1 Index: linux-2.6.9-full/include/linux/ext3_fs_sb.h
2 ===================================================================
3 --- linux-2.6.9-full.orig/include/linux/ext3_fs_sb.h 2006-05-18 23:57:04.000000000 +0400
4 +++ linux-2.6.9-full/include/linux/ext3_fs_sb.h 2007-03-28 15:42:16.000000000 +0400
5 @@ -81,6 +81,61 @@ struct ext3_sb_info {
6 char *s_qf_names[MAXQUOTAS]; /* Names of quota files with journalled quota */
7 int s_jquota_fmt; /* Format of quota to use */
10 + /* for buddy allocator */
11 + struct ext3_group_info ***s_group_info;
12 + struct inode *s_buddy_cache;
13 + long s_blocks_reserved;
14 + spinlock_t s_reserve_lock;
15 + struct list_head s_active_transaction;
16 + struct list_head s_closed_transaction;
17 + struct list_head s_committed_transaction;
18 + spinlock_t s_md_lock;
19 + tid_t s_last_transaction;
20 + unsigned short *s_mb_offsets, *s_mb_maxs;
23 + unsigned long s_mb_factor;
24 + unsigned long s_stripe;
25 + unsigned long s_mb_stream_request;
26 + unsigned long s_mb_max_to_scan;
27 + unsigned long s_mb_min_to_scan;
28 + unsigned long s_mb_max_groups_to_scan;
29 + unsigned long s_mb_stats;
30 + unsigned long s_mb_order2_reqs;
32 + /* history to debug policy */
33 + struct ext3_mb_history *s_mb_history;
34 + int s_mb_history_cur;
35 + int s_mb_history_max;
36 + int s_mb_history_num;
37 + struct proc_dir_entry *s_mb_proc;
38 + spinlock_t s_mb_history_lock;
39 + int s_mb_history_filter;
41 + /* stats for buddy allocator */
42 + spinlock_t s_mb_pa_lock;
43 + atomic_t s_bal_reqs; /* number of reqs with len > 1 */
44 + atomic_t s_bal_success; /* we found long enough chunks */
45 + atomic_t s_bal_allocated; /* in blocks */
46 + atomic_t s_bal_ex_scanned; /* total extents scanned */
47 + atomic_t s_bal_goals; /* goal hits */
48 + atomic_t s_bal_breaks; /* too long searches */
49 + atomic_t s_bal_2orders; /* 2^order hits */
50 + spinlock_t s_bal_lock;
51 + unsigned long s_mb_buddies_generated;
52 + unsigned long long s_mb_generation_time;
53 + atomic_t s_mb_lost_chunks;
54 + atomic_t s_mb_preallocated;
55 + atomic_t s_mb_discarded;
57 + /* locality groups */
58 + struct ext3_locality_group *s_locality_groups;
62 +#define EXT3_GROUP_INFO(sb, group) \
63 + EXT3_SB(sb)->s_group_info[(group) >> EXT3_DESC_PER_BLOCK_BITS(sb)] \
64 + [(group) & (EXT3_DESC_PER_BLOCK(sb) - 1)]
66 #endif /* _LINUX_EXT3_FS_SB */
67 Index: linux-2.6.9-full/include/linux/ext3_fs.h
68 ===================================================================
69 --- linux-2.6.9-full.orig/include/linux/ext3_fs.h 2007-03-28 01:29:39.000000000 +0400
70 +++ linux-2.6.9-full/include/linux/ext3_fs.h 2007-03-28 15:45:07.000000000 +0400
71 @@ -57,6 +57,30 @@ struct statfs;
72 #define ext3_debug(f, a...) do {} while (0)
75 +#define EXT3_MULTIBLOCK_ALLOCATOR 1
77 +#define EXT3_MB_HINT_MERGE 1 /* prefer goal again. length */
78 +#define EXT3_MB_HINT_RESERVED 2 /* blocks already reserved */
79 +#define EXT3_MB_HINT_METADATA 4 /* metadata is being allocated */
80 +#define EXT3_MB_HINT_FIRST 8 /* first blocks in the file */
81 +#define EXT3_MB_HINT_BEST 16 /* search for the best chunk */
82 +#define EXT3_MB_HINT_DATA 32 /* data is being allocated */
83 +#define EXT3_MB_HINT_NOPREALLOC 64 /* don't preallocate (for tails) */
84 +#define EXT3_MB_HINT_GROUP_ALLOC 128 /* allocate for locality group */
85 +#define EXT3_MB_HINT_GOAL_ONLY 256 /* allocate goal blocks or none */
87 +struct ext3_allocation_request {
88 + struct inode *inode; /* target inode for block we're allocating */
89 + unsigned long logical; /* logical block in target inode */
90 + unsigned long goal; /* phys. target (a hint) */
91 + unsigned long lleft; /* the closest logical allocated block to the left */
92 + unsigned long pleft; /* phys. block for ^^^ */
93 + unsigned long lright; /* the closest logical allocated block to the right */
94 + unsigned long pright; /* phys. block for ^^^ */
95 + unsigned long len; /* how many blocks we want to allocate */
96 + unsigned long flags; /* flags. see above EXT3_MB_HINT_* */
100 * Special inodes numbers
103 #define ext3_find_first_zero_bit ext2_find_first_zero_bit
104 #define ext3_find_next_zero_bit ext2_find_next_zero_bit
106 +#ifndef ext2_find_next_le_bit
107 +#ifdef __LITTLE_ENDIAN
108 +#define ext2_find_next_le_bit(addr, size, off) find_next_bit((addr), (size), (off))
110 +#error "mballoc needs a patch for big-endian systems - Lustre bug 10634"
111 +#endif /* __LITTLE_ENDIAN */
112 +#endif /* !ext2_find_next_le_bit */
115 * Maximal mount counts between two filesystem checks
117 @@ -763,6 +787,20 @@ extern unsigned long ext3_count_dirs (st
118 extern void ext3_check_inodes_bitmap (struct super_block *);
119 extern unsigned long ext3_count_free (struct buffer_head *, unsigned);
122 +extern long ext3_mb_stats;
123 +extern long ext3_mb_max_to_scan;
124 +extern int ext3_mb_init(struct super_block *, int);
125 +extern int ext3_mb_release(struct super_block *);
126 +extern unsigned long ext3_mb_new_blocks(handle_t *, struct ext3_allocation_request *, int *);
127 +extern int ext3_mb_reserve_blocks(struct super_block *, int);
128 +extern void ext3_mb_release_blocks(struct super_block *, int);
129 +extern void ext3_mb_release_blocks(struct super_block *, int);
130 +extern void ext3_mb_discard_inode_preallocations(struct inode *);
131 +extern int __init init_ext3_proc(void);
132 +extern void exit_ext3_proc(void);
133 +extern void ext3_mb_free_blocks(handle_t *, struct inode *, unsigned long, unsigned long, int, int *);
137 extern int ext3_block_truncate_page(handle_t *, struct page *,
138 Index: linux-2.6.9-full/fs/ext3/super.c
139 ===================================================================
140 --- linux-2.6.9-full.orig/fs/ext3/super.c 2007-03-28 01:29:38.000000000 +0400
141 +++ linux-2.6.9-full/fs/ext3/super.c 2007-03-28 15:42:16.000000000 +0400
142 @@ -394,6 +394,7 @@ void ext3_put_super (struct super_block
143 struct ext3_super_block *es = sbi->s_es;
146 + ext3_mb_release(sb);
147 ext3_ext_release(sb);
148 ext3_xattr_put_super(sb);
149 journal_destroy(sbi->s_journal);
150 @@ -463,6 +464,8 @@ static struct inode *ext3_alloc_inode(st
151 ei->vfs_inode.i_version = 1;
153 memset(&ei->i_cached_extent, 0, sizeof(ei->i_cached_extent));
154 + INIT_LIST_HEAD(&ei->i_prealloc_list);
155 + spin_lock_init(&ei->i_prealloc_lock);
156 return &ei->vfs_inode;
159 @@ -2433,7 +2436,13 @@ static struct file_system_type ext3_fs_t
161 static int __init init_ext3_fs(void)
163 - int err = init_ext3_xattr();
166 + err = init_ext3_proc();
170 + err = init_ext3_xattr();
173 err = init_inodecache();
174 @@ -2455,6 +2464,7 @@ static void __exit exit_ext3_fs(void)
175 unregister_filesystem(&ext3_fs_type);
176 destroy_inodecache();
181 int ext3_prep_san_write(struct inode *inode, long *blocks,
182 Index: linux-2.6.9-full/fs/ext3/mballoc.c
183 ===================================================================
184 --- linux-2.6.9-full.orig/fs/ext3/mballoc.c 2007-02-13 18:39:59.640066087 +0300
185 +++ linux-2.6.9-full/fs/ext3/mballoc.c 2007-03-29 00:28:40.000000000 +0400
188 + * Copyright 2008 Sun Microsystems, Inc.
189 + * Written by Alex Tomas <alex@clusterfs.com>
191 + * This program is free software; you can redistribute it and/or modify
192 + * it under the terms of the GNU General Public License version 2 as
193 + * published by the Free Software Foundation.
195 + * This program is distributed in the hope that it will be useful,
196 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
197 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
198 + * GNU General Public License for more details.
200 + * You should have received a copy of the GNU General Public Licens
201 + * along with this program; if not, write to the Free Software
202 + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
207 + * mballoc.c contains the multiblocks allocation routines
210 +#include <linux/time.h>
211 +#include <linux/fs.h>
212 +#include <linux/namei.h>
213 +#include <linux/ext3_jbd.h>
214 +#include <linux/jbd.h>
215 +#include <linux/ext3_fs.h>
216 +#include <linux/quotaops.h>
217 +#include <linux/buffer_head.h>
218 +#include <linux/module.h>
219 +#include <linux/swap.h>
220 +#include <linux/proc_fs.h>
221 +#include <linux/pagemap.h>
222 +#include <linux/seq_file.h>
223 +#include <linux/version.h>
227 + * - test ext3_ext_search_left() and ext3_ext_search_right()
228 + * - search for metadata in few groups
231 + * - normalization should take into account whether file is still open
232 + * - discard preallocations if no free space left (policy?)
233 + * - don't normalize tails
235 + * - reservation for superuser
238 + * - bitmap read-ahead (proposed by Oleg Drokin aka green)
239 + * - track min/max extents in each group for better group selection
240 + * - mb_mark_used() may allocate chunk right after splitting buddy
241 + * - tree of groups sorted by number of free blocks
246 + * mballoc operates on the following data:
248 + * - in-core buddy (actually includes buddy and bitmap)
249 + * - preallocation descriptors (PAs)
251 + * there are two types of preallocations:
253 + * assiged to specific inode and can be used for this inode only.
254 + * it describes part of inode's space preallocated to specific
255 + * physical blocks. any block from that preallocated can be used
256 + * independent. the descriptor just tracks number of blocks left
257 + * unused. so, before taking some block from descriptor, one must
258 + * make sure corresponded logical block isn't allocated yet. this
259 + * also means that freeing any block within descriptor's range
260 + * must discard all preallocated blocks.
262 + * assigned to specific locality group which does not translate to
263 + * permanent set of inodes: inode can join and leave group. space
264 + * from this type of preallocation can be used for any inode. thus
265 + * it's consumed from the beginning to the end.
267 + * relation between them can be expressed as:
268 + * in-core buddy = on-disk bitmap + preallocation descriptors
270 + * this mean blocks mballoc considers used are:
271 + * - allocated blocks (persistent)
272 + * - preallocated blocks (non-persistent)
274 + * consistency in mballoc world means that at any time a block is either
275 + * free or used in ALL structures. notice: "any time" should not be read
276 + * literally -- time is discrete and delimited by locks.
278 + * to keep it simple, we don't use block numbers, instead we count number of
279 + * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
281 + * all operations can be expressed as:
282 + * - init buddy: buddy = on-disk + PAs
283 + * - new PA: buddy += N; PA = N
284 + * - use inode PA: on-disk += N; PA -= N
285 + * - discard inode PA buddy -= on-disk - PA; PA = 0
286 + * - use locality group PA on-disk += N; PA -= N
287 + * - discard locality group PA buddy -= PA; PA = 0
288 + * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
289 + * is used in real operation because we can't know actual used
290 + * bits from PA, only from on-disk bitmap
292 + * if we follow this strict logic, then all operations above should be atomic.
293 + * given some of them can block, we'd have to use something like semaphores
294 + * killing performance on high-end SMP hardware. let's try to relax it using
295 + * the following knowledge:
296 + * 1) if buddy is referenced, it's already initialized
297 + * 2) while block is used in buddy and the buddy is referenced,
298 + * nobody can re-allocate that block
299 + * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
300 + * bit set and PA claims same block, it's OK. IOW, one can set bit in
301 + * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
304 + * so, now we're building a concurrency table:
307 + * blocks for PA are allocated in the buddy, buddy must be referenced
308 + * until PA is linked to allocation group to avoid concurrent buddy init
310 + * we need to make sure that either on-disk bitmap or PA has uptodate data
311 + * given (3) we care that PA-=N operation doesn't interfere with init
312 + * - discard inode PA
313 + * the simplest way would be to have buddy initialized by the discard
314 + * - use locality group PA
315 + * again PA-=N must be serialized with init
316 + * - discard locality group PA
317 + * the simplest way would be to have buddy initialized by the discard
320 + * i_truncate_mutex serializes them
321 + * - discard inode PA
322 + * discard process must wait until PA isn't used by another process
323 + * - use locality group PA
324 + * some mutex should serialize them
325 + * - discard locality group PA
326 + * discard process must wait until PA isn't used by another process
329 + * i_truncate_mutex or another mutex should serializes them
330 + * - discard inode PA
331 + * discard process must wait until PA isn't used by another process
332 + * - use locality group PA
333 + * nothing wrong here -- they're different PAs covering different blocks
334 + * - discard locality group PA
335 + * discard process must wait until PA isn't used by another process
337 + * now we're ready to make few consequences:
338 + * - PA is referenced and while it is no discard is possible
339 + * - PA is referenced until block isn't marked in on-disk bitmap
340 + * - PA changes only after on-disk bitmap
341 + * - discard must not compete with init. either init is done before
342 + * any discard or they're serialized somehow
343 + * - buddy init as sum of on-disk bitmap and PAs is done atomically
345 + * a special case when we've used PA to emptiness. no need to modify buddy
346 + * in this case, but we should care about concurrent init
351 + * Logic in few words:
356 + * mark bits in on-disk bitmap
359 + * - use preallocation:
360 + * find proper PA (per-inode or group)
362 + * mark bits in on-disk bitmap
368 + * mark bits in on-disk bitmap
371 + * - discard preallocations in group:
373 + * move them onto local list
374 + * load on-disk bitmap
376 + * remove PA from object (inode or locality group)
377 + * mark free blocks in-core
379 + * - discard inode's preallocations:
386 + * - bitlock on a group (group)
387 + * - object (inode/locality) (object)
388 + * - per-pa lock (pa)
395 + * - find and use pa:
398 + * - release consumed pa:
403 + * - generate in-core bitmap:
407 + * - discard all for given object (inode, locality group):
412 + * - discard all for given group:
421 + * with AGGRESSIVE_CHECK allocator runs consistency checks over
422 + * structures. these checks slow things down a lot
424 +#define AGGRESSIVE_CHECK__
427 + * with DOUBLE_CHECK defined mballoc creates persistent in-core
428 + * bitmaps, maintains and uses them to check for double allocations
430 +#define DOUBLE_CHECK__
436 +#define mb_debug(fmt,a...) printk(fmt, ##a)
438 +#define mb_debug(fmt,a...)
442 + * with EXT3_MB_HISTORY mballoc stores last N allocations in memory
443 + * and you can monitor it in /proc/fs/ext3/<dev>/mb_history
445 +#define EXT3_MB_HISTORY
446 +#define EXT3_MB_HISTORY_ALLOC 1 /* allocation */
447 +#define EXT3_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
448 +#define EXT3_MB_HISTORY_DISCARD 4 /* preallocation discarded */
449 +#define EXT3_MB_HISTORY_FREE 8 /* free */
451 +#define EXT3_MB_HISTORY_DEFAULT (EXT3_MB_HISTORY_ALLOC | \
452 + EXT3_MB_HISTORY_PREALLOC | \
453 + EXT3_MB_HISTORY_DISCARD | \
454 + EXT3_MB_HISTORY_FREE)
457 + * How long mballoc can look for a best extent (in found extents)
459 +#define MB_DEFAULT_MAX_TO_SCAN 200
462 + * How long mballoc must look for a best extent
464 +#define MB_DEFAULT_MIN_TO_SCAN 10
467 + * How many groups mballoc will scan looking for the best chunk
469 +#define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
472 + * with 'ext3_mb_stats' allocator will collect stats that will be
473 + * shown at umount. The collecting costs though!
