x86/mm/pat: Don't report PAT on CPUs that don't support it
[linux/fpc-iii.git] / fs / ext4 / mballoc.c
blob354dc1a894c29bb86dcafcd1b2a5d8a53ca7119d
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 License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
25 #include "mballoc.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/backing-dev.h>
30 #include <trace/events/ext4.h>
32 #ifdef CONFIG_EXT4_DEBUG
33 ushort ext4_mballoc_debug __read_mostly;
35 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
36 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
37 #endif
40 * MUSTDO:
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
44 * TODO v4:
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
48 * - quota
49 * - reservation for superuser
51 * TODO v3:
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
56 * - error handling
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
63 * During initialization phase of the allocator we decide to use the
64 * group preallocation or inode preallocation depending on the size of
65 * the file. The size of the file could be the resulting file size we
66 * would have after allocation, or the current file size, which ever
67 * is larger. If the size is less than sbi->s_mb_stream_request we
68 * select to use the group preallocation. The default value of
69 * s_mb_stream_request is 16 blocks. This can also be tuned via
70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
71 * terms of number of blocks.
73 * The main motivation for having small file use group preallocation is to
74 * ensure that we have small files closer together on the disk.
76 * First stage the allocator looks at the inode prealloc list,
77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
78 * spaces for this particular inode. The inode prealloc space is
79 * represented as:
81 * pa_lstart -> the logical start block for this prealloc space
82 * pa_pstart -> the physical start block for this prealloc space
83 * pa_len -> length for this prealloc space (in clusters)
84 * pa_free -> free space available in this prealloc space (in clusters)
86 * The inode preallocation space is used looking at the _logical_ start
87 * block. If only the logical file block falls within the range of prealloc
88 * space we will consume the particular prealloc space. This makes sure that
89 * we have contiguous physical blocks representing the file blocks
91 * The important thing to be noted in case of inode prealloc space is that
92 * we don't modify the values associated to inode prealloc space except
93 * pa_free.
95 * If we are not able to find blocks in the inode prealloc space and if we
96 * have the group allocation flag set then we look at the locality group
97 * prealloc space. These are per CPU prealloc list represented as
99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
101 * The reason for having a per cpu locality group is to reduce the contention
102 * between CPUs. It is possible to get scheduled at this point.
104 * The locality group prealloc space is used looking at whether we have
105 * enough free space (pa_free) within the prealloc space.
107 * If we can't allocate blocks via inode prealloc or/and locality group
108 * prealloc then we look at the buddy cache. The buddy cache is represented
109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
110 * mapped to the buddy and bitmap information regarding different
111 * groups. The buddy information is attached to buddy cache inode so that
112 * we can access them through the page cache. The information regarding
113 * each group is loaded via ext4_mb_load_buddy. The information involve
114 * block bitmap and buddy information. The information are stored in the
115 * inode as:
117 * { page }
118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
121 * one block each for bitmap and buddy information. So for each group we
122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
141 * 512 blocks. This can be tuned via
142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
148 * The regular allocator (using the buddy cache) supports a few tunables.
150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
154 * The regular allocator uses buddy scan only if the request len is power of
155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
162 * min_to_scan indicate how long the mballoc __must__ look for a best
163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
167 * can be used for allocation. ext4_mb_good_group explains how the groups are
168 * checked.
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
177 * mballoc operates on the following data:
178 * - on-disk bitmap
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
182 * there are two types of preallocations:
183 * - inode
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
192 * - locality group
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
233 * block
235 * so, now we're building a concurrency table:
236 * - init buddy vs.
237 * - new PA
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
240 * - use inode PA
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
243 * - discard inode PA
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
249 * - new PA vs.
250 * - use inode PA
251 * i_data_sem serializes them
252 * - discard inode PA
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
258 * - use inode PA
259 * - use inode PA
260 * i_data_sem or another mutex should serializes them
261 * - discard inode PA
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
282 * Logic in few words:
284 * - allocation:
285 * load group
286 * find blocks
287 * mark bits in on-disk bitmap
288 * release group
290 * - use preallocation:
291 * find proper PA (per-inode or group)
292 * load group
293 * mark bits in on-disk bitmap
294 * release group
295 * release PA
297 * - free:
298 * load group
299 * mark bits in on-disk bitmap
300 * release group
302 * - discard preallocations in group:
303 * mark PAs deleted
304 * move them onto local list
305 * load on-disk bitmap
306 * load group
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
310 * - discard inode's preallocations:
314 * Locking rules
316 * Locks:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
319 * - per-pa lock (pa)
321 * Paths:
322 * - new pa
323 * object
324 * group
326 * - find and use pa:
327 * pa
329 * - release consumed pa:
330 * pa
331 * group
332 * object
334 * - generate in-core bitmap:
335 * group
336 * pa
338 * - discard all for given object (inode, locality group):
339 * object
340 * pa
341 * group
343 * - discard all for given group:
344 * group
345 * pa
346 * group
347 * object
350 static struct kmem_cache *ext4_pspace_cachep;
351 static struct kmem_cache *ext4_ac_cachep;
352 static struct kmem_cache *ext4_free_data_cachep;
354 /* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
360 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
366 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367 ext4_group_t group);
368 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369 ext4_group_t group);
370 static void ext4_free_data_callback(struct super_block *sb,
371 struct ext4_journal_cb_entry *jce, int rc);
373 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
375 #if BITS_PER_LONG == 64
376 *bit += ((unsigned long) addr & 7UL) << 3;
377 addr = (void *) ((unsigned long) addr & ~7UL);
378 #elif BITS_PER_LONG == 32
379 *bit += ((unsigned long) addr & 3UL) << 3;
380 addr = (void *) ((unsigned long) addr & ~3UL);
381 #else
382 #error "how many bits you are?!"
383 #endif
384 return addr;
387 static inline int mb_test_bit(int bit, void *addr)
390 * ext4_test_bit on architecture like powerpc
391 * needs unsigned long aligned address
393 addr = mb_correct_addr_and_bit(&bit, addr);
394 return ext4_test_bit(bit, addr);
397 static inline void mb_set_bit(int bit, void *addr)
399 addr = mb_correct_addr_and_bit(&bit, addr);
400 ext4_set_bit(bit, addr);
403 static inline void mb_clear_bit(int bit, void *addr)
405 addr = mb_correct_addr_and_bit(&bit, addr);
406 ext4_clear_bit(bit, addr);
409 static inline int mb_test_and_clear_bit(int bit, void *addr)
411 addr = mb_correct_addr_and_bit(&bit, addr);
412 return ext4_test_and_clear_bit(bit, addr);
415 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
417 int fix = 0, ret, tmpmax;
418 addr = mb_correct_addr_and_bit(&fix, addr);
419 tmpmax = max + fix;
420 start += fix;
422 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
423 if (ret > max)
424 return max;
425 return ret;
428 static inline int mb_find_next_bit(void *addr, int max, int start)
430 int fix = 0, ret, tmpmax;
431 addr = mb_correct_addr_and_bit(&fix, addr);
432 tmpmax = max + fix;
433 start += fix;
435 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
436 if (ret > max)
437 return max;
438 return ret;
441 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
443 char *bb;
445 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
446 BUG_ON(max == NULL);
448 if (order > e4b->bd_blkbits + 1) {
449 *max = 0;
450 return NULL;
453 /* at order 0 we see each particular block */
454 if (order == 0) {
455 *max = 1 << (e4b->bd_blkbits + 3);
456 return e4b->bd_bitmap;
459 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
460 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
462 return bb;
465 #ifdef DOUBLE_CHECK
466 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
467 int first, int count)
469 int i;
470 struct super_block *sb = e4b->bd_sb;
472 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 return;
474 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
475 for (i = 0; i < count; i++) {
476 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
477 ext4_fsblk_t blocknr;
479 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
480 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
481 ext4_grp_locked_error(sb, e4b->bd_group,
482 inode ? inode->i_ino : 0,
483 blocknr,
484 "freeing block already freed "
485 "(bit %u)",
486 first + i);
488 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
492 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
494 int i;
496 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
497 return;
498 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
499 for (i = 0; i < count; i++) {
500 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
501 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
505 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
508 unsigned char *b1, *b2;
509 int i;
510 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
511 b2 = (unsigned char *) bitmap;
512 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
513 if (b1[i] != b2[i]) {
514 ext4_msg(e4b->bd_sb, KERN_ERR,
515 "corruption in group %u "
516 "at byte %u(%u): %x in copy != %x "
517 "on disk/prealloc",
518 e4b->bd_group, i, i * 8, b1[i], b2[i]);
519 BUG();
525 #else
526 static inline void mb_free_blocks_double(struct inode *inode,
527 struct ext4_buddy *e4b, int first, int count)
529 return;
531 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
532 int first, int count)
534 return;
536 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
538 return;
540 #endif
542 #ifdef AGGRESSIVE_CHECK
544 #define MB_CHECK_ASSERT(assert) \
545 do { \
546 if (!(assert)) { \
547 printk(KERN_EMERG \
548 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
549 function, file, line, # assert); \
550 BUG(); \
552 } while (0)
554 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
555 const char *function, int line)
557 struct super_block *sb = e4b->bd_sb;
558 int order = e4b->bd_blkbits + 1;
559 int max;
560 int max2;
561 int i;
562 int j;
563 int k;
564 int count;
565 struct ext4_group_info *grp;
566 int fragments = 0;
567 int fstart;
568 struct list_head *cur;
569 void *buddy;
570 void *buddy2;
573 static int mb_check_counter;
574 if (mb_check_counter++ % 100 != 0)
575 return 0;
578 while (order > 1) {
579 buddy = mb_find_buddy(e4b, order, &max);
580 MB_CHECK_ASSERT(buddy);
581 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
582 MB_CHECK_ASSERT(buddy2);
583 MB_CHECK_ASSERT(buddy != buddy2);
584 MB_CHECK_ASSERT(max * 2 == max2);
586 count = 0;
587 for (i = 0; i < max; i++) {
589 if (mb_test_bit(i, buddy)) {
590 /* only single bit in buddy2 may be 1 */
591 if (!mb_test_bit(i << 1, buddy2)) {
592 MB_CHECK_ASSERT(
593 mb_test_bit((i<<1)+1, buddy2));
594 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
595 MB_CHECK_ASSERT(
596 mb_test_bit(i << 1, buddy2));
598 continue;
601 /* both bits in buddy2 must be 1 */
602 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
603 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
605 for (j = 0; j < (1 << order); j++) {
606 k = (i * (1 << order)) + j;
607 MB_CHECK_ASSERT(
608 !mb_test_bit(k, e4b->bd_bitmap));
610 count++;
612 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
613 order--;
616 fstart = -1;
617 buddy = mb_find_buddy(e4b, 0, &max);
618 for (i = 0; i < max; i++) {
619 if (!mb_test_bit(i, buddy)) {
620 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
621 if (fstart == -1) {
622 fragments++;
623 fstart = i;
625 continue;
627 fstart = -1;
628 /* check used bits only */
629 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
630 buddy2 = mb_find_buddy(e4b, j, &max2);
631 k = i >> j;
632 MB_CHECK_ASSERT(k < max2);
633 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
636 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
637 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
639 grp = ext4_get_group_info(sb, e4b->bd_group);
640 list_for_each(cur, &grp->bb_prealloc_list) {
641 ext4_group_t groupnr;
642 struct ext4_prealloc_space *pa;
643 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
644 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
645 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
646 for (i = 0; i < pa->pa_len; i++)
647 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
649 return 0;
651 #undef MB_CHECK_ASSERT
652 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
653 __FILE__, __func__, __LINE__)
654 #else
655 #define mb_check_buddy(e4b)
656 #endif
659 * Divide blocks started from @first with length @len into
660 * smaller chunks with power of 2 blocks.
661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
662 * then increase bb_counters[] for corresponded chunk size.
664 static void ext4_mb_mark_free_simple(struct super_block *sb,
665 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
666 struct ext4_group_info *grp)
668 struct ext4_sb_info *sbi = EXT4_SB(sb);
669 ext4_grpblk_t min;
670 ext4_grpblk_t max;
671 ext4_grpblk_t chunk;
672 unsigned int border;
674 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
676 border = 2 << sb->s_blocksize_bits;
678 while (len > 0) {
679 /* find how many blocks can be covered since this position */
680 max = ffs(first | border) - 1;
682 /* find how many blocks of power 2 we need to mark */
683 min = fls(len) - 1;
685 if (max < min)
686 min = max;
687 chunk = 1 << min;
689 /* mark multiblock chunks only */
690 grp->bb_counters[min]++;
691 if (min > 0)
692 mb_clear_bit(first >> min,
693 buddy + sbi->s_mb_offsets[min]);
695 len -= chunk;
696 first += chunk;
701 * Cache the order of the largest free extent we have available in this block
702 * group.
704 static void
705 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
707 int i;
708 int bits;
710 grp->bb_largest_free_order = -1; /* uninit */
712 bits = sb->s_blocksize_bits + 1;
713 for (i = bits; i >= 0; i--) {
714 if (grp->bb_counters[i] > 0) {
715 grp->bb_largest_free_order = i;
716 break;
721 static noinline_for_stack
722 void ext4_mb_generate_buddy(struct super_block *sb,
723 void *buddy, void *bitmap, ext4_group_t group)
725 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
726 struct ext4_sb_info *sbi = EXT4_SB(sb);
727 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
728 ext4_grpblk_t i = 0;
729 ext4_grpblk_t first;
730 ext4_grpblk_t len;
731 unsigned free = 0;
732 unsigned fragments = 0;
733 unsigned long long period = get_cycles();
735 /* initialize buddy from bitmap which is aggregation
736 * of on-disk bitmap and preallocations */
737 i = mb_find_next_zero_bit(bitmap, max, 0);
738 grp->bb_first_free = i;
739 while (i < max) {
740 fragments++;
741 first = i;
742 i = mb_find_next_bit(bitmap, max, i);
743 len = i - first;
744 free += len;
745 if (len > 1)
746 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
747 else
748 grp->bb_counters[0]++;
749 if (i < max)
750 i = mb_find_next_zero_bit(bitmap, max, i);
752 grp->bb_fragments = fragments;
754 if (free != grp->bb_free) {
755 ext4_grp_locked_error(sb, group, 0, 0,
756 "block bitmap and bg descriptor "
757 "inconsistent: %u vs %u free clusters",
758 free, grp->bb_free);
760 * If we intend to continue, we consider group descriptor
761 * corrupt and update bb_free using bitmap value
763 grp->bb_free = free;
764 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
765 percpu_counter_sub(&sbi->s_freeclusters_counter,
766 grp->bb_free);
767 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
769 mb_set_largest_free_order(sb, grp);
771 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
773 period = get_cycles() - period;
774 spin_lock(&EXT4_SB(sb)->s_bal_lock);
775 EXT4_SB(sb)->s_mb_buddies_generated++;
776 EXT4_SB(sb)->s_mb_generation_time += period;
777 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
780 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
782 int count;
783 int order = 1;
784 void *buddy;
786 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
787 ext4_set_bits(buddy, 0, count);
789 e4b->bd_info->bb_fragments = 0;
790 memset(e4b->bd_info->bb_counters, 0,
791 sizeof(*e4b->bd_info->bb_counters) *
792 (e4b->bd_sb->s_blocksize_bits + 2));
794 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
795 e4b->bd_bitmap, e4b->bd_group);
798 /* The buddy information is attached the buddy cache inode
799 * for convenience. The information regarding each group
800 * is loaded via ext4_mb_load_buddy. The information involve
801 * block bitmap and buddy information. The information are
802 * stored in the inode as
804 * { page }
805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
808 * one block each for bitmap and buddy information.
809 * So for each group we take up 2 blocks. A page can
810 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
811 * So it can have information regarding groups_per_page which
812 * is blocks_per_page/2
814 * Locking note: This routine takes the block group lock of all groups
815 * for this page; do not hold this lock when calling this routine!
818 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
820 ext4_group_t ngroups;
821 int blocksize;
822 int blocks_per_page;
823 int groups_per_page;
824 int err = 0;
825 int i;
826 ext4_group_t first_group, group;
827 int first_block;
828 struct super_block *sb;
829 struct buffer_head *bhs;
830 struct buffer_head **bh = NULL;
831 struct inode *inode;
832 char *data;
833 char *bitmap;
834 struct ext4_group_info *grinfo;
836 mb_debug(1, "init page %lu\n", page->index);
838 inode = page->mapping->host;
839 sb = inode->i_sb;
840 ngroups = ext4_get_groups_count(sb);
841 blocksize = i_blocksize(inode);
842 blocks_per_page = PAGE_SIZE / blocksize;
844 groups_per_page = blocks_per_page >> 1;
845 if (groups_per_page == 0)
846 groups_per_page = 1;
848 /* allocate buffer_heads to read bitmaps */
849 if (groups_per_page > 1) {
850 i = sizeof(struct buffer_head *) * groups_per_page;
851 bh = kzalloc(i, gfp);
852 if (bh == NULL) {
853 err = -ENOMEM;
854 goto out;
856 } else
857 bh = &bhs;
859 first_group = page->index * blocks_per_page / 2;
861 /* read all groups the page covers into the cache */
862 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
863 if (group >= ngroups)
864 break;
866 grinfo = ext4_get_group_info(sb, group);
868 * If page is uptodate then we came here after online resize
869 * which added some new uninitialized group info structs, so
870 * we must skip all initialized uptodate buddies on the page,
871 * which may be currently in use by an allocating task.
