1 Index: linux-2.6.9-42.0.10.EL_lustre.1.4.10/include/linux/ext3_fs_sb.h
2 ===================================================================
3 --- linux-2.6.9-42.0.10.EL_lustre.1.4.10.orig/include/linux/ext3_fs_sb.h 2007-06-14 13:59:04.000000000 +0200
4 +++ linux-2.6.9-42.0.10.EL_lustre.1.4.10/include/linux/ext3_fs_sb.h 2007-06-14 14:16:57.000000000 +0200
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-2.6.9-42.0.10.EL_lustre.1.4.10/include/linux/ext3_fs.h
68 ===================================================================
69 --- linux-2.6.9-42.0.10.EL_lustre.1.4.10.orig/include/linux/ext3_fs.h 2007-06-14 13:59:04.000000000 +0200
70 +++ linux-2.6.9-42.0.10.EL_lustre.1.4.10/include/linux/ext3_fs.h 2007-06-14 14:16:57.000000000 +0200
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
102 @@ -387,6 +411,14 @@ struct ext3_inode {
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 - CFS bug 10634"
111 +#endif /* __LITTLE_ENDIAN */
112 +#endif /* !ext2_find_next_le_bit */
115 * Maximal mount counts between two filesystem checks
117 @@ -763,6 +795,20 @@ 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-2.6.9-42.0.10.EL_lustre.1.4.10/fs/ext3/super.c
139 ===================================================================
140 --- linux-2.6.9-42.0.10.EL_lustre.1.4.10.orig/fs/ext3/super.c 2007-06-14 13:59:04.000000000 +0200
141 +++ linux-2.6.9-42.0.10.EL_lustre.1.4.10/fs/ext3/super.c 2007-06-14 14:16:57.000000000 +0200
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 @@ -2576,7 +2579,13 @@ 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 @@ -2598,6 +2607,7 @@ 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-2.6.9-42.0.10.EL_lustre.1.4.10/fs/ext3/mballoc.c
183 ===================================================================
184 --- /dev/null 1970-01-01 00:00:00.000000000 +0000
185 +++ linux-2.6.9-42.0.10.EL_lustre.1.4.10/fs/ext3/mballoc.c 2007-06-14 14:19:27.000000000 +0200
188 + * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
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 inode *inode, struct ext3_buddy *e3b,
826 + int first, int count)
829 + struct super_block *sb = e3b->bd_sb;
831 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
833 + BUG_ON(!ext3_is_group_locked(sb, e3b->bd_group));
834 + for (i = 0; i < count; i++) {
835 + if (!mb_test_bit(first + i, e3b->bd_info->bb_bitmap)) {
836 + unsigned long blocknr;
837 + blocknr = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb);
838 + blocknr += first + i;
840 + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
842 + ext3_error(sb, __FUNCTION__, "double-free of inode"
843 + " %lu's block %lu(bit %u in group %u)\n",
844 + inode ? inode->i_ino : 0, blocknr,
845 + first + i, e3b->bd_group);
847 + mb_clear_bit(first + i, e3b->bd_info->bb_bitmap);
851 +void mb_mark_used_double(struct ext3_buddy *e3b, int first, int count)
854 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
856 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
857 + for (i = 0; i < count; i++) {
858 + BUG_ON(mb_test_bit(first + i, e3b->bd_info->bb_bitmap));
859 + mb_set_bit(first + i, e3b->bd_info->bb_bitmap);
863 +void mb_cmp_bitmaps(struct ext3_buddy *e3b, void *bitmap)
865 + if (memcmp(e3b->bd_info->bb_bitmap, bitmap, e3b->bd_sb->s_blocksize)) {
866 + unsigned char *b1, *b2;
868 + b1 = (unsigned char *) e3b->bd_info->bb_bitmap;
869 + b2 = (unsigned char *) bitmap;
870 + for (i = 0; i < e3b->bd_sb->s_blocksize; i++) {
871 + if (b1[i] != b2[i]) {
872 + printk("corruption in group %u at byte %u(%u): "
873 + "%x in copy != %x on disk/prealloc\n",
874 + e3b->bd_group, i, i * 8, b1[i], b2[i]);
882 +#define mb_free_blocks_double(a,b,c,d)
883 +#define mb_mark_used_double(a,b,c)
884 +#define mb_cmp_bitmaps(a,b)
887 +#ifdef AGGRESSIVE_CHECK
889 +#define MB_CHECK_ASSERT(assert) \
892 + printk (KERN_EMERG \
893 + "Assertion failure in %s() at %s:%d: \"%s\"\n", \
894 + function, file, line, # assert); \
899 +static int __mb_check_buddy(struct ext3_buddy *e3b, char *file,
900 + const char *function, int line)
902 + struct super_block *sb = e3b->bd_sb;
903 + int order = e3b->bd_blkbits + 1;
904 + int max, max2, i, j, k, count;
905 + struct ext3_group_info *grp;
906 + int fragments = 0, fstart;
907 + struct list_head *cur;
908 + void *buddy, *buddy2;
910 + if (!test_opt(sb, MBALLOC))
914 + static int mb_check_counter = 0;
915 + if (mb_check_counter++ % 100 != 0)
919 + while (order > 1) {
920 + buddy = mb_find_buddy(e3b, order, &max);
921 + MB_CHECK_ASSERT(buddy);
922 + buddy2 = mb_find_buddy(e3b, order - 1, &max2);
923 + MB_CHECK_ASSERT(buddy2);
924 + MB_CHECK_ASSERT(buddy != buddy2);
925 + MB_CHECK_ASSERT(max * 2 == max2);
928 + for (i = 0; i < max; i++) {
930 + if (mb_test_bit(i, buddy)) {
931 + /* only single bit in buddy2 may be 1 */
932 + if (!mb_test_bit(i << 1, buddy2))
933 + MB_CHECK_ASSERT(mb_test_bit((i<<1)+1, buddy2));
934 + else if (!mb_test_bit((i << 1) + 1, buddy2))
935 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
939 + /* both bits in buddy2 must be 0 */
940 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
941 + MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
943 + for (j = 0; j < (1 << order); j++) {
944 + k = (i * (1 << order)) + j;
945 + MB_CHECK_ASSERT(!mb_test_bit(k, EXT3_MB_BITMAP(e3b)));
949 + MB_CHECK_ASSERT(e3b->bd_info->bb_counters[order] == count);
954 + buddy = mb_find_buddy(e3b, 0, &max);
955 + for (i = 0; i < max; i++) {
956 + if (!mb_test_bit(i, buddy)) {
957 + MB_CHECK_ASSERT(i >= e3b->bd_info->bb_first_free);
958 + if (fstart == -1) {
965 + /* check used bits only */
966 + for (j = 0; j < e3b->bd_blkbits + 1; j++) {
967 + buddy2 = mb_find_buddy(e3b, j, &max2);
969 + MB_CHECK_ASSERT(k < max2);
970 + MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
973 + MB_CHECK_ASSERT(!EXT3_MB_GRP_NEED_INIT(e3b->bd_info));
974 + MB_CHECK_ASSERT(e3b->bd_info->bb_fragments == fragments);
976 + grp = EXT3_GROUP_INFO(sb, e3b->bd_group);
977 + buddy = mb_find_buddy(e3b, 0, &max);
978 + list_for_each(cur, &grp->bb_prealloc_list) {
979 + unsigned long groupnr;
980 + struct ext3_prealloc_space *pa;
981 + pa = list_entry(cur, struct ext3_prealloc_space, group_list);
982 + ext3_get_group_no_and_offset(sb, pa->pstart, &groupnr, &k);
983 + MB_CHECK_ASSERT(groupnr == e3b->bd_group);
984 + for (i = 0; i < pa->len; i++)
985 + MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
989 +#undef MB_CHECK_ASSERT
990 +#define mb_check_buddy(e3b) __mb_check_buddy(e3b,__FILE__,__FUNCTION__,__LINE__)
992 +#define mb_check_buddy(e3b)
995 +/* find most significant bit */
996 +static int inline fmsb(unsigned short word)
1010 + } while (word != 0);
1016 +ext3_mb_mark_free_simple(struct super_block *sb, void *buddy, unsigned first,
1017 + int len, struct ext3_group_info *grp)
1019 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1020 + unsigned short min, max, chunk, border;
1022 + BUG_ON(len >= EXT3_BLOCKS_PER_GROUP(sb));
1024 + border = 2 << sb->s_blocksize_bits;
1027 + /* find how many blocks can be covered since this position */
1028 + max = ffs(first | border) - 1;
1030 + /* find how many blocks of power 2 we need to mark */
1037 + /* mark multiblock chunks only */
1038 + grp->bb_counters[min]++;
1040 + mb_clear_bit(first >> min, buddy + sbi->s_mb_offsets[min]);
1048 +ext3_mb_generate_buddy(struct super_block *sb, void *buddy, void *bitmap,
1051 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
1052 + unsigned short max = EXT3_BLOCKS_PER_GROUP(sb);
1053 + unsigned short i = 0, first, len;
1054 + unsigned free = 0, fragments = 0;
1055 + unsigned long long period = get_cycles();
1057 + /* initialize buddy from bitmap which is aggregation
1058 + * of on-disk bitmap and preallocations */
1059 + i = mb_find_next_zero_bit(bitmap, max, 0);
1060 + grp->bb_first_free = i;
1064 + i = ext2_find_next_le_bit(bitmap, max, i);
1068 + ext3_mb_mark_free_simple(sb, buddy, first, len, grp);
1070 + grp->bb_counters[0]++;
1072 + i = mb_find_next_zero_bit(bitmap, max, i);
1074 + grp->bb_fragments = fragments;
1076 + if (free != grp->bb_free) {
1077 + printk("EXT3-fs: group %u: %u blocks in bitmap, %u in gd\n",
1078 + group, free, grp->bb_free);
1079 + grp->bb_free = free;
1082 + clear_bit(EXT3_GROUP_INFO_NEED_INIT_BIT, &grp->bb_state);
1084 + period = get_cycles() - period;
1085 + spin_lock(&EXT3_SB(sb)->s_bal_lock);
1086 + EXT3_SB(sb)->s_mb_buddies_generated++;
1087 + EXT3_SB(sb)->s_mb_generation_time += period;
1088 + spin_unlock(&EXT3_SB(sb)->s_bal_lock);
1091 +static int ext3_mb_init_cache(struct page *page, char *incore)
1093 + int blocksize, blocks_per_page, groups_per_page;
1094 + int err = 0, i, first_group, first_block;
1095 + struct super_block *sb;
1096 + struct buffer_head *bhs;
1097 + struct buffer_head **bh;
1098 + struct inode *inode;
1099 + char *data, *bitmap;
1101 + mb_debug("init page %lu\n", page->index);
1103 + inode = page->mapping->host;
1105 + blocksize = 1 << inode->i_blkbits;
1106 + blocks_per_page = PAGE_CACHE_SIZE / blocksize;
1108 + groups_per_page = blocks_per_page >> 1;
1109 + if (groups_per_page == 0)
1110 + groups_per_page = 1;
1112 + /* allocate buffer_heads to read bitmaps */
1113 + if (groups_per_page > 1) {
1115 + i = sizeof(struct buffer_head *) * groups_per_page;
1116 + bh = kmalloc(i, GFP_NOFS);
1123 + first_group = page->index * blocks_per_page / 2;
1125 + /* read all groups the page covers into the cache */
1126 + for (i = 0; i < groups_per_page; i++) {
1127 + struct ext3_group_desc * desc;
1129 + if (first_group + i >= EXT3_SB(sb)->s_groups_count)
1133 + desc = ext3_get_group_desc(sb, first_group + i, NULL);
1138 + bh[i] = sb_getblk(sb, le32_to_cpu(desc->bg_block_bitmap));
1139 + if (bh[i] == NULL)
1142 + if (buffer_uptodate(bh[i]))
1145 + lock_buffer(bh[i]);
1146 + if (buffer_uptodate(bh[i])) {
1147 + unlock_buffer(bh[i]);
1152 + bh[i]->b_end_io = end_buffer_read_sync;
1153 + submit_bh(READ, bh[i]);
1154 + mb_debug("read bitmap for group %u\n", first_group + i);
1157 + /* wait for I/O completion */
1158 + for (i = 0; i < groups_per_page && bh[i]; i++)
1159 + wait_on_buffer(bh[i]);
1162 + for (i = 0; i < groups_per_page && bh[i]; i++)
1163 + if (!buffer_uptodate(bh[i]))
1166 + first_block = page->index * blocks_per_page;
1167 + for (i = 0; i < blocks_per_page; i++) {
1170 + group = (first_block + i) >> 1;
1171 + if (group >= EXT3_SB(sb)->s_groups_count)
1174 + data = page_address(page) + (i * blocksize);
1175 + bitmap = bh[group - first_group]->b_data;
1177 + if ((first_block + i) & 1) {
1178 + /* this is block of buddy */
1179 + BUG_ON(incore == NULL);
1180 + mb_debug("put buddy for group %u in page %lu/%x\n",
1181 + group, page->index, i * blocksize);
1182 + memset(data, 0xff, blocksize);
1183 + EXT3_GROUP_INFO(sb, group)->bb_fragments = 0;
1184 + memset(EXT3_GROUP_INFO(sb, group)->bb_counters, 0,
1185 + sizeof(unsigned short)*(sb->s_blocksize_bits+2));
1186 + ext3_mb_generate_buddy(sb, data, incore, group);
1189 + /* this is block of bitmap */
1190 + BUG_ON(incore != NULL);
1191 + mb_debug("put bitmap for group %u in page %lu/%x\n",
1192 + group, page->index, i * blocksize);
1194 + /* see comments in ext3_mb_put_pa() */
1195 + ext3_lock_group(sb, group);
1196 + memcpy(data, bitmap, blocksize);
