 ", computed = "E3FSBLK", "E3FSBLK"\n",
1788                le32_to_cpu(es->s_free_blocks_count),
1789                 desc_count, bitmap_count);
1790         return bitmap_count;
1791 #else
1792         desc_count = 0;
1793         smp_rmb();
1794         for (i = 0; i < ngroups; i++) {
1795                 gdp = ext3_get_group_desc(sb, i, NULL);
1796                 if (!gdp)
1797                         continue;
1798                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1799         }
1800
1801         return desc_count;
1802 #endif
1803 }
1804
1805 static inline int test_root(int a, int b)
1806 {
1807         int num = b;
1808
1809         while (a > num)
1810                 num *= b;
1811         return num == a;
1812 }
1813
1814 static int ext3_group_sparse(int group)
1815 {
1816         if (group <= 1)
1817                 return 1;
1818         if (!(group & 1))
1819                 return 0;
1820         return (test_root(group, 7) || test_root(group, 5) ||
1821                 test_root(group, 3));
1822 }
1823
1824 /**
1825  *      ext3_bg_has_super - number of blocks used by the superblock in group
1826  *      @sb: superblock for filesystem
1827  *      @group: group number to check
1828  *
1829  *      Return the number of blocks used by the superblock (primary or backup)
1830  *      in this group.  Currently this will be only 0 or 1.
1831  */
1832 int ext3_bg_has_super(struct super_block *sb, int group)
1833 {
1834         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1835                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1836                         !ext3_group_sparse(group))
1837                 return 0;
1838         return 1;
1839 }
1840
1841 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1842 {
1843         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1844         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1845         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1846
1847         if (group == first || group == first + 1 || group == last)
1848                 return 1;
1849         return 0;
1850 }
1851
1852 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1853 {
1854         return ext3_bg_has_super(sb, group) ? EXT3_SB(sb)->s_gdb_count : 0;
1855 }
1856
1857 /**
1858  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1859  *      @sb: superblock for filesystem
1860  *      @group: group number to check
1861  *
1862  *      Return the number of blocks used by the group descriptor table
1863  *      (primary or backup) in this group.  In the future there may be a
1864  *      different number of descriptor blocks in each group.
1865  */
1866 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1867 {
1868         unsigned long first_meta_bg =
1869                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1870         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1871
1872         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1873                         metagroup < first_meta_bg)
1874                 return ext3_bg_num_gdb_nometa(sb,group);
1875
1876         return ext3_bg_num_gdb_meta(sb,group);
1877
1878 }