1 Index: linux-2.6.22.19/include/linux/ext3_fs.h
2 ===================================================================
3 --- linux-2.6.22.19.orig/include/linux/ext3_fs.h
4 +++ linux-2.6.22.19/include/linux/ext3_fs.h
6 #define ext3_debug(f, a...) do {} while (0)
9 +#define EXT3_MULTIBLOCK_ALLOCATOR 1
11 +#define EXT3_MB_HINT_MERGE 1 /* prefer goal again. length */
12 +#define EXT3_MB_HINT_RESERVED 2 /* blocks already reserved */
13 +#define EXT3_MB_HINT_METADATA 4 /* metadata is being allocated */
14 +#define EXT3_MB_HINT_FIRST 8 /* first blocks in the file */
15 +#define EXT3_MB_HINT_BEST 16 /* search for the best chunk */
16 +#define EXT3_MB_HINT_DATA 32 /* data is being allocated */
17 +#define EXT3_MB_HINT_NOPREALLOC 64 /* don't preallocate (for tails) */
18 +#define EXT3_MB_HINT_GROUP_ALLOC 128 /* allocate for locality group */
19 +#define EXT3_MB_HINT_GOAL_ONLY 256 /* allocate goal blocks or none */
20 +#define EXT3_MB_HINT_TRY_GOAL 512 /* goal is meaningful */
22 +struct ext3_allocation_request {
23 + struct inode *inode; /* target inode for block we're allocating */
24 + unsigned long logical; /* logical block in target inode */
25 + unsigned long goal; /* phys. target (a hint) */
26 + unsigned long lleft; /* the closest logical allocated block to the left */
27 + unsigned long pleft; /* phys. block for ^^^ */
28 + unsigned long lright; /* the closest logical allocated block to the right */
29 + unsigned long pright; /* phys. block for ^^^ */
30 + unsigned long len; /* how many blocks we want to allocate */
31 + unsigned long flags; /* flags. see above EXT3_MB_HINT_* */
35 * Special inodes numbers
37 @@ -412,6 +437,14 @@ struct ext3_inode {
38 #define ext3_find_first_zero_bit ext2_find_first_zero_bit
39 #define ext3_find_next_zero_bit ext2_find_next_zero_bit
41 +#ifndef ext2_find_next_le_bit
42 +#ifdef __LITTLE_ENDIAN
43 +#define ext2_find_next_le_bit(addr, size, off) find_next_bit((addr), (size), (off))
45 +#error "mballoc needs a patch for big-endian systems - CFS bug 10634"
46 +#endif /* __LITTLE_ENDIAN */
47 +#endif /* !ext2_find_next_le_bit */
50 * Maximal mount counts between two filesystem checks
52 @@ -813,6 +846,20 @@ extern unsigned long ext3_count_dirs (st
53 extern void ext3_check_inodes_bitmap (struct super_block *);
54 extern unsigned long ext3_count_free (struct buffer_head *, unsigned);
57 +extern long ext3_mb_stats;
58 +extern long ext3_mb_max_to_scan;
59 +extern int ext3_mb_init(struct super_block *, int);
60 +extern int ext3_mb_release(struct super_block *);
61 +extern unsigned long ext3_mb_new_blocks(handle_t *, struct ext3_allocation_request *, int *);
62 +extern int ext3_mb_reserve_blocks(struct super_block *, int);
63 +extern void ext3_mb_release_blocks(struct super_block *, int);
64 +extern void ext3_mb_release_blocks(struct super_block *, int);
65 +extern void ext3_mb_discard_inode_preallocations(struct inode *);
66 +extern int __init init_ext3_mb_proc(void);
67 +extern void exit_ext3_mb_proc(void);
68 +extern void ext3_mb_free_blocks(handle_t *, struct inode *, unsigned long, unsigned long, int, int *);
72 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
73 @@ -859,6 +906,10 @@ extern int ext3_group_extend(struct supe
74 ext3_fsblk_t n_blocks_count);
77 +extern struct proc_dir_entry *proc_root_ext3;
78 +extern int __init init_ext3_proc(void);
79 +extern void exit_ext3_proc(void);
81 extern void ext3_error (struct super_block *, const char *, const char *, ...)
82 __attribute__ ((format (printf, 3, 4)));
83 extern void __ext3_std_error (struct super_block *, const char *, int);
84 Index: linux-2.6.22.19/include/linux/ext3_fs_sb.h
85 ===================================================================
86 --- linux-2.6.22.19.orig/include/linux/ext3_fs_sb.h
87 +++ linux-2.6.22.19/include/linux/ext3_fs_sb.h
88 @@ -88,6 +88,68 @@ struct ext3_sb_info {
89 unsigned long s_ext_blocks;
90 unsigned long s_ext_extents;
93 + /* for buddy allocator */
94 + struct ext3_group_info ***s_group_info;
95 + struct inode *s_buddy_cache;
96 + long s_blocks_reserved;
97 + spinlock_t s_reserve_lock;
98 + struct list_head s_active_transaction;
99 + struct list_head s_closed_transaction;
100 + struct list_head s_committed_transaction;
101 + spinlock_t s_md_lock;
102 + tid_t s_last_transaction;
103 + unsigned short *s_mb_offsets, *s_mb_maxs;
106 + unsigned long s_mb_factor;
107 + unsigned long s_stripe;
108 + unsigned long s_mb_small_req;
109 + unsigned long s_mb_large_req;
110 + unsigned long s_mb_max_to_scan;
111 + unsigned long s_mb_min_to_scan;
112 + unsigned long s_mb_max_groups_to_scan;
113 + unsigned long s_mb_stats;
114 + unsigned long s_mb_order2_reqs;
115 + unsigned long *s_mb_prealloc_table;
116 + unsigned long s_mb_prealloc_table_size;
117 + unsigned long s_mb_group_prealloc;
118 + /* where last allocation was done - for stream allocation */
119 + unsigned long s_mb_last_group;
120 + unsigned long s_mb_last_start;
122 + /* history to debug policy */
123 + struct ext3_mb_history *s_mb_history;
124 + int s_mb_history_cur;
125 + int s_mb_history_max;
126 + int s_mb_history_num;
127 + struct proc_dir_entry *s_dev_proc;
128 + spinlock_t s_mb_history_lock;
129 + int s_mb_history_filter;
131 + /* stats for buddy allocator */
132 + spinlock_t s_mb_pa_lock;
133 + atomic_t s_bal_reqs; /* number of reqs with len > 1 */
134 + atomic_t s_bal_success; /* we found long enough chunks */
135 + atomic_t s_bal_allocated; /* in blocks */
136 + atomic_t s_bal_ex_scanned; /* total extents scanned */
137 + atomic_t s_bal_goals; /* goal hits */
138 + atomic_t s_bal_breaks; /* too long searches */
139 + atomic_t s_bal_2orders; /* 2^order hits */
140 + spinlock_t s_bal_lock;
141 + unsigned long s_mb_buddies_generated;
142 + unsigned long long s_mb_generation_time;
143 + atomic_t s_mb_lost_chunks;
144 + atomic_t s_mb_preallocated;
145 + atomic_t s_mb_discarded;
147 + /* locality groups */
148 + struct ext3_locality_group *s_locality_groups;
152 +#define EXT3_GROUP_INFO(sb, group) \
153 + EXT3_SB(sb)->s_group_info[(group) >> EXT3_DESC_PER_BLOCK_BITS(sb)] \
154 + [(group) & (EXT3_DESC_PER_BLOCK(sb) - 1)]
156 #endif /* _LINUX_EXT3_FS_SB */
157 Index: linux-2.6.22.19/fs/ext3/super.c
158 ===================================================================
159 --- linux-2.6.22.19.orig/fs/ext3/super.c
160 +++ linux-2.6.22.19/fs/ext3/super.c
161 @@ -392,6 +392,7 @@ static void ext3_put_super (struct super
162 struct ext3_super_block *es = sbi->s_es;
165 + ext3_mb_release(sb);
166 ext3_ext_release(sb);
167 ext3_xattr_put_super(sb);
168 journal_destroy(sbi->s_journal);
169 @@ -434,6 +435,10 @@ static void ext3_put_super (struct super
170 invalidate_bdev(sbi->journal_bdev);
171 ext3_blkdev_remove(sbi);
173 + if (sbi->s_dev_proc) {
174 + remove_proc_entry(sbi->s_dev_proc->name, proc_root_ext3);
175 + sbi->s_dev_proc = NULL;
177 sb->s_fs_info = NULL;
180 @@ -459,6 +464,8 @@ static struct inode *ext3_alloc_inode(st
181 ei->vfs_inode.i_version = 1;
183 memset(&ei->i_cached_extent, 0, sizeof(ei->i_cached_extent));
184 + INIT_LIST_HEAD(&ei->i_prealloc_list);
185 + spin_lock_init(&ei->i_prealloc_lock);
186 return &ei->vfs_inode;
189 @@ -1434,6 +1441,7 @@ static int ext3_fill_super (struct super
190 unsigned long journal_devnum = 0;
191 unsigned long def_mount_opts;
197 @@ -1448,6 +1456,22 @@ static int ext3_fill_super (struct super
198 sbi->s_mount_opt = 0;
199 sbi->s_resuid = EXT3_DEF_RESUID;
200 sbi->s_resgid = EXT3_DEF_RESGID;
201 + devname = kstrdup(sb->s_id, GFP_KERNEL);
204 + while ((p = strchr(p, '/')))
206 + sbi->s_dev_proc = proc_mkdir(devname, proc_root_ext3);
207 + if (sbi->s_dev_proc == NULL)
208 + printk(KERN_WARNING "EXT3-fs warning: unable to create "
209 + "procfs entry for %s(%s)\n",
210 + sb->s_id, devname);
213 + printk(KERN_WARNING "EXT3-fs warning: cannot allocate memory "
214 + "to create procfs entry for %s\n",
220 @@ -1857,6 +1881,10 @@ failed_mount:
221 ext3_blkdev_remove(sbi);
224 + if (sbi->s_dev_proc) {
225 + remove_proc_entry(sbi->s_dev_proc->name, proc_root_ext3);
226 + sbi->s_dev_proc = NULL;
228 sb->s_fs_info = NULL;
231 @@ -2787,9 +2815,46 @@ static struct file_system_type ext3_fs_t
232 .fs_flags = FS_REQUIRES_DEV,
235 +#define EXT3_ROOT "ext3"
236 +struct proc_dir_entry *proc_root_ext3;
238 +int __init init_ext3_proc(void)
242 + if ((ret = init_ext3_mb_proc()))
245 + proc_root_ext3 = proc_mkdir(EXT3_ROOT, proc_root_fs);
246 + if (proc_root_ext3 == NULL) {
247 + printk(KERN_ERR "EXT3-fs: Unable to create %s\n", EXT3_ROOT);
255 + exit_ext3_mb_proc();
260 +void exit_ext3_proc(void)
262 + exit_ext3_mb_proc();
263 + remove_proc_entry(EXT3_ROOT, proc_root_fs);
266 static int __init init_ext3_fs(void)
268 - int err = init_ext3_xattr();
271 + err = init_ext3_proc();
275 + err = init_ext3_xattr();
278 err = init_inodecache();
279 @@ -2811,6 +2876,7 @@ static void __exit exit_ext3_fs(void)
280 unregister_filesystem(&ext3_fs_type);
281 destroy_inodecache();
286 int ext3_map_inode_page(struct inode *inode, struct page *page,
287 Index: linux-2.6.22.19/fs/ext3/mballoc.c
288 ===================================================================
290 +++ linux-2.6.22.19/fs/ext3/mballoc.c
293 + * Copyright 2008 Sun Microsystems, Inc.
294 + * Written by Alex Tomas <alex@clusterfs.com>
296 + * This program is free software; you can redistribute it and/or modify
297 + * it under the terms of the GNU General Public License version 2 as
298 + * published by the Free Software Foundation.
300 + * This program is distributed in the hope that it will be useful,
301 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
302 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
303 + * GNU General Public License for more details.
305 + * You should have received a copy of the GNU General Public Licens
306 + * along with this program; if not, write to the Free Software
307 + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
312 + * mballoc.c contains the multiblocks allocation routines
315 +#include <linux/time.h>
316 +#include <linux/fs.h>
317 +#include <linux/namei.h>
318 +#include <linux/ext3_jbd.h>
319 +#include <linux/jbd.h>
320 +#include <linux/ext3_fs.h>
321 +#include <linux/quotaops.h>
322 +#include <linux/buffer_head.h>
323 +#include <linux/module.h>
324 +#include <linux/swap.h>
325 +#include <linux/proc_fs.h>
326 +#include <linux/pagemap.h>
327 +#include <linux/seq_file.h>
328 +#include <linux/version.h>
332 + * - test ext3_ext_search_left() and ext3_ext_search_right()
333 + * - search for metadata in few groups
336 + * - normalization should take into account whether file is still open
337 + * - discard preallocations if no free space left (policy?)
338 + * - don't normalize tails
340 + * - reservation for superuser
343 + * - bitmap read-ahead (proposed by Oleg Drokin aka green)
344 + * - track min/max extents in each group for better group selection
345 + * - mb_mark_used() may allocate chunk right after splitting buddy
346 + * - tree of groups sorted by number of free blocks
351 + * mballoc operates on the following data:
353 + * - in-core buddy (actually includes buddy and bitmap)
354 + * - preallocation descriptors (PAs)
356 + * there are two types of preallocations:
358 + * assiged to specific inode and can be used for this inode only.
359 + * it describes part of inode's space preallocated to specific
360 + * physical blocks. any block from that preallocated can be used
361 + * independent. the descriptor just tracks number of blocks left
362 + * unused. so, before taking some block from descriptor, one must
363 + * make sure corresponded logical block isn't allocated yet. this
364 + * also means that freeing any block within descriptor's range
365 + * must discard all preallocated blocks.
367 + * assigned to specific locality group which does not translate to
368 + * permanent set of inodes: inode can join and leave group. space
369 + * from this type of preallocation can be used for any inode. thus
370 + * it's consumed from the beginning to the end.
372 + * relation between them can be expressed as:
373 + * in-core buddy = on-disk bitmap + preallocation descriptors
375 + * this mean blocks mballoc considers used are:
376 + * - allocated blocks (persistent)
377 + * - preallocated blocks (non-persistent)
379 + * consistency in mballoc world means that at any time a block is either
380 + * free or used in ALL structures. notice: "any time" should not be read
381 + * literally -- time is discrete and delimited by locks.
383 + * to keep it simple, we don't use block numbers, instead we count number of
384 + * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
386 + * all operations can be expressed as:
387 + * - init buddy: buddy = on-disk + PAs
388 + * - new PA: buddy += N; PA = N
389 + * - use inode PA: on-disk += N; PA -= N
390 + * - discard inode PA buddy -= on-disk - PA; PA = 0
391 + * - use locality group PA on-disk += N; PA -= N
392 + * - discard locality group PA buddy -= PA; PA = 0
393 + * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
394 + * is used in real operation because we can't know actual used
395 + * bits from PA, only from on-disk bitmap
397 + * if we follow this strict logic, then all operations above should be atomic.
398 + * given some of them can block, we'd have to use something like semaphores
399 + * killing performance on high-end SMP hardware. let's try to relax it using
400 + * the following knowledge:
401 + * 1) if buddy is referenced, it's already initialized
402 + * 2) while block is used in buddy and the buddy is referenced,
403 + * nobody can re-allocate that block
404 + * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
405 + * bit set and PA claims same block, it's OK. IOW, one can set bit in
406 + * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
409 + * so, now we're building a concurrency table:
412 + * blocks for PA are allocated in the buddy, buddy must be referenced
413 + * until PA is linked to allocation group to avoid concurrent buddy init
415 + * we need to make sure that either on-disk bitmap or PA has uptodate data
416 + * given (3) we care that PA-=N operation doesn't interfere with init
417 + * - discard inode PA
418 + * the simplest way would be to have buddy initialized by the discard
419 + * - use locality group PA
420 + * again PA-=N must be serialized with init
421 + * - discard locality group PA
422 + * the simplest way would be to have buddy initialized by the discard
425 + * i_truncate_mutex serializes them
426 + * - discard inode PA
427 + * discard process must wait until PA isn't used by another process
428 + * - use locality group PA
429 + * some mutex should serialize them
430 + * - discard locality group PA
431 + * discard process must wait until PA isn't used by another process
434 + * i_truncate_mutex or another mutex should serializes them
435 + * - discard inode PA
436 + * discard process must wait until PA isn't used by another process
437 + * - use locality group PA
438 + * nothing wrong here -- they're different PAs covering different blocks
439 + * - discard locality group PA
440 + * discard process must wait until PA isn't used by another process
442 + * now we're ready to make few consequences:
443 + * - PA is referenced and while it is no discard is possible
444 + * - PA is referenced until block isn't marked in on-disk bitmap
445 + * - PA changes only after on-disk bitmap
446 + * - discard must not compete with init. either init is done before
447 + * any discard or they're serialized somehow
448 + * - buddy init as sum of on-disk bitmap and PAs is done atomically
450 + * a special case when we've used PA to emptiness. no need to modify buddy
451 + * in this case, but we should care about concurrent init
456 + * Logic in few words:
461 + * mark bits in on-disk bitmap
464 + * - use preallocation:
465 + * find proper PA (per-inode or group)
467 + * mark bits in on-disk bitmap
473 + * mark bits in on-disk bitmap
476 + * - discard preallocations in group:
478 + * move them onto local list
479 + * load on-disk bitmap
481 + * remove PA from object (inode or locality group)
482 + * mark free blocks in-core
484 + * - discard inode's preallocations:
491 + * - bitlock on a group (group)
492 + * - object (inode/locality) (object)
493 + * - per-pa lock (pa)
500 + * - find and use pa:
503 + * - release consumed pa:
508 + * - generate in-core bitmap:
512 + * - discard all for given object (inode, locality group):
517 + * - discard all for given group:
526 + * with AGGRESSIVE_CHECK allocator runs consistency checks over
527 + * structures. these checks slow things down a lot
529 +#define AGGRESSIVE_CHECK__
532 + * with DOUBLE_CHECK defined mballoc creates persistent in-core
533 + * bitmaps, maintains and uses them to check for double allocations
535 +#define DOUBLE_CHECK__
541 +#define mb_debug(fmt,a...) printk(fmt, ##a)
543 +#define mb_debug(fmt,a...)
547 + * with EXT3_MB_HISTORY mballoc stores last N allocations in memory
548 + * and you can monitor it in /proc/fs/ext3/<dev>/mb_history
550 +#define EXT3_MB_HISTORY
551 +#define EXT3_MB_HISTORY_ALLOC 1 /* allocation */
552 +#define EXT3_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
553 +#define EXT3_MB_HISTORY_DISCARD 4 /* preallocation discarded */
554 +#define EXT3_MB_HISTORY_FREE 8 /* free */
556 +#define EXT3_MB_HISTORY_DEFAULT (EXT3_MB_HISTORY_ALLOC | \
557 + EXT3_MB_HISTORY_PREALLOC | \
558 + EXT3_MB_HISTORY_DISCARD | \
559 + EXT3_MB_HISTORY_FREE)
562 + * How long mballoc can look for a best extent (in found extents)
564 +#define MB_DEFAULT_MAX_TO_SCAN 200
567 + * How long mballoc must look for a best extent
569 +#define MB_DEFAULT_MIN_TO_SCAN 10
572 + * How many groups mballoc will scan looking for the best chunk
574 +#define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
577 + * with 'ext3_mb_stats' allocator will collect stats that will be
578 + * shown at umount. The collecting costs though!
