aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ext4/mballoc.c
diff options
context:
space:
mode:
authorAlex Tomas <alex@clusterfs.com>2008-01-29 00:19:52 -0500
committerTheodore Ts'o <tytso@mit.edu>2008-01-29 00:19:52 -0500
commitc9de560ded61faa5b754137b7753da252391c55a (patch)
tree2c4311377c4aa72450e27f531e198fe3e1c67db0 /fs/ext4/mballoc.c
parent1988b51e476bd097d910c9245b53f2e38aedaf0d (diff)
ext4: Add multi block allocator for ext4
Signed-off-by: Alex Tomas <alex@clusterfs.com> Signed-off-by: Andreas Dilger <adilger@clusterfs.com> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Eric Sandeen <sandeen@redhat.com> Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Diffstat (limited to 'fs/ext4/mballoc.c')
-rw-r--r--fs/ext4/mballoc.c4552
1 files changed, 4552 insertions, 0 deletions
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
new file mode 100644
index 000000000000..76e5fedc0a0b
--- /dev/null
+++ b/fs/ext4/mballoc.c
@@ -0,0 +1,4552 @@
1/*
2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
17 */
18
19
20/*
21 * mballoc.c contains the multiblocks allocation routines
22 */
23
24#include <linux/time.h>
25#include <linux/fs.h>
26#include <linux/namei.h>
27#include <linux/ext4_jbd2.h>
28#include <linux/ext4_fs.h>
29#include <linux/quotaops.h>
30#include <linux/buffer_head.h>
31#include <linux/module.h>
32#include <linux/swap.h>
33#include <linux/proc_fs.h>
34#include <linux/pagemap.h>
35#include <linux/seq_file.h>
36#include <linux/version.h>
37#include "group.h"
38
39/*
40 * MUSTDO:
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
43 *
44 * TODO v4:
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
48 * - quota
49 * - reservation for superuser
50 *
51 * TODO v3:
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
56 * - error handling
57 */
58
59/*
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
62 *
63 * During initialization phase of the allocator we decide to use the group
64 * preallocation or inode preallocation depending on the size file. The
65 * size of the file could be the resulting file size we would have after
66 * allocation or the current file size which ever is larger. If the size is
67 * less that sbi->s_mb_stream_request we select the group
68 * preallocation. The default value of s_mb_stream_request is 16
69 * blocks. This can also be tuned via
70 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
71 * of number of blocks.
72 *
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small file closer in the disk.
75 *
76 * First stage the allocator looks at the inode prealloc list
77 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
78 * this particular inode. The inode prealloc space is represented as:
79 *
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> lenght for this prealloc space
83 * pa_free -> free space available in this prealloc space
84 *
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This make sure that
88 * that the we have contiguous physical blocks representing the file blocks
89 *
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
92 * pa_free.
93 *
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list repreasented as
97 *
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
99 *
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
102 *
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) withing the prealloc space.
105 *
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
114 * inode as:
115 *
116 * { page }
117 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
118 *
119 *
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
124 *
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
127 *
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
131 *
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
139 * 512 blocks. This can be tuned via
140 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
141 * terms of number of blocks. If we have mounted the file system with -O
142 * stripe=<value> option the group prealloc request is normalized to the
143 * stripe value (sbi->s_stripe)
144 *
145 * The regular allocator(using the buddy cache) support few tunables.
146 *
147 * /proc/fs/ext4/<partition>/min_to_scan
148 * /proc/fs/ext4/<partition>/max_to_scan
149 * /proc/fs/ext4/<partition>/order2_req
150 *
151 * The regular allocator use buddy scan only if the request len is power of
152 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
153 * value of s_mb_order2_reqs can be tuned via
154 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
155 * stripe size (sbi->s_stripe), we try to search for contigous block in
156 * stripe size. This should result in better allocation on RAID setup. If
157 * not we search in the specific group using bitmap for best extents. The
158 * tunable min_to_scan and max_to_scan controll the behaviour here.
159 * min_to_scan indicate how long the mballoc __must__ look for a best
160 * extent and max_to_scanindicate how long the mballoc __can__ look for a
161 * best extent in the found extents. Searching for the blocks starts with
162 * the group specified as the goal value in allocation context via
163 * ac_g_ex. Each group is first checked based on the criteria whether it
164 * can used for allocation. ext4_mb_good_group explains how the groups are
165 * checked.
166 *
167 * Both the prealloc space are getting populated as above. So for the first
168 * request we will hit the buddy cache which will result in this prealloc
169 * space getting filled. The prealloc space is then later used for the
170 * subsequent request.
171 */
172
173/*
174 * mballoc operates on the following data:
175 * - on-disk bitmap
176 * - in-core buddy (actually includes buddy and bitmap)
177 * - preallocation descriptors (PAs)
178 *
179 * there are two types of preallocations:
180 * - inode
181 * assiged to specific inode and can be used for this inode only.
182 * it describes part of inode's space preallocated to specific
183 * physical blocks. any block from that preallocated can be used
184 * independent. the descriptor just tracks number of blocks left
185 * unused. so, before taking some block from descriptor, one must
186 * make sure corresponded logical block isn't allocated yet. this
187 * also means that freeing any block within descriptor's range
188 * must discard all preallocated blocks.
189 * - locality group
190 * assigned to specific locality group which does not translate to
191 * permanent set of inodes: inode can join and leave group. space
192 * from this type of preallocation can be used for any inode. thus
193 * it's consumed from the beginning to the end.
194 *
195 * relation between them can be expressed as:
196 * in-core buddy = on-disk bitmap + preallocation descriptors
197 *
198 * this mean blocks mballoc considers used are:
199 * - allocated blocks (persistent)
200 * - preallocated blocks (non-persistent)
201 *
202 * consistency in mballoc world means that at any time a block is either
203 * free or used in ALL structures. notice: "any time" should not be read
204 * literally -- time is discrete and delimited by locks.
205 *
206 * to keep it simple, we don't use block numbers, instead we count number of
207 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
208 *
209 * all operations can be expressed as:
210 * - init buddy: buddy = on-disk + PAs
211 * - new PA: buddy += N; PA = N
212 * - use inode PA: on-disk += N; PA -= N
213 * - discard inode PA buddy -= on-disk - PA; PA = 0
214 * - use locality group PA on-disk += N; PA -= N
215 * - discard locality group PA buddy -= PA; PA = 0
216 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
217 * is used in real operation because we can't know actual used
218 * bits from PA, only from on-disk bitmap
219 *
220 * if we follow this strict logic, then all operations above should be atomic.
221 * given some of them can block, we'd have to use something like semaphores
222 * killing performance on high-end SMP hardware. let's try to relax it using
223 * the following knowledge:
224 * 1) if buddy is referenced, it's already initialized
225 * 2) while block is used in buddy and the buddy is referenced,
226 * nobody can re-allocate that block
227 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
228 * bit set and PA claims same block, it's OK. IOW, one can set bit in
229 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
230 * block
231 *
232 * so, now we're building a concurrency table:
233 * - init buddy vs.
234 * - new PA
235 * blocks for PA are allocated in the buddy, buddy must be referenced
236 * until PA is linked to allocation group to avoid concurrent buddy init
237 * - use inode PA
238 * we need to make sure that either on-disk bitmap or PA has uptodate data
239 * given (3) we care that PA-=N operation doesn't interfere with init
240 * - discard inode PA
241 * the simplest way would be to have buddy initialized by the discard
242 * - use locality group PA
243 * again PA-=N must be serialized with init
244 * - discard locality group PA
245 * the simplest way would be to have buddy initialized by the discard
246 * - new PA vs.
247 * - use inode PA
248 * i_data_sem serializes them
249 * - discard inode PA
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * some mutex should serialize them
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
255 * - use inode PA
256 * - use inode PA
257 * i_data_sem or another mutex should serializes them
258 * - discard inode PA
259 * discard process must wait until PA isn't used by another process
260 * - use locality group PA
261 * nothing wrong here -- they're different PAs covering different blocks
262 * - discard locality group PA
263 * discard process must wait until PA isn't used by another process
264 *
265 * now we're ready to make few consequences:
266 * - PA is referenced and while it is no discard is possible
267 * - PA is referenced until block isn't marked in on-disk bitmap
268 * - PA changes only after on-disk bitmap
269 * - discard must not compete with init. either init is done before
270 * any discard or they're serialized somehow
271 * - buddy init as sum of on-disk bitmap and PAs is done atomically
272 *
273 * a special case when we've used PA to emptiness. no need to modify buddy
274 * in this case, but we should care about concurrent init
275 *
276 */
277
278 /*
279 * Logic in few words:
280 *
281 * - allocation:
282 * load group
283 * find blocks
284 * mark bits in on-disk bitmap
285 * release group
286 *
287 * - use preallocation:
288 * find proper PA (per-inode or group)
289 * load group
290 * mark bits in on-disk bitmap
291 * release group
292 * release PA
293 *
294 * - free:
295 * load group
296 * mark bits in on-disk bitmap
297 * release group
298 *
299 * - discard preallocations in group:
300 * mark PAs deleted
301 * move them onto local list
302 * load on-disk bitmap
303 * load group
304 * remove PA from object (inode or locality group)
305 * mark free blocks in-core
306 *
307 * - discard inode's preallocations:
308 */
309
310/*
311 * Locking rules
312 *
313 * Locks:
314 * - bitlock on a group (group)
315 * - object (inode/locality) (object)
316 * - per-pa lock (pa)
317 *
318 * Paths:
319 * - new pa
320 * object
321 * group
322 *
323 * - find and use pa:
324 * pa
325 *
326 * - release consumed pa:
327 * pa
328 * group
329 * object
330 *
331 * - generate in-core bitmap:
332 * group
333 * pa
334 *
335 * - discard all for given object (inode, locality group):
336 * object
337 * pa
338 * group
339 *
340 * - discard all for given group:
341 * group
342 * pa
343 * group
344 * object
345 *
346 */
347
348/*
349 * with AGGRESSIVE_CHECK allocator runs consistency checks over
350 * structures. these checks slow things down a lot
351 */
352#define AGGRESSIVE_CHECK__
353
354/*
355 * with DOUBLE_CHECK defined mballoc creates persistent in-core
356 * bitmaps, maintains and uses them to check for double allocations
357 */
358#define DOUBLE_CHECK__
359
360/*
361 */
362#define MB_DEBUG__
363#ifdef MB_DEBUG
364#define mb_debug(fmt, a...) printk(fmt, ##a)
365#else
366#define mb_debug(fmt, a...)
367#endif
368
369/*
370 * with EXT4_MB_HISTORY mballoc stores last N allocations in memory
371 * and you can monitor it in /proc/fs/ext4/<dev>/mb_history
372 */
373#define EXT4_MB_HISTORY
374#define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
375#define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
376#define EXT4_MB_HISTORY_DISCARD 4 /* preallocation discarded */
377#define EXT4_MB_HISTORY_FREE 8 /* free */
378
379#define EXT4_MB_HISTORY_DEFAULT (EXT4_MB_HISTORY_ALLOC | \
380 EXT4_MB_HISTORY_PREALLOC)
381
382/*
383 * How long mballoc can look for a best extent (in found extents)
384 */
385#define MB_DEFAULT_MAX_TO_SCAN 200
386
387/*
388 * How long mballoc must look for a best extent
389 */
390#define MB_DEFAULT_MIN_TO_SCAN 10
391
392/*
393 * How many groups mballoc will scan looking for the best chunk
394 */
395#define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
396
397/*
398 * with 'ext4_mb_stats' allocator will collect stats that will be
399 * shown at umount. The collecting costs though!
400 */
401#define MB_DEFAULT_STATS 1
402
403/*
404 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
405 * by the stream allocator, which purpose is to pack requests
406 * as close each to other as possible to produce smooth I/O traffic
407 * We use locality group prealloc space for stream request.
408 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
409 */
410#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
411
412/*
413 * for which requests use 2^N search using buddies
414 */
415#define MB_DEFAULT_ORDER2_REQS 2
416
417/*
418 * default group prealloc size 512 blocks
419 */
420#define MB_DEFAULT_GROUP_PREALLOC 512
421
422static struct kmem_cache *ext4_pspace_cachep;
423
424#ifdef EXT4_BB_MAX_BLOCKS
425#undef EXT4_BB_MAX_BLOCKS
426#endif
427#define EXT4_BB_MAX_BLOCKS 30
428
429struct ext4_free_metadata {
430 ext4_group_t group;
431 unsigned short num;
432 ext4_grpblk_t blocks[EXT4_BB_MAX_BLOCKS];
433 struct list_head list;
434};
435
436struct ext4_group_info {
437 unsigned long bb_state;
438 unsigned long bb_tid;
439 struct ext4_free_metadata *bb_md_cur;
440 unsigned short bb_first_free;
441 unsigned short bb_free;
442 unsigned short bb_fragments;
443 struct list_head bb_prealloc_list;
444#ifdef DOUBLE_CHECK
445 void *bb_bitmap;
446#endif
447 unsigned short bb_counters[];
448};
449
450#define EXT4_GROUP_INFO_NEED_INIT_BIT 0
451#define EXT4_GROUP_INFO_LOCKED_BIT 1
452
453#define EXT4_MB_GRP_NEED_INIT(grp) \
454 (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
455
456
457struct ext4_prealloc_space {
458 struct list_head pa_inode_list;
459 struct list_head pa_group_list;
460 union {
461 struct list_head pa_tmp_list;
462 struct rcu_head pa_rcu;
463 } u;
464 spinlock_t pa_lock;
465 atomic_t pa_count;
466 unsigned pa_deleted;
467 ext4_fsblk_t pa_pstart; /* phys. block */
468 ext4_lblk_t pa_lstart; /* log. block */
469 unsigned short pa_len; /* len of preallocated chunk */
470 unsigned short pa_free; /* how many blocks are free */
471 unsigned short pa_linear; /* consumed in one direction
472 * strictly, for grp prealloc */
473 spinlock_t *pa_obj_lock;
474 struct inode *pa_inode; /* hack, for history only */
475};
476
477
478struct ext4_free_extent {
479 ext4_lblk_t fe_logical;
480 ext4_grpblk_t fe_start;
481 ext4_group_t fe_group;
482 int fe_len;
483};
484
485/*
486 * Locality group:
487 * we try to group all related changes together
488 * so that writeback can flush/allocate them together as well
489 */
490struct ext4_locality_group {
491 /* for allocator */
492 struct mutex lg_mutex; /* to serialize allocates */
493 struct list_head lg_prealloc_list;/* list of preallocations */
494 spinlock_t lg_prealloc_lock;
495};
496
497struct ext4_allocation_context {
498 struct inode *ac_inode;
499 struct super_block *ac_sb;
500
501 /* original request */
502 struct ext4_free_extent ac_o_ex;
503
504 /* goal request (after normalization) */
505 struct ext4_free_extent ac_g_ex;
506
507 /* the best found extent */
508 struct ext4_free_extent ac_b_ex;
509
510 /* copy of the bext found extent taken before preallocation efforts */
511 struct ext4_free_extent ac_f_ex;
512
513 /* number of iterations done. we have to track to limit searching */
514 unsigned long ac_ex_scanned;
515 __u16 ac_groups_scanned;
516 __u16 ac_found;
517 __u16 ac_tail;
518 __u16 ac_buddy;
519 __u16 ac_flags; /* allocation hints */
520 __u8 ac_status;
521 __u8 ac_criteria;
522 __u8 ac_repeats;
523 __u8 ac_2order; /* if request is to allocate 2^N blocks and
524 * N > 0, the field stores N, otherwise 0 */
525 __u8 ac_op; /* operation, for history only */
526 struct page *ac_bitmap_page;
527 struct page *ac_buddy_page;
528 struct ext4_prealloc_space *ac_pa;
529 struct ext4_locality_group *ac_lg;
530};
531
532#define AC_STATUS_CONTINUE 1
533#define AC_STATUS_FOUND 2
534#define AC_STATUS_BREAK 3
535
536struct ext4_mb_history {
537 struct ext4_free_extent orig; /* orig allocation */
538 struct ext4_free_extent goal; /* goal allocation */
539 struct ext4_free_extent result; /* result allocation */
540 unsigned pid;
541 unsigned ino;
542 __u16 found; /* how many extents have been found */
543 __u16 groups; /* how many groups have been scanned */
544 __u16 tail; /* what tail broke some buddy */
545 __u16 buddy; /* buddy the tail ^^^ broke */
546 __u16 flags;
547 __u8 cr:3; /* which phase the result extent was found at */
548 __u8 op:4;
549 __u8 merged:1;
550};
551
552struct ext4_buddy {
553 struct page *bd_buddy_page;
554 void *bd_buddy;
555 struct page *bd_bitmap_page;
556 void *bd_bitmap;
557 struct ext4_group_info *bd_info;
558 struct super_block *bd_sb;
559 __u16 bd_blkbits;
560 ext4_group_t bd_group;
561};
562#define EXT4_MB_BITMAP(e4b) ((e4b)->bd_bitmap)
563#define EXT4_MB_BUDDY(e4b) ((e4b)->bd_buddy)
564
565#ifndef EXT4_MB_HISTORY
566static inline void ext4_mb_store_history(struct ext4_allocation_context *ac)
567{
568 return;
569}
570#else
571static void ext4_mb_store_history(struct ext4_allocation_context *ac);
572#endif
573
574#define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
575
576static struct proc_dir_entry *proc_root_ext4;
577struct buffer_head *read_block_bitmap(struct super_block *, ext4_group_t);
578ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
579 ext4_fsblk_t goal, unsigned long *count, int *errp);
580
581static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
582 ext4_group_t group);
583static void ext4_mb_poll_new_transaction(struct super_block *, handle_t *);
584static void ext4_mb_free_committed_blocks(struct super_block *);
585static void ext4_mb_return_to_preallocation(struct inode *inode,
586 struct ext4_buddy *e4b, sector_t block,
587 int count);
588static void ext4_mb_put_pa(struct ext4_allocation_context *,
589 struct super_block *, struct ext4_prealloc_space *pa);
590static int ext4_mb_init_per_dev_proc(struct super_block *sb);
591static int ext4_mb_destroy_per_dev_proc(struct super_block *sb);
592
593
594static inline void ext4_lock_group(struct super_block *sb, ext4_group_t group)
595{
596 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
597
598 bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
599}
600
601static inline void ext4_unlock_group(struct super_block *sb,
602 ext4_group_t group)
603{
604 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
605
606 bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
607}
608
609static inline int ext4_is_group_locked(struct super_block *sb,
610 ext4_group_t group)
611{
612 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
613
614 return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT,
615 &(grinfo->bb_state));
616}
617
618static ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
619 struct ext4_free_extent *fex)
620{
621 ext4_fsblk_t block;
622
623 block = (ext4_fsblk_t) fex->fe_group * EXT4_BLOCKS_PER_GROUP(sb)
624 + fex->fe_start
625 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
626 return block;
627}
628
629#if BITS_PER_LONG == 64
630#define mb_correct_addr_and_bit(bit, addr) \
631{ \
632 bit += ((unsigned long) addr & 7UL) << 3; \
633 addr = (void *) ((unsigned long) addr & ~7UL); \
634}
635#elif BITS_PER_LONG == 32
636#define mb_correct_addr_and_bit(bit, addr) \
637{ \
638 bit += ((unsigned long) addr & 3UL) << 3; \
639 addr = (void *) ((unsigned long) addr & ~3UL); \
640}
641#else
642#error "how many bits you are?!"
