Merge branch 'linux-next' of git://git.kernel.org/pub/scm/linux/kernel/git/konrad...
[zen-stable.git] / fs / ext4 / mballoc.c
blobe2d8be8f28bfb8555644bef5100b2f9b8c2cbe9a
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 "mballoc.h"
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
30 * MUSTDO:
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
34 * TODO v4:
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
38 * - quota
39 * - reservation for superuser
41 * TODO v3:
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
46 * - error handling
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
69 * represented as:
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space (in clusters)
74 * pa_free -> free space available in this prealloc space (in clusters)
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This makes sure that
79 * we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
83 * pa_free.
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list represented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) within the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
105 * inode as:
107 * { page }
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
130 * dependent on the cluster size; for non-bigalloc file systems, it is
131 * 512 blocks. This can be tuned via
132 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
133 * terms of number of blocks. If we have mounted the file system with -O
134 * stripe=<value> option the group prealloc request is normalized to the
135 * the smallest multiple of the stripe value (sbi->s_stripe) which is
136 * greater than the default mb_group_prealloc.
138 * The regular allocator (using the buddy cache) supports a few tunables.
140 * /sys/fs/ext4/<partition>/mb_min_to_scan
141 * /sys/fs/ext4/<partition>/mb_max_to_scan
142 * /sys/fs/ext4/<partition>/mb_order2_req
144 * The regular allocator uses buddy scan only if the request len is power of
145 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
146 * value of s_mb_order2_reqs can be tuned via
147 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
148 * stripe size (sbi->s_stripe), we try to search for contiguous block in
149 * stripe size. This should result in better allocation on RAID setups. If
150 * not, we search in the specific group using bitmap for best extents. The
151 * tunable min_to_scan and max_to_scan control the behaviour here.
152 * min_to_scan indicate how long the mballoc __must__ look for a best
153 * extent and max_to_scan indicates how long the mballoc __can__ look for a
154 * best extent in the found extents. Searching for the blocks starts with
155 * the group specified as the goal value in allocation context via
156 * ac_g_ex. Each group is first checked based on the criteria whether it
157 * can be used for allocation. ext4_mb_good_group explains how the groups are
158 * checked.
160 * Both the prealloc space are getting populated as above. So for the first
161 * request we will hit the buddy cache which will result in this prealloc
162 * space getting filled. The prealloc space is then later used for the
163 * subsequent request.
167 * mballoc operates on the following data:
168 * - on-disk bitmap
169 * - in-core buddy (actually includes buddy and bitmap)
170 * - preallocation descriptors (PAs)
172 * there are two types of preallocations:
173 * - inode
174 * assiged to specific inode and can be used for this inode only.
175 * it describes part of inode's space preallocated to specific
176 * physical blocks. any block from that preallocated can be used
177 * independent. the descriptor just tracks number of blocks left
178 * unused. so, before taking some block from descriptor, one must
179 * make sure corresponded logical block isn't allocated yet. this
180 * also means that freeing any block within descriptor's range
181 * must discard all preallocated blocks.
182 * - locality group
183 * assigned to specific locality group which does not translate to
184 * permanent set of inodes: inode can join and leave group. space
185 * from this type of preallocation can be used for any inode. thus
186 * it's consumed from the beginning to the end.
188 * relation between them can be expressed as:
189 * in-core buddy = on-disk bitmap + preallocation descriptors
191 * this mean blocks mballoc considers used are:
192 * - allocated blocks (persistent)
193 * - preallocated blocks (non-persistent)
195 * consistency in mballoc world means that at any time a block is either
196 * free or used in ALL structures. notice: "any time" should not be read
197 * literally -- time is discrete and delimited by locks.
199 * to keep it simple, we don't use block numbers, instead we count number of
200 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
202 * all operations can be expressed as:
203 * - init buddy: buddy = on-disk + PAs
204 * - new PA: buddy += N; PA = N
205 * - use inode PA: on-disk += N; PA -= N
206 * - discard inode PA buddy -= on-disk - PA; PA = 0
207 * - use locality group PA on-disk += N; PA -= N
208 * - discard locality group PA buddy -= PA; PA = 0
209 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
210 * is used in real operation because we can't know actual used
211 * bits from PA, only from on-disk bitmap
213 * if we follow this strict logic, then all operations above should be atomic.
214 * given some of them can block, we'd have to use something like semaphores
215 * killing performance on high-end SMP hardware. let's try to relax it using
216 * the following knowledge:
217 * 1) if buddy is referenced, it's already initialized
218 * 2) while block is used in buddy and the buddy is referenced,
219 * nobody can re-allocate that block
220 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
221 * bit set and PA claims same block, it's OK. IOW, one can set bit in
222 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * block
225 * so, now we're building a concurrency table:
226 * - init buddy vs.
227 * - new PA
228 * blocks for PA are allocated in the buddy, buddy must be referenced
229 * until PA is linked to allocation group to avoid concurrent buddy init
230 * - use inode PA
231 * we need to make sure that either on-disk bitmap or PA has uptodate data
232 * given (3) we care that PA-=N operation doesn't interfere with init
233 * - discard inode PA
234 * the simplest way would be to have buddy initialized by the discard
235 * - use locality group PA
236 * again PA-=N must be serialized with init
237 * - discard locality group PA
238 * the simplest way would be to have buddy initialized by the discard
239 * - new PA vs.
240 * - use inode PA
241 * i_data_sem serializes them
242 * - discard inode PA
243 * discard process must wait until PA isn't used by another process
244 * - use locality group PA
245 * some mutex should serialize them
246 * - discard locality group PA
247 * discard process must wait until PA isn't used by another process
248 * - use inode PA
249 * - use inode PA
250 * i_data_sem or another mutex should 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 * nothing wrong here -- they're different PAs covering different blocks
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
258 * now we're ready to make few consequences:
259 * - PA is referenced and while it is no discard is possible
260 * - PA is referenced until block isn't marked in on-disk bitmap
261 * - PA changes only after on-disk bitmap
262 * - discard must not compete with init. either init is done before
263 * any discard or they're serialized somehow
264 * - buddy init as sum of on-disk bitmap and PAs is done atomically
266 * a special case when we've used PA to emptiness. no need to modify buddy
267 * in this case, but we should care about concurrent init
272 * Logic in few words:
274 * - allocation:
275 * load group
276 * find blocks
277 * mark bits in on-disk bitmap
278 * release group
280 * - use preallocation:
281 * find proper PA (per-inode or group)
282 * load group
283 * mark bits in on-disk bitmap
284 * release group
285 * release PA
287 * - free:
288 * load group
289 * mark bits in on-disk bitmap
290 * release group
292 * - discard preallocations in group:
293 * mark PAs deleted
294 * move them onto local list
295 * load on-disk bitmap
296 * load group
297 * remove PA from object (inode or locality group)
298 * mark free blocks in-core
300 * - discard inode's preallocations:
304 * Locking rules
306 * Locks:
307 * - bitlock on a group (group)
308 * - object (inode/locality) (object)
309 * - per-pa lock (pa)
311 * Paths:
312 * - new pa
313 * object
314 * group
316 * - find and use pa:
317 * pa
319 * - release consumed pa:
320 * pa
321 * group
322 * object
324 * - generate in-core bitmap:
325 * group
326 * pa
328 * - discard all for given object (inode, locality group):
329 * object
330 * pa
331 * group
333 * - discard all for given group:
334 * group
335 * pa
336 * group
337 * object
340 static struct kmem_cache *ext4_pspace_cachep;
341 static struct kmem_cache *ext4_ac_cachep;
342 static struct kmem_cache *ext4_free_ext_cachep;
344 /* We create slab caches for groupinfo data structures based on the
345 * superblock block size. There will be one per mounted filesystem for
346 * each unique s_blocksize_bits */
347 #define NR_GRPINFO_CACHES 8
348 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
350 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
351 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
352 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
353 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
356 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
357 ext4_group_t group);
358 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
359 ext4_group_t group);
360 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
362 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
364 #if BITS_PER_LONG == 64
365 *bit += ((unsigned long) addr & 7UL) << 3;
366 addr = (void *) ((unsigned long) addr & ~7UL);
367 #elif BITS_PER_LONG == 32
368 *bit += ((unsigned long) addr & 3UL) << 3;
369 addr = (void *) ((unsigned long) addr & ~3UL);
370 #else
371 #error "how many bits you are?!"
372 #endif
373 return addr;
376 static inline int mb_test_bit(int bit, void *addr)
379 * ext4_test_bit on architecture like powerpc
380 * needs unsigned long aligned address
382 addr = mb_correct_addr_and_bit(&bit, addr);
383 return ext4_test_bit(bit, addr);
386 static inline void mb_set_bit(int bit, void *addr)
388 addr = mb_correct_addr_and_bit(&bit, addr);
389 ext4_set_bit(bit, addr);
392 static inline void mb_clear_bit(int bit, void *addr)
394 addr = mb_correct_addr_and_bit(&bit, addr);
395 ext4_clear_bit(bit, addr);
398 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
400 int fix = 0, ret, tmpmax;
401 addr = mb_correct_addr_and_bit(&fix, addr);
402 tmpmax = max + fix;
403 start += fix;
405 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
406 if (ret > max)
407 return max;
408 return ret;
411 static inline int mb_find_next_bit(void *addr, int max, int start)
413 int fix = 0, ret, tmpmax;
414 addr = mb_correct_addr_and_bit(&fix, addr);
415 tmpmax = max + fix;
416 start += fix;
418 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
419 if (ret > max)
420 return max;
421 return ret;
424 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
426 char *bb;
428 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
429 BUG_ON(max == NULL);
431 if (order > e4b->bd_blkbits + 1) {
432 *max = 0;
433 return NULL;
436 /* at order 0 we see each particular block */
437 if (order == 0) {
438 *max = 1 << (e4b->bd_blkbits + 3);
439 return EXT4_MB_BITMAP(e4b);
442 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
443 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
445 return bb;
448 #ifdef DOUBLE_CHECK
449 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
450 int first, int count)
452 int i;
453 struct super_block *sb = e4b->bd_sb;
455 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
456 return;
457 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
458 for (i = 0; i < count; i++) {
459 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
460 ext4_fsblk_t blocknr;
462 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
463 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
464 ext4_grp_locked_error(sb, e4b->bd_group,
465 inode ? inode->i_ino : 0,
466 blocknr,
467 "freeing block already freed "
468 "(bit %u)",
469 first + i);
471 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
475 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
477 int i;
479 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
480 return;
481 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
482 for (i = 0; i < count; i++) {
483 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
484 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
488 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
490 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
491 unsigned char *b1, *b2;
492 int i;
493 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
494 b2 = (unsigned char *) bitmap;
495 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
496 if (b1[i] != b2[i]) {
497 ext4_msg(e4b->bd_sb, KERN_ERR,
498 "corruption in group %u "
499 "at byte %u(%u): %x in copy != %x "
500 "on disk/prealloc",
501 e4b->bd_group, i, i * 8, b1[i], b2[i]);
502 BUG();
508 #else
509 static inline void mb_free_blocks_double(struct inode *inode,
510 struct ext4_buddy *e4b, int first, int count)
512 return;
514 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
515 int first, int count)
517 return;
519 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
521 return;
523 #endif
525 #ifdef AGGRESSIVE_CHECK
527 #define MB_CHECK_ASSERT(assert) \
528 do { \
529 if (!(assert)) { \
530 printk(KERN_EMERG \
531 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
532 function, file, line, # assert); \
533 BUG(); \
535 } while (0)
537 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
538 const char *function, int line)
540 struct super_block *sb = e4b->bd_sb;
541 int order = e4b->bd_blkbits + 1;
542 int max;
543 int max2;
544 int i;
545 int j;
546 int k;
547 int count;
548 struct ext4_group_info *grp;
549 int fragments = 0;
550 int fstart;
551 struct list_head *cur;
552 void *buddy;
553 void *buddy2;
556 static int mb_check_counter;
557 if (mb_check_counter++ % 100 != 0)
558 return 0;
561 while (order > 1) {
562 buddy = mb_find_buddy(e4b, order, &max);
563 MB_CHECK_ASSERT(buddy);
564 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
565 MB_CHECK_ASSERT(buddy2);
566 MB_CHECK_ASSERT(buddy != buddy2);
567 MB_CHECK_ASSERT(max * 2 == max2);
569 count = 0;
570 for (i = 0; i < max; i++) {
572 if (mb_test_bit(i, buddy)) {
573 /* only single bit in buddy2 may be 1 */
574 if (!mb_test_bit(i << 1, buddy2)) {
575 MB_CHECK_ASSERT(
576 mb_test_bit((i<<1)+1, buddy2));
577 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
578 MB_CHECK_ASSERT(
579 mb_test_bit(i << 1, buddy2));
581 continue;
584 /* both bits in buddy2 must be 1 */
585 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
586 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
588 for (j = 0; j < (1 << order); j++) {
589 k = (i * (1 << order)) + j;
590 MB_CHECK_ASSERT(
591 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
593 count++;
595 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
596 order--;
599 fstart = -1;
600 buddy = mb_find_buddy(e4b, 0, &max);
601 for (i = 0; i < max; i++) {
602 if (!mb_test_bit(i, buddy)) {
603 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
604 if (fstart == -1) {
605 fragments++;
606 fstart = i;
608 continue;
610 fstart = -1;
611 /* check used bits only */
612 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
613 buddy2 = mb_find_buddy(e4b, j, &max2);
614 k = i >> j;
615 MB_CHECK_ASSERT(k < max2);
616 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
619 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
620 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
622 grp = ext4_get_group_info(sb, e4b->bd_group);
623 list_for_each(cur, &grp->bb_prealloc_list) {
624 ext4_group_t groupnr;
625 struct ext4_prealloc_space *pa;
626 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
627 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
628 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
629 for (i = 0; i < pa->pa_len; i++)
630 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
632 return 0;
634 #undef MB_CHECK_ASSERT
635 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
636 __FILE__, __func__, __LINE__)
637 #else
638 #define mb_check_buddy(e4b)
639 #endif
642 * Divide blocks started from @first with length @len into
643 * smaller chunks with power of 2 blocks.
644 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
645 * then increase bb_counters[] for corresponded chunk size.
647 static void ext4_mb_mark_free_simple(struct super_block *sb,
648 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
649 struct ext4_group_info *grp)
651 struct ext4_sb_info *sbi = EXT4_SB(sb);
652 ext4_grpblk_t min;
653 ext4_grpblk_t max;
654 ext4_grpblk_t chunk;
655 unsigned short border;
657 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
659 border = 2 << sb->s_blocksize_bits;
661 while (len > 0) {
662 /* find how many blocks can be covered since this position */
663 max = ffs(first | border) - 1;
665 /* find how many blocks of power 2 we need to mark */
666 min = fls(len) - 1;
668 if (max < min)
669 min = max;
670 chunk = 1 << min;
672 /* mark multiblock chunks only */
673 grp->bb_counters[min]++;
674 if (min > 0)
675 mb_clear_bit(first >> min,
676 buddy + sbi->s_mb_offsets[min]);
678 len -= chunk;
679 first += chunk;
684 * Cache the order of the largest free extent we have available in this block
685 * group.
687 static void
688 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
690 int i;
691 int bits;
693 grp->bb_largest_free_order = -1; /* uninit */
695 bits = sb->s_blocksize_bits + 1;
696 for (i = bits; i >= 0; i--) {
697 if (grp->bb_counters[i] > 0) {
698 grp->bb_largest_free_order = i;
699 break;
704 static noinline_for_stack
705 void ext4_mb_generate_buddy(struct super_block *sb,
706 void *buddy, void *bitmap, ext4_group_t group)
708 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
709 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
710 ext4_grpblk_t i = 0;
711 ext4_grpblk_t first;
712 ext4_grpblk_t len;
713 unsigned free = 0;
714 unsigned fragments = 0;
715 unsigned long long period = get_cycles();
717 /* initialize buddy from bitmap which is aggregation
718 * of on-disk bitmap and preallocations */
719 i = mb_find_next_zero_bit(bitmap, max, 0);
720 grp->bb_first_free = i;
721 while (i < max) {
722 fragments++;
723 first = i;
724 i = mb_find_next_bit(bitmap, max, i);
725 len = i - first;
726 free += len;
727 if (len > 1)
728 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
729 else
730 grp->bb_counters[0]++;
731 if (i < max)
732 i = mb_find_next_zero_bit(bitmap, max, i);
734 grp->bb_fragments = fragments;
736 if (free != grp->bb_free) {
737 ext4_grp_locked_error(sb, group, 0, 0,
738 "%u clusters in bitmap, %u in gd",
739 free, grp->bb_free);
741 * If we intent to continue, we consider group descritor
742 * corrupt and update bb_free using bitmap value
744 grp->bb_free = free;
746 mb_set_largest_free_order(sb, grp);
748 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
750 period = get_cycles() - period;
751 spin_lock(&EXT4_SB(sb)->s_bal_lock);
752 EXT4_SB(sb)->s_mb_buddies_generated++;
753 EXT4_SB(sb)->s_mb_generation_time += period;
754 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
757 /* The buddy information is attached the buddy cache inode
758 * for convenience. The information regarding each group
759 * is loaded via ext4_mb_load_buddy. The information involve
760 * block bitmap and buddy information. The information are
761 * stored in the inode as
763 * { page }
764 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
767 * one block each for bitmap and buddy information.