873 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
874 bh[i] = NULL;
875 continue;
877 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
878 if (IS_ERR(bh[i])) {
879 err = PTR_ERR(bh[i]);
880 bh[i] = NULL;
881 goto out;
883 mb_debug(1, "read bitmap for group %u\n", group);
886 /* wait for I/O completion */
887 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
888 int err2;
890 if (!bh[i])
891 continue;
892 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
893 if (!err)
894 err = err2;
897 first_block = page->index * blocks_per_page;
898 for (i = 0; i < blocks_per_page; i++) {
899 group = (first_block + i) >> 1;
900 if (group >= ngroups)
901 break;
903 if (!bh[group - first_group])
904 /* skip initialized uptodate buddy */
905 continue;
907 if (!buffer_verified(bh[group - first_group]))
908 /* Skip faulty bitmaps */
909 continue;
910 err = 0;
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, gfp_t gfp)
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_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);
1006 if (!page)
1007 return -ENOMEM;
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 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1020 if (!page)
1021 return -ENOMEM;
1022 BUG_ON(page->mapping != inode->i_mapping);
1023 e4b->bd_buddy_page = page;
1024 return 0;
1027 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1029 if (e4b->bd_bitmap_page) {
1030 unlock_page(e4b->bd_bitmap_page);
1031 put_page(e4b->bd_bitmap_page);
1033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 put_page(e4b->bd_buddy_page);
1040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1044 static noinline_for_stack
1045 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1048 struct ext4_group_info *this_grp;
1049 struct ext4_buddy e4b;
1050 struct page *page;
1051 int ret = 0;
1053 might_sleep();
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.
1062 * The call to ext4_mb_get_buddy_page_lock will mark the
1063 * page accessed.
1065 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1066 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1068 * somebody initialized the group
1069 * return without doing anything
1071 goto err;
1074 page = e4b.bd_bitmap_page;
1075 ret = ext4_mb_init_cache(page, NULL, gfp);
1076 if (ret)
1077 goto err;
1078 if (!PageUptodate(page)) {
1079 ret = -EIO;
1080 goto err;
1083 if (e4b.bd_buddy_page == NULL) {
1085 * If both the bitmap and buddy are in
1086 * the same page we don't need to force
1087 * init the buddy
1089 ret = 0;
1090 goto err;
1092 /* init buddy cache */
1093 page = e4b.bd_buddy_page;
1094 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1095 if (ret)
1096 goto err;
1097 if (!PageUptodate(page)) {
1098 ret = -EIO;
1099 goto err;
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_gfp(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b, gfp_t gfp)
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 might_sleep();
1126 mb_debug(1, "load group %u\n", group);
1128 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1129 grp = ext4_get_group_info(sb, group);
1131 e4b->bd_blkbits = sb->s_blocksize_bits;
1132 e4b->bd_info = grp;
1133 e4b->bd_sb = sb;
1134 e4b->bd_group = group;
1135 e4b->bd_buddy_page = NULL;
1136 e4b->bd_bitmap_page = NULL;
1138 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1140 * we need full data about the group
1141 * to make a good selection
1143 ret = ext4_mb_init_group(sb, group, gfp);
1144 if (ret)
1145 return ret;
1149 * the buddy cache inode stores the block bitmap
1150 * and buddy information in consecutive blocks.
1151 * So for each group we need two blocks.
1153 block = group * 2;
1154 pnum = block / blocks_per_page;
1155 poff = block % blocks_per_page;
1157 /* we could use find_or_create_page(), but it locks page
1158 * what we'd like to avoid in fast path ... */
1159 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1160 if (page == NULL || !PageUptodate(page)) {
1161 if (page)
1163 * drop the page reference and try
1164 * to get the page with lock. If we
1165 * are not uptodate that implies
1166 * somebody just created the page but
1167 * is yet to initialize the same. So
1168 * wait for it to initialize.
1170 put_page(page);
1171 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1172 if (page) {
1173 BUG_ON(page->mapping != inode->i_mapping);
1174 if (!PageUptodate(page)) {
1175 ret = ext4_mb_init_cache(page, NULL, gfp);
1176 if (ret) {
1177 unlock_page(page);
1178 goto err;
1180 mb_cmp_bitmaps(e4b, page_address(page) +
1181 (poff * sb->s_blocksize));
1183 unlock_page(page);
1186 if (page == NULL) {
1187 ret = -ENOMEM;
1188 goto err;
1190 if (!PageUptodate(page)) {
1191 ret = -EIO;
1192 goto err;
1195 /* Pages marked accessed already */
1196 e4b->bd_bitmap_page = page;
1197 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1199 block++;
1200 pnum = block / blocks_per_page;
1201 poff = block % blocks_per_page;
1203 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1204 if (page == NULL || !PageUptodate(page)) {
1205 if (page)
1206 put_page(page);
1207 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1208 if (page) {
1209 BUG_ON(page->mapping != inode->i_mapping);
1210 if (!PageUptodate(page)) {
1211 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1212 gfp);
1213 if (ret) {
1214 unlock_page(page);
1215 goto err;
1218 unlock_page(page);
1221 if (page == NULL) {
1222 ret = -ENOMEM;
1223 goto err;
1225 if (!PageUptodate(page)) {
1226 ret = -EIO;
1227 goto err;
1230 /* Pages marked accessed already */
1231 e4b->bd_buddy_page = page;
1232 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1234 BUG_ON(e4b->bd_bitmap_page == NULL);
1235 BUG_ON(e4b->bd_buddy_page == NULL);
1237 return 0;
1239 err:
1240 if (page)
1241 put_page(page);
1242 if (e4b->bd_bitmap_page)
1243 put_page(e4b->bd_bitmap_page);
1244 if (e4b->bd_buddy_page)
1245 put_page(e4b->bd_buddy_page);
1246 e4b->bd_buddy = NULL;
1247 e4b->bd_bitmap = NULL;
1248 return ret;
1251 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1252 struct ext4_buddy *e4b)
1254 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1257 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1259 if (e4b->bd_bitmap_page)
1260 put_page(e4b->bd_bitmap_page);
1261 if (e4b->bd_buddy_page)
1262 put_page(e4b->bd_buddy_page);
1266 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1268 int order = 1;
1269 int bb_incr = 1 << (e4b->bd_blkbits - 1);
1270 void *bb;
1272 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1273 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1275 bb = e4b->bd_buddy;
1276 while (order <= e4b->bd_blkbits + 1) {
1277 block = block >> 1;
1278 if (!mb_test_bit(block, bb)) {
1279 /* this block is part of buddy of order 'order' */
1280 return order;
1282 bb += bb_incr;
1283 bb_incr >>= 1;
1284 order++;
1286 return 0;
1289 static void mb_clear_bits(void *bm, int cur, int len)
1291 __u32 *addr;
1293 len = cur + len;
1294 while (cur < len) {
1295 if ((cur & 31) == 0 && (len - cur) >= 32) {
1296 /* fast path: clear whole word at once */
1297 addr = bm + (cur >> 3);
1298 *addr = 0;
1299 cur += 32;
1300 continue;
1302 mb_clear_bit(cur, bm);
1303 cur++;
1307 /* clear bits in given range
1308 * will return first found zero bit if any, -1 otherwise
1310 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1312 __u32 *addr;
1313 int zero_bit = -1;
1315 len = cur + len;
1316 while (cur < len) {
1317 if ((cur & 31) == 0 && (len - cur) >= 32) {
1318 /* fast path: clear whole word at once */
1319 addr = bm + (cur >> 3);
1320 if (*addr != (__u32)(-1) && zero_bit == -1)
1321 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1322 *addr = 0;
1323 cur += 32;
1324 continue;
1326 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1327 zero_bit = cur;
1328 cur++;
1331 return zero_bit;
1334 void ext4_set_bits(void *bm, int cur, int len)
1336 __u32 *addr;
1338 len = cur + len;
1339 while (cur < len) {
1340 if ((cur & 31) == 0 && (len - cur) >= 32) {
1341 /* fast path: set whole word at once */
1342 addr = bm + (cur >> 3);
1343 *addr = 0xffffffff;
1344 cur += 32;
1345 continue;
1347 mb_set_bit(cur, bm);
1348 cur++;
1353 * _________________________________________________________________ */
1355 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1357 if (mb_test_bit(*bit + side, bitmap)) {
1358 mb_clear_bit(*bit, bitmap);
1359 (*bit) -= side;
1360 return 1;
1362 else {
1363 (*bit) += side;
1364 mb_set_bit(*bit, bitmap);
1365 return -1;
1369 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1371 int max;
1372 int order = 1;
1373 void *buddy = mb_find_buddy(e4b, order, &max);
1375 while (buddy) {
1376 void *buddy2;
1378 /* Bits in range [first; last] are known to be set since
1379 * corresponding blocks were allocated. Bits in range
1380 * (first; last) will stay set because they form buddies on
1381 * upper layer. We just deal with borders if they don't
1382 * align with upper layer and then go up.
1383 * Releasing entire group is all about clearing
1384 * single bit of highest order buddy.
1387 /* Example:
1388 * ---------------------------------
1389 * | 1 | 1 | 1 | 1 |
1390 * ---------------------------------
1391 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1392 * ---------------------------------
1393 * 0 1 2 3 4 5 6 7
1394 * \_____________________/
1396 * Neither [1] nor [6] is aligned to above layer.
1397 * Left neighbour [0] is free, so mark it busy,
1398 * decrease bb_counters and extend range to
1399 * [0; 6]
1400 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1401 * mark [6] free, increase bb_counters and shrink range to
1402 * [0; 5].
1403 * Then shift range to [0; 2], go up and do the same.
1407 if (first & 1)
1408 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1409 if (!(last & 1))
1410 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1411 if (first > last)
1412 break;
1413 order++;
1415 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1416 mb_clear_bits(buddy, first, last - first + 1);
1417 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1418 break;
1420 first >>= 1;
1421 last >>= 1;
1422 buddy = buddy2;
1426 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1427 int first, int count)
1429 int left_is_free = 0;
1430 int right_is_free = 0;
1431 int block;
1432 int last = first + count - 1;
1433 struct super_block *sb = e4b->bd_sb;
1435 if (WARN_ON(count == 0))
1436 return;
1437 BUG_ON(last >= (sb->s_blocksize << 3));
1438 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1439 /* Don't bother if the block group is corrupt. */
1440 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1441 return;
1443 mb_check_buddy(e4b);
1444 mb_free_blocks_double(inode, e4b, first, count);
1446 e4b->bd_info->bb_free += count;
1447 if (first < e4b->bd_info->bb_first_free)
1448 e4b->bd_info->bb_first_free = first;
1450 /* access memory sequentially: check left neighbour,
1451 * clear range and then check right neighbour
1453 if (first != 0)
1454 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1455 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1456 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1457 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1459 if (unlikely(block != -1)) {
1460 struct ext4_sb_info *sbi = EXT4_SB(sb);
1461 ext4_fsblk_t blocknr;
1463 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1464 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1465 ext4_grp_locked_error(sb, e4b->bd_group,
1466 inode ? inode->i_ino : 0,
1467 blocknr,
1468 "freeing already freed block "
1469 "(bit %u); block bitmap corrupt.",
1470 block);
1471 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1472 percpu_counter_sub(&sbi->s_freeclusters_counter,
1473 e4b->bd_info->bb_free);
1474 /* Mark the block group as corrupt. */
1475 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1476 &e4b->bd_info->bb_state);
1477 mb_regenerate_buddy(e4b);
1478 goto done;
1481 /* let's maintain fragments counter */
1482 if (left_is_free && right_is_free)
1483 e4b->bd_info->bb_fragments--;
1484 else if (!left_is_free && !right_is_free)
1485 e4b->bd_info->bb_fragments++;
1487 /* buddy[0] == bd_bitmap is a special case, so handle
1488 * it right away and let mb_buddy_mark_free stay free of
1489 * zero order checks.
1490 * Check if neighbours are to be coaleasced,
1491 * adjust bitmap bb_counters and borders appropriately.
1493 if (first & 1) {
1494 first += !left_is_free;
1495 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1497 if (!(last & 1)) {
1498 last -= !right_is_free;
1499 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1502 if (first <= last)
1503 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1505 done:
1506 mb_set_largest_free_order(sb, e4b->bd_info);
1507 mb_check_buddy(e4b);
1510 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1511 int needed, struct ext4_free_extent *ex)
1513 int next = block;
1514 int max, order;
1515 void *buddy;
1517 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1518 BUG_ON(ex == NULL);
1520 buddy = mb_find_buddy(e4b, 0, &max);
1521 BUG_ON(buddy == NULL);
1522 BUG_ON(block >= max);
1523 if (mb_test_bit(block, buddy)) {
1524 ex->fe_len = 0;
1525 ex->fe_start = 0;
1526 ex->fe_group = 0;
1527 return 0;
1530 /* find actual order */
1531 order = mb_find_order_for_block(e4b, block);
1532 block = block >> order;
1534 ex->fe_len = 1 << order;
1535 ex->fe_start = block << order;
1536 ex->fe_group = e4b->bd_group;
1538 /* calc difference from given start */
1539 next = next - ex->fe_start;
1540 ex->fe_len -= next;
1541 ex->fe_start += next;
1543 while (needed > ex->fe_len &&
1544 mb_find_buddy(e4b, order, &max)) {
1546 if (block + 1 >= max)
1547 break;
1549 next = (block + 1) * (1 << order);
1550 if (mb_test_bit(next, e4b->bd_bitmap))
1551 break;
1553 order = mb_find_order_for_block(e4b, next);
1555 block = next >> order;
1556 ex->fe_len += 1 << order;
1559 if (ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))) {
1560 /* Should never happen! (but apparently sometimes does?!?) */
1561 WARN_ON(1);
1562 ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1563 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1564 block, order, needed, ex->fe_group, ex->fe_start,
1565 ex->fe_len, ex->fe_logical);
1566 ex->fe_len = 0;
1567 ex->fe_start = 0;
1568 ex->fe_group = 0;
1570 return ex->fe_len;
1573 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1575 int ord;
1576 int mlen = 0;
1577 int max = 0;
1578 int cur;
1579 int start = ex->fe_start;
1580 int len = ex->fe_len;
1581 unsigned ret = 0;
1582 int len0 = len;
1583 void *buddy;
1585 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1586 BUG_ON(e4b->bd_group != ex->fe_group);
1587 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1588 mb_check_buddy(e4b);
1589 mb_mark_used_double(e4b, start, len);
1591 e4b->bd_info->bb_free -= len;
1592 if (e4b->bd_info->bb_first_free == start)
1593 e4b->bd_info->bb_first_free += len;
1595 /* let's maintain fragments counter */
1596 if (start != 0)
1597 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1598 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1599 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1600 if (mlen && max)
1601 e4b->bd_info->bb_fragments++;
1602 else if (!mlen && !max)
1603 e4b->bd_info->bb_fragments--;
1605 /* let's maintain buddy itself */
1606 while (len) {
1607 ord = mb_find_order_for_block(e4b, start);
1609 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1610 /* the whole chunk may be allocated at once! */
1611 mlen = 1 << ord;
1612 buddy = mb_find_buddy(e4b, ord, &max);
1613 BUG_ON((start >> ord) >= max);
1614 mb_set_bit(start >> ord, buddy);
1615 e4b->bd_info->bb_counters[ord]--;
1616 start += mlen;
1617 len -= mlen;
1618 BUG_ON(len < 0);
1619 continue;
1622 /* store for history */
1623 if (ret == 0)
1624 ret = len | (ord << 16);
1626 /* we have to split large buddy */
1627 BUG_ON(ord <= 0);
1628 buddy = mb_find_buddy(e4b, ord, &max);
1629 mb_set_bit(start >> ord, buddy);
1630 e4b->bd_info->bb_counters[ord]--;
1632 ord--;
1633 cur = (start >> ord) & ~1U;
1634 buddy = mb_find_buddy(e4b, ord, &max);
1635 mb_clear_bit(cur, buddy);
1636 mb_clear_bit(cur + 1, buddy);
1637 e4b->bd_info->bb_counters[ord]++;
1638 e4b->bd_info->bb_counters[ord]++;
1640 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1642 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1643 mb_check_buddy(e4b);
1645 return ret;
1649 * Must be called under group lock!