643#endif
644
645static inline int mb_test_bit(int bit, void *addr)
646{
647 /*
648 * ext4_test_bit on architecture like powerpc
649 * needs unsigned long aligned address
650 */
651 mb_correct_addr_and_bit(bit, addr);
652 return ext4_test_bit(bit, addr);
653}
654
655static inline void mb_set_bit(int bit, void *addr)
656{
657 mb_correct_addr_and_bit(bit, addr);
658 ext4_set_bit(bit, addr);
659}
660
661static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
662{
663 mb_correct_addr_and_bit(bit, addr);
664 ext4_set_bit_atomic(lock, bit, addr);
665}
666
667static inline void mb_clear_bit(int bit, void *addr)
668{
669 mb_correct_addr_and_bit(bit, addr);
670 ext4_clear_bit(bit, addr);
671}
672
673static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
674{
675 mb_correct_addr_and_bit(bit, addr);
676 ext4_clear_bit_atomic(lock, bit, addr);
677}
678
679static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
680{
681 char *bb;
682
683 /* FIXME!! is this needed */
684 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
685 BUG_ON(max == NULL);
686
687 if (order > e4b->bd_blkbits + 1) {
688 *max = 0;
689 return NULL;
690 }
691
692 /* at order 0 we see each particular block */
693 *max = 1 << (e4b->bd_blkbits + 3);
694 if (order == 0)
695 return EXT4_MB_BITMAP(e4b);
696
697 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
698 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
699
700 return bb;
701}
702
703#ifdef DOUBLE_CHECK
704static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
705 int first, int count)
706{
707 int i;
708 struct super_block *sb = e4b->bd_sb;
709
710 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
711 return;
712 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
713 for (i = 0; i < count; i++) {
714 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
715 ext4_fsblk_t blocknr;
716 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
717 blocknr += first + i;
718 blocknr +=
719 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
720
721 ext4_error(sb, __FUNCTION__, "double-free of inode"
722 " %lu's block %llu(bit %u in group %lu)\n",
723 inode ? inode->i_ino : 0, blocknr,
724 first + i, e4b->bd_group);
725 }
726 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
727 }
728}
729
730static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
731{
732 int i;
733
734 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
735 return;
736 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
737 for (i = 0; i < count; i++) {
738 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
739 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
740 }
741}
742
743static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
744{
745 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
746 unsigned char *b1, *b2;
747 int i;
748 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
749 b2 = (unsigned char *) bitmap;
750 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
751 if (b1[i] != b2[i]) {
752 printk("corruption in group %lu at byte %u(%u):"
753 " %x in copy != %x on disk/prealloc\n",
754 e4b->bd_group, i, i * 8, b1[i], b2[i]);
755 BUG();
756 }
757 }
758 }
759}
760
761#else
762static inline void mb_free_blocks_double(struct inode *inode,
763 struct ext4_buddy *e4b, int first, int count)
764{
765 return;
766}
767static inline void mb_mark_used_double(struct ext4_buddy *e4b,
768 int first, int count)
769{
770 return;
771}
772static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
773{
774 return;
775}
776#endif
777
778#ifdef AGGRESSIVE_CHECK
779
780#define MB_CHECK_ASSERT(assert) \
781do { \
782 if (!(assert)) { \
783 printk(KERN_EMERG \
784 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
785 function, file, line, # assert); \
786 BUG(); \
787 } \
788} while (0)
789
790static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
791 const char *function, int line)
792{
793 struct super_block *sb = e4b->bd_sb;
794 int order = e4b->bd_blkbits + 1;
795 int max;
796 int max2;
797 int i;
798 int j;
799 int k;
800 int count;
801 struct ext4_group_info *grp;
802 int fragments = 0;
803 int fstart;
804 struct list_head *cur;
805 void *buddy;
806 void *buddy2;
807
808 if (!test_opt(sb, MBALLOC))
809 return 0;
810
811 {
812 static int mb_check_counter;
813 if (mb_check_counter++ % 100 != 0)
814 return 0;
815 }
816
817 while (order > 1) {
818 buddy = mb_find_buddy(e4b, order, &max);
819 MB_CHECK_ASSERT(buddy);
820 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
821 MB_CHECK_ASSERT(buddy2);
822 MB_CHECK_ASSERT(buddy != buddy2);
823 MB_CHECK_ASSERT(max * 2 == max2);
824
825 count = 0;
826 for (i = 0; i < max; i++) {
827
828 if (mb_test_bit(i, buddy)) {
829 /* only single bit in buddy2 may be 1 */
830 if (!mb_test_bit(i << 1, buddy2)) {
831 MB_CHECK_ASSERT(
832 mb_test_bit((i<<1)+1, buddy2));
833 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
834 MB_CHECK_ASSERT(
835 mb_test_bit(i << 1, buddy2));
836 }
837 continue;
838 }
839
840 /* both bits in buddy2 must be 0 */
841 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
842 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
843
844 for (j = 0; j < (1 << order); j++) {
845 k = (i * (1 << order)) + j;
846 MB_CHECK_ASSERT(
847 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
848 }
849 count++;
850 }
851 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
852 order--;
853 }
854
855 fstart = -1;
856 buddy = mb_find_buddy(e4b, 0, &max);
857 for (i = 0; i < max; i++) {
858 if (!mb_test_bit(i, buddy)) {
859 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
860 if (fstart == -1) {
861 fragments++;
862 fstart = i;
863 }
864 continue;
865 }
866 fstart = -1;
867 /* check used bits only */
868 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
869 buddy2 = mb_find_buddy(e4b, j, &max2);
870 k = i >> j;
871 MB_CHECK_ASSERT(k < max2);
872 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
873 }
874 }
875 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
876 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
877
878 grp = ext4_get_group_info(sb, e4b->bd_group);
879 buddy = mb_find_buddy(e4b, 0, &max);
880 list_for_each(cur, &grp->bb_prealloc_list) {
881 ext4_group_t groupnr;
882 struct ext4_prealloc_space *pa;
883 pa = list_entry(cur, struct ext4_prealloc_space, group_list);
884 ext4_get_group_no_and_offset(sb, pa->pstart, &groupnr, &k);
885 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
886 for (i = 0; i < pa->len; i++)
887 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
888 }
889 return 0;
890}
891#undef MB_CHECK_ASSERT
892#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
893 __FILE__, __FUNCTION__, __LINE__)
894#else
895#define mb_check_buddy(e4b)
896#endif
897
898/* FIXME!! need more doc */
899static void ext4_mb_mark_free_simple(struct super_block *sb,
900 void *buddy, unsigned first, int len,
901 struct ext4_group_info *grp)
902{
903 struct ext4_sb_info *sbi = EXT4_SB(sb);
904 unsigned short min;
905 unsigned short max;
906 unsigned short chunk;
907 unsigned short border;
908
909 BUG_ON(len >= EXT4_BLOCKS_PER_GROUP(sb));
910
911 border = 2 << sb->s_blocksize_bits;
912
913 while (len > 0) {
914 /* find how many blocks can be covered since this position */
915 max = ffs(first | border) - 1;
916
917 /* find how many blocks of power 2 we need to mark */
918 min = fls(len) - 1;
919
920 if (max < min)
921 min = max;
922 chunk = 1 << min;
923
924 /* mark multiblock chunks only */
925 grp->bb_counters[min]++;
926 if (min > 0)
927 mb_clear_bit(first >> min,
928 buddy + sbi->s_mb_offsets[min]);
929
930 len -= chunk;
931 first += chunk;
932 }
933}
934
935static void ext4_mb_generate_buddy(struct super_block *sb,
936 void *buddy, void *bitmap, ext4_group_t group)
937{
938 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
939 unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
940 unsigned short i = 0;
941 unsigned short first;
942 unsigned short len;
943 unsigned free = 0;
944 unsigned fragments = 0;
945 unsigned long long period = get_cycles();
946
947 /* initialize buddy from bitmap which is aggregation
948 * of on-disk bitmap and preallocations */
949 i = ext4_find_next_zero_bit(bitmap, max, 0);
950 grp->bb_first_free = i;
951 while (i < max) {
952 fragments++;
953 first = i;
954 i = ext4_find_next_bit(bitmap, max, i);
955 len = i - first;
956 free += len;
957 if (len > 1)
958 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
959 else
960 grp->bb_counters[0]++;
961 if (i < max)
962 i = ext4_find_next_zero_bit(bitmap, max, i);
963 }
964 grp->bb_fragments = fragments;
965
966 if (free != grp->bb_free) {
967 printk(KERN_DEBUG
968 "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
969 group, free, grp->bb_free);
970 grp->bb_free = free;
971 }
972
973 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
974
975 period = get_cycles() - period;
976 spin_lock(&EXT4_SB(sb)->s_bal_lock);
977 EXT4_SB(sb)->s_mb_buddies_generated++;
978 EXT4_SB(sb)->s_mb_generation_time += period;
979 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
980}
981
982/* The buddy information is attached the buddy cache inode
983 * for convenience. The information regarding each group
984 * is loaded via ext4_mb_load_buddy. The information involve
985 * block bitmap and buddy information. The information are
986 * stored in the inode as
987 *
988 * { page }
989 * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
990 *
991 *
992 * one block each for bitmap and buddy information.
993 * So for each group we take up 2 blocks. A page can
994 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
995 * So it can have information regarding groups_per_page which
996 * is blocks_per_page/2
997 */
998
999static int ext4_mb_init_cache(struct page *page, char *incore)
1000{
1001 int blocksize;
1002 int blocks_per_page;
1003 int groups_per_page;
1004 int err = 0;
1005 int i;
1006 ext4_group_t first_group;
1007 int first_block;
1008 struct super_block *sb;
1009 struct buffer_head *bhs;
1010 struct buffer_head **bh;
1011 struct inode *inode;
1012 char *data;
1013 char *bitmap;
1014
1015 mb_debug("init page %lu\n", page->index);
1016
1017 inode = page->mapping->host;
1018 sb = inode->i_sb;
1019 blocksize = 1 << inode->i_blkbits;
1020 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
1021
1022 groups_per_page = blocks_per_page >> 1;
1023 if (groups_per_page == 0)
1024 groups_per_page = 1;
1025
1026 /* allocate buffer_heads to read bitmaps */
1027 if (groups_per_page > 1) {
1028 err = -ENOMEM;
1029 i = sizeof(struct buffer_head *) * groups_per_page;
1030 bh = kzalloc(i, GFP_NOFS);
1031 if (bh == NULL)
1032 goto out;
1033 } else
1034 bh = &bhs;
1035
1036 first_group = page->index * blocks_per_page / 2;
1037
1038 /* read all groups the page covers into the cache */
1039 for (i = 0; i < groups_per_page; i++) {
1040 struct ext4_group_desc *desc;
1041
1042 if (first_group + i >= EXT4_SB(sb)->s_groups_count)
1043 break;
1044
1045 err = -EIO;
1046 desc = ext4_get_group_desc(sb, first_group + i, NULL);
1047 if (desc == NULL)
1048 goto out;
1049
1050 err = -ENOMEM;
1051 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
1052 if (bh[i] == NULL)
1053 goto out;
1054
1055 if (bh_uptodate_or_lock(bh[i]))
1056 continue;
1057
1058 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
1059 ext4_init_block_bitmap(sb, bh[i],
1060 first_group + i, desc);
1061 set_buffer_uptodate(bh[i]);
1062 unlock_buffer(bh[i]);
1063 continue;
1064 }
1065 get_bh(bh[i]);
1066 bh[i]->b_end_io = end_buffer_read_sync;
1067 submit_bh(READ, bh[i]);
1068 mb_debug("read bitmap for group %lu\n", first_group + i);
1069 }
1070
1071 /* wait for I/O completion */
1072 for (i = 0; i < groups_per_page && bh[i]; i++)
1073 wait_on_buffer(bh[i]);
1074
1075 err = -EIO;
1076 for (i = 0; i < groups_per_page && bh[i]; i++)
1077 if (!buffer_uptodate(bh[i]))
1078 goto out;
1079
1080 first_block = page->index * blocks_per_page;
1081 for (i = 0; i < blocks_per_page; i++) {
1082 int group;
1083 struct ext4_group_info *grinfo;
1084
1085 group = (first_block + i) >> 1;
1086 if (group >= EXT4_SB(sb)->s_groups_count)
1087 break;
1088
1089 /*
1090 * data carry information regarding this
1091 * particular group in the format specified
1092 * above
1093 *
1094 */
1095 data = page_address(page) + (i * blocksize);
1096 bitmap = bh[group - first_group]->b_data;
1097
1098 /*
1099 * We place the buddy block and bitmap block
1100 * close together
1101 */
1102 if ((first_block + i) & 1) {
1103 /* this is block of buddy */
1104 BUG_ON(incore == NULL);
1105 mb_debug("put buddy for group %u in page %lu/%x\n",
1106 group, page->index, i * blocksize);
1107 memset(data, 0xff, blocksize);
1108 grinfo = ext4_get_group_info(sb, group);
1109 grinfo->bb_fragments = 0;
1110 memset(grinfo->bb_counters, 0,
1111 sizeof(unsigned short)*(sb->s_blocksize_bits+2));
1112 /*
1113 * incore got set to the group block bitmap below
1114 */
1115 ext4_mb_generate_buddy(sb, data, incore, group);
1116 incore = NULL;
1117 } else {
1118 /* this is block of bitmap */
1119 BUG_ON(incore != NULL);
1120 mb_debug("put bitmap for group %u in page %lu/%x\n",
1121 group, page->index, i * blocksize);
1122
1123 /* see comments in ext4_mb_put_pa() */
1124 ext4_lock_group(sb, group);
1125 memcpy(data, bitmap, blocksize);
1126
1127 /* mark all preallocated blks used in in-core bitmap */
1128 ext4_mb_generate_from_pa(sb, data, group);
1129 ext4_unlock_group(sb, group);
1130
1131 /* set incore so that the buddy information can be
1132 * generated using this
1133 */
1134 incore = data;
1135 }
1136 }
1137 SetPageUptodate(page);
1138
1139out:
1140 if (bh) {
1141 for (i = 0; i < groups_per_page && bh[i]; i++)
1142 brelse(bh[i]);
1143 if (bh != &bhs)
1144 kfree(bh);
1145 }
1146 return err;
1147}
1148
1149static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1150 struct ext4_buddy *e4b)
1151{
1152 struct ext4_sb_info *sbi = EXT4_SB(sb);
1153 struct inode *inode = sbi->s_buddy_cache;
1154 int blocks_per_page;
1155 int block;
1156 int pnum;
1157 int poff;
1158 struct page *page;
1159
1160 mb_debug("load group %lu\n", group);
1161
1162 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1163
1164 e4b->bd_blkbits = sb->s_blocksize_bits;
1165 e4b->bd_info = ext4_get_group_info(sb, group);
1166 e4b->bd_sb = sb;
1167 e4b->bd_group = group;
1168 e4b->bd_buddy_page = NULL;
1169 e4b->bd_bitmap_page = NULL;
1170
1171 /*
1172 * the buddy cache inode stores the block bitmap
1173 * and buddy information in consecutive blocks.
1174 * So for each group we need two blocks.
