Linux 3.12.5
[linux/fpc-iii.git] / fs / ext4 / mballoc.c
bloba41e3ba8cfaa416b50f639e8e7b184f0439ab1fb
1 /*
2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
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.
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.
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-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
25 #include "mballoc.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly;
34 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
36 #endif
39 * MUSTDO:
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
43 * TODO v4:
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
47 * - quota
48 * - reservation for superuser
50 * TODO v3:
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
55 * - error handling
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
78 * represented as:
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 -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
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 makes sure that
88 * we have contiguous physical blocks representing the file blocks
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.
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 represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
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.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
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:
116 * { page }
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
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
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
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
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. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
167 * checked.
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
177 * - on-disk bitmap
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
182 * - inode
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
191 * - locality group
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
232 * block
234 * so, now we're building a concurrency table:
235 * - init buddy vs.
236 * - new PA
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
239 * - use inode PA
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
242 * - discard inode PA
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
248 * - new PA vs.
249 * - use inode PA
250 * i_data_sem serializes them
251 * - discard inode PA
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
257 * - use inode PA
258 * - use inode PA
259 * i_data_sem or another mutex should serializes them
260 * - discard inode PA
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
283 * - allocation:
284 * load group
285 * find blocks
286 * mark bits in on-disk bitmap
287 * release group
289 * - use preallocation:
290 * find proper PA (per-inode or group)
291 * load group
292 * mark bits in on-disk bitmap
293 * release group
294 * release PA
296 * - free:
297 * load group
298 * mark bits in on-disk bitmap
299 * release group
301 * - discard preallocations in group:
302 * mark PAs deleted
303 * move them onto local list
304 * load on-disk bitmap
305 * load group
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
313 * Locking rules
315 * Locks:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
318 * - per-pa lock (pa)
320 * Paths:
321 * - new pa
322 * object
323 * group
325 * - find and use pa:
326 * pa
328 * - release consumed pa:
329 * pa
330 * group
331 * object
333 * - generate in-core bitmap:
334 * group
335 * pa
337 * - discard all for given object (inode, locality group):
338 * object
339 * pa
340 * group
342 * - discard all for given group:
343 * group
344 * pa
345 * group
346 * object
349 static struct kmem_cache *ext4_pspace_cachep;
350 static struct kmem_cache *ext4_ac_cachep;
351 static struct kmem_cache *ext4_free_data_cachep;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
366 ext4_group_t group);
367 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
368 ext4_group_t group);
369 static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
372 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374 #if BITS_PER_LONG == 64
375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
380 #else
381 #error "how many bits you are?!"
382 #endif
383 return addr;
386 static inline int mb_test_bit(int bit, void *addr)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr = mb_correct_addr_and_bit(&bit, addr);
393 return ext4_test_bit(bit, addr);
396 static inline void mb_set_bit(int bit, void *addr)
398 addr = mb_correct_addr_and_bit(&bit, addr);
399 ext4_set_bit(bit, addr);
402 static inline void mb_clear_bit(int bit, void *addr)
404 addr = mb_correct_addr_and_bit(&bit, addr);
405 ext4_clear_bit(bit, addr);
408 static inline int mb_test_and_clear_bit(int bit, void *addr)
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
414 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416 int fix = 0, ret, tmpmax;
417 addr = mb_correct_addr_and_bit(&fix, addr);
418 tmpmax = max + fix;
419 start += fix;
421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
422 if (ret > max)
423 return max;
424 return ret;
427 static inline int mb_find_next_bit(void *addr, int max, int start)
429 int fix = 0, ret, tmpmax;
430 addr = mb_correct_addr_and_bit(&fix, addr);
431 tmpmax = max + fix;
432 start += fix;
434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
435 if (ret > max)
436 return max;
437 return ret;
440 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
442 char *bb;
444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
445 BUG_ON(max == NULL);
447 if (order > e4b->bd_blkbits + 1) {
448 *max = 0;
449 return NULL;
452 /* at order 0 we see each particular block */
453 if (order == 0) {
454 *max = 1 << (e4b->bd_blkbits + 3);
455 return e4b->bd_bitmap;
458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
461 return bb;
464 #ifdef DOUBLE_CHECK
465 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
468 int i;
469 struct super_block *sb = e4b->bd_sb;
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
472 return;
473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
480 ext4_grp_locked_error(sb, e4b->bd_group,
481 inode ? inode->i_ino : 0,
482 blocknr,
483 "freeing block already freed "
484 "(bit %u)",
485 first + i);
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
491 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
493 int i;
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
496 return;
497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
504 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
508 int i;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
516 "on disk/prealloc",
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
518 BUG();
524 #else
525 static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
528 return;
530 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
533 return;
535 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
537 return;
539 #endif
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
544 do { \
545 if (!(assert)) { \
546 printk(KERN_EMERG \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
549 BUG(); \
551 } while (0)
553 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
558 int max;
559 int max2;
560 int i;
561 int j;
562 int k;
563 int count;
564 struct ext4_group_info *grp;
565 int fragments = 0;
566 int fstart;
567 struct list_head *cur;
568 void *buddy;
569 void *buddy2;
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
574 return 0;
577 while (order > 1) {
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
585 count = 0;
586 for (i = 0; i < max; i++) {
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
591 MB_CHECK_ASSERT(
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
594 MB_CHECK_ASSERT(
595 mb_test_bit(i << 1, buddy2));
597 continue;
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
606 MB_CHECK_ASSERT(
607 !mb_test_bit(k, e4b->bd_bitmap));
609 count++;
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
612 order--;
615 fstart = -1;
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
620 if (fstart == -1) {
621 fragments++;
622 fstart = i;
624 continue;
626 fstart = -1;
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
630 k = i >> j;
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
638 grp = ext4_get_group_info(sb, e4b->bd_group);
639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
645 for (i = 0; i < pa->pa_len; i++)
646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
648 return 0;
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
653 #else
654 #define mb_check_buddy(e4b)
655 #endif
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block *sb,
664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
665 struct ext4_group_info *grp)
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
668 ext4_grpblk_t min;
669 ext4_grpblk_t max;
670 ext4_grpblk_t chunk;
671 unsigned short border;
673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
675 border = 2 << sb->s_blocksize_bits;
677 while (len > 0) {
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
681 /* find how many blocks of power 2 we need to mark */
682 min = fls(len) - 1;
684 if (max < min)
685 min = max;
686 chunk = 1 << min;
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
690 if (min > 0)
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
694 len -= chunk;
695 first += chunk;
700 * Cache the order of the largest free extent we have available in this block
701 * group.
703 static void
704 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
706 int i;
707 int bits;
709 grp->bb_largest_free_order = -1; /* uninit */
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
715 break;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block *sb,
722 void *buddy, void *bitmap, ext4_group_t group)
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
725 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
726 ext4_grpblk_t i = 0;
727 ext4_grpblk_t first;
728 ext4_grpblk_t len;
729 unsigned free = 0;
730 unsigned fragments = 0;
731 unsigned long long period = get_cycles();
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
735 i = mb_find_next_zero_bit(bitmap, max, 0);
736 grp->bb_first_free = i;
737 while (i < max) {
738 fragments++;
739 first = i;
740 i = mb_find_next_bit(bitmap, max, i);
741 len = i - first;
742 free += len;
743 if (len > 1)
744 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
745 else
746 grp->bb_counters[0]++;
747 if (i < max)
748 i = mb_find_next_zero_bit(bitmap, max, i);
750 grp->bb_fragments = fragments;
752 if (free != grp->bb_free) {
753 ext4_grp_locked_error(sb, group, 0, 0,
754 "%u clusters in bitmap, %u in gd; "
755 "block bitmap corrupt.",
756 free, grp->bb_free);
758 * If we intend to continue, we consider group descriptor
759 * corrupt and update bb_free using bitmap value
761 grp->bb_free = free;
762 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
764 mb_set_largest_free_order(sb, grp);
766 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
768 period = get_cycles() - period;
769 spin_lock(&EXT4_SB(sb)->s_bal_lock);
770 EXT4_SB(sb)->s_mb_buddies_generated++;
771 EXT4_SB(sb)->s_mb_generation_time += period;
772 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
775 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
777 int count;
778 int order = 1;
779 void *buddy;
781 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
782 ext4_set_bits(buddy, 0, count);
784 e4b->bd_info->bb_fragments = 0;
785 memset(e4b->bd_info->bb_counters, 0,
786 sizeof(*e4b->bd_info->bb_counters) *
787 (e4b->bd_sb->s_blocksize_bits + 2));
789 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
790 e4b->bd_bitmap, e4b->bd_group);
793 /* The buddy information is attached the buddy cache inode
794 * for convenience. The information regarding each group
795 * is loaded via ext4_mb_load_buddy. The information involve
796 * block bitmap and buddy information. The information are
797 * stored in the inode as
799 * { page }
800 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
803 * one block each for bitmap and buddy information.
804 * So for each group we take up 2 blocks. A page can
805 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
806 * So it can have information regarding groups_per_page which
807 * is blocks_per_page/2
809 * Locking note: This routine takes the block group lock of all groups
810 * for this page; do not hold this lock when calling this routine!
813 static int ext4_mb_init_cache(struct page *page, char *incore)
815 ext4_group_t ngroups;
816 int blocksize;
817 int blocks_per_page;
818 int groups_per_page;
819 int err = 0;
820 int i;
821 ext4_group_t first_group, group;
822 int first_block;
823 struct super_block *sb;
824 struct buffer_head *bhs;
825 struct buffer_head **bh = NULL;
826 struct inode *inode;
827 char *data;
828 char *bitmap;
829 struct ext4_group_info *grinfo;
831 mb_debug(1, "init page %lu\n", page->index);
833 inode = page->mapping->host;
834 sb = inode->i_sb;
835 ngroups = ext4_get_groups_count(sb);
836 blocksize = 1 << inode->i_blkbits;
837 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
839 groups_per_page = blocks_per_page >> 1;
840 if (groups_per_page == 0)
841 groups_per_page = 1;
843 /* allocate buffer_heads to read bitmaps */
844 if (groups_per_page > 1) {
845 i = sizeof(struct buffer_head *) * groups_per_page;
846 bh = kzalloc(i, GFP_NOFS);
847 if (bh == NULL) {
848 err = -ENOMEM;
849 goto out;
851 } else
852 bh = &bhs;
854 first_group = page->index * blocks_per_page / 2;
856 /* read all groups the page covers into the cache */
857 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
858 if (group >= ngroups)
859 break;
861 grinfo = ext4_get_group_info(sb, group);
863 * If page is uptodate then we came here after online resize
864 * which added some new uninitialized group info structs, so
865 * we must skip all initialized uptodate buddies on the page,
866 * which may be currently in use by an allocating task.
868 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
869 bh[i] = NULL;
870 continue;
872 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
873 err = -ENOMEM;
874 goto out;
876 mb_debug(1, "read bitmap for group %u\n", group);
879 /* wait for I/O completion */
880 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
881 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
882 err = -EIO;
883 goto out;
887 first_block = page->index * blocks_per_page;
888 for (i = 0; i < blocks_per_page; i++) {
889 group = (first_block + i) >> 1;
890 if (group >= ngroups)
891 break;
893 if (!bh[group - first_group])
894 /* skip initialized uptodate buddy */
895 continue;
898 * data carry information regarding this
899 * particular group in the format specified
900 * above
903 data = page_address(page) + (i * blocksize);
904 bitmap = bh[group - first_group]->b_data;
907 * We place the buddy block and bitmap block
908 * close together
910 if ((first_block + i) & 1) {
911 /* this is block of buddy */
912 BUG_ON(incore == NULL);
913 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
914 group, page->index, i * blocksize);
915 trace_ext4_mb_buddy_bitmap_load(sb, group);
916 grinfo = ext4_get_group_info(sb, group);
917 grinfo->bb_fragments = 0;
918 memset(grinfo->bb_counters, 0,
919 sizeof(*grinfo->bb_counters) *
920 (sb->s_blocksize_bits+2));
922 * incore got set to the group block bitmap below
924 ext4_lock_group(sb, group);
925 /* init the buddy */
926 memset(data, 0xff, blocksize);
927 ext4_mb_generate_buddy(sb, data, incore, group);
928 ext4_unlock_group(sb, group);
929 incore = NULL;
930 } else {
931 /* this is block of bitmap */
932 BUG_ON(incore != NULL);
933 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
934 group, page->index, i * blocksize);
935 trace_ext4_mb_bitmap_load(sb, group);
937 /* see comments in ext4_mb_put_pa() */
938 ext4_lock_group(sb, group);
939 memcpy(data, bitmap, blocksize);
941 /* mark all preallocated blks used in in-core bitmap */
942 ext4_mb_generate_from_pa(sb, data, group);
943 ext4_mb_generate_from_freelist(sb, data, group);
944 ext4_unlock_group(sb, group);
946 /* set incore so that the buddy information can be
947 * generated using this
949 incore = data;
952 SetPageUptodate(page);
954 out:
955 if (bh) {
956 for (i = 0; i < groups_per_page; i++)
957 brelse(bh[i]);
958 if (bh != &bhs)
959 kfree(bh);
961 return err;
965 * Lock the buddy and bitmap pages. This make sure other parallel init_group
966 * on the same buddy page doesn't happen whild holding the buddy page lock.
967 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
968 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
970 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
971 ext4_group_t group, struct ext4_buddy *e4b)
973 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
974 int block, pnum, poff;
975 int blocks_per_page;
976 struct page *page;
978 e4b->bd_buddy_page = NULL;
979 e4b->bd_bitmap_page = NULL;
981 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
983 * the buddy cache inode stores the block bitmap
984 * and buddy information in consecutive blocks.
985 * So for each group we need two blocks.
987 block = group * 2;
988 pnum = block / blocks_per_page;
989 poff = block % blocks_per_page;
990 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
991 if (!page)
992 return -EIO;
993 BUG_ON(page->mapping != inode->i_mapping);
994 e4b->bd_bitmap_page = page;
995 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
997 if (blocks_per_page >= 2) {
998 /* buddy and bitmap are on the same page */
999 return 0;
1002 block++;
1003 pnum = block / blocks_per_page;
1004 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1005 if (!page)
1006 return -EIO;
1007 BUG_ON(page->mapping != inode->i_mapping);
1008 e4b->bd_buddy_page = page;
1009 return 0;
1012 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1014 if (e4b->bd_bitmap_page) {
1015 unlock_page(e4b->bd_bitmap_page);
1016 page_cache_release(e4b->bd_bitmap_page);
1018 if (e4b->bd_buddy_page) {
1019 unlock_page(e4b->bd_buddy_page);
1020 page_cache_release(e4b->bd_buddy_page);
1025 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1026 * block group lock of all groups for this page; do not hold the BG lock when
1027 * calling this routine!
1029 static noinline_for_stack
1030 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1033 struct ext4_group_info *this_grp;
1034 struct ext4_buddy e4b;
1035 struct page *page;
1036 int ret = 0;
1038 might_sleep();
1039 mb_debug(1, "init group %u\n", group);
1040 this_grp = ext4_get_group_info(sb, group);
1042 * This ensures that we don't reinit the buddy cache
1043 * page which map to the group from which we are already
1044 * allocating. If we are looking at the buddy cache we would
1045 * have taken a reference using ext4_mb_load_buddy and that
1046 * would have pinned buddy page to page cache.
1048 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1049 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1051 * somebody initialized the group
1052 * return without doing anything
1054 goto err;
1057 page = e4b.bd_bitmap_page;
1058 ret = ext4_mb_init_cache(page, NULL);
1059 if (ret)
1060 goto err;
1061 if (!PageUptodate(page)) {
1062 ret = -EIO;
1063 goto err;
1065 mark_page_accessed(page);
1067 if (e4b.bd_buddy_page == NULL) {
1069 * If both the bitmap and buddy are in
1070 * the same page we don't need to force
1071 * init the buddy
1073 ret = 0;
1074 goto err;
1076 /* init buddy cache */
1077 page = e4b.bd_buddy_page;
1078 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1079 if (ret)
1080 goto err;
1081 if (!PageUptodate(page)) {
1082 ret = -EIO;
1083 goto err;
1085 mark_page_accessed(page);
1086 err:
1087 ext4_mb_put_buddy_page_lock(&e4b);
1088 return ret;
1092 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1093 * block group lock of all groups for this page; do not hold the BG lock when
1094 * calling this routine!