768 * So for each group we take up 2 blocks. A page can
769 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
770 * So it can have information regarding groups_per_page which
771 * is blocks_per_page/2
773 * Locking note: This routine takes the block group lock of all groups
774 * for this page; do not hold this lock when calling this routine!
777 static int ext4_mb_init_cache(struct page *page, char *incore)
779 ext4_group_t ngroups;
780 int blocksize;
781 int blocks_per_page;
782 int groups_per_page;
783 int err = 0;
784 int i;
785 ext4_group_t first_group;
786 int first_block;
787 struct super_block *sb;
788 struct buffer_head *bhs;
789 struct buffer_head **bh;
790 struct inode *inode;
791 char *data;
792 char *bitmap;
793 struct ext4_group_info *grinfo;
795 mb_debug(1, "init page %lu\n", page->index);
797 inode = page->mapping->host;
798 sb = inode->i_sb;
799 ngroups = ext4_get_groups_count(sb);
800 blocksize = 1 << inode->i_blkbits;
801 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
803 groups_per_page = blocks_per_page >> 1;
804 if (groups_per_page == 0)
805 groups_per_page = 1;
807 /* allocate buffer_heads to read bitmaps */
808 if (groups_per_page > 1) {
809 err = -ENOMEM;
810 i = sizeof(struct buffer_head *) * groups_per_page;
811 bh = kzalloc(i, GFP_NOFS);
812 if (bh == NULL)
813 goto out;
814 } else
815 bh = &bhs;
817 first_group = page->index * blocks_per_page / 2;
819 /* read all groups the page covers into the cache */
820 for (i = 0; i < groups_per_page; i++) {
821 struct ext4_group_desc *desc;
823 if (first_group + i >= ngroups)
824 break;
826 grinfo = ext4_get_group_info(sb, first_group + i);
828 * If page is uptodate then we came here after online resize
829 * which added some new uninitialized group info structs, so
830 * we must skip all initialized uptodate buddies on the page,
831 * which may be currently in use by an allocating task.
833 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
834 bh[i] = NULL;
835 continue;
838 err = -EIO;
839 desc = ext4_get_group_desc(sb, first_group + i, NULL);
840 if (desc == NULL)
841 goto out;
843 err = -ENOMEM;
844 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
845 if (bh[i] == NULL)
846 goto out;
848 if (bitmap_uptodate(bh[i]))
849 continue;
851 lock_buffer(bh[i]);
852 if (bitmap_uptodate(bh[i])) {
853 unlock_buffer(bh[i]);
854 continue;
856 ext4_lock_group(sb, first_group + i);
857 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
858 ext4_init_block_bitmap(sb, bh[i],
859 first_group + i, desc);
860 set_bitmap_uptodate(bh[i]);
861 set_buffer_uptodate(bh[i]);
862 ext4_unlock_group(sb, first_group + i);
863 unlock_buffer(bh[i]);
864 continue;
866 ext4_unlock_group(sb, first_group + i);
867 if (buffer_uptodate(bh[i])) {
869 * if not uninit if bh is uptodate,
870 * bitmap is also uptodate
872 set_bitmap_uptodate(bh[i]);
873 unlock_buffer(bh[i]);
874 continue;
876 get_bh(bh[i]);
878 * submit the buffer_head for read. We can
879 * safely mark the bitmap as uptodate now.
880 * We do it here so the bitmap uptodate bit
881 * get set with buffer lock held.
883 set_bitmap_uptodate(bh[i]);
884 bh[i]->b_end_io = end_buffer_read_sync;
885 submit_bh(READ, bh[i]);
886 mb_debug(1, "read bitmap for group %u\n", first_group + i);
889 /* wait for I/O completion */
890 for (i = 0; i < groups_per_page; i++)
891 if (bh[i])
892 wait_on_buffer(bh[i]);
894 err = -EIO;
895 for (i = 0; i < groups_per_page; i++)
896 if (bh[i] && !buffer_uptodate(bh[i]))
897 goto out;
899 err = 0;
900 first_block = page->index * blocks_per_page;
901 for (i = 0; i < blocks_per_page; i++) {
902 int group;
904 group = (first_block + i) >> 1;
905 if (group >= ngroups)
906 break;
908 if (!bh[group - first_group])
909 /* skip initialized uptodate buddy */
910 continue;
913 * data carry information regarding this
914 * particular group in the format specified
915 * above
918 data = page_address(page) + (i * blocksize);
919 bitmap = bh[group - first_group]->b_data;
922 * We place the buddy block and bitmap block
923 * close together
925 if ((first_block + i) & 1) {
926 /* this is block of buddy */
927 BUG_ON(incore == NULL);
928 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
929 group, page->index, i * blocksize);
930 trace_ext4_mb_buddy_bitmap_load(sb, group);
931 grinfo = ext4_get_group_info(sb, group);
932 grinfo->bb_fragments = 0;
933 memset(grinfo->bb_counters, 0,
934 sizeof(*grinfo->bb_counters) *
935 (sb->s_blocksize_bits+2));
937 * incore got set to the group block bitmap below
939 ext4_lock_group(sb, group);
940 /* init the buddy */
941 memset(data, 0xff, blocksize);
942 ext4_mb_generate_buddy(sb, data, incore, group);
943 ext4_unlock_group(sb, group);
944 incore = NULL;
945 } else {
946 /* this is block of bitmap */
947 BUG_ON(incore != NULL);
948 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
949 group, page->index, i * blocksize);
950 trace_ext4_mb_bitmap_load(sb, group);
952 /* see comments in ext4_mb_put_pa() */
953 ext4_lock_group(sb, group);
954 memcpy(data, bitmap, blocksize);
956 /* mark all preallocated blks used in in-core bitmap */
957 ext4_mb_generate_from_pa(sb, data, group);
958 ext4_mb_generate_from_freelist(sb, data, group);
959 ext4_unlock_group(sb, group);
961 /* set incore so that the buddy information can be
962 * generated using this
964 incore = data;
967 SetPageUptodate(page);
969 out:
970 if (bh) {
971 for (i = 0; i < groups_per_page; i++)
972 brelse(bh[i]);
973 if (bh != &bhs)
974 kfree(bh);
976 return err;
980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
981 * on the same buddy page doesn't happen whild holding the buddy page lock.
982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
985 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
986 ext4_group_t group, struct ext4_buddy *e4b)
988 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
989 int block, pnum, poff;
990 int blocks_per_page;
991 struct page *page;
993 e4b->bd_buddy_page = NULL;
994 e4b->bd_bitmap_page = NULL;
996 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
998 * the buddy cache inode stores the block bitmap
999 * and buddy information in consecutive blocks.
1000 * So for each group we need two blocks.
1002 block = group * 2;
1003 pnum = block / blocks_per_page;
1004 poff = block % blocks_per_page;
1005 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1006 if (!page)
1007 return -EIO;
1008 BUG_ON(page->mapping != inode->i_mapping);
1009 e4b->bd_bitmap_page = page;
1010 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1012 if (blocks_per_page >= 2) {
1013 /* buddy and bitmap are on the same page */
1014 return 0;
1017 block++;
1018 pnum = block / blocks_per_page;
1019 poff = block % blocks_per_page;
1020 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1021 if (!page)
1022 return -EIO;
1023 BUG_ON(page->mapping != inode->i_mapping);
1024 e4b->bd_buddy_page = page;
1025 return 0;
1028 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1030 if (e4b->bd_bitmap_page) {
1031 unlock_page(e4b->bd_bitmap_page);
1032 page_cache_release(e4b->bd_bitmap_page);
1034 if (e4b->bd_buddy_page) {
1035 unlock_page(e4b->bd_buddy_page);
1036 page_cache_release(e4b->bd_buddy_page);
1041 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1042 * block group lock of all groups for this page; do not hold the BG lock when
1043 * calling this routine!
1045 static noinline_for_stack
1046 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1049 struct ext4_group_info *this_grp;
1050 struct ext4_buddy e4b;
1051 struct page *page;
1052 int ret = 0;
1054 mb_debug(1, "init group %u\n", group);
1055 this_grp = ext4_get_group_info(sb, group);
1057 * This ensures that we don't reinit the buddy cache
1058 * page which map to the group from which we are already
1059 * allocating. If we are looking at the buddy cache we would
1060 * have taken a reference using ext4_mb_load_buddy and that
1061 * would have pinned buddy page to page cache.
1063 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1064 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1066 * somebody initialized the group
1067 * return without doing anything
1069 goto err;
1072 page = e4b.bd_bitmap_page;
1073 ret = ext4_mb_init_cache(page, NULL);
1074 if (ret)
1075 goto err;
1076 if (!PageUptodate(page)) {
1077 ret = -EIO;
1078 goto err;
1080 mark_page_accessed(page);
1082 if (e4b.bd_buddy_page == NULL) {
1084 * If both the bitmap and buddy are in
1085 * the same page we don't need to force
1086 * init the buddy
1088 ret = 0;
1089 goto err;
1091 /* init buddy cache */
1092 page = e4b.bd_buddy_page;
1093 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1094 if (ret)
1095 goto err;
1096 if (!PageUptodate(page)) {
1097 ret = -EIO;
1098 goto err;
1100 mark_page_accessed(page);
1101 err:
1102 ext4_mb_put_buddy_page_lock(&e4b);
1103 return ret;
1107 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1111 static noinline_for_stack int
1112 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b)
1115 int blocks_per_page;
1116 int block;
1117 int pnum;
1118 int poff;
1119 struct page *page;
1120 int ret;
1121 struct ext4_group_info *grp;
1122 struct ext4_sb_info *sbi = EXT4_SB(sb);
1123 struct inode *inode = sbi->s_buddy_cache;
1125 mb_debug(1, "load group %u\n", group);
1127 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1128 grp = ext4_get_group_info(sb, group);
1130 e4b->bd_blkbits = sb->s_blocksize_bits;
1131 e4b->bd_info = grp;
1132 e4b->bd_sb = sb;
1133 e4b->bd_group = group;
1134 e4b->bd_buddy_page = NULL;
1135 e4b->bd_bitmap_page = NULL;
1137 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1139 * we need full data about the group
1140 * to make a good selection
1142 ret = ext4_mb_init_group(sb, group);
1143 if (ret)
1144 return ret;
1148 * the buddy cache inode stores the block bitmap
1149 * and buddy information in consecutive blocks.
1150 * So for each group we need two blocks.
1152 block = group * 2;
1153 pnum = block / blocks_per_page;
1154 poff = block % blocks_per_page;
1156 /* we could use find_or_create_page(), but it locks page
1157 * what we'd like to avoid in fast path ... */
1158 page = find_get_page(inode->i_mapping, pnum);
1159 if (page == NULL || !PageUptodate(page)) {
1160 if (page)
1162 * drop the page reference and try
1163 * to get the page with lock. If we
1164 * are not uptodate that implies
1165 * somebody just created the page but
1166 * is yet to initialize the same. So
1167 * wait for it to initialize.
1169 page_cache_release(page);
1170 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1171 if (page) {
1172 BUG_ON(page->mapping != inode->i_mapping);
1173 if (!PageUptodate(page)) {
1174 ret = ext4_mb_init_cache(page, NULL);
1175 if (ret) {
1176 unlock_page(page);
1177 goto err;
1179 mb_cmp_bitmaps(e4b, page_address(page) +
1180 (poff * sb->s_blocksize));
1182 unlock_page(page);
1185 if (page == NULL || !PageUptodate(page)) {
1186 ret = -EIO;
1187 goto err;
1189 e4b->bd_bitmap_page = page;
1190 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1191 mark_page_accessed(page);
1193 block++;
1194 pnum = block / blocks_per_page;
1195 poff = block % blocks_per_page;
1197 page = find_get_page(inode->i_mapping, pnum);
1198 if (page == NULL || !PageUptodate(page)) {
1199 if (page)
1200 page_cache_release(page);
1201 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1202 if (page) {
1203 BUG_ON(page->mapping != inode->i_mapping);
1204 if (!PageUptodate(page)) {
1205 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1206 if (ret) {
1207 unlock_page(page);
1208 goto err;
1211 unlock_page(page);
1214 if (page == NULL || !PageUptodate(page)) {
1215 ret = -EIO;
1216 goto err;
1218 e4b->bd_buddy_page = page;
1219 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1220 mark_page_accessed(page);
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1225 return 0;
1227 err:
1228 if (page)
1229 page_cache_release(page);
1230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
1236 return ret;
1239 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1248 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1250 int order = 1;
1251 void *bb;
1253 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1256 bb = EXT4_MB_BUDDY(e4b);
1257 while (order <= e4b->bd_blkbits + 1) {
1258 block = block >> 1;
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1261 return order;
1263 bb += 1 << (e4b->bd_blkbits - order);
1264 order++;
1266 return 0;
1269 static void mb_clear_bits(void *bm, int cur, int len)
1271 __u32 *addr;
1273 len = cur + len;
1274 while (cur < len) {
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1278 *addr = 0;
1279 cur += 32;
1280 continue;
1282 mb_clear_bit(cur, bm);
1283 cur++;
1287 void ext4_set_bits(void *bm, int cur, int len)
1289 __u32 *addr;
1291 len = cur + len;
1292 while (cur < len) {
1293 if ((cur & 31) == 0 && (len - cur) >= 32) {
1294 /* fast path: set whole word at once */
1295 addr = bm + (cur >> 3);
1296 *addr = 0xffffffff;
1297 cur += 32;
1298 continue;
1300 mb_set_bit(cur, bm);
1301 cur++;
1305 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1306 int first, int count)
1308 int block = 0;
1309 int max = 0;
1310 int order;
1311 void *buddy;
1312 void *buddy2;
1313 struct super_block *sb = e4b->bd_sb;
1315 BUG_ON(first + count > (sb->s_blocksize << 3));
1316 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1317 mb_check_buddy(e4b);
1318 mb_free_blocks_double(inode, e4b, first, count);
1320 e4b->bd_info->bb_free += count;
1321 if (first < e4b->bd_info->bb_first_free)
1322 e4b->bd_info->bb_first_free = first;
1324 /* let's maintain fragments counter */
1325 if (first != 0)
1326 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1327 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1328 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1329 if (block && max)
1330 e4b->bd_info->bb_fragments--;
1331 else if (!block && !max)
1332 e4b->bd_info->bb_fragments++;
1334 /* let's maintain buddy itself */
1335 while (count-- > 0) {
1336 block = first++;
1337 order = 0;
1339 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1340 ext4_fsblk_t blocknr;
1342 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1343 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1344 ext4_grp_locked_error(sb, e4b->bd_group,
1345 inode ? inode->i_ino : 0,
1346 blocknr,
1347 "freeing already freed block "
1348 "(bit %u)", block);
1350 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1351 e4b->bd_info->bb_counters[order]++;
1353 /* start of the buddy */
1354 buddy = mb_find_buddy(e4b, order, &max);
1356 do {
1357 block &= ~1UL;
1358 if (mb_test_bit(block, buddy) ||
1359 mb_test_bit(block + 1, buddy))
1360 break;
1362 /* both the buddies are free, try to coalesce them */
1363 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1365 if (!buddy2)
1366 break;
1368 if (order > 0) {
1369 /* for special purposes, we don't set
1370 * free bits in bitmap */
1371 mb_set_bit(block, buddy);
1372 mb_set_bit(block + 1, buddy);
1374 e4b->bd_info->bb_counters[order]--;
1375 e4b->bd_info->bb_counters[order]--;
1377 block = block >> 1;
1378 order++;
1379 e4b->bd_info->bb_counters[order]++;
1381 mb_clear_bit(block, buddy2);
1382 buddy = buddy2;
1383 } while (1);
1385 mb_set_largest_free_order(sb, e4b->bd_info);
1386 mb_check_buddy(e4b);
1389 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1390 int needed, struct ext4_free_extent *ex)
1392 int next = block;
1393 int max;
1394 void *buddy;
1396 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1397 BUG_ON(ex == NULL);
1399 buddy = mb_find_buddy(e4b, order, &max);
1400 BUG_ON(buddy == NULL);
1401 BUG_ON(block >= max);
1402 if (mb_test_bit(block, buddy)) {
1403 ex->fe_len = 0;
1404 ex->fe_start = 0;
1405 ex->fe_group = 0;
1406 return 0;
1409 /* FIXME dorp order completely ? */
1410 if (likely(order == 0)) {
1411 /* find actual order */
1412 order = mb_find_order_for_block(e4b, block);
1413 block = block >> order;
1416 ex->fe_len = 1 << order;
1417 ex->fe_start = block << order;
1418 ex->fe_group = e4b->bd_group;
1420 /* calc difference from given start */
1421 next = next - ex->fe_start;
1422 ex->fe_len -= next;
1423 ex->fe_start += next;
1425 while (needed > ex->fe_len &&
1426 (buddy = mb_find_buddy(e4b, order, &max))) {
1428 if (block + 1 >= max)
1429 break;
1431 next = (block + 1) * (1 << order);
1432 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1433 break;
1435 order = mb_find_order_for_block(e4b, next);
1437 block = next >> order;
1438 ex->fe_len += 1 << order;
1441 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1442 return ex->fe_len;
1445 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1447 int ord;
1448 int mlen = 0;
1449 int max = 0;
1450 int cur;
1451 int start = ex->fe_start;
1452 int len = ex->fe_len;
1453 unsigned ret = 0;
1454 int len0 = len;
1455 void *buddy;
1457 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1458 BUG_ON(e4b->bd_group != ex->fe_group);
1459 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1460 mb_check_buddy(e4b);
1461 mb_mark_used_double(e4b, start, len);
1463 e4b->bd_info->bb_free -= len;
1464 if (e4b->bd_info->bb_first_free == start)
1465 e4b->bd_info->bb_first_free += len;
1467 /* let's maintain fragments counter */
1468 if (start != 0)
1469 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1470 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1471 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1472 if (mlen && max)
1473 e4b->bd_info->bb_fragments++;
1474 else if (!mlen && !max)
1475 e4b->bd_info->bb_fragments--;
1477 /* let's maintain buddy itself */
1478 while (len) {
1479 ord = mb_find_order_for_block(e4b, start);
1481 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1482 /* the whole chunk may be allocated at once! */
1483 mlen = 1 << ord;
1484 buddy = mb_find_buddy(e4b, ord, &max);
1485 BUG_ON((start >> ord) >= max);
1486 mb_set_bit(start >> ord, buddy);
1487 e4b->bd_info->bb_counters[ord]--;
1488 start += mlen;
1489 len -= mlen;
1490 BUG_ON(len < 0);
1491 continue;
1494 /* store for history */
1495 if (ret == 0)
1496 ret = len | (ord << 16);
1498 /* we have to split large buddy */
1499 BUG_ON(ord <= 0);
1500 buddy = mb_find_buddy(e4b, ord, &max);
1501 mb_set_bit(start >> ord, buddy);
1502 e4b->bd_info->bb_counters[ord]--;
1504 ord--;
1505 cur = (start >> ord) & ~1U;
1506 buddy = mb_find_buddy(e4b, ord, &max);
1507 mb_clear_bit(cur, buddy);
1508 mb_clear_bit(cur + 1, buddy);
1509 e4b->bd_info->bb_counters[ord]++;
1510 e4b->bd_info->bb_counters[ord]++;
1512 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1514 ext4_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1515 mb_check_buddy(e4b);
1517 return ret;
1521 * Must be called under group lock!
