Loading...
1/*
2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
17 */
18
19
20/*
21 * mballoc.c contains the multiblocks allocation routines
22 */
23
24#include "mballoc.h"
25#include <linux/debugfs.h>
26#include <linux/slab.h>
27#include <trace/events/ext4.h>
28
29/*
30 * MUSTDO:
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
33 *
34 * TODO v4:
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
38 * - quota
39 * - reservation for superuser
40 *
41 * TODO v3:
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
46 * - error handling
47 */
48
49/*
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
52 *
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
62 *
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
65 *
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
69 * represented as:
70 *
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
75 *
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This makes sure that
79 * we have contiguous physical blocks representing the file blocks
80 *
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
83 * pa_free.
84 *
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list represented as
88 *
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 *
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
93 *
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) within the prealloc space.
96 *
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
105 * inode as:
106 *
107 * { page }
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
109 *
110 *
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
115 *
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
118 *
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
122 *
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * the smallest multiple of the stripe value (sbi->s_stripe) which is
135 * greater than the default mb_group_prealloc.
136 *
137 * The regular allocator (using the buddy cache) supports a few tunables.
138 *
139 * /sys/fs/ext4/<partition>/mb_min_to_scan
140 * /sys/fs/ext4/<partition>/mb_max_to_scan
141 * /sys/fs/ext4/<partition>/mb_order2_req
142 *
143 * The regular allocator uses buddy scan only if the request len is power of
144 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
145 * value of s_mb_order2_reqs can be tuned via
146 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
147 * stripe size (sbi->s_stripe), we try to search for contiguous block in
148 * stripe size. This should result in better allocation on RAID setups. If
149 * not, we search in the specific group using bitmap for best extents. The
150 * tunable min_to_scan and max_to_scan control the behaviour here.
151 * min_to_scan indicate how long the mballoc __must__ look for a best
152 * extent and max_to_scan indicates how long the mballoc __can__ look for a
153 * best extent in the found extents. Searching for the blocks starts with
154 * the group specified as the goal value in allocation context via
155 * ac_g_ex. Each group is first checked based on the criteria whether it
156 * can be used for allocation. ext4_mb_good_group explains how the groups are
157 * checked.
158 *
159 * Both the prealloc space are getting populated as above. So for the first
160 * request we will hit the buddy cache which will result in this prealloc
161 * space getting filled. The prealloc space is then later used for the
162 * subsequent request.
163 */
164
165/*
166 * mballoc operates on the following data:
167 * - on-disk bitmap
168 * - in-core buddy (actually includes buddy and bitmap)
169 * - preallocation descriptors (PAs)
170 *
171 * there are two types of preallocations:
172 * - inode
173 * assiged to specific inode and can be used for this inode only.
174 * it describes part of inode's space preallocated to specific
175 * physical blocks. any block from that preallocated can be used
176 * independent. the descriptor just tracks number of blocks left
177 * unused. so, before taking some block from descriptor, one must
178 * make sure corresponded logical block isn't allocated yet. this
179 * also means that freeing any block within descriptor's range
180 * must discard all preallocated blocks.
181 * - locality group
182 * assigned to specific locality group which does not translate to
183 * permanent set of inodes: inode can join and leave group. space
184 * from this type of preallocation can be used for any inode. thus
185 * it's consumed from the beginning to the end.
186 *
187 * relation between them can be expressed as:
188 * in-core buddy = on-disk bitmap + preallocation descriptors
189 *
190 * this mean blocks mballoc considers used are:
191 * - allocated blocks (persistent)
192 * - preallocated blocks (non-persistent)
193 *
194 * consistency in mballoc world means that at any time a block is either
195 * free or used in ALL structures. notice: "any time" should not be read
196 * literally -- time is discrete and delimited by locks.
197 *
198 * to keep it simple, we don't use block numbers, instead we count number of
199 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 *
201 * all operations can be expressed as:
202 * - init buddy: buddy = on-disk + PAs
203 * - new PA: buddy += N; PA = N
204 * - use inode PA: on-disk += N; PA -= N
205 * - discard inode PA buddy -= on-disk - PA; PA = 0
206 * - use locality group PA on-disk += N; PA -= N
207 * - discard locality group PA buddy -= PA; PA = 0
208 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
209 * is used in real operation because we can't know actual used
210 * bits from PA, only from on-disk bitmap
211 *
212 * if we follow this strict logic, then all operations above should be atomic.
213 * given some of them can block, we'd have to use something like semaphores
214 * killing performance on high-end SMP hardware. let's try to relax it using
215 * the following knowledge:
216 * 1) if buddy is referenced, it's already initialized
217 * 2) while block is used in buddy and the buddy is referenced,
218 * nobody can re-allocate that block
219 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
220 * bit set and PA claims same block, it's OK. IOW, one can set bit in
221 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
222 * block
223 *
224 * so, now we're building a concurrency table:
225 * - init buddy vs.
226 * - new PA
227 * blocks for PA are allocated in the buddy, buddy must be referenced
228 * until PA is linked to allocation group to avoid concurrent buddy init
229 * - use inode PA
230 * we need to make sure that either on-disk bitmap or PA has uptodate data
231 * given (3) we care that PA-=N operation doesn't interfere with init
232 * - discard inode PA
233 * the simplest way would be to have buddy initialized by the discard
234 * - use locality group PA
235 * again PA-=N must be serialized with init
236 * - discard locality group PA
237 * the simplest way would be to have buddy initialized by the discard
238 * - new PA vs.
239 * - use inode PA
240 * i_data_sem serializes them
241 * - discard inode PA
242 * discard process must wait until PA isn't used by another process
243 * - use locality group PA
244 * some mutex should serialize them
245 * - discard locality group PA
246 * discard process must wait until PA isn't used by another process
247 * - use inode PA
248 * - use inode PA
249 * i_data_sem or another mutex should serializes them
250 * - discard inode PA
251 * discard process must wait until PA isn't used by another process
252 * - use locality group PA
253 * nothing wrong here -- they're different PAs covering different blocks
254 * - discard locality group PA
255 * discard process must wait until PA isn't used by another process
256 *
257 * now we're ready to make few consequences:
258 * - PA is referenced and while it is no discard is possible
259 * - PA is referenced until block isn't marked in on-disk bitmap
260 * - PA changes only after on-disk bitmap
261 * - discard must not compete with init. either init is done before
262 * any discard or they're serialized somehow
263 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 *
265 * a special case when we've used PA to emptiness. no need to modify buddy
266 * in this case, but we should care about concurrent init
267 *
268 */
269
270 /*
271 * Logic in few words:
272 *
273 * - allocation:
274 * load group
275 * find blocks
276 * mark bits in on-disk bitmap
277 * release group
278 *
279 * - use preallocation:
280 * find proper PA (per-inode or group)
281 * load group
282 * mark bits in on-disk bitmap
283 * release group
284 * release PA
285 *
286 * - free:
287 * load group
288 * mark bits in on-disk bitmap
289 * release group
290 *
291 * - discard preallocations in group:
292 * mark PAs deleted
293 * move them onto local list
294 * load on-disk bitmap
295 * load group
296 * remove PA from object (inode or locality group)
297 * mark free blocks in-core
298 *
299 * - discard inode's preallocations:
300 */
301
302/*
303 * Locking rules
304 *
305 * Locks:
306 * - bitlock on a group (group)
307 * - object (inode/locality) (object)
308 * - per-pa lock (pa)
309 *
310 * Paths:
311 * - new pa
312 * object
313 * group
314 *
315 * - find and use pa:
316 * pa
317 *
318 * - release consumed pa:
319 * pa
320 * group
321 * object
322 *
323 * - generate in-core bitmap:
324 * group
325 * pa
326 *
327 * - discard all for given object (inode, locality group):
328 * object
329 * pa
330 * group
331 *
332 * - discard all for given group:
333 * group
334 * pa
335 * group
336 * object
337 *
338 */
339static struct kmem_cache *ext4_pspace_cachep;
340static struct kmem_cache *ext4_ac_cachep;
341static struct kmem_cache *ext4_free_ext_cachep;
342
343/* We create slab caches for groupinfo data structures based on the
344 * superblock block size. There will be one per mounted filesystem for
345 * each unique s_blocksize_bits */
346#define NR_GRPINFO_CACHES 8
347static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
348
349static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
350 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
351 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
352 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
353};
354
355static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
356 ext4_group_t group);
357static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
358 ext4_group_t group);
359static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
360
361static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
362{
363#if BITS_PER_LONG == 64
364 *bit += ((unsigned long) addr & 7UL) << 3;
365 addr = (void *) ((unsigned long) addr & ~7UL);
366#elif BITS_PER_LONG == 32
367 *bit += ((unsigned long) addr & 3UL) << 3;
368 addr = (void *) ((unsigned long) addr & ~3UL);
369#else
370#error "how many bits you are?!"
371#endif
372 return addr;
373}
374
375static inline int mb_test_bit(int bit, void *addr)
376{
377 /*
378 * ext4_test_bit on architecture like powerpc
379 * needs unsigned long aligned address
380 */
381 addr = mb_correct_addr_and_bit(&bit, addr);
382 return ext4_test_bit(bit, addr);
383}
384
385static inline void mb_set_bit(int bit, void *addr)
386{
387 addr = mb_correct_addr_and_bit(&bit, addr);
388 ext4_set_bit(bit, addr);
389}
390
391static inline void mb_clear_bit(int bit, void *addr)
392{
393 addr = mb_correct_addr_and_bit(&bit, addr);
394 ext4_clear_bit(bit, addr);
395}
396
397static inline int mb_find_next_zero_bit(void *addr, int max, int start)
398{
399 int fix = 0, ret, tmpmax;
400 addr = mb_correct_addr_and_bit(&fix, addr);
401 tmpmax = max + fix;
402 start += fix;
403
404 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
405 if (ret > max)
406 return max;
407 return ret;
408}
409
410static inline int mb_find_next_bit(void *addr, int max, int start)
411{
412 int fix = 0, ret, tmpmax;
413 addr = mb_correct_addr_and_bit(&fix, addr);
414 tmpmax = max + fix;
415 start += fix;
416
417 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
418 if (ret > max)
419 return max;
420 return ret;
421}
422
423static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
424{
425 char *bb;
426
427 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
428 BUG_ON(max == NULL);
429
430 if (order > e4b->bd_blkbits + 1) {
431 *max = 0;
432 return NULL;
433 }
434
435 /* at order 0 we see each particular block */
436 if (order == 0) {
437 *max = 1 << (e4b->bd_blkbits + 3);
438 return EXT4_MB_BITMAP(e4b);
439 }
440
441 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
442 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
443
444 return bb;
445}
446
447#ifdef DOUBLE_CHECK
448static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
449 int first, int count)
450{
451 int i;
452 struct super_block *sb = e4b->bd_sb;
453
454 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
455 return;
456 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
457 for (i = 0; i < count; i++) {
458 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
459 ext4_fsblk_t blocknr;
460
461 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
462 blocknr += first + i;
463 ext4_grp_locked_error(sb, e4b->bd_group,
464 inode ? inode->i_ino : 0,
465 blocknr,
466 "freeing block already freed "
467 "(bit %u)",
468 first + i);
469 }
470 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
471 }
472}
473
474static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
475{
476 int i;
477
478 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
479 return;
480 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
481 for (i = 0; i < count; i++) {
482 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
483 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
484 }
485}
486
487static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
488{
489 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
490 unsigned char *b1, *b2;
491 int i;
492 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
493 b2 = (unsigned char *) bitmap;
494 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
495 if (b1[i] != b2[i]) {
496 ext4_msg(e4b->bd_sb, KERN_ERR,
497 "corruption in group %u "
498 "at byte %u(%u): %x in copy != %x "
499 "on disk/prealloc",
500 e4b->bd_group, i, i * 8, b1[i], b2[i]);
501 BUG();
502 }
503 }
504 }
505}
506
507#else
508static inline void mb_free_blocks_double(struct inode *inode,
509 struct ext4_buddy *e4b, int first, int count)
510{
511 return;
512}
513static inline void mb_mark_used_double(struct ext4_buddy *e4b,
514 int first, int count)
515{
516 return;
517}
518static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
519{
520 return;
521}
522#endif
523
524#ifdef AGGRESSIVE_CHECK
525
526#define MB_CHECK_ASSERT(assert) \
527do { \
528 if (!(assert)) { \
529 printk(KERN_EMERG \
530 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
531 function, file, line, # assert); \
532 BUG(); \
533 } \
534} while (0)
535
536static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
537 const char *function, int line)
538{
539 struct super_block *sb = e4b->bd_sb;
540 int order = e4b->bd_blkbits + 1;
541 int max;
542 int max2;
543 int i;
544 int j;
545 int k;
546 int count;
547 struct ext4_group_info *grp;
548 int fragments = 0;
549 int fstart;
550 struct list_head *cur;
551 void *buddy;
552 void *buddy2;
553
554 {
555 static int mb_check_counter;
556 if (mb_check_counter++ % 100 != 0)
557 return 0;
558 }
559
560 while (order > 1) {
561 buddy = mb_find_buddy(e4b, order, &max);
562 MB_CHECK_ASSERT(buddy);
563 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
564 MB_CHECK_ASSERT(buddy2);
565 MB_CHECK_ASSERT(buddy != buddy2);
566 MB_CHECK_ASSERT(max * 2 == max2);
567
568 count = 0;
569 for (i = 0; i < max; i++) {
570
571 if (mb_test_bit(i, buddy)) {
572 /* only single bit in buddy2 may be 1 */
573 if (!mb_test_bit(i << 1, buddy2)) {
574 MB_CHECK_ASSERT(
575 mb_test_bit((i<<1)+1, buddy2));
576 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
577 MB_CHECK_ASSERT(
578 mb_test_bit(i << 1, buddy2));
579 }
580 continue;
581 }
582
583 /* both bits in buddy2 must be 0 */
584 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
585 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
586
587 for (j = 0; j < (1 << order); j++) {
588 k = (i * (1 << order)) + j;
589 MB_CHECK_ASSERT(
590 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
591 }
592 count++;
593 }
594 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
595 order--;
596 }
597
598 fstart = -1;
599 buddy = mb_find_buddy(e4b, 0, &max);
600 for (i = 0; i < max; i++) {
601 if (!mb_test_bit(i, buddy)) {
602 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
603 if (fstart == -1) {
604 fragments++;
605 fstart = i;
606 }
607 continue;
608 }
609 fstart = -1;
610 /* check used bits only */
611 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
612 buddy2 = mb_find_buddy(e4b, j, &max2);
613 k = i >> j;
614 MB_CHECK_ASSERT(k < max2);
615 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
616 }
617 }
618 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
619 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
620
621 grp = ext4_get_group_info(sb, e4b->bd_group);
622 list_for_each(cur, &grp->bb_prealloc_list) {
623 ext4_group_t groupnr;
624 struct ext4_prealloc_space *pa;
625 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
626 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
627 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
628 for (i = 0; i < pa->pa_len; i++)
629 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
630 }
631 return 0;
632}
633#undef MB_CHECK_ASSERT
634#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
635 __FILE__, __func__, __LINE__)
636#else
637#define mb_check_buddy(e4b)
638#endif
639
640/*
641 * Divide blocks started from @first with length @len into
642 * smaller chunks with power of 2 blocks.
643 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
644 * then increase bb_counters[] for corresponded chunk size.
645 */
646static void ext4_mb_mark_free_simple(struct super_block *sb,
647 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
648 struct ext4_group_info *grp)
649{
650 struct ext4_sb_info *sbi = EXT4_SB(sb);
651 ext4_grpblk_t min;
652 ext4_grpblk_t max;
653 ext4_grpblk_t chunk;
654 unsigned short border;
655
656 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
657
658 border = 2 << sb->s_blocksize_bits;
659
660 while (len > 0) {
661 /* find how many blocks can be covered since this position */
662 max = ffs(first | border) - 1;
663
664 /* find how many blocks of power 2 we need to mark */
665 min = fls(len) - 1;
666
667 if (max < min)
668 min = max;
669 chunk = 1 << min;
670
671 /* mark multiblock chunks only */
672 grp->bb_counters[min]++;
673 if (min > 0)
674 mb_clear_bit(first >> min,
675 buddy + sbi->s_mb_offsets[min]);
676
677 len -= chunk;
678 first += chunk;
679 }
680}
681
682/*
683 * Cache the order of the largest free extent we have available in this block
684 * group.
685 */
686static void
687mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
688{
689 int i;
690 int bits;
691
692 grp->bb_largest_free_order = -1; /* uninit */
693
694 bits = sb->s_blocksize_bits + 1;
695 for (i = bits; i >= 0; i--) {
696 if (grp->bb_counters[i] > 0) {
697 grp->bb_largest_free_order = i;
698 break;
699 }
700 }
701}
702
703static noinline_for_stack
704void ext4_mb_generate_buddy(struct super_block *sb,
705 void *buddy, void *bitmap, ext4_group_t group)
706{
707 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
708 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
709 ext4_grpblk_t i = 0;
710 ext4_grpblk_t first;
711 ext4_grpblk_t len;
712 unsigned free = 0;
713 unsigned fragments = 0;
714 unsigned long long period = get_cycles();
715
716 /* initialize buddy from bitmap which is aggregation
717 * of on-disk bitmap and preallocations */
718 i = mb_find_next_zero_bit(bitmap, max, 0);
719 grp->bb_first_free = i;
720 while (i < max) {
721 fragments++;
722 first = i;
723 i = mb_find_next_bit(bitmap, max, i);
724 len = i - first;
725 free += len;
726 if (len > 1)
727 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
728 else
729 grp->bb_counters[0]++;
730 if (i < max)
731 i = mb_find_next_zero_bit(bitmap, max, i);
732 }
733 grp->bb_fragments = fragments;
734
735 if (free != grp->bb_free) {
736 ext4_grp_locked_error(sb, group, 0, 0,
737 "%u blocks in bitmap, %u in gd",
738 free, grp->bb_free);
739 /*
740 * If we intent to continue, we consider group descritor
741 * corrupt and update bb_free using bitmap value
742 */
743 grp->bb_free = free;
744 }
745 mb_set_largest_free_order(sb, grp);
746
747 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
748
749 period = get_cycles() - period;
750 spin_lock(&EXT4_SB(sb)->s_bal_lock);
751 EXT4_SB(sb)->s_mb_buddies_generated++;
752 EXT4_SB(sb)->s_mb_generation_time += period;
753 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
754}
755
756/* The buddy information is attached the buddy cache inode
757 * for convenience. The information regarding each group
758 * is loaded via ext4_mb_load_buddy. The information involve
759 * block bitmap and buddy information. The information are
760 * stored in the inode as
761 *
762 * { page }
763 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
764 *
765 *
766 * one block each for bitmap and buddy information.
767 * So for each group we take up 2 blocks. A page can
768 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
769 * So it can have information regarding groups_per_page which
770 * is blocks_per_page/2
771 *
772 * Locking note: This routine takes the block group lock of all groups
773 * for this page; do not hold this lock when calling this routine!
774 */
775
776static int ext4_mb_init_cache(struct page *page, char *incore)
777{
778 ext4_group_t ngroups;
779 int blocksize;
780 int blocks_per_page;
781 int groups_per_page;
782 int err = 0;
783 int i;
784 ext4_group_t first_group;
785 int first_block;
786 struct super_block *sb;
787 struct buffer_head *bhs;
788 struct buffer_head **bh;
789 struct inode *inode;
790 char *data;
791 char *bitmap;
792 struct ext4_group_info *grinfo;
793
794 mb_debug(1, "init page %lu\n", page->index);
795
796 inode = page->mapping->host;
797 sb = inode->i_sb;
798 ngroups = ext4_get_groups_count(sb);
799 blocksize = 1 << inode->i_blkbits;
800 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
801
802 groups_per_page = blocks_per_page >> 1;
803 if (groups_per_page == 0)
804 groups_per_page = 1;
805
806 /* allocate buffer_heads to read bitmaps */
807 if (groups_per_page > 1) {
808 err = -ENOMEM;
809 i = sizeof(struct buffer_head *) * groups_per_page;
810 bh = kzalloc(i, GFP_NOFS);
811 if (bh == NULL)
812 goto out;
813 } else
814 bh = &bhs;
815
816 first_group = page->index * blocks_per_page / 2;
817
818 /* read all groups the page covers into the cache */
819 for (i = 0; i < groups_per_page; i++) {
820 struct ext4_group_desc *desc;
821
822 if (first_group + i >= ngroups)
823 break;
824
825 grinfo = ext4_get_group_info(sb, first_group + i);
826 /*
827 * If page is uptodate then we came here after online resize
828 * which added some new uninitialized group info structs, so
829 * we must skip all initialized uptodate buddies on the page,
830 * which may be currently in use by an allocating task.
831 */
832 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
833 bh[i] = NULL;
834 continue;
835 }
836
837 err = -EIO;
838 desc = ext4_get_group_desc(sb, first_group + i, NULL);
839 if (desc == NULL)
840 goto out;
841
842 err = -ENOMEM;
843 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
844 if (bh[i] == NULL)
845 goto out;
846
847 if (bitmap_uptodate(bh[i]))
848 continue;
849
850 lock_buffer(bh[i]);
851 if (bitmap_uptodate(bh[i])) {
852 unlock_buffer(bh[i]);
853 continue;
854 }
855 ext4_lock_group(sb, first_group + i);
856 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
857 ext4_init_block_bitmap(sb, bh[i],
858 first_group + i, desc);
859 set_bitmap_uptodate(bh[i]);
860 set_buffer_uptodate(bh[i]);
861 ext4_unlock_group(sb, first_group + i);
862 unlock_buffer(bh[i]);
863 continue;
864 }
865 ext4_unlock_group(sb, first_group + i);
866 if (buffer_uptodate(bh[i])) {
867 /*
868 * if not uninit if bh is uptodate,
869 * bitmap is also uptodate
870 */
871 set_bitmap_uptodate(bh[i]);
872 unlock_buffer(bh[i]);
873 continue;
874 }
875 get_bh(bh[i]);
876 /*
877 * submit the buffer_head for read. We can
878 * safely mark the bitmap as uptodate now.
879 * We do it here so the bitmap uptodate bit
880 * get set with buffer lock held.
881 */
882 set_bitmap_uptodate(bh[i]);
883 bh[i]->b_end_io = end_buffer_read_sync;
884 submit_bh(READ, bh[i]);
885 mb_debug(1, "read bitmap for group %u\n", first_group + i);
886 }
887
888 /* wait for I/O completion */
889 for (i = 0; i < groups_per_page; i++)
890 if (bh[i])
891 wait_on_buffer(bh[i]);
892
893 err = -EIO;
894 for (i = 0; i < groups_per_page; i++)
895 if (bh[i] && !buffer_uptodate(bh[i]))
896 goto out;
897
898 err = 0;
899 first_block = page->index * blocks_per_page;
900 for (i = 0; i < blocks_per_page; i++) {
901 int group;
902
903 group = (first_block + i) >> 1;
904 if (group >= ngroups)
905 break;
906
907 if (!bh[group - first_group])
908 /* skip initialized uptodate buddy */
909 continue;
910
911 /*
912 * data carry information regarding this
913 * particular group in the format specified
914 * above
915 *
916 */
917 data = page_address(page) + (i * blocksize);
918 bitmap = bh[group - first_group]->b_data;
919
920 /*
921 * We place the buddy block and bitmap block
922 * close together
923 */
924 if ((first_block + i) & 1) {
925 /* this is block of buddy */
926 BUG_ON(incore == NULL);
927 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
928 group, page->index, i * blocksize);
929 trace_ext4_mb_buddy_bitmap_load(sb, group);
930 grinfo = ext4_get_group_info(sb, group);
931 grinfo->bb_fragments = 0;
932 memset(grinfo->bb_counters, 0,
933 sizeof(*grinfo->bb_counters) *
934 (sb->s_blocksize_bits+2));
935 /*
936 * incore got set to the group block bitmap below
937 */
938 ext4_lock_group(sb, group);
939 /* init the buddy */
940 memset(data, 0xff, blocksize);
941 ext4_mb_generate_buddy(sb, data, incore, group);
942 ext4_unlock_group(sb, group);
943 incore = NULL;
944 } else {
945 /* this is block of bitmap */
946 BUG_ON(incore != NULL);
947 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
948 group, page->index, i * blocksize);
949 trace_ext4_mb_bitmap_load(sb, group);
950
951 /* see comments in ext4_mb_put_pa() */
952 ext4_lock_group(sb, group);
953 memcpy(data, bitmap, blocksize);
954
955 /* mark all preallocated blks used in in-core bitmap */
956 ext4_mb_generate_from_pa(sb, data, group);
957 ext4_mb_generate_from_freelist(sb, data, group);
958 ext4_unlock_group(sb, group);
959
960 /* set incore so that the buddy information can be
961 * generated using this
962 */
963 incore = data;
964 }
965 }
966 SetPageUptodate(page);
967
968out:
969 if (bh) {
970 for (i = 0; i < groups_per_page; i++)
971 brelse(bh[i]);
972 if (bh != &bhs)
973 kfree(bh);
974 }
975 return err;
976}
977
978/*
979 * Lock the buddy and bitmap pages. This make sure other parallel init_group
980 * on the same buddy page doesn't happen whild holding the buddy page lock.
981 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
982 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
983 */
984static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
985 ext4_group_t group, struct ext4_buddy *e4b)
986{
987 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
988 int block, pnum, poff;
989 int blocks_per_page;
990 struct page *page;
991
992 e4b->bd_buddy_page = NULL;
993 e4b->bd_bitmap_page = NULL;
994
995 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
996 /*
997 * the buddy cache inode stores the block bitmap
998 * and buddy information in consecutive blocks.
999 * So for each group we need two blocks.
1000 */
1001 block = group * 2;
1002 pnum = block / blocks_per_page;
1003 poff = block % blocks_per_page;
1004 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1005 if (!page)
1006 return -EIO;
1007 BUG_ON(page->mapping != inode->i_mapping);
1008 e4b->bd_bitmap_page = page;
1009 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1010
1011 if (blocks_per_page >= 2) {
1012 /* buddy and bitmap are on the same page */
1013 return 0;
1014 }
1015
1016 block++;
1017 pnum = block / blocks_per_page;
1018 poff = block % blocks_per_page;
1019 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1020 if (!page)
1021 return -EIO;
1022 BUG_ON(page->mapping != inode->i_mapping);
1023 e4b->bd_buddy_page = page;
1024 return 0;
1025}
1026
1027static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1028{
1029 if (e4b->bd_bitmap_page) {
1030 unlock_page(e4b->bd_bitmap_page);
1031 page_cache_release(e4b->bd_bitmap_page);
1032 }
1033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 page_cache_release(e4b->bd_buddy_page);
1036 }
1037}
1038
1039/*
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!
1043 */
1044static noinline_for_stack
1045int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1046{
1047
1048 struct ext4_group_info *this_grp;
1049 struct ext4_buddy e4b;
1050 struct page *page;
1051 int ret = 0;
1052
1053 mb_debug(1, "init group %u\n", group);
1054 this_grp = ext4_get_group_info(sb, group);
1055 /*
1056 * This ensures that we don't reinit the buddy cache
1057 * page which map to the group from which we are already
1058 * allocating. If we are looking at the buddy cache we would
1059 * have taken a reference using ext4_mb_load_buddy and that
1060 * would have pinned buddy page to page cache.
1061 */
1062 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1063 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1064 /*
1065 * somebody initialized the group
1066 * return without doing anything
1067 */
1068 goto err;
1069 }
1070
1071 page = e4b.bd_bitmap_page;
1072 ret = ext4_mb_init_cache(page, NULL);
1073 if (ret)
1074 goto err;
1075 if (!PageUptodate(page)) {
1076 ret = -EIO;
1077 goto err;
1078 }
1079 mark_page_accessed(page);
1080
1081 if (e4b.bd_buddy_page == NULL) {
1082 /*
1083 * If both the bitmap and buddy are in
1084 * the same page we don't need to force
1085 * init the buddy
1086 */
1087 ret = 0;
1088 goto err;
1089 }
1090 /* init buddy cache */
1091 page = e4b.bd_buddy_page;
1092 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1093 if (ret)
1094 goto err;
1095 if (!PageUptodate(page)) {
1096 ret = -EIO;
1097 goto err;
1098 }
1099 mark_page_accessed(page);
1100err:
1101 ext4_mb_put_buddy_page_lock(&e4b);
1102 return ret;
1103}
1104
1105/*
1106 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1107 * block group lock of all groups for this page; do not hold the BG lock when
1108 * calling this routine!
1109 */
1110static noinline_for_stack int
1111ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1112 struct ext4_buddy *e4b)
1113{
1114 int blocks_per_page;
1115 int block;
1116 int pnum;
1117 int poff;
1118 struct page *page;
1119 int ret;
1120 struct ext4_group_info *grp;
1121 struct ext4_sb_info *sbi = EXT4_SB(sb);
1122 struct inode *inode = sbi->s_buddy_cache;
1123
1124 mb_debug(1, "load group %u\n", group);
1125
1126 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1127 grp = ext4_get_group_info(sb, group);
1128
1129 e4b->bd_blkbits = sb->s_blocksize_bits;
1130 e4b->bd_info = grp;
1131 e4b->bd_sb = sb;
1132 e4b->bd_group = group;
1133 e4b->bd_buddy_page = NULL;
1134 e4b->bd_bitmap_page = NULL;
1135
1136 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1137 /*
1138 * we need full data about the group
1139 * to make a good selection
1140 */
1141 ret = ext4_mb_init_group(sb, group);
1142 if (ret)
1143 return ret;
1144 }
1145
1146 /*
1147 * the buddy cache inode stores the block bitmap
1148 * and buddy information in consecutive blocks.
1149 * So for each group we need two blocks.
1150 */
1151 block = group * 2;
1152 pnum = block / blocks_per_page;
1153 poff = block % blocks_per_page;
1154
1155 /* we could use find_or_create_page(), but it locks page
1156 * what we'd like to avoid in fast path ... */
1157 page = find_get_page(inode->i_mapping, pnum);
1158 if (page == NULL || !PageUptodate(page)) {
1159 if (page)
1160 /*
1161 * drop the page reference and try
1162 * to get the page with lock. If we
1163 * are not uptodate that implies
1164 * somebody just created the page but
1165 * is yet to initialize the same. So
1166 * wait for it to initialize.
1167 */
1168 page_cache_release(page);
1169 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1170 if (page) {
1171 BUG_ON(page->mapping != inode->i_mapping);
1172 if (!PageUptodate(page)) {
1173 ret = ext4_mb_init_cache(page, NULL);
1174 if (ret) {
1175 unlock_page(page);
1176 goto err;
1177 }
1178 mb_cmp_bitmaps(e4b, page_address(page) +
1179 (poff * sb->s_blocksize));
1180 }
1181 unlock_page(page);
1182 }
1183 }
1184 if (page == NULL || !PageUptodate(page)) {
1185 ret = -EIO;
1186 goto err;
1187 }
1188 e4b->bd_bitmap_page = page;
1189 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1190 mark_page_accessed(page);
1191
1192 block++;
1193 pnum = block / blocks_per_page;
1194 poff = block % blocks_per_page;
1195
1196 page = find_get_page(inode->i_mapping, pnum);
1197 if (page == NULL || !PageUptodate(page)) {
1198 if (page)
1199 page_cache_release(page);
1200 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1201 if (page) {
1202 BUG_ON(page->mapping != inode->i_mapping);
1203 if (!PageUptodate(page)) {
1204 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1205 if (ret) {
1206 unlock_page(page);
1207 goto err;
1208 }
1209 }
1210 unlock_page(page);
1211 }
1212 }
1213 if (page == NULL || !PageUptodate(page)) {
1214 ret = -EIO;
1215 goto err;
1216 }
1217 e4b->bd_buddy_page = page;
1218 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1219 mark_page_accessed(page);
1220
1221 BUG_ON(e4b->bd_bitmap_page == NULL);
1222 BUG_ON(e4b->bd_buddy_page == NULL);
1223
1224 return 0;
1225
1226err:
1227 if (page)
1228 page_cache_release(page);
1229 if (e4b->bd_bitmap_page)
1230 page_cache_release(e4b->bd_bitmap_page);
1231 if (e4b->bd_buddy_page)
1232 page_cache_release(e4b->bd_buddy_page);
1233 e4b->bd_buddy = NULL;
1234 e4b->bd_bitmap = NULL;
1235 return ret;
1236}
1237
1238static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1239{
1240 if (e4b->bd_bitmap_page)
1241 page_cache_release(e4b->bd_bitmap_page);
1242 if (e4b->bd_buddy_page)
1243 page_cache_release(e4b->bd_buddy_page);
1244}
1245
1246
1247static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1248{
1249 int order = 1;
1250 void *bb;
1251
1252 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1253 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1254
1255 bb = EXT4_MB_BUDDY(e4b);
1256 while (order <= e4b->bd_blkbits + 1) {
1257 block = block >> 1;
1258 if (!mb_test_bit(block, bb)) {
1259 /* this block is part of buddy of order 'order' */
1260 return order;
1261 }
1262 bb += 1 << (e4b->bd_blkbits - order);
1263 order++;
1264 }
1265 return 0;
1266}
1267
1268static void mb_clear_bits(void *bm, int cur, int len)
1269{
1270 __u32 *addr;
1271
1272 len = cur + len;
1273 while (cur < len) {
1274 if ((cur & 31) == 0 && (len - cur) >= 32) {
1275 /* fast path: clear whole word at once */
1276 addr = bm + (cur >> 3);
1277 *addr = 0;
1278 cur += 32;
1279 continue;
1280 }
1281 mb_clear_bit(cur, bm);
1282 cur++;
1283 }
1284}
1285
1286void ext4_set_bits(void *bm, int cur, int len)
1287{
1288 __u32 *addr;
1289
1290 len = cur + len;
1291 while (cur < len) {
1292 if ((cur & 31) == 0 && (len - cur) >= 32) {
1293 /* fast path: set whole word at once */
1294 addr = bm + (cur >> 3);
1295 *addr = 0xffffffff;
1296 cur += 32;
1297 continue;
1298 }
1299 mb_set_bit(cur, bm);
1300 cur++;
1301 }
1302}
1303
1304static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1305 int first, int count)
1306{
1307 int block = 0;
1308 int max = 0;
1309 int order;
1310 void *buddy;
1311 void *buddy2;
1312 struct super_block *sb = e4b->bd_sb;
1313
1314 BUG_ON(first + count > (sb->s_blocksize << 3));
1315 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1316 mb_check_buddy(e4b);
1317 mb_free_blocks_double(inode, e4b, first, count);
1318
1319 e4b->bd_info->bb_free += count;
1320 if (first < e4b->bd_info->bb_first_free)
1321 e4b->bd_info->bb_first_free = first;
1322
1323 /* let's maintain fragments counter */
1324 if (first != 0)
1325 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1326 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1327 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1328 if (block && max)
1329 e4b->bd_info->bb_fragments--;
1330 else if (!block && !max)
1331 e4b->bd_info->bb_fragments++;
1332
1333 /* let's maintain buddy itself */
1334 while (count-- > 0) {
1335 block = first++;
1336 order = 0;
1337
1338 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1339 ext4_fsblk_t blocknr;
1340
1341 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1342 blocknr += block;
1343 ext4_grp_locked_error(sb, e4b->bd_group,
1344 inode ? inode->i_ino : 0,
1345 blocknr,
1346 "freeing already freed block "
1347 "(bit %u)", block);
1348 }
1349 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1350 e4b->bd_info->bb_counters[order]++;
1351
1352 /* start of the buddy */
1353 buddy = mb_find_buddy(e4b, order, &max);
1354
1355 do {
1356 block &= ~1UL;
1357 if (mb_test_bit(block, buddy) ||
1358 mb_test_bit(block + 1, buddy))
1359 break;
1360
1361 /* both the buddies are free, try to coalesce them */
1362 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1363
1364 if (!buddy2)
1365 break;
1366
1367 if (order > 0) {
1368 /* for special purposes, we don't set
1369 * free bits in bitmap */
1370 mb_set_bit(block, buddy);
1371 mb_set_bit(block + 1, buddy);
1372 }
1373 e4b->bd_info->bb_counters[order]--;
1374 e4b->bd_info->bb_counters[order]--;
1375
1376 block = block >> 1;
1377 order++;
1378 e4b->bd_info->bb_counters[order]++;
1379
1380 mb_clear_bit(block, buddy2);
1381 buddy = buddy2;
1382 } while (1);
1383 }
1384 mb_set_largest_free_order(sb, e4b->bd_info);
1385 mb_check_buddy(e4b);
1386}
1387
1388static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1389 int needed, struct ext4_free_extent *ex)
1390{
1391 int next = block;
1392 int max;
1393 int ord;
1394 void *buddy;
1395
1396 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1397 BUG_ON(ex == NULL);
1398
1399 buddy = mb_find_buddy(e4b, order, &max);
1400 BUG_ON(buddy == NULL);
1401 BUG_ON(block >= max);
1402 if (mb_test_bit(block, buddy)) {
1403 ex->fe_len = 0;
1404 ex->fe_start = 0;
1405 ex->fe_group = 0;
1406 return 0;
1407 }
1408
1409 /* FIXME dorp order completely ? */
1410 if (likely(order == 0)) {
1411 /* find actual order */
1412 order = mb_find_order_for_block(e4b, block);
1413 block = block >> order;
1414 }
1415
1416 ex->fe_len = 1 << order;
1417 ex->fe_start = block << order;
1418 ex->fe_group = e4b->bd_group;
1419
1420 /* calc difference from given start */
1421 next = next - ex->fe_start;
1422 ex->fe_len -= next;
1423 ex->fe_start += next;
1424
1425 while (needed > ex->fe_len &&
1426 (buddy = mb_find_buddy(e4b, order, &max))) {
1427
1428 if (block + 1 >= max)
1429 break;
1430
1431 next = (block + 1) * (1 << order);
1432 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1433 break;
1434
1435 ord = mb_find_order_for_block(e4b, next);
1436
1437 order = ord;
1438 block = next >> order;
1439 ex->fe_len += 1 << order;
1440 }
1441
1442 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1443 return ex->fe_len;
1444}
1445
1446static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1447{
1448 int ord;
1449 int mlen = 0;
1450 int max = 0;
1451 int cur;
1452 int start = ex->fe_start;
1453 int len = ex->fe_len;
1454 unsigned ret = 0;
1455 int len0 = len;
1456 void *buddy;
1457
1458 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1459 BUG_ON(e4b->bd_group != ex->fe_group);
1460 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1461 mb_check_buddy(e4b);
1462 mb_mark_used_double(e4b, start, len);
1463
1464 e4b->bd_info->bb_free -= len;
1465 if (e4b->bd_info->bb_first_free == start)
1466 e4b->bd_info->bb_first_free += len;
1467
1468 /* let's maintain fragments counter */
1469 if (start != 0)
1470 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1471 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1472 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1473 if (mlen && max)
1474 e4b->bd_info->bb_fragments++;
1475 else if (!mlen && !max)
1476 e4b->bd_info->bb_fragments--;
1477
1478 /* let's maintain buddy itself */
1479 while (len) {
1480 ord = mb_find_order_for_block(e4b, start);
1481
1482 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1483 /* the whole chunk may be allocated at once! */
1484 mlen = 1 << ord;
1485 buddy = mb_find_buddy(e4b, ord, &max);
1486 BUG_ON((start >> ord) >= max);
1487 mb_set_bit(start >> ord, buddy);
1488 e4b->bd_info->bb_counters[ord]--;
1489 start += mlen;
1490 len -= mlen;
1491 BUG_ON(len < 0);
1492 continue;
1493 }
1494
1495 /* store for history */
1496 if (ret == 0)
1497 ret = len | (ord << 16);
1498
1499 /* we have to split large buddy */
1500 BUG_ON(ord <= 0);
1501 buddy = mb_find_buddy(e4b, ord, &max);
1502 mb_set_bit(start >> ord, buddy);
1503 e4b->bd_info->bb_counters[ord]--;
1504
1505 ord--;
1506 cur = (start >> ord) & ~1U;
1507 buddy = mb_find_buddy(e4b, ord, &max);
1508 mb_clear_bit(cur, buddy);
1509 mb_clear_bit(cur + 1, buddy);
1510 e4b->bd_info->bb_counters[ord]++;
1511 e4b->bd_info->bb_counters[ord]++;
1512 }
1513 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1514
1515 ext4_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1516 mb_check_buddy(e4b);
1517
1518 return ret;
1519}
1520
1521/*
1522 * Must be called under group lock!