1096 static noinline_for_stack int
1097 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1098 struct ext4_buddy *e4b)
1100 int blocks_per_page;
1101 int block;
1102 int pnum;
1103 int poff;
1104 struct page *page;
1105 int ret;
1106 struct ext4_group_info *grp;
1107 struct ext4_sb_info *sbi = EXT4_SB(sb);
1108 struct inode *inode = sbi->s_buddy_cache;
1110 might_sleep();
1111 mb_debug(1, "load group %u\n", group);
1113 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1114 grp = ext4_get_group_info(sb, group);
1116 e4b->bd_blkbits = sb->s_blocksize_bits;
1117 e4b->bd_info = grp;
1118 e4b->bd_sb = sb;
1119 e4b->bd_group = group;
1120 e4b->bd_buddy_page = NULL;
1121 e4b->bd_bitmap_page = NULL;
1123 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1125 * we need full data about the group
1126 * to make a good selection
1128 ret = ext4_mb_init_group(sb, group);
1129 if (ret)
1130 return ret;
1134 * the buddy cache inode stores the block bitmap
1135 * and buddy information in consecutive blocks.
1136 * So for each group we need two blocks.
1138 block = group * 2;
1139 pnum = block / blocks_per_page;
1140 poff = block % blocks_per_page;
1142 /* we could use find_or_create_page(), but it locks page
1143 * what we'd like to avoid in fast path ... */
1144 page = find_get_page(inode->i_mapping, pnum);
1145 if (page == NULL || !PageUptodate(page)) {
1146 if (page)
1148 * drop the page reference and try
1149 * to get the page with lock. If we
1150 * are not uptodate that implies
1151 * somebody just created the page but
1152 * is yet to initialize the same. So
1153 * wait for it to initialize.
1155 page_cache_release(page);
1156 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1157 if (page) {
1158 BUG_ON(page->mapping != inode->i_mapping);
1159 if (!PageUptodate(page)) {
1160 ret = ext4_mb_init_cache(page, NULL);
1161 if (ret) {
1162 unlock_page(page);
1163 goto err;
1165 mb_cmp_bitmaps(e4b, page_address(page) +
1166 (poff * sb->s_blocksize));
1168 unlock_page(page);
1171 if (page == NULL || !PageUptodate(page)) {
1172 ret = -EIO;
1173 goto err;
1175 e4b->bd_bitmap_page = page;
1176 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1177 mark_page_accessed(page);
1179 block++;
1180 pnum = block / blocks_per_page;
1181 poff = block % blocks_per_page;
1183 page = find_get_page(inode->i_mapping, pnum);
1184 if (page == NULL || !PageUptodate(page)) {
1185 if (page)
1186 page_cache_release(page);
1187 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1188 if (page) {
1189 BUG_ON(page->mapping != inode->i_mapping);
1190 if (!PageUptodate(page)) {
1191 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1192 if (ret) {
1193 unlock_page(page);
1194 goto err;
1197 unlock_page(page);
1200 if (page == NULL || !PageUptodate(page)) {
1201 ret = -EIO;
1202 goto err;
1204 e4b->bd_buddy_page = page;
1205 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1206 mark_page_accessed(page);
1208 BUG_ON(e4b->bd_bitmap_page == NULL);
1209 BUG_ON(e4b->bd_buddy_page == NULL);
1211 return 0;
1213 err:
1214 if (page)
1215 page_cache_release(page);
1216 if (e4b->bd_bitmap_page)
1217 page_cache_release(e4b->bd_bitmap_page);
1218 if (e4b->bd_buddy_page)
1219 page_cache_release(e4b->bd_buddy_page);
1220 e4b->bd_buddy = NULL;
1221 e4b->bd_bitmap = NULL;
1222 return ret;
1225 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1227 if (e4b->bd_bitmap_page)
1228 page_cache_release(e4b->bd_bitmap_page);
1229 if (e4b->bd_buddy_page)
1230 page_cache_release(e4b->bd_buddy_page);
1234 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1236 int order = 1;
1237 void *bb;
1239 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1240 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1242 bb = e4b->bd_buddy;
1243 while (order <= e4b->bd_blkbits + 1) {
1244 block = block >> 1;
1245 if (!mb_test_bit(block, bb)) {
1246 /* this block is part of buddy of order 'order' */
1247 return order;
1249 bb += 1 << (e4b->bd_blkbits - order);
1250 order++;
1252 return 0;
1255 static void mb_clear_bits(void *bm, int cur, int len)
1257 __u32 *addr;
1259 len = cur + len;
1260 while (cur < len) {
1261 if ((cur & 31) == 0 && (len - cur) >= 32) {
1262 /* fast path: clear whole word at once */
1263 addr = bm + (cur >> 3);
1264 *addr = 0;
1265 cur += 32;
1266 continue;
1268 mb_clear_bit(cur, bm);
1269 cur++;
1273 /* clear bits in given range
1274 * will return first found zero bit if any, -1 otherwise
1276 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1278 __u32 *addr;
1279 int zero_bit = -1;
1281 len = cur + len;
1282 while (cur < len) {
1283 if ((cur & 31) == 0 && (len - cur) >= 32) {
1284 /* fast path: clear whole word at once */
1285 addr = bm + (cur >> 3);
1286 if (*addr != (__u32)(-1) && zero_bit == -1)
1287 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1288 *addr = 0;
1289 cur += 32;
1290 continue;
1292 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1293 zero_bit = cur;
1294 cur++;
1297 return zero_bit;
1300 void ext4_set_bits(void *bm, int cur, int len)
1302 __u32 *addr;
1304 len = cur + len;
1305 while (cur < len) {
1306 if ((cur & 31) == 0 && (len - cur) >= 32) {
1307 /* fast path: set whole word at once */
1308 addr = bm + (cur >> 3);
1309 *addr = 0xffffffff;
1310 cur += 32;
1311 continue;
1313 mb_set_bit(cur, bm);
1314 cur++;
1319 * _________________________________________________________________ */
1321 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1323 if (mb_test_bit(*bit + side, bitmap)) {
1324 mb_clear_bit(*bit, bitmap);
1325 (*bit) -= side;
1326 return 1;
1328 else {
1329 (*bit) += side;
1330 mb_set_bit(*bit, bitmap);
1331 return -1;
1335 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1337 int max;
1338 int order = 1;
1339 void *buddy = mb_find_buddy(e4b, order, &max);
1341 while (buddy) {
1342 void *buddy2;
1344 /* Bits in range [first; last] are known to be set since
1345 * corresponding blocks were allocated. Bits in range
1346 * (first; last) will stay set because they form buddies on
1347 * upper layer. We just deal with borders if they don't
1348 * align with upper layer and then go up.
1349 * Releasing entire group is all about clearing
1350 * single bit of highest order buddy.
1353 /* Example:
1354 * ---------------------------------
1355 * | 1 | 1 | 1 | 1 |
1356 * ---------------------------------
1357 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1358 * ---------------------------------
1359 * 0 1 2 3 4 5 6 7
1360 * \_____________________/
1362 * Neither [1] nor [6] is aligned to above layer.
1363 * Left neighbour [0] is free, so mark it busy,
1364 * decrease bb_counters and extend range to
1365 * [0; 6]
1366 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1367 * mark [6] free, increase bb_counters and shrink range to
1368 * [0; 5].
1369 * Then shift range to [0; 2], go up and do the same.
1373 if (first & 1)
1374 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1375 if (!(last & 1))
1376 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1377 if (first > last)
1378 break;
1379 order++;
1381 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1382 mb_clear_bits(buddy, first, last - first + 1);
1383 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1384 break;
1386 first >>= 1;
1387 last >>= 1;
1388 buddy = buddy2;
1392 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1393 int first, int count)
1395 int left_is_free = 0;
1396 int right_is_free = 0;
1397 int block;
1398 int last = first + count - 1;
1399 struct super_block *sb = e4b->bd_sb;
1401 BUG_ON(last >= (sb->s_blocksize << 3));
1402 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1403 /* Don't bother if the block group is corrupt. */
1404 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1405 return;
1407 mb_check_buddy(e4b);
1408 mb_free_blocks_double(inode, e4b, first, count);
1410 e4b->bd_info->bb_free += count;
1411 if (first < e4b->bd_info->bb_first_free)
1412 e4b->bd_info->bb_first_free = first;
1414 /* access memory sequentially: check left neighbour,
1415 * clear range and then check right neighbour
1417 if (first != 0)
1418 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1419 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1420 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1421 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1423 if (unlikely(block != -1)) {
1424 ext4_fsblk_t blocknr;
1426 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1427 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1428 ext4_grp_locked_error(sb, e4b->bd_group,
1429 inode ? inode->i_ino : 0,
1430 blocknr,
1431 "freeing already freed block "
1432 "(bit %u); block bitmap corrupt.",
1433 block);
1434 /* Mark the block group as corrupt. */
1435 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1436 &e4b->bd_info->bb_state);
1437 mb_regenerate_buddy(e4b);
1438 goto done;
1441 /* let's maintain fragments counter */
1442 if (left_is_free && right_is_free)
1443 e4b->bd_info->bb_fragments--;
1444 else if (!left_is_free && !right_is_free)
1445 e4b->bd_info->bb_fragments++;
1447 /* buddy[0] == bd_bitmap is a special case, so handle
1448 * it right away and let mb_buddy_mark_free stay free of
1449 * zero order checks.
1450 * Check if neighbours are to be coaleasced,
1451 * adjust bitmap bb_counters and borders appropriately.
1453 if (first & 1) {
1454 first += !left_is_free;
1455 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1457 if (!(last & 1)) {
1458 last -= !right_is_free;
1459 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1462 if (first <= last)
1463 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1465 done:
1466 mb_set_largest_free_order(sb, e4b->bd_info);
1467 mb_check_buddy(e4b);
1470 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1471 int needed, struct ext4_free_extent *ex)
1473 int next = block;
1474 int max, order;
1475 void *buddy;
1477 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1478 BUG_ON(ex == NULL);
1480 buddy = mb_find_buddy(e4b, 0, &max);
1481 BUG_ON(buddy == NULL);
1482 BUG_ON(block >= max);
1483 if (mb_test_bit(block, buddy)) {
1484 ex->fe_len = 0;
1485 ex->fe_start = 0;
1486 ex->fe_group = 0;
1487 return 0;
1490 /* find actual order */
1491 order = mb_find_order_for_block(e4b, block);
1492 block = block >> order;
1494 ex->fe_len = 1 << order;
1495 ex->fe_start = block << order;
1496 ex->fe_group = e4b->bd_group;
1498 /* calc difference from given start */
1499 next = next - ex->fe_start;
1500 ex->fe_len -= next;
1501 ex->fe_start += next;
1503 while (needed > ex->fe_len &&
1504 mb_find_buddy(e4b, order, &max)) {
1506 if (block + 1 >= max)
1507 break;
1509 next = (block + 1) * (1 << order);
1510 if (mb_test_bit(next, e4b->bd_bitmap))
1511 break;
1513 order = mb_find_order_for_block(e4b, next);
1515 block = next >> order;
1516 ex->fe_len += 1 << order;
1519 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1520 return ex->fe_len;
1523 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1525 int ord;
1526 int mlen = 0;
1527 int max = 0;
1528 int cur;
1529 int start = ex->fe_start;
1530 int len = ex->fe_len;
1531 unsigned ret = 0;
1532 int len0 = len;
1533 void *buddy;
1535 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1536 BUG_ON(e4b->bd_group != ex->fe_group);
1537 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1538 mb_check_buddy(e4b);
1539 mb_mark_used_double(e4b, start, len);
1541 e4b->bd_info->bb_free -= len;
1542 if (e4b->bd_info->bb_first_free == start)
1543 e4b->bd_info->bb_first_free += len;
1545 /* let's maintain fragments counter */
1546 if (start != 0)
1547 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1548 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1549 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1550 if (mlen && max)
1551 e4b->bd_info->bb_fragments++;
1552 else if (!mlen && !max)
1553 e4b->bd_info->bb_fragments--;
1555 /* let's maintain buddy itself */
1556 while (len) {
1557 ord = mb_find_order_for_block(e4b, start);
1559 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1560 /* the whole chunk may be allocated at once! */
1561 mlen = 1 << ord;
1562 buddy = mb_find_buddy(e4b, ord, &max);
1563 BUG_ON((start >> ord) >= max);
1564 mb_set_bit(start >> ord, buddy);
1565 e4b->bd_info->bb_counters[ord]--;
1566 start += mlen;
1567 len -= mlen;
1568 BUG_ON(len < 0);
1569 continue;
1572 /* store for history */
1573 if (ret == 0)
1574 ret = len | (ord << 16);
1576 /* we have to split large buddy */
1577 BUG_ON(ord <= 0);
1578 buddy = mb_find_buddy(e4b, ord, &max);
1579 mb_set_bit(start >> ord, buddy);
1580 e4b->bd_info->bb_counters[ord]--;
1582 ord--;
1583 cur = (start >> ord) & ~1U;
1584 buddy = mb_find_buddy(e4b, ord, &max);
1585 mb_clear_bit(cur, buddy);
1586 mb_clear_bit(cur + 1, buddy);
1587 e4b->bd_info->bb_counters[ord]++;
1588 e4b->bd_info->bb_counters[ord]++;
1590 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1592 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1593 mb_check_buddy(e4b);
1595 return ret;
1599 * Must be called under group lock!
1601 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1602 struct ext4_buddy *e4b)
1604 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1605 int ret;
1607 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1608 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1610 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1611 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1612 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1614 /* preallocation can change ac_b_ex, thus we store actually
1615 * allocated blocks for history */
1616 ac->ac_f_ex = ac->ac_b_ex;
1618 ac->ac_status = AC_STATUS_FOUND;
1619 ac->ac_tail = ret & 0xffff;
1620 ac->ac_buddy = ret >> 16;
1623 * take the page reference. We want the page to be pinned
1624 * so that we don't get a ext4_mb_init_cache_call for this
1625 * group until we update the bitmap. That would mean we
1626 * double allocate blocks. The reference is dropped
1627 * in ext4_mb_release_context
1629 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1630 get_page(ac->ac_bitmap_page);
1631 ac->ac_buddy_page = e4b->bd_buddy_page;
1632 get_page(ac->ac_buddy_page);
1633 /* store last allocated for subsequent stream allocation */
1634 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1635 spin_lock(&sbi->s_md_lock);
1636 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1637 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1638 spin_unlock(&sbi->s_md_lock);
1643 * regular allocator, for general purposes allocation
1646 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1647 struct ext4_buddy *e4b,
1648 int finish_group)
1650 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1651 struct ext4_free_extent *bex = &ac->ac_b_ex;
1652 struct ext4_free_extent *gex = &ac->ac_g_ex;
1653 struct ext4_free_extent ex;
1654 int max;
1656 if (ac->ac_status == AC_STATUS_FOUND)
1657 return;
1659 * We don't want to scan for a whole year
1661 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1662 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1663 ac->ac_status = AC_STATUS_BREAK;
1664 return;
1668 * Haven't found good chunk so far, let's continue
1670 if (bex->fe_len < gex->fe_len)
1671 return;
1673 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1674 && bex->fe_group == e4b->bd_group) {
1675 /* recheck chunk's availability - we don't know
1676 * when it was found (within this lock-unlock
1677 * period or not) */
1678 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1679 if (max >= gex->fe_len) {
1680 ext4_mb_use_best_found(ac, e4b);
1681 return;
1687 * The routine checks whether found extent is good enough. If it is,
1688 * then the extent gets marked used and flag is set to the context
1689 * to stop scanning. Otherwise, the extent is compared with the
1690 * previous found extent and if new one is better, then it's stored
1691 * in the context. Later, the best found extent will be used, if
1692 * mballoc can't find good enough extent.
1694 * FIXME: real allocation policy is to be designed yet!