1523 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1524 struct ext4_buddy *e4b)
1526 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1527 int ret;
1529 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1530 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1532 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1533 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1534 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1536 /* preallocation can change ac_b_ex, thus we store actually
1537 * allocated blocks for history */
1538 ac->ac_f_ex = ac->ac_b_ex;
1540 ac->ac_status = AC_STATUS_FOUND;
1541 ac->ac_tail = ret & 0xffff;
1542 ac->ac_buddy = ret >> 16;
1545 * take the page reference. We want the page to be pinned
1546 * so that we don't get a ext4_mb_init_cache_call for this
1547 * group until we update the bitmap. That would mean we
1548 * double allocate blocks. The reference is dropped
1549 * in ext4_mb_release_context
1551 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1552 get_page(ac->ac_bitmap_page);
1553 ac->ac_buddy_page = e4b->bd_buddy_page;
1554 get_page(ac->ac_buddy_page);
1555 /* store last allocated for subsequent stream allocation */
1556 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1557 spin_lock(&sbi->s_md_lock);
1558 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1559 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1560 spin_unlock(&sbi->s_md_lock);
1565 * regular allocator, for general purposes allocation
1568 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1569 struct ext4_buddy *e4b,
1570 int finish_group)
1572 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1573 struct ext4_free_extent *bex = &ac->ac_b_ex;
1574 struct ext4_free_extent *gex = &ac->ac_g_ex;
1575 struct ext4_free_extent ex;
1576 int max;
1578 if (ac->ac_status == AC_STATUS_FOUND)
1579 return;
1581 * We don't want to scan for a whole year
1583 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1584 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1585 ac->ac_status = AC_STATUS_BREAK;
1586 return;
1590 * Haven't found good chunk so far, let's continue
1592 if (bex->fe_len < gex->fe_len)
1593 return;
1595 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1596 && bex->fe_group == e4b->bd_group) {
1597 /* recheck chunk's availability - we don't know
1598 * when it was found (within this lock-unlock
1599 * period or not) */
1600 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1601 if (max >= gex->fe_len) {
1602 ext4_mb_use_best_found(ac, e4b);
1603 return;
1609 * The routine checks whether found extent is good enough. If it is,
1610 * then the extent gets marked used and flag is set to the context
1611 * to stop scanning. Otherwise, the extent is compared with the
1612 * previous found extent and if new one is better, then it's stored
1613 * in the context. Later, the best found extent will be used, if
1614 * mballoc can't find good enough extent.
1616 * FIXME: real allocation policy is to be designed yet!
1618 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1619 struct ext4_free_extent *ex,
1620 struct ext4_buddy *e4b)
1622 struct ext4_free_extent *bex = &ac->ac_b_ex;
1623 struct ext4_free_extent *gex = &ac->ac_g_ex;
1625 BUG_ON(ex->fe_len <= 0);
1626 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1627 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1628 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1630 ac->ac_found++;
1633 * The special case - take what you catch first
1635 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1636 *bex = *ex;
1637 ext4_mb_use_best_found(ac, e4b);
1638 return;
1642 * Let's check whether the chuck is good enough
1644 if (ex->fe_len == gex->fe_len) {
1645 *bex = *ex;
1646 ext4_mb_use_best_found(ac, e4b);
1647 return;
1651 * If this is first found extent, just store it in the context
1653 if (bex->fe_len == 0) {
1654 *bex = *ex;
1655 return;
1659 * If new found extent is better, store it in the context
1661 if (bex->fe_len < gex->fe_len) {
1662 /* if the request isn't satisfied, any found extent
1663 * larger than previous best one is better */
1664 if (ex->fe_len > bex->fe_len)
1665 *bex = *ex;
1666 } else if (ex->fe_len > gex->fe_len) {
1667 /* if the request is satisfied, then we try to find
1668 * an extent that still satisfy the request, but is
1669 * smaller than previous one */
1670 if (ex->fe_len < bex->fe_len)
1671 *bex = *ex;
1674 ext4_mb_check_limits(ac, e4b, 0);
1677 static noinline_for_stack
1678 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1679 struct ext4_buddy *e4b)
1681 struct ext4_free_extent ex = ac->ac_b_ex;
1682 ext4_group_t group = ex.fe_group;
1683 int max;
1684 int err;
1686 BUG_ON(ex.fe_len <= 0);
1687 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1688 if (err)
1689 return err;
1691 ext4_lock_group(ac->ac_sb, group);
1692 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1694 if (max > 0) {
1695 ac->ac_b_ex = ex;
1696 ext4_mb_use_best_found(ac, e4b);
1699 ext4_unlock_group(ac->ac_sb, group);
1700 ext4_mb_unload_buddy(e4b);
1702 return 0;
1705 static noinline_for_stack
1706 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1707 struct ext4_buddy *e4b)
1709 ext4_group_t group = ac->ac_g_ex.fe_group;
1710 int max;
1711 int err;
1712 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1713 struct ext4_free_extent ex;
1715 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1716 return 0;
1718 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1719 if (err)
1720 return err;
1722 ext4_lock_group(ac->ac_sb, group);
1723 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1724 ac->ac_g_ex.fe_len, &ex);
1726 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1727 ext4_fsblk_t start;
1729 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1730 ex.fe_start;
1731 /* use do_div to get remainder (would be 64-bit modulo) */
1732 if (do_div(start, sbi->s_stripe) == 0) {
1733 ac->ac_found++;
1734 ac->ac_b_ex = ex;
1735 ext4_mb_use_best_found(ac, e4b);
1737 } else if (max >= ac->ac_g_ex.fe_len) {
1738 BUG_ON(ex.fe_len <= 0);
1739 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1740 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1741 ac->ac_found++;
1742 ac->ac_b_ex = ex;
1743 ext4_mb_use_best_found(ac, e4b);
1744 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1745 /* Sometimes, caller may want to merge even small
1746 * number of blocks to an existing extent */
1747 BUG_ON(ex.fe_len <= 0);
1748 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1749 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1750 ac->ac_found++;
1751 ac->ac_b_ex = ex;
1752 ext4_mb_use_best_found(ac, e4b);
1754 ext4_unlock_group(ac->ac_sb, group);
1755 ext4_mb_unload_buddy(e4b);
1757 return 0;
1761 * The routine scans buddy structures (not bitmap!) from given order
1762 * to max order and tries to find big enough chunk to satisfy the req
1764 static noinline_for_stack
1765 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1766 struct ext4_buddy *e4b)
1768 struct super_block *sb = ac->ac_sb;
1769 struct ext4_group_info *grp = e4b->bd_info;
1770 void *buddy;
1771 int i;
1772 int k;
1773 int max;
1775 BUG_ON(ac->ac_2order <= 0);
1776 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1777 if (grp->bb_counters[i] == 0)
1778 continue;
1780 buddy = mb_find_buddy(e4b, i, &max);
1781 BUG_ON(buddy == NULL);
1783 k = mb_find_next_zero_bit(buddy, max, 0);
1784 BUG_ON(k >= max);
1786 ac->ac_found++;
1788 ac->ac_b_ex.fe_len = 1 << i;
1789 ac->ac_b_ex.fe_start = k << i;
1790 ac->ac_b_ex.fe_group = e4b->bd_group;
1792 ext4_mb_use_best_found(ac, e4b);
1794 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1796 if (EXT4_SB(sb)->s_mb_stats)
1797 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1799 break;
1804 * The routine scans the group and measures all found extents.
1805 * In order to optimize scanning, caller must pass number of
1806 * free blocks in the group, so the routine can know upper limit.
1808 static noinline_for_stack
1809 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1810 struct ext4_buddy *e4b)
1812 struct super_block *sb = ac->ac_sb;
1813 void *bitmap = EXT4_MB_BITMAP(e4b);
1814 struct ext4_free_extent ex;
1815 int i;
1816 int free;
1818 free = e4b->bd_info->bb_free;
1819 BUG_ON(free <= 0);
1821 i = e4b->bd_info->bb_first_free;
1823 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1824 i = mb_find_next_zero_bit(bitmap,
1825 EXT4_CLUSTERS_PER_GROUP(sb), i);
1826 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1828 * IF we have corrupt bitmap, we won't find any
1829 * free blocks even though group info says we
1830 * we have free blocks
1832 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1833 "%d free clusters as per "
1834 "group info. But bitmap says 0",
1835 free);
1836 break;
1839 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1840 BUG_ON(ex.fe_len <= 0);
1841 if (free < ex.fe_len) {
1842 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1843 "%d free clusters as per "
1844 "group info. But got %d blocks",
1845 free, ex.fe_len);
1847 * The number of free blocks differs. This mostly
1848 * indicate that the bitmap is corrupt. So exit
1849 * without claiming the space.
1851 break;
1854 ext4_mb_measure_extent(ac, &ex, e4b);
1856 i += ex.fe_len;
1857 free -= ex.fe_len;
1860 ext4_mb_check_limits(ac, e4b, 1);
1864 * This is a special case for storages like raid5
1865 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1867 static noinline_for_stack
1868 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1869 struct ext4_buddy *e4b)
1871 struct super_block *sb = ac->ac_sb;
1872 struct ext4_sb_info *sbi = EXT4_SB(sb);
1873 void *bitmap = EXT4_MB_BITMAP(e4b);
1874 struct ext4_free_extent ex;
1875 ext4_fsblk_t first_group_block;
1876 ext4_fsblk_t a;
1877 ext4_grpblk_t i;
1878 int max;
1880 BUG_ON(sbi->s_stripe == 0);
1882 /* find first stripe-aligned block in group */
1883 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1885 a = first_group_block + sbi->s_stripe - 1;
1886 do_div(a, sbi->s_stripe);
1887 i = (a * sbi->s_stripe) - first_group_block;
1889 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1890 if (!mb_test_bit(i, bitmap)) {
1891 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1892 if (max >= sbi->s_stripe) {
1893 ac->ac_found++;
1894 ac->ac_b_ex = ex;
1895 ext4_mb_use_best_found(ac, e4b);
1896 break;
1899 i += sbi->s_stripe;
1903 /* This is now called BEFORE we load the buddy bitmap. */
1904 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1905 ext4_group_t group, int cr)
1907 unsigned free, fragments;
1908 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1909 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1911 BUG_ON(cr < 0 || cr >= 4);
1913 /* We only do this if the grp has never been initialized */
1914 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1915 int ret = ext4_mb_init_group(ac->ac_sb, group);
1916 if (ret)
1917 return 0;
1920 free = grp->bb_free;
1921 fragments = grp->bb_fragments;
1922 if (free == 0)
1923 return 0;
1924 if (fragments == 0)
1925 return 0;
1927 switch (cr) {
1928 case 0:
1929 BUG_ON(ac->ac_2order == 0);
1931 if (grp->bb_largest_free_order < ac->ac_2order)
1932 return 0;
1934 /* Avoid using the first bg of a flexgroup for data files */
1935 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1936 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1937 ((group % flex_size) == 0))
1938 return 0;
1940 return 1;
1941 case 1:
1942 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1943 return 1;
1944 break;
1945 case 2:
1946 if (free >= ac->ac_g_ex.fe_len)
1947 return 1;
1948 break;
1949 case 3:
1950 return 1;
1951 default:
1952 BUG();
1955 return 0;
1958 static noinline_for_stack int
1959 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1961 ext4_group_t ngroups, group, i;
1962 int cr;
1963 int err = 0;
1964 struct ext4_sb_info *sbi;
1965 struct super_block *sb;
1966 struct ext4_buddy e4b;
1968 sb = ac->ac_sb;
1969 sbi = EXT4_SB(sb);
1970 ngroups = ext4_get_groups_count(sb);
1971 /* non-extent files are limited to low blocks/groups */
1972 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1973 ngroups = sbi->s_blockfile_groups;
1975 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1977 /* first, try the goal */
1978 err = ext4_mb_find_by_goal(ac, &e4b);
1979 if (err || ac->ac_status == AC_STATUS_FOUND)
1980 goto out;
1982 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1983 goto out;
1986 * ac->ac2_order is set only if the fe_len is a power of 2
1987 * if ac2_order is set we also set criteria to 0 so that we
1988 * try exact allocation using buddy.
1990 i = fls(ac->ac_g_ex.fe_len);
1991 ac->ac_2order = 0;
1993 * We search using buddy data only if the order of the request
1994 * is greater than equal to the sbi_s_mb_order2_reqs
1995 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1997 if (i >= sbi->s_mb_order2_reqs) {
1999 * This should tell if fe_len is exactly power of 2
2001 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2002 ac->ac_2order = i - 1;
2005 /* if stream allocation is enabled, use global goal */
2006 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2007 /* TBD: may be hot point */
2008 spin_lock(&sbi->s_md_lock);
2009 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2010 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2011 spin_unlock(&sbi->s_md_lock);
2014 /* Let's just scan groups to find more-less suitable blocks */
2015 cr = ac->ac_2order ? 0 : 1;
2017 * cr == 0 try to get exact allocation,
2018 * cr == 3 try to get anything
2020 repeat:
2021 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2022 ac->ac_criteria = cr;
2024 * searching for the right group start
2025 * from the goal value specified
2027 group = ac->ac_g_ex.fe_group;
2029 for (i = 0; i < ngroups; group++, i++) {
2030 if (group == ngroups)
2031 group = 0;
2033 /* This now checks without needing the buddy page */
2034 if (!ext4_mb_good_group(ac, group, cr))
2035 continue;
2037 err = ext4_mb_load_buddy(sb, group, &e4b);
2038 if (err)
2039 goto out;
2041 ext4_lock_group(sb, group);
2044 * We need to check again after locking the
2045 * block group
2047 if (!ext4_mb_good_group(ac, group, cr)) {
2048 ext4_unlock_group(sb, group);
2049 ext4_mb_unload_buddy(&e4b);
2050 continue;
2053 ac->ac_groups_scanned++;
2054 if (cr == 0)
2055 ext4_mb_simple_scan_group(ac, &e4b);
2056 else if (cr == 1 && sbi->s_stripe &&
2057 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2058 ext4_mb_scan_aligned(ac, &e4b);
2059 else
2060 ext4_mb_complex_scan_group(ac, &e4b);
2062 ext4_unlock_group(sb, group);
2063 ext4_mb_unload_buddy(&e4b);
2065 if (ac->ac_status != AC_STATUS_CONTINUE)
2066 break;
2070 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2071 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2073 * We've been searching too long. Let's try to allocate
2074 * the best chunk we've found so far
2077 ext4_mb_try_best_found(ac, &e4b);
2078 if (ac->ac_status != AC_STATUS_FOUND) {
2080 * Someone more lucky has already allocated it.