475 +#define MB_DEFAULT_STATS 1
478 + * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
479 + * by the stream allocator, which purpose is to pack requests
480 + * as close each to other as possible to produce smooth I/O traffic
482 +#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
485 + * for which requests use 2^N search using buddies
487 +#define MB_DEFAULT_ORDER2_REQS 8
490 + * default stripe size = 1MB
492 +#define MB_DEFAULT_STRIPE 256
494 +static kmem_cache_t *ext3_pspace_cachep = NULL;
496 +#ifdef EXT3_BB_MAX_BLOCKS
497 +#undef EXT3_BB_MAX_BLOCKS
499 +#define EXT3_BB_MAX_BLOCKS 30
501 +struct ext3_free_metadata {
502 + unsigned short group;
503 + unsigned short num;
504 + unsigned short blocks[EXT3_BB_MAX_BLOCKS];
505 + struct list_head list;
508 +struct ext3_group_info {
509 + unsigned long bb_state;
510 + unsigned long bb_tid;
511 + struct ext3_free_metadata *bb_md_cur;
512 + unsigned short bb_first_free;
513 + unsigned short bb_free;
514 + unsigned short bb_fragments;
515 + struct list_head bb_prealloc_list;
519 + unsigned short bb_counters[];
522 +#define EXT3_GROUP_INFO_NEED_INIT_BIT 0
523 +#define EXT3_GROUP_INFO_LOCKED_BIT 1
525 +#define EXT3_MB_GRP_NEED_INIT(grp) \
526 + (test_bit(EXT3_GROUP_INFO_NEED_INIT_BIT, &(grp)->bb_state))
529 +struct ext3_prealloc_space {
530 + struct list_head pa_inode_list;
531 + struct list_head pa_group_list;
533 + struct list_head pa_tmp_list;
534 + struct rcu_head pa_rcu;
536 + spinlock_t pa_lock;
538 + unsigned pa_deleted;
539 + unsigned long pa_pstart; /* phys. block */
540 + unsigned long pa_lstart; /* log. block */
541 + unsigned short pa_len; /* len of preallocated chunk */
542 + unsigned short pa_free; /* how many blocks are free */
543 + unsigned short pa_linear; /* consumed in one direction
544 + * strictly, for group prealloc */
545 + spinlock_t *pa_obj_lock;
546 + struct inode *pa_inode; /* hack, for history only */
550 +struct ext3_free_extent {
551 + unsigned long fe_logical;
552 + unsigned long fe_start;
553 + unsigned long fe_group;
554 + unsigned long fe_len;
559 + * we try to group all related changes together
560 + * so that writeback can flush/allocate them together as well
562 +struct ext3_locality_group {
563 + /* for allocator */
564 + struct semaphore lg_sem; /* to serialize allocates */
565 + struct list_head lg_prealloc_list;/* list of preallocations */
566 + spinlock_t lg_prealloc_lock;
569 +struct ext3_allocation_context {
570 + struct inode *ac_inode;
571 + struct super_block *ac_sb;
573 + /* original request */
574 + struct ext3_free_extent ac_o_ex;
576 + /* goal request (after normalization) */
577 + struct ext3_free_extent ac_g_ex;
579 + /* the best found extent */
580 + struct ext3_free_extent ac_b_ex;
582 + /* copy of the bext found extent taken before preallocation efforts */
583 + struct ext3_free_extent ac_f_ex;
585 + /* number of iterations done. we have to track to limit searching */
586 + unsigned long ac_ex_scanned;
587 + __u16 ac_groups_scanned;
591 + __u16 ac_flags; /* allocation hints */
595 + __u8 ac_2order; /* if request is to allocate 2^N blocks and
596 + * N > 0, the field stores N, otherwise 0 */
597 + __u8 ac_op; /* operation, for history only */
598 + struct page *ac_bitmap_page;
599 + struct page *ac_buddy_page;
600 + struct ext3_prealloc_space *ac_pa;
601 + struct ext3_locality_group *ac_lg;
604 +#define AC_STATUS_CONTINUE 1
605 +#define AC_STATUS_FOUND 2
606 +#define AC_STATUS_BREAK 3
608 +struct ext3_mb_history {
609 + struct ext3_free_extent orig; /* orig allocation */
610 + struct ext3_free_extent goal; /* goal allocation */
611 + struct ext3_free_extent result; /* result allocation */
614 + __u16 found; /* how many extents have been found */
615 + __u16 groups; /* how many groups have been scanned */
616 + __u16 tail; /* what tail broke some buddy */
617 + __u16 buddy; /* buddy the tail ^^^ broke */
619 + __u8 cr:3; /* which phase the result extent was found at */
625 + struct page *bd_buddy_page;
627 + struct page *bd_bitmap_page;
629 + struct ext3_group_info *bd_info;
630 + struct super_block *bd_sb;
634 +#define EXT3_MB_BITMAP(e3b) ((e3b)->bd_bitmap)
635 +#define EXT3_MB_BUDDY(e3b) ((e3b)->bd_buddy)
637 +#ifndef EXT3_MB_HISTORY
638 +#define ext3_mb_store_history(ac)
640 +static void ext3_mb_store_history(struct ext3_allocation_context *ac);
643 +#define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
645 +static struct proc_dir_entry *proc_root_ext3;
647 +int ext3_create (struct inode *, struct dentry *, int, struct nameidata *);
648 +struct buffer_head * read_block_bitmap(struct super_block *, unsigned int);
649 +unsigned long ext3_new_blocks_old(handle_t *handle, struct inode *inode,
650 + unsigned long goal, unsigned long *count, int *errp);
651 +void ext3_mb_release_blocks(struct super_block *, int);
652 +void ext3_mb_poll_new_transaction(struct super_block *, handle_t *);
653 +void ext3_mb_free_committed_blocks(struct super_block *);
654 +void ext3_mb_generate_from_pa(struct super_block *sb, void *bitmap, int group);
655 +void ext3_mb_free_consumed_preallocations(struct ext3_allocation_context *ac);
656 +void ext3_mb_return_to_preallocation(struct inode *inode, struct ext3_buddy *e3b,
657 + sector_t block, int count);
658 +void ext3_mb_show_ac(struct ext3_allocation_context *ac);
659 +void ext3_mb_check_with_pa(struct ext3_buddy *e3b, int first, int count);
660 +void ext3_mb_put_pa(struct ext3_allocation_context *, struct super_block *, struct ext3_prealloc_space *pa);
661 +int ext3_mb_init_per_dev_proc(struct super_block *sb);
662 +int ext3_mb_destroy_per_dev_proc(struct super_block *sb);
665 + * Calculate the block group number and offset, given a block number
667 +static void ext3_get_group_no_and_offset(struct super_block *sb,
668 + unsigned long blocknr,
669 + unsigned long *blockgrpp,
670 + unsigned long *offsetp)
672 + struct ext3_super_block *es = EXT3_SB(sb)->s_es;
673 + unsigned long offset;
675 + blocknr = blocknr - le32_to_cpu(es->s_first_data_block);
676 + offset = blocknr % EXT3_BLOCKS_PER_GROUP(sb);
677 + blocknr = blocknr / EXT3_BLOCKS_PER_GROUP(sb);
681 + *blockgrpp = blocknr;
686 +ext3_lock_group(struct super_block *sb, int group)
688 + bit_spin_lock(EXT3_GROUP_INFO_LOCKED_BIT,
689 + &EXT3_GROUP_INFO(sb, group)->bb_state);
693 +ext3_unlock_group(struct super_block *sb, int group)
695 + bit_spin_unlock(EXT3_GROUP_INFO_LOCKED_BIT,
696 + &EXT3_GROUP_INFO(sb, group)->bb_state);
700 +ext3_is_group_locked(struct super_block *sb, int group)
702 + return bit_spin_is_locked(EXT3_GROUP_INFO_LOCKED_BIT,
703 + &EXT3_GROUP_INFO(sb, group)->bb_state);
706 +unsigned long ext3_grp_offs_to_block(struct super_block *sb,
707 + struct ext3_free_extent *fex)
709 + unsigned long block;
711 + block = (unsigned long) fex->fe_group * EXT3_BLOCKS_PER_GROUP(sb)
713 + + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
717 +#if BITS_PER_LONG == 64
718 +#define mb_correct_addr_and_bit(bit,addr) \
720 + bit += ((unsigned long) addr & 7UL) << 3; \
721 + addr = (void *) ((unsigned long) addr & ~7UL); \
723 +#elif BITS_PER_LONG == 32
724 +#define mb_correct_addr_and_bit(bit,addr) \
726 + bit += ((unsigned long) addr & 3UL) << 3; \
727 + addr = (void *) ((unsigned long) addr & ~3UL); \
730 +#error "how many bits you are?!"
733 +static inline int mb_test_bit(int bit, void *addr)
735 + mb_correct_addr_and_bit(bit,addr);
736 + return ext2_test_bit(bit, addr);
739 +static inline void mb_set_bit(int bit, void *addr)
741 + mb_correct_addr_and_bit(bit,addr);
742 + ext2_set_bit(bit, addr);
745 +static inline void mb_set_bit_atomic(int bit, void *addr)
747 + mb_correct_addr_and_bit(bit,addr);
748 + ext2_set_bit_atomic(NULL, bit, addr);
751 +static inline void mb_clear_bit(int bit, void *addr)
753 + mb_correct_addr_and_bit(bit,addr);
754 + ext2_clear_bit(bit, addr);
757 +static inline void mb_clear_bit_atomic(int bit, void *addr)
759 + mb_correct_addr_and_bit(bit,addr);
760 + ext2_clear_bit_atomic(NULL, bit, addr);
763 +static inline int mb_find_next_zero_bit(void *addr, int max, int start)
766 +#if BITS_PER_LONG == 64
767 + fix = ((unsigned long) addr & 7UL) << 3;
768 + addr = (void *) ((unsigned long) addr & ~7UL);
769 +#elif BITS_PER_LONG == 32
770 + fix = ((unsigned long) addr & 3UL) << 3;
771 + addr = (void *) ((unsigned long) addr & ~3UL);
773 +#error "how many bits you are?!"
777 + return ext2_find_next_zero_bit(addr, max, start) - fix;
780 +static inline int mb_find_next_bit(void *addr, int max, int start)
783 +#if BITS_PER_LONG == 64
784 + fix = ((unsigned long) addr & 7UL) << 3;
785 + addr = (void *) ((unsigned long) addr & ~7UL);
786 +#elif BITS_PER_LONG == 32
787 + fix = ((unsigned long) addr & 3UL) << 3;
788 + addr = (void *) ((unsigned long) addr & ~3UL);
790 +#error "how many bits you are?!"
797 + return find_next_bit(addr, max, start) - fix;
801 +static inline void *mb_find_buddy(struct ext3_buddy *e3b, int order, int *max)
805 + BUG_ON(EXT3_MB_BITMAP(e3b) == EXT3_MB_BUDDY(e3b));
806 + BUG_ON(max == NULL);
808 + if (order > e3b->bd_blkbits + 1) {
813 + /* at order 0 we see each particular block */
814 + *max = 1 << (e3b->bd_blkbits + 3);
816 + return EXT3_MB_BITMAP(e3b);
818 + bb = EXT3_MB_BUDDY(e3b) + EXT3_SB(e3b->bd_sb)->s_mb_offsets[order];
819 + *max = EXT3_SB(e3b->bd_sb)->s_mb_maxs[order];
825 +void mb_free_blocks_double(struct ext3_buddy *e3b, int first, int count)
828 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
830 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
831 + for (i = 0; i < count; i++) {
832 + BUG_ON(!mb_test_bit(first + i, e3b->bd_info->bb_bitmap));
833 + mb_clear_bit(first + i, e3b->bd_info->bb_bitmap);
837 +void mb_mark_used_double(struct ext3_buddy *e3b, int first, int count)
840 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
842 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
843 + for (i = 0; i < count; i++) {
844 + BUG_ON(mb_test_bit(first + i, e3b->bd_info->bb_bitmap));
845 + mb_set_bit(first + i, e3b->bd_info->bb_bitmap);
849 +void mb_cmp_bitmaps(struct ext3_buddy *e3b, void *bitmap)
851 + if (memcmp(e3b->bd_info->bb_bitmap, bitmap, e3b->bd_sb->s_blocksize)) {
852 + unsigned char *b1, *b2;
854 + b1 = (unsigned char *) e3b->bd_info->bb_bitmap;
855 + b2 = (unsigned char *) bitmap;
856 + for (i = 0; i < e3b->bd_sb->s_blocksize; i++) {
857 + if (b1[i] != b2[i]) {
858 + printk("corruption in group %u at byte %u(%u): "
859 + "%x in copy != %x on disk/prealloc\n",
860 + e3b->bd_group, i, i * 8, b1[i], b2[i]);
868 +#define mb_free_blocks_double(a,b,c)
869 +#define mb_mark_used_double(a,b,c)
870 +#define mb_cmp_bitmaps(a,b)
873 +#ifdef AGGRESSIVE_CHECK
875 +#define MB_CHECK_ASSERT(assert) \
878 + printk (KERN_EMERG \
879 + "Assertion failure in %s() at %s:%d: \"%s\"\n", \
880 + function, file, line, # assert); \
885 +static int __mb_check_buddy(struct ext3_buddy *e3b, char *file,
886 + const char *function, int line)
888 + struct super_block *sb = e3b->bd_sb;
889 + int order = e3b->bd_blkbits + 1;
890 + int max, max2, i, j, k, count;
891 + struct ext3_group_info *grp;
892 + int fragments = 0, fstart;
893 + struct list_head *cur;
894 + void *buddy, *buddy2;
896 + if (!test_opt(sb, MBALLOC))
900 + static int mb_check_counter = 0;
901 + if (mb_check_counter++ % 100 != 0)
905 + while (order > 1) {
906 + buddy = mb_find_buddy(e3b, order, &max);
907 + MB_CHECK_ASSERT(buddy);
908 + buddy2 = mb_find_buddy(e3b, order - 1, &max2);
909 + MB_CHECK_ASSERT(buddy2);
910 + MB_CHECK_ASSERT(buddy != buddy2);
911 + MB_CHECK_ASSERT(max * 2 == max2);
914 + for (i = 0; i < max; i++) {
916 + if (mb_test_bit(i, buddy)) {
917 + /* only single bit in buddy2 may be 1 */
918 + if (!mb_test_bit(i << 1, buddy2))
919 + MB_CHECK_ASSERT(mb_test_bit((i<<1)+1, buddy2));
920 + else if (!mb_test_bit((i << 1) + 1, buddy2))
921 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
925 + /* both bits in buddy2 must be 0 */
926 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
927 + MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
929 + for (j = 0; j < (1 << order); j++) {
930 + k = (i * (1 << order)) + j;
931 + MB_CHECK_ASSERT(!mb_test_bit(k, EXT3_MB_BITMAP(e3b)));
935 + MB_CHECK_ASSERT(e3b->bd_info->bb_counters[order] == count);
940 + buddy = mb_find_buddy(e3b, 0, &max);
941 + for (i = 0; i < max; i++) {
942 + if (!mb_test_bit(i, buddy)) {
943 + MB_CHECK_ASSERT(i >= e3b->bd_info->bb_first_free);
944 + if (fstart == -1) {
951 + /* check used bits only */
952 + for (j = 0; j < e3b->bd_blkbits + 1; j++) {
953 + buddy2 = mb_find_buddy(e3b, j, &max2);
955 + MB_CHECK_ASSERT(k < max2);
956 + MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
959 + MB_CHECK_ASSERT(!EXT3_MB_GRP_NEED_INIT(e3b->bd_info));
960 + MB_CHECK_ASSERT(e3b->bd_info->bb_fragments == fragments);
962 + grp = EXT3_GROUP_INFO(sb, e3b->bd_group);
963 + buddy = mb_find_buddy(e3b, 0, &max);
964 + list_for_each(cur, &grp->bb_prealloc_list) {
965 + unsigned long groupnr;
966 + struct ext3_prealloc_space *pa;
967 + pa = list_entry(cur, struct ext3_prealloc_space, group_list);
968 + ext3_get_group_no_and_offset(sb, pa->pstart, &groupnr, &k);
969 + MB_CHECK_ASSERT(groupnr == e3b->bd_group);
970 + for (i = 0; i < pa->len; i++)
971 + MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
975 +#undef MB_CHECK_ASSERT
976 +#define mb_check_buddy(e3b) __mb_check_buddy(e3b,__FILE__,__FUNCTION__,__LINE__)
978 +#define mb_check_buddy(e3b)
981 +/* find most significant bit */
982 +static int inline fmsb(unsigned short word)
996 + } while (word != 0);
1002 +ext3_mb_mark_free_simple(struct super_block *sb, void *buddy, unsigned first,
1003 + int len, struct ext3_group_info *grp)
1005 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1006 + unsigned short min, max, chunk, border;
1008 + BUG_ON(len >= EXT3_BLOCKS_PER_GROUP(sb));
1010 + border = 2 << sb->s_blocksize_bits;
1013 + /* find how many blocks can be covered since this position */
1014 + max = ffs(first | border) - 1;
1016 + /* find how many blocks of power 2 we need to mark */
1023 + /* mark multiblock chunks only */
1024 + grp->bb_counters[min]++;
1026 + mb_clear_bit(first >> min, buddy + sbi->s_mb_offsets[min]);
1034 +ext3_mb_generate_buddy(struct super_block *sb, void *buddy, void *bitmap,
1037 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
1038 + unsigned short max = EXT3_BLOCKS_PER_GROUP(sb);
1039 + unsigned short i = 0, first, len;
1040 + unsigned free = 0, fragments = 0;
1041 + unsigned long long period = get_cycles();