1696 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1697 struct ext4_free_extent *ex,
1698 struct ext4_buddy *e4b)
1700 struct ext4_free_extent *bex = &ac->ac_b_ex;
1701 struct ext4_free_extent *gex = &ac->ac_g_ex;
1703 BUG_ON(ex->fe_len <= 0);
1704 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1705 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1706 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1708 ac->ac_found++;
1711 * The special case - take what you catch first
1713 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1714 *bex = *ex;
1715 ext4_mb_use_best_found(ac, e4b);
1716 return;
1720 * Let's check whether the chuck is good enough
1722 if (ex->fe_len == gex->fe_len) {
1723 *bex = *ex;
1724 ext4_mb_use_best_found(ac, e4b);
1725 return;
1729 * If this is first found extent, just store it in the context
1731 if (bex->fe_len == 0) {
1732 *bex = *ex;
1733 return;
1737 * If new found extent is better, store it in the context
1739 if (bex->fe_len < gex->fe_len) {
1740 /* if the request isn't satisfied, any found extent
1741 * larger than previous best one is better */
1742 if (ex->fe_len > bex->fe_len)
1743 *bex = *ex;
1744 } else if (ex->fe_len > gex->fe_len) {
1745 /* if the request is satisfied, then we try to find
1746 * an extent that still satisfy the request, but is
1747 * smaller than previous one */
1748 if (ex->fe_len < bex->fe_len)
1749 *bex = *ex;
1752 ext4_mb_check_limits(ac, e4b, 0);
1755 static noinline_for_stack
1756 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1757 struct ext4_buddy *e4b)
1759 struct ext4_free_extent ex = ac->ac_b_ex;
1760 ext4_group_t group = ex.fe_group;
1761 int max;
1762 int err;
1764 BUG_ON(ex.fe_len <= 0);
1765 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1766 if (err)
1767 return err;
1769 ext4_lock_group(ac->ac_sb, group);
1770 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1772 if (max > 0) {
1773 ac->ac_b_ex = ex;
1774 ext4_mb_use_best_found(ac, e4b);
1777 ext4_unlock_group(ac->ac_sb, group);
1778 ext4_mb_unload_buddy(e4b);
1780 return 0;
1783 static noinline_for_stack
1784 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1785 struct ext4_buddy *e4b)
1787 ext4_group_t group = ac->ac_g_ex.fe_group;
1788 int max;
1789 int err;
1790 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1791 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1792 struct ext4_free_extent ex;
1794 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1795 return 0;
1796 if (grp->bb_free == 0)
1797 return 0;
1799 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1800 if (err)
1801 return err;
1803 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1804 ext4_mb_unload_buddy(e4b);
1805 return 0;
1808 ext4_lock_group(ac->ac_sb, group);
1809 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1810 ac->ac_g_ex.fe_len, &ex);
1812 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1813 ext4_fsblk_t start;
1815 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1816 ex.fe_start;
1817 /* use do_div to get remainder (would be 64-bit modulo) */
1818 if (do_div(start, sbi->s_stripe) == 0) {
1819 ac->ac_found++;
1820 ac->ac_b_ex = ex;
1821 ext4_mb_use_best_found(ac, e4b);
1823 } else if (max >= ac->ac_g_ex.fe_len) {
1824 BUG_ON(ex.fe_len <= 0);
1825 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1826 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1827 ac->ac_found++;
1828 ac->ac_b_ex = ex;
1829 ext4_mb_use_best_found(ac, e4b);
1830 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1831 /* Sometimes, caller may want to merge even small
1832 * number of blocks to an existing extent */
1833 BUG_ON(ex.fe_len <= 0);
1834 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1835 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1836 ac->ac_found++;
1837 ac->ac_b_ex = ex;
1838 ext4_mb_use_best_found(ac, e4b);
1840 ext4_unlock_group(ac->ac_sb, group);
1841 ext4_mb_unload_buddy(e4b);
1843 return 0;
1847 * The routine scans buddy structures (not bitmap!) from given order
1848 * to max order and tries to find big enough chunk to satisfy the req
1850 static noinline_for_stack
1851 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1852 struct ext4_buddy *e4b)
1854 struct super_block *sb = ac->ac_sb;
1855 struct ext4_group_info *grp = e4b->bd_info;
1856 void *buddy;
1857 int i;
1858 int k;
1859 int max;
1861 BUG_ON(ac->ac_2order <= 0);
1862 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1863 if (grp->bb_counters[i] == 0)
1864 continue;
1866 buddy = mb_find_buddy(e4b, i, &max);
1867 BUG_ON(buddy == NULL);
1869 k = mb_find_next_zero_bit(buddy, max, 0);
1870 BUG_ON(k >= max);
1872 ac->ac_found++;
1874 ac->ac_b_ex.fe_len = 1 << i;
1875 ac->ac_b_ex.fe_start = k << i;
1876 ac->ac_b_ex.fe_group = e4b->bd_group;
1878 ext4_mb_use_best_found(ac, e4b);
1880 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1882 if (EXT4_SB(sb)->s_mb_stats)
1883 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1885 break;
1890 * The routine scans the group and measures all found extents.
1891 * In order to optimize scanning, caller must pass number of
1892 * free blocks in the group, so the routine can know upper limit.
1894 static noinline_for_stack
1895 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1896 struct ext4_buddy *e4b)
1898 struct super_block *sb = ac->ac_sb;
1899 void *bitmap = e4b->bd_bitmap;
1900 struct ext4_free_extent ex;
1901 int i;
1902 int free;
1904 free = e4b->bd_info->bb_free;
1905 BUG_ON(free <= 0);
1907 i = e4b->bd_info->bb_first_free;
1909 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1910 i = mb_find_next_zero_bit(bitmap,
1911 EXT4_CLUSTERS_PER_GROUP(sb), i);
1912 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1914 * IF we have corrupt bitmap, we won't find any
1915 * free blocks even though group info says we
1916 * we have free blocks
1918 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1919 "%d free clusters as per "
1920 "group info. But bitmap says 0",
1921 free);
1922 break;
1925 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1926 BUG_ON(ex.fe_len <= 0);
1927 if (free < ex.fe_len) {
1928 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1929 "%d free clusters as per "
1930 "group info. But got %d blocks",
1931 free, ex.fe_len);
1933 * The number of free blocks differs. This mostly
1934 * indicate that the bitmap is corrupt. So exit
1935 * without claiming the space.
1937 break;
1940 ext4_mb_measure_extent(ac, &ex, e4b);
1942 i += ex.fe_len;
1943 free -= ex.fe_len;
1946 ext4_mb_check_limits(ac, e4b, 1);
1950 * This is a special case for storages like raid5
1951 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1953 static noinline_for_stack
1954 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1955 struct ext4_buddy *e4b)
1957 struct super_block *sb = ac->ac_sb;
1958 struct ext4_sb_info *sbi = EXT4_SB(sb);
1959 void *bitmap = e4b->bd_bitmap;
1960 struct ext4_free_extent ex;
1961 ext4_fsblk_t first_group_block;
1962 ext4_fsblk_t a;
1963 ext4_grpblk_t i;
1964 int max;
1966 BUG_ON(sbi->s_stripe == 0);
1968 /* find first stripe-aligned block in group */
1969 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1971 a = first_group_block + sbi->s_stripe - 1;
1972 do_div(a, sbi->s_stripe);
1973 i = (a * sbi->s_stripe) - first_group_block;
1975 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1976 if (!mb_test_bit(i, bitmap)) {
1977 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1978 if (max >= sbi->s_stripe) {
1979 ac->ac_found++;
1980 ac->ac_b_ex = ex;
1981 ext4_mb_use_best_found(ac, e4b);
1982 break;
1985 i += sbi->s_stripe;
1989 /* This is now called BEFORE we load the buddy bitmap. */
1990 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1991 ext4_group_t group, int cr)
1993 unsigned free, fragments;
1994 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1995 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1997 BUG_ON(cr < 0 || cr >= 4);
1999 free = grp->bb_free;
2000 if (free == 0)
2001 return 0;
2002 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2003 return 0;
2005 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2006 return 0;
2008 /* We only do this if the grp has never been initialized */
2009 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2010 int ret = ext4_mb_init_group(ac->ac_sb, group);
2011 if (ret)
2012 return 0;
2015 fragments = grp->bb_fragments;
2016 if (fragments == 0)
2017 return 0;
2019 switch (cr) {
2020 case 0:
2021 BUG_ON(ac->ac_2order == 0);
2023 /* Avoid using the first bg of a flexgroup for data files */
2024 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2025 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2026 ((group % flex_size) == 0))
2027 return 0;
2029 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2030 (free / fragments) >= ac->ac_g_ex.fe_len)
2031 return 1;
2033 if (grp->bb_largest_free_order < ac->ac_2order)
2034 return 0;
2036 return 1;
2037 case 1:
2038 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2039 return 1;
2040 break;
2041 case 2:
2042 if (free >= ac->ac_g_ex.fe_len)
2043 return 1;
2044 break;
2045 case 3:
2046 return 1;
2047 default:
2048 BUG();
2051 return 0;
2054 static noinline_for_stack int
2055 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2057 ext4_group_t ngroups, group, i;
2058 int cr;
2059 int err = 0;
2060 struct ext4_sb_info *sbi;
2061 struct super_block *sb;
2062 struct ext4_buddy e4b;
2064 sb = ac->ac_sb;
2065 sbi = EXT4_SB(sb);
2066 ngroups = ext4_get_groups_count(sb);
2067 /* non-extent files are limited to low blocks/groups */
2068 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2069 ngroups = sbi->s_blockfile_groups;
2071 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2073 /* first, try the goal */
2074 err = ext4_mb_find_by_goal(ac, &e4b);
2075 if (err || ac->ac_status == AC_STATUS_FOUND)
2076 goto out;
2078 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2079 goto out;
2082 * ac->ac2_order is set only if the fe_len is a power of 2
2083 * if ac2_order is set we also set criteria to 0 so that we
2084 * try exact allocation using buddy.
2086 i = fls(ac->ac_g_ex.fe_len);
2087 ac->ac_2order = 0;
2089 * We search using buddy data only if the order of the request
2090 * is greater than equal to the sbi_s_mb_order2_reqs
2091 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2093 if (i >= sbi->s_mb_order2_reqs) {
2095 * This should tell if fe_len is exactly power of 2
2097 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2098 ac->ac_2order = i - 1;
2101 /* if stream allocation is enabled, use global goal */
2102 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2103 /* TBD: may be hot point */
2104 spin_lock(&sbi->s_md_lock);
2105 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2106 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2107 spin_unlock(&sbi->s_md_lock);
2110 /* Let's just scan groups to find more-less suitable blocks */
2111 cr = ac->ac_2order ? 0 : 1;
2113 * cr == 0 try to get exact allocation,
2114 * cr == 3 try to get anything
2116 repeat:
2117 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2118 ac->ac_criteria = cr;
2120 * searching for the right group start
2121 * from the goal value specified
2123 group = ac->ac_g_ex.fe_group;
2125 for (i = 0; i < ngroups; group++, i++) {
2126 cond_resched();
2128 * Artificially restricted ngroups for non-extent
2129 * files makes group > ngroups possible on first loop.
2131 if (group >= ngroups)
2132 group = 0;
2134 /* This now checks without needing the buddy page */
2135 if (!ext4_mb_good_group(ac, group, cr))
2136 continue;
2138 err = ext4_mb_load_buddy(sb, group, &e4b);
2139 if (err)
2140 goto out;
2142 ext4_lock_group(sb, group);
2145 * We need to check again after locking the
2146 * block group
2148 if (!ext4_mb_good_group(ac, group, cr)) {
2149 ext4_unlock_group(sb, group);
2150 ext4_mb_unload_buddy(&e4b);
2151 continue;
2154 ac->ac_groups_scanned++;
2155 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2156 ext4_mb_simple_scan_group(ac, &e4b);
2157 else if (cr == 1 && sbi->s_stripe &&
2158 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2159 ext4_mb_scan_aligned(ac, &e4b);
2160 else
2161 ext4_mb_complex_scan_group(ac, &e4b);
2163 ext4_unlock_group(sb, group);
2164 ext4_mb_unload_buddy(&e4b);
2166 if (ac->ac_status != AC_STATUS_CONTINUE)
2167 break;
2171 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2172 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2174 * We've been searching too long. Let's try to allocate
2175 * the best chunk we've found so far
2178 ext4_mb_try_best_found(ac, &e4b);
2179 if (ac->ac_status != AC_STATUS_FOUND) {
2181 * Someone more lucky has already allocated it.
2182 * The only thing we can do is just take first
2183 * found block(s)
2184 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2186 ac->ac_b_ex.fe_group = 0;
2187 ac->ac_b_ex.fe_start = 0;
2188 ac->ac_b_ex.fe_len = 0;
2189 ac->ac_status = AC_STATUS_CONTINUE;
2190 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2191 cr = 3;
2192 atomic_inc(&sbi->s_mb_lost_chunks);
2193 goto repeat;
2196 out:
2197 return err;
2200 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2202 struct super_block *sb = seq->private;
2203 ext4_group_t group;
2205 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2206 return NULL;
2207 group = *pos + 1;
2208 return (void *) ((unsigned long) group);
2211 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2213 struct super_block *sb = seq->private;
2214 ext4_group_t group;
2216 ++*pos;
2217 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2218 return NULL;
2219 group = *pos + 1;
2220 return (void *) ((unsigned long) group);
2223 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2225 struct super_block *sb = seq->private;
2226 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2227 int i;
2228 int err, buddy_loaded = 0;
2229 struct ext4_buddy e4b;
2230 struct ext4_group_info *grinfo;
2231 struct sg {
2232 struct ext4_group_info info;
2233 ext4_grpblk_t counters[16];
2234 } sg;
2236 group--;
2237 if (group == 0)
2238 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2239 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2240 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2241 "group", "free", "frags", "first",
2242 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2243 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2245 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2246 sizeof(struct ext4_group_info);
2247 grinfo = ext4_get_group_info(sb, group);
2248 /* Load the group info in memory only if not already loaded. */
2249 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2250 err = ext4_mb_load_buddy(sb, group, &e4b);
2251 if (err) {
2252 seq_printf(seq, "#%-5u: I/O error\n", group);
2253 return 0;
2255 buddy_loaded = 1;
2258 memcpy(&sg, ext4_get_group_info(sb, group), i);
2260 if (buddy_loaded)
2261 ext4_mb_unload_buddy(&e4b);
2263 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2264 sg.info.bb_fragments, sg.info.bb_first_free);
2265 for (i = 0; i <= 13; i++)
2266 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2267 sg.info.bb_counters[i] : 0);
2268 seq_printf(seq, " ]\n");
2270 return 0;
2273 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2277 static const struct seq_operations ext4_mb_seq_groups_ops = {
2278 .start = ext4_mb_seq_groups_start,
2279 .next = ext4_mb_seq_groups_next,
2280 .stop = ext4_mb_seq_groups_stop,
2281 .show = ext4_mb_seq_groups_show,
2284 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2286 struct super_block *sb = PDE_DATA(inode);
2287 int rc;
2289 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2290 if (rc == 0) {
2291 struct seq_file *m = file->private_data;
2292 m->private = sb;
2294 return rc;
2298 static const struct file_operations ext4_mb_seq_groups_fops = {
2299 .owner = THIS_MODULE,
2300 .open = ext4_mb_seq_groups_open,
2301 .read = seq_read,
2302 .llseek = seq_lseek,
2303 .release = seq_release,
2306 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2308 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2309 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2311 BUG_ON(!cachep);
2312 return cachep;
2316 * Allocate the top-level s_group_info array for the specified number
2317 * of groups
2319 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2321 struct ext4_sb_info *sbi = EXT4_SB(sb);
2322 unsigned size;
2323 struct ext4_group_info ***new_groupinfo;
2325 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2326 EXT4_DESC_PER_BLOCK_BITS(sb);
2327 if (size <= sbi->s_group_info_size)
2328 return 0;
2330 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2331 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2332 if (!new_groupinfo) {
2333 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2334 return -ENOMEM;
2336 if (sbi->s_group_info) {
2337 memcpy(new_groupinfo, sbi->s_group_info,
2338 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2339 ext4_kvfree(sbi->s_group_info);
2341 sbi->s_group_info = new_groupinfo;
2342 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2343 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2344 sbi->s_group_info_size);
2345 return 0;
2348 /* Create and initialize ext4_group_info data for the given group. */
2349 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2350 struct ext4_group_desc *desc)
2352 int i;
2353 int metalen = 0;
2354 struct ext4_sb_info *sbi = EXT4_SB(sb);
2355 struct ext4_group_info **meta_group_info;
2356 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2359 * First check if this group is the first of a reserved block.