2081 * The only thing we can do is just take first
2082 * found block(s)
2083 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2085 ac->ac_b_ex.fe_group = 0;
2086 ac->ac_b_ex.fe_start = 0;
2087 ac->ac_b_ex.fe_len = 0;
2088 ac->ac_status = AC_STATUS_CONTINUE;
2089 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2090 cr = 3;
2091 atomic_inc(&sbi->s_mb_lost_chunks);
2092 goto repeat;
2095 out:
2096 return err;
2099 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2101 struct super_block *sb = seq->private;
2102 ext4_group_t group;
2104 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2105 return NULL;
2106 group = *pos + 1;
2107 return (void *) ((unsigned long) group);
2110 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2112 struct super_block *sb = seq->private;
2113 ext4_group_t group;
2115 ++*pos;
2116 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2117 return NULL;
2118 group = *pos + 1;
2119 return (void *) ((unsigned long) group);
2122 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2124 struct super_block *sb = seq->private;
2125 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2126 int i;
2127 int err;
2128 struct ext4_buddy e4b;
2129 struct sg {
2130 struct ext4_group_info info;
2131 ext4_grpblk_t counters[16];
2132 } sg;
2134 group--;
2135 if (group == 0)
2136 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2137 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2138 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2139 "group", "free", "frags", "first",
2140 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2141 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2143 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2144 sizeof(struct ext4_group_info);
2145 err = ext4_mb_load_buddy(sb, group, &e4b);
2146 if (err) {
2147 seq_printf(seq, "#%-5u: I/O error\n", group);
2148 return 0;
2150 ext4_lock_group(sb, group);
2151 memcpy(&sg, ext4_get_group_info(sb, group), i);
2152 ext4_unlock_group(sb, group);
2153 ext4_mb_unload_buddy(&e4b);
2155 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2156 sg.info.bb_fragments, sg.info.bb_first_free);
2157 for (i = 0; i <= 13; i++)
2158 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2159 sg.info.bb_counters[i] : 0);
2160 seq_printf(seq, " ]\n");
2162 return 0;
2165 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2169 static const struct seq_operations ext4_mb_seq_groups_ops = {
2170 .start = ext4_mb_seq_groups_start,
2171 .next = ext4_mb_seq_groups_next,
2172 .stop = ext4_mb_seq_groups_stop,
2173 .show = ext4_mb_seq_groups_show,
2176 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2178 struct super_block *sb = PDE(inode)->data;
2179 int rc;
2181 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2182 if (rc == 0) {
2183 struct seq_file *m = file->private_data;
2184 m->private = sb;
2186 return rc;
2190 static const struct file_operations ext4_mb_seq_groups_fops = {
2191 .owner = THIS_MODULE,
2192 .open = ext4_mb_seq_groups_open,
2193 .read = seq_read,
2194 .llseek = seq_lseek,
2195 .release = seq_release,
2198 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2200 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2201 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2203 BUG_ON(!cachep);
2204 return cachep;
2207 /* Create and initialize ext4_group_info data for the given group. */
2208 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2209 struct ext4_group_desc *desc)
2211 int i;
2212 int metalen = 0;
2213 struct ext4_sb_info *sbi = EXT4_SB(sb);
2214 struct ext4_group_info **meta_group_info;
2215 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2218 * First check if this group is the first of a reserved block.
2219 * If it's true, we have to allocate a new table of pointers
2220 * to ext4_group_info structures
2222 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2223 metalen = sizeof(*meta_group_info) <<
2224 EXT4_DESC_PER_BLOCK_BITS(sb);
2225 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2226 if (meta_group_info == NULL) {
2227 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem "
2228 "for a buddy group");
2229 goto exit_meta_group_info;
2231 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2232 meta_group_info;
2235 meta_group_info =
2236 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2237 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2239 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2240 if (meta_group_info[i] == NULL) {
2241 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem");
2242 goto exit_group_info;
2244 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2245 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2246 &(meta_group_info[i]->bb_state));
2249 * initialize bb_free to be able to skip
2250 * empty groups without initialization
2252 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2253 meta_group_info[i]->bb_free =
2254 ext4_free_clusters_after_init(sb, group, desc);
2255 } else {
2256 meta_group_info[i]->bb_free =
2257 ext4_free_group_clusters(sb, desc);
2260 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2261 init_rwsem(&meta_group_info[i]->alloc_sem);
2262 meta_group_info[i]->bb_free_root = RB_ROOT;
2263 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2265 #ifdef DOUBLE_CHECK
2267 struct buffer_head *bh;
2268 meta_group_info[i]->bb_bitmap =
2269 kmalloc(sb->s_blocksize, GFP_KERNEL);
2270 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2271 bh = ext4_read_block_bitmap(sb, group);
2272 BUG_ON(bh == NULL);
2273 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2274 sb->s_blocksize);
2275 put_bh(bh);
2277 #endif
2279 return 0;
2281 exit_group_info:
2282 /* If a meta_group_info table has been allocated, release it now */
2283 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2284 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2285 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2287 exit_meta_group_info:
2288 return -ENOMEM;
2289 } /* ext4_mb_add_groupinfo */
2291 static int ext4_mb_init_backend(struct super_block *sb)
2293 ext4_group_t ngroups = ext4_get_groups_count(sb);
2294 ext4_group_t i;
2295 struct ext4_sb_info *sbi = EXT4_SB(sb);
2296 struct ext4_super_block *es = sbi->s_es;
2297 int num_meta_group_infos;
2298 int num_meta_group_infos_max;
2299 int array_size;
2300 struct ext4_group_desc *desc;
2301 struct kmem_cache *cachep;
2303 /* This is the number of blocks used by GDT */
2304 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2305 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2308 * This is the total number of blocks used by GDT including
2309 * the number of reserved blocks for GDT.
2310 * The s_group_info array is allocated with this value
2311 * to allow a clean online resize without a complex
2312 * manipulation of pointer.
2313 * The drawback is the unused memory when no resize
2314 * occurs but it's very low in terms of pages
2315 * (see comments below)
2316 * Need to handle this properly when META_BG resizing is allowed
2318 num_meta_group_infos_max = num_meta_group_infos +
2319 le16_to_cpu(es->s_reserved_gdt_blocks);
2322 * array_size is the size of s_group_info array. We round it
2323 * to the next power of two because this approximation is done
2324 * internally by kmalloc so we can have some more memory
2325 * for free here (e.g. may be used for META_BG resize).
2327 array_size = 1;
2328 while (array_size < sizeof(*sbi->s_group_info) *
2329 num_meta_group_infos_max)
2330 array_size = array_size << 1;
2331 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2332 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2333 * So a two level scheme suffices for now. */
2334 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2335 if (sbi->s_group_info == NULL) {
2336 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2337 return -ENOMEM;
2339 sbi->s_buddy_cache = new_inode(sb);
2340 if (sbi->s_buddy_cache == NULL) {
2341 ext4_msg(sb, KERN_ERR, "can't get new inode");
2342 goto err_freesgi;
2344 /* To avoid potentially colliding with an valid on-disk inode number,
2345 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2346 * not in the inode hash, so it should never be found by iget(), but
2347 * this will avoid confusion if it ever shows up during debugging. */
2348 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2349 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2350 for (i = 0; i < ngroups; i++) {
2351 desc = ext4_get_group_desc(sb, i, NULL);
2352 if (desc == NULL) {
2353 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2354 goto err_freebuddy;
2356 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2357 goto err_freebuddy;
2360 return 0;
2362 err_freebuddy:
2363 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2364 while (i-- > 0)
2365 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2366 i = num_meta_group_infos;
2367 while (i-- > 0)
2368 kfree(sbi->s_group_info[i]);
2369 iput(sbi->s_buddy_cache);
2370 err_freesgi:
2371 ext4_kvfree(sbi->s_group_info);
2372 return -ENOMEM;
2375 static void ext4_groupinfo_destroy_slabs(void)
2377 int i;
2379 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2380 if (ext4_groupinfo_caches[i])
2381 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2382 ext4_groupinfo_caches[i] = NULL;
2386 static int ext4_groupinfo_create_slab(size_t size)
2388 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2389 int slab_size;
2390 int blocksize_bits = order_base_2(size);
2391 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2392 struct kmem_cache *cachep;
2394 if (cache_index >= NR_GRPINFO_CACHES)
2395 return -EINVAL;
2397 if (unlikely(cache_index < 0))
2398 cache_index = 0;
2400 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2401 if (ext4_groupinfo_caches[cache_index]) {
2402 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2403 return 0; /* Already created */
2406 slab_size = offsetof(struct ext4_group_info,
2407 bb_counters[blocksize_bits + 2]);
2409 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2410 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2411 NULL);
2413 ext4_groupinfo_caches[cache_index] = cachep;
2415 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2416 if (!cachep) {
2417 printk(KERN_EMERG
2418 "EXT4-fs: no memory for groupinfo slab cache\n");
2419 return -ENOMEM;
2422 return 0;
2425 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2427 struct ext4_sb_info *sbi = EXT4_SB(sb);
2428 unsigned i, j;
2429 unsigned offset;
2430 unsigned max;
2431 int ret;
2433 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2435 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2436 if (sbi->s_mb_offsets == NULL) {
2437 ret = -ENOMEM;
2438 goto out;
2441 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2442 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2443 if (sbi->s_mb_maxs == NULL) {
2444 ret = -ENOMEM;
2445 goto out;
2448 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2449 if (ret < 0)
2450 goto out;
2452 /* order 0 is regular bitmap */
2453 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2454 sbi->s_mb_offsets[0] = 0;
2456 i = 1;
2457 offset = 0;
2458 max = sb->s_blocksize << 2;
2459 do {
2460 sbi->s_mb_offsets[i] = offset;
2461 sbi->s_mb_maxs[i] = max;
2462 offset += 1 << (sb->s_blocksize_bits - i);
2463 max = max >> 1;
2464 i++;
2465 } while (i <= sb->s_blocksize_bits + 1);
2467 spin_lock_init(&sbi->s_md_lock);
2468 spin_lock_init(&sbi->s_bal_lock);
2470 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2471 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2472 sbi->s_mb_stats = MB_DEFAULT_STATS;
2473 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2474 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2476 * The default group preallocation is 512, which for 4k block
2477 * sizes translates to 2 megabytes. However for bigalloc file
2478 * systems, this is probably too big (i.e, if the cluster size
2479 * is 1 megabyte, then group preallocation size becomes half a
2480 * gigabyte!). As a default, we will keep a two megabyte
2481 * group pralloc size for cluster sizes up to 64k, and after
2482 * that, we will force a minimum group preallocation size of
2483 * 32 clusters. This translates to 8 megs when the cluster
2484 * size is 256k, and 32 megs when the cluster size is 1 meg,
2485 * which seems reasonable as a default.
2487 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2488 sbi->s_cluster_bits, 32);
2490 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2491 * to the lowest multiple of s_stripe which is bigger than
2492 * the s_mb_group_prealloc as determined above. We want
2493 * the preallocation size to be an exact multiple of the
2494 * RAID stripe size so that preallocations don't fragment
2495 * the stripes.
2497 if (sbi->s_stripe > 1) {
2498 sbi->s_mb_group_prealloc = roundup(
2499 sbi->s_mb_group_prealloc, sbi->s_stripe);
2502 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2503 if (sbi->s_locality_groups == NULL) {
2504 ret = -ENOMEM;
2505 goto out_free_groupinfo_slab;
2507 for_each_possible_cpu(i) {
2508 struct ext4_locality_group *lg;
2509 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2510 mutex_init(&lg->lg_mutex);
2511 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2512 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2513 spin_lock_init(&lg->lg_prealloc_lock);
2516 /* init file for buddy data */
2517 ret = ext4_mb_init_backend(sb);
2518 if (ret != 0)
2519 goto out_free_locality_groups;
2521 if (sbi->s_proc)
2522 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2523 &ext4_mb_seq_groups_fops, sb);
2525 if (sbi->s_journal)
2526 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2528 return 0;
2530 out_free_locality_groups:
2531 free_percpu(sbi->s_locality_groups);
2532 sbi->s_locality_groups = NULL;
2533 out_free_groupinfo_slab:
2534 ext4_groupinfo_destroy_slabs();
2535 out:
2536 kfree(sbi->s_mb_offsets);
2537 sbi->s_mb_offsets = NULL;
2538 kfree(sbi->s_mb_maxs);
2539 sbi->s_mb_maxs = NULL;
2540 return ret;
2543 /* need to called with the ext4 group lock held */
2544 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2546 struct ext4_prealloc_space *pa;
2547 struct list_head *cur, *tmp;
2548 int count = 0;
2550 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2551 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2552 list_del(&pa->pa_group_list);
2553 count++;
2554 kmem_cache_free(ext4_pspace_cachep, pa);
2556 if (count)
2557 mb_debug(1, "mballoc: %u PAs left\n", count);
2561 int ext4_mb_release(struct super_block *sb)
2563 ext4_group_t ngroups = ext4_get_groups_count(sb);
2564 ext4_group_t i;
2565 int num_meta_group_infos;
2566 struct ext4_group_info *grinfo;
2567 struct ext4_sb_info *sbi = EXT4_SB(sb);
2568 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2570 if (sbi->s_group_info) {
2571 for (i = 0; i < ngroups; i++) {
2572 grinfo = ext4_get_group_info(sb, i);
2573 #ifdef DOUBLE_CHECK
2574 kfree(grinfo->bb_bitmap);
2575 #endif
2576 ext4_lock_group(sb, i);
2577 ext4_mb_cleanup_pa(grinfo);
2578 ext4_unlock_group(sb, i);
2579 kmem_cache_free(cachep, grinfo);
2581 num_meta_group_infos = (ngroups +
2582 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2583 EXT4_DESC_PER_BLOCK_BITS(sb);
2584 for (i = 0; i < num_meta_group_infos; i++)
2585 kfree(sbi->s_group_info[i]);
2586 ext4_kvfree(sbi->s_group_info);
2588 kfree(sbi->s_mb_offsets);
2589 kfree(sbi->s_mb_maxs);
2590 if (sbi->s_buddy_cache)
2591 iput(sbi->s_buddy_cache);
2592 if (sbi->s_mb_stats) {
2593 ext4_msg(sb, KERN_INFO,
2594 "mballoc: %u blocks %u reqs (%u success)",
2595 atomic_read(&sbi->s_bal_allocated),
2596 atomic_read(&sbi->s_bal_reqs),
2597 atomic_read(&sbi->s_bal_success));
2598 ext4_msg(sb, KERN_INFO,
2599 "mballoc: %u extents scanned, %u goal hits, "
2600 "%u 2^N hits, %u breaks, %u lost",
2601 atomic_read(&sbi->s_bal_ex_scanned),
2602 atomic_read(&sbi->s_bal_goals),
2603 atomic_read(&sbi->s_bal_2orders),
2604 atomic_read(&sbi->s_bal_breaks),
2605 atomic_read(&sbi->s_mb_lost_chunks));
2606 ext4_msg(sb, KERN_INFO,
2607 "mballoc: %lu generated and it took %Lu",
2608 sbi->s_mb_buddies_generated,
2609 sbi->s_mb_generation_time);
2610 ext4_msg(sb, KERN_INFO,
2611 "mballoc: %u preallocated, %u discarded",
2612 atomic_read(&sbi->s_mb_preallocated),
2613 atomic_read(&sbi->s_mb_discarded));
2616 free_percpu(sbi->s_locality_groups);
2617 if (sbi->s_proc)
2618 remove_proc_entry("mb_groups", sbi->s_proc);
2620 return 0;
2623 static inline int ext4_issue_discard(struct super_block *sb,
2624 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2626 ext4_fsblk_t discard_block;
2628 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2629 ext4_group_first_block_no(sb, block_group));
2630 count = EXT4_C2B(EXT4_SB(sb), count);
2631 trace_ext4_discard_blocks(sb,
2632 (unsigned long long) discard_block, count);
2633 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2637 * This function is called by the jbd2 layer once the commit has finished,
2638 * so we know we can free the blocks that were released with that commit.
2640 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2642 struct super_block *sb = journal->j_private;
2643 struct ext4_buddy e4b;
2644 struct ext4_group_info *db;
2645 int err, count = 0, count2 = 0;
2646 struct ext4_free_data *entry;
2647 struct list_head *l, *ltmp;
2649 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2650 entry = list_entry(l, struct ext4_free_data, list);
2652 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2653 entry->count, entry->group, entry);
2655 if (test_opt(sb, DISCARD))
2656 ext4_issue_discard(sb, entry->group,
2657 entry->start_cluster, entry->count);
2659 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2660 /* we expect to find existing buddy because it's pinned */
2661 BUG_ON(err != 0);
2663 db = e4b.bd_info;
2664 /* there are blocks to put in buddy to make them really free */
2665 count += entry->count;
2666 count2++;
2667 ext4_lock_group(sb, entry->group);
2668 /* Take it out of per group rb tree */
2669 rb_erase(&entry->node, &(db->bb_free_root));
2670 mb_free_blocks(NULL, &e4b, entry->start_cluster, entry->count);
2673 * Clear the trimmed flag for the group so that the next
2674 * ext4_trim_fs can trim it.