1198 + /* mark all preallocated blocks used in in-core bitmap */
1199 + ext3_mb_generate_from_pa(sb, data, group);
1200 + ext3_unlock_group(sb, group);
1205 + SetPageUptodate(page);
1209 + for (i = 0; i < groups_per_page && bh[i]; i++)
1217 +static int ext3_mb_load_buddy(struct super_block *sb, int group,
1218 + struct ext3_buddy *e3b)
1220 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1221 + struct inode *inode = sbi->s_buddy_cache;
1222 + int blocks_per_page, block, pnum, poff;
1223 + struct page *page;
1225 + mb_debug("load group %u\n", group);
1227 + blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1229 + e3b->bd_blkbits = sb->s_blocksize_bits;
1230 + e3b->bd_info = EXT3_GROUP_INFO(sb, group);
1232 + e3b->bd_group = group;
1233 + e3b->bd_buddy_page = NULL;
1234 + e3b->bd_bitmap_page = NULL;
1236 + block = group * 2;
1237 + pnum = block / blocks_per_page;
1238 + poff = block % blocks_per_page;
1240 + /* we could use find_or_create_page(), but it locks page
1241 + * what we'd like to avoid in fast path ... */
1242 + page = find_get_page(inode->i_mapping, pnum);
1243 + if (page == NULL || !PageUptodate(page)) {
1245 + page_cache_release(page);
1246 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1248 + BUG_ON(page->mapping != inode->i_mapping);
1249 + if (!PageUptodate(page)) {
1250 + ext3_mb_init_cache(page, NULL);
1251 + mb_cmp_bitmaps(e3b, page_address(page) +
1252 + (poff * sb->s_blocksize));
1254 + unlock_page(page);
1257 + if (page == NULL || !PageUptodate(page))
1259 + e3b->bd_bitmap_page = page;
1260 + e3b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1261 + mark_page_accessed(page);
1264 + pnum = block / blocks_per_page;
1265 + poff = block % blocks_per_page;
1267 + page = find_get_page(inode->i_mapping, pnum);
1268 + if (page == NULL || !PageUptodate(page)) {
1270 + page_cache_release(page);
1271 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1273 + BUG_ON(page->mapping != inode->i_mapping);
1274 + if (!PageUptodate(page))
1275 + ext3_mb_init_cache(page, e3b->bd_bitmap);
1277 + unlock_page(page);
1280 + if (page == NULL || !PageUptodate(page))
1282 + e3b->bd_buddy_page = page;
1283 + e3b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1284 + mark_page_accessed(page);
1286 + BUG_ON(e3b->bd_bitmap_page == NULL);
1287 + BUG_ON(e3b->bd_buddy_page == NULL);
1292 + if (e3b->bd_bitmap_page)
1293 + page_cache_release(e3b->bd_bitmap_page);
1294 + if (e3b->bd_buddy_page)
1295 + page_cache_release(e3b->bd_buddy_page);
1296 + e3b->bd_buddy = NULL;
1297 + e3b->bd_bitmap = NULL;
1301 +static void ext3_mb_release_desc(struct ext3_buddy *e3b)
1303 + if (e3b->bd_bitmap_page)
1304 + page_cache_release(e3b->bd_bitmap_page);
1305 + if (e3b->bd_buddy_page)
1306 + page_cache_release(e3b->bd_buddy_page);
1310 +static int mb_find_order_for_block(struct ext3_buddy *e3b, int block)
1315 + BUG_ON(EXT3_MB_BITMAP(e3b) == EXT3_MB_BUDDY(e3b));
1316 + BUG_ON(block >= (1 << (e3b->bd_blkbits + 3)));
1318 + bb = EXT3_MB_BUDDY(e3b);
1319 + while (order <= e3b->bd_blkbits + 1) {
1320 + block = block >> 1;
1321 + if (!mb_test_bit(block, bb)) {
1322 + /* this block is part of buddy of order 'order' */
1325 + bb += 1 << (e3b->bd_blkbits - order);
1331 +static inline void mb_clear_bits(void *bm, int cur, int len)
1336 + while (cur < len) {
1337 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1338 + /* fast path: clear whole word at once */
1339 + addr = bm + (cur >> 3);
1344 + mb_clear_bit_atomic(cur, bm);
1349 +static inline void mb_set_bits(void *bm, int cur, int len)
1354 + while (cur < len) {
1355 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1356 + /* fast path: clear whole word at once */
1357 + addr = bm + (cur >> 3);
1358 + *addr = 0xffffffff;
1362 + mb_set_bit_atomic(cur, bm);
1367 +static int mb_free_blocks(struct inode *inode, struct ext3_buddy *e3b,
1368 + int first, int count)
1370 + int block = 0, max = 0, order;
1371 + void *buddy, *buddy2;
1372 + struct super_block *sb = e3b->bd_sb;
1374 + BUG_ON(first + count > (sb->s_blocksize << 3));
1375 + BUG_ON(!ext3_is_group_locked(sb, e3b->bd_group));
1376 + mb_check_buddy(e3b);
1377 + mb_free_blocks_double(inode, e3b, first, count);
1379 + e3b->bd_info->bb_free += count;
1380 + if (first < e3b->bd_info->bb_first_free)
1381 + e3b->bd_info->bb_first_free = first;
1383 + /* let's maintain fragments counter */
1385 + block = !mb_test_bit(first - 1, EXT3_MB_BITMAP(e3b));
1386 + if (first + count < EXT3_SB(sb)->s_mb_maxs[0])
1387 + max = !mb_test_bit(first + count, EXT3_MB_BITMAP(e3b));
1389 + e3b->bd_info->bb_fragments--;
1390 + else if (!block && !max)
1391 + e3b->bd_info->bb_fragments++;
1393 + /* let's maintain buddy itself */
1394 + while (count-- > 0) {
1398 + if (!mb_test_bit(block, EXT3_MB_BITMAP(e3b))) {
1399 + unsigned long blocknr;
1400 + blocknr = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb);
1403 + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
1405 + ext3_error(sb, __FUNCTION__, "double-free of inode"
1406 + " %lu's block %lu(bit %u in group %u)\n",
1407 + inode ? inode->i_ino : 0, blocknr, block,
1410 + mb_clear_bit(block, EXT3_MB_BITMAP(e3b));
1411 + e3b->bd_info->bb_counters[order]++;
1413 + /* start of the buddy */
1414 + buddy = mb_find_buddy(e3b, order, &max);
1418 + if (mb_test_bit(block, buddy) ||
1419 + mb_test_bit(block + 1, buddy))
1422 + /* both the buddies are free, try to coalesce them */
1423 + buddy2 = mb_find_buddy(e3b, order + 1, &max);
1429 + /* for special purposes, we don't set
1430 + * free bits in bitmap */
1431 + mb_set_bit(block, buddy);
1432 + mb_set_bit(block + 1, buddy);
1434 + e3b->bd_info->bb_counters[order]--;
1435 + e3b->bd_info->bb_counters[order]--;
1437 + block = block >> 1;
1439 + e3b->bd_info->bb_counters[order]++;
1441 + mb_clear_bit(block, buddy2);
1445 + mb_check_buddy(e3b);
1450 +static int mb_find_extent(struct ext3_buddy *e3b, int order, int block,
1451 + int needed, struct ext3_free_extent *ex)
1453 + int next = block, max, ord;
1456 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1457 + BUG_ON(ex == NULL);
1459 + buddy = mb_find_buddy(e3b, order, &max);
1460 + BUG_ON(buddy == NULL);
1461 + BUG_ON(block >= max);
1462 + if (mb_test_bit(block, buddy)) {
1469 + if (likely(order == 0)) {
1470 + /* find actual order */
1471 + order = mb_find_order_for_block(e3b, block);
1472 + block = block >> order;
1475 + ex->fe_len = 1 << order;
1476 + ex->fe_start = block << order;
1477 + ex->fe_group = e3b->bd_group;
1479 + /* calc difference from given start */
1480 + next = next - ex->fe_start;
1481 + ex->fe_len -= next;
1482 + ex->fe_start += next;
1484 + while (needed > ex->fe_len && (buddy = mb_find_buddy(e3b, order, &max))) {
1486 + if (block + 1 >= max)
1489 + next = (block + 1) * (1 << order);
1490 + if (mb_test_bit(next, EXT3_MB_BITMAP(e3b)))
1493 + ord = mb_find_order_for_block(e3b, next);
1496 + block = next >> order;
1497 + ex->fe_len += 1 << order;
1500 + BUG_ON(ex->fe_start + ex->fe_len > (1 << (e3b->bd_blkbits + 3)));
1501 + return ex->fe_len;
1504 +static int mb_mark_used(struct ext3_buddy *e3b, struct ext3_free_extent *ex)
1506 + int ord, mlen = 0, max = 0, cur;
1507 + int start = ex->fe_start;
1508 + int len = ex->fe_len;
1513 + BUG_ON(start + len > (e3b->bd_sb->s_blocksize << 3));
1514 + BUG_ON(e3b->bd_group != ex->fe_group);
1515 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1516 + mb_check_buddy(e3b);
1517 + mb_mark_used_double(e3b, start, len);
1519 + e3b->bd_info->bb_free -= len;
1520 + if (e3b->bd_info->bb_first_free == start)
1521 + e3b->bd_info->bb_first_free += len;
1523 + /* let's maintain fragments counter */
1525 + mlen = !mb_test_bit(start - 1, EXT3_MB_BITMAP(e3b));
1526 + if (start + len < EXT3_SB(e3b->bd_sb)->s_mb_maxs[0])
1527 + max = !mb_test_bit(start + len, EXT3_MB_BITMAP(e3b));
1529 + e3b->bd_info->bb_fragments++;
1530 + else if (!mlen && !max)
1531 + e3b->bd_info->bb_fragments--;
1533 + /* let's maintain buddy itself */
1535 + ord = mb_find_order_for_block(e3b, start);
1537 + if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1538 + /* the whole chunk may be allocated at once! */
1540 + buddy = mb_find_buddy(e3b, ord, &max);
1541 + BUG_ON((start >> ord) >= max);
1542 + mb_set_bit(start >> ord, buddy);
1543 + e3b->bd_info->bb_counters[ord]--;
1550 + /* store for history */
1552 + ret = len | (ord << 16);
1554 + /* we have to split large buddy */
1556 + buddy = mb_find_buddy(e3b, ord, &max);
1557 + mb_set_bit(start >> ord, buddy);
1558 + e3b->bd_info->bb_counters[ord]--;
1561 + cur = (start >> ord) & ~1U;
1562 + buddy = mb_find_buddy(e3b, ord, &max);
1563 + mb_clear_bit(cur, buddy);
1564 + mb_clear_bit(cur + 1, buddy);
1565 + e3b->bd_info->bb_counters[ord]++;
1566 + e3b->bd_info->bb_counters[ord]++;
1569 + mb_set_bits(EXT3_MB_BITMAP(e3b), ex->fe_start, len0);
1570 + mb_check_buddy(e3b);
1576 + * Must be called under group lock!
1578 +static void ext3_mb_use_best_found(struct ext3_allocation_context *ac,
1579 + struct ext3_buddy *e3b)
1581 + unsigned long ret;
1583 + BUG_ON(ac->ac_b_ex.fe_group != e3b->bd_group);
1584 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1586 + ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1587 + ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1588 + ret = mb_mark_used(e3b, &ac->ac_b_ex);
1590 + /* preallocation can change ac_b_ex, thus we store actually
1591 + * allocated blocks for history */
1592 + ac->ac_f_ex = ac->ac_b_ex;
1594 + ac->ac_status = AC_STATUS_FOUND;
1595 + ac->ac_tail = ret & 0xffff;
1596 + ac->ac_buddy = ret >> 16;
1598 + /* XXXXXXX: SUCH A HORRIBLE **CK */
1599 + ac->ac_bitmap_page = e3b->bd_bitmap_page;
1600 + get_page(ac->ac_bitmap_page);
1601 + ac->ac_buddy_page = e3b->bd_buddy_page;
1602 + get_page(ac->ac_buddy_page);
1606 + * regular allocator, for general purposes allocation
1609 +void ext3_mb_check_limits(struct ext3_allocation_context *ac,
1610 + struct ext3_buddy *e3b,
1613 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1614 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1615 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1616 + struct ext3_free_extent ex;
1620 + * We don't want to scan for a whole year
1622 + if (ac->ac_found > sbi->s_mb_max_to_scan &&
1623 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1624 + ac->ac_status = AC_STATUS_BREAK;
1629 + * Haven't found good chunk so far, let's continue
1631 + if (bex->fe_len < gex->fe_len)
1634 + if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1635 + && bex->fe_group == e3b->bd_group) {
1636 + /* recheck chunk's availability - we don't know
1637 + * when it was found (within this lock-unlock
1638 + * period or not) */
1639 + max = mb_find_extent(e3b, 0, bex->fe_start, gex->fe_len, &ex);
1640 + if (max >= gex->fe_len) {
1641 + ext3_mb_use_best_found(ac, e3b);
1648 + * The routine checks whether found extent is good enough. If it is,
1649 + * then the extent gets marked used and flag is set to the context
1650 + * to stop scanning. Otherwise, the extent is compared with the
1651 + * previous found extent and if new one is better, then it's stored
1652 + * in the context. Later, the best found extent will be used, if
1653 + * mballoc can't find good enough extent.
1655 + * FIXME: real allocation policy is to be designed yet!