1651 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1652 struct ext4_buddy *e4b)
1654 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1655 int ret;
1657 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1658 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1660 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1661 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1662 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1664 /* preallocation can change ac_b_ex, thus we store actually
1665 * allocated blocks for history */
1666 ac->ac_f_ex = ac->ac_b_ex;
1668 ac->ac_status = AC_STATUS_FOUND;
1669 ac->ac_tail = ret & 0xffff;
1670 ac->ac_buddy = ret >> 16;
1673 * take the page reference. We want the page to be pinned
1674 * so that we don't get a ext4_mb_init_cache_call for this
1675 * group until we update the bitmap. That would mean we
1676 * double allocate blocks. The reference is dropped
1677 * in ext4_mb_release_context
1679 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1680 get_page(ac->ac_bitmap_page);
1681 ac->ac_buddy_page = e4b->bd_buddy_page;
1682 get_page(ac->ac_buddy_page);
1683 /* store last allocated for subsequent stream allocation */
1684 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1685 spin_lock(&sbi->s_md_lock);
1686 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1687 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1688 spin_unlock(&sbi->s_md_lock);
1693 * regular allocator, for general purposes allocation
1696 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1697 struct ext4_buddy *e4b,
1698 int finish_group)
1700 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1701 struct ext4_free_extent *bex = &ac->ac_b_ex;
1702 struct ext4_free_extent *gex = &ac->ac_g_ex;
1703 struct ext4_free_extent ex;
1704 int max;
1706 if (ac->ac_status == AC_STATUS_FOUND)
1707 return;
1709 * We don't want to scan for a whole year
1711 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1712 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1713 ac->ac_status = AC_STATUS_BREAK;
1714 return;
1718 * Haven't found good chunk so far, let's continue
1720 if (bex->fe_len < gex->fe_len)
1721 return;
1723 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1724 && bex->fe_group == e4b->bd_group) {
1725 /* recheck chunk's availability - we don't know
1726 * when it was found (within this lock-unlock
1727 * period or not) */
1728 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1729 if (max >= gex->fe_len) {
1730 ext4_mb_use_best_found(ac, e4b);
1731 return;
1737 * The routine checks whether found extent is good enough. If it is,
1738 * then the extent gets marked used and flag is set to the context
1739 * to stop scanning. Otherwise, the extent is compared with the
1740 * previous found extent and if new one is better, then it's stored
1741 * in the context. Later, the best found extent will be used, if
1742 * mballoc can't find good enough extent.
1744 * FIXME: real allocation policy is to be designed yet!
1746 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1747 struct ext4_free_extent *ex,
1748 struct ext4_buddy *e4b)
1750 struct ext4_free_extent *bex = &ac->ac_b_ex;
1751 struct ext4_free_extent *gex = &ac->ac_g_ex;
1753 BUG_ON(ex->fe_len <= 0);
1754 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1755 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1756 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1758 ac->ac_found++;
1761 * The special case - take what you catch first
1763 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1764 *bex = *ex;
1765 ext4_mb_use_best_found(ac, e4b);
1766 return;
1770 * Let's check whether the chuck is good enough
1772 if (ex->fe_len == gex->fe_len) {
1773 *bex = *ex;
1774 ext4_mb_use_best_found(ac, e4b);
1775 return;
1779 * If this is first found extent, just store it in the context
1781 if (bex->fe_len == 0) {
1782 *bex = *ex;
1783 return;
1787 * If new found extent is better, store it in the context
1789 if (bex->fe_len < gex->fe_len) {
1790 /* if the request isn't satisfied, any found extent
1791 * larger than previous best one is better */
1792 if (ex->fe_len > bex->fe_len)
1793 *bex = *ex;
1794 } else if (ex->fe_len > gex->fe_len) {
1795 /* if the request is satisfied, then we try to find
1796 * an extent that still satisfy the request, but is
1797 * smaller than previous one */
1798 if (ex->fe_len < bex->fe_len)
1799 *bex = *ex;
1802 ext4_mb_check_limits(ac, e4b, 0);
1805 static noinline_for_stack
1806 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1807 struct ext4_buddy *e4b)
1809 struct ext4_free_extent ex = ac->ac_b_ex;
1810 ext4_group_t group = ex.fe_group;
1811 int max;
1812 int err;
1814 BUG_ON(ex.fe_len <= 0);
1815 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1816 if (err)
1817 return err;
1819 ext4_lock_group(ac->ac_sb, group);
1820 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1822 if (max > 0) {
1823 ac->ac_b_ex = ex;
1824 ext4_mb_use_best_found(ac, e4b);
1827 ext4_unlock_group(ac->ac_sb, group);
1828 ext4_mb_unload_buddy(e4b);
1830 return 0;
1833 static noinline_for_stack
1834 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1835 struct ext4_buddy *e4b)
1837 ext4_group_t group = ac->ac_g_ex.fe_group;
1838 int max;
1839 int err;
1840 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1841 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1842 struct ext4_free_extent ex;
1844 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1845 return 0;
1846 if (grp->bb_free == 0)
1847 return 0;
1849 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1850 if (err)
1851 return err;
1853 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1854 ext4_mb_unload_buddy(e4b);
1855 return 0;
1858 ext4_lock_group(ac->ac_sb, group);
1859 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1860 ac->ac_g_ex.fe_len, &ex);
1861 ex.fe_logical = 0xDEADFA11; /* debug value */
1863 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1864 ext4_fsblk_t start;
1866 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1867 ex.fe_start;
1868 /* use do_div to get remainder (would be 64-bit modulo) */
1869 if (do_div(start, sbi->s_stripe) == 0) {
1870 ac->ac_found++;
1871 ac->ac_b_ex = ex;
1872 ext4_mb_use_best_found(ac, e4b);
1874 } else if (max >= ac->ac_g_ex.fe_len) {
1875 BUG_ON(ex.fe_len <= 0);
1876 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1877 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1878 ac->ac_found++;
1879 ac->ac_b_ex = ex;
1880 ext4_mb_use_best_found(ac, e4b);
1881 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1882 /* Sometimes, caller may want to merge even small
1883 * number of blocks to an existing extent */
1884 BUG_ON(ex.fe_len <= 0);
1885 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1886 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1887 ac->ac_found++;
1888 ac->ac_b_ex = ex;
1889 ext4_mb_use_best_found(ac, e4b);
1891 ext4_unlock_group(ac->ac_sb, group);
1892 ext4_mb_unload_buddy(e4b);
1894 return 0;
1898 * The routine scans buddy structures (not bitmap!) from given order
1899 * to max order and tries to find big enough chunk to satisfy the req
1901 static noinline_for_stack
1902 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1903 struct ext4_buddy *e4b)
1905 struct super_block *sb = ac->ac_sb;
1906 struct ext4_group_info *grp = e4b->bd_info;
1907 void *buddy;
1908 int i;
1909 int k;
1910 int max;
1912 BUG_ON(ac->ac_2order <= 0);
1913 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1914 if (grp->bb_counters[i] == 0)
1915 continue;
1917 buddy = mb_find_buddy(e4b, i, &max);
1918 BUG_ON(buddy == NULL);
1920 k = mb_find_next_zero_bit(buddy, max, 0);
1921 BUG_ON(k >= max);
1923 ac->ac_found++;
1925 ac->ac_b_ex.fe_len = 1 << i;
1926 ac->ac_b_ex.fe_start = k << i;
1927 ac->ac_b_ex.fe_group = e4b->bd_group;
1929 ext4_mb_use_best_found(ac, e4b);
1931 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1933 if (EXT4_SB(sb)->s_mb_stats)
1934 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1936 break;
1941 * The routine scans the group and measures all found extents.
1942 * In order to optimize scanning, caller must pass number of
1943 * free blocks in the group, so the routine can know upper limit.
1945 static noinline_for_stack
1946 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1947 struct ext4_buddy *e4b)
1949 struct super_block *sb = ac->ac_sb;
1950 void *bitmap = e4b->bd_bitmap;
1951 struct ext4_free_extent ex;
1952 int i;
1953 int free;
1955 free = e4b->bd_info->bb_free;
1956 BUG_ON(free <= 0);
1958 i = e4b->bd_info->bb_first_free;
1960 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1961 i = mb_find_next_zero_bit(bitmap,
1962 EXT4_CLUSTERS_PER_GROUP(sb), i);
1963 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1965 * IF we have corrupt bitmap, we won't find any
1966 * free blocks even though group info says we
1967 * we have free blocks
1969 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1970 "%d free clusters as per "
1971 "group info. But bitmap says 0",
1972 free);
1973 break;
1976 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1977 BUG_ON(ex.fe_len <= 0);
1978 if (free < ex.fe_len) {
1979 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1980 "%d free clusters as per "
1981 "group info. But got %d blocks",
1982 free, ex.fe_len);
1984 * The number of free blocks differs. This mostly
1985 * indicate that the bitmap is corrupt. So exit
1986 * without claiming the space.
1988 break;
1990 ex.fe_logical = 0xDEADC0DE; /* debug value */
1991 ext4_mb_measure_extent(ac, &ex, e4b);
1993 i += ex.fe_len;
1994 free -= ex.fe_len;
1997 ext4_mb_check_limits(ac, e4b, 1);
2001 * This is a special case for storages like raid5
2002 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2004 static noinline_for_stack
2005 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2006 struct ext4_buddy *e4b)
2008 struct super_block *sb = ac->ac_sb;
2009 struct ext4_sb_info *sbi = EXT4_SB(sb);
2010 void *bitmap = e4b->bd_bitmap;
2011 struct ext4_free_extent ex;
2012 ext4_fsblk_t first_group_block;
2013 ext4_fsblk_t a;
2014 ext4_grpblk_t i;
2015 int max;
2017 BUG_ON(sbi->s_stripe == 0);
2019 /* find first stripe-aligned block in group */
2020 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2022 a = first_group_block + sbi->s_stripe - 1;
2023 do_div(a, sbi->s_stripe);
2024 i = (a * sbi->s_stripe) - first_group_block;
2026 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2027 if (!mb_test_bit(i, bitmap)) {
2028 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2029 if (max >= sbi->s_stripe) {
2030 ac->ac_found++;
2031 ex.fe_logical = 0xDEADF00D; /* debug value */
2032 ac->ac_b_ex = ex;
2033 ext4_mb_use_best_found(ac, e4b);
2034 break;
2037 i += sbi->s_stripe;
2042 * This is now called BEFORE we load the buddy bitmap.
2043 * Returns either 1 or 0 indicating that the group is either suitable
2044 * for the allocation or not. In addition it can also return negative
2045 * error code when something goes wrong.
2047 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2048 ext4_group_t group, int cr)
2050 unsigned free, fragments;
2051 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2052 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2054 BUG_ON(cr < 0 || cr >= 4);
2056 free = grp->bb_free;
2057 if (free == 0)
2058 return 0;
2059 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2060 return 0;
2062 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2063 return 0;
2065 /* We only do this if the grp has never been initialized */
2066 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2067 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2068 if (ret)
2069 return ret;
2072 fragments = grp->bb_fragments;
2073 if (fragments == 0)
2074 return 0;
2076 switch (cr) {
2077 case 0:
2078 BUG_ON(ac->ac_2order == 0);
2080 /* Avoid using the first bg of a flexgroup for data files */
2081 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2082 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2083 ((group % flex_size) == 0))
2084 return 0;
2086 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2087 (free / fragments) >= ac->ac_g_ex.fe_len)
2088 return 1;
2090 if (grp->bb_largest_free_order < ac->ac_2order)
2091 return 0;
2093 return 1;
2094 case 1:
2095 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2096 return 1;
2097 break;
2098 case 2:
2099 if (free >= ac->ac_g_ex.fe_len)
2100 return 1;
2101 break;
2102 case 3:
2103 return 1;
2104 default:
2105 BUG();
2108 return 0;
2111 static noinline_for_stack int
2112 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2114 ext4_group_t ngroups, group, i;
2115 int cr;
2116 int err = 0, first_err = 0;
2117 struct ext4_sb_info *sbi;
2118 struct super_block *sb;
2119 struct ext4_buddy e4b;
2121 sb = ac->ac_sb;
2122 sbi = EXT4_SB(sb);
2123 ngroups = ext4_get_groups_count(sb);
2124 /* non-extent files are limited to low blocks/groups */
2125 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2126 ngroups = sbi->s_blockfile_groups;
2128 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2130 /* first, try the goal */
2131 err = ext4_mb_find_by_goal(ac, &e4b);
2132 if (err || ac->ac_status == AC_STATUS_FOUND)
2133 goto out;
2135 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2136 goto out;
2139 * ac->ac2_order is set only if the fe_len is a power of 2
2140 * if ac2_order is set we also set criteria to 0 so that we
2141 * try exact allocation using buddy.
2143 i = fls(ac->ac_g_ex.fe_len);
2144 ac->ac_2order = 0;
2146 * We search using buddy data only if the order of the request
2147 * is greater than equal to the sbi_s_mb_order2_reqs
2148 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2149 * We also support searching for power-of-two requests only for
2150 * requests upto maximum buddy size we have constructed.
2152 if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2154 * This should tell if fe_len is exactly power of 2
2156 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2157 ac->ac_2order = i - 1;
2160 /* if stream allocation is enabled, use global goal */
2161 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2162 /* TBD: may be hot point */
2163 spin_lock(&sbi->s_md_lock);
2164 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2165 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2166 spin_unlock(&sbi->s_md_lock);
2169 /* Let's just scan groups to find more-less suitable blocks */
2170 cr = ac->ac_2order ? 0 : 1;
2172 * cr == 0 try to get exact allocation,
2173 * cr == 3 try to get anything
2175 repeat:
2176 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2177 ac->ac_criteria = cr;
2179 * searching for the right group start
2180 * from the goal value specified
2182 group = ac->ac_g_ex.fe_group;
2184 for (i = 0; i < ngroups; group++, i++) {
2185 int ret = 0;
2186 cond_resched();
2188 * Artificially restricted ngroups for non-extent
2189 * files makes group > ngroups possible on first loop.
2191 if (group >= ngroups)
2192 group = 0;
2194 /* This now checks without needing the buddy page */
2195 ret = ext4_mb_good_group(ac, group, cr);
2196 if (ret <= 0) {
2197 if (!first_err)
2198 first_err = ret;
2199 continue;
2202 err = ext4_mb_load_buddy(sb, group, &e4b);
2203 if (err)
2204 goto out;
2206 ext4_lock_group(sb, group);
2209 * We need to check again after locking the
2210 * block group
2212 ret = ext4_mb_good_group(ac, group, cr);
2213 if (ret <= 0) {
2214 ext4_unlock_group(sb, group);
2215 ext4_mb_unload_buddy(&e4b);
2216 if (!first_err)
2217 first_err = ret;
2218 continue;
2221 ac->ac_groups_scanned++;
2222 if (cr == 0)
2223 ext4_mb_simple_scan_group(ac, &e4b);
2224 else if (cr == 1 && sbi->s_stripe &&
2225 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2226 ext4_mb_scan_aligned(ac, &e4b);
2227 else
2228 ext4_mb_complex_scan_group(ac, &e4b);
2230 ext4_unlock_group(sb, group);
2231 ext4_mb_unload_buddy(&e4b);
2233 if (ac->ac_status != AC_STATUS_CONTINUE)
2234 break;
2238 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2239 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2241 * We've been searching too long. Let's try to allocate
2242 * the best chunk we've found so far
2245 ext4_mb_try_best_found(ac, &e4b);
2246 if (ac->ac_status != AC_STATUS_FOUND) {
2248 * Someone more lucky has already allocated it.