2675 * If the volume is mounted with -o discard, online discard
2676 * is supported and the free blocks will be trimmed online.
2678 if (!test_opt(sb, DISCARD))
2679 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2681 if (!db->bb_free_root.rb_node) {
2682 /* No more items in the per group rb tree
2683 * balance refcounts from ext4_mb_free_metadata()
2685 page_cache_release(e4b.bd_buddy_page);
2686 page_cache_release(e4b.bd_bitmap_page);
2688 ext4_unlock_group(sb, entry->group);
2689 kmem_cache_free(ext4_free_ext_cachep, entry);
2690 ext4_mb_unload_buddy(&e4b);
2693 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2696 #ifdef CONFIG_EXT4_DEBUG
2697 u8 mb_enable_debug __read_mostly;
2699 static struct dentry *debugfs_dir;
2700 static struct dentry *debugfs_debug;
2702 static void __init ext4_create_debugfs_entry(void)
2704 debugfs_dir = debugfs_create_dir("ext4", NULL);
2705 if (debugfs_dir)
2706 debugfs_debug = debugfs_create_u8("mballoc-debug",
2707 S_IRUGO | S_IWUSR,
2708 debugfs_dir,
2709 &mb_enable_debug);
2712 static void ext4_remove_debugfs_entry(void)
2714 debugfs_remove(debugfs_debug);
2715 debugfs_remove(debugfs_dir);
2718 #else
2720 static void __init ext4_create_debugfs_entry(void)
2724 static void ext4_remove_debugfs_entry(void)
2728 #endif
2730 int __init ext4_init_mballoc(void)
2732 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2733 SLAB_RECLAIM_ACCOUNT);
2734 if (ext4_pspace_cachep == NULL)
2735 return -ENOMEM;
2737 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2738 SLAB_RECLAIM_ACCOUNT);
2739 if (ext4_ac_cachep == NULL) {
2740 kmem_cache_destroy(ext4_pspace_cachep);
2741 return -ENOMEM;
2744 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2745 SLAB_RECLAIM_ACCOUNT);
2746 if (ext4_free_ext_cachep == NULL) {
2747 kmem_cache_destroy(ext4_pspace_cachep);
2748 kmem_cache_destroy(ext4_ac_cachep);
2749 return -ENOMEM;
2751 ext4_create_debugfs_entry();
2752 return 0;
2755 void ext4_exit_mballoc(void)
2758 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2759 * before destroying the slab cache.
2761 rcu_barrier();
2762 kmem_cache_destroy(ext4_pspace_cachep);
2763 kmem_cache_destroy(ext4_ac_cachep);
2764 kmem_cache_destroy(ext4_free_ext_cachep);
2765 ext4_groupinfo_destroy_slabs();
2766 ext4_remove_debugfs_entry();
2771 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2772 * Returns 0 if success or error code
2774 static noinline_for_stack int
2775 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2776 handle_t *handle, unsigned int reserv_clstrs)
2778 struct buffer_head *bitmap_bh = NULL;
2779 struct ext4_group_desc *gdp;
2780 struct buffer_head *gdp_bh;
2781 struct ext4_sb_info *sbi;
2782 struct super_block *sb;
2783 ext4_fsblk_t block;
2784 int err, len;
2786 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2787 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2789 sb = ac->ac_sb;
2790 sbi = EXT4_SB(sb);
2792 err = -EIO;
2793 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2794 if (!bitmap_bh)
2795 goto out_err;
2797 err = ext4_journal_get_write_access(handle, bitmap_bh);
2798 if (err)
2799 goto out_err;
2801 err = -EIO;
2802 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2803 if (!gdp)
2804 goto out_err;
2806 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2807 ext4_free_group_clusters(sb, gdp));
2809 err = ext4_journal_get_write_access(handle, gdp_bh);
2810 if (err)
2811 goto out_err;
2813 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2815 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2816 if (!ext4_data_block_valid(sbi, block, len)) {
2817 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2818 "fs metadata\n", block, block+len);
2819 /* File system mounted not to panic on error
2820 * Fix the bitmap and repeat the block allocation
2821 * We leak some of the blocks here.
2823 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2824 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2825 ac->ac_b_ex.fe_len);
2826 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2827 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2828 if (!err)
2829 err = -EAGAIN;
2830 goto out_err;
2833 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2834 #ifdef AGGRESSIVE_CHECK
2836 int i;
2837 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2838 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2839 bitmap_bh->b_data));
2842 #endif
2843 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2844 ac->ac_b_ex.fe_len);
2845 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2846 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2847 ext4_free_group_clusters_set(sb, gdp,
2848 ext4_free_clusters_after_init(sb,
2849 ac->ac_b_ex.fe_group, gdp));
2851 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2852 ext4_free_group_clusters_set(sb, gdp, len);
2853 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2855 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2856 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2858 * Now reduce the dirty block count also. Should not go negative
2860 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2861 /* release all the reserved blocks if non delalloc */
2862 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2863 reserv_clstrs);
2865 if (sbi->s_log_groups_per_flex) {
2866 ext4_group_t flex_group = ext4_flex_group(sbi,
2867 ac->ac_b_ex.fe_group);
2868 atomic_sub(ac->ac_b_ex.fe_len,
2869 &sbi->s_flex_groups[flex_group].free_clusters);
2872 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2873 if (err)
2874 goto out_err;
2875 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2877 out_err:
2878 ext4_mark_super_dirty(sb);
2879 brelse(bitmap_bh);
2880 return err;
2884 * here we normalize request for locality group
2885 * Group request are normalized to s_mb_group_prealloc, which goes to
2886 * s_strip if we set the same via mount option.
2887 * s_mb_group_prealloc can be configured via
2888 * /sys/fs/ext4/<partition>/mb_group_prealloc
2890 * XXX: should we try to preallocate more than the group has now?
2892 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2894 struct super_block *sb = ac->ac_sb;
2895 struct ext4_locality_group *lg = ac->ac_lg;
2897 BUG_ON(lg == NULL);
2898 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2899 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2900 current->pid, ac->ac_g_ex.fe_len);
2904 * Normalization means making request better in terms of
2905 * size and alignment
2907 static noinline_for_stack void
2908 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2909 struct ext4_allocation_request *ar)
2911 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2912 int bsbits, max;
2913 ext4_lblk_t end;
2914 loff_t size, orig_size, start_off;
2915 ext4_lblk_t start;
2916 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2917 struct ext4_prealloc_space *pa;
2919 /* do normalize only data requests, metadata requests
2920 do not need preallocation */
2921 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2922 return;
2924 /* sometime caller may want exact blocks */
2925 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2926 return;
2928 /* caller may indicate that preallocation isn't
2929 * required (it's a tail, for example) */
2930 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2931 return;
2933 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2934 ext4_mb_normalize_group_request(ac);
2935 return ;
2938 bsbits = ac->ac_sb->s_blocksize_bits;
2940 /* first, let's learn actual file size
2941 * given current request is allocated */
2942 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2943 size = size << bsbits;
2944 if (size < i_size_read(ac->ac_inode))
2945 size = i_size_read(ac->ac_inode);
2946 orig_size = size;
2948 /* max size of free chunks */
2949 max = 2 << bsbits;
2951 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2952 (req <= (size) || max <= (chunk_size))
2954 /* first, try to predict filesize */
2955 /* XXX: should this table be tunable? */
2956 start_off = 0;
2957 if (size <= 16 * 1024) {
2958 size = 16 * 1024;
2959 } else if (size <= 32 * 1024) {
2960 size = 32 * 1024;
2961 } else if (size <= 64 * 1024) {
2962 size = 64 * 1024;
2963 } else if (size <= 128 * 1024) {
2964 size = 128 * 1024;
2965 } else if (size <= 256 * 1024) {
2966 size = 256 * 1024;
2967 } else if (size <= 512 * 1024) {
2968 size = 512 * 1024;
2969 } else if (size <= 1024 * 1024) {
2970 size = 1024 * 1024;
2971 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2972 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2973 (21 - bsbits)) << 21;
2974 size = 2 * 1024 * 1024;
2975 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2976 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2977 (22 - bsbits)) << 22;
2978 size = 4 * 1024 * 1024;
2979 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2980 (8<<20)>>bsbits, max, 8 * 1024)) {
2981 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2982 (23 - bsbits)) << 23;
2983 size = 8 * 1024 * 1024;
2984 } else {
2985 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2986 size = ac->ac_o_ex.fe_len << bsbits;
2988 size = size >> bsbits;
2989 start = start_off >> bsbits;
2991 /* don't cover already allocated blocks in selected range */
2992 if (ar->pleft && start <= ar->lleft) {
2993 size -= ar->lleft + 1 - start;
2994 start = ar->lleft + 1;
2996 if (ar->pright && start + size - 1 >= ar->lright)
2997 size -= start + size - ar->lright;
2999 end = start + size;
3001 /* check we don't cross already preallocated blocks */
3002 rcu_read_lock();
3003 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3004 ext4_lblk_t pa_end;
3006 if (pa->pa_deleted)
3007 continue;
3008 spin_lock(&pa->pa_lock);
3009 if (pa->pa_deleted) {
3010 spin_unlock(&pa->pa_lock);
3011 continue;
3014 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3015 pa->pa_len);
3017 /* PA must not overlap original request */
3018 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3019 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3021 /* skip PAs this normalized request doesn't overlap with */
3022 if (pa->pa_lstart >= end || pa_end <= start) {
3023 spin_unlock(&pa->pa_lock);
3024 continue;
3026 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3028 /* adjust start or end to be adjacent to this pa */
3029 if (pa_end <= ac->ac_o_ex.fe_logical) {
3030 BUG_ON(pa_end < start);
3031 start = pa_end;
3032 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3033 BUG_ON(pa->pa_lstart > end);
3034 end = pa->pa_lstart;
3036 spin_unlock(&pa->pa_lock);
3038 rcu_read_unlock();
3039 size = end - start;
3041 /* XXX: extra loop to check we really don't overlap preallocations */
3042 rcu_read_lock();
3043 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3044 ext4_lblk_t pa_end;
3046 spin_lock(&pa->pa_lock);
3047 if (pa->pa_deleted == 0) {
3048 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3049 pa->pa_len);
3050 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3052 spin_unlock(&pa->pa_lock);
3054 rcu_read_unlock();
3056 if (start + size <= ac->ac_o_ex.fe_logical &&
3057 start > ac->ac_o_ex.fe_logical) {
3058 ext4_msg(ac->ac_sb, KERN_ERR,
3059 "start %lu, size %lu, fe_logical %lu",
3060 (unsigned long) start, (unsigned long) size,
3061 (unsigned long) ac->ac_o_ex.fe_logical);
3063 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3064 start > ac->ac_o_ex.fe_logical);
3065 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3067 /* now prepare goal request */
3069 /* XXX: is it better to align blocks WRT to logical
3070 * placement or satisfy big request as is */
3071 ac->ac_g_ex.fe_logical = start;
3072 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3074 /* define goal start in order to merge */
3075 if (ar->pright && (ar->lright == (start + size))) {
3076 /* merge to the right */
3077 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3078 &ac->ac_f_ex.fe_group,
3079 &ac->ac_f_ex.fe_start);
3080 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3082 if (ar->pleft && (ar->lleft + 1 == start)) {
3083 /* merge to the left */
3084 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3085 &ac->ac_f_ex.fe_group,
3086 &ac->ac_f_ex.fe_start);
3087 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3090 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3091 (unsigned) orig_size, (unsigned) start);
3094 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3096 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3098 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3099 atomic_inc(&sbi->s_bal_reqs);
3100 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3101 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3102 atomic_inc(&sbi->s_bal_success);
3103 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3104 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3105 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3106 atomic_inc(&sbi->s_bal_goals);
3107 if (ac->ac_found > sbi->s_mb_max_to_scan)
3108 atomic_inc(&sbi->s_bal_breaks);
3111 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3112 trace_ext4_mballoc_alloc(ac);
3113 else
3114 trace_ext4_mballoc_prealloc(ac);
3118 * Called on failure; free up any blocks from the inode PA for this
3119 * context. We don't need this for MB_GROUP_PA because we only change
3120 * pa_free in ext4_mb_release_context(), but on failure, we've already
3121 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3123 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3125 struct ext4_prealloc_space *pa = ac->ac_pa;
3126 int len;
3128 if (pa && pa->pa_type == MB_INODE_PA) {
3129 len = ac->ac_b_ex.fe_len;
3130 pa->pa_free += len;
3136 * use blocks preallocated to inode
3138 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3139 struct ext4_prealloc_space *pa)
3141 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3142 ext4_fsblk_t start;
3143 ext4_fsblk_t end;
3144 int len;
3146 /* found preallocated blocks, use them */
3147 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3148 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3149 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3150 len = EXT4_NUM_B2C(sbi, end - start);
3151 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3152 &ac->ac_b_ex.fe_start);
3153 ac->ac_b_ex.fe_len = len;
3154 ac->ac_status = AC_STATUS_FOUND;
3155 ac->ac_pa = pa;
3157 BUG_ON(start < pa->pa_pstart);
3158 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3159 BUG_ON(pa->pa_free < len);
3160 pa->pa_free -= len;
3162 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3166 * use blocks preallocated to locality group
3168 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3169 struct ext4_prealloc_space *pa)
3171 unsigned int len = ac->ac_o_ex.fe_len;
3173 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3174 &ac->ac_b_ex.fe_group,
3175 &ac->ac_b_ex.fe_start);
3176 ac->ac_b_ex.fe_len = len;
3177 ac->ac_status = AC_STATUS_FOUND;
3178 ac->ac_pa = pa;
3180 /* we don't correct pa_pstart or pa_plen here to avoid
3181 * possible race when the group is being loaded concurrently
3182 * instead we correct pa later, after blocks are marked
3183 * in on-disk bitmap -- see ext4_mb_release_context()
3184 * Other CPUs are prevented from allocating from this pa by lg_mutex
3186 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3190 * Return the prealloc space that have minimal distance
3191 * from the goal block. @cpa is the prealloc
3192 * space that is having currently known minimal distance
3193 * from the goal block.
3195 static struct ext4_prealloc_space *
3196 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3197 struct ext4_prealloc_space *pa,
3198 struct ext4_prealloc_space *cpa)
3200 ext4_fsblk_t cur_distance, new_distance;
3202 if (cpa == NULL) {
3203 atomic_inc(&pa->pa_count);
3204 return pa;
3206 cur_distance = abs(goal_block - cpa->pa_pstart);
3207 new_distance = abs(goal_block - pa->pa_pstart);
3209 if (cur_distance <= new_distance)
3210 return cpa;
3212 /* drop the previous reference */
3213 atomic_dec(&cpa->pa_count);
3214 atomic_inc(&pa->pa_count);
3215 return pa;
3219 * search goal blocks in preallocated space
3221 static noinline_for_stack int
3222 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3224 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3225 int order, i;
3226 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3227 struct ext4_locality_group *lg;
3228 struct ext4_prealloc_space *pa, *cpa = NULL;
3229 ext4_fsblk_t goal_block;
3231 /* only data can be preallocated */
3232 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3233 return 0;
3235 /* first, try per-file preallocation */
3236 rcu_read_lock();
3237 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3239 /* all fields in this condition don't change,
3240 * so we can skip locking for them */
3241 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3242 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3243 EXT4_C2B(sbi, pa->pa_len)))
3244 continue;
3246 /* non-extent files can't have physical blocks past 2^32 */
3247 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3248 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3249 EXT4_MAX_BLOCK_FILE_PHYS))
3250 continue;
3252 /* found preallocated blocks, use them */
3253 spin_lock(&pa->pa_lock);
3254 if (pa->pa_deleted == 0 && pa->pa_free) {
3255 atomic_inc(&pa->pa_count);
3256 ext4_mb_use_inode_pa(ac, pa);
3257 spin_unlock(&pa->pa_lock);
3258 ac->ac_criteria = 10;
3259 rcu_read_unlock();
3260 return 1;
3262 spin_unlock(&pa->pa_lock);
3264 rcu_read_unlock();
3266 /* can we use group allocation? */
3267 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3268 return 0;
3270 /* inode may have no locality group for some reason */
3271 lg = ac->ac_lg;
3272 if (lg == NULL)
3273 return 0;
3274 order = fls(ac->ac_o_ex.fe_len) - 1;
3275 if (order > PREALLOC_TB_SIZE - 1)
3276 /* The max size of hash table is PREALLOC_TB_SIZE */
3277 order = PREALLOC_TB_SIZE - 1;
3279 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3281 * search for the prealloc space that is having
3282 * minimal distance from the goal block.
3284 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3285 rcu_read_lock();
3286 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3287 pa_inode_list) {
3288 spin_lock(&pa->pa_lock);
3289 if (pa->pa_deleted == 0 &&
3290 pa->pa_free >= ac->ac_o_ex.fe_len) {
3292 cpa = ext4_mb_check_group_pa(goal_block,
3293 pa, cpa);
3295 spin_unlock(&pa->pa_lock);
3297 rcu_read_unlock();
3299 if (cpa) {
3300 ext4_mb_use_group_pa(ac, cpa);
3301 ac->ac_criteria = 20;
3302 return 1;
3304 return 0;
3308 * the function goes through all block freed in the group
3309 * but not yet committed and marks them used in in-core bitmap.