1043 + /* initialize buddy from bitmap which is aggregation
1044 + * of on-disk bitmap and preallocations */
1045 + i = mb_find_next_zero_bit(bitmap, max, 0);
1046 + grp->bb_first_free = i;
1050 + i = ext2_find_next_le_bit(bitmap, max, i);
1054 + ext3_mb_mark_free_simple(sb, buddy, first, len, grp);
1056 + grp->bb_counters[0]++;
1058 + i = mb_find_next_zero_bit(bitmap, max, i);
1060 + grp->bb_fragments = fragments;
1062 + if (free != grp->bb_free) {
1063 + printk("EXT3-fs: group %u: %u blocks in bitmap, %u in gd\n",
1064 + group, free, grp->bb_free);
1065 + grp->bb_free = free;
1068 + clear_bit(EXT3_GROUP_INFO_NEED_INIT_BIT, &grp->bb_state);
1070 + period = get_cycles() - period;
1071 + spin_lock(&EXT3_SB(sb)->s_bal_lock);
1072 + EXT3_SB(sb)->s_mb_buddies_generated++;
1073 + EXT3_SB(sb)->s_mb_generation_time += period;
1074 + spin_unlock(&EXT3_SB(sb)->s_bal_lock);
1077 +static int ext3_mb_init_cache(struct page *page, char *incore)
1079 + int blocksize, blocks_per_page, groups_per_page;
1080 + int err = 0, i, first_group, first_block;
1081 + struct super_block *sb;
1082 + struct buffer_head *bhs;
1083 + struct buffer_head **bh;
1084 + struct inode *inode;
1085 + char *data, *bitmap;
1087 + mb_debug("init page %lu\n", page->index);
1089 + inode = page->mapping->host;
1091 + blocksize = 1 << inode->i_blkbits;
1092 + blocks_per_page = PAGE_CACHE_SIZE / blocksize;
1094 + groups_per_page = blocks_per_page >> 1;
1095 + if (groups_per_page == 0)
1096 + groups_per_page = 1;
1098 + /* allocate buffer_heads to read bitmaps */
1099 + if (groups_per_page > 1) {
1101 + i = sizeof(struct buffer_head *) * groups_per_page;
1102 + bh = kmalloc(i, GFP_NOFS);
1109 + first_group = page->index * blocks_per_page / 2;
1111 + /* read all groups the page covers into the cache */
1112 + for (i = 0; i < groups_per_page; i++) {
1113 + struct ext3_group_desc * desc;
1115 + if (first_group + i >= EXT3_SB(sb)->s_groups_count)
1119 + desc = ext3_get_group_desc(sb, first_group + i, NULL);
1124 + bh[i] = sb_getblk(sb, le32_to_cpu(desc->bg_block_bitmap));
1125 + if (bh[i] == NULL)
1128 + if (buffer_uptodate(bh[i]))
1131 + lock_buffer(bh[i]);
1132 + if (buffer_uptodate(bh[i])) {
1133 + unlock_buffer(bh[i]);
1138 + bh[i]->b_end_io = end_buffer_read_sync;
1139 + submit_bh(READ, bh[i]);
1140 + mb_debug("read bitmap for group %u\n", first_group + i);
1143 + /* wait for I/O completion */
1144 + for (i = 0; i < groups_per_page && bh[i]; i++)
1145 + wait_on_buffer(bh[i]);
1148 + for (i = 0; i < groups_per_page && bh[i]; i++)
1149 + if (!buffer_uptodate(bh[i]))
1152 + first_block = page->index * blocks_per_page;
1153 + for (i = 0; i < blocks_per_page; i++) {
1156 + group = (first_block + i) >> 1;
1157 + if (group >= EXT3_SB(sb)->s_groups_count)
1160 + data = page_address(page) + (i * blocksize);
1161 + bitmap = bh[group - first_group]->b_data;
1163 + if ((first_block + i) & 1) {
1164 + /* this is block of buddy */
1165 + BUG_ON(incore == NULL);
1166 + mb_debug("put buddy for group %u in page %lu/%x\n",
1167 + group, page->index, i * blocksize);
1168 + memset(data, 0xff, blocksize);
1169 + EXT3_GROUP_INFO(sb, group)->bb_fragments = 0;
1170 + memset(EXT3_GROUP_INFO(sb, group)->bb_counters, 0,
1171 + sizeof(unsigned short)*(sb->s_blocksize_bits+2));
1172 + ext3_mb_generate_buddy(sb, data, incore, group);
1175 + /* this is block of bitmap */
1176 + BUG_ON(incore != NULL);
1177 + mb_debug("put bitmap for group %u in page %lu/%x\n",
1178 + group, page->index, i * blocksize);
1180 + /* see comments in ext3_mb_put_pa() */
1181 + ext3_lock_group(sb, group);
1182 + memcpy(data, bitmap, blocksize);
1184 + /* mark all preallocated blocks used in in-core bitmap */
1185 + ext3_mb_generate_from_pa(sb, data, group);
1186 + ext3_unlock_group(sb, group);
1191 + SetPageUptodate(page);
1195 + for (i = 0; i < groups_per_page && bh[i]; i++)
1203 +static int ext3_mb_load_buddy(struct super_block *sb, int group,
1204 + struct ext3_buddy *e3b)
1206 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1207 + struct inode *inode = sbi->s_buddy_cache;
1208 + int blocks_per_page, block, pnum, poff;
1209 + struct page *page;
1211 + mb_debug("load group %u\n", group);
1213 + blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1215 + e3b->bd_blkbits = sb->s_blocksize_bits;
1216 + e3b->bd_info = EXT3_GROUP_INFO(sb, group);
1218 + e3b->bd_group = group;
1219 + e3b->bd_buddy_page = NULL;
1220 + e3b->bd_bitmap_page = NULL;
1222 + block = group * 2;
1223 + pnum = block / blocks_per_page;
1224 + poff = block % blocks_per_page;
1226 + /* we could use find_or_create_page(), but it locks page
1227 + * what we'd like to avoid in fast path ... */
1228 + page = find_get_page(inode->i_mapping, pnum);
1229 + if (page == NULL || !PageUptodate(page)) {
1231 + page_cache_release(page);
1232 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1234 + BUG_ON(page->mapping != inode->i_mapping);
1235 + if (!PageUptodate(page)) {
1236 + ext3_mb_init_cache(page, NULL);
1237 + mb_cmp_bitmaps(e3b, page_address(page) +
1238 + (poff * sb->s_blocksize));
1240 + unlock_page(page);
1243 + if (page == NULL || !PageUptodate(page))
1245 + e3b->bd_bitmap_page = page;
1246 + e3b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1247 + mark_page_accessed(page);
1250 + pnum = block / blocks_per_page;
1251 + poff = block % blocks_per_page;
1253 + page = find_get_page(inode->i_mapping, pnum);
1254 + if (page == NULL || !PageUptodate(page)) {
1256 + page_cache_release(page);
1257 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1259 + BUG_ON(page->mapping != inode->i_mapping);
1260 + if (!PageUptodate(page))
1261 + ext3_mb_init_cache(page, e3b->bd_bitmap);
1263 + unlock_page(page);
1266 + if (page == NULL || !PageUptodate(page))
1268 + e3b->bd_buddy_page = page;
1269 + e3b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1270 + mark_page_accessed(page);
1272 + BUG_ON(e3b->bd_bitmap_page == NULL);
1273 + BUG_ON(e3b->bd_buddy_page == NULL);
1278 + if (e3b->bd_bitmap_page)
1279 + page_cache_release(e3b->bd_bitmap_page);
1280 + if (e3b->bd_buddy_page)
1281 + page_cache_release(e3b->bd_buddy_page);
1282 + e3b->bd_buddy = NULL;
1283 + e3b->bd_bitmap = NULL;
1287 +static void ext3_mb_release_desc(struct ext3_buddy *e3b)
1289 + if (e3b->bd_bitmap_page)
1290 + page_cache_release(e3b->bd_bitmap_page);
1291 + if (e3b->bd_buddy_page)
1292 + page_cache_release(e3b->bd_buddy_page);
1296 +static int mb_find_order_for_block(struct ext3_buddy *e3b, int block)
1301 + BUG_ON(EXT3_MB_BITMAP(e3b) == EXT3_MB_BUDDY(e3b));
1302 + BUG_ON(block >= (1 << (e3b->bd_blkbits + 3)));
1304 + bb = EXT3_MB_BUDDY(e3b);
1305 + while (order <= e3b->bd_blkbits + 1) {
1306 + block = block >> 1;
1307 + if (!mb_test_bit(block, bb)) {
1308 + /* this block is part of buddy of order 'order' */
1311 + bb += 1 << (e3b->bd_blkbits - order);
1317 +static inline void mb_clear_bits(void *bm, int cur, int len)
1322 + while (cur < len) {
1323 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1324 + /* fast path: clear whole word at once */
1325 + addr = bm + (cur >> 3);
1330 + mb_clear_bit_atomic(cur, bm);
1335 +static inline void mb_set_bits(void *bm, int cur, int len)
1340 + while (cur < len) {
1341 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1342 + /* fast path: clear whole word at once */
1343 + addr = bm + (cur >> 3);
1344 + *addr = 0xffffffff;
1348 + mb_set_bit_atomic(cur, bm);
1353 +static int mb_free_blocks(struct ext3_buddy *e3b, int first, int count)
1355 + int block = 0, max = 0, order;
1356 + void *buddy, *buddy2;
1358 + BUG_ON(first + count > (e3b->bd_sb->s_blocksize << 3));
1359 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1360 + mb_check_buddy(e3b);
1361 + mb_free_blocks_double(e3b, first, count);
1363 + e3b->bd_info->bb_free += count;
1364 + if (first < e3b->bd_info->bb_first_free)
1365 + e3b->bd_info->bb_first_free = first;
1367 + /* let's maintain fragments counter */
1369 + block = !mb_test_bit(first - 1, EXT3_MB_BITMAP(e3b));
1370 + if (first + count < EXT3_SB(e3b->bd_sb)->s_mb_maxs[0])
1371 + max = !mb_test_bit(first + count, EXT3_MB_BITMAP(e3b));
1373 + e3b->bd_info->bb_fragments--;
1374 + else if (!block && !max)
1375 + e3b->bd_info->bb_fragments++;
1377 + /* let's maintain buddy itself */
1378 + while (count-- > 0) {
1382 + BUG_ON(!mb_test_bit(block, EXT3_MB_BITMAP(e3b)));
1383 + mb_clear_bit(block, EXT3_MB_BITMAP(e3b));
1384 + e3b->bd_info->bb_counters[order]++;
1386 + /* start of the buddy */
1387 + buddy = mb_find_buddy(e3b, order, &max);
1391 + if (mb_test_bit(block, buddy) ||
1392 + mb_test_bit(block + 1, buddy))
1395 + /* both the buddies are free, try to coalesce them */
1396 + buddy2 = mb_find_buddy(e3b, order + 1, &max);
1402 + /* for special purposes, we don't set
1403 + * free bits in bitmap */
1404 + mb_set_bit(block, buddy);
1405 + mb_set_bit(block + 1, buddy);
1407 + e3b->bd_info->bb_counters[order]--;
1408 + e3b->bd_info->bb_counters[order]--;
1410 + block = block >> 1;
1412 + e3b->bd_info->bb_counters[order]++;
1414 + mb_clear_bit(block, buddy2);
1418 + mb_check_buddy(e3b);
1423 +static int mb_find_extent(struct ext3_buddy *e3b, int order, int block,
1424 + int needed, struct ext3_free_extent *ex)
1426 + int next = block, max, ord;
1429 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1430 + BUG_ON(ex == NULL);
1432 + buddy = mb_find_buddy(e3b, order, &max);
1433 + BUG_ON(buddy == NULL);
1434 + BUG_ON(block >= max);
1435 + if (mb_test_bit(block, buddy)) {
1442 + if (likely(order == 0)) {
1443 + /* find actual order */
1444 + order = mb_find_order_for_block(e3b, block);
1445 + block = block >> order;
1448 + ex->fe_len = 1 << order;
1449 + ex->fe_start = block << order;
1450 + ex->fe_group = e3b->bd_group;
1452 + /* calc difference from given start */
1453 + next = next - ex->fe_start;
1454 + ex->fe_len -= next;
1455 + ex->fe_start += next;
1457 + while (needed > ex->fe_len && (buddy = mb_find_buddy(e3b, order, &max))) {
1459 + if (block + 1 >= max)
1462 + next = (block + 1) * (1 << order);
1463 + if (mb_test_bit(next, EXT3_MB_BITMAP(e3b)))
1466 + ord = mb_find_order_for_block(e3b, next);
1469 + block = next >> order;
1470 + ex->fe_len += 1 << order;
1473 + BUG_ON(ex->fe_start + ex->fe_len > (1 << (e3b->bd_blkbits + 3)));
1474 + return ex->fe_len;
1477 +static int mb_mark_used(struct ext3_buddy *e3b, struct ext3_free_extent *ex)
1479 + int ord, mlen = 0, max = 0, cur;
1480 + int start = ex->fe_start;
1481 + int len = ex->fe_len;
1486 + BUG_ON(start + len > (e3b->bd_sb->s_blocksize << 3));
1487 + BUG_ON(e3b->bd_group != ex->fe_group);
1488 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1489 + mb_check_buddy(e3b);
1490 + mb_mark_used_double(e3b, start, len);
1492 + e3b->bd_info->bb_free -= len;
1493 + if (e3b->bd_info->bb_first_free == start)
1494 + e3b->bd_info->bb_first_free += len;
1496 + /* let's maintain fragments counter */
1498 + mlen = !mb_test_bit(start - 1, EXT3_MB_BITMAP(e3b));
1499 + if (start + len < EXT3_SB(e3b->bd_sb)->s_mb_maxs[0])
1500 + max = !mb_test_bit(start + len, EXT3_MB_BITMAP(e3b));
1502 + e3b->bd_info->bb_fragments++;
1503 + else if (!mlen && !max)
1504 + e3b->bd_info->bb_fragments--;
1506 + /* let's maintain buddy itself */
1508 + ord = mb_find_order_for_block(e3b, start);
1510 + if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1511 + /* the whole chunk may be allocated at once! */
1513 + buddy = mb_find_buddy(e3b, ord, &max);
1514 + BUG_ON((start >> ord) >= max);
1515 + mb_set_bit(start >> ord, buddy);
1516 + e3b->bd_info->bb_counters[ord]--;
1523 + /* store for history */
1525 + ret = len | (ord << 16);
1527 + /* we have to split large buddy */
1529 + buddy = mb_find_buddy(e3b, ord, &max);
1530 + mb_set_bit(start >> ord, buddy);
1531 + e3b->bd_info->bb_counters[ord]--;
1534 + cur = (start >> ord) & ~1U;
1535 + buddy = mb_find_buddy(e3b, ord, &max);
1536 + mb_clear_bit(cur, buddy);
1537 + mb_clear_bit(cur + 1, buddy);
1538 + e3b->bd_info->bb_counters[ord]++;
1539 + e3b->bd_info->bb_counters[ord]++;
1542 + mb_set_bits(EXT3_MB_BITMAP(e3b), ex->fe_start, len0);
1543 + mb_check_buddy(e3b);
1549 + * Must be called under group lock!
1551 +static void ext3_mb_use_best_found(struct ext3_allocation_context *ac,
1552 + struct ext3_buddy *e3b)
1554 + unsigned long ret;
1556 + BUG_ON(ac->ac_b_ex.fe_group != e3b->bd_group);
1557 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1559 + ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1560 + ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1561 + ret = mb_mark_used(e3b, &ac->ac_b_ex);
1563 + /* preallocation can change ac_b_ex, thus we store actually
1564 + * allocated blocks for history */
1565 + ac->ac_f_ex = ac->ac_b_ex;
1567 + ac->ac_status = AC_STATUS_FOUND;
1568 + ac->ac_tail = ret & 0xffff;
1569 + ac->ac_buddy = ret >> 16;
1571 + /* XXXXXXX: SUCH A HORRIBLE **CK */
1572 + ac->ac_bitmap_page = e3b->bd_bitmap_page;
1573 + get_page(ac->ac_bitmap_page);
1574 + ac->ac_buddy_page = e3b->bd_buddy_page;
1575 + get_page(ac->ac_buddy_page);
1579 + * regular allocator, for general purposes allocation
1582 +void ext3_mb_check_limits(struct ext3_allocation_context *ac,
1583 + struct ext3_buddy *e3b,
1586 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1587 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1588 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1589 + struct ext3_free_extent ex;
1593 + * We don't want to scan for a whole year
1595 + if (ac->ac_found > sbi->s_mb_max_to_scan &&
1596 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1597 + ac->ac_status = AC_STATUS_BREAK;
1602 + * Haven't found good chunk so far, let's continue
1604 + if (bex->fe_len < gex->fe_len)
1607 + if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1608 + && bex->fe_group == e3b->bd_group) {
1609 + /* recheck chunk's availability - we don't know
1610 + * when it was found (within this lock-unlock
1611 + * period or not) */
1612 + max = mb_find_extent(e3b, 0, bex->fe_start, gex->fe_len, &ex);
1613 + if (max >= gex->fe_len) {
1614 + ext3_mb_use_best_found(ac, e3b);
1621 + * The routine checks whether found extent is good enough. If it is,
1622 + * then the extent gets marked used and flag is set to the context
1623 + * to stop scanning. Otherwise, the extent is compared with the
1624 + * previous found extent and if new one is better, then it's stored
1625 + * in the context. Later, the best found extent will be used, if
1626 + * mballoc can't find good enough extent.
1628 + * FIXME: real allocation policy is to be designed yet!