2360 * If it's true, we have to allocate a new table of pointers
2361 * to ext4_group_info structures
2363 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2364 metalen = sizeof(*meta_group_info) <<
2365 EXT4_DESC_PER_BLOCK_BITS(sb);
2366 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2367 if (meta_group_info == NULL) {
2368 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2369 "for a buddy group");
2370 goto exit_meta_group_info;
2372 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2373 meta_group_info;
2376 meta_group_info =
2377 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2378 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2380 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2381 if (meta_group_info[i] == NULL) {
2382 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2383 goto exit_group_info;
2385 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2386 &(meta_group_info[i]->bb_state));
2389 * initialize bb_free to be able to skip
2390 * empty groups without initialization
2392 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2393 meta_group_info[i]->bb_free =
2394 ext4_free_clusters_after_init(sb, group, desc);
2395 } else {
2396 meta_group_info[i]->bb_free =
2397 ext4_free_group_clusters(sb, desc);
2400 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2401 init_rwsem(&meta_group_info[i]->alloc_sem);
2402 meta_group_info[i]->bb_free_root = RB_ROOT;
2403 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2405 #ifdef DOUBLE_CHECK
2407 struct buffer_head *bh;
2408 meta_group_info[i]->bb_bitmap =
2409 kmalloc(sb->s_blocksize, GFP_KERNEL);
2410 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2411 bh = ext4_read_block_bitmap(sb, group);
2412 BUG_ON(bh == NULL);
2413 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2414 sb->s_blocksize);
2415 put_bh(bh);
2417 #endif
2419 return 0;
2421 exit_group_info:
2422 /* If a meta_group_info table has been allocated, release it now */
2423 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2424 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2425 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2427 exit_meta_group_info:
2428 return -ENOMEM;
2429 } /* ext4_mb_add_groupinfo */
2431 static int ext4_mb_init_backend(struct super_block *sb)
2433 ext4_group_t ngroups = ext4_get_groups_count(sb);
2434 ext4_group_t i;
2435 struct ext4_sb_info *sbi = EXT4_SB(sb);
2436 int err;
2437 struct ext4_group_desc *desc;
2438 struct kmem_cache *cachep;
2440 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2441 if (err)
2442 return err;
2444 sbi->s_buddy_cache = new_inode(sb);
2445 if (sbi->s_buddy_cache == NULL) {
2446 ext4_msg(sb, KERN_ERR, "can't get new inode");
2447 goto err_freesgi;
2449 /* To avoid potentially colliding with an valid on-disk inode number,
2450 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2451 * not in the inode hash, so it should never be found by iget(), but
2452 * this will avoid confusion if it ever shows up during debugging. */
2453 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2454 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2455 for (i = 0; i < ngroups; i++) {
2456 desc = ext4_get_group_desc(sb, i, NULL);
2457 if (desc == NULL) {
2458 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2459 goto err_freebuddy;
2461 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2462 goto err_freebuddy;
2465 return 0;
2467 err_freebuddy:
2468 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2469 while (i-- > 0)
2470 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2471 i = sbi->s_group_info_size;
2472 while (i-- > 0)
2473 kfree(sbi->s_group_info[i]);
2474 iput(sbi->s_buddy_cache);
2475 err_freesgi:
2476 ext4_kvfree(sbi->s_group_info);
2477 return -ENOMEM;
2480 static void ext4_groupinfo_destroy_slabs(void)
2482 int i;
2484 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2485 if (ext4_groupinfo_caches[i])
2486 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2487 ext4_groupinfo_caches[i] = NULL;
2491 static int ext4_groupinfo_create_slab(size_t size)
2493 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2494 int slab_size;
2495 int blocksize_bits = order_base_2(size);
2496 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2497 struct kmem_cache *cachep;
2499 if (cache_index >= NR_GRPINFO_CACHES)
2500 return -EINVAL;
2502 if (unlikely(cache_index < 0))
2503 cache_index = 0;
2505 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2506 if (ext4_groupinfo_caches[cache_index]) {
2507 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2508 return 0; /* Already created */
2511 slab_size = offsetof(struct ext4_group_info,
2512 bb_counters[blocksize_bits + 2]);
2514 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2515 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2516 NULL);
2518 ext4_groupinfo_caches[cache_index] = cachep;
2520 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2521 if (!cachep) {
2522 printk(KERN_EMERG
2523 "EXT4-fs: no memory for groupinfo slab cache\n");
2524 return -ENOMEM;
2527 return 0;
2530 int ext4_mb_init(struct super_block *sb)
2532 struct ext4_sb_info *sbi = EXT4_SB(sb);
2533 unsigned i, j;
2534 unsigned offset;
2535 unsigned max;
2536 int ret;
2538 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2540 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2541 if (sbi->s_mb_offsets == NULL) {
2542 ret = -ENOMEM;
2543 goto out;
2546 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2547 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2548 if (sbi->s_mb_maxs == NULL) {
2549 ret = -ENOMEM;
2550 goto out;
2553 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2554 if (ret < 0)
2555 goto out;
2557 /* order 0 is regular bitmap */
2558 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2559 sbi->s_mb_offsets[0] = 0;
2561 i = 1;
2562 offset = 0;
2563 max = sb->s_blocksize << 2;
2564 do {
2565 sbi->s_mb_offsets[i] = offset;
2566 sbi->s_mb_maxs[i] = max;
2567 offset += 1 << (sb->s_blocksize_bits - i);
2568 max = max >> 1;
2569 i++;
2570 } while (i <= sb->s_blocksize_bits + 1);
2572 spin_lock_init(&sbi->s_md_lock);
2573 spin_lock_init(&sbi->s_bal_lock);
2575 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2576 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2577 sbi->s_mb_stats = MB_DEFAULT_STATS;
2578 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2579 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2581 * The default group preallocation is 512, which for 4k block
2582 * sizes translates to 2 megabytes. However for bigalloc file
2583 * systems, this is probably too big (i.e, if the cluster size
2584 * is 1 megabyte, then group preallocation size becomes half a
2585 * gigabyte!). As a default, we will keep a two megabyte
2586 * group pralloc size for cluster sizes up to 64k, and after
2587 * that, we will force a minimum group preallocation size of
2588 * 32 clusters. This translates to 8 megs when the cluster
2589 * size is 256k, and 32 megs when the cluster size is 1 meg,
2590 * which seems reasonable as a default.
2592 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2593 sbi->s_cluster_bits, 32);
2595 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2596 * to the lowest multiple of s_stripe which is bigger than
2597 * the s_mb_group_prealloc as determined above. We want
2598 * the preallocation size to be an exact multiple of the
2599 * RAID stripe size so that preallocations don't fragment
2600 * the stripes.
2602 if (sbi->s_stripe > 1) {
2603 sbi->s_mb_group_prealloc = roundup(
2604 sbi->s_mb_group_prealloc, sbi->s_stripe);
2607 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2608 if (sbi->s_locality_groups == NULL) {
2609 ret = -ENOMEM;
2610 goto out_free_groupinfo_slab;
2612 for_each_possible_cpu(i) {
2613 struct ext4_locality_group *lg;
2614 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2615 mutex_init(&lg->lg_mutex);
2616 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2617 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2618 spin_lock_init(&lg->lg_prealloc_lock);
2621 /* init file for buddy data */
2622 ret = ext4_mb_init_backend(sb);
2623 if (ret != 0)
2624 goto out_free_locality_groups;
2626 if (sbi->s_proc)
2627 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2628 &ext4_mb_seq_groups_fops, sb);
2630 return 0;
2632 out_free_locality_groups:
2633 free_percpu(sbi->s_locality_groups);
2634 sbi->s_locality_groups = NULL;
2635 out_free_groupinfo_slab:
2636 ext4_groupinfo_destroy_slabs();
2637 out:
2638 kfree(sbi->s_mb_offsets);
2639 sbi->s_mb_offsets = NULL;
2640 kfree(sbi->s_mb_maxs);
2641 sbi->s_mb_maxs = NULL;
2642 return ret;
2645 /* need to called with the ext4 group lock held */
2646 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2648 struct ext4_prealloc_space *pa;
2649 struct list_head *cur, *tmp;
2650 int count = 0;
2652 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2653 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2654 list_del(&pa->pa_group_list);
2655 count++;
2656 kmem_cache_free(ext4_pspace_cachep, pa);
2658 if (count)
2659 mb_debug(1, "mballoc: %u PAs left\n", count);
2663 int ext4_mb_release(struct super_block *sb)
2665 ext4_group_t ngroups = ext4_get_groups_count(sb);
2666 ext4_group_t i;
2667 int num_meta_group_infos;
2668 struct ext4_group_info *grinfo;
2669 struct ext4_sb_info *sbi = EXT4_SB(sb);
2670 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2672 if (sbi->s_proc)
2673 remove_proc_entry("mb_groups", sbi->s_proc);
2675 if (sbi->s_group_info) {
2676 for (i = 0; i < ngroups; i++) {
2677 grinfo = ext4_get_group_info(sb, i);
2678 #ifdef DOUBLE_CHECK
2679 kfree(grinfo->bb_bitmap);
2680 #endif
2681 ext4_lock_group(sb, i);
2682 ext4_mb_cleanup_pa(grinfo);
2683 ext4_unlock_group(sb, i);
2684 kmem_cache_free(cachep, grinfo);
2686 num_meta_group_infos = (ngroups +
2687 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2688 EXT4_DESC_PER_BLOCK_BITS(sb);
2689 for (i = 0; i < num_meta_group_infos; i++)
2690 kfree(sbi->s_group_info[i]);
2691 ext4_kvfree(sbi->s_group_info);
2693 kfree(sbi->s_mb_offsets);
2694 kfree(sbi->s_mb_maxs);
2695 if (sbi->s_buddy_cache)
2696 iput(sbi->s_buddy_cache);
2697 if (sbi->s_mb_stats) {
2698 ext4_msg(sb, KERN_INFO,
2699 "mballoc: %u blocks %u reqs (%u success)",
2700 atomic_read(&sbi->s_bal_allocated),
2701 atomic_read(&sbi->s_bal_reqs),
2702 atomic_read(&sbi->s_bal_success));
2703 ext4_msg(sb, KERN_INFO,
2704 "mballoc: %u extents scanned, %u goal hits, "
2705 "%u 2^N hits, %u breaks, %u lost",
2706 atomic_read(&sbi->s_bal_ex_scanned),
2707 atomic_read(&sbi->s_bal_goals),
2708 atomic_read(&sbi->s_bal_2orders),
2709 atomic_read(&sbi->s_bal_breaks),
2710 atomic_read(&sbi->s_mb_lost_chunks));
2711 ext4_msg(sb, KERN_INFO,
2712 "mballoc: %lu generated and it took %Lu",
2713 sbi->s_mb_buddies_generated,
2714 sbi->s_mb_generation_time);
2715 ext4_msg(sb, KERN_INFO,
2716 "mballoc: %u preallocated, %u discarded",
2717 atomic_read(&sbi->s_mb_preallocated),
2718 atomic_read(&sbi->s_mb_discarded));
2721 free_percpu(sbi->s_locality_groups);
2723 return 0;
2726 static inline int ext4_issue_discard(struct super_block *sb,
2727 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2729 ext4_fsblk_t discard_block;
2731 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2732 ext4_group_first_block_no(sb, block_group));
2733 count = EXT4_C2B(EXT4_SB(sb), count);
2734 trace_ext4_discard_blocks(sb,
2735 (unsigned long long) discard_block, count);
2736 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2740 * This function is called by the jbd2 layer once the commit has finished,
2741 * so we know we can free the blocks that were released with that commit.
2743 static void ext4_free_data_callback(struct super_block *sb,
2744 struct ext4_journal_cb_entry *jce,
2745 int rc)
2747 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2748 struct ext4_buddy e4b;
2749 struct ext4_group_info *db;
2750 int err, count = 0, count2 = 0;
2752 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2753 entry->efd_count, entry->efd_group, entry);
2755 if (test_opt(sb, DISCARD)) {
2756 err = ext4_issue_discard(sb, entry->efd_group,
2757 entry->efd_start_cluster,
2758 entry->efd_count);
2759 if (err && err != -EOPNOTSUPP)
2760 ext4_msg(sb, KERN_WARNING, "discard request in"
2761 " group:%d block:%d count:%d failed"
2762 " with %d", entry->efd_group,
2763 entry->efd_start_cluster,
2764 entry->efd_count, err);
2767 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2768 /* we expect to find existing buddy because it's pinned */
2769 BUG_ON(err != 0);
2772 db = e4b.bd_info;
2773 /* there are blocks to put in buddy to make them really free */
2774 count += entry->efd_count;
2775 count2++;
2776 ext4_lock_group(sb, entry->efd_group);
2777 /* Take it out of per group rb tree */
2778 rb_erase(&entry->efd_node, &(db->bb_free_root));
2779 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2782 * Clear the trimmed flag for the group so that the next
2783 * ext4_trim_fs can trim it.
2784 * If the volume is mounted with -o discard, online discard
2785 * is supported and the free blocks will be trimmed online.
2787 if (!test_opt(sb, DISCARD))
2788 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2790 if (!db->bb_free_root.rb_node) {
2791 /* No more items in the per group rb tree
2792 * balance refcounts from ext4_mb_free_metadata()
2794 page_cache_release(e4b.bd_buddy_page);
2795 page_cache_release(e4b.bd_bitmap_page);
2797 ext4_unlock_group(sb, entry->efd_group);
2798 kmem_cache_free(ext4_free_data_cachep, entry);
2799 ext4_mb_unload_buddy(&e4b);
2801 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2804 int __init ext4_init_mballoc(void)
2806 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2807 SLAB_RECLAIM_ACCOUNT);
2808 if (ext4_pspace_cachep == NULL)
2809 return -ENOMEM;
2811 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2812 SLAB_RECLAIM_ACCOUNT);
2813 if (ext4_ac_cachep == NULL) {
2814 kmem_cache_destroy(ext4_pspace_cachep);
2815 return -ENOMEM;
2818 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2819 SLAB_RECLAIM_ACCOUNT);
2820 if (ext4_free_data_cachep == NULL) {
2821 kmem_cache_destroy(ext4_pspace_cachep);
2822 kmem_cache_destroy(ext4_ac_cachep);
2823 return -ENOMEM;
2825 return 0;
2828 void ext4_exit_mballoc(void)
2831 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2832 * before destroying the slab cache.
2834 rcu_barrier();
2835 kmem_cache_destroy(ext4_pspace_cachep);
2836 kmem_cache_destroy(ext4_ac_cachep);
2837 kmem_cache_destroy(ext4_free_data_cachep);
2838 ext4_groupinfo_destroy_slabs();
2843 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2844 * Returns 0 if success or error code
2846 static noinline_for_stack int
2847 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2848 handle_t *handle, unsigned int reserv_clstrs)
2850 struct buffer_head *bitmap_bh = NULL;
2851 struct ext4_group_desc *gdp;
2852 struct buffer_head *gdp_bh;
2853 struct ext4_sb_info *sbi;
2854 struct super_block *sb;
2855 ext4_fsblk_t block;
2856 int err, len;
2858 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2859 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2861 sb = ac->ac_sb;
2862 sbi = EXT4_SB(sb);
2864 err = -EIO;
2865 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2866 if (!bitmap_bh)
2867 goto out_err;
2869 err = ext4_journal_get_write_access(handle, bitmap_bh);
2870 if (err)
2871 goto out_err;
2873 err = -EIO;
2874 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2875 if (!gdp)
2876 goto out_err;
2878 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2879 ext4_free_group_clusters(sb, gdp));
2881 err = ext4_journal_get_write_access(handle, gdp_bh);
2882 if (err)
2883 goto out_err;
2885 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2887 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2888 if (!ext4_data_block_valid(sbi, block, len)) {
2889 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2890 "fs metadata", block, block+len);
2891 /* File system mounted not to panic on error
2892 * Fix the bitmap and repeat the block allocation
2893 * We leak some of the blocks here.