1657 +static void ext3_mb_measure_extent(struct ext3_allocation_context *ac,
1658 + struct ext3_free_extent *ex,
1659 + struct ext3_buddy *e3b)
1661 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1662 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1664 + BUG_ON(ex->fe_len <= 0);
1665 + BUG_ON(ex->fe_len >= (1 << ac->ac_sb->s_blocksize_bits) * 8);
1666 + BUG_ON(ex->fe_start >= (1 << ac->ac_sb->s_blocksize_bits) * 8);
1667 + BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1672 + * The special case - take what you catch first
1674 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1676 + ext3_mb_use_best_found(ac, e3b);
1681 + * Let's check whether the chuck is good enough
1683 + if (ex->fe_len == gex->fe_len) {
1685 + ext3_mb_use_best_found(ac, e3b);
1690 + * If this is first found extent, just store it in the context
1692 + if (bex->fe_len == 0) {
1698 + * If new found extent is better, store it in the context
1700 + if (bex->fe_len < gex->fe_len) {
1701 + /* if the request isn't satisfied, any found extent
1702 + * larger than previous best one is better */
1703 + if (ex->fe_len > bex->fe_len)
1705 + } else if (ex->fe_len > gex->fe_len) {
1706 + /* if the request is satisfied, then we try to find
1707 + * an extent that still satisfy the request, but is
1708 + * smaller than previous one */
1712 + ext3_mb_check_limits(ac, e3b, 0);
1715 +static int ext3_mb_try_best_found(struct ext3_allocation_context *ac,
1716 + struct ext3_buddy *e3b)
1718 + struct ext3_free_extent ex = ac->ac_b_ex;
1719 + int group = ex.fe_group, max, err;
1721 + BUG_ON(ex.fe_len <= 0);
1722 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1726 + ext3_lock_group(ac->ac_sb, group);
1727 + max = mb_find_extent(e3b, 0, ex.fe_start, ex.fe_len, &ex);
1731 + ext3_mb_use_best_found(ac, e3b);
1734 + ext3_unlock_group(ac->ac_sb, group);
1735 + ext3_mb_release_desc(e3b);
1740 +static int ext3_mb_find_by_goal(struct ext3_allocation_context *ac,
1741 + struct ext3_buddy *e3b)
1743 + int group = ac->ac_g_ex.fe_group, max, err;
1744 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1745 + struct ext3_super_block *es = sbi->s_es;
1746 + struct ext3_free_extent ex;
1748 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1752 + ext3_lock_group(ac->ac_sb, group);
1753 + max = mb_find_extent(e3b, 0, ac->ac_g_ex.fe_start,
1754 + ac->ac_g_ex.fe_len, &ex);
1756 + if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1757 + unsigned long start;
1758 + start = (e3b->bd_group * EXT3_BLOCKS_PER_GROUP(ac->ac_sb) +
1759 + ex.fe_start + le32_to_cpu(es->s_first_data_block));
1760 + if (start % sbi->s_stripe == 0) {
1763 + ext3_mb_use_best_found(ac, e3b);
1765 + } else if (max >= ac->ac_g_ex.fe_len) {
1766 + BUG_ON(ex.fe_len <= 0);
1767 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1768 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1771 + ext3_mb_use_best_found(ac, e3b);
1772 + } else if (max > 0 && (ac->ac_flags & EXT3_MB_HINT_MERGE)) {
1773 + /* Sometimes, caller may want to merge even small
1774 + * number of blocks to an existing extent */
1775 + BUG_ON(ex.fe_len <= 0);
1776 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1777 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1780 + ext3_mb_use_best_found(ac, e3b);
1782 + ext3_unlock_group(ac->ac_sb, group);
1783 + ext3_mb_release_desc(e3b);
1789 + * The routine scans buddy structures (not bitmap!) from given order
1790 + * to max order and tries to find big enough chunk to satisfy the req
1792 +static void ext3_mb_simple_scan_group(struct ext3_allocation_context *ac,
1793 + struct ext3_buddy *e3b)
1795 + struct super_block *sb = ac->ac_sb;
1796 + struct ext3_group_info *grp = e3b->bd_info;
1800 + BUG_ON(ac->ac_2order <= 0);
1801 + for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1802 + if (grp->bb_counters[i] == 0)
1805 + buddy = mb_find_buddy(e3b, i, &max);
1806 + BUG_ON(buddy == NULL);
1808 + k = mb_find_next_zero_bit(buddy, max, 0);
1813 + ac->ac_b_ex.fe_len = 1 << i;
1814 + ac->ac_b_ex.fe_start = k << i;
1815 + ac->ac_b_ex.fe_group = e3b->bd_group;
1817 + ext3_mb_use_best_found(ac, e3b);
1819 + BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1821 + if (EXT3_SB(sb)->s_mb_stats)
1822 + atomic_inc(&EXT3_SB(sb)->s_bal_2orders);
1829 + * The routine scans the group and measures all found extents.
1830 + * In order to optimize scanning, caller must pass number of
1831 + * free blocks in the group, so the routine can know upper limit.
1833 +static void ext3_mb_complex_scan_group(struct ext3_allocation_context *ac,
1834 + struct ext3_buddy *e3b)
1836 + struct super_block *sb = ac->ac_sb;
1837 + void *bitmap = EXT3_MB_BITMAP(e3b);
1838 + struct ext3_free_extent ex;
1841 + free = e3b->bd_info->bb_free;
1842 + BUG_ON(free <= 0);
1844 + i = e3b->bd_info->bb_first_free;
1846 + while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1847 + i = mb_find_next_zero_bit(bitmap, sb->s_blocksize * 8, i);
1848 + if (i >= sb->s_blocksize * 8) {
1849 + BUG_ON(free != 0);
1853 + mb_find_extent(e3b, 0, i, ac->ac_g_ex.fe_len, &ex);
1854 + BUG_ON(ex.fe_len <= 0);
1855 + BUG_ON(free < ex.fe_len);
1857 + ext3_mb_measure_extent(ac, &ex, e3b);
1860 + free -= ex.fe_len;
1863 + ext3_mb_check_limits(ac, e3b, 1);
1867 + * This is a special case for storages like raid5
1868 + * we try to find stripe-aligned chunks for stripe-size requests
1870 +static void ext3_mb_scan_aligned(struct ext3_allocation_context *ac,
1871 + struct ext3_buddy *e3b)
1873 + struct super_block *sb = ac->ac_sb;
1874 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1875 + void *bitmap = EXT3_MB_BITMAP(e3b);
1876 + struct ext3_free_extent ex;
1877 + unsigned long i, max;
1879 + BUG_ON(sbi->s_stripe == 0);
1881 + /* find first stripe-aligned block */
1882 + i = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb)
1883 + + le32_to_cpu(sbi->s_es->s_first_data_block);
1884 + i = ((i + sbi->s_stripe - 1) / sbi->s_stripe) * sbi->s_stripe;
1885 + i = (i - le32_to_cpu(sbi->s_es->s_first_data_block))
1886 + % EXT3_BLOCKS_PER_GROUP(sb);
1888 + while (i < sb->s_blocksize * 8) {
1889 + if (!mb_test_bit(i, bitmap)) {
1890 + max = mb_find_extent(e3b, 0, i, sbi->s_stripe, &ex);
1891 + if (max >= sbi->s_stripe) {
1894 + ext3_mb_use_best_found(ac, e3b);
1898 + i += sbi->s_stripe;
1902 +static int ext3_mb_good_group(struct ext3_allocation_context *ac,
1903 + int group, int cr)
1905 + struct ext3_group_info *grp = EXT3_GROUP_INFO(ac->ac_sb, group);
1906 + unsigned free, fragments, i, bits;
1908 + BUG_ON(cr < 0 || cr >= 4);
1909 + BUG_ON(EXT3_MB_GRP_NEED_INIT(grp));
1911 + free = grp->bb_free;
1912 + fragments = grp->bb_fragments;
1915 + if (fragments == 0)
1920 + BUG_ON(ac->ac_2order == 0);
1921 + bits = ac->ac_sb->s_blocksize_bits + 1;
1922 + for (i = ac->ac_2order; i <= bits; i++)
1923 + if (grp->bb_counters[i] > 0)
1927 + if ((free / fragments) >= ac->ac_g_ex.fe_len)
1931 + if (free >= ac->ac_g_ex.fe_len)
1943 +int ext3_mb_regular_allocator(struct ext3_allocation_context *ac)
1945 + int group, i, cr, err = 0;
1946 + struct ext3_sb_info *sbi;
1947 + struct super_block *sb;
1948 + struct ext3_buddy e3b;
1951 + sbi = EXT3_SB(sb);
1952 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1954 + /* first, try the goal */
1955 + err = ext3_mb_find_by_goal(ac, &e3b);
1956 + if (err || ac->ac_status == AC_STATUS_FOUND)
1959 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
1962 + i = ffs(ac->ac_g_ex.fe_len);
1963 + ac->ac_2order = 0;
1964 + if (i >= sbi->s_mb_order2_reqs) {
1966 + if ((ac->ac_g_ex.fe_len & (~(1 << i))) == 0)
1967 + ac->ac_2order = i;
1970 + group = ac->ac_g_ex.fe_group;
1972 + /* Let's just scan groups to find more-less suitable blocks */
1973 + cr = ac->ac_2order ? 0 : 1;
1975 + for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1976 + ac->ac_criteria = cr;
1977 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; group++, i++) {
1978 + struct ext3_group_info *grp;
1980 + if (group == EXT3_SB(sb)->s_groups_count)
1983 + /* quick check to skip empty groups */
1984 + grp = EXT3_GROUP_INFO(ac->ac_sb, group);
1985 + if (grp->bb_free == 0)
1988 + if (EXT3_MB_GRP_NEED_INIT(EXT3_GROUP_INFO(sb, group))) {
1989 + /* we need full data about the group
1990 + * to make a good selection */
1991 + err = ext3_mb_load_buddy(sb, group, &e3b);
1994 + ext3_mb_release_desc(&e3b);
1997 + /* check is group good for our criteries */
1998 + if (!ext3_mb_good_group(ac, group, cr))
2001 + err = ext3_mb_load_buddy(sb, group, &e3b);
2005 + ext3_lock_group(sb, group);
2006 + if (!ext3_mb_good_group(ac, group, cr)) {
2007 + /* someone did allocation from this group */
2008 + ext3_unlock_group(sb, group);
2009 + ext3_mb_release_desc(&e3b);
2013 + ac->ac_groups_scanned++;
2015 + ext3_mb_simple_scan_group(ac, &e3b);
2016 + else if (cr == 1 && ac->ac_g_ex.fe_len == sbi->s_stripe)
2017 + ext3_mb_scan_aligned(ac, &e3b);
2019 + ext3_mb_complex_scan_group(ac, &e3b);
2021 + ext3_unlock_group(sb, group);
2022 + ext3_mb_release_desc(&e3b);
2024 + if (ac->ac_status != AC_STATUS_CONTINUE)
2029 + if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2030 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
2032 + * We've been searching too long. Let's try to allocate
2033 + * the best chunk we've found so far
2036 + ext3_mb_try_best_found(ac, &e3b);
2037 + if (ac->ac_status != AC_STATUS_FOUND) {
2039 + * Someone more lucky has already allocated it.
2040 + * The only thing we can do is just take first
2042 + printk(KERN_DEBUG "EXT3-fs: someone won our chunk\n");
2044 + ac->ac_b_ex.fe_group = 0;
2045 + ac->ac_b_ex.fe_start = 0;
2046 + ac->ac_b_ex.fe_len = 0;
2047 + ac->ac_status = AC_STATUS_CONTINUE;
2048 + ac->ac_flags |= EXT3_MB_HINT_FIRST;
2050 + atomic_inc(&sbi->s_mb_lost_chunks);
2058 +#ifdef EXT3_MB_HISTORY
2059 +struct ext3_mb_proc_session {
2060 + struct ext3_mb_history *history;
2061 + struct super_block *sb;
2066 +static void *ext3_mb_history_skip_empty(struct ext3_mb_proc_session *s,
2067 + struct ext3_mb_history *hs,
2070 + if (hs == s->history + s->max)
2072 + if (!first && hs == s->history + s->start)
2074 + while (hs->orig.fe_len == 0) {
2076 + if (hs == s->history + s->max)
2078 + if (hs == s->history + s->start)
2084 +static void *ext3_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2086 + struct ext3_mb_proc_session *s = seq->private;
2087 + struct ext3_mb_history *hs;
2091 + return SEQ_START_TOKEN;
2092 + hs = ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2095 + while (--l && (hs = ext3_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2099 +static void *ext3_mb_seq_history_next(struct seq_file *seq, void *v, loff_t *pos)
2101 + struct ext3_mb_proc_session *s = seq->private;
2102 + struct ext3_mb_history *hs = v;
2105 + if (v == SEQ_START_TOKEN)
2106 + return ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2108 + return ext3_mb_history_skip_empty(s, ++hs, 0);
2111 +static int ext3_mb_seq_history_show(struct seq_file *seq, void *v)
2113 + char buf[25], buf2[25], buf3[25], *fmt;
2114 + struct ext3_mb_history *hs = v;
2116 + if (v == SEQ_START_TOKEN) {
2117 + seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2118 + "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2119 + "pid", "inode", "original", "goal", "result","found",
2120 + "grps", "cr", "flags", "merge", "tail", "broken");
2124 + if (hs->op == EXT3_MB_HISTORY_ALLOC) {
2125 + fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2126 + "%-5u %-5s %-5u %-6u\n";
2127 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2128 + hs->result.fe_start, hs->result.fe_len,
2129 + hs->result.fe_logical);
2130 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2131 + hs->orig.fe_start, hs->orig.fe_len,
2132 + hs->orig.fe_logical);
2133 + sprintf(buf3, "%lu/%lu/%lu@%lu", hs->goal.fe_group,
2134 + hs->goal.fe_start, hs->goal.fe_len,
2135 + hs->goal.fe_logical);
2136 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2137 + hs->found, hs->groups, hs->cr, hs->flags,
2138 + hs->merged ? "M" : "", hs->tail,
2139 + hs->buddy ? 1 << hs->buddy : 0);
2140 + } else if (hs->op == EXT3_MB_HISTORY_PREALLOC) {
2141 + fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2142 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2143 + hs->result.fe_start, hs->result.fe_len,
2144 + hs->result.fe_logical);
2145 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2146 + hs->orig.fe_start, hs->orig.fe_len,
2147 + hs->orig.fe_logical);
2148 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2149 + } else if (hs->op == EXT3_MB_HISTORY_DISCARD) {
2150 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2151 + hs->result.fe_start, hs->result.fe_len);
2152 + seq_printf(seq, "%-5u %-8u %-23s discard\n",
2153 + hs->pid, hs->ino, buf2);
2154 + } else if (hs->op == EXT3_MB_HISTORY_FREE) {
2155 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2156 + hs->result.fe_start, hs->result.fe_len);
2157 + seq_printf(seq, "%-5u %-8u %-23s free\n",