1630 +static void ext3_mb_measure_extent(struct ext3_allocation_context *ac,
1631 + struct ext3_free_extent *ex,
1632 + struct ext3_buddy *e3b)
1634 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1635 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1637 + BUG_ON(ex->fe_len <= 0);
1638 + BUG_ON(ex->fe_len >= (1 << ac->ac_sb->s_blocksize_bits) * 8);
1639 + BUG_ON(ex->fe_start >= (1 << ac->ac_sb->s_blocksize_bits) * 8);
1640 + BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1645 + * The special case - take what you catch first
1647 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1649 + ext3_mb_use_best_found(ac, e3b);
1654 + * Let's check whether the chuck is good enough
1656 + if (ex->fe_len == gex->fe_len) {
1658 + ext3_mb_use_best_found(ac, e3b);
1663 + * If this is first found extent, just store it in the context
1665 + if (bex->fe_len == 0) {
1671 + * If new found extent is better, store it in the context
1673 + if (bex->fe_len < gex->fe_len) {
1674 + /* if the request isn't satisfied, any found extent
1675 + * larger than previous best one is better */
1676 + if (ex->fe_len > bex->fe_len)
1678 + } else if (ex->fe_len > gex->fe_len) {
1679 + /* if the request is satisfied, then we try to find
1680 + * an extent that still satisfy the request, but is
1681 + * smaller than previous one */
1685 + ext3_mb_check_limits(ac, e3b, 0);
1688 +static int ext3_mb_try_best_found(struct ext3_allocation_context *ac,
1689 + struct ext3_buddy *e3b)
1691 + struct ext3_free_extent ex = ac->ac_b_ex;
1692 + int group = ex.fe_group, max, err;
1694 + BUG_ON(ex.fe_len <= 0);
1695 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1699 + ext3_lock_group(ac->ac_sb, group);
1700 + max = mb_find_extent(e3b, 0, ex.fe_start, ex.fe_len, &ex);
1704 + ext3_mb_use_best_found(ac, e3b);
1707 + ext3_unlock_group(ac->ac_sb, group);
1708 + ext3_mb_release_desc(e3b);
1713 +static int ext3_mb_find_by_goal(struct ext3_allocation_context *ac,
1714 + struct ext3_buddy *e3b)
1716 + int group = ac->ac_g_ex.fe_group, max, err;
1717 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1718 + struct ext3_super_block *es = sbi->s_es;
1719 + struct ext3_free_extent ex;
1721 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1725 + ext3_lock_group(ac->ac_sb, group);
1726 + max = mb_find_extent(e3b, 0, ac->ac_g_ex.fe_start,
1727 + ac->ac_g_ex.fe_len, &ex);
1729 + if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1730 + unsigned long start;
1731 + start = (e3b->bd_group * EXT3_BLOCKS_PER_GROUP(ac->ac_sb) +
1732 + ex.fe_start + le32_to_cpu(es->s_first_data_block));
1733 + if (start % sbi->s_stripe == 0) {
1736 + ext3_mb_use_best_found(ac, e3b);
1738 + } else if (max >= ac->ac_g_ex.fe_len) {
1739 + BUG_ON(ex.fe_len <= 0);
1740 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1741 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1744 + ext3_mb_use_best_found(ac, e3b);
1745 + } else if (max > 0 && (ac->ac_flags & EXT3_MB_HINT_MERGE)) {
1746 + /* Sometimes, caller may want to merge even small
1747 + * number of blocks to an existing extent */
1748 + BUG_ON(ex.fe_len <= 0);
1749 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1750 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1753 + ext3_mb_use_best_found(ac, e3b);
1755 + ext3_unlock_group(ac->ac_sb, group);
1756 + ext3_mb_release_desc(e3b);
1762 + * The routine scans buddy structures (not bitmap!) from given order
1763 + * to max order and tries to find big enough chunk to satisfy the req
1765 +static void ext3_mb_simple_scan_group(struct ext3_allocation_context *ac,
1766 + struct ext3_buddy *e3b)
1768 + struct super_block *sb = ac->ac_sb;
1769 + struct ext3_group_info *grp = e3b->bd_info;
1773 + BUG_ON(ac->ac_2order <= 0);
1774 + for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1775 + if (grp->bb_counters[i] == 0)
1778 + buddy = mb_find_buddy(e3b, i, &max);
1779 + BUG_ON(buddy == NULL);
1781 + k = mb_find_next_zero_bit(buddy, max, 0);
1786 + ac->ac_b_ex.fe_len = 1 << i;
1787 + ac->ac_b_ex.fe_start = k << i;
1788 + ac->ac_b_ex.fe_group = e3b->bd_group;
1790 + ext3_mb_use_best_found(ac, e3b);
1792 + BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1794 + if (EXT3_SB(sb)->s_mb_stats)
1795 + atomic_inc(&EXT3_SB(sb)->s_bal_2orders);
1802 + * The routine scans the group and measures all found extents.
1803 + * In order to optimize scanning, caller must pass number of
1804 + * free blocks in the group, so the routine can know upper limit.
1806 +static void ext3_mb_complex_scan_group(struct ext3_allocation_context *ac,
1807 + struct ext3_buddy *e3b)
1809 + struct super_block *sb = ac->ac_sb;
1810 + void *bitmap = EXT3_MB_BITMAP(e3b);
1811 + struct ext3_free_extent ex;
1814 + free = e3b->bd_info->bb_free;
1815 + BUG_ON(free <= 0);
1817 + i = e3b->bd_info->bb_first_free;
1819 + while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1820 + i = mb_find_next_zero_bit(bitmap, sb->s_blocksize * 8, i);
1821 + if (i >= sb->s_blocksize * 8) {
1822 + BUG_ON(free != 0);
1826 + mb_find_extent(e3b, 0, i, ac->ac_g_ex.fe_len, &ex);
1827 + BUG_ON(ex.fe_len <= 0);
1828 + BUG_ON(free < ex.fe_len);
1830 + ext3_mb_measure_extent(ac, &ex, e3b);
1833 + free -= ex.fe_len;
1836 + ext3_mb_check_limits(ac, e3b, 1);
1840 + * This is a special case for storages like raid5
1841 + * we try to find stripe-aligned chunks for stripe-size requests
1843 +static void ext3_mb_scan_aligned(struct ext3_allocation_context *ac,
1844 + struct ext3_buddy *e3b)
1846 + struct super_block *sb = ac->ac_sb;
1847 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1848 + void *bitmap = EXT3_MB_BITMAP(e3b);
1849 + struct ext3_free_extent ex;
1850 + unsigned long i, max;
1852 + BUG_ON(sbi->s_stripe == 0);
1854 + /* find first stripe-aligned block */
1855 + i = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb)
1856 + + le32_to_cpu(sbi->s_es->s_first_data_block);
1857 + i = ((i + sbi->s_stripe - 1) / sbi->s_stripe) * sbi->s_stripe;
1858 + i = (i - le32_to_cpu(sbi->s_es->s_first_data_block))
1859 + % EXT3_BLOCKS_PER_GROUP(sb);
1861 + while (i < sb->s_blocksize * 8) {
1862 + if (!mb_test_bit(i, bitmap)) {
1863 + max = mb_find_extent(e3b, 0, i, sbi->s_stripe, &ex);
1864 + if (max >= sbi->s_stripe) {
1867 + ext3_mb_use_best_found(ac, e3b);
1871 + i += sbi->s_stripe;
1875 +static int ext3_mb_good_group(struct ext3_allocation_context *ac,
1876 + int group, int cr)
1878 + struct ext3_group_info *grp = EXT3_GROUP_INFO(ac->ac_sb, group);
1879 + unsigned free, fragments, i, bits;
1881 + BUG_ON(cr < 0 || cr >= 4);
1882 + BUG_ON(EXT3_MB_GRP_NEED_INIT(grp));
1884 + free = grp->bb_free;
1885 + fragments = grp->bb_fragments;
1888 + if (fragments == 0)
1893 + BUG_ON(ac->ac_2order == 0);
1894 + bits = ac->ac_sb->s_blocksize_bits + 1;
1895 + for (i = ac->ac_2order; i <= bits; i++)
1896 + if (grp->bb_counters[i] > 0)
1900 + if ((free / fragments) >= ac->ac_g_ex.fe_len)
1904 + if (free >= ac->ac_g_ex.fe_len)
1916 +int ext3_mb_regular_allocator(struct ext3_allocation_context *ac)
1918 + int group, i, cr, err = 0;
1919 + struct ext3_sb_info *sbi;
1920 + struct super_block *sb;
1921 + struct ext3_buddy e3b;
1924 + sbi = EXT3_SB(sb);
1925 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1927 + /* first, try the goal */
1928 + err = ext3_mb_find_by_goal(ac, &e3b);
1929 + if (err || ac->ac_status == AC_STATUS_FOUND)
1932 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
1935 + i = ffs(ac->ac_g_ex.fe_len);
1936 + ac->ac_2order = 0;
1937 + if (i >= sbi->s_mb_order2_reqs) {
1939 + if ((ac->ac_g_ex.fe_len & (~(1 << i))) == 0)
1940 + ac->ac_2order = i;
1943 + group = ac->ac_g_ex.fe_group;
1945 + /* Let's just scan groups to find more-less suitable blocks */
1946 + cr = ac->ac_2order ? 0 : 1;
1948 + for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1949 + ac->ac_criteria = cr;
1950 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; group++, i++) {
1951 + struct ext3_group_info *grp;
1953 + if (group == EXT3_SB(sb)->s_groups_count)
1956 + /* quick check to skip empty groups */
1957 + grp = EXT3_GROUP_INFO(ac->ac_sb, group);
1958 + if (grp->bb_free == 0)
1961 + if (EXT3_MB_GRP_NEED_INIT(EXT3_GROUP_INFO(sb, group))) {
1962 + /* we need full data about the group
1963 + * to make a good selection */
1964 + err = ext3_mb_load_buddy(sb, group, &e3b);
1967 + ext3_mb_release_desc(&e3b);
1970 + /* check is group good for our criteries */
1971 + if (!ext3_mb_good_group(ac, group, cr))
1974 + err = ext3_mb_load_buddy(sb, group, &e3b);
1978 + ext3_lock_group(sb, group);
1979 + if (!ext3_mb_good_group(ac, group, cr)) {
1980 + /* someone did allocation from this group */
1981 + ext3_unlock_group(sb, group);
1982 + ext3_mb_release_desc(&e3b);
1986 + ac->ac_groups_scanned++;
1988 + ext3_mb_simple_scan_group(ac, &e3b);
1989 + else if (cr == 1 && ac->ac_g_ex.fe_len == sbi->s_stripe)
1990 + ext3_mb_scan_aligned(ac, &e3b);
1992 + ext3_mb_complex_scan_group(ac, &e3b);
1994 + ext3_unlock_group(sb, group);
1995 + ext3_mb_release_desc(&e3b);
1997 + if (ac->ac_status != AC_STATUS_CONTINUE)
2002 + if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2003 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
2005 + * We've been searching too long. Let's try to allocate
2006 + * the best chunk we've found so far
2009 + ext3_mb_try_best_found(ac, &e3b);
2010 + if (ac->ac_status != AC_STATUS_FOUND) {
2012 + * Someone more lucky has already allocated it.
2013 + * The only thing we can do is just take first
2015 + printk(KERN_DEBUG "EXT3-fs: someone won our chunk\n");
2017 + ac->ac_b_ex.fe_group = 0;
2018 + ac->ac_b_ex.fe_start = 0;
2019 + ac->ac_b_ex.fe_len = 0;
2020 + ac->ac_status = AC_STATUS_CONTINUE;
2021 + ac->ac_flags |= EXT3_MB_HINT_FIRST;
2023 + atomic_inc(&sbi->s_mb_lost_chunks);
2031 +#ifdef EXT3_MB_HISTORY
2032 +struct ext3_mb_proc_session {
2033 + struct ext3_mb_history *history;
2034 + struct super_block *sb;
2039 +static void *ext3_mb_history_skip_empty(struct ext3_mb_proc_session *s,
2040 + struct ext3_mb_history *hs,
2043 + if (hs == s->history + s->max)
2045 + if (!first && hs == s->history + s->start)
2047 + while (hs->orig.fe_len == 0) {
2049 + if (hs == s->history + s->max)
2051 + if (hs == s->history + s->start)
2057 +static void *ext3_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2059 + struct ext3_mb_proc_session *s = seq->private;
2060 + struct ext3_mb_history *hs;
2064 + return SEQ_START_TOKEN;
2065 + hs = ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2068 + while (--l && (hs = ext3_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2072 +static void *ext3_mb_seq_history_next(struct seq_file *seq, void *v, loff_t *pos)
2074 + struct ext3_mb_proc_session *s = seq->private;
2075 + struct ext3_mb_history *hs = v;
2078 + if (v == SEQ_START_TOKEN)
2079 + return ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2081 + return ext3_mb_history_skip_empty(s, ++hs, 0);
2084 +static int ext3_mb_seq_history_show(struct seq_file *seq, void *v)
2086 + char buf[25], buf2[25], buf3[25], *fmt;
2087 + struct ext3_mb_history *hs = v;
2089 + if (v == SEQ_START_TOKEN) {
2090 + seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2091 + "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2092 + "pid", "inode", "original", "goal", "result","found",
2093 + "grps", "cr", "flags", "merge", "tail", "broken");
2097 + if (hs->op == EXT3_MB_HISTORY_ALLOC) {
2098 + fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2099 + "%-5u %-5s %-5u %-6u\n";
2100 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2101 + hs->result.fe_start, hs->result.fe_len,
2102 + hs->result.fe_logical);
2103 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2104 + hs->orig.fe_start, hs->orig.fe_len,
2105 + hs->orig.fe_logical);
2106 + sprintf(buf3, "%lu/%lu/%lu@%lu", hs->goal.fe_group,
2107 + hs->goal.fe_start, hs->goal.fe_len,
2108 + hs->goal.fe_logical);
2109 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2110 + hs->found, hs->groups, hs->cr, hs->flags,
2111 + hs->merged ? "M" : "", hs->tail,
2112 + hs->buddy ? 1 << hs->buddy : 0);
2113 + } else if (hs->op == EXT3_MB_HISTORY_PREALLOC) {
2114 + fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2115 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2116 + hs->result.fe_start, hs->result.fe_len,
2117 + hs->result.fe_logical);
2118 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2119 + hs->orig.fe_start, hs->orig.fe_len,
2120 + hs->orig.fe_logical);
2121 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2122 + } else if (hs->op == EXT3_MB_HISTORY_DISCARD) {
2123 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2124 + hs->result.fe_start, hs->result.fe_len);
2125 + seq_printf(seq, "%-5u %-8u %-23s discard\n",
2126 + hs->pid, hs->ino, buf2);
2127 + } else if (hs->op == EXT3_MB_HISTORY_FREE) {
2128 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2129 + hs->result.fe_start, hs->result.fe_len);
2130 + seq_printf(seq, "%-5u %-8u %-23s free\n",
2131 + hs->pid, hs->ino, buf2);
2136 +static void ext3_mb_seq_history_stop(struct seq_file *seq, void *v)
2140 +static struct seq_operations ext3_mb_seq_history_ops = {
2141 + .start = ext3_mb_seq_history_start,
2142 + .next = ext3_mb_seq_history_next,
2143 + .stop = ext3_mb_seq_history_stop,
2144 + .show = ext3_mb_seq_history_show,
2147 +static int ext3_mb_seq_history_open(struct inode *inode, struct file *file)
2149 + struct super_block *sb = PDE(inode)->data;
2150 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2151 + struct ext3_mb_proc_session *s;
2154 + s = kmalloc(sizeof(*s), GFP_KERNEL);
2158 + size = sizeof(struct ext3_mb_history) * sbi->s_mb_history_max;
2159 + s->history = kmalloc(size, GFP_KERNEL);
2160 + if (s->history == NULL) {
2165 + spin_lock(&sbi->s_mb_history_lock);
2166 + memcpy(s->history, sbi->s_mb_history, size);
2167 + s->max = sbi->s_mb_history_max;
2168 + s->start = sbi->s_mb_history_cur % s->max;
2169 + spin_unlock(&sbi->s_mb_history_lock);
2171 + rc = seq_open(file, &ext3_mb_seq_history_ops);
2173 + struct seq_file *m = (struct seq_file *)file->private_data;
2176 + kfree(s->history);
2183 +static int ext3_mb_seq_history_release(struct inode *inode, struct file *file)
2185 + struct seq_file *seq = (struct seq_file *)file->private_data;
2186 + struct ext3_mb_proc_session *s = seq->private;
2187 + kfree(s->history);
2189 + return seq_release(inode, file);
2192 +static ssize_t ext3_mb_seq_history_write(struct file *file,
2193 + const char __user *buffer,
2194 + size_t count, loff_t *ppos)
2196 + struct seq_file *seq = (struct seq_file *)file->private_data;
2197 + struct ext3_mb_proc_session *s = seq->private;
2198 + struct super_block *sb = s->sb;
2202 + if (count >= sizeof(str)) {
2203 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2204 + "mb_history", (int)sizeof(str));
2205 + return -EOVERFLOW;
2208 + if (copy_from_user(str, buffer, count))
2211 + value = simple_strtol(str, NULL, 0);
2214 + EXT3_SB(sb)->s_mb_history_filter = value;
2219 +static struct file_operations ext3_mb_seq_history_fops = {
2220 + .owner = THIS_MODULE,
2221 + .open = ext3_mb_seq_history_open,
2223 + .write = ext3_mb_seq_history_write,
2224 + .llseek = seq_lseek,
2225 + .release = ext3_mb_seq_history_release,
2228 +static void *ext3_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2230 + struct super_block *sb = seq->private;
2231 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2234 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2238 + return (void *) group;
2241 +static void *ext3_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2243 + struct super_block *sb = seq->private;
2244 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2248 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2251 + return (void *) group;;
2254 +static int ext3_mb_seq_groups_show(struct seq_file *seq, void *v)
2256 + struct super_block *sb = seq->private;
2257 + long group = (long) v;
2259 + struct ext3_buddy e3b;
2261 + struct ext3_group_info info;
2262 + unsigned short counters[16];
2267 + seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2268 + "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2269 + "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2270 + "group", "free", "frags", "first",
2271 + "2^0", "2^1", "2^2", "2^3", "2^4", "2^5","2^6",