2895 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2896 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2897 ac->ac_b_ex.fe_len);
2898 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2899 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2900 if (!err)
2901 err = -EAGAIN;
2902 goto out_err;
2905 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2906 #ifdef AGGRESSIVE_CHECK
2908 int i;
2909 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2910 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2911 bitmap_bh->b_data));
2914 #endif
2915 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2916 ac->ac_b_ex.fe_len);
2917 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2918 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2919 ext4_free_group_clusters_set(sb, gdp,
2920 ext4_free_clusters_after_init(sb,
2921 ac->ac_b_ex.fe_group, gdp));
2923 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2924 ext4_free_group_clusters_set(sb, gdp, len);
2925 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2926 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2928 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2929 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2931 * Now reduce the dirty block count also. Should not go negative
2933 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2934 /* release all the reserved blocks if non delalloc */
2935 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2936 reserv_clstrs);
2938 if (sbi->s_log_groups_per_flex) {
2939 ext4_group_t flex_group = ext4_flex_group(sbi,
2940 ac->ac_b_ex.fe_group);
2941 atomic64_sub(ac->ac_b_ex.fe_len,
2942 &sbi->s_flex_groups[flex_group].free_clusters);
2945 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2946 if (err)
2947 goto out_err;
2948 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2950 out_err:
2951 brelse(bitmap_bh);
2952 return err;
2956 * here we normalize request for locality group
2957 * Group request are normalized to s_mb_group_prealloc, which goes to
2958 * s_strip if we set the same via mount option.
2959 * s_mb_group_prealloc can be configured via
2960 * /sys/fs/ext4/<partition>/mb_group_prealloc
2962 * XXX: should we try to preallocate more than the group has now?
2964 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2966 struct super_block *sb = ac->ac_sb;
2967 struct ext4_locality_group *lg = ac->ac_lg;
2969 BUG_ON(lg == NULL);
2970 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2971 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2972 current->pid, ac->ac_g_ex.fe_len);
2976 * Normalization means making request better in terms of
2977 * size and alignment
2979 static noinline_for_stack void
2980 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2981 struct ext4_allocation_request *ar)
2983 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2984 int bsbits, max;
2985 ext4_lblk_t end;
2986 loff_t size, start_off;
2987 loff_t orig_size __maybe_unused;
2988 ext4_lblk_t start;
2989 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2990 struct ext4_prealloc_space *pa;
2992 /* do normalize only data requests, metadata requests
2993 do not need preallocation */
2994 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2995 return;
2997 /* sometime caller may want exact blocks */
2998 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2999 return;
3001 /* caller may indicate that preallocation isn't
3002 * required (it's a tail, for example) */
3003 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3004 return;
3006 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3007 ext4_mb_normalize_group_request(ac);
3008 return ;
3011 bsbits = ac->ac_sb->s_blocksize_bits;
3013 /* first, let's learn actual file size
3014 * given current request is allocated */
3015 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3016 size = size << bsbits;
3017 if (size < i_size_read(ac->ac_inode))
3018 size = i_size_read(ac->ac_inode);
3019 orig_size = size;
3021 /* max size of free chunks */
3022 max = 2 << bsbits;
3024 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3025 (req <= (size) || max <= (chunk_size))
3027 /* first, try to predict filesize */
3028 /* XXX: should this table be tunable? */
3029 start_off = 0;
3030 if (size <= 16 * 1024) {
3031 size = 16 * 1024;
3032 } else if (size <= 32 * 1024) {
3033 size = 32 * 1024;
3034 } else if (size <= 64 * 1024) {
3035 size = 64 * 1024;
3036 } else if (size <= 128 * 1024) {
3037 size = 128 * 1024;
3038 } else if (size <= 256 * 1024) {
3039 size = 256 * 1024;
3040 } else if (size <= 512 * 1024) {
3041 size = 512 * 1024;
3042 } else if (size <= 1024 * 1024) {
3043 size = 1024 * 1024;
3044 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3045 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3046 (21 - bsbits)) << 21;
3047 size = 2 * 1024 * 1024;
3048 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3049 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3050 (22 - bsbits)) << 22;
3051 size = 4 * 1024 * 1024;
3052 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3053 (8<<20)>>bsbits, max, 8 * 1024)) {
3054 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3055 (23 - bsbits)) << 23;
3056 size = 8 * 1024 * 1024;
3057 } else {
3058 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3059 size = ac->ac_o_ex.fe_len << bsbits;
3061 size = size >> bsbits;
3062 start = start_off >> bsbits;
3064 /* don't cover already allocated blocks in selected range */
3065 if (ar->pleft && start <= ar->lleft) {
3066 size -= ar->lleft + 1 - start;
3067 start = ar->lleft + 1;
3069 if (ar->pright && start + size - 1 >= ar->lright)
3070 size -= start + size - ar->lright;
3072 end = start + size;
3074 /* check we don't cross already preallocated blocks */
3075 rcu_read_lock();
3076 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3077 ext4_lblk_t pa_end;
3079 if (pa->pa_deleted)
3080 continue;
3081 spin_lock(&pa->pa_lock);
3082 if (pa->pa_deleted) {
3083 spin_unlock(&pa->pa_lock);
3084 continue;
3087 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3088 pa->pa_len);
3090 /* PA must not overlap original request */
3091 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3092 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3094 /* skip PAs this normalized request doesn't overlap with */
3095 if (pa->pa_lstart >= end || pa_end <= start) {
3096 spin_unlock(&pa->pa_lock);
3097 continue;
3099 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3101 /* adjust start or end to be adjacent to this pa */
3102 if (pa_end <= ac->ac_o_ex.fe_logical) {
3103 BUG_ON(pa_end < start);
3104 start = pa_end;
3105 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3106 BUG_ON(pa->pa_lstart > end);
3107 end = pa->pa_lstart;
3109 spin_unlock(&pa->pa_lock);
3111 rcu_read_unlock();
3112 size = end - start;
3114 /* XXX: extra loop to check we really don't overlap preallocations */
3115 rcu_read_lock();
3116 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3117 ext4_lblk_t pa_end;
3119 spin_lock(&pa->pa_lock);
3120 if (pa->pa_deleted == 0) {
3121 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3122 pa->pa_len);
3123 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3125 spin_unlock(&pa->pa_lock);
3127 rcu_read_unlock();
3129 if (start + size <= ac->ac_o_ex.fe_logical &&
3130 start > ac->ac_o_ex.fe_logical) {
3131 ext4_msg(ac->ac_sb, KERN_ERR,
3132 "start %lu, size %lu, fe_logical %lu",
3133 (unsigned long) start, (unsigned long) size,
3134 (unsigned long) ac->ac_o_ex.fe_logical);
3136 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3137 start > ac->ac_o_ex.fe_logical);
3138 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3140 /* now prepare goal request */
3142 /* XXX: is it better to align blocks WRT to logical
3143 * placement or satisfy big request as is */
3144 ac->ac_g_ex.fe_logical = start;
3145 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3147 /* define goal start in order to merge */
3148 if (ar->pright && (ar->lright == (start + size))) {
3149 /* merge to the right */
3150 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3151 &ac->ac_f_ex.fe_group,
3152 &ac->ac_f_ex.fe_start);
3153 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3155 if (ar->pleft && (ar->lleft + 1 == start)) {
3156 /* merge to the left */
3157 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3158 &ac->ac_f_ex.fe_group,
3159 &ac->ac_f_ex.fe_start);
3160 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3163 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3164 (unsigned) orig_size, (unsigned) start);
3167 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3169 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3171 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3172 atomic_inc(&sbi->s_bal_reqs);
3173 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3174 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3175 atomic_inc(&sbi->s_bal_success);
3176 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3177 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3178 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3179 atomic_inc(&sbi->s_bal_goals);
3180 if (ac->ac_found > sbi->s_mb_max_to_scan)
3181 atomic_inc(&sbi->s_bal_breaks);
3184 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3185 trace_ext4_mballoc_alloc(ac);
3186 else
3187 trace_ext4_mballoc_prealloc(ac);
3191 * Called on failure; free up any blocks from the inode PA for this
3192 * context. We don't need this for MB_GROUP_PA because we only change
3193 * pa_free in ext4_mb_release_context(), but on failure, we've already
3194 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3196 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3198 struct ext4_prealloc_space *pa = ac->ac_pa;
3200 if (pa && pa->pa_type == MB_INODE_PA)
3201 pa->pa_free += ac->ac_b_ex.fe_len;
3205 * use blocks preallocated to inode
3207 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3208 struct ext4_prealloc_space *pa)
3210 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3211 ext4_fsblk_t start;
3212 ext4_fsblk_t end;
3213 int len;
3215 /* found preallocated blocks, use them */
3216 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3217 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3218 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3219 len = EXT4_NUM_B2C(sbi, end - start);
3220 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3221 &ac->ac_b_ex.fe_start);
3222 ac->ac_b_ex.fe_len = len;
3223 ac->ac_status = AC_STATUS_FOUND;
3224 ac->ac_pa = pa;
3226 BUG_ON(start < pa->pa_pstart);
3227 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3228 BUG_ON(pa->pa_free < len);
3229 pa->pa_free -= len;
3231 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3235 * use blocks preallocated to locality group
3237 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3238 struct ext4_prealloc_space *pa)
3240 unsigned int len = ac->ac_o_ex.fe_len;
3242 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3243 &ac->ac_b_ex.fe_group,
3244 &ac->ac_b_ex.fe_start);
3245 ac->ac_b_ex.fe_len = len;
3246 ac->ac_status = AC_STATUS_FOUND;
3247 ac->ac_pa = pa;
3249 /* we don't correct pa_pstart or pa_plen here to avoid
3250 * possible race when the group is being loaded concurrently
3251 * instead we correct pa later, after blocks are marked
3252 * in on-disk bitmap -- see ext4_mb_release_context()
3253 * Other CPUs are prevented from allocating from this pa by lg_mutex
3255 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3259 * Return the prealloc space that have minimal distance
3260 * from the goal block. @cpa is the prealloc
3261 * space that is having currently known minimal distance
3262 * from the goal block.
3264 static struct ext4_prealloc_space *
3265 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3266 struct ext4_prealloc_space *pa,
3267 struct ext4_prealloc_space *cpa)
3269 ext4_fsblk_t cur_distance, new_distance;
3271 if (cpa == NULL) {
3272 atomic_inc(&pa->pa_count);
3273 return pa;
3275 cur_distance = abs(goal_block - cpa->pa_pstart);
3276 new_distance = abs(goal_block - pa->pa_pstart);
3278 if (cur_distance <= new_distance)
3279 return cpa;
3281 /* drop the previous reference */
3282 atomic_dec(&cpa->pa_count);
3283 atomic_inc(&pa->pa_count);
3284 return pa;
3288 * search goal blocks in preallocated space
3290 static noinline_for_stack int
3291 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3293 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3294 int order, i;
3295 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3296 struct ext4_locality_group *lg;
3297 struct ext4_prealloc_space *pa, *cpa = NULL;
3298 ext4_fsblk_t goal_block;
3300 /* only data can be preallocated */
3301 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3302 return 0;
3304 /* first, try per-file preallocation */
3305 rcu_read_lock();
3306 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3308 /* all fields in this condition don't change,
3309 * so we can skip locking for them */
3310 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3311 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3312 EXT4_C2B(sbi, pa->pa_len)))
3313 continue;
3315 /* non-extent files can't have physical blocks past 2^32 */
3316 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3317 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3318 EXT4_MAX_BLOCK_FILE_PHYS))
3319 continue;
3321 /* found preallocated blocks, use them */
3322 spin_lock(&pa->pa_lock);
3323 if (pa->pa_deleted == 0 && pa->pa_free) {
3324 atomic_inc(&pa->pa_count);
3325 ext4_mb_use_inode_pa(ac, pa);
3326 spin_unlock(&pa->pa_lock);
3327 ac->ac_criteria = 10;
3328 rcu_read_unlock();
3329 return 1;
3331 spin_unlock(&pa->pa_lock);
3333 rcu_read_unlock();
3335 /* can we use group allocation? */
3336 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3337 return 0;
3339 /* inode may have no locality group for some reason */
3340 lg = ac->ac_lg;
3341 if (lg == NULL)
3342 return 0;
3343 order = fls(ac->ac_o_ex.fe_len) - 1;
3344 if (order > PREALLOC_TB_SIZE - 1)
3345 /* The max size of hash table is PREALLOC_TB_SIZE */
3346 order = PREALLOC_TB_SIZE - 1;
3348 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3350 * search for the prealloc space that is having
3351 * minimal distance from the goal block.
3353 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3354 rcu_read_lock();
3355 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3356 pa_inode_list) {
3357 spin_lock(&pa->pa_lock);
3358 if (pa->pa_deleted == 0 &&
3359 pa->pa_free >= ac->ac_o_ex.fe_len) {
3361 cpa = ext4_mb_check_group_pa(goal_block,
3362 pa, cpa);
3364 spin_unlock(&pa->pa_lock);
3366 rcu_read_unlock();
3368 if (cpa) {
3369 ext4_mb_use_group_pa(ac, cpa);
3370 ac->ac_criteria = 20;
3371 return 1;
3373 return 0;
3377 * the function goes through all block freed in the group
3378 * but not yet committed and marks them used in in-core bitmap.
3379 * buddy must be generated from this bitmap
3380 * Need to be called with the ext4 group lock held
3382 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3383 ext4_group_t group)
3385 struct rb_node *n;
3386 struct ext4_group_info *grp;
3387 struct ext4_free_data *entry;
3389 grp = ext4_get_group_info(sb, group);
3390 n = rb_first(&(grp->bb_free_root));
3392 while (n) {
3393 entry = rb_entry(n, struct ext4_free_data, efd_node);
3394 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3395 n = rb_next(n);
3397 return;
3401 * the function goes through all preallocation in this group and marks them
3402 * used in in-core bitmap. buddy must be generated from this bitmap
3403 * Need to be called with ext4 group lock held
3405 static noinline_for_stack
3406 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3407 ext4_group_t group)
3409 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3410 struct ext4_prealloc_space *pa;
3411 struct list_head *cur;
3412 ext4_group_t groupnr;
3413 ext4_grpblk_t start;
3414 int preallocated = 0;
3415 int len;
3417 /* all form of preallocation discards first load group,
3418 * so the only competing code is preallocation use.
3419 * we don't need any locking here
3420 * notice we do NOT ignore preallocations with pa_deleted
3421 * otherwise we could leave used blocks available for
3422 * allocation in buddy when concurrent ext4_mb_put_pa()
3423 * is dropping preallocation
3425 list_for_each(cur, &grp->bb_prealloc_list) {
3426 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3427 spin_lock(&pa->pa_lock);
3428 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3429 &groupnr, &start);
3430 len = pa->pa_len;
3431 spin_unlock(&pa->pa_lock);
3432 if (unlikely(len == 0))
3433 continue;
3434 BUG_ON(groupnr != group);
3435 ext4_set_bits(bitmap, start, len);
3436 preallocated += len;
3438 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3441 static void ext4_mb_pa_callback(struct rcu_head *head)
3443 struct ext4_prealloc_space *pa;
3444 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3445 kmem_cache_free(ext4_pspace_cachep, pa);
3449 * drops a reference to preallocated space descriptor
3450 * if this was the last reference and the space is consumed
3452 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3453 struct super_block *sb, struct ext4_prealloc_space *pa)
3455 ext4_group_t grp;
3456 ext4_fsblk_t grp_blk;
3458 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3459 return;
3461 /* in this short window concurrent discard can set pa_deleted */
3462 spin_lock(&pa->pa_lock);
3463 if (pa->pa_deleted == 1) {
3464 spin_unlock(&pa->pa_lock);
3465 return;
3468 pa->pa_deleted = 1;
3469 spin_unlock(&pa->pa_lock);
3471 grp_blk = pa->pa_pstart;
3473 * If doing group-based preallocation, pa_pstart may be in the
3474 * next group when pa is used up
3476 if (pa->pa_type == MB_GROUP_PA)
3477 grp_blk--;
3479 grp = ext4_get_group_number(sb, grp_blk);
3482 * possible race:
3484 * P1 (buddy init) P2 (regular allocation)
3485 * find block B in PA
3486 * copy on-disk bitmap to buddy
3487 * mark B in on-disk bitmap
3488 * drop PA from group
3489 * mark all PAs in buddy
3491 * thus, P1 initializes buddy with B available. to prevent this
3492 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3493 * against that pair
3495 ext4_lock_group(sb, grp);
3496 list_del(&pa->pa_group_list);
3497 ext4_unlock_group(sb, grp);
3499 spin_lock(pa->pa_obj_lock);
3500 list_del_rcu(&pa->pa_inode_list);
3501 spin_unlock(pa->pa_obj_lock);
3503 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3507 * creates new preallocated space for given inode
3509 static noinline_for_stack int
3510 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3512 struct super_block *sb = ac->ac_sb;
3513 struct ext4_sb_info *sbi = EXT4_SB(sb);
3514 struct ext4_prealloc_space *pa;
3515 struct ext4_group_info *grp;
3516 struct ext4_inode_info *ei;
3518 /* preallocate only when found space is larger then requested */
3519 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3520 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3521 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3523 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3524 if (pa == NULL)
3525 return -ENOMEM;
3527 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3528 int winl;
3529 int wins;
3530 int win;
3531 int offs;
3533 /* we can't allocate as much as normalizer wants.