1523 */
1524static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1525 struct ext4_buddy *e4b)
1526{
1527 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1528 int ret;
1529
1530 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1531 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1532
1533 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1534 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1535 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1536
1537 /* preallocation can change ac_b_ex, thus we store actually
1538 * allocated blocks for history */
1539 ac->ac_f_ex = ac->ac_b_ex;
1540
1541 ac->ac_status = AC_STATUS_FOUND;
1542 ac->ac_tail = ret & 0xffff;
1543 ac->ac_buddy = ret >> 16;
1544
1545 /*
1546 * take the page reference. We want the page to be pinned
1547 * so that we don't get a ext4_mb_init_cache_call for this
1548 * group until we update the bitmap. That would mean we
1549 * double allocate blocks. The reference is dropped
1550 * in ext4_mb_release_context
1551 */
1552 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1553 get_page(ac->ac_bitmap_page);
1554 ac->ac_buddy_page = e4b->bd_buddy_page;
1555 get_page(ac->ac_buddy_page);
1556 /* store last allocated for subsequent stream allocation */
1557 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1558 spin_lock(&sbi->s_md_lock);
1559 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1560 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1561 spin_unlock(&sbi->s_md_lock);
1562 }
1563}
1564
1565/*
1566 * regular allocator, for general purposes allocation
1567 */
1568
1569static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1570 struct ext4_buddy *e4b,
1571 int finish_group)
1572{
1573 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1574 struct ext4_free_extent *bex = &ac->ac_b_ex;
1575 struct ext4_free_extent *gex = &ac->ac_g_ex;
1576 struct ext4_free_extent ex;
1577 int max;
1578
1579 if (ac->ac_status == AC_STATUS_FOUND)
1580 return;
1581 /*
1582 * We don't want to scan for a whole year
1583 */
1584 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1585 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1586 ac->ac_status = AC_STATUS_BREAK;
1587 return;
1588 }
1589
1590 /*
1591 * Haven't found good chunk so far, let's continue
1592 */
1593 if (bex->fe_len < gex->fe_len)
1594 return;
1595
1596 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1597 && bex->fe_group == e4b->bd_group) {
1598 /* recheck chunk's availability - we don't know
1599 * when it was found (within this lock-unlock
1600 * period or not) */
1601 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1602 if (max >= gex->fe_len) {
1603 ext4_mb_use_best_found(ac, e4b);
1604 return;
1605 }
1606 }
1607}
1608
1609/*
1610 * The routine checks whether found extent is good enough. If it is,
1611 * then the extent gets marked used and flag is set to the context
1612 * to stop scanning. Otherwise, the extent is compared with the
1613 * previous found extent and if new one is better, then it's stored
1614 * in the context. Later, the best found extent will be used, if
1615 * mballoc can't find good enough extent.
1616 *
1617 * FIXME: real allocation policy is to be designed yet!
1618 */
1619static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1620 struct ext4_free_extent *ex,
1621 struct ext4_buddy *e4b)
1622{
1623 struct ext4_free_extent *bex = &ac->ac_b_ex;
1624 struct ext4_free_extent *gex = &ac->ac_g_ex;
1625
1626 BUG_ON(ex->fe_len <= 0);
1627 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1628 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1629 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1630
1631 ac->ac_found++;
1632
1633 /*
1634 * The special case - take what you catch first
1635 */
1636 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1637 *bex = *ex;
1638 ext4_mb_use_best_found(ac, e4b);
1639 return;
1640 }
1641
1642 /*
1643 * Let's check whether the chuck is good enough
1644 */
1645 if (ex->fe_len == gex->fe_len) {
1646 *bex = *ex;
1647 ext4_mb_use_best_found(ac, e4b);
1648 return;
1649 }
1650
1651 /*
1652 * If this is first found extent, just store it in the context
1653 */
1654 if (bex->fe_len == 0) {
1655 *bex = *ex;
1656 return;
1657 }
1658
1659 /*
1660 * If new found extent is better, store it in the context
1661 */
1662 if (bex->fe_len < gex->fe_len) {
1663 /* if the request isn't satisfied, any found extent
1664 * larger than previous best one is better */
1665 if (ex->fe_len > bex->fe_len)
1666 *bex = *ex;
1667 } else if (ex->fe_len > gex->fe_len) {
1668 /* if the request is satisfied, then we try to find
1669 * an extent that still satisfy the request, but is
1670 * smaller than previous one */
1671 if (ex->fe_len < bex->fe_len)
1672 *bex = *ex;
1673 }
1674
1675 ext4_mb_check_limits(ac, e4b, 0);
1676}
1677
1678static noinline_for_stack
1679int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1680 struct ext4_buddy *e4b)
1681{
1682 struct ext4_free_extent ex = ac->ac_b_ex;
1683 ext4_group_t group = ex.fe_group;
1684 int max;
1685 int err;
1686
1687 BUG_ON(ex.fe_len <= 0);
1688 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1689 if (err)
1690 return err;
1691
1692 ext4_lock_group(ac->ac_sb, group);
1693 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1694
1695 if (max > 0) {
1696 ac->ac_b_ex = ex;
1697 ext4_mb_use_best_found(ac, e4b);
1698 }
1699
1700 ext4_unlock_group(ac->ac_sb, group);
1701 ext4_mb_unload_buddy(e4b);
1702
1703 return 0;
1704}
1705
1706static noinline_for_stack
1707int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1708 struct ext4_buddy *e4b)
1709{
1710 ext4_group_t group = ac->ac_g_ex.fe_group;
1711 int max;
1712 int err;
1713 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1714 struct ext4_free_extent ex;
1715
1716 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1717 return 0;
1718
1719 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1720 if (err)
1721 return err;
1722
1723 ext4_lock_group(ac->ac_sb, group);
1724 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1725 ac->ac_g_ex.fe_len, &ex);
1726
1727 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1728 ext4_fsblk_t start;
1729
1730 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1731 ex.fe_start;
1732 /* use do_div to get remainder (would be 64-bit modulo) */
1733 if (do_div(start, sbi->s_stripe) == 0) {
1734 ac->ac_found++;
1735 ac->ac_b_ex = ex;
1736 ext4_mb_use_best_found(ac, e4b);
1737 }
1738 } else if (max >= ac->ac_g_ex.fe_len) {
1739 BUG_ON(ex.fe_len <= 0);
1740 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1741 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1742 ac->ac_found++;
1743 ac->ac_b_ex = ex;
1744 ext4_mb_use_best_found(ac, e4b);
1745 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1746 /* Sometimes, caller may want to merge even small
1747 * number of blocks to an existing extent */
1748 BUG_ON(ex.fe_len <= 0);
1749 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1750 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1751 ac->ac_found++;
1752 ac->ac_b_ex = ex;
1753 ext4_mb_use_best_found(ac, e4b);
1754 }
1755 ext4_unlock_group(ac->ac_sb, group);
1756 ext4_mb_unload_buddy(e4b);
1757
1758 return 0;
1759}
1760
1761/*
1762 * The routine scans buddy structures (not bitmap!) from given order
1763 * to max order and tries to find big enough chunk to satisfy the req
1764 */
1765static noinline_for_stack
1766void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1767 struct ext4_buddy *e4b)
1768{
1769 struct super_block *sb = ac->ac_sb;
1770 struct ext4_group_info *grp = e4b->bd_info;
1771 void *buddy;
1772 int i;
1773 int k;
1774 int max;
1775
1776 BUG_ON(ac->ac_2order <= 0);
1777 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1778 if (grp->bb_counters[i] == 0)
1779 continue;
1780
1781 buddy = mb_find_buddy(e4b, i, &max);
1782 BUG_ON(buddy == NULL);
1783
1784 k = mb_find_next_zero_bit(buddy, max, 0);
1785 BUG_ON(k >= max);
1786
1787 ac->ac_found++;
1788
1789 ac->ac_b_ex.fe_len = 1 << i;
1790 ac->ac_b_ex.fe_start = k << i;
1791 ac->ac_b_ex.fe_group = e4b->bd_group;
1792
1793 ext4_mb_use_best_found(ac, e4b);
1794
1795 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1796
1797 if (EXT4_SB(sb)->s_mb_stats)
1798 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1799
1800 break;
1801 }
1802}
1803
1804/*
1805 * The routine scans the group and measures all found extents.
1806 * In order to optimize scanning, caller must pass number of
1807 * free blocks in the group, so the routine can know upper limit.
1808 */
1809static noinline_for_stack
1810void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1811 struct ext4_buddy *e4b)
1812{
1813 struct super_block *sb = ac->ac_sb;
1814 void *bitmap = EXT4_MB_BITMAP(e4b);
1815 struct ext4_free_extent ex;
1816 int i;
1817 int free;
1818
1819 free = e4b->bd_info->bb_free;
1820 BUG_ON(free <= 0);
1821
1822 i = e4b->bd_info->bb_first_free;
1823
1824 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1825 i = mb_find_next_zero_bit(bitmap,
1826 EXT4_BLOCKS_PER_GROUP(sb), i);
1827 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1828 /*
1829 * IF we have corrupt bitmap, we won't find any
1830 * free blocks even though group info says we
1831 * we have free blocks
1832 */
1833 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1834 "%d free blocks as per "
1835 "group info. But bitmap says 0",
1836 free);
1837 break;
1838 }
1839
1840 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1841 BUG_ON(ex.fe_len <= 0);
1842 if (free < ex.fe_len) {
1843 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1844 "%d free blocks as per "
1845 "group info. But got %d blocks",
1846 free, ex.fe_len);
1847 /*
1848 * The number of free blocks differs. This mostly
1849 * indicate that the bitmap is corrupt. So exit
1850 * without claiming the space.
1851 */
1852 break;
1853 }
1854
1855 ext4_mb_measure_extent(ac, &ex, e4b);
1856
1857 i += ex.fe_len;
1858 free -= ex.fe_len;
1859 }
1860
1861 ext4_mb_check_limits(ac, e4b, 1);
1862}
1863
1864/*
1865 * This is a special case for storages like raid5
1866 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1867 */
1868static noinline_for_stack
1869void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1870 struct ext4_buddy *e4b)
1871{
1872 struct super_block *sb = ac->ac_sb;
1873 struct ext4_sb_info *sbi = EXT4_SB(sb);
1874 void *bitmap = EXT4_MB_BITMAP(e4b);
1875 struct ext4_free_extent ex;
1876 ext4_fsblk_t first_group_block;
1877 ext4_fsblk_t a;
1878 ext4_grpblk_t i;
1879 int max;
1880
1881 BUG_ON(sbi->s_stripe == 0);
1882
1883 /* find first stripe-aligned block in group */
1884 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1885
1886 a = first_group_block + sbi->s_stripe - 1;
1887 do_div(a, sbi->s_stripe);
1888 i = (a * sbi->s_stripe) - first_group_block;
1889
1890 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1891 if (!mb_test_bit(i, bitmap)) {
1892 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1893 if (max >= sbi->s_stripe) {
1894 ac->ac_found++;
1895 ac->ac_b_ex = ex;
1896 ext4_mb_use_best_found(ac, e4b);
1897 break;
1898 }
1899 }
1900 i += sbi->s_stripe;
1901 }
1902}
1903
1904/* This is now called BEFORE we load the buddy bitmap. */
1905static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1906 ext4_group_t group, int cr)
1907{
1908 unsigned free, fragments;
1909 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1910 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1911
1912 BUG_ON(cr < 0 || cr >= 4);
1913
1914 /* We only do this if the grp has never been initialized */
1915 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1916 int ret = ext4_mb_init_group(ac->ac_sb, group);
1917 if (ret)
1918 return 0;
1919 }
1920
1921 free = grp->bb_free;
1922 fragments = grp->bb_fragments;
1923 if (free == 0)
1924 return 0;
1925 if (fragments == 0)
1926 return 0;
1927
1928 switch (cr) {
1929 case 0:
1930 BUG_ON(ac->ac_2order == 0);
1931
1932 if (grp->bb_largest_free_order < ac->ac_2order)
1933 return 0;
1934
1935 /* Avoid using the first bg of a flexgroup for data files */
1936 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1937 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1938 ((group % flex_size) == 0))
1939 return 0;
1940
1941 return 1;
1942 case 1:
1943 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1944 return 1;
1945 break;
1946 case 2:
1947 if (free >= ac->ac_g_ex.fe_len)
1948 return 1;
1949 break;
1950 case 3:
1951 return 1;
1952 default:
1953 BUG();
1954 }
1955
1956 return 0;
1957}
1958
1959static noinline_for_stack int
1960ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1961{
1962 ext4_group_t ngroups, group, i;
1963 int cr;
1964 int err = 0;
1965 struct ext4_sb_info *sbi;
1966 struct super_block *sb;
1967 struct ext4_buddy e4b;
1968
1969 sb = ac->ac_sb;
1970 sbi = EXT4_SB(sb);
1971 ngroups = ext4_get_groups_count(sb);
1972 /* non-extent files are limited to low blocks/groups */
1973 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1974 ngroups = sbi->s_blockfile_groups;
1975
1976 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1977
1978 /* first, try the goal */
1979 err = ext4_mb_find_by_goal(ac, &e4b);
1980 if (err || ac->ac_status == AC_STATUS_FOUND)
1981 goto out;
1982
1983 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1984 goto out;
1985
1986 /*
1987 * ac->ac2_order is set only if the fe_len is a power of 2
1988 * if ac2_order is set we also set criteria to 0 so that we
1989 * try exact allocation using buddy.
1990 */
1991 i = fls(ac->ac_g_ex.fe_len);
1992 ac->ac_2order = 0;
1993 /*
1994 * We search using buddy data only if the order of the request
1995 * is greater than equal to the sbi_s_mb_order2_reqs
1996 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1997 */
1998 if (i >= sbi->s_mb_order2_reqs) {
1999 /*
2000 * This should tell if fe_len is exactly power of 2
2001 */
2002 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2003 ac->ac_2order = i - 1;
2004 }
2005
2006 /* if stream allocation is enabled, use global goal */
2007 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2008 /* TBD: may be hot point */
2009 spin_lock(&sbi->s_md_lock);
2010 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2011 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2012 spin_unlock(&sbi->s_md_lock);
2013 }
2014
2015 /* Let's just scan groups to find more-less suitable blocks */
2016 cr = ac->ac_2order ? 0 : 1;
2017 /*
2018 * cr == 0 try to get exact allocation,
2019 * cr == 3 try to get anything
2020 */
2021repeat:
2022 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2023 ac->ac_criteria = cr;
2024 /*
2025 * searching for the right group start
2026 * from the goal value specified
2027 */
2028 group = ac->ac_g_ex.fe_group;
2029
2030 for (i = 0; i < ngroups; group++, i++) {
2031 if (group == ngroups)
2032 group = 0;
2033
2034 /* This now checks without needing the buddy page */
2035 if (!ext4_mb_good_group(ac, group, cr))
2036 continue;
2037
2038 err = ext4_mb_load_buddy(sb, group, &e4b);
2039 if (err)
2040 goto out;
2041
2042 ext4_lock_group(sb, group);
2043
2044 /*
2045 * We need to check again after locking the
2046 * block group
2047 */
2048 if (!ext4_mb_good_group(ac, group, cr)) {
2049 ext4_unlock_group(sb, group);
2050 ext4_mb_unload_buddy(&e4b);
2051 continue;
2052 }
2053
2054 ac->ac_groups_scanned++;
2055 if (cr == 0)
2056 ext4_mb_simple_scan_group(ac, &e4b);
2057 else if (cr == 1 && sbi->s_stripe &&
2058 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2059 ext4_mb_scan_aligned(ac, &e4b);
2060 else
2061 ext4_mb_complex_scan_group(ac, &e4b);
2062
2063 ext4_unlock_group(sb, group);
2064 ext4_mb_unload_buddy(&e4b);
2065
2066 if (ac->ac_status != AC_STATUS_CONTINUE)
2067 break;
2068 }
2069 }
2070
2071 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2072 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2073 /*
2074 * We've been searching too long. Let's try to allocate
2075 * the best chunk we've found so far
2076 */
2077
2078 ext4_mb_try_best_found(ac, &e4b);
2079 if (ac->ac_status != AC_STATUS_FOUND) {
2080 /*
2081 * Someone more lucky has already allocated it.
2082 * The only thing we can do is just take first
2083 * found block(s)
2084 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2085 */
2086 ac->ac_b_ex.fe_group = 0;
2087 ac->ac_b_ex.fe_start = 0;
2088 ac->ac_b_ex.fe_len = 0;
2089 ac->ac_status = AC_STATUS_CONTINUE;
2090 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2091 cr = 3;
2092 atomic_inc(&sbi->s_mb_lost_chunks);
2093 goto repeat;
2094 }
2095 }
2096out:
2097 return err;
2098}
2099
2100static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2101{
2102 struct super_block *sb = seq->private;
2103 ext4_group_t group;
2104
2105 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2106 return NULL;
2107 group = *pos + 1;
2108 return (void *) ((unsigned long) group);
2109}
2110
2111static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2112{
2113 struct super_block *sb = seq->private;
2114 ext4_group_t group;
2115
2116 ++*pos;
2117 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2118 return NULL;
2119 group = *pos + 1;
2120 return (void *) ((unsigned long) group);
2121}
2122
2123static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2124{
2125 struct super_block *sb = seq->private;
2126 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2127 int i;
2128 int err;
2129 struct ext4_buddy e4b;
2130 struct sg {
2131 struct ext4_group_info info;
2132 ext4_grpblk_t counters[16];
2133 } sg;
2134
2135 group--;
2136 if (group == 0)
2137 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2138 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2139 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2140 "group", "free", "frags", "first",
2141 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2142 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2143
2144 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2145 sizeof(struct ext4_group_info);
2146 err = ext4_mb_load_buddy(sb, group, &e4b);
2147 if (err) {
2148 seq_printf(seq, "#%-5u: I/O error\n", group);
2149 return 0;
2150 }
2151 ext4_lock_group(sb, group);
2152 memcpy(&sg, ext4_get_group_info(sb, group), i);
2153 ext4_unlock_group(sb, group);
2154 ext4_mb_unload_buddy(&e4b);
2155
2156 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2157 sg.info.bb_fragments, sg.info.bb_first_free);
2158 for (i = 0; i <= 13; i++)
2159 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2160 sg.info.bb_counters[i] : 0);
2161 seq_printf(seq, " ]\n");
2162
2163 return 0;
2164}
2165
2166static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2167{
2168}
2169
2170static const struct seq_operations ext4_mb_seq_groups_ops = {
2171 .start = ext4_mb_seq_groups_start,
2172 .next = ext4_mb_seq_groups_next,
2173 .stop = ext4_mb_seq_groups_stop,
2174 .show = ext4_mb_seq_groups_show,
2175};
2176
2177static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2178{
2179 struct super_block *sb = PDE(inode)->data;
2180 int rc;
2181
2182 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2183 if (rc == 0) {
2184 struct seq_file *m = file->private_data;
2185 m->private = sb;
2186 }
2187 return rc;
2188
2189}
2190
2191static const struct file_operations ext4_mb_seq_groups_fops = {
2192 .owner = THIS_MODULE,
2193 .open = ext4_mb_seq_groups_open,
2194 .read = seq_read,
2195 .llseek = seq_lseek,
2196 .release = seq_release,
2197};
2198
2199static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2200{
2201 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2202 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2203
2204 BUG_ON(!cachep);
2205 return cachep;
2206}
2207
2208/* Create and initialize ext4_group_info data for the given group. */
2209int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2210 struct ext4_group_desc *desc)
2211{
2212 int i;
2213 int metalen = 0;
2214 struct ext4_sb_info *sbi = EXT4_SB(sb);
2215 struct ext4_group_info **meta_group_info;
2216 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2217
2218 /*
2219 * First check if this group is the first of a reserved block.
2220 * If it's true, we have to allocate a new table of pointers
2221 * to ext4_group_info structures
2222 */
2223 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2224 metalen = sizeof(*meta_group_info) <<
2225 EXT4_DESC_PER_BLOCK_BITS(sb);
2226 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2227 if (meta_group_info == NULL) {
2228 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem "
2229 "for a buddy group");
2230 goto exit_meta_group_info;
2231 }
2232 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2233 meta_group_info;
2234 }
2235
2236 meta_group_info =
2237 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2238 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2239
2240 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2241 if (meta_group_info[i] == NULL) {
2242 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem");
2243 goto exit_group_info;
2244 }
2245 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2246 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2247 &(meta_group_info[i]->bb_state));
2248
2249 /*
2250 * initialize bb_free to be able to skip
2251 * empty groups without initialization
2252 */
2253 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2254 meta_group_info[i]->bb_free =
2255 ext4_free_blocks_after_init(sb, group, desc);
2256 } else {
2257 meta_group_info[i]->bb_free =
2258 ext4_free_blks_count(sb, desc);
2259 }
2260
2261 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2262 init_rwsem(&meta_group_info[i]->alloc_sem);
2263 meta_group_info[i]->bb_free_root = RB_ROOT;
2264 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2265
2266#ifdef DOUBLE_CHECK
2267 {
2268 struct buffer_head *bh;
2269 meta_group_info[i]->bb_bitmap =
2270 kmalloc(sb->s_blocksize, GFP_KERNEL);
2271 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2272 bh = ext4_read_block_bitmap(sb, group);
2273 BUG_ON(bh == NULL);
2274 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2275 sb->s_blocksize);
2276 put_bh(bh);
2277 }
2278#endif
2279
2280 return 0;
2281
2282exit_group_info:
2283 /* If a meta_group_info table has been allocated, release it now */
2284 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2285 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2286 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2287 }
2288exit_meta_group_info:
2289 return -ENOMEM;
2290} /* ext4_mb_add_groupinfo */
2291
2292static int ext4_mb_init_backend(struct super_block *sb)
2293{
2294 ext4_group_t ngroups = ext4_get_groups_count(sb);
2295 ext4_group_t i;
2296 struct ext4_sb_info *sbi = EXT4_SB(sb);
2297 struct ext4_super_block *es = sbi->s_es;
2298 int num_meta_group_infos;
2299 int num_meta_group_infos_max;
2300 int array_size;
2301 struct ext4_group_desc *desc;
2302 struct kmem_cache *cachep;
2303
2304 /* This is the number of blocks used by GDT */
2305 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2306 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2307
2308 /*
2309 * This is the total number of blocks used by GDT including
2310 * the number of reserved blocks for GDT.
2311 * The s_group_info array is allocated with this value
2312 * to allow a clean online resize without a complex
2313 * manipulation of pointer.
2314 * The drawback is the unused memory when no resize
2315 * occurs but it's very low in terms of pages
2316 * (see comments below)
2317 * Need to handle this properly when META_BG resizing is allowed
2318 */
2319 num_meta_group_infos_max = num_meta_group_infos +
2320 le16_to_cpu(es->s_reserved_gdt_blocks);
2321
2322 /*
2323 * array_size is the size of s_group_info array. We round it
2324 * to the next power of two because this approximation is done
2325 * internally by kmalloc so we can have some more memory
2326 * for free here (e.g. may be used for META_BG resize).
2327 */
2328 array_size = 1;
2329 while (array_size < sizeof(*sbi->s_group_info) *
2330 num_meta_group_infos_max)
2331 array_size = array_size << 1;
2332 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2333 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2334 * So a two level scheme suffices for now. */
2335 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2336 if (sbi->s_group_info == NULL) {
2337 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2338 return -ENOMEM;
2339 }
2340 sbi->s_buddy_cache = new_inode(sb);
2341 if (sbi->s_buddy_cache == NULL) {
2342 ext4_msg(sb, KERN_ERR, "can't get new inode");
2343 goto err_freesgi;
2344 }
2345 /* To avoid potentially colliding with an valid on-disk inode number,
2346 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2347 * not in the inode hash, so it should never be found by iget(), but
2348 * this will avoid confusion if it ever shows up during debugging. */
2349 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2350 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2351 for (i = 0; i < ngroups; i++) {
2352 desc = ext4_get_group_desc(sb, i, NULL);
2353 if (desc == NULL) {
2354 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2355 goto err_freebuddy;
2356 }
2357 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2358 goto err_freebuddy;
2359 }
2360
2361 return 0;
2362
2363err_freebuddy:
2364 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2365 while (i-- > 0)
2366 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2367 i = num_meta_group_infos;
2368 while (i-- > 0)
2369 kfree(sbi->s_group_info[i]);
2370 iput(sbi->s_buddy_cache);
2371err_freesgi:
2372 ext4_kvfree(sbi->s_group_info);
2373 return -ENOMEM;
2374}
2375
2376static void ext4_groupinfo_destroy_slabs(void)
2377{
2378 int i;
2379
2380 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2381 if (ext4_groupinfo_caches[i])
2382 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2383 ext4_groupinfo_caches[i] = NULL;
2384 }
2385}
2386
2387static int ext4_groupinfo_create_slab(size_t size)
2388{
2389 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2390 int slab_size;
2391 int blocksize_bits = order_base_2(size);
2392 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2393 struct kmem_cache *cachep;
2394
2395 if (cache_index >= NR_GRPINFO_CACHES)
2396 return -EINVAL;
2397
2398 if (unlikely(cache_index < 0))
2399 cache_index = 0;
2400
2401 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2402 if (ext4_groupinfo_caches[cache_index]) {
2403 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2404 return 0; /* Already created */
2405 }
2406
2407 slab_size = offsetof(struct ext4_group_info,
2408 bb_counters[blocksize_bits + 2]);
2409
2410 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2411 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2412 NULL);
2413
2414 ext4_groupinfo_caches[cache_index] = cachep;
2415
2416 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2417 if (!cachep) {
2418 printk(KERN_EMERG
2419 "EXT4-fs: no memory for groupinfo slab cache\n");
2420 return -ENOMEM;
2421 }
2422
2423 return 0;
2424}
2425
2426int ext4_mb_init(struct super_block *sb, int needs_recovery)
2427{
2428 struct ext4_sb_info *sbi = EXT4_SB(sb);
2429 unsigned i, j;
2430 unsigned offset;
2431 unsigned max;
2432 int ret;
2433
2434 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2435
2436 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2437 if (sbi->s_mb_offsets == NULL) {
2438 ret = -ENOMEM;
2439 goto out;
2440 }
2441
2442 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2443 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2444 if (sbi->s_mb_maxs == NULL) {
2445 ret = -ENOMEM;
2446 goto out;
2447 }
2448
2449 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2450 if (ret < 0)
2451 goto out;
2452
2453 /* order 0 is regular bitmap */
2454 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2455 sbi->s_mb_offsets[0] = 0;
2456
2457 i = 1;
2458 offset = 0;
2459 max = sb->s_blocksize << 2;
2460 do {
2461 sbi->s_mb_offsets[i] = offset;
2462 sbi->s_mb_maxs[i] = max;
2463 offset += 1 << (sb->s_blocksize_bits - i);
2464 max = max >> 1;
2465 i++;
2466 } while (i <= sb->s_blocksize_bits + 1);
2467
2468 spin_lock_init(&sbi->s_md_lock);
2469 spin_lock_init(&sbi->s_bal_lock);
2470
2471 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2472 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2473 sbi->s_mb_stats = MB_DEFAULT_STATS;
2474 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2475 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2476 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2477 /*
2478 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2479 * to the lowest multiple of s_stripe which is bigger than
2480 * the s_mb_group_prealloc as determined above. We want
2481 * the preallocation size to be an exact multiple of the
2482 * RAID stripe size so that preallocations don't fragment
2483 * the stripes.
2484 */
2485 if (sbi->s_stripe > 1) {
2486 sbi->s_mb_group_prealloc = roundup(
2487 sbi->s_mb_group_prealloc, sbi->s_stripe);
2488 }
2489
2490 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2491 if (sbi->s_locality_groups == NULL) {
2492 ret = -ENOMEM;
2493 goto out;
2494 }
2495 for_each_possible_cpu(i) {
2496 struct ext4_locality_group *lg;
2497 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2498 mutex_init(&lg->lg_mutex);
2499 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2500 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2501 spin_lock_init(&lg->lg_prealloc_lock);
2502 }
2503
2504 /* init file for buddy data */
2505 ret = ext4_mb_init_backend(sb);
2506 if (ret != 0) {
2507 goto out;
2508 }
2509
2510 if (sbi->s_proc)
2511 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2512 &ext4_mb_seq_groups_fops, sb);
2513
2514 if (sbi->s_journal)
2515 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2516out:
2517 if (ret) {
2518 kfree(sbi->s_mb_offsets);
2519 kfree(sbi->s_mb_maxs);
2520 }
2521 return ret;
2522}
2523
2524/* need to called with the ext4 group lock held */
2525static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2526{
2527 struct ext4_prealloc_space *pa;
2528 struct list_head *cur, *tmp;
2529 int count = 0;
2530
2531 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2532 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2533 list_del(&pa->pa_group_list);
2534 count++;
2535 kmem_cache_free(ext4_pspace_cachep, pa);
2536 }
2537 if (count)
2538 mb_debug(1, "mballoc: %u PAs left\n", count);
2539
2540}
2541
2542int ext4_mb_release(struct super_block *sb)
2543{
2544 ext4_group_t ngroups = ext4_get_groups_count(sb);
2545 ext4_group_t i;
2546 int num_meta_group_infos;
2547 struct ext4_group_info *grinfo;
2548 struct ext4_sb_info *sbi = EXT4_SB(sb);
2549 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2550
2551 if (sbi->s_group_info) {
2552 for (i = 0; i < ngroups; i++) {
2553 grinfo = ext4_get_group_info(sb, i);
2554#ifdef DOUBLE_CHECK
2555 kfree(grinfo->bb_bitmap);
2556#endif
2557 ext4_lock_group(sb, i);
2558 ext4_mb_cleanup_pa(grinfo);
2559 ext4_unlock_group(sb, i);
2560 kmem_cache_free(cachep, grinfo);
2561 }
2562 num_meta_group_infos = (ngroups +
2563 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2564 EXT4_DESC_PER_BLOCK_BITS(sb);
2565 for (i = 0; i < num_meta_group_infos; i++)
2566 kfree(sbi->s_group_info[i]);
2567 ext4_kvfree(sbi->s_group_info);
2568 }
2569 kfree(sbi->s_mb_offsets);
2570 kfree(sbi->s_mb_maxs);
2571 if (sbi->s_buddy_cache)
2572 iput(sbi->s_buddy_cache);
2573 if (sbi->s_mb_stats) {
2574 ext4_msg(sb, KERN_INFO,
2575 "mballoc: %u blocks %u reqs (%u success)",
2576 atomic_read(&sbi->s_bal_allocated),
2577 atomic_read(&sbi->s_bal_reqs),
2578 atomic_read(&sbi->s_bal_success));
2579 ext4_msg(sb, KERN_INFO,
2580 "mballoc: %u extents scanned, %u goal hits, "
2581 "%u 2^N hits, %u breaks, %u lost",
2582 atomic_read(&sbi->s_bal_ex_scanned),
2583 atomic_read(&sbi->s_bal_goals),
2584 atomic_read(&sbi->s_bal_2orders),
2585 atomic_read(&sbi->s_bal_breaks),
2586 atomic_read(&sbi->s_mb_lost_chunks));
2587 ext4_msg(sb, KERN_INFO,
2588 "mballoc: %lu generated and it took %Lu",
2589 sbi->s_mb_buddies_generated,
2590 sbi->s_mb_generation_time);
2591 ext4_msg(sb, KERN_INFO,
2592 "mballoc: %u preallocated, %u discarded",
2593 atomic_read(&sbi->s_mb_preallocated),
2594 atomic_read(&sbi->s_mb_discarded));
2595 }
2596
2597 free_percpu(sbi->s_locality_groups);
2598 if (sbi->s_proc)
2599 remove_proc_entry("mb_groups", sbi->s_proc);
2600
2601 return 0;
2602}
2603
2604static inline int ext4_issue_discard(struct super_block *sb,
2605 ext4_group_t block_group, ext4_grpblk_t block, int count)
2606{
2607 ext4_fsblk_t discard_block;
2608
2609 discard_block = block + ext4_group_first_block_no(sb, block_group);
2610 trace_ext4_discard_blocks(sb,
2611 (unsigned long long) discard_block, count);
2612 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2613}
2614
2615/*
2616 * This function is called by the jbd2 layer once the commit has finished,
2617 * so we know we can free the blocks that were released with that commit.
2618 */
2619static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2620{
2621 struct super_block *sb = journal->j_private;
2622 struct ext4_buddy e4b;
2623 struct ext4_group_info *db;
2624 int err, count = 0, count2 = 0;
2625 struct ext4_free_data *entry;
2626 struct list_head *l, *ltmp;
2627
2628 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2629 entry = list_entry(l, struct ext4_free_data, list);
2630
2631 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2632 entry->count, entry->group, entry);
2633
2634 if (test_opt(sb, DISCARD))
2635 ext4_issue_discard(sb, entry->group,
2636 entry->start_blk, entry->count);
2637
2638 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2639 /* we expect to find existing buddy because it's pinned */
2640 BUG_ON(err != 0);
2641
2642 db = e4b.bd_info;
2643 /* there are blocks to put in buddy to make them really free */
2644 count += entry->count;
2645 count2++;
2646 ext4_lock_group(sb, entry->group);
2647 /* Take it out of per group rb tree */
2648 rb_erase(&entry->node, &(db->bb_free_root));
2649 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2650
2651 /*
2652 * Clear the trimmed flag for the group so that the next
2653 * ext4_trim_fs can trim it.
2654 * If the volume is mounted with -o discard, online discard
2655 * is supported and the free blocks will be trimmed online.
2656 */
2657 if (!test_opt(sb, DISCARD))
2658 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2659
2660 if (!db->bb_free_root.rb_node) {
2661 /* No more items in the per group rb tree
2662 * balance refcounts from ext4_mb_free_metadata()
2663 */
2664 page_cache_release(e4b.bd_buddy_page);
2665 page_cache_release(e4b.bd_bitmap_page);
2666 }
2667 ext4_unlock_group(sb, entry->group);
2668 kmem_cache_free(ext4_free_ext_cachep, entry);
2669 ext4_mb_unload_buddy(&e4b);
2670 }
2671
2672 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2673}
2674
2675#ifdef CONFIG_EXT4_DEBUG
2676u8 mb_enable_debug __read_mostly;
2677
2678static struct dentry *debugfs_dir;
2679static struct dentry *debugfs_debug;
2680
2681static void __init ext4_create_debugfs_entry(void)
2682{
2683 debugfs_dir = debugfs_create_dir("ext4", NULL);
2684 if (debugfs_dir)
2685 debugfs_debug = debugfs_create_u8("mballoc-debug",
2686 S_IRUGO | S_IWUSR,
2687 debugfs_dir,
2688 &mb_enable_debug);
2689}
2690
2691static void ext4_remove_debugfs_entry(void)
2692{
2693 debugfs_remove(debugfs_debug);
2694 debugfs_remove(debugfs_dir);
2695}
2696
2697#else
2698
2699static void __init ext4_create_debugfs_entry(void)
2700{
2701}
2702
2703static void ext4_remove_debugfs_entry(void)
2704{
2705}
2706
2707#endif
2708
2709int __init ext4_init_mballoc(void)
2710{
2711 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2712 SLAB_RECLAIM_ACCOUNT);
2713 if (ext4_pspace_cachep == NULL)
2714 return -ENOMEM;
2715
2716 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2717 SLAB_RECLAIM_ACCOUNT);
2718 if (ext4_ac_cachep == NULL) {
2719 kmem_cache_destroy(ext4_pspace_cachep);
2720 return -ENOMEM;
2721 }
2722
2723 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2724 SLAB_RECLAIM_ACCOUNT);
2725 if (ext4_free_ext_cachep == NULL) {
2726 kmem_cache_destroy(ext4_pspace_cachep);
2727 kmem_cache_destroy(ext4_ac_cachep);
2728 return -ENOMEM;
2729 }
2730 ext4_create_debugfs_entry();
2731 return 0;
2732}
2733
2734void ext4_exit_mballoc(void)
2735{
2736 /*
2737 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2738 * before destroying the slab cache.
2739 */
2740 rcu_barrier();
2741 kmem_cache_destroy(ext4_pspace_cachep);
2742 kmem_cache_destroy(ext4_ac_cachep);
2743 kmem_cache_destroy(ext4_free_ext_cachep);
2744 ext4_groupinfo_destroy_slabs();
2745 ext4_remove_debugfs_entry();
2746}
2747
2748
2749/*
2750 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2751 * Returns 0 if success or error code
2752 */
2753static noinline_for_stack int
2754ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2755 handle_t *handle, unsigned int reserv_blks)
2756{
2757 struct buffer_head *bitmap_bh = NULL;
2758 struct ext4_group_desc *gdp;
2759 struct buffer_head *gdp_bh;
2760 struct ext4_sb_info *sbi;
2761 struct super_block *sb;
2762 ext4_fsblk_t block;
2763 int err, len;
2764
2765 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2766 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2767
2768 sb = ac->ac_sb;
2769 sbi = EXT4_SB(sb);
2770
2771 err = -EIO;
2772 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2773 if (!bitmap_bh)
2774 goto out_err;
2775
2776 err = ext4_journal_get_write_access(handle, bitmap_bh);
2777 if (err)
2778 goto out_err;
2779
2780 err = -EIO;
2781 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2782 if (!gdp)
2783 goto out_err;
2784
2785 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2786 ext4_free_blks_count(sb, gdp));
2787
2788 err = ext4_journal_get_write_access(handle, gdp_bh);
2789 if (err)
2790 goto out_err;
2791
2792 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2793
2794 len = ac->ac_b_ex.fe_len;
2795 if (!ext4_data_block_valid(sbi, block, len)) {
2796 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2797 "fs metadata\n", block, block+len);
2798 /* File system mounted not to panic on error
2799 * Fix the bitmap and repeat the block allocation
2800 * We leak some of the blocks here.
2801 */
2802 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2803 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2804 ac->ac_b_ex.fe_len);
2805 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2806 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2807 if (!err)
2808 err = -EAGAIN;
2809 goto out_err;
2810 }
2811
2812 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2813#ifdef AGGRESSIVE_CHECK
2814 {
2815 int i;
2816 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2817 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2818 bitmap_bh->b_data));
2819 }
2820 }
2821#endif
2822 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2823 ac->ac_b_ex.fe_len);
2824 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2825 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2826 ext4_free_blks_set(sb, gdp,
2827 ext4_free_blocks_after_init(sb,
2828 ac->ac_b_ex.fe_group, gdp));
2829 }
2830 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2831 ext4_free_blks_set(sb, gdp, len);
2832 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2833
2834 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2835 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2836 /*
2837 * Now reduce the dirty block count also. Should not go negative
2838 */
2839 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2840 /* release all the reserved blocks if non delalloc */
2841 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2842
2843 if (sbi->s_log_groups_per_flex) {
2844 ext4_group_t flex_group = ext4_flex_group(sbi,
2845 ac->ac_b_ex.fe_group);
2846 atomic_sub(ac->ac_b_ex.fe_len,
2847 &sbi->s_flex_groups[flex_group].free_blocks);
2848 }
2849
2850 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2851 if (err)
2852 goto out_err;
2853 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2854
2855out_err:
2856 ext4_mark_super_dirty(sb);
2857 brelse(bitmap_bh);
2858 return err;
2859}
2860
2861/*
2862 * here we normalize request for locality group
2863 * Group request are normalized to s_mb_group_prealloc, which goes to
2864 * s_strip if we set the same via mount option.
2865 * s_mb_group_prealloc can be configured via
2866 * /sys/fs/ext4/<partition>/mb_group_prealloc
2867 *
2868 * XXX: should we try to preallocate more than the group has now?