1175 */
1176 block = group * 2;
1177 pnum = block / blocks_per_page;
1178 poff = block % blocks_per_page;
1179
1180 /* we could use find_or_create_page(), but it locks page
1181 * what we'd like to avoid in fast path ... */
1182 page = find_get_page(inode->i_mapping, pnum);
1183 if (page == NULL || !PageUptodate(page)) {
1184 if (page)
1185 page_cache_release(page);
1186 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1187 if (page) {
1188 BUG_ON(page->mapping != inode->i_mapping);
1189 if (!PageUptodate(page)) {
1190 ext4_mb_init_cache(page, NULL);
1191 mb_cmp_bitmaps(e4b, page_address(page) +
1192 (poff * sb->s_blocksize));
1193 }
1194 unlock_page(page);
1195 }
1196 }
1197 if (page == NULL || !PageUptodate(page))
1198 goto err;
1199 e4b->bd_bitmap_page = page;
1200 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1201 mark_page_accessed(page);
1202
1203 block++;
1204 pnum = block / blocks_per_page;
1205 poff = block % blocks_per_page;
1206
1207 page = find_get_page(inode->i_mapping, pnum);
1208 if (page == NULL || !PageUptodate(page)) {
1209 if (page)
1210 page_cache_release(page);
1211 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1212 if (page) {
1213 BUG_ON(page->mapping != inode->i_mapping);
1214 if (!PageUptodate(page))
1215 ext4_mb_init_cache(page, e4b->bd_bitmap);
1216
1217 unlock_page(page);
1218 }
1219 }
1220 if (page == NULL || !PageUptodate(page))
1221 goto err;
1222 e4b->bd_buddy_page = page;
1223 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1224 mark_page_accessed(page);
1225
1226 BUG_ON(e4b->bd_bitmap_page == NULL);
1227 BUG_ON(e4b->bd_buddy_page == NULL);
1228
1229 return 0;
1230
1231err:
1232 if (e4b->bd_bitmap_page)
1233 page_cache_release(e4b->bd_bitmap_page);
1234 if (e4b->bd_buddy_page)
1235 page_cache_release(e4b->bd_buddy_page);
1236 e4b->bd_buddy = NULL;
1237 e4b->bd_bitmap = NULL;
1238 return -EIO;
1239}
1240
1241static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1242{
1243 if (e4b->bd_bitmap_page)
1244 page_cache_release(e4b->bd_bitmap_page);
1245 if (e4b->bd_buddy_page)
1246 page_cache_release(e4b->bd_buddy_page);
1247}
1248
1249
1250static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1251{
1252 int order = 1;
1253 void *bb;
1254
1255 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1256 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1257
1258 bb = EXT4_MB_BUDDY(e4b);
1259 while (order <= e4b->bd_blkbits + 1) {
1260 block = block >> 1;
1261 if (!mb_test_bit(block, bb)) {
1262 /* this block is part of buddy of order 'order' */
1263 return order;
1264 }
1265 bb += 1 << (e4b->bd_blkbits - order);
1266 order++;
1267 }
1268 return 0;
1269}
1270
1271static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
1272{
1273 __u32 *addr;
1274
1275 len = cur + len;
1276 while (cur < len) {
1277 if ((cur & 31) == 0 && (len - cur) >= 32) {
1278 /* fast path: clear whole word at once */
1279 addr = bm + (cur >> 3);
1280 *addr = 0;
1281 cur += 32;
1282 continue;
1283 }
1284 mb_clear_bit_atomic(lock, cur, bm);
1285 cur++;
1286 }
1287}
1288
1289static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
1290{
1291 __u32 *addr;
1292
1293 len = cur + len;
1294 while (cur < len) {
1295 if ((cur & 31) == 0 && (len - cur) >= 32) {
1296 /* fast path: set whole word at once */
1297 addr = bm + (cur >> 3);
1298 *addr = 0xffffffff;
1299 cur += 32;
1300 continue;
1301 }
1302 mb_set_bit_atomic(lock, cur, bm);
1303 cur++;
1304 }
1305}
1306
1307static int mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1308 int first, int count)
1309{
1310 int block = 0;
1311 int max = 0;
1312 int order;
1313 void *buddy;
1314 void *buddy2;
1315 struct super_block *sb = e4b->bd_sb;
1316
1317 BUG_ON(first + count > (sb->s_blocksize << 3));
1318 BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
1319 mb_check_buddy(e4b);
1320 mb_free_blocks_double(inode, e4b, first, count);
1321
1322 e4b->bd_info->bb_free += count;
1323 if (first < e4b->bd_info->bb_first_free)
1324 e4b->bd_info->bb_first_free = first;
1325
1326 /* let's maintain fragments counter */
1327 if (first != 0)
1328 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1329 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1330 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1331 if (block && max)
1332 e4b->bd_info->bb_fragments--;
1333 else if (!block && !max)
1334 e4b->bd_info->bb_fragments++;
1335
1336 /* let's maintain buddy itself */
1337 while (count-- > 0) {
1338 block = first++;
1339 order = 0;
1340
1341 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1342 ext4_fsblk_t blocknr;
1343 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1344 blocknr += block;
1345 blocknr +=
1346 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
1347
1348 ext4_error(sb, __FUNCTION__, "double-free of inode"
1349 " %lu's block %llu(bit %u in group %lu)\n",
1350 inode ? inode->i_ino : 0, blocknr, block,
1351 e4b->bd_group);
1352 }
1353 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1354 e4b->bd_info->bb_counters[order]++;
1355
1356 /* start of the buddy */
1357 buddy = mb_find_buddy(e4b, order, &max);
1358
1359 do {
1360 block &= ~1UL;
1361 if (mb_test_bit(block, buddy) ||
1362 mb_test_bit(block + 1, buddy))
1363 break;
1364
1365 /* both the buddies are free, try to coalesce them */
1366 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1367
1368 if (!buddy2)
1369 break;
1370
1371 if (order > 0) {
1372 /* for special purposes, we don't set
1373 * free bits in bitmap */
1374 mb_set_bit(block, buddy);
1375 mb_set_bit(block + 1, buddy);
1376 }
1377 e4b->bd_info->bb_counters[order]--;
1378 e4b->bd_info->bb_counters[order]--;
1379
1380 block = block >> 1;
1381 order++;
1382 e4b->bd_info->bb_counters[order]++;
1383
1384 mb_clear_bit(block, buddy2);
1385 buddy = buddy2;
1386 } while (1);
1387 }
1388 mb_check_buddy(e4b);
1389
1390 return 0;
1391}
1392
1393static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1394 int needed, struct ext4_free_extent *ex)
1395{
1396 int next = block;
1397 int max;
1398 int ord;
1399 void *buddy;
1400
1401 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1402 BUG_ON(ex == NULL);
1403
1404 buddy = mb_find_buddy(e4b, order, &max);
1405 BUG_ON(buddy == NULL);
1406 BUG_ON(block >= max);
1407 if (mb_test_bit(block, buddy)) {
1408 ex->fe_len = 0;
1409 ex->fe_start = 0;
1410 ex->fe_group = 0;
1411 return 0;
1412 }
1413
1414 /* FIXME dorp order completely ? */
1415 if (likely(order == 0)) {
1416 /* find actual order */
1417 order = mb_find_order_for_block(e4b, block);
1418 block = block >> order;
1419 }
1420
1421 ex->fe_len = 1 << order;
1422 ex->fe_start = block << order;
1423 ex->fe_group = e4b->bd_group;
1424
1425 /* calc difference from given start */
1426 next = next - ex->fe_start;
1427 ex->fe_len -= next;
1428 ex->fe_start += next;
1429
1430 while (needed > ex->fe_len &&
1431 (buddy = mb_find_buddy(e4b, order, &max))) {
1432
1433 if (block + 1 >= max)
1434 break;
1435
1436 next = (block + 1) * (1 << order);
1437 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1438 break;
1439
1440 ord = mb_find_order_for_block(e4b, next);
1441
1442 order = ord;
1443 block = next >> order;
1444 ex->fe_len += 1 << order;
1445 }
1446
1447 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1448 return ex->fe_len;
1449}
1450
1451static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1452{
1453 int ord;
1454 int mlen = 0;
1455 int max = 0;
1456 int cur;
1457 int start = ex->fe_start;
1458 int len = ex->fe_len;
1459 unsigned ret = 0;
1460 int len0 = len;
1461 void *buddy;
1462
1463 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1464 BUG_ON(e4b->bd_group != ex->fe_group);
1465 BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
1466 mb_check_buddy(e4b);
1467 mb_mark_used_double(e4b, start, len);
1468
1469 e4b->bd_info->bb_free -= len;
1470 if (e4b->bd_info->bb_first_free == start)
1471 e4b->bd_info->bb_first_free += len;
1472
1473 /* let's maintain fragments counter */
1474 if (start != 0)
1475 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1476 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1477 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1478 if (mlen && max)
1479 e4b->bd_info->bb_fragments++;
1480 else if (!mlen && !max)
1481 e4b->bd_info->bb_fragments--;
1482
1483 /* let's maintain buddy itself */
1484 while (len) {
1485 ord = mb_find_order_for_block(e4b, start);
1486
1487 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1488 /* the whole chunk may be allocated at once! */
1489 mlen = 1 << ord;
1490 buddy = mb_find_buddy(e4b, ord, &max);
1491 BUG_ON((start >> ord) >= max);
1492 mb_set_bit(start >> ord, buddy);
1493 e4b->bd_info->bb_counters[ord]--;
1494 start += mlen;
1495 len -= mlen;
1496 BUG_ON(len < 0);
1497 continue;
1498 }
1499
1500 /* store for history */
1501 if (ret == 0)
1502 ret = len | (ord << 16);
1503
1504 /* we have to split large buddy */
1505 BUG_ON(ord <= 0);
1506 buddy = mb_find_buddy(e4b, ord, &max);
1507 mb_set_bit(start >> ord, buddy);
1508 e4b->bd_info->bb_counters[ord]--;
1509
1510 ord--;
1511 cur = (start >> ord) & ~1U;
1512 buddy = mb_find_buddy(e4b, ord, &max);
1513 mb_clear_bit(cur, buddy);
1514 mb_clear_bit(cur + 1, buddy);
1515 e4b->bd_info->bb_counters[ord]++;
1516 e4b->bd_info->bb_counters[ord]++;
1517 }
1518
1519 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
1520 EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1521 mb_check_buddy(e4b);
1522
1523 return ret;
1524}
1525
1526/*
1527 * Must be called under group lock!
1528 */
1529static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1530 struct ext4_buddy *e4b)
1531{
1532 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1533 int ret;
1534
1535 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1536 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1537
1538 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1539 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1540 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1541
1542 /* preallocation can change ac_b_ex, thus we store actually
1543 * allocated blocks for history */
1544 ac->ac_f_ex = ac->ac_b_ex;
1545
1546 ac->ac_status = AC_STATUS_FOUND;
1547 ac->ac_tail = ret & 0xffff;
1548 ac->ac_buddy = ret >> 16;
1549
1550 /* XXXXXXX: SUCH A HORRIBLE **CK */
1551 /*FIXME!! Why ? */
1552 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1553 get_page(ac->ac_bitmap_page);
1554 ac->ac_buddy_page = e4b->bd_buddy_page;
1555 get_page(ac->ac_buddy_page);
1556
1557 /* store last allocated for subsequent stream allocation */
1558 if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
1559 spin_lock(&sbi->s_md_lock);
1560 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1561 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1562 spin_unlock(&sbi->s_md_lock);
1563 }
1564}
1565
1566/*
1567 * regular allocator, for general purposes allocation
1568 */
1569
1570static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1571 struct ext4_buddy *e4b,
1572 int finish_group)
1573{
1574 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1575 struct ext4_free_extent *bex = &ac->ac_b_ex;
1576 struct ext4_free_extent *gex = &ac->ac_g_ex;
1577 struct ext4_free_extent ex;
1578 int max;
1579
1580 /*
1581 * We don't want to scan for a whole year
1582 */
1583 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1584 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1585 ac->ac_status = AC_STATUS_BREAK;
1586 return;
1587 }
1588
1589 /*
1590 * Haven't found good chunk so far, let's continue
1591 */
1592 if (bex->fe_len < gex->fe_len)
1593 return;
1594
1595 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1596 && bex->fe_group == e4b->bd_group) {
1597 /* recheck chunk's availability - we don't know
1598 * when it was found (within this lock-unlock
1599 * period or not) */
1600 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1601 if (max >= gex->fe_len) {
1602 ext4_mb_use_best_found(ac, e4b);
1603 return;
1604 }
1605 }
1606}
1607
1608/*
1609 * The routine checks whether found extent is good enough. If it is,
1610 * then the extent gets marked used and flag is set to the context
1611 * to stop scanning. Otherwise, the extent is compared with the
1612 * previous found extent and if new one is better, then it's stored
1613 * in the context. Later, the best found extent will be used, if
1614 * mballoc can't find good enough extent.
1615 *
1616 * FIXME: real allocation policy is to be designed yet!
1617 */
1618static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1619 struct ext4_free_extent *ex,
1620 struct ext4_buddy *e4b)
1621{
1622 struct ext4_free_extent *bex = &ac->ac_b_ex;
1623 struct ext4_free_extent *gex = &ac->ac_g_ex;
1624
1625 BUG_ON(ex->fe_len <= 0);
1626 BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1627 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1628 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1629
1630 ac->ac_found++;
1631
1632 /*
1633 * The special case - take what you catch first
1634 */
1635 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1636 *bex = *ex;
1637 ext4_mb_use_best_found(ac, e4b);
1638 return;
1639 }
1640
1641 /*
1642 * Let's check whether the chuck is good enough
1643 */
1644 if (ex->fe_len == gex->fe_len) {
1645 *bex = *ex;
1646 ext4_mb_use_best_found(ac, e4b);
1647 return;
1648 }
1649
1650 /*
1651 * If this is first found extent, just store it in the context
1652 */
1653 if (bex->fe_len == 0) {
1654 *bex = *ex;
1655 return;
1656 }
1657
1658 /*
1659 * If new found extent is better, store it in the context
1660 */
1661 if (bex->fe_len < gex->fe_len) {
1662 /* if the request isn't satisfied, any found extent
1663 * larger than previous best one is better */
1664 if (ex->fe_len > bex->fe_len)
1665 *bex = *ex;
1666 } else if (ex->fe_len > gex->fe_len) {
1667 /* if the request is satisfied, then we try to find
1668 * an extent that still satisfy the request, but is
1669 * smaller than previous one */
1670 if (ex->fe_len < bex->fe_len)
1671 *bex = *ex;
1672 }
1673
1674 ext4_mb_check_limits(ac, e4b, 0);
1675}
1676
1677static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1678 struct ext4_buddy *e4b)
1679{
1680 struct ext4_free_extent ex = ac->ac_b_ex;
1681 ext4_group_t group = ex.fe_group;
1682 int max;
1683 int err;
1684
1685 BUG_ON(ex.fe_len <= 0);
1686 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1687 if (err)
1688 return err;
1689
1690 ext4_lock_group(ac->ac_sb, group);
1691 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1692
1693 if (max > 0) {
1694 ac->ac_b_ex = ex;
1695 ext4_mb_use_best_found(ac, e4b);
1696 }
1697
1698 ext4_unlock_group(ac->ac_sb, group);
1699 ext4_mb_release_desc(e4b);
1700
1701 return 0;
1702}
1703
1704static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1705 struct ext4_buddy *e4b)
1706{
1707 ext4_group_t group = ac->ac_g_ex.fe_group;
1708 int max;
1709 int err;
1710 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1711 struct ext4_super_block *es = sbi->s_es;
1712 struct ext4_free_extent ex;
1713
1714 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1715 return 0;
1716
1717 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1718 if (err)
1719 return err;
1720
1721 ext4_lock_group(ac->ac_sb, group);
1722 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1723 ac->ac_g_ex.fe_len, &ex);
1724
1725 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1726 ext4_fsblk_t start;
1727
1728 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1729 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1730 /* use do_div to get remainder (would be 64-bit modulo) */
1731 if (do_div(start, sbi->s_stripe) == 0) {
1732 ac->ac_found++;
1733 ac->ac_b_ex = ex;
1734 ext4_mb_use_best_found(ac, e4b);
1735 }
1736 } else if (max >= ac->ac_g_ex.fe_len) {
1737 BUG_ON(ex.fe_len <= 0);
1738 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1739 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1740 ac->ac_found++;
1741 ac->ac_b_ex = ex;
1742 ext4_mb_use_best_found(ac, e4b);
1743 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1744 /* Sometimes, caller may want to merge even small
1745 * number of blocks to an existing extent */
1746 BUG_ON(ex.fe_len <= 0);
1747 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1748 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1749 ac->ac_found++;
1750 ac->ac_b_ex = ex;
1751 ext4_mb_use_best_found(ac, e4b);
1752 }
1753 ext4_unlock_group(ac->ac_sb, group);
1754 ext4_mb_release_desc(e4b);
1755
1756 return 0;
1757}
1758
1759/*
1760 * The routine scans buddy structures (not bitmap!) from given order
1761 * to max order and tries to find big enough chunk to satisfy the req
1762 */
1763static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1764 struct ext4_buddy *e4b)
1765{
1766 struct super_block *sb = ac->ac_sb;
1767 struct ext4_group_info *grp = e4b->bd_info;
1768 void *buddy;
1769 int i;
1770 int k;
1771 int max;
1772
1773 BUG_ON(ac->ac_2order <= 0);
1774 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1775 if (grp->bb_counters[i] == 0)
1776 continue;
1777
1778 buddy = mb_find_buddy(e4b, i, &max);
1779 BUG_ON(buddy == NULL);
1780
1781 k = ext4_find_next_zero_bit(buddy, max, 0);
1782 BUG_ON(k >= max);
1783
1784 ac->ac_found++;
1785
1786 ac->ac_b_ex.fe_len = 1 << i;
1787 ac->ac_b_ex.fe_start = k << i;
1788 ac->ac_b_ex.fe_group = e4b->bd_group;
1789
1790 ext4_mb_use_best_found(ac, e4b);
1791
1792 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1793
1794 if (EXT4_SB(sb)->s_mb_stats)
1795 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1796
1797 break;
1798 }
1799}
1800
1801/*
1802 * The routine scans the group and measures all found extents.
1803 * In order to optimize scanning, caller must pass number of
1804 * free blocks in the group, so the routine can know upper limit.