580 +#define MB_DEFAULT_STATS 1
583 + * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
584 + * by the stream allocator, which purpose is to pack requests
585 + * as close each to other as possible to produce smooth I/O traffic
587 +#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
590 + * for which requests use 2^N search using buddies
592 +#define MB_DEFAULT_ORDER2_REQS 8
595 + * default stripe size = 1MB
597 +#define MB_DEFAULT_STRIPE 256
599 +static kmem_cache_t *ext3_pspace_cachep = NULL;
601 +#ifdef EXT3_BB_MAX_BLOCKS
602 +#undef EXT3_BB_MAX_BLOCKS
604 +#define EXT3_BB_MAX_BLOCKS 30
606 +struct ext3_free_metadata {
608 + unsigned short num;
609 + unsigned short blocks[EXT3_BB_MAX_BLOCKS];
610 + struct list_head list;
613 +struct ext3_group_info {
614 + unsigned long bb_state;
615 + unsigned long bb_tid;
616 + struct ext3_free_metadata *bb_md_cur;
617 + unsigned short bb_first_free;
618 + unsigned short bb_free;
619 + unsigned short bb_fragments;
620 + struct list_head bb_prealloc_list;
621 + unsigned long bb_prealloc_nr;
625 + unsigned short bb_counters[];
628 +#define EXT3_GROUP_INFO_NEED_INIT_BIT 0
629 +#define EXT3_GROUP_INFO_LOCKED_BIT 1
631 +#define EXT3_MB_GRP_NEED_INIT(grp) \
632 + (test_bit(EXT3_GROUP_INFO_NEED_INIT_BIT, &(grp)->bb_state))
635 +struct ext3_prealloc_space {
636 + struct list_head pa_inode_list;
637 + struct list_head pa_group_list;
639 + struct list_head pa_tmp_list;
640 + struct rcu_head pa_rcu;
642 + spinlock_t pa_lock;
644 + unsigned pa_deleted;
645 + unsigned long pa_pstart; /* phys. block */
646 + unsigned long pa_lstart; /* log. block */
647 + unsigned short pa_len; /* len of preallocated chunk */
648 + unsigned short pa_free; /* how many blocks are free */
649 + unsigned short pa_linear; /* consumed in one direction
650 + * strictly, for group prealloc */
651 + spinlock_t *pa_obj_lock;
652 + struct inode *pa_inode; /* hack, for history only */
656 +struct ext3_free_extent {
657 + unsigned long fe_logical;
658 + unsigned long fe_start;
659 + unsigned long fe_group;
660 + unsigned long fe_len;
665 + * we try to group all related changes together
666 + * so that writeback can flush/allocate them together as well
668 +struct ext3_locality_group {
669 + /* for allocator */
670 + struct semaphore lg_sem; /* to serialize allocates */
671 + struct list_head lg_prealloc_list;/* list of preallocations */
672 + spinlock_t lg_prealloc_lock;
675 +struct ext3_allocation_context {
676 + struct inode *ac_inode;
677 + struct super_block *ac_sb;
679 + /* original request */
680 + struct ext3_free_extent ac_o_ex;
682 + /* goal request (after normalization) */
683 + struct ext3_free_extent ac_g_ex;
685 + /* the best found extent */
686 + struct ext3_free_extent ac_b_ex;
688 + /* copy of the bext found extent taken before preallocation efforts */
689 + struct ext3_free_extent ac_f_ex;
691 + /* number of iterations done. we have to track to limit searching */
692 + unsigned long ac_ex_scanned;
693 + __u16 ac_groups_scanned;
697 + __u16 ac_flags; /* allocation hints */
701 + __u8 ac_2order; /* if request is to allocate 2^N blocks and
702 + * N > 0, the field stores N, otherwise 0 */
703 + __u8 ac_op; /* operation, for history only */
704 + struct page *ac_bitmap_page;
705 + struct page *ac_buddy_page;
706 + struct ext3_prealloc_space *ac_pa;
707 + struct ext3_locality_group *ac_lg;
710 +#define AC_STATUS_CONTINUE 1
711 +#define AC_STATUS_FOUND 2
712 +#define AC_STATUS_BREAK 3
714 +struct ext3_mb_history {
715 + struct ext3_free_extent orig; /* orig allocation */
716 + struct ext3_free_extent goal; /* goal allocation */
717 + struct ext3_free_extent result; /* result allocation */
720 + __u16 found; /* how many extents have been found */
721 + __u16 groups; /* how many groups have been scanned */
722 + __u16 tail; /* what tail broke some buddy */
723 + __u16 buddy; /* buddy the tail ^^^ broke */
725 + __u8 cr:8; /* which phase the result extent was found at */
731 + struct page *bd_buddy_page;
733 + struct page *bd_bitmap_page;
735 + struct ext3_group_info *bd_info;
736 + struct super_block *bd_sb;
738 + unsigned bd_blkbits;
740 +#define EXT3_MB_BITMAP(e3b) ((e3b)->bd_bitmap)
741 +#define EXT3_MB_BUDDY(e3b) ((e3b)->bd_buddy)
743 +#ifndef EXT3_MB_HISTORY
744 +#define ext3_mb_store_history(ac)
746 +static void ext3_mb_store_history(struct ext3_allocation_context *ac);
749 +#define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
751 +int ext3_create (struct inode *, struct dentry *, int, struct nameidata *);
752 +struct buffer_head * read_block_bitmap(struct super_block *, unsigned int);
753 +unsigned long ext3_new_blocks_old(handle_t *handle, struct inode *inode,
754 + unsigned long goal, unsigned long *count, int *errp);
755 +void ext3_mb_release_blocks(struct super_block *, int);
756 +void ext3_mb_poll_new_transaction(struct super_block *, handle_t *);
757 +void ext3_mb_free_committed_blocks(struct super_block *);
758 +int ext3_mb_generate_from_pa(struct super_block *sb, void *bitmap, int group);
759 +void ext3_mb_free_consumed_preallocations(struct ext3_allocation_context *ac);
760 +void ext3_mb_return_to_preallocation(struct inode *inode, struct ext3_buddy *e3b,
761 + sector_t block, int count);
762 +void ext3_mb_show_ac(struct ext3_allocation_context *ac);
763 +void ext3_mb_check_with_pa(struct ext3_buddy *e3b, int first, int count);
764 +void ext3_mb_put_pa(struct ext3_allocation_context *, struct super_block *, struct ext3_prealloc_space *pa);
765 +int ext3_mb_init_per_dev_proc(struct super_block *sb);
766 +int ext3_mb_destroy_per_dev_proc(struct super_block *sb);
769 + * Calculate the block group number and offset, given a block number
771 +static void ext3_get_group_no_and_offset(struct super_block *sb,
772 + unsigned long blocknr,
773 + unsigned long *blockgrpp,
774 + unsigned long *offsetp)
776 + struct ext3_super_block *es = EXT3_SB(sb)->s_es;
777 + unsigned long offset;
779 + blocknr = blocknr - le32_to_cpu(es->s_first_data_block);
780 + offset = blocknr % EXT3_BLOCKS_PER_GROUP(sb);
781 + blocknr = blocknr / EXT3_BLOCKS_PER_GROUP(sb);
785 + *blockgrpp = blocknr;
790 +ext3_lock_group(struct super_block *sb, int group)
792 + bit_spin_lock(EXT3_GROUP_INFO_LOCKED_BIT,
793 + &EXT3_GROUP_INFO(sb, group)->bb_state);
797 +ext3_unlock_group(struct super_block *sb, int group)
799 + bit_spin_unlock(EXT3_GROUP_INFO_LOCKED_BIT,
800 + &EXT3_GROUP_INFO(sb, group)->bb_state);
804 +ext3_is_group_locked(struct super_block *sb, int group)
806 + return bit_spin_is_locked(EXT3_GROUP_INFO_LOCKED_BIT,
807 + &EXT3_GROUP_INFO(sb, group)->bb_state);
810 +unsigned long ext3_grp_offs_to_block(struct super_block *sb,
811 + struct ext3_free_extent *fex)
813 + unsigned long block;
815 + block = (unsigned long) fex->fe_group * EXT3_BLOCKS_PER_GROUP(sb)
817 + + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
821 +#if BITS_PER_LONG == 64
822 +#define mb_correct_addr_and_bit(bit,addr) \
824 + bit += ((unsigned long) addr & 7UL) << 3; \
825 + addr = (void *) ((unsigned long) addr & ~7UL); \
827 +#elif BITS_PER_LONG == 32
828 +#define mb_correct_addr_and_bit(bit,addr) \
830 + bit += ((unsigned long) addr & 3UL) << 3; \
831 + addr = (void *) ((unsigned long) addr & ~3UL); \
834 +#error "how many bits you are?!"
837 +static inline int mb_test_bit(int bit, void *addr)
839 + mb_correct_addr_and_bit(bit,addr);
840 + return ext2_test_bit(bit, addr);
843 +static inline void mb_set_bit(int bit, void *addr)
845 + mb_correct_addr_and_bit(bit,addr);
846 + ext2_set_bit(bit, addr);
849 +static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
851 + mb_correct_addr_and_bit(bit,addr);
852 + ext2_set_bit_atomic(lock, bit, addr);
855 +static inline void mb_clear_bit(int bit, void *addr)
857 + mb_correct_addr_and_bit(bit,addr);
858 + ext2_clear_bit(bit, addr);
861 +static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
863 + mb_correct_addr_and_bit(bit,addr);
864 + ext2_clear_bit_atomic(lock, bit, addr);
867 +static inline int mb_find_next_zero_bit(void *addr, int max, int start)
870 +#if BITS_PER_LONG == 64
871 + fix = ((unsigned long) addr & 7UL) << 3;
872 + addr = (void *) ((unsigned long) addr & ~7UL);
873 +#elif BITS_PER_LONG == 32
874 + fix = ((unsigned long) addr & 3UL) << 3;
875 + addr = (void *) ((unsigned long) addr & ~3UL);
877 +#error "how many bits you are?!"
881 + return ext2_find_next_zero_bit(addr, max, start) - fix;
884 +static inline int mb_find_next_bit(void *addr, int max, int start)
887 +#if BITS_PER_LONG == 64
888 + fix = ((unsigned long) addr & 7UL) << 3;
889 + addr = (void *) ((unsigned long) addr & ~7UL);
890 +#elif BITS_PER_LONG == 32
891 + fix = ((unsigned long) addr & 3UL) << 3;
892 + addr = (void *) ((unsigned long) addr & ~3UL);
894 +#error "how many bits you are?!"
901 + return find_next_bit(addr, max, start) - fix;
905 +static inline void *mb_find_buddy(struct ext3_buddy *e3b, int order, int *max)
909 + BUG_ON(EXT3_MB_BITMAP(e3b) == EXT3_MB_BUDDY(e3b));
910 + BUG_ON(max == NULL);
912 + if (order > e3b->bd_blkbits + 1) {
917 + /* at order 0 we see each particular block */
918 + *max = 1 << (e3b->bd_blkbits + 3);
920 + return EXT3_MB_BITMAP(e3b);
922 + bb = EXT3_MB_BUDDY(e3b) + EXT3_SB(e3b->bd_sb)->s_mb_offsets[order];
923 + *max = EXT3_SB(e3b->bd_sb)->s_mb_maxs[order];
929 +void mb_free_blocks_double(struct inode *inode, struct ext3_buddy *e3b,
930 + int first, int count)
933 + struct super_block *sb = e3b->bd_sb;
935 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
937 + BUG_ON(!ext3_is_group_locked(sb, e3b->bd_group));
938 + for (i = 0; i < count; i++) {
939 + if (!mb_test_bit(first + i, e3b->bd_info->bb_bitmap)) {
940 + unsigned long blocknr;
941 + blocknr = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb);
942 + blocknr += first + i;
944 + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
946 + ext3_error(sb, __FUNCTION__, "double-free of inode"
947 + " %lu's block %lu(bit %u in group %u)\n",
948 + inode ? inode->i_ino : 0, blocknr,
949 + first + i, e3b->bd_group);
951 + mb_clear_bit(first + i, e3b->bd_info->bb_bitmap);
955 +void mb_mark_used_double(struct ext3_buddy *e3b, int first, int count)
958 + if (unlikely(e3b->bd_info->bb_bitmap == NULL))
960 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
961 + for (i = 0; i < count; i++) {
962 + BUG_ON(mb_test_bit(first + i, e3b->bd_info->bb_bitmap));
963 + mb_set_bit(first + i, e3b->bd_info->bb_bitmap);
967 +void mb_cmp_bitmaps(struct ext3_buddy *e3b, void *bitmap)
969 + if (memcmp(e3b->bd_info->bb_bitmap, bitmap, e3b->bd_sb->s_blocksize)) {
970 + unsigned char *b1, *b2;
972 + b1 = (unsigned char *) e3b->bd_info->bb_bitmap;
973 + b2 = (unsigned char *) bitmap;
974 + for (i = 0; i < e3b->bd_sb->s_blocksize; i++) {
975 + if (b1[i] != b2[i]) {
976 + printk("corruption in group %u at byte %u(%u): "
977 + "%x in copy != %x on disk/prealloc\n",
978 + e3b->bd_group, i, i * 8, b1[i], b2[i]);
986 +#define mb_free_blocks_double(a,b,c,d)
987 +#define mb_mark_used_double(a,b,c)
988 +#define mb_cmp_bitmaps(a,b)
991 +#ifdef AGGRESSIVE_CHECK
993 +#define MB_CHECK_ASSERT(assert) \
996 + printk (KERN_EMERG \
997 + "Assertion failure in %s() at %s:%d: \"%s\"\n", \
998 + function, file, line, # assert); \
1003 +static int __mb_check_buddy(struct ext3_buddy *e3b, char *file,
1004 + const char *function, int line)
1006 + struct super_block *sb = e3b->bd_sb;
1007 + int order = e3b->bd_blkbits + 1;
1008 + int max, max2, i, j, k, count;
1009 + struct ext3_group_info *grp;
1010 + int fragments = 0, fstart;
1011 + struct list_head *cur;
1012 + void *buddy, *buddy2;
1014 + if (!test_opt(sb, MBALLOC))
1018 + static int mb_check_counter = 0;
1019 + if (mb_check_counter++ % 100 != 0)
1023 + while (order > 1) {
1024 + buddy = mb_find_buddy(e3b, order, &max);
1025 + MB_CHECK_ASSERT(buddy);
1026 + buddy2 = mb_find_buddy(e3b, order - 1, &max2);
1027 + MB_CHECK_ASSERT(buddy2);
1028 + MB_CHECK_ASSERT(buddy != buddy2);
1029 + MB_CHECK_ASSERT(max * 2 == max2);
1032 + for (i = 0; i < max; i++) {
1034 + if (mb_test_bit(i, buddy)) {
1035 + /* only single bit in buddy2 may be 1 */
1036 + if (!mb_test_bit(i << 1, buddy2))
1037 + MB_CHECK_ASSERT(mb_test_bit((i<<1)+1, buddy2));
1038 + else if (!mb_test_bit((i << 1) + 1, buddy2))
1039 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
1043 + /* both bits in buddy2 must be 0 */
1044 + MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
1045 + MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
1047 + for (j = 0; j < (1 << order); j++) {
1048 + k = (i * (1 << order)) + j;
1049 + MB_CHECK_ASSERT(!mb_test_bit(k, EXT3_MB_BITMAP(e3b)));
1053 + MB_CHECK_ASSERT(e3b->bd_info->bb_counters[order] == count);
1058 + buddy = mb_find_buddy(e3b, 0, &max);
1059 + for (i = 0; i < max; i++) {
1060 + if (!mb_test_bit(i, buddy)) {
1061 + MB_CHECK_ASSERT(i >= e3b->bd_info->bb_first_free);
1062 + if (fstart == -1) {
1069 + /* check used bits only */
1070 + for (j = 0; j < e3b->bd_blkbits + 1; j++) {
1071 + buddy2 = mb_find_buddy(e3b, j, &max2);
1073 + MB_CHECK_ASSERT(k < max2);
1074 + MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
1077 + MB_CHECK_ASSERT(!EXT3_MB_GRP_NEED_INIT(e3b->bd_info));
1078 + MB_CHECK_ASSERT(e3b->bd_info->bb_fragments == fragments);
1080 + grp = EXT3_GROUP_INFO(sb, e3b->bd_group);
1081 + buddy = mb_find_buddy(e3b, 0, &max);
1082 + list_for_each(cur, &grp->bb_prealloc_list) {
1083 + unsigned long groupnr;
1084 + struct ext3_prealloc_space *pa;
1085 + pa = list_entry(cur, struct ext3_prealloc_space, group_list);
1086 + ext3_get_group_no_and_offset(sb, pa->pstart, &groupnr, &k);
1087 + MB_CHECK_ASSERT(groupnr == e3b->bd_group);
1088 + for (i = 0; i < pa->len; i++)
1089 + MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
1093 +#undef MB_CHECK_ASSERT
1094 +#define mb_check_buddy(e3b) __mb_check_buddy(e3b,__FILE__,__FUNCTION__,__LINE__)
1096 +#define mb_check_buddy(e3b)
1099 +/* find most significant bit */
1100 +static int inline fmsb(unsigned short word)
1114 + } while (word != 0);
1120 +ext3_mb_mark_free_simple(struct super_block *sb, void *buddy, unsigned first,
1121 + int len, struct ext3_group_info *grp)
1123 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1124 + unsigned short min, max, chunk, border;
1126 + BUG_ON(len >= EXT3_BLOCKS_PER_GROUP(sb));
1128 + border = 2 << sb->s_blocksize_bits;
1131 + /* find how many blocks can be covered since this position */
1132 + max = ffs(first | border) - 1;
1134 + /* find how many blocks of power 2 we need to mark */
1141 + /* mark multiblock chunks only */
1142 + grp->bb_counters[min]++;
1144 + mb_clear_bit(first >> min, buddy + sbi->s_mb_offsets[min]);
1152 +ext3_mb_generate_buddy(struct super_block *sb, void *buddy, void *bitmap,
1155 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
1156 + unsigned short max = EXT3_BLOCKS_PER_GROUP(sb);
1157 + unsigned short i = 0, first, len;
1158 + unsigned free = 0, fragments = 0;
1159 + unsigned long long period = get_cycles();
1161 + /* initialize buddy from bitmap which is aggregation
1162 + * of on-disk bitmap and preallocations */
1163 + i = mb_find_next_zero_bit(bitmap, max, 0);
1164 + grp->bb_first_free = i;
1168 + i = ext2_find_next_le_bit(bitmap, max, i);
1174 + ext3_mb_mark_free_simple(sb, buddy, first, len, grp);
1176 + grp->bb_counters[0]++;