2869 */
2870static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2871{
2872 struct super_block *sb = ac->ac_sb;
2873 struct ext4_locality_group *lg = ac->ac_lg;
2874
2875 BUG_ON(lg == NULL);
2876 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2877 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2878 current->pid, ac->ac_g_ex.fe_len);
2879}
2880
2881/*
2882 * Normalization means making request better in terms of
2883 * size and alignment
2884 */
2885static noinline_for_stack void
2886ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2887 struct ext4_allocation_request *ar)
2888{
2889 int bsbits, max;
2890 ext4_lblk_t end;
2891 loff_t size, orig_size, start_off;
2892 ext4_lblk_t start;
2893 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2894 struct ext4_prealloc_space *pa;
2895
2896 /* do normalize only data requests, metadata requests
2897 do not need preallocation */
2898 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2899 return;
2900
2901 /* sometime caller may want exact blocks */
2902 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2903 return;
2904
2905 /* caller may indicate that preallocation isn't
2906 * required (it's a tail, for example) */
2907 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2908 return;
2909
2910 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2911 ext4_mb_normalize_group_request(ac);
2912 return ;
2913 }
2914
2915 bsbits = ac->ac_sb->s_blocksize_bits;
2916
2917 /* first, let's learn actual file size
2918 * given current request is allocated */
2919 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2920 size = size << bsbits;
2921 if (size < i_size_read(ac->ac_inode))
2922 size = i_size_read(ac->ac_inode);
2923 orig_size = size;
2924
2925 /* max size of free chunks */
2926 max = 2 << bsbits;
2927
2928#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2929 (req <= (size) || max <= (chunk_size))
2930
2931 /* first, try to predict filesize */
2932 /* XXX: should this table be tunable? */
2933 start_off = 0;
2934 if (size <= 16 * 1024) {
2935 size = 16 * 1024;
2936 } else if (size <= 32 * 1024) {
2937 size = 32 * 1024;
2938 } else if (size <= 64 * 1024) {
2939 size = 64 * 1024;
2940 } else if (size <= 128 * 1024) {
2941 size = 128 * 1024;
2942 } else if (size <= 256 * 1024) {
2943 size = 256 * 1024;
2944 } else if (size <= 512 * 1024) {
2945 size = 512 * 1024;
2946 } else if (size <= 1024 * 1024) {
2947 size = 1024 * 1024;
2948 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2949 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2950 (21 - bsbits)) << 21;
2951 size = 2 * 1024 * 1024;
2952 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2953 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2954 (22 - bsbits)) << 22;
2955 size = 4 * 1024 * 1024;
2956 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2957 (8<<20)>>bsbits, max, 8 * 1024)) {
2958 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2959 (23 - bsbits)) << 23;
2960 size = 8 * 1024 * 1024;
2961 } else {
2962 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2963 size = ac->ac_o_ex.fe_len << bsbits;
2964 }
2965 size = size >> bsbits;
2966 start = start_off >> bsbits;
2967
2968 /* don't cover already allocated blocks in selected range */
2969 if (ar->pleft && start <= ar->lleft) {
2970 size -= ar->lleft + 1 - start;
2971 start = ar->lleft + 1;
2972 }
2973 if (ar->pright && start + size - 1 >= ar->lright)
2974 size -= start + size - ar->lright;
2975
2976 end = start + size;
2977
2978 /* check we don't cross already preallocated blocks */
2979 rcu_read_lock();
2980 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2981 ext4_lblk_t pa_end;
2982
2983 if (pa->pa_deleted)
2984 continue;
2985 spin_lock(&pa->pa_lock);
2986 if (pa->pa_deleted) {
2987 spin_unlock(&pa->pa_lock);
2988 continue;
2989 }
2990
2991 pa_end = pa->pa_lstart + pa->pa_len;
2992
2993 /* PA must not overlap original request */
2994 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2995 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2996
2997 /* skip PAs this normalized request doesn't overlap with */
2998 if (pa->pa_lstart >= end || pa_end <= start) {
2999 spin_unlock(&pa->pa_lock);
3000 continue;
3001 }
3002 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3003
3004 /* adjust start or end to be adjacent to this pa */
3005 if (pa_end <= ac->ac_o_ex.fe_logical) {
3006 BUG_ON(pa_end < start);
3007 start = pa_end;
3008 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3009 BUG_ON(pa->pa_lstart > end);
3010 end = pa->pa_lstart;
3011 }
3012 spin_unlock(&pa->pa_lock);
3013 }
3014 rcu_read_unlock();
3015 size = end - start;
3016
3017 /* XXX: extra loop to check we really don't overlap preallocations */
3018 rcu_read_lock();
3019 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3020 ext4_lblk_t pa_end;
3021 spin_lock(&pa->pa_lock);
3022 if (pa->pa_deleted == 0) {
3023 pa_end = pa->pa_lstart + pa->pa_len;
3024 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3025 }
3026 spin_unlock(&pa->pa_lock);
3027 }
3028 rcu_read_unlock();
3029
3030 if (start + size <= ac->ac_o_ex.fe_logical &&
3031 start > ac->ac_o_ex.fe_logical) {
3032 ext4_msg(ac->ac_sb, KERN_ERR,
3033 "start %lu, size %lu, fe_logical %lu",
3034 (unsigned long) start, (unsigned long) size,
3035 (unsigned long) ac->ac_o_ex.fe_logical);
3036 }
3037 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3038 start > ac->ac_o_ex.fe_logical);
3039 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3040
3041 /* now prepare goal request */
3042
3043 /* XXX: is it better to align blocks WRT to logical
3044 * placement or satisfy big request as is */
3045 ac->ac_g_ex.fe_logical = start;
3046 ac->ac_g_ex.fe_len = size;
3047
3048 /* define goal start in order to merge */
3049 if (ar->pright && (ar->lright == (start + size))) {
3050 /* merge to the right */
3051 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3052 &ac->ac_f_ex.fe_group,
3053 &ac->ac_f_ex.fe_start);
3054 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3055 }
3056 if (ar->pleft && (ar->lleft + 1 == start)) {
3057 /* merge to the left */
3058 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3059 &ac->ac_f_ex.fe_group,
3060 &ac->ac_f_ex.fe_start);
3061 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3062 }
3063
3064 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3065 (unsigned) orig_size, (unsigned) start);
3066}
3067
3068static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3069{
3070 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3071
3072 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3073 atomic_inc(&sbi->s_bal_reqs);
3074 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3075 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3076 atomic_inc(&sbi->s_bal_success);
3077 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3078 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3079 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3080 atomic_inc(&sbi->s_bal_goals);
3081 if (ac->ac_found > sbi->s_mb_max_to_scan)
3082 atomic_inc(&sbi->s_bal_breaks);
3083 }
3084
3085 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3086 trace_ext4_mballoc_alloc(ac);
3087 else
3088 trace_ext4_mballoc_prealloc(ac);
3089}
3090
3091/*
3092 * Called on failure; free up any blocks from the inode PA for this
3093 * context. We don't need this for MB_GROUP_PA because we only change
3094 * pa_free in ext4_mb_release_context(), but on failure, we've already
3095 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3096 */
3097static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3098{
3099 struct ext4_prealloc_space *pa = ac->ac_pa;
3100 int len;
3101
3102 if (pa && pa->pa_type == MB_INODE_PA) {
3103 len = ac->ac_b_ex.fe_len;
3104 pa->pa_free += len;
3105 }
3106
3107}
3108
3109/*
3110 * use blocks preallocated to inode
3111 */
3112static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3113 struct ext4_prealloc_space *pa)
3114{
3115 ext4_fsblk_t start;
3116 ext4_fsblk_t end;
3117 int len;
3118
3119 /* found preallocated blocks, use them */
3120 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3121 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3122 len = end - start;
3123 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3124 &ac->ac_b_ex.fe_start);
3125 ac->ac_b_ex.fe_len = len;
3126 ac->ac_status = AC_STATUS_FOUND;
3127 ac->ac_pa = pa;
3128
3129 BUG_ON(start < pa->pa_pstart);
3130 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3131 BUG_ON(pa->pa_free < len);
3132 pa->pa_free -= len;
3133
3134 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3135}
3136
3137/*
3138 * use blocks preallocated to locality group
3139 */
3140static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3141 struct ext4_prealloc_space *pa)
3142{
3143 unsigned int len = ac->ac_o_ex.fe_len;
3144
3145 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3146 &ac->ac_b_ex.fe_group,
3147 &ac->ac_b_ex.fe_start);
3148 ac->ac_b_ex.fe_len = len;
3149 ac->ac_status = AC_STATUS_FOUND;
3150 ac->ac_pa = pa;
3151
3152 /* we don't correct pa_pstart or pa_plen here to avoid
3153 * possible race when the group is being loaded concurrently
3154 * instead we correct pa later, after blocks are marked
3155 * in on-disk bitmap -- see ext4_mb_release_context()
3156 * Other CPUs are prevented from allocating from this pa by lg_mutex
3157 */
3158 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3159}
3160
3161/*
3162 * Return the prealloc space that have minimal distance
3163 * from the goal block. @cpa is the prealloc
3164 * space that is having currently known minimal distance
3165 * from the goal block.
3166 */
3167static struct ext4_prealloc_space *
3168ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3169 struct ext4_prealloc_space *pa,
3170 struct ext4_prealloc_space *cpa)
3171{
3172 ext4_fsblk_t cur_distance, new_distance;
3173
3174 if (cpa == NULL) {
3175 atomic_inc(&pa->pa_count);
3176 return pa;
3177 }
3178 cur_distance = abs(goal_block - cpa->pa_pstart);
3179 new_distance = abs(goal_block - pa->pa_pstart);
3180
3181 if (cur_distance <= new_distance)
3182 return cpa;
3183
3184 /* drop the previous reference */
3185 atomic_dec(&cpa->pa_count);
3186 atomic_inc(&pa->pa_count);
3187 return pa;
3188}
3189
3190/*
3191 * search goal blocks in preallocated space
3192 */
3193static noinline_for_stack int
3194ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3195{
3196 int order, i;
3197 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3198 struct ext4_locality_group *lg;
3199 struct ext4_prealloc_space *pa, *cpa = NULL;
3200 ext4_fsblk_t goal_block;
3201
3202 /* only data can be preallocated */
3203 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3204 return 0;
3205
3206 /* first, try per-file preallocation */
3207 rcu_read_lock();
3208 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3209
3210 /* all fields in this condition don't change,
3211 * so we can skip locking for them */
3212 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3213 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3214 continue;
3215
3216 /* non-extent files can't have physical blocks past 2^32 */
3217 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3218 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3219 continue;
3220
3221 /* found preallocated blocks, use them */
3222 spin_lock(&pa->pa_lock);
3223 if (pa->pa_deleted == 0 && pa->pa_free) {
3224 atomic_inc(&pa->pa_count);
3225 ext4_mb_use_inode_pa(ac, pa);
3226 spin_unlock(&pa->pa_lock);
3227 ac->ac_criteria = 10;
3228 rcu_read_unlock();
3229 return 1;
3230 }
3231 spin_unlock(&pa->pa_lock);
3232 }
3233 rcu_read_unlock();
3234
3235 /* can we use group allocation? */
3236 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3237 return 0;
3238
3239 /* inode may have no locality group for some reason */
3240 lg = ac->ac_lg;
3241 if (lg == NULL)
3242 return 0;
3243 order = fls(ac->ac_o_ex.fe_len) - 1;
3244 if (order > PREALLOC_TB_SIZE - 1)
3245 /* The max size of hash table is PREALLOC_TB_SIZE */
3246 order = PREALLOC_TB_SIZE - 1;
3247
3248 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3249 /*
3250 * search for the prealloc space that is having
3251 * minimal distance from the goal block.
3252 */
3253 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3254 rcu_read_lock();
3255 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3256 pa_inode_list) {
3257 spin_lock(&pa->pa_lock);
3258 if (pa->pa_deleted == 0 &&
3259 pa->pa_free >= ac->ac_o_ex.fe_len) {
3260
3261 cpa = ext4_mb_check_group_pa(goal_block,
3262 pa, cpa);
3263 }
3264 spin_unlock(&pa->pa_lock);
3265 }
3266 rcu_read_unlock();
3267 }
3268 if (cpa) {
3269 ext4_mb_use_group_pa(ac, cpa);
3270 ac->ac_criteria = 20;
3271 return 1;
3272 }
3273 return 0;
3274}
3275
3276/*
3277 * the function goes through all block freed in the group
3278 * but not yet committed and marks them used in in-core bitmap.
3279 * buddy must be generated from this bitmap
3280 * Need to be called with the ext4 group lock held
3281 */
3282static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3283 ext4_group_t group)
3284{
3285 struct rb_node *n;
3286 struct ext4_group_info *grp;
3287 struct ext4_free_data *entry;
3288
3289 grp = ext4_get_group_info(sb, group);
3290 n = rb_first(&(grp->bb_free_root));
3291
3292 while (n) {
3293 entry = rb_entry(n, struct ext4_free_data, node);
3294 ext4_set_bits(bitmap, entry->start_blk, entry->count);
3295 n = rb_next(n);
3296 }
3297 return;
3298}
3299
3300/*
3301 * the function goes through all preallocation in this group and marks them
3302 * used in in-core bitmap. buddy must be generated from this bitmap
3303 * Need to be called with ext4 group lock held
3304 */
3305static noinline_for_stack
3306void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3307 ext4_group_t group)
3308{
3309 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3310 struct ext4_prealloc_space *pa;
3311 struct list_head *cur;
3312 ext4_group_t groupnr;
3313 ext4_grpblk_t start;
3314 int preallocated = 0;
3315 int count = 0;
3316 int len;
3317
3318 /* all form of preallocation discards first load group,
3319 * so the only competing code is preallocation use.
3320 * we don't need any locking here
3321 * notice we do NOT ignore preallocations with pa_deleted
3322 * otherwise we could leave used blocks available for
3323 * allocation in buddy when concurrent ext4_mb_put_pa()
3324 * is dropping preallocation
3325 */
3326 list_for_each(cur, &grp->bb_prealloc_list) {
3327 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3328 spin_lock(&pa->pa_lock);
3329 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3330 &groupnr, &start);
3331 len = pa->pa_len;
3332 spin_unlock(&pa->pa_lock);
3333 if (unlikely(len == 0))
3334 continue;
3335 BUG_ON(groupnr != group);
3336 ext4_set_bits(bitmap, start, len);
3337 preallocated += len;
3338 count++;
3339 }
3340 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3341}
3342
3343static void ext4_mb_pa_callback(struct rcu_head *head)
3344{
3345 struct ext4_prealloc_space *pa;
3346 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3347 kmem_cache_free(ext4_pspace_cachep, pa);
3348}
3349
3350/*
3351 * drops a reference to preallocated space descriptor
3352 * if this was the last reference and the space is consumed
3353 */
3354static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3355 struct super_block *sb, struct ext4_prealloc_space *pa)
3356{
3357 ext4_group_t grp;
3358 ext4_fsblk_t grp_blk;
3359
3360 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3361 return;
3362
3363 /* in this short window concurrent discard can set pa_deleted */
3364 spin_lock(&pa->pa_lock);
3365 if (pa->pa_deleted == 1) {
3366 spin_unlock(&pa->pa_lock);
3367 return;
3368 }
3369
3370 pa->pa_deleted = 1;
3371 spin_unlock(&pa->pa_lock);
3372
3373 grp_blk = pa->pa_pstart;
3374 /*
3375 * If doing group-based preallocation, pa_pstart may be in the
3376 * next group when pa is used up
3377 */
3378 if (pa->pa_type == MB_GROUP_PA)
3379 grp_blk--;
3380
3381 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3382
3383 /*
3384 * possible race:
3385 *
3386 * P1 (buddy init) P2 (regular allocation)
3387 * find block B in PA
3388 * copy on-disk bitmap to buddy
3389 * mark B in on-disk bitmap
3390 * drop PA from group
3391 * mark all PAs in buddy
3392 *
3393 * thus, P1 initializes buddy with B available. to prevent this
3394 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3395 * against that pair
3396 */
3397 ext4_lock_group(sb, grp);
3398 list_del(&pa->pa_group_list);
3399 ext4_unlock_group(sb, grp);
3400
3401 spin_lock(pa->pa_obj_lock);
3402 list_del_rcu(&pa->pa_inode_list);
3403 spin_unlock(pa->pa_obj_lock);
3404
3405 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3406}
3407
3408/*
3409 * creates new preallocated space for given inode
3410 */
3411static noinline_for_stack int
3412ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3413{
3414 struct super_block *sb = ac->ac_sb;
3415 struct ext4_prealloc_space *pa;
3416 struct ext4_group_info *grp;
3417 struct ext4_inode_info *ei;
3418
3419 /* preallocate only when found space is larger then requested */
3420 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3421 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3422 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3423
3424 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3425 if (pa == NULL)
3426 return -ENOMEM;
3427
3428 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3429 int winl;
3430 int wins;
3431 int win;
3432 int offs;
3433
3434 /* we can't allocate as much as normalizer wants.
3435 * so, found space must get proper lstart
3436 * to cover original request */
3437 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3438 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3439
3440 /* we're limited by original request in that
3441 * logical block must be covered any way
3442 * winl is window we can move our chunk within */
3443 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3444
3445 /* also, we should cover whole original request */
3446 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3447
3448 /* the smallest one defines real window */
3449 win = min(winl, wins);
3450
3451 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3452 if (offs && offs < win)
3453 win = offs;
3454
3455 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3456 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3457 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3458 }
3459
3460 /* preallocation can change ac_b_ex, thus we store actually
3461 * allocated blocks for history */
3462 ac->ac_f_ex = ac->ac_b_ex;
3463
3464 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3465 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3466 pa->pa_len = ac->ac_b_ex.fe_len;
3467 pa->pa_free = pa->pa_len;
3468 atomic_set(&pa->pa_count, 1);
3469 spin_lock_init(&pa->pa_lock);
3470 INIT_LIST_HEAD(&pa->pa_inode_list);
3471 INIT_LIST_HEAD(&pa->pa_group_list);
3472 pa->pa_deleted = 0;
3473 pa->pa_type = MB_INODE_PA;
3474
3475 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3476 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3477 trace_ext4_mb_new_inode_pa(ac, pa);
3478
3479 ext4_mb_use_inode_pa(ac, pa);
3480 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3481
3482 ei = EXT4_I(ac->ac_inode);
3483 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3484
3485 pa->pa_obj_lock = &ei->i_prealloc_lock;
3486 pa->pa_inode = ac->ac_inode;
3487
3488 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3489 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3490 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3491
3492 spin_lock(pa->pa_obj_lock);
3493 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3494 spin_unlock(pa->pa_obj_lock);
3495
3496 return 0;
3497}
3498
3499/*
3500 * creates new preallocated space for locality group inodes belongs to
3501 */
3502static noinline_for_stack int
3503ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3504{
3505 struct super_block *sb = ac->ac_sb;
3506 struct ext4_locality_group *lg;
3507 struct ext4_prealloc_space *pa;
3508 struct ext4_group_info *grp;
3509
3510 /* preallocate only when found space is larger then requested */
3511 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3512 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3513 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3514
3515 BUG_ON(ext4_pspace_cachep == NULL);
3516 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3517 if (pa == NULL)
3518 return -ENOMEM;
3519
3520 /* preallocation can change ac_b_ex, thus we store actually
3521 * allocated blocks for history */
3522 ac->ac_f_ex = ac->ac_b_ex;
3523
3524 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3525 pa->pa_lstart = pa->pa_pstart;
3526 pa->pa_len = ac->ac_b_ex.fe_len;
3527 pa->pa_free = pa->pa_len;
3528 atomic_set(&pa->pa_count, 1);
3529 spin_lock_init(&pa->pa_lock);
3530 INIT_LIST_HEAD(&pa->pa_inode_list);
3531 INIT_LIST_HEAD(&pa->pa_group_list);
3532 pa->pa_deleted = 0;
3533 pa->pa_type = MB_GROUP_PA;
3534
3535 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3536 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3537 trace_ext4_mb_new_group_pa(ac, pa);
3538
3539 ext4_mb_use_group_pa(ac, pa);
3540 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3541
3542 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3543 lg = ac->ac_lg;
3544 BUG_ON(lg == NULL);
3545
3546 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3547 pa->pa_inode = NULL;
3548
3549 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3550 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3551 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3552
3553 /*
3554 * We will later add the new pa to the right bucket
3555 * after updating the pa_free in ext4_mb_release_context
3556 */
3557 return 0;
3558}
3559
3560static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3561{
3562 int err;
3563
3564 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3565 err = ext4_mb_new_group_pa(ac);
3566 else
3567 err = ext4_mb_new_inode_pa(ac);
3568 return err;
3569}
3570
3571/*
3572 * finds all unused blocks in on-disk bitmap, frees them in
3573 * in-core bitmap and buddy.
3574 * @pa must be unlinked from inode and group lists, so that
3575 * nobody else can find/use it.
3576 * the caller MUST hold group/inode locks.
3577 * TODO: optimize the case when there are no in-core structures yet
3578 */
3579static noinline_for_stack int
3580ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3581 struct ext4_prealloc_space *pa)
3582{
3583 struct super_block *sb = e4b->bd_sb;
3584 struct ext4_sb_info *sbi = EXT4_SB(sb);
3585 unsigned int end;
3586 unsigned int next;
3587 ext4_group_t group;
3588 ext4_grpblk_t bit;
3589 unsigned long long grp_blk_start;
3590 int err = 0;
3591 int free = 0;
3592
3593 BUG_ON(pa->pa_deleted == 0);
3594 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3595 grp_blk_start = pa->pa_pstart - bit;
3596 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3597 end = bit + pa->pa_len;
3598
3599 while (bit < end) {
3600 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3601 if (bit >= end)
3602 break;
3603 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3604 mb_debug(1, " free preallocated %u/%u in group %u\n",
3605 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3606 (unsigned) next - bit, (unsigned) group);
3607 free += next - bit;
3608
3609 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3610 trace_ext4_mb_release_inode_pa(pa, grp_blk_start + bit,
3611 next - bit);
3612 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3613 bit = next + 1;
3614 }
3615 if (free != pa->pa_free) {
3616 ext4_msg(e4b->bd_sb, KERN_CRIT,
3617 "pa %p: logic %lu, phys. %lu, len %lu",
3618 pa, (unsigned long) pa->pa_lstart,
3619 (unsigned long) pa->pa_pstart,
3620 (unsigned long) pa->pa_len);
3621 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3622 free, pa->pa_free);
3623 /*
3624 * pa is already deleted so we use the value obtained
3625 * from the bitmap and continue.
3626 */
3627 }
3628 atomic_add(free, &sbi->s_mb_discarded);
3629
3630 return err;
3631}
3632
3633static noinline_for_stack int
3634ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3635 struct ext4_prealloc_space *pa)
3636{
3637 struct super_block *sb = e4b->bd_sb;
3638 ext4_group_t group;
3639 ext4_grpblk_t bit;
3640
3641 trace_ext4_mb_release_group_pa(pa);
3642 BUG_ON(pa->pa_deleted == 0);
3643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3644 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3645 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3646 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3647 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3648
3649 return 0;
3650}
3651
3652/*
3653 * releases all preallocations in given group
3654 *
3655 * first, we need to decide discard policy:
3656 * - when do we discard
3657 * 1) ENOSPC
3658 * - how many do we discard
3659 * 1) how many requested
3660 */
3661static noinline_for_stack int
3662ext4_mb_discard_group_preallocations(struct super_block *sb,
3663 ext4_group_t group, int needed)
3664{
3665 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3666 struct buffer_head *bitmap_bh = NULL;
3667 struct ext4_prealloc_space *pa, *tmp;
3668 struct list_head list;
3669 struct ext4_buddy e4b;
3670 int err;
3671 int busy = 0;
3672 int free = 0;
3673
3674 mb_debug(1, "discard preallocation for group %u\n", group);
3675
3676 if (list_empty(&grp->bb_prealloc_list))
3677 return 0;
3678
3679 bitmap_bh = ext4_read_block_bitmap(sb, group);
3680 if (bitmap_bh == NULL) {
3681 ext4_error(sb, "Error reading block bitmap for %u", group);
3682 return 0;
3683 }
3684
3685 err = ext4_mb_load_buddy(sb, group, &e4b);
3686 if (err) {
3687 ext4_error(sb, "Error loading buddy information for %u", group);
3688 put_bh(bitmap_bh);
3689 return 0;
3690 }
3691
3692 if (needed == 0)
3693 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3694
3695 INIT_LIST_HEAD(&list);
3696repeat:
3697 ext4_lock_group(sb, group);
3698 list_for_each_entry_safe(pa, tmp,
3699 &grp->bb_prealloc_list, pa_group_list) {
3700 spin_lock(&pa->pa_lock);
3701 if (atomic_read(&pa->pa_count)) {
3702 spin_unlock(&pa->pa_lock);
3703 busy = 1;
3704 continue;
3705 }
3706 if (pa->pa_deleted) {
3707 spin_unlock(&pa->pa_lock);
3708 continue;
3709 }
3710
3711 /* seems this one can be freed ... */
3712 pa->pa_deleted = 1;
3713
3714 /* we can trust pa_free ... */
3715 free += pa->pa_free;
3716
3717 spin_unlock(&pa->pa_lock);
3718
3719 list_del(&pa->pa_group_list);
3720 list_add(&pa->u.pa_tmp_list, &list);
3721 }
3722
3723 /* if we still need more blocks and some PAs were used, try again */
3724 if (free < needed && busy) {
3725 busy = 0;
3726 ext4_unlock_group(sb, group);
3727 /*
3728 * Yield the CPU here so that we don't get soft lockup
3729 * in non preempt case.
3730 */
3731 yield();
3732 goto repeat;
3733 }
3734
3735 /* found anything to free? */
3736 if (list_empty(&list)) {
3737 BUG_ON(free != 0);
3738 goto out;
3739 }
3740
3741 /* now free all selected PAs */
3742 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3743
3744 /* remove from object (inode or locality group) */
3745 spin_lock(pa->pa_obj_lock);
3746 list_del_rcu(&pa->pa_inode_list);
3747 spin_unlock(pa->pa_obj_lock);
3748
3749 if (pa->pa_type == MB_GROUP_PA)
3750 ext4_mb_release_group_pa(&e4b, pa);
3751 else
3752 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3753
3754 list_del(&pa->u.pa_tmp_list);
3755 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3756 }
3757
3758out:
3759 ext4_unlock_group(sb, group);
3760 ext4_mb_unload_buddy(&e4b);
3761 put_bh(bitmap_bh);
3762 return free;
3763}
3764
3765/*
3766 * releases all non-used preallocated blocks for given inode
3767 *
3768 * It's important to discard preallocations under i_data_sem
3769 * We don't want another block to be served from the prealloc
3770 * space when we are discarding the inode prealloc space.
3771 *
3772 * FIXME!! Make sure it is valid at all the call sites
3773 */
3774void ext4_discard_preallocations(struct inode *inode)
3775{
3776 struct ext4_inode_info *ei = EXT4_I(inode);
3777 struct super_block *sb = inode->i_sb;
3778 struct buffer_head *bitmap_bh = NULL;
3779 struct ext4_prealloc_space *pa, *tmp;
3780 ext4_group_t group = 0;
3781 struct list_head list;
3782 struct ext4_buddy e4b;
3783 int err;
3784
3785 if (!S_ISREG(inode->i_mode)) {
3786 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3787 return;
3788 }
3789
3790 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3791 trace_ext4_discard_preallocations(inode);
3792
3793 INIT_LIST_HEAD(&list);
3794
3795repeat:
3796 /* first, collect all pa's in the inode */
3797 spin_lock(&ei->i_prealloc_lock);
3798 while (!list_empty(&ei->i_prealloc_list)) {
3799 pa = list_entry(ei->i_prealloc_list.next,
3800 struct ext4_prealloc_space, pa_inode_list);
3801 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3802 spin_lock(&pa->pa_lock);
3803 if (atomic_read(&pa->pa_count)) {
3804 /* this shouldn't happen often - nobody should
3805 * use preallocation while we're discarding it */
3806 spin_unlock(&pa->pa_lock);
3807 spin_unlock(&ei->i_prealloc_lock);
3808 ext4_msg(sb, KERN_ERR,
3809 "uh-oh! used pa while discarding");
3810 WARN_ON(1);
3811 schedule_timeout_uninterruptible(HZ);
3812 goto repeat;
3813
3814 }
3815 if (pa->pa_deleted == 0) {
3816 pa->pa_deleted = 1;
3817 spin_unlock(&pa->pa_lock);
3818 list_del_rcu(&pa->pa_inode_list);
3819 list_add(&pa->u.pa_tmp_list, &list);
3820 continue;
3821 }
3822
3823 /* someone is deleting pa right now */
3824 spin_unlock(&pa->pa_lock);
3825 spin_unlock(&ei->i_prealloc_lock);
3826
3827 /* we have to wait here because pa_deleted
3828 * doesn't mean pa is already unlinked from
3829 * the list. as we might be called from
3830 * ->clear_inode() the inode will get freed
3831 * and concurrent thread which is unlinking
3832 * pa from inode's list may access already
3833 * freed memory, bad-bad-bad */
3834
3835 /* XXX: if this happens too often, we can
3836 * add a flag to force wait only in case
3837 * of ->clear_inode(), but not in case of
3838 * regular truncate */
3839 schedule_timeout_uninterruptible(HZ);
3840 goto repeat;
3841 }
3842 spin_unlock(&ei->i_prealloc_lock);
3843
3844 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3845 BUG_ON(pa->pa_type != MB_INODE_PA);
3846 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3847
3848 err = ext4_mb_load_buddy(sb, group, &e4b);
3849 if (err) {
3850 ext4_error(sb, "Error loading buddy information for %u",
3851 group);
3852 continue;
3853 }
3854
3855 bitmap_bh = ext4_read_block_bitmap(sb, group);
3856 if (bitmap_bh == NULL) {
3857 ext4_error(sb, "Error reading block bitmap for %u",
3858 group);
3859 ext4_mb_unload_buddy(&e4b);
3860 continue;
3861 }
3862
3863 ext4_lock_group(sb, group);
3864 list_del(&pa->pa_group_list);
3865 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3866 ext4_unlock_group(sb, group);
3867
3868 ext4_mb_unload_buddy(&e4b);
3869 put_bh(bitmap_bh);
3870
3871 list_del(&pa->u.pa_tmp_list);
3872 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3873 }
3874}
3875
3876#ifdef CONFIG_EXT4_DEBUG
3877static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3878{
3879 struct super_block *sb = ac->ac_sb;
3880 ext4_group_t ngroups, i;
3881
3882 if (!mb_enable_debug ||
3883 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3884 return;
3885
3886 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:"
3887 " Allocation context details:");
3888 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d",
3889 ac->ac_status, ac->ac_flags);
3890 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, "
3891 "goal %lu/%lu/%lu@%lu, "
3892 "best %lu/%lu/%lu@%lu cr %d",
3893 (unsigned long)ac->ac_o_ex.fe_group,
3894 (unsigned long)ac->ac_o_ex.fe_start,
3895 (unsigned long)ac->ac_o_ex.fe_len,
3896 (unsigned long)ac->ac_o_ex.fe_logical,
3897 (unsigned long)ac->ac_g_ex.fe_group,
3898 (unsigned long)ac->ac_g_ex.fe_start,
3899 (unsigned long)ac->ac_g_ex.fe_len,
3900 (unsigned long)ac->ac_g_ex.fe_logical,
3901 (unsigned long)ac->ac_b_ex.fe_group,
3902 (unsigned long)ac->ac_b_ex.fe_start,
3903 (unsigned long)ac->ac_b_ex.fe_len,
3904 (unsigned long)ac->ac_b_ex.fe_logical,
3905 (int)ac->ac_criteria);
3906 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found",
3907 ac->ac_ex_scanned, ac->ac_found);
3908 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: ");
3909 ngroups = ext4_get_groups_count(sb);
3910 for (i = 0; i < ngroups; i++) {
3911 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3912 struct ext4_prealloc_space *pa;
3913 ext4_grpblk_t start;
3914 struct list_head *cur;
3915 ext4_lock_group(sb, i);
3916 list_for_each(cur, &grp->bb_prealloc_list) {
3917 pa = list_entry(cur, struct ext4_prealloc_space,
3918 pa_group_list);
3919 spin_lock(&pa->pa_lock);
3920 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3921 NULL, &start);
3922 spin_unlock(&pa->pa_lock);
3923 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3924 start, pa->pa_len);
3925 }
3926 ext4_unlock_group(sb, i);
3927
3928 if (grp->bb_free == 0)
3929 continue;
3930 printk(KERN_ERR "%u: %d/%d \n",
3931 i, grp->bb_free, grp->bb_fragments);
3932 }
3933 printk(KERN_ERR "\n");
3934}
3935#else
3936static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3937{
3938 return;
3939}
3940#endif
3941
3942/*
3943 * We use locality group preallocation for small size file. The size of the
3944 * file is determined by the current size or the resulting size after
3945 * allocation which ever is larger
3946 *
3947 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3948 */
3949static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3950{
3951 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3952 int bsbits = ac->ac_sb->s_blocksize_bits;
3953 loff_t size, isize;
3954
3955 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3956 return;
3957
3958 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3959 return;
3960
3961 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3962 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3963 >> bsbits;
3964
3965 if ((size == isize) &&
3966 !ext4_fs_is_busy(sbi) &&
3967 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3968 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3969 return;
3970 }
3971
3972 /* don't use group allocation for large files */
3973 size = max(size, isize);
3974 if (size > sbi->s_mb_stream_request) {
3975 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3976 return;
3977 }
3978
3979 BUG_ON(ac->ac_lg != NULL);
3980 /*
3981 * locality group prealloc space are per cpu. The reason for having
3982 * per cpu locality group is to reduce the contention between block
3983 * request from multiple CPUs.
3984 */
3985 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3986
3987 /* we're going to use group allocation */
3988 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3989
3990 /* serialize all allocations in the group */
3991 mutex_lock(&ac->ac_lg->lg_mutex);
3992}
3993
3994static noinline_for_stack int
3995ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3996 struct ext4_allocation_request *ar)
3997{
3998 struct super_block *sb = ar->inode->i_sb;
3999 struct ext4_sb_info *sbi = EXT4_SB(sb);
4000 struct ext4_super_block *es = sbi->s_es;
4001 ext4_group_t group;
4002 unsigned int len;
4003 ext4_fsblk_t goal;
4004 ext4_grpblk_t block;
4005
4006 /* we can't allocate > group size */
4007 len = ar->len;
4008
4009 /* just a dirty hack to filter too big requests */
4010 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4011 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4012
4013 /* start searching from the goal */
4014 goal = ar->goal;
4015 if (goal < le32_to_cpu(es->s_first_data_block) ||
4016 goal >= ext4_blocks_count(es))
4017 goal = le32_to_cpu(es->s_first_data_block);
4018 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4019
4020 /* set up allocation goals */
4021 memset(ac, 0, sizeof(struct ext4_allocation_context));
4022 ac->ac_b_ex.fe_logical = ar->logical;
4023 ac->ac_status = AC_STATUS_CONTINUE;
4024 ac->ac_sb = sb;
4025 ac->ac_inode = ar->inode;
4026 ac->ac_o_ex.fe_logical = ar->logical;
4027 ac->ac_o_ex.fe_group = group;
4028 ac->ac_o_ex.fe_start = block;
4029 ac->ac_o_ex.fe_len = len;
4030 ac->ac_g_ex.fe_logical = ar->logical;
4031 ac->ac_g_ex.fe_group = group;
4032 ac->ac_g_ex.fe_start = block;
4033 ac->ac_g_ex.fe_len = len;
4034 ac->ac_flags = ar->flags;
4035
4036 /* we have to define context: we'll we work with a file or
4037 * locality group. this is a policy, actually */
4038 ext4_mb_group_or_file(ac);
4039
4040 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4041 "left: %u/%u, right %u/%u to %swritable\n",
4042 (unsigned) ar->len, (unsigned) ar->logical,
4043 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4044 (unsigned) ar->lleft, (unsigned) ar->pleft,
4045 (unsigned) ar->lright, (unsigned) ar->pright,
4046 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4047 return 0;
4048
4049}
4050
4051static noinline_for_stack void
4052ext4_mb_discard_lg_preallocations(struct super_block *sb,
4053 struct ext4_locality_group *lg,
4054 int order, int total_entries)
4055{
4056 ext4_group_t group = 0;
4057 struct ext4_buddy e4b;
4058 struct list_head discard_list;
4059 struct ext4_prealloc_space *pa, *tmp;
4060
4061 mb_debug(1, "discard locality group preallocation\n");
4062
4063 INIT_LIST_HEAD(&discard_list);
4064
4065 spin_lock(&lg->lg_prealloc_lock);
4066 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4067 pa_inode_list) {
4068 spin_lock(&pa->pa_lock);
4069 if (atomic_read(&pa->pa_count)) {
4070 /*
4071 * This is the pa that we just used
4072 * for block allocation. So don't
4073 * free that
4074 */
4075 spin_unlock(&pa->pa_lock);
4076 continue;
4077 }
4078 if (pa->pa_deleted) {
4079 spin_unlock(&pa->pa_lock);
4080 continue;
4081 }
4082 /* only lg prealloc space */
4083 BUG_ON(pa->pa_type != MB_GROUP_PA);
4084
4085 /* seems this one can be freed ... */
4086 pa->pa_deleted = 1;
4087 spin_unlock(&pa->pa_lock);
4088
4089 list_del_rcu(&pa->pa_inode_list);
4090 list_add(&pa->u.pa_tmp_list, &discard_list);
4091
4092 total_entries--;
4093 if (total_entries <= 5) {
4094 /*
4095 * we want to keep only 5 entries
4096 * allowing it to grow to 8. This
4097 * mak sure we don't call discard
4098 * soon for this list.
4099 */
4100 break;
4101 }
4102 }
4103 spin_unlock(&lg->lg_prealloc_lock);
4104
4105 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4106
4107 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4108 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4109 ext4_error(sb, "Error loading buddy information for %u",
4110 group);
4111 continue;
4112 }
4113 ext4_lock_group(sb, group);
4114 list_del(&pa->pa_group_list);
4115 ext4_mb_release_group_pa(&e4b, pa);
4116 ext4_unlock_group(sb, group);
4117
4118 ext4_mb_unload_buddy(&e4b);
4119 list_del(&pa->u.pa_tmp_list);
4120 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4121 }
4122}
4123
4124/*
4125 * We have incremented pa_count. So it cannot be freed at this
4126 * point. Also we hold lg_mutex. So no parallel allocation is
4127 * possible from this lg. That means pa_free cannot be updated.
4128 *
4129 * A parallel ext4_mb_discard_group_preallocations is possible.
4130 * which can cause the lg_prealloc_list to be updated.
4131 */
4132
4133static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4134{
4135 int order, added = 0, lg_prealloc_count = 1;
4136 struct super_block *sb = ac->ac_sb;
4137 struct ext4_locality_group *lg = ac->ac_lg;
4138 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4139
4140 order = fls(pa->pa_free) - 1;
4141 if (order > PREALLOC_TB_SIZE - 1)
4142 /* The max size of hash table is PREALLOC_TB_SIZE */
4143 order = PREALLOC_TB_SIZE - 1;
4144 /* Add the prealloc space to lg */
4145 rcu_read_lock();
4146 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4147 pa_inode_list) {
4148 spin_lock(&tmp_pa->pa_lock);
4149 if (tmp_pa->pa_deleted) {
4150 spin_unlock(&tmp_pa->pa_lock);
4151 continue;
4152 }
4153 if (!added && pa->pa_free < tmp_pa->pa_free) {
4154 /* Add to the tail of the previous entry */
4155 list_add_tail_rcu(&pa->pa_inode_list,
4156 &tmp_pa->pa_inode_list);
4157 added = 1;
4158 /*
4159 * we want to count the total
4160 * number of entries in the list
4161 */
4162 }
4163 spin_unlock(&tmp_pa->pa_lock);
4164 lg_prealloc_count++;
4165 }
4166 if (!added)
4167 list_add_tail_rcu(&pa->pa_inode_list,
4168 &lg->lg_prealloc_list[order]);
4169 rcu_read_unlock();
4170
4171 /* Now trim the list to be not more than 8 elements */
4172 if (lg_prealloc_count > 8) {
4173 ext4_mb_discard_lg_preallocations(sb, lg,
4174 order, lg_prealloc_count);
4175 return;
4176 }
4177 return ;
4178}
4179
4180/*
4181 * release all resource we used in allocation
4182 */
4183static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4184{
4185 struct ext4_prealloc_space *pa = ac->ac_pa;
4186 if (pa) {
4187 if (pa->pa_type == MB_GROUP_PA) {
4188 /* see comment in ext4_mb_use_group_pa() */
4189 spin_lock(&pa->pa_lock);
4190 pa->pa_pstart += ac->ac_b_ex.fe_len;
4191 pa->pa_lstart += ac->ac_b_ex.fe_len;
4192 pa->pa_free -= ac->ac_b_ex.fe_len;
4193 pa->pa_len -= ac->ac_b_ex.fe_len;
4194 spin_unlock(&pa->pa_lock);
4195 }
4196 }
4197 if (pa) {
4198 /*
4199 * We want to add the pa to the right bucket.
4200 * Remove it from the list and while adding
4201 * make sure the list to which we are adding
4202 * doesn't grow big.