2249 * The only thing we can do is just take first
2250 * found block(s)
2251 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2253 ac->ac_b_ex.fe_group = 0;
2254 ac->ac_b_ex.fe_start = 0;
2255 ac->ac_b_ex.fe_len = 0;
2256 ac->ac_status = AC_STATUS_CONTINUE;
2257 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2258 cr = 3;
2259 atomic_inc(&sbi->s_mb_lost_chunks);
2260 goto repeat;
2263 out:
2264 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2265 err = first_err;
2266 return err;
2269 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2271 struct super_block *sb = seq->private;
2272 ext4_group_t group;
2274 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2275 return NULL;
2276 group = *pos + 1;
2277 return (void *) ((unsigned long) group);
2280 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2282 struct super_block *sb = seq->private;
2283 ext4_group_t group;
2285 ++*pos;
2286 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2287 return NULL;
2288 group = *pos + 1;
2289 return (void *) ((unsigned long) group);
2292 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2294 struct super_block *sb = seq->private;
2295 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2296 int i;
2297 int err, buddy_loaded = 0;
2298 struct ext4_buddy e4b;
2299 struct ext4_group_info *grinfo;
2300 struct sg {
2301 struct ext4_group_info info;
2302 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2303 } sg;
2305 group--;
2306 if (group == 0)
2307 seq_puts(seq, "#group: free frags first ["
2308 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2309 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2311 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2312 sizeof(struct ext4_group_info);
2313 grinfo = ext4_get_group_info(sb, group);
2314 /* Load the group info in memory only if not already loaded. */
2315 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2316 err = ext4_mb_load_buddy(sb, group, &e4b);
2317 if (err) {
2318 seq_printf(seq, "#%-5u: I/O error\n", group);
2319 return 0;
2321 buddy_loaded = 1;
2324 memcpy(&sg, ext4_get_group_info(sb, group), i);
2326 if (buddy_loaded)
2327 ext4_mb_unload_buddy(&e4b);
2329 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2330 sg.info.bb_fragments, sg.info.bb_first_free);
2331 for (i = 0; i <= 13; i++)
2332 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2333 sg.info.bb_counters[i] : 0);
2334 seq_printf(seq, " ]\n");
2336 return 0;
2339 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2343 static const struct seq_operations ext4_mb_seq_groups_ops = {
2344 .start = ext4_mb_seq_groups_start,
2345 .next = ext4_mb_seq_groups_next,
2346 .stop = ext4_mb_seq_groups_stop,
2347 .show = ext4_mb_seq_groups_show,
2350 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2352 struct super_block *sb = PDE_DATA(inode);
2353 int rc;
2355 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2356 if (rc == 0) {
2357 struct seq_file *m = file->private_data;
2358 m->private = sb;
2360 return rc;
2364 const struct file_operations ext4_seq_mb_groups_fops = {
2365 .open = ext4_mb_seq_groups_open,
2366 .read = seq_read,
2367 .llseek = seq_lseek,
2368 .release = seq_release,
2371 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2373 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2374 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2376 BUG_ON(!cachep);
2377 return cachep;
2381 * Allocate the top-level s_group_info array for the specified number
2382 * of groups
2384 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2386 struct ext4_sb_info *sbi = EXT4_SB(sb);
2387 unsigned size;
2388 struct ext4_group_info ***new_groupinfo;
2390 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2391 EXT4_DESC_PER_BLOCK_BITS(sb);
2392 if (size <= sbi->s_group_info_size)
2393 return 0;
2395 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2396 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2397 if (!new_groupinfo) {
2398 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2399 return -ENOMEM;
2401 if (sbi->s_group_info) {
2402 memcpy(new_groupinfo, sbi->s_group_info,
2403 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2404 kvfree(sbi->s_group_info);
2406 sbi->s_group_info = new_groupinfo;
2407 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2408 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2409 sbi->s_group_info_size);
2410 return 0;
2413 /* Create and initialize ext4_group_info data for the given group. */
2414 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2415 struct ext4_group_desc *desc)
2417 int i;
2418 int metalen = 0;
2419 struct ext4_sb_info *sbi = EXT4_SB(sb);
2420 struct ext4_group_info **meta_group_info;
2421 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2424 * First check if this group is the first of a reserved block.
2425 * If it's true, we have to allocate a new table of pointers
2426 * to ext4_group_info structures
2428 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2429 metalen = sizeof(*meta_group_info) <<
2430 EXT4_DESC_PER_BLOCK_BITS(sb);
2431 meta_group_info = kmalloc(metalen, GFP_NOFS);
2432 if (meta_group_info == NULL) {
2433 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2434 "for a buddy group");
2435 goto exit_meta_group_info;
2437 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2438 meta_group_info;
2441 meta_group_info =
2442 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2443 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2445 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2446 if (meta_group_info[i] == NULL) {
2447 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2448 goto exit_group_info;
2450 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2451 &(meta_group_info[i]->bb_state));
2454 * initialize bb_free to be able to skip
2455 * empty groups without initialization
2457 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2458 meta_group_info[i]->bb_free =
2459 ext4_free_clusters_after_init(sb, group, desc);
2460 } else {
2461 meta_group_info[i]->bb_free =
2462 ext4_free_group_clusters(sb, desc);
2465 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2466 init_rwsem(&meta_group_info[i]->alloc_sem);
2467 meta_group_info[i]->bb_free_root = RB_ROOT;
2468 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2470 #ifdef DOUBLE_CHECK
2472 struct buffer_head *bh;
2473 meta_group_info[i]->bb_bitmap =
2474 kmalloc(sb->s_blocksize, GFP_NOFS);
2475 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2476 bh = ext4_read_block_bitmap(sb, group);
2477 BUG_ON(IS_ERR_OR_NULL(bh));
2478 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2479 sb->s_blocksize);
2480 put_bh(bh);
2482 #endif
2484 return 0;
2486 exit_group_info:
2487 /* If a meta_group_info table has been allocated, release it now */
2488 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2489 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2490 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2492 exit_meta_group_info:
2493 return -ENOMEM;
2494 } /* ext4_mb_add_groupinfo */
2496 static int ext4_mb_init_backend(struct super_block *sb)
2498 ext4_group_t ngroups = ext4_get_groups_count(sb);
2499 ext4_group_t i;
2500 struct ext4_sb_info *sbi = EXT4_SB(sb);
2501 int err;
2502 struct ext4_group_desc *desc;
2503 struct kmem_cache *cachep;
2505 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2506 if (err)
2507 return err;
2509 sbi->s_buddy_cache = new_inode(sb);
2510 if (sbi->s_buddy_cache == NULL) {
2511 ext4_msg(sb, KERN_ERR, "can't get new inode");
2512 goto err_freesgi;
2514 /* To avoid potentially colliding with an valid on-disk inode number,
2515 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2516 * not in the inode hash, so it should never be found by iget(), but
2517 * this will avoid confusion if it ever shows up during debugging. */
2518 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2519 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2520 for (i = 0; i < ngroups; i++) {
2521 desc = ext4_get_group_desc(sb, i, NULL);
2522 if (desc == NULL) {
2523 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2524 goto err_freebuddy;
2526 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2527 goto err_freebuddy;
2530 return 0;
2532 err_freebuddy:
2533 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2534 while (i-- > 0)
2535 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2536 i = sbi->s_group_info_size;
2537 while (i-- > 0)
2538 kfree(sbi->s_group_info[i]);
2539 iput(sbi->s_buddy_cache);
2540 err_freesgi:
2541 kvfree(sbi->s_group_info);
2542 return -ENOMEM;
2545 static void ext4_groupinfo_destroy_slabs(void)
2547 int i;
2549 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2550 if (ext4_groupinfo_caches[i])
2551 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2552 ext4_groupinfo_caches[i] = NULL;
2556 static int ext4_groupinfo_create_slab(size_t size)
2558 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2559 int slab_size;
2560 int blocksize_bits = order_base_2(size);
2561 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2562 struct kmem_cache *cachep;
2564 if (cache_index >= NR_GRPINFO_CACHES)
2565 return -EINVAL;
2567 if (unlikely(cache_index < 0))
2568 cache_index = 0;
2570 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2571 if (ext4_groupinfo_caches[cache_index]) {
2572 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2573 return 0; /* Already created */
2576 slab_size = offsetof(struct ext4_group_info,
2577 bb_counters[blocksize_bits + 2]);
2579 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2580 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2581 NULL);
2583 ext4_groupinfo_caches[cache_index] = cachep;
2585 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2586 if (!cachep) {
2587 printk(KERN_EMERG
2588 "EXT4-fs: no memory for groupinfo slab cache\n");
2589 return -ENOMEM;
2592 return 0;
2595 int ext4_mb_init(struct super_block *sb)
2597 struct ext4_sb_info *sbi = EXT4_SB(sb);
2598 unsigned i, j;
2599 unsigned offset, offset_incr;
2600 unsigned max;
2601 int ret;
2603 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2605 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2606 if (sbi->s_mb_offsets == NULL) {
2607 ret = -ENOMEM;
2608 goto out;
2611 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2612 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2613 if (sbi->s_mb_maxs == NULL) {
2614 ret = -ENOMEM;
2615 goto out;
2618 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2619 if (ret < 0)
2620 goto out;
2622 /* order 0 is regular bitmap */
2623 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2624 sbi->s_mb_offsets[0] = 0;
2626 i = 1;
2627 offset = 0;
2628 offset_incr = 1 << (sb->s_blocksize_bits - 1);
2629 max = sb->s_blocksize << 2;
2630 do {
2631 sbi->s_mb_offsets[i] = offset;
2632 sbi->s_mb_maxs[i] = max;
2633 offset += offset_incr;
2634 offset_incr = offset_incr >> 1;
2635 max = max >> 1;
2636 i++;
2637 } while (i <= sb->s_blocksize_bits + 1);
2639 spin_lock_init(&sbi->s_md_lock);
2640 spin_lock_init(&sbi->s_bal_lock);
2641 sbi->s_mb_free_pending = 0;
2643 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2644 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2645 sbi->s_mb_stats = MB_DEFAULT_STATS;
2646 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2647 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2649 * The default group preallocation is 512, which for 4k block
2650 * sizes translates to 2 megabytes. However for bigalloc file
2651 * systems, this is probably too big (i.e, if the cluster size
2652 * is 1 megabyte, then group preallocation size becomes half a
2653 * gigabyte!). As a default, we will keep a two megabyte
2654 * group pralloc size for cluster sizes up to 64k, and after
2655 * that, we will force a minimum group preallocation size of
2656 * 32 clusters. This translates to 8 megs when the cluster
2657 * size is 256k, and 32 megs when the cluster size is 1 meg,
2658 * which seems reasonable as a default.
2660 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2661 sbi->s_cluster_bits, 32);
2663 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2664 * to the lowest multiple of s_stripe which is bigger than
2665 * the s_mb_group_prealloc as determined above. We want
2666 * the preallocation size to be an exact multiple of the
2667 * RAID stripe size so that preallocations don't fragment
2668 * the stripes.
2670 if (sbi->s_stripe > 1) {
2671 sbi->s_mb_group_prealloc = roundup(
2672 sbi->s_mb_group_prealloc, sbi->s_stripe);
2675 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2676 if (sbi->s_locality_groups == NULL) {
2677 ret = -ENOMEM;
2678 goto out;
2680 for_each_possible_cpu(i) {
2681 struct ext4_locality_group *lg;
2682 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2683 mutex_init(&lg->lg_mutex);
2684 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2685 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2686 spin_lock_init(&lg->lg_prealloc_lock);
2689 /* init file for buddy data */
2690 ret = ext4_mb_init_backend(sb);
2691 if (ret != 0)
2692 goto out_free_locality_groups;
2694 return 0;
2696 out_free_locality_groups:
2697 free_percpu(sbi->s_locality_groups);
2698 sbi->s_locality_groups = NULL;
2699 out:
2700 kfree(sbi->s_mb_offsets);
2701 sbi->s_mb_offsets = NULL;
2702 kfree(sbi->s_mb_maxs);
2703 sbi->s_mb_maxs = NULL;
2704 return ret;
2707 /* need to called with the ext4 group lock held */
2708 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2710 struct ext4_prealloc_space *pa;
2711 struct list_head *cur, *tmp;
2712 int count = 0;
2714 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2715 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2716 list_del(&pa->pa_group_list);
2717 count++;
2718 kmem_cache_free(ext4_pspace_cachep, pa);
2720 if (count)
2721 mb_debug(1, "mballoc: %u PAs left\n", count);
2725 int ext4_mb_release(struct super_block *sb)
2727 ext4_group_t ngroups = ext4_get_groups_count(sb);
2728 ext4_group_t i;
2729 int num_meta_group_infos;
2730 struct ext4_group_info *grinfo;
2731 struct ext4_sb_info *sbi = EXT4_SB(sb);
2732 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2734 if (sbi->s_group_info) {
2735 for (i = 0; i < ngroups; i++) {
2736 grinfo = ext4_get_group_info(sb, i);
2737 #ifdef DOUBLE_CHECK
2738 kfree(grinfo->bb_bitmap);
2739 #endif
2740 ext4_lock_group(sb, i);
2741 ext4_mb_cleanup_pa(grinfo);
2742 ext4_unlock_group(sb, i);
2743 kmem_cache_free(cachep, grinfo);
2745 num_meta_group_infos = (ngroups +
2746 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2747 EXT4_DESC_PER_BLOCK_BITS(sb);
2748 for (i = 0; i < num_meta_group_infos; i++)
2749 kfree(sbi->s_group_info[i]);
2750 kvfree(sbi->s_group_info);
2752 kfree(sbi->s_mb_offsets);
2753 kfree(sbi->s_mb_maxs);
2754 iput(sbi->s_buddy_cache);
2755 if (sbi->s_mb_stats) {
2756 ext4_msg(sb, KERN_INFO,
2757 "mballoc: %u blocks %u reqs (%u success)",
2758 atomic_read(&sbi->s_bal_allocated),
2759 atomic_read(&sbi->s_bal_reqs),
2760 atomic_read(&sbi->s_bal_success));
2761 ext4_msg(sb, KERN_INFO,
2762 "mballoc: %u extents scanned, %u goal hits, "
2763 "%u 2^N hits, %u breaks, %u lost",
2764 atomic_read(&sbi->s_bal_ex_scanned),
2765 atomic_read(&sbi->s_bal_goals),
2766 atomic_read(&sbi->s_bal_2orders),
2767 atomic_read(&sbi->s_bal_breaks),
2768 atomic_read(&sbi->s_mb_lost_chunks));
2769 ext4_msg(sb, KERN_INFO,
2770 "mballoc: %lu generated and it took %Lu",
2771 sbi->s_mb_buddies_generated,
2772 sbi->s_mb_generation_time);
2773 ext4_msg(sb, KERN_INFO,
2774 "mballoc: %u preallocated, %u discarded",
2775 atomic_read(&sbi->s_mb_preallocated),
2776 atomic_read(&sbi->s_mb_discarded));
2779 free_percpu(sbi->s_locality_groups);
2781 return 0;
2784 static inline int ext4_issue_discard(struct super_block *sb,
2785 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2787 ext4_fsblk_t discard_block;
2789 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2790 ext4_group_first_block_no(sb, block_group));
2791 count = EXT4_C2B(EXT4_SB(sb), count);
2792 trace_ext4_discard_blocks(sb,
2793 (unsigned long long) discard_block, count);
2794 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2798 * This function is called by the jbd2 layer once the commit has finished,
2799 * so we know we can free the blocks that were released with that commit.
2801 static void ext4_free_data_callback(struct super_block *sb,
2802 struct ext4_journal_cb_entry *jce,
2803 int rc)
2805 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2806 struct ext4_buddy e4b;
2807 struct ext4_group_info *db;
2808 int err, count = 0, count2 = 0;
2810 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2811 entry->efd_count, entry->efd_group, entry);
2813 if (test_opt(sb, DISCARD)) {
2814 err = ext4_issue_discard(sb, entry->efd_group,
2815 entry->efd_start_cluster,
2816 entry->efd_count);
2817 if (err && err != -EOPNOTSUPP)
2818 ext4_msg(sb, KERN_WARNING, "discard request in"
2819 " group:%d block:%d count:%d failed"
2820 " with %d", entry->efd_group,
2821 entry->efd_start_cluster,
2822 entry->efd_count, err);
2825 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2826 /* we expect to find existing buddy because it's pinned */
2827 BUG_ON(err != 0);
2829 spin_lock(&EXT4_SB(sb)->s_md_lock);
2830 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
2831 spin_unlock(&EXT4_SB(sb)->s_md_lock);
2833 db = e4b.bd_info;
2834 /* there are blocks to put in buddy to make them really free */
2835 count += entry->efd_count;
2836 count2++;
2837 ext4_lock_group(sb, entry->efd_group);
2838 /* Take it out of per group rb tree */
2839 rb_erase(&entry->efd_node, &(db->bb_free_root));
2840 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2843 * Clear the trimmed flag for the group so that the next
2844 * ext4_trim_fs can trim it.
2845 * If the volume is mounted with -o discard, online discard
2846 * is supported and the free blocks will be trimmed online.
2848 if (!test_opt(sb, DISCARD))
2849 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2851 if (!db->bb_free_root.rb_node) {
2852 /* No more items in the per group rb tree
2853 * balance refcounts from ext4_mb_free_metadata()
2855 put_page(e4b.bd_buddy_page);
2856 put_page(e4b.bd_bitmap_page);
2858 ext4_unlock_group(sb, entry->efd_group);
2859 kmem_cache_free(ext4_free_data_cachep, entry);
2860 ext4_mb_unload_buddy(&e4b);
2862 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2865 int __init ext4_init_mballoc(void)
2867 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2868 SLAB_RECLAIM_ACCOUNT);
2869 if (ext4_pspace_cachep == NULL)
2870 return -ENOMEM;
2872 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2873 SLAB_RECLAIM_ACCOUNT);
2874 if (ext4_ac_cachep == NULL) {
2875 kmem_cache_destroy(ext4_pspace_cachep);
2876 return -ENOMEM;
2879 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2880 SLAB_RECLAIM_ACCOUNT);
2881 if (ext4_free_data_cachep == NULL) {
2882 kmem_cache_destroy(ext4_pspace_cachep);
2883 kmem_cache_destroy(ext4_ac_cachep);
2884 return -ENOMEM;
2886 return 0;
2889 void ext4_exit_mballoc(void)
2892 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2893 * before destroying the slab cache.