1178 + i = mb_find_next_zero_bit(bitmap, max, i);
1180 + grp->bb_fragments = fragments;
1182 + if (free != grp->bb_free) {
1183 + struct ext3_group_desc *gdp;
1184 + gdp = ext3_get_group_desc (sb, group, NULL);
1185 + ext3_error(sb, __FUNCTION__,
1186 + "group %u: %u blocks in bitmap, %u in bb, "
1187 + "%u in gd, %lu pa's\n", group, free, grp->bb_free,
1188 + le16_to_cpu(gdp->bg_free_blocks_count),
1189 + grp->bb_prealloc_nr);
1193 + clear_bit(EXT3_GROUP_INFO_NEED_INIT_BIT, &grp->bb_state);
1195 + period = get_cycles() - period;
1196 + spin_lock(&EXT3_SB(sb)->s_bal_lock);
1197 + EXT3_SB(sb)->s_mb_buddies_generated++;
1198 + EXT3_SB(sb)->s_mb_generation_time += period;
1199 + spin_unlock(&EXT3_SB(sb)->s_bal_lock);
1204 +static int ext3_mb_init_cache(struct page *page, char *incore)
1206 + int blocksize, blocks_per_page, groups_per_page;
1207 + int err = 0, i, first_group, first_block;
1208 + struct super_block *sb;
1209 + struct buffer_head *bhs;
1210 + struct buffer_head **bh;
1211 + struct inode *inode;
1212 + char *data, *bitmap;
1214 + mb_debug("init page %lu\n", page->index);
1216 + inode = page->mapping->host;
1218 + blocksize = 1 << inode->i_blkbits;
1219 + blocks_per_page = PAGE_CACHE_SIZE / blocksize;
1221 + groups_per_page = blocks_per_page >> 1;
1222 + if (groups_per_page == 0)
1223 + groups_per_page = 1;
1225 + /* allocate buffer_heads to read bitmaps */
1226 + if (groups_per_page > 1) {
1228 + i = sizeof(struct buffer_head *) * groups_per_page;
1229 + bh = kmalloc(i, GFP_NOFS);
1236 + first_group = page->index * blocks_per_page / 2;
1238 + /* read all groups the page covers into the cache */
1239 + for (i = 0; i < groups_per_page; i++) {
1240 + struct ext3_group_desc * desc;
1242 + if (first_group + i >= EXT3_SB(sb)->s_groups_count)
1246 + desc = ext3_get_group_desc(sb, first_group + i, NULL);
1251 + bh[i] = sb_getblk(sb, le32_to_cpu(desc->bg_block_bitmap));
1252 + if (bh[i] == NULL)
1255 + if (buffer_uptodate(bh[i]))
1258 + lock_buffer(bh[i]);
1259 + if (buffer_uptodate(bh[i])) {
1260 + unlock_buffer(bh[i]);
1265 + bh[i]->b_end_io = end_buffer_read_sync;
1266 + submit_bh(READ, bh[i]);
1267 + mb_debug("read bitmap for group %u\n", first_group + i);
1270 + /* wait for I/O completion */
1271 + for (i = 0; i < groups_per_page && bh[i]; i++)
1272 + wait_on_buffer(bh[i]);
1275 + for (i = 0; i < groups_per_page && bh[i]; i++)
1276 + if (!buffer_uptodate(bh[i]))
1280 + first_block = page->index * blocks_per_page;
1281 + for (i = 0; i < blocks_per_page && err == 0; i++) {
1284 + group = (first_block + i) >> 1;
1285 + if (group >= EXT3_SB(sb)->s_groups_count)
1288 + data = page_address(page) + (i * blocksize);
1289 + bitmap = bh[group - first_group]->b_data;
1291 + if ((first_block + i) & 1) {
1292 + /* this is block of buddy */
1293 + BUG_ON(incore == NULL);
1294 + mb_debug("put buddy for group %u in page %lu/%x\n",
1295 + group, page->index, i * blocksize);
1296 + memset(data, 0xff, blocksize);
1297 + EXT3_GROUP_INFO(sb, group)->bb_fragments = 0;
1298 + memset(EXT3_GROUP_INFO(sb, group)->bb_counters, 0,
1299 + sizeof(unsigned short)*(sb->s_blocksize_bits+2));
1300 + err = ext3_mb_generate_buddy(sb, data, incore, group);
1303 + /* this is block of bitmap */
1304 + BUG_ON(incore != NULL);
1305 + mb_debug("put bitmap for group %u in page %lu/%x\n",
1306 + group, page->index, i * blocksize);
1308 + /* see comments in ext3_mb_put_pa() */
1309 + ext3_lock_group(sb, group);
1310 + memcpy(data, bitmap, blocksize);
1312 + /* mark all preallocated blocks used in in-core bitmap */
1313 + err = ext3_mb_generate_from_pa(sb, data, group);
1314 + ext3_unlock_group(sb, group);
1319 + if (likely(err == 0))
1320 + SetPageUptodate(page);
1324 + for (i = 0; i < groups_per_page && bh[i]; i++)
1332 +static int ext3_mb_load_buddy(struct super_block *sb, int group,
1333 + struct ext3_buddy *e3b)
1335 + struct ext3_sb_info *sbi = EXT3_SB(sb);
1336 + struct inode *inode = sbi->s_buddy_cache;
1337 + int blocks_per_page, block, pnum, poff;
1338 + struct page *page;
1340 + mb_debug("load group %u\n", group);
1342 + blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1344 + e3b->bd_blkbits = sb->s_blocksize_bits;
1345 + e3b->bd_info = EXT3_GROUP_INFO(sb, group);
1347 + e3b->bd_group = group;
1348 + e3b->bd_buddy_page = NULL;
1349 + e3b->bd_bitmap_page = NULL;
1351 + block = group * 2;
1352 + pnum = block / blocks_per_page;
1353 + poff = block % blocks_per_page;
1355 + /* we could use find_or_create_page(), but it locks page
1356 + * what we'd like to avoid in fast path ... */
1357 + page = find_get_page(inode->i_mapping, pnum);
1358 + if (page == NULL || !PageUptodate(page)) {
1360 + page_cache_release(page);
1361 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1363 + BUG_ON(page->mapping != inode->i_mapping);
1364 + if (!PageUptodate(page)) {
1365 + ext3_mb_init_cache(page, NULL);
1366 + mb_cmp_bitmaps(e3b, page_address(page) +
1367 + (poff * sb->s_blocksize));
1369 + unlock_page(page);
1372 + if (page == NULL || !PageUptodate(page))
1374 + e3b->bd_bitmap_page = page;
1375 + e3b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1376 + mark_page_accessed(page);
1379 + pnum = block / blocks_per_page;
1380 + poff = block % blocks_per_page;
1382 + page = find_get_page(inode->i_mapping, pnum);
1383 + if (page == NULL || !PageUptodate(page)) {
1385 + page_cache_release(page);
1386 + page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1388 + BUG_ON(page->mapping != inode->i_mapping);
1389 + if (!PageUptodate(page))
1390 + ext3_mb_init_cache(page, e3b->bd_bitmap);
1392 + unlock_page(page);
1395 + if (page == NULL || !PageUptodate(page))
1397 + e3b->bd_buddy_page = page;
1398 + e3b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1399 + mark_page_accessed(page);
1401 + BUG_ON(e3b->bd_bitmap_page == NULL);
1402 + BUG_ON(e3b->bd_buddy_page == NULL);
1407 + if (e3b->bd_bitmap_page)
1408 + page_cache_release(e3b->bd_bitmap_page);
1409 + if (e3b->bd_buddy_page)
1410 + page_cache_release(e3b->bd_buddy_page);
1411 + e3b->bd_buddy = NULL;
1412 + e3b->bd_bitmap = NULL;
1416 +static void ext3_mb_release_desc(struct ext3_buddy *e3b)
1418 + if (e3b->bd_bitmap_page)
1419 + page_cache_release(e3b->bd_bitmap_page);
1420 + if (e3b->bd_buddy_page)
1421 + page_cache_release(e3b->bd_buddy_page);
1425 +static int mb_find_order_for_block(struct ext3_buddy *e3b, int block)
1430 + BUG_ON(EXT3_MB_BITMAP(e3b) == EXT3_MB_BUDDY(e3b));
1431 + BUG_ON(block >= (1 << (e3b->bd_blkbits + 3)));
1433 + bb = EXT3_MB_BUDDY(e3b);
1434 + while (order <= e3b->bd_blkbits + 1) {
1435 + block = block >> 1;
1436 + if (!mb_test_bit(block, bb)) {
1437 + /* this block is part of buddy of order 'order' */
1440 + bb += 1 << (e3b->bd_blkbits - order);
1446 +static inline void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1451 + while (cur < len) {
1452 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1453 + /* fast path: clear whole word at once */
1454 + addr = bm + (cur >> 3);
1459 + mb_clear_bit_atomic(lock, cur, bm);
1464 +static inline void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1469 + while (cur < len) {
1470 + if ((cur & 31) == 0 && (len - cur) >= 32) {
1471 + /* fast path: clear whole word at once */
1472 + addr = bm + (cur >> 3);
1473 + *addr = 0xffffffff;
1477 + mb_set_bit_atomic(lock, cur, bm);
1482 +static int mb_free_blocks(struct inode *inode, struct ext3_buddy *e3b,
1483 + int first, int count)
1485 + int block = 0, max = 0, order;
1486 + void *buddy, *buddy2;
1487 + struct super_block *sb = e3b->bd_sb;
1489 + BUG_ON(first + count > (sb->s_blocksize << 3));
1490 + BUG_ON(!ext3_is_group_locked(sb, e3b->bd_group));
1491 + mb_check_buddy(e3b);
1492 + mb_free_blocks_double(inode, e3b, first, count);
1494 + e3b->bd_info->bb_free += count;
1495 + if (first < e3b->bd_info->bb_first_free)
1496 + e3b->bd_info->bb_first_free = first;
1498 + /* let's maintain fragments counter */
1500 + block = !mb_test_bit(first - 1, EXT3_MB_BITMAP(e3b));
1501 + if (first + count < EXT3_SB(sb)->s_mb_maxs[0])
1502 + max = !mb_test_bit(first + count, EXT3_MB_BITMAP(e3b));
1504 + e3b->bd_info->bb_fragments--;
1505 + else if (!block && !max)
1506 + e3b->bd_info->bb_fragments++;
1508 + /* let's maintain buddy itself */
1509 + while (count-- > 0) {
1513 + if (!mb_test_bit(block, EXT3_MB_BITMAP(e3b))) {
1514 + unsigned long blocknr;
1515 + blocknr = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb);
1518 + le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block);
1520 + ext3_error(sb, __FUNCTION__, "double-free of inode"
1521 + " %lu's block %lu(bit %u in group %u)\n",
1522 + inode ? inode->i_ino : 0, blocknr, block,
1525 + mb_clear_bit(block, EXT3_MB_BITMAP(e3b));
1526 + e3b->bd_info->bb_counters[order]++;
1528 + /* start of the buddy */
1529 + buddy = mb_find_buddy(e3b, order, &max);
1533 + if (mb_test_bit(block, buddy) ||
1534 + mb_test_bit(block + 1, buddy))
1537 + /* both the buddies are free, try to coalesce them */
1538 + buddy2 = mb_find_buddy(e3b, order + 1, &max);
1544 + /* for special purposes, we don't set
1545 + * free bits in bitmap */
1546 + mb_set_bit(block, buddy);
1547 + mb_set_bit(block + 1, buddy);
1549 + e3b->bd_info->bb_counters[order]--;
1550 + e3b->bd_info->bb_counters[order]--;
1552 + block = block >> 1;
1554 + e3b->bd_info->bb_counters[order]++;
1556 + mb_clear_bit(block, buddy2);
1560 + mb_check_buddy(e3b);
1565 +static int mb_find_extent(struct ext3_buddy *e3b, int order, int block,
1566 + int needed, struct ext3_free_extent *ex)
1568 + int next = block, max, ord;
1571 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1572 + BUG_ON(ex == NULL);
1574 + buddy = mb_find_buddy(e3b, order, &max);
1575 + BUG_ON(buddy == NULL);
1576 + BUG_ON(block >= max);
1577 + if (mb_test_bit(block, buddy)) {
1584 + if (likely(order == 0)) {
1585 + /* find actual order */
1586 + order = mb_find_order_for_block(e3b, block);
1587 + block = block >> order;
1590 + ex->fe_len = 1 << order;
1591 + ex->fe_start = block << order;
1592 + ex->fe_group = e3b->bd_group;
1594 + /* calc difference from given start */
1595 + next = next - ex->fe_start;
1596 + ex->fe_len -= next;
1597 + ex->fe_start += next;
1599 + while (needed > ex->fe_len && (buddy = mb_find_buddy(e3b, order, &max))) {
1601 + if (block + 1 >= max)
1604 + next = (block + 1) * (1 << order);
1605 + if (mb_test_bit(next, EXT3_MB_BITMAP(e3b)))
1608 + ord = mb_find_order_for_block(e3b, next);
1611 + block = next >> order;
1612 + ex->fe_len += 1 << order;
1615 + BUG_ON(ex->fe_start + ex->fe_len > (1 << (e3b->bd_blkbits + 3)));
1616 + return ex->fe_len;
1619 +static int mb_mark_used(struct ext3_buddy *e3b, struct ext3_free_extent *ex)
1621 + int ord, mlen = 0, max = 0, cur;
1622 + int start = ex->fe_start;
1623 + int len = ex->fe_len;
1628 + BUG_ON(start + len > (e3b->bd_sb->s_blocksize << 3));
1629 + BUG_ON(e3b->bd_group != ex->fe_group);
1630 + BUG_ON(!ext3_is_group_locked(e3b->bd_sb, e3b->bd_group));
1631 + mb_check_buddy(e3b);
1632 + mb_mark_used_double(e3b, start, len);
1634 + e3b->bd_info->bb_free -= len;
1635 + if (e3b->bd_info->bb_first_free == start)
1636 + e3b->bd_info->bb_first_free += len;
1638 + /* let's maintain fragments counter */
1640 + mlen = !mb_test_bit(start - 1, EXT3_MB_BITMAP(e3b));
1641 + if (start + len < EXT3_SB(e3b->bd_sb)->s_mb_maxs[0])
1642 + max = !mb_test_bit(start + len, EXT3_MB_BITMAP(e3b));
1644 + e3b->bd_info->bb_fragments++;
1645 + else if (!mlen && !max)
1646 + e3b->bd_info->bb_fragments--;
1648 + /* let's maintain buddy itself */
1650 + ord = mb_find_order_for_block(e3b, start);
1652 + if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1653 + /* the whole chunk may be allocated at once! */
1655 + buddy = mb_find_buddy(e3b, ord, &max);
1656 + BUG_ON((start >> ord) >= max);
1657 + mb_set_bit(start >> ord, buddy);
1658 + e3b->bd_info->bb_counters[ord]--;
1665 + /* store for history */
1667 + ret = len | (ord << 16);
1669 + /* we have to split large buddy */
1671 + buddy = mb_find_buddy(e3b, ord, &max);
1672 + mb_set_bit(start >> ord, buddy);
1673 + e3b->bd_info->bb_counters[ord]--;
1676 + cur = (start >> ord) & ~1U;
1677 + buddy = mb_find_buddy(e3b, ord, &max);
1678 + mb_clear_bit(cur, buddy);
1679 + mb_clear_bit(cur + 1, buddy);
1680 + e3b->bd_info->bb_counters[ord]++;
1681 + e3b->bd_info->bb_counters[ord]++;
1684 + mb_set_bits(sb_bgl_lock(EXT3_SB(e3b->bd_sb), ex->fe_group),
1685 + EXT3_MB_BITMAP(e3b), ex->fe_start, len0);
1686 + mb_check_buddy(e3b);
1692 + * Must be called under group lock!
1694 +static void ext3_mb_use_best_found(struct ext3_allocation_context *ac,
1695 + struct ext3_buddy *e3b)
1697 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1698 + unsigned long ret;
1700 + BUG_ON(ac->ac_b_ex.fe_group != e3b->bd_group);
1701 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1703 + ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1704 + ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1705 + ret = mb_mark_used(e3b, &ac->ac_b_ex);
1707 + /* preallocation can change ac_b_ex, thus we store actually
1708 + * allocated blocks for history */
1709 + ac->ac_f_ex = ac->ac_b_ex;
1711 + ac->ac_status = AC_STATUS_FOUND;
1712 + ac->ac_tail = ret & 0xffff;
1713 + ac->ac_buddy = ret >> 16;
1715 + /* XXXXXXX: SUCH A HORRIBLE **CK */
1716 + ac->ac_bitmap_page = e3b->bd_bitmap_page;
1717 + get_page(ac->ac_bitmap_page);
1718 + ac->ac_buddy_page = e3b->bd_buddy_page;
1719 + get_page(ac->ac_buddy_page);
1721 + /* store last allocated for subsequent stream allocation */
1722 + if ((ac->ac_flags & EXT3_MB_HINT_DATA)) {
1723 + spin_lock(&sbi->s_md_lock);
1724 + sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1725 + sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1726 + spin_unlock(&sbi->s_md_lock);
1731 + * regular allocator, for general purposes allocation
1734 +void ext3_mb_check_limits(struct ext3_allocation_context *ac,
1735 + struct ext3_buddy *e3b,
1738 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1739 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1740 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1741 + struct ext3_free_extent ex;
1745 + * We don't want to scan for a whole year
1747 + if (ac->ac_found > sbi->s_mb_max_to_scan &&
1748 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1749 + ac->ac_status = AC_STATUS_BREAK;
1754 + * Haven't found good chunk so far, let's continue
1756 + if (bex->fe_len < gex->fe_len)
1759 + if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1760 + && bex->fe_group == e3b->bd_group) {
1761 + /* recheck chunk's availability - we don't know
1762 + * when it was found (within this lock-unlock
1763 + * period or not) */
1764 + max = mb_find_extent(e3b, 0, bex->fe_start, gex->fe_len, &ex);
1765 + if (max >= gex->fe_len) {
1766 + ext3_mb_use_best_found(ac, e3b);
1773 + * The routine checks whether found extent is good enough. If it is,
1774 + * then the extent gets marked used and flag is set to the context
1775 + * to stop scanning. Otherwise, the extent is compared with the
1776 + * previous found extent and if new one is better, then it's stored
1777 + * in the context. Later, the best found extent will be used, if
1778 + * mballoc can't find good enough extent.
1780 + * FIXME: real allocation policy is to be designed yet!