4203 */
4204 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4205 spin_lock(pa->pa_obj_lock);
4206 list_del_rcu(&pa->pa_inode_list);
4207 spin_unlock(pa->pa_obj_lock);
4208 ext4_mb_add_n_trim(ac);
4209 }
4210 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4211 }
4212 if (ac->ac_bitmap_page)
4213 page_cache_release(ac->ac_bitmap_page);
4214 if (ac->ac_buddy_page)
4215 page_cache_release(ac->ac_buddy_page);
4216 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4217 mutex_unlock(&ac->ac_lg->lg_mutex);
4218 ext4_mb_collect_stats(ac);
4219 return 0;
4220}
4221
4222static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4223{
4224 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4225 int ret;
4226 int freed = 0;
4227
4228 trace_ext4_mb_discard_preallocations(sb, needed);
4229 for (i = 0; i < ngroups && needed > 0; i++) {
4230 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4231 freed += ret;
4232 needed -= ret;
4233 }
4234
4235 return freed;
4236}
4237
4238/*
4239 * Main entry point into mballoc to allocate blocks
4240 * it tries to use preallocation first, then falls back
4241 * to usual allocation
4242 */
4243ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4244 struct ext4_allocation_request *ar, int *errp)
4245{
4246 int freed;
4247 struct ext4_allocation_context *ac = NULL;
4248 struct ext4_sb_info *sbi;
4249 struct super_block *sb;
4250 ext4_fsblk_t block = 0;
4251 unsigned int inquota = 0;
4252 unsigned int reserv_blks = 0;
4253
4254 sb = ar->inode->i_sb;
4255 sbi = EXT4_SB(sb);
4256
4257 trace_ext4_request_blocks(ar);
4258
4259 /*
4260 * For delayed allocation, we could skip the ENOSPC and
4261 * EDQUOT check, as blocks and quotas have been already
4262 * reserved when data being copied into pagecache.
4263 */
4264 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4265 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4266 else {
4267 /* Without delayed allocation we need to verify
4268 * there is enough free blocks to do block allocation
4269 * and verify allocation doesn't exceed the quota limits.
4270 */
4271 while (ar->len &&
4272 ext4_claim_free_blocks(sbi, ar->len, ar->flags)) {
4273
4274 /* let others to free the space */
4275 yield();
4276 ar->len = ar->len >> 1;
4277 }
4278 if (!ar->len) {
4279 *errp = -ENOSPC;
4280 return 0;
4281 }
4282 reserv_blks = ar->len;
4283 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4284 dquot_alloc_block_nofail(ar->inode, ar->len);
4285 } else {
4286 while (ar->len &&
4287 dquot_alloc_block(ar->inode, ar->len)) {
4288
4289 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4290 ar->len--;
4291 }
4292 }
4293 inquota = ar->len;
4294 if (ar->len == 0) {
4295 *errp = -EDQUOT;
4296 goto out;
4297 }
4298 }
4299
4300 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4301 if (!ac) {
4302 ar->len = 0;
4303 *errp = -ENOMEM;
4304 goto out;
4305 }
4306
4307 *errp = ext4_mb_initialize_context(ac, ar);
4308 if (*errp) {
4309 ar->len = 0;
4310 goto out;
4311 }
4312
4313 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4314 if (!ext4_mb_use_preallocated(ac)) {
4315 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4316 ext4_mb_normalize_request(ac, ar);
4317repeat:
4318 /* allocate space in core */
4319 *errp = ext4_mb_regular_allocator(ac);
4320 if (*errp)
4321 goto errout;
4322
4323 /* as we've just preallocated more space than
4324 * user requested orinally, we store allocated
4325 * space in a special descriptor */
4326 if (ac->ac_status == AC_STATUS_FOUND &&
4327 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4328 ext4_mb_new_preallocation(ac);
4329 }
4330 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4331 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4332 if (*errp == -EAGAIN) {
4333 /*
4334 * drop the reference that we took
4335 * in ext4_mb_use_best_found
4336 */
4337 ext4_mb_release_context(ac);
4338 ac->ac_b_ex.fe_group = 0;
4339 ac->ac_b_ex.fe_start = 0;
4340 ac->ac_b_ex.fe_len = 0;
4341 ac->ac_status = AC_STATUS_CONTINUE;
4342 goto repeat;
4343 } else if (*errp)
4344 errout:
4345 ext4_discard_allocated_blocks(ac);
4346 else {
4347 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4348 ar->len = ac->ac_b_ex.fe_len;
4349 }
4350 } else {
4351 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4352 if (freed)
4353 goto repeat;
4354 *errp = -ENOSPC;
4355 }
4356
4357 if (*errp) {
4358 ac->ac_b_ex.fe_len = 0;
4359 ar->len = 0;
4360 ext4_mb_show_ac(ac);
4361 }
4362 ext4_mb_release_context(ac);
4363out:
4364 if (ac)
4365 kmem_cache_free(ext4_ac_cachep, ac);
4366 if (inquota && ar->len < inquota)
4367 dquot_free_block(ar->inode, inquota - ar->len);
4368 if (!ar->len) {
4369 if (!ext4_test_inode_state(ar->inode,
4370 EXT4_STATE_DELALLOC_RESERVED))
4371 /* release all the reserved blocks if non delalloc */
4372 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4373 reserv_blks);
4374 }
4375
4376 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4377
4378 return block;
4379}
4380
4381/*
4382 * We can merge two free data extents only if the physical blocks
4383 * are contiguous, AND the extents were freed by the same transaction,
4384 * AND the blocks are associated with the same group.
4385 */
4386static int can_merge(struct ext4_free_data *entry1,
4387 struct ext4_free_data *entry2)
4388{
4389 if ((entry1->t_tid == entry2->t_tid) &&
4390 (entry1->group == entry2->group) &&
4391 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4392 return 1;
4393 return 0;
4394}
4395
4396static noinline_for_stack int
4397ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4398 struct ext4_free_data *new_entry)
4399{
4400 ext4_group_t group = e4b->bd_group;
4401 ext4_grpblk_t block;
4402 struct ext4_free_data *entry;
4403 struct ext4_group_info *db = e4b->bd_info;
4404 struct super_block *sb = e4b->bd_sb;
4405 struct ext4_sb_info *sbi = EXT4_SB(sb);
4406 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4407 struct rb_node *parent = NULL, *new_node;
4408
4409 BUG_ON(!ext4_handle_valid(handle));
4410 BUG_ON(e4b->bd_bitmap_page == NULL);
4411 BUG_ON(e4b->bd_buddy_page == NULL);
4412
4413 new_node = &new_entry->node;
4414 block = new_entry->start_blk;
4415
4416 if (!*n) {
4417 /* first free block exent. We need to
4418 protect buddy cache from being freed,
4419 * otherwise we'll refresh it from
4420 * on-disk bitmap and lose not-yet-available
4421 * blocks */
4422 page_cache_get(e4b->bd_buddy_page);
4423 page_cache_get(e4b->bd_bitmap_page);
4424 }
4425 while (*n) {
4426 parent = *n;
4427 entry = rb_entry(parent, struct ext4_free_data, node);
4428 if (block < entry->start_blk)
4429 n = &(*n)->rb_left;
4430 else if (block >= (entry->start_blk + entry->count))
4431 n = &(*n)->rb_right;
4432 else {
4433 ext4_grp_locked_error(sb, group, 0,
4434 ext4_group_first_block_no(sb, group) + block,
4435 "Block already on to-be-freed list");
4436 return 0;
4437 }
4438 }
4439
4440 rb_link_node(new_node, parent, n);
4441 rb_insert_color(new_node, &db->bb_free_root);
4442
4443 /* Now try to see the extent can be merged to left and right */
4444 node = rb_prev(new_node);
4445 if (node) {
4446 entry = rb_entry(node, struct ext4_free_data, node);
4447 if (can_merge(entry, new_entry)) {
4448 new_entry->start_blk = entry->start_blk;
4449 new_entry->count += entry->count;
4450 rb_erase(node, &(db->bb_free_root));
4451 spin_lock(&sbi->s_md_lock);
4452 list_del(&entry->list);
4453 spin_unlock(&sbi->s_md_lock);
4454 kmem_cache_free(ext4_free_ext_cachep, entry);
4455 }
4456 }
4457
4458 node = rb_next(new_node);
4459 if (node) {
4460 entry = rb_entry(node, struct ext4_free_data, node);
4461 if (can_merge(new_entry, entry)) {
4462 new_entry->count += entry->count;
4463 rb_erase(node, &(db->bb_free_root));
4464 spin_lock(&sbi->s_md_lock);
4465 list_del(&entry->list);
4466 spin_unlock(&sbi->s_md_lock);
4467 kmem_cache_free(ext4_free_ext_cachep, entry);
4468 }
4469 }
4470 /* Add the extent to transaction's private list */
4471 spin_lock(&sbi->s_md_lock);
4472 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4473 spin_unlock(&sbi->s_md_lock);
4474 return 0;
4475}
4476
4477/**
4478 * ext4_free_blocks() -- Free given blocks and update quota
4479 * @handle: handle for this transaction
4480 * @inode: inode
4481 * @block: start physical block to free
4482 * @count: number of blocks to count
4483 * @flags: flags used by ext4_free_blocks
4484 */
4485void ext4_free_blocks(handle_t *handle, struct inode *inode,
4486 struct buffer_head *bh, ext4_fsblk_t block,
4487 unsigned long count, int flags)
4488{
4489 struct buffer_head *bitmap_bh = NULL;
4490 struct super_block *sb = inode->i_sb;
4491 struct ext4_group_desc *gdp;
4492 unsigned long freed = 0;
4493 unsigned int overflow;
4494 ext4_grpblk_t bit;
4495 struct buffer_head *gd_bh;
4496 ext4_group_t block_group;
4497 struct ext4_sb_info *sbi;
4498 struct ext4_buddy e4b;
4499 int err = 0;
4500 int ret;
4501
4502 if (bh) {
4503 if (block)
4504 BUG_ON(block != bh->b_blocknr);
4505 else
4506 block = bh->b_blocknr;
4507 }
4508
4509 sbi = EXT4_SB(sb);
4510 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4511 !ext4_data_block_valid(sbi, block, count)) {
4512 ext4_error(sb, "Freeing blocks not in datazone - "
4513 "block = %llu, count = %lu", block, count);
4514 goto error_return;
4515 }
4516
4517 ext4_debug("freeing block %llu\n", block);
4518 trace_ext4_free_blocks(inode, block, count, flags);
4519
4520 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4521 struct buffer_head *tbh = bh;
4522 int i;
4523
4524 BUG_ON(bh && (count > 1));
4525
4526 for (i = 0; i < count; i++) {
4527 if (!bh)
4528 tbh = sb_find_get_block(inode->i_sb,
4529 block + i);
4530 if (unlikely(!tbh))
4531 continue;
4532 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4533 inode, tbh, block + i);
4534 }
4535 }
4536
4537 /*
4538 * We need to make sure we don't reuse the freed block until
4539 * after the transaction is committed, which we can do by
4540 * treating the block as metadata, below. We make an
4541 * exception if the inode is to be written in writeback mode
4542 * since writeback mode has weak data consistency guarantees.
4543 */
4544 if (!ext4_should_writeback_data(inode))
4545 flags |= EXT4_FREE_BLOCKS_METADATA;
4546
4547do_more:
4548 overflow = 0;
4549 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4550
4551 /*
4552 * Check to see if we are freeing blocks across a group
4553 * boundary.
4554 */
4555 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4556 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4557 count -= overflow;
4558 }
4559 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4560 if (!bitmap_bh) {
4561 err = -EIO;
4562 goto error_return;
4563 }
4564 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4565 if (!gdp) {
4566 err = -EIO;
4567 goto error_return;
4568 }
4569
4570 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4571 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4572 in_range(block, ext4_inode_table(sb, gdp),
4573 EXT4_SB(sb)->s_itb_per_group) ||
4574 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4575 EXT4_SB(sb)->s_itb_per_group)) {
4576
4577 ext4_error(sb, "Freeing blocks in system zone - "
4578 "Block = %llu, count = %lu", block, count);
4579 /* err = 0. ext4_std_error should be a no op */
4580 goto error_return;
4581 }
4582
4583 BUFFER_TRACE(bitmap_bh, "getting write access");
4584 err = ext4_journal_get_write_access(handle, bitmap_bh);
4585 if (err)
4586 goto error_return;
4587
4588 /*
4589 * We are about to modify some metadata. Call the journal APIs
4590 * to unshare ->b_data if a currently-committing transaction is
4591 * using it
4592 */
4593 BUFFER_TRACE(gd_bh, "get_write_access");
4594 err = ext4_journal_get_write_access(handle, gd_bh);
4595 if (err)
4596 goto error_return;
4597#ifdef AGGRESSIVE_CHECK
4598 {
4599 int i;
4600 for (i = 0; i < count; i++)
4601 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4602 }
4603#endif
4604 trace_ext4_mballoc_free(sb, inode, block_group, bit, count);
4605
4606 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4607 if (err)
4608 goto error_return;
4609
4610 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4611 struct ext4_free_data *new_entry;
4612 /*
4613 * blocks being freed are metadata. these blocks shouldn't
4614 * be used until this transaction is committed
4615 */
4616 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4617 if (!new_entry) {
4618 err = -ENOMEM;
4619 goto error_return;
4620 }
4621 new_entry->start_blk = bit;
4622 new_entry->group = block_group;
4623 new_entry->count = count;
4624 new_entry->t_tid = handle->h_transaction->t_tid;
4625
4626 ext4_lock_group(sb, block_group);
4627 mb_clear_bits(bitmap_bh->b_data, bit, count);
4628 ext4_mb_free_metadata(handle, &e4b, new_entry);
4629 } else {
4630 /* need to update group_info->bb_free and bitmap
4631 * with group lock held. generate_buddy look at
4632 * them with group lock_held
4633 */
4634 ext4_lock_group(sb, block_group);
4635 mb_clear_bits(bitmap_bh->b_data, bit, count);
4636 mb_free_blocks(inode, &e4b, bit, count);
4637 }
4638
4639 ret = ext4_free_blks_count(sb, gdp) + count;
4640 ext4_free_blks_set(sb, gdp, ret);
4641 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4642 ext4_unlock_group(sb, block_group);
4643 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4644
4645 if (sbi->s_log_groups_per_flex) {
4646 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4647 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4648 }
4649
4650 ext4_mb_unload_buddy(&e4b);
4651
4652 freed += count;
4653
4654 /* We dirtied the bitmap block */
4655 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4656 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4657
4658 /* And the group descriptor block */
4659 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4660 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4661 if (!err)
4662 err = ret;
4663
4664 if (overflow && !err) {
4665 block += count;
4666 count = overflow;
4667 put_bh(bitmap_bh);
4668 goto do_more;
4669 }
4670 ext4_mark_super_dirty(sb);
4671error_return:
4672 if (freed && !(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4673 dquot_free_block(inode, freed);
4674 brelse(bitmap_bh);
4675 ext4_std_error(sb, err);
4676 return;
4677}
4678
4679/**
4680 * ext4_group_add_blocks() -- Add given blocks to an existing group
4681 * @handle: handle to this transaction
4682 * @sb: super block
4683 * @block: start physcial block to add to the block group
4684 * @count: number of blocks to free
4685 *
4686 * This marks the blocks as free in the bitmap and buddy.
4687 */
4688int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4689 ext4_fsblk_t block, unsigned long count)
4690{
4691 struct buffer_head *bitmap_bh = NULL;
4692 struct buffer_head *gd_bh;
4693 ext4_group_t block_group;
4694 ext4_grpblk_t bit;
4695 unsigned int i;
4696 struct ext4_group_desc *desc;
4697 struct ext4_sb_info *sbi = EXT4_SB(sb);
4698 struct ext4_buddy e4b;
4699 int err = 0, ret, blk_free_count;
4700 ext4_grpblk_t blocks_freed;
4701
4702 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4703
4704 if (count == 0)
4705 return 0;
4706
4707 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4708 /*
4709 * Check to see if we are freeing blocks across a group
4710 * boundary.
4711 */
4712 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4713 ext4_warning(sb, "too much blocks added to group %u\n",
4714 block_group);
4715 err = -EINVAL;
4716 goto error_return;
4717 }
4718
4719 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4720 if (!bitmap_bh) {
4721 err = -EIO;
4722 goto error_return;
4723 }
4724
4725 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4726 if (!desc) {
4727 err = -EIO;
4728 goto error_return;
4729 }
4730
4731 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4732 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4733 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4734 in_range(block + count - 1, ext4_inode_table(sb, desc),
4735 sbi->s_itb_per_group)) {
4736 ext4_error(sb, "Adding blocks in system zones - "
4737 "Block = %llu, count = %lu",
4738 block, count);
4739 err = -EINVAL;
4740 goto error_return;
4741 }
4742
4743 BUFFER_TRACE(bitmap_bh, "getting write access");
4744 err = ext4_journal_get_write_access(handle, bitmap_bh);
4745 if (err)
4746 goto error_return;
4747
4748 /*
4749 * We are about to modify some metadata. Call the journal APIs
4750 * to unshare ->b_data if a currently-committing transaction is
4751 * using it
4752 */
4753 BUFFER_TRACE(gd_bh, "get_write_access");
4754 err = ext4_journal_get_write_access(handle, gd_bh);
4755 if (err)
4756 goto error_return;
4757
4758 for (i = 0, blocks_freed = 0; i < count; i++) {
4759 BUFFER_TRACE(bitmap_bh, "clear bit");
4760 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4761 ext4_error(sb, "bit already cleared for block %llu",
4762 (ext4_fsblk_t)(block + i));
4763 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4764 } else {
4765 blocks_freed++;
4766 }
4767 }
4768
4769 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4770 if (err)
4771 goto error_return;
4772
4773 /*
4774 * need to update group_info->bb_free and bitmap
4775 * with group lock held. generate_buddy look at
4776 * them with group lock_held
4777 */
4778 ext4_lock_group(sb, block_group);
4779 mb_clear_bits(bitmap_bh->b_data, bit, count);
4780 mb_free_blocks(NULL, &e4b, bit, count);
4781 blk_free_count = blocks_freed + ext4_free_blks_count(sb, desc);
4782 ext4_free_blks_set(sb, desc, blk_free_count);
4783 desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
4784 ext4_unlock_group(sb, block_group);
4785 percpu_counter_add(&sbi->s_freeblocks_counter, blocks_freed);
4786
4787 if (sbi->s_log_groups_per_flex) {
4788 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4789 atomic_add(blocks_freed,
4790 &sbi->s_flex_groups[flex_group].free_blocks);
4791 }
4792
4793 ext4_mb_unload_buddy(&e4b);
4794
4795 /* We dirtied the bitmap block */
4796 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4797 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4798
4799 /* And the group descriptor block */
4800 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4801 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4802 if (!err)
4803 err = ret;
4804
4805error_return:
4806 brelse(bitmap_bh);
4807 ext4_std_error(sb, err);
4808 return err;
4809}
4810
4811/**
4812 * ext4_trim_extent -- function to TRIM one single free extent in the group
4813 * @sb: super block for the file system
4814 * @start: starting block of the free extent in the alloc. group
4815 * @count: number of blocks to TRIM
4816 * @group: alloc. group we are working with
4817 * @e4b: ext4 buddy for the group
4818 *
4819 * Trim "count" blocks starting at "start" in the "group". To assure that no
4820 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4821 * be called with under the group lock.
4822 */
4823static void ext4_trim_extent(struct super_block *sb, int start, int count,
4824 ext4_group_t group, struct ext4_buddy *e4b)
4825{
4826 struct ext4_free_extent ex;
4827
4828 trace_ext4_trim_extent(sb, group, start, count);
4829
4830 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4831
4832 ex.fe_start = start;
4833 ex.fe_group = group;
4834 ex.fe_len = count;
4835
4836 /*
4837 * Mark blocks used, so no one can reuse them while
4838 * being trimmed.
4839 */
4840 mb_mark_used(e4b, &ex);
4841 ext4_unlock_group(sb, group);
4842 ext4_issue_discard(sb, group, start, count);
4843 ext4_lock_group(sb, group);
4844 mb_free_blocks(NULL, e4b, start, ex.fe_len);
4845}
4846
4847/**
4848 * ext4_trim_all_free -- function to trim all free space in alloc. group
4849 * @sb: super block for file system
4850 * @group: group to be trimmed
4851 * @start: first group block to examine
4852 * @max: last group block to examine
4853 * @minblocks: minimum extent block count
4854 *
4855 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4856 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4857 * the extent.
4858 *
4859 *
4860 * ext4_trim_all_free walks through group's block bitmap searching for free
4861 * extents. When the free extent is found, mark it as used in group buddy
4862 * bitmap. Then issue a TRIM command on this extent and free the extent in
4863 * the group buddy bitmap. This is done until whole group is scanned.
4864 */
4865static ext4_grpblk_t
4866ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4867 ext4_grpblk_t start, ext4_grpblk_t max,
4868 ext4_grpblk_t minblocks)
4869{
4870 void *bitmap;
4871 ext4_grpblk_t next, count = 0, free_count = 0;
4872 struct ext4_buddy e4b;
4873 int ret;
4874
4875 trace_ext4_trim_all_free(sb, group, start, max);
4876
4877 ret = ext4_mb_load_buddy(sb, group, &e4b);
4878 if (ret) {
4879 ext4_error(sb, "Error in loading buddy "
4880 "information for %u", group);
4881 return ret;
4882 }
4883 bitmap = e4b.bd_bitmap;
4884
4885 ext4_lock_group(sb, group);
4886 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4887 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4888 goto out;
4889
4890 start = (e4b.bd_info->bb_first_free > start) ?
4891 e4b.bd_info->bb_first_free : start;
4892
4893 while (start < max) {
4894 start = mb_find_next_zero_bit(bitmap, max, start);
4895 if (start >= max)
4896 break;
4897 next = mb_find_next_bit(bitmap, max, start);
4898
4899 if ((next - start) >= minblocks) {
4900 ext4_trim_extent(sb, start,
4901 next - start, group, &e4b);
4902 count += next - start;
4903 }
4904 free_count += next - start;
4905 start = next + 1;
4906
4907 if (fatal_signal_pending(current)) {
4908 count = -ERESTARTSYS;
4909 break;
4910 }
4911
4912 if (need_resched()) {
4913 ext4_unlock_group(sb, group);
4914 cond_resched();
4915 ext4_lock_group(sb, group);
4916 }
4917
4918 if ((e4b.bd_info->bb_free - free_count) < minblocks)
4919 break;
4920 }
4921
4922 if (!ret)
4923 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4924out:
4925 ext4_unlock_group(sb, group);
4926 ext4_mb_unload_buddy(&e4b);
4927
4928 ext4_debug("trimmed %d blocks in the group %d\n",
4929 count, group);
4930
4931 return count;
4932}
4933
4934/**
4935 * ext4_trim_fs() -- trim ioctl handle function
4936 * @sb: superblock for filesystem
4937 * @range: fstrim_range structure
4938 *
4939 * start: First Byte to trim
4940 * len: number of Bytes to trim from start
4941 * minlen: minimum extent length in Bytes
4942 * ext4_trim_fs goes through all allocation groups containing Bytes from
4943 * start to start+len. For each such a group ext4_trim_all_free function
4944 * is invoked to trim all free space.
4945 */
4946int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4947{
4948 struct ext4_group_info *grp;
4949 ext4_group_t first_group, last_group;
4950 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
4951 ext4_grpblk_t cnt = 0, first_block, last_block;
4952 uint64_t start, len, minlen, trimmed = 0;
4953 ext4_fsblk_t first_data_blk =
4954 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
4955 int ret = 0;
4956
4957 start = range->start >> sb->s_blocksize_bits;
4958 len = range->len >> sb->s_blocksize_bits;
4959 minlen = range->minlen >> sb->s_blocksize_bits;
4960
4961 if (unlikely(minlen > EXT4_BLOCKS_PER_GROUP(sb)))
4962 return -EINVAL;
4963 if (start + len <= first_data_blk)
4964 goto out;
4965 if (start < first_data_blk) {
4966 len -= first_data_blk - start;
4967 start = first_data_blk;
4968 }
4969
4970 /* Determine first and last group to examine based on start and len */
4971 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
4972 &first_group, &first_block);
4973 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
4974 &last_group, &last_block);
4975 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
4976 last_block = EXT4_BLOCKS_PER_GROUP(sb);
4977
4978 if (first_group > last_group)
4979 return -EINVAL;
4980
4981 for (group = first_group; group <= last_group; group++) {
4982 grp = ext4_get_group_info(sb, group);
4983 /* We only do this if the grp has never been initialized */
4984 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
4985 ret = ext4_mb_init_group(sb, group);
4986 if (ret)
4987 break;
4988 }
4989
4990 /*
4991 * For all the groups except the last one, last block will
4992 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
4993 * change it for the last group in which case start +
4994 * len < EXT4_BLOCKS_PER_GROUP(sb).
4995 */
4996 if (first_block + len < EXT4_BLOCKS_PER_GROUP(sb))
4997 last_block = first_block + len;
4998 len -= last_block - first_block;
4999
5000 if (grp->bb_free >= minlen) {
5001 cnt = ext4_trim_all_free(sb, group, first_block,
5002 last_block, minlen);
5003 if (cnt < 0) {
5004 ret = cnt;
5005 break;
5006 }
5007 }
5008 trimmed += cnt;
5009 first_block = 0;
5010 }
5011 range->len = trimmed * sb->s_blocksize;
5012
5013 if (!ret)
5014 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5015
5016out:
5017 return ret;
5018}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4 * Written by Alex Tomas <alex@clusterfs.com>
5 */
6
7
8/*
9 * mballoc.c contains the multiblocks allocation routines
10 */
11
12#include "ext4_jbd2.h"
13#include "mballoc.h"
14#include <linux/log2.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/nospec.h>
18#include <linux/backing-dev.h>
19#include <trace/events/ext4.h>
20
21/*
22 * MUSTDO:
23 * - test ext4_ext_search_left() and ext4_ext_search_right()
24 * - search for metadata in few groups
25 *
26 * TODO v4:
27 * - normalization should take into account whether file is still open
28 * - discard preallocations if no free space left (policy?)
29 * - don't normalize tails
30 * - quota
31 * - reservation for superuser
32 *
33 * TODO v3:
34 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
35 * - track min/max extents in each group for better group selection
36 * - mb_mark_used() may allocate chunk right after splitting buddy
37 * - tree of groups sorted by number of free blocks
38 * - error handling
39 */
40
41/*
42 * The allocation request involve request for multiple number of blocks
43 * near to the goal(block) value specified.
44 *
45 * During initialization phase of the allocator we decide to use the
46 * group preallocation or inode preallocation depending on the size of
47 * the file. The size of the file could be the resulting file size we
48 * would have after allocation, or the current file size, which ever
49 * is larger. If the size is less than sbi->s_mb_stream_request we
50 * select to use the group preallocation. The default value of
51 * s_mb_stream_request is 16 blocks. This can also be tuned via
52 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
53 * terms of number of blocks.
54 *
55 * The main motivation for having small file use group preallocation is to
56 * ensure that we have small files closer together on the disk.
57 *
58 * First stage the allocator looks at the inode prealloc list,
59 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
60 * spaces for this particular inode. The inode prealloc space is
61 * represented as:
62 *
63 * pa_lstart -> the logical start block for this prealloc space
64 * pa_pstart -> the physical start block for this prealloc space
65 * pa_len -> length for this prealloc space (in clusters)
66 * pa_free -> free space available in this prealloc space (in clusters)
67 *
68 * The inode preallocation space is used looking at the _logical_ start
69 * block. If only the logical file block falls within the range of prealloc
70 * space we will consume the particular prealloc space. This makes sure that
71 * we have contiguous physical blocks representing the file blocks
72 *
73 * The important thing to be noted in case of inode prealloc space is that
74 * we don't modify the values associated to inode prealloc space except
75 * pa_free.
76 *
77 * If we are not able to find blocks in the inode prealloc space and if we
78 * have the group allocation flag set then we look at the locality group
79 * prealloc space. These are per CPU prealloc list represented as
80 *
81 * ext4_sb_info.s_locality_groups[smp_processor_id()]
82 *
83 * The reason for having a per cpu locality group is to reduce the contention
84 * between CPUs. It is possible to get scheduled at this point.
85 *
86 * The locality group prealloc space is used looking at whether we have
87 * enough free space (pa_free) within the prealloc space.
88 *
89 * If we can't allocate blocks via inode prealloc or/and locality group
90 * prealloc then we look at the buddy cache. The buddy cache is represented
91 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
92 * mapped to the buddy and bitmap information regarding different
93 * groups. The buddy information is attached to buddy cache inode so that
94 * we can access them through the page cache. The information regarding
95 * each group is loaded via ext4_mb_load_buddy. The information involve
96 * block bitmap and buddy information. The information are stored in the
97 * inode as:
98 *
99 * { page }
100 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
101 *
102 *
103 * one block each for bitmap and buddy information. So for each group we
104 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
105 * blocksize) blocks. So it can have information regarding groups_per_page
106 * which is blocks_per_page/2
107 *
108 * The buddy cache inode is not stored on disk. The inode is thrown
109 * away when the filesystem is unmounted.
110 *
111 * We look for count number of blocks in the buddy cache. If we were able
112 * to locate that many free blocks we return with additional information
113 * regarding rest of the contiguous physical block available
114 *
115 * Before allocating blocks via buddy cache we normalize the request
116 * blocks. This ensure we ask for more blocks that we needed. The extra
117 * blocks that we get after allocation is added to the respective prealloc
118 * list. In case of inode preallocation we follow a list of heuristics
119 * based on file size. This can be found in ext4_mb_normalize_request. If
120 * we are doing a group prealloc we try to normalize the request to
121 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
122 * dependent on the cluster size; for non-bigalloc file systems, it is
123 * 512 blocks. This can be tuned via
124 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
125 * terms of number of blocks. If we have mounted the file system with -O
126 * stripe=<value> option the group prealloc request is normalized to the
127 * smallest multiple of the stripe value (sbi->s_stripe) which is
128 * greater than the default mb_group_prealloc.
129 *
130 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
131 * structures in two data structures:
132 *
133 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
134 *
135 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
136 *
137 * This is an array of lists where the index in the array represents the
138 * largest free order in the buddy bitmap of the participating group infos of
139 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
140 * number of buddy bitmap orders possible) number of lists. Group-infos are
141 * placed in appropriate lists.
142 *
143 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
144 *
145 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
146 *
147 * This is an array of lists where in the i-th list there are groups with
148 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
149 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
150 * Note that we don't bother with a special list for completely empty groups
151 * so we only have MB_NUM_ORDERS(sb) lists.
152 *
153 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
154 * structures to decide the order in which groups are to be traversed for
155 * fulfilling an allocation request.
156 *
157 * At CR = 0, we look for groups which have the largest_free_order >= the order
158 * of the request. We directly look at the largest free order list in the data
159 * structure (1) above where largest_free_order = order of the request. If that
160 * list is empty, we look at remaining list in the increasing order of
161 * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time.
162 *
163 * At CR = 1, we only consider groups where average fragment size > request
164 * size. So, we lookup a group which has average fragment size just above or
165 * equal to request size using our average fragment size group lists (data
166 * structure 2) in O(1) time.
167 *
168 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
169 * linear order which requires O(N) search time for each CR 0 and CR 1 phase.
170 *
171 * The regular allocator (using the buddy cache) supports a few tunables.
172 *
173 * /sys/fs/ext4/<partition>/mb_min_to_scan
174 * /sys/fs/ext4/<partition>/mb_max_to_scan
175 * /sys/fs/ext4/<partition>/mb_order2_req
176 * /sys/fs/ext4/<partition>/mb_linear_limit
177 *
178 * The regular allocator uses buddy scan only if the request len is power of
179 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
180 * value of s_mb_order2_reqs can be tuned via
181 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
182 * stripe size (sbi->s_stripe), we try to search for contiguous block in
183 * stripe size. This should result in better allocation on RAID setups. If
184 * not, we search in the specific group using bitmap for best extents. The
185 * tunable min_to_scan and max_to_scan control the behaviour here.
186 * min_to_scan indicate how long the mballoc __must__ look for a best
187 * extent and max_to_scan indicates how long the mballoc __can__ look for a
188 * best extent in the found extents. Searching for the blocks starts with
189 * the group specified as the goal value in allocation context via
190 * ac_g_ex. Each group is first checked based on the criteria whether it
191 * can be used for allocation. ext4_mb_good_group explains how the groups are
192 * checked.
193 *
194 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
195 * get traversed linearly. That may result in subsequent allocations being not
196 * close to each other. And so, the underlying device may get filled up in a
197 * non-linear fashion. While that may not matter on non-rotational devices, for
198 * rotational devices that may result in higher seek times. "mb_linear_limit"
199 * tells mballoc how many groups mballoc should search linearly before
200 * performing consulting above data structures for more efficient lookups. For
201 * non rotational devices, this value defaults to 0 and for rotational devices
202 * this is set to MB_DEFAULT_LINEAR_LIMIT.
203 *
204 * Both the prealloc space are getting populated as above. So for the first
205 * request we will hit the buddy cache which will result in this prealloc
206 * space getting filled. The prealloc space is then later used for the
207 * subsequent request.
208 */
209
210/*
211 * mballoc operates on the following data:
212 * - on-disk bitmap
213 * - in-core buddy (actually includes buddy and bitmap)
214 * - preallocation descriptors (PAs)
215 *
216 * there are two types of preallocations:
217 * - inode
218 * assiged to specific inode and can be used for this inode only.
219 * it describes part of inode's space preallocated to specific
220 * physical blocks. any block from that preallocated can be used
221 * independent. the descriptor just tracks number of blocks left
222 * unused. so, before taking some block from descriptor, one must
223 * make sure corresponded logical block isn't allocated yet. this
224 * also means that freeing any block within descriptor's range
225 * must discard all preallocated blocks.
226 * - locality group
227 * assigned to specific locality group which does not translate to
228 * permanent set of inodes: inode can join and leave group. space
229 * from this type of preallocation can be used for any inode. thus
230 * it's consumed from the beginning to the end.
231 *
232 * relation between them can be expressed as:
233 * in-core buddy = on-disk bitmap + preallocation descriptors
234 *
235 * this mean blocks mballoc considers used are:
236 * - allocated blocks (persistent)
237 * - preallocated blocks (non-persistent)
238 *
239 * consistency in mballoc world means that at any time a block is either
240 * free or used in ALL structures. notice: "any time" should not be read
241 * literally -- time is discrete and delimited by locks.
242 *
243 * to keep it simple, we don't use block numbers, instead we count number of
244 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
245 *
246 * all operations can be expressed as:
247 * - init buddy: buddy = on-disk + PAs
248 * - new PA: buddy += N; PA = N
249 * - use inode PA: on-disk += N; PA -= N
250 * - discard inode PA buddy -= on-disk - PA; PA = 0
251 * - use locality group PA on-disk += N; PA -= N
252 * - discard locality group PA buddy -= PA; PA = 0
253 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
254 * is used in real operation because we can't know actual used
255 * bits from PA, only from on-disk bitmap
256 *
257 * if we follow this strict logic, then all operations above should be atomic.
258 * given some of them can block, we'd have to use something like semaphores
259 * killing performance on high-end SMP hardware. let's try to relax it using
260 * the following knowledge:
261 * 1) if buddy is referenced, it's already initialized
262 * 2) while block is used in buddy and the buddy is referenced,
263 * nobody can re-allocate that block
264 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
265 * bit set and PA claims same block, it's OK. IOW, one can set bit in
266 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
267 * block
268 *
269 * so, now we're building a concurrency table:
270 * - init buddy vs.
271 * - new PA
272 * blocks for PA are allocated in the buddy, buddy must be referenced
273 * until PA is linked to allocation group to avoid concurrent buddy init
274 * - use inode PA
275 * we need to make sure that either on-disk bitmap or PA has uptodate data
276 * given (3) we care that PA-=N operation doesn't interfere with init
277 * - discard inode PA
278 * the simplest way would be to have buddy initialized by the discard
279 * - use locality group PA
280 * again PA-=N must be serialized with init
281 * - discard locality group PA
282 * the simplest way would be to have buddy initialized by the discard
283 * - new PA vs.
284 * - use inode PA
285 * i_data_sem serializes them
286 * - discard inode PA
287 * discard process must wait until PA isn't used by another process
288 * - use locality group PA
289 * some mutex should serialize them
290 * - discard locality group PA
291 * discard process must wait until PA isn't used by another process
292 * - use inode PA
293 * - use inode PA
294 * i_data_sem or another mutex should serializes them
295 * - discard inode PA
296 * discard process must wait until PA isn't used by another process
297 * - use locality group PA
298 * nothing wrong here -- they're different PAs covering different blocks
299 * - discard locality group PA
300 * discard process must wait until PA isn't used by another process
301 *
302 * now we're ready to make few consequences:
303 * - PA is referenced and while it is no discard is possible
304 * - PA is referenced until block isn't marked in on-disk bitmap
305 * - PA changes only after on-disk bitmap
306 * - discard must not compete with init. either init is done before
307 * any discard or they're serialized somehow
308 * - buddy init as sum of on-disk bitmap and PAs is done atomically
309 *
310 * a special case when we've used PA to emptiness. no need to modify buddy
311 * in this case, but we should care about concurrent init
312 *
313 */
314
315 /*
316 * Logic in few words:
317 *
318 * - allocation:
319 * load group
320 * find blocks
321 * mark bits in on-disk bitmap
322 * release group
323 *
324 * - use preallocation:
325 * find proper PA (per-inode or group)
326 * load group
327 * mark bits in on-disk bitmap
328 * release group
329 * release PA
330 *
331 * - free:
332 * load group
333 * mark bits in on-disk bitmap
334 * release group
335 *
336 * - discard preallocations in group:
337 * mark PAs deleted
338 * move them onto local list
339 * load on-disk bitmap
340 * load group
341 * remove PA from object (inode or locality group)
342 * mark free blocks in-core
343 *
344 * - discard inode's preallocations:
345 */
346
347/*
348 * Locking rules
349 *
350 * Locks:
351 * - bitlock on a group (group)
352 * - object (inode/locality) (object)
353 * - per-pa lock (pa)
354 * - cr0 lists lock (cr0)
355 * - cr1 tree lock (cr1)
356 *
357 * Paths:
358 * - new pa
359 * object
360 * group
361 *
362 * - find and use pa:
363 * pa
364 *
365 * - release consumed pa:
366 * pa
367 * group
368 * object
369 *
370 * - generate in-core bitmap:
371 * group
372 * pa
373 *
374 * - discard all for given object (inode, locality group):
375 * object
376 * pa
377 * group
378 *
379 * - discard all for given group:
380 * group
381 * pa
382 * group
383 * object
384 *
385 * - allocation path (ext4_mb_regular_allocator)
386 * group
387 * cr0/cr1
388 */
389static struct kmem_cache *ext4_pspace_cachep;
390static struct kmem_cache *ext4_ac_cachep;
391static struct kmem_cache *ext4_free_data_cachep;
392
393/* We create slab caches for groupinfo data structures based on the
394 * superblock block size. There will be one per mounted filesystem for
395 * each unique s_blocksize_bits */
396#define NR_GRPINFO_CACHES 8
397static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
398
399static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
400 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
401 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
402 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
403};
404
405static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
406 ext4_group_t group);
407static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
408 ext4_group_t group);
409static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
410
411static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
412 ext4_group_t group, int cr);
413
414static int ext4_try_to_trim_range(struct super_block *sb,
415 struct ext4_buddy *e4b, ext4_grpblk_t start,
416 ext4_grpblk_t max, ext4_grpblk_t minblocks);
417
418/*
419 * The algorithm using this percpu seq counter goes below:
420 * 1. We sample the percpu discard_pa_seq counter before trying for block
421 * allocation in ext4_mb_new_blocks().
422 * 2. We increment this percpu discard_pa_seq counter when we either allocate
423 * or free these blocks i.e. while marking those blocks as used/free in
424 * mb_mark_used()/mb_free_blocks().
425 * 3. We also increment this percpu seq counter when we successfully identify
426 * that the bb_prealloc_list is not empty and hence proceed for discarding
427 * of those PAs inside ext4_mb_discard_group_preallocations().
428 *
429 * Now to make sure that the regular fast path of block allocation is not
430 * affected, as a small optimization we only sample the percpu seq counter
431 * on that cpu. Only when the block allocation fails and when freed blocks
432 * found were 0, that is when we sample percpu seq counter for all cpus using
433 * below function ext4_get_discard_pa_seq_sum(). This happens after making
434 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
435 */
436static DEFINE_PER_CPU(u64, discard_pa_seq);
437static inline u64 ext4_get_discard_pa_seq_sum(void)
438{
439 int __cpu;
440 u64 __seq = 0;
441
442 for_each_possible_cpu(__cpu)
443 __seq += per_cpu(discard_pa_seq, __cpu);
444 return __seq;
445}
446
447static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
448{
449#if BITS_PER_LONG == 64
450 *bit += ((unsigned long) addr & 7UL) << 3;
451 addr = (void *) ((unsigned long) addr & ~7UL);
452#elif BITS_PER_LONG == 32
453 *bit += ((unsigned long) addr & 3UL) << 3;
454 addr = (void *) ((unsigned long) addr & ~3UL);
455#else
456#error "how many bits you are?!"