2895 rcu_barrier();
2896 kmem_cache_destroy(ext4_pspace_cachep);
2897 kmem_cache_destroy(ext4_ac_cachep);
2898 kmem_cache_destroy(ext4_free_data_cachep);
2899 ext4_groupinfo_destroy_slabs();
2904 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2905 * Returns 0 if success or error code
2907 static noinline_for_stack int
2908 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2909 handle_t *handle, unsigned int reserv_clstrs)
2911 struct buffer_head *bitmap_bh = NULL;
2912 struct ext4_group_desc *gdp;
2913 struct buffer_head *gdp_bh;
2914 struct ext4_sb_info *sbi;
2915 struct super_block *sb;
2916 ext4_fsblk_t block;
2917 int err, len;
2919 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2920 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2922 sb = ac->ac_sb;
2923 sbi = EXT4_SB(sb);
2925 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2926 if (IS_ERR(bitmap_bh)) {
2927 err = PTR_ERR(bitmap_bh);
2928 bitmap_bh = NULL;
2929 goto out_err;
2932 BUFFER_TRACE(bitmap_bh, "getting write access");
2933 err = ext4_journal_get_write_access(handle, bitmap_bh);
2934 if (err)
2935 goto out_err;
2937 err = -EIO;
2938 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2939 if (!gdp)
2940 goto out_err;
2942 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2943 ext4_free_group_clusters(sb, gdp));
2945 BUFFER_TRACE(gdp_bh, "get_write_access");
2946 err = ext4_journal_get_write_access(handle, gdp_bh);
2947 if (err)
2948 goto out_err;
2950 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2952 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2953 if (!ext4_data_block_valid(sbi, block, len)) {
2954 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2955 "fs metadata", block, block+len);
2956 /* File system mounted not to panic on error
2957 * Fix the bitmap and return EFSCORRUPTED
2958 * We leak some of the blocks here.
2960 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2961 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2962 ac->ac_b_ex.fe_len);
2963 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2964 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2965 if (!err)
2966 err = -EFSCORRUPTED;
2967 goto out_err;
2970 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2971 #ifdef AGGRESSIVE_CHECK
2973 int i;
2974 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2975 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2976 bitmap_bh->b_data));
2979 #endif
2980 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2981 ac->ac_b_ex.fe_len);
2982 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2983 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2984 ext4_free_group_clusters_set(sb, gdp,
2985 ext4_free_clusters_after_init(sb,
2986 ac->ac_b_ex.fe_group, gdp));
2988 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2989 ext4_free_group_clusters_set(sb, gdp, len);
2990 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2991 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2993 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2994 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2996 * Now reduce the dirty block count also. Should not go negative
2998 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2999 /* release all the reserved blocks if non delalloc */
3000 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3001 reserv_clstrs);
3003 if (sbi->s_log_groups_per_flex) {
3004 ext4_group_t flex_group = ext4_flex_group(sbi,
3005 ac->ac_b_ex.fe_group);
3006 atomic64_sub(ac->ac_b_ex.fe_len,
3007 &sbi->s_flex_groups[flex_group].free_clusters);
3010 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3011 if (err)
3012 goto out_err;
3013 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3015 out_err:
3016 brelse(bitmap_bh);
3017 return err;
3021 * here we normalize request for locality group
3022 * Group request are normalized to s_mb_group_prealloc, which goes to
3023 * s_strip if we set the same via mount option.
3024 * s_mb_group_prealloc can be configured via
3025 * /sys/fs/ext4/<partition>/mb_group_prealloc
3027 * XXX: should we try to preallocate more than the group has now?
3029 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3031 struct super_block *sb = ac->ac_sb;
3032 struct ext4_locality_group *lg = ac->ac_lg;
3034 BUG_ON(lg == NULL);
3035 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3036 mb_debug(1, "#%u: goal %u blocks for locality group\n",
3037 current->pid, ac->ac_g_ex.fe_len);
3041 * Normalization means making request better in terms of
3042 * size and alignment
3044 static noinline_for_stack void
3045 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3046 struct ext4_allocation_request *ar)
3048 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3049 int bsbits, max;
3050 ext4_lblk_t end;
3051 loff_t size, start_off;
3052 loff_t orig_size __maybe_unused;
3053 ext4_lblk_t start;
3054 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3055 struct ext4_prealloc_space *pa;
3057 /* do normalize only data requests, metadata requests
3058 do not need preallocation */
3059 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3060 return;
3062 /* sometime caller may want exact blocks */
3063 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3064 return;
3066 /* caller may indicate that preallocation isn't
3067 * required (it's a tail, for example) */
3068 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3069 return;
3071 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3072 ext4_mb_normalize_group_request(ac);
3073 return ;
3076 bsbits = ac->ac_sb->s_blocksize_bits;
3078 /* first, let's learn actual file size
3079 * given current request is allocated */
3080 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3081 size = size << bsbits;
3082 if (size < i_size_read(ac->ac_inode))
3083 size = i_size_read(ac->ac_inode);
3084 orig_size = size;
3086 /* max size of free chunks */
3087 max = 2 << bsbits;
3089 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3090 (req <= (size) || max <= (chunk_size))
3092 /* first, try to predict filesize */
3093 /* XXX: should this table be tunable? */
3094 start_off = 0;
3095 if (size <= 16 * 1024) {
3096 size = 16 * 1024;
3097 } else if (size <= 32 * 1024) {
3098 size = 32 * 1024;
3099 } else if (size <= 64 * 1024) {
3100 size = 64 * 1024;
3101 } else if (size <= 128 * 1024) {
3102 size = 128 * 1024;
3103 } else if (size <= 256 * 1024) {
3104 size = 256 * 1024;
3105 } else if (size <= 512 * 1024) {
3106 size = 512 * 1024;
3107 } else if (size <= 1024 * 1024) {
3108 size = 1024 * 1024;
3109 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3110 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3111 (21 - bsbits)) << 21;
3112 size = 2 * 1024 * 1024;
3113 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3114 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3115 (22 - bsbits)) << 22;
3116 size = 4 * 1024 * 1024;
3117 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3118 (8<<20)>>bsbits, max, 8 * 1024)) {
3119 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3120 (23 - bsbits)) << 23;
3121 size = 8 * 1024 * 1024;
3122 } else {
3123 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3124 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3125 ac->ac_o_ex.fe_len) << bsbits;
3127 size = size >> bsbits;
3128 start = start_off >> bsbits;
3130 /* don't cover already allocated blocks in selected range */
3131 if (ar->pleft && start <= ar->lleft) {
3132 size -= ar->lleft + 1 - start;
3133 start = ar->lleft + 1;
3135 if (ar->pright && start + size - 1 >= ar->lright)
3136 size -= start + size - ar->lright;
3139 * Trim allocation request for filesystems with artificially small
3140 * groups.
3142 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3143 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3145 end = start + size;
3147 /* check we don't cross already preallocated blocks */
3148 rcu_read_lock();
3149 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3150 ext4_lblk_t pa_end;
3152 if (pa->pa_deleted)
3153 continue;
3154 spin_lock(&pa->pa_lock);
3155 if (pa->pa_deleted) {
3156 spin_unlock(&pa->pa_lock);
3157 continue;
3160 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3161 pa->pa_len);
3163 /* PA must not overlap original request */
3164 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3165 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3167 /* skip PAs this normalized request doesn't overlap with */
3168 if (pa->pa_lstart >= end || pa_end <= start) {
3169 spin_unlock(&pa->pa_lock);
3170 continue;
3172 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3174 /* adjust start or end to be adjacent to this pa */
3175 if (pa_end <= ac->ac_o_ex.fe_logical) {
3176 BUG_ON(pa_end < start);
3177 start = pa_end;
3178 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3179 BUG_ON(pa->pa_lstart > end);
3180 end = pa->pa_lstart;
3182 spin_unlock(&pa->pa_lock);
3184 rcu_read_unlock();
3185 size = end - start;
3187 /* XXX: extra loop to check we really don't overlap preallocations */
3188 rcu_read_lock();
3189 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3190 ext4_lblk_t pa_end;
3192 spin_lock(&pa->pa_lock);
3193 if (pa->pa_deleted == 0) {
3194 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3195 pa->pa_len);
3196 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3198 spin_unlock(&pa->pa_lock);
3200 rcu_read_unlock();
3202 if (start + size <= ac->ac_o_ex.fe_logical &&
3203 start > ac->ac_o_ex.fe_logical) {
3204 ext4_msg(ac->ac_sb, KERN_ERR,
3205 "start %lu, size %lu, fe_logical %lu",
3206 (unsigned long) start, (unsigned long) size,
3207 (unsigned long) ac->ac_o_ex.fe_logical);
3208 BUG();
3210 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3212 /* now prepare goal request */
3214 /* XXX: is it better to align blocks WRT to logical
3215 * placement or satisfy big request as is */
3216 ac->ac_g_ex.fe_logical = start;
3217 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3219 /* define goal start in order to merge */
3220 if (ar->pright && (ar->lright == (start + size))) {
3221 /* merge to the right */
3222 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3223 &ac->ac_f_ex.fe_group,
3224 &ac->ac_f_ex.fe_start);
3225 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3227 if (ar->pleft && (ar->lleft + 1 == start)) {
3228 /* merge to the left */
3229 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3230 &ac->ac_f_ex.fe_group,
3231 &ac->ac_f_ex.fe_start);
3232 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3235 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3236 (unsigned) orig_size, (unsigned) start);
3239 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3241 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3243 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3244 atomic_inc(&sbi->s_bal_reqs);
3245 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3246 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3247 atomic_inc(&sbi->s_bal_success);
3248 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3249 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3250 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3251 atomic_inc(&sbi->s_bal_goals);
3252 if (ac->ac_found > sbi->s_mb_max_to_scan)
3253 atomic_inc(&sbi->s_bal_breaks);
3256 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3257 trace_ext4_mballoc_alloc(ac);
3258 else
3259 trace_ext4_mballoc_prealloc(ac);
3263 * Called on failure; free up any blocks from the inode PA for this
3264 * context. We don't need this for MB_GROUP_PA because we only change
3265 * pa_free in ext4_mb_release_context(), but on failure, we've already
3266 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3268 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3270 struct ext4_prealloc_space *pa = ac->ac_pa;
3271 struct ext4_buddy e4b;
3272 int err;
3274 if (pa == NULL) {
3275 if (ac->ac_f_ex.fe_len == 0)
3276 return;
3277 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3278 if (err) {
3280 * This should never happen since we pin the
3281 * pages in the ext4_allocation_context so
3282 * ext4_mb_load_buddy() should never fail.
3284 WARN(1, "mb_load_buddy failed (%d)", err);
3285 return;
3287 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3288 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3289 ac->ac_f_ex.fe_len);
3290 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3291 ext4_mb_unload_buddy(&e4b);
3292 return;
3294 if (pa->pa_type == MB_INODE_PA)
3295 pa->pa_free += ac->ac_b_ex.fe_len;
3299 * use blocks preallocated to inode
3301 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3302 struct ext4_prealloc_space *pa)
3304 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3305 ext4_fsblk_t start;
3306 ext4_fsblk_t end;
3307 int len;
3309 /* found preallocated blocks, use them */
3310 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3311 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3312 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3313 len = EXT4_NUM_B2C(sbi, end - start);
3314 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3315 &ac->ac_b_ex.fe_start);
3316 ac->ac_b_ex.fe_len = len;
3317 ac->ac_status = AC_STATUS_FOUND;
3318 ac->ac_pa = pa;
3320 BUG_ON(start < pa->pa_pstart);
3321 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3322 BUG_ON(pa->pa_free < len);
3323 pa->pa_free -= len;
3325 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3329 * use blocks preallocated to locality group
3331 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3332 struct ext4_prealloc_space *pa)
3334 unsigned int len = ac->ac_o_ex.fe_len;
3336 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3337 &ac->ac_b_ex.fe_group,
3338 &ac->ac_b_ex.fe_start);
3339 ac->ac_b_ex.fe_len = len;
3340 ac->ac_status = AC_STATUS_FOUND;
3341 ac->ac_pa = pa;
3343 /* we don't correct pa_pstart or pa_plen here to avoid
3344 * possible race when the group is being loaded concurrently
3345 * instead we correct pa later, after blocks are marked
3346 * in on-disk bitmap -- see ext4_mb_release_context()
3347 * Other CPUs are prevented from allocating from this pa by lg_mutex
3349 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3353 * Return the prealloc space that have minimal distance
3354 * from the goal block. @cpa is the prealloc
3355 * space that is having currently known minimal distance
3356 * from the goal block.
3358 static struct ext4_prealloc_space *
3359 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3360 struct ext4_prealloc_space *pa,
3361 struct ext4_prealloc_space *cpa)
3363 ext4_fsblk_t cur_distance, new_distance;
3365 if (cpa == NULL) {
3366 atomic_inc(&pa->pa_count);
3367 return pa;
3369 cur_distance = abs(goal_block - cpa->pa_pstart);
3370 new_distance = abs(goal_block - pa->pa_pstart);
3372 if (cur_distance <= new_distance)
3373 return cpa;
3375 /* drop the previous reference */
3376 atomic_dec(&cpa->pa_count);
3377 atomic_inc(&pa->pa_count);
3378 return pa;
3382 * search goal blocks in preallocated space
3384 static noinline_for_stack int
3385 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3387 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3388 int order, i;
3389 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3390 struct ext4_locality_group *lg;
3391 struct ext4_prealloc_space *pa, *cpa = NULL;
3392 ext4_fsblk_t goal_block;
3394 /* only data can be preallocated */
3395 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3396 return 0;
3398 /* first, try per-file preallocation */
3399 rcu_read_lock();
3400 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3402 /* all fields in this condition don't change,
3403 * so we can skip locking for them */
3404 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3405 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3406 EXT4_C2B(sbi, pa->pa_len)))
3407 continue;
3409 /* non-extent files can't have physical blocks past 2^32 */
3410 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3411 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3412 EXT4_MAX_BLOCK_FILE_PHYS))
3413 continue;
3415 /* found preallocated blocks, use them */
3416 spin_lock(&pa->pa_lock);
3417 if (pa->pa_deleted == 0 && pa->pa_free) {
3418 atomic_inc(&pa->pa_count);
3419 ext4_mb_use_inode_pa(ac, pa);
3420 spin_unlock(&pa->pa_lock);
3421 ac->ac_criteria = 10;
3422 rcu_read_unlock();
3423 return 1;
3425 spin_unlock(&pa->pa_lock);
3427 rcu_read_unlock();
3429 /* can we use group allocation? */
3430 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3431 return 0;
3433 /* inode may have no locality group for some reason */
3434 lg = ac->ac_lg;
3435 if (lg == NULL)
3436 return 0;
3437 order = fls(ac->ac_o_ex.fe_len) - 1;
3438 if (order > PREALLOC_TB_SIZE - 1)
3439 /* The max size of hash table is PREALLOC_TB_SIZE */
3440 order = PREALLOC_TB_SIZE - 1;
3442 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3444 * search for the prealloc space that is having
3445 * minimal distance from the goal block.
3447 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3448 rcu_read_lock();
3449 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3450 pa_inode_list) {
3451 spin_lock(&pa->pa_lock);
3452 if (pa->pa_deleted == 0 &&
3453 pa->pa_free >= ac->ac_o_ex.fe_len) {
3455 cpa = ext4_mb_check_group_pa(goal_block,
3456 pa, cpa);
3458 spin_unlock(&pa->pa_lock);
3460 rcu_read_unlock();
3462 if (cpa) {
3463 ext4_mb_use_group_pa(ac, cpa);
3464 ac->ac_criteria = 20;
3465 return 1;
3467 return 0;
3471 * the function goes through all block freed in the group
3472 * but not yet committed and marks them used in in-core bitmap.
3473 * buddy must be generated from this bitmap
3474 * Need to be called with the ext4 group lock held
3476 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3477 ext4_group_t group)
3479 struct rb_node *n;
3480 struct ext4_group_info *grp;
3481 struct ext4_free_data *entry;
3483 grp = ext4_get_group_info(sb, group);
3484 n = rb_first(&(grp->bb_free_root));
3486 while (n) {
3487 entry = rb_entry(n, struct ext4_free_data, efd_node);
3488 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3489 n = rb_next(n);
3491 return;
3495 * the function goes through all preallocation in this group and marks them
3496 * used in in-core bitmap. buddy must be generated from this bitmap
3497 * Need to be called with ext4 group lock held
3499 static noinline_for_stack
3500 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3501 ext4_group_t group)
3503 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3504 struct ext4_prealloc_space *pa;
3505 struct list_head *cur;
3506 ext4_group_t groupnr;
3507 ext4_grpblk_t start;
3508 int preallocated = 0;
3509 int len;
3511 /* all form of preallocation discards first load group,
3512 * so the only competing code is preallocation use.