2158 + hs->pid, hs->ino, buf2);
2163 +static void ext3_mb_seq_history_stop(struct seq_file *seq, void *v)
2167 +static struct seq_operations ext3_mb_seq_history_ops = {
2168 + .start = ext3_mb_seq_history_start,
2169 + .next = ext3_mb_seq_history_next,
2170 + .stop = ext3_mb_seq_history_stop,
2171 + .show = ext3_mb_seq_history_show,
2174 +static int ext3_mb_seq_history_open(struct inode *inode, struct file *file)
2176 + struct super_block *sb = PDE(inode)->data;
2177 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2178 + struct ext3_mb_proc_session *s;
2181 + s = kmalloc(sizeof(*s), GFP_KERNEL);
2185 + size = sizeof(struct ext3_mb_history) * sbi->s_mb_history_max;
2186 + s->history = kmalloc(size, GFP_KERNEL);
2187 + if (s->history == NULL) {
2192 + spin_lock(&sbi->s_mb_history_lock);
2193 + memcpy(s->history, sbi->s_mb_history, size);
2194 + s->max = sbi->s_mb_history_max;
2195 + s->start = sbi->s_mb_history_cur % s->max;
2196 + spin_unlock(&sbi->s_mb_history_lock);
2198 + rc = seq_open(file, &ext3_mb_seq_history_ops);
2200 + struct seq_file *m = (struct seq_file *)file->private_data;
2203 + kfree(s->history);
2210 +static int ext3_mb_seq_history_release(struct inode *inode, struct file *file)
2212 + struct seq_file *seq = (struct seq_file *)file->private_data;
2213 + struct ext3_mb_proc_session *s = seq->private;
2214 + kfree(s->history);
2216 + return seq_release(inode, file);
2219 +static ssize_t ext3_mb_seq_history_write(struct file *file,
2220 + const char __user *buffer,
2221 + size_t count, loff_t *ppos)
2223 + struct seq_file *seq = (struct seq_file *)file->private_data;
2224 + struct ext3_mb_proc_session *s = seq->private;
2225 + struct super_block *sb = s->sb;
2229 + if (count >= sizeof(str)) {
2230 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2231 + "mb_history", (int)sizeof(str));
2232 + return -EOVERFLOW;
2235 + if (copy_from_user(str, buffer, count))
2238 + value = simple_strtol(str, NULL, 0);
2241 + EXT3_SB(sb)->s_mb_history_filter = value;
2246 +static struct file_operations ext3_mb_seq_history_fops = {
2247 + .owner = THIS_MODULE,
2248 + .open = ext3_mb_seq_history_open,
2250 + .write = ext3_mb_seq_history_write,
2251 + .llseek = seq_lseek,
2252 + .release = ext3_mb_seq_history_release,
2255 +static void *ext3_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2257 + struct super_block *sb = seq->private;
2258 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2261 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2265 + return (void *) group;
2268 +static void *ext3_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2270 + struct super_block *sb = seq->private;
2271 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2275 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2278 + return (void *) group;;
2281 +static int ext3_mb_seq_groups_show(struct seq_file *seq, void *v)
2283 + struct super_block *sb = seq->private;
2284 + long group = (long) v;
2286 + struct ext3_buddy e3b;
2288 + struct ext3_group_info info;
2289 + unsigned short counters[16];
2294 + seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2295 + "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2296 + "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2297 + "group", "free", "frags", "first",
2298 + "2^0", "2^1", "2^2", "2^3", "2^4", "2^5","2^6",
2299 + "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2301 + i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2302 + sizeof(struct ext3_group_info);
2303 + err = ext3_mb_load_buddy(sb, group, &e3b);
2305 + seq_printf(seq, "#%-5lu: I/O error\n", group);
2308 + ext3_lock_group(sb, group);
2309 + memcpy(&sg, EXT3_GROUP_INFO(sb, group), i);
2310 + ext3_unlock_group(sb, group);
2311 + ext3_mb_release_desc(&e3b);
2313 + seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2314 + sg.info.bb_fragments, sg.info.bb_first_free);
2315 + for (i = 0; i <= 13; i++)
2316 + seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2317 + sg.info.bb_counters[i] : 0);
2318 + seq_printf(seq, " ]\n");
2323 +static void ext3_mb_seq_groups_stop(struct seq_file *seq, void *v)
2327 +static struct seq_operations ext3_mb_seq_groups_ops = {
2328 + .start = ext3_mb_seq_groups_start,
2329 + .next = ext3_mb_seq_groups_next,
2330 + .stop = ext3_mb_seq_groups_stop,
2331 + .show = ext3_mb_seq_groups_show,
2334 +static int ext3_mb_seq_groups_open(struct inode *inode, struct file *file)
2336 + struct super_block *sb = PDE(inode)->data;
2339 + rc = seq_open(file, &ext3_mb_seq_groups_ops);
2341 + struct seq_file *m = (struct seq_file *)file->private_data;
2348 +static struct file_operations ext3_mb_seq_groups_fops = {
2349 + .owner = THIS_MODULE,
2350 + .open = ext3_mb_seq_groups_open,
2352 + .llseek = seq_lseek,
2353 + .release = seq_release,
2356 +static void ext3_mb_history_release(struct super_block *sb)
2358 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2360 + remove_proc_entry("mb_groups", sbi->s_mb_proc);
2361 + remove_proc_entry("mb_history", sbi->s_mb_proc);
2363 + if (sbi->s_mb_history)
2364 + kfree(sbi->s_mb_history);
2367 +static void ext3_mb_history_init(struct super_block *sb)
2369 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2372 + if (sbi->s_mb_proc != NULL) {
2373 + struct proc_dir_entry *p;
2374 + p = create_proc_entry("mb_history", S_IRUGO, sbi->s_mb_proc);
2376 + p->proc_fops = &ext3_mb_seq_history_fops;
2379 + p = create_proc_entry("mb_groups", S_IRUGO, sbi->s_mb_proc);
2381 + p->proc_fops = &ext3_mb_seq_groups_fops;
2386 + sbi->s_mb_history_max = 1000;
2387 + sbi->s_mb_history_cur = 0;
2388 + spin_lock_init(&sbi->s_mb_history_lock);
2389 + i = sbi->s_mb_history_max * sizeof(struct ext3_mb_history);
2390 + sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
2391 + memset(sbi->s_mb_history, 0, i);
2392 + /* if we can't allocate history, then we simple won't use it */
2396 +ext3_mb_store_history(struct ext3_allocation_context *ac)
2398 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
2399 + struct ext3_mb_history h;
2401 + if (likely(sbi->s_mb_history == NULL))
2404 + if (!(ac->ac_op & sbi->s_mb_history_filter))
2408 + h.pid = current->pid;
2409 + h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2410 + h.orig = ac->ac_o_ex;
2411 + h.result = ac->ac_b_ex;
2412 + h.flags = ac->ac_flags;
2414 + if (ac->ac_op == EXT3_MB_HISTORY_ALLOC) {
2415 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2416 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2418 + h.goal = ac->ac_g_ex;
2419 + h.result = ac->ac_f_ex;
2422 + spin_lock(&sbi->s_mb_history_lock);
2423 + memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2424 + if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2425 + sbi->s_mb_history_cur = 0;
2426 + spin_unlock(&sbi->s_mb_history_lock);
2430 +#define ext3_mb_history_release(sb)
2431 +#define ext3_mb_history_init(sb)
2434 +int ext3_mb_init_backend(struct super_block *sb)
2436 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2437 + int i, j, len, metalen;
2438 + int num_meta_group_infos =
2439 + (sbi->s_groups_count + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2440 + EXT3_DESC_PER_BLOCK_BITS(sb);
2441 + struct ext3_group_info **meta_group_info;
2443 + /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2444 + * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2445 + * So a two level scheme suffices for now. */
2446 + sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
2447 + num_meta_group_infos, GFP_KERNEL);
2448 + if (sbi->s_group_info == NULL) {
2449 + printk(KERN_ERR "EXT3-fs: can't allocate buddy meta group\n");
2452 + sbi->s_buddy_cache = new_inode(sb);
2453 + if (sbi->s_buddy_cache == NULL) {
2454 + printk(KERN_ERR "EXT3-fs: can't get new inode\n");
2457 + EXT3_I(sbi->s_buddy_cache)->i_disksize = 0;
2459 + metalen = sizeof(*meta_group_info) << EXT3_DESC_PER_BLOCK_BITS(sb);
2460 + for (i = 0; i < num_meta_group_infos; i++) {
2461 + if ((i + 1) == num_meta_group_infos)
2462 + metalen = sizeof(*meta_group_info) *
2463 + (sbi->s_groups_count -
2464 + (i << EXT3_DESC_PER_BLOCK_BITS(sb)));
2465 + meta_group_info = kmalloc(metalen, GFP_KERNEL);
2466 + if (meta_group_info == NULL) {
2467 + printk(KERN_ERR "EXT3-fs: can't allocate mem for a "
2469 + goto err_freemeta;
2471 + sbi->s_group_info[i] = meta_group_info;
2475 + * calculate needed size. if change bb_counters size,
2476 + * don't forget about ext3_mb_generate_buddy()
2478 + len = sizeof(struct ext3_group_info);
2479 + len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
2480 + for (i = 0; i < sbi->s_groups_count; i++) {
2481 + struct ext3_group_desc * desc;
2484 + sbi->s_group_info[i >> EXT3_DESC_PER_BLOCK_BITS(sb)];
2485 + j = i & (EXT3_DESC_PER_BLOCK(sb) - 1);
2487 + meta_group_info[j] = kmalloc(len, GFP_KERNEL);
2488 + if (meta_group_info[j] == NULL) {
2489 + printk(KERN_ERR "EXT3-fs: can't allocate buddy mem\n");
2491 + goto err_freebuddy;
2493 + desc = ext3_get_group_desc(sb, i, NULL);
2494 + if (desc == NULL) {
2495 + printk(KERN_ERR"EXT3-fs: can't read descriptor %u\n",i);
2496 + goto err_freebuddy;
2498 + memset(meta_group_info[j], 0, len);
2499 + set_bit(EXT3_GROUP_INFO_NEED_INIT_BIT,
2500 + &meta_group_info[j]->bb_state);
2502 + /* initialize bb_free to be able to skip
2503 + * empty groups without initialization */
2504 + meta_group_info[j]->bb_free =
2505 + le16_to_cpu(desc->bg_free_blocks_count);
2507 + INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
2509 +#ifdef DOUBLE_CHECK
2511 + struct buffer_head *bh;
2512 + meta_group_info[j]->bb_bitmap =
2513 + kmalloc(sb->s_blocksize, GFP_KERNEL);
2514 + BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
2515 + bh = read_block_bitmap(sb, i);
2516 + BUG_ON(bh == NULL);
2517 + memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
2529 + kfree(EXT3_GROUP_INFO(sb, i));
2532 + i = num_meta_group_infos;
2535 + kfree(sbi->s_group_info[i]);
2536 + iput(sbi->s_buddy_cache);
2538 + kfree(sbi->s_group_info);
2542 +int ext3_mb_init(struct super_block *sb, int needs_recovery)
2544 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2545 + unsigned i, offset, max;
2547 + if (!test_opt(sb, MBALLOC))
2550 + i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2552 + sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2553 + if (sbi->s_mb_offsets == NULL) {
2554 + clear_opt(sbi->s_mount_opt, MBALLOC);
2557 + sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2558 + if (sbi->s_mb_maxs == NULL) {
2559 + clear_opt(sbi->s_mount_opt, MBALLOC);
2560 + kfree(sbi->s_mb_maxs);
2564 + /* order 0 is regular bitmap */
2565 + sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2566 + sbi->s_mb_offsets[0] = 0;
2570 + max = sb->s_blocksize << 2;
2572 + sbi->s_mb_offsets[i] = offset;
2573 + sbi->s_mb_maxs[i] = max;
2574 + offset += 1 << (sb->s_blocksize_bits - i);
2577 + } while (i <= sb->s_blocksize_bits + 1);
2579 + /* init file for buddy data */
2580 + if ((i = ext3_mb_init_backend(sb))) {
2581 + clear_opt(sbi->s_mount_opt, MBALLOC);
2582 + kfree(sbi->s_mb_offsets);
2583 + kfree(sbi->s_mb_maxs);
2587 + spin_lock_init(&sbi->s_md_lock);
2588 + INIT_LIST_HEAD(&sbi->s_active_transaction);
2589 + INIT_LIST_HEAD(&sbi->s_closed_transaction);
2590 + INIT_LIST_HEAD(&sbi->s_committed_transaction);
2591 + spin_lock_init(&sbi->s_bal_lock);
2593 + sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2594 + sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2595 + sbi->s_mb_max_groups_to_scan = MB_DEFAULT_MAX_GROUPS_TO_SCAN;
2596 + sbi->s_mb_stats = MB_DEFAULT_STATS;
2597 + sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2598 + sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2599 + sbi->s_mb_history_filter = EXT3_MB_HISTORY_DEFAULT;
2601 + i = sizeof(struct ext3_locality_group) * NR_CPUS;
2602 + sbi->s_locality_groups = kmalloc(i, GFP_NOFS);
2603 + if (sbi->s_locality_groups == NULL) {
2604 + clear_opt(sbi->s_mount_opt, MBALLOC);
2605 + kfree(sbi->s_mb_offsets);
2606 + kfree(sbi->s_mb_maxs);
2609 + for (i = 0; i < NR_CPUS; i++) {
2610 + struct ext3_locality_group *lg;
2611 + lg = &sbi->s_locality_groups[i];
2612 + sema_init(&lg->lg_sem, 1);
2613 + INIT_LIST_HEAD(&lg->lg_prealloc_list);
2614 + spin_lock_init(&lg->lg_prealloc_lock);
2617 + ext3_mb_init_per_dev_proc(sb);
2618 + ext3_mb_history_init(sb);
2620 + printk("EXT3-fs: mballoc enabled\n");
2624 +void ext3_mb_cleanup_pa(struct ext3_group_info *grp)
2626 + struct ext3_prealloc_space *pa;
2627 + struct list_head *cur, *tmp;
2630 + list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2631 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
2632 + list_del_rcu(&pa->pa_group_list);
2637 + mb_debug("mballoc: %u PAs left\n", count);
2641 +int ext3_mb_release(struct super_block *sb)
2643 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2644 + int i, num_meta_group_infos;
2646 + if (!test_opt(sb, MBALLOC))
2649 + /* release freed, non-committed blocks */
2650 + spin_lock(&sbi->s_md_lock);