2272 + "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2274 + i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2275 + sizeof(struct ext3_group_info);
2276 + err = ext3_mb_load_buddy(sb, group, &e3b);
2278 + seq_printf(seq, "#%-5lu: I/O error\n", group);
2281 + ext3_lock_group(sb, group);
2282 + memcpy(&sg, EXT3_GROUP_INFO(sb, group), i);
2283 + ext3_unlock_group(sb, group);
2284 + ext3_mb_release_desc(&e3b);
2286 + seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2287 + sg.info.bb_fragments, sg.info.bb_first_free);
2288 + for (i = 0; i <= 13; i++)
2289 + seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2290 + sg.info.bb_counters[i] : 0);
2291 + seq_printf(seq, " ]\n");
2296 +static void ext3_mb_seq_groups_stop(struct seq_file *seq, void *v)
2300 +static struct seq_operations ext3_mb_seq_groups_ops = {
2301 + .start = ext3_mb_seq_groups_start,
2302 + .next = ext3_mb_seq_groups_next,
2303 + .stop = ext3_mb_seq_groups_stop,
2304 + .show = ext3_mb_seq_groups_show,
2307 +static int ext3_mb_seq_groups_open(struct inode *inode, struct file *file)
2309 + struct super_block *sb = PDE(inode)->data;
2312 + rc = seq_open(file, &ext3_mb_seq_groups_ops);
2314 + struct seq_file *m = (struct seq_file *)file->private_data;
2321 +static struct file_operations ext3_mb_seq_groups_fops = {
2322 + .owner = THIS_MODULE,
2323 + .open = ext3_mb_seq_groups_open,
2325 + .llseek = seq_lseek,
2326 + .release = seq_release,
2329 +static void ext3_mb_history_release(struct super_block *sb)
2331 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2333 + remove_proc_entry("mb_groups", sbi->s_mb_proc);
2334 + remove_proc_entry("mb_history", sbi->s_mb_proc);
2336 + if (sbi->s_mb_history)
2337 + kfree(sbi->s_mb_history);
2340 +static void ext3_mb_history_init(struct super_block *sb)
2342 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2345 + if (sbi->s_mb_proc != NULL) {
2346 + struct proc_dir_entry *p;
2347 + p = create_proc_entry("mb_history", S_IRUGO, sbi->s_mb_proc);
2349 + p->proc_fops = &ext3_mb_seq_history_fops;
2352 + p = create_proc_entry("mb_groups", S_IRUGO, sbi->s_mb_proc);
2354 + p->proc_fops = &ext3_mb_seq_groups_fops;
2359 + sbi->s_mb_history_max = 1000;
2360 + sbi->s_mb_history_cur = 0;
2361 + spin_lock_init(&sbi->s_mb_history_lock);
2362 + i = sbi->s_mb_history_max * sizeof(struct ext3_mb_history);
2363 + sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
2364 + if (likely(sbi->s_mb_history != NULL))
2365 + memset(sbi->s_mb_history, 0, i);
2366 + /* if we can't allocate history, then we simple won't use it */
2370 +ext3_mb_store_history(struct ext3_allocation_context *ac)
2372 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
2373 + struct ext3_mb_history h;
2375 + if (unlikely(sbi->s_mb_history == NULL))
2378 + if (!(ac->ac_op & sbi->s_mb_history_filter))
2382 + h.pid = current->pid;
2383 + h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2384 + h.orig = ac->ac_o_ex;
2385 + h.result = ac->ac_b_ex;
2386 + h.flags = ac->ac_flags;
2388 + if (ac->ac_op == EXT3_MB_HISTORY_ALLOC) {
2389 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2390 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2392 + h.goal = ac->ac_g_ex;
2393 + h.result = ac->ac_f_ex;
2396 + spin_lock(&sbi->s_mb_history_lock);
2397 + memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2398 + if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2399 + sbi->s_mb_history_cur = 0;
2400 + spin_unlock(&sbi->s_mb_history_lock);
2404 +#define ext3_mb_history_release(sb)
2405 +#define ext3_mb_history_init(sb)
2408 +int ext3_mb_init_backend(struct super_block *sb)
2410 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2411 + int i, j, len, metalen;
2412 + int num_meta_group_infos =
2413 + (sbi->s_groups_count + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2414 + EXT3_DESC_PER_BLOCK_BITS(sb);
2415 + struct ext3_group_info **meta_group_info;
2417 + /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2418 + * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2419 + * So a two level scheme suffices for now. */
2420 + sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
2421 + num_meta_group_infos, GFP_KERNEL);
2422 + if (sbi->s_group_info == NULL) {
2423 + printk(KERN_ERR "EXT3-fs: can't allocate buddy meta group\n");
2426 + sbi->s_buddy_cache = new_inode(sb);
2427 + if (sbi->s_buddy_cache == NULL) {
2428 + printk(KERN_ERR "EXT3-fs: can't get new inode\n");
2431 + EXT3_I(sbi->s_buddy_cache)->i_disksize = 0;
2433 + metalen = sizeof(*meta_group_info) << EXT3_DESC_PER_BLOCK_BITS(sb);
2434 + for (i = 0; i < num_meta_group_infos; i++) {
2435 + if ((i + 1) == num_meta_group_infos)
2436 + metalen = sizeof(*meta_group_info) *
2437 + (sbi->s_groups_count -
2438 + (i << EXT3_DESC_PER_BLOCK_BITS(sb)));
2439 + meta_group_info = kmalloc(metalen, GFP_KERNEL);
2440 + if (meta_group_info == NULL) {
2441 + printk(KERN_ERR "EXT3-fs: can't allocate mem for a "
2443 + goto err_freemeta;
2445 + sbi->s_group_info[i] = meta_group_info;
2449 + * calculate needed size. if change bb_counters size,
2450 + * don't forget about ext3_mb_generate_buddy()
2452 + len = sizeof(struct ext3_group_info);
2453 + len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
2454 + for (i = 0; i < sbi->s_groups_count; i++) {
2455 + struct ext3_group_desc * desc;
2458 + sbi->s_group_info[i >> EXT3_DESC_PER_BLOCK_BITS(sb)];
2459 + j = i & (EXT3_DESC_PER_BLOCK(sb) - 1);
2461 + meta_group_info[j] = kmalloc(len, GFP_KERNEL);
2462 + if (meta_group_info[j] == NULL) {
2463 + printk(KERN_ERR "EXT3-fs: can't allocate buddy mem\n");
2465 + goto err_freebuddy;
2467 + desc = ext3_get_group_desc(sb, i, NULL);
2468 + if (desc == NULL) {
2469 + printk(KERN_ERR"EXT3-fs: can't read descriptor %u\n",i);
2470 + goto err_freebuddy;
2472 + memset(meta_group_info[j], 0, len);
2473 + set_bit(EXT3_GROUP_INFO_NEED_INIT_BIT,
2474 + &meta_group_info[j]->bb_state);
2476 + /* initialize bb_free to be able to skip
2477 + * empty groups without initialization */
2478 + meta_group_info[j]->bb_free =
2479 + le16_to_cpu(desc->bg_free_blocks_count);
2481 + INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
2483 +#ifdef DOUBLE_CHECK
2485 + struct buffer_head *bh;
2486 + meta_group_info[j]->bb_bitmap =
2487 + kmalloc(sb->s_blocksize, GFP_KERNEL);
2488 + BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
2489 + bh = read_block_bitmap(sb, i);
2490 + BUG_ON(bh == NULL);
2491 + memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
2503 + kfree(EXT3_GROUP_INFO(sb, i));
2506 + i = num_meta_group_infos;
2509 + kfree(sbi->s_group_info[i]);
2510 + iput(sbi->s_buddy_cache);
2512 + kfree(sbi->s_group_info);
2516 +int ext3_mb_init(struct super_block *sb, int needs_recovery)
2518 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2519 + unsigned i, offset, max;
2521 + if (!test_opt(sb, MBALLOC))
2524 + i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2526 + sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2527 + if (sbi->s_mb_offsets == NULL) {
2528 + clear_opt(sbi->s_mount_opt, MBALLOC);
2531 + sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2532 + if (sbi->s_mb_maxs == NULL) {
2533 + clear_opt(sbi->s_mount_opt, MBALLOC);
2534 + kfree(sbi->s_mb_maxs);
2538 + /* order 0 is regular bitmap */
2539 + sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2540 + sbi->s_mb_offsets[0] = 0;
2544 + max = sb->s_blocksize << 2;
2546 + sbi->s_mb_offsets[i] = offset;
2547 + sbi->s_mb_maxs[i] = max;
2548 + offset += 1 << (sb->s_blocksize_bits - i);
2551 + } while (i <= sb->s_blocksize_bits + 1);
2553 + /* init file for buddy data */
2554 + if ((i = ext3_mb_init_backend(sb))) {
2555 + clear_opt(sbi->s_mount_opt, MBALLOC);
2556 + kfree(sbi->s_mb_offsets);
2557 + kfree(sbi->s_mb_maxs);
2561 + spin_lock_init(&sbi->s_md_lock);
2562 + INIT_LIST_HEAD(&sbi->s_active_transaction);
2563 + INIT_LIST_HEAD(&sbi->s_closed_transaction);
2564 + INIT_LIST_HEAD(&sbi->s_committed_transaction);
2565 + spin_lock_init(&sbi->s_bal_lock);
2567 + sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2568 + sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2569 + sbi->s_mb_max_groups_to_scan = MB_DEFAULT_MAX_GROUPS_TO_SCAN;
2570 + sbi->s_mb_stats = MB_DEFAULT_STATS;
2571 + sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2572 + sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2573 + sbi->s_mb_history_filter = EXT3_MB_HISTORY_DEFAULT;
2575 + i = sizeof(struct ext3_locality_group) * NR_CPUS;
2576 + sbi->s_locality_groups = kmalloc(i, GFP_NOFS);
2577 + if (sbi->s_locality_groups == NULL) {
2578 + clear_opt(sbi->s_mount_opt, MBALLOC);
2579 + kfree(sbi->s_mb_offsets);
2580 + kfree(sbi->s_mb_maxs);
2583 + for (i = 0; i < NR_CPUS; i++) {
2584 + struct ext3_locality_group *lg;
2585 + lg = &sbi->s_locality_groups[i];
2586 + sema_init(&lg->lg_sem, 1);
2587 + INIT_LIST_HEAD(&lg->lg_prealloc_list);
2588 + spin_lock_init(&lg->lg_prealloc_lock);
2591 + ext3_mb_init_per_dev_proc(sb);
2592 + ext3_mb_history_init(sb);
2594 + printk("EXT3-fs: mballoc enabled\n");
2598 +void ext3_mb_cleanup_pa(struct ext3_group_info *grp)
2600 + struct ext3_prealloc_space *pa;
2601 + struct list_head *cur, *tmp;
2604 + list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2605 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
2606 + list_del_rcu(&pa->pa_group_list);
2611 + mb_debug("mballoc: %u PAs left\n", count);
2615 +int ext3_mb_release(struct super_block *sb)
2617 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2618 + int i, num_meta_group_infos;
2620 + if (!test_opt(sb, MBALLOC))
2623 + /* release freed, non-committed blocks */
2624 + spin_lock(&sbi->s_md_lock);
2625 + list_splice_init(&sbi->s_closed_transaction,
2626 + &sbi->s_committed_transaction);
2627 + list_splice_init(&sbi->s_active_transaction,
2628 + &sbi->s_committed_transaction);
2629 + spin_unlock(&sbi->s_md_lock);
2630 + ext3_mb_free_committed_blocks(sb);
2632 + if (sbi->s_group_info) {
2633 + for (i = 0; i < sbi->s_groups_count; i++) {
2634 +#ifdef DOUBLE_CHECK
2635 + if (EXT3_GROUP_INFO(sb, i)->bb_bitmap)
2636 + kfree(EXT3_GROUP_INFO(sb, i)->bb_bitmap);
2638 + ext3_mb_cleanup_pa(EXT3_GROUP_INFO(sb, i));
2639 + kfree(EXT3_GROUP_INFO(sb, i));
2641 + num_meta_group_infos = (sbi->s_groups_count +
2642 + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2643 + EXT3_DESC_PER_BLOCK_BITS(sb);
2644 + for (i = 0; i < num_meta_group_infos; i++)
2645 + kfree(sbi->s_group_info[i]);
2646 + kfree(sbi->s_group_info);
2648 + if (sbi->s_mb_offsets)
2649 + kfree(sbi->s_mb_offsets);
2650 + if (sbi->s_mb_maxs)
2651 + kfree(sbi->s_mb_maxs);
2652 + if (sbi->s_buddy_cache)
2653 + iput(sbi->s_buddy_cache);
2654 + if (sbi->s_mb_stats) {
2655 + printk("EXT3-fs: mballoc: %u blocks %u reqs (%u success)\n",
2656 + atomic_read(&sbi->s_bal_allocated),
2657 + atomic_read(&sbi->s_bal_reqs),
2658 + atomic_read(&sbi->s_bal_success));
2659 + printk("EXT3-fs: mballoc: %u extents scanned, %u goal hits, "
2660 + "%u 2^N hits, %u breaks, %u lost\n",
2661 + atomic_read(&sbi->s_bal_ex_scanned),
2662 + atomic_read(&sbi->s_bal_goals),
2663 + atomic_read(&sbi->s_bal_2orders),
2664 + atomic_read(&sbi->s_bal_breaks),
2665 + atomic_read(&sbi->s_mb_lost_chunks));
2666 + printk("EXT3-fs: mballoc: %lu generated and it took %Lu\n",
2667 + sbi->s_mb_buddies_generated++,
2668 + sbi->s_mb_generation_time);
2669 + printk("EXT3-fs: mballoc: %u preallocated, %u discarded\n",
2670 + atomic_read(&sbi->s_mb_preallocated),
2671 + atomic_read(&sbi->s_mb_discarded));
2674 + if (sbi->s_locality_groups)
2675 + kfree(sbi->s_locality_groups);
2677 + ext3_mb_history_release(sb);
2678 + ext3_mb_destroy_per_dev_proc(sb);
2683 +void ext3_mb_free_committed_blocks(struct super_block *sb)
2685 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2686 + int err, i, count = 0, count2 = 0;
2687 + struct ext3_free_metadata *md;
2688 + struct ext3_buddy e3b;
2690 + if (list_empty(&sbi->s_committed_transaction))
2693 + /* there is committed blocks to be freed yet */
2695 + /* get next array of blocks */
2697 + spin_lock(&sbi->s_md_lock);
2698 + if (!list_empty(&sbi->s_committed_transaction)) {
2699 + md = list_entry(sbi->s_committed_transaction.next,
2700 + struct ext3_free_metadata, list);
2701 + list_del(&md->list);
2703 + spin_unlock(&sbi->s_md_lock);
2708 + mb_debug("gonna free %u blocks in group %u (0x%p):",
2709 + md->num, md->group, md);
2711 + err = ext3_mb_load_buddy(sb, md->group, &e3b);
2712 + /* we expect to find existing buddy because it's pinned */
2715 + /* there are blocks to put in buddy to make them really free */
2718 + ext3_lock_group(sb, md->group);
2719 + for (i = 0; i < md->num; i++) {
2720 + mb_debug(" %u", md->blocks[i]);
2721 + err = mb_free_blocks(&e3b, md->blocks[i], 1);
2725 + ext3_unlock_group(sb, md->group);
2727 + /* balance refcounts from ext3_mb_free_metadata() */
2728 + page_cache_release(e3b.bd_buddy_page);
2729 + page_cache_release(e3b.bd_bitmap_page);
2732 + ext3_mb_release_desc(&e3b);
2736 + mb_debug("freed %u blocks in %u structures\n", count, count2);
2739 +#define EXT3_ROOT "ext3"
2740 +#define EXT3_MB_STATS_NAME "stats"
2741 +#define EXT3_MB_MAX_TO_SCAN_NAME "max_to_scan"
2742 +#define EXT3_MB_MIN_TO_SCAN_NAME "min_to_scan"
2743 +#define EXT3_MB_ORDER2_REQ "order2_req"
2744 +#define EXT3_MB_STREAM_REQ "stream_req"
2746 +static int ext3_mb_stats_read(char *page, char **start, off_t off,
2747 + int count, int *eof, void *data)
2749 + struct ext3_sb_info *sbi = data;
2756 + len = sprintf(page, "%ld\n", sbi->s_mb_stats);
2761 +static int ext3_mb_stats_write(struct file *file, const char *buffer,
2762 + unsigned long count, void *data)
2764 + struct ext3_sb_info *sbi = data;
2767 + if (count >= sizeof(str)) {
2768 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2769 + EXT3_MB_STATS_NAME, (int)sizeof(str));
2770 + return -EOVERFLOW;
2773 + if (copy_from_user(str, buffer, count))
2776 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2777 + sbi->s_mb_stats = (simple_strtol(str, NULL, 0) != 0);
2781 +static int ext3_mb_max_to_scan_read(char *page, char **start, off_t off,
2782 + int count, int *eof, void *data)
2784 + struct ext3_sb_info *sbi = data;
2791 + len = sprintf(page, "%ld\n", sbi->s_mb_max_to_scan);
2796 +static int ext3_mb_max_to_scan_write(struct file *file, const char *buffer,
2797 + unsigned long count, void *data)
2799 + struct ext3_sb_info *sbi = data;
2803 + if (count >= sizeof(str)) {
2804 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2805 + EXT3_MB_MAX_TO_SCAN_NAME, (int)sizeof(str));
2806 + return -EOVERFLOW;
2809 + if (copy_from_user(str, buffer, count))
2812 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2813 + value = simple_strtol(str, NULL, 0);
2817 + sbi->s_mb_max_to_scan = value;
2822 +static int ext3_mb_min_to_scan_read(char *page, char **start, off_t off,
2823 + int count, int *eof, void *data)
2825 + struct ext3_sb_info *sbi = data;
2832 + len = sprintf(page, "%ld\n", sbi->s_mb_min_to_scan);
2837 +static int ext3_mb_order2_req_write(struct file *file, const char *buffer,
2838 + unsigned long count, void *data)
2840 + struct ext3_sb_info *sbi = data;
2844 + if (count >= sizeof(str)) {
2845 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2846 + EXT3_MB_MIN_TO_SCAN_NAME, (int)sizeof(str));
2847 + return -EOVERFLOW;
2850 + if (copy_from_user(str, buffer, count))
2853 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2854 + value = simple_strtol(str, NULL, 0);
2858 + sbi->s_mb_order2_reqs = value;
2863 +static int ext3_mb_order2_req_read(char *page, char **start, off_t off,
2864 + int count, int *eof, void *data)
2866 + struct ext3_sb_info *sbi = data;
2873 + len = sprintf(page, "%ld\n", sbi->s_mb_order2_reqs);
2878 +static int ext3_mb_min_to_scan_write(struct file *file, const char *buffer,
2879 + unsigned long count, void *data)
2881 + struct ext3_sb_info *sbi = data;
2885 + if (count >= sizeof(str)) {
2886 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2887 + EXT3_MB_MIN_TO_SCAN_NAME, (int)sizeof(str));
2888 + return -EOVERFLOW;
2891 + if (copy_from_user(str, buffer, count))
2894 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2895 + value = simple_strtol(str, NULL, 0);
2899 + sbi->s_mb_min_to_scan = value;