457#endif
458 return addr;
459}
460
461static inline int mb_test_bit(int bit, void *addr)
462{
463 /*
464 * ext4_test_bit on architecture like powerpc
465 * needs unsigned long aligned address
466 */
467 addr = mb_correct_addr_and_bit(&bit, addr);
468 return ext4_test_bit(bit, addr);
469}
470
471static inline void mb_set_bit(int bit, void *addr)
472{
473 addr = mb_correct_addr_and_bit(&bit, addr);
474 ext4_set_bit(bit, addr);
475}
476
477static inline void mb_clear_bit(int bit, void *addr)
478{
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 ext4_clear_bit(bit, addr);
481}
482
483static inline int mb_test_and_clear_bit(int bit, void *addr)
484{
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 return ext4_test_and_clear_bit(bit, addr);
487}
488
489static inline int mb_find_next_zero_bit(void *addr, int max, int start)
490{
491 int fix = 0, ret, tmpmax;
492 addr = mb_correct_addr_and_bit(&fix, addr);
493 tmpmax = max + fix;
494 start += fix;
495
496 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
497 if (ret > max)
498 return max;
499 return ret;
500}
501
502static inline int mb_find_next_bit(void *addr, int max, int start)
503{
504 int fix = 0, ret, tmpmax;
505 addr = mb_correct_addr_and_bit(&fix, addr);
506 tmpmax = max + fix;
507 start += fix;
508
509 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
510 if (ret > max)
511 return max;
512 return ret;
513}
514
515static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
516{
517 char *bb;
518
519 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
520 BUG_ON(max == NULL);
521
522 if (order > e4b->bd_blkbits + 1) {
523 *max = 0;
524 return NULL;
525 }
526
527 /* at order 0 we see each particular block */
528 if (order == 0) {
529 *max = 1 << (e4b->bd_blkbits + 3);
530 return e4b->bd_bitmap;
531 }
532
533 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
534 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
535
536 return bb;
537}
538
539#ifdef DOUBLE_CHECK
540static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
541 int first, int count)
542{
543 int i;
544 struct super_block *sb = e4b->bd_sb;
545
546 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
547 return;
548 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
549 for (i = 0; i < count; i++) {
550 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
551 ext4_fsblk_t blocknr;
552
553 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
554 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
555 ext4_grp_locked_error(sb, e4b->bd_group,
556 inode ? inode->i_ino : 0,
557 blocknr,
558 "freeing block already freed "
559 "(bit %u)",
560 first + i);
561 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
562 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
563 }
564 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
565 }
566}
567
568static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
569{
570 int i;
571
572 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
573 return;
574 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
575 for (i = 0; i < count; i++) {
576 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
577 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
578 }
579}
580
581static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
582{
583 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
584 return;
585 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
586 unsigned char *b1, *b2;
587 int i;
588 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
589 b2 = (unsigned char *) bitmap;
590 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
591 if (b1[i] != b2[i]) {
592 ext4_msg(e4b->bd_sb, KERN_ERR,
593 "corruption in group %u "
594 "at byte %u(%u): %x in copy != %x "
595 "on disk/prealloc",
596 e4b->bd_group, i, i * 8, b1[i], b2[i]);
597 BUG();
598 }
599 }
600 }
601}
602
603static void mb_group_bb_bitmap_alloc(struct super_block *sb,
604 struct ext4_group_info *grp, ext4_group_t group)
605{
606 struct buffer_head *bh;
607
608 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
609 if (!grp->bb_bitmap)
610 return;
611
612 bh = ext4_read_block_bitmap(sb, group);
613 if (IS_ERR_OR_NULL(bh)) {
614 kfree(grp->bb_bitmap);
615 grp->bb_bitmap = NULL;
616 return;
617 }
618
619 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
620 put_bh(bh);
621}
622
623static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
624{
625 kfree(grp->bb_bitmap);
626}
627
628#else
629static inline void mb_free_blocks_double(struct inode *inode,
630 struct ext4_buddy *e4b, int first, int count)
631{
632 return;
633}
634static inline void mb_mark_used_double(struct ext4_buddy *e4b,
635 int first, int count)
636{
637 return;
638}
639static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
640{
641 return;
642}
643
644static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
645 struct ext4_group_info *grp, ext4_group_t group)
646{
647 return;
648}
649
650static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
651{
652 return;
653}
654#endif
655
656#ifdef AGGRESSIVE_CHECK
657
658#define MB_CHECK_ASSERT(assert) \
659do { \
660 if (!(assert)) { \
661 printk(KERN_EMERG \
662 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
663 function, file, line, # assert); \
664 BUG(); \
665 } \
666} while (0)
667
668static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
669 const char *function, int line)
670{
671 struct super_block *sb = e4b->bd_sb;
672 int order = e4b->bd_blkbits + 1;
673 int max;
674 int max2;
675 int i;
676 int j;
677 int k;
678 int count;
679 struct ext4_group_info *grp;
680 int fragments = 0;
681 int fstart;
682 struct list_head *cur;
683 void *buddy;
684 void *buddy2;
685
686 if (e4b->bd_info->bb_check_counter++ % 10)
687 return 0;
688
689 while (order > 1) {
690 buddy = mb_find_buddy(e4b, order, &max);
691 MB_CHECK_ASSERT(buddy);
692 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
693 MB_CHECK_ASSERT(buddy2);
694 MB_CHECK_ASSERT(buddy != buddy2);
695 MB_CHECK_ASSERT(max * 2 == max2);
696
697 count = 0;
698 for (i = 0; i < max; i++) {
699
700 if (mb_test_bit(i, buddy)) {
701 /* only single bit in buddy2 may be 0 */
702 if (!mb_test_bit(i << 1, buddy2)) {
703 MB_CHECK_ASSERT(
704 mb_test_bit((i<<1)+1, buddy2));
705 }
706 continue;
707 }
708
709 /* both bits in buddy2 must be 1 */
710 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
711 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
712
713 for (j = 0; j < (1 << order); j++) {
714 k = (i * (1 << order)) + j;
715 MB_CHECK_ASSERT(
716 !mb_test_bit(k, e4b->bd_bitmap));
717 }
718 count++;
719 }
720 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
721 order--;
722 }
723
724 fstart = -1;
725 buddy = mb_find_buddy(e4b, 0, &max);
726 for (i = 0; i < max; i++) {
727 if (!mb_test_bit(i, buddy)) {
728 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
729 if (fstart == -1) {
730 fragments++;
731 fstart = i;
732 }
733 continue;
734 }
735 fstart = -1;
736 /* check used bits only */
737 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
738 buddy2 = mb_find_buddy(e4b, j, &max2);
739 k = i >> j;
740 MB_CHECK_ASSERT(k < max2);
741 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
742 }
743 }
744 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
745 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
746
747 grp = ext4_get_group_info(sb, e4b->bd_group);
748 list_for_each(cur, &grp->bb_prealloc_list) {
749 ext4_group_t groupnr;
750 struct ext4_prealloc_space *pa;
751 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
752 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
753 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
754 for (i = 0; i < pa->pa_len; i++)
755 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
756 }
757 return 0;
758}
759#undef MB_CHECK_ASSERT
760#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
761 __FILE__, __func__, __LINE__)
762#else
763#define mb_check_buddy(e4b)
764#endif
765
766/*
767 * Divide blocks started from @first with length @len into
768 * smaller chunks with power of 2 blocks.
769 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
770 * then increase bb_counters[] for corresponded chunk size.
771 */
772static void ext4_mb_mark_free_simple(struct super_block *sb,
773 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
774 struct ext4_group_info *grp)
775{
776 struct ext4_sb_info *sbi = EXT4_SB(sb);
777 ext4_grpblk_t min;
778 ext4_grpblk_t max;
779 ext4_grpblk_t chunk;
780 unsigned int border;
781
782 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
783
784 border = 2 << sb->s_blocksize_bits;
785
786 while (len > 0) {
787 /* find how many blocks can be covered since this position */
788 max = ffs(first | border) - 1;
789
790 /* find how many blocks of power 2 we need to mark */
791 min = fls(len) - 1;
792
793 if (max < min)
794 min = max;
795 chunk = 1 << min;
796
797 /* mark multiblock chunks only */
798 grp->bb_counters[min]++;
799 if (min > 0)
800 mb_clear_bit(first >> min,
801 buddy + sbi->s_mb_offsets[min]);
802
803 len -= chunk;
804 first += chunk;
805 }
806}
807
808static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
809{
810 int order;
811
812 /*
813 * We don't bother with a special lists groups with only 1 block free
814 * extents and for completely empty groups.
815 */
816 order = fls(len) - 2;
817 if (order < 0)
818 return 0;
819 if (order == MB_NUM_ORDERS(sb))
820 order--;
821 return order;
822}
823
824/* Move group to appropriate avg_fragment_size list */
825static void
826mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
827{
828 struct ext4_sb_info *sbi = EXT4_SB(sb);
829 int new_order;
830
831 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
832 return;
833
834 new_order = mb_avg_fragment_size_order(sb,
835 grp->bb_free / grp->bb_fragments);
836 if (new_order == grp->bb_avg_fragment_size_order)
837 return;
838
839 if (grp->bb_avg_fragment_size_order != -1) {
840 write_lock(&sbi->s_mb_avg_fragment_size_locks[
841 grp->bb_avg_fragment_size_order]);
842 list_del(&grp->bb_avg_fragment_size_node);
843 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
844 grp->bb_avg_fragment_size_order]);
845 }
846 grp->bb_avg_fragment_size_order = new_order;
847 write_lock(&sbi->s_mb_avg_fragment_size_locks[
848 grp->bb_avg_fragment_size_order]);
849 list_add_tail(&grp->bb_avg_fragment_size_node,
850 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
851 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
852 grp->bb_avg_fragment_size_order]);
853}
854
855/*
856 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
857 * cr level needs an update.
858 */
859static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac,
860 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
861{
862 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
863 struct ext4_group_info *iter, *grp;
864 int i;
865
866 if (ac->ac_status == AC_STATUS_FOUND)
867 return;
868
869 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED))
870 atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
871
872 grp = NULL;
873 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
874 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
875 continue;
876 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
877 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
878 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
879 continue;
880 }
881 grp = NULL;
882 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
883 bb_largest_free_order_node) {
884 if (sbi->s_mb_stats)
885 atomic64_inc(&sbi->s_bal_cX_groups_considered[0]);
886 if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) {
887 grp = iter;
888 break;
889 }
890 }
891 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
892 if (grp)
893 break;
894 }
895
896 if (!grp) {
897 /* Increment cr and search again */
898 *new_cr = 1;
899 } else {
900 *group = grp->bb_group;
901 ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED;
902 }
903}
904
905/*
906 * Choose next group by traversing average fragment size list of suitable
907 * order. Updates *new_cr if cr level needs an update.
908 */
909static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
910 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
911{
912 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
913 struct ext4_group_info *grp = NULL, *iter;
914 int i;
915
916 if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
917 if (sbi->s_mb_stats)
918 atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
919 }
920
921 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
922 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
923 if (list_empty(&sbi->s_mb_avg_fragment_size[i]))
924 continue;
925 read_lock(&sbi->s_mb_avg_fragment_size_locks[i]);
926 if (list_empty(&sbi->s_mb_avg_fragment_size[i])) {
927 read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]);
928 continue;
929 }
930 list_for_each_entry(iter, &sbi->s_mb_avg_fragment_size[i],
931 bb_avg_fragment_size_node) {
932 if (sbi->s_mb_stats)
933 atomic64_inc(&sbi->s_bal_cX_groups_considered[1]);
934 if (likely(ext4_mb_good_group(ac, iter->bb_group, 1))) {
935 grp = iter;
936 break;
937 }
938 }
939 read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]);
940 if (grp)
941 break;
942 }
943
944 if (grp) {
945 *group = grp->bb_group;
946 ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED;
947 } else {
948 *new_cr = 2;
949 }
950}
951
952static inline int should_optimize_scan(struct ext4_allocation_context *ac)
953{
954 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
955 return 0;
956 if (ac->ac_criteria >= 2)
957 return 0;
958 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
959 return 0;
960 return 1;
961}
962
963/*
964 * Return next linear group for allocation. If linear traversal should not be
965 * performed, this function just returns the same group
966 */
967static int
968next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups)
969{
970 if (!should_optimize_scan(ac))
971 goto inc_and_return;
972
973 if (ac->ac_groups_linear_remaining) {
974 ac->ac_groups_linear_remaining--;
975 goto inc_and_return;
976 }
977
978 return group;
979inc_and_return:
980 /*
981 * Artificially restricted ngroups for non-extent
982 * files makes group > ngroups possible on first loop.
983 */
984 return group + 1 >= ngroups ? 0 : group + 1;
985}
986
987/*
988 * ext4_mb_choose_next_group: choose next group for allocation.
989 *
990 * @ac Allocation Context
991 * @new_cr This is an output parameter. If the there is no good group
992 * available at current CR level, this field is updated to indicate
993 * the new cr level that should be used.
994 * @group This is an input / output parameter. As an input it indicates the
995 * next group that the allocator intends to use for allocation. As
996 * output, this field indicates the next group that should be used as
997 * determined by the optimization functions.
998 * @ngroups Total number of groups
999 */
1000static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1001 int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1002{
1003 *new_cr = ac->ac_criteria;
1004
1005 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) {
1006 *group = next_linear_group(ac, *group, ngroups);
1007 return;
1008 }
1009
1010 if (*new_cr == 0) {
1011 ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1012 } else if (*new_cr == 1) {
1013 ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1014 } else {
1015 /*
1016 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1017 * bb_free. But until that happens, we should never come here.
1018 */
1019 WARN_ON(1);
1020 }
1021}
1022
1023/*
1024 * Cache the order of the largest free extent we have available in this block
1025 * group.
1026 */
1027static void
1028mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1029{
1030 struct ext4_sb_info *sbi = EXT4_SB(sb);
1031 int i;
1032
1033 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1034 if (grp->bb_counters[i] > 0)
1035 break;
1036 /* No need to move between order lists? */
1037 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1038 i == grp->bb_largest_free_order) {
1039 grp->bb_largest_free_order = i;
1040 return;
1041 }
1042
1043 if (grp->bb_largest_free_order >= 0) {
1044 write_lock(&sbi->s_mb_largest_free_orders_locks[
1045 grp->bb_largest_free_order]);
1046 list_del_init(&grp->bb_largest_free_order_node);
1047 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1048 grp->bb_largest_free_order]);
1049 }
1050 grp->bb_largest_free_order = i;
1051 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1052 write_lock(&sbi->s_mb_largest_free_orders_locks[
1053 grp->bb_largest_free_order]);
1054 list_add_tail(&grp->bb_largest_free_order_node,
1055 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1056 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1057 grp->bb_largest_free_order]);
1058 }
1059}
1060
1061static noinline_for_stack
1062void ext4_mb_generate_buddy(struct super_block *sb,
1063 void *buddy, void *bitmap, ext4_group_t group)
1064{
1065 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1066 struct ext4_sb_info *sbi = EXT4_SB(sb);
1067 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1068 ext4_grpblk_t i = 0;
1069 ext4_grpblk_t first;
1070 ext4_grpblk_t len;
1071 unsigned free = 0;
1072 unsigned fragments = 0;
1073 unsigned long long period = get_cycles();
1074
1075 /* initialize buddy from bitmap which is aggregation
1076 * of on-disk bitmap and preallocations */
1077 i = mb_find_next_zero_bit(bitmap, max, 0);
1078 grp->bb_first_free = i;
1079 while (i < max) {
1080 fragments++;
1081 first = i;
1082 i = mb_find_next_bit(bitmap, max, i);
1083 len = i - first;
1084 free += len;
1085 if (len > 1)
1086 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1087 else
1088 grp->bb_counters[0]++;
1089 if (i < max)
1090 i = mb_find_next_zero_bit(bitmap, max, i);
1091 }
1092 grp->bb_fragments = fragments;
1093
1094 if (free != grp->bb_free) {
1095 ext4_grp_locked_error(sb, group, 0, 0,
1096 "block bitmap and bg descriptor "
1097 "inconsistent: %u vs %u free clusters",
1098 free, grp->bb_free);
1099 /*
1100 * If we intend to continue, we consider group descriptor
1101 * corrupt and update bb_free using bitmap value
1102 */
1103 grp->bb_free = free;
1104 ext4_mark_group_bitmap_corrupted(sb, group,
1105 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1106 }
1107 mb_set_largest_free_order(sb, grp);
1108 mb_update_avg_fragment_size(sb, grp);
1109
1110 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1111
1112 period = get_cycles() - period;
1113 atomic_inc(&sbi->s_mb_buddies_generated);
1114 atomic64_add(period, &sbi->s_mb_generation_time);
1115}
1116
1117/* The buddy information is attached the buddy cache inode
1118 * for convenience. The information regarding each group
1119 * is loaded via ext4_mb_load_buddy. The information involve
1120 * block bitmap and buddy information. The information are
1121 * stored in the inode as
1122 *
1123 * { page }
1124 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1125 *
1126 *
1127 * one block each for bitmap and buddy information.
1128 * So for each group we take up 2 blocks. A page can
1129 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1130 * So it can have information regarding groups_per_page which
1131 * is blocks_per_page/2
1132 *
1133 * Locking note: This routine takes the block group lock of all groups
1134 * for this page; do not hold this lock when calling this routine!
1135 */
1136
1137static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1138{
1139 ext4_group_t ngroups;
1140 int blocksize;
1141 int blocks_per_page;
1142 int groups_per_page;
1143 int err = 0;
1144 int i;
1145 ext4_group_t first_group, group;
1146 int first_block;
1147 struct super_block *sb;
1148 struct buffer_head *bhs;
1149 struct buffer_head **bh = NULL;
1150 struct inode *inode;
1151 char *data;
1152 char *bitmap;
1153 struct ext4_group_info *grinfo;
1154
1155 inode = page->mapping->host;
1156 sb = inode->i_sb;
1157 ngroups = ext4_get_groups_count(sb);
1158 blocksize = i_blocksize(inode);
1159 blocks_per_page = PAGE_SIZE / blocksize;
1160
1161 mb_debug(sb, "init page %lu\n", page->index);
1162
1163 groups_per_page = blocks_per_page >> 1;
1164 if (groups_per_page == 0)
1165 groups_per_page = 1;
1166
1167 /* allocate buffer_heads to read bitmaps */
1168 if (groups_per_page > 1) {
1169 i = sizeof(struct buffer_head *) * groups_per_page;
1170 bh = kzalloc(i, gfp);
1171 if (bh == NULL) {
1172 err = -ENOMEM;
1173 goto out;
1174 }
1175 } else
1176 bh = &bhs;
1177
1178 first_group = page->index * blocks_per_page / 2;
1179
1180 /* read all groups the page covers into the cache */
1181 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1182 if (group >= ngroups)
1183 break;
1184
1185 grinfo = ext4_get_group_info(sb, group);
1186 /*
1187 * If page is uptodate then we came here after online resize
1188 * which added some new uninitialized group info structs, so
1189 * we must skip all initialized uptodate buddies on the page,
1190 * which may be currently in use by an allocating task.
1191 */
1192 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1193 bh[i] = NULL;
1194 continue;
1195 }
1196 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1197 if (IS_ERR(bh[i])) {
1198 err = PTR_ERR(bh[i]);
1199 bh[i] = NULL;
1200 goto out;
1201 }
1202 mb_debug(sb, "read bitmap for group %u\n", group);
1203 }
1204
1205 /* wait for I/O completion */
1206 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1207 int err2;
1208
1209 if (!bh[i])
1210 continue;
1211 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1212 if (!err)
1213 err = err2;
1214 }
1215
1216 first_block = page->index * blocks_per_page;
1217 for (i = 0; i < blocks_per_page; i++) {
1218 group = (first_block + i) >> 1;
1219 if (group >= ngroups)
1220 break;
1221
1222 if (!bh[group - first_group])
1223 /* skip initialized uptodate buddy */
1224 continue;
1225
1226 if (!buffer_verified(bh[group - first_group]))
1227 /* Skip faulty bitmaps */
1228 continue;
1229 err = 0;
1230
1231 /*
1232 * data carry information regarding this
1233 * particular group in the format specified
1234 * above
1235 *
1236 */
1237 data = page_address(page) + (i * blocksize);
1238 bitmap = bh[group - first_group]->b_data;
1239
1240 /*
1241 * We place the buddy block and bitmap block
1242 * close together
1243 */
1244 if ((first_block + i) & 1) {
1245 /* this is block of buddy */
1246 BUG_ON(incore == NULL);
1247 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1248 group, page->index, i * blocksize);
1249 trace_ext4_mb_buddy_bitmap_load(sb, group);
1250 grinfo = ext4_get_group_info(sb, group);
1251 grinfo->bb_fragments = 0;
1252 memset(grinfo->bb_counters, 0,
1253 sizeof(*grinfo->bb_counters) *
1254 (MB_NUM_ORDERS(sb)));
1255 /*
1256 * incore got set to the group block bitmap below
1257 */
1258 ext4_lock_group(sb, group);
1259 /* init the buddy */
1260 memset(data, 0xff, blocksize);
1261 ext4_mb_generate_buddy(sb, data, incore, group);
1262 ext4_unlock_group(sb, group);
1263 incore = NULL;
1264 } else {
1265 /* this is block of bitmap */
1266 BUG_ON(incore != NULL);
1267 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1268 group, page->index, i * blocksize);
1269 trace_ext4_mb_bitmap_load(sb, group);
1270
1271 /* see comments in ext4_mb_put_pa() */
1272 ext4_lock_group(sb, group);
1273 memcpy(data, bitmap, blocksize);
1274
1275 /* mark all preallocated blks used in in-core bitmap */
1276 ext4_mb_generate_from_pa(sb, data, group);
1277 ext4_mb_generate_from_freelist(sb, data, group);
1278 ext4_unlock_group(sb, group);
1279
1280 /* set incore so that the buddy information can be
1281 * generated using this
1282 */
1283 incore = data;
1284 }
1285 }
1286 SetPageUptodate(page);
1287
1288out:
1289 if (bh) {
1290 for (i = 0; i < groups_per_page; i++)
1291 brelse(bh[i]);
1292 if (bh != &bhs)
1293 kfree(bh);
1294 }
1295 return err;
1296}
1297
1298/*
1299 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1300 * on the same buddy page doesn't happen whild holding the buddy page lock.
1301 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1302 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1303 */
1304static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1305 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1306{
1307 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1308 int block, pnum, poff;
1309 int blocks_per_page;
1310 struct page *page;
1311
1312 e4b->bd_buddy_page = NULL;
1313 e4b->bd_bitmap_page = NULL;
1314
1315 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1316 /*
1317 * the buddy cache inode stores the block bitmap
1318 * and buddy information in consecutive blocks.
1319 * So for each group we need two blocks.
1320 */
1321 block = group * 2;
1322 pnum = block / blocks_per_page;
1323 poff = block % blocks_per_page;
1324 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1325 if (!page)
1326 return -ENOMEM;
1327 BUG_ON(page->mapping != inode->i_mapping);
1328 e4b->bd_bitmap_page = page;
1329 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1330
1331 if (blocks_per_page >= 2) {
1332 /* buddy and bitmap are on the same page */
1333 return 0;
1334 }
1335
1336 block++;
1337 pnum = block / blocks_per_page;
1338 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1339 if (!page)
1340 return -ENOMEM;
1341 BUG_ON(page->mapping != inode->i_mapping);
1342 e4b->bd_buddy_page = page;
1343 return 0;
1344}
1345
1346static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1347{
1348 if (e4b->bd_bitmap_page) {
1349 unlock_page(e4b->bd_bitmap_page);
1350 put_page(e4b->bd_bitmap_page);
1351 }
1352 if (e4b->bd_buddy_page) {
1353 unlock_page(e4b->bd_buddy_page);
1354 put_page(e4b->bd_buddy_page);
1355 }
1356}
1357
1358/*
1359 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1360 * block group lock of all groups for this page; do not hold the BG lock when
1361 * calling this routine!
1362 */
1363static noinline_for_stack
1364int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1365{
1366
1367 struct ext4_group_info *this_grp;
1368 struct ext4_buddy e4b;
1369 struct page *page;
1370 int ret = 0;
1371
1372 might_sleep();
1373 mb_debug(sb, "init group %u\n", group);
1374 this_grp = ext4_get_group_info(sb, group);
1375 /*
1376 * This ensures that we don't reinit the buddy cache
1377 * page which map to the group from which we are already
1378 * allocating. If we are looking at the buddy cache we would
1379 * have taken a reference using ext4_mb_load_buddy and that
1380 * would have pinned buddy page to page cache.
1381 * The call to ext4_mb_get_buddy_page_lock will mark the
1382 * page accessed.
1383 */
1384 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1385 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1386 /*
1387 * somebody initialized the group
1388 * return without doing anything
1389 */
1390 goto err;
1391 }
1392
1393 page = e4b.bd_bitmap_page;
1394 ret = ext4_mb_init_cache(page, NULL, gfp);
1395 if (ret)
1396 goto err;
1397 if (!PageUptodate(page)) {
1398 ret = -EIO;
1399 goto err;
1400 }
1401
1402 if (e4b.bd_buddy_page == NULL) {
1403 /*
1404 * If both the bitmap and buddy are in
1405 * the same page we don't need to force
1406 * init the buddy
1407 */
1408 ret = 0;
1409 goto err;
1410 }
1411 /* init buddy cache */
1412 page = e4b.bd_buddy_page;
1413 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1414 if (ret)
1415 goto err;
1416 if (!PageUptodate(page)) {
1417 ret = -EIO;
1418 goto err;
1419 }
1420err:
1421 ext4_mb_put_buddy_page_lock(&e4b);
1422 return ret;
1423}
1424
1425/*
1426 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1427 * block group lock of all groups for this page; do not hold the BG lock when
1428 * calling this routine!
1429 */
1430static noinline_for_stack int
1431ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1432 struct ext4_buddy *e4b, gfp_t gfp)
1433{
1434 int blocks_per_page;
1435 int block;
1436 int pnum;
1437 int poff;
1438 struct page *page;
1439 int ret;
1440 struct ext4_group_info *grp;
1441 struct ext4_sb_info *sbi = EXT4_SB(sb);
1442 struct inode *inode = sbi->s_buddy_cache;
1443
1444 might_sleep();
1445 mb_debug(sb, "load group %u\n", group);
1446
1447 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1448 grp = ext4_get_group_info(sb, group);
1449
1450 e4b->bd_blkbits = sb->s_blocksize_bits;
1451 e4b->bd_info = grp;
1452 e4b->bd_sb = sb;
1453 e4b->bd_group = group;
1454 e4b->bd_buddy_page = NULL;
1455 e4b->bd_bitmap_page = NULL;
1456
1457 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1458 /*
1459 * we need full data about the group
1460 * to make a good selection
1461 */
1462 ret = ext4_mb_init_group(sb, group, gfp);
1463 if (ret)
1464 return ret;
1465 }
1466
1467 /*
1468 * the buddy cache inode stores the block bitmap
1469 * and buddy information in consecutive blocks.
1470 * So for each group we need two blocks.
1471 */
1472 block = group * 2;
1473 pnum = block / blocks_per_page;
1474 poff = block % blocks_per_page;
1475
1476 /* we could use find_or_create_page(), but it locks page
1477 * what we'd like to avoid in fast path ... */
1478 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1479 if (page == NULL || !PageUptodate(page)) {
1480 if (page)
1481 /*
1482 * drop the page reference and try
1483 * to get the page with lock. If we
1484 * are not uptodate that implies
1485 * somebody just created the page but
1486 * is yet to initialize the same. So
1487 * wait for it to initialize.
1488 */
1489 put_page(page);
1490 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1491 if (page) {
1492 BUG_ON(page->mapping != inode->i_mapping);
1493 if (!PageUptodate(page)) {
1494 ret = ext4_mb_init_cache(page, NULL, gfp);
1495 if (ret) {
1496 unlock_page(page);
1497 goto err;
1498 }
1499 mb_cmp_bitmaps(e4b, page_address(page) +
1500 (poff * sb->s_blocksize));
1501 }
1502 unlock_page(page);
1503 }
1504 }
1505 if (page == NULL) {
1506 ret = -ENOMEM;
1507 goto err;
1508 }
1509 if (!PageUptodate(page)) {
1510 ret = -EIO;
1511 goto err;
1512 }
1513
1514 /* Pages marked accessed already */
1515 e4b->bd_bitmap_page = page;
1516 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1517
1518 block++;
1519 pnum = block / blocks_per_page;
1520 poff = block % blocks_per_page;
1521
1522 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1523 if (page == NULL || !PageUptodate(page)) {
1524 if (page)
1525 put_page(page);
1526 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1527 if (page) {
1528 BUG_ON(page->mapping != inode->i_mapping);
1529 if (!PageUptodate(page)) {
1530 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1531 gfp);
1532 if (ret) {
1533 unlock_page(page);
1534 goto err;
1535 }
1536 }
1537 unlock_page(page);
1538 }
1539 }
1540 if (page == NULL) {
1541 ret = -ENOMEM;
1542 goto err;
1543 }
1544 if (!PageUptodate(page)) {
1545 ret = -EIO;
1546 goto err;
1547 }
1548
1549 /* Pages marked accessed already */
1550 e4b->bd_buddy_page = page;
1551 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1552
1553 return 0;
1554
1555err:
1556 if (page)
1557 put_page(page);
1558 if (e4b->bd_bitmap_page)
1559 put_page(e4b->bd_bitmap_page);
1560 if (e4b->bd_buddy_page)
1561 put_page(e4b->bd_buddy_page);
1562 e4b->bd_buddy = NULL;
1563 e4b->bd_bitmap = NULL;
1564 return ret;
1565}
1566
1567static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1568 struct ext4_buddy *e4b)
1569{
1570 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1571}
1572
1573static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1574{
1575 if (e4b->bd_bitmap_page)
1576 put_page(e4b->bd_bitmap_page);
1577 if (e4b->bd_buddy_page)
1578 put_page(e4b->bd_buddy_page);
1579}
1580
1581
1582static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1583{
1584 int order = 1, max;
1585 void *bb;
1586
1587 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1588 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1589
1590 while (order <= e4b->bd_blkbits + 1) {
1591 bb = mb_find_buddy(e4b, order, &max);
1592 if (!mb_test_bit(block >> order, bb)) {
1593 /* this block is part of buddy of order 'order' */
1594 return order;
1595 }
1596 order++;
1597 }
1598 return 0;
1599}
1600
1601static void mb_clear_bits(void *bm, int cur, int len)
1602{
1603 __u32 *addr;
1604
1605 len = cur + len;
1606 while (cur < len) {
1607 if ((cur & 31) == 0 && (len - cur) >= 32) {
1608 /* fast path: clear whole word at once */
1609 addr = bm + (cur >> 3);
1610 *addr = 0;
1611 cur += 32;
1612 continue;
1613 }
1614 mb_clear_bit(cur, bm);
1615 cur++;
1616 }
1617}
1618
1619/* clear bits in given range
1620 * will return first found zero bit if any, -1 otherwise
1621 */
1622static int mb_test_and_clear_bits(void *bm, int cur, int len)
1623{
1624 __u32 *addr;
1625 int zero_bit = -1;
1626
1627 len = cur + len;
1628 while (cur < len) {
1629 if ((cur & 31) == 0 && (len - cur) >= 32) {
1630 /* fast path: clear whole word at once */
1631 addr = bm + (cur >> 3);
1632 if (*addr != (__u32)(-1) && zero_bit == -1)
1633 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1634 *addr = 0;
1635 cur += 32;
1636 continue;
1637 }
1638 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1639 zero_bit = cur;
1640 cur++;
1641 }
1642
1643 return zero_bit;
1644}
1645
1646void mb_set_bits(void *bm, int cur, int len)
1647{
1648 __u32 *addr;
1649
1650 len = cur + len;
1651 while (cur < len) {
1652 if ((cur & 31) == 0 && (len - cur) >= 32) {
1653 /* fast path: set whole word at once */
1654 addr = bm + (cur >> 3);
1655 *addr = 0xffffffff;
1656 cur += 32;
1657 continue;
1658 }
1659 mb_set_bit(cur, bm);
1660 cur++;
1661 }
1662}
1663
1664static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1665{
1666 if (mb_test_bit(*bit + side, bitmap)) {
1667 mb_clear_bit(*bit, bitmap);
1668 (*bit) -= side;
1669 return 1;
1670 }
1671 else {
1672 (*bit) += side;
1673 mb_set_bit(*bit, bitmap);
1674 return -1;
1675 }
1676}
1677
1678static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1679{
1680 int max;
1681 int order = 1;
1682 void *buddy = mb_find_buddy(e4b, order, &max);
1683
1684 while (buddy) {
1685 void *buddy2;
1686
1687 /* Bits in range [first; last] are known to be set since
1688 * corresponding blocks were allocated. Bits in range
1689 * (first; last) will stay set because they form buddies on
1690 * upper layer. We just deal with borders if they don't
1691 * align with upper layer and then go up.
1692 * Releasing entire group is all about clearing
1693 * single bit of highest order buddy.
1694 */
1695
1696 /* Example:
1697 * ---------------------------------
1698 * | 1 | 1 | 1 | 1 |
1699 * ---------------------------------
1700 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1701 * ---------------------------------
1702 * 0 1 2 3 4 5 6 7
1703 * \_____________________/
1704 *
1705 * Neither [1] nor [6] is aligned to above layer.
1706 * Left neighbour [0] is free, so mark it busy,
1707 * decrease bb_counters and extend range to
1708 * [0; 6]
1709 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1710 * mark [6] free, increase bb_counters and shrink range to
1711 * [0; 5].
1712 * Then shift range to [0; 2], go up and do the same.
1713 */
1714
1715
1716 if (first & 1)
1717 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1718 if (!(last & 1))
1719 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1720 if (first > last)
1721 break;
1722 order++;
1723
1724 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1725 mb_clear_bits(buddy, first, last - first + 1);
1726 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1727 break;
1728 }
1729 first >>= 1;
1730 last >>= 1;
1731 buddy = buddy2;
1732 }
1733}
1734
1735static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1736 int first, int count)
1737{
1738 int left_is_free = 0;
1739 int right_is_free = 0;
1740 int block;
1741 int last = first + count - 1;
1742 struct super_block *sb = e4b->bd_sb;
1743
1744 if (WARN_ON(count == 0))
1745 return;
1746 BUG_ON(last >= (sb->s_blocksize << 3));
1747 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1748 /* Don't bother if the block group is corrupt. */
1749 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1750 return;
1751
1752 mb_check_buddy(e4b);
1753 mb_free_blocks_double(inode, e4b, first, count);
1754
1755 this_cpu_inc(discard_pa_seq);
1756 e4b->bd_info->bb_free += count;
1757 if (first < e4b->bd_info->bb_first_free)
1758 e4b->bd_info->bb_first_free = first;
1759
1760 /* access memory sequentially: check left neighbour,
1761 * clear range and then check right neighbour
1762 */
1763 if (first != 0)
1764 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1765 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1766 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1767 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1768
1769 if (unlikely(block != -1)) {
1770 struct ext4_sb_info *sbi = EXT4_SB(sb);
1771 ext4_fsblk_t blocknr;
1772
1773 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1774 blocknr += EXT4_C2B(sbi, block);
1775 if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1776 ext4_grp_locked_error(sb, e4b->bd_group,
1777 inode ? inode->i_ino : 0,
1778 blocknr,
1779 "freeing already freed block (bit %u); block bitmap corrupt.",
1780 block);
1781 ext4_mark_group_bitmap_corrupted(
1782 sb, e4b->bd_group,
1783 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1784 }
1785 goto done;
1786 }
1787
1788 /* let's maintain fragments counter */
1789 if (left_is_free && right_is_free)
1790 e4b->bd_info->bb_fragments--;
1791 else if (!left_is_free && !right_is_free)
1792 e4b->bd_info->bb_fragments++;
1793
1794 /* buddy[0] == bd_bitmap is a special case, so handle
1795 * it right away and let mb_buddy_mark_free stay free of
1796 * zero order checks.
1797 * Check if neighbours are to be coaleasced,
1798 * adjust bitmap bb_counters and borders appropriately.
1799 */
1800 if (first & 1) {
1801 first += !left_is_free;
1802 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1803 }
1804 if (!(last & 1)) {
1805 last -= !right_is_free;
1806 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1807 }
1808
1809 if (first <= last)
1810 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1811
1812done:
1813 mb_set_largest_free_order(sb, e4b->bd_info);
1814 mb_update_avg_fragment_size(sb, e4b->bd_info);
1815 mb_check_buddy(e4b);
1816}
1817
1818static int mb_find_extent(struct ext4_buddy *e4b, int block,
1819 int needed, struct ext4_free_extent *ex)
1820{
1821 int next = block;
1822 int max, order;
1823 void *buddy;
1824
1825 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1826 BUG_ON(ex == NULL);
1827
1828 buddy = mb_find_buddy(e4b, 0, &max);
1829 BUG_ON(buddy == NULL);
1830 BUG_ON(block >= max);
1831 if (mb_test_bit(block, buddy)) {
1832 ex->fe_len = 0;
1833 ex->fe_start = 0;
1834 ex->fe_group = 0;
1835 return 0;
1836 }
1837
1838 /* find actual order */
1839 order = mb_find_order_for_block(e4b, block);
1840 block = block >> order;
1841
1842 ex->fe_len = 1 << order;
1843 ex->fe_start = block << order;
1844 ex->fe_group = e4b->bd_group;
1845
1846 /* calc difference from given start */
1847 next = next - ex->fe_start;
1848 ex->fe_len -= next;
1849 ex->fe_start += next;
1850
1851 while (needed > ex->fe_len &&
1852 mb_find_buddy(e4b, order, &max)) {
1853
1854 if (block + 1 >= max)
1855 break;
1856
1857 next = (block + 1) * (1 << order);
1858 if (mb_test_bit(next, e4b->bd_bitmap))
1859 break;
1860
1861 order = mb_find_order_for_block(e4b, next);
1862
1863 block = next >> order;
1864 ex->fe_len += 1 << order;
1865 }
1866
1867 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1868 /* Should never happen! (but apparently sometimes does?!?) */
1869 WARN_ON(1);
1870 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1871 "corruption or bug in mb_find_extent "
1872 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1873 block, order, needed, ex->fe_group, ex->fe_start,
1874 ex->fe_len, ex->fe_logical);
1875 ex->fe_len = 0;
1876 ex->fe_start = 0;
1877 ex->fe_group = 0;
1878 }
1879 return ex->fe_len;
1880}
1881
1882static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1883{
1884 int ord;
1885 int mlen = 0;
1886 int max = 0;
1887 int cur;
1888 int start = ex->fe_start;
1889 int len = ex->fe_len;
1890 unsigned ret = 0;
1891 int len0 = len;
1892 void *buddy;
1893 bool split = false;
1894
1895 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1896 BUG_ON(e4b->bd_group != ex->fe_group);
1897 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1898 mb_check_buddy(e4b);
1899 mb_mark_used_double(e4b, start, len);
1900
1901 this_cpu_inc(discard_pa_seq);
1902 e4b->bd_info->bb_free -= len;
1903 if (e4b->bd_info->bb_first_free == start)
1904 e4b->bd_info->bb_first_free += len;
1905
1906 /* let's maintain fragments counter */
1907 if (start != 0)
1908 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1909 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1910 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1911 if (mlen && max)
1912 e4b->bd_info->bb_fragments++;
1913 else if (!mlen && !max)
1914 e4b->bd_info->bb_fragments--;
1915
1916 /* let's maintain buddy itself */
1917 while (len) {
1918 if (!split)
1919 ord = mb_find_order_for_block(e4b, start);
1920
1921 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1922 /* the whole chunk may be allocated at once! */
1923 mlen = 1 << ord;
1924 if (!split)
1925 buddy = mb_find_buddy(e4b, ord, &max);
1926 else
1927 split = false;
1928 BUG_ON((start >> ord) >= max);
1929 mb_set_bit(start >> ord, buddy);
1930 e4b->bd_info->bb_counters[ord]--;
1931 start += mlen;
1932 len -= mlen;
1933 BUG_ON(len < 0);
1934 continue;
1935 }
1936
1937 /* store for history */
1938 if (ret == 0)
1939 ret = len | (ord << 16);
1940
1941 /* we have to split large buddy */
1942 BUG_ON(ord <= 0);
1943 buddy = mb_find_buddy(e4b, ord, &max);
1944 mb_set_bit(start >> ord, buddy);
1945 e4b->bd_info->bb_counters[ord]--;
1946
1947 ord--;
1948 cur = (start >> ord) & ~1U;
1949 buddy = mb_find_buddy(e4b, ord, &max);
1950 mb_clear_bit(cur, buddy);
1951 mb_clear_bit(cur + 1, buddy);
1952 e4b->bd_info->bb_counters[ord]++;
1953 e4b->bd_info->bb_counters[ord]++;
1954 split = true;
1955 }
1956 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1957
1958 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1959 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1960 mb_check_buddy(e4b);
1961
1962 return ret;
1963}
1964
1965/*
1966 * Must be called under group lock!