3513 * we don't need any locking here
3514 * notice we do NOT ignore preallocations with pa_deleted
3515 * otherwise we could leave used blocks available for
3516 * allocation in buddy when concurrent ext4_mb_put_pa()
3517 * is dropping preallocation
3519 list_for_each(cur, &grp->bb_prealloc_list) {
3520 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3521 spin_lock(&pa->pa_lock);
3522 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3523 &groupnr, &start);
3524 len = pa->pa_len;
3525 spin_unlock(&pa->pa_lock);
3526 if (unlikely(len == 0))
3527 continue;
3528 BUG_ON(groupnr != group);
3529 ext4_set_bits(bitmap, start, len);
3530 preallocated += len;
3532 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3535 static void ext4_mb_pa_callback(struct rcu_head *head)
3537 struct ext4_prealloc_space *pa;
3538 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3540 BUG_ON(atomic_read(&pa->pa_count));
3541 BUG_ON(pa->pa_deleted == 0);
3542 kmem_cache_free(ext4_pspace_cachep, pa);
3546 * drops a reference to preallocated space descriptor
3547 * if this was the last reference and the space is consumed
3549 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3550 struct super_block *sb, struct ext4_prealloc_space *pa)
3552 ext4_group_t grp;
3553 ext4_fsblk_t grp_blk;
3555 /* in this short window concurrent discard can set pa_deleted */
3556 spin_lock(&pa->pa_lock);
3557 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3558 spin_unlock(&pa->pa_lock);
3559 return;
3562 if (pa->pa_deleted == 1) {
3563 spin_unlock(&pa->pa_lock);
3564 return;
3567 pa->pa_deleted = 1;
3568 spin_unlock(&pa->pa_lock);
3570 grp_blk = pa->pa_pstart;
3572 * If doing group-based preallocation, pa_pstart may be in the
3573 * next group when pa is used up
3575 if (pa->pa_type == MB_GROUP_PA)
3576 grp_blk--;
3578 grp = ext4_get_group_number(sb, grp_blk);
3581 * possible race:
3583 * P1 (buddy init) P2 (regular allocation)
3584 * find block B in PA
3585 * copy on-disk bitmap to buddy
3586 * mark B in on-disk bitmap
3587 * drop PA from group
3588 * mark all PAs in buddy
3590 * thus, P1 initializes buddy with B available. to prevent this
3591 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3592 * against that pair
3594 ext4_lock_group(sb, grp);
3595 list_del(&pa->pa_group_list);
3596 ext4_unlock_group(sb, grp);
3598 spin_lock(pa->pa_obj_lock);
3599 list_del_rcu(&pa->pa_inode_list);
3600 spin_unlock(pa->pa_obj_lock);
3602 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3606 * creates new preallocated space for given inode
3608 static noinline_for_stack int
3609 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3611 struct super_block *sb = ac->ac_sb;
3612 struct ext4_sb_info *sbi = EXT4_SB(sb);
3613 struct ext4_prealloc_space *pa;
3614 struct ext4_group_info *grp;
3615 struct ext4_inode_info *ei;
3617 /* preallocate only when found space is larger then requested */
3618 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3619 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3620 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3622 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3623 if (pa == NULL)
3624 return -ENOMEM;
3626 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3627 int winl;
3628 int wins;
3629 int win;
3630 int offs;
3632 /* we can't allocate as much as normalizer wants.
3633 * so, found space must get proper lstart
3634 * to cover original request */
3635 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3636 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3638 /* we're limited by original request in that
3639 * logical block must be covered any way
3640 * winl is window we can move our chunk within */
3641 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3643 /* also, we should cover whole original request */
3644 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3646 /* the smallest one defines real window */
3647 win = min(winl, wins);
3649 offs = ac->ac_o_ex.fe_logical %
3650 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3651 if (offs && offs < win)
3652 win = offs;
3654 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3655 EXT4_NUM_B2C(sbi, win);
3656 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3657 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3660 /* preallocation can change ac_b_ex, thus we store actually
3661 * allocated blocks for history */
3662 ac->ac_f_ex = ac->ac_b_ex;
3664 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3665 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3666 pa->pa_len = ac->ac_b_ex.fe_len;
3667 pa->pa_free = pa->pa_len;
3668 atomic_set(&pa->pa_count, 1);
3669 spin_lock_init(&pa->pa_lock);
3670 INIT_LIST_HEAD(&pa->pa_inode_list);
3671 INIT_LIST_HEAD(&pa->pa_group_list);
3672 pa->pa_deleted = 0;
3673 pa->pa_type = MB_INODE_PA;
3675 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3676 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3677 trace_ext4_mb_new_inode_pa(ac, pa);
3679 ext4_mb_use_inode_pa(ac, pa);
3680 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3682 ei = EXT4_I(ac->ac_inode);
3683 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3685 pa->pa_obj_lock = &ei->i_prealloc_lock;
3686 pa->pa_inode = ac->ac_inode;
3688 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3689 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3690 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3692 spin_lock(pa->pa_obj_lock);
3693 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3694 spin_unlock(pa->pa_obj_lock);
3696 return 0;
3700 * creates new preallocated space for locality group inodes belongs to
3702 static noinline_for_stack int
3703 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3705 struct super_block *sb = ac->ac_sb;
3706 struct ext4_locality_group *lg;
3707 struct ext4_prealloc_space *pa;
3708 struct ext4_group_info *grp;
3710 /* preallocate only when found space is larger then requested */
3711 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3712 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3713 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3715 BUG_ON(ext4_pspace_cachep == NULL);
3716 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3717 if (pa == NULL)
3718 return -ENOMEM;
3720 /* preallocation can change ac_b_ex, thus we store actually
3721 * allocated blocks for history */
3722 ac->ac_f_ex = ac->ac_b_ex;
3724 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3725 pa->pa_lstart = pa->pa_pstart;
3726 pa->pa_len = ac->ac_b_ex.fe_len;
3727 pa->pa_free = pa->pa_len;
3728 atomic_set(&pa->pa_count, 1);
3729 spin_lock_init(&pa->pa_lock);
3730 INIT_LIST_HEAD(&pa->pa_inode_list);
3731 INIT_LIST_HEAD(&pa->pa_group_list);
3732 pa->pa_deleted = 0;
3733 pa->pa_type = MB_GROUP_PA;
3735 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3736 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3737 trace_ext4_mb_new_group_pa(ac, pa);
3739 ext4_mb_use_group_pa(ac, pa);
3740 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3742 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3743 lg = ac->ac_lg;
3744 BUG_ON(lg == NULL);
3746 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3747 pa->pa_inode = NULL;
3749 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3750 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3751 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3754 * We will later add the new pa to the right bucket
3755 * after updating the pa_free in ext4_mb_release_context
3757 return 0;
3760 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3762 int err;
3764 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3765 err = ext4_mb_new_group_pa(ac);
3766 else
3767 err = ext4_mb_new_inode_pa(ac);
3768 return err;
3772 * finds all unused blocks in on-disk bitmap, frees them in
3773 * in-core bitmap and buddy.
3774 * @pa must be unlinked from inode and group lists, so that
3775 * nobody else can find/use it.
3776 * the caller MUST hold group/inode locks.
3777 * TODO: optimize the case when there are no in-core structures yet
3779 static noinline_for_stack int
3780 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3781 struct ext4_prealloc_space *pa)
3783 struct super_block *sb = e4b->bd_sb;
3784 struct ext4_sb_info *sbi = EXT4_SB(sb);
3785 unsigned int end;
3786 unsigned int next;
3787 ext4_group_t group;
3788 ext4_grpblk_t bit;
3789 unsigned long long grp_blk_start;
3790 int err = 0;
3791 int free = 0;
3793 BUG_ON(pa->pa_deleted == 0);
3794 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3795 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3796 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3797 end = bit + pa->pa_len;
3799 while (bit < end) {
3800 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3801 if (bit >= end)
3802 break;
3803 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3804 mb_debug(1, " free preallocated %u/%u in group %u\n",
3805 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3806 (unsigned) next - bit, (unsigned) group);
3807 free += next - bit;
3809 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3810 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3811 EXT4_C2B(sbi, bit)),
3812 next - bit);
3813 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3814 bit = next + 1;
3816 if (free != pa->pa_free) {
3817 ext4_msg(e4b->bd_sb, KERN_CRIT,
3818 "pa %p: logic %lu, phys. %lu, len %lu",
3819 pa, (unsigned long) pa->pa_lstart,
3820 (unsigned long) pa->pa_pstart,
3821 (unsigned long) pa->pa_len);
3822 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3823 free, pa->pa_free);
3825 * pa is already deleted so we use the value obtained
3826 * from the bitmap and continue.
3829 atomic_add(free, &sbi->s_mb_discarded);
3831 return err;
3834 static noinline_for_stack int
3835 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3836 struct ext4_prealloc_space *pa)
3838 struct super_block *sb = e4b->bd_sb;
3839 ext4_group_t group;
3840 ext4_grpblk_t bit;
3842 trace_ext4_mb_release_group_pa(sb, pa);
3843 BUG_ON(pa->pa_deleted == 0);
3844 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3845 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3846 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3847 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3848 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3850 return 0;
3854 * releases all preallocations in given group
3856 * first, we need to decide discard policy:
3857 * - when do we discard
3858 * 1) ENOSPC
3859 * - how many do we discard
3860 * 1) how many requested
3862 static noinline_for_stack int
3863 ext4_mb_discard_group_preallocations(struct super_block *sb,
3864 ext4_group_t group, int needed)
3866 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3867 struct buffer_head *bitmap_bh = NULL;
3868 struct ext4_prealloc_space *pa, *tmp;
3869 struct list_head list;
3870 struct ext4_buddy e4b;
3871 int err;
3872 int busy = 0;
3873 int free = 0;
3875 mb_debug(1, "discard preallocation for group %u\n", group);
3877 if (list_empty(&grp->bb_prealloc_list))
3878 return 0;
3880 bitmap_bh = ext4_read_block_bitmap(sb, group);
3881 if (IS_ERR(bitmap_bh)) {
3882 err = PTR_ERR(bitmap_bh);
3883 ext4_error(sb, "Error %d reading block bitmap for %u",
3884 err, group);
3885 return 0;
3888 err = ext4_mb_load_buddy(sb, group, &e4b);
3889 if (err) {
3890 ext4_error(sb, "Error loading buddy information for %u", group);
3891 put_bh(bitmap_bh);
3892 return 0;
3895 if (needed == 0)
3896 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3898 INIT_LIST_HEAD(&list);
3899 repeat:
3900 ext4_lock_group(sb, group);
3901 list_for_each_entry_safe(pa, tmp,
3902 &grp->bb_prealloc_list, pa_group_list) {
3903 spin_lock(&pa->pa_lock);
3904 if (atomic_read(&pa->pa_count)) {
3905 spin_unlock(&pa->pa_lock);
3906 busy = 1;
3907 continue;
3909 if (pa->pa_deleted) {
3910 spin_unlock(&pa->pa_lock);
3911 continue;
3914 /* seems this one can be freed ... */
3915 pa->pa_deleted = 1;
3917 /* we can trust pa_free ... */
3918 free += pa->pa_free;
3920 spin_unlock(&pa->pa_lock);
3922 list_del(&pa->pa_group_list);
3923 list_add(&pa->u.pa_tmp_list, &list);
3926 /* if we still need more blocks and some PAs were used, try again */
3927 if (free < needed && busy) {
3928 busy = 0;
3929 ext4_unlock_group(sb, group);
3930 cond_resched();
3931 goto repeat;
3934 /* found anything to free? */
3935 if (list_empty(&list)) {
3936 BUG_ON(free != 0);
3937 goto out;
3940 /* now free all selected PAs */
3941 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3943 /* remove from object (inode or locality group) */
3944 spin_lock(pa->pa_obj_lock);
3945 list_del_rcu(&pa->pa_inode_list);
3946 spin_unlock(pa->pa_obj_lock);
3948 if (pa->pa_type == MB_GROUP_PA)
3949 ext4_mb_release_group_pa(&e4b, pa);
3950 else
3951 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3953 list_del(&pa->u.pa_tmp_list);
3954 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3957 out:
3958 ext4_unlock_group(sb, group);
3959 ext4_mb_unload_buddy(&e4b);
3960 put_bh(bitmap_bh);
3961 return free;
3965 * releases all non-used preallocated blocks for given inode
3967 * It's important to discard preallocations under i_data_sem
3968 * We don't want another block to be served from the prealloc
3969 * space when we are discarding the inode prealloc space.
3971 * FIXME!! Make sure it is valid at all the call sites
3973 void ext4_discard_preallocations(struct inode *inode)
3975 struct ext4_inode_info *ei = EXT4_I(inode);
3976 struct super_block *sb = inode->i_sb;
3977 struct buffer_head *bitmap_bh = NULL;
3978 struct ext4_prealloc_space *pa, *tmp;
3979 ext4_group_t group = 0;
3980 struct list_head list;
3981 struct ext4_buddy e4b;
3982 int err;
3984 if (!S_ISREG(inode->i_mode)) {
3985 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3986 return;
3989 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3990 trace_ext4_discard_preallocations(inode);
3992 INIT_LIST_HEAD(&list);
3994 repeat:
3995 /* first, collect all pa's in the inode */
3996 spin_lock(&ei->i_prealloc_lock);
3997 while (!list_empty(&ei->i_prealloc_list)) {
3998 pa = list_entry(ei->i_prealloc_list.next,
3999 struct ext4_prealloc_space, pa_inode_list);
4000 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4001 spin_lock(&pa->pa_lock);
4002 if (atomic_read(&pa->pa_count)) {
4003 /* this shouldn't happen often - nobody should
4004 * use preallocation while we're discarding it */
4005 spin_unlock(&pa->pa_lock);
4006 spin_unlock(&ei->i_prealloc_lock);
4007 ext4_msg(sb, KERN_ERR,
4008 "uh-oh! used pa while discarding");
4009 WARN_ON(1);
4010 schedule_timeout_uninterruptible(HZ);
4011 goto repeat;
4014 if (pa->pa_deleted == 0) {
4015 pa->pa_deleted = 1;
4016 spin_unlock(&pa->pa_lock);
4017 list_del_rcu(&pa->pa_inode_list);
4018 list_add(&pa->u.pa_tmp_list, &list);
4019 continue;
4022 /* someone is deleting pa right now */
4023 spin_unlock(&pa->pa_lock);
4024 spin_unlock(&ei->i_prealloc_lock);
4026 /* we have to wait here because pa_deleted
4027 * doesn't mean pa is already unlinked from
4028 * the list. as we might be called from
4029 * ->clear_inode() the inode will get freed
4030 * and concurrent thread which is unlinking
4031 * pa from inode's list may access already
4032 * freed memory, bad-bad-bad */
4034 /* XXX: if this happens too often, we can
4035 * add a flag to force wait only in case
4036 * of ->clear_inode(), but not in case of
4037 * regular truncate */
4038 schedule_timeout_uninterruptible(HZ);
4039 goto repeat;
4041 spin_unlock(&ei->i_prealloc_lock);
4043 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4044 BUG_ON(pa->pa_type != MB_INODE_PA);
4045 group = ext4_get_group_number(sb, pa->pa_pstart);
4047 err = ext4_mb_load_buddy(sb, group, &e4b);
4048 if (err) {
4049 ext4_error(sb, "Error loading buddy information for %u",
4050 group);
4051 continue;
4054 bitmap_bh = ext4_read_block_bitmap(sb, group);
4055 if (IS_ERR(bitmap_bh)) {
4056 err = PTR_ERR(bitmap_bh);
4057 ext4_error(sb, "Error %d reading block bitmap for %u",
4058 err, group);
4059 ext4_mb_unload_buddy(&e4b);
4060 continue;
4063 ext4_lock_group(sb, group);
4064 list_del(&pa->pa_group_list);
4065 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4066 ext4_unlock_group(sb, group);
4068 ext4_mb_unload_buddy(&e4b);
4069 put_bh(bitmap_bh);
4071 list_del(&pa->u.pa_tmp_list);
4072 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4076 #ifdef CONFIG_EXT4_DEBUG
4077 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4079 struct super_block *sb = ac->ac_sb;
4080 ext4_group_t ngroups, i;
4082 if (!ext4_mballoc_debug ||
4083 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4084 return;
4086 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4087 " Allocation context details:");
4088 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4089 ac->ac_status, ac->ac_flags);
4090 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4091 "goal %lu/%lu/%lu@%lu, "
4092 "best %lu/%lu/%lu@%lu cr %d",
4093 (unsigned long)ac->ac_o_ex.fe_group,
4094 (unsigned long)ac->ac_o_ex.fe_start,
4095 (unsigned long)ac->ac_o_ex.fe_len,
4096 (unsigned long)ac->ac_o_ex.fe_logical,
4097 (unsigned long)ac->ac_g_ex.fe_group,
4098 (unsigned long)ac->ac_g_ex.fe_start,
4099 (unsigned long)ac->ac_g_ex.fe_len,
4100 (unsigned long)ac->ac_g_ex.fe_logical,
4101 (unsigned long)ac->ac_b_ex.fe_group,
4102 (unsigned long)ac->ac_b_ex.fe_start,
4103 (unsigned long)ac->ac_b_ex.fe_len,
4104 (unsigned long)ac->ac_b_ex.fe_logical,
4105 (int)ac->ac_criteria);
4106 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4107 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4108 ngroups = ext4_get_groups_count(sb);
4109 for (i = 0; i < ngroups; i++) {
4110 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4111 struct ext4_prealloc_space *pa;
4112 ext4_grpblk_t start;
4113 struct list_head *cur;
4114 ext4_lock_group(sb, i);
4115 list_for_each(cur, &grp->bb_prealloc_list) {
4116 pa = list_entry(cur, struct ext4_prealloc_space,
4117 pa_group_list);
4118 spin_lock(&pa->pa_lock);
4119 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4120 NULL, &start);
4121 spin_unlock(&pa->pa_lock);
4122 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4123 start, pa->pa_len);
4125 ext4_unlock_group(sb, i);
4127 if (grp->bb_free == 0)
4128 continue;
4129 printk(KERN_ERR "%u: %d/%d \n",
4130 i, grp->bb_free, grp->bb_fragments);
4132 printk(KERN_ERR "\n");
4134 #else
4135 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4137 return;
4139 #endif
4142 * We use locality group preallocation for small size file. The size of the
4143 * file is determined by the current size or the resulting size after
4144 * allocation which ever is larger
4146 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4148 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4150 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4151 int bsbits = ac->ac_sb->s_blocksize_bits;
4152 loff_t size, isize;
4154 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4155 return;
4157 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4158 return;
4160 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4161 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4162 >> bsbits;
4164 if ((size == isize) &&
4165 !ext4_fs_is_busy(sbi) &&
4166 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4167 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4168 return;
4171 if (sbi->s_mb_group_prealloc <= 0) {
4172 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4173 return;
4176 /* don't use group allocation for large files */
4177 size = max(size, isize);
4178 if (size > sbi->s_mb_stream_request) {
4179 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4180 return;
4183 BUG_ON(ac->ac_lg != NULL);
4185 * locality group prealloc space are per cpu. The reason for having
4186 * per cpu locality group is to reduce the contention between block
4187 * request from multiple CPUs.