1782 +static void ext3_mb_measure_extent(struct ext3_allocation_context *ac,
1783 + struct ext3_free_extent *ex,
1784 + struct ext3_buddy *e3b)
1786 + struct ext3_free_extent *bex = &ac->ac_b_ex;
1787 + struct ext3_free_extent *gex = &ac->ac_g_ex;
1789 + BUG_ON(ex->fe_len <= 0);
1790 + BUG_ON(ex->fe_len >= EXT3_BLOCKS_PER_GROUP(ac->ac_sb));
1791 + BUG_ON(ex->fe_start >= EXT3_BLOCKS_PER_GROUP(ac->ac_sb));
1792 + BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1797 + * The special case - take what you catch first
1799 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
1801 + ext3_mb_use_best_found(ac, e3b);
1806 + * Let's check whether the chuck is good enough
1808 + if (ex->fe_len == gex->fe_len) {
1810 + ext3_mb_use_best_found(ac, e3b);
1815 + * If this is first found extent, just store it in the context
1817 + if (bex->fe_len == 0) {
1823 + * If new found extent is better, store it in the context
1825 + if (bex->fe_len < gex->fe_len) {
1826 + /* if the request isn't satisfied, any found extent
1827 + * larger than previous best one is better */
1828 + if (ex->fe_len > bex->fe_len)
1830 + } else if (ex->fe_len > gex->fe_len) {
1831 + /* if the request is satisfied, then we try to find
1832 + * an extent that still satisfy the request, but is
1833 + * smaller than previous one */
1834 + if (ex->fe_len < bex->fe_len)
1838 + ext3_mb_check_limits(ac, e3b, 0);
1841 +static int ext3_mb_try_best_found(struct ext3_allocation_context *ac,
1842 + struct ext3_buddy *e3b)
1844 + struct ext3_free_extent ex = ac->ac_b_ex;
1845 + int group = ex.fe_group, max, err;
1847 + BUG_ON(ex.fe_len <= 0);
1848 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1852 + ext3_lock_group(ac->ac_sb, group);
1853 + max = mb_find_extent(e3b, 0, ex.fe_start, ex.fe_len, &ex);
1857 + ext3_mb_use_best_found(ac, e3b);
1860 + ext3_unlock_group(ac->ac_sb, group);
1861 + ext3_mb_release_desc(e3b);
1866 +static int ext3_mb_find_by_goal(struct ext3_allocation_context *ac,
1867 + struct ext3_buddy *e3b)
1869 + int group = ac->ac_g_ex.fe_group, max, err;
1870 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
1871 + struct ext3_super_block *es = sbi->s_es;
1872 + struct ext3_free_extent ex;
1874 + if (!(ac->ac_flags & EXT3_MB_HINT_TRY_GOAL))
1877 + err = ext3_mb_load_buddy(ac->ac_sb, group, e3b);
1881 + ext3_lock_group(ac->ac_sb, group);
1882 + max = mb_find_extent(e3b, 0, ac->ac_g_ex.fe_start,
1883 + ac->ac_g_ex.fe_len, &ex);
1885 + if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1886 + unsigned long start;
1887 + start = (e3b->bd_group * EXT3_BLOCKS_PER_GROUP(ac->ac_sb) +
1888 + ex.fe_start + le32_to_cpu(es->s_first_data_block));
1889 + if (start % sbi->s_stripe == 0) {
1892 + ext3_mb_use_best_found(ac, e3b);
1894 + } else if (max >= ac->ac_g_ex.fe_len) {
1895 + BUG_ON(ex.fe_len <= 0);
1896 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1897 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1900 + ext3_mb_use_best_found(ac, e3b);
1901 + } else if (max > 0 && (ac->ac_flags & EXT3_MB_HINT_MERGE)) {
1902 + /* Sometimes, caller may want to merge even small
1903 + * number of blocks to an existing extent */
1904 + BUG_ON(ex.fe_len <= 0);
1905 + BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1906 + BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1909 + ext3_mb_use_best_found(ac, e3b);
1911 + ext3_unlock_group(ac->ac_sb, group);
1912 + ext3_mb_release_desc(e3b);
1918 + * The routine scans buddy structures (not bitmap!) from given order
1919 + * to max order and tries to find big enough chunk to satisfy the req
1921 +static void ext3_mb_simple_scan_group(struct ext3_allocation_context *ac,
1922 + struct ext3_buddy *e3b)
1924 + struct super_block *sb = ac->ac_sb;
1925 + struct ext3_group_info *grp = e3b->bd_info;
1929 + BUG_ON(ac->ac_2order <= 0);
1930 + for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1931 + if (grp->bb_counters[i] == 0)
1934 + buddy = mb_find_buddy(e3b, i, &max);
1935 + BUG_ON(buddy == NULL);
1937 + k = mb_find_next_zero_bit(buddy, max, 0);
1942 + ac->ac_b_ex.fe_len = 1 << i;
1943 + ac->ac_b_ex.fe_start = k << i;
1944 + ac->ac_b_ex.fe_group = e3b->bd_group;
1946 + ext3_mb_use_best_found(ac, e3b);
1948 + BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1950 + if (EXT3_SB(sb)->s_mb_stats)
1951 + atomic_inc(&EXT3_SB(sb)->s_bal_2orders);
1958 + * The routine scans the group and measures all found extents.
1959 + * In order to optimize scanning, caller must pass number of
1960 + * free blocks in the group, so the routine can know upper limit.
1962 +static void ext3_mb_complex_scan_group(struct ext3_allocation_context *ac,
1963 + struct ext3_buddy *e3b)
1965 + struct super_block *sb = ac->ac_sb;
1966 + void *bitmap = EXT3_MB_BITMAP(e3b);
1967 + struct ext3_free_extent ex;
1970 + free = e3b->bd_info->bb_free;
1971 + BUG_ON(free <= 0);
1973 + i = e3b->bd_info->bb_first_free;
1975 + while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1976 + i = mb_find_next_zero_bit(bitmap, EXT3_BLOCKS_PER_GROUP(sb), i);
1977 + if (i >= EXT3_BLOCKS_PER_GROUP(sb)) {
1978 + BUG_ON(free != 0);
1982 + mb_find_extent(e3b, 0, i, ac->ac_g_ex.fe_len, &ex);
1983 + BUG_ON(ex.fe_len <= 0);
1984 + BUG_ON(free < ex.fe_len);
1986 + ext3_mb_measure_extent(ac, &ex, e3b);
1989 + free -= ex.fe_len;
1992 + ext3_mb_check_limits(ac, e3b, 1);
1996 + * This is a special case for storages like raid5
1997 + * we try to find stripe-aligned chunks for stripe-size requests
1999 +static void ext3_mb_scan_aligned(struct ext3_allocation_context *ac,
2000 + struct ext3_buddy *e3b)
2002 + struct super_block *sb = ac->ac_sb;
2003 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2004 + void *bitmap = EXT3_MB_BITMAP(e3b);
2005 + struct ext3_free_extent ex;
2006 + unsigned long i, max;
2008 + BUG_ON(sbi->s_stripe == 0);
2010 + /* find first stripe-aligned block */
2011 + i = e3b->bd_group * EXT3_BLOCKS_PER_GROUP(sb)
2012 + + le32_to_cpu(sbi->s_es->s_first_data_block);
2013 + i = ((i + sbi->s_stripe - 1) / sbi->s_stripe) * sbi->s_stripe;
2014 + i = (i - le32_to_cpu(sbi->s_es->s_first_data_block))
2015 + % EXT3_BLOCKS_PER_GROUP(sb);
2017 + while (i < EXT3_BLOCKS_PER_GROUP(sb)) {
2018 + if (!mb_test_bit(i, bitmap)) {
2019 + max = mb_find_extent(e3b, 0, i, sbi->s_stripe, &ex);
2020 + if (max >= sbi->s_stripe) {
2023 + ext3_mb_use_best_found(ac, e3b);
2027 + i += sbi->s_stripe;
2031 +static int ext3_mb_good_group(struct ext3_allocation_context *ac,
2032 + int group, int cr)
2034 + struct ext3_group_info *grp = EXT3_GROUP_INFO(ac->ac_sb, group);
2035 + unsigned free, fragments, i, bits;
2037 + BUG_ON(cr < 0 || cr >= 4);
2038 + BUG_ON(EXT3_MB_GRP_NEED_INIT(grp));
2040 + free = grp->bb_free;
2041 + fragments = grp->bb_fragments;
2044 + if (fragments == 0)
2049 + BUG_ON(ac->ac_2order == 0);
2050 + bits = ac->ac_sb->s_blocksize_bits + 1;
2051 + for (i = ac->ac_2order; i <= bits; i++)
2052 + if (grp->bb_counters[i] > 0)
2056 + if ((free / fragments) >= ac->ac_g_ex.fe_len)
2060 + if (free >= ac->ac_g_ex.fe_len)
2072 +int ext3_mb_regular_allocator(struct ext3_allocation_context *ac)
2074 + int group, i, cr, err = 0;
2075 + struct ext3_sb_info *sbi;
2076 + struct super_block *sb;
2077 + struct ext3_buddy e3b;
2080 + sbi = EXT3_SB(sb);
2081 + BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2083 + /* first, try the goal */
2084 + err = ext3_mb_find_by_goal(ac, &e3b);
2085 + if (err || ac->ac_status == AC_STATUS_FOUND)
2088 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
2091 + i = ffs(ac->ac_g_ex.fe_len);
2092 + ac->ac_2order = 0;
2093 + if (i >= sbi->s_mb_order2_reqs) {
2095 + if ((ac->ac_g_ex.fe_len & (~(1 << i))) == 0)
2096 + ac->ac_2order = i;
2099 + /* if stream allocation is enabled, use global goal */
2100 + if ((ac->ac_g_ex.fe_len < sbi->s_mb_large_req) &&
2101 + (ac->ac_flags & EXT3_MB_HINT_DATA)) {
2102 + /* TBD: may be hot point */
2103 + spin_lock(&sbi->s_md_lock);
2104 + ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2105 + ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2106 + spin_unlock(&sbi->s_md_lock);
2109 + group = ac->ac_g_ex.fe_group;
2111 + /* Let's just scan groups to find more-less suitable blocks */
2112 + cr = ac->ac_2order ? 0 : 1;
2114 + for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2115 + ac->ac_criteria = cr;
2116 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; group++, i++) {
2117 + struct ext3_group_info *grp;
2119 + if (group == EXT3_SB(sb)->s_groups_count)
2122 + /* quick check to skip empty groups */
2123 + grp = EXT3_GROUP_INFO(ac->ac_sb, group);
2124 + if (grp->bb_free == 0)
2127 + if (EXT3_MB_GRP_NEED_INIT(EXT3_GROUP_INFO(sb, group))) {
2128 + /* we need full data about the group
2129 + * to make a good selection */
2130 + err = ext3_mb_load_buddy(sb, group, &e3b);
2133 + ext3_mb_release_desc(&e3b);
2136 + /* check is group good for our criteries */
2137 + if (!ext3_mb_good_group(ac, group, cr))
2140 + err = ext3_mb_load_buddy(sb, group, &e3b);
2144 + ext3_lock_group(sb, group);
2145 + if (!ext3_mb_good_group(ac, group, cr)) {
2146 + /* someone did allocation from this group */
2147 + ext3_unlock_group(sb, group);
2148 + ext3_mb_release_desc(&e3b);
2152 + ac->ac_groups_scanned++;
2154 + ext3_mb_simple_scan_group(ac, &e3b);
2155 + else if (cr == 1 && ac->ac_g_ex.fe_len == sbi->s_stripe)
2156 + ext3_mb_scan_aligned(ac, &e3b);
2158 + ext3_mb_complex_scan_group(ac, &e3b);
2160 + ext3_unlock_group(sb, group);
2161 + ext3_mb_release_desc(&e3b);
2163 + if (ac->ac_status != AC_STATUS_CONTINUE)
2168 + if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2169 + !(ac->ac_flags & EXT3_MB_HINT_FIRST)) {
2171 + * We've been searching too long. Let's try to allocate
2172 + * the best chunk we've found so far
2175 + ext3_mb_try_best_found(ac, &e3b);
2176 + if (ac->ac_status != AC_STATUS_FOUND) {
2178 + * Someone more lucky has already allocated it.
2179 + * The only thing we can do is just take first
2181 + printk(KERN_DEBUG "EXT3-fs: someone won our chunk\n");
2183 + ac->ac_b_ex.fe_group = 0;
2184 + ac->ac_b_ex.fe_start = 0;
2185 + ac->ac_b_ex.fe_len = 0;
2186 + ac->ac_status = AC_STATUS_CONTINUE;
2187 + ac->ac_flags |= EXT3_MB_HINT_FIRST;
2189 + atomic_inc(&sbi->s_mb_lost_chunks);
2197 +#ifdef EXT3_MB_HISTORY
2198 +struct ext3_mb_proc_session {
2199 + struct ext3_mb_history *history;
2200 + struct super_block *sb;
2205 +static void *ext3_mb_history_skip_empty(struct ext3_mb_proc_session *s,
2206 + struct ext3_mb_history *hs,
2209 + if (hs == s->history + s->max)
2211 + if (!first && hs == s->history + s->start)
2213 + while (hs->orig.fe_len == 0) {
2215 + if (hs == s->history + s->max)
2217 + if (hs == s->history + s->start)
2223 +static void *ext3_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2225 + struct ext3_mb_proc_session *s = seq->private;
2226 + struct ext3_mb_history *hs;
2230 + return SEQ_START_TOKEN;
2231 + hs = ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2234 + while (--l && (hs = ext3_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2238 +static void *ext3_mb_seq_history_next(struct seq_file *seq, void *v, loff_t *pos)
2240 + struct ext3_mb_proc_session *s = seq->private;
2241 + struct ext3_mb_history *hs = v;
2244 + if (v == SEQ_START_TOKEN)
2245 + return ext3_mb_history_skip_empty(s, s->history + s->start, 1);
2247 + return ext3_mb_history_skip_empty(s, ++hs, 0);
2250 +static int ext3_mb_seq_history_show(struct seq_file *seq, void *v)
2252 + char buf[25], buf2[25], buf3[25], *fmt;
2253 + struct ext3_mb_history *hs = v;
2255 + if (v == SEQ_START_TOKEN) {
2256 + seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2257 + "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2258 + "pid", "inode", "original", "goal", "result","found",
2259 + "grps", "cr", "flags", "merge", "tail", "broken");
2263 + if (hs->op == EXT3_MB_HISTORY_ALLOC) {
2264 + fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2265 + "%-5u %-5s %-5u %-6u\n";
2266 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2267 + hs->result.fe_start, hs->result.fe_len,
2268 + hs->result.fe_logical);
2269 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2270 + hs->orig.fe_start, hs->orig.fe_len,
2271 + hs->orig.fe_logical);
2272 + sprintf(buf3, "%lu/%lu/%lu@%lu", hs->goal.fe_group,
2273 + hs->goal.fe_start, hs->goal.fe_len,
2274 + hs->goal.fe_logical);
2275 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2276 + hs->found, hs->groups, hs->cr, hs->flags,
2277 + hs->merged ? "M" : "", hs->tail,
2278 + hs->buddy ? 1 << hs->buddy : 0);
2279 + } else if (hs->op == EXT3_MB_HISTORY_PREALLOC) {
2280 + fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2281 + sprintf(buf2, "%lu/%lu/%lu@%lu", hs->result.fe_group,
2282 + hs->result.fe_start, hs->result.fe_len,
2283 + hs->result.fe_logical);
2284 + sprintf(buf, "%lu/%lu/%lu@%lu", hs->orig.fe_group,
2285 + hs->orig.fe_start, hs->orig.fe_len,
2286 + hs->orig.fe_logical);
2287 + seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2288 + } else if (hs->op == EXT3_MB_HISTORY_DISCARD) {
2289 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2290 + hs->result.fe_start, hs->result.fe_len);
2291 + seq_printf(seq, "%-5u %-8u %-23s discard\n",
2292 + hs->pid, hs->ino, buf2);
2293 + } else if (hs->op == EXT3_MB_HISTORY_FREE) {
2294 + sprintf(buf2, "%lu/%lu/%lu", hs->result.fe_group,
2295 + hs->result.fe_start, hs->result.fe_len);
2296 + seq_printf(seq, "%-5u %-8u %-23s free\n",
2297 + hs->pid, hs->ino, buf2);
2299 + seq_printf(seq, "unknown op %d\n", hs->op);
2304 +static void ext3_mb_seq_history_stop(struct seq_file *seq, void *v)
2308 +static struct seq_operations ext3_mb_seq_history_ops = {
2309 + .start = ext3_mb_seq_history_start,
2310 + .next = ext3_mb_seq_history_next,
2311 + .stop = ext3_mb_seq_history_stop,
2312 + .show = ext3_mb_seq_history_show,
2315 +static int ext3_mb_seq_history_open(struct inode *inode, struct file *file)
2317 + struct super_block *sb = PDE(inode)->data;
2318 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2319 + struct ext3_mb_proc_session *s;
2322 + s = kmalloc(sizeof(*s), GFP_KERNEL);
2326 + size = sizeof(struct ext3_mb_history) * sbi->s_mb_history_max;
2327 + s->history = kmalloc(size, GFP_KERNEL);
2328 + if (s->history == NULL) {
2333 + spin_lock(&sbi->s_mb_history_lock);
2334 + memcpy(s->history, sbi->s_mb_history, size);
2335 + s->max = sbi->s_mb_history_max;
2336 + s->start = sbi->s_mb_history_cur % s->max;
2337 + spin_unlock(&sbi->s_mb_history_lock);
2339 + rc = seq_open(file, &ext3_mb_seq_history_ops);
2341 + struct seq_file *m = (struct seq_file *)file->private_data;
2344 + kfree(s->history);
2351 +static int ext3_mb_seq_history_release(struct inode *inode, struct file *file)
2353 + struct seq_file *seq = (struct seq_file *)file->private_data;
2354 + struct ext3_mb_proc_session *s = seq->private;
2355 + kfree(s->history);
2357 + return seq_release(inode, file);
2360 +static ssize_t ext3_mb_seq_history_write(struct file *file,
2361 + const char __user *buffer,
2362 + size_t count, loff_t *ppos)
2364 + struct seq_file *seq = (struct seq_file *)file->private_data;
2365 + struct ext3_mb_proc_session *s = seq->private;
2366 + struct super_block *sb = s->sb;
2370 + if (count >= sizeof(str)) {
2371 + printk(KERN_ERR "EXT3-fs: %s string too long, max %u bytes\n",
2372 + "mb_history", (int)sizeof(str));
2373 + return -EOVERFLOW;
2376 + if (copy_from_user(str, buffer, count))
2379 + value = simple_strtol(str, NULL, 0);
2382 + EXT3_SB(sb)->s_mb_history_filter = value;
2387 +static struct file_operations ext3_mb_seq_history_fops = {
2388 + .owner = THIS_MODULE,
2389 + .open = ext3_mb_seq_history_open,
2391 + .write = ext3_mb_seq_history_write,
2392 + .llseek = seq_lseek,
2393 + .release = ext3_mb_seq_history_release,
2396 +static void *ext3_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2398 + struct super_block *sb = seq->private;
2399 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2402 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2406 + return (void *) group;
2409 +static void *ext3_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2411 + struct super_block *sb = seq->private;
2412 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2416 + if (*pos < 0 || *pos >= sbi->s_groups_count)
2419 + return (void *) group;;
2422 +static int ext3_mb_seq_groups_show(struct seq_file *seq, void *v)
2424 + struct super_block *sb = seq->private;
2425 + struct ext3_group_desc *gdp;
2426 + long group = (long) v;
2427 + int i, err, free = 0;
2428 + struct ext3_buddy e3b;
2430 + struct ext3_group_info info;
2431 + unsigned short counters[16];
2436 + seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s "
2437 + "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2438 + "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2439 + "group", "free", "ingd", "frags", "first", "pa",
2440 + "2^0", "2^1", "2^2", "2^3", "2^4", "2^5","2^6",
2441 + "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2443 + i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2444 + sizeof(struct ext3_group_info);