1805 */
1806static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1807 struct ext4_buddy *e4b)
1808{
1809 struct super_block *sb = ac->ac_sb;
1810 void *bitmap = EXT4_MB_BITMAP(e4b);
1811 struct ext4_free_extent ex;
1812 int i;
1813 int free;
1814
1815 free = e4b->bd_info->bb_free;
1816 BUG_ON(free <= 0);
1817
1818 i = e4b->bd_info->bb_first_free;
1819
1820 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1821 i = ext4_find_next_zero_bit(bitmap,
1822 EXT4_BLOCKS_PER_GROUP(sb), i);
1823 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1824 BUG_ON(free != 0);
1825 break;
1826 }
1827
1828 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1829 BUG_ON(ex.fe_len <= 0);
1830 BUG_ON(free < ex.fe_len);
1831
1832 ext4_mb_measure_extent(ac, &ex, e4b);
1833
1834 i += ex.fe_len;
1835 free -= ex.fe_len;
1836 }
1837
1838 ext4_mb_check_limits(ac, e4b, 1);
1839}
1840
1841/*
1842 * This is a special case for storages like raid5
1843 * we try to find stripe-aligned chunks for stripe-size requests
1844 * XXX should do so at least for multiples of stripe size as well
1845 */
1846static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1847 struct ext4_buddy *e4b)
1848{
1849 struct super_block *sb = ac->ac_sb;
1850 struct ext4_sb_info *sbi = EXT4_SB(sb);
1851 void *bitmap = EXT4_MB_BITMAP(e4b);
1852 struct ext4_free_extent ex;
1853 ext4_fsblk_t first_group_block;
1854 ext4_fsblk_t a;
1855 ext4_grpblk_t i;
1856 int max;
1857
1858 BUG_ON(sbi->s_stripe == 0);
1859
1860 /* find first stripe-aligned block in group */
1861 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1862 + le32_to_cpu(sbi->s_es->s_first_data_block);
1863 a = first_group_block + sbi->s_stripe - 1;
1864 do_div(a, sbi->s_stripe);
1865 i = (a * sbi->s_stripe) - first_group_block;
1866
1867 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1868 if (!mb_test_bit(i, bitmap)) {
1869 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1870 if (max >= sbi->s_stripe) {
1871 ac->ac_found++;
1872 ac->ac_b_ex = ex;
1873 ext4_mb_use_best_found(ac, e4b);
1874 break;
1875 }
1876 }
1877 i += sbi->s_stripe;
1878 }
1879}
1880
1881static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1882 ext4_group_t group, int cr)
1883{
1884 unsigned free, fragments;
1885 unsigned i, bits;
1886 struct ext4_group_desc *desc;
1887 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1888
1889 BUG_ON(cr < 0 || cr >= 4);
1890 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1891
1892 free = grp->bb_free;
1893 fragments = grp->bb_fragments;
1894 if (free == 0)
1895 return 0;
1896 if (fragments == 0)
1897 return 0;
1898
1899 switch (cr) {
1900 case 0:
1901 BUG_ON(ac->ac_2order == 0);
1902 /* If this group is uninitialized, skip it initially */
1903 desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
1904 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
1905 return 0;
1906
1907 bits = ac->ac_sb->s_blocksize_bits + 1;
1908 for (i = ac->ac_2order; i <= bits; i++)
1909 if (grp->bb_counters[i] > 0)
1910 return 1;
1911 break;
1912 case 1:
1913 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1914 return 1;
1915 break;
1916 case 2:
1917 if (free >= ac->ac_g_ex.fe_len)
1918 return 1;
1919 break;
1920 case 3:
1921 return 1;
1922 default:
1923 BUG();
1924 }
1925
1926 return 0;
1927}
1928
1929static int ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1930{
1931 ext4_group_t group;
1932 ext4_group_t i;
1933 int cr;
1934 int err = 0;
1935 int bsbits;
1936 struct ext4_sb_info *sbi;
1937 struct super_block *sb;
1938 struct ext4_buddy e4b;
1939 loff_t size, isize;
1940
1941 sb = ac->ac_sb;
1942 sbi = EXT4_SB(sb);
1943 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1944
1945 /* first, try the goal */
1946 err = ext4_mb_find_by_goal(ac, &e4b);
1947 if (err || ac->ac_status == AC_STATUS_FOUND)
1948 goto out;
1949
1950 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1951 goto out;
1952
1953 /*
1954 * ac->ac2_order is set only if the fe_len is a power of 2
1955 * if ac2_order is set we also set criteria to 0 so that we
1956 * try exact allocation using buddy.
1957 */
1958 i = fls(ac->ac_g_ex.fe_len);
1959 ac->ac_2order = 0;
1960 /*
1961 * We search using buddy data only if the order of the request
1962 * is greater than equal to the sbi_s_mb_order2_reqs
1963 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1964 */
1965 if (i >= sbi->s_mb_order2_reqs) {
1966 /*
1967 * This should tell if fe_len is exactly power of 2
1968 */
1969 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1970 ac->ac_2order = i - 1;
1971 }
1972
1973 bsbits = ac->ac_sb->s_blocksize_bits;
1974 /* if stream allocation is enabled, use global goal */
1975 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
1976 isize = i_size_read(ac->ac_inode) >> bsbits;
1977 if (size < isize)
1978 size = isize;
1979
1980 if (size < sbi->s_mb_stream_request &&
1981 (ac->ac_flags & EXT4_MB_HINT_DATA)) {
1982 /* TBD: may be hot point */
1983 spin_lock(&sbi->s_md_lock);
1984 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1985 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1986 spin_unlock(&sbi->s_md_lock);
1987 }
1988
1989 /* searching for the right group start from the goal value specified */
1990 group = ac->ac_g_ex.fe_group;
1991
1992 /* Let's just scan groups to find more-less suitable blocks */
1993 cr = ac->ac_2order ? 0 : 1;
1994 /*
1995 * cr == 0 try to get exact allocation,
1996 * cr == 3 try to get anything
1997 */
1998repeat:
1999 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2000 ac->ac_criteria = cr;
2001 for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
2002 struct ext4_group_info *grp;
2003 struct ext4_group_desc *desc;
2004
2005 if (group == EXT4_SB(sb)->s_groups_count)
2006 group = 0;
2007
2008 /* quick check to skip empty groups */
2009 grp = ext4_get_group_info(ac->ac_sb, group);
2010 if (grp->bb_free == 0)
2011 continue;
2012
2013 /*
2014 * if the group is already init we check whether it is
2015 * a good group and if not we don't load the buddy
2016 */
2017 if (EXT4_MB_GRP_NEED_INIT(grp)) {
2018 /*
2019 * we need full data about the group
2020 * to make a good selection
2021 */
2022 err = ext4_mb_load_buddy(sb, group, &e4b);
2023 if (err)
2024 goto out;
2025 ext4_mb_release_desc(&e4b);
2026 }
2027
2028 /*
2029 * If the particular group doesn't satisfy our
2030 * criteria we continue with the next group
2031 */
2032 if (!ext4_mb_good_group(ac, group, cr))
2033 continue;
2034
2035 err = ext4_mb_load_buddy(sb, group, &e4b);
2036 if (err)
2037 goto out;
2038
2039 ext4_lock_group(sb, group);
2040 if (!ext4_mb_good_group(ac, group, cr)) {
2041 /* someone did allocation from this group */
2042 ext4_unlock_group(sb, group);
2043 ext4_mb_release_desc(&e4b);
2044 continue;
2045 }
2046
2047 ac->ac_groups_scanned++;
2048 desc = ext4_get_group_desc(sb, group, NULL);
2049 if (cr == 0 || (desc->bg_flags &
2050 cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
2051 ac->ac_2order != 0))
2052 ext4_mb_simple_scan_group(ac, &e4b);
2053 else if (cr == 1 &&
2054 ac->ac_g_ex.fe_len == sbi->s_stripe)
2055 ext4_mb_scan_aligned(ac, &e4b);
2056 else
2057 ext4_mb_complex_scan_group(ac, &e4b);
2058
2059 ext4_unlock_group(sb, group);
2060 ext4_mb_release_desc(&e4b);
2061
2062 if (ac->ac_status != AC_STATUS_CONTINUE)
2063 break;
2064 }
2065 }
2066
2067 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2068 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2069 /*
2070 * We've been searching too long. Let's try to allocate
2071 * the best chunk we've found so far
2072 */
2073
2074 ext4_mb_try_best_found(ac, &e4b);
2075 if (ac->ac_status != AC_STATUS_FOUND) {
2076 /*
2077 * Someone more lucky has already allocated it.
2078 * The only thing we can do is just take first
2079 * found block(s)
2080 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2081 */
2082 ac->ac_b_ex.fe_group = 0;
2083 ac->ac_b_ex.fe_start = 0;
2084 ac->ac_b_ex.fe_len = 0;
2085 ac->ac_status = AC_STATUS_CONTINUE;
2086 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2087 cr = 3;
2088 atomic_inc(&sbi->s_mb_lost_chunks);
2089 goto repeat;
2090 }
2091 }
2092out:
2093 return err;
2094}
2095
2096#ifdef EXT4_MB_HISTORY
2097struct ext4_mb_proc_session {
2098 struct ext4_mb_history *history;
2099 struct super_block *sb;
2100 int start;
2101 int max;
2102};
2103
2104static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
2105 struct ext4_mb_history *hs,
2106 int first)
2107{
2108 if (hs == s->history + s->max)
2109 hs = s->history;
2110 if (!first && hs == s->history + s->start)
2111 return NULL;
2112 while (hs->orig.fe_len == 0) {
2113 hs++;
2114 if (hs == s->history + s->max)
2115 hs = s->history;
2116 if (hs == s->history + s->start)
2117 return NULL;
2118 }
2119 return hs;
2120}
2121
2122static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
2123{
2124 struct ext4_mb_proc_session *s = seq->private;
2125 struct ext4_mb_history *hs;
2126 int l = *pos;
2127
2128 if (l == 0)
2129 return SEQ_START_TOKEN;
2130 hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2131 if (!hs)
2132 return NULL;
2133 while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
2134 return hs;
2135}
2136
2137static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
2138 loff_t *pos)
2139{
2140 struct ext4_mb_proc_session *s = seq->private;
2141 struct ext4_mb_history *hs = v;
2142
2143 ++*pos;
2144 if (v == SEQ_START_TOKEN)
2145 return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
2146 else
2147 return ext4_mb_history_skip_empty(s, ++hs, 0);
2148}
2149
2150static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
2151{
2152 char buf[25], buf2[25], buf3[25], *fmt;
2153 struct ext4_mb_history *hs = v;
2154
2155 if (v == SEQ_START_TOKEN) {
2156 seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
2157 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2158 "pid", "inode", "original", "goal", "result", "found",
2159 "grps", "cr", "flags", "merge", "tail", "broken");
2160 return 0;
2161 }
2162
2163 if (hs->op == EXT4_MB_HISTORY_ALLOC) {
2164 fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2165 "%-5u %-5s %-5u %-6u\n";
2166 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
2167 hs->result.fe_start, hs->result.fe_len,
2168 hs->result.fe_logical);
2169 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
2170 hs->orig.fe_start, hs->orig.fe_len,
2171 hs->orig.fe_logical);
2172 sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
2173 hs->goal.fe_start, hs->goal.fe_len,
2174 hs->goal.fe_logical);
2175 seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
2176 hs->found, hs->groups, hs->cr, hs->flags,
2177 hs->merged ? "M" : "", hs->tail,
2178 hs->buddy ? 1 << hs->buddy : 0);
2179 } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
2180 fmt = "%-5u %-8u %-23s %-23s %-23s\n";
2181 sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
2182 hs->result.fe_start, hs->result.fe_len,
2183 hs->result.fe_logical);
2184 sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
2185 hs->orig.fe_start, hs->orig.fe_len,
2186 hs->orig.fe_logical);
2187 seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
2188 } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
2189 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
2190 hs->result.fe_start, hs->result.fe_len);
2191 seq_printf(seq, "%-5u %-8u %-23s discard\n",
2192 hs->pid, hs->ino, buf2);
2193 } else if (hs->op == EXT4_MB_HISTORY_FREE) {
2194 sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
2195 hs->result.fe_start, hs->result.fe_len);
2196 seq_printf(seq, "%-5u %-8u %-23s free\n",
2197 hs->pid, hs->ino, buf2);
2198 }
2199 return 0;
2200}
2201
2202static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
2203{
2204}
2205
2206static struct seq_operations ext4_mb_seq_history_ops = {
2207 .start = ext4_mb_seq_history_start,
2208 .next = ext4_mb_seq_history_next,
2209 .stop = ext4_mb_seq_history_stop,
2210 .show = ext4_mb_seq_history_show,
2211};
2212
2213static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
2214{
2215 struct super_block *sb = PDE(inode)->data;
2216 struct ext4_sb_info *sbi = EXT4_SB(sb);
2217 struct ext4_mb_proc_session *s;
2218 int rc;
2219 int size;
2220
2221 s = kmalloc(sizeof(*s), GFP_KERNEL);
2222 if (s == NULL)
2223 return -ENOMEM;
2224 s->sb = sb;
2225 size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
2226 s->history = kmalloc(size, GFP_KERNEL);
2227 if (s->history == NULL) {
2228 kfree(s);
2229 return -ENOMEM;
2230 }
2231
2232 spin_lock(&sbi->s_mb_history_lock);
2233 memcpy(s->history, sbi->s_mb_history, size);
2234 s->max = sbi->s_mb_history_max;
2235 s->start = sbi->s_mb_history_cur % s->max;
2236 spin_unlock(&sbi->s_mb_history_lock);
2237
2238 rc = seq_open(file, &ext4_mb_seq_history_ops);
2239 if (rc == 0) {
2240 struct seq_file *m = (struct seq_file *)file->private_data;
2241 m->private = s;
2242 } else {
2243 kfree(s->history);
2244 kfree(s);
2245 }
2246 return rc;
2247
2248}
2249
2250static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
2251{
2252 struct seq_file *seq = (struct seq_file *)file->private_data;
2253 struct ext4_mb_proc_session *s = seq->private;
2254 kfree(s->history);
2255 kfree(s);
2256 return seq_release(inode, file);
2257}
2258
2259static ssize_t ext4_mb_seq_history_write(struct file *file,
2260 const char __user *buffer,
2261 size_t count, loff_t *ppos)
2262{
2263 struct seq_file *seq = (struct seq_file *)file->private_data;
2264 struct ext4_mb_proc_session *s = seq->private;
2265 struct super_block *sb = s->sb;
2266 char str[32];
2267 int value;
2268
2269 if (count >= sizeof(str)) {
2270 printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
2271 "mb_history", (int)sizeof(str));
2272 return -EOVERFLOW;
2273 }
2274
2275 if (copy_from_user(str, buffer, count))
2276 return -EFAULT;
2277
2278 value = simple_strtol(str, NULL, 0);
2279 if (value < 0)
2280 return -ERANGE;
2281 EXT4_SB(sb)->s_mb_history_filter = value;
2282
2283 return count;
2284}
2285
2286static struct file_operations ext4_mb_seq_history_fops = {
2287 .owner = THIS_MODULE,
2288 .open = ext4_mb_seq_history_open,
2289 .read = seq_read,
2290 .write = ext4_mb_seq_history_write,
2291 .llseek = seq_lseek,
2292 .release = ext4_mb_seq_history_release,
2293};
2294
2295static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2296{
2297 struct super_block *sb = seq->private;
2298 struct ext4_sb_info *sbi = EXT4_SB(sb);
2299 ext4_group_t group;
2300
2301 if (*pos < 0 || *pos >= sbi->s_groups_count)
2302 return NULL;
2303
2304 group = *pos + 1;
2305 return (void *) group;
2306}
2307
2308static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2309{
2310 struct super_block *sb = seq->private;
2311 struct ext4_sb_info *sbi = EXT4_SB(sb);
2312 ext4_group_t group;
2313
2314 ++*pos;
2315 if (*pos < 0 || *pos >= sbi->s_groups_count)
2316 return NULL;
2317 group = *pos + 1;
2318 return (void *) group;;
2319}
2320
2321static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2322{
2323 struct super_block *sb = seq->private;
2324 long group = (long) v;
2325 int i;
2326 int err;
2327 struct ext4_buddy e4b;
2328 struct sg {
2329 struct ext4_group_info info;
2330 unsigned short counters[16];
2331 } sg;
2332
2333 group--;
2334 if (group == 0)
2335 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2336 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2337 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2338 "group", "free", "frags", "first",
2339 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2340 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2341
2342 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2343 sizeof(struct ext4_group_info);
2344 err = ext4_mb_load_buddy(sb, group, &e4b);
2345 if (err) {
2346 seq_printf(seq, "#%-5lu: I/O error\n", group);
2347 return 0;
2348 }
2349 ext4_lock_group(sb, group);
2350 memcpy(&sg, ext4_get_group_info(sb, group), i);
2351 ext4_unlock_group(sb, group);
2352 ext4_mb_release_desc(&e4b);
2353
2354 seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
2355 sg.info.bb_fragments, sg.info.bb_first_free);
2356 for (i = 0; i <= 13; i++)
2357 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2358 sg.info.bb_counters[i] : 0);
2359 seq_printf(seq, " ]\n");
2360
2361 return 0;
2362}
2363
2364static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2365{
2366}
2367
2368static struct seq_operations ext4_mb_seq_groups_ops = {
2369 .start = ext4_mb_seq_groups_start,
2370 .next = ext4_mb_seq_groups_next,
2371 .stop = ext4_mb_seq_groups_stop,
2372 .show = ext4_mb_seq_groups_show,
2373};
2374
2375static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2376{
2377 struct super_block *sb = PDE(inode)->data;
2378 int rc;
2379