4189 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4191 /* we're going to use group allocation */
4192 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4194 /* serialize all allocations in the group */
4195 mutex_lock(&ac->ac_lg->lg_mutex);
4198 static noinline_for_stack int
4199 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4200 struct ext4_allocation_request *ar)
4202 struct super_block *sb = ar->inode->i_sb;
4203 struct ext4_sb_info *sbi = EXT4_SB(sb);
4204 struct ext4_super_block *es = sbi->s_es;
4205 ext4_group_t group;
4206 unsigned int len;
4207 ext4_fsblk_t goal;
4208 ext4_grpblk_t block;
4210 /* we can't allocate > group size */
4211 len = ar->len;
4213 /* just a dirty hack to filter too big requests */
4214 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4215 len = EXT4_CLUSTERS_PER_GROUP(sb);
4217 /* start searching from the goal */
4218 goal = ar->goal;
4219 if (goal < le32_to_cpu(es->s_first_data_block) ||
4220 goal >= ext4_blocks_count(es))
4221 goal = le32_to_cpu(es->s_first_data_block);
4222 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4224 /* set up allocation goals */
4225 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4226 ac->ac_status = AC_STATUS_CONTINUE;
4227 ac->ac_sb = sb;
4228 ac->ac_inode = ar->inode;
4229 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4230 ac->ac_o_ex.fe_group = group;
4231 ac->ac_o_ex.fe_start = block;
4232 ac->ac_o_ex.fe_len = len;
4233 ac->ac_g_ex = ac->ac_o_ex;
4234 ac->ac_flags = ar->flags;
4236 /* we have to define context: we'll we work with a file or
4237 * locality group. this is a policy, actually */
4238 ext4_mb_group_or_file(ac);
4240 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4241 "left: %u/%u, right %u/%u to %swritable\n",
4242 (unsigned) ar->len, (unsigned) ar->logical,
4243 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4244 (unsigned) ar->lleft, (unsigned) ar->pleft,
4245 (unsigned) ar->lright, (unsigned) ar->pright,
4246 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4247 return 0;
4251 static noinline_for_stack void
4252 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4253 struct ext4_locality_group *lg,
4254 int order, int total_entries)
4256 ext4_group_t group = 0;
4257 struct ext4_buddy e4b;
4258 struct list_head discard_list;
4259 struct ext4_prealloc_space *pa, *tmp;
4261 mb_debug(1, "discard locality group preallocation\n");
4263 INIT_LIST_HEAD(&discard_list);
4265 spin_lock(&lg->lg_prealloc_lock);
4266 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4267 pa_inode_list) {
4268 spin_lock(&pa->pa_lock);
4269 if (atomic_read(&pa->pa_count)) {
4271 * This is the pa that we just used
4272 * for block allocation. So don't
4273 * free that
4275 spin_unlock(&pa->pa_lock);
4276 continue;
4278 if (pa->pa_deleted) {
4279 spin_unlock(&pa->pa_lock);
4280 continue;
4282 /* only lg prealloc space */
4283 BUG_ON(pa->pa_type != MB_GROUP_PA);
4285 /* seems this one can be freed ... */
4286 pa->pa_deleted = 1;
4287 spin_unlock(&pa->pa_lock);
4289 list_del_rcu(&pa->pa_inode_list);
4290 list_add(&pa->u.pa_tmp_list, &discard_list);
4292 total_entries--;
4293 if (total_entries <= 5) {
4295 * we want to keep only 5 entries
4296 * allowing it to grow to 8. This
4297 * mak sure we don't call discard
4298 * soon for this list.
4300 break;
4303 spin_unlock(&lg->lg_prealloc_lock);
4305 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4307 group = ext4_get_group_number(sb, pa->pa_pstart);
4308 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4309 ext4_error(sb, "Error loading buddy information for %u",
4310 group);
4311 continue;
4313 ext4_lock_group(sb, group);
4314 list_del(&pa->pa_group_list);
4315 ext4_mb_release_group_pa(&e4b, pa);
4316 ext4_unlock_group(sb, group);
4318 ext4_mb_unload_buddy(&e4b);
4319 list_del(&pa->u.pa_tmp_list);
4320 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4325 * We have incremented pa_count. So it cannot be freed at this
4326 * point. Also we hold lg_mutex. So no parallel allocation is
4327 * possible from this lg. That means pa_free cannot be updated.
4329 * A parallel ext4_mb_discard_group_preallocations is possible.
4330 * which can cause the lg_prealloc_list to be updated.
4333 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4335 int order, added = 0, lg_prealloc_count = 1;
4336 struct super_block *sb = ac->ac_sb;
4337 struct ext4_locality_group *lg = ac->ac_lg;
4338 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4340 order = fls(pa->pa_free) - 1;
4341 if (order > PREALLOC_TB_SIZE - 1)
4342 /* The max size of hash table is PREALLOC_TB_SIZE */
4343 order = PREALLOC_TB_SIZE - 1;
4344 /* Add the prealloc space to lg */
4345 spin_lock(&lg->lg_prealloc_lock);
4346 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4347 pa_inode_list) {
4348 spin_lock(&tmp_pa->pa_lock);
4349 if (tmp_pa->pa_deleted) {
4350 spin_unlock(&tmp_pa->pa_lock);
4351 continue;
4353 if (!added && pa->pa_free < tmp_pa->pa_free) {
4354 /* Add to the tail of the previous entry */
4355 list_add_tail_rcu(&pa->pa_inode_list,
4356 &tmp_pa->pa_inode_list);
4357 added = 1;
4359 * we want to count the total
4360 * number of entries in the list
4363 spin_unlock(&tmp_pa->pa_lock);
4364 lg_prealloc_count++;
4366 if (!added)
4367 list_add_tail_rcu(&pa->pa_inode_list,
4368 &lg->lg_prealloc_list[order]);
4369 spin_unlock(&lg->lg_prealloc_lock);
4371 /* Now trim the list to be not more than 8 elements */
4372 if (lg_prealloc_count > 8) {
4373 ext4_mb_discard_lg_preallocations(sb, lg,
4374 order, lg_prealloc_count);
4375 return;
4377 return ;
4381 * release all resource we used in allocation
4383 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4385 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4386 struct ext4_prealloc_space *pa = ac->ac_pa;
4387 if (pa) {
4388 if (pa->pa_type == MB_GROUP_PA) {
4389 /* see comment in ext4_mb_use_group_pa() */
4390 spin_lock(&pa->pa_lock);
4391 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4392 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4393 pa->pa_free -= ac->ac_b_ex.fe_len;
4394 pa->pa_len -= ac->ac_b_ex.fe_len;
4395 spin_unlock(&pa->pa_lock);
4398 if (pa) {
4400 * We want to add the pa to the right bucket.
4401 * Remove it from the list and while adding
4402 * make sure the list to which we are adding
4403 * doesn't grow big.
4405 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4406 spin_lock(pa->pa_obj_lock);
4407 list_del_rcu(&pa->pa_inode_list);
4408 spin_unlock(pa->pa_obj_lock);
4409 ext4_mb_add_n_trim(ac);
4411 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4413 if (ac->ac_bitmap_page)
4414 put_page(ac->ac_bitmap_page);
4415 if (ac->ac_buddy_page)
4416 put_page(ac->ac_buddy_page);
4417 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4418 mutex_unlock(&ac->ac_lg->lg_mutex);
4419 ext4_mb_collect_stats(ac);
4420 return 0;
4423 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4425 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4426 int ret;
4427 int freed = 0;
4429 trace_ext4_mb_discard_preallocations(sb, needed);
4430 for (i = 0; i < ngroups && needed > 0; i++) {
4431 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4432 freed += ret;
4433 needed -= ret;
4436 return freed;
4440 * Main entry point into mballoc to allocate blocks
4441 * it tries to use preallocation first, then falls back
4442 * to usual allocation
4444 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4445 struct ext4_allocation_request *ar, int *errp)
4447 int freed;
4448 struct ext4_allocation_context *ac = NULL;
4449 struct ext4_sb_info *sbi;
4450 struct super_block *sb;
4451 ext4_fsblk_t block = 0;
4452 unsigned int inquota = 0;
4453 unsigned int reserv_clstrs = 0;
4455 might_sleep();
4456 sb = ar->inode->i_sb;
4457 sbi = EXT4_SB(sb);
4459 trace_ext4_request_blocks(ar);
4461 /* Allow to use superuser reservation for quota file */
4462 if (IS_NOQUOTA(ar->inode))
4463 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4465 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4466 /* Without delayed allocation we need to verify
4467 * there is enough free blocks to do block allocation
4468 * and verify allocation doesn't exceed the quota limits.
4470 while (ar->len &&
4471 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4473 /* let others to free the space */
4474 cond_resched();
4475 ar->len = ar->len >> 1;
4477 if (!ar->len) {
4478 *errp = -ENOSPC;
4479 return 0;
4481 reserv_clstrs = ar->len;
4482 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4483 dquot_alloc_block_nofail(ar->inode,
4484 EXT4_C2B(sbi, ar->len));
4485 } else {
4486 while (ar->len &&
4487 dquot_alloc_block(ar->inode,
4488 EXT4_C2B(sbi, ar->len))) {
4490 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4491 ar->len--;
4494 inquota = ar->len;
4495 if (ar->len == 0) {
4496 *errp = -EDQUOT;
4497 goto out;
4501 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4502 if (!ac) {
4503 ar->len = 0;
4504 *errp = -ENOMEM;
4505 goto out;
4508 *errp = ext4_mb_initialize_context(ac, ar);
4509 if (*errp) {
4510 ar->len = 0;
4511 goto out;
4514 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4515 if (!ext4_mb_use_preallocated(ac)) {
4516 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4517 ext4_mb_normalize_request(ac, ar);
4518 repeat:
4519 /* allocate space in core */
4520 *errp = ext4_mb_regular_allocator(ac);
4521 if (*errp)
4522 goto discard_and_exit;
4524 /* as we've just preallocated more space than
4525 * user requested originally, we store allocated
4526 * space in a special descriptor */
4527 if (ac->ac_status == AC_STATUS_FOUND &&
4528 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4529 *errp = ext4_mb_new_preallocation(ac);
4530 if (*errp) {
4531 discard_and_exit:
4532 ext4_discard_allocated_blocks(ac);
4533 goto errout;
4536 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4537 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4538 if (*errp) {
4539 ext4_discard_allocated_blocks(ac);
4540 goto errout;
4541 } else {
4542 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4543 ar->len = ac->ac_b_ex.fe_len;
4545 } else {
4546 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4547 if (freed)
4548 goto repeat;
4549 *errp = -ENOSPC;
4552 errout:
4553 if (*errp) {
4554 ac->ac_b_ex.fe_len = 0;
4555 ar->len = 0;
4556 ext4_mb_show_ac(ac);
4558 ext4_mb_release_context(ac);
4559 out:
4560 if (ac)
4561 kmem_cache_free(ext4_ac_cachep, ac);
4562 if (inquota && ar->len < inquota)
4563 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4564 if (!ar->len) {
4565 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4566 /* release all the reserved blocks if non delalloc */
4567 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4568 reserv_clstrs);
4571 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4573 return block;
4577 * We can merge two free data extents only if the physical blocks
4578 * are contiguous, AND the extents were freed by the same transaction,
4579 * AND the blocks are associated with the same group.
4581 static int can_merge(struct ext4_free_data *entry1,
4582 struct ext4_free_data *entry2)
4584 if ((entry1->efd_tid == entry2->efd_tid) &&
4585 (entry1->efd_group == entry2->efd_group) &&
4586 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4587 return 1;
4588 return 0;
4591 static noinline_for_stack int
4592 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4593 struct ext4_free_data *new_entry)
4595 ext4_group_t group = e4b->bd_group;
4596 ext4_grpblk_t cluster;
4597 ext4_grpblk_t clusters = new_entry->efd_count;
4598 struct ext4_free_data *entry;
4599 struct ext4_group_info *db = e4b->bd_info;
4600 struct super_block *sb = e4b->bd_sb;
4601 struct ext4_sb_info *sbi = EXT4_SB(sb);
4602 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4603 struct rb_node *parent = NULL, *new_node;
4605 BUG_ON(!ext4_handle_valid(handle));
4606 BUG_ON(e4b->bd_bitmap_page == NULL);
4607 BUG_ON(e4b->bd_buddy_page == NULL);
4609 new_node = &new_entry->efd_node;
4610 cluster = new_entry->efd_start_cluster;
4612 if (!*n) {
4613 /* first free block exent. We need to
4614 protect buddy cache from being freed,
4615 * otherwise we'll refresh it from
4616 * on-disk bitmap and lose not-yet-available
4617 * blocks */
4618 get_page(e4b->bd_buddy_page);
4619 get_page(e4b->bd_bitmap_page);
4621 while (*n) {
4622 parent = *n;
4623 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4624 if (cluster < entry->efd_start_cluster)
4625 n = &(*n)->rb_left;
4626 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4627 n = &(*n)->rb_right;
4628 else {
4629 ext4_grp_locked_error(sb, group, 0,
4630 ext4_group_first_block_no(sb, group) +
4631 EXT4_C2B(sbi, cluster),
4632 "Block already on to-be-freed list");
4633 return 0;
4637 rb_link_node(new_node, parent, n);
4638 rb_insert_color(new_node, &db->bb_free_root);
4640 /* Now try to see the extent can be merged to left and right */
4641 node = rb_prev(new_node);
4642 if (node) {
4643 entry = rb_entry(node, struct ext4_free_data, efd_node);
4644 if (can_merge(entry, new_entry) &&
4645 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4646 new_entry->efd_start_cluster = entry->efd_start_cluster;
4647 new_entry->efd_count += entry->efd_count;
4648 rb_erase(node, &(db->bb_free_root));
4649 kmem_cache_free(ext4_free_data_cachep, entry);
4653 node = rb_next(new_node);
4654 if (node) {
4655 entry = rb_entry(node, struct ext4_free_data, efd_node);
4656 if (can_merge(new_entry, entry) &&
4657 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4658 new_entry->efd_count += entry->efd_count;
4659 rb_erase(node, &(db->bb_free_root));
4660 kmem_cache_free(ext4_free_data_cachep, entry);
4663 /* Add the extent to transaction's private list */
4664 new_entry->efd_jce.jce_func = ext4_free_data_callback;
4665 spin_lock(&sbi->s_md_lock);
4666 _ext4_journal_callback_add(handle, &new_entry->efd_jce);
4667 sbi->s_mb_free_pending += clusters;
4668 spin_unlock(&sbi->s_md_lock);
4669 return 0;
4673 * ext4_free_blocks() -- Free given blocks and update quota
4674 * @handle: handle for this transaction
4675 * @inode: inode
4676 * @block: start physical block to free
4677 * @count: number of blocks to count
4678 * @flags: flags used by ext4_free_blocks
4680 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4681 struct buffer_head *bh, ext4_fsblk_t block,
4682 unsigned long count, int flags)
4684 struct buffer_head *bitmap_bh = NULL;
4685 struct super_block *sb = inode->i_sb;
4686 struct ext4_group_desc *gdp;
4687 unsigned int overflow;
4688 ext4_grpblk_t bit;
4689 struct buffer_head *gd_bh;
4690 ext4_group_t block_group;
4691 struct ext4_sb_info *sbi;
4692 struct ext4_buddy e4b;
4693 unsigned int count_clusters;
4694 int err = 0;
4695 int ret;
4697 might_sleep();
4698 if (bh) {
4699 if (block)
4700 BUG_ON(block != bh->b_blocknr);
4701 else
4702 block = bh->b_blocknr;
4705 sbi = EXT4_SB(sb);
4706 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4707 !ext4_data_block_valid(sbi, block, count)) {
4708 ext4_error(sb, "Freeing blocks not in datazone - "
4709 "block = %llu, count = %lu", block, count);
4710 goto error_return;
4713 ext4_debug("freeing block %llu\n", block);
4714 trace_ext4_free_blocks(inode, block, count, flags);
4716 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4717 BUG_ON(count > 1);
4719 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4720 inode, bh, block);
4724 * If the extent to be freed does not begin on a cluster
4725 * boundary, we need to deal with partial clusters at the
4726 * beginning and end of the extent. Normally we will free
4727 * blocks at the beginning or the end unless we are explicitly
4728 * requested to avoid doing so.