2445 + err = ext3_mb_load_buddy(sb, group, &e3b);
2447 + seq_printf(seq, "#%-5lu: I/O error\n", group);
2451 + gdp = ext3_get_group_desc(sb, group, NULL);
2453 + free = le16_to_cpu(gdp->bg_free_blocks_count);
2455 + ext3_lock_group(sb, group);
2456 + memcpy(&sg, EXT3_GROUP_INFO(sb, group), i);
2457 + ext3_unlock_group(sb, group);
2458 + ext3_mb_release_desc(&e3b);
2460 + seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [", group,
2461 + sg.info.bb_free, free,
2462 + sg.info.bb_fragments, sg.info.bb_first_free,
2463 + sg.info.bb_prealloc_nr);
2464 + for (i = 0; i <= 13; i++)
2465 + seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2466 + sg.info.bb_counters[i] : 0);
2467 + seq_printf(seq, " ]\n");
2472 +static void ext3_mb_seq_groups_stop(struct seq_file *seq, void *v)
2476 +static struct seq_operations ext3_mb_seq_groups_ops = {
2477 + .start = ext3_mb_seq_groups_start,
2478 + .next = ext3_mb_seq_groups_next,
2479 + .stop = ext3_mb_seq_groups_stop,
2480 + .show = ext3_mb_seq_groups_show,
2483 +static int ext3_mb_seq_groups_open(struct inode *inode, struct file *file)
2485 + struct super_block *sb = PDE(inode)->data;
2488 + rc = seq_open(file, &ext3_mb_seq_groups_ops);
2490 + struct seq_file *m = (struct seq_file *)file->private_data;
2497 +static struct file_operations ext3_mb_seq_groups_fops = {
2498 + .owner = THIS_MODULE,
2499 + .open = ext3_mb_seq_groups_open,
2501 + .llseek = seq_lseek,
2502 + .release = seq_release,
2505 +static void ext3_mb_history_release(struct super_block *sb)
2507 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2509 + remove_proc_entry("mb_groups", sbi->s_dev_proc);
2510 + remove_proc_entry("mb_history", sbi->s_dev_proc);
2512 + if (sbi->s_mb_history)
2513 + kfree(sbi->s_mb_history);
2516 +static void ext3_mb_history_init(struct super_block *sb)
2518 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2521 + if (sbi->s_dev_proc != NULL) {
2522 + struct proc_dir_entry *p;
2523 + p = create_proc_entry("mb_history", S_IRUGO, sbi->s_dev_proc);
2525 + p->proc_fops = &ext3_mb_seq_history_fops;
2528 + p = create_proc_entry("mb_groups", S_IRUGO, sbi->s_dev_proc);
2530 + p->proc_fops = &ext3_mb_seq_groups_fops;
2535 + sbi->s_mb_history_max = 1000;
2536 + sbi->s_mb_history_cur = 0;
2537 + spin_lock_init(&sbi->s_mb_history_lock);
2538 + i = sbi->s_mb_history_max * sizeof(struct ext3_mb_history);
2539 + sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
2540 + if (likely(sbi->s_mb_history != NULL))
2541 + memset(sbi->s_mb_history, 0, i);
2542 + /* if we can't allocate history, then we simple won't use it */
2546 +ext3_mb_store_history(struct ext3_allocation_context *ac)
2548 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
2549 + struct ext3_mb_history h;
2551 + if (unlikely(sbi->s_mb_history == NULL))
2554 + if (!(ac->ac_op & sbi->s_mb_history_filter))
2558 + h.pid = current->pid;
2559 + h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2560 + h.orig = ac->ac_o_ex;
2561 + h.result = ac->ac_b_ex;
2562 + h.flags = ac->ac_flags;
2563 + h.found = ac->ac_found;
2564 + h.groups = ac->ac_groups_scanned;
2565 + h.cr = ac->ac_criteria;
2566 + h.tail = ac->ac_tail;
2567 + h.buddy = ac->ac_buddy;
2569 + h.cr = ac->ac_criteria;
2570 + if (ac->ac_op == EXT3_MB_HISTORY_ALLOC) {
2571 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2572 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2574 + h.goal = ac->ac_g_ex;
2575 + h.result = ac->ac_f_ex;
2578 + spin_lock(&sbi->s_mb_history_lock);
2579 + memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2580 + if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2581 + sbi->s_mb_history_cur = 0;
2582 + spin_unlock(&sbi->s_mb_history_lock);
2586 +#define ext3_mb_history_release(sb)
2587 +#define ext3_mb_history_init(sb)
2590 +int ext3_mb_init_backend(struct super_block *sb)
2592 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2593 + int i, j, len, metalen;
2594 + int num_meta_group_infos =
2595 + (sbi->s_groups_count + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2596 + EXT3_DESC_PER_BLOCK_BITS(sb);
2597 + struct ext3_group_info **meta_group_info;
2599 + /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2600 + * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2601 + * So a two level scheme suffices for now. */
2602 + sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
2603 + num_meta_group_infos, GFP_KERNEL);
2604 + if (sbi->s_group_info == NULL) {
2605 + printk(KERN_ERR "EXT3-fs: can't allocate buddy meta group\n");
2608 + sbi->s_buddy_cache = new_inode(sb);
2609 + if (sbi->s_buddy_cache == NULL) {
2610 + printk(KERN_ERR "EXT3-fs: can't get new inode\n");
2613 + EXT3_I(sbi->s_buddy_cache)->i_disksize = 0;
2615 + metalen = sizeof(*meta_group_info) << EXT3_DESC_PER_BLOCK_BITS(sb);
2616 + for (i = 0; i < num_meta_group_infos; i++) {
2617 + if ((i + 1) == num_meta_group_infos)
2618 + metalen = sizeof(*meta_group_info) *
2619 + (sbi->s_groups_count -
2620 + (i << EXT3_DESC_PER_BLOCK_BITS(sb)));
2621 + meta_group_info = kmalloc(metalen, GFP_KERNEL);
2622 + if (meta_group_info == NULL) {
2623 + printk(KERN_ERR "EXT3-fs: can't allocate mem for a "
2625 + goto err_freemeta;
2627 + sbi->s_group_info[i] = meta_group_info;
2631 + * calculate needed size. if change bb_counters size,
2632 + * don't forget about ext3_mb_generate_buddy()
2634 + len = sizeof(struct ext3_group_info);
2635 + len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
2636 + for (i = 0; i < sbi->s_groups_count; i++) {
2637 + struct ext3_group_desc * desc;
2640 + sbi->s_group_info[i >> EXT3_DESC_PER_BLOCK_BITS(sb)];
2641 + j = i & (EXT3_DESC_PER_BLOCK(sb) - 1);
2643 + meta_group_info[j] = kmalloc(len, GFP_KERNEL);
2644 + if (meta_group_info[j] == NULL) {
2645 + printk(KERN_ERR "EXT3-fs: can't allocate buddy mem\n");
2647 + goto err_freebuddy;
2649 + desc = ext3_get_group_desc(sb, i, NULL);
2650 + if (desc == NULL) {
2651 + printk(KERN_ERR"EXT3-fs: can't read descriptor %u\n",i);
2652 + goto err_freebuddy;
2654 + memset(meta_group_info[j], 0, len);
2655 + set_bit(EXT3_GROUP_INFO_NEED_INIT_BIT,
2656 + &meta_group_info[j]->bb_state);
2658 + /* initialize bb_free to be able to skip
2659 + * empty groups without initialization */
2660 + meta_group_info[j]->bb_free =
2661 + le16_to_cpu(desc->bg_free_blocks_count);
2663 + INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
2665 +#ifdef DOUBLE_CHECK
2667 + struct buffer_head *bh;
2668 + meta_group_info[j]->bb_bitmap =
2669 + kmalloc(sb->s_blocksize, GFP_KERNEL);
2670 + BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
2671 + bh = read_block_bitmap(sb, i);
2672 + BUG_ON(bh == NULL);
2673 + memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
2685 + kfree(EXT3_GROUP_INFO(sb, i));
2688 + i = num_meta_group_infos;
2691 + kfree(sbi->s_group_info[i]);
2692 + iput(sbi->s_buddy_cache);
2694 + kfree(sbi->s_group_info);
2698 +static void ext3_mb_prealloc_table_add(struct ext3_sb_info *sbi, int value)
2702 + if (value > (sbi->s_blocks_per_group - 1 - 1 - sbi->s_itb_per_group))
2705 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
2706 + if (sbi->s_mb_prealloc_table[i] == 0) {
2707 + sbi->s_mb_prealloc_table[i] = value;
2711 + /* they should add values in order */
2712 + if (value <= sbi->s_mb_prealloc_table[i])
2717 +int ext3_mb_init(struct super_block *sb, int needs_recovery)
2719 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2720 + unsigned i, offset, max;
2722 + if (!test_opt(sb, MBALLOC))
2725 + i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2727 + sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2728 + if (sbi->s_mb_offsets == NULL) {
2729 + clear_opt(sbi->s_mount_opt, MBALLOC);
2732 + sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2733 + if (sbi->s_mb_maxs == NULL) {
2734 + clear_opt(sbi->s_mount_opt, MBALLOC);
2735 + kfree(sbi->s_mb_maxs);
2739 + /* order 0 is regular bitmap */
2740 + sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2741 + sbi->s_mb_offsets[0] = 0;
2745 + max = sb->s_blocksize << 2;
2747 + sbi->s_mb_offsets[i] = offset;
2748 + sbi->s_mb_maxs[i] = max;
2749 + offset += 1 << (sb->s_blocksize_bits - i);
2752 + } while (i <= sb->s_blocksize_bits + 1);
2754 + /* init file for buddy data */
2755 + if ((i = ext3_mb_init_backend(sb))) {
2756 + clear_opt(sbi->s_mount_opt, MBALLOC);
2757 + kfree(sbi->s_mb_offsets);
2758 + kfree(sbi->s_mb_maxs);
2762 + spin_lock_init(&sbi->s_md_lock);
2763 + INIT_LIST_HEAD(&sbi->s_active_transaction);
2764 + INIT_LIST_HEAD(&sbi->s_closed_transaction);
2765 + INIT_LIST_HEAD(&sbi->s_committed_transaction);
2766 + spin_lock_init(&sbi->s_bal_lock);
2768 + sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2769 + sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2770 + sbi->s_mb_max_groups_to_scan = MB_DEFAULT_MAX_GROUPS_TO_SCAN;
2771 + sbi->s_mb_stats = MB_DEFAULT_STATS;
2772 + sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2773 + sbi->s_mb_history_filter = EXT3_MB_HISTORY_DEFAULT;
2775 + if (sbi->s_stripe == 0) {
2776 + sbi->s_mb_prealloc_table_size = 8;
2777 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
2778 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
2779 + if (sbi->s_mb_prealloc_table == NULL) {
2780 + clear_opt(sbi->s_mount_opt, MBALLOC);
2781 + kfree(sbi->s_mb_offsets);
2782 + kfree(sbi->s_mb_maxs);
2785 + memset(sbi->s_mb_prealloc_table, 0, i);
2787 + ext3_mb_prealloc_table_add(sbi, 4);
2788 + ext3_mb_prealloc_table_add(sbi, 8);
2789 + ext3_mb_prealloc_table_add(sbi, 16);
2790 + ext3_mb_prealloc_table_add(sbi, 32);
2791 + ext3_mb_prealloc_table_add(sbi, 64);
2792 + ext3_mb_prealloc_table_add(sbi, 128);
2793 + ext3_mb_prealloc_table_add(sbi, 256);
2794 + ext3_mb_prealloc_table_add(sbi, 512);
2796 + sbi->s_mb_small_req = 256;
2797 + sbi->s_mb_large_req = 1024;
2798 + sbi->s_mb_group_prealloc = 512;
2800 + sbi->s_mb_prealloc_table_size = 3;
2801 + i = sbi->s_mb_prealloc_table_size * sizeof(unsigned long);
2802 + sbi->s_mb_prealloc_table = kmalloc(i, GFP_NOFS);
2803 + if (sbi->s_mb_prealloc_table == NULL) {
2804 + clear_opt(sbi->s_mount_opt, MBALLOC);
2805 + kfree(sbi->s_mb_offsets);
2806 + kfree(sbi->s_mb_maxs);
2809 + memset(sbi->s_mb_prealloc_table, 0, i);
2811 + ext3_mb_prealloc_table_add(sbi, sbi->s_stripe);
2812 + ext3_mb_prealloc_table_add(sbi, sbi->s_stripe * 2);
2813 + ext3_mb_prealloc_table_add(sbi, sbi->s_stripe * 4);
2815 + sbi->s_mb_small_req = sbi->s_stripe;
2816 + sbi->s_mb_large_req = sbi->s_stripe * 8;
2817 + sbi->s_mb_group_prealloc = sbi->s_stripe * 4;
2820 + i = sizeof(struct ext3_locality_group) * num_possible_cpus();
2821 + sbi->s_locality_groups = kmalloc(i, GFP_NOFS);
2822 + if (sbi->s_locality_groups == NULL) {
2823 + clear_opt(sbi->s_mount_opt, MBALLOC);
2824 + kfree(sbi->s_mb_prealloc_table);
2825 + kfree(sbi->s_mb_offsets);
2826 + kfree(sbi->s_mb_maxs);
2829 + for (i = 0; i < num_possible_cpus(); i++) {
2830 + struct ext3_locality_group *lg;
2831 + lg = &sbi->s_locality_groups[i];
2832 + sema_init(&lg->lg_sem, 1);
2833 + INIT_LIST_HEAD(&lg->lg_prealloc_list);
2834 + spin_lock_init(&lg->lg_prealloc_lock);
2837 + ext3_mb_init_per_dev_proc(sb);
2838 + ext3_mb_history_init(sb);
2840 + printk("EXT3-fs: mballoc enabled\n");
2844 +void ext3_mb_cleanup_pa(struct ext3_group_info *grp)
2846 + struct ext3_prealloc_space *pa;
2847 + struct list_head *cur, *tmp;
2850 + list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2851 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
2852 + list_del_rcu(&pa->pa_group_list);
2857 + mb_debug("mballoc: %u PAs left\n", count);
2861 +int ext3_mb_release(struct super_block *sb)
2863 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2864 + int i, num_meta_group_infos;
2866 + if (!test_opt(sb, MBALLOC))
2869 + /* release freed, non-committed blocks */
2870 + spin_lock(&sbi->s_md_lock);
2871 + list_splice_init(&sbi->s_closed_transaction,
2872 + &sbi->s_committed_transaction);
2873 + list_splice_init(&sbi->s_active_transaction,
2874 + &sbi->s_committed_transaction);
2875 + spin_unlock(&sbi->s_md_lock);
2876 + ext3_mb_free_committed_blocks(sb);
2878 + if (sbi->s_group_info) {
2879 + for (i = 0; i < sbi->s_groups_count; i++) {
2880 +#ifdef DOUBLE_CHECK
2881 + if (EXT3_GROUP_INFO(sb, i)->bb_bitmap)
2882 + kfree(EXT3_GROUP_INFO(sb, i)->bb_bitmap);
2884 + ext3_mb_cleanup_pa(EXT3_GROUP_INFO(sb, i));
2885 + kfree(EXT3_GROUP_INFO(sb, i));
2887 + num_meta_group_infos = (sbi->s_groups_count +
2888 + EXT3_DESC_PER_BLOCK(sb) - 1) >>
2889 + EXT3_DESC_PER_BLOCK_BITS(sb);
2890 + for (i = 0; i < num_meta_group_infos; i++)
2891 + kfree(sbi->s_group_info[i]);
2892 + kfree(sbi->s_group_info);
2894 + if (sbi->s_mb_offsets)
2895 + kfree(sbi->s_mb_offsets);
2896 + if (sbi->s_mb_maxs)
2897 + kfree(sbi->s_mb_maxs);
2898 + if (sbi->s_buddy_cache)
2899 + iput(sbi->s_buddy_cache);
2900 + if (sbi->s_mb_stats) {
2901 + printk("EXT3-fs: mballoc: %u blocks %u reqs (%u success)\n",
2902 + atomic_read(&sbi->s_bal_allocated),
2903 + atomic_read(&sbi->s_bal_reqs),
2904 + atomic_read(&sbi->s_bal_success));
2905 + printk("EXT3-fs: mballoc: %u extents scanned, %u goal hits, "
2906 + "%u 2^N hits, %u breaks, %u lost\n",
2907 + atomic_read(&sbi->s_bal_ex_scanned),
2908 + atomic_read(&sbi->s_bal_goals),
2909 + atomic_read(&sbi->s_bal_2orders),
2910 + atomic_read(&sbi->s_bal_breaks),
2911 + atomic_read(&sbi->s_mb_lost_chunks));
2912 + printk("EXT3-fs: mballoc: %lu generated and it took %Lu\n",
2913 + sbi->s_mb_buddies_generated++,
2914 + sbi->s_mb_generation_time);
2915 + printk("EXT3-fs: mballoc: %u preallocated, %u discarded\n",
2916 + atomic_read(&sbi->s_mb_preallocated),
2917 + atomic_read(&sbi->s_mb_discarded));
2920 + if (sbi->s_locality_groups)
2921 + kfree(sbi->s_locality_groups);
2923 + ext3_mb_history_release(sb);
2924 + ext3_mb_destroy_per_dev_proc(sb);
2929 +void ext3_mb_free_committed_blocks(struct super_block *sb)
2931 + struct ext3_sb_info *sbi = EXT3_SB(sb);
2932 + int err, i, count = 0, count2 = 0;
2933 + struct ext3_free_metadata *md;
2934 + struct ext3_buddy e3b;
2936 + if (list_empty(&sbi->s_committed_transaction))
2939 + /* there is committed blocks to be freed yet */
2941 + /* get next array of blocks */
2943 + spin_lock(&sbi->s_md_lock);
2944 + if (!list_empty(&sbi->s_committed_transaction)) {
2945 + md = list_entry(sbi->s_committed_transaction.next,
2946 + struct ext3_free_metadata, list);
2947 + list_del(&md->list);
2949 + spin_unlock(&sbi->s_md_lock);
2954 + mb_debug("gonna free %u blocks in group %u (0x%p):",
2955 + md->num, md->group, md);
2957 + err = ext3_mb_load_buddy(sb, md->group, &e3b);
2958 + /* we expect to find existing buddy because it's pinned */
2961 + /* there are blocks to put in buddy to make them really free */
2964 + ext3_lock_group(sb, md->group);
2965 + for (i = 0; i < md->num; i++) {
2966 + mb_debug(" %u", md->blocks[i]);
2967 + err = mb_free_blocks(NULL, &e3b, md->blocks[i], 1);
2971 + ext3_unlock_group(sb, md->group);
2973 + /* balance refcounts from ext3_mb_free_metadata() */
2974 + page_cache_release(e3b.bd_buddy_page);
2975 + page_cache_release(e3b.bd_bitmap_page);
2978 + ext3_mb_release_desc(&e3b);
2982 + mb_debug("freed %u blocks in %u structures\n", count, count2);
2985 +#define EXT3_MB_STATS_NAME "stats"
2986 +#define EXT3_MB_MAX_TO_SCAN_NAME "max_to_scan"
2987 +#define EXT3_MB_MIN_TO_SCAN_NAME "min_to_scan"
2988 +#define EXT3_MB_ORDER2_REQ "order2_req"
2989 +#define EXT3_MB_SMALL_REQ "small_req"
2990 +#define EXT3_MB_LARGE_REQ "large_req"
2991 +#define EXT3_MB_PREALLOC_TABLE "prealloc_table"
2992 +#define EXT3_MB_GROUP_PREALLOC "group_prealloc"
2994 +static int ext3_mb_read_prealloc_table(char *page, char **start, off_t off,
2995 + int count, int *eof, void *data)
2997 + struct ext3_sb_info *sbi = data;
3005 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++)
3006 + len += sprintf(page + len, "%ld ",
3007 + sbi->s_mb_prealloc_table[i]);
3008 + len += sprintf(page + len, "\n");
3014 +static int ext3_mb_write_prealloc_table(struct file *file,
3015 + const char __user *buf,
3016 + unsigned long cnt, void *data)
3018 + struct ext3_sb_info *sbi = data;
3019 + unsigned long value;
3020 + unsigned long prev = 0;
3024 + unsigned long *new_table;
3028 + if (cnt >= sizeof(str))
3030 + if (copy_from_user(str, buf, cnt))
3036 + while (cur < end) {
3037 + while ((cur < end) && (*cur == ' ')) cur++;
3038 + value = simple_strtol(cur, &cur, 0);
3041 + if (value <= prev)
3047 + new_table = kmalloc(num * sizeof(*new_table), GFP_KERNEL);
3048 + if (new_table == NULL)
3050 + kfree(sbi->s_mb_prealloc_table);
3051 + memset(new_table, 0, num * sizeof(*new_table));
3052 + sbi->s_mb_prealloc_table = new_table;
3053 + sbi->s_mb_prealloc_table_size = num;
3056 + while (cur < end && i < num) {
3057 + while ((cur < end) && (*cur == ' ')) cur++;
3058 + value = simple_strtol(cur, &cur, 0);
3059 + ext3_mb_prealloc_table_add(sbi, value);
3066 +#define MB_PROC_VALUE_READ(name) \
3067 +static int ext3_mb_read_##name(char *page, char **start, \
3068 + off_t off, int count, int *eof, void *data) \
3070 + struct ext3_sb_info *sbi = data; \