2380 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2381 if (rc == 0) {
2382 struct seq_file *m = (struct seq_file *)file->private_data;
2383 m->private = sb;
2384 }
2385 return rc;
2386
2387}
2388
2389static struct file_operations ext4_mb_seq_groups_fops = {
2390 .owner = THIS_MODULE,
2391 .open = ext4_mb_seq_groups_open,
2392 .read = seq_read,
2393 .llseek = seq_lseek,
2394 .release = seq_release,
2395};
2396
2397static void ext4_mb_history_release(struct super_block *sb)
2398{
2399 struct ext4_sb_info *sbi = EXT4_SB(sb);
2400
2401 remove_proc_entry("mb_groups", sbi->s_mb_proc);
2402 remove_proc_entry("mb_history", sbi->s_mb_proc);
2403
2404 kfree(sbi->s_mb_history);
2405}
2406
2407static void ext4_mb_history_init(struct super_block *sb)
2408{
2409 struct ext4_sb_info *sbi = EXT4_SB(sb);
2410 int i;
2411
2412 if (sbi->s_mb_proc != NULL) {
2413 struct proc_dir_entry *p;
2414 p = create_proc_entry("mb_history", S_IRUGO, sbi->s_mb_proc);
2415 if (p) {
2416 p->proc_fops = &ext4_mb_seq_history_fops;
2417 p->data = sb;
2418 }
2419 p = create_proc_entry("mb_groups", S_IRUGO, sbi->s_mb_proc);
2420 if (p) {
2421 p->proc_fops = &ext4_mb_seq_groups_fops;
2422 p->data = sb;
2423 }
2424 }
2425
2426 sbi->s_mb_history_max = 1000;
2427 sbi->s_mb_history_cur = 0;
2428 spin_lock_init(&sbi->s_mb_history_lock);
2429 i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
2430 sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
2431 if (likely(sbi->s_mb_history != NULL))
2432 memset(sbi->s_mb_history, 0, i);
2433 /* if we can't allocate history, then we simple won't use it */
2434}
2435
2436static void ext4_mb_store_history(struct ext4_allocation_context *ac)
2437{
2438 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2439 struct ext4_mb_history h;
2440
2441 if (unlikely(sbi->s_mb_history == NULL))
2442 return;
2443
2444 if (!(ac->ac_op & sbi->s_mb_history_filter))
2445 return;
2446
2447 h.op = ac->ac_op;
2448 h.pid = current->pid;
2449 h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
2450 h.orig = ac->ac_o_ex;
2451 h.result = ac->ac_b_ex;
2452 h.flags = ac->ac_flags;
2453 h.found = ac->ac_found;
2454 h.groups = ac->ac_groups_scanned;
2455 h.cr = ac->ac_criteria;
2456 h.tail = ac->ac_tail;
2457 h.buddy = ac->ac_buddy;
2458 h.merged = 0;
2459 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
2460 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2461 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2462 h.merged = 1;
2463 h.goal = ac->ac_g_ex;
2464 h.result = ac->ac_f_ex;
2465 }
2466
2467 spin_lock(&sbi->s_mb_history_lock);
2468 memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
2469 if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
2470 sbi->s_mb_history_cur = 0;
2471 spin_unlock(&sbi->s_mb_history_lock);
2472}
2473
2474#else
2475#define ext4_mb_history_release(sb)
2476#define ext4_mb_history_init(sb)
2477#endif
2478
2479static int ext4_mb_init_backend(struct super_block *sb)
2480{
2481 ext4_group_t i;
2482 int j, len, metalen;
2483 struct ext4_sb_info *sbi = EXT4_SB(sb);
2484 int num_meta_group_infos =
2485 (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2486 EXT4_DESC_PER_BLOCK_BITS(sb);
2487 struct ext4_group_info **meta_group_info;
2488
2489 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2490 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2491 * So a two level scheme suffices for now. */
2492 sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
2493 num_meta_group_infos, GFP_KERNEL);
2494 if (sbi->s_group_info == NULL) {
2495 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2496 return -ENOMEM;
2497 }
2498 sbi->s_buddy_cache = new_inode(sb);
2499 if (sbi->s_buddy_cache == NULL) {
2500 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2501 goto err_freesgi;
2502 }
2503 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2504
2505 metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
2506 for (i = 0; i < num_meta_group_infos; i++) {
2507 if ((i + 1) == num_meta_group_infos)
2508 metalen = sizeof(*meta_group_info) *
2509 (sbi->s_groups_count -
2510 (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
2511 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2512 if (meta_group_info == NULL) {
2513 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2514 "buddy group\n");
2515 goto err_freemeta;
2516 }
2517 sbi->s_group_info[i] = meta_group_info;
2518 }
2519
2520 /*
2521 * calculate needed size. if change bb_counters size,
2522 * don't forget about ext4_mb_generate_buddy()
2523 */
2524 len = sizeof(struct ext4_group_info);
2525 len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
2526 for (i = 0; i < sbi->s_groups_count; i++) {
2527 struct ext4_group_desc *desc;
2528
2529 meta_group_info =
2530 sbi->s_group_info[i >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2531 j = i & (EXT4_DESC_PER_BLOCK(sb) - 1);
2532
2533 meta_group_info[j] = kzalloc(len, GFP_KERNEL);
2534 if (meta_group_info[j] == NULL) {
2535 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2536 i--;
2537 goto err_freebuddy;
2538 }
2539 desc = ext4_get_group_desc(sb, i, NULL);
2540 if (desc == NULL) {
2541 printk(KERN_ERR
2542 "EXT4-fs: can't read descriptor %lu\n", i);
2543 goto err_freebuddy;
2544 }
2545 memset(meta_group_info[j], 0, len);
2546 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2547 &(meta_group_info[j]->bb_state));
2548
2549 /*
2550 * initialize bb_free to be able to skip
2551 * empty groups without initialization
2552 */
2553 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2554 meta_group_info[j]->bb_free =
2555 ext4_free_blocks_after_init(sb, i, desc);
2556 } else {
2557 meta_group_info[j]->bb_free =
2558 le16_to_cpu(desc->bg_free_blocks_count);
2559 }
2560
2561 INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
2562
2563#ifdef DOUBLE_CHECK
2564 {
2565 struct buffer_head *bh;
2566 meta_group_info[j]->bb_bitmap =
2567 kmalloc(sb->s_blocksize, GFP_KERNEL);
2568 BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
2569 bh = read_block_bitmap(sb, i);
2570 BUG_ON(bh == NULL);
2571 memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
2572 sb->s_blocksize);
2573 put_bh(bh);
2574 }
2575#endif
2576
2577 }
2578
2579 return 0;
2580
2581err_freebuddy:
2582 while (i >= 0) {
2583 kfree(ext4_get_group_info(sb, i));
2584 i--;
2585 }
2586 i = num_meta_group_infos;
2587err_freemeta:
2588 while (--i >= 0)
2589 kfree(sbi->s_group_info[i]);
2590 iput(sbi->s_buddy_cache);
2591err_freesgi:
2592 kfree(sbi->s_group_info);
2593 return -ENOMEM;
2594}
2595
2596int ext4_mb_init(struct super_block *sb, int needs_recovery)
2597{
2598 struct ext4_sb_info *sbi = EXT4_SB(sb);
2599 unsigned i;
2600 unsigned offset;
2601 unsigned max;
2602
2603 if (!test_opt(sb, MBALLOC))
2604 return 0;
2605
2606 i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
2607
2608 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2609 if (sbi->s_mb_offsets == NULL) {
2610 clear_opt(sbi->s_mount_opt, MBALLOC);
2611 return -ENOMEM;
2612 }
2613 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2614 if (sbi->s_mb_maxs == NULL) {
2615 clear_opt(sbi->s_mount_opt, MBALLOC);
2616 kfree(sbi->s_mb_maxs);
2617 return -ENOMEM;
2618 }
2619
2620 /* order 0 is regular bitmap */
2621 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2622 sbi->s_mb_offsets[0] = 0;
2623
2624 i = 1;
2625 offset = 0;
2626 max = sb->s_blocksize << 2;
2627 do {
2628 sbi->s_mb_offsets[i] = offset;
2629 sbi->s_mb_maxs[i] = max;
2630 offset += 1 << (sb->s_blocksize_bits - i);
2631 max = max >> 1;
2632 i++;
2633 } while (i <= sb->s_blocksize_bits + 1);
2634
2635 /* init file for buddy data */
2636 i = ext4_mb_init_backend(sb);
2637 if (i) {
2638 clear_opt(sbi->s_mount_opt, MBALLOC);
2639 kfree(sbi->s_mb_offsets);
2640 kfree(sbi->s_mb_maxs);
2641 return i;
2642 }
2643
2644 spin_lock_init(&sbi->s_md_lock);
2645 INIT_LIST_HEAD(&sbi->s_active_transaction);
2646 INIT_LIST_HEAD(&sbi->s_closed_transaction);
2647 INIT_LIST_HEAD(&sbi->s_committed_transaction);
2648 spin_lock_init(&sbi->s_bal_lock);
2649
2650 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2651 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2652 sbi->s_mb_stats = MB_DEFAULT_STATS;
2653 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2654 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2655 sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
2656 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2657
2658 i = sizeof(struct ext4_locality_group) * NR_CPUS;
2659 sbi->s_locality_groups = kmalloc(i, GFP_KERNEL);
2660 if (sbi->s_locality_groups == NULL) {
2661 clear_opt(sbi->s_mount_opt, MBALLOC);
2662 kfree(sbi->s_mb_offsets);
2663 kfree(sbi->s_mb_maxs);
2664 return -ENOMEM;
2665 }
2666 for (i = 0; i < NR_CPUS; i++) {
2667 struct ext4_locality_group *lg;
2668 lg = &sbi->s_locality_groups[i];
2669 mutex_init(&lg->lg_mutex);
2670 INIT_LIST_HEAD(&lg->lg_prealloc_list);
2671 spin_lock_init(&lg->lg_prealloc_lock);
2672 }
2673
2674 ext4_mb_init_per_dev_proc(sb);
2675 ext4_mb_history_init(sb);
2676
2677 printk("EXT4-fs: mballoc enabled\n");
2678 return 0;
2679}
2680
2681/* need to called with ext4 group lock (ext4_lock_group) */
2682static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2683{
2684 struct ext4_prealloc_space *pa;
2685 struct list_head *cur, *tmp;
2686 int count = 0;
2687
2688 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2689 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2690 list_del(&pa->pa_group_list);
2691 count++;
2692 kfree(pa);
2693 }
2694 if (count)
2695 mb_debug("mballoc: %u PAs left\n", count);
2696
2697}
2698
2699int ext4_mb_release(struct super_block *sb)
2700{
2701 ext4_group_t i;
2702 int num_meta_group_infos;
2703 struct ext4_group_info *grinfo;
2704 struct ext4_sb_info *sbi = EXT4_SB(sb);
2705
2706 if (!test_opt(sb, MBALLOC))
2707 return 0;
2708
2709 /* release freed, non-committed blocks */
2710 spin_lock(&sbi->s_md_lock);
2711 list_splice_init(&sbi->s_closed_transaction,
2712 &sbi->s_committed_transaction);
2713 list_splice_init(&sbi->s_active_transaction,
2714 &sbi->s_committed_transaction);
2715 spin_unlock(&sbi->s_md_lock);
2716 ext4_mb_free_committed_blocks(sb);
2717
2718 if (sbi->s_group_info) {
2719 for (i = 0; i < sbi->s_groups_count; i++) {
2720 grinfo = ext4_get_group_info(sb, i);
2721#ifdef DOUBLE_CHECK
2722 kfree(grinfo->bb_bitmap);
2723#endif
2724 ext4_lock_group(sb, i);
2725 ext4_mb_cleanup_pa(grinfo);
2726 ext4_unlock_group(sb, i);
2727 kfree(grinfo);
2728 }
2729 num_meta_group_infos = (sbi->s_groups_count +
2730 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2731 EXT4_DESC_PER_BLOCK_BITS(sb);
2732 for (i = 0; i < num_meta_group_infos; i++)
2733 kfree(sbi->s_group_info[i]);
2734 kfree(sbi->s_group_info);
2735 }
2736 kfree(sbi->s_mb_offsets);
2737 kfree(sbi->s_mb_maxs);
2738 if (sbi->s_buddy_cache)
2739 iput(sbi->s_buddy_cache);
2740 if (sbi->s_mb_stats) {
2741 printk(KERN_INFO
2742 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2743 atomic_read(&sbi->s_bal_allocated),
2744 atomic_read(&sbi->s_bal_reqs),
2745 atomic_read(&sbi->s_bal_success));
2746 printk(KERN_INFO
2747 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2748 "%u 2^N hits, %u breaks, %u lost\n",
2749 atomic_read(&sbi->s_bal_ex_scanned),
2750 atomic_read(&sbi->s_bal_goals),
2751 atomic_read(&sbi->s_bal_2orders),
2752 atomic_read(&sbi->s_bal_breaks),
2753 atomic_read(&sbi->s_mb_lost_chunks));
2754 printk(KERN_INFO
2755 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2756 sbi->s_mb_buddies_generated++,
2757 sbi->s_mb_generation_time);
2758 printk(KERN_INFO
2759 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2760 atomic_read(&sbi->s_mb_preallocated),
2761 atomic_read(&sbi->s_mb_discarded));
2762 }
2763
2764 kfree(sbi->s_locality_groups);
2765
2766 ext4_mb_history_release(sb);
2767 ext4_mb_destroy_per_dev_proc(sb);
2768
2769 return 0;
2770}
2771
2772static void ext4_mb_free_committed_blocks(struct super_block *sb)
2773{
2774 struct ext4_sb_info *sbi = EXT4_SB(sb);
2775 int err;
2776 int i;
2777 int count = 0;
2778 int count2 = 0;
2779 struct ext4_free_metadata *md;
2780 struct ext4_buddy e4b;
2781
2782 if (list_empty(&sbi->s_committed_transaction))
2783 return;
2784
2785 /* there is committed blocks to be freed yet */
2786 do {
2787 /* get next array of blocks */
2788 md = NULL;
2789 spin_lock(&sbi->s_md_lock);
2790 if (!list_empty(&sbi->s_committed_transaction)) {
2791 md = list_entry(sbi->s_committed_transaction.next,
2792 struct ext4_free_metadata, list);
2793 list_del(&md->list);
2794 }
2795 spin_unlock(&sbi->s_md_lock);
2796
2797 if (md == NULL)
2798 break;
2799
2800 mb_debug("gonna free %u blocks in group %lu (0x%p):",
2801 md->num, md->group, md);
2802
2803 err = ext4_mb_load_buddy(sb, md->group, &e4b);
2804 /* we expect to find existing buddy because it's pinned */
2805 BUG_ON(err != 0);
2806
2807 /* there are blocks to put in buddy to make them really free */
2808 count += md->num;
2809 count2++;
2810 ext4_lock_group(sb, md->group);
2811 for (i = 0; i < md->num; i++) {
2812 mb_debug(" %u", md->blocks[i]);
2813 err = mb_free_blocks(NULL, &e4b, md->blocks[i], 1);
2814 BUG_ON(err != 0);
2815 }
2816 mb_debug("\n");
2817 ext4_unlock_group(sb, md->group);
2818
2819 /* balance refcounts from ext4_mb_free_metadata() */
2820 page_cache_release(e4b.bd_buddy_page);
2821 page_cache_release(e4b.bd_bitmap_page);
2822
2823 kfree(md);
2824 ext4_mb_release_desc(&e4b);
2825
2826 } while (md);
2827
2828 mb_debug("freed %u blocks in %u structures\n", count, count2);
2829}
2830
2831#define EXT4_ROOT "ext4"
2832#define EXT4_MB_STATS_NAME "stats"
2833#define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2834#define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2835#define EXT4_MB_ORDER2_REQ "order2_req"
2836#define EXT4_MB_STREAM_REQ "stream_req"
2837#define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2838
2839
2840
2841#define MB_PROC_VALUE_READ(name) \
2842static int ext4_mb_read_##name(char *page, char **start, \
2843 off_t off, int count, int *eof, void *data) \
2844{ \
2845 struct ext4_sb_info *sbi = data; \
2846 int len; \
2847 *eof = 1; \
2848 if (off != 0) \
2849 return 0; \
2850 len = sprintf(page, "%ld\n", sbi->s_mb_##name); \
2851 *start = page; \
2852 return len; \
2853}
2854
2855#define MB_PROC_VALUE_WRITE(name) \
2856static int ext4_mb_write_##name(struct file *file, \
2857 const char __user *buf, unsigned long cnt, void *data) \
2858{ \
2859 struct ext4_sb_info *sbi = data; \
2860 char str[32]; \
2861 long value; \
2862 if (cnt >= sizeof(str)) \
2863 return -EINVAL; \
2864 if (copy_from_user(str, buf, cnt)) \
2865 return -EFAULT; \
2866 value = simple_strtol(str, NULL, 0); \
2867 if (value <= 0) \
2868 return -ERANGE; \
2869 sbi->s_mb_##name = value; \
2870 return cnt; \
2871}
2872
2873MB_PROC_VALUE_READ(stats);
2874MB_PROC_VALUE_WRITE(stats);
2875MB_PROC_VALUE_READ(max_to_scan);
2876MB_PROC_VALUE_WRITE(max_to_scan);
2877MB_PROC_VALUE_READ(min_to_scan);
2878MB_PROC_VALUE_WRITE(min_to_scan);
2879MB_PROC_VALUE_READ(order2_reqs);
2880MB_PROC_VALUE_WRITE(order2_reqs);
2881MB_PROC_VALUE_READ(stream_request);
2882MB_PROC_VALUE_WRITE(stream_request);
2883MB_PROC_VALUE_READ(group_prealloc);
2884MB_PROC_VALUE_WRITE(group_prealloc);
2885
2886#define MB_PROC_HANDLER(name, var) \
2887do { \
2888 proc = create_proc_entry(name, mode, sbi->s_mb_proc); \
2889 if (proc == NULL) { \
2890 printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
2891 goto err_out; \
2892 } \
2893 proc->data = sbi; \
2894 proc->read_proc = ext4_mb_read_##var ; \
2895 proc->write_proc = ext4_mb_write_##var; \
2896} while (0)
2897
2898static int ext4_mb_init_per_dev_proc(struct super_block *sb)
2899{
2900 mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
2901 struct ext4_sb_info *sbi = EXT4_SB(sb);
2902 struct proc_dir_entry *proc;
2903 char devname[64];
2904
2905 snprintf(devname, sizeof(devname) - 1, "%s",
2906 bdevname(sb->s_bdev, devname));
2907 sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext4);
2908
2909 MB_PROC_HANDLER(EXT4_MB_STATS_NAME, stats);
2910 MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, max_to_scan);
2911 MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, min_to_scan);
2912 MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ, order2_reqs);
2913 MB_PROC_HANDLER(EXT4_MB_STREAM_REQ, stream_request);
2914 MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, group_prealloc);
2915
2916 return 0;
2917
2918err_out:
2919 printk(KERN_ERR "EXT4-fs: Unable to create %s\n", devname);
2920 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2921 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2922 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2923 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2924 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2925 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2926 remove_proc_entry(devname, proc_root_ext4);