3534 * so, found space must get proper lstart
3535 * to cover original request */
3536 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3537 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3539 /* we're limited by original request in that
3540 * logical block must be covered any way
3541 * winl is window we can move our chunk within */
3542 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3544 /* also, we should cover whole original request */
3545 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3547 /* the smallest one defines real window */
3548 win = min(winl, wins);
3550 offs = ac->ac_o_ex.fe_logical %
3551 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3552 if (offs && offs < win)
3553 win = offs;
3555 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3556 EXT4_NUM_B2C(sbi, win);
3557 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3558 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3561 /* preallocation can change ac_b_ex, thus we store actually
3562 * allocated blocks for history */
3563 ac->ac_f_ex = ac->ac_b_ex;
3565 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3566 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3567 pa->pa_len = ac->ac_b_ex.fe_len;
3568 pa->pa_free = pa->pa_len;
3569 atomic_set(&pa->pa_count, 1);
3570 spin_lock_init(&pa->pa_lock);
3571 INIT_LIST_HEAD(&pa->pa_inode_list);
3572 INIT_LIST_HEAD(&pa->pa_group_list);
3573 pa->pa_deleted = 0;
3574 pa->pa_type = MB_INODE_PA;
3576 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3577 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3578 trace_ext4_mb_new_inode_pa(ac, pa);
3580 ext4_mb_use_inode_pa(ac, pa);
3581 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3583 ei = EXT4_I(ac->ac_inode);
3584 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3586 pa->pa_obj_lock = &ei->i_prealloc_lock;
3587 pa->pa_inode = ac->ac_inode;
3589 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3590 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3591 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3593 spin_lock(pa->pa_obj_lock);
3594 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3595 spin_unlock(pa->pa_obj_lock);
3597 return 0;
3601 * creates new preallocated space for locality group inodes belongs to
3603 static noinline_for_stack int
3604 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3606 struct super_block *sb = ac->ac_sb;
3607 struct ext4_locality_group *lg;
3608 struct ext4_prealloc_space *pa;
3609 struct ext4_group_info *grp;
3611 /* preallocate only when found space is larger then requested */
3612 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3613 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3614 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3616 BUG_ON(ext4_pspace_cachep == NULL);
3617 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3618 if (pa == NULL)
3619 return -ENOMEM;
3621 /* preallocation can change ac_b_ex, thus we store actually
3622 * allocated blocks for history */
3623 ac->ac_f_ex = ac->ac_b_ex;
3625 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3626 pa->pa_lstart = pa->pa_pstart;
3627 pa->pa_len = ac->ac_b_ex.fe_len;
3628 pa->pa_free = pa->pa_len;
3629 atomic_set(&pa->pa_count, 1);
3630 spin_lock_init(&pa->pa_lock);
3631 INIT_LIST_HEAD(&pa->pa_inode_list);
3632 INIT_LIST_HEAD(&pa->pa_group_list);
3633 pa->pa_deleted = 0;
3634 pa->pa_type = MB_GROUP_PA;
3636 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3637 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3638 trace_ext4_mb_new_group_pa(ac, pa);
3640 ext4_mb_use_group_pa(ac, pa);
3641 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3643 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3644 lg = ac->ac_lg;
3645 BUG_ON(lg == NULL);
3647 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3648 pa->pa_inode = NULL;
3650 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3651 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3652 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3655 * We will later add the new pa to the right bucket
3656 * after updating the pa_free in ext4_mb_release_context
3658 return 0;
3661 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3663 int err;
3665 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3666 err = ext4_mb_new_group_pa(ac);
3667 else
3668 err = ext4_mb_new_inode_pa(ac);
3669 return err;
3673 * finds all unused blocks in on-disk bitmap, frees them in
3674 * in-core bitmap and buddy.
3675 * @pa must be unlinked from inode and group lists, so that
3676 * nobody else can find/use it.
3677 * the caller MUST hold group/inode locks.
3678 * TODO: optimize the case when there are no in-core structures yet
3680 static noinline_for_stack int
3681 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3682 struct ext4_prealloc_space *pa)
3684 struct super_block *sb = e4b->bd_sb;
3685 struct ext4_sb_info *sbi = EXT4_SB(sb);
3686 unsigned int end;
3687 unsigned int next;
3688 ext4_group_t group;
3689 ext4_grpblk_t bit;
3690 unsigned long long grp_blk_start;
3691 int err = 0;
3692 int free = 0;
3694 BUG_ON(pa->pa_deleted == 0);
3695 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3696 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3697 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3698 end = bit + pa->pa_len;
3700 while (bit < end) {
3701 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3702 if (bit >= end)
3703 break;
3704 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3705 mb_debug(1, " free preallocated %u/%u in group %u\n",
3706 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3707 (unsigned) next - bit, (unsigned) group);
3708 free += next - bit;
3710 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3711 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3712 EXT4_C2B(sbi, bit)),
3713 next - bit);
3714 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3715 bit = next + 1;
3717 if (free != pa->pa_free) {
3718 ext4_msg(e4b->bd_sb, KERN_CRIT,
3719 "pa %p: logic %lu, phys. %lu, len %lu",
3720 pa, (unsigned long) pa->pa_lstart,
3721 (unsigned long) pa->pa_pstart,
3722 (unsigned long) pa->pa_len);
3723 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3724 free, pa->pa_free);
3726 * pa is already deleted so we use the value obtained
3727 * from the bitmap and continue.
3730 atomic_add(free, &sbi->s_mb_discarded);
3732 return err;
3735 static noinline_for_stack int
3736 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3737 struct ext4_prealloc_space *pa)
3739 struct super_block *sb = e4b->bd_sb;
3740 ext4_group_t group;
3741 ext4_grpblk_t bit;
3743 trace_ext4_mb_release_group_pa(sb, pa);
3744 BUG_ON(pa->pa_deleted == 0);
3745 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3746 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3747 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3748 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3749 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3751 return 0;
3755 * releases all preallocations in given group
3757 * first, we need to decide discard policy:
3758 * - when do we discard
3759 * 1) ENOSPC
3760 * - how many do we discard
3761 * 1) how many requested
3763 static noinline_for_stack int
3764 ext4_mb_discard_group_preallocations(struct super_block *sb,
3765 ext4_group_t group, int needed)
3767 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3768 struct buffer_head *bitmap_bh = NULL;
3769 struct ext4_prealloc_space *pa, *tmp;
3770 struct list_head list;
3771 struct ext4_buddy e4b;
3772 int err;
3773 int busy = 0;
3774 int free = 0;
3776 mb_debug(1, "discard preallocation for group %u\n", group);
3778 if (list_empty(&grp->bb_prealloc_list))
3779 return 0;
3781 bitmap_bh = ext4_read_block_bitmap(sb, group);
3782 if (bitmap_bh == NULL) {
3783 ext4_error(sb, "Error reading block bitmap for %u", group);
3784 return 0;
3787 err = ext4_mb_load_buddy(sb, group, &e4b);
3788 if (err) {
3789 ext4_error(sb, "Error loading buddy information for %u", group);
3790 put_bh(bitmap_bh);
3791 return 0;
3794 if (needed == 0)
3795 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3797 INIT_LIST_HEAD(&list);
3798 repeat:
3799 ext4_lock_group(sb, group);
3800 list_for_each_entry_safe(pa, tmp,
3801 &grp->bb_prealloc_list, pa_group_list) {
3802 spin_lock(&pa->pa_lock);
3803 if (atomic_read(&pa->pa_count)) {
3804 spin_unlock(&pa->pa_lock);
3805 busy = 1;
3806 continue;
3808 if (pa->pa_deleted) {
3809 spin_unlock(&pa->pa_lock);
3810 continue;
3813 /* seems this one can be freed ... */
3814 pa->pa_deleted = 1;
3816 /* we can trust pa_free ... */
3817 free += pa->pa_free;
3819 spin_unlock(&pa->pa_lock);
3821 list_del(&pa->pa_group_list);
3822 list_add(&pa->u.pa_tmp_list, &list);
3825 /* if we still need more blocks and some PAs were used, try again */
3826 if (free < needed && busy) {
3827 busy = 0;
3828 ext4_unlock_group(sb, group);
3829 cond_resched();
3830 goto repeat;
3833 /* found anything to free? */
3834 if (list_empty(&list)) {
3835 BUG_ON(free != 0);
3836 goto out;
3839 /* now free all selected PAs */
3840 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3842 /* remove from object (inode or locality group) */
3843 spin_lock(pa->pa_obj_lock);
3844 list_del_rcu(&pa->pa_inode_list);
3845 spin_unlock(pa->pa_obj_lock);
3847 if (pa->pa_type == MB_GROUP_PA)
3848 ext4_mb_release_group_pa(&e4b, pa);
3849 else
3850 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3852 list_del(&pa->u.pa_tmp_list);
3853 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3856 out:
3857 ext4_unlock_group(sb, group);
3858 ext4_mb_unload_buddy(&e4b);
3859 put_bh(bitmap_bh);
3860 return free;
3864 * releases all non-used preallocated blocks for given inode
3866 * It's important to discard preallocations under i_data_sem
3867 * We don't want another block to be served from the prealloc
3868 * space when we are discarding the inode prealloc space.
3870 * FIXME!! Make sure it is valid at all the call sites
3872 void ext4_discard_preallocations(struct inode *inode)
3874 struct ext4_inode_info *ei = EXT4_I(inode);
3875 struct super_block *sb = inode->i_sb;
3876 struct buffer_head *bitmap_bh = NULL;
3877 struct ext4_prealloc_space *pa, *tmp;
3878 ext4_group_t group = 0;
3879 struct list_head list;
3880 struct ext4_buddy e4b;
3881 int err;
3883 if (!S_ISREG(inode->i_mode)) {
3884 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3885 return;
3888 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3889 trace_ext4_discard_preallocations(inode);
3891 INIT_LIST_HEAD(&list);
3893 repeat:
3894 /* first, collect all pa's in the inode */
3895 spin_lock(&ei->i_prealloc_lock);
3896 while (!list_empty(&ei->i_prealloc_list)) {
3897 pa = list_entry(ei->i_prealloc_list.next,
3898 struct ext4_prealloc_space, pa_inode_list);
3899 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3900 spin_lock(&pa->pa_lock);
3901 if (atomic_read(&pa->pa_count)) {
3902 /* this shouldn't happen often - nobody should
3903 * use preallocation while we're discarding it */
3904 spin_unlock(&pa->pa_lock);
3905 spin_unlock(&ei->i_prealloc_lock);
3906 ext4_msg(sb, KERN_ERR,
3907 "uh-oh! used pa while discarding");
3908 WARN_ON(1);
3909 schedule_timeout_uninterruptible(HZ);
3910 goto repeat;
3913 if (pa->pa_deleted == 0) {
3914 pa->pa_deleted = 1;
3915 spin_unlock(&pa->pa_lock);
3916 list_del_rcu(&pa->pa_inode_list);
3917 list_add(&pa->u.pa_tmp_list, &list);
3918 continue;
3921 /* someone is deleting pa right now */
3922 spin_unlock(&pa->pa_lock);
3923 spin_unlock(&ei->i_prealloc_lock);
3925 /* we have to wait here because pa_deleted
3926 * doesn't mean pa is already unlinked from
3927 * the list. as we might be called from
3928 * ->clear_inode() the inode will get freed
3929 * and concurrent thread which is unlinking
3930 * pa from inode's list may access already
3931 * freed memory, bad-bad-bad */
3933 /* XXX: if this happens too often, we can
3934 * add a flag to force wait only in case
3935 * of ->clear_inode(), but not in case of
3936 * regular truncate */
3937 schedule_timeout_uninterruptible(HZ);
3938 goto repeat;
3940 spin_unlock(&ei->i_prealloc_lock);
3942 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3943 BUG_ON(pa->pa_type != MB_INODE_PA);
3944 group = ext4_get_group_number(sb, pa->pa_pstart);
3946 err = ext4_mb_load_buddy(sb, group, &e4b);
3947 if (err) {
3948 ext4_error(sb, "Error loading buddy information for %u",
3949 group);
3950 continue;
3953 bitmap_bh = ext4_read_block_bitmap(sb, group);
3954 if (bitmap_bh == NULL) {
3955 ext4_error(sb, "Error reading block bitmap for %u",
3956 group);
3957 ext4_mb_unload_buddy(&e4b);
3958 continue;
3961 ext4_lock_group(sb, group);
3962 list_del(&pa->pa_group_list);
3963 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3964 ext4_unlock_group(sb, group);
3966 ext4_mb_unload_buddy(&e4b);
3967 put_bh(bitmap_bh);
3969 list_del(&pa->u.pa_tmp_list);
3970 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3974 #ifdef CONFIG_EXT4_DEBUG
3975 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3977 struct super_block *sb = ac->ac_sb;
3978 ext4_group_t ngroups, i;
3980 if (!ext4_mballoc_debug ||
3981 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3982 return;
3984 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3985 " Allocation context details:");
3986 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3987 ac->ac_status, ac->ac_flags);
3988 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3989 "goal %lu/%lu/%lu@%lu, "
3990 "best %lu/%lu/%lu@%lu cr %d",
3991 (unsigned long)ac->ac_o_ex.fe_group,
3992 (unsigned long)ac->ac_o_ex.fe_start,
3993 (unsigned long)ac->ac_o_ex.fe_len,
3994 (unsigned long)ac->ac_o_ex.fe_logical,
3995 (unsigned long)ac->ac_g_ex.fe_group,
3996 (unsigned long)ac->ac_g_ex.fe_start,
3997 (unsigned long)ac->ac_g_ex.fe_len,
3998 (unsigned long)ac->ac_g_ex.fe_logical,
3999 (unsigned long)ac->ac_b_ex.fe_group,
4000 (unsigned long)ac->ac_b_ex.fe_start,
4001 (unsigned long)ac->ac_b_ex.fe_len,
4002 (unsigned long)ac->ac_b_ex.fe_logical,
4003 (int)ac->ac_criteria);
4004 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
4005 ac->ac_ex_scanned, ac->ac_found);
4006 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4007 ngroups = ext4_get_groups_count(sb);
4008 for (i = 0; i < ngroups; i++) {
4009 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4010 struct ext4_prealloc_space *pa;
4011 ext4_grpblk_t start;
4012 struct list_head *cur;
4013 ext4_lock_group(sb, i);
4014 list_for_each(cur, &grp->bb_prealloc_list) {
4015 pa = list_entry(cur, struct ext4_prealloc_space,
4016 pa_group_list);
4017 spin_lock(&pa->pa_lock);
4018 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4019 NULL, &start);
4020 spin_unlock(&pa->pa_lock);
4021 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4022 start, pa->pa_len);
4024 ext4_unlock_group(sb, i);
4026 if (grp->bb_free == 0)
4027 continue;
4028 printk(KERN_ERR "%u: %d/%d \n",
4029 i, grp->bb_free, grp->bb_fragments);
4031 printk(KERN_ERR "\n");
4033 #else
4034 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4036 return;
4038 #endif
4041 * We use locality group preallocation for small size file. The size of the
4042 * file is determined by the current size or the resulting size after
4043 * allocation which ever is larger
4045 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4047 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4049 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4050 int bsbits = ac->ac_sb->s_blocksize_bits;
4051 loff_t size, isize;
4053 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4054 return;
4056 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4057 return;
4059 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4060 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4061 >> bsbits;
4063 if ((size == isize) &&
4064 !ext4_fs_is_busy(sbi) &&
4065 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4066 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4067 return;
4070 if (sbi->s_mb_group_prealloc <= 0) {
4071 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4072 return;
4075 /* don't use group allocation for large files */
4076 size = max(size, isize);
4077 if (size > sbi->s_mb_stream_request) {
4078 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4079 return;
4082 BUG_ON(ac->ac_lg != NULL);
4084 * locality group prealloc space are per cpu. The reason for having
4085 * per cpu locality group is to reduce the contention between block
4086 * request from multiple CPUs.