1967 */
1968static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1969 struct ext4_buddy *e4b)
1970{
1971 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1972 int ret;
1973
1974 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1975 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1976
1977 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1978 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1979 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1980
1981 /* preallocation can change ac_b_ex, thus we store actually
1982 * allocated blocks for history */
1983 ac->ac_f_ex = ac->ac_b_ex;
1984
1985 ac->ac_status = AC_STATUS_FOUND;
1986 ac->ac_tail = ret & 0xffff;
1987 ac->ac_buddy = ret >> 16;
1988
1989 /*
1990 * take the page reference. We want the page to be pinned
1991 * so that we don't get a ext4_mb_init_cache_call for this
1992 * group until we update the bitmap. That would mean we
1993 * double allocate blocks. The reference is dropped
1994 * in ext4_mb_release_context
1995 */
1996 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1997 get_page(ac->ac_bitmap_page);
1998 ac->ac_buddy_page = e4b->bd_buddy_page;
1999 get_page(ac->ac_buddy_page);
2000 /* store last allocated for subsequent stream allocation */
2001 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2002 spin_lock(&sbi->s_md_lock);
2003 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2004 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2005 spin_unlock(&sbi->s_md_lock);
2006 }
2007 /*
2008 * As we've just preallocated more space than
2009 * user requested originally, we store allocated
2010 * space in a special descriptor.
2011 */
2012 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2013 ext4_mb_new_preallocation(ac);
2014
2015}
2016
2017static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2018 struct ext4_buddy *e4b,
2019 int finish_group)
2020{
2021 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2022 struct ext4_free_extent *bex = &ac->ac_b_ex;
2023 struct ext4_free_extent *gex = &ac->ac_g_ex;
2024 struct ext4_free_extent ex;
2025 int max;
2026
2027 if (ac->ac_status == AC_STATUS_FOUND)
2028 return;
2029 /*
2030 * We don't want to scan for a whole year
2031 */
2032 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2033 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2034 ac->ac_status = AC_STATUS_BREAK;
2035 return;
2036 }
2037
2038 /*
2039 * Haven't found good chunk so far, let's continue
2040 */
2041 if (bex->fe_len < gex->fe_len)
2042 return;
2043
2044 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2045 && bex->fe_group == e4b->bd_group) {
2046 /* recheck chunk's availability - we don't know
2047 * when it was found (within this lock-unlock
2048 * period or not) */
2049 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2050 if (max >= gex->fe_len) {
2051 ext4_mb_use_best_found(ac, e4b);
2052 return;
2053 }
2054 }
2055}
2056
2057/*
2058 * The routine checks whether found extent is good enough. If it is,
2059 * then the extent gets marked used and flag is set to the context
2060 * to stop scanning. Otherwise, the extent is compared with the
2061 * previous found extent and if new one is better, then it's stored
2062 * in the context. Later, the best found extent will be used, if
2063 * mballoc can't find good enough extent.
2064 *
2065 * FIXME: real allocation policy is to be designed yet!
2066 */
2067static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2068 struct ext4_free_extent *ex,
2069 struct ext4_buddy *e4b)
2070{
2071 struct ext4_free_extent *bex = &ac->ac_b_ex;
2072 struct ext4_free_extent *gex = &ac->ac_g_ex;
2073
2074 BUG_ON(ex->fe_len <= 0);
2075 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2076 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2077 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2078
2079 ac->ac_found++;
2080
2081 /*
2082 * The special case - take what you catch first
2083 */
2084 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2085 *bex = *ex;
2086 ext4_mb_use_best_found(ac, e4b);
2087 return;
2088 }
2089
2090 /*
2091 * Let's check whether the chuck is good enough
2092 */
2093 if (ex->fe_len == gex->fe_len) {
2094 *bex = *ex;
2095 ext4_mb_use_best_found(ac, e4b);
2096 return;
2097 }
2098
2099 /*
2100 * If this is first found extent, just store it in the context
2101 */
2102 if (bex->fe_len == 0) {
2103 *bex = *ex;
2104 return;
2105 }
2106
2107 /*
2108 * If new found extent is better, store it in the context
2109 */
2110 if (bex->fe_len < gex->fe_len) {
2111 /* if the request isn't satisfied, any found extent
2112 * larger than previous best one is better */
2113 if (ex->fe_len > bex->fe_len)
2114 *bex = *ex;
2115 } else if (ex->fe_len > gex->fe_len) {
2116 /* if the request is satisfied, then we try to find
2117 * an extent that still satisfy the request, but is
2118 * smaller than previous one */
2119 if (ex->fe_len < bex->fe_len)
2120 *bex = *ex;
2121 }
2122
2123 ext4_mb_check_limits(ac, e4b, 0);
2124}
2125
2126static noinline_for_stack
2127int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2128 struct ext4_buddy *e4b)
2129{
2130 struct ext4_free_extent ex = ac->ac_b_ex;
2131 ext4_group_t group = ex.fe_group;
2132 int max;
2133 int err;
2134
2135 BUG_ON(ex.fe_len <= 0);
2136 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2137 if (err)
2138 return err;
2139
2140 ext4_lock_group(ac->ac_sb, group);
2141 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2142
2143 if (max > 0) {
2144 ac->ac_b_ex = ex;
2145 ext4_mb_use_best_found(ac, e4b);
2146 }
2147
2148 ext4_unlock_group(ac->ac_sb, group);
2149 ext4_mb_unload_buddy(e4b);
2150
2151 return 0;
2152}
2153
2154static noinline_for_stack
2155int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2156 struct ext4_buddy *e4b)
2157{
2158 ext4_group_t group = ac->ac_g_ex.fe_group;
2159 int max;
2160 int err;
2161 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2162 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2163 struct ext4_free_extent ex;
2164
2165 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2166 return 0;
2167 if (grp->bb_free == 0)
2168 return 0;
2169
2170 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2171 if (err)
2172 return err;
2173
2174 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2175 ext4_mb_unload_buddy(e4b);
2176 return 0;
2177 }
2178
2179 ext4_lock_group(ac->ac_sb, group);
2180 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2181 ac->ac_g_ex.fe_len, &ex);
2182 ex.fe_logical = 0xDEADFA11; /* debug value */
2183
2184 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2185 ext4_fsblk_t start;
2186
2187 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2188 ex.fe_start;
2189 /* use do_div to get remainder (would be 64-bit modulo) */
2190 if (do_div(start, sbi->s_stripe) == 0) {
2191 ac->ac_found++;
2192 ac->ac_b_ex = ex;
2193 ext4_mb_use_best_found(ac, e4b);
2194 }
2195 } else if (max >= ac->ac_g_ex.fe_len) {
2196 BUG_ON(ex.fe_len <= 0);
2197 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2198 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2199 ac->ac_found++;
2200 ac->ac_b_ex = ex;
2201 ext4_mb_use_best_found(ac, e4b);
2202 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2203 /* Sometimes, caller may want to merge even small
2204 * number of blocks to an existing extent */
2205 BUG_ON(ex.fe_len <= 0);
2206 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2207 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2208 ac->ac_found++;
2209 ac->ac_b_ex = ex;
2210 ext4_mb_use_best_found(ac, e4b);
2211 }
2212 ext4_unlock_group(ac->ac_sb, group);
2213 ext4_mb_unload_buddy(e4b);
2214
2215 return 0;
2216}
2217
2218/*
2219 * The routine scans buddy structures (not bitmap!) from given order
2220 * to max order and tries to find big enough chunk to satisfy the req
2221 */
2222static noinline_for_stack
2223void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2224 struct ext4_buddy *e4b)
2225{
2226 struct super_block *sb = ac->ac_sb;
2227 struct ext4_group_info *grp = e4b->bd_info;
2228 void *buddy;
2229 int i;
2230 int k;
2231 int max;
2232
2233 BUG_ON(ac->ac_2order <= 0);
2234 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2235 if (grp->bb_counters[i] == 0)
2236 continue;
2237
2238 buddy = mb_find_buddy(e4b, i, &max);
2239 BUG_ON(buddy == NULL);
2240
2241 k = mb_find_next_zero_bit(buddy, max, 0);
2242 if (k >= max) {
2243 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2244 "%d free clusters of order %d. But found 0",
2245 grp->bb_counters[i], i);
2246 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2247 e4b->bd_group,
2248 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2249 break;
2250 }
2251 ac->ac_found++;
2252
2253 ac->ac_b_ex.fe_len = 1 << i;
2254 ac->ac_b_ex.fe_start = k << i;
2255 ac->ac_b_ex.fe_group = e4b->bd_group;
2256
2257 ext4_mb_use_best_found(ac, e4b);
2258
2259 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2260
2261 if (EXT4_SB(sb)->s_mb_stats)
2262 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2263
2264 break;
2265 }
2266}
2267
2268/*
2269 * The routine scans the group and measures all found extents.
2270 * In order to optimize scanning, caller must pass number of
2271 * free blocks in the group, so the routine can know upper limit.
2272 */
2273static noinline_for_stack
2274void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2275 struct ext4_buddy *e4b)
2276{
2277 struct super_block *sb = ac->ac_sb;
2278 void *bitmap = e4b->bd_bitmap;
2279 struct ext4_free_extent ex;
2280 int i;
2281 int free;
2282
2283 free = e4b->bd_info->bb_free;
2284 if (WARN_ON(free <= 0))
2285 return;
2286
2287 i = e4b->bd_info->bb_first_free;
2288
2289 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2290 i = mb_find_next_zero_bit(bitmap,
2291 EXT4_CLUSTERS_PER_GROUP(sb), i);
2292 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2293 /*
2294 * IF we have corrupt bitmap, we won't find any
2295 * free blocks even though group info says we
2296 * have free blocks
2297 */
2298 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2299 "%d free clusters as per "
2300 "group info. But bitmap says 0",
2301 free);
2302 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2303 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2304 break;
2305 }
2306
2307 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2308 if (WARN_ON(ex.fe_len <= 0))
2309 break;
2310 if (free < ex.fe_len) {
2311 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2312 "%d free clusters as per "
2313 "group info. But got %d blocks",
2314 free, ex.fe_len);
2315 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2316 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2317 /*
2318 * The number of free blocks differs. This mostly
2319 * indicate that the bitmap is corrupt. So exit
2320 * without claiming the space.
2321 */
2322 break;
2323 }
2324 ex.fe_logical = 0xDEADC0DE; /* debug value */
2325 ext4_mb_measure_extent(ac, &ex, e4b);
2326
2327 i += ex.fe_len;
2328 free -= ex.fe_len;
2329 }
2330
2331 ext4_mb_check_limits(ac, e4b, 1);
2332}
2333
2334/*
2335 * This is a special case for storages like raid5
2336 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2337 */
2338static noinline_for_stack
2339void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2340 struct ext4_buddy *e4b)
2341{
2342 struct super_block *sb = ac->ac_sb;
2343 struct ext4_sb_info *sbi = EXT4_SB(sb);
2344 void *bitmap = e4b->bd_bitmap;
2345 struct ext4_free_extent ex;
2346 ext4_fsblk_t first_group_block;
2347 ext4_fsblk_t a;
2348 ext4_grpblk_t i;
2349 int max;
2350
2351 BUG_ON(sbi->s_stripe == 0);
2352
2353 /* find first stripe-aligned block in group */
2354 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2355
2356 a = first_group_block + sbi->s_stripe - 1;
2357 do_div(a, sbi->s_stripe);
2358 i = (a * sbi->s_stripe) - first_group_block;
2359
2360 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2361 if (!mb_test_bit(i, bitmap)) {
2362 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2363 if (max >= sbi->s_stripe) {
2364 ac->ac_found++;
2365 ex.fe_logical = 0xDEADF00D; /* debug value */
2366 ac->ac_b_ex = ex;
2367 ext4_mb_use_best_found(ac, e4b);
2368 break;
2369 }
2370 }
2371 i += sbi->s_stripe;
2372 }
2373}
2374
2375/*
2376 * This is also called BEFORE we load the buddy bitmap.
2377 * Returns either 1 or 0 indicating that the group is either suitable
2378 * for the allocation or not.
2379 */
2380static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2381 ext4_group_t group, int cr)
2382{
2383 ext4_grpblk_t free, fragments;
2384 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2385 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2386
2387 BUG_ON(cr < 0 || cr >= 4);
2388
2389 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2390 return false;
2391
2392 free = grp->bb_free;
2393 if (free == 0)
2394 return false;
2395
2396 fragments = grp->bb_fragments;
2397 if (fragments == 0)
2398 return false;
2399
2400 switch (cr) {
2401 case 0:
2402 BUG_ON(ac->ac_2order == 0);
2403
2404 /* Avoid using the first bg of a flexgroup for data files */
2405 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2406 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2407 ((group % flex_size) == 0))
2408 return false;
2409
2410 if (free < ac->ac_g_ex.fe_len)
2411 return false;
2412
2413 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2414 return true;
2415
2416 if (grp->bb_largest_free_order < ac->ac_2order)
2417 return false;
2418
2419 return true;
2420 case 1:
2421 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2422 return true;
2423 break;
2424 case 2:
2425 if (free >= ac->ac_g_ex.fe_len)
2426 return true;
2427 break;
2428 case 3:
2429 return true;
2430 default:
2431 BUG();
2432 }
2433
2434 return false;
2435}
2436
2437/*
2438 * This could return negative error code if something goes wrong
2439 * during ext4_mb_init_group(). This should not be called with
2440 * ext4_lock_group() held.
2441 *
2442 * Note: because we are conditionally operating with the group lock in
2443 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2444 * function using __acquire and __release. This means we need to be
2445 * super careful before messing with the error path handling via "goto
2446 * out"!
2447 */
2448static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2449 ext4_group_t group, int cr)
2450{
2451 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2452 struct super_block *sb = ac->ac_sb;
2453 struct ext4_sb_info *sbi = EXT4_SB(sb);
2454 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2455 ext4_grpblk_t free;
2456 int ret = 0;
2457
2458 if (sbi->s_mb_stats)
2459 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2460 if (should_lock) {
2461 ext4_lock_group(sb, group);
2462 __release(ext4_group_lock_ptr(sb, group));
2463 }
2464 free = grp->bb_free;
2465 if (free == 0)
2466 goto out;
2467 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2468 goto out;
2469 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2470 goto out;
2471 if (should_lock) {
2472 __acquire(ext4_group_lock_ptr(sb, group));
2473 ext4_unlock_group(sb, group);
2474 }
2475
2476 /* We only do this if the grp has never been initialized */
2477 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2478 struct ext4_group_desc *gdp =
2479 ext4_get_group_desc(sb, group, NULL);
2480 int ret;
2481
2482 /* cr=0/1 is a very optimistic search to find large
2483 * good chunks almost for free. If buddy data is not
2484 * ready, then this optimization makes no sense. But
2485 * we never skip the first block group in a flex_bg,
2486 * since this gets used for metadata block allocation,
2487 * and we want to make sure we locate metadata blocks
2488 * in the first block group in the flex_bg if possible.
2489 */
2490 if (cr < 2 &&
2491 (!sbi->s_log_groups_per_flex ||
2492 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2493 !(ext4_has_group_desc_csum(sb) &&
2494 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2495 return 0;
2496 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2497 if (ret)
2498 return ret;
2499 }
2500
2501 if (should_lock) {
2502 ext4_lock_group(sb, group);
2503 __release(ext4_group_lock_ptr(sb, group));
2504 }
2505 ret = ext4_mb_good_group(ac, group, cr);
2506out:
2507 if (should_lock) {
2508 __acquire(ext4_group_lock_ptr(sb, group));
2509 ext4_unlock_group(sb, group);
2510 }
2511 return ret;
2512}
2513
2514/*
2515 * Start prefetching @nr block bitmaps starting at @group.
2516 * Return the next group which needs to be prefetched.
2517 */
2518ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2519 unsigned int nr, int *cnt)
2520{
2521 ext4_group_t ngroups = ext4_get_groups_count(sb);
2522 struct buffer_head *bh;
2523 struct blk_plug plug;
2524
2525 blk_start_plug(&plug);
2526 while (nr-- > 0) {
2527 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2528 NULL);
2529 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2530
2531 /*
2532 * Prefetch block groups with free blocks; but don't
2533 * bother if it is marked uninitialized on disk, since
2534 * it won't require I/O to read. Also only try to
2535 * prefetch once, so we avoid getblk() call, which can
2536 * be expensive.
2537 */
2538 if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2539 EXT4_MB_GRP_NEED_INIT(grp) &&
2540 ext4_free_group_clusters(sb, gdp) > 0 &&
2541 !(ext4_has_group_desc_csum(sb) &&
2542 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2543 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2544 if (bh && !IS_ERR(bh)) {
2545 if (!buffer_uptodate(bh) && cnt)
2546 (*cnt)++;
2547 brelse(bh);
2548 }
2549 }
2550 if (++group >= ngroups)
2551 group = 0;
2552 }
2553 blk_finish_plug(&plug);
2554 return group;
2555}
2556
2557/*
2558 * Prefetching reads the block bitmap into the buffer cache; but we
2559 * need to make sure that the buddy bitmap in the page cache has been
2560 * initialized. Note that ext4_mb_init_group() will block if the I/O
2561 * is not yet completed, or indeed if it was not initiated by
2562 * ext4_mb_prefetch did not start the I/O.
2563 *
2564 * TODO: We should actually kick off the buddy bitmap setup in a work
2565 * queue when the buffer I/O is completed, so that we don't block
2566 * waiting for the block allocation bitmap read to finish when
2567 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2568 */
2569void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2570 unsigned int nr)
2571{
2572 while (nr-- > 0) {
2573 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2574 NULL);
2575 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2576
2577 if (!group)
2578 group = ext4_get_groups_count(sb);
2579 group--;
2580 grp = ext4_get_group_info(sb, group);
2581
2582 if (EXT4_MB_GRP_NEED_INIT(grp) &&
2583 ext4_free_group_clusters(sb, gdp) > 0 &&
2584 !(ext4_has_group_desc_csum(sb) &&
2585 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2586 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2587 break;
2588 }
2589 }
2590}
2591
2592static noinline_for_stack int
2593ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2594{
2595 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2596 int cr = -1, new_cr;
2597 int err = 0, first_err = 0;
2598 unsigned int nr = 0, prefetch_ios = 0;
2599 struct ext4_sb_info *sbi;
2600 struct super_block *sb;
2601 struct ext4_buddy e4b;
2602 int lost;
2603
2604 sb = ac->ac_sb;
2605 sbi = EXT4_SB(sb);
2606 ngroups = ext4_get_groups_count(sb);
2607 /* non-extent files are limited to low blocks/groups */
2608 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2609 ngroups = sbi->s_blockfile_groups;
2610
2611 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2612
2613 /* first, try the goal */
2614 err = ext4_mb_find_by_goal(ac, &e4b);
2615 if (err || ac->ac_status == AC_STATUS_FOUND)
2616 goto out;
2617
2618 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2619 goto out;
2620
2621 /*
2622 * ac->ac_2order is set only if the fe_len is a power of 2
2623 * if ac->ac_2order is set we also set criteria to 0 so that we
2624 * try exact allocation using buddy.
2625 */
2626 i = fls(ac->ac_g_ex.fe_len);
2627 ac->ac_2order = 0;
2628 /*
2629 * We search using buddy data only if the order of the request
2630 * is greater than equal to the sbi_s_mb_order2_reqs
2631 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2632 * We also support searching for power-of-two requests only for
2633 * requests upto maximum buddy size we have constructed.
2634 */
2635 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2636 /*
2637 * This should tell if fe_len is exactly power of 2
2638 */
2639 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2640 ac->ac_2order = array_index_nospec(i - 1,
2641 MB_NUM_ORDERS(sb));
2642 }
2643
2644 /* if stream allocation is enabled, use global goal */
2645 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2646 /* TBD: may be hot point */
2647 spin_lock(&sbi->s_md_lock);
2648 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2649 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2650 spin_unlock(&sbi->s_md_lock);
2651 }
2652
2653 /* Let's just scan groups to find more-less suitable blocks */
2654 cr = ac->ac_2order ? 0 : 1;
2655 /*
2656 * cr == 0 try to get exact allocation,
2657 * cr == 3 try to get anything
2658 */
2659repeat:
2660 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2661 ac->ac_criteria = cr;
2662 /*
2663 * searching for the right group start
2664 * from the goal value specified
2665 */
2666 group = ac->ac_g_ex.fe_group;
2667 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2668 prefetch_grp = group;
2669
2670 for (i = 0, new_cr = cr; i < ngroups; i++,
2671 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2672 int ret = 0;
2673
2674 cond_resched();
2675 if (new_cr != cr) {
2676 cr = new_cr;
2677 goto repeat;
2678 }
2679
2680 /*
2681 * Batch reads of the block allocation bitmaps
2682 * to get multiple READs in flight; limit
2683 * prefetching at cr=0/1, otherwise mballoc can
2684 * spend a lot of time loading imperfect groups
2685 */
2686 if ((prefetch_grp == group) &&
2687 (cr > 1 ||
2688 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2689 unsigned int curr_ios = prefetch_ios;
2690
2691 nr = sbi->s_mb_prefetch;
2692 if (ext4_has_feature_flex_bg(sb)) {
2693 nr = 1 << sbi->s_log_groups_per_flex;
2694 nr -= group & (nr - 1);
2695 nr = min(nr, sbi->s_mb_prefetch);
2696 }
2697 prefetch_grp = ext4_mb_prefetch(sb, group,
2698 nr, &prefetch_ios);
2699 if (prefetch_ios == curr_ios)
2700 nr = 0;
2701 }
2702
2703 /* This now checks without needing the buddy page */
2704 ret = ext4_mb_good_group_nolock(ac, group, cr);
2705 if (ret <= 0) {
2706 if (!first_err)
2707 first_err = ret;
2708 continue;
2709 }
2710
2711 err = ext4_mb_load_buddy(sb, group, &e4b);
2712 if (err)
2713 goto out;
2714
2715 ext4_lock_group(sb, group);
2716
2717 /*
2718 * We need to check again after locking the
2719 * block group
2720 */
2721 ret = ext4_mb_good_group(ac, group, cr);
2722 if (ret == 0) {
2723 ext4_unlock_group(sb, group);
2724 ext4_mb_unload_buddy(&e4b);
2725 continue;
2726 }
2727
2728 ac->ac_groups_scanned++;
2729 if (cr == 0)
2730 ext4_mb_simple_scan_group(ac, &e4b);
2731 else if (cr == 1 && sbi->s_stripe &&
2732 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2733 ext4_mb_scan_aligned(ac, &e4b);
2734 else
2735 ext4_mb_complex_scan_group(ac, &e4b);
2736
2737 ext4_unlock_group(sb, group);
2738 ext4_mb_unload_buddy(&e4b);
2739
2740 if (ac->ac_status != AC_STATUS_CONTINUE)
2741 break;
2742 }
2743 /* Processed all groups and haven't found blocks */
2744 if (sbi->s_mb_stats && i == ngroups)
2745 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2746 }
2747
2748 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2749 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2750 /*
2751 * We've been searching too long. Let's try to allocate
2752 * the best chunk we've found so far
2753 */
2754 ext4_mb_try_best_found(ac, &e4b);
2755 if (ac->ac_status != AC_STATUS_FOUND) {
2756 /*
2757 * Someone more lucky has already allocated it.
2758 * The only thing we can do is just take first
2759 * found block(s)
2760 */
2761 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2762 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2763 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2764 ac->ac_b_ex.fe_len, lost);
2765
2766 ac->ac_b_ex.fe_group = 0;
2767 ac->ac_b_ex.fe_start = 0;
2768 ac->ac_b_ex.fe_len = 0;
2769 ac->ac_status = AC_STATUS_CONTINUE;
2770 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2771 cr = 3;
2772 goto repeat;
2773 }
2774 }
2775
2776 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2777 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2778out:
2779 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2780 err = first_err;
2781
2782 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2783 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2784 ac->ac_flags, cr, err);
2785
2786 if (nr)
2787 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2788
2789 return err;
2790}
2791
2792static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2793{
2794 struct super_block *sb = pde_data(file_inode(seq->file));
2795 ext4_group_t group;
2796
2797 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2798 return NULL;
2799 group = *pos + 1;
2800 return (void *) ((unsigned long) group);
2801}
2802
2803static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2804{
2805 struct super_block *sb = pde_data(file_inode(seq->file));
2806 ext4_group_t group;
2807
2808 ++*pos;
2809 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2810 return NULL;
2811 group = *pos + 1;
2812 return (void *) ((unsigned long) group);
2813}
2814
2815static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2816{
2817 struct super_block *sb = pde_data(file_inode(seq->file));
2818 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2819 int i;
2820 int err, buddy_loaded = 0;
2821 struct ext4_buddy e4b;
2822 struct ext4_group_info *grinfo;
2823 unsigned char blocksize_bits = min_t(unsigned char,
2824 sb->s_blocksize_bits,
2825 EXT4_MAX_BLOCK_LOG_SIZE);
2826 struct sg {
2827 struct ext4_group_info info;
2828 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2829 } sg;
2830
2831 group--;
2832 if (group == 0)
2833 seq_puts(seq, "#group: free frags first ["
2834 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2835 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2836
2837 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2838 sizeof(struct ext4_group_info);
2839
2840 grinfo = ext4_get_group_info(sb, group);
2841 /* Load the group info in memory only if not already loaded. */
2842 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2843 err = ext4_mb_load_buddy(sb, group, &e4b);
2844 if (err) {
2845 seq_printf(seq, "#%-5u: I/O error\n", group);
2846 return 0;
2847 }
2848 buddy_loaded = 1;
2849 }
2850
2851 memcpy(&sg, ext4_get_group_info(sb, group), i);
2852
2853 if (buddy_loaded)
2854 ext4_mb_unload_buddy(&e4b);
2855
2856 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2857 sg.info.bb_fragments, sg.info.bb_first_free);
2858 for (i = 0; i <= 13; i++)
2859 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2860 sg.info.bb_counters[i] : 0);
2861 seq_puts(seq, " ]\n");
2862
2863 return 0;
2864}
2865
2866static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2867{
2868}
2869
2870const struct seq_operations ext4_mb_seq_groups_ops = {
2871 .start = ext4_mb_seq_groups_start,
2872 .next = ext4_mb_seq_groups_next,
2873 .stop = ext4_mb_seq_groups_stop,
2874 .show = ext4_mb_seq_groups_show,
2875};
2876
2877int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2878{
2879 struct super_block *sb = seq->private;
2880 struct ext4_sb_info *sbi = EXT4_SB(sb);
2881
2882 seq_puts(seq, "mballoc:\n");
2883 if (!sbi->s_mb_stats) {
2884 seq_puts(seq, "\tmb stats collection turned off.\n");
2885 seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2886 return 0;
2887 }
2888 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2889 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2890
2891 seq_printf(seq, "\tgroups_scanned: %u\n", atomic_read(&sbi->s_bal_groups_scanned));
2892
2893 seq_puts(seq, "\tcr0_stats:\n");
2894 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2895 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2896 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2897 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2898 atomic64_read(&sbi->s_bal_cX_failed[0]));
2899 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2900 atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2901
2902 seq_puts(seq, "\tcr1_stats:\n");
2903 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2904 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2905 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2906 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2907 atomic64_read(&sbi->s_bal_cX_failed[1]));
2908 seq_printf(seq, "\t\tbad_suggestions: %u\n",
2909 atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2910
2911 seq_puts(seq, "\tcr2_stats:\n");
2912 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2913 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2914 atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2915 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2916 atomic64_read(&sbi->s_bal_cX_failed[2]));
2917
2918 seq_puts(seq, "\tcr3_stats:\n");
2919 seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2920 seq_printf(seq, "\t\tgroups_considered: %llu\n",
2921 atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2922 seq_printf(seq, "\t\tuseless_loops: %llu\n",
2923 atomic64_read(&sbi->s_bal_cX_failed[3]));
2924 seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2925 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2926 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2927 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2928 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2929
2930 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2931 atomic_read(&sbi->s_mb_buddies_generated),
2932 ext4_get_groups_count(sb));
2933 seq_printf(seq, "\tbuddies_time_used: %llu\n",
2934 atomic64_read(&sbi->s_mb_generation_time));
2935 seq_printf(seq, "\tpreallocated: %u\n",
2936 atomic_read(&sbi->s_mb_preallocated));
2937 seq_printf(seq, "\tdiscarded: %u\n",
2938 atomic_read(&sbi->s_mb_discarded));
2939 return 0;
2940}
2941
2942static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2943__acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2944{
2945 struct super_block *sb = pde_data(file_inode(seq->file));
2946 unsigned long position;
2947
2948 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
2949 return NULL;
2950 position = *pos + 1;
2951 return (void *) ((unsigned long) position);
2952}
2953
2954static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2955{
2956 struct super_block *sb = pde_data(file_inode(seq->file));
2957 unsigned long position;
2958
2959 ++*pos;
2960 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
2961 return NULL;
2962 position = *pos + 1;
2963 return (void *) ((unsigned long) position);
2964}
2965
2966static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
2967{
2968 struct super_block *sb = pde_data(file_inode(seq->file));
2969 struct ext4_sb_info *sbi = EXT4_SB(sb);
2970 unsigned long position = ((unsigned long) v);
2971 struct ext4_group_info *grp;
2972 unsigned int count;
2973
2974 position--;
2975 if (position >= MB_NUM_ORDERS(sb)) {
2976 position -= MB_NUM_ORDERS(sb);
2977 if (position == 0)
2978 seq_puts(seq, "avg_fragment_size_lists:\n");
2979
2980 count = 0;
2981 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
2982 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
2983 bb_avg_fragment_size_node)
2984 count++;
2985 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
2986 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
2987 (unsigned int)position, count);
2988 return 0;
2989 }
2990
2991 if (position == 0) {
2992 seq_printf(seq, "optimize_scan: %d\n",
2993 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
2994 seq_puts(seq, "max_free_order_lists:\n");
2995 }
2996 count = 0;
2997 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
2998 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
2999 bb_largest_free_order_node)
3000 count++;
3001 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
3002 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3003 (unsigned int)position, count);
3004
3005 return 0;
3006}
3007
3008static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3009{
3010}
3011
3012const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3013 .start = ext4_mb_seq_structs_summary_start,
3014 .next = ext4_mb_seq_structs_summary_next,
3015 .stop = ext4_mb_seq_structs_summary_stop,
3016 .show = ext4_mb_seq_structs_summary_show,
3017};
3018
3019static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3020{
3021 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3022 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3023
3024 BUG_ON(!cachep);
3025 return cachep;
3026}
3027
3028/*
3029 * Allocate the top-level s_group_info array for the specified number
3030 * of groups
3031 */
3032int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3033{
3034 struct ext4_sb_info *sbi = EXT4_SB(sb);
3035 unsigned size;
3036 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3037
3038 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3039 EXT4_DESC_PER_BLOCK_BITS(sb);
3040 if (size <= sbi->s_group_info_size)
3041 return 0;
3042
3043 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3044 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3045 if (!new_groupinfo) {
3046 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3047 return -ENOMEM;
3048 }
3049 rcu_read_lock();
3050 old_groupinfo = rcu_dereference(sbi->s_group_info);
3051 if (old_groupinfo)
3052 memcpy(new_groupinfo, old_groupinfo,
3053 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3054 rcu_read_unlock();
3055 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3056 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3057 if (old_groupinfo)
3058 ext4_kvfree_array_rcu(old_groupinfo);
3059 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3060 sbi->s_group_info_size);
3061 return 0;
3062}
3063
3064/* Create and initialize ext4_group_info data for the given group. */
3065int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3066 struct ext4_group_desc *desc)
3067{
3068 int i;
3069 int metalen = 0;
3070 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3071 struct ext4_sb_info *sbi = EXT4_SB(sb);
3072 struct ext4_group_info **meta_group_info;
3073 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3074
3075 /*
3076 * First check if this group is the first of a reserved block.
3077 * If it's true, we have to allocate a new table of pointers
3078 * to ext4_group_info structures
3079 */
3080 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3081 metalen = sizeof(*meta_group_info) <<
3082 EXT4_DESC_PER_BLOCK_BITS(sb);
3083 meta_group_info = kmalloc(metalen, GFP_NOFS);
3084 if (meta_group_info == NULL) {
3085 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3086 "for a buddy group");
3087 goto exit_meta_group_info;
3088 }
3089 rcu_read_lock();
3090 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3091 rcu_read_unlock();
3092 }
3093
3094 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3095 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3096
3097 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3098 if (meta_group_info[i] == NULL) {
3099 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3100 goto exit_group_info;
3101 }
3102 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3103 &(meta_group_info[i]->bb_state));
3104
3105 /*
3106 * initialize bb_free to be able to skip
3107 * empty groups without initialization
3108 */
3109 if (ext4_has_group_desc_csum(sb) &&
3110 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3111 meta_group_info[i]->bb_free =
3112 ext4_free_clusters_after_init(sb, group, desc);
3113 } else {
3114 meta_group_info[i]->bb_free =
3115 ext4_free_group_clusters(sb, desc);
3116 }
3117
3118 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3119 init_rwsem(&meta_group_info[i]->alloc_sem);
3120 meta_group_info[i]->bb_free_root = RB_ROOT;
3121 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3122 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3123 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3124 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3125 meta_group_info[i]->bb_group = group;
3126
3127 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3128 return 0;
3129
3130exit_group_info:
3131 /* If a meta_group_info table has been allocated, release it now */
3132 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3133 struct ext4_group_info ***group_info;
3134
3135 rcu_read_lock();
3136 group_info = rcu_dereference(sbi->s_group_info);
3137 kfree(group_info[idx]);
3138 group_info[idx] = NULL;
3139 rcu_read_unlock();
3140 }
3141exit_meta_group_info:
3142 return -ENOMEM;
3143} /* ext4_mb_add_groupinfo */
3144
3145static int ext4_mb_init_backend(struct super_block *sb)
3146{
3147 ext4_group_t ngroups = ext4_get_groups_count(sb);
3148 ext4_group_t i;
3149 struct ext4_sb_info *sbi = EXT4_SB(sb);
3150 int err;
3151 struct ext4_group_desc *desc;
3152 struct ext4_group_info ***group_info;
3153 struct kmem_cache *cachep;
3154
3155 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3156 if (err)
3157 return err;
3158
3159 sbi->s_buddy_cache = new_inode(sb);
3160 if (sbi->s_buddy_cache == NULL) {
3161 ext4_msg(sb, KERN_ERR, "can't get new inode");
3162 goto err_freesgi;
3163 }
3164 /* To avoid potentially colliding with an valid on-disk inode number,
3165 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3166 * not in the inode hash, so it should never be found by iget(), but
3167 * this will avoid confusion if it ever shows up during debugging. */
3168 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3169 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3170 for (i = 0; i < ngroups; i++) {
3171 cond_resched();
3172 desc = ext4_get_group_desc(sb, i, NULL);
3173 if (desc == NULL) {
3174 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3175 goto err_freebuddy;
3176 }
3177 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3178 goto err_freebuddy;
3179 }
3180
3181 if (ext4_has_feature_flex_bg(sb)) {
3182 /* a single flex group is supposed to be read by a single IO.
3183 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3184 * unsigned integer, so the maximum shift is 32.
3185 */
3186 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3187 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3188 goto err_freebuddy;
3189 }
3190 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3191 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3192 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3193 } else {
3194 sbi->s_mb_prefetch = 32;
3195 }
3196 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3197 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3198 /* now many real IOs to prefetch within a single allocation at cr=0
3199 * given cr=0 is an CPU-related optimization we shouldn't try to
3200 * load too many groups, at some point we should start to use what
3201 * we've got in memory.
3202 * with an average random access time 5ms, it'd take a second to get
3203 * 200 groups (* N with flex_bg), so let's make this limit 4
3204 */
3205 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3206 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3207 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3208
3209 return 0;
3210
3211err_freebuddy:
3212 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3213 while (i-- > 0)
3214 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3215 i = sbi->s_group_info_size;
3216 rcu_read_lock();
3217 group_info = rcu_dereference(sbi->s_group_info);
3218 while (i-- > 0)
3219 kfree(group_info[i]);
3220 rcu_read_unlock();
3221 iput(sbi->s_buddy_cache);
3222err_freesgi:
3223 rcu_read_lock();
3224 kvfree(rcu_dereference(sbi->s_group_info));
3225 rcu_read_unlock();
3226 return -ENOMEM;
3227}
3228
3229static void ext4_groupinfo_destroy_slabs(void)
3230{
3231 int i;
3232
3233 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3234 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3235 ext4_groupinfo_caches[i] = NULL;
3236 }
3237}
3238
3239static int ext4_groupinfo_create_slab(size_t size)
3240{
3241 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3242 int slab_size;
3243 int blocksize_bits = order_base_2(size);
3244 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3245 struct kmem_cache *cachep;
3246
3247 if (cache_index >= NR_GRPINFO_CACHES)
3248 return -EINVAL;
3249
3250 if (unlikely(cache_index < 0))
3251 cache_index = 0;
3252
3253 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3254 if (ext4_groupinfo_caches[cache_index]) {
3255 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3256 return 0; /* Already created */
3257 }
3258
3259 slab_size = offsetof(struct ext4_group_info,
3260 bb_counters[blocksize_bits + 2]);
3261
3262 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3263 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3264 NULL);
3265
3266 ext4_groupinfo_caches[cache_index] = cachep;
3267
3268 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3269 if (!cachep) {
3270 printk(KERN_EMERG
3271 "EXT4-fs: no memory for groupinfo slab cache\n");
3272 return -ENOMEM;
3273 }
3274
3275 return 0;
3276}
3277
3278static void ext4_discard_work(struct work_struct *work)
3279{
3280 struct ext4_sb_info *sbi = container_of(work,
3281 struct ext4_sb_info, s_discard_work);
3282 struct super_block *sb = sbi->s_sb;
3283 struct ext4_free_data *fd, *nfd;
3284 struct ext4_buddy e4b;
3285 struct list_head discard_list;
3286 ext4_group_t grp, load_grp;
3287 int err = 0;
3288
3289 INIT_LIST_HEAD(&discard_list);
3290 spin_lock(&sbi->s_md_lock);
3291 list_splice_init(&sbi->s_discard_list, &discard_list);
3292 spin_unlock(&sbi->s_md_lock);
3293
3294 load_grp = UINT_MAX;
3295 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3296 /*
3297 * If filesystem is umounting or no memory or suffering
3298 * from no space, give up the discard
3299 */
3300 if ((sb->s_flags & SB_ACTIVE) && !err &&
3301 !atomic_read(&sbi->s_retry_alloc_pending)) {
3302 grp = fd->efd_group;
3303 if (grp != load_grp) {
3304 if (load_grp != UINT_MAX)
3305 ext4_mb_unload_buddy(&e4b);
3306
3307 err = ext4_mb_load_buddy(sb, grp, &e4b);
3308 if (err) {
3309 kmem_cache_free(ext4_free_data_cachep, fd);
3310 load_grp = UINT_MAX;
3311 continue;
3312 } else {
3313 load_grp = grp;
3314 }
3315 }
3316
3317 ext4_lock_group(sb, grp);
3318 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3319 fd->efd_start_cluster + fd->efd_count - 1, 1);
3320 ext4_unlock_group(sb, grp);
3321 }
3322 kmem_cache_free(ext4_free_data_cachep, fd);
3323 }
3324
3325 if (load_grp != UINT_MAX)
3326 ext4_mb_unload_buddy(&e4b);
3327}
3328
3329int ext4_mb_init(struct super_block *sb)
3330{
3331 struct ext4_sb_info *sbi = EXT4_SB(sb);
3332 unsigned i, j;
3333 unsigned offset, offset_incr;
3334 unsigned max;
3335 int ret;
3336
3337 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3338
3339 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3340 if (sbi->s_mb_offsets == NULL) {
3341 ret = -ENOMEM;
3342 goto out;
3343 }
3344
3345 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3346 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3347 if (sbi->s_mb_maxs == NULL) {
3348 ret = -ENOMEM;
3349 goto out;
3350 }
3351
3352 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3353 if (ret < 0)
3354 goto out;
3355
3356 /* order 0 is regular bitmap */
3357 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3358 sbi->s_mb_offsets[0] = 0;
3359
3360 i = 1;
3361 offset = 0;
3362 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3363 max = sb->s_blocksize << 2;
3364 do {
3365 sbi->s_mb_offsets[i] = offset;
3366 sbi->s_mb_maxs[i] = max;
3367 offset += offset_incr;
3368 offset_incr = offset_incr >> 1;
3369 max = max >> 1;
3370 i++;
3371 } while (i < MB_NUM_ORDERS(sb));
3372
3373 sbi->s_mb_avg_fragment_size =
3374 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3375 GFP_KERNEL);
3376 if (!sbi->s_mb_avg_fragment_size) {
3377 ret = -ENOMEM;
3378 goto out;
3379 }
3380 sbi->s_mb_avg_fragment_size_locks =
3381 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3382 GFP_KERNEL);
3383 if (!sbi->s_mb_avg_fragment_size_locks) {
3384 ret = -ENOMEM;
3385 goto out;
3386 }
3387 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3388 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3389 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3390 }
3391 sbi->s_mb_largest_free_orders =
3392 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3393 GFP_KERNEL);
3394 if (!sbi->s_mb_largest_free_orders) {
3395 ret = -ENOMEM;
3396 goto out;
3397 }
3398 sbi->s_mb_largest_free_orders_locks =
3399 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3400 GFP_KERNEL);
3401 if (!sbi->s_mb_largest_free_orders_locks) {
3402 ret = -ENOMEM;
3403 goto out;
3404 }
3405 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3406 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3407 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3408 }
3409
3410 spin_lock_init(&sbi->s_md_lock);
3411 sbi->s_mb_free_pending = 0;
3412 INIT_LIST_HEAD(&sbi->s_freed_data_list);
3413 INIT_LIST_HEAD(&sbi->s_discard_list);
3414 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3415 atomic_set(&sbi->s_retry_alloc_pending, 0);
3416
3417 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3418 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3419 sbi->s_mb_stats = MB_DEFAULT_STATS;
3420 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3421 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3422 sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3423 /*
3424 * The default group preallocation is 512, which for 4k block
3425 * sizes translates to 2 megabytes. However for bigalloc file
3426 * systems, this is probably too big (i.e, if the cluster size
3427 * is 1 megabyte, then group preallocation size becomes half a
3428 * gigabyte!). As a default, we will keep a two megabyte
3429 * group pralloc size for cluster sizes up to 64k, and after
3430 * that, we will force a minimum group preallocation size of
3431 * 32 clusters. This translates to 8 megs when the cluster
3432 * size is 256k, and 32 megs when the cluster size is 1 meg,
3433 * which seems reasonable as a default.