2651 + list_splice_init(&sbi->s_closed_transaction,
2652 + &sbi->s_committed_transaction);
2653 + list_splice_init(&sbi->s_active_transaction,
2654 + &sbi->s_committed_transaction);
2655 + spin_unlock(&sbi->s_md_lock);
2656 + ext3_mb_free_committed_blocks(sb);
2658 + if (sbi->s_group_info) {
2659 + for (i = 0; i < sbi->s_groups_count; i++) {
2660 +#ifdef DOUBLE_CHECK
2661 + if (EXT3_GROUP_INFO(sb, i)->bb_bitmap)
2662 + kfree(EXT3_GROUP_INFO(sb, i)->bb_bitmap);
2664 + ext3_mb_cleanup_pa(EXT3_GROUP_INFO(sb, i));
2665 + kfree(EXT3_GROUP_INFO(sb, i));
2667 + num_meta_group_infos = (sbi->s_groups_count +
2668 + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2669 + EXT3_DESC_PER_BLOCK_BITS(sb);
2670 + for (i = 0; i < num_meta_group_infos; i++)
2671 + kfree(sbi->s_group_info[i]);
2672 + kfree(sbi->s_group_info);
2674 + if (sbi->s_mb_offsets)
2675 + kfree(sbi->s_mb_offsets);
2676 + if (sbi->s_mb_maxs)
2677 + kfree(sbi->s_mb_maxs);
2678 + if (sbi->s_buddy_cache)
2679 + iput(sbi->s_buddy_cache);
2680 + if (sbi->s_mb_stats) {
2681 + printk("EXT3-fs: mballoc: %u blocks %u reqs (%u success)\n",
2682 + atomic_read(&sbi->s_bal_allocated),
2683 + atomic_read(&sbi->s_bal_reqs),
2684 + atomic_read(&sbi->s_bal_success));
2685 + printk("EXT3-fs: mballoc: %u extents scanned, %u goal hits, "
2686 + "%u 2^N hits, %u breaks, %u lost\n",
2687 + atomic_read(&sbi->s_bal_ex_scanned),
2688 + atomic_read(&sbi->s_bal_goals),
2689 + atomic_read(&sbi->s_bal_2orders),
2690 + atomic_read(&sbi->s_bal_breaks),
2691 + atomic_read(&sbi->s_mb_lost_chunks));
2692 + printk("EXT3-fs: mballoc: %lu generated and it took %Lu\n",
2693 + sbi->s_mb_buddies_generated++,
2694 + sbi->s_mb_generation_time);
2695 + printk("EXT3-fs: mballoc: %u preallocated, %u discarded\n",
2696 + atomic_read(&sbi->s_mb_preallocated),
2697 + atomic_read(&sbi->s_mb_discarded));
2700 + if (sbi->s_locality_groups)
2701 + kfree(sbi->s_locality_groups);
2703 + ext3_mb_history_release(sb);
2704 + ext3_mb_destroy_per_dev_proc(sb);
2709 +void ext3_mb_free_committed_blocks(struct super_block *sb)
2711 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2712 + int err, i, count = 0, count2 = 0;
2713 + struct ext3_free_metadata *md;
2714 + struct ext3_buddy e3b;
2716 + if (list_empty(&sbi->s_committed_transaction))
2719 + /* there is committed blocks to be freed yet */
2721 + /* get next array of blocks */
2723 + spin_lock(&sbi->s_md_lock);
2724 + if (!list_empty(&sbi->s_committed_transaction)) {
2725 + md = list_entry(sbi->s_committed_transaction.next,
2726 + struct ext3_free_metadata, list);
2727 + list_del(&md->list);
2729 + spin_unlock(&sbi->s_md_lock);
2734 + mb_debug("gonna free %u blocks in group %u (0x%p):",
2735 + md->num, md->group, md);
2737 + err = ext3_mb_load_buddy(sb, md->group, &e3b);
2738 + /* we expect to find existing buddy because it's pinned */
2741 + /* there are blocks to put in buddy to make them really free */
2744 + ext3_lock_group(sb, md->group);
2745 + for (i = 0; i < md->num; i++) {
2746 + mb_debug(" %u", md->blocks[i]);
2747 + err = mb_free_blocks(NULL, &e3b, md->blocks[i], 1);
2751 + ext3_unlock_group(sb, md->group);
2753 + /* balance refcounts from ext3_mb_free_metadata() */
2754 + page_cache_release(e3b.bd_buddy_page);
2755 + page_cache_release(e3b.bd_bitmap_page);
2758 + ext3_mb_release_desc(&e3b);
2762 + mb_debug("freed %u blocks in %u structures\n", count, count2);
2765 +#define EXT3_ROOT "ext3"
2766 +#define EXT3_MB_STATS_NAME "stats"
2767 +#define EXT3_MB_MAX_TO_SCAN_NAME "max_to_scan"
2768 +#define EXT3_MB_MIN_TO_SCAN_NAME "min_to_scan"
2769 +#define EXT3_MB_ORDER2_REQ "order2_req"
2770 +#define EXT3_MB_STREAM_REQ "stream_req"
2772 +static int ext3_mb_stats_read(char *page, char **start, off_t off,
2773 + int count, int *eof, void *data)
2775 + struct ext3_sb_info *sbi = data;
2782 + len = sprintf(page, "%ld\n", sbi->s_mb_stats);
2787 +static int ext3_mb_stats_write(struct file *file, const char *buffer,
2788 + unsigned long count, void *data)
2790 + struct ext3_sb_info *sbi = data;
2793 + if (count >= sizeof(str)) {
2794 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2795 + EXT3_MB_STATS_NAME, (int)sizeof(str));
2796 + return -EOVERFLOW;
2799 + if (copy_from_user(str, buffer, count))
2802 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2803 + sbi->s_mb_stats = (simple_strtol(str, NULL, 0) != 0);
2807 +static int ext3_mb_max_to_scan_read(char *page, char **start, off_t off,
2808 + int count, int *eof, void *data)
2810 + struct ext3_sb_info *sbi = data;
2817 + len = sprintf(page, "%ld\n", sbi->s_mb_max_to_scan);
2822 +static int ext3_mb_max_to_scan_write(struct file *file, const char *buffer,
2823 + unsigned long count, void *data)
2825 + struct ext3_sb_info *sbi = data;
2829 + if (count >= sizeof(str)) {
2830 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2831 + EXT3_MB_MAX_TO_SCAN_NAME, (int)sizeof(str));
2832 + return -EOVERFLOW;
2835 + if (copy_from_user(str, buffer, count))
2838 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2839 + value = simple_strtol(str, NULL, 0);
2843 + sbi->s_mb_max_to_scan = value;
2848 +static int ext3_mb_min_to_scan_read(char *page, char **start, off_t off,
2849 + int count, int *eof, void *data)
2851 + struct ext3_sb_info *sbi = data;
2858 + len = sprintf(page, "%ld\n", sbi->s_mb_min_to_scan);
2863 +static int ext3_mb_order2_req_write(struct file *file, const char *buffer,
2864 + unsigned long count, void *data)
2866 + struct ext3_sb_info *sbi = data;
2870 + if (count >= sizeof(str)) {
2871 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2872 + EXT3_MB_MIN_TO_SCAN_NAME, (int)sizeof(str));
2873 + return -EOVERFLOW;
2876 + if (copy_from_user(str, buffer, count))
2879 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2880 + value = simple_strtol(str, NULL, 0);
2884 + sbi->s_mb_order2_reqs = value;
2889 +static int ext3_mb_order2_req_read(char *page, char **start, off_t off,
2890 + int count, int *eof, void *data)
2892 + struct ext3_sb_info *sbi = data;
2899 + len = sprintf(page, "%ld\n", sbi->s_mb_order2_reqs);
2904 +static int ext3_mb_min_to_scan_write(struct file *file, const char *buffer,
2905 + unsigned long count, void *data)
2907 + struct ext3_sb_info *sbi = data;
2911 + if (count >= sizeof(str)) {
2912 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2913 + EXT3_MB_MIN_TO_SCAN_NAME, (int)sizeof(str));
2914 + return -EOVERFLOW;
2917 + if (copy_from_user(str, buffer, count))
2920 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2921 + value = simple_strtol(str, NULL, 0);
2925 + sbi->s_mb_min_to_scan = value;
2930 +static int ext3_mb_stream_req_read(char *page, char **start, off_t off,
2931 + int count, int *eof, void *data)
2933 + struct ext3_sb_info *sbi = data;
2940 + len = sprintf(page, "%ld\n", sbi->s_mb_stream_request);
2945 +static int ext3_mb_stream_req_write(struct file *file, const char *buffer,
2946 + unsigned long count, void *data)
2948 + struct ext3_sb_info *sbi = data;
2952 + if (count >= sizeof(str)) {
2953 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2954 + EXT3_MB_STREAM_REQ, (int)sizeof(str));
2955 + return -EOVERFLOW;
2958 + if (copy_from_user(str, buffer, count))
2961 + /* Only set to 0 or 1 respectively; zero->0; non-zero->1 */
2962 + value = simple_strtol(str, NULL, 0);
2966 + sbi->s_mb_stream_request = value;
2971 +int ext3_mb_init_per_dev_proc(struct super_block *sb)
2973 + mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2974 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2975 + struct proc_dir_entry *proc;
2976 + char devname[64], *name;
2978 + snprintf(devname, sizeof(devname) - 1, "%s",
2979 + bdevname(sb->s_bdev, devname));
2980 + sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext3);
2982 + name = EXT3_MB_STATS_NAME;
2983 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2987 + proc->read_proc = ext3_mb_stats_read;
2988 + proc->write_proc = ext3_mb_stats_write;
2990 + name = EXT3_MB_MAX_TO_SCAN_NAME;
2991 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
2995 + proc->read_proc = ext3_mb_max_to_scan_read;
2996 + proc->write_proc = ext3_mb_max_to_scan_write;
2998 + name = EXT3_MB_MIN_TO_SCAN_NAME;
2999 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
3003 + proc->read_proc = ext3_mb_min_to_scan_read;
3004 + proc->write_proc = ext3_mb_min_to_scan_write;
3006 + name = EXT3_MB_ORDER2_REQ;
3007 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
3011 + proc->read_proc = ext3_mb_order2_req_read;
3012 + proc->write_proc = ext3_mb_order2_req_write;
3014 + name = EXT3_MB_STREAM_REQ;
3015 + proc = create_proc_entry(name, mode, sbi->s_mb_proc);
3019 + proc->read_proc = ext3_mb_stream_req_read;
3020 + proc->write_proc = ext3_mb_stream_req_write;
3025 + printk(KERN_ERR "EXT3-fs: Unable to create %s\n", name);
3026 + remove_proc_entry(EXT3_MB_STREAM_REQ, sbi->s_mb_proc);
3027 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_mb_proc);
3028 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
3029 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
3030 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_mb_proc);
3031 + remove_proc_entry(devname, proc_root_ext3);
3032 + sbi->s_mb_proc = NULL;
3037 +int ext3_mb_destroy_per_dev_proc(struct super_block *sb)
3039 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3042 + if (sbi->s_mb_proc == NULL)
3045 + snprintf(devname, sizeof(devname) - 1, "%s",
3046 + bdevname(sb->s_bdev, devname));
3047 + remove_proc_entry(EXT3_MB_STREAM_REQ, sbi->s_mb_proc);
3048 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_mb_proc);
3049 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
3050 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
3051 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_mb_proc);
3052 + remove_proc_entry(devname, proc_root_ext3);
3057 +int __init init_ext3_proc(void)
3059 + ext3_pspace_cachep =
3060 + kmem_cache_create("ext3_prealloc_space",
3061 + sizeof(struct ext3_prealloc_space),
3062 + 0, SLAB_RECLAIM_ACCOUNT, NULL, NULL);
3063 + if (ext3_pspace_cachep == NULL)
3066 + proc_root_ext3 = proc_mkdir(EXT3_ROOT, proc_root_fs);
3067 + if (proc_root_ext3 == NULL)
3068 + printk(KERN_ERR "EXT3-fs: Unable to create %s\n", EXT3_ROOT);
3073 +void exit_ext3_proc(void)
3075 + /* XXX: synchronize_rcu(); */
3076 + kmem_cache_destroy(ext3_pspace_cachep);
3077 + remove_proc_entry(EXT3_ROOT, proc_root_fs);
3082 + * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
3083 + * Returns 0 if success or error code
3085 +int ext3_mb_mark_diskspace_used(struct ext3_allocation_context *ac, handle_t *handle)
3087 + struct buffer_head *bitmap_bh = NULL;
3088 + struct ext3_super_block *es;
3089 + struct ext3_group_desc *gdp;
3090 + struct buffer_head *gdp_bh;
3091 + struct ext3_sb_info *sbi;
3092 + struct super_block *sb;
3096 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3097 + BUG_ON(ac->ac_b_ex.fe_len <= 0);
3100 + sbi = EXT3_SB(sb);
3103 + ext3_debug("using block group %d(%d)\n", ac->ac_b_group.group,
3104 + gdp->bg_free_blocks_count);
3106 + /* time to check quota, we can't do this before because
3107 + * having quota spent on preallocated-unused-yet blocks
3108 + * would be wrong */
3109 + len = ac->ac_b_ex.fe_len;
3110 + while (len && DQUOT_ALLOC_BLOCK(ac->ac_inode, len)) len--;
3111 + if (ac->ac_b_ex.fe_len != len) {
3112 + /* some blocks can't be allocated due to quota
3113 + * we have to return them back */
3121 + bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3125 + err = ext3_journal_get_write_access(handle, bitmap_bh);
3130 + gdp = ext3_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3134 + err = ext3_journal_get_write_access(handle, gdp_bh);
3138 + block = ac->ac_b_ex.fe_group * EXT3_BLOCKS_PER_GROUP(sb)
3139 + + ac->ac_b_ex.fe_start
3140 + + le32_to_cpu(es->s_first_data_block);
3142 + if (block == le32_to_cpu(gdp->bg_block_bitmap) ||
3143 + block == le32_to_cpu(gdp->bg_inode_bitmap) ||
3144 + in_range(block, le32_to_cpu(gdp->bg_inode_table),
3145 + EXT3_SB(sb)->s_itb_per_group))
3146 + ext3_error(sb, __FUNCTION__,
3147 + "Allocating block in system zone - block = %lu",
3148 + (unsigned long) block);
3149 +#ifdef AGGRESSIVE_CHECK
3152 + for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3153 + BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3154 + bitmap_bh->b_data));
3158 + mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
3160 + spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3161 + gdp->bg_free_blocks_count =
3162 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
3163 + - ac->ac_b_ex.fe_len);
3164 + spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3165 + percpu_counter_mod(&sbi->s_freeblocks_counter, - ac->ac_b_ex.fe_len);
3167 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
3170 + err = ext3_journal_dirty_metadata(handle, gdp_bh);
3174 + brelse(bitmap_bh);
3179 + * here we normalize request for locality group
3180 + * XXX: should we try to preallocate more than the group has now?