3310 * buddy must be generated from this bitmap
3311 * Need to be called with the ext4 group lock held
3313 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3314 ext4_group_t group)
3316 struct rb_node *n;
3317 struct ext4_group_info *grp;
3318 struct ext4_free_data *entry;
3320 grp = ext4_get_group_info(sb, group);
3321 n = rb_first(&(grp->bb_free_root));
3323 while (n) {
3324 entry = rb_entry(n, struct ext4_free_data, node);
3325 ext4_set_bits(bitmap, entry->start_cluster, entry->count);
3326 n = rb_next(n);
3328 return;
3332 * the function goes through all preallocation in this group and marks them
3333 * used in in-core bitmap. buddy must be generated from this bitmap
3334 * Need to be called with ext4 group lock held
3336 static noinline_for_stack
3337 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3338 ext4_group_t group)
3340 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3341 struct ext4_prealloc_space *pa;
3342 struct list_head *cur;
3343 ext4_group_t groupnr;
3344 ext4_grpblk_t start;
3345 int preallocated = 0;
3346 int len;
3348 /* all form of preallocation discards first load group,
3349 * so the only competing code is preallocation use.
3350 * we don't need any locking here
3351 * notice we do NOT ignore preallocations with pa_deleted
3352 * otherwise we could leave used blocks available for
3353 * allocation in buddy when concurrent ext4_mb_put_pa()
3354 * is dropping preallocation
3356 list_for_each(cur, &grp->bb_prealloc_list) {
3357 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3358 spin_lock(&pa->pa_lock);
3359 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3360 &groupnr, &start);
3361 len = pa->pa_len;
3362 spin_unlock(&pa->pa_lock);
3363 if (unlikely(len == 0))
3364 continue;
3365 BUG_ON(groupnr != group);
3366 ext4_set_bits(bitmap, start, len);
3367 preallocated += len;
3369 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3372 static void ext4_mb_pa_callback(struct rcu_head *head)
3374 struct ext4_prealloc_space *pa;
3375 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3376 kmem_cache_free(ext4_pspace_cachep, pa);
3380 * drops a reference to preallocated space descriptor
3381 * if this was the last reference and the space is consumed
3383 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3384 struct super_block *sb, struct ext4_prealloc_space *pa)
3386 ext4_group_t grp;
3387 ext4_fsblk_t grp_blk;
3389 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3390 return;
3392 /* in this short window concurrent discard can set pa_deleted */
3393 spin_lock(&pa->pa_lock);
3394 if (pa->pa_deleted == 1) {
3395 spin_unlock(&pa->pa_lock);
3396 return;
3399 pa->pa_deleted = 1;
3400 spin_unlock(&pa->pa_lock);
3402 grp_blk = pa->pa_pstart;
3404 * If doing group-based preallocation, pa_pstart may be in the
3405 * next group when pa is used up
3407 if (pa->pa_type == MB_GROUP_PA)
3408 grp_blk--;
3410 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3413 * possible race:
3415 * P1 (buddy init) P2 (regular allocation)
3416 * find block B in PA
3417 * copy on-disk bitmap to buddy
3418 * mark B in on-disk bitmap
3419 * drop PA from group
3420 * mark all PAs in buddy
3422 * thus, P1 initializes buddy with B available. to prevent this
3423 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3424 * against that pair
3426 ext4_lock_group(sb, grp);
3427 list_del(&pa->pa_group_list);
3428 ext4_unlock_group(sb, grp);
3430 spin_lock(pa->pa_obj_lock);
3431 list_del_rcu(&pa->pa_inode_list);
3432 spin_unlock(pa->pa_obj_lock);
3434 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3438 * creates new preallocated space for given inode
3440 static noinline_for_stack int
3441 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3443 struct super_block *sb = ac->ac_sb;
3444 struct ext4_sb_info *sbi = EXT4_SB(sb);
3445 struct ext4_prealloc_space *pa;
3446 struct ext4_group_info *grp;
3447 struct ext4_inode_info *ei;
3449 /* preallocate only when found space is larger then requested */
3450 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3451 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3452 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3454 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3455 if (pa == NULL)
3456 return -ENOMEM;
3458 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3459 int winl;
3460 int wins;
3461 int win;
3462 int offs;
3464 /* we can't allocate as much as normalizer wants.
3465 * so, found space must get proper lstart
3466 * to cover original request */
3467 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3468 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3470 /* we're limited by original request in that
3471 * logical block must be covered any way
3472 * winl is window we can move our chunk within */
3473 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3475 /* also, we should cover whole original request */
3476 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3478 /* the smallest one defines real window */
3479 win = min(winl, wins);
3481 offs = ac->ac_o_ex.fe_logical %
3482 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3483 if (offs && offs < win)
3484 win = offs;
3486 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3487 EXT4_B2C(sbi, win);
3488 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3489 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3492 /* preallocation can change ac_b_ex, thus we store actually
3493 * allocated blocks for history */
3494 ac->ac_f_ex = ac->ac_b_ex;
3496 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3497 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3498 pa->pa_len = ac->ac_b_ex.fe_len;
3499 pa->pa_free = pa->pa_len;
3500 atomic_set(&pa->pa_count, 1);
3501 spin_lock_init(&pa->pa_lock);
3502 INIT_LIST_HEAD(&pa->pa_inode_list);
3503 INIT_LIST_HEAD(&pa->pa_group_list);
3504 pa->pa_deleted = 0;
3505 pa->pa_type = MB_INODE_PA;
3507 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3508 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3509 trace_ext4_mb_new_inode_pa(ac, pa);
3511 ext4_mb_use_inode_pa(ac, pa);
3512 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3514 ei = EXT4_I(ac->ac_inode);
3515 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3517 pa->pa_obj_lock = &ei->i_prealloc_lock;
3518 pa->pa_inode = ac->ac_inode;
3520 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3521 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3522 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3524 spin_lock(pa->pa_obj_lock);
3525 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3526 spin_unlock(pa->pa_obj_lock);
3528 return 0;
3532 * creates new preallocated space for locality group inodes belongs to
3534 static noinline_for_stack int
3535 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3537 struct super_block *sb = ac->ac_sb;
3538 struct ext4_locality_group *lg;
3539 struct ext4_prealloc_space *pa;
3540 struct ext4_group_info *grp;
3542 /* preallocate only when found space is larger then requested */
3543 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3544 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3545 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3547 BUG_ON(ext4_pspace_cachep == NULL);
3548 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3549 if (pa == NULL)
3550 return -ENOMEM;
3552 /* preallocation can change ac_b_ex, thus we store actually
3553 * allocated blocks for history */
3554 ac->ac_f_ex = ac->ac_b_ex;
3556 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3557 pa->pa_lstart = pa->pa_pstart;
3558 pa->pa_len = ac->ac_b_ex.fe_len;
3559 pa->pa_free = pa->pa_len;
3560 atomic_set(&pa->pa_count, 1);
3561 spin_lock_init(&pa->pa_lock);
3562 INIT_LIST_HEAD(&pa->pa_inode_list);
3563 INIT_LIST_HEAD(&pa->pa_group_list);
3564 pa->pa_deleted = 0;
3565 pa->pa_type = MB_GROUP_PA;
3567 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3568 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3569 trace_ext4_mb_new_group_pa(ac, pa);
3571 ext4_mb_use_group_pa(ac, pa);
3572 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3574 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3575 lg = ac->ac_lg;
3576 BUG_ON(lg == NULL);
3578 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3579 pa->pa_inode = NULL;
3581 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3582 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3583 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3586 * We will later add the new pa to the right bucket
3587 * after updating the pa_free in ext4_mb_release_context
3589 return 0;
3592 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3594 int err;
3596 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3597 err = ext4_mb_new_group_pa(ac);
3598 else
3599 err = ext4_mb_new_inode_pa(ac);
3600 return err;
3604 * finds all unused blocks in on-disk bitmap, frees them in
3605 * in-core bitmap and buddy.
3606 * @pa must be unlinked from inode and group lists, so that
3607 * nobody else can find/use it.
3608 * the caller MUST hold group/inode locks.
3609 * TODO: optimize the case when there are no in-core structures yet
3611 static noinline_for_stack int
3612 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3613 struct ext4_prealloc_space *pa)
3615 struct super_block *sb = e4b->bd_sb;
3616 struct ext4_sb_info *sbi = EXT4_SB(sb);
3617 unsigned int end;
3618 unsigned int next;
3619 ext4_group_t group;
3620 ext4_grpblk_t bit;
3621 unsigned long long grp_blk_start;
3622 int err = 0;
3623 int free = 0;
3625 BUG_ON(pa->pa_deleted == 0);
3626 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3627 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3628 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3629 end = bit + pa->pa_len;
3631 while (bit < end) {
3632 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3633 if (bit >= end)
3634 break;
3635 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3636 mb_debug(1, " free preallocated %u/%u in group %u\n",
3637 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3638 (unsigned) next - bit, (unsigned) group);
3639 free += next - bit;
3641 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3642 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3643 EXT4_C2B(sbi, bit)),
3644 next - bit);
3645 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3646 bit = next + 1;
3648 if (free != pa->pa_free) {
3649 ext4_msg(e4b->bd_sb, KERN_CRIT,
3650 "pa %p: logic %lu, phys. %lu, len %lu",
3651 pa, (unsigned long) pa->pa_lstart,
3652 (unsigned long) pa->pa_pstart,
3653 (unsigned long) pa->pa_len);
3654 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3655 free, pa->pa_free);
3657 * pa is already deleted so we use the value obtained
3658 * from the bitmap and continue.
3661 atomic_add(free, &sbi->s_mb_discarded);
3663 return err;
3666 static noinline_for_stack int
3667 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3668 struct ext4_prealloc_space *pa)
3670 struct super_block *sb = e4b->bd_sb;
3671 ext4_group_t group;
3672 ext4_grpblk_t bit;
3674 trace_ext4_mb_release_group_pa(pa);
3675 BUG_ON(pa->pa_deleted == 0);
3676 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3677 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3678 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3679 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3680 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3682 return 0;
3686 * releases all preallocations in given group
3688 * first, we need to decide discard policy:
3689 * - when do we discard
3690 * 1) ENOSPC
3691 * - how many do we discard
3692 * 1) how many requested
3694 static noinline_for_stack int
3695 ext4_mb_discard_group_preallocations(struct super_block *sb,
3696 ext4_group_t group, int needed)
3698 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3699 struct buffer_head *bitmap_bh = NULL;
3700 struct ext4_prealloc_space *pa, *tmp;
3701 struct list_head list;
3702 struct ext4_buddy e4b;
3703 int err;
3704 int busy = 0;
3705 int free = 0;
3707 mb_debug(1, "discard preallocation for group %u\n", group);
3709 if (list_empty(&grp->bb_prealloc_list))
3710 return 0;
3712 bitmap_bh = ext4_read_block_bitmap(sb, group);
3713 if (bitmap_bh == NULL) {
3714 ext4_error(sb, "Error reading block bitmap for %u", group);
3715 return 0;
3718 err = ext4_mb_load_buddy(sb, group, &e4b);
3719 if (err) {
3720 ext4_error(sb, "Error loading buddy information for %u", group);
3721 put_bh(bitmap_bh);
3722 return 0;
3725 if (needed == 0)
3726 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3728 INIT_LIST_HEAD(&list);
3729 repeat:
3730 ext4_lock_group(sb, group);
3731 list_for_each_entry_safe(pa, tmp,
3732 &grp->bb_prealloc_list, pa_group_list) {
3733 spin_lock(&pa->pa_lock);
3734 if (atomic_read(&pa->pa_count)) {
3735 spin_unlock(&pa->pa_lock);
3736 busy = 1;
3737 continue;
3739 if (pa->pa_deleted) {
3740 spin_unlock(&pa->pa_lock);
3741 continue;
3744 /* seems this one can be freed ... */
3745 pa->pa_deleted = 1;
3747 /* we can trust pa_free ... */
3748 free += pa->pa_free;
3750 spin_unlock(&pa->pa_lock);
3752 list_del(&pa->pa_group_list);
3753 list_add(&pa->u.pa_tmp_list, &list);
3756 /* if we still need more blocks and some PAs were used, try again */
3757 if (free < needed && busy) {
3758 busy = 0;
3759 ext4_unlock_group(sb, group);
3761 * Yield the CPU here so that we don't get soft lockup
3762 * in non preempt case.
3764 yield();
3765 goto repeat;
3768 /* found anything to free? */
3769 if (list_empty(&list)) {
3770 BUG_ON(free != 0);
3771 goto out;
3774 /* now free all selected PAs */
3775 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3777 /* remove from object (inode or locality group) */
3778 spin_lock(pa->pa_obj_lock);
3779 list_del_rcu(&pa->pa_inode_list);
3780 spin_unlock(pa->pa_obj_lock);
3782 if (pa->pa_type == MB_GROUP_PA)
3783 ext4_mb_release_group_pa(&e4b, pa);
3784 else
3785 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3787 list_del(&pa->u.pa_tmp_list);
3788 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3791 out:
3792 ext4_unlock_group(sb, group);
3793 ext4_mb_unload_buddy(&e4b);
3794 put_bh(bitmap_bh);
3795 return free;
3799 * releases all non-used preallocated blocks for given inode
3801 * It's important to discard preallocations under i_data_sem
3802 * We don't want another block to be served from the prealloc
3803 * space when we are discarding the inode prealloc space.
3805 * FIXME!! Make sure it is valid at all the call sites
3807 void ext4_discard_preallocations(struct inode *inode)
3809 struct ext4_inode_info *ei = EXT4_I(inode);
3810 struct super_block *sb = inode->i_sb;
3811 struct buffer_head *bitmap_bh = NULL;
3812 struct ext4_prealloc_space *pa, *tmp;
3813 ext4_group_t group = 0;
3814 struct list_head list;
3815 struct ext4_buddy e4b;
3816 int err;
3818 if (!S_ISREG(inode->i_mode)) {
3819 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3820 return;
3823 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3824 trace_ext4_discard_preallocations(inode);
3826 INIT_LIST_HEAD(&list);
3828 repeat:
3829 /* first, collect all pa's in the inode */
3830 spin_lock(&ei->i_prealloc_lock);
3831 while (!list_empty(&ei->i_prealloc_list)) {
3832 pa = list_entry(ei->i_prealloc_list.next,
3833 struct ext4_prealloc_space, pa_inode_list);
3834 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3835 spin_lock(&pa->pa_lock);
3836 if (atomic_read(&pa->pa_count)) {
3837 /* this shouldn't happen often - nobody should
3838 * use preallocation while we're discarding it */
3839 spin_unlock(&pa->pa_lock);
3840 spin_unlock(&ei->i_prealloc_lock);
3841 ext4_msg(sb, KERN_ERR,
3842 "uh-oh! used pa while discarding");
3843 WARN_ON(1);
3844 schedule_timeout_uninterruptible(HZ);
3845 goto repeat;
3848 if (pa->pa_deleted == 0) {
3849 pa->pa_deleted = 1;
3850 spin_unlock(&pa->pa_lock);
3851 list_del_rcu(&pa->pa_inode_list);
3852 list_add(&pa->u.pa_tmp_list, &list);
3853 continue;
3856 /* someone is deleting pa right now */
3857 spin_unlock(&pa->pa_lock);
3858 spin_unlock(&ei->i_prealloc_lock);
3860 /* we have to wait here because pa_deleted
3861 * doesn't mean pa is already unlinked from
3862 * the list. as we might be called from
3863 * ->clear_inode() the inode will get freed
3864 * and concurrent thread which is unlinking
3865 * pa from inode's list may access already
3866 * freed memory, bad-bad-bad */
3868 /* XXX: if this happens too often, we can
3869 * add a flag to force wait only in case
3870 * of ->clear_inode(), but not in case of
3871 * regular truncate */
3872 schedule_timeout_uninterruptible(HZ);
3873 goto repeat;
3875 spin_unlock(&ei->i_prealloc_lock);
3877 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3878 BUG_ON(pa->pa_type != MB_INODE_PA);
3879 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3881 err = ext4_mb_load_buddy(sb, group, &e4b);
3882 if (err) {
3883 ext4_error(sb, "Error loading buddy information for %u",
3884 group);
3885 continue;
3888 bitmap_bh = ext4_read_block_bitmap(sb, group);
3889 if (bitmap_bh == NULL) {
3890 ext4_error(sb, "Error reading block bitmap for %u",
3891 group);
3892 ext4_mb_unload_buddy(&e4b);
3893 continue;
3896 ext4_lock_group(sb, group);
3897 list_del(&pa->pa_group_list);
3898 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3899 ext4_unlock_group(sb, group);
3901 ext4_mb_unload_buddy(&e4b);
3902 put_bh(bitmap_bh);
3904 list_del(&pa->u.pa_tmp_list);
3905 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3909 #ifdef CONFIG_EXT4_DEBUG
3910 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3912 struct super_block *sb = ac->ac_sb;
3913 ext4_group_t ngroups, i;
3915 if (!mb_enable_debug ||
3916 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3917 return;
3919 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:"
3920 " Allocation context details:");
3921 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d",
3922 ac->ac_status, ac->ac_flags);
3923 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, "
3924 "goal %lu/%lu/%lu@%lu, "
3925 "best %lu/%lu/%lu@%lu cr %d",
3926 (unsigned long)ac->ac_o_ex.fe_group,
3927 (unsigned long)ac->ac_o_ex.fe_start,
3928 (unsigned long)ac->ac_o_ex.fe_len,
3929 (unsigned long)ac->ac_o_ex.fe_logical,
3930 (unsigned long)ac->ac_g_ex.fe_group,
3931 (unsigned long)ac->ac_g_ex.fe_start,
3932 (unsigned long)ac->ac_g_ex.fe_len,
3933 (unsigned long)ac->ac_g_ex.fe_logical,
3934 (unsigned long)ac->ac_b_ex.fe_group,
3935 (unsigned long)ac->ac_b_ex.fe_start,
3936 (unsigned long)ac->ac_b_ex.fe_len,
3937 (unsigned long)ac->ac_b_ex.fe_logical,
3938 (int)ac->ac_criteria);
3939 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found",
3940 ac->ac_ex_scanned, ac->ac_found);
3941 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: ");
3942 ngroups = ext4_get_groups_count(sb);
3943 for (i = 0; i < ngroups; i++) {
3944 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3945 struct ext4_prealloc_space *pa;
3946 ext4_grpblk_t start;
3947 struct list_head *cur;
3948 ext4_lock_group(sb, i);
3949 list_for_each(cur, &grp->bb_prealloc_list) {
3950 pa = list_entry(cur, struct ext4_prealloc_space,
3951 pa_group_list);
3952 spin_lock(&pa->pa_lock);
3953 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3954 NULL, &start);
3955 spin_unlock(&pa->pa_lock);
3956 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3957 start, pa->pa_len);
3959 ext4_unlock_group(sb, i);
3961 if (grp->bb_free == 0)
3962 continue;
3963 printk(KERN_ERR "%u: %d/%d \n",
3964 i, grp->bb_free, grp->bb_fragments);
3966 printk(KERN_ERR "\n");
3968 #else
3969 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3971 return;
3973 #endif
3976 * We use locality group preallocation for small size file. The size of the
3977 * file is determined by the current size or the resulting size after
3978 * allocation which ever is larger
3980 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3982 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3984 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3985 int bsbits = ac->ac_sb->s_blocksize_bits;
3986 loff_t size, isize;
3988 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3989 return;
3991 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3992 return;
3994 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3995 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3996 >> bsbits;
3998 if ((size == isize) &&
3999 !ext4_fs_is_busy(sbi) &&
4000 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4001 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4002 return;
4005 if (sbi->s_mb_group_prealloc <= 0) {
4006 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4007 return;
4010 /* don't use group allocation for large files */
4011 size = max(size, isize);
4012 if (size > sbi->s_mb_stream_request) {
4013 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4014 return;
4017 BUG_ON(ac->ac_lg != NULL);
4019 * locality group prealloc space are per cpu. The reason for having
4020 * per cpu locality group is to reduce the contention between block
4021 * request from multiple CPUs.