2927 sbi->s_mb_proc = NULL;
2928
2929 return -ENOMEM;
2930}
2931
2932static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
2933{
2934 struct ext4_sb_info *sbi = EXT4_SB(sb);
2935 char devname[64];
2936
2937 if (sbi->s_mb_proc == NULL)
2938 return -EINVAL;
2939
2940 snprintf(devname, sizeof(devname) - 1, "%s",
2941 bdevname(sb->s_bdev, devname));
2942 remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
2943 remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
2944 remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
2945 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
2946 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
2947 remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
2948 remove_proc_entry(devname, proc_root_ext4);
2949
2950 return 0;
2951}
2952
2953int __init init_ext4_mballoc(void)
2954{
2955 ext4_pspace_cachep =
2956 kmem_cache_create("ext4_prealloc_space",
2957 sizeof(struct ext4_prealloc_space),
2958 0, SLAB_RECLAIM_ACCOUNT, NULL);
2959 if (ext4_pspace_cachep == NULL)
2960 return -ENOMEM;
2961
2962#ifdef CONFIG_PROC_FS
2963 proc_root_ext4 = proc_mkdir(EXT4_ROOT, proc_root_fs);
2964 if (proc_root_ext4 == NULL)
2965 printk(KERN_ERR "EXT4-fs: Unable to create %s\n", EXT4_ROOT);
2966#endif
2967
2968 return 0;
2969}
2970
2971void exit_ext4_mballoc(void)
2972{
2973 /* XXX: synchronize_rcu(); */
2974 kmem_cache_destroy(ext4_pspace_cachep);
2975#ifdef CONFIG_PROC_FS
2976 remove_proc_entry(EXT4_ROOT, proc_root_fs);
2977#endif
2978}
2979
2980
2981/*
2982 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2983 * Returns 0 if success or error code
2984 */
2985static int ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2986 handle_t *handle)
2987{
2988 struct buffer_head *bitmap_bh = NULL;
2989 struct ext4_super_block *es;
2990 struct ext4_group_desc *gdp;
2991 struct buffer_head *gdp_bh;
2992 struct ext4_sb_info *sbi;
2993 struct super_block *sb;
2994 ext4_fsblk_t block;
2995 int err;
2996
2997 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2998 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2999
3000 sb = ac->ac_sb;
3001 sbi = EXT4_SB(sb);
3002 es = sbi->s_es;
3003
3004 ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
3005 gdp->bg_free_blocks_count);
3006
3007 err = -EIO;
3008 bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3009 if (!bitmap_bh)
3010 goto out_err;
3011
3012 err = ext4_journal_get_write_access(handle, bitmap_bh);
3013 if (err)
3014 goto out_err;
3015
3016 err = -EIO;
3017 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3018 if (!gdp)
3019 goto out_err;
3020
3021 err = ext4_journal_get_write_access(handle, gdp_bh);
3022 if (err)
3023 goto out_err;
3024
3025 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
3026 + ac->ac_b_ex.fe_start
3027 + le32_to_cpu(es->s_first_data_block);
3028
3029 if (block == ext4_block_bitmap(sb, gdp) ||
3030 block == ext4_inode_bitmap(sb, gdp) ||
3031 in_range(block, ext4_inode_table(sb, gdp),
3032 EXT4_SB(sb)->s_itb_per_group)) {
3033
3034 ext4_error(sb, __FUNCTION__,
3035 "Allocating block in system zone - block = %llu",
3036 block);
3037 }
3038#ifdef AGGRESSIVE_CHECK
3039 {
3040 int i;
3041 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3042 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3043 bitmap_bh->b_data));
3044 }
3045 }
3046#endif
3047 mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
3048 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
3049
3050 spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3051 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
3052 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3053 gdp->bg_free_blocks_count =
3054 cpu_to_le16(ext4_free_blocks_after_init(sb,
3055 ac->ac_b_ex.fe_group,
3056 gdp));
3057 }
3058 gdp->bg_free_blocks_count =
3059 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
3060 - ac->ac_b_ex.fe_len);
3061 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
3062 spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
3063 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
3064
3065 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
3066 if (err)
3067 goto out_err;
3068 err = ext4_journal_dirty_metadata(handle, gdp_bh);
3069
3070out_err:
3071 sb->s_dirt = 1;
3072 put_bh(bitmap_bh);
3073 return err;
3074}
3075
3076/*
3077 * here we normalize request for locality group
3078 * Group request are normalized to s_strip size if we set the same via mount
3079 * option. If not we set it to s_mb_group_prealloc which can be configured via
3080 * /proc/fs/ext4/<partition>/group_prealloc
3081 *
3082 * XXX: should we try to preallocate more than the group has now?
3083 */
3084static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3085{
3086 struct super_block *sb = ac->ac_sb;
3087 struct ext4_locality_group *lg = ac->ac_lg;
3088
3089 BUG_ON(lg == NULL);
3090 if (EXT4_SB(sb)->s_stripe)
3091 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
3092 else
3093 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3094 mb_debug("#%u: goal %lu blocks for locality group\n",
3095 current->pid, ac->ac_g_ex.fe_len);
3096}
3097
3098/*
3099 * Normalization means making request better in terms of
3100 * size and alignment
3101 */
3102static void ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3103 struct ext4_allocation_request *ar)
3104{
3105 int bsbits, max;
3106 ext4_lblk_t end;
3107 struct list_head *cur;
3108 loff_t size, orig_size, start_off;
3109 ext4_lblk_t start, orig_start;
3110 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3111
3112 /* do normalize only data requests, metadata requests
3113 do not need preallocation */
3114 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3115 return;
3116
3117 /* sometime caller may want exact blocks */
3118 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3119 return;
3120
3121 /* caller may indicate that preallocation isn't
3122 * required (it's a tail, for example) */
3123 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3124 return;
3125
3126 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3127 ext4_mb_normalize_group_request(ac);
3128 return ;
3129 }
3130
3131 bsbits = ac->ac_sb->s_blocksize_bits;
3132
3133 /* first, let's learn actual file size
3134 * given current request is allocated */
3135 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3136 size = size << bsbits;
3137 if (size < i_size_read(ac->ac_inode))
3138 size = i_size_read(ac->ac_inode);
3139
3140 /* max available blocks in a free group */
3141 max = EXT4_BLOCKS_PER_GROUP(ac->ac_sb) - 1 - 1 -
3142 EXT4_SB(ac->ac_sb)->s_itb_per_group;
3143
3144#define NRL_CHECK_SIZE(req, size, max,bits) \
3145 (req <= (size) || max <= ((size) >> bits))
3146
3147 /* first, try to predict filesize */
3148 /* XXX: should this table be tunable? */
3149 start_off = 0;
3150 if (size <= 16 * 1024) {
3151 size = 16 * 1024;
3152 } else if (size <= 32 * 1024) {
3153 size = 32 * 1024;
3154 } else if (size <= 64 * 1024) {
3155 size = 64 * 1024;
3156 } else if (size <= 128 * 1024) {
3157 size = 128 * 1024;
3158 } else if (size <= 256 * 1024) {
3159 size = 256 * 1024;
3160 } else if (size <= 512 * 1024) {
3161 size = 512 * 1024;
3162 } else if (size <= 1024 * 1024) {
3163 size = 1024 * 1024;
3164 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, bsbits)) {
3165 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3166 (20 - bsbits)) << 20;
3167 size = 1024 * 1024;
3168 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, bsbits)) {
3169 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3170 (22 - bsbits)) << 22;
3171 size = 4 * 1024 * 1024;
3172 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3173 (8<<20)>>bsbits, max, bsbits)) {
3174 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3175 (23 - bsbits)) << 23;
3176 size = 8 * 1024 * 1024;
3177 } else {
3178 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3179 size = ac->ac_o_ex.fe_len << bsbits;
3180 }
3181 orig_size = size = size >> bsbits;
3182 orig_start = start = start_off >> bsbits;
3183
3184 /* don't cover already allocated blocks in selected range */
3185 if (ar->pleft && start <= ar->lleft) {
3186 size -= ar->lleft + 1 - start;
3187 start = ar->lleft + 1;
3188 }
3189 if (ar->pright && start + size - 1 >= ar->lright)
3190 size -= start + size - ar->lright;
3191
3192 end = start + size;
3193
3194 /* check we don't cross already preallocated blocks */
3195 rcu_read_lock();
3196 list_for_each_rcu(cur, &ei->i_prealloc_list) {
3197 struct ext4_prealloc_space *pa;
3198 unsigned long pa_end;
3199
3200 pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
3201
3202 if (pa->pa_deleted)
3203 continue;
3204 spin_lock(&pa->pa_lock);
3205 if (pa->pa_deleted) {
3206 spin_unlock(&pa->pa_lock);
3207 continue;
3208 }
3209
3210 pa_end = pa->pa_lstart + pa->pa_len;
3211
3212 /* PA must not overlap original request */
3213 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3214 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3215
3216 /* skip PA normalized request doesn't overlap with */
3217 if (pa->pa_lstart >= end) {
3218 spin_unlock(&pa->pa_lock);
3219 continue;
3220 }
3221 if (pa_end <= start) {
3222 spin_unlock(&pa->pa_lock);
3223 continue;
3224 }
3225 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3226
3227 if (pa_end <= ac->ac_o_ex.fe_logical) {
3228 BUG_ON(pa_end < start);
3229 start = pa_end;
3230 }
3231
3232 if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3233 BUG_ON(pa->pa_lstart > end);
3234 end = pa->pa_lstart;
3235 }
3236 spin_unlock(&pa->pa_lock);
3237 }
3238 rcu_read_unlock();
3239 size = end - start;
3240
3241 /* XXX: extra loop to check we really don't overlap preallocations */
3242 rcu_read_lock();
3243 list_for_each_rcu(cur, &ei->i_prealloc_list) {
3244 struct ext4_prealloc_space *pa;
3245 unsigned long pa_end;
3246 pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
3247 spin_lock(&pa->pa_lock);
3248 if (pa->pa_deleted == 0) {
3249 pa_end = pa->pa_lstart + pa->pa_len;
3250 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3251 }
3252 spin_unlock(&pa->pa_lock);
3253 }
3254 rcu_read_unlock();
3255
3256 if (start + size <= ac->ac_o_ex.fe_logical &&
3257 start > ac->ac_o_ex.fe_logical) {
3258 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
3259 (unsigned long) start, (unsigned long) size,
3260 (unsigned long) ac->ac_o_ex.fe_logical);
3261 }
3262 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3263 start > ac->ac_o_ex.fe_logical);
3264 BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3265
3266 /* now prepare goal request */
3267
3268 /* XXX: is it better to align blocks WRT to logical
3269 * placement or satisfy big request as is */
3270 ac->ac_g_ex.fe_logical = start;
3271 ac->ac_g_ex.fe_len = size;
3272
3273 /* define goal start in order to merge */
3274 if (ar->pright && (ar->lright == (start + size))) {
3275 /* merge to the right */
3276 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3277 &ac->ac_f_ex.fe_group,
3278 &ac->ac_f_ex.fe_start);
3279 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3280 }
3281 if (ar->pleft && (ar->lleft + 1 == start)) {
3282 /* merge to the left */
3283 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3284 &ac->ac_f_ex.fe_group,
3285 &ac->ac_f_ex.fe_start);
3286 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3287 }
3288
3289 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
3290 (unsigned) orig_size, (unsigned) start);
3291}
3292
3293static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3294{
3295 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3296
3297 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3298 atomic_inc(&sbi->s_bal_reqs);
3299 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3300 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
3301 atomic_inc(&sbi->s_bal_success);
3302 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3303 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3304 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3305 atomic_inc(&sbi->s_bal_goals);
3306 if (ac->ac_found > sbi->s_mb_max_to_scan)
3307 atomic_inc(&sbi->s_bal_breaks);
3308 }
3309
3310 ext4_mb_store_history(ac);
3311}
3312
3313/*
3314 * use blocks preallocated to inode
3315 */
3316static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3317 struct ext4_prealloc_space *pa)
3318{
3319 ext4_fsblk_t start;
3320 ext4_fsblk_t end;
3321 int len;
3322
3323 /* found preallocated blocks, use them */
3324 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3325 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3326 len = end - start;
3327 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3328 &ac->ac_b_ex.fe_start);
3329 ac->ac_b_ex.fe_len = len;
3330 ac->ac_status = AC_STATUS_FOUND;
3331 ac->ac_pa = pa;
3332
3333 BUG_ON(start < pa->pa_pstart);
3334 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3335 BUG_ON(pa->pa_free < len);
3336 pa->pa_free -= len;
3337
3338 mb_debug("use %llu/%lu from inode pa %p\n", start, len, pa);
3339}
3340
3341/*
3342 * use blocks preallocated to locality group
3343 */
3344static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3345 struct ext4_prealloc_space *pa)
3346{
3347 unsigned len = ac->ac_o_ex.fe_len;
3348
3349 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3350 &ac->ac_b_ex.fe_group,
3351 &ac->ac_b_ex.fe_start);
3352 ac->ac_b_ex.fe_len = len;
3353 ac->ac_status = AC_STATUS_FOUND;
3354 ac->ac_pa = pa;
3355
3356 /* we don't correct pa_pstart or pa_plen here to avoid
3357 * possible race when tte group is being loaded concurrently
3358 * instead we correct pa later, after blocks are marked
3359 * in on-disk bitmap -- see ext4_mb_release_context() */
3360 /*
3361 * FIXME!! but the other CPUs can look at this particular
3362 * pa and think that it have enought free blocks if we
3363 * don't update pa_free here right ?
3364 */
3365 mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3366}
3367
3368/*
3369 * search goal blocks in preallocated space
3370 */
3371static int ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3372{
3373 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3374 struct ext4_locality_group *lg;
3375 struct ext4_prealloc_space *pa;
3376 struct list_head *cur;
3377
3378 /* only data can be preallocated */
3379 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3380 return 0;
3381
3382 /* first, try per-file preallocation */
3383 rcu_read_lock();
3384 list_for_each_rcu(cur, &ei->i_prealloc_list) {
3385 pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
3386
3387 /* all fields in this condition don't change,
3388 * so we can skip locking for them */
3389 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3390 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3391 continue;
3392
3393 /* found preallocated blocks, use them */
3394 spin_lock(&pa->pa_lock);
3395 if (pa->pa_deleted == 0 && pa->pa_free) {
3396 atomic_inc(&pa->pa_count);
3397 ext4_mb_use_inode_pa(ac, pa);
3398 spin_unlock(&pa->pa_lock);
3399 ac->ac_criteria = 10;
3400 rcu_read_unlock();
3401 return 1;
3402 }
3403 spin_unlock(&pa->pa_lock);
3404 }
3405 rcu_read_unlock();
3406
3407 /* can we use group allocation? */
3408 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3409 return 0;
3410
3411 /* inode may have no locality group for some reason */
3412 lg = ac->ac_lg;
3413 if (lg == NULL)
3414 return 0;
3415
3416 rcu_read_lock();
3417 list_for_each_rcu(cur, &lg->lg_prealloc_list) {
3418 pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
3419 spin_lock(&pa->pa_lock);
3420 if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
3421 atomic_inc(&pa->pa_count);
3422 ext4_mb_use_group_pa(ac, pa);
3423 spin_unlock(&pa->pa_lock);
3424 ac->ac_criteria = 20;
3425 rcu_read_unlock();
3426 return 1;
3427 }
3428 spin_unlock(&pa->pa_lock);
3429 }
3430 rcu_read_unlock();
3431
3432 return 0;
3433}
3434
3435/*
3436 * the function goes through all preallocation in this group and marks them
3437 * used in in-core bitmap. buddy must be generated from this bitmap
3438 * Need to be called with ext4 group lock (ext4_lock_group)
3439 */
3440static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3441 ext4_group_t group)
3442{
3443 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3444 struct ext4_prealloc_space *pa;
3445 struct list_head *cur;
3446 ext4_group_t groupnr;
3447 ext4_grpblk_t start;
3448 int preallocated = 0;
3449 int count = 0;
3450 int len;
3451
3452 /* all form of preallocation discards first load group,
3453 * so the only competing code is preallocation use.