3182 +void ext3_mb_normalize_group_request(struct ext3_allocation_context *ac)
3184 + struct super_block *sb = ac->ac_sb;
3185 + struct ext3_locality_group *lg = ac->ac_lg;
3187 + BUG_ON(lg == NULL);
3188 + if (EXT3_SB(sb)->s_stripe)
3189 + ac->ac_g_ex.fe_len = EXT3_SB(sb)->s_stripe;
3191 + ac->ac_g_ex.fe_len = (1024 * 1024) >> sb->s_blocksize_bits;
3193 + mb_debug("#%u: goal %u blocks for locality group\n",
3194 + current->pid, ac->ac_g_ex.fe_len);
3198 + * Normalization means making request better in terms of
3199 + * size and alignment
3201 +void ext3_mb_normalize_request(struct ext3_allocation_context *ac,
3202 + struct ext3_allocation_request *ar)
3204 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3205 + loff_t start, end, size, orig_size, orig_start;
3206 + struct list_head *cur;
3209 + /* do normalize only data requests, metadata requests
3210 + do not need preallocation */
3211 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3214 + /* sometime caller may want exact blocks */
3215 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
3218 + /* caller may indicate that preallocation isn't
3219 + * required (it's a tail, for example) */
3220 + if (ac->ac_flags & EXT3_MB_HINT_NOPREALLOC)
3223 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3224 + return ext3_mb_normalize_group_request(ac);
3226 + bsbits = ac->ac_sb->s_blocksize_bits;
3228 + /* first, let's learn actual file size
3229 + * given current request is allocated */
3230 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3231 + size = size << bsbits;
3232 + if (size < i_size_read(ac->ac_inode))
3233 + size = i_size_read(ac->ac_inode);
3235 + /* first, try to predict filesize */
3236 + /* XXX: should this table be tunable? */
3238 + if (size <= 16 * 1024) {
3240 + } else if (size <= 32 * 1024) {
3242 + } else if (size <= 64 * 1024) {
3244 + } else if (size <= 128 * 1024) {
3245 + size = 128 * 1024;
3246 + } else if (size <= 256 * 1024) {
3247 + size = 256 * 1024;
3248 + } else if (size <= 512 * 1024) {
3249 + size = 512 * 1024;
3250 + } else if (size <= 1024 * 1024) {
3251 + size = 1024 * 1024;
3252 + } else if (size < 4 * 1024 * 1024) {
3253 + start = ac->ac_o_ex.fe_logical << bsbits;
3254 + start = (start / (1024 * 1024)) * (1024 * 1024);
3255 + size = 1024 * 1024;
3256 + } else if (size < 8 * 1024 * 1024) {
3257 + start = ac->ac_o_ex.fe_logical << bsbits;
3258 + start = (start / (4 * (1024 * 1024))) * 4 * (1024 * 1024);
3259 + size = 4 * 1024 * 1024;
3260 + } else if (ac->ac_o_ex.fe_len < ((8 << 20) >> bsbits)) {
3261 + start = ac->ac_o_ex.fe_logical;
3262 + start = start << bsbits;
3263 + start = (start / (8 * (1024 * 1024))) * 8 * (1024 * 1024);
3264 + size = 8 * 1024 * 1024;
3266 + start = ac->ac_o_ex.fe_logical;
3267 + start = start << bsbits;
3268 + size = ac->ac_o_ex.fe_len << bsbits;
3270 + orig_size = size = size >> bsbits;
3271 + orig_start = start = start >> bsbits;
3273 + /* don't cover already allocated blocks in selected range */
3274 + if (ar->pleft && start <= ar->lleft) {
3275 + size -= ar->lleft + 1 - start;
3276 + start = ar->lleft + 1;
3278 + if (ar->pright && start + size - 1 >= ar->lright)
3279 + size -= start + size - ar->lright;
3281 + end = start + size;
3283 + /* check we don't cross already preallocated blocks */
3285 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3286 + struct ext3_prealloc_space *pa;
3287 + unsigned long pa_end;
3289 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3290 + pa_end = pa->pa_lstart + pa->pa_len;
3292 + /* PA must not overlap original request */
3293 + BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3294 + ac->ac_o_ex.fe_logical < pa->pa_lstart));
3296 + /* skip PA normalized request doesn't overlap with */
3297 + if (pa->pa_lstart >= end)
3299 + if (pa_end <= start)
3301 + BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3303 + if (pa_end <= ac->ac_o_ex.fe_logical) {
3304 + BUG_ON(pa_end < start);
3308 + if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3309 + BUG_ON(pa->pa_lstart > end);
3310 + end = pa->pa_lstart;
3313 + rcu_read_unlock();
3314 + size = end - start;
3316 + /* XXX: extra loop to check we really don't overlap preallocations */
3318 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3319 + struct ext3_prealloc_space *pa;
3320 + unsigned long pa_end;
3321 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3322 + pa_end = pa->pa_lstart + pa->pa_len;
3323 + BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3325 + rcu_read_unlock();
3327 + if (start + size <= ac->ac_o_ex.fe_logical &&
3328 + start > ac->ac_o_ex.fe_logical) {
3329 + printk("start %lu, size %lu, fe_logical %lu\n",
3330 + (unsigned long) start, (unsigned long) size,
3331 + (unsigned long) ac->ac_o_ex.fe_logical);
3333 + BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3334 + start > ac->ac_o_ex.fe_logical);
3336 + /* now prepare goal request */
3337 + BUG_ON(size <= 0 || size >= EXT3_BLOCKS_PER_GROUP(ac->ac_sb));
3338 + if (size < ac->ac_o_ex.fe_len) {
3339 + /* XXX: don't normalize tails? */
3342 + /* XXX: is it better to align blocks WRT to logical placement
3343 + * or satisfy big request as is */
3344 + ac->ac_g_ex.fe_logical = start;
3345 + ac->ac_g_ex.fe_len = size;
3347 + mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3348 + (unsigned) orig_size, (unsigned) start);
3351 +void ext3_mb_collect_stats(struct ext3_allocation_context *ac)
3353 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
3355 + if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3356 + atomic_inc(&sbi->s_bal_reqs);
3357 + atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3358 + if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3359 + atomic_inc(&sbi->s_bal_success);
3360 + atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3361 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3362 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3363 + atomic_inc(&sbi->s_bal_goals);
3364 + if (ac->ac_found > sbi->s_mb_max_to_scan)
3365 + atomic_inc(&sbi->s_bal_breaks);
3368 + ext3_mb_store_history(ac);
3372 + * use blocks preallocated to inode
3374 +void ext3_mb_use_inode_pa(struct ext3_allocation_context *ac,
3375 + struct ext3_prealloc_space *pa)
3377 + unsigned long start, len;
3379 + /* found preallocated blocks, use them */
3380 + start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3381 + len = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3382 + len = len - start;
3383 + ext3_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3384 + &ac->ac_b_ex.fe_start);
3385 + ac->ac_b_ex.fe_len = len;
3386 + ac->ac_status = AC_STATUS_FOUND;
3389 + BUG_ON(start < pa->pa_pstart);
3390 + BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3391 + BUG_ON(pa->pa_free < len);
3392 + pa->pa_free -= len;
3394 + mb_debug("use %lu/%lu from inode pa %p\n", start, len, pa);
3398 + * use blocks preallocated to locality group
3400 +void ext3_mb_use_group_pa(struct ext3_allocation_context *ac,
3401 + struct ext3_prealloc_space *pa)
3403 + unsigned len = ac->ac_o_ex.fe_len;
3405 + ext3_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3406 + &ac->ac_b_ex.fe_group,
3407 + &ac->ac_b_ex.fe_start);
3408 + ac->ac_b_ex.fe_len = len;
3409 + ac->ac_status = AC_STATUS_FOUND;
3412 + /* we don't correct pa_pstart or pa_plen here to avoid
3413 + * possible race when tte group is being loaded concurrently
3414 + * instead we correct pa later, after blocks are marked
3415 + * in on-disk bitmap -- see ext3_mb_release_context() */
3416 + mb_debug("use %lu/%lu from group pa %p\n", pa->pa_lstart-len, len, pa);
3420 + * search goal blocks in preallocated space
3422 +int ext3_mb_use_preallocated(struct ext3_allocation_context *ac)
3424 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3425 + struct ext3_locality_group *lg;
3426 + struct ext3_prealloc_space *pa;
3427 + struct list_head *cur;
3429 + /* only data can be preallocated */
3430 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3433 + /* first, try per-file preallocation */
3435 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3436 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3438 + /* all fields in this condition don't change,
3439 + * so we can skip locking for them */
3440 + if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3441 + ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3444 + /* found preallocated blocks, use them */
3445 + spin_lock(&pa->pa_lock);
3446 + if (pa->pa_deleted == 0 && pa->pa_free) {
3447 + atomic_inc(&pa->pa_count);
3448 + ext3_mb_use_inode_pa(ac, pa);
3449 + spin_unlock(&pa->pa_lock);
3450 + ac->ac_criteria = 10;
3451 + rcu_read_unlock();
3454 + spin_unlock(&pa->pa_lock);
3456 + rcu_read_unlock();
3458 + /* can we use group allocation? */
3459 + if (!(ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC))
3462 + /* inode may have no locality group for some reason */
3468 + list_for_each_rcu(cur, &lg->lg_prealloc_list) {
3469 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3470 + spin_lock(&pa->pa_lock);
3471 + if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3472 + atomic_inc(&pa->pa_count);
3473 + ext3_mb_use_group_pa(ac, pa);
3474 + spin_unlock(&pa->pa_lock);
3475 + ac->ac_criteria = 20;
3476 + rcu_read_unlock();
3479 + spin_unlock(&pa->pa_lock);
3481 + rcu_read_unlock();
3487 + * the function goes through all preallocation in this group and marks them
3488 + * used in in-core bitmap. buddy must be generated from this bitmap
3490 +void ext3_mb_generate_from_pa(struct super_block *sb, void *bitmap, int group)
3492 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
3493 + struct ext3_prealloc_space *pa;
3494 + struct list_head *cur;
3495 + unsigned long groupnr;
3496 + unsigned long start;
3497 + int preallocated = 0, count = 0, len;
3499 + /* all form of preallocation discards first load group,
3500 + * so the only competing code is preallocation use.
3501 + * we don't need any locking here
3502 + * notice we do NOT ignore preallocations with pa_deleted
3503 + * otherwise we could leave used blocks available for
3504 + * allocation in buddy when concurrent ext3_mb_put_pa()
3505 + * is dropping preallocation
3507 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
3508 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
3509 + spin_lock(&pa->pa_lock);
3510 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &start);
3512 + spin_unlock(&pa->pa_lock);
3513 + BUG_ON(groupnr != group);
3514 + mb_set_bits(bitmap, start, len);
3515 + preallocated += len;
3518 + mb_debug("prellocated %u for group %u\n", preallocated, group);
3521 +#if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,5)
3522 +static void ext3_mb_pa_callback(struct rcu_head *head)
3524 + struct ext3_prealloc_space *pa;
3525 + pa = container_of(head, struct ext3_prealloc_space, u.pa_rcu);
3526 + kmem_cache_free(ext3_pspace_cachep, pa);
3528 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback)
3530 +static void ext3_mb_pa_callback(void *pa)
3532 + kmem_cache_free(ext3_pspace_cachep, pa);
3534 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback, pa)
3538 + * drops a reference to preallocated space descriptor
3539 + * if this was the last reference and the space is consumed
3541 +void ext3_mb_put_pa(struct ext3_allocation_context *ac,
3542 + struct super_block *sb, struct ext3_prealloc_space *pa)
3544 + unsigned long grp;
3546 + if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3549 + /* in this short window concurrent discard can set pa_deleted */
3550 + spin_lock(&pa->pa_lock);
3551 + if (pa->pa_deleted == 0) {
3552 + spin_unlock(&pa->pa_lock);
3556 + pa->pa_deleted = 1;
3557 + spin_unlock(&pa->pa_lock);
3559 + /* -1 is to protect from crossing allocation group */
3560 + ext3_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3565 + * P1 (buddy init) P2 (regular allocation)
3566 + * find block B in PA
3567 + * copy on-disk bitmap to buddy
3568 + * mark B in on-disk bitmap
3569 + * drop PA from group
3570 + * mark all PAs in buddy
3572 + * thus, P1 initializes buddy with B available. to prevent this
3573 + * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3574 + * against that pair
3576 + ext3_lock_group(sb, grp);
3577 + list_del_rcu(&pa->pa_group_list);
3578 + ext3_unlock_group(sb, grp);
3580 + spin_lock(pa->pa_obj_lock);
3581 + list_del_rcu(&pa->pa_inode_list);
3582 + spin_unlock(pa->pa_obj_lock);
3588 + * creates new preallocated space for given inode
3590 +int ext3_mb_new_inode_pa(struct ext3_allocation_context *ac)
3592 + struct super_block *sb = ac->ac_sb;
3593 + struct ext3_prealloc_space *pa;
3594 + struct ext3_group_info *grp;
3595 + struct ext3_inode_info *ei;
3597 + /* preallocate only when found space is larger then requested */
3598 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3599 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3600 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3602 + pa = kmem_cache_alloc(ext3_pspace_cachep, SLAB_NOFS);
3606 + if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3607 + int winl, wins, win, offs;
3609 + /* we can't allocate as much as normalizer wants.