4088 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4090 /* we're going to use group allocation */
4091 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4093 /* serialize all allocations in the group */
4094 mutex_lock(&ac->ac_lg->lg_mutex);
4097 static noinline_for_stack int
4098 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4099 struct ext4_allocation_request *ar)
4101 struct super_block *sb = ar->inode->i_sb;
4102 struct ext4_sb_info *sbi = EXT4_SB(sb);
4103 struct ext4_super_block *es = sbi->s_es;
4104 ext4_group_t group;
4105 unsigned int len;
4106 ext4_fsblk_t goal;
4107 ext4_grpblk_t block;
4109 /* we can't allocate > group size */
4110 len = ar->len;
4112 /* just a dirty hack to filter too big requests */
4113 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4114 len = EXT4_CLUSTERS_PER_GROUP(sb);
4116 /* start searching from the goal */
4117 goal = ar->goal;
4118 if (goal < le32_to_cpu(es->s_first_data_block) ||
4119 goal >= ext4_blocks_count(es))
4120 goal = le32_to_cpu(es->s_first_data_block);
4121 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4123 /* set up allocation goals */
4124 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4125 ac->ac_status = AC_STATUS_CONTINUE;
4126 ac->ac_sb = sb;
4127 ac->ac_inode = ar->inode;
4128 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4129 ac->ac_o_ex.fe_group = group;
4130 ac->ac_o_ex.fe_start = block;
4131 ac->ac_o_ex.fe_len = len;
4132 ac->ac_g_ex = ac->ac_o_ex;
4133 ac->ac_flags = ar->flags;
4135 /* we have to define context: we'll we work with a file or
4136 * locality group. this is a policy, actually */
4137 ext4_mb_group_or_file(ac);
4139 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4140 "left: %u/%u, right %u/%u to %swritable\n",
4141 (unsigned) ar->len, (unsigned) ar->logical,
4142 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4143 (unsigned) ar->lleft, (unsigned) ar->pleft,
4144 (unsigned) ar->lright, (unsigned) ar->pright,
4145 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4146 return 0;
4150 static noinline_for_stack void
4151 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4152 struct ext4_locality_group *lg,
4153 int order, int total_entries)
4155 ext4_group_t group = 0;
4156 struct ext4_buddy e4b;
4157 struct list_head discard_list;
4158 struct ext4_prealloc_space *pa, *tmp;
4160 mb_debug(1, "discard locality group preallocation\n");
4162 INIT_LIST_HEAD(&discard_list);
4164 spin_lock(&lg->lg_prealloc_lock);
4165 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4166 pa_inode_list) {
4167 spin_lock(&pa->pa_lock);
4168 if (atomic_read(&pa->pa_count)) {
4170 * This is the pa that we just used
4171 * for block allocation. So don't
4172 * free that
4174 spin_unlock(&pa->pa_lock);
4175 continue;
4177 if (pa->pa_deleted) {
4178 spin_unlock(&pa->pa_lock);
4179 continue;
4181 /* only lg prealloc space */
4182 BUG_ON(pa->pa_type != MB_GROUP_PA);
4184 /* seems this one can be freed ... */
4185 pa->pa_deleted = 1;
4186 spin_unlock(&pa->pa_lock);
4188 list_del_rcu(&pa->pa_inode_list);
4189 list_add(&pa->u.pa_tmp_list, &discard_list);
4191 total_entries--;
4192 if (total_entries <= 5) {
4194 * we want to keep only 5 entries
4195 * allowing it to grow to 8. This
4196 * mak sure we don't call discard
4197 * soon for this list.
4199 break;
4202 spin_unlock(&lg->lg_prealloc_lock);
4204 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4206 group = ext4_get_group_number(sb, pa->pa_pstart);
4207 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4208 ext4_error(sb, "Error loading buddy information for %u",
4209 group);
4210 continue;
4212 ext4_lock_group(sb, group);
4213 list_del(&pa->pa_group_list);
4214 ext4_mb_release_group_pa(&e4b, pa);
4215 ext4_unlock_group(sb, group);
4217 ext4_mb_unload_buddy(&e4b);
4218 list_del(&pa->u.pa_tmp_list);
4219 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4224 * We have incremented pa_count. So it cannot be freed at this
4225 * point. Also we hold lg_mutex. So no parallel allocation is
4226 * possible from this lg. That means pa_free cannot be updated.
4228 * A parallel ext4_mb_discard_group_preallocations is possible.
4229 * which can cause the lg_prealloc_list to be updated.
4232 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4234 int order, added = 0, lg_prealloc_count = 1;
4235 struct super_block *sb = ac->ac_sb;
4236 struct ext4_locality_group *lg = ac->ac_lg;
4237 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4239 order = fls(pa->pa_free) - 1;
4240 if (order > PREALLOC_TB_SIZE - 1)
4241 /* The max size of hash table is PREALLOC_TB_SIZE */
4242 order = PREALLOC_TB_SIZE - 1;
4243 /* Add the prealloc space to lg */
4244 spin_lock(&lg->lg_prealloc_lock);
4245 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4246 pa_inode_list) {
4247 spin_lock(&tmp_pa->pa_lock);
4248 if (tmp_pa->pa_deleted) {
4249 spin_unlock(&tmp_pa->pa_lock);
4250 continue;
4252 if (!added && pa->pa_free < tmp_pa->pa_free) {
4253 /* Add to the tail of the previous entry */
4254 list_add_tail_rcu(&pa->pa_inode_list,
4255 &tmp_pa->pa_inode_list);
4256 added = 1;
4258 * we want to count the total
4259 * number of entries in the list
4262 spin_unlock(&tmp_pa->pa_lock);
4263 lg_prealloc_count++;
4265 if (!added)
4266 list_add_tail_rcu(&pa->pa_inode_list,
4267 &lg->lg_prealloc_list[order]);
4268 spin_unlock(&lg->lg_prealloc_lock);
4270 /* Now trim the list to be not more than 8 elements */
4271 if (lg_prealloc_count > 8) {
4272 ext4_mb_discard_lg_preallocations(sb, lg,
4273 order, lg_prealloc_count);
4274 return;
4276 return ;
4280 * release all resource we used in allocation
4282 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4284 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4285 struct ext4_prealloc_space *pa = ac->ac_pa;
4286 if (pa) {
4287 if (pa->pa_type == MB_GROUP_PA) {
4288 /* see comment in ext4_mb_use_group_pa() */
4289 spin_lock(&pa->pa_lock);
4290 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4291 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4292 pa->pa_free -= ac->ac_b_ex.fe_len;
4293 pa->pa_len -= ac->ac_b_ex.fe_len;
4294 spin_unlock(&pa->pa_lock);
4297 if (pa) {
4299 * We want to add the pa to the right bucket.
4300 * Remove it from the list and while adding
4301 * make sure the list to which we are adding
4302 * doesn't grow big.
4304 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4305 spin_lock(pa->pa_obj_lock);
4306 list_del_rcu(&pa->pa_inode_list);
4307 spin_unlock(pa->pa_obj_lock);
4308 ext4_mb_add_n_trim(ac);
4310 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4312 if (ac->ac_bitmap_page)
4313 page_cache_release(ac->ac_bitmap_page);
4314 if (ac->ac_buddy_page)
4315 page_cache_release(ac->ac_buddy_page);
4316 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4317 mutex_unlock(&ac->ac_lg->lg_mutex);
4318 ext4_mb_collect_stats(ac);
4319 return 0;
4322 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4324 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4325 int ret;
4326 int freed = 0;
4328 trace_ext4_mb_discard_preallocations(sb, needed);
4329 for (i = 0; i < ngroups && needed > 0; i++) {
4330 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4331 freed += ret;
4332 needed -= ret;
4335 return freed;
4339 * Main entry point into mballoc to allocate blocks
4340 * it tries to use preallocation first, then falls back
4341 * to usual allocation
4343 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4344 struct ext4_allocation_request *ar, int *errp)
4346 int freed;
4347 struct ext4_allocation_context *ac = NULL;
4348 struct ext4_sb_info *sbi;
4349 struct super_block *sb;
4350 ext4_fsblk_t block = 0;
4351 unsigned int inquota = 0;
4352 unsigned int reserv_clstrs = 0;
4354 might_sleep();
4355 sb = ar->inode->i_sb;
4356 sbi = EXT4_SB(sb);
4358 trace_ext4_request_blocks(ar);
4360 /* Allow to use superuser reservation for quota file */
4361 if (IS_NOQUOTA(ar->inode))
4362 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4365 * For delayed allocation, we could skip the ENOSPC and
4366 * EDQUOT check, as blocks and quotas have been already
4367 * reserved when data being copied into pagecache.
4369 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4370 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4371 else {
4372 /* Without delayed allocation we need to verify
4373 * there is enough free blocks to do block allocation
4374 * and verify allocation doesn't exceed the quota limits.
4376 while (ar->len &&
4377 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4379 /* let others to free the space */
4380 cond_resched();
4381 ar->len = ar->len >> 1;
4383 if (!ar->len) {
4384 *errp = -ENOSPC;
4385 return 0;
4387 reserv_clstrs = ar->len;
4388 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4389 dquot_alloc_block_nofail(ar->inode,
4390 EXT4_C2B(sbi, ar->len));
4391 } else {
4392 while (ar->len &&
4393 dquot_alloc_block(ar->inode,
4394 EXT4_C2B(sbi, ar->len))) {
4396 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4397 ar->len--;
4400 inquota = ar->len;
4401 if (ar->len == 0) {
4402 *errp = -EDQUOT;
4403 goto out;
4407 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4408 if (!ac) {
4409 ar->len = 0;
4410 *errp = -ENOMEM;
4411 goto out;
4414 *errp = ext4_mb_initialize_context(ac, ar);
4415 if (*errp) {
4416 ar->len = 0;
4417 goto out;
4420 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4421 if (!ext4_mb_use_preallocated(ac)) {
4422 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4423 ext4_mb_normalize_request(ac, ar);
4424 repeat:
4425 /* allocate space in core */
4426 *errp = ext4_mb_regular_allocator(ac);
4427 if (*errp)
4428 goto discard_and_exit;
4430 /* as we've just preallocated more space than
4431 * user requested originally, we store allocated
4432 * space in a special descriptor */
4433 if (ac->ac_status == AC_STATUS_FOUND &&
4434 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4435 *errp = ext4_mb_new_preallocation(ac);
4436 if (*errp) {
4437 discard_and_exit:
4438 ext4_discard_allocated_blocks(ac);
4439 goto errout;
4442 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4443 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4444 if (*errp == -EAGAIN) {
4446 * drop the reference that we took
4447 * in ext4_mb_use_best_found
4449 ext4_mb_release_context(ac);
4450 ac->ac_b_ex.fe_group = 0;
4451 ac->ac_b_ex.fe_start = 0;
4452 ac->ac_b_ex.fe_len = 0;
4453 ac->ac_status = AC_STATUS_CONTINUE;
4454 goto repeat;
4455 } else if (*errp) {
4456 ext4_discard_allocated_blocks(ac);
4457 goto errout;
4458 } else {
4459 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4460 ar->len = ac->ac_b_ex.fe_len;
4462 } else {
4463 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4464 if (freed)
4465 goto repeat;
4466 *errp = -ENOSPC;
4469 errout:
4470 if (*errp) {
4471 ac->ac_b_ex.fe_len = 0;
4472 ar->len = 0;
4473 ext4_mb_show_ac(ac);
4475 ext4_mb_release_context(ac);
4476 out:
4477 if (ac)
4478 kmem_cache_free(ext4_ac_cachep, ac);
4479 if (inquota && ar->len < inquota)
4480 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4481 if (!ar->len) {
4482 if (!ext4_test_inode_state(ar->inode,
4483 EXT4_STATE_DELALLOC_RESERVED))
4484 /* release all the reserved blocks if non delalloc */
4485 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4486 reserv_clstrs);
4489 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4491 return block;
4495 * We can merge two free data extents only if the physical blocks
4496 * are contiguous, AND the extents were freed by the same transaction,
4497 * AND the blocks are associated with the same group.
4499 static int can_merge(struct ext4_free_data *entry1,
4500 struct ext4_free_data *entry2)
4502 if ((entry1->efd_tid == entry2->efd_tid) &&
4503 (entry1->efd_group == entry2->efd_group) &&
4504 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4505 return 1;
4506 return 0;
4509 static noinline_for_stack int
4510 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4511 struct ext4_free_data *new_entry)
4513 ext4_group_t group = e4b->bd_group;
4514 ext4_grpblk_t cluster;
4515 struct ext4_free_data *entry;
4516 struct ext4_group_info *db = e4b->bd_info;
4517 struct super_block *sb = e4b->bd_sb;
4518 struct ext4_sb_info *sbi = EXT4_SB(sb);
4519 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4520 struct rb_node *parent = NULL, *new_node;
4522 BUG_ON(!ext4_handle_valid(handle));
4523 BUG_ON(e4b->bd_bitmap_page == NULL);
4524 BUG_ON(e4b->bd_buddy_page == NULL);
4526 new_node = &new_entry->efd_node;
4527 cluster = new_entry->efd_start_cluster;
4529 if (!*n) {
4530 /* first free block exent. We need to
4531 protect buddy cache from being freed,
4532 * otherwise we'll refresh it from
4533 * on-disk bitmap and lose not-yet-available
4534 * blocks */
4535 page_cache_get(e4b->bd_buddy_page);
4536 page_cache_get(e4b->bd_bitmap_page);
4538 while (*n) {
4539 parent = *n;
4540 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4541 if (cluster < entry->efd_start_cluster)
4542 n = &(*n)->rb_left;
4543 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4544 n = &(*n)->rb_right;
4545 else {
4546 ext4_grp_locked_error(sb, group, 0,
4547 ext4_group_first_block_no(sb, group) +
4548 EXT4_C2B(sbi, cluster),
4549 "Block already on to-be-freed list");
4550 return 0;
4554 rb_link_node(new_node, parent, n);
4555 rb_insert_color(new_node, &db->bb_free_root);
4557 /* Now try to see the extent can be merged to left and right */
4558 node = rb_prev(new_node);
4559 if (node) {
4560 entry = rb_entry(node, struct ext4_free_data, efd_node);
4561 if (can_merge(entry, new_entry) &&
4562 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4563 new_entry->efd_start_cluster = entry->efd_start_cluster;
4564 new_entry->efd_count += entry->efd_count;
4565 rb_erase(node, &(db->bb_free_root));
4566 kmem_cache_free(ext4_free_data_cachep, entry);
4570 node = rb_next(new_node);
4571 if (node) {
4572 entry = rb_entry(node, struct ext4_free_data, efd_node);
4573 if (can_merge(new_entry, entry) &&
4574 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4575 new_entry->efd_count += entry->efd_count;
4576 rb_erase(node, &(db->bb_free_root));
4577 kmem_cache_free(ext4_free_data_cachep, entry);
4580 /* Add the extent to transaction's private list */
4581 ext4_journal_callback_add(handle, ext4_free_data_callback,
4582 &new_entry->efd_jce);
4583 return 0;
4587 * ext4_free_blocks() -- Free given blocks and update quota
4588 * @handle: handle for this transaction
4589 * @inode: inode
4590 * @block: start physical block to free
4591 * @count: number of blocks to count
4592 * @flags: flags used by ext4_free_blocks
4594 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4595 struct buffer_head *bh, ext4_fsblk_t block,
4596 unsigned long count, int flags)
4598 struct buffer_head *bitmap_bh = NULL;
4599 struct super_block *sb = inode->i_sb;
4600 struct ext4_group_desc *gdp;
4601 unsigned int overflow;
4602 ext4_grpblk_t bit;
4603 struct buffer_head *gd_bh;
4604 ext4_group_t block_group;
4605 struct ext4_sb_info *sbi;
4606 struct ext4_inode_info *ei = EXT4_I(inode);
4607 struct ext4_buddy e4b;
4608 unsigned int count_clusters;
4609 int err = 0;
4610 int ret;
4612 might_sleep();
4613 if (bh) {
4614 if (block)
4615 BUG_ON(block != bh->b_blocknr);
4616 else
4617 block = bh->b_blocknr;
4620 sbi = EXT4_SB(sb);
4621 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4622 !ext4_data_block_valid(sbi, block, count)) {
4623 ext4_error(sb, "Freeing blocks not in datazone - "
4624 "block = %llu, count = %lu", block, count);
4625 goto error_return;
4628 ext4_debug("freeing block %llu\n", block);
4629 trace_ext4_free_blocks(inode, block, count, flags);
4631 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4632 struct buffer_head *tbh = bh;
4633 int i;
4635 BUG_ON(bh && (count > 1));
4637 for (i = 0; i < count; i++) {
4638 cond_resched();
4639 if (!bh)
4640 tbh = sb_find_get_block(inode->i_sb,
4641 block + i);
4642 if (!tbh)
4643 continue;
4644 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4645 inode, tbh, block + i);
4650 * We need to make sure we don't reuse the freed block until
4651 * after the transaction is committed, which we can do by
4652 * treating the block as metadata, below. We make an
4653 * exception if the inode is to be written in writeback mode
4654 * since writeback mode has weak data consistency guarantees.