4023 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4025 /* we're going to use group allocation */
4026 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4028 /* serialize all allocations in the group */
4029 mutex_lock(&ac->ac_lg->lg_mutex);
4032 static noinline_for_stack int
4033 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4034 struct ext4_allocation_request *ar)
4036 struct super_block *sb = ar->inode->i_sb;
4037 struct ext4_sb_info *sbi = EXT4_SB(sb);
4038 struct ext4_super_block *es = sbi->s_es;
4039 ext4_group_t group;
4040 unsigned int len;
4041 ext4_fsblk_t goal;
4042 ext4_grpblk_t block;
4044 /* we can't allocate > group size */
4045 len = ar->len;
4047 /* just a dirty hack to filter too big requests */
4048 if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4049 len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4051 /* start searching from the goal */
4052 goal = ar->goal;
4053 if (goal < le32_to_cpu(es->s_first_data_block) ||
4054 goal >= ext4_blocks_count(es))
4055 goal = le32_to_cpu(es->s_first_data_block);
4056 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4058 /* set up allocation goals */
4059 memset(ac, 0, sizeof(struct ext4_allocation_context));
4060 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4061 ac->ac_status = AC_STATUS_CONTINUE;
4062 ac->ac_sb = sb;
4063 ac->ac_inode = ar->inode;
4064 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4065 ac->ac_o_ex.fe_group = group;
4066 ac->ac_o_ex.fe_start = block;
4067 ac->ac_o_ex.fe_len = len;
4068 ac->ac_g_ex = ac->ac_o_ex;
4069 ac->ac_flags = ar->flags;
4071 /* we have to define context: we'll we work with a file or
4072 * locality group. this is a policy, actually */
4073 ext4_mb_group_or_file(ac);
4075 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4076 "left: %u/%u, right %u/%u to %swritable\n",
4077 (unsigned) ar->len, (unsigned) ar->logical,
4078 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4079 (unsigned) ar->lleft, (unsigned) ar->pleft,
4080 (unsigned) ar->lright, (unsigned) ar->pright,
4081 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4082 return 0;
4086 static noinline_for_stack void
4087 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4088 struct ext4_locality_group *lg,
4089 int order, int total_entries)
4091 ext4_group_t group = 0;
4092 struct ext4_buddy e4b;
4093 struct list_head discard_list;
4094 struct ext4_prealloc_space *pa, *tmp;
4096 mb_debug(1, "discard locality group preallocation\n");
4098 INIT_LIST_HEAD(&discard_list);
4100 spin_lock(&lg->lg_prealloc_lock);
4101 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4102 pa_inode_list) {
4103 spin_lock(&pa->pa_lock);
4104 if (atomic_read(&pa->pa_count)) {
4106 * This is the pa that we just used
4107 * for block allocation. So don't
4108 * free that
4110 spin_unlock(&pa->pa_lock);
4111 continue;
4113 if (pa->pa_deleted) {
4114 spin_unlock(&pa->pa_lock);
4115 continue;
4117 /* only lg prealloc space */
4118 BUG_ON(pa->pa_type != MB_GROUP_PA);
4120 /* seems this one can be freed ... */
4121 pa->pa_deleted = 1;
4122 spin_unlock(&pa->pa_lock);
4124 list_del_rcu(&pa->pa_inode_list);
4125 list_add(&pa->u.pa_tmp_list, &discard_list);
4127 total_entries--;
4128 if (total_entries <= 5) {
4130 * we want to keep only 5 entries
4131 * allowing it to grow to 8. This
4132 * mak sure we don't call discard
4133 * soon for this list.
4135 break;
4138 spin_unlock(&lg->lg_prealloc_lock);
4140 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4142 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4143 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4144 ext4_error(sb, "Error loading buddy information for %u",
4145 group);
4146 continue;
4148 ext4_lock_group(sb, group);
4149 list_del(&pa->pa_group_list);
4150 ext4_mb_release_group_pa(&e4b, pa);
4151 ext4_unlock_group(sb, group);
4153 ext4_mb_unload_buddy(&e4b);
4154 list_del(&pa->u.pa_tmp_list);
4155 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4160 * We have incremented pa_count. So it cannot be freed at this
4161 * point. Also we hold lg_mutex. So no parallel allocation is
4162 * possible from this lg. That means pa_free cannot be updated.
4164 * A parallel ext4_mb_discard_group_preallocations is possible.
4165 * which can cause the lg_prealloc_list to be updated.
4168 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4170 int order, added = 0, lg_prealloc_count = 1;
4171 struct super_block *sb = ac->ac_sb;
4172 struct ext4_locality_group *lg = ac->ac_lg;
4173 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4175 order = fls(pa->pa_free) - 1;
4176 if (order > PREALLOC_TB_SIZE - 1)
4177 /* The max size of hash table is PREALLOC_TB_SIZE */
4178 order = PREALLOC_TB_SIZE - 1;
4179 /* Add the prealloc space to lg */
4180 rcu_read_lock();
4181 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4182 pa_inode_list) {
4183 spin_lock(&tmp_pa->pa_lock);
4184 if (tmp_pa->pa_deleted) {
4185 spin_unlock(&tmp_pa->pa_lock);
4186 continue;
4188 if (!added && pa->pa_free < tmp_pa->pa_free) {
4189 /* Add to the tail of the previous entry */
4190 list_add_tail_rcu(&pa->pa_inode_list,
4191 &tmp_pa->pa_inode_list);
4192 added = 1;
4194 * we want to count the total
4195 * number of entries in the list
4198 spin_unlock(&tmp_pa->pa_lock);
4199 lg_prealloc_count++;
4201 if (!added)
4202 list_add_tail_rcu(&pa->pa_inode_list,
4203 &lg->lg_prealloc_list[order]);
4204 rcu_read_unlock();
4206 /* Now trim the list to be not more than 8 elements */
4207 if (lg_prealloc_count > 8) {
4208 ext4_mb_discard_lg_preallocations(sb, lg,
4209 order, lg_prealloc_count);
4210 return;
4212 return ;
4216 * release all resource we used in allocation
4218 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4220 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4221 struct ext4_prealloc_space *pa = ac->ac_pa;
4222 if (pa) {
4223 if (pa->pa_type == MB_GROUP_PA) {
4224 /* see comment in ext4_mb_use_group_pa() */
4225 spin_lock(&pa->pa_lock);
4226 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4227 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4228 pa->pa_free -= ac->ac_b_ex.fe_len;
4229 pa->pa_len -= ac->ac_b_ex.fe_len;
4230 spin_unlock(&pa->pa_lock);
4233 if (pa) {
4235 * We want to add the pa to the right bucket.
4236 * Remove it from the list and while adding
4237 * make sure the list to which we are adding
4238 * doesn't grow big.
4240 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4241 spin_lock(pa->pa_obj_lock);
4242 list_del_rcu(&pa->pa_inode_list);
4243 spin_unlock(pa->pa_obj_lock);
4244 ext4_mb_add_n_trim(ac);
4246 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4248 if (ac->ac_bitmap_page)
4249 page_cache_release(ac->ac_bitmap_page);
4250 if (ac->ac_buddy_page)
4251 page_cache_release(ac->ac_buddy_page);
4252 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4253 mutex_unlock(&ac->ac_lg->lg_mutex);
4254 ext4_mb_collect_stats(ac);
4255 return 0;
4258 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4260 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4261 int ret;
4262 int freed = 0;
4264 trace_ext4_mb_discard_preallocations(sb, needed);
4265 for (i = 0; i < ngroups && needed > 0; i++) {
4266 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4267 freed += ret;
4268 needed -= ret;
4271 return freed;
4275 * Main entry point into mballoc to allocate blocks
4276 * it tries to use preallocation first, then falls back
4277 * to usual allocation
4279 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4280 struct ext4_allocation_request *ar, int *errp)
4282 int freed;
4283 struct ext4_allocation_context *ac = NULL;
4284 struct ext4_sb_info *sbi;
4285 struct super_block *sb;
4286 ext4_fsblk_t block = 0;
4287 unsigned int inquota = 0;
4288 unsigned int reserv_clstrs = 0;
4290 sb = ar->inode->i_sb;
4291 sbi = EXT4_SB(sb);
4293 trace_ext4_request_blocks(ar);
4295 /* Allow to use superuser reservation for quota file */
4296 if (IS_NOQUOTA(ar->inode))
4297 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4300 * For delayed allocation, we could skip the ENOSPC and
4301 * EDQUOT check, as blocks and quotas have been already
4302 * reserved when data being copied into pagecache.
4304 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4305 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4306 else {
4307 /* Without delayed allocation we need to verify
4308 * there is enough free blocks to do block allocation
4309 * and verify allocation doesn't exceed the quota limits.
4311 while (ar->len &&
4312 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4314 /* let others to free the space */
4315 yield();
4316 ar->len = ar->len >> 1;
4318 if (!ar->len) {
4319 *errp = -ENOSPC;
4320 return 0;
4322 reserv_clstrs = ar->len;
4323 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4324 dquot_alloc_block_nofail(ar->inode,
4325 EXT4_C2B(sbi, ar->len));
4326 } else {
4327 while (ar->len &&
4328 dquot_alloc_block(ar->inode,
4329 EXT4_C2B(sbi, ar->len))) {
4331 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4332 ar->len--;
4335 inquota = ar->len;
4336 if (ar->len == 0) {
4337 *errp = -EDQUOT;
4338 goto out;
4342 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4343 if (!ac) {
4344 ar->len = 0;
4345 *errp = -ENOMEM;
4346 goto out;
4349 *errp = ext4_mb_initialize_context(ac, ar);
4350 if (*errp) {
4351 ar->len = 0;
4352 goto out;
4355 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4356 if (!ext4_mb_use_preallocated(ac)) {
4357 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4358 ext4_mb_normalize_request(ac, ar);
4359 repeat:
4360 /* allocate space in core */
4361 *errp = ext4_mb_regular_allocator(ac);
4362 if (*errp)
4363 goto errout;
4365 /* as we've just preallocated more space than
4366 * user requested orinally, we store allocated
4367 * space in a special descriptor */
4368 if (ac->ac_status == AC_STATUS_FOUND &&
4369 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4370 ext4_mb_new_preallocation(ac);
4372 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4373 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4374 if (*errp == -EAGAIN) {
4376 * drop the reference that we took
4377 * in ext4_mb_use_best_found
4379 ext4_mb_release_context(ac);
4380 ac->ac_b_ex.fe_group = 0;
4381 ac->ac_b_ex.fe_start = 0;
4382 ac->ac_b_ex.fe_len = 0;
4383 ac->ac_status = AC_STATUS_CONTINUE;
4384 goto repeat;
4385 } else if (*errp)
4386 errout:
4387 ext4_discard_allocated_blocks(ac);
4388 else {
4389 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4390 ar->len = ac->ac_b_ex.fe_len;
4392 } else {
4393 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4394 if (freed)
4395 goto repeat;
4396 *errp = -ENOSPC;
4399 if (*errp) {
4400 ac->ac_b_ex.fe_len = 0;
4401 ar->len = 0;
4402 ext4_mb_show_ac(ac);
4404 ext4_mb_release_context(ac);
4405 out:
4406 if (ac)
4407 kmem_cache_free(ext4_ac_cachep, ac);
4408 if (inquota && ar->len < inquota)
4409 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4410 if (!ar->len) {
4411 if (!ext4_test_inode_state(ar->inode,
4412 EXT4_STATE_DELALLOC_RESERVED))
4413 /* release all the reserved blocks if non delalloc */
4414 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4415 reserv_clstrs);
4418 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4420 return block;
4424 * We can merge two free data extents only if the physical blocks
4425 * are contiguous, AND the extents were freed by the same transaction,
4426 * AND the blocks are associated with the same group.