3454 * we don't need any locking here
3455 * notice we do NOT ignore preallocations with pa_deleted
3456 * otherwise we could leave used blocks available for
3457 * allocation in buddy when concurrent ext4_mb_put_pa()
3458 * is dropping preallocation
3459 */
3460 list_for_each(cur, &grp->bb_prealloc_list) {
3461 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3462 spin_lock(&pa->pa_lock);
3463 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3464 &groupnr, &start);
3465 len = pa->pa_len;
3466 spin_unlock(&pa->pa_lock);
3467 if (unlikely(len == 0))
3468 continue;
3469 BUG_ON(groupnr != group);
3470 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
3471 bitmap, start, len);
3472 preallocated += len;
3473 count++;
3474 }
3475 mb_debug("prellocated %u for group %lu\n", preallocated, group);
3476}
3477
3478static void ext4_mb_pa_callback(struct rcu_head *head)
3479{
3480 struct ext4_prealloc_space *pa;
3481 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3482 kmem_cache_free(ext4_pspace_cachep, pa);
3483}
3484
3485/*
3486 * drops a reference to preallocated space descriptor
3487 * if this was the last reference and the space is consumed
3488 */
3489static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3490 struct super_block *sb, struct ext4_prealloc_space *pa)
3491{
3492 unsigned long grp;
3493
3494 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3495 return;
3496
3497 /* in this short window concurrent discard can set pa_deleted */
3498 spin_lock(&pa->pa_lock);
3499 if (pa->pa_deleted == 1) {
3500 spin_unlock(&pa->pa_lock);
3501 return;
3502 }
3503
3504 pa->pa_deleted = 1;
3505 spin_unlock(&pa->pa_lock);
3506
3507 /* -1 is to protect from crossing allocation group */
3508 ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
3509
3510 /*
3511 * possible race:
3512 *
3513 * P1 (buddy init) P2 (regular allocation)
3514 * find block B in PA
3515 * copy on-disk bitmap to buddy
3516 * mark B in on-disk bitmap
3517 * drop PA from group
3518 * mark all PAs in buddy
3519 *
3520 * thus, P1 initializes buddy with B available. to prevent this
3521 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3522 * against that pair
3523 */
3524 ext4_lock_group(sb, grp);
3525 list_del(&pa->pa_group_list);
3526 ext4_unlock_group(sb, grp);
3527
3528 spin_lock(pa->pa_obj_lock);
3529 list_del_rcu(&pa->pa_inode_list);
3530 spin_unlock(pa->pa_obj_lock);
3531
3532 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3533}
3534
3535/*
3536 * creates new preallocated space for given inode
3537 */
3538static int ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3539{
3540 struct super_block *sb = ac->ac_sb;
3541 struct ext4_prealloc_space *pa;
3542 struct ext4_group_info *grp;
3543 struct ext4_inode_info *ei;
3544
3545 /* preallocate only when found space is larger then requested */
3546 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3547 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3548 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3549
3550 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3551 if (pa == NULL)
3552 return -ENOMEM;
3553
3554 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3555 int winl;
3556 int wins;
3557 int win;
3558 int offs;
3559
3560 /* we can't allocate as much as normalizer wants.
3561 * so, found space must get proper lstart
3562 * to cover original request */
3563 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3564 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3565
3566 /* we're limited by original request in that
3567 * logical block must be covered any way
3568 * winl is window we can move our chunk within */
3569 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3570
3571 /* also, we should cover whole original request */
3572 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3573
3574 /* the smallest one defines real window */
3575 win = min(winl, wins);
3576
3577 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3578 if (offs && offs < win)
3579 win = offs;
3580
3581 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3582 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3583 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3584 }
3585
3586 /* preallocation can change ac_b_ex, thus we store actually
3587 * allocated blocks for history */
3588 ac->ac_f_ex = ac->ac_b_ex;
3589
3590 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3591 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3592 pa->pa_len = ac->ac_b_ex.fe_len;
3593 pa->pa_free = pa->pa_len;
3594 atomic_set(&pa->pa_count, 1);
3595 spin_lock_init(&pa->pa_lock);
3596 pa->pa_deleted = 0;
3597 pa->pa_linear = 0;
3598
3599 mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
3600 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3601
3602 ext4_mb_use_inode_pa(ac, pa);
3603 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3604
3605 ei = EXT4_I(ac->ac_inode);
3606 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3607
3608 pa->pa_obj_lock = &ei->i_prealloc_lock;
3609 pa->pa_inode = ac->ac_inode;
3610
3611 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3612 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3613 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3614
3615 spin_lock(pa->pa_obj_lock);
3616 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3617 spin_unlock(pa->pa_obj_lock);
3618
3619 return 0;
3620}
3621
3622/*
3623 * creates new preallocated space for locality group inodes belongs to
3624 */
3625static int ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3626{
3627 struct super_block *sb = ac->ac_sb;
3628 struct ext4_locality_group *lg;
3629 struct ext4_prealloc_space *pa;
3630 struct ext4_group_info *grp;
3631
3632 /* preallocate only when found space is larger then requested */
3633 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3634 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3635 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3636
3637 BUG_ON(ext4_pspace_cachep == NULL);
3638 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3639 if (pa == NULL)
3640 return -ENOMEM;
3641
3642 /* preallocation can change ac_b_ex, thus we store actually
3643 * allocated blocks for history */
3644 ac->ac_f_ex = ac->ac_b_ex;
3645
3646 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3647 pa->pa_lstart = pa->pa_pstart;
3648 pa->pa_len = ac->ac_b_ex.fe_len;
3649 pa->pa_free = pa->pa_len;
3650 atomic_set(&pa->pa_count, 1);
3651 spin_lock_init(&pa->pa_lock);
3652 pa->pa_deleted = 0;
3653 pa->pa_linear = 1;
3654
3655 mb_debug("new group pa %p: %llu/%u for %u\n", pa,
3656 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3657
3658 ext4_mb_use_group_pa(ac, pa);
3659 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3660
3661 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3662 lg = ac->ac_lg;
3663 BUG_ON(lg == NULL);
3664
3665 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3666 pa->pa_inode = NULL;
3667
3668 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3669 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3670 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3671
3672 spin_lock(pa->pa_obj_lock);
3673 list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
3674 spin_unlock(pa->pa_obj_lock);
3675
3676 return 0;
3677}
3678
3679static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3680{
3681 int err;
3682
3683 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3684 err = ext4_mb_new_group_pa(ac);
3685 else
3686 err = ext4_mb_new_inode_pa(ac);
3687 return err;
3688}
3689
3690/*
3691 * finds all unused blocks in on-disk bitmap, frees them in
3692 * in-core bitmap and buddy.
3693 * @pa must be unlinked from inode and group lists, so that
3694 * nobody else can find/use it.
3695 * the caller MUST hold group/inode locks.
3696 * TODO: optimize the case when there are no in-core structures yet
3697 */
3698static int ext4_mb_release_inode_pa(struct ext4_buddy *e4b,
3699 struct buffer_head *bitmap_bh,
3700 struct ext4_prealloc_space *pa)
3701{
3702 struct ext4_allocation_context ac;
3703 struct super_block *sb = e4b->bd_sb;
3704 struct ext4_sb_info *sbi = EXT4_SB(sb);
3705 unsigned long end;
3706 unsigned long next;
3707 ext4_group_t group;
3708 ext4_grpblk_t bit;
3709 sector_t start;
3710 int err = 0;
3711 int free = 0;
3712
3713 BUG_ON(pa->pa_deleted == 0);
3714 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3715 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3716 end = bit + pa->pa_len;
3717
3718 ac.ac_sb = sb;
3719 ac.ac_inode = pa->pa_inode;
3720 ac.ac_op = EXT4_MB_HISTORY_DISCARD;
3721
3722 while (bit < end) {
3723 bit = ext4_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3724 if (bit >= end)
3725 break;
3726 next = ext4_find_next_bit(bitmap_bh->b_data, end, bit);
3727 if (next > end)
3728 next = end;
3729 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3730 le32_to_cpu(sbi->s_es->s_first_data_block);
3731 mb_debug(" free preallocated %u/%u in group %u\n",
3732 (unsigned) start, (unsigned) next - bit,
3733 (unsigned) group);
3734 free += next - bit;
3735
3736 ac.ac_b_ex.fe_group = group;
3737 ac.ac_b_ex.fe_start = bit;
3738 ac.ac_b_ex.fe_len = next - bit;
3739 ac.ac_b_ex.fe_logical = 0;
3740 ext4_mb_store_history(&ac);
3741
3742 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3743 bit = next + 1;
3744 }
3745 if (free != pa->pa_free) {
3746 printk(KERN_ERR "pa %p: logic %lu, phys. %lu, len %lu\n",
3747 pa, (unsigned long) pa->pa_lstart,
3748 (unsigned long) pa->pa_pstart,
3749 (unsigned long) pa->pa_len);
3750 printk(KERN_ERR "free %u, pa_free %u\n", free, pa->pa_free);
3751 }
3752 BUG_ON(free != pa->pa_free);
3753 atomic_add(free, &sbi->s_mb_discarded);
3754
3755 return err;
3756}
3757
3758static int ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3759 struct ext4_prealloc_space *pa)
3760{
3761 struct ext4_allocation_context ac;
3762 struct super_block *sb = e4b->bd_sb;
3763 ext4_group_t group;
3764 ext4_grpblk_t bit;
3765
3766 ac.ac_op = EXT4_MB_HISTORY_DISCARD;
3767
3768 BUG_ON(pa->pa_deleted == 0);
3769 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3770 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3771 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3772 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3773
3774 ac.ac_sb = sb;
3775 ac.ac_inode = NULL;
3776 ac.ac_b_ex.fe_group = group;
3777 ac.ac_b_ex.fe_start = bit;
3778 ac.ac_b_ex.fe_len = pa->pa_len;
3779 ac.ac_b_ex.fe_logical = 0;
3780 ext4_mb_store_history(&ac);
3781
3782 return 0;
3783}
3784
3785/*
3786 * releases all preallocations in given group
3787 *
3788 * first, we need to decide discard policy:
3789 * - when do we discard
3790 * 1) ENOSPC
3791 * - how many do we discard
3792 * 1) how many requested
3793 */
3794static int ext4_mb_discard_group_preallocations(struct super_block *sb,
3795 ext4_group_t group, int needed)
3796{
3797 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3798 struct buffer_head *bitmap_bh = NULL;
3799 struct ext4_prealloc_space *pa, *tmp;
3800 struct list_head list;
3801 struct ext4_buddy e4b;
3802 int err;
3803 int busy = 0;
3804 int free = 0;
3805
3806 mb_debug("discard preallocation for group %lu\n", group);
3807
3808 if (list_empty(&grp->bb_prealloc_list))
3809 return 0;
3810
3811 bitmap_bh = read_block_bitmap(sb, group);
3812 if (bitmap_bh == NULL) {
3813 /* error handling here */
3814 ext4_mb_release_desc(&e4b);
3815 BUG_ON(bitmap_bh == NULL);
3816 }
3817
3818 err = ext4_mb_load_buddy(sb, group, &e4b);
3819 BUG_ON(err != 0); /* error handling here */
3820
3821 if (needed == 0)
3822 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3823
3824 grp = ext4_get_group_info(sb, group);
3825 INIT_LIST_HEAD(&list);
3826
3827repeat:
3828 ext4_lock_group(sb, group);
3829 list_for_each_entry_safe(pa, tmp,
3830 &grp->bb_prealloc_list, pa_group_list) {
3831 spin_lock(&pa->pa_lock);
3832 if (atomic_read(&pa->pa_count)) {
3833 spin_unlock(&pa->pa_lock);
3834 busy = 1;
3835 continue;
3836 }
3837 if (pa->pa_deleted) {
3838 spin_unlock(&pa->pa_lock);
3839 continue;
3840 }
3841
3842 /* seems this one can be freed ... */
3843 pa->pa_deleted = 1;
3844
3845 /* we can trust pa_free ... */
3846 free += pa->pa_free;
3847
3848 spin_unlock(&pa->pa_lock);
3849
3850 list_del(&pa->pa_group_list);
3851 list_add(&pa->u.pa_tmp_list, &list);
3852 }
3853
3854 /* if we still need more blocks and some PAs were used, try again */
3855 if (free < needed && busy) {
3856 busy = 0;
3857 ext4_unlock_group(sb, group);
3858 /*
3859 * Yield the CPU here so that we don't get soft lockup
3860 * in non preempt case.
3861 */
3862 yield();
3863 goto repeat;
3864 }
3865
3866 /* found anything to free? */
3867 if (list_empty(&list)) {
3868 BUG_ON(free != 0);
3869 goto out;
3870 }
3871
3872 /* now free all selected PAs */
3873 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3874
3875 /* remove from object (inode or locality group) */
3876 spin_lock(pa->pa_obj_lock);
3877 list_del_rcu(&pa->pa_inode_list);
3878 spin_unlock(pa->pa_obj_lock);
3879
3880 if (pa->pa_linear)
3881 ext4_mb_release_group_pa(&e4b, pa);
3882 else
3883 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3884
3885 list_del(&pa->u.pa_tmp_list);
3886 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3887 }
3888
3889out:
3890 ext4_unlock_group(sb, group);
3891 ext4_mb_release_desc(&e4b);
3892 put_bh(bitmap_bh);
3893 return free;
3894}
3895
3896/*
3897 * releases all non-used preallocated blocks for given inode
3898 *
3899 * It's important to discard preallocations under i_data_sem
3900 * We don't want another block to be served from the prealloc
3901 * space when we are discarding the inode prealloc space.
3902 *
3903 * FIXME!! Make sure it is valid at all the call sites
3904 */
3905void ext4_mb_discard_inode_preallocations(struct inode *inode)
3906{
3907 struct ext4_inode_info *ei = EXT4_I(inode);
3908 struct super_block *sb = inode->i_sb;
3909 struct buffer_head *bitmap_bh = NULL;
3910 struct ext4_prealloc_space *pa, *tmp;
3911 ext4_group_t group = 0;
3912 struct list_head list;
3913 struct ext4_buddy e4b;
3914 int err;
3915
3916 if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
3917 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3918 return;
3919 }
3920
3921 mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
3922
3923 INIT_LIST_HEAD(&list);
3924
3925repeat:
3926 /* first, collect all pa's in the inode */
3927 spin_lock(&ei->i_prealloc_lock);
3928 while (!list_empty(&ei->i_prealloc_list)) {
3929 pa = list_entry(ei->i_prealloc_list.next,
3930 struct ext4_prealloc_space, pa_inode_list);
3931 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3932 spin_lock(&pa->pa_lock);
3933 if (atomic_read(&pa->pa_count)) {
3934 /* this shouldn't happen often - nobody should
3935 * use preallocation while we're discarding it */
3936 spin_unlock(&pa->pa_lock);
3937 spin_unlock(&ei->i_prealloc_lock);
3938 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3939 WARN_ON(1);
3940 schedule_timeout_uninterruptible(HZ);
3941 goto repeat;
3942
3943 }
3944 if (pa->pa_deleted == 0) {
3945 pa->pa_deleted = 1;
3946 spin_unlock(&pa->pa_lock);
3947 list_del_rcu(&pa->pa_inode_list);
3948 list_add(&pa->u.pa_tmp_list, &list);
3949 continue;
3950 }
3951
3952 /* someone is deleting pa right now */
3953 spin_unlock(&pa->pa_lock);
3954 spin_unlock(&ei->i_prealloc_lock);
3955
3956 /* we have to wait here because pa_deleted
3957 * doesn't mean pa is already unlinked from
3958 * the list. as we might be called from
3959 * ->clear_inode() the inode will get freed
3960 * and concurrent thread which is unlinking
3961 * pa from inode's list may access already
3962 * freed memory, bad-bad-bad */
3963
3964 /* XXX: if this happens too often, we can
3965 * add a flag to force wait only in case
3966 * of ->clear_inode(), but not in case of
3967 * regular truncate */
3968 schedule_timeout_uninterruptible(HZ);
3969 goto repeat;
3970 }
3971 spin_unlock(&ei->i_prealloc_lock);
3972
3973 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3974 BUG_ON(pa->pa_linear != 0);
3975 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3976
3977 err = ext4_mb_load_buddy(sb, group, &e4b);
3978 BUG_ON(err != 0); /* error handling here */
3979
3980 bitmap_bh = read_block_bitmap(sb, group);
3981 if (bitmap_bh == NULL) {
3982 /* error handling here */
3983 ext4_mb_release_desc(&e4b);
3984 BUG_ON(bitmap_bh == NULL);
3985 }
3986
3987 ext4_lock_group(sb, group);
3988 list_del(&pa->pa_group_list);
3989 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3990 ext4_unlock_group(sb, group);
3991
3992 ext4_mb_release_desc(&e4b);
3993 put_bh(bitmap_bh);
3994
3995 list_del(&pa->u.pa_tmp_list);
3996 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3997 }
3998}
3999
4000/*
4001 * finds all preallocated spaces and return blocks being freed to them
4002 * if preallocated space becomes full (no block is used from the space)
4003 * then the function frees space in buddy
4004 * XXX: at the moment, truncate (which is the only way to free blocks)
4005 * discards all preallocations
4006 */
4007static void ext4_mb_return_to_preallocation(struct inode *inode,
4008 struct ext4_buddy *e4b,
4009 sector_t block, int count)
4010{
4011 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
4012}
4013#ifdef MB_DEBUG
4014static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4015{
4016 struct super_block *sb = ac->ac_sb;
4017 ext4_group_t i;
4018
4019 printk(KERN_ERR "EXT4-fs: Can't allocate:"
4020 " Allocation context details:\n");
4021 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
4022 ac->ac_status, ac->ac_flags);
4023 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4024 "best %lu/%lu/%lu@%lu cr %d\n",
4025 (unsigned long)ac->ac_o_ex.fe_group,
4026 (unsigned long)ac->ac_o_ex.fe_start,
4027 (unsigned long)ac->ac_o_ex.fe_len,
4028 (unsigned long)ac->ac_o_ex.fe_logical,
4029 (unsigned long)ac->ac_g_ex.fe_group,
4030 (unsigned long)ac->ac_g_ex.fe_start,
4031 (unsigned long)ac->ac_g_ex.fe_len,
4032 (unsigned long)ac->ac_g_ex.fe_logical,
4033 (unsigned long)ac->ac_b_ex.fe_group,
4034 (unsigned long)ac->ac_b_ex.fe_start,
4035 (unsigned long)ac->ac_b_ex.fe_len,
4036 (unsigned long)ac->ac_b_ex.fe_logical,
4037 (int)ac->ac_criteria);
4038 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
4039 ac->ac_found);
4040 printk(KERN_ERR "EXT4-fs: groups: \n");
4041 for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
4042 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4043 struct ext4_prealloc_space *pa;
4044 ext4_grpblk_t start;
4045 struct list_head *cur;
4046 ext4_lock_group(sb, i);
4047 list_for_each(cur, &grp->bb_prealloc_list) {
4048 pa = list_entry(cur, struct ext4_prealloc_space,
4049 pa_group_list);
4050 spin_lock(&pa->pa_lock);
4051 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4052 NULL, &start);
4053 spin_unlock(&pa->pa_lock);
4054 printk(KERN_ERR "PA:%lu:%d:%u \n", i,
4055 start, pa->pa_len);
4056 }
4057 ext4_lock_group(sb, i);
4058
4059 if (grp->bb_free == 0)
4060 continue;
4061 printk(KERN_ERR "%lu: %d/%d \n",
4062 i, grp->bb_free, grp->bb_fragments);
4063 }
4064 printk(KERN_ERR "\n");
4065}
4066#else
4067static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4068{
4069 return;
4070}
4071#endif
4072
4073/*
4074 * We use locality group preallocation for small size file. The size of the
4075 * file is determined by the current size or the resulting size after
4076 * allocation which ever is larger
4077 *
4078 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4079 */
4080static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4081{
4082 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4083 int bsbits = ac->ac_sb->s_blocksize_bits;
4084 loff_t size, isize;
4085
4086 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4087 return;
4088
4089 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
4090 isize = i_size_read(ac->ac_inode) >> bsbits;
4091 size = max(size, isize);
4092
4093 /* don't use group allocation for large files */
4094 if (size >= sbi->s_mb_stream_request)
4095 return;
4096
4097 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4098 return;
4099
4100 BUG_ON(ac->ac_lg != NULL);
4101 /*
4102 * locality group prealloc space are per cpu. The reason for having
4103 * per cpu locality group is to reduce the contention between block
4104 * request from multiple CPUs.