2904 +static int ext3_mb_stream_req_read(char *page, char **start, off_t off,
2905 + int count, int *eof, void *data)
2907 + struct ext3_sb_info *sbi = data;
2914 + len = sprintf(page, "%ld\n", sbi->s_mb_stream_request);
2919 +static int ext3_mb_stream_req_write(struct file *file, const char *buffer,
2920 + unsigned long count, void *data)
2922 + struct ext3_sb_info *sbi = data;
2926 + if (count >= sizeof(str)) {
2927 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2928 + EXT3_MB_STREAM_REQ, (int)sizeof(str));
2929 + return -EOVERFLOW;
2932 + if (copy_from_user(str, buffer, count))
2935 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2936 + value = simple_strtol(str, NULL, 0);
2940 + sbi->s_mb_stream_request = value;
2945 +int ext3_mb_init_per_dev_proc(struct super_block *sb)
2947 + mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2948 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2949 + struct proc_dir_entry *proc;
2950 + char devname[64], *name;
2952 + snprintf(devname, sizeof(devname) - 1, "%s",
2953 + bdevname(sb->s_bdev, devname));
2954 + sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext3);
2956 + name = EXT3_MB_STATS_NAME;
2957 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2961 + proc->read_proc = ext3_mb_stats_read;
2962 + proc->write_proc = ext3_mb_stats_write;
2964 + name = EXT3_MB_MAX_TO_SCAN_NAME;
2965 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2969 + proc->read_proc = ext3_mb_max_to_scan_read;
2970 + proc->write_proc = ext3_mb_max_to_scan_write;
2972 + name = EXT3_MB_MIN_TO_SCAN_NAME;
2973 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2977 + proc->read_proc = ext3_mb_min_to_scan_read;
2978 + proc->write_proc = ext3_mb_min_to_scan_write;
2980 + name = EXT3_MB_ORDER2_REQ;
2981 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2985 + proc->read_proc = ext3_mb_order2_req_read;
2986 + proc->write_proc = ext3_mb_order2_req_write;
2988 + name = EXT3_MB_STREAM_REQ;
2989 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2993 + proc->read_proc = ext3_mb_stream_req_read;
2994 + proc->write_proc = ext3_mb_stream_req_write;
2999 + printk(KERN_ERR "EXT3-fs: Unable to create %s\n", name);
3000 + remove_proc_entry(EXT3_MB_STREAM_REQ, sbi->s_mb_proc);
3001 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_mb_proc);
3002 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
3003 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
3004 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_mb_proc);
3005 + remove_proc_entry(devname, proc_root_ext3);
3006 + sbi->s_mb_proc = NULL;
3011 +int ext3_mb_destroy_per_dev_proc(struct super_block *sb)
3013 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3016 + if (sbi->s_mb_proc == NULL)
3019 + snprintf(devname, sizeof(devname) - 1, "%s",
3020 + bdevname(sb->s_bdev, devname));
3021 + remove_proc_entry(EXT3_MB_STREAM_REQ, sbi->s_mb_proc);
3022 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_mb_proc);
3023 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
3024 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
3025 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_mb_proc);
3026 + remove_proc_entry(devname, proc_root_ext3);
3031 +int __init init_ext3_proc(void)
3033 + ext3_pspace_cachep =
3034 + kmem_cache_create("ext3_prealloc_space",
3035 + sizeof(struct ext3_prealloc_space),
3036 + 0, SLAB_RECLAIM_ACCOUNT, NULL, NULL);
3037 + if (ext3_pspace_cachep == NULL)
3040 + proc_root_ext3 = proc_mkdir(EXT3_ROOT, proc_root_fs);
3041 + if (proc_root_ext3 == NULL)
3042 + printk(KERN_ERR "EXT3-fs: Unable to create %s\n", EXT3_ROOT);
3047 +void exit_ext3_proc(void)
3049 + /* XXX: synchronize_rcu(); */
3050 + kmem_cache_destroy(ext3_pspace_cachep);
3051 + remove_proc_entry(EXT3_ROOT, proc_root_fs);
3056 + * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
3057 + * Returns 0 if success or error code
3059 +int ext3_mb_mark_diskspace_used(struct ext3_allocation_context *ac, handle_t *handle)
3061 + struct buffer_head *bitmap_bh = NULL;
3062 + struct ext3_super_block *es;
3063 + struct ext3_group_desc *gdp;
3064 + struct buffer_head *gdp_bh;
3065 + struct ext3_sb_info *sbi;
3066 + struct super_block *sb;
3070 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3071 + BUG_ON(ac->ac_b_ex.fe_len <= 0);
3074 + sbi = EXT3_SB(sb);
3077 + ext3_debug("using block group %d(%d)\n", ac->ac_b_group.group,
3078 + gdp->bg_free_blocks_count);
3080 + /* time to check quota, we can't do this before because
3081 + * having quota spent on preallocated-unused-yet blocks
3082 + * would be wrong */
3083 + len = ac->ac_b_ex.fe_len;
3084 + while (len && DQUOT_ALLOC_BLOCK(ac->ac_inode, len)) len--;
3085 + if (ac->ac_b_ex.fe_len != len) {
3086 + /* some blocks can't be allocated due to quota
3087 + * we have to return them back */
3095 + bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3099 + err = ext3_journal_get_write_access(handle, bitmap_bh);
3104 + gdp = ext3_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3108 + err = ext3_journal_get_write_access(handle, gdp_bh);
3112 + block = ac->ac_b_ex.fe_group * EXT3_BLOCKS_PER_GROUP(sb)
3113 + + ac->ac_b_ex.fe_start
3114 + + le32_to_cpu(es->s_first_data_block);
3116 + if (block == le32_to_cpu(gdp->bg_block_bitmap) ||
3117 + block == le32_to_cpu(gdp->bg_inode_bitmap) ||
3118 + in_range(block, le32_to_cpu(gdp->bg_inode_table),
3119 + EXT3_SB(sb)->s_itb_per_group))
3120 + ext3_error(sb, __FUNCTION__,
3121 + "Allocating block in system zone - block = %lu",
3122 + (unsigned long) block);
3123 +#ifdef AGGRESSIVE_CHECK
3126 + for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3127 + BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3128 + bitmap_bh->b_data));
3132 + mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
3134 + spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3135 + gdp->bg_free_blocks_count =
3136 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
3137 + - ac->ac_b_ex.fe_len);
3138 + spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3139 + percpu_counter_mod(&sbi->s_freeblocks_counter, - ac->ac_b_ex.fe_len);
3141 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
3144 + err = ext3_journal_dirty_metadata(handle, gdp_bh);
3148 + brelse(bitmap_bh);
3153 + * here we normalize request for locality group
3154 + * XXX: should we try to preallocate more than the group has now?
3156 +void ext3_mb_normalize_group_request(struct ext3_allocation_context *ac)
3158 + struct super_block *sb = ac->ac_sb;
3159 + struct ext3_locality_group *lg = ac->ac_lg;
3161 + BUG_ON(lg == NULL);
3162 + if (EXT3_SB(sb)->s_stripe)
3163 + ac->ac_g_ex.fe_len = EXT3_SB(sb)->s_stripe;
3165 + ac->ac_g_ex.fe_len = (1024 * 1024) >> sb->s_blocksize_bits;
3167 + mb_debug("#%u: goal %u blocks for locality group\n",
3168 + current->pid, ac->ac_g_ex.fe_len);
3172 + * Normalization means making request better in terms of
3173 + * size and alignment
3175 +void ext3_mb_normalize_request(struct ext3_allocation_context *ac,
3176 + struct ext3_allocation_request *ar)
3178 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3179 + loff_t start, end, size, orig_size, orig_start;
3180 + struct list_head *cur;
3183 + /* do normalize only data requests, metadata requests
3184 + do not need preallocation */
3185 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3188 + /* sometime caller may want exact blocks */
3189 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
3192 + /* caller may indicate that preallocation isn't
3193 + * required (it's a tail, for example) */
3194 + if (ac->ac_flags & EXT3_MB_HINT_NOPREALLOC)
3197 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3198 + return ext3_mb_normalize_group_request(ac);
3200 + bsbits = ac->ac_sb->s_blocksize_bits;
3202 + /* first, let's learn actual file size
3203 + * given current request is allocated */
3204 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3205 + size = size << bsbits;
3206 + if (size < i_size_read(ac->ac_inode))
3207 + size = i_size_read(ac->ac_inode);
3209 + /* first, try to predict filesize */
3210 + /* XXX: should this table be tunable? */
3212 + if (size <= 16 * 1024) {
3214 + } else if (size <= 32 * 1024) {
3216 + } else if (size <= 64 * 1024) {
3218 + } else if (size <= 128 * 1024) {
3219 + size = 128 * 1024;
3220 + } else if (size <= 256 * 1024) {
3221 + size = 256 * 1024;
3222 + } else if (size <= 512 * 1024) {
3223 + size = 512 * 1024;
3224 + } else if (size <= 1024 * 1024) {
3225 + size = 1024 * 1024;
3226 + } else if (size < 4 * 1024 * 1024) {
3227 + start = ac->ac_o_ex.fe_logical << bsbits;
3228 + start = (start / (1024 * 1024)) * (1024 * 1024);
3229 + size = 1024 * 1024;
3230 + } else if (size < 8 * 1024 * 1024) {
3231 + start = ac->ac_o_ex.fe_logical << bsbits;
3232 + start = (start / (4 * (1024 * 1024))) * 4 * (1024 * 1024);
3233 + size = 4 * 1024 * 1024;
3234 + } else if (ac->ac_o_ex.fe_len < ((8 << 20) >> bsbits)) {
3235 + start = ac->ac_o_ex.fe_logical;
3236 + start = start << bsbits;
3237 + start = (start / (8 * (1024 * 1024))) * 8 * (1024 * 1024);
3238 + size = 8 * 1024 * 1024;
3240 + start = ac->ac_o_ex.fe_logical;
3241 + start = start << bsbits;
3242 + size = ac->ac_o_ex.fe_len << bsbits;
3244 + orig_size = size = size >> bsbits;
3245 + orig_start = start = start >> bsbits;
3247 + /* don't cover already allocated blocks in selected range */
3248 + if (ar->pleft && start <= ar->lleft) {
3249 + size -= ar->lleft + 1 - start;
3250 + start = ar->lleft + 1;
3252 + if (ar->pright && start + size - 1 >= ar->lright)
3253 + size -= start + size - ar->lright;
3255 + end = start + size;
3257 + /* check we don't cross already preallocated blocks */
3259 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3260 + struct ext3_prealloc_space *pa;
3261 + unsigned long pa_end;
3263 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3264 + pa_end = pa->pa_lstart + pa->pa_len;
3266 + /* PA must not overlap original request */
3267 + BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3268 + ac->ac_o_ex.fe_logical < pa->pa_lstart));
3270 + /* skip PA normalized request doesn't overlap with */
3271 + if (pa->pa_lstart >= end)
3273 + if (pa_end <= start)
3275 + BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3277 + if (pa_end <= ac->ac_o_ex.fe_logical) {
3278 + BUG_ON(pa_end < start);
3282 + if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3283 + BUG_ON(pa->pa_lstart > end);
3284 + end = pa->pa_lstart;
3287 + rcu_read_unlock();
3288 + size = end - start;
3290 + /* XXX: extra loop to check we really don't overlap preallocations */
3292 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3293 + struct ext3_prealloc_space *pa;
3294 + unsigned long pa_end;
3295 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3296 + pa_end = pa->pa_lstart + pa->pa_len;
3297 + BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3299 + rcu_read_unlock();
3301 + if (start + size <= ac->ac_o_ex.fe_logical &&
3302 + start > ac->ac_o_ex.fe_logical) {
3303 + printk("start %lu, size %lu, fe_logical %lu\n",
3304 + (unsigned long) start, (unsigned long) size,
3305 + (unsigned long) ac->ac_o_ex.fe_logical);
3307 + BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3308 + start > ac->ac_o_ex.fe_logical);
3310 + /* now prepare goal request */
3311 + BUG_ON(size <= 0 || size >= EXT3_BLOCKS_PER_GROUP(ac->ac_sb));
3312 + if (size < ac->ac_o_ex.fe_len) {
3313 + /* XXX: don't normalize tails? */
3316 + /* XXX: is it better to align blocks WRT to logical placement
3317 + * or satisfy big request as is */
3318 + ac->ac_g_ex.fe_logical = start;
3319 + ac->ac_g_ex.fe_len = size;
3321 + mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3322 + (unsigned) orig_size, (unsigned) start);
3325 +void ext3_mb_collect_stats(struct ext3_allocation_context *ac)
3327 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
3329 + if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3330 + atomic_inc(&sbi->s_bal_reqs);
3331 + atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3332 + if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3333 + atomic_inc(&sbi->s_bal_success);
3334 + atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3335 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3336 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3337 + atomic_inc(&sbi->s_bal_goals);
3338 + if (ac->ac_found > sbi->s_mb_max_to_scan)
3339 + atomic_inc(&sbi->s_bal_breaks);
3342 + ext3_mb_store_history(ac);
3346 + * use blocks preallocated to inode
3348 +void ext3_mb_use_inode_pa(struct ext3_allocation_context *ac,
3349 + struct ext3_prealloc_space *pa)
3351 + unsigned long start, len;
3353 + /* found preallocated blocks, use them */
3354 + start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3355 + len = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3356 + len = len - start;
3357 + ext3_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3358 + &ac->ac_b_ex.fe_start);
3359 + ac->ac_b_ex.fe_len = len;
3360 + ac->ac_status = AC_STATUS_FOUND;
3363 + BUG_ON(start < pa->pa_pstart);
3364 + BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3365 + BUG_ON(pa->pa_free < len);
3366 + pa->pa_free -= len;
3368 + mb_debug("use %lu/%lu from inode pa %p\n", start, len, pa);
3372 + * use blocks preallocated to locality group
3374 +void ext3_mb_use_group_pa(struct ext3_allocation_context *ac,
3375 + struct ext3_prealloc_space *pa)
3377 + unsigned len = ac->ac_o_ex.fe_len;
3379 + ext3_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3380 + &ac->ac_b_ex.fe_group,
3381 + &ac->ac_b_ex.fe_start);
3382 + ac->ac_b_ex.fe_len = len;
3383 + ac->ac_status = AC_STATUS_FOUND;
3386 + /* we don't correct pa_pstart or pa_plen here to avoid
3387 + * possible race when tte group is being loaded concurrently
3388 + * instead we correct pa later, after blocks are marked
3389 + * in on-disk bitmap -- see ext3_mb_release_context() */
3390 + mb_debug("use %lu/%lu from group pa %p\n", pa->pa_lstart-len, len, pa);
3394 + * search goal blocks in preallocated space
3396 +int ext3_mb_use_preallocated(struct ext3_allocation_context *ac)
3398 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3399 + struct ext3_locality_group *lg;
3400 + struct ext3_prealloc_space *pa;
3401 + struct list_head *cur;
3403 + /* only data can be preallocated */
3404 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3407 + /* first, try per-file preallocation */
3409 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3410 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3412 + /* all fields in this condition don't change,
3413 + * so we can skip locking for them */
3414 + if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3415 + ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3418 + /* found preallocated blocks, use them */
3419 + spin_lock(&pa->pa_lock);
3420 + if (pa->pa_deleted == 0 && pa->pa_free) {
3421 + atomic_inc(&pa->pa_count);
3422 + ext3_mb_use_inode_pa(ac, pa);
3423 + spin_unlock(&pa->pa_lock);
3424 + ac->ac_criteria = 10;
3425 + rcu_read_unlock();
3428 + spin_unlock(&pa->pa_lock);
3430 + rcu_read_unlock();
3432 + /* can we use group allocation? */
3433 + if (!(ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC))
3436 + /* inode may have no locality group for some reason */
3442 + list_for_each_rcu(cur, &lg->lg_prealloc_list) {
3443 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3444 + spin_lock(&pa->pa_lock);
3445 + if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3446 + atomic_inc(&pa->pa_count);
3447 + ext3_mb_use_group_pa(ac, pa);
3448 + spin_unlock(&pa->pa_lock);
3449 + ac->ac_criteria = 20;
3450 + rcu_read_unlock();
3453 + spin_unlock(&pa->pa_lock);
3455 + rcu_read_unlock();
3461 + * the function goes through all preallocation in this group and marks them
3462 + * used in in-core bitmap. buddy must be generated from this bitmap
3464 +void ext3_mb_generate_from_pa(struct super_block *sb, void *bitmap, int group)
3466 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
3467 + struct ext3_prealloc_space *pa;
3468 + struct list_head *cur;
3469 + unsigned long groupnr;
3470 + unsigned long start;
3471 + int preallocated = 0, count = 0, len;
3473 + /* all form of preallocation discards first load group,
3474 + * so the only competing code is preallocation use.