4428 static int can_merge(struct ext4_free_data *entry1,
4429 struct ext4_free_data *entry2)
4431 if ((entry1->t_tid == entry2->t_tid) &&
4432 (entry1->group == entry2->group) &&
4433 ((entry1->start_cluster + entry1->count) == entry2->start_cluster))
4434 return 1;
4435 return 0;
4438 static noinline_for_stack int
4439 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4440 struct ext4_free_data *new_entry)
4442 ext4_group_t group = e4b->bd_group;
4443 ext4_grpblk_t cluster;
4444 struct ext4_free_data *entry;
4445 struct ext4_group_info *db = e4b->bd_info;
4446 struct super_block *sb = e4b->bd_sb;
4447 struct ext4_sb_info *sbi = EXT4_SB(sb);
4448 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4449 struct rb_node *parent = NULL, *new_node;
4451 BUG_ON(!ext4_handle_valid(handle));
4452 BUG_ON(e4b->bd_bitmap_page == NULL);
4453 BUG_ON(e4b->bd_buddy_page == NULL);
4455 new_node = &new_entry->node;
4456 cluster = new_entry->start_cluster;
4458 if (!*n) {
4459 /* first free block exent. We need to
4460 protect buddy cache from being freed,
4461 * otherwise we'll refresh it from
4462 * on-disk bitmap and lose not-yet-available
4463 * blocks */
4464 page_cache_get(e4b->bd_buddy_page);
4465 page_cache_get(e4b->bd_bitmap_page);
4467 while (*n) {
4468 parent = *n;
4469 entry = rb_entry(parent, struct ext4_free_data, node);
4470 if (cluster < entry->start_cluster)
4471 n = &(*n)->rb_left;
4472 else if (cluster >= (entry->start_cluster + entry->count))
4473 n = &(*n)->rb_right;
4474 else {
4475 ext4_grp_locked_error(sb, group, 0,
4476 ext4_group_first_block_no(sb, group) +
4477 EXT4_C2B(sbi, cluster),
4478 "Block already on to-be-freed list");
4479 return 0;
4483 rb_link_node(new_node, parent, n);
4484 rb_insert_color(new_node, &db->bb_free_root);
4486 /* Now try to see the extent can be merged to left and right */
4487 node = rb_prev(new_node);
4488 if (node) {
4489 entry = rb_entry(node, struct ext4_free_data, node);
4490 if (can_merge(entry, new_entry)) {
4491 new_entry->start_cluster = entry->start_cluster;
4492 new_entry->count += entry->count;
4493 rb_erase(node, &(db->bb_free_root));
4494 spin_lock(&sbi->s_md_lock);
4495 list_del(&entry->list);
4496 spin_unlock(&sbi->s_md_lock);
4497 kmem_cache_free(ext4_free_ext_cachep, entry);
4501 node = rb_next(new_node);
4502 if (node) {
4503 entry = rb_entry(node, struct ext4_free_data, node);
4504 if (can_merge(new_entry, entry)) {
4505 new_entry->count += entry->count;
4506 rb_erase(node, &(db->bb_free_root));
4507 spin_lock(&sbi->s_md_lock);
4508 list_del(&entry->list);
4509 spin_unlock(&sbi->s_md_lock);
4510 kmem_cache_free(ext4_free_ext_cachep, entry);
4513 /* Add the extent to transaction's private list */
4514 spin_lock(&sbi->s_md_lock);
4515 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4516 spin_unlock(&sbi->s_md_lock);
4517 return 0;
4521 * ext4_free_blocks() -- Free given blocks and update quota
4522 * @handle: handle for this transaction
4523 * @inode: inode
4524 * @block: start physical block to free
4525 * @count: number of blocks to count
4526 * @flags: flags used by ext4_free_blocks
4528 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4529 struct buffer_head *bh, ext4_fsblk_t block,
4530 unsigned long count, int flags)
4532 struct buffer_head *bitmap_bh = NULL;
4533 struct super_block *sb = inode->i_sb;
4534 struct ext4_group_desc *gdp;
4535 unsigned long freed = 0;
4536 unsigned int overflow;
4537 ext4_grpblk_t bit;
4538 struct buffer_head *gd_bh;
4539 ext4_group_t block_group;
4540 struct ext4_sb_info *sbi;
4541 struct ext4_buddy e4b;
4542 unsigned int count_clusters;
4543 int err = 0;
4544 int ret;
4546 if (bh) {
4547 if (block)
4548 BUG_ON(block != bh->b_blocknr);
4549 else
4550 block = bh->b_blocknr;
4553 sbi = EXT4_SB(sb);
4554 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4555 !ext4_data_block_valid(sbi, block, count)) {
4556 ext4_error(sb, "Freeing blocks not in datazone - "
4557 "block = %llu, count = %lu", block, count);
4558 goto error_return;
4561 ext4_debug("freeing block %llu\n", block);
4562 trace_ext4_free_blocks(inode, block, count, flags);
4564 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4565 struct buffer_head *tbh = bh;
4566 int i;
4568 BUG_ON(bh && (count > 1));
4570 for (i = 0; i < count; i++) {
4571 if (!bh)
4572 tbh = sb_find_get_block(inode->i_sb,
4573 block + i);
4574 if (unlikely(!tbh))
4575 continue;
4576 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4577 inode, tbh, block + i);
4582 * We need to make sure we don't reuse the freed block until
4583 * after the transaction is committed, which we can do by
4584 * treating the block as metadata, below. We make an
4585 * exception if the inode is to be written in writeback mode
4586 * since writeback mode has weak data consistency guarantees.
4588 if (!ext4_should_writeback_data(inode))
4589 flags |= EXT4_FREE_BLOCKS_METADATA;
4592 * If the extent to be freed does not begin on a cluster
4593 * boundary, we need to deal with partial clusters at the
4594 * beginning and end of the extent. Normally we will free
4595 * blocks at the beginning or the end unless we are explicitly
4596 * requested to avoid doing so.
4598 overflow = block & (sbi->s_cluster_ratio - 1);
4599 if (overflow) {
4600 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4601 overflow = sbi->s_cluster_ratio - overflow;
4602 block += overflow;
4603 if (count > overflow)
4604 count -= overflow;
4605 else
4606 return;
4607 } else {
4608 block -= overflow;
4609 count += overflow;
4612 overflow = count & (sbi->s_cluster_ratio - 1);
4613 if (overflow) {
4614 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4615 if (count > overflow)
4616 count -= overflow;
4617 else
4618 return;
4619 } else
4620 count += sbi->s_cluster_ratio - overflow;
4623 do_more:
4624 overflow = 0;
4625 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4628 * Check to see if we are freeing blocks across a group
4629 * boundary.
4631 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4632 overflow = EXT4_C2B(sbi, bit) + count -
4633 EXT4_BLOCKS_PER_GROUP(sb);
4634 count -= overflow;
4636 count_clusters = EXT4_B2C(sbi, count);
4637 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4638 if (!bitmap_bh) {
4639 err = -EIO;
4640 goto error_return;
4642 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4643 if (!gdp) {
4644 err = -EIO;
4645 goto error_return;
4648 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4649 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4650 in_range(block, ext4_inode_table(sb, gdp),
4651 EXT4_SB(sb)->s_itb_per_group) ||
4652 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4653 EXT4_SB(sb)->s_itb_per_group)) {
4655 ext4_error(sb, "Freeing blocks in system zone - "
4656 "Block = %llu, count = %lu", block, count);
4657 /* err = 0. ext4_std_error should be a no op */
4658 goto error_return;
4661 BUFFER_TRACE(bitmap_bh, "getting write access");
4662 err = ext4_journal_get_write_access(handle, bitmap_bh);
4663 if (err)
4664 goto error_return;
4667 * We are about to modify some metadata. Call the journal APIs
4668 * to unshare ->b_data if a currently-committing transaction is
4669 * using it
4671 BUFFER_TRACE(gd_bh, "get_write_access");
4672 err = ext4_journal_get_write_access(handle, gd_bh);
4673 if (err)
4674 goto error_return;
4675 #ifdef AGGRESSIVE_CHECK
4677 int i;
4678 for (i = 0; i < count_clusters; i++)
4679 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4681 #endif
4682 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4684 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4685 if (err)
4686 goto error_return;
4688 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4689 struct ext4_free_data *new_entry;
4691 * blocks being freed are metadata. these blocks shouldn't
4692 * be used until this transaction is committed
4694 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4695 if (!new_entry) {
4696 err = -ENOMEM;
4697 goto error_return;
4699 new_entry->start_cluster = bit;
4700 new_entry->group = block_group;
4701 new_entry->count = count_clusters;
4702 new_entry->t_tid = handle->h_transaction->t_tid;
4704 ext4_lock_group(sb, block_group);
4705 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4706 ext4_mb_free_metadata(handle, &e4b, new_entry);
4707 } else {
4708 /* need to update group_info->bb_free and bitmap
4709 * with group lock held. generate_buddy look at
4710 * them with group lock_held
4712 ext4_lock_group(sb, block_group);
4713 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4714 mb_free_blocks(inode, &e4b, bit, count_clusters);
4717 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4718 ext4_free_group_clusters_set(sb, gdp, ret);
4719 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4720 ext4_unlock_group(sb, block_group);
4721 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4723 if (sbi->s_log_groups_per_flex) {
4724 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4725 atomic_add(count_clusters,
4726 &sbi->s_flex_groups[flex_group].free_clusters);
4729 ext4_mb_unload_buddy(&e4b);
4731 freed += count;
4733 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4734 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4736 /* We dirtied the bitmap block */
4737 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4738 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4740 /* And the group descriptor block */
4741 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4742 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4743 if (!err)
4744 err = ret;
4746 if (overflow && !err) {
4747 block += count;
4748 count = overflow;
4749 put_bh(bitmap_bh);
4750 goto do_more;
4752 ext4_mark_super_dirty(sb);
4753 error_return:
4754 brelse(bitmap_bh);
4755 ext4_std_error(sb, err);
4756 return;
4760 * ext4_group_add_blocks() -- Add given blocks to an existing group
4761 * @handle: handle to this transaction
4762 * @sb: super block
4763 * @block: start physcial block to add to the block group
4764 * @count: number of blocks to free
4766 * This marks the blocks as free in the bitmap and buddy.
4768 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4769 ext4_fsblk_t block, unsigned long count)
4771 struct buffer_head *bitmap_bh = NULL;
4772 struct buffer_head *gd_bh;
4773 ext4_group_t block_group;
4774 ext4_grpblk_t bit;
4775 unsigned int i;
4776 struct ext4_group_desc *desc;
4777 struct ext4_sb_info *sbi = EXT4_SB(sb);
4778 struct ext4_buddy e4b;
4779 int err = 0, ret, blk_free_count;
4780 ext4_grpblk_t blocks_freed;
4782 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4784 if (count == 0)
4785 return 0;
4787 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4789 * Check to see if we are freeing blocks across a group
4790 * boundary.
4792 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4793 ext4_warning(sb, "too much blocks added to group %u\n",
4794 block_group);
4795 err = -EINVAL;
4796 goto error_return;
4799 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4800 if (!bitmap_bh) {
4801 err = -EIO;
4802 goto error_return;
4805 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4806 if (!desc) {
4807 err = -EIO;
4808 goto error_return;
4811 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4812 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4813 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4814 in_range(block + count - 1, ext4_inode_table(sb, desc),
4815 sbi->s_itb_per_group)) {
4816 ext4_error(sb, "Adding blocks in system zones - "
4817 "Block = %llu, count = %lu",
4818 block, count);
4819 err = -EINVAL;
4820 goto error_return;
4823 BUFFER_TRACE(bitmap_bh, "getting write access");
4824 err = ext4_journal_get_write_access(handle, bitmap_bh);
4825 if (err)
4826 goto error_return;
4829 * We are about to modify some metadata. Call the journal APIs
4830 * to unshare ->b_data if a currently-committing transaction is
4831 * using it
4833 BUFFER_TRACE(gd_bh, "get_write_access");
4834 err = ext4_journal_get_write_access(handle, gd_bh);
4835 if (err)
4836 goto error_return;
4838 for (i = 0, blocks_freed = 0; i < count; i++) {
4839 BUFFER_TRACE(bitmap_bh, "clear bit");
4840 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4841 ext4_error(sb, "bit already cleared for block %llu",
4842 (ext4_fsblk_t)(block + i));
4843 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4844 } else {
4845 blocks_freed++;
4849 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4850 if (err)
4851 goto error_return;
4854 * need to update group_info->bb_free and bitmap
4855 * with group lock held. generate_buddy look at
4856 * them with group lock_held
4858 ext4_lock_group(sb, block_group);
4859 mb_clear_bits(bitmap_bh->b_data, bit, count);
4860 mb_free_blocks(NULL, &e4b, bit, count);
4861 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4862 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4863 desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
4864 ext4_unlock_group(sb, block_group);
4865 percpu_counter_add(&sbi->s_freeclusters_counter,
4866 EXT4_B2C(sbi, blocks_freed));
4868 if (sbi->s_log_groups_per_flex) {
4869 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4870 atomic_add(EXT4_B2C(sbi, blocks_freed),
4871 &sbi->s_flex_groups[flex_group].free_clusters);
4874 ext4_mb_unload_buddy(&e4b);
4876 /* We dirtied the bitmap block */
4877 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4878 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4880 /* And the group descriptor block */
4881 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4882 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4883 if (!err)
4884 err = ret;
4886 error_return:
4887 brelse(bitmap_bh);
4888 ext4_std_error(sb, err);
4889 return err;
4893 * ext4_trim_extent -- function to TRIM one single free extent in the group
4894 * @sb: super block for the file system
4895 * @start: starting block of the free extent in the alloc. group
4896 * @count: number of blocks to TRIM
4897 * @group: alloc. group we are working with
4898 * @e4b: ext4 buddy for the group
4900 * Trim "count" blocks starting at "start" in the "group". To assure that no
4901 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4902 * be called with under the group lock.
4904 static void ext4_trim_extent(struct super_block *sb, int start, int count,
4905 ext4_group_t group, struct ext4_buddy *e4b)
4907 struct ext4_free_extent ex;
4909 trace_ext4_trim_extent(sb, group, start, count);
4911 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4913 ex.fe_start = start;
4914 ex.fe_group = group;
4915 ex.fe_len = count;
4918 * Mark blocks used, so no one can reuse them while
4919 * being trimmed.
4921 mb_mark_used(e4b, &ex);
4922 ext4_unlock_group(sb, group);
4923 ext4_issue_discard(sb, group, start, count);
4924 ext4_lock_group(sb, group);
4925 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4929 * ext4_trim_all_free -- function to trim all free space in alloc. group
4930 * @sb: super block for file system
4931 * @group: group to be trimmed
4932 * @start: first group block to examine
4933 * @max: last group block to examine
4934 * @minblocks: minimum extent block count
4936 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4937 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4938 * the extent.
4941 * ext4_trim_all_free walks through group's block bitmap searching for free
4942 * extents. When the free extent is found, mark it as used in group buddy
4943 * bitmap. Then issue a TRIM command on this extent and free the extent in
4944 * the group buddy bitmap. This is done until whole group is scanned.
4946 static ext4_grpblk_t
4947 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4948 ext4_grpblk_t start, ext4_grpblk_t max,
4949 ext4_grpblk_t minblocks)
4951 void *bitmap;
4952 ext4_grpblk_t next, count = 0, free_count = 0;
4953 struct ext4_buddy e4b;
4954 int ret;
4956 trace_ext4_trim_all_free(sb, group, start, max);
4958 ret = ext4_mb_load_buddy(sb, group, &e4b);
4959 if (ret) {
4960 ext4_error(sb, "Error in loading buddy "
4961 "information for %u", group);
4962 return ret;
4964 bitmap = e4b.bd_bitmap;
4966 ext4_lock_group(sb, group);
4967 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4968 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4969 goto out;
4971 start = (e4b.bd_info->bb_first_free > start) ?
4972 e4b.bd_info->bb_first_free : start;
4974 while (start < max) {
4975 start = mb_find_next_zero_bit(bitmap, max, start);
4976 if (start >= max)
4977 break;
4978 next = mb_find_next_bit(bitmap, max, start);
4980 if ((next - start) >= minblocks) {
4981 ext4_trim_extent(sb, start,
4982 next - start, group, &e4b);
4983 count += next - start;
4985 free_count += next - start;
4986 start = next + 1;
4988 if (fatal_signal_pending(current)) {
4989 count = -ERESTARTSYS;
4990 break;
4993 if (need_resched()) {
4994 ext4_unlock_group(sb, group);
4995 cond_resched();
4996 ext4_lock_group(sb, group);
4999 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5000 break;
5003 if (!ret)
5004 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5005 out:
5006 ext4_unlock_group(sb, group);
5007 ext4_mb_unload_buddy(&e4b);
5009 ext4_debug("trimmed %d blocks in the group %d\n",
5010 count, group);
5012 return count;
5016 * ext4_trim_fs() -- trim ioctl handle function
5017 * @sb: superblock for filesystem
5018 * @range: fstrim_range structure
5020 * start: First Byte to trim
5021 * len: number of Bytes to trim from start
5022 * minlen: minimum extent length in Bytes
5023 * ext4_trim_fs goes through all allocation groups containing Bytes from
5024 * start to start+len. For each such a group ext4_trim_all_free function
5025 * is invoked to trim all free space.
5027 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5029 struct ext4_group_info *grp;
5030 ext4_group_t first_group, last_group;
5031 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
5032 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5033 uint64_t start, len, minlen, trimmed = 0;
5034 ext4_fsblk_t first_data_blk =
5035 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5036 int ret = 0;
5038 start = range->start >> sb->s_blocksize_bits;
5039 len = range->len >> sb->s_blocksize_bits;
5040 minlen = range->minlen >> sb->s_blocksize_bits;
5042 if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)))
5043 return -EINVAL;
5044 if (start + len <= first_data_blk)
5045 goto out;
5046 if (start < first_data_blk) {
5047 len -= first_data_blk - start;
5048 start = first_data_blk;
5051 /* Determine first and last group to examine based on start and len */
5052 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5053 &first_group, &first_cluster);
5054 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
5055 &last_group, &last_cluster);
5056 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
5057 last_cluster = EXT4_CLUSTERS_PER_GROUP(sb);
5059 if (first_group > last_group)
5060 return -EINVAL;
5062 for (group = first_group; group <= last_group; group++) {
5063 grp = ext4_get_group_info(sb, group);
5064 /* We only do this if the grp has never been initialized */
5065 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5066 ret = ext4_mb_init_group(sb, group);
5067 if (ret)
5068 break;
5072 * For all the groups except the last one, last block will
5073 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
5074 * change it for the last group in which case start +
5075 * len < EXT4_BLOCKS_PER_GROUP(sb).
5077 if (first_cluster + len < EXT4_CLUSTERS_PER_GROUP(sb))
5078 last_cluster = first_cluster + len;
5079 len -= last_cluster - first_cluster;
5081 if (grp->bb_free >= minlen) {
5082 cnt = ext4_trim_all_free(sb, group, first_cluster,
5083 last_cluster, minlen);
5084 if (cnt < 0) {
5085 ret = cnt;
5086 break;
5089 trimmed += cnt;
5090 first_cluster = 0;
5092 range->len = trimmed * sb->s_blocksize;
5094 if (!ret)
5095 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5097 out:
5098 return ret;