3075 + len = sprintf(page, "%ld\n", sbi->s_mb_##name); \
3080 +#define MB_PROC_VALUE_WRITE(name) \
3081 +static int ext3_mb_write_##name(struct file *file, \
3082 + const char __user *buf, unsigned long cnt, void *data) \
3084 + struct ext3_sb_info *sbi = data; \
3087 + if (cnt >= sizeof(str)) \
3089 + if (copy_from_user(str, buf, cnt)) \
3091 + value = simple_strtol(str, NULL, 0); \
3094 + sbi->s_mb_##name = value; \
3098 +MB_PROC_VALUE_READ(stats);
3099 +MB_PROC_VALUE_WRITE(stats);
3100 +MB_PROC_VALUE_READ(max_to_scan);
3101 +MB_PROC_VALUE_WRITE(max_to_scan);
3102 +MB_PROC_VALUE_READ(min_to_scan);
3103 +MB_PROC_VALUE_WRITE(min_to_scan);
3104 +MB_PROC_VALUE_READ(order2_reqs);
3105 +MB_PROC_VALUE_WRITE(order2_reqs);
3106 +MB_PROC_VALUE_READ(small_req);
3107 +MB_PROC_VALUE_WRITE(small_req);
3108 +MB_PROC_VALUE_READ(large_req);
3109 +MB_PROC_VALUE_WRITE(large_req);
3110 +MB_PROC_VALUE_READ(group_prealloc);
3111 +MB_PROC_VALUE_WRITE(group_prealloc);
3113 +#define MB_PROC_HANDLER(name, var) \
3115 + proc = create_proc_entry(name, mode, sbi->s_dev_proc); \
3116 + if (proc == NULL) { \
3117 + printk(KERN_ERR "EXT3-fs: can't to create %s\n", name); \
3120 + proc->data = sbi; \
3121 + proc->read_proc = ext3_mb_read_##var ; \
3122 + proc->write_proc = ext3_mb_write_##var; \
3125 +int ext3_mb_init_per_dev_proc(struct super_block *sb)
3127 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3128 + mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
3129 + struct proc_dir_entry *proc;
3131 + MB_PROC_HANDLER(EXT3_MB_STATS_NAME, stats);
3132 + MB_PROC_HANDLER(EXT3_MB_MAX_TO_SCAN_NAME, max_to_scan);
3133 + MB_PROC_HANDLER(EXT3_MB_MIN_TO_SCAN_NAME, min_to_scan);
3134 + MB_PROC_HANDLER(EXT3_MB_ORDER2_REQ, order2_reqs);
3135 + MB_PROC_HANDLER(EXT3_MB_SMALL_REQ, small_req);
3136 + MB_PROC_HANDLER(EXT3_MB_LARGE_REQ, large_req);
3137 + MB_PROC_HANDLER(EXT3_MB_PREALLOC_TABLE, prealloc_table);
3138 + MB_PROC_HANDLER(EXT3_MB_GROUP_PREALLOC, group_prealloc);
3143 + remove_proc_entry(EXT3_MB_GROUP_PREALLOC, sbi->s_dev_proc);
3144 + remove_proc_entry(EXT3_MB_PREALLOC_TABLE, sbi->s_dev_proc);
3145 + remove_proc_entry(EXT3_MB_LARGE_REQ, sbi->s_dev_proc);
3146 + remove_proc_entry(EXT3_MB_SMALL_REQ, sbi->s_dev_proc);
3147 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_dev_proc);
3148 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_dev_proc);
3149 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_dev_proc);
3150 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_dev_proc);
3155 +int ext3_mb_destroy_per_dev_proc(struct super_block *sb)
3157 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3159 + if (sbi->s_dev_proc == NULL)
3162 + remove_proc_entry(EXT3_MB_GROUP_PREALLOC, sbi->s_dev_proc);
3163 + remove_proc_entry(EXT3_MB_PREALLOC_TABLE, sbi->s_dev_proc);
3164 + remove_proc_entry(EXT3_MB_SMALL_REQ, sbi->s_dev_proc);
3165 + remove_proc_entry(EXT3_MB_LARGE_REQ, sbi->s_dev_proc);
3166 + remove_proc_entry(EXT3_MB_ORDER2_REQ, sbi->s_dev_proc);
3167 + remove_proc_entry(EXT3_MB_MIN_TO_SCAN_NAME, sbi->s_dev_proc);
3168 + remove_proc_entry(EXT3_MB_MAX_TO_SCAN_NAME, sbi->s_dev_proc);
3169 + remove_proc_entry(EXT3_MB_STATS_NAME, sbi->s_dev_proc);
3174 +int __init init_ext3_mb_proc(void)
3176 + ext3_pspace_cachep =
3177 + kmem_cache_create("ext3_prealloc_space",
3178 + sizeof(struct ext3_prealloc_space),
3179 + 0, SLAB_RECLAIM_ACCOUNT, NULL, NULL);
3180 + if (ext3_pspace_cachep == NULL)
3186 +void exit_ext3_mb_proc(void)
3188 + /* XXX: synchronize_rcu(); */
3189 + kmem_cache_destroy(ext3_pspace_cachep);
3194 + * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
3195 + * Returns 0 if success or error code
3197 +int ext3_mb_mark_diskspace_used(struct ext3_allocation_context *ac, handle_t *handle)
3199 + struct buffer_head *bitmap_bh = NULL;
3200 + struct ext3_super_block *es;
3201 + struct ext3_group_desc *gdp;
3202 + struct buffer_head *gdp_bh;
3203 + struct ext3_sb_info *sbi;
3204 + struct super_block *sb;
3208 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3209 + BUG_ON(ac->ac_b_ex.fe_len <= 0);
3212 + sbi = EXT3_SB(sb);
3215 + ext3_debug("using block group %d(%d)\n", ac->ac_b_group.group,
3216 + gdp->bg_free_blocks_count);
3219 + bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3223 + err = ext3_journal_get_write_access(handle, bitmap_bh);
3228 + gdp = ext3_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3232 + err = ext3_journal_get_write_access(handle, gdp_bh);
3236 + block = ac->ac_b_ex.fe_group * EXT3_BLOCKS_PER_GROUP(sb)
3237 + + ac->ac_b_ex.fe_start
3238 + + le32_to_cpu(es->s_first_data_block);
3240 + if (block == le32_to_cpu(gdp->bg_block_bitmap) ||
3241 + block == le32_to_cpu(gdp->bg_inode_bitmap) ||
3242 + in_range(block, le32_to_cpu(gdp->bg_inode_table),
3243 + EXT3_SB(sb)->s_itb_per_group))
3244 + ext3_error(sb, __FUNCTION__,
3245 + "Allocating block in system zone - block = %lu",
3246 + (unsigned long) block);
3247 +#ifdef AGGRESSIVE_CHECK
3250 + for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3251 + BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3252 + bitmap_bh->b_data));
3256 + mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
3257 + ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
3259 + spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3260 + gdp->bg_free_blocks_count =
3261 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
3262 + - ac->ac_b_ex.fe_len);
3263 + spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3264 + percpu_counter_mod(&sbi->s_freeblocks_counter, - ac->ac_b_ex.fe_len);
3266 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
3269 + err = ext3_journal_dirty_metadata(handle, gdp_bh);
3273 + brelse(bitmap_bh);
3278 + * here we normalize request for locality group
3279 + * XXX: should we try to preallocate more than the group has now?
3281 +void ext3_mb_normalize_group_request(struct ext3_allocation_context *ac)
3283 + struct super_block *sb = ac->ac_sb;
3284 + struct ext3_locality_group *lg = ac->ac_lg;
3286 + BUG_ON(lg == NULL);
3287 + ac->ac_g_ex.fe_len = EXT3_SB(sb)->s_mb_group_prealloc;
3289 + mb_debug("#%u: goal %u blocks for locality group\n",
3290 + current->pid, ac->ac_g_ex.fe_len);
3294 + * Normalization means making request better in terms of
3295 + * size and alignment
3297 +void ext3_mb_normalize_request(struct ext3_allocation_context *ac,
3298 + struct ext3_allocation_request *ar)
3300 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3301 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
3302 + loff_t start, end, size, orig_size, orig_start;
3303 + struct list_head *cur;
3304 + int bsbits, i, wind;
3306 + /* do normalize only data requests, metadata requests
3307 + do not need preallocation */
3308 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3311 + /* sometime caller may want exact blocks */
3312 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
3315 + /* caller may indicate that preallocation isn't
3316 + * required (it's a tail, for example) */
3317 + if (ac->ac_flags & EXT3_MB_HINT_NOPREALLOC)
3320 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3321 + return ext3_mb_normalize_group_request(ac);
3323 + bsbits = ac->ac_sb->s_blocksize_bits;
3325 + /* first, let's learn actual file size
3326 + * given current request is allocated */
3327 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3328 + size = size << bsbits;
3329 + if (size < i_size_read(ac->ac_inode))
3330 + size = i_size_read(ac->ac_inode);
3331 + size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
3336 + /* let's choose preallocation window depending on file size */
3337 + for (i = 0; i < sbi->s_mb_prealloc_table_size; i++) {
3338 + if (size <= sbi->s_mb_prealloc_table[i]) {
3339 + wind = sbi->s_mb_prealloc_table[i];
3346 + __u64 tstart, tend;
3347 + /* file is quite large, we now preallocate with
3348 + * the biggest configured window with regart to
3349 + * logical offset */
3350 + wind = sbi->s_mb_prealloc_table[i - 1];
3351 + tstart = ac->ac_o_ex.fe_logical;
3352 + do_div(tstart, wind);
3353 + start = tstart * wind;
3354 + tend = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len - 1;
3355 + do_div(tend, wind);
3356 + tend = tend * wind + wind;
3357 + size = tend - start;
3360 + orig_start = start;
3362 + /* don't cover already allocated blocks in selected range */
3363 + if (ar->pleft && start <= ar->lleft) {
3364 + size -= ar->lleft + 1 - start;
3365 + start = ar->lleft + 1;
3367 + if (ar->pright && start + size - 1 >= ar->lright)
3368 + size -= start + size - ar->lright;
3370 + end = start + size;
3372 + /* check we don't cross already preallocated blocks */
3374 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3375 + struct ext3_prealloc_space *pa;
3376 + unsigned long pa_end;
3378 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3380 + if (pa->pa_deleted)
3382 + spin_lock(&pa->pa_lock);
3383 + if (pa->pa_deleted) {
3384 + spin_unlock(&pa->pa_lock);
3388 + pa_end = pa->pa_lstart + pa->pa_len;
3390 + /* PA must not overlap original request */
3391 + BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3392 + ac->ac_o_ex.fe_logical < pa->pa_lstart));
3394 + /* skip PA normalized request doesn't overlap with */
3395 + if (pa->pa_lstart >= end) {
3396 + spin_unlock(&pa->pa_lock);
3399 + if (pa_end <= start) {
3400 + spin_unlock(&pa->pa_lock);
3403 + BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3405 + if (pa_end <= ac->ac_o_ex.fe_logical) {
3406 + BUG_ON(pa_end < start);
3410 + if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3411 + BUG_ON(pa->pa_lstart > end);
3412 + end = pa->pa_lstart;
3414 + spin_unlock(&pa->pa_lock);
3416 + rcu_read_unlock();
3417 + size = end - start;
3419 + /* XXX: extra loop to check we really don't overlap preallocations */
3421 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3422 + struct ext3_prealloc_space *pa;
3423 + unsigned long pa_end;
3424 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3425 + spin_lock(&pa->pa_lock);
3426 + if (pa->pa_deleted == 0) {
3427 + pa_end = pa->pa_lstart + pa->pa_len;
3428 + BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3430 + spin_unlock(&pa->pa_lock);
3432 + rcu_read_unlock();
3434 + if (start + size <= ac->ac_o_ex.fe_logical &&
3435 + start > ac->ac_o_ex.fe_logical) {
3436 + printk("start %lu, size %lu, fe_logical %lu\n",
3437 + (unsigned long) start, (unsigned long) size,
3438 + (unsigned long) ac->ac_o_ex.fe_logical);
3440 + BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3441 + start > ac->ac_o_ex.fe_logical);
3443 + /* now prepare goal request */
3445 + /* XXX: is it better to align blocks WRT to logical
3446 + * placement or satisfy big request as is */
3447 + ac->ac_g_ex.fe_logical = start;
3448 + ac->ac_g_ex.fe_len = size;
3450 + /* define goal start in order to merge */
3451 + if (ar->pright && (ar->lright == (start + size))) {
3452 + /* merge to the right */
3453 + ext3_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3454 + &ac->ac_f_ex.fe_group,
3455 + &ac->ac_f_ex.fe_start);
3456 + ac->ac_flags |= EXT3_MB_HINT_TRY_GOAL;
3458 + if (ar->pleft && (ar->lleft + 1 == start)) {
3459 + /* merge to the left */
3460 + ext3_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3461 + &ac->ac_f_ex.fe_group,
3462 + &ac->ac_f_ex.fe_start);
3463 + ac->ac_flags |= EXT3_MB_HINT_TRY_GOAL;
3466 + mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3467 + (unsigned) orig_size, (unsigned) start);
3470 +void ext3_mb_collect_stats(struct ext3_allocation_context *ac)
3472 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
3474 + if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3475 + atomic_inc(&sbi->s_bal_reqs);
3476 + atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3477 + if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3478 + atomic_inc(&sbi->s_bal_success);
3479 + atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3480 + if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3481 + ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3482 + atomic_inc(&sbi->s_bal_goals);
3483 + if (ac->ac_found > sbi->s_mb_max_to_scan)
3484 + atomic_inc(&sbi->s_bal_breaks);
3487 + ext3_mb_store_history(ac);
3491 + * use blocks preallocated to inode
3493 +void ext3_mb_use_inode_pa(struct ext3_allocation_context *ac,
3494 + struct ext3_prealloc_space *pa)
3496 + unsigned long start, len;
3498 + /* found preallocated blocks, use them */
3499 + start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3500 + len = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3501 + len = len - start;
3502 + ext3_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3503 + &ac->ac_b_ex.fe_start);
3504 + ac->ac_b_ex.fe_len = len;
3505 + ac->ac_status = AC_STATUS_FOUND;
3508 + BUG_ON(start < pa->pa_pstart);
3509 + BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3510 + BUG_ON(pa->pa_free < len);
3511 + pa->pa_free -= len;
3513 + mb_debug("use %lu/%lu from inode pa %p\n", start, len, pa);
3517 + * use blocks preallocated to locality group
3519 +void ext3_mb_use_group_pa(struct ext3_allocation_context *ac,
3520 + struct ext3_prealloc_space *pa)
3522 + unsigned len = ac->ac_o_ex.fe_len;
3524 + ext3_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3525 + &ac->ac_b_ex.fe_group,
3526 + &ac->ac_b_ex.fe_start);
3527 + ac->ac_b_ex.fe_len = len;
3528 + ac->ac_status = AC_STATUS_FOUND;
3531 + /* we don't correct pa_pstart or pa_plen here to avoid
3532 + * possible race when tte group is being loaded concurrently
3533 + * instead we correct pa later, after blocks are marked
3534 + * in on-disk bitmap -- see ext3_mb_release_context() */
3535 + mb_debug("use %lu/%lu from group pa %p\n", pa->pa_lstart-len, len, pa);
3539 + * search goal blocks in preallocated space
3541 +int ext3_mb_use_preallocated(struct ext3_allocation_context *ac)
3543 + struct ext3_inode_info *ei = EXT3_I(ac->ac_inode);
3544 + struct ext3_locality_group *lg;
3545 + struct ext3_prealloc_space *pa;
3546 + struct list_head *cur;
3548 + /* only data can be preallocated */
3549 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
3552 + /* first, try per-file preallocation */
3554 + list_for_each_rcu(cur, &ei->i_prealloc_list) {
3555 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3557 + /* all fields in this condition don't change,
3558 + * so we can skip locking for them */
3559 + if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3560 + ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3563 + /* found preallocated blocks, use them */
3564 + spin_lock(&pa->pa_lock);
3565 + if (pa->pa_deleted == 0 && pa->pa_free) {
3566 + atomic_inc(&pa->pa_count);
3567 + ext3_mb_use_inode_pa(ac, pa);
3568 + spin_unlock(&pa->pa_lock);
3569 + ac->ac_criteria = 10;
3570 + rcu_read_unlock();
3573 + spin_unlock(&pa->pa_lock);
3575 + rcu_read_unlock();
3577 + /* can we use group allocation? */
3578 + if (!(ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC))
3581 + /* inode may have no locality group for some reason */
3587 + list_for_each_rcu(cur, &lg->lg_prealloc_list) {
3588 + pa = list_entry(cur, struct ext3_prealloc_space, pa_inode_list);
3589 + spin_lock(&pa->pa_lock);
3590 + if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3591 + atomic_inc(&pa->pa_count);
3592 + ext3_mb_use_group_pa(ac, pa);
3593 + spin_unlock(&pa->pa_lock);
3594 + ac->ac_criteria = 20;
3595 + rcu_read_unlock();
3598 + spin_unlock(&pa->pa_lock);
3600 + rcu_read_unlock();
3606 + * check free blocks in bitmap match free block in group descriptor
3607 + * do this before taking preallocated blocks into account to be able
3608 + * to detect on-disk corruptions
3610 +int ext3_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
3611 + struct ext3_group_desc *gdp, int group)
3613 + unsigned short max = EXT3_BLOCKS_PER_GROUP(sb);
3614 + unsigned short i, first, free = 0;
3616 + i = mb_find_next_zero_bit(bitmap, max, 0);
3620 + i = ext2_find_next_le_bit(bitmap, max, i);
3623 + free += i - first;
3625 + i = mb_find_next_zero_bit(bitmap, max, i);
3628 + if (free != le16_to_cpu(gdp->bg_free_blocks_count)) {
3629 + ext3_error(sb, __FUNCTION__, "on-disk bitmap for group %d"
3630 + "corrupted: %u blocks free in bitmap, %u - in gd\n",
3631 + group, free, le16_to_cpu(gdp->bg_free_blocks_count));
3638 + * the function goes through all preallocation in this group and marks them
3639 + * used in in-core bitmap. buddy must be generated from this bitmap
3641 +int ext3_mb_generate_from_pa(struct super_block *sb, void *bitmap, int group)
3643 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
3644 + struct ext3_prealloc_space *pa;
3645 + struct ext3_group_desc *gdp;
3646 + struct list_head *cur;
3647 + unsigned long groupnr;
3648 + unsigned long start;
3649 + int preallocated = 0, count = 0, len, skip = 0, err;
3651 + gdp = ext3_get_group_desc (sb, group, NULL);
3655 + /* before applying preallocations, check bitmap consistency */
3656 + err = ext3_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
3660 + /* all form of preallocation discards first load group,
3661 + * so the only competing code is preallocation use.