4105 */
4106 ac->ac_lg = &sbi->s_locality_groups[get_cpu()];
4107 put_cpu();
4108
4109 /* we're going to use group allocation */
4110 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4111
4112 /* serialize all allocations in the group */
4113 mutex_lock(&ac->ac_lg->lg_mutex);
4114}
4115
4116static int ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4117 struct ext4_allocation_request *ar)
4118{
4119 struct super_block *sb = ar->inode->i_sb;
4120 struct ext4_sb_info *sbi = EXT4_SB(sb);
4121 struct ext4_super_block *es = sbi->s_es;
4122 ext4_group_t group;
4123 unsigned long len;
4124 unsigned long goal;
4125 ext4_grpblk_t block;
4126
4127 /* we can't allocate > group size */
4128 len = ar->len;
4129
4130 /* just a dirty hack to filter too big requests */
4131 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4132 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4133
4134 /* start searching from the goal */
4135 goal = ar->goal;
4136 if (goal < le32_to_cpu(es->s_first_data_block) ||
4137 goal >= ext4_blocks_count(es))
4138 goal = le32_to_cpu(es->s_first_data_block);
4139 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4140
4141 /* set up allocation goals */
4142 ac->ac_b_ex.fe_logical = ar->logical;
4143 ac->ac_b_ex.fe_group = 0;
4144 ac->ac_b_ex.fe_start = 0;
4145 ac->ac_b_ex.fe_len = 0;
4146 ac->ac_status = AC_STATUS_CONTINUE;
4147 ac->ac_groups_scanned = 0;
4148 ac->ac_ex_scanned = 0;
4149 ac->ac_found = 0;
4150 ac->ac_sb = sb;
4151 ac->ac_inode = ar->inode;
4152 ac->ac_o_ex.fe_logical = ar->logical;
4153 ac->ac_o_ex.fe_group = group;
4154 ac->ac_o_ex.fe_start = block;
4155 ac->ac_o_ex.fe_len = len;
4156 ac->ac_g_ex.fe_logical = ar->logical;
4157 ac->ac_g_ex.fe_group = group;
4158 ac->ac_g_ex.fe_start = block;
4159 ac->ac_g_ex.fe_len = len;
4160 ac->ac_f_ex.fe_len = 0;
4161 ac->ac_flags = ar->flags;
4162 ac->ac_2order = 0;
4163 ac->ac_criteria = 0;
4164 ac->ac_pa = NULL;
4165 ac->ac_bitmap_page = NULL;
4166 ac->ac_buddy_page = NULL;
4167 ac->ac_lg = NULL;
4168
4169 /* we have to define context: we'll we work with a file or
4170 * locality group. this is a policy, actually */
4171 ext4_mb_group_or_file(ac);
4172
4173 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4174 "left: %u/%u, right %u/%u to %swritable\n",
4175 (unsigned) ar->len, (unsigned) ar->logical,
4176 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4177 (unsigned) ar->lleft, (unsigned) ar->pleft,
4178 (unsigned) ar->lright, (unsigned) ar->pright,
4179 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4180 return 0;
4181
4182}
4183
4184/*
4185 * release all resource we used in allocation
4186 */
4187static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4188{
4189 if (ac->ac_pa) {
4190 if (ac->ac_pa->pa_linear) {
4191 /* see comment in ext4_mb_use_group_pa() */
4192 spin_lock(&ac->ac_pa->pa_lock);
4193 ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
4194 ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
4195 ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
4196 ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
4197 spin_unlock(&ac->ac_pa->pa_lock);
4198 }
4199 ext4_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
4200 }
4201 if (ac->ac_bitmap_page)
4202 page_cache_release(ac->ac_bitmap_page);
4203 if (ac->ac_buddy_page)
4204 page_cache_release(ac->ac_buddy_page);
4205 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4206 mutex_unlock(&ac->ac_lg->lg_mutex);
4207 ext4_mb_collect_stats(ac);
4208 return 0;
4209}
4210
4211static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4212{
4213 ext4_group_t i;
4214 int ret;
4215 int freed = 0;
4216
4217 for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
4218 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4219 freed += ret;
4220 needed -= ret;
4221 }
4222
4223 return freed;
4224}
4225
4226/*
4227 * Main entry point into mballoc to allocate blocks
4228 * it tries to use preallocation first, then falls back
4229 * to usual allocation
4230 */
4231ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4232 struct ext4_allocation_request *ar, int *errp)
4233{
4234 struct ext4_allocation_context ac;
4235 struct ext4_sb_info *sbi;
4236 struct super_block *sb;
4237 ext4_fsblk_t block = 0;
4238 int freed;
4239 int inquota;
4240
4241 sb = ar->inode->i_sb;
4242 sbi = EXT4_SB(sb);
4243
4244 if (!test_opt(sb, MBALLOC)) {
4245 block = ext4_new_blocks_old(handle, ar->inode, ar->goal,
4246 &(ar->len), errp);
4247 return block;
4248 }
4249
4250 while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
4251 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4252 ar->len--;
4253 }
4254 if (ar->len == 0) {
4255 *errp = -EDQUOT;
4256 return 0;
4257 }
4258 inquota = ar->len;
4259
4260 ext4_mb_poll_new_transaction(sb, handle);
4261
4262 *errp = ext4_mb_initialize_context(&ac, ar);
4263 if (*errp) {
4264 ar->len = 0;
4265 goto out;
4266 }
4267
4268 ac.ac_op = EXT4_MB_HISTORY_PREALLOC;
4269 if (!ext4_mb_use_preallocated(&ac)) {
4270
4271 ac.ac_op = EXT4_MB_HISTORY_ALLOC;
4272 ext4_mb_normalize_request(&ac, ar);
4273
4274repeat:
4275 /* allocate space in core */
4276 ext4_mb_regular_allocator(&ac);
4277
4278 /* as we've just preallocated more space than
4279 * user requested orinally, we store allocated
4280 * space in a special descriptor */
4281 if (ac.ac_status == AC_STATUS_FOUND &&
4282 ac.ac_o_ex.fe_len < ac.ac_b_ex.fe_len)
4283 ext4_mb_new_preallocation(&ac);
4284 }
4285
4286 if (likely(ac.ac_status == AC_STATUS_FOUND)) {
4287 ext4_mb_mark_diskspace_used(&ac, handle);
4288 *errp = 0;
4289 block = ext4_grp_offs_to_block(sb, &ac.ac_b_ex);
4290 ar->len = ac.ac_b_ex.fe_len;
4291 } else {
4292 freed = ext4_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
4293 if (freed)
4294 goto repeat;
4295 *errp = -ENOSPC;
4296 ac.ac_b_ex.fe_len = 0;
4297 ar->len = 0;
4298 ext4_mb_show_ac(&ac);
4299 }
4300
4301 ext4_mb_release_context(&ac);
4302
4303out:
4304 if (ar->len < inquota)
4305 DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
4306
4307 return block;
4308}
4309static void ext4_mb_poll_new_transaction(struct super_block *sb,
4310 handle_t *handle)
4311{
4312 struct ext4_sb_info *sbi = EXT4_SB(sb);
4313
4314 if (sbi->s_last_transaction == handle->h_transaction->t_tid)
4315 return;
4316
4317 /* new transaction! time to close last one and free blocks for
4318 * committed transaction. we know that only transaction can be
4319 * active, so previos transaction can be being logged and we
4320 * know that transaction before previous is known to be already
4321 * logged. this means that now we may free blocks freed in all
4322 * transactions before previous one. hope I'm clear enough ... */
4323
4324 spin_lock(&sbi->s_md_lock);
4325 if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
4326 mb_debug("new transaction %lu, old %lu\n",
4327 (unsigned long) handle->h_transaction->t_tid,
4328 (unsigned long) sbi->s_last_transaction);
4329 list_splice_init(&sbi->s_closed_transaction,
4330 &sbi->s_committed_transaction);
4331 list_splice_init(&sbi->s_active_transaction,
4332 &sbi->s_closed_transaction);
4333 sbi->s_last_transaction = handle->h_transaction->t_tid;
4334 }
4335 spin_unlock(&sbi->s_md_lock);
4336
4337 ext4_mb_free_committed_blocks(sb);
4338}
4339
4340static int ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4341 ext4_group_t group, ext4_grpblk_t block, int count)
4342{
4343 struct ext4_group_info *db = e4b->bd_info;
4344 struct super_block *sb = e4b->bd_sb;
4345 struct ext4_sb_info *sbi = EXT4_SB(sb);
4346 struct ext4_free_metadata *md;
4347 int i;
4348
4349 BUG_ON(e4b->bd_bitmap_page == NULL);
4350 BUG_ON(e4b->bd_buddy_page == NULL);
4351
4352 ext4_lock_group(sb, group);
4353 for (i = 0; i < count; i++) {
4354 md = db->bb_md_cur;
4355 if (md && db->bb_tid != handle->h_transaction->t_tid) {
4356 db->bb_md_cur = NULL;
4357 md = NULL;
4358 }
4359
4360 if (md == NULL) {
4361 ext4_unlock_group(sb, group);
4362 md = kmalloc(sizeof(*md), GFP_NOFS);
4363 if (md == NULL)
4364 return -ENOMEM;
4365 md->num = 0;
4366 md->group = group;
4367
4368 ext4_lock_group(sb, group);
4369 if (db->bb_md_cur == NULL) {
4370 spin_lock(&sbi->s_md_lock);
4371 list_add(&md->list, &sbi->s_active_transaction);
4372 spin_unlock(&sbi->s_md_lock);
4373 /* protect buddy cache from being freed,
4374 * otherwise we'll refresh it from
4375 * on-disk bitmap and lose not-yet-available
4376 * blocks */
4377 page_cache_get(e4b->bd_buddy_page);
4378 page_cache_get(e4b->bd_bitmap_page);
4379 db->bb_md_cur = md;
4380 db->bb_tid = handle->h_transaction->t_tid;
4381 mb_debug("new md 0x%p for group %lu\n",
4382 md, md->group);
4383 } else {
4384 kfree(md);
4385 md = db->bb_md_cur;
4386 }
4387 }
4388
4389 BUG_ON(md->num >= EXT4_BB_MAX_BLOCKS);
4390 md->blocks[md->num] = block + i;
4391 md->num++;
4392 if (md->num == EXT4_BB_MAX_BLOCKS) {
4393 /* no more space, put full container on a sb's list */
4394 db->bb_md_cur = NULL;
4395 }
4396 }
4397 ext4_unlock_group(sb, group);
4398 return 0;
4399}
4400
4401/*
4402 * Main entry point into mballoc to free blocks
4403 */
4404void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
4405 unsigned long block, unsigned long count,
4406 int metadata, unsigned long *freed)
4407{
4408 struct buffer_head *bitmap_bh = 0;
4409 struct super_block *sb = inode->i_sb;
4410 struct ext4_allocation_context ac;
4411 struct ext4_group_desc *gdp;
4412 struct ext4_super_block *es;
4413 unsigned long overflow;
4414 ext4_grpblk_t bit;
4415 struct buffer_head *gd_bh;
4416 ext4_group_t block_group;
4417 struct ext4_sb_info *sbi;
4418 struct ext4_buddy e4b;
4419 int err = 0;
4420 int ret;
4421
4422 *freed = 0;
4423
4424 ext4_mb_poll_new_transaction(sb, handle);
4425
4426 sbi = EXT4_SB(sb);
4427 es = EXT4_SB(sb)->s_es;
4428 if (block < le32_to_cpu(es->s_first_data_block) ||
4429 block + count < block ||
4430 block + count > ext4_blocks_count(es)) {
4431 ext4_error(sb, __FUNCTION__,
4432 "Freeing blocks not in datazone - "
4433 "block = %lu, count = %lu", block, count);
4434 goto error_return;
4435 }
4436
4437 ext4_debug("freeing block %lu\n", block);
4438
4439 ac.ac_op = EXT4_MB_HISTORY_FREE;
4440 ac.ac_inode = inode;
4441 ac.ac_sb = sb;
4442
4443do_more:
4444 overflow = 0;
4445 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4446
4447 /*
4448 * Check to see if we are freeing blocks across a group
4449 * boundary.
4450 */
4451 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4452 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4453 count -= overflow;
4454 }
4455 bitmap_bh = read_block_bitmap(sb, block_group);
4456 if (!bitmap_bh)
4457 goto error_return;
4458 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4459 if (!gdp)
4460 goto error_return;
4461
4462 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4463 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4464 in_range(block, ext4_inode_table(sb, gdp),
4465 EXT4_SB(sb)->s_itb_per_group) ||
4466 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4467 EXT4_SB(sb)->s_itb_per_group)) {
4468
4469 ext4_error(sb, __FUNCTION__,
4470 "Freeing blocks in system zone - "
4471 "Block = %lu, count = %lu", block, count);
4472 }
4473
4474 BUFFER_TRACE(bitmap_bh, "getting write access");
4475 err = ext4_journal_get_write_access(handle, bitmap_bh);
4476 if (err)
4477 goto error_return;
4478
4479 /*
4480 * We are about to modify some metadata. Call the journal APIs
4481 * to unshare ->b_data if a currently-committing transaction is
4482 * using it
4483 */
4484 BUFFER_TRACE(gd_bh, "get_write_access");
4485 err = ext4_journal_get_write_access(handle, gd_bh);
4486 if (err)
4487 goto error_return;
4488
4489 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4490 if (err)
4491 goto error_return;
4492
4493#ifdef AGGRESSIVE_CHECK
4494 {
4495 int i;
4496 for (i = 0; i < count; i++)
4497 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4498 }
4499#endif
4500 mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
4501 bit, count);
4502
4503 /* We dirtied the bitmap block */
4504 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4505 err = ext4_journal_dirty_metadata(handle, bitmap_bh);
4506
4507 ac.ac_b_ex.fe_group = block_group;
4508 ac.ac_b_ex.fe_start = bit;
4509 ac.ac_b_ex.fe_len = count;
4510 ext4_mb_store_history(&ac);
4511
4512 if (metadata) {
4513 /* blocks being freed are metadata. these blocks shouldn't
4514 * be used until this transaction is committed */
4515 ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
4516 } else {
4517 ext4_lock_group(sb, block_group);
4518 err = mb_free_blocks(inode, &e4b, bit, count);
4519 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4520 ext4_unlock_group(sb, block_group);
4521 BUG_ON(err != 0);
4522 }
4523
4524 spin_lock(sb_bgl_lock(sbi, block_group));
4525 gdp->bg_free_blocks_count =
4526 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
4527 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4528 spin_unlock(sb_bgl_lock(sbi, block_group));
4529 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4530
4531 ext4_mb_release_desc(&e4b);
4532
4533 *freed += count;
4534
4535 /* And the group descriptor block */
4536 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4537 ret = ext4_journal_dirty_metadata(handle, gd_bh);
4538 if (!err)
4539 err = ret;
4540
4541 if (overflow && !err) {
4542 block += count;
4543 count = overflow;
4544 put_bh(bitmap_bh);
4545 goto do_more;
4546 }
4547 sb->s_dirt = 1;
4548error_return:
4549 brelse(bitmap_bh);
4550 ext4_std_error(sb, err);
4551 return;
4552}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-06 08:25:53 -0500