3475 + * we don't need any locking here
3476 + * notice we do NOT ignore preallocations with pa_deleted
3477 + * otherwise we could leave used blocks available for
3478 + * allocation in buddy when concurrent ext3_mb_put_pa()
3479 + * is dropping preallocation
3481 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
3482 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
3483 + spin_lock(&pa->pa_lock);
3484 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &start);
3486 + spin_unlock(&pa->pa_lock);
3487 + BUG_ON(groupnr != group);
3488 + mb_set_bits(bitmap, start, len);
3489 + preallocated += len;
3492 + mb_debug("prellocated %u for group %u\n", preallocated, group);
3495 +#if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,5)
3496 +static void ext3_mb_pa_callback(struct rcu_head *head)
3498 + struct ext3_prealloc_space *pa;
3499 + pa = container_of(head, struct ext3_prealloc_space, u.pa_rcu);
3500 + kmem_cache_free(ext3_pspace_cachep, pa);
3502 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback)
3504 +static void ext3_mb_pa_callback(void *pa)
3506 + kmem_cache_free(ext3_pspace_cachep, pa);
3508 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback, pa)
3512 + * drops a reference to preallocated space descriptor
3513 + * if this was the last reference and the space is consumed
3515 +void ext3_mb_put_pa(struct ext3_allocation_context *ac,
3516 + struct super_block *sb, struct ext3_prealloc_space *pa)
3518 + unsigned long grp;
3520 + if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3523 + /* in this short window concurrent discard can set pa_deleted */
3524 + spin_lock(&pa->pa_lock);
3525 + if (pa->pa_deleted == 0) {
3526 + spin_unlock(&pa->pa_lock);
3530 + pa->pa_deleted = 1;
3531 + spin_unlock(&pa->pa_lock);
3533 + /* -1 is to protect from crossing allocation group */
3534 + ext3_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3539 + * P1 (buddy init) P2 (regular allocation)
3540 + * find block B in PA
3541 + * copy on-disk bitmap to buddy
3542 + * mark B in on-disk bitmap
3543 + * drop PA from group
3544 + * mark all PAs in buddy
3546 + * thus, P1 initializes buddy with B available. to prevent this
3547 + * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3548 + * against that pair
3550 + ext3_lock_group(sb, grp);
3551 + list_del_rcu(&pa->pa_group_list);
3552 + ext3_unlock_group(sb, grp);
3554 + spin_lock(pa->pa_obj_lock);
3555 + list_del_rcu(&pa->pa_inode_list);
3556 + spin_unlock(pa->pa_obj_lock);
3562 + * creates new preallocated space for given inode
3564 +int ext3_mb_new_inode_pa(struct ext3_allocation_context *ac)
3566 + struct super_block *sb = ac->ac_sb;
3567 + struct ext3_prealloc_space *pa;
3568 + struct ext3_group_info *grp;
3569 + struct ext3_inode_info *ei;
3571 + /* preallocate only when found space is larger then requested */
3572 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3573 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3574 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3576 + pa = kmem_cache_alloc(ext3_pspace_cachep, SLAB_NOFS);
3580 + if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3581 + int winl, wins, win, offs;
3583 + /* we can't allocate as much as normalizer wants.
3584 + * so, found space must get proper lstart
3585 + * to cover original request */
3586 + BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3587 + BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3589 + /* we're limited by original request in that
3590 + * logical block must be covered any way
3591 + * winl is window we can move our chunk within */
3592 + winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3594 + /* also, we should cover whole original request */
3595 + wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3597 + /* the smallest one defines real window */
3598 + win = min(winl, wins);
3600 + offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3601 + if (offs && offs < win)
3604 + ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3605 + BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3606 + BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3609 + /* preallocation can change ac_b_ex, thus we store actually
3610 + * allocated blocks for history */
3611 + ac->ac_f_ex = ac->ac_b_ex;
3613 + pa->pa_lstart = ac->ac_b_ex.fe_logical;
3614 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3615 + pa->pa_len = ac->ac_b_ex.fe_len;
3616 + pa->pa_free = pa->pa_len;
3617 + atomic_set(&pa->pa_count, 1);
3618 + spin_lock_init(&pa->pa_lock);
3619 + pa->pa_deleted = 0;
3620 + pa->pa_linear = 0;
3622 + mb_debug("new inode pa %p: %lu/%lu for %lu\n", pa,
3623 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3625 + ext3_mb_use_inode_pa(ac, pa);
3626 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3628 + ei = EXT3_I(ac->ac_inode);
3629 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3631 + pa->pa_obj_lock = &ei->i_prealloc_lock;
3632 + pa->pa_inode = ac->ac_inode;
3634 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3635 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3636 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3638 + spin_lock(pa->pa_obj_lock);
3639 + list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3640 + spin_unlock(pa->pa_obj_lock);
3646 + * creates new preallocated space for locality group inodes belongs to
3648 +int ext3_mb_new_group_pa(struct ext3_allocation_context *ac)
3650 + struct super_block *sb = ac->ac_sb;
3651 + struct ext3_locality_group *lg;
3652 + struct ext3_prealloc_space *pa;
3653 + struct ext3_group_info *grp;
3655 + /* preallocate only when found space is larger then requested */
3656 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3657 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3658 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3660 + BUG_ON(ext3_pspace_cachep == NULL);
3661 + pa = kmem_cache_alloc(ext3_pspace_cachep, SLAB_NOFS);
3665 + /* preallocation can change ac_b_ex, thus we store actually
3666 + * allocated blocks for history */
3667 + ac->ac_f_ex = ac->ac_b_ex;
3669 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3670 + pa->pa_lstart = pa->pa_pstart;
3671 + pa->pa_len = ac->ac_b_ex.fe_len;
3672 + pa->pa_free = pa->pa_len;
3673 + atomic_set(&pa->pa_count, 1);
3674 + spin_lock_init(&pa->pa_lock);
3675 + pa->pa_deleted = 0;
3676 + pa->pa_linear = 1;
3678 + mb_debug("new group pa %p: %lu/%lu for %lu\n", pa,
3679 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3681 + ext3_mb_use_group_pa(ac, pa);
3682 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3684 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3686 + BUG_ON(lg == NULL);
3688 + pa->pa_obj_lock = &lg->lg_prealloc_lock;
3689 + pa->pa_inode = NULL;
3691 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3692 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3693 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3695 + spin_lock(pa->pa_obj_lock);
3696 + list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3697 + spin_unlock(pa->pa_obj_lock);
3702 +int ext3_mb_new_preallocation(struct ext3_allocation_context *ac)
3706 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3707 + err = ext3_mb_new_group_pa(ac);
3709 + err = ext3_mb_new_inode_pa(ac);
3714 + * finds all unused blocks in on-disk bitmap, frees them in
3715 + * in-core bitmap and buddy.
3716 + * @pa must be unlinked from inode and group lists, so that
3717 + * nobody else can find/use it.
3718 + * the caller MUST hold group/inode locks.
3719 + * TODO: optimize the case when there are no in-core structures yet
3721 +int ext3_mb_release_inode_pa(struct ext3_buddy *e3b,
3722 + struct buffer_head *bitmap_bh,
3723 + struct ext3_prealloc_space *pa)
3725 + struct ext3_allocation_context ac;
3726 + struct super_block *sb = e3b->bd_sb;
3727 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3728 + unsigned long bit, end, next, group;
3730 + int err = 0, free = 0;
3732 + BUG_ON(pa->pa_deleted == 0);
3733 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3734 + BUG_ON(group != e3b->bd_group);
3735 + end = bit + pa->pa_len;
3738 + ac.ac_inode = pa->pa_inode;
3739 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3741 + while (bit < end) {
3742 + bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3745 + next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3748 + start = group * EXT3_BLOCKS_PER_GROUP(sb) + bit +
3749 + le32_to_cpu(sbi->s_es->s_first_data_block);
3750 + mb_debug(" free preallocated %u/%u in group %u\n",
3751 + (unsigned) start, (unsigned) next - bit,
3752 + (unsigned) group);
3753 + free += next - bit;
3755 + ac.ac_b_ex.fe_group = group;
3756 + ac.ac_b_ex.fe_start = bit;
3757 + ac.ac_b_ex.fe_len = next - bit;
3758 + ac.ac_b_ex.fe_logical = 0;
3759 + ext3_mb_store_history(&ac);
3761 + mb_free_blocks(e3b, bit, next - bit);
3764 + if (free != pa->pa_free) {
3765 + printk("pa %p: logic %lu, phys. %lu, len %lu\n",
3766 + pa, (unsigned long) pa->pa_lstart,
3767 + (unsigned long) pa->pa_pstart,
3768 + (unsigned long) pa->pa_len);
3769 + printk("free %u, pa_free %u\n", free, pa->pa_free);
3771 + BUG_ON(free != pa->pa_free);
3772 + atomic_add(free, &sbi->s_mb_discarded);
3777 +int ext3_mb_release_group_pa(struct ext3_buddy *e3b,
3778 + struct ext3_prealloc_space *pa)
3780 + struct ext3_allocation_context ac;
3781 + struct super_block *sb = e3b->bd_sb;
3782 + unsigned long bit, group;
3784 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3786 + BUG_ON(pa->pa_deleted == 0);
3787 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3788 + BUG_ON(group != e3b->bd_group);
3789 + mb_free_blocks(e3b, bit, pa->pa_len);
3790 + atomic_add(pa->pa_len, &EXT3_SB(sb)->s_mb_discarded);
3793 + ac.ac_inode = NULL;
3794 + ac.ac_b_ex.fe_group = group;
3795 + ac.ac_b_ex.fe_start = bit;
3796 + ac.ac_b_ex.fe_len = pa->pa_len;
3797 + ac.ac_b_ex.fe_logical = 0;
3798 + ext3_mb_store_history(&ac);
3804 + * releases all preallocations in given group
3806 + * first, we need to decide discard policy:
3807 + * - when do we discard
3809 + * - how many do we discard
3810 + * 1) how many requested
3812 +int ext3_mb_discard_group_preallocations(struct super_block *sb,
3813 + int group, int needed)
3815 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
3816 + struct buffer_head *bitmap_bh = NULL;
3817 + struct ext3_prealloc_space *pa, *tmp;
3818 + struct list_head list;
3819 + struct ext3_buddy e3b;
3820 + int err, busy, free = 0;
3822 + mb_debug("discard preallocation for group %lu\n", group);
3824 + if (list_empty(&grp->bb_prealloc_list))
3827 + bitmap_bh = read_block_bitmap(sb, group);
3828 + if (bitmap_bh == NULL) {
3829 + /* error handling here */
3830 + ext3_mb_release_desc(&e3b);
3831 + BUG_ON(bitmap_bh == NULL);
3834 + err = ext3_mb_load_buddy(sb, group, &e3b);
3835 + BUG_ON(err != 0); /* error handling here */
3838 + needed = EXT3_BLOCKS_PER_GROUP(sb) + 1;
3840 + grp = EXT3_GROUP_INFO(sb, group);
3841 + INIT_LIST_HEAD(&list);
3845 + ext3_lock_group(sb, group);
3846 + list_for_each_entry_safe (pa, tmp, &grp->bb_prealloc_list, pa_group_list) {
3847 + spin_lock(&pa->pa_lock);
3848 + if (atomic_read(&pa->pa_count)) {
3849 + spin_unlock(&pa->pa_lock);
3850 + printk("uh! busy PA\n");
3855 + if (pa->pa_deleted) {
3856 + spin_unlock(&pa->pa_lock);
3860 + /* seems this one can be freed ... */
3861 + pa->pa_deleted = 1;
3863 + /* we can trust pa_free ... */
3864 + free += pa->pa_free;
3866 + spin_unlock(&pa->pa_lock);
3868 + list_del_rcu(&pa->pa_group_list);
3869 + list_add(&pa->u.pa_tmp_list, &list);
3872 + /* if we still need more blocks and some PAs were used, try again */
3873 + if (free < needed && busy)
3876 + /* found anything to free? */
3877 + if (list_empty(&list)) {
3878 + BUG_ON(free != 0);
3882 + /* now free all selected PAs */
3883 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3885 + /* remove from object (inode or locality group) */
3886 + spin_lock(pa->pa_obj_lock);
3887 + list_del_rcu(&pa->pa_inode_list);
3888 + spin_unlock(pa->pa_obj_lock);
3890 + if (pa->pa_linear)
3891 + ext3_mb_release_group_pa(&e3b, pa);
3893 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
3895 + list_del(&pa->u.pa_tmp_list);
3900 + ext3_unlock_group(sb, group);
3901 + ext3_mb_release_desc(&e3b);
3902 + brelse(bitmap_bh);
3907 + * releases all non-used preallocated blocks for given inode
3909 +void ext3_mb_discard_inode_preallocations(struct inode *inode)
3911 + struct ext3_inode_info *ei = EXT3_I(inode);
3912 + struct super_block *sb = inode->i_sb;
3913 + struct buffer_head *bitmap_bh = NULL;
3914 + struct ext3_prealloc_space *pa, *tmp;
3915 + unsigned long group = 0;
3916 + struct list_head list;
3917 + struct ext3_buddy e3b;
3920 + if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3921 + /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3925 + mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3927 + INIT_LIST_HEAD(&list);
3930 + /* first, collect all pa's in the inode */
3931 + spin_lock(&ei->i_prealloc_lock);
3932 + while (!list_empty(&ei->i_prealloc_list)) {
3933 + pa = list_entry(ei->i_prealloc_list.next,
3934 + struct ext3_prealloc_space, pa_inode_list);
3935 + BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3936 + spin_lock(&pa->pa_lock);
3937 + if (atomic_read(&pa->pa_count)) {
3938 + /* this shouldn't happen often - nobody should
3939 + * use preallocation while we're discarding it */
3940 + spin_unlock(&pa->pa_lock);
3941 + spin_unlock(&ei->i_prealloc_lock);
3942 + printk("uh-oh! used pa while discarding\n");
3944 + current->state = TASK_UNINTERRUPTIBLE;
3945 + schedule_timeout(HZ);
3949 + if (pa->pa_deleted == 0) {
3950 + pa->pa_deleted = 1;
3951 + spin_unlock(&pa->pa_lock);
3952 + list_del_rcu(&pa->pa_inode_list);
3953 + list_add(&pa->u.pa_tmp_list, &list);
3957 + /* someone is deleting pa right now */
3958 + spin_unlock(&pa->pa_lock);
3959 + spin_unlock(&ei->i_prealloc_lock);
3961 + /* we have to wait here because pa_deleted
3962 + * doesn't mean pa is already unlinked from
3963 + * the list. as we might be called from
3964 + * ->clear_inode() the inode will get freed
3965 + * and concurrent thread which is unlinking
3966 + * pa from inode's list may access already
3967 + * freed memory, bad-bad-bad */
3969 + /* XXX: if this happens too often, we can
3970 + * add a flag to force wait only in case
3971 + * of ->clear_inode(), but not in case of
3972 + * regular truncate */
3973 + printk("uh-oh! some one just deleted it\n");
3975 + current->state = TASK_UNINTERRUPTIBLE;
3976 + schedule_timeout(HZ);
3979 + spin_unlock(&ei->i_prealloc_lock);
3981 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3982 + BUG_ON(pa->pa_linear != 0);
3983 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3985 + err = ext3_mb_load_buddy(sb, group, &e3b);
3986 + BUG_ON(err != 0); /* error handling here */
3988 + bitmap_bh = read_block_bitmap(sb, group);
3989 + if (bitmap_bh == NULL) {
3990 + /* error handling here */
3991 + ext3_mb_release_desc(&e3b);
3992 + BUG_ON(bitmap_bh == NULL);
3995 + ext3_lock_group(sb, group);