3662 + * we don't need any locking here
3663 + * notice we do NOT ignore preallocations with pa_deleted
3664 + * otherwise we could leave used blocks available for
3665 + * allocation in buddy when concurrent ext3_mb_put_pa()
3666 + * is dropping preallocation
3668 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
3669 + pa = list_entry(cur, struct ext3_prealloc_space, pa_group_list);
3670 + spin_lock(&pa->pa_lock);
3671 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &start);
3673 + spin_unlock(&pa->pa_lock);
3674 + if (unlikely(len == 0)) {
3678 + BUG_ON(groupnr != group && len != 0);
3679 + mb_set_bits(sb_bgl_lock(EXT3_SB(sb), group), bitmap, start,len);
3680 + preallocated += len;
3683 + if (count + skip != grp->bb_prealloc_nr) {
3684 + ext3_error(sb, __FUNCTION__, "lost preallocations: "
3685 + "count %d, bb_prealloc_nr %lu, skip %d\n",
3686 + count, grp->bb_prealloc_nr, skip);
3689 + mb_debug("prellocated %u for group %u\n", preallocated, group);
3693 +#if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,5)
3694 +static void ext3_mb_pa_callback(struct rcu_head *head)
3696 + struct ext3_prealloc_space *pa;
3697 + pa = container_of(head, struct ext3_prealloc_space, u.pa_rcu);
3698 + kmem_cache_free(ext3_pspace_cachep, pa);
3700 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback)
3702 +static void ext3_mb_pa_callback(void *pa)
3704 + kmem_cache_free(ext3_pspace_cachep, pa);
3706 +#define mb_call_rcu(__pa) call_rcu(&(__pa)->u.pa_rcu, ext3_mb_pa_callback, pa)
3710 + * drops a reference to preallocated space descriptor
3711 + * if this was the last reference and the space is consumed
3713 +void ext3_mb_put_pa(struct ext3_allocation_context *ac,
3714 + struct super_block *sb, struct ext3_prealloc_space *pa)
3716 + unsigned long grp;
3718 + if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3721 + /* in this short window concurrent discard can set pa_deleted */
3722 + spin_lock(&pa->pa_lock);
3723 + if (pa->pa_deleted == 1) {
3724 + spin_unlock(&pa->pa_lock);
3728 + pa->pa_deleted = 1;
3729 + spin_unlock(&pa->pa_lock);
3731 + /* -1 is to protect from crossing allocation group */
3732 + ext3_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3737 + * P1 (buddy init) P2 (regular allocation)
3738 + * find block B in PA
3739 + * copy on-disk bitmap to buddy
3740 + * mark B in on-disk bitmap
3741 + * drop PA from group
3742 + * mark all PAs in buddy
3744 + * thus, P1 initializes buddy with B available. to prevent this
3745 + * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3746 + * against that pair
3748 + ext3_lock_group(sb, grp);
3749 + list_del_rcu(&pa->pa_group_list);
3750 + EXT3_GROUP_INFO(sb, grp)->bb_prealloc_nr--;
3751 + ext3_unlock_group(sb, grp);
3753 + spin_lock(pa->pa_obj_lock);
3754 + list_del_rcu(&pa->pa_inode_list);
3755 + spin_unlock(pa->pa_obj_lock);
3761 + * creates new preallocated space for given inode
3763 +int ext3_mb_new_inode_pa(struct ext3_allocation_context *ac)
3765 + struct super_block *sb = ac->ac_sb;
3766 + struct ext3_prealloc_space *pa;
3767 + struct ext3_group_info *grp;
3768 + struct ext3_inode_info *ei;
3770 + /* preallocate only when found space is larger then requested */
3771 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3772 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3773 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3775 + pa = kmem_cache_alloc(ext3_pspace_cachep, GFP_NOFS);
3779 + if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3780 + int winl, wins, win, offs;
3782 + /* we can't allocate as much as normalizer wants.
3783 + * so, found space must get proper lstart
3784 + * to cover original request */
3785 + BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3786 + BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3788 + /* we're limited by original request in that
3789 + * logical block must be covered any way
3790 + * winl is window we can move our chunk within */
3791 + winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3793 + /* also, we should cover whole original request */
3794 + wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3796 + /* the smallest one defines real window */
3797 + win = min(winl, wins);
3799 + offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3800 + if (offs && offs < win)
3803 + ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3804 + BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3805 + BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3808 + /* preallocation can change ac_b_ex, thus we store actually
3809 + * allocated blocks for history */
3810 + ac->ac_f_ex = ac->ac_b_ex;
3812 + pa->pa_lstart = ac->ac_b_ex.fe_logical;
3813 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3814 + pa->pa_len = ac->ac_b_ex.fe_len;
3815 + pa->pa_free = pa->pa_len;
3816 + atomic_set(&pa->pa_count, 1);
3817 + spin_lock_init(&pa->pa_lock);
3818 + pa->pa_deleted = 0;
3819 + pa->pa_linear = 0;
3821 + mb_debug("new inode pa %p: %lu/%lu for %lu\n", pa,
3822 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3824 + ext3_mb_use_inode_pa(ac, pa);
3825 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3827 + ei = EXT3_I(ac->ac_inode);
3828 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3830 + pa->pa_obj_lock = &ei->i_prealloc_lock;
3831 + pa->pa_inode = ac->ac_inode;
3833 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3834 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3835 + grp->bb_prealloc_nr++;
3836 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3838 + spin_lock(pa->pa_obj_lock);
3839 + list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3840 + spin_unlock(pa->pa_obj_lock);
3846 + * creates new preallocated space for locality group inodes belongs to
3848 +int ext3_mb_new_group_pa(struct ext3_allocation_context *ac)
3850 + struct super_block *sb = ac->ac_sb;
3851 + struct ext3_locality_group *lg;
3852 + struct ext3_prealloc_space *pa;
3853 + struct ext3_group_info *grp;
3855 + /* preallocate only when found space is larger then requested */
3856 + BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3857 + BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3858 + BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3860 + BUG_ON(ext3_pspace_cachep == NULL);
3861 + pa = kmem_cache_alloc(ext3_pspace_cachep, GFP_NOFS);
3865 + /* preallocation can change ac_b_ex, thus we store actually
3866 + * allocated blocks for history */
3867 + ac->ac_f_ex = ac->ac_b_ex;
3869 + pa->pa_pstart = ext3_grp_offs_to_block(sb, &ac->ac_b_ex);
3870 + pa->pa_lstart = pa->pa_pstart;
3871 + pa->pa_len = ac->ac_b_ex.fe_len;
3872 + pa->pa_free = pa->pa_len;
3873 + atomic_set(&pa->pa_count, 1);
3874 + spin_lock_init(&pa->pa_lock);
3875 + pa->pa_deleted = 0;
3876 + pa->pa_linear = 1;
3878 + mb_debug("new group pa %p: %lu/%lu for %lu\n", pa,
3879 + pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3881 + ext3_mb_use_group_pa(ac, pa);
3882 + atomic_add(pa->pa_free, &EXT3_SB(sb)->s_mb_preallocated);
3884 + grp = EXT3_GROUP_INFO(sb, ac->ac_b_ex.fe_group);
3886 + BUG_ON(lg == NULL);
3888 + pa->pa_obj_lock = &lg->lg_prealloc_lock;
3889 + pa->pa_inode = NULL;
3891 + ext3_lock_group(sb, ac->ac_b_ex.fe_group);
3892 + list_add_rcu(&pa->pa_group_list, &grp->bb_prealloc_list);
3893 + grp->bb_prealloc_nr++;
3894 + ext3_unlock_group(sb, ac->ac_b_ex.fe_group);
3896 + spin_lock(pa->pa_obj_lock);
3897 + list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3898 + spin_unlock(pa->pa_obj_lock);
3903 +int ext3_mb_new_preallocation(struct ext3_allocation_context *ac)
3907 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
3908 + err = ext3_mb_new_group_pa(ac);
3910 + err = ext3_mb_new_inode_pa(ac);
3915 + * finds all unused blocks in on-disk bitmap, frees them in
3916 + * in-core bitmap and buddy.
3917 + * @pa must be unlinked from inode and group lists, so that
3918 + * nobody else can find/use it.
3919 + * the caller MUST hold group/inode locks.
3920 + * TODO: optimize the case when there are no in-core structures yet
3922 +int ext3_mb_release_inode_pa(struct ext3_buddy *e3b,
3923 + struct buffer_head *bitmap_bh,
3924 + struct ext3_prealloc_space *pa)
3926 + struct ext3_allocation_context ac;
3927 + struct super_block *sb = e3b->bd_sb;
3928 + struct ext3_sb_info *sbi = EXT3_SB(sb);
3929 + unsigned long bit, end, next, group;
3931 + int err = 0, free = 0;
3933 + BUG_ON(pa->pa_deleted == 0);
3934 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3935 + BUG_ON(group != e3b->bd_group && pa->pa_len != 0);
3936 + end = bit + pa->pa_len;
3939 + ac.ac_inode = pa->pa_inode;
3940 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3941 + ac.ac_o_ex.fe_len = 1;
3943 + while (bit < end) {
3944 + bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3947 + next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3950 + start = group * EXT3_BLOCKS_PER_GROUP(sb) + bit +
3951 + le32_to_cpu(sbi->s_es->s_first_data_block);
3952 + mb_debug(" free preallocated %u/%u in group %u\n",
3953 + (unsigned) start, (unsigned) next - bit,
3954 + (unsigned) group);
3955 + free += next - bit;
3957 + ac.ac_b_ex.fe_group = group;
3958 + ac.ac_b_ex.fe_start = bit;
3959 + ac.ac_b_ex.fe_len = next - bit;
3960 + ac.ac_b_ex.fe_logical = 0;
3961 + ext3_mb_store_history(&ac);
3963 + mb_free_blocks(pa->pa_inode, e3b, bit, next - bit);
3966 + if (free != pa->pa_free) {
3967 + printk("pa %p: logic %lu, phys. %lu, len %lu\n",
3968 + pa, (unsigned long) pa->pa_lstart,
3969 + (unsigned long) pa->pa_pstart,
3970 + (unsigned long) pa->pa_len);
3971 + printk("free %u, pa_free %u\n", free, pa->pa_free);
3973 + BUG_ON(free != pa->pa_free);
3974 + atomic_add(free, &sbi->s_mb_discarded);
3979 +int ext3_mb_release_group_pa(struct ext3_buddy *e3b,
3980 + struct ext3_prealloc_space *pa)
3982 + struct ext3_allocation_context ac;
3983 + struct super_block *sb = e3b->bd_sb;
3984 + unsigned long bit, group;
3986 + ac.ac_op = EXT3_MB_HISTORY_DISCARD;
3988 + BUG_ON(pa->pa_deleted == 0);
3989 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3990 + BUG_ON(group != e3b->bd_group && pa->pa_len != 0);
3991 + mb_free_blocks(pa->pa_inode, e3b, bit, pa->pa_len);
3992 + atomic_add(pa->pa_len, &EXT3_SB(sb)->s_mb_discarded);
3995 + ac.ac_inode = NULL;
3996 + ac.ac_b_ex.fe_group = group;
3997 + ac.ac_b_ex.fe_start = bit;
3998 + ac.ac_b_ex.fe_len = pa->pa_len;
3999 + ac.ac_b_ex.fe_logical = 0;
4000 + ext3_mb_store_history(&ac);
4006 + * releases all preallocations in given group
4008 + * first, we need to decide discard policy:
4009 + * - when do we discard
4011 + * - how many do we discard
4012 + * 1) how many requested
4014 +int ext3_mb_discard_group_preallocations(struct super_block *sb,
4015 + int group, int needed)
4017 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, group);
4018 + struct buffer_head *bitmap_bh = NULL;
4019 + struct ext3_prealloc_space *pa, *tmp;
4020 + struct list_head list;
4021 + struct ext3_buddy e3b;
4022 + int err, busy, free = 0;
4024 + mb_debug("discard preallocation for group %lu\n", group);
4026 + if (list_empty(&grp->bb_prealloc_list))
4029 + bitmap_bh = read_block_bitmap(sb, group);
4030 + if (bitmap_bh == NULL) {
4031 + /* error handling here */
4032 + ext3_mb_release_desc(&e3b);
4033 + BUG_ON(bitmap_bh == NULL);
4036 + err = ext3_mb_load_buddy(sb, group, &e3b);
4038 + brelse(bitmap_bh);
4043 + needed = EXT3_BLOCKS_PER_GROUP(sb) + 1;
4045 + grp = EXT3_GROUP_INFO(sb, group);
4046 + INIT_LIST_HEAD(&list);
4050 + ext3_lock_group(sb, group);
4051 + list_for_each_entry_safe (pa, tmp, &grp->bb_prealloc_list, pa_group_list) {
4052 + spin_lock(&pa->pa_lock);
4053 + if (atomic_read(&pa->pa_count)) {
4054 + spin_unlock(&pa->pa_lock);
4058 + if (pa->pa_deleted) {
4059 + spin_unlock(&pa->pa_lock);
4063 + /* seems this one can be freed ... */
4064 + pa->pa_deleted = 1;
4066 + /* we can trust pa_free ... */
4067 + free += pa->pa_free;
4069 + spin_unlock(&pa->pa_lock);
4071 + BUG_ON(grp->bb_prealloc_nr == 0);
4072 + grp->bb_prealloc_nr--;
4073 + list_del_rcu(&pa->pa_group_list);
4074 + list_add(&pa->u.pa_tmp_list, &list);
4077 + /* if we still need more blocks and some PAs were used, try again */
4078 + if (free < needed && busy) {
4079 + ext3_unlock_group(sb, group);
4083 + /* found anything to free? */
4084 + if (list_empty(&list)) {
4085 + BUG_ON(free != 0);
4089 + /* now free all selected PAs */
4090 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4092 + /* remove from object (inode or locality group) */
4093 + spin_lock(pa->pa_obj_lock);
4094 + list_del_rcu(&pa->pa_inode_list);
4095 + spin_unlock(pa->pa_obj_lock);
4097 + if (pa->pa_linear)
4098 + ext3_mb_release_group_pa(&e3b, pa);
4100 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
4102 + list_del(&pa->u.pa_tmp_list);
4107 + ext3_unlock_group(sb, group);
4108 + ext3_mb_release_desc(&e3b);
4109 + brelse(bitmap_bh);
4114 + * releases all non-used preallocated blocks for given inode
4116 +void ext3_mb_discard_inode_preallocations(struct inode *inode)
4118 + struct ext3_inode_info *ei = EXT3_I(inode);
4119 + struct super_block *sb = inode->i_sb;
4120 + struct buffer_head *bitmap_bh = NULL;
4121 + struct ext3_prealloc_space *pa, *tmp;
4122 + unsigned long group = 0;
4123 + struct list_head list;
4124 + struct ext3_buddy e3b;
4127 + if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
4128 + /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4132 + mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
4134 + INIT_LIST_HEAD(&list);
4137 + /* first, collect all pa's in the inode */
4138 + spin_lock(&ei->i_prealloc_lock);
4139 + while (!list_empty(&ei->i_prealloc_list)) {
4140 + pa = list_entry(ei->i_prealloc_list.next,
4141 + struct ext3_prealloc_space, pa_inode_list);
4142 + BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4143 + spin_lock(&pa->pa_lock);
4144 + if (atomic_read(&pa->pa_count)) {
4145 + /* this shouldn't happen often - nobody should
4146 + * use preallocation while we're discarding it */
4147 + spin_unlock(&pa->pa_lock);
4148 + spin_unlock(&ei->i_prealloc_lock);
4149 + current->state = TASK_UNINTERRUPTIBLE;
4150 + schedule_timeout(HZ);
4154 + if (pa->pa_deleted == 0) {
4155 + pa->pa_deleted = 1;
4156 + spin_unlock(&pa->pa_lock);
4157 + list_del_rcu(&pa->pa_inode_list);
4158 + list_add(&pa->u.pa_tmp_list, &list);
4162 + /* someone is deleting pa right now */
4163 + spin_unlock(&pa->pa_lock);
4164 + spin_unlock(&ei->i_prealloc_lock);
4166 + /* we have to wait here because pa_deleted
4167 + * doesn't mean pa is already unlinked from
4168 + * the list. as we might be called from
4169 + * ->clear_inode() the inode will get freed
4170 + * and concurrent thread which is unlinking
4171 + * pa from inode's list may access already
4172 + * freed memory, bad-bad-bad */
4174 + /* XXX: if this happens too often, we can
4175 + * add a flag to force wait only in case
4176 + * of ->clear_inode(), but not in case of
4177 + * regular truncate */
4178 + current->state = TASK_UNINTERRUPTIBLE;
4179 + schedule_timeout(HZ);
4182 + spin_unlock(&ei->i_prealloc_lock);
4184 + list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4185 + BUG_ON(pa->pa_linear != 0);
4186 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4188 + err = ext3_mb_load_buddy(sb, group, &e3b);
4192 + bitmap_bh = read_block_bitmap(sb, group);
4194 + ext3_lock_group(sb, group);
4195 + BUG_ON(e3b.bd_info->bb_prealloc_nr == 0);
4196 + e3b.bd_info->bb_prealloc_nr--;
4197 + list_del_rcu(&pa->pa_group_list);
4199 + /* can be NULL due to IO error, at worst
4200 + * we leave some free blocks unavailable
4201 + * do not go RO - no need for */
4202 + if (bitmap_bh != NULL)