4656 if (!ext4_should_writeback_data(inode))
4657 flags |= EXT4_FREE_BLOCKS_METADATA;
4660 * If the extent to be freed does not begin on a cluster
4661 * boundary, we need to deal with partial clusters at the
4662 * beginning and end of the extent. Normally we will free
4663 * blocks at the beginning or the end unless we are explicitly
4664 * requested to avoid doing so.
4666 overflow = block & (sbi->s_cluster_ratio - 1);
4667 if (overflow) {
4668 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4669 overflow = sbi->s_cluster_ratio - overflow;
4670 block += overflow;
4671 if (count > overflow)
4672 count -= overflow;
4673 else
4674 return;
4675 } else {
4676 block -= overflow;
4677 count += overflow;
4680 overflow = count & (sbi->s_cluster_ratio - 1);
4681 if (overflow) {
4682 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4683 if (count > overflow)
4684 count -= overflow;
4685 else
4686 return;
4687 } else
4688 count += sbi->s_cluster_ratio - overflow;
4691 do_more:
4692 overflow = 0;
4693 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4695 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4696 ext4_get_group_info(sb, block_group))))
4697 return;
4700 * Check to see if we are freeing blocks across a group
4701 * boundary.
4703 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4704 overflow = EXT4_C2B(sbi, bit) + count -
4705 EXT4_BLOCKS_PER_GROUP(sb);
4706 count -= overflow;
4708 count_clusters = EXT4_NUM_B2C(sbi, count);
4709 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4710 if (!bitmap_bh) {
4711 err = -EIO;
4712 goto error_return;
4714 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4715 if (!gdp) {
4716 err = -EIO;
4717 goto error_return;
4720 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4721 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4722 in_range(block, ext4_inode_table(sb, gdp),
4723 EXT4_SB(sb)->s_itb_per_group) ||
4724 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4725 EXT4_SB(sb)->s_itb_per_group)) {
4727 ext4_error(sb, "Freeing blocks in system zone - "
4728 "Block = %llu, count = %lu", block, count);
4729 /* err = 0. ext4_std_error should be a no op */
4730 goto error_return;
4733 BUFFER_TRACE(bitmap_bh, "getting write access");
4734 err = ext4_journal_get_write_access(handle, bitmap_bh);
4735 if (err)
4736 goto error_return;
4739 * We are about to modify some metadata. Call the journal APIs
4740 * to unshare ->b_data if a currently-committing transaction is
4741 * using it
4743 BUFFER_TRACE(gd_bh, "get_write_access");
4744 err = ext4_journal_get_write_access(handle, gd_bh);
4745 if (err)
4746 goto error_return;
4747 #ifdef AGGRESSIVE_CHECK
4749 int i;
4750 for (i = 0; i < count_clusters; i++)
4751 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4753 #endif
4754 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4756 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4757 if (err)
4758 goto error_return;
4760 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4761 struct ext4_free_data *new_entry;
4763 * blocks being freed are metadata. these blocks shouldn't
4764 * be used until this transaction is committed
4766 retry:
4767 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4768 if (!new_entry) {
4770 * We use a retry loop because
4771 * ext4_free_blocks() is not allowed to fail.
4773 cond_resched();
4774 congestion_wait(BLK_RW_ASYNC, HZ/50);
4775 goto retry;
4777 new_entry->efd_start_cluster = bit;
4778 new_entry->efd_group = block_group;
4779 new_entry->efd_count = count_clusters;
4780 new_entry->efd_tid = handle->h_transaction->t_tid;
4782 ext4_lock_group(sb, block_group);
4783 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4784 ext4_mb_free_metadata(handle, &e4b, new_entry);
4785 } else {
4786 /* need to update group_info->bb_free and bitmap
4787 * with group lock held. generate_buddy look at
4788 * them with group lock_held
4790 if (test_opt(sb, DISCARD)) {
4791 err = ext4_issue_discard(sb, block_group, bit, count);
4792 if (err && err != -EOPNOTSUPP)
4793 ext4_msg(sb, KERN_WARNING, "discard request in"
4794 " group:%d block:%d count:%lu failed"
4795 " with %d", block_group, bit, count,
4796 err);
4800 ext4_lock_group(sb, block_group);
4801 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4802 mb_free_blocks(inode, &e4b, bit, count_clusters);
4805 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4806 ext4_free_group_clusters_set(sb, gdp, ret);
4807 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4808 ext4_group_desc_csum_set(sb, block_group, gdp);
4809 ext4_unlock_group(sb, block_group);
4811 if (sbi->s_log_groups_per_flex) {
4812 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4813 atomic64_add(count_clusters,
4814 &sbi->s_flex_groups[flex_group].free_clusters);
4817 if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) {
4818 percpu_counter_add(&sbi->s_dirtyclusters_counter,
4819 count_clusters);
4820 spin_lock(&ei->i_block_reservation_lock);
4821 if (flags & EXT4_FREE_BLOCKS_METADATA)
4822 ei->i_reserved_meta_blocks += count_clusters;
4823 else
4824 ei->i_reserved_data_blocks += count_clusters;
4825 spin_unlock(&ei->i_block_reservation_lock);
4826 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4827 dquot_reclaim_block(inode,
4828 EXT4_C2B(sbi, count_clusters));
4829 } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4830 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4831 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4833 ext4_mb_unload_buddy(&e4b);
4835 /* We dirtied the bitmap block */
4836 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4837 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4839 /* And the group descriptor block */
4840 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4841 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4842 if (!err)
4843 err = ret;
4845 if (overflow && !err) {
4846 block += count;
4847 count = overflow;
4848 put_bh(bitmap_bh);
4849 goto do_more;
4851 error_return:
4852 brelse(bitmap_bh);
4853 ext4_std_error(sb, err);
4854 return;
4858 * ext4_group_add_blocks() -- Add given blocks to an existing group
4859 * @handle: handle to this transaction
4860 * @sb: super block
4861 * @block: start physical block to add to the block group
4862 * @count: number of blocks to free
4864 * This marks the blocks as free in the bitmap and buddy.
4866 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4867 ext4_fsblk_t block, unsigned long count)
4869 struct buffer_head *bitmap_bh = NULL;
4870 struct buffer_head *gd_bh;
4871 ext4_group_t block_group;
4872 ext4_grpblk_t bit;
4873 unsigned int i;
4874 struct ext4_group_desc *desc;
4875 struct ext4_sb_info *sbi = EXT4_SB(sb);
4876 struct ext4_buddy e4b;
4877 int err = 0, ret, blk_free_count;
4878 ext4_grpblk_t blocks_freed;
4880 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4882 if (count == 0)
4883 return 0;
4885 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4887 * Check to see if we are freeing blocks across a group
4888 * boundary.
4890 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4891 ext4_warning(sb, "too much blocks added to group %u\n",
4892 block_group);
4893 err = -EINVAL;
4894 goto error_return;
4897 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4898 if (!bitmap_bh) {
4899 err = -EIO;
4900 goto error_return;
4903 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4904 if (!desc) {
4905 err = -EIO;
4906 goto error_return;
4909 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4910 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4911 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4912 in_range(block + count - 1, ext4_inode_table(sb, desc),
4913 sbi->s_itb_per_group)) {
4914 ext4_error(sb, "Adding blocks in system zones - "
4915 "Block = %llu, count = %lu",
4916 block, count);
4917 err = -EINVAL;
4918 goto error_return;
4921 BUFFER_TRACE(bitmap_bh, "getting write access");
4922 err = ext4_journal_get_write_access(handle, bitmap_bh);
4923 if (err)
4924 goto error_return;
4927 * We are about to modify some metadata. Call the journal APIs
4928 * to unshare ->b_data if a currently-committing transaction is
4929 * using it
4931 BUFFER_TRACE(gd_bh, "get_write_access");
4932 err = ext4_journal_get_write_access(handle, gd_bh);
4933 if (err)
4934 goto error_return;
4936 for (i = 0, blocks_freed = 0; i < count; i++) {
4937 BUFFER_TRACE(bitmap_bh, "clear bit");
4938 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4939 ext4_error(sb, "bit already cleared for block %llu",
4940 (ext4_fsblk_t)(block + i));
4941 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4942 } else {
4943 blocks_freed++;
4947 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4948 if (err)
4949 goto error_return;
4952 * need to update group_info->bb_free and bitmap
4953 * with group lock held. generate_buddy look at
4954 * them with group lock_held
4956 ext4_lock_group(sb, block_group);
4957 mb_clear_bits(bitmap_bh->b_data, bit, count);
4958 mb_free_blocks(NULL, &e4b, bit, count);
4959 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4960 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4961 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4962 ext4_group_desc_csum_set(sb, block_group, desc);
4963 ext4_unlock_group(sb, block_group);
4964 percpu_counter_add(&sbi->s_freeclusters_counter,
4965 EXT4_NUM_B2C(sbi, blocks_freed));
4967 if (sbi->s_log_groups_per_flex) {
4968 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4969 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4970 &sbi->s_flex_groups[flex_group].free_clusters);
4973 ext4_mb_unload_buddy(&e4b);
4975 /* We dirtied the bitmap block */
4976 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4977 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4979 /* And the group descriptor block */
4980 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4981 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4982 if (!err)
4983 err = ret;
4985 error_return:
4986 brelse(bitmap_bh);
4987 ext4_std_error(sb, err);
4988 return err;
4992 * ext4_trim_extent -- function to TRIM one single free extent in the group
4993 * @sb: super block for the file system
4994 * @start: starting block of the free extent in the alloc. group
4995 * @count: number of blocks to TRIM
4996 * @group: alloc. group we are working with
4997 * @e4b: ext4 buddy for the group
4999 * Trim "count" blocks starting at "start" in the "group". To assure that no
5000 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5001 * be called with under the group lock.
5003 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5004 ext4_group_t group, struct ext4_buddy *e4b)
5006 struct ext4_free_extent ex;
5007 int ret = 0;
5009 trace_ext4_trim_extent(sb, group, start, count);
5011 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5013 ex.fe_start = start;
5014 ex.fe_group = group;
5015 ex.fe_len = count;
5018 * Mark blocks used, so no one can reuse them while
5019 * being trimmed.
5021 mb_mark_used(e4b, &ex);
5022 ext4_unlock_group(sb, group);
5023 ret = ext4_issue_discard(sb, group, start, count);
5024 ext4_lock_group(sb, group);
5025 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5026 return ret;
5030 * ext4_trim_all_free -- function to trim all free space in alloc. group
5031 * @sb: super block for file system
5032 * @group: group to be trimmed
5033 * @start: first group block to examine
5034 * @max: last group block to examine
5035 * @minblocks: minimum extent block count
5037 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5038 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5039 * the extent.
5042 * ext4_trim_all_free walks through group's block bitmap searching for free
5043 * extents. When the free extent is found, mark it as used in group buddy
5044 * bitmap. Then issue a TRIM command on this extent and free the extent in
5045 * the group buddy bitmap. This is done until whole group is scanned.
5047 static ext4_grpblk_t
5048 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5049 ext4_grpblk_t start, ext4_grpblk_t max,
5050 ext4_grpblk_t minblocks)
5052 void *bitmap;
5053 ext4_grpblk_t next, count = 0, free_count = 0;
5054 struct ext4_buddy e4b;
5055 int ret = 0;
5057 trace_ext4_trim_all_free(sb, group, start, max);
5059 ret = ext4_mb_load_buddy(sb, group, &e4b);
5060 if (ret) {
5061 ext4_error(sb, "Error in loading buddy "
5062 "information for %u", group);
5063 return ret;
5065 bitmap = e4b.bd_bitmap;
5067 ext4_lock_group(sb, group);
5068 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5069 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5070 goto out;
5072 start = (e4b.bd_info->bb_first_free > start) ?
5073 e4b.bd_info->bb_first_free : start;
5075 while (start <= max) {
5076 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5077 if (start > max)
5078 break;
5079 next = mb_find_next_bit(bitmap, max + 1, start);
5081 if ((next - start) >= minblocks) {
5082 ret = ext4_trim_extent(sb, start,
5083 next - start, group, &e4b);
5084 if (ret && ret != -EOPNOTSUPP)
5085 break;
5086 ret = 0;
5087 count += next - start;
5089 free_count += next - start;
5090 start = next + 1;
5092 if (fatal_signal_pending(current)) {
5093 count = -ERESTARTSYS;
5094 break;
5097 if (need_resched()) {
5098 ext4_unlock_group(sb, group);
5099 cond_resched();
5100 ext4_lock_group(sb, group);
5103 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5104 break;
5107 if (!ret) {
5108 ret = count;
5109 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5111 out:
5112 ext4_unlock_group(sb, group);
5113 ext4_mb_unload_buddy(&e4b);
5115 ext4_debug("trimmed %d blocks in the group %d\n",
5116 count, group);
5118 return ret;
5122 * ext4_trim_fs() -- trim ioctl handle function
5123 * @sb: superblock for filesystem
5124 * @range: fstrim_range structure
5126 * start: First Byte to trim
5127 * len: number of Bytes to trim from start
5128 * minlen: minimum extent length in Bytes
5129 * ext4_trim_fs goes through all allocation groups containing Bytes from
5130 * start to start+len. For each such a group ext4_trim_all_free function
5131 * is invoked to trim all free space.
5133 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5135 struct ext4_group_info *grp;
5136 ext4_group_t group, first_group, last_group;
5137 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5138 uint64_t start, end, minlen, trimmed = 0;
5139 ext4_fsblk_t first_data_blk =
5140 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5141 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5142 int ret = 0;
5144 start = range->start >> sb->s_blocksize_bits;
5145 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5146 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5147 range->minlen >> sb->s_blocksize_bits);
5149 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5150 start >= max_blks ||
5151 range->len < sb->s_blocksize)
5152 return -EINVAL;
5153 if (end >= max_blks)
5154 end = max_blks - 1;
5155 if (end <= first_data_blk)
5156 goto out;
5157 if (start < first_data_blk)
5158 start = first_data_blk;
5160 /* Determine first and last group to examine based on start and end */
5161 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5162 &first_group, &first_cluster);
5163 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5164 &last_group, &last_cluster);
5166 /* end now represents the last cluster to discard in this group */
5167 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5169 for (group = first_group; group <= last_group; group++) {
5170 grp = ext4_get_group_info(sb, group);
5171 /* We only do this if the grp has never been initialized */
5172 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5173 ret = ext4_mb_init_group(sb, group);
5174 if (ret)
5175 break;
5179 * For all the groups except the last one, last cluster will
5180 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5181 * change it for the last group, note that last_cluster is
5182 * already computed earlier by ext4_get_group_no_and_offset()
5184 if (group == last_group)
5185 end = last_cluster;
5187 if (grp->bb_free >= minlen) {
5188 cnt = ext4_trim_all_free(sb, group, first_cluster,
5189 end, minlen);
5190 if (cnt < 0) {
5191 ret = cnt;
5192 break;
5194 trimmed += cnt;
5198 * For every group except the first one, we are sure
5199 * that the first cluster to discard will be cluster #0.
5201 first_cluster = 0;
5204 if (!ret)
5205 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5207 out:
5208 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5209 return ret;