3610 + * so, found space must get proper lstart
3611 + * to cover original request */
3612 + BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3613 + BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3615 + /* we're limited by original request in that
3616 + * logical block must be covered any way
3617 + * winl is window we can move our chunk within */
3618 + winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3620 + /* also, we should cover whole original request */
3621 + wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3623 + /* the smallest one defines real window */
3624 + win = min(winl, wins);
3626 + offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3627 + if (offs && offs < win)
3630 + ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3631 + BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3632 + BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3635 + /* preallocation can change ac_b_ex, thus we store actually
3636 + * allocated blocks for history */
3637 + ac->ac_f_ex = ac->ac_b_ex;
3639 + pa->pa_lstart = ac->ac_b_ex.fe_logical;
3640 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3641 + pa->pa_len = ac->ac_b_ex.fe_len;
3642 + pa->pa_free = pa->pa_len;
3643 + atomic_set(&pa->pa_count, 1);
3644 + spin_lock_init(&pa->pa_lock);
3645 + pa->pa_deleted = 0;
3646 + pa->pa_linear = 0;
3648 + mb_debug("new inode pa %p: %lu/%lu for %lu\n", pa,
3649 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3651 + ext3_mb_use_inode_pa(ac, pa);
3652 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3654 + ei = EXT3_I(ac->ac_inode);
3655 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3657 + pa->pa_obj_lock = &ei->i_prealloc_lock;
3658 + pa->pa_inode = ac->ac_inode;
3660 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3661 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3662 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3664 + spin_lock(pa->pa_obj_lock);
3665 + list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3666 + spin_unlock(pa->pa_obj_lock);
3672 + * creates new preallocated space for locality group inodes belongs to
3674 +int ext3_mb_new_group_pa(struct ext3_allocation_context *ac)
3676 + struct super_block *sb = ac->ac_sb;
3677 + struct ext3_locality_group *lg;
3678 + struct ext3_prealloc_space *pa;
3679 + struct ext3_group_info *grp;
3681 + /* preallocate only when found space is larger then requested */
3682 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3683 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3684 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3686 + BUG_ON(ext3_pspace_cachep == NULL);
3687 + pa = kmem_cache_alloc(ext3_pspace_cachep, SLAB_NOFS);
3691 + /* preallocation can change ac_b_ex, thus we store actually
3692 + * allocated blocks for history */
3693 + ac->ac_f_ex = ac->ac_b_ex;
3695 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3696 + pa->pa_lstart = pa->pa_pstart;
3697 + pa->pa_len = ac->ac_b_ex.fe_len;
3698 + pa->pa_free = pa->pa_len;
3699 + atomic_set(&pa->pa_count, 1);
3700 + spin_lock_init(&pa->pa_lock);
3701 + pa->pa_deleted = 0;
3702 + pa->pa_linear = 1;
3704 + mb_debug("new group pa %p: %lu/%lu for %lu\n", pa,
3705 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3707 + ext3_mb_use_group_pa(ac, pa);
3708 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3710 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3712 + BUG_ON(lg == NULL);
3714 + pa->pa_obj_lock = &lg->lg_prealloc_lock;
3715 + pa->pa_inode = NULL;
3717 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3718 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3719 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3721 + spin_lock(pa->pa_obj_lock);
3722 + list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3723 + spin_unlock(pa->pa_obj_lock);
3728 +int ext3_mb_new_preallocation(struct ext3_allocation_context *ac)
3732 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3733 + err = ext3_mb_new_group_pa(ac);
3735 + err = ext3_mb_new_inode_pa(ac);
3740 + * finds all unused blocks in on-disk bitmap, frees them in
3741 + * in-core bitmap and buddy.
3742 + * @pa must be unlinked from inode and group lists, so that
3743 + * nobody else can find/use it.
3744 + * the caller MUST hold group/inode locks.
3745 + * TODO: optimize the case when there are no in-core structures yet
3747 +int ext3_mb_release_inode_pa(struct ext3_buddy *e3b,
3748 + struct buffer_head *bitmap_bh,
3749 + struct ext3_prealloc_space *pa)
3751 + struct ext3_allocation_context ac;
3752 + struct super_block *sb = e3b->bd_sb;
3753 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3754 + unsigned long bit, end, next, group;
3756 + int err = 0, free = 0;
3758 + BUG_ON(pa->pa_deleted == 0);
3759 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3760 + BUG_ON(group != e3b->bd_group);
3761 + end = bit + pa->pa_len;
3764 + ac.ac_inode = pa->pa_inode;
3765 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3767 + while (bit < end) {
3768 + bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3771 + next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3774 + start = group * EXT3_BLOCKS_PER_GROUP(sb) + bit +
3775 + le32_to_cpu(sbi->s_es->s_first_data_block);
3776 + mb_debug(" free preallocated %u/%u in group %u\n",
3777 + (unsigned) start, (unsigned) next - bit,
3778 + (unsigned) group);
3779 + free += next - bit;
3781 + ac.ac_b_ex.fe_group = group;
3782 + ac.ac_b_ex.fe_start = bit;
3783 + ac.ac_b_ex.fe_len = next - bit;
3784 + ac.ac_b_ex.fe_logical = 0;
3785 + ext3_mb_store_history(&ac);
3787 + mb_free_blocks(pa->pa_inode, e3b, bit, next - bit);
3790 + if (free != pa->pa_free) {
3791 + printk("pa %p: logic %lu, phys. %lu, len %lu\n",
3792 + pa, (unsigned long) pa->pa_lstart,
3793 + (unsigned long) pa->pa_pstart,
3794 + (unsigned long) pa->pa_len);
3795 + printk("free %u, pa_free %u\n", free, pa->pa_free);
3797 + BUG_ON(free != pa->pa_free);
3798 + atomic_add(free, &sbi->s_mb_discarded);
3803 +int ext3_mb_release_group_pa(struct ext3_buddy *e3b,
3804 + struct ext3_prealloc_space *pa)
3806 + struct ext3_allocation_context ac;
3807 + struct super_block *sb = e3b->bd_sb;
3808 + unsigned long bit, group;
3810 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3812 + BUG_ON(pa->pa_deleted == 0);
3813 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3814 + BUG_ON(group != e3b->bd_group);
3815 + mb_free_blocks(pa->pa_inode, e3b, bit, pa->pa_len);
3816 + atomic_add(pa->pa_len, &EXT3_SB(sb)->s_mb_discarded);
3819 + ac.ac_inode = NULL;
3820 + ac.ac_b_ex.fe_group = group;
3821 + ac.ac_b_ex.fe_start = bit;
3822 + ac.ac_b_ex.fe_len = pa->pa_len;
3823 + ac.ac_b_ex.fe_logical = 0;
3824 + ext3_mb_store_history(&ac);
3830 + * releases all preallocations in given group
3832 + * first, we need to decide discard policy:
3833 + * - when do we discard
3835 + * - how many do we discard
3836 + * 1) how many requested
3838 +int ext3_mb_discard_group_preallocations(struct super_block *sb,
3839 + int group, int needed)
3841 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
3842 + struct buffer_head *bitmap_bh = NULL;
3843 + struct ext3_prealloc_space *pa, *tmp;
3844 + struct list_head list;
3845 + struct ext3_buddy e3b;
3846 + int err, busy, free = 0;
3848 + mb_debug("discard preallocation for group %lu\n", group);
3850 + if (list_empty(&grp->bb_prealloc_list))
3853 + bitmap_bh = read_block_bitmap(sb, group);
3854 + if (bitmap_bh == NULL) {
3855 + /* error handling here */
3856 + ext3_mb_release_desc(&e3b);
3857 + BUG_ON(bitmap_bh == NULL);
3860 + err = ext3_mb_load_buddy(sb, group, &e3b);
3861 + BUG_ON(err != 0); /* error handling here */
3864 + needed = EXT3_BLOCKS_PER_GROUP(sb) + 1;
3866 + grp = EXT3_GROUP_INFO(sb, group);
3867 + INIT_LIST_HEAD(&list);
3871 + ext3_lock_group(sb, group);
3872 + list_for_each_entry_safe (pa, tmp, &grp->bb_prealloc_list, pa_group_list) {
3873 + spin_lock(&pa->pa_lock);
3874 + if (atomic_read(&pa->pa_count)) {
3875 + spin_unlock(&pa->pa_lock);
3876 + printk("uh! busy PA\n");
3881 + if (pa->pa_deleted) {
3882 + spin_unlock(&pa->pa_lock);
3886 + /* seems this one can be freed ... */
3887 + pa->pa_deleted = 1;
3889 + /* we can trust pa_free ... */
3890 + free += pa->pa_free;
3892 + spin_unlock(&pa->pa_lock);
3894 + list_del_rcu(&pa->pa_group_list);
3895 + list_add(&pa->u.pa_tmp_list, &list);
3898 + /* if we still need more blocks and some PAs were used, try again */
3899 + if (free < needed && busy)
3902 + /* found anything to free? */
3903 + if (list_empty(&list)) {
3904 + BUG_ON(free != 0);
3908 + /* now free all selected PAs */
3909 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3911 + /* remove from object (inode or locality group) */
3912 + spin_lock(pa->pa_obj_lock);
3913 + list_del_rcu(&pa->pa_inode_list);
3914 + spin_unlock(pa->pa_obj_lock);
3916 + if (pa->pa_linear)
3917 + ext3_mb_release_group_pa(&e3b, pa);
3919 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
3921 + list_del(&pa->u.pa_tmp_list);
3926 + ext3_unlock_group(sb, group);
3927 + ext3_mb_release_desc(&e3b);
3928 + brelse(bitmap_bh);
3933 + * releases all non-used preallocated blocks for given inode
3935 +void ext3_mb_discard_inode_preallocations(struct inode *inode)
3937 + struct ext3_inode_info *ei = EXT3_I(inode);
3938 + struct super_block *sb = inode->i_sb;
3939 + struct buffer_head *bitmap_bh = NULL;
3940 + struct ext3_prealloc_space *pa, *tmp;
3941 + unsigned long group = 0;
3942 + struct list_head list;
3943 + struct ext3_buddy e3b;
3946 + if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3947 + /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3951 + mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3953 + INIT_LIST_HEAD(&list);
3956 + /* first, collect all pa's in the inode */
3957 + spin_lock(&ei->i_prealloc_lock);
3958 + while (!list_empty(&ei->i_prealloc_list)) {
3959 + pa = list_entry(ei->i_prealloc_list.next,
3960 + struct ext3_prealloc_space, pa_inode_list);
3961 + BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3962 + spin_lock(&pa->pa_lock);
3963 + if (atomic_read(&pa->pa_count)) {
3964 + /* this shouldn't happen often - nobody should
3965 + * use preallocation while we're discarding it */
3966 + spin_unlock(&pa->pa_lock);
3967 + spin_unlock(&ei->i_prealloc_lock);
3968 + printk("uh-oh! used pa while discarding\n");
3970 + current->state = TASK_UNINTERRUPTIBLE;
3971 + schedule_timeout(HZ);
3975 + if (pa->pa_deleted == 0) {
3976 + pa->pa_deleted = 1;
3977 + spin_unlock(&pa->pa_lock);
3978 + list_del_rcu(&pa->pa_inode_list);
3979 + list_add(&pa->u.pa_tmp_list, &list);
3983 + /* someone is deleting pa right now */
3984 + spin_unlock(&pa->pa_lock);
3985 + spin_unlock(&ei->i_prealloc_lock);
3987 + /* we have to wait here because pa_deleted
3988 + * doesn't mean pa is already unlinked from
3989 + * the list. as we might be called from
3990 + * ->clear_inode() the inode will get freed
3991 + * and concurrent thread which is unlinking
3992 + * pa from inode's list may access already
3993 + * freed memory, bad-bad-bad */
3995 + /* XXX: if this happens too often, we can
3996 + * add a flag to force wait only in case
3997 + * of ->clear_inode(), but not in case of
3998 + * regular truncate */
3999 + printk("uh-oh! some one just deleted it\n");
4001 + current->state = TASK_UNINTERRUPTIBLE;
4002 + schedule_timeout(HZ);
4005 + spin_unlock(&ei->i_prealloc_lock);
4007 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4008 + BUG_ON(pa->pa_linear != 0);
4009 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4011 + err = ext3_mb_load_buddy(sb, group, &e3b);
4012 + BUG_ON(err != 0); /* error handling here */
4014 + bitmap_bh = read_block_bitmap(sb, group);
4015 + if (bitmap_bh == NULL) {
4016 + /* error handling here */
4017 + ext3_mb_release_desc(&e3b);
4018 + BUG_ON(bitmap_bh == NULL);
4021 + ext3_lock_group(sb, group);
4022 + list_del_rcu(&pa->pa_group_list);
4023 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
4024 + ext3_unlock_group(sb, group);
4026 + ext3_mb_release_desc(&e3b);
4027 + brelse(bitmap_bh);
4029 + list_del(&pa->u.pa_tmp_list);
4035 + * finds all preallocated spaces and return blocks being freed to them
4036 + * if preallocated space becomes full (no block is used from the space)
4037 + * then the function frees space in buddy
4038 + * XXX: at the moment, truncate (which is the only way to free blocks)
4039 + * discards all preallocations
4041 +void ext3_mb_return_to_preallocation(struct inode *inode, struct ext3_buddy *e3b,
4042 + sector_t block, int count)
4044 + BUG_ON(!list_empty(&EXT3_I(inode)->i_prealloc_list));
4047 +void ext3_mb_show_ac(struct ext3_allocation_context *ac)
4050 + struct super_block *sb = ac->ac_sb;
4053 + printk(KERN_ERR "EXT3-fs: can't allocate: status %d flags %d\n",
4054 + ac->ac_status, ac->ac_flags);
4055 + printk(KERN_ERR "EXT3-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4056 + "best %lu/%lu/%lu@%lu cr %d\n",
4057 + ac->ac_o_ex.fe_group, ac->ac_o_ex.fe_start,
4058 + ac->ac_o_ex.fe_len, ac->ac_o_ex.fe_logical,
4059 + ac->ac_g_ex.fe_group, ac->ac_g_ex.fe_start,
4060 + ac->ac_g_ex.fe_len, ac->ac_g_ex.fe_logical,
4061 + ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
4062 + ac->ac_b_ex.fe_len, ac->ac_b_ex.fe_logical,
4064 + printk(KERN_ERR "EXT3-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4066 + printk("EXT3-fs: groups: ");
4067 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