3434 */
3435 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3436 sbi->s_cluster_bits, 32);
3437 /*
3438 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3439 * to the lowest multiple of s_stripe which is bigger than
3440 * the s_mb_group_prealloc as determined above. We want
3441 * the preallocation size to be an exact multiple of the
3442 * RAID stripe size so that preallocations don't fragment
3443 * the stripes.
3444 */
3445 if (sbi->s_stripe > 1) {
3446 sbi->s_mb_group_prealloc = roundup(
3447 sbi->s_mb_group_prealloc, sbi->s_stripe);
3448 }
3449
3450 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3451 if (sbi->s_locality_groups == NULL) {
3452 ret = -ENOMEM;
3453 goto out;
3454 }
3455 for_each_possible_cpu(i) {
3456 struct ext4_locality_group *lg;
3457 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3458 mutex_init(&lg->lg_mutex);
3459 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3460 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3461 spin_lock_init(&lg->lg_prealloc_lock);
3462 }
3463
3464 if (bdev_nonrot(sb->s_bdev))
3465 sbi->s_mb_max_linear_groups = 0;
3466 else
3467 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3468 /* init file for buddy data */
3469 ret = ext4_mb_init_backend(sb);
3470 if (ret != 0)
3471 goto out_free_locality_groups;
3472
3473 return 0;
3474
3475out_free_locality_groups:
3476 free_percpu(sbi->s_locality_groups);
3477 sbi->s_locality_groups = NULL;
3478out:
3479 kfree(sbi->s_mb_avg_fragment_size);
3480 kfree(sbi->s_mb_avg_fragment_size_locks);
3481 kfree(sbi->s_mb_largest_free_orders);
3482 kfree(sbi->s_mb_largest_free_orders_locks);
3483 kfree(sbi->s_mb_offsets);
3484 sbi->s_mb_offsets = NULL;
3485 kfree(sbi->s_mb_maxs);
3486 sbi->s_mb_maxs = NULL;
3487 return ret;
3488}
3489
3490/* need to called with the ext4 group lock held */
3491static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3492{
3493 struct ext4_prealloc_space *pa;
3494 struct list_head *cur, *tmp;
3495 int count = 0;
3496
3497 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3498 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3499 list_del(&pa->pa_group_list);
3500 count++;
3501 kmem_cache_free(ext4_pspace_cachep, pa);
3502 }
3503 return count;
3504}
3505
3506int ext4_mb_release(struct super_block *sb)
3507{
3508 ext4_group_t ngroups = ext4_get_groups_count(sb);
3509 ext4_group_t i;
3510 int num_meta_group_infos;
3511 struct ext4_group_info *grinfo, ***group_info;
3512 struct ext4_sb_info *sbi = EXT4_SB(sb);
3513 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3514 int count;
3515
3516 if (test_opt(sb, DISCARD)) {
3517 /*
3518 * wait the discard work to drain all of ext4_free_data
3519 */
3520 flush_work(&sbi->s_discard_work);
3521 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3522 }
3523
3524 if (sbi->s_group_info) {
3525 for (i = 0; i < ngroups; i++) {
3526 cond_resched();
3527 grinfo = ext4_get_group_info(sb, i);
3528 mb_group_bb_bitmap_free(grinfo);
3529 ext4_lock_group(sb, i);
3530 count = ext4_mb_cleanup_pa(grinfo);
3531 if (count)
3532 mb_debug(sb, "mballoc: %d PAs left\n",
3533 count);
3534 ext4_unlock_group(sb, i);
3535 kmem_cache_free(cachep, grinfo);
3536 }
3537 num_meta_group_infos = (ngroups +
3538 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3539 EXT4_DESC_PER_BLOCK_BITS(sb);
3540 rcu_read_lock();
3541 group_info = rcu_dereference(sbi->s_group_info);
3542 for (i = 0; i < num_meta_group_infos; i++)
3543 kfree(group_info[i]);
3544 kvfree(group_info);
3545 rcu_read_unlock();
3546 }
3547 kfree(sbi->s_mb_avg_fragment_size);
3548 kfree(sbi->s_mb_avg_fragment_size_locks);
3549 kfree(sbi->s_mb_largest_free_orders);
3550 kfree(sbi->s_mb_largest_free_orders_locks);
3551 kfree(sbi->s_mb_offsets);
3552 kfree(sbi->s_mb_maxs);
3553 iput(sbi->s_buddy_cache);
3554 if (sbi->s_mb_stats) {
3555 ext4_msg(sb, KERN_INFO,
3556 "mballoc: %u blocks %u reqs (%u success)",
3557 atomic_read(&sbi->s_bal_allocated),
3558 atomic_read(&sbi->s_bal_reqs),
3559 atomic_read(&sbi->s_bal_success));
3560 ext4_msg(sb, KERN_INFO,
3561 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3562 "%u 2^N hits, %u breaks, %u lost",
3563 atomic_read(&sbi->s_bal_ex_scanned),
3564 atomic_read(&sbi->s_bal_groups_scanned),
3565 atomic_read(&sbi->s_bal_goals),
3566 atomic_read(&sbi->s_bal_2orders),
3567 atomic_read(&sbi->s_bal_breaks),
3568 atomic_read(&sbi->s_mb_lost_chunks));
3569 ext4_msg(sb, KERN_INFO,
3570 "mballoc: %u generated and it took %llu",
3571 atomic_read(&sbi->s_mb_buddies_generated),
3572 atomic64_read(&sbi->s_mb_generation_time));
3573 ext4_msg(sb, KERN_INFO,
3574 "mballoc: %u preallocated, %u discarded",
3575 atomic_read(&sbi->s_mb_preallocated),
3576 atomic_read(&sbi->s_mb_discarded));
3577 }
3578
3579 free_percpu(sbi->s_locality_groups);
3580
3581 return 0;
3582}
3583
3584static inline int ext4_issue_discard(struct super_block *sb,
3585 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3586 struct bio **biop)
3587{
3588 ext4_fsblk_t discard_block;
3589
3590 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3591 ext4_group_first_block_no(sb, block_group));
3592 count = EXT4_C2B(EXT4_SB(sb), count);
3593 trace_ext4_discard_blocks(sb,
3594 (unsigned long long) discard_block, count);
3595 if (biop) {
3596 return __blkdev_issue_discard(sb->s_bdev,
3597 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3598 (sector_t)count << (sb->s_blocksize_bits - 9),
3599 GFP_NOFS, biop);
3600 } else
3601 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3602}
3603
3604static void ext4_free_data_in_buddy(struct super_block *sb,
3605 struct ext4_free_data *entry)
3606{
3607 struct ext4_buddy e4b;
3608 struct ext4_group_info *db;
3609 int err, count = 0, count2 = 0;
3610
3611 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3612 entry->efd_count, entry->efd_group, entry);
3613
3614 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3615 /* we expect to find existing buddy because it's pinned */
3616 BUG_ON(err != 0);
3617
3618 spin_lock(&EXT4_SB(sb)->s_md_lock);
3619 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3620 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3621
3622 db = e4b.bd_info;
3623 /* there are blocks to put in buddy to make them really free */
3624 count += entry->efd_count;
3625 count2++;
3626 ext4_lock_group(sb, entry->efd_group);
3627 /* Take it out of per group rb tree */
3628 rb_erase(&entry->efd_node, &(db->bb_free_root));
3629 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3630
3631 /*
3632 * Clear the trimmed flag for the group so that the next
3633 * ext4_trim_fs can trim it.
3634 * If the volume is mounted with -o discard, online discard
3635 * is supported and the free blocks will be trimmed online.
3636 */
3637 if (!test_opt(sb, DISCARD))
3638 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3639
3640 if (!db->bb_free_root.rb_node) {
3641 /* No more items in the per group rb tree
3642 * balance refcounts from ext4_mb_free_metadata()
3643 */
3644 put_page(e4b.bd_buddy_page);
3645 put_page(e4b.bd_bitmap_page);
3646 }
3647 ext4_unlock_group(sb, entry->efd_group);
3648 ext4_mb_unload_buddy(&e4b);
3649
3650 mb_debug(sb, "freed %d blocks in %d structures\n", count,
3651 count2);
3652}
3653
3654/*
3655 * This function is called by the jbd2 layer once the commit has finished,
3656 * so we know we can free the blocks that were released with that commit.
3657 */
3658void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3659{
3660 struct ext4_sb_info *sbi = EXT4_SB(sb);
3661 struct ext4_free_data *entry, *tmp;
3662 struct list_head freed_data_list;
3663 struct list_head *cut_pos = NULL;
3664 bool wake;
3665
3666 INIT_LIST_HEAD(&freed_data_list);
3667
3668 spin_lock(&sbi->s_md_lock);
3669 list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3670 if (entry->efd_tid != commit_tid)
3671 break;
3672 cut_pos = &entry->efd_list;
3673 }
3674 if (cut_pos)
3675 list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3676 cut_pos);
3677 spin_unlock(&sbi->s_md_lock);
3678
3679 list_for_each_entry(entry, &freed_data_list, efd_list)
3680 ext4_free_data_in_buddy(sb, entry);
3681
3682 if (test_opt(sb, DISCARD)) {
3683 spin_lock(&sbi->s_md_lock);
3684 wake = list_empty(&sbi->s_discard_list);
3685 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3686 spin_unlock(&sbi->s_md_lock);
3687 if (wake)
3688 queue_work(system_unbound_wq, &sbi->s_discard_work);
3689 } else {
3690 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3691 kmem_cache_free(ext4_free_data_cachep, entry);
3692 }
3693}
3694
3695int __init ext4_init_mballoc(void)
3696{
3697 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3698 SLAB_RECLAIM_ACCOUNT);
3699 if (ext4_pspace_cachep == NULL)
3700 goto out;
3701
3702 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3703 SLAB_RECLAIM_ACCOUNT);
3704 if (ext4_ac_cachep == NULL)
3705 goto out_pa_free;
3706
3707 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3708 SLAB_RECLAIM_ACCOUNT);
3709 if (ext4_free_data_cachep == NULL)
3710 goto out_ac_free;
3711
3712 return 0;
3713
3714out_ac_free:
3715 kmem_cache_destroy(ext4_ac_cachep);
3716out_pa_free:
3717 kmem_cache_destroy(ext4_pspace_cachep);
3718out:
3719 return -ENOMEM;
3720}
3721
3722void ext4_exit_mballoc(void)
3723{
3724 /*
3725 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3726 * before destroying the slab cache.
3727 */
3728 rcu_barrier();
3729 kmem_cache_destroy(ext4_pspace_cachep);
3730 kmem_cache_destroy(ext4_ac_cachep);
3731 kmem_cache_destroy(ext4_free_data_cachep);
3732 ext4_groupinfo_destroy_slabs();
3733}
3734
3735
3736/*
3737 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3738 * Returns 0 if success or error code
3739 */
3740static noinline_for_stack int
3741ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3742 handle_t *handle, unsigned int reserv_clstrs)
3743{
3744 struct buffer_head *bitmap_bh = NULL;
3745 struct ext4_group_desc *gdp;
3746 struct buffer_head *gdp_bh;
3747 struct ext4_sb_info *sbi;
3748 struct super_block *sb;
3749 ext4_fsblk_t block;
3750 int err, len;
3751
3752 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3753 BUG_ON(ac->ac_b_ex.fe_len <= 0);
3754
3755 sb = ac->ac_sb;
3756 sbi = EXT4_SB(sb);
3757
3758 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3759 if (IS_ERR(bitmap_bh)) {
3760 err = PTR_ERR(bitmap_bh);
3761 bitmap_bh = NULL;
3762 goto out_err;
3763 }
3764
3765 BUFFER_TRACE(bitmap_bh, "getting write access");
3766 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3767 EXT4_JTR_NONE);
3768 if (err)
3769 goto out_err;
3770
3771 err = -EIO;
3772 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3773 if (!gdp)
3774 goto out_err;
3775
3776 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3777 ext4_free_group_clusters(sb, gdp));
3778
3779 BUFFER_TRACE(gdp_bh, "get_write_access");
3780 err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3781 if (err)
3782 goto out_err;
3783
3784 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3785
3786 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3787 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3788 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3789 "fs metadata", block, block+len);
3790 /* File system mounted not to panic on error
3791 * Fix the bitmap and return EFSCORRUPTED
3792 * We leak some of the blocks here.
3793 */
3794 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3795 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3796 ac->ac_b_ex.fe_len);
3797 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3798 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3799 if (!err)
3800 err = -EFSCORRUPTED;
3801 goto out_err;
3802 }
3803
3804 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3805#ifdef AGGRESSIVE_CHECK
3806 {
3807 int i;
3808 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3809 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3810 bitmap_bh->b_data));
3811 }
3812 }
3813#endif
3814 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3815 ac->ac_b_ex.fe_len);
3816 if (ext4_has_group_desc_csum(sb) &&
3817 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3818 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3819 ext4_free_group_clusters_set(sb, gdp,
3820 ext4_free_clusters_after_init(sb,
3821 ac->ac_b_ex.fe_group, gdp));
3822 }
3823 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3824 ext4_free_group_clusters_set(sb, gdp, len);
3825 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3826 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3827
3828 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3829 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3830 /*
3831 * Now reduce the dirty block count also. Should not go negative
3832 */
3833 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3834 /* release all the reserved blocks if non delalloc */
3835 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3836 reserv_clstrs);
3837
3838 if (sbi->s_log_groups_per_flex) {
3839 ext4_group_t flex_group = ext4_flex_group(sbi,
3840 ac->ac_b_ex.fe_group);
3841 atomic64_sub(ac->ac_b_ex.fe_len,
3842 &sbi_array_rcu_deref(sbi, s_flex_groups,
3843 flex_group)->free_clusters);
3844 }
3845
3846 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3847 if (err)
3848 goto out_err;
3849 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3850
3851out_err:
3852 brelse(bitmap_bh);
3853 return err;
3854}
3855
3856/*
3857 * Idempotent helper for Ext4 fast commit replay path to set the state of
3858 * blocks in bitmaps and update counters.
3859 */
3860void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3861 int len, int state)
3862{
3863 struct buffer_head *bitmap_bh = NULL;
3864 struct ext4_group_desc *gdp;
3865 struct buffer_head *gdp_bh;
3866 struct ext4_sb_info *sbi = EXT4_SB(sb);
3867 ext4_group_t group;
3868 ext4_grpblk_t blkoff;
3869 int i, err;
3870 int already;
3871 unsigned int clen, clen_changed, thisgrp_len;
3872
3873 while (len > 0) {
3874 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3875
3876 /*
3877 * Check to see if we are freeing blocks across a group
3878 * boundary.
3879 * In case of flex_bg, this can happen that (block, len) may
3880 * span across more than one group. In that case we need to
3881 * get the corresponding group metadata to work with.
3882 * For this we have goto again loop.
3883 */
3884 thisgrp_len = min_t(unsigned int, (unsigned int)len,
3885 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3886 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3887
3888 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3889 ext4_error(sb, "Marking blocks in system zone - "
3890 "Block = %llu, len = %u",
3891 block, thisgrp_len);
3892 bitmap_bh = NULL;
3893 break;
3894 }
3895
3896 bitmap_bh = ext4_read_block_bitmap(sb, group);
3897 if (IS_ERR(bitmap_bh)) {
3898 err = PTR_ERR(bitmap_bh);
3899 bitmap_bh = NULL;
3900 break;
3901 }
3902
3903 err = -EIO;
3904 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3905 if (!gdp)
3906 break;
3907
3908 ext4_lock_group(sb, group);
3909 already = 0;
3910 for (i = 0; i < clen; i++)
3911 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3912 !state)
3913 already++;
3914
3915 clen_changed = clen - already;
3916 if (state)
3917 mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3918 else
3919 mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3920 if (ext4_has_group_desc_csum(sb) &&
3921 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3922 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3923 ext4_free_group_clusters_set(sb, gdp,
3924 ext4_free_clusters_after_init(sb, group, gdp));
3925 }
3926 if (state)
3927 clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3928 else
3929 clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3930
3931 ext4_free_group_clusters_set(sb, gdp, clen);
3932 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3933 ext4_group_desc_csum_set(sb, group, gdp);
3934
3935 ext4_unlock_group(sb, group);
3936
3937 if (sbi->s_log_groups_per_flex) {
3938 ext4_group_t flex_group = ext4_flex_group(sbi, group);
3939 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3940 s_flex_groups, flex_group);
3941
3942 if (state)
3943 atomic64_sub(clen_changed, &fg->free_clusters);
3944 else
3945 atomic64_add(clen_changed, &fg->free_clusters);
3946
3947 }
3948
3949 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3950 if (err)
3951 break;
3952 sync_dirty_buffer(bitmap_bh);
3953 err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3954 sync_dirty_buffer(gdp_bh);
3955 if (err)
3956 break;
3957
3958 block += thisgrp_len;
3959 len -= thisgrp_len;
3960 brelse(bitmap_bh);
3961 BUG_ON(len < 0);
3962 }
3963
3964 if (err)
3965 brelse(bitmap_bh);
3966}
3967
3968/*
3969 * here we normalize request for locality group
3970 * Group request are normalized to s_mb_group_prealloc, which goes to
3971 * s_strip if we set the same via mount option.
3972 * s_mb_group_prealloc can be configured via
3973 * /sys/fs/ext4/<partition>/mb_group_prealloc
3974 *
3975 * XXX: should we try to preallocate more than the group has now?
3976 */
3977static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3978{
3979 struct super_block *sb = ac->ac_sb;
3980 struct ext4_locality_group *lg = ac->ac_lg;
3981
3982 BUG_ON(lg == NULL);
3983 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3984 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3985}
3986
3987/*
3988 * Normalization means making request better in terms of
3989 * size and alignment
3990 */
3991static noinline_for_stack void
3992ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3993 struct ext4_allocation_request *ar)
3994{
3995 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3996 int bsbits, max;
3997 ext4_lblk_t end;
3998 loff_t size, start_off;
3999 loff_t orig_size __maybe_unused;
4000 ext4_lblk_t start;
4001 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4002 struct ext4_prealloc_space *pa;
4003
4004 /* do normalize only data requests, metadata requests
4005 do not need preallocation */
4006 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4007 return;
4008
4009 /* sometime caller may want exact blocks */
4010 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4011 return;
4012
4013 /* caller may indicate that preallocation isn't
4014 * required (it's a tail, for example) */
4015 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4016 return;
4017
4018 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4019 ext4_mb_normalize_group_request(ac);
4020 return ;
4021 }
4022
4023 bsbits = ac->ac_sb->s_blocksize_bits;
4024
4025 /* first, let's learn actual file size
4026 * given current request is allocated */
4027 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4028 size = size << bsbits;
4029 if (size < i_size_read(ac->ac_inode))
4030 size = i_size_read(ac->ac_inode);
4031 orig_size = size;
4032
4033 /* max size of free chunks */
4034 max = 2 << bsbits;
4035
4036#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4037 (req <= (size) || max <= (chunk_size))
4038
4039 /* first, try to predict filesize */
4040 /* XXX: should this table be tunable? */
4041 start_off = 0;
4042 if (size <= 16 * 1024) {
4043 size = 16 * 1024;
4044 } else if (size <= 32 * 1024) {
4045 size = 32 * 1024;
4046 } else if (size <= 64 * 1024) {
4047 size = 64 * 1024;
4048 } else if (size <= 128 * 1024) {
4049 size = 128 * 1024;
4050 } else if (size <= 256 * 1024) {
4051 size = 256 * 1024;
4052 } else if (size <= 512 * 1024) {
4053 size = 512 * 1024;
4054 } else if (size <= 1024 * 1024) {
4055 size = 1024 * 1024;
4056 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4057 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4058 (21 - bsbits)) << 21;
4059 size = 2 * 1024 * 1024;
4060 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4061 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4062 (22 - bsbits)) << 22;
4063 size = 4 * 1024 * 1024;
4064 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4065 (8<<20)>>bsbits, max, 8 * 1024)) {
4066 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4067 (23 - bsbits)) << 23;
4068 size = 8 * 1024 * 1024;
4069 } else {
4070 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4071 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4072 ac->ac_o_ex.fe_len) << bsbits;
4073 }
4074 size = size >> bsbits;
4075 start = start_off >> bsbits;
4076
4077 /*
4078 * For tiny groups (smaller than 8MB) the chosen allocation
4079 * alignment may be larger than group size. Make sure the
4080 * alignment does not move allocation to a different group which
4081 * makes mballoc fail assertions later.
4082 */
4083 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4084 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4085
4086 /* don't cover already allocated blocks in selected range */
4087 if (ar->pleft && start <= ar->lleft) {
4088 size -= ar->lleft + 1 - start;
4089 start = ar->lleft + 1;
4090 }
4091 if (ar->pright && start + size - 1 >= ar->lright)
4092 size -= start + size - ar->lright;
4093
4094 /*
4095 * Trim allocation request for filesystems with artificially small
4096 * groups.
4097 */
4098 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4099 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4100
4101 end = start + size;
4102
4103 /* check we don't cross already preallocated blocks */
4104 rcu_read_lock();
4105 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4106 ext4_lblk_t pa_end;
4107
4108 if (pa->pa_deleted)
4109 continue;
4110 spin_lock(&pa->pa_lock);
4111 if (pa->pa_deleted) {
4112 spin_unlock(&pa->pa_lock);
4113 continue;
4114 }
4115
4116 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4117 pa->pa_len);
4118
4119 /* PA must not overlap original request */
4120 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4121 ac->ac_o_ex.fe_logical < pa->pa_lstart));
4122
4123 /* skip PAs this normalized request doesn't overlap with */
4124 if (pa->pa_lstart >= end || pa_end <= start) {
4125 spin_unlock(&pa->pa_lock);
4126 continue;
4127 }
4128 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4129
4130 /* adjust start or end to be adjacent to this pa */
4131 if (pa_end <= ac->ac_o_ex.fe_logical) {
4132 BUG_ON(pa_end < start);
4133 start = pa_end;
4134 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4135 BUG_ON(pa->pa_lstart > end);
4136 end = pa->pa_lstart;
4137 }
4138 spin_unlock(&pa->pa_lock);
4139 }
4140 rcu_read_unlock();
4141 size = end - start;
4142
4143 /* XXX: extra loop to check we really don't overlap preallocations */
4144 rcu_read_lock();
4145 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4146 ext4_lblk_t pa_end;
4147
4148 spin_lock(&pa->pa_lock);
4149 if (pa->pa_deleted == 0) {
4150 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4151 pa->pa_len);
4152 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4153 }
4154 spin_unlock(&pa->pa_lock);
4155 }
4156 rcu_read_unlock();
4157
4158 /*
4159 * In this function "start" and "size" are normalized for better
4160 * alignment and length such that we could preallocate more blocks.
4161 * This normalization is done such that original request of
4162 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4163 * "size" boundaries.
4164 * (Note fe_len can be relaxed since FS block allocation API does not
4165 * provide gurantee on number of contiguous blocks allocation since that
4166 * depends upon free space left, etc).
4167 * In case of inode pa, later we use the allocated blocks
4168 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4169 * range of goal/best blocks [start, size] to put it at the
4170 * ac_o_ex.fe_logical extent of this inode.
4171 * (See ext4_mb_use_inode_pa() for more details)
4172 */
4173 if (start + size <= ac->ac_o_ex.fe_logical ||
4174 start > ac->ac_o_ex.fe_logical) {
4175 ext4_msg(ac->ac_sb, KERN_ERR,
4176 "start %lu, size %lu, fe_logical %lu",
4177 (unsigned long) start, (unsigned long) size,
4178 (unsigned long) ac->ac_o_ex.fe_logical);
4179 BUG();
4180 }
4181 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4182
4183 /* now prepare goal request */
4184
4185 /* XXX: is it better to align blocks WRT to logical
4186 * placement or satisfy big request as is */
4187 ac->ac_g_ex.fe_logical = start;
4188 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4189
4190 /* define goal start in order to merge */
4191 if (ar->pright && (ar->lright == (start + size))) {
4192 /* merge to the right */
4193 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4194 &ac->ac_f_ex.fe_group,
4195 &ac->ac_f_ex.fe_start);
4196 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4197 }
4198 if (ar->pleft && (ar->lleft + 1 == start)) {
4199 /* merge to the left */
4200 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4201 &ac->ac_f_ex.fe_group,
4202 &ac->ac_f_ex.fe_start);
4203 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4204 }
4205
4206 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4207 orig_size, start);
4208}
4209
4210static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4211{
4212 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4213
4214 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4215 atomic_inc(&sbi->s_bal_reqs);
4216 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4217 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4218 atomic_inc(&sbi->s_bal_success);
4219 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4220 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4221 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4222 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4223 atomic_inc(&sbi->s_bal_goals);
4224 if (ac->ac_found > sbi->s_mb_max_to_scan)
4225 atomic_inc(&sbi->s_bal_breaks);
4226 }
4227
4228 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4229 trace_ext4_mballoc_alloc(ac);
4230 else
4231 trace_ext4_mballoc_prealloc(ac);
4232}
4233
4234/*
4235 * Called on failure; free up any blocks from the inode PA for this
4236 * context. We don't need this for MB_GROUP_PA because we only change
4237 * pa_free in ext4_mb_release_context(), but on failure, we've already
4238 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4239 */
4240static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4241{
4242 struct ext4_prealloc_space *pa = ac->ac_pa;
4243 struct ext4_buddy e4b;
4244 int err;
4245
4246 if (pa == NULL) {
4247 if (ac->ac_f_ex.fe_len == 0)
4248 return;
4249 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4250 if (err) {
4251 /*
4252 * This should never happen since we pin the
4253 * pages in the ext4_allocation_context so
4254 * ext4_mb_load_buddy() should never fail.
4255 */
4256 WARN(1, "mb_load_buddy failed (%d)", err);
4257 return;
4258 }
4259 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4260 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4261 ac->ac_f_ex.fe_len);
4262 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4263 ext4_mb_unload_buddy(&e4b);
4264 return;
4265 }
4266 if (pa->pa_type == MB_INODE_PA)
4267 pa->pa_free += ac->ac_b_ex.fe_len;
4268}
4269
4270/*
4271 * use blocks preallocated to inode
4272 */
4273static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4274 struct ext4_prealloc_space *pa)
4275{
4276 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4277 ext4_fsblk_t start;
4278 ext4_fsblk_t end;
4279 int len;
4280
4281 /* found preallocated blocks, use them */
4282 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4283 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4284 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4285 len = EXT4_NUM_B2C(sbi, end - start);
4286 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4287 &ac->ac_b_ex.fe_start);
4288 ac->ac_b_ex.fe_len = len;
4289 ac->ac_status = AC_STATUS_FOUND;
4290 ac->ac_pa = pa;
4291
4292 BUG_ON(start < pa->pa_pstart);
4293 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4294 BUG_ON(pa->pa_free < len);
4295 pa->pa_free -= len;
4296
4297 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4298}
4299
4300/*
4301 * use blocks preallocated to locality group
4302 */
4303static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4304 struct ext4_prealloc_space *pa)
4305{
4306 unsigned int len = ac->ac_o_ex.fe_len;
4307
4308 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4309 &ac->ac_b_ex.fe_group,
4310 &ac->ac_b_ex.fe_start);
4311 ac->ac_b_ex.fe_len = len;
4312 ac->ac_status = AC_STATUS_FOUND;
4313 ac->ac_pa = pa;
4314
4315 /* we don't correct pa_pstart or pa_plen here to avoid
4316 * possible race when the group is being loaded concurrently
4317 * instead we correct pa later, after blocks are marked
4318 * in on-disk bitmap -- see ext4_mb_release_context()
4319 * Other CPUs are prevented from allocating from this pa by lg_mutex
4320 */
4321 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4322 pa->pa_lstart-len, len, pa);
4323}
4324
4325/*
4326 * Return the prealloc space that have minimal distance
4327 * from the goal block. @cpa is the prealloc
4328 * space that is having currently known minimal distance
4329 * from the goal block.
4330 */
4331static struct ext4_prealloc_space *
4332ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4333 struct ext4_prealloc_space *pa,
4334 struct ext4_prealloc_space *cpa)
4335{
4336 ext4_fsblk_t cur_distance, new_distance;
4337
4338 if (cpa == NULL) {
4339 atomic_inc(&pa->pa_count);
4340 return pa;
4341 }
4342 cur_distance = abs(goal_block - cpa->pa_pstart);
4343 new_distance = abs(goal_block - pa->pa_pstart);
4344
4345 if (cur_distance <= new_distance)
4346 return cpa;
4347
4348 /* drop the previous reference */
4349 atomic_dec(&cpa->pa_count);
4350 atomic_inc(&pa->pa_count);
4351 return pa;
4352}
4353
4354/*
4355 * search goal blocks in preallocated space
4356 */
4357static noinline_for_stack bool
4358ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4359{
4360 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4361 int order, i;
4362 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4363 struct ext4_locality_group *lg;
4364 struct ext4_prealloc_space *pa, *cpa = NULL;
4365 ext4_fsblk_t goal_block;
4366
4367 /* only data can be preallocated */
4368 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4369 return false;
4370
4371 /* first, try per-file preallocation */
4372 rcu_read_lock();
4373 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4374
4375 /* all fields in this condition don't change,
4376 * so we can skip locking for them */
4377 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4378 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4379 EXT4_C2B(sbi, pa->pa_len)))
4380 continue;
4381
4382 /* non-extent files can't have physical blocks past 2^32 */
4383 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4384 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4385 EXT4_MAX_BLOCK_FILE_PHYS))
4386 continue;
4387
4388 /* found preallocated blocks, use them */
4389 spin_lock(&pa->pa_lock);
4390 if (pa->pa_deleted == 0 && pa->pa_free) {
4391 atomic_inc(&pa->pa_count);
4392 ext4_mb_use_inode_pa(ac, pa);
4393 spin_unlock(&pa->pa_lock);
4394 ac->ac_criteria = 10;
4395 rcu_read_unlock();
4396 return true;
4397 }
4398 spin_unlock(&pa->pa_lock);
4399 }
4400 rcu_read_unlock();
4401
4402 /* can we use group allocation? */
4403 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4404 return false;
4405
4406 /* inode may have no locality group for some reason */
4407 lg = ac->ac_lg;
4408 if (lg == NULL)
4409 return false;
4410 order = fls(ac->ac_o_ex.fe_len) - 1;
4411 if (order > PREALLOC_TB_SIZE - 1)
4412 /* The max size of hash table is PREALLOC_TB_SIZE */
4413 order = PREALLOC_TB_SIZE - 1;
4414
4415 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4416 /*
4417 * search for the prealloc space that is having
4418 * minimal distance from the goal block.
4419 */
4420 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4421 rcu_read_lock();
4422 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4423 pa_inode_list) {
4424 spin_lock(&pa->pa_lock);
4425 if (pa->pa_deleted == 0 &&
4426 pa->pa_free >= ac->ac_o_ex.fe_len) {
4427
4428 cpa = ext4_mb_check_group_pa(goal_block,
4429 pa, cpa);
4430 }
4431 spin_unlock(&pa->pa_lock);
4432 }
4433 rcu_read_unlock();
4434 }
4435 if (cpa) {
4436 ext4_mb_use_group_pa(ac, cpa);
4437 ac->ac_criteria = 20;
4438 return true;
4439 }
4440 return false;
4441}
4442
4443/*
4444 * the function goes through all block freed in the group
4445 * but not yet committed and marks them used in in-core bitmap.
4446 * buddy must be generated from this bitmap
4447 * Need to be called with the ext4 group lock held
4448 */
4449static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4450 ext4_group_t group)
4451{
4452 struct rb_node *n;
4453 struct ext4_group_info *grp;
4454 struct ext4_free_data *entry;
4455
4456 grp = ext4_get_group_info(sb, group);
4457 n = rb_first(&(grp->bb_free_root));
4458
4459 while (n) {
4460 entry = rb_entry(n, struct ext4_free_data, efd_node);
4461 mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4462 n = rb_next(n);
4463 }
4464 return;
4465}
4466
4467/*
4468 * the function goes through all preallocation in this group and marks them
4469 * used in in-core bitmap. buddy must be generated from this bitmap
4470 * Need to be called with ext4 group lock held
4471 */
4472static noinline_for_stack
4473void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4474 ext4_group_t group)
4475{
4476 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4477 struct ext4_prealloc_space *pa;
4478 struct list_head *cur;
4479 ext4_group_t groupnr;
4480 ext4_grpblk_t start;
4481 int preallocated = 0;
4482 int len;
4483
4484 /* all form of preallocation discards first load group,
4485 * so the only competing code is preallocation use.
4486 * we don't need any locking here
4487 * notice we do NOT ignore preallocations with pa_deleted
4488 * otherwise we could leave used blocks available for
4489 * allocation in buddy when concurrent ext4_mb_put_pa()
4490 * is dropping preallocation
4491 */
4492 list_for_each(cur, &grp->bb_prealloc_list) {
4493 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4494 spin_lock(&pa->pa_lock);
4495 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4496 &groupnr, &start);
4497 len = pa->pa_len;
4498 spin_unlock(&pa->pa_lock);
4499 if (unlikely(len == 0))
4500 continue;
4501 BUG_ON(groupnr != group);
4502 mb_set_bits(bitmap, start, len);
4503 preallocated += len;
4504 }
4505 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4506}
4507
4508static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4509 struct ext4_prealloc_space *pa)
4510{
4511 struct ext4_inode_info *ei;
4512
4513 if (pa->pa_deleted) {
4514 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4515 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4516 pa->pa_len);
4517 return;
4518 }
4519
4520 pa->pa_deleted = 1;
4521
4522 if (pa->pa_type == MB_INODE_PA) {
4523 ei = EXT4_I(pa->pa_inode);
4524 atomic_dec(&ei->i_prealloc_active);
4525 }
4526}
4527
4528static void ext4_mb_pa_callback(struct rcu_head *head)
4529{
4530 struct ext4_prealloc_space *pa;
4531 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4532
4533 BUG_ON(atomic_read(&pa->pa_count));
4534 BUG_ON(pa->pa_deleted == 0);
4535 kmem_cache_free(ext4_pspace_cachep, pa);
4536}
4537
4538/*
4539 * drops a reference to preallocated space descriptor
4540 * if this was the last reference and the space is consumed
4541 */
4542static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4543 struct super_block *sb, struct ext4_prealloc_space *pa)
4544{
4545 ext4_group_t grp;
4546 ext4_fsblk_t grp_blk;
4547
4548 /* in this short window concurrent discard can set pa_deleted */
4549 spin_lock(&pa->pa_lock);
4550 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4551 spin_unlock(&pa->pa_lock);
4552 return;
4553 }
4554
4555 if (pa->pa_deleted == 1) {
4556 spin_unlock(&pa->pa_lock);
4557 return;
4558 }
4559
4560 ext4_mb_mark_pa_deleted(sb, pa);
4561 spin_unlock(&pa->pa_lock);
4562
4563 grp_blk = pa->pa_pstart;
4564 /*
4565 * If doing group-based preallocation, pa_pstart may be in the
4566 * next group when pa is used up
4567 */
4568 if (pa->pa_type == MB_GROUP_PA)
4569 grp_blk--;
4570
4571 grp = ext4_get_group_number(sb, grp_blk);
4572
4573 /*
4574 * possible race:
4575 *
4576 * P1 (buddy init) P2 (regular allocation)
4577 * find block B in PA
4578 * copy on-disk bitmap to buddy
4579 * mark B in on-disk bitmap
4580 * drop PA from group
4581 * mark all PAs in buddy
4582 *
4583 * thus, P1 initializes buddy with B available. to prevent this
4584 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4585 * against that pair
4586 */
4587 ext4_lock_group(sb, grp);
4588 list_del(&pa->pa_group_list);
4589 ext4_unlock_group(sb, grp);
4590
4591 spin_lock(pa->pa_obj_lock);
4592 list_del_rcu(&pa->pa_inode_list);
4593 spin_unlock(pa->pa_obj_lock);
4594
4595 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4596}
4597
4598/*
4599 * creates new preallocated space for given inode
4600 */
4601static noinline_for_stack void
4602ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4603{
4604 struct super_block *sb = ac->ac_sb;
4605 struct ext4_sb_info *sbi = EXT4_SB(sb);
4606 struct ext4_prealloc_space *pa;
4607 struct ext4_group_info *grp;
4608 struct ext4_inode_info *ei;
4609
4610 /* preallocate only when found space is larger then requested */
4611 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4612 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4613 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4614 BUG_ON(ac->ac_pa == NULL);
4615
4616 pa = ac->ac_pa;
4617
4618 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4619 int winl;
4620 int wins;
4621 int win;
4622 int offs;
4623
4624 /* we can't allocate as much as normalizer wants.