4203 + ext3_mb_release_inode_pa(&e3b, bitmap_bh, pa);
4204 + ext3_unlock_group(sb, group);
4206 + ext3_mb_release_desc(&e3b);
4207 + brelse(bitmap_bh);
4209 + list_del(&pa->u.pa_tmp_list);
4215 + * finds all preallocated spaces and return blocks being freed to them
4216 + * if preallocated space becomes full (no block is used from the space)
4217 + * then the function frees space in buddy
4218 + * XXX: at the moment, truncate (which is the only way to free blocks)
4219 + * discards all preallocations
4221 +void ext3_mb_return_to_preallocation(struct inode *inode, struct ext3_buddy *e3b,
4222 + sector_t block, int count)
4224 + BUG_ON(!list_empty(&EXT3_I(inode)->i_prealloc_list));
4227 +void ext3_mb_show_ac(struct ext3_allocation_context *ac)
4230 + struct super_block *sb = ac->ac_sb;
4233 + printk(KERN_ERR "EXT3-fs: can't allocate: status %d flags %d\n",
4234 + ac->ac_status, ac->ac_flags);
4235 + printk(KERN_ERR "EXT3-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4236 + "best %lu/%lu/%lu@%lu cr %d\n",
4237 + ac->ac_o_ex.fe_group, ac->ac_o_ex.fe_start,
4238 + ac->ac_o_ex.fe_len, ac->ac_o_ex.fe_logical,
4239 + ac->ac_g_ex.fe_group, ac->ac_g_ex.fe_start,
4240 + ac->ac_g_ex.fe_len, ac->ac_g_ex.fe_logical,
4241 + ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
4242 + ac->ac_b_ex.fe_len, ac->ac_b_ex.fe_logical,
4244 + printk(KERN_ERR "EXT3-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4246 + printk("EXT3-fs: groups: ");
4247 + for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
4248 + struct ext3_group_info *grp = EXT3_GROUP_INFO(sb, i);
4249 + struct ext3_prealloc_space *pa;
4250 + unsigned long start;
4251 + struct list_head *cur;
4252 + list_for_each_rcu(cur, &grp->bb_prealloc_list) {
4253 + pa = list_entry(cur, struct ext3_prealloc_space,
4255 + spin_lock(&pa->pa_lock);
4256 + ext3_get_group_no_and_offset(sb, pa->pa_pstart, NULL, &start);
4257 + spin_unlock(&pa->pa_lock);
4258 + printk("PA:%u:%lu:%u ", i, start, pa->pa_len);
4261 + if (grp->bb_free == 0)
4263 + printk("%d: %d/%d ", i, grp->bb_free, grp->bb_fragments);
4270 +void ext3_mb_group_or_file(struct ext3_allocation_context *ac)
4272 + struct ext3_sb_info *sbi = EXT3_SB(ac->ac_sb);
4276 + if (!(ac->ac_flags & EXT3_MB_HINT_DATA))
4279 + if (ac->ac_o_ex.fe_len >= sbi->s_mb_small_req)
4282 + if (unlikely(ac->ac_flags & EXT3_MB_HINT_GOAL_ONLY))
4285 + /* request is so large that we don't care about
4286 + * streaming - it overweights any possible seek */
4287 + if (ac->ac_o_ex.fe_len >= sbi->s_mb_large_req)
4290 + bsbits = ac->ac_sb->s_blocksize_bits;
4292 + size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4293 + size = size << bsbits;
4294 + if (size < i_size_read(ac->ac_inode))
4295 + size = i_size_read(ac->ac_inode);
4296 + size = (size + ac->ac_sb->s_blocksize - 1) >> bsbits;
4298 + /* don't use group allocation for large files */
4299 + if (size >= sbi->s_mb_large_req)
4302 + BUG_ON(ac->ac_lg != NULL);
4303 + ac->ac_lg = &sbi->s_locality_groups[smp_processor_id()];
4305 + /* we're going to use group allocation */
4306 + ac->ac_flags |= EXT3_MB_HINT_GROUP_ALLOC;
4308 + /* serialize all allocations in the group */
4309 + down(&ac->ac_lg->lg_sem);
4312 +int ext3_mb_initialize_context(struct ext3_allocation_context *ac,
4313 + struct ext3_allocation_request *ar)
4315 + struct super_block *sb = ar->inode->i_sb;
4316 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4317 + struct ext3_super_block *es = sbi->s_es;
4318 + unsigned long group, len, goal;
4319 + unsigned long block;
4321 + /* we can't allocate > group size */
4323 + if (len >= EXT3_BLOCKS_PER_GROUP(sb) - 10)
4324 + len = EXT3_BLOCKS_PER_GROUP(sb) - 10;
4326 + /* start searching from the goal */
4328 + if (goal < le32_to_cpu(es->s_first_data_block) ||
4329 + goal >= le32_to_cpu(es->s_blocks_count))
4330 + goal = le32_to_cpu(es->s_first_data_block);
4331 + ext3_get_group_no_and_offset(sb, goal, &group, &block);
4333 + /* set up allocation goals */
4334 + ac->ac_b_ex.fe_logical = ar->logical;
4335 + ac->ac_b_ex.fe_group = 0;
4336 + ac->ac_b_ex.fe_start = 0;
4337 + ac->ac_b_ex.fe_len = 0;
4338 + ac->ac_status = AC_STATUS_CONTINUE;
4339 + ac->ac_groups_scanned = 0;
4340 + ac->ac_ex_scanned = 0;
4343 + ac->ac_inode = ar->inode;
4344 + ac->ac_o_ex.fe_logical = ar->logical;
4345 + ac->ac_o_ex.fe_group = group;
4346 + ac->ac_o_ex.fe_start = block;
4347 + ac->ac_o_ex.fe_len = len;
4348 + ac->ac_g_ex.fe_logical = ar->logical;
4349 + ac->ac_g_ex.fe_group = group;
4350 + ac->ac_g_ex.fe_start = block;
4351 + ac->ac_g_ex.fe_len = len;
4352 + ac->ac_f_ex.fe_len = 0;
4353 + ac->ac_flags = ar->flags;
4354 + ac->ac_2order = 0;
4355 + ac->ac_criteria = 0;
4357 + ac->ac_bitmap_page = NULL;
4358 + ac->ac_buddy_page = NULL;
4361 + /* we have to define context: we'll we work with a file or
4362 + * locality group. this is a policy, actually */
4363 + ext3_mb_group_or_file(ac);
4365 + mb_debug("init ac: %u blocks @ %llu, goal %llu, flags %x, 2^%d, "
4366 + "left: %llu/%llu, right %llu/%llu to %swritable\n",
4367 + (unsigned) ar->len, (unsigned) ar->logical,
4368 + (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4369 + (unsigned) ar->lleft, (unsigned) ar->pleft,
4370 + (unsigned) ar->lright, (unsigned) ar->pright,
4371 + atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4377 + * release all resource we used in allocation
4379 +int ext3_mb_release_context(struct ext3_allocation_context *ac)
4382 + if (ac->ac_pa->pa_linear) {
4383 + /* see comment in ext3_mb_use_group_pa() */
4384 + spin_lock(&ac->ac_pa->pa_lock);
4385 + ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
4386 + ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
4387 + ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
4388 + ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
4389 + spin_unlock(&ac->ac_pa->pa_lock);
4391 + ext3_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
4393 + if (ac->ac_bitmap_page)
4394 + page_cache_release(ac->ac_bitmap_page);
4395 + if (ac->ac_buddy_page)
4396 + page_cache_release(ac->ac_buddy_page);
4397 + if (ac->ac_flags & EXT3_MB_HINT_GROUP_ALLOC)
4398 + up(&ac->ac_lg->lg_sem);
4399 + ext3_mb_collect_stats(ac);
4403 +int ext3_mb_discard_preallocations(struct super_block *sb, int needed)
4405 + int i, ret, freed = 0;
4407 + for (i = 0; i < EXT3_SB(sb)->s_groups_count && needed > 0; i++) {
4408 + ret = ext3_mb_discard_group_preallocations(sb, i, needed);
4417 + * Main entry point into mballoc to allocate blocks
4418 + * it tries to use preallocation first, then falls back
4419 + * to usual allocation
4421 +unsigned long ext3_mb_new_blocks(handle_t *handle,
4422 + struct ext3_allocation_request *ar, int *errp)
4424 + struct ext3_allocation_context ac;
4425 + struct ext3_sb_info *sbi;
4426 + struct super_block *sb;
4427 + unsigned long block = 0;
4428 + int freed, inquota;
4430 + sb = ar->inode->i_sb;
4431 + sbi = EXT3_SB(sb);
4433 + if (!test_opt(sb, MBALLOC)) {
4434 + static int ext3_mballoc_warning = 0;
4435 + if (ext3_mballoc_warning++ == 0)
4436 + printk(KERN_ERR "EXT3-fs: multiblock request with "
4437 + "mballoc disabled!\n");
4439 + block = ext3_new_block_old(handle, ar->inode, ar->goal, errp);
4443 + while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4444 + ar->flags |= EXT3_MB_HINT_NOPREALLOC;
4447 + if (ar->len == 0) {
4451 + inquota = ar->len;
4453 + ext3_mb_poll_new_transaction(sb, handle);
4455 + if ((*errp = ext3_mb_initialize_context(&ac, ar))) {
4460 + ac.ac_op = EXT3_MB_HISTORY_PREALLOC;
4461 + if (!ext3_mb_use_preallocated(&ac)) {
4463 + ac.ac_op = EXT3_MB_HISTORY_ALLOC;
4464 + ext3_mb_normalize_request(&ac, ar);
4467 + /* allocate space in core */
4468 + ext3_mb_regular_allocator(&ac);
4470 + /* as we've just preallocated more space than
4471 + * user requested orinally, we store allocated
4472 + * space in a special descriptor */
4473 + if (ac.ac_status == AC_STATUS_FOUND &&
4474 + ac.ac_o_ex.fe_len < ac.ac_b_ex.fe_len)
4475 + ext3_mb_new_preallocation(&ac);
4478 + if (likely(ac.ac_status == AC_STATUS_FOUND)) {
4479 + ext3_mb_mark_diskspace_used(&ac, handle);
4481 + block = ext3_grp_offs_to_block(sb, &ac.ac_b_ex);
4482 + ar->len = ac.ac_b_ex.fe_len;
4484 + freed = ext3_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
4488 + ac.ac_b_ex.fe_len = 0;
4490 + ext3_mb_show_ac(&ac);
4493 + ext3_mb_release_context(&ac);
4496 + if (ar->len < inquota)
4497 + DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4501 +EXPORT_SYMBOL(ext3_mb_new_blocks);
4503 +int ext3_new_block(handle_t *handle, struct inode *inode,
4504 + unsigned long goal, int *errp)
4506 + struct ext3_allocation_request ar;
4507 + unsigned long ret;
4509 + if (!test_opt(inode->i_sb, MBALLOC)) {
4510 + ret = ext3_new_block_old(handle, inode, goal, errp);
4523 + ret = ext3_mb_new_blocks(handle, &ar, errp);
4527 +void ext3_mb_poll_new_transaction(struct super_block *sb, handle_t *handle)
4529 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4531 + if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4534 + /* new transaction! time to close last one and free blocks for
4535 + * committed transaction. we know that only transaction can be
4536 + * active, so previos transaction can be being logged and we
4537 + * know that transaction before previous is known to be already
4538 + * logged. this means that now we may free blocks freed in all
4539 + * transactions before previous one. hope I'm clear enough ... */
4541 + spin_lock(&sbi->s_md_lock);
4542 + if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4543 + mb_debug("new transaction %lu, old %lu\n",
4544 + (unsigned long) handle->h_transaction->t_tid,
4545 + (unsigned long) sbi->s_last_transaction);
4546 + list_splice_init(&sbi->s_closed_transaction,
4547 + &sbi->s_committed_transaction);
4548 + list_splice_init(&sbi->s_active_transaction,
4549 + &sbi->s_closed_transaction);
4550 + sbi->s_last_transaction = handle->h_transaction->t_tid;
4552 + spin_unlock(&sbi->s_md_lock);
4554 + ext3_mb_free_committed_blocks(sb);
4557 +int ext3_mb_free_metadata(handle_t *handle, struct ext3_buddy *e3b,
4558 + int group, int block, int count)
4560 + struct ext3_group_info *db = e3b->bd_info;
4561 + struct super_block *sb = e3b->bd_sb;
4562 + struct ext3_sb_info *sbi = EXT3_SB(sb);
4563 + struct ext3_free_metadata *md;
4566 + BUG_ON(e3b->bd_bitmap_page == NULL);
4567 + BUG_ON(e3b->bd_buddy_page == NULL);
4569 + ext3_lock_group(sb, group);
4570 + for (i = 0; i < count; i++) {
4571 + md = db->bb_md_cur;
4572 + if (md && db->bb_tid != handle->h_transaction->t_tid) {
4573 + db->bb_md_cur = NULL;
4578 + ext3_unlock_group(sb, group);
4579 + md = kmalloc(sizeof(*md), GFP_KERNEL);
4583 + md->group = group;
4585 + ext3_lock_group(sb, group);
4586 + if (db->bb_md_cur == NULL) {
4587 + spin_lock(&sbi->s_md_lock);
4588 + list_add(&md->list, &sbi->s_active_transaction);
4589 + spin_unlock(&sbi->s_md_lock);
4590 + /* protect buddy cache from being freed,
4591 + * otherwise we'll refresh it from
4592 + * on-disk bitmap and lose not-yet-available
4594 + page_cache_get(e3b->bd_buddy_page);
4595 + page_cache_get(e3b->bd_bitmap_page);
4596 + db->bb_md_cur = md;
4597 + db->bb_tid = handle->h_transaction->t_tid;
4598 + mb_debug("new md 0x%p for group %u\n",
4602 + md = db->bb_md_cur;
4606 + BUG_ON(md->num >= EXT3_BB_MAX_BLOCKS);
4607 + md->blocks[md->num] = block + i;
4609 + if (md->num == EXT3_BB_MAX_BLOCKS) {
4610 + /* no more space, put full container on a sb's list */
4611 + db->bb_md_cur = NULL;
4614 + ext3_unlock_group(sb, group);
4619 + * Main entry point into mballoc to free blocks
4621 +void ext3_mb_free_blocks(handle_t *handle, struct inode *inode,
4622 + unsigned long block, unsigned long count,
4623 + int metadata, int *freed)
4625 + struct buffer_head *bitmap_bh = NULL;
4626 + struct super_block *sb = inode->i_sb;
4627 + struct ext3_allocation_context ac;
4628 + struct ext3_group_desc *gdp;
4629 + struct ext3_super_block *es;
4630 + unsigned long bit, overflow;
4631 + struct buffer_head *gd_bh;
4632 + unsigned long block_group;
4633 + struct ext3_sb_info *sbi;
4634 + struct ext3_buddy e3b;
4639 + ext3_mb_poll_new_transaction(sb, handle);
4641 + sbi = EXT3_SB(sb);
4642 + es = EXT3_SB(sb)->s_es;
4643 + if (block < le32_to_cpu(es->s_first_data_block) ||
4644 + block + count < block ||
4645 + block + count > le32_to_cpu(es->s_blocks_count)) {
4646 + ext3_error (sb, __FUNCTION__,
4647 + "Freeing blocks not in datazone - "
4648 + "block = %lu, count = %lu", block, count);
4649 + goto error_return;
4652 + ext3_debug("freeing block %lu\n", block);
4654 + ac.ac_op = EXT3_MB_HISTORY_FREE;
4655 + ac.ac_inode = inode;
4660 + ext3_get_group_no_and_offset(sb, block, &block_group, &bit);
4663 + * Check to see if we are freeing blocks across a group
4666 + if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
4667 + overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
4668 + count -= overflow;
4670 + brelse(bitmap_bh);
4671 + bitmap_bh = read_block_bitmap(sb, block_group);
4673 + goto error_return;
4674 + gdp = ext3_get_group_desc (sb, block_group, &gd_bh);
4676 + goto error_return;
4678 + if (in_range (le32_to_cpu(gdp->bg_block_bitmap), block, count) ||
4679 + in_range (le32_to_cpu(gdp->bg_inode_bitmap), block, count) ||
4680 + in_range (block, le32_to_cpu(gdp->bg_inode_table),
4681 + EXT3_SB(sb)->s_itb_per_group) ||
4682 + in_range (block + count - 1, le32_to_cpu(gdp->bg_inode_table),
4683 + EXT3_SB(sb)->s_itb_per_group))
4684 + ext3_error(sb, __FUNCTION__,
4685 + "Freeing blocks in system zone - "
4686 + "Block = %lu, count = %lu", block, count);
4688 + BUFFER_TRACE(bitmap_bh, "getting write access");
4689 + err = ext3_journal_get_write_access(handle, bitmap_bh);
4691 + goto error_return;
4694 + * We are about to modify some metadata. Call the journal APIs
4695 + * to unshare ->b_data if a currently-committing transaction is
4698 + BUFFER_TRACE(gd_bh, "get_write_access");
4699 + err = ext3_journal_get_write_access(handle, gd_bh);
4701 + goto error_return;
4703 + err = ext3_mb_load_buddy(sb, block_group, &e3b);
4705 + goto error_return;
4707 +#ifdef AGGRESSIVE_CHECK
4710 + for (i = 0; i < count; i++)
4711 + BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4714 + mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data, bit,
4717 + /* We dirtied the bitmap block */
4718 + BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4719 + err = ext3_journal_dirty_metadata(handle, bitmap_bh);
4721 + ac.ac_b_ex.fe_group = block_group;
4722 + ac.ac_b_ex.fe_start = bit;
4723 + ac.ac_b_ex.fe_len = count;
4724 + ext3_mb_store_history(&ac);
4727 + /* blocks being freed are metadata. these blocks shouldn't
4728 + * be used until this transaction is committed */
4729 + ext3_mb_free_metadata(handle, &e3b, block_group, bit, count);
4731 + ext3_lock_group(sb, block_group);
4732 + err = mb_free_blocks(inode, &e3b, bit, count);
4733 + ext3_mb_return_to_preallocation(inode, &e3b, block, count);
4734 + ext3_unlock_group(sb, block_group);
4738 + spin_lock(sb_bgl_lock(sbi, block_group));
4739 + gdp->bg_free_blocks_count =
4740 + cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
4741 + spin_unlock(sb_bgl_lock(sbi, block_group));
4742 + percpu_counter_mod(&sbi->s_freeblocks_counter, count);
4744 + ext3_mb_release_desc(&e3b);
4748 + /* And the group descriptor block */
4749 + BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4750 + ret = ext3_journal_dirty_metadata(handle, gd_bh);
4751 + if (!err) err = ret;
4753 + if (overflow && !err) {
4760 + brelse(bitmap_bh);
4761 + ext3_std_error(sb, err);
4765 +EXPORT_SYMBOL(ext3_free_blocks);
4766 +EXPORT_SYMBOL(ext3_mb_discard_inode_preallocations);