4068 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, i);
4069 + struct ext3_prealloc_space *pa;
4070 + unsigned long start;
4071 + struct list_head *cur;
4072 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
4073 + pa = list_entry(cur, struct ext3_prealloc_space,
4075 + spin_lock(&pa->pa_lock);
4076 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, NULL, &start);
4077 + spin_unlock(&pa->pa_lock);
4078 + printk("PA:%u:%lu:%u ", i, start, pa->pa_len);
4081 + if (grp->bb_free == 0)
4083 + printk("%d: %d/%d ", i, grp->bb_free, grp->bb_fragments);
4090 +void ext3_mb_group_or_file(struct ext3_allocation_context *ac)
4092 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
4093 + int bsbits = ac->ac_sb->s_blocksize_bits;
4094 + loff_t size, isize;
4096 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
4099 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4100 + isize = i_size_read(ac->ac_inode) >> bsbits;
4104 + /* don't use group allocation for large files */
4105 + if (size >= sbi->s_mb_stream_request)
4108 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
4111 + BUG_ON(ac->ac_lg != NULL);
4112 + ac->ac_lg = &sbi->s_locality_groups[smp_processor_id()];
4114 + /* we're going to use group allocation */
4115 + ac->ac_flags |= EXT3_MB_HINT_GROUP_ALLOC;
4117 + /* serialize all allocations in the group */
4118 + down(&ac->ac_lg->lg_sem);
4121 +int ext3_mb_initialize_context(struct ext3_allocation_context *ac,
4122 + struct ext3_allocation_request *ar)
4124 + struct super_block *sb = ar->inode->i_sb;
4125 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4126 + struct ext3_super_block *es = sbi->s_es;
4127 + unsigned long group, len, goal;
4128 + unsigned long block;
4130 + /* we can't allocate > group size */
4132 + if (len >= EXT3_BLOCKS_PER_GROUP(sb) - 10)
4133 + len = EXT3_BLOCKS_PER_GROUP(sb) - 10;
4135 + /* start searching from the goal */
4137 + if (goal < le32_to_cpu(es->s_first_data_block) ||
4138 + goal >= le32_to_cpu(es->s_blocks_count))
4139 + goal = le32_to_cpu(es->s_first_data_block);
4140 + ext3_get_group_no_and_offset(sb, goal, &group, &block);
4142 + /* set up allocation goals */
4143 + ac->ac_b_ex.fe_logical = ar->logical;
4144 + ac->ac_b_ex.fe_group = 0;
4145 + ac->ac_b_ex.fe_start = 0;
4146 + ac->ac_b_ex.fe_len = 0;
4147 + ac->ac_status = AC_STATUS_CONTINUE;
4148 + ac->ac_groups_scanned = 0;
4149 + ac->ac_ex_scanned = 0;
4152 + ac->ac_inode = ar->inode;
4153 + ac->ac_o_ex.fe_logical = ar->logical;
4154 + ac->ac_o_ex.fe_group = group;
4155 + ac->ac_o_ex.fe_start = block;
4156 + ac->ac_o_ex.fe_len = len;
4157 + ac->ac_g_ex.fe_logical = ar->logical;
4158 + ac->ac_g_ex.fe_group = group;
4159 + ac->ac_g_ex.fe_start = block;
4160 + ac->ac_g_ex.fe_len = len;
4161 + ac->ac_f_ex.fe_len = 0;
4162 + ac->ac_flags = ar->flags;
4163 + ac->ac_2order = 0;
4164 + ac->ac_criteria = 0;
4166 + ac->ac_bitmap_page = NULL;
4167 + ac->ac_buddy_page = NULL;
4170 + /* we have to define context: we'll we work with a file or
4171 + * locality group. this is a policy, actually */
4172 + ext3_mb_group_or_file(ac);
4174 + mb_debug("init ac: %u blocks @ %llu, goal %llu, flags %x, 2^%d, "
4175 + "left: %llu/%llu, right %llu/%llu to %swritable\n",
4176 + (unsigned) ar->len, (unsigned) ar->logical,
4177 + (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4178 + (unsigned) ar->lleft, (unsigned) ar->pleft,
4179 + (unsigned) ar->lright, (unsigned) ar->pright,
4180 + atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4186 + * release all resource we used in allocation
4188 +int ext3_mb_release_context(struct ext3_allocation_context *ac)
4191 + if (ac->ac_pa->pa_linear) {
4192 + /* see comment in ext3_mb_use_group_pa() */
4193 + spin_lock(&ac->ac_pa->pa_lock);
4194 + ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
4195 + ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
4196 + ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
4197 + ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
4198 + spin_unlock(&ac->ac_pa->pa_lock);
4200 + ext3_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
4202 + if (ac->ac_bitmap_page)
4203 + page_cache_release(ac->ac_bitmap_page);
4204 + if (ac->ac_buddy_page)
4205 + page_cache_release(ac->ac_buddy_page);
4206 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
4207 + up(&ac->ac_lg->lg_sem);
4208 + ext3_mb_collect_stats(ac);
4212 +int ext3_mb_discard_preallocations(struct super_block *sb, int needed)
4214 + int i, ret, freed = 0;
4216 + for (i = 0; i < EXT3_SB(sb)->s_groups_count && needed > 0; i++) {
4217 + ret = ext3_mb_discard_group_preallocations(sb, i, needed);
4226 + * Main entry point into mballoc to allocate blocks
4227 + * it tries to use preallocation first, then falls back
4228 + * to usual allocation
4230 +unsigned long ext3_mb_new_blocks(handle_t *handle,
4231 + struct ext3_allocation_request *ar, int *errp)
4233 + struct ext3_allocation_context ac;
4234 + struct ext3_sb_info *sbi;
4235 + struct super_block *sb;
4236 + unsigned long block;
4239 + sb = ar->inode->i_sb;
4240 + sbi = EXT3_SB(sb);
4242 + if (!test_opt(sb, MBALLOC)) {
4243 + static int ext3_mballoc_warning = 0;
4244 + if (ext3_mballoc_warning++ == 0)
4245 + printk(KERN_ERR "EXT3-fs: multiblock request with "
4246 + "mballoc disabled!\n");
4248 + err = ext3_new_block_old(handle, ar->inode, ar->goal, errp);
4252 + ext3_mb_poll_new_transaction(sb, handle);
4254 + if ((err = ext3_mb_initialize_context(&ac, ar)))
4257 + ac.ac_op = EXT3_MB_HISTORY_PREALLOC;
4258 + if (!ext3_mb_use_preallocated(&ac)) {
4260 + ac.ac_op = EXT3_MB_HISTORY_ALLOC;
4261 + ext3_mb_normalize_request(&ac, ar);
4264 + /* allocate space in core */
4265 + ext3_mb_regular_allocator(&ac);
4267 + /* as we've just preallocated more space than
4268 + * user requested orinally, we store allocated
4269 + * space in a special descriptor */
4270 + if (ac.ac_status == AC_STATUS_FOUND &&
4271 + ac.ac_o_ex.fe_len < ac.ac_b_ex.fe_len)
4272 + ext3_mb_new_preallocation(&ac);
4275 + if (likely(ac.ac_status == AC_STATUS_FOUND)) {
4276 + ext3_mb_mark_diskspace_used(&ac, handle);
4278 + block = ext3_grp_offs_to_block(sb, &ac.ac_b_ex);
4279 + ar->len = ac.ac_b_ex.fe_len;
4281 + freed = ext3_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
4285 + ac.ac_b_ex.fe_len = 0;
4287 + ext3_mb_show_ac(&ac);
4290 + ext3_mb_release_context(&ac);
4294 +EXPORT_SYMBOL(ext3_mb_new_blocks);
4296 +int ext3_new_block(handle_t *handle, struct inode *inode,
4297 + unsigned long goal, int *errp)
4299 + struct ext3_allocation_request ar;
4300 + unsigned long ret;
4302 + if (!test_opt(inode->i_sb, MBALLOC)) {
4303 + ret = ext3_new_block_old(handle, inode, goal, errp);
4316 + ret = ext3_mb_new_blocks(handle, &ar, errp);
4320 +void ext3_mb_poll_new_transaction(struct super_block *sb, handle_t *handle)
4322 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4324 + if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4327 + /* new transaction! time to close last one and free blocks for
4328 + * committed transaction. we know that only transaction can be
4329 + * active, so previos transaction can be being logged and we
4330 + * know that transaction before previous is known to be already
4331 + * logged. this means that now we may free blocks freed in all
4332 + * transactions before previous one. hope I'm clear enough ... */
4334 + spin_lock(&sbi->s_md_lock);
4335 + if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4336 + mb_debug("new transaction %lu, old %lu\n",
4337 + (unsigned long) handle->h_transaction->t_tid,
4338 + (unsigned long) sbi->s_last_transaction);
4339 + list_splice_init(&sbi->s_closed_transaction,
4340 + &sbi->s_committed_transaction);
4341 + list_splice_init(&sbi->s_active_transaction,
4342 + &sbi->s_closed_transaction);
4343 + sbi->s_last_transaction = handle->h_transaction->t_tid;
4345 + spin_unlock(&sbi->s_md_lock);
4347 + ext3_mb_free_committed_blocks(sb);
4350 +int ext3_mb_free_metadata(handle_t *handle, struct ext3_buddy *e3b,
4351 + int group, int block, int count)
4353 + struct ext3_group_info *db = e3b->bd_info;
4354 + struct super_block *sb = e3b->bd_sb;
4355 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4356 + struct ext3_free_metadata *md;
4359 + BUG_ON(e3b->bd_bitmap_page == NULL);
4360 + BUG_ON(e3b->bd_buddy_page == NULL);
4362 + ext3_lock_group(sb, group);
4363 + for (i = 0; i < count; i++) {
4364 + md = db->bb_md_cur;
4365 + if (md && db->bb_tid != handle->h_transaction->t_tid) {
4366 + db->bb_md_cur = NULL;
4371 + ext3_unlock_group(sb, group);
4372 + md = kmalloc(sizeof(*md), GFP_KERNEL);
4376 + md->group = group;
4378 + ext3_lock_group(sb, group);
4379 + if (db->bb_md_cur == NULL) {
4380 + spin_lock(&sbi->s_md_lock);
4381 + list_add(&md->list, &sbi->s_active_transaction);
4382 + spin_unlock(&sbi->s_md_lock);
4383 + /* protect buddy cache from being freed,
4384 + * otherwise we'll refresh it from
4385 + * on-disk bitmap and lose not-yet-available
4387 + page_cache_get(e3b->bd_buddy_page);
4388 + page_cache_get(e3b->bd_bitmap_page);
4389 + db->bb_md_cur = md;
4390 + db->bb_tid = handle->h_transaction->t_tid;
4391 + mb_debug("new md 0x%p for group %u\n",
4395 + md = db->bb_md_cur;
4399 + BUG_ON(md->num >= EXT3_BB_MAX_BLOCKS);
4400 + md->blocks[md->num] = block + i;
4402 + if (md->num == EXT3_BB_MAX_BLOCKS) {
4403 + /* no more space, put full container on a sb's list */
4404 + db->bb_md_cur = NULL;
4407 + ext3_unlock_group(sb, group);
4412 + * Main entry point into mballoc to free blocks
4414 +void ext3_mb_free_blocks(handle_t *handle, struct inode *inode,
4415 + unsigned long block, unsigned long count,
4416 + int metadata, int *freed)
4418 + struct buffer_head *bitmap_bh = NULL;
4419 + struct super_block *sb = inode->i_sb;
4420 + struct ext3_allocation_context ac;
4421 + struct ext3_group_desc *gdp;
4422 + struct ext3_super_block *es;
4423 + unsigned long bit, overflow;
4424 + struct buffer_head *gd_bh;
4425 + unsigned long block_group;
4426 + struct ext3_sb_info *sbi;
4427 + struct ext3_buddy e3b;
4432 + ext3_mb_poll_new_transaction(sb, handle);
4434 + sbi = EXT3_SB(sb);
4435 + es = EXT3_SB(sb)->s_es;
4436 + if (block < le32_to_cpu(es->s_first_data_block) ||
4437 + block + count < block ||
4438 + block + count > le32_to_cpu(es->s_blocks_count)) {
4439 + ext3_error (sb, __FUNCTION__,
4440 + "Freeing blocks not in datazone - "
4441 + "block = %lu, count = %lu", block, count);
4442 + goto error_return;
4445 + ext3_debug("freeing block %lu\n", block);
4447 + ac.ac_op = EXT3_MB_HISTORY_FREE;
4448 + ac.ac_inode = inode;
4453 + ext3_get_group_no_and_offset(sb, block, &block_group, &bit);
4456 + * Check to see if we are freeing blocks across a group
4459 + if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
4460 + overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
4461 + count -= overflow;
4463 + brelse(bitmap_bh);
4464 + bitmap_bh = read_block_bitmap(sb, block_group);
4466 + goto error_return;
4467 + gdp = ext3_get_group_desc (sb, block_group, &gd_bh);
4469 + goto error_return;
4471 + if (in_range (le32_to_cpu(gdp->bg_block_bitmap), block, count) ||
4472 + in_range (le32_to_cpu(gdp->bg_inode_bitmap), block, count) ||
4473 + in_range (block, le32_to_cpu(gdp->bg_inode_table),
4474 + EXT3_SB(sb)->s_itb_per_group) ||
4475 + in_range (block + count - 1, le32_to_cpu(gdp->bg_inode_table),
4476 + EXT3_SB(sb)->s_itb_per_group))
4477 + ext3_error(sb, __FUNCTION__,
4478 + "Freeing blocks in system zone - "
4479 + "Block = %lu, count = %lu", block, count);
4481 + BUFFER_TRACE(bitmap_bh, "getting write access");
4482 + err = ext3_journal_get_write_access(handle, bitmap_bh);
4484 + goto error_return;
4487 + * We are about to modify some metadata. Call the journal APIs
4488 + * to unshare ->b_data if a currently-committing transaction is
4491 + BUFFER_TRACE(gd_bh, "get_write_access");
4492 + err = ext3_journal_get_write_access(handle, gd_bh);
4494 + goto error_return;
4496 + err = ext3_mb_load_buddy(sb, block_group, &e3b);
4498 + goto error_return;
4500 +#ifdef AGGRESSIVE_CHECK
4503 + for (i = 0; i < count; i++)
4504 + BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4507 + mb_clear_bits(bitmap_bh->b_data, bit, count);
4509 + /* We dirtied the bitmap block */
4510 + BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4511 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
4513 + ac.ac_b_ex.fe_group = block_group;
4514 + ac.ac_b_ex.fe_start = bit;
4515 + ac.ac_b_ex.fe_len = count;
4516 + ext3_mb_store_history(&ac);
4519 + /* blocks being freed are metadata. these blocks shouldn't
4520 + * be used until this transaction is committed */
4521 + ext3_mb_free_metadata(handle, &e3b, block_group, bit, count);
4523 + ext3_lock_group(sb, block_group);
4524 + err = mb_free_blocks(inode, &e3b, bit, count);
4525 + ext3_mb_return_to_preallocation(inode, &e3b, block, count);
4526 + ext3_unlock_group(sb, block_group);
4530 + spin_lock(sb_bgl_lock(sbi, block_group));
4531 + gdp->bg_free_blocks_count =
4532 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
4533 + spin_unlock(sb_bgl_lock(sbi, block_group));
4534 + percpu_counter_mod(&sbi->s_freeblocks_counter, count);
4536 + ext3_mb_release_desc(&e3b);
4540 + /* And the group descriptor block */
4541 + BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4542 + ret = ext3_journal_dirty_metadata(handle, gd_bh);
4543 + if (!err) err = ret;
4545 + if (overflow && !err) {
4552 + brelse(bitmap_bh);
4553 + ext3_std_error(sb, err);