3996 + list_del_rcu(&pa->pa_group_list);
3997 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
3998 + ext3_unlock_group(sb, group);
4000 + ext3_mb_release_desc(&e3b);
4001 + brelse(bitmap_bh);
4003 + list_del(&pa->u.pa_tmp_list);
4009 + * finds all preallocated spaces and return blocks being freed to them
4010 + * if preallocated space becomes full (no block is used from the space)
4011 + * then the function frees space in buddy
4012 + * XXX: at the moment, truncate (which is the only way to free blocks)
4013 + * discards all preallocations
4015 +void ext3_mb_return_to_preallocation(struct inode *inode, struct ext3_buddy *e3b,
4016 + sector_t block, int count)
4018 + BUG_ON(!list_empty(&EXT3_I(inode)->i_prealloc_list));
4021 +void ext3_mb_show_ac(struct ext3_allocation_context *ac)
4024 + struct super_block *sb = ac->ac_sb;
4027 + printk(KERN_ERR "EXT3-fs: can't allocate: status %d flags %d\n",
4028 + ac->ac_status, ac->ac_flags);
4029 + printk(KERN_ERR "EXT3-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4030 + "best %lu/%lu/%lu@%lu cr %d\n",
4031 + ac->ac_o_ex.fe_group, ac->ac_o_ex.fe_start,
4032 + ac->ac_o_ex.fe_len, ac->ac_o_ex.fe_logical,
4033 + ac->ac_g_ex.fe_group, ac->ac_g_ex.fe_start,
4034 + ac->ac_g_ex.fe_len, ac->ac_g_ex.fe_logical,
4035 + ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
4036 + ac->ac_b_ex.fe_len, ac->ac_b_ex.fe_logical,
4038 + printk(KERN_ERR "EXT3-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4040 + printk("EXT3-fs: groups: ");
4041 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
4042 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, i);
4043 + struct ext3_prealloc_space *pa;
4044 + unsigned long start;
4045 + struct list_head *cur;
4046 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
4047 + pa = list_entry(cur, struct ext3_prealloc_space,
4049 + spin_lock(&pa->pa_lock);
4050 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, NULL, &start);
4051 + spin_unlock(&pa->pa_lock);
4052 + printk("PA:%u:%lu:%u ", i, start, pa->pa_len);
4055 + if (grp->bb_free == 0)
4057 + printk("%d: %d/%d ", i, grp->bb_free, grp->bb_fragments);
4064 +void ext3_mb_group_or_file(struct ext3_allocation_context *ac)
4066 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
4067 + int bsbits = ac->ac_sb->s_blocksize_bits;
4068 + loff_t size, isize;
4070 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
4073 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4074 + isize = i_size_read(ac->ac_inode) >> bsbits;
4078 + /* don't use group allocation for large files */
4079 + if (size >= sbi->s_mb_stream_request)
4082 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
4085 + BUG_ON(ac->ac_lg != NULL);
4086 + ac->ac_lg = &sbi->s_locality_groups[smp_processor_id()];
4088 + /* we're going to use group allocation */
4089 + ac->ac_flags |= EXT3_MB_HINT_GROUP_ALLOC;
4091 + /* serialize all allocations in the group */
4092 + down(&ac->ac_lg->lg_sem);
4095 +int ext3_mb_initialize_context(struct ext3_allocation_context *ac,
4096 + struct ext3_allocation_request *ar)
4098 + struct super_block *sb = ar->inode->i_sb;
4099 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4100 + struct ext3_super_block *es = sbi->s_es;
4101 + unsigned long group, len, goal;
4102 + unsigned long block;
4104 + /* we can't allocate > group size */
4106 + if (len >= EXT3_BLOCKS_PER_GROUP(sb) - 10)
4107 + len = EXT3_BLOCKS_PER_GROUP(sb) - 10;
4109 + /* start searching from the goal */
4111 + if (goal < le32_to_cpu(es->s_first_data_block) ||
4112 + goal >= le32_to_cpu(es->s_blocks_count))
4113 + goal = le32_to_cpu(es->s_first_data_block);
4114 + ext3_get_group_no_and_offset(sb, goal, &group, &block);
4116 + /* set up allocation goals */
4117 + ac->ac_b_ex.fe_logical = ar->logical;
4118 + ac->ac_b_ex.fe_group = 0;
4119 + ac->ac_b_ex.fe_start = 0;
4120 + ac->ac_b_ex.fe_len = 0;
4121 + ac->ac_status = AC_STATUS_CONTINUE;
4122 + ac->ac_groups_scanned = 0;
4123 + ac->ac_ex_scanned = 0;
4126 + ac->ac_inode = ar->inode;
4127 + ac->ac_o_ex.fe_logical = ar->logical;
4128 + ac->ac_o_ex.fe_group = group;
4129 + ac->ac_o_ex.fe_start = block;
4130 + ac->ac_o_ex.fe_len = len;
4131 + ac->ac_g_ex.fe_logical = ar->logical;
4132 + ac->ac_g_ex.fe_group = group;
4133 + ac->ac_g_ex.fe_start = block;
4134 + ac->ac_g_ex.fe_len = len;
4135 + ac->ac_f_ex.fe_len = 0;
4136 + ac->ac_flags = ar->flags;
4137 + ac->ac_2order = 0;
4138 + ac->ac_criteria = 0;
4140 + ac->ac_bitmap_page = NULL;
4141 + ac->ac_buddy_page = NULL;
4144 + /* we have to define context: we'll we work with a file or
4145 + * locality group. this is a policy, actually */
4146 + ext3_mb_group_or_file(ac);
4148 + mb_debug("init ac: %u blocks @ %llu, goal %llu, flags %x, 2^%d, "
4149 + "left: %llu/%llu, right %llu/%llu to %swritable\n",
4150 + (unsigned) ar->len, (unsigned) ar->logical,
4151 + (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4152 + (unsigned) ar->lleft, (unsigned) ar->pleft,
4153 + (unsigned) ar->lright, (unsigned) ar->pright,
4154 + atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4160 + * release all resource we used in allocation
4162 +int ext3_mb_release_context(struct ext3_allocation_context *ac)
4165 + if (ac->ac_pa->pa_linear) {
4166 + /* see comment in ext3_mb_use_group_pa() */
4167 + spin_lock(&ac->ac_pa->pa_lock);
4168 + ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
4169 + ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
4170 + ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
4171 + ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
4172 + spin_unlock(&ac->ac_pa->pa_lock);
4174 + ext3_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
4176 + if (ac->ac_bitmap_page)
4177 + page_cache_release(ac->ac_bitmap_page);
4178 + if (ac->ac_buddy_page)
4179 + page_cache_release(ac->ac_buddy_page);
4180 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
4181 + up(&ac->ac_lg->lg_sem);
4182 + ext3_mb_collect_stats(ac);
4186 +int ext3_mb_discard_preallocations(struct super_block *sb, int needed)
4188 + int i, ret, freed = 0;
4190 + for (i = 0; i < EXT3_SB(sb)->s_groups_count && needed > 0; i++) {
4191 + ret = ext3_mb_discard_group_preallocations(sb, i, needed);
4200 + * Main entry point into mballoc to allocate blocks
4201 + * it tries to use preallocation first, then falls back
4202 + * to usual allocation
4204 +unsigned long ext3_mb_new_blocks(handle_t *handle,
4205 + struct ext3_allocation_request *ar, int *errp)
4207 + struct ext3_allocation_context ac;
4208 + struct ext3_sb_info *sbi;
4209 + struct super_block *sb;
4210 + unsigned long block;
4213 + sb = ar->inode->i_sb;
4214 + sbi = EXT3_SB(sb);
4216 + if (!test_opt(sb, MBALLOC)) {
4217 + static int ext3_mballoc_warning = 0;
4218 + if (ext3_mballoc_warning++ == 0)
4219 + printk(KERN_ERR "EXT3-fs: multiblock request with "
4220 + "mballoc disabled!\n");
4222 + err = ext3_new_block_old(handle, ar->inode, ar->goal, errp);
4226 + ext3_mb_poll_new_transaction(sb, handle);
4228 + if ((err = ext3_mb_initialize_context(&ac, ar)))
4231 + ac.ac_op = EXT3_MB_HISTORY_PREALLOC;
4232 + if (!ext3_mb_use_preallocated(&ac)) {
4234 + ac.ac_op = EXT3_MB_HISTORY_ALLOC;
4235 + ext3_mb_normalize_request(&ac, ar);
4238 + /* allocate space in core */
4239 + ext3_mb_regular_allocator(&ac);
4241 + /* as we've just preallocated more space than
4242 + * user requested orinally, we store allocated
4243 + * space in a special descriptor */
4244 + if (ac.ac_status == AC_STATUS_FOUND &&
4245 + ac.ac_o_ex.fe_len < ac.ac_b_ex.fe_len)
4246 + ext3_mb_new_preallocation(&ac);
4249 + if (likely(ac.ac_status == AC_STATUS_FOUND)) {
4250 + ext3_mb_mark_diskspace_used(&ac, handle);
4252 + block = ext3_grp_offs_to_block(sb, &ac.ac_b_ex);
4253 + ar->len = ac.ac_b_ex.fe_len;
4255 + freed = ext3_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
4259 + ac.ac_b_ex.fe_len = 0;
4261 + ext3_mb_show_ac(&ac);
4264 + ext3_mb_release_context(&ac);
4268 +EXPORT_SYMBOL(ext3_mb_new_blocks);
4270 +int ext3_new_block(handle_t *handle, struct inode *inode,
4271 + unsigned long goal, int *errp)
4273 + struct ext3_allocation_request ar;
4274 + unsigned long ret;
4276 + if (!test_opt(inode->i_sb, MBALLOC)) {
4277 + ret = ext3_new_block_old(handle, inode, goal, errp);
4290 + ret = ext3_mb_new_blocks(handle, &ar, errp);
4294 +void ext3_mb_poll_new_transaction(struct super_block *sb, handle_t *handle)
4296 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4298 + if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4301 + /* new transaction! time to close last one and free blocks for
4302 + * committed transaction. we know that only transaction can be
4303 + * active, so previos transaction can be being logged and we
4304 + * know that transaction before previous is known to be already
4305 + * logged. this means that now we may free blocks freed in all
4306 + * transactions before previous one. hope I'm clear enough ... */
4308 + spin_lock(&sbi->s_md_lock);
4309 + if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4310 + mb_debug("new transaction %lu, old %lu\n",
4311 + (unsigned long) handle->h_transaction->t_tid,
4312 + (unsigned long) sbi->s_last_transaction);
4313 + list_splice_init(&sbi->s_closed_transaction,
4314 + &sbi->s_committed_transaction);
4315 + list_splice_init(&sbi->s_active_transaction,
4316 + &sbi->s_closed_transaction);
4317 + sbi->s_last_transaction = handle->h_transaction->t_tid;
4319 + spin_unlock(&sbi->s_md_lock);
4321 + ext3_mb_free_committed_blocks(sb);
4324 +int ext3_mb_free_metadata(handle_t *handle, struct ext3_buddy *e3b,
4325 + int group, int block, int count)
4327 + struct ext3_group_info *db = e3b->bd_info;
4328 + struct super_block *sb = e3b->bd_sb;
4329 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4330 + struct ext3_free_metadata *md;
4333 + BUG_ON(e3b->bd_bitmap_page == NULL);
4334 + BUG_ON(e3b->bd_buddy_page == NULL);
4336 + ext3_lock_group(sb, group);
4337 + for (i = 0; i < count; i++) {
4338 + md = db->bb_md_cur;
4339 + if (md && db->bb_tid != handle->h_transaction->t_tid) {
4340 + db->bb_md_cur = NULL;
4345 + ext3_unlock_group(sb, group);
4346 + md = kmalloc(sizeof(*md), GFP_KERNEL);
4350 + md->group = group;
4352 + ext3_lock_group(sb, group);
4353 + if (db->bb_md_cur == NULL) {
4354 + spin_lock(&sbi->s_md_lock);
4355 + list_add(&md->list, &sbi->s_active_transaction);
4356 + spin_unlock(&sbi->s_md_lock);
4357 + /* protect buddy cache from being freed,
4358 + * otherwise we'll refresh it from
4359 + * on-disk bitmap and lose not-yet-available
4361 + page_cache_get(e3b->bd_buddy_page);
4362 + page_cache_get(e3b->bd_bitmap_page);
4363 + db->bb_md_cur = md;
4364 + db->bb_tid = handle->h_transaction->t_tid;
4365 + mb_debug("new md 0x%p for group %u\n",
4369 + md = db->bb_md_cur;
4373 + BUG_ON(md->num >= EXT3_BB_MAX_BLOCKS);
4374 + md->blocks[md->num] = block + i;
4376 + if (md->num == EXT3_BB_MAX_BLOCKS) {
4377 + /* no more space, put full container on a sb's list */
4378 + db->bb_md_cur = NULL;
4381 + ext3_unlock_group(sb, group);
4386 + * Main entry point into mballoc to free blocks
4388 +void ext3_mb_free_blocks(handle_t *handle, struct inode *inode,
4389 + unsigned long block, unsigned long count,
4390 + int metadata, int *freed)
4392 + struct buffer_head *bitmap_bh = NULL;
4393 + struct super_block *sb = inode->i_sb;
4394 + struct ext3_allocation_context ac;
4395 + struct ext3_group_desc *gdp;
4396 + struct ext3_super_block *es;
4397 + unsigned long bit, overflow;
4398 + struct buffer_head *gd_bh;
4399 + unsigned long block_group;
4400 + struct ext3_sb_info *sbi;
4401 + struct ext3_buddy e3b;
4406 + ext3_mb_poll_new_transaction(sb, handle);
4408 + sbi = EXT3_SB(sb);
4409 + es = EXT3_SB(sb)->s_es;
4410 + if (block < le32_to_cpu(es->s_first_data_block) ||
4411 + block + count < block ||
4412 + block + count > le32_to_cpu(es->s_blocks_count)) {
4413 + ext3_error (sb, __FUNCTION__,
4414 + "Freeing blocks not in datazone - "
4415 + "block = %lu, count = %lu", block, count);
4416 + goto error_return;
4419 + ext3_debug("freeing block %lu\n", block);
4421 + ac.ac_op = EXT3_MB_HISTORY_FREE;
4422 + ac.ac_inode = inode;
4427 + ext3_get_group_no_and_offset(sb, block, &block_group, &bit);
4430 + * Check to see if we are freeing blocks across a group
4433 + if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
4434 + overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
4435 + count -= overflow;
4437 + brelse(bitmap_bh);
4438 + bitmap_bh = read_block_bitmap(sb, block_group);
4440 + goto error_return;
4441 + gdp = ext3_get_group_desc (sb, block_group, &gd_bh);
4443 + goto error_return;
4445 + if (in_range (le32_to_cpu(gdp->bg_block_bitmap), block, count) ||
4446 + in_range (le32_to_cpu(gdp->bg_inode_bitmap), block, count) ||
4447 + in_range (block, le32_to_cpu(gdp->bg_inode_table),
4448 + EXT3_SB(sb)->s_itb_per_group) ||
4449 + in_range (block + count - 1, le32_to_cpu(gdp->bg_inode_table),
4450 + EXT3_SB(sb)->s_itb_per_group))
4451 + ext3_error(sb, __FUNCTION__,
4452 + "Freeing blocks in system zone - "
4453 + "Block = %lu, count = %lu", block, count);
4455 + BUFFER_TRACE(bitmap_bh, "getting write access");
4456 + err = ext3_journal_get_write_access(handle, bitmap_bh);
4458 + goto error_return;
4461 + * We are about to modify some metadata. Call the journal APIs
4462 + * to unshare ->b_data if a currently-committing transaction is
4465 + BUFFER_TRACE(gd_bh, "get_write_access");
4466 + err = ext3_journal_get_write_access(handle, gd_bh);
4468 + goto error_return;
4470 + err = ext3_mb_load_buddy(sb, block_group, &e3b);
4472 + goto error_return;
4474 +#ifdef AGGRESSIVE_CHECK
4477 + for (i = 0; i < count; i++)
4478 + BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4481 + mb_clear_bits(bitmap_bh->b_data, bit, count);
4483 + /* We dirtied the bitmap block */
4484 + BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4485 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
4487 + ac.ac_b_ex.fe_group = block_group;
4488 + ac.ac_b_ex.fe_start = bit;
4489 + ac.ac_b_ex.fe_len = count;
4490 + ext3_mb_store_history(&ac);
4493 + /* blocks being freed are metadata. these blocks shouldn't
4494 + * be used until this transaction is committed */
4495 + ext3_mb_free_metadata(handle, &e3b, block_group, bit, count);
4497 + ext3_lock_group(sb, block_group);
4498 + err = mb_free_blocks(&e3b, bit, count);
4499 + ext3_mb_return_to_preallocation(inode, &e3b, block, count);
4500 + ext3_unlock_group(sb, block_group);
4504 + spin_lock(sb_bgl_lock(sbi, block_group));
4505 + gdp->bg_free_blocks_count =
4506 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
4507 + spin_unlock(sb_bgl_lock(sbi, block_group));
4508 + percpu_counter_mod(&sbi->s_freeblocks_counter, count);
4510 + ext3_mb_release_desc(&e3b);
4514 + /* And the group descriptor block */
4515 + BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4516 + ret = ext3_journal_dirty_metadata(handle, gd_bh);
4517 + if (!err) err = ret;
4519 + if (overflow && !err) {
4526 + brelse(bitmap_bh);
4527 + ext3_std_error(sb, err);