4730 overflow = EXT4_PBLK_COFF(sbi, block);
4731 if (overflow) {
4732 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4733 overflow = sbi->s_cluster_ratio - overflow;
4734 block += overflow;
4735 if (count > overflow)
4736 count -= overflow;
4737 else
4738 return;
4739 } else {
4740 block -= overflow;
4741 count += overflow;
4744 overflow = EXT4_LBLK_COFF(sbi, count);
4745 if (overflow) {
4746 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4747 if (count > overflow)
4748 count -= overflow;
4749 else
4750 return;
4751 } else
4752 count += sbi->s_cluster_ratio - overflow;
4755 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4756 int i;
4757 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4759 for (i = 0; i < count; i++) {
4760 cond_resched();
4761 if (is_metadata)
4762 bh = sb_find_get_block(inode->i_sb, block + i);
4763 ext4_forget(handle, is_metadata, inode, bh, block + i);
4767 do_more:
4768 overflow = 0;
4769 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4771 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4772 ext4_get_group_info(sb, block_group))))
4773 return;
4776 * Check to see if we are freeing blocks across a group
4777 * boundary.
4779 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4780 overflow = EXT4_C2B(sbi, bit) + count -
4781 EXT4_BLOCKS_PER_GROUP(sb);
4782 count -= overflow;
4784 count_clusters = EXT4_NUM_B2C(sbi, count);
4785 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4786 if (IS_ERR(bitmap_bh)) {
4787 err = PTR_ERR(bitmap_bh);
4788 bitmap_bh = NULL;
4789 goto error_return;
4791 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4792 if (!gdp) {
4793 err = -EIO;
4794 goto error_return;
4797 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4798 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4799 in_range(block, ext4_inode_table(sb, gdp),
4800 EXT4_SB(sb)->s_itb_per_group) ||
4801 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4802 EXT4_SB(sb)->s_itb_per_group)) {
4804 ext4_error(sb, "Freeing blocks in system zone - "
4805 "Block = %llu, count = %lu", block, count);
4806 /* err = 0. ext4_std_error should be a no op */
4807 goto error_return;
4810 BUFFER_TRACE(bitmap_bh, "getting write access");
4811 err = ext4_journal_get_write_access(handle, bitmap_bh);
4812 if (err)
4813 goto error_return;
4816 * We are about to modify some metadata. Call the journal APIs
4817 * to unshare ->b_data if a currently-committing transaction is
4818 * using it
4820 BUFFER_TRACE(gd_bh, "get_write_access");
4821 err = ext4_journal_get_write_access(handle, gd_bh);
4822 if (err)
4823 goto error_return;
4824 #ifdef AGGRESSIVE_CHECK
4826 int i;
4827 for (i = 0; i < count_clusters; i++)
4828 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4830 #endif
4831 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4833 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4834 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4835 GFP_NOFS|__GFP_NOFAIL);
4836 if (err)
4837 goto error_return;
4840 * We need to make sure we don't reuse the freed block until after the
4841 * transaction is committed. We make an exception if the inode is to be
4842 * written in writeback mode since writeback mode has weak data
4843 * consistency guarantees.
4845 if (ext4_handle_valid(handle) &&
4846 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4847 !ext4_should_writeback_data(inode))) {
4848 struct ext4_free_data *new_entry;
4850 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4851 * to fail.
4853 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4854 GFP_NOFS|__GFP_NOFAIL);
4855 new_entry->efd_start_cluster = bit;
4856 new_entry->efd_group = block_group;
4857 new_entry->efd_count = count_clusters;
4858 new_entry->efd_tid = handle->h_transaction->t_tid;
4860 ext4_lock_group(sb, block_group);
4861 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4862 ext4_mb_free_metadata(handle, &e4b, new_entry);
4863 } else {
4864 /* need to update group_info->bb_free and bitmap
4865 * with group lock held. generate_buddy look at
4866 * them with group lock_held
4868 if (test_opt(sb, DISCARD)) {
4869 err = ext4_issue_discard(sb, block_group, bit, count);
4870 if (err && err != -EOPNOTSUPP)
4871 ext4_msg(sb, KERN_WARNING, "discard request in"
4872 " group:%d block:%d count:%lu failed"
4873 " with %d", block_group, bit, count,
4874 err);
4875 } else
4876 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4878 ext4_lock_group(sb, block_group);
4879 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4880 mb_free_blocks(inode, &e4b, bit, count_clusters);
4883 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4884 ext4_free_group_clusters_set(sb, gdp, ret);
4885 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4886 ext4_group_desc_csum_set(sb, block_group, gdp);
4887 ext4_unlock_group(sb, block_group);
4889 if (sbi->s_log_groups_per_flex) {
4890 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4891 atomic64_add(count_clusters,
4892 &sbi->s_flex_groups[flex_group].free_clusters);
4895 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4896 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4897 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4899 ext4_mb_unload_buddy(&e4b);
4901 /* We dirtied the bitmap block */
4902 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4903 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4905 /* And the group descriptor block */
4906 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4907 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4908 if (!err)
4909 err = ret;
4911 if (overflow && !err) {
4912 block += count;
4913 count = overflow;
4914 put_bh(bitmap_bh);
4915 goto do_more;
4917 error_return:
4918 brelse(bitmap_bh);
4919 ext4_std_error(sb, err);
4920 return;
4924 * ext4_group_add_blocks() -- Add given blocks to an existing group
4925 * @handle: handle to this transaction
4926 * @sb: super block
4927 * @block: start physical block to add to the block group
4928 * @count: number of blocks to free
4930 * This marks the blocks as free in the bitmap and buddy.
4932 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4933 ext4_fsblk_t block, unsigned long count)
4935 struct buffer_head *bitmap_bh = NULL;
4936 struct buffer_head *gd_bh;
4937 ext4_group_t block_group;
4938 ext4_grpblk_t bit;
4939 unsigned int i;
4940 struct ext4_group_desc *desc;
4941 struct ext4_sb_info *sbi = EXT4_SB(sb);
4942 struct ext4_buddy e4b;
4943 int err = 0, ret, blk_free_count;
4944 ext4_grpblk_t blocks_freed;
4946 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4948 if (count == 0)
4949 return 0;
4951 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4953 * Check to see if we are freeing blocks across a group
4954 * boundary.
4956 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4957 ext4_warning(sb, "too much blocks added to group %u",
4958 block_group);
4959 err = -EINVAL;
4960 goto error_return;
4963 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4964 if (IS_ERR(bitmap_bh)) {
4965 err = PTR_ERR(bitmap_bh);
4966 bitmap_bh = NULL;
4967 goto error_return;
4970 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4971 if (!desc) {
4972 err = -EIO;
4973 goto error_return;
4976 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4977 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4978 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4979 in_range(block + count - 1, ext4_inode_table(sb, desc),
4980 sbi->s_itb_per_group)) {
4981 ext4_error(sb, "Adding blocks in system zones - "
4982 "Block = %llu, count = %lu",
4983 block, count);
4984 err = -EINVAL;
4985 goto error_return;
4988 BUFFER_TRACE(bitmap_bh, "getting write access");
4989 err = ext4_journal_get_write_access(handle, bitmap_bh);
4990 if (err)
4991 goto error_return;
4994 * We are about to modify some metadata. Call the journal APIs
4995 * to unshare ->b_data if a currently-committing transaction is
4996 * using it
4998 BUFFER_TRACE(gd_bh, "get_write_access");
4999 err = ext4_journal_get_write_access(handle, gd_bh);
5000 if (err)
5001 goto error_return;
5003 for (i = 0, blocks_freed = 0; i < count; i++) {
5004 BUFFER_TRACE(bitmap_bh, "clear bit");
5005 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5006 ext4_error(sb, "bit already cleared for block %llu",
5007 (ext4_fsblk_t)(block + i));
5008 BUFFER_TRACE(bitmap_bh, "bit already cleared");
5009 } else {
5010 blocks_freed++;
5014 err = ext4_mb_load_buddy(sb, block_group, &e4b);
5015 if (err)
5016 goto error_return;
5019 * need to update group_info->bb_free and bitmap
5020 * with group lock held. generate_buddy look at
5021 * them with group lock_held
5023 ext4_lock_group(sb, block_group);
5024 mb_clear_bits(bitmap_bh->b_data, bit, count);
5025 mb_free_blocks(NULL, &e4b, bit, count);
5026 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5027 ext4_free_group_clusters_set(sb, desc, blk_free_count);
5028 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5029 ext4_group_desc_csum_set(sb, block_group, desc);
5030 ext4_unlock_group(sb, block_group);
5031 percpu_counter_add(&sbi->s_freeclusters_counter,
5032 EXT4_NUM_B2C(sbi, blocks_freed));
5034 if (sbi->s_log_groups_per_flex) {
5035 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5036 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5037 &sbi->s_flex_groups[flex_group].free_clusters);
5040 ext4_mb_unload_buddy(&e4b);
5042 /* We dirtied the bitmap block */
5043 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5044 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5046 /* And the group descriptor block */
5047 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5048 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5049 if (!err)
5050 err = ret;
5052 error_return:
5053 brelse(bitmap_bh);
5054 ext4_std_error(sb, err);
5055 return err;
5059 * ext4_trim_extent -- function to TRIM one single free extent in the group
5060 * @sb: super block for the file system
5061 * @start: starting block of the free extent in the alloc. group
5062 * @count: number of blocks to TRIM
5063 * @group: alloc. group we are working with
5064 * @e4b: ext4 buddy for the group
5066 * Trim "count" blocks starting at "start" in the "group". To assure that no
5067 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5068 * be called with under the group lock.
5070 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5071 ext4_group_t group, struct ext4_buddy *e4b)
5072 __releases(bitlock)
5073 __acquires(bitlock)
5075 struct ext4_free_extent ex;
5076 int ret = 0;
5078 trace_ext4_trim_extent(sb, group, start, count);
5080 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5082 ex.fe_start = start;
5083 ex.fe_group = group;
5084 ex.fe_len = count;
5087 * Mark blocks used, so no one can reuse them while
5088 * being trimmed.
5090 mb_mark_used(e4b, &ex);
5091 ext4_unlock_group(sb, group);
5092 ret = ext4_issue_discard(sb, group, start, count);
5093 ext4_lock_group(sb, group);
5094 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5095 return ret;
5099 * ext4_trim_all_free -- function to trim all free space in alloc. group
5100 * @sb: super block for file system
5101 * @group: group to be trimmed
5102 * @start: first group block to examine
5103 * @max: last group block to examine
5104 * @minblocks: minimum extent block count
5106 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5107 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5108 * the extent.
5111 * ext4_trim_all_free walks through group's block bitmap searching for free
5112 * extents. When the free extent is found, mark it as used in group buddy
5113 * bitmap. Then issue a TRIM command on this extent and free the extent in
5114 * the group buddy bitmap. This is done until whole group is scanned.
5116 static ext4_grpblk_t
5117 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5118 ext4_grpblk_t start, ext4_grpblk_t max,
5119 ext4_grpblk_t minblocks)
5121 void *bitmap;
5122 ext4_grpblk_t next, count = 0, free_count = 0;
5123 struct ext4_buddy e4b;
5124 int ret = 0;
5126 trace_ext4_trim_all_free(sb, group, start, max);
5128 ret = ext4_mb_load_buddy(sb, group, &e4b);
5129 if (ret) {
5130 ext4_error(sb, "Error in loading buddy "
5131 "information for %u", group);
5132 return ret;
5134 bitmap = e4b.bd_bitmap;
5136 ext4_lock_group(sb, group);
5137 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5138 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5139 goto out;
5141 start = (e4b.bd_info->bb_first_free > start) ?
5142 e4b.bd_info->bb_first_free : start;
5144 while (start <= max) {
5145 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5146 if (start > max)
5147 break;
5148 next = mb_find_next_bit(bitmap, max + 1, start);
5150 if ((next - start) >= minblocks) {
5151 ret = ext4_trim_extent(sb, start,
5152 next - start, group, &e4b);
5153 if (ret && ret != -EOPNOTSUPP)
5154 break;
5155 ret = 0;
5156 count += next - start;
5158 free_count += next - start;
5159 start = next + 1;
5161 if (fatal_signal_pending(current)) {
5162 count = -ERESTARTSYS;
5163 break;
5166 if (need_resched()) {
5167 ext4_unlock_group(sb, group);
5168 cond_resched();
5169 ext4_lock_group(sb, group);
5172 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5173 break;
5176 if (!ret) {
5177 ret = count;
5178 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5180 out:
5181 ext4_unlock_group(sb, group);
5182 ext4_mb_unload_buddy(&e4b);
5184 ext4_debug("trimmed %d blocks in the group %d\n",
5185 count, group);
5187 return ret;
5191 * ext4_trim_fs() -- trim ioctl handle function
5192 * @sb: superblock for filesystem
5193 * @range: fstrim_range structure
5195 * start: First Byte to trim
5196 * len: number of Bytes to trim from start
5197 * minlen: minimum extent length in Bytes
5198 * ext4_trim_fs goes through all allocation groups containing Bytes from
5199 * start to start+len. For each such a group ext4_trim_all_free function
5200 * is invoked to trim all free space.
5202 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5204 struct ext4_group_info *grp;
5205 ext4_group_t group, first_group, last_group;
5206 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5207 uint64_t start, end, minlen, trimmed = 0;
5208 ext4_fsblk_t first_data_blk =
5209 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5210 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5211 int ret = 0;
5213 start = range->start >> sb->s_blocksize_bits;
5214 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5215 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5216 range->minlen >> sb->s_blocksize_bits);
5218 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5219 start >= max_blks ||
5220 range->len < sb->s_blocksize)
5221 return -EINVAL;
5222 if (end >= max_blks)
5223 end = max_blks - 1;
5224 if (end <= first_data_blk)
5225 goto out;
5226 if (start < first_data_blk)
5227 start = first_data_blk;
5229 /* Determine first and last group to examine based on start and end */
5230 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5231 &first_group, &first_cluster);
5232 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5233 &last_group, &last_cluster);
5235 /* end now represents the last cluster to discard in this group */
5236 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5238 for (group = first_group; group <= last_group; group++) {
5239 grp = ext4_get_group_info(sb, group);
5240 /* We only do this if the grp has never been initialized */
5241 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5242 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5243 if (ret)
5244 break;
5248 * For all the groups except the last one, last cluster will
5249 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5250 * change it for the last group, note that last_cluster is
5251 * already computed earlier by ext4_get_group_no_and_offset()
5253 if (group == last_group)
5254 end = last_cluster;
5256 if (grp->bb_free >= minlen) {
5257 cnt = ext4_trim_all_free(sb, group, first_cluster,
5258 end, minlen);
5259 if (cnt < 0) {
5260 ret = cnt;
5261 break;
5263 trimmed += cnt;
5267 * For every group except the first one, we are sure
5268 * that the first cluster to discard will be cluster #0.
5270 first_cluster = 0;
5273 if (!ret)
5274 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5276 out:
5277 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5278 return ret;