4625 * so, found space must get proper lstart
4626 * to cover original request */
4627 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4628 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4629
4630 /* we're limited by original request in that
4631 * logical block must be covered any way
4632 * winl is window we can move our chunk within */
4633 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4634
4635 /* also, we should cover whole original request */
4636 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4637
4638 /* the smallest one defines real window */
4639 win = min(winl, wins);
4640
4641 offs = ac->ac_o_ex.fe_logical %
4642 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4643 if (offs && offs < win)
4644 win = offs;
4645
4646 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4647 EXT4_NUM_B2C(sbi, win);
4648 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4649 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4650 }
4651
4652 /* preallocation can change ac_b_ex, thus we store actually
4653 * allocated blocks for history */
4654 ac->ac_f_ex = ac->ac_b_ex;
4655
4656 pa->pa_lstart = ac->ac_b_ex.fe_logical;
4657 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4658 pa->pa_len = ac->ac_b_ex.fe_len;
4659 pa->pa_free = pa->pa_len;
4660 spin_lock_init(&pa->pa_lock);
4661 INIT_LIST_HEAD(&pa->pa_inode_list);
4662 INIT_LIST_HEAD(&pa->pa_group_list);
4663 pa->pa_deleted = 0;
4664 pa->pa_type = MB_INODE_PA;
4665
4666 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4667 pa->pa_len, pa->pa_lstart);
4668 trace_ext4_mb_new_inode_pa(ac, pa);
4669
4670 ext4_mb_use_inode_pa(ac, pa);
4671 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4672
4673 ei = EXT4_I(ac->ac_inode);
4674 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4675
4676 pa->pa_obj_lock = &ei->i_prealloc_lock;
4677 pa->pa_inode = ac->ac_inode;
4678
4679 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4680
4681 spin_lock(pa->pa_obj_lock);
4682 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4683 spin_unlock(pa->pa_obj_lock);
4684 atomic_inc(&ei->i_prealloc_active);
4685}
4686
4687/*
4688 * creates new preallocated space for locality group inodes belongs to
4689 */
4690static noinline_for_stack void
4691ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4692{
4693 struct super_block *sb = ac->ac_sb;
4694 struct ext4_locality_group *lg;
4695 struct ext4_prealloc_space *pa;
4696 struct ext4_group_info *grp;
4697
4698 /* preallocate only when found space is larger then requested */
4699 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4700 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4701 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4702 BUG_ON(ac->ac_pa == NULL);
4703
4704 pa = ac->ac_pa;
4705
4706 /* preallocation can change ac_b_ex, thus we store actually
4707 * allocated blocks for history */
4708 ac->ac_f_ex = ac->ac_b_ex;
4709
4710 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4711 pa->pa_lstart = pa->pa_pstart;
4712 pa->pa_len = ac->ac_b_ex.fe_len;
4713 pa->pa_free = pa->pa_len;
4714 spin_lock_init(&pa->pa_lock);
4715 INIT_LIST_HEAD(&pa->pa_inode_list);
4716 INIT_LIST_HEAD(&pa->pa_group_list);
4717 pa->pa_deleted = 0;
4718 pa->pa_type = MB_GROUP_PA;
4719
4720 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4721 pa->pa_len, pa->pa_lstart);
4722 trace_ext4_mb_new_group_pa(ac, pa);
4723
4724 ext4_mb_use_group_pa(ac, pa);
4725 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4726
4727 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4728 lg = ac->ac_lg;
4729 BUG_ON(lg == NULL);
4730
4731 pa->pa_obj_lock = &lg->lg_prealloc_lock;
4732 pa->pa_inode = NULL;
4733
4734 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4735
4736 /*
4737 * We will later add the new pa to the right bucket
4738 * after updating the pa_free in ext4_mb_release_context
4739 */
4740}
4741
4742static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4743{
4744 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4745 ext4_mb_new_group_pa(ac);
4746 else
4747 ext4_mb_new_inode_pa(ac);
4748}
4749
4750/*
4751 * finds all unused blocks in on-disk bitmap, frees them in
4752 * in-core bitmap and buddy.
4753 * @pa must be unlinked from inode and group lists, so that
4754 * nobody else can find/use it.
4755 * the caller MUST hold group/inode locks.
4756 * TODO: optimize the case when there are no in-core structures yet
4757 */
4758static noinline_for_stack int
4759ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4760 struct ext4_prealloc_space *pa)
4761{
4762 struct super_block *sb = e4b->bd_sb;
4763 struct ext4_sb_info *sbi = EXT4_SB(sb);
4764 unsigned int end;
4765 unsigned int next;
4766 ext4_group_t group;
4767 ext4_grpblk_t bit;
4768 unsigned long long grp_blk_start;
4769 int free = 0;
4770
4771 BUG_ON(pa->pa_deleted == 0);
4772 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4773 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4774 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4775 end = bit + pa->pa_len;
4776
4777 while (bit < end) {
4778 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4779 if (bit >= end)
4780 break;
4781 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4782 mb_debug(sb, "free preallocated %u/%u in group %u\n",
4783 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4784 (unsigned) next - bit, (unsigned) group);
4785 free += next - bit;
4786
4787 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4788 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4789 EXT4_C2B(sbi, bit)),
4790 next - bit);
4791 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4792 bit = next + 1;
4793 }
4794 if (free != pa->pa_free) {
4795 ext4_msg(e4b->bd_sb, KERN_CRIT,
4796 "pa %p: logic %lu, phys. %lu, len %d",
4797 pa, (unsigned long) pa->pa_lstart,
4798 (unsigned long) pa->pa_pstart,
4799 pa->pa_len);
4800 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4801 free, pa->pa_free);
4802 /*
4803 * pa is already deleted so we use the value obtained
4804 * from the bitmap and continue.
4805 */
4806 }
4807 atomic_add(free, &sbi->s_mb_discarded);
4808
4809 return 0;
4810}
4811
4812static noinline_for_stack int
4813ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4814 struct ext4_prealloc_space *pa)
4815{
4816 struct super_block *sb = e4b->bd_sb;
4817 ext4_group_t group;
4818 ext4_grpblk_t bit;
4819
4820 trace_ext4_mb_release_group_pa(sb, pa);
4821 BUG_ON(pa->pa_deleted == 0);
4822 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4823 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4824 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4825 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4826 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4827
4828 return 0;
4829}
4830
4831/*
4832 * releases all preallocations in given group
4833 *
4834 * first, we need to decide discard policy:
4835 * - when do we discard
4836 * 1) ENOSPC
4837 * - how many do we discard
4838 * 1) how many requested
4839 */
4840static noinline_for_stack int
4841ext4_mb_discard_group_preallocations(struct super_block *sb,
4842 ext4_group_t group, int *busy)
4843{
4844 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4845 struct buffer_head *bitmap_bh = NULL;
4846 struct ext4_prealloc_space *pa, *tmp;
4847 struct list_head list;
4848 struct ext4_buddy e4b;
4849 int err;
4850 int free = 0;
4851
4852 mb_debug(sb, "discard preallocation for group %u\n", group);
4853 if (list_empty(&grp->bb_prealloc_list))
4854 goto out_dbg;
4855
4856 bitmap_bh = ext4_read_block_bitmap(sb, group);
4857 if (IS_ERR(bitmap_bh)) {
4858 err = PTR_ERR(bitmap_bh);
4859 ext4_error_err(sb, -err,
4860 "Error %d reading block bitmap for %u",
4861 err, group);
4862 goto out_dbg;
4863 }
4864
4865 err = ext4_mb_load_buddy(sb, group, &e4b);
4866 if (err) {
4867 ext4_warning(sb, "Error %d loading buddy information for %u",
4868 err, group);
4869 put_bh(bitmap_bh);
4870 goto out_dbg;
4871 }
4872
4873 INIT_LIST_HEAD(&list);
4874 ext4_lock_group(sb, group);
4875 list_for_each_entry_safe(pa, tmp,
4876 &grp->bb_prealloc_list, pa_group_list) {
4877 spin_lock(&pa->pa_lock);
4878 if (atomic_read(&pa->pa_count)) {
4879 spin_unlock(&pa->pa_lock);
4880 *busy = 1;
4881 continue;
4882 }
4883 if (pa->pa_deleted) {
4884 spin_unlock(&pa->pa_lock);
4885 continue;
4886 }
4887
4888 /* seems this one can be freed ... */
4889 ext4_mb_mark_pa_deleted(sb, pa);
4890
4891 if (!free)
4892 this_cpu_inc(discard_pa_seq);
4893
4894 /* we can trust pa_free ... */
4895 free += pa->pa_free;
4896
4897 spin_unlock(&pa->pa_lock);
4898
4899 list_del(&pa->pa_group_list);
4900 list_add(&pa->u.pa_tmp_list, &list);
4901 }
4902
4903 /* now free all selected PAs */
4904 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4905
4906 /* remove from object (inode or locality group) */
4907 spin_lock(pa->pa_obj_lock);
4908 list_del_rcu(&pa->pa_inode_list);
4909 spin_unlock(pa->pa_obj_lock);
4910
4911 if (pa->pa_type == MB_GROUP_PA)
4912 ext4_mb_release_group_pa(&e4b, pa);
4913 else
4914 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4915
4916 list_del(&pa->u.pa_tmp_list);
4917 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4918 }
4919
4920 ext4_unlock_group(sb, group);
4921 ext4_mb_unload_buddy(&e4b);
4922 put_bh(bitmap_bh);
4923out_dbg:
4924 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4925 free, group, grp->bb_free);
4926 return free;
4927}
4928
4929/*
4930 * releases all non-used preallocated blocks for given inode
4931 *
4932 * It's important to discard preallocations under i_data_sem
4933 * We don't want another block to be served from the prealloc
4934 * space when we are discarding the inode prealloc space.
4935 *
4936 * FIXME!! Make sure it is valid at all the call sites
4937 */
4938void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4939{
4940 struct ext4_inode_info *ei = EXT4_I(inode);
4941 struct super_block *sb = inode->i_sb;
4942 struct buffer_head *bitmap_bh = NULL;
4943 struct ext4_prealloc_space *pa, *tmp;
4944 ext4_group_t group = 0;
4945 struct list_head list;
4946 struct ext4_buddy e4b;
4947 int err;
4948
4949 if (!S_ISREG(inode->i_mode)) {
4950 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4951 return;
4952 }
4953
4954 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4955 return;
4956
4957 mb_debug(sb, "discard preallocation for inode %lu\n",
4958 inode->i_ino);
4959 trace_ext4_discard_preallocations(inode,
4960 atomic_read(&ei->i_prealloc_active), needed);
4961
4962 INIT_LIST_HEAD(&list);
4963
4964 if (needed == 0)
4965 needed = UINT_MAX;
4966
4967repeat:
4968 /* first, collect all pa's in the inode */
4969 spin_lock(&ei->i_prealloc_lock);
4970 while (!list_empty(&ei->i_prealloc_list) && needed) {
4971 pa = list_entry(ei->i_prealloc_list.prev,
4972 struct ext4_prealloc_space, pa_inode_list);
4973 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4974 spin_lock(&pa->pa_lock);
4975 if (atomic_read(&pa->pa_count)) {
4976 /* this shouldn't happen often - nobody should
4977 * use preallocation while we're discarding it */
4978 spin_unlock(&pa->pa_lock);
4979 spin_unlock(&ei->i_prealloc_lock);
4980 ext4_msg(sb, KERN_ERR,
4981 "uh-oh! used pa while discarding");
4982 WARN_ON(1);
4983 schedule_timeout_uninterruptible(HZ);
4984 goto repeat;
4985
4986 }
4987 if (pa->pa_deleted == 0) {
4988 ext4_mb_mark_pa_deleted(sb, pa);
4989 spin_unlock(&pa->pa_lock);
4990 list_del_rcu(&pa->pa_inode_list);
4991 list_add(&pa->u.pa_tmp_list, &list);
4992 needed--;
4993 continue;
4994 }
4995
4996 /* someone is deleting pa right now */
4997 spin_unlock(&pa->pa_lock);
4998 spin_unlock(&ei->i_prealloc_lock);
4999
5000 /* we have to wait here because pa_deleted
5001 * doesn't mean pa is already unlinked from
5002 * the list. as we might be called from
5003 * ->clear_inode() the inode will get freed
5004 * and concurrent thread which is unlinking
5005 * pa from inode's list may access already
5006 * freed memory, bad-bad-bad */
5007
5008 /* XXX: if this happens too often, we can
5009 * add a flag to force wait only in case
5010 * of ->clear_inode(), but not in case of
5011 * regular truncate */
5012 schedule_timeout_uninterruptible(HZ);
5013 goto repeat;
5014 }
5015 spin_unlock(&ei->i_prealloc_lock);
5016
5017 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5018 BUG_ON(pa->pa_type != MB_INODE_PA);
5019 group = ext4_get_group_number(sb, pa->pa_pstart);
5020
5021 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5022 GFP_NOFS|__GFP_NOFAIL);
5023 if (err) {
5024 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5025 err, group);
5026 continue;
5027 }
5028
5029 bitmap_bh = ext4_read_block_bitmap(sb, group);
5030 if (IS_ERR(bitmap_bh)) {
5031 err = PTR_ERR(bitmap_bh);
5032 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5033 err, group);
5034 ext4_mb_unload_buddy(&e4b);
5035 continue;
5036 }
5037
5038 ext4_lock_group(sb, group);
5039 list_del(&pa->pa_group_list);
5040 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5041 ext4_unlock_group(sb, group);
5042
5043 ext4_mb_unload_buddy(&e4b);
5044 put_bh(bitmap_bh);
5045
5046 list_del(&pa->u.pa_tmp_list);
5047 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5048 }
5049}
5050
5051static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5052{
5053 struct ext4_prealloc_space *pa;
5054
5055 BUG_ON(ext4_pspace_cachep == NULL);
5056 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5057 if (!pa)
5058 return -ENOMEM;
5059 atomic_set(&pa->pa_count, 1);
5060 ac->ac_pa = pa;
5061 return 0;
5062}
5063
5064static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5065{
5066 struct ext4_prealloc_space *pa = ac->ac_pa;
5067
5068 BUG_ON(!pa);
5069 ac->ac_pa = NULL;
5070 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5071 kmem_cache_free(ext4_pspace_cachep, pa);
5072}
5073
5074#ifdef CONFIG_EXT4_DEBUG
5075static inline void ext4_mb_show_pa(struct super_block *sb)
5076{
5077 ext4_group_t i, ngroups;
5078
5079 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5080 return;
5081
5082 ngroups = ext4_get_groups_count(sb);
5083 mb_debug(sb, "groups: ");
5084 for (i = 0; i < ngroups; i++) {
5085 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5086 struct ext4_prealloc_space *pa;
5087 ext4_grpblk_t start;
5088 struct list_head *cur;
5089 ext4_lock_group(sb, i);
5090 list_for_each(cur, &grp->bb_prealloc_list) {
5091 pa = list_entry(cur, struct ext4_prealloc_space,
5092 pa_group_list);
5093 spin_lock(&pa->pa_lock);
5094 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5095 NULL, &start);
5096 spin_unlock(&pa->pa_lock);
5097 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5098 pa->pa_len);
5099 }
5100 ext4_unlock_group(sb, i);
5101 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5102 grp->bb_fragments);
5103 }
5104}
5105
5106static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5107{
5108 struct super_block *sb = ac->ac_sb;
5109
5110 if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5111 return;
5112
5113 mb_debug(sb, "Can't allocate:"
5114 " Allocation context details:");
5115 mb_debug(sb, "status %u flags 0x%x",
5116 ac->ac_status, ac->ac_flags);
5117 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5118 "goal %lu/%lu/%lu@%lu, "
5119 "best %lu/%lu/%lu@%lu cr %d",
5120 (unsigned long)ac->ac_o_ex.fe_group,
5121 (unsigned long)ac->ac_o_ex.fe_start,
5122 (unsigned long)ac->ac_o_ex.fe_len,
5123 (unsigned long)ac->ac_o_ex.fe_logical,
5124 (unsigned long)ac->ac_g_ex.fe_group,
5125 (unsigned long)ac->ac_g_ex.fe_start,
5126 (unsigned long)ac->ac_g_ex.fe_len,
5127 (unsigned long)ac->ac_g_ex.fe_logical,
5128 (unsigned long)ac->ac_b_ex.fe_group,
5129 (unsigned long)ac->ac_b_ex.fe_start,
5130 (unsigned long)ac->ac_b_ex.fe_len,
5131 (unsigned long)ac->ac_b_ex.fe_logical,
5132 (int)ac->ac_criteria);
5133 mb_debug(sb, "%u found", ac->ac_found);
5134 ext4_mb_show_pa(sb);
5135}
5136#else
5137static inline void ext4_mb_show_pa(struct super_block *sb)
5138{
5139 return;
5140}
5141static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5142{
5143 ext4_mb_show_pa(ac->ac_sb);
5144 return;
5145}
5146#endif
5147
5148/*
5149 * We use locality group preallocation for small size file. The size of the
5150 * file is determined by the current size or the resulting size after
5151 * allocation which ever is larger
5152 *
5153 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5154 */
5155static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5156{
5157 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5158 int bsbits = ac->ac_sb->s_blocksize_bits;
5159 loff_t size, isize;
5160 bool inode_pa_eligible, group_pa_eligible;
5161
5162 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5163 return;
5164
5165 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5166 return;
5167
5168 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5169 inode_pa_eligible = true;
5170 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5171 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5172 >> bsbits;
5173
5174 /* No point in using inode preallocation for closed files */
5175 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5176 !inode_is_open_for_write(ac->ac_inode))
5177 inode_pa_eligible = false;
5178
5179 size = max(size, isize);
5180 /* Don't use group allocation for large files */
5181 if (size > sbi->s_mb_stream_request)
5182 group_pa_eligible = false;
5183
5184 if (!group_pa_eligible) {
5185 if (inode_pa_eligible)
5186 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5187 else
5188 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5189 return;
5190 }
5191
5192 BUG_ON(ac->ac_lg != NULL);
5193 /*
5194 * locality group prealloc space are per cpu. The reason for having
5195 * per cpu locality group is to reduce the contention between block
5196 * request from multiple CPUs.
5197 */
5198 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5199
5200 /* we're going to use group allocation */
5201 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5202
5203 /* serialize all allocations in the group */
5204 mutex_lock(&ac->ac_lg->lg_mutex);
5205}
5206
5207static noinline_for_stack void
5208ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5209 struct ext4_allocation_request *ar)
5210{
5211 struct super_block *sb = ar->inode->i_sb;
5212 struct ext4_sb_info *sbi = EXT4_SB(sb);
5213 struct ext4_super_block *es = sbi->s_es;
5214 ext4_group_t group;
5215 unsigned int len;
5216 ext4_fsblk_t goal;
5217 ext4_grpblk_t block;
5218
5219 /* we can't allocate > group size */
5220 len = ar->len;
5221
5222 /* just a dirty hack to filter too big requests */
5223 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5224 len = EXT4_CLUSTERS_PER_GROUP(sb);
5225
5226 /* start searching from the goal */
5227 goal = ar->goal;
5228 if (goal < le32_to_cpu(es->s_first_data_block) ||
5229 goal >= ext4_blocks_count(es))
5230 goal = le32_to_cpu(es->s_first_data_block);
5231 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5232
5233 /* set up allocation goals */
5234 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5235 ac->ac_status = AC_STATUS_CONTINUE;
5236 ac->ac_sb = sb;
5237 ac->ac_inode = ar->inode;
5238 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5239 ac->ac_o_ex.fe_group = group;
5240 ac->ac_o_ex.fe_start = block;
5241 ac->ac_o_ex.fe_len = len;
5242 ac->ac_g_ex = ac->ac_o_ex;
5243 ac->ac_flags = ar->flags;
5244
5245 /* we have to define context: we'll work with a file or
5246 * locality group. this is a policy, actually */
5247 ext4_mb_group_or_file(ac);
5248
5249 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5250 "left: %u/%u, right %u/%u to %swritable\n",
5251 (unsigned) ar->len, (unsigned) ar->logical,
5252 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5253 (unsigned) ar->lleft, (unsigned) ar->pleft,
5254 (unsigned) ar->lright, (unsigned) ar->pright,
5255 inode_is_open_for_write(ar->inode) ? "" : "non-");
5256}
5257
5258static noinline_for_stack void
5259ext4_mb_discard_lg_preallocations(struct super_block *sb,
5260 struct ext4_locality_group *lg,
5261 int order, int total_entries)
5262{
5263 ext4_group_t group = 0;
5264 struct ext4_buddy e4b;
5265 struct list_head discard_list;
5266 struct ext4_prealloc_space *pa, *tmp;
5267
5268 mb_debug(sb, "discard locality group preallocation\n");
5269
5270 INIT_LIST_HEAD(&discard_list);
5271
5272 spin_lock(&lg->lg_prealloc_lock);
5273 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5274 pa_inode_list,
5275 lockdep_is_held(&lg->lg_prealloc_lock)) {
5276 spin_lock(&pa->pa_lock);
5277 if (atomic_read(&pa->pa_count)) {
5278 /*
5279 * This is the pa that we just used
5280 * for block allocation. So don't
5281 * free that
5282 */
5283 spin_unlock(&pa->pa_lock);
5284 continue;
5285 }
5286 if (pa->pa_deleted) {
5287 spin_unlock(&pa->pa_lock);
5288 continue;
5289 }
5290 /* only lg prealloc space */
5291 BUG_ON(pa->pa_type != MB_GROUP_PA);
5292
5293 /* seems this one can be freed ... */
5294 ext4_mb_mark_pa_deleted(sb, pa);
5295 spin_unlock(&pa->pa_lock);
5296
5297 list_del_rcu(&pa->pa_inode_list);
5298 list_add(&pa->u.pa_tmp_list, &discard_list);
5299
5300 total_entries--;
5301 if (total_entries <= 5) {
5302 /*
5303 * we want to keep only 5 entries
5304 * allowing it to grow to 8. This
5305 * mak sure we don't call discard
5306 * soon for this list.
5307 */
5308 break;
5309 }
5310 }
5311 spin_unlock(&lg->lg_prealloc_lock);
5312
5313 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5314 int err;
5315
5316 group = ext4_get_group_number(sb, pa->pa_pstart);
5317 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5318 GFP_NOFS|__GFP_NOFAIL);
5319 if (err) {
5320 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5321 err, group);
5322 continue;
5323 }
5324 ext4_lock_group(sb, group);
5325 list_del(&pa->pa_group_list);
5326 ext4_mb_release_group_pa(&e4b, pa);
5327 ext4_unlock_group(sb, group);
5328
5329 ext4_mb_unload_buddy(&e4b);
5330 list_del(&pa->u.pa_tmp_list);
5331 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5332 }
5333}
5334
5335/*
5336 * We have incremented pa_count. So it cannot be freed at this
5337 * point. Also we hold lg_mutex. So no parallel allocation is
5338 * possible from this lg. That means pa_free cannot be updated.
5339 *
5340 * A parallel ext4_mb_discard_group_preallocations is possible.
5341 * which can cause the lg_prealloc_list to be updated.
5342 */
5343
5344static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5345{
5346 int order, added = 0, lg_prealloc_count = 1;
5347 struct super_block *sb = ac->ac_sb;
5348 struct ext4_locality_group *lg = ac->ac_lg;
5349 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5350
5351 order = fls(pa->pa_free) - 1;
5352 if (order > PREALLOC_TB_SIZE - 1)
5353 /* The max size of hash table is PREALLOC_TB_SIZE */
5354 order = PREALLOC_TB_SIZE - 1;
5355 /* Add the prealloc space to lg */
5356 spin_lock(&lg->lg_prealloc_lock);
5357 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5358 pa_inode_list,
5359 lockdep_is_held(&lg->lg_prealloc_lock)) {
5360 spin_lock(&tmp_pa->pa_lock);
5361 if (tmp_pa->pa_deleted) {
5362 spin_unlock(&tmp_pa->pa_lock);
5363 continue;
5364 }
5365 if (!added && pa->pa_free < tmp_pa->pa_free) {
5366 /* Add to the tail of the previous entry */
5367 list_add_tail_rcu(&pa->pa_inode_list,
5368 &tmp_pa->pa_inode_list);
5369 added = 1;
5370 /*
5371 * we want to count the total
5372 * number of entries in the list
5373 */
5374 }
5375 spin_unlock(&tmp_pa->pa_lock);
5376 lg_prealloc_count++;
5377 }
5378 if (!added)
5379 list_add_tail_rcu(&pa->pa_inode_list,
5380 &lg->lg_prealloc_list[order]);
5381 spin_unlock(&lg->lg_prealloc_lock);
5382
5383 /* Now trim the list to be not more than 8 elements */
5384 if (lg_prealloc_count > 8) {
5385 ext4_mb_discard_lg_preallocations(sb, lg,
5386 order, lg_prealloc_count);
5387 return;
5388 }
5389 return ;
5390}
5391
5392/*
5393 * if per-inode prealloc list is too long, trim some PA
5394 */
5395static void ext4_mb_trim_inode_pa(struct inode *inode)
5396{
5397 struct ext4_inode_info *ei = EXT4_I(inode);
5398 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5399 int count, delta;
5400
5401 count = atomic_read(&ei->i_prealloc_active);
5402 delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5403 if (count > sbi->s_mb_max_inode_prealloc + delta) {
5404 count -= sbi->s_mb_max_inode_prealloc;
5405 ext4_discard_preallocations(inode, count);
5406 }
5407}
5408
5409/*
5410 * release all resource we used in allocation
5411 */
5412static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5413{
5414 struct inode *inode = ac->ac_inode;
5415 struct ext4_inode_info *ei = EXT4_I(inode);
5416 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5417 struct ext4_prealloc_space *pa = ac->ac_pa;
5418 if (pa) {
5419 if (pa->pa_type == MB_GROUP_PA) {
5420 /* see comment in ext4_mb_use_group_pa() */
5421 spin_lock(&pa->pa_lock);
5422 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5423 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5424 pa->pa_free -= ac->ac_b_ex.fe_len;
5425 pa->pa_len -= ac->ac_b_ex.fe_len;
5426 spin_unlock(&pa->pa_lock);
5427
5428 /*
5429 * We want to add the pa to the right bucket.
5430 * Remove it from the list and while adding
5431 * make sure the list to which we are adding
5432 * doesn't grow big.
5433 */
5434 if (likely(pa->pa_free)) {
5435 spin_lock(pa->pa_obj_lock);
5436 list_del_rcu(&pa->pa_inode_list);
5437 spin_unlock(pa->pa_obj_lock);
5438 ext4_mb_add_n_trim(ac);
5439 }
5440 }
5441
5442 if (pa->pa_type == MB_INODE_PA) {
5443 /*
5444 * treat per-inode prealloc list as a lru list, then try
5445 * to trim the least recently used PA.
5446 */
5447 spin_lock(pa->pa_obj_lock);
5448 list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5449 spin_unlock(pa->pa_obj_lock);
5450 }
5451
5452 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5453 }
5454 if (ac->ac_bitmap_page)
5455 put_page(ac->ac_bitmap_page);
5456 if (ac->ac_buddy_page)
5457 put_page(ac->ac_buddy_page);
5458 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5459 mutex_unlock(&ac->ac_lg->lg_mutex);
5460 ext4_mb_collect_stats(ac);
5461 ext4_mb_trim_inode_pa(inode);
5462 return 0;
5463}
5464
5465static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5466{
5467 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5468 int ret;
5469 int freed = 0, busy = 0;
5470 int retry = 0;
5471
5472 trace_ext4_mb_discard_preallocations(sb, needed);
5473
5474 if (needed == 0)
5475 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5476 repeat:
5477 for (i = 0; i < ngroups && needed > 0; i++) {
5478 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5479 freed += ret;
5480 needed -= ret;
5481 cond_resched();
5482 }
5483
5484 if (needed > 0 && busy && ++retry < 3) {
5485 busy = 0;
5486 goto repeat;
5487 }
5488
5489 return freed;
5490}
5491
5492static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5493 struct ext4_allocation_context *ac, u64 *seq)
5494{
5495 int freed;
5496 u64 seq_retry = 0;
5497 bool ret = false;
5498
5499 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5500 if (freed) {
5501 ret = true;
5502 goto out_dbg;
5503 }
5504 seq_retry = ext4_get_discard_pa_seq_sum();
5505 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5506 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5507 *seq = seq_retry;
5508 ret = true;
5509 }
5510
5511out_dbg:
5512 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5513 return ret;
5514}
5515
5516static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5517 struct ext4_allocation_request *ar, int *errp);
5518
5519/*
5520 * Main entry point into mballoc to allocate blocks
5521 * it tries to use preallocation first, then falls back
5522 * to usual allocation
5523 */
5524ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5525 struct ext4_allocation_request *ar, int *errp)
5526{
5527 struct ext4_allocation_context *ac = NULL;
5528 struct ext4_sb_info *sbi;
5529 struct super_block *sb;
5530 ext4_fsblk_t block = 0;
5531 unsigned int inquota = 0;
5532 unsigned int reserv_clstrs = 0;
5533 int retries = 0;
5534 u64 seq;
5535
5536 might_sleep();
5537 sb = ar->inode->i_sb;
5538 sbi = EXT4_SB(sb);
5539
5540 trace_ext4_request_blocks(ar);
5541 if (sbi->s_mount_state & EXT4_FC_REPLAY)
5542 return ext4_mb_new_blocks_simple(handle, ar, errp);
5543
5544 /* Allow to use superuser reservation for quota file */
5545 if (ext4_is_quota_file(ar->inode))
5546 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5547
5548 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5549 /* Without delayed allocation we need to verify
5550 * there is enough free blocks to do block allocation
5551 * and verify allocation doesn't exceed the quota limits.
5552 */
5553 while (ar->len &&
5554 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5555
5556 /* let others to free the space */
5557 cond_resched();
5558 ar->len = ar->len >> 1;
5559 }
5560 if (!ar->len) {
5561 ext4_mb_show_pa(sb);
5562 *errp = -ENOSPC;
5563 return 0;
5564 }
5565 reserv_clstrs = ar->len;
5566 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5567 dquot_alloc_block_nofail(ar->inode,
5568 EXT4_C2B(sbi, ar->len));
5569 } else {
5570 while (ar->len &&
5571 dquot_alloc_block(ar->inode,
5572 EXT4_C2B(sbi, ar->len))) {
5573
5574 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5575 ar->len--;
5576 }
5577 }
5578 inquota = ar->len;
5579 if (ar->len == 0) {
5580 *errp = -EDQUOT;
5581 goto out;
5582 }
5583 }
5584
5585 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5586 if (!ac) {
5587 ar->len = 0;
5588 *errp = -ENOMEM;
5589 goto out;
5590 }
5591
5592 ext4_mb_initialize_context(ac, ar);
5593
5594 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5595 seq = this_cpu_read(discard_pa_seq);
5596 if (!ext4_mb_use_preallocated(ac)) {
5597 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5598 ext4_mb_normalize_request(ac, ar);
5599
5600 *errp = ext4_mb_pa_alloc(ac);
5601 if (*errp)
5602 goto errout;
5603repeat:
5604 /* allocate space in core */
5605 *errp = ext4_mb_regular_allocator(ac);
5606 /*
5607 * pa allocated above is added to grp->bb_prealloc_list only
5608 * when we were able to allocate some block i.e. when
5609 * ac->ac_status == AC_STATUS_FOUND.
5610 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5611 * So we have to free this pa here itself.
5612 */
5613 if (*errp) {
5614 ext4_mb_pa_free(ac);
5615 ext4_discard_allocated_blocks(ac);
5616 goto errout;
5617 }
5618 if (ac->ac_status == AC_STATUS_FOUND &&
5619 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5620 ext4_mb_pa_free(ac);
5621 }
5622 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5623 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5624 if (*errp) {
5625 ext4_discard_allocated_blocks(ac);
5626 goto errout;
5627 } else {
5628 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5629 ar->len = ac->ac_b_ex.fe_len;
5630 }
5631 } else {
5632 if (++retries < 3 &&
5633 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5634 goto repeat;
5635 /*
5636 * If block allocation fails then the pa allocated above
5637 * needs to be freed here itself.
5638 */
5639 ext4_mb_pa_free(ac);
5640 *errp = -ENOSPC;
5641 }
5642
5643errout:
5644 if (*errp) {
5645 ac->ac_b_ex.fe_len = 0;
5646 ar->len = 0;
5647 ext4_mb_show_ac(ac);
5648 }
5649 ext4_mb_release_context(ac);
5650out:
5651 if (ac)
5652 kmem_cache_free(ext4_ac_cachep, ac);
5653 if (inquota && ar->len < inquota)
5654 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5655 if (!ar->len) {
5656 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5657 /* release all the reserved blocks if non delalloc */
5658 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5659 reserv_clstrs);
5660 }
5661
5662 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5663
5664 return block;
5665}
5666
5667/*
5668 * We can merge two free data extents only if the physical blocks
5669 * are contiguous, AND the extents were freed by the same transaction,
5670 * AND the blocks are associated with the same group.
5671 */
5672static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5673 struct ext4_free_data *entry,
5674 struct ext4_free_data *new_entry,
5675 struct rb_root *entry_rb_root)
5676{
5677 if ((entry->efd_tid != new_entry->efd_tid) ||
5678 (entry->efd_group != new_entry->efd_group))
5679 return;
5680 if (entry->efd_start_cluster + entry->efd_count ==
5681 new_entry->efd_start_cluster) {
5682 new_entry->efd_start_cluster = entry->efd_start_cluster;
5683 new_entry->efd_count += entry->efd_count;
5684 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5685 entry->efd_start_cluster) {
5686 new_entry->efd_count += entry->efd_count;
5687 } else
5688 return;
5689 spin_lock(&sbi->s_md_lock);
5690 list_del(&entry->efd_list);
5691 spin_unlock(&sbi->s_md_lock);
5692 rb_erase(&entry->efd_node, entry_rb_root);
5693 kmem_cache_free(ext4_free_data_cachep, entry);
5694}
5695
5696static noinline_for_stack int
5697ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5698 struct ext4_free_data *new_entry)
5699{
5700 ext4_group_t group = e4b->bd_group;
5701 ext4_grpblk_t cluster;
5702 ext4_grpblk_t clusters = new_entry->efd_count;
5703 struct ext4_free_data *entry;
5704 struct ext4_group_info *db = e4b->bd_info;
5705 struct super_block *sb = e4b->bd_sb;
5706 struct ext4_sb_info *sbi = EXT4_SB(sb);
5707 struct rb_node **n = &db->bb_free_root.rb_node, *node;
5708 struct rb_node *parent = NULL, *new_node;
5709
5710 BUG_ON(!ext4_handle_valid(handle));
5711 BUG_ON(e4b->bd_bitmap_page == NULL);
5712 BUG_ON(e4b->bd_buddy_page == NULL);
5713
5714 new_node = &new_entry->efd_node;
5715 cluster = new_entry->efd_start_cluster;
5716
5717 if (!*n) {
5718 /* first free block exent. We need to
5719 protect buddy cache from being freed,
5720 * otherwise we'll refresh it from
5721 * on-disk bitmap and lose not-yet-available
5722 * blocks */
5723 get_page(e4b->bd_buddy_page);
5724 get_page(e4b->bd_bitmap_page);
5725 }
5726 while (*n) {
5727 parent = *n;
5728 entry = rb_entry(parent, struct ext4_free_data, efd_node);
5729 if (cluster < entry->efd_start_cluster)
5730 n = &(*n)->rb_left;
5731 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5732 n = &(*n)->rb_right;
5733 else {
5734 ext4_grp_locked_error(sb, group, 0,
5735 ext4_group_first_block_no(sb, group) +
5736 EXT4_C2B(sbi, cluster),
5737 "Block already on to-be-freed list");
5738 kmem_cache_free(ext4_free_data_cachep, new_entry);
5739 return 0;
5740 }
5741 }
5742
5743 rb_link_node(new_node, parent, n);
5744 rb_insert_color(new_node, &db->bb_free_root);
5745
5746 /* Now try to see the extent can be merged to left and right */
5747 node = rb_prev(new_node);
5748 if (node) {
5749 entry = rb_entry(node, struct ext4_free_data, efd_node);
5750 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5751 &(db->bb_free_root));
5752 }
5753
5754 node = rb_next(new_node);
5755 if (node) {
5756 entry = rb_entry(node, struct ext4_free_data, efd_node);
5757 ext4_try_merge_freed_extent(sbi, entry, new_entry,
5758 &(db->bb_free_root));
5759 }
5760
5761 spin_lock(&sbi->s_md_lock);
5762 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5763 sbi->s_mb_free_pending += clusters;
5764 spin_unlock(&sbi->s_md_lock);
5765 return 0;
5766}
5767
5768/*
5769 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5770 * linearly starting at the goal block and also excludes the blocks which
5771 * are going to be in use after fast commit replay.
5772 */
5773static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5774 struct ext4_allocation_request *ar, int *errp)
5775{
5776 struct buffer_head *bitmap_bh;
5777 struct super_block *sb = ar->inode->i_sb;
5778 ext4_group_t group;
5779 ext4_grpblk_t blkoff;
5780 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5781 ext4_grpblk_t i = 0;
5782 ext4_fsblk_t goal, block;
5783 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5784
5785 goal = ar->goal;
5786 if (goal < le32_to_cpu(es->s_first_data_block) ||
5787 goal >= ext4_blocks_count(es))
5788 goal = le32_to_cpu(es->s_first_data_block);
5789
5790 ar->len = 0;
5791 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5792 for (; group < ext4_get_groups_count(sb); group++) {
5793 bitmap_bh = ext4_read_block_bitmap(sb, group);
5794 if (IS_ERR(bitmap_bh)) {
5795 *errp = PTR_ERR(bitmap_bh);
5796 pr_warn("Failed to read block bitmap\n");
5797 return 0;
5798 }
5799
5800 ext4_get_group_no_and_offset(sb,
5801 max(ext4_group_first_block_no(sb, group), goal),
5802 NULL, &blkoff);
5803 while (1) {
5804 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5805 blkoff);
5806 if (i >= max)
5807 break;
5808 if (ext4_fc_replay_check_excluded(sb,
5809 ext4_group_first_block_no(sb, group) + i)) {
5810 blkoff = i + 1;
5811 } else
5812 break;
5813 }
5814 brelse(bitmap_bh);
5815 if (i < max)
5816 break;
5817 }
5818
5819 if (group >= ext4_get_groups_count(sb) || i >= max) {
5820 *errp = -ENOSPC;
5821 return 0;
5822 }
5823
5824 block = ext4_group_first_block_no(sb, group) + i;
5825 ext4_mb_mark_bb(sb, block, 1, 1);
5826 ar->len = 1;
5827
5828 return block;
5829}
5830
5831static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5832 unsigned long count)
5833{
5834 struct buffer_head *bitmap_bh;
5835 struct super_block *sb = inode->i_sb;
5836 struct ext4_group_desc *gdp;
5837 struct buffer_head *gdp_bh;
5838 ext4_group_t group;
5839 ext4_grpblk_t blkoff;
5840 int already_freed = 0, err, i;
5841
5842 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5843 bitmap_bh = ext4_read_block_bitmap(sb, group);
5844 if (IS_ERR(bitmap_bh)) {
5845 err = PTR_ERR(bitmap_bh);
5846 pr_warn("Failed to read block bitmap\n");
5847 return;
5848 }
5849 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5850 if (!gdp)
5851 return;
5852
5853 for (i = 0; i < count; i++) {
5854 if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5855 already_freed++;
5856 }
5857 mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5858 err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5859 if (err)
5860 return;
5861 ext4_free_group_clusters_set(
5862 sb, gdp, ext4_free_group_clusters(sb, gdp) +
5863 count - already_freed);
5864 ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5865 ext4_group_desc_csum_set(sb, group, gdp);
5866 ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5867 sync_dirty_buffer(bitmap_bh);
5868 sync_dirty_buffer(gdp_bh);
5869 brelse(bitmap_bh);
5870}
5871
5872/**
5873 * ext4_mb_clear_bb() -- helper function for freeing blocks.
5874 * Used by ext4_free_blocks()
5875 * @handle: handle for this transaction
5876 * @inode: inode
5877 * @block: starting physical block to be freed
5878 * @count: number of blocks to be freed
5879 * @flags: flags used by ext4_free_blocks
5880 */
5881static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5882 ext4_fsblk_t block, unsigned long count,
5883 int flags)
5884{
5885 struct buffer_head *bitmap_bh = NULL;
5886 struct super_block *sb = inode->i_sb;
5887 struct ext4_group_desc *gdp;
5888 unsigned int overflow;
5889 ext4_grpblk_t bit;
5890 struct buffer_head *gd_bh;
5891 ext4_group_t block_group;
5892 struct ext4_sb_info *sbi;
5893 struct ext4_buddy e4b;
5894 unsigned int count_clusters;
5895 int err = 0;
5896 int ret;
5897
5898 sbi = EXT4_SB(sb);
5899
5900 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5901 !ext4_inode_block_valid(inode, block, count)) {
5902 ext4_error(sb, "Freeing blocks in system zone - "
5903 "Block = %llu, count = %lu", block, count);
5904 /* err = 0. ext4_std_error should be a no op */
5905 goto error_return;
5906 }
5907 flags |= EXT4_FREE_BLOCKS_VALIDATED;
5908
5909do_more:
5910 overflow = 0;
5911 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5912
5913 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5914 ext4_get_group_info(sb, block_group))))
5915 return;
5916
5917 /*
5918 * Check to see if we are freeing blocks across a group
5919 * boundary.
5920 */
5921 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5922 overflow = EXT4_C2B(sbi, bit) + count -
5923 EXT4_BLOCKS_PER_GROUP(sb);
5924 count -= overflow;
5925 /* The range changed so it's no longer validated */
5926 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
5927 }
5928 count_clusters = EXT4_NUM_B2C(sbi, count);
5929 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5930 if (IS_ERR(bitmap_bh)) {
5931 err = PTR_ERR(bitmap_bh);
5932 bitmap_bh = NULL;
5933 goto error_return;
5934 }
5935 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5936 if (!gdp) {
5937 err = -EIO;
5938 goto error_return;
5939 }
5940
5941 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5942 !ext4_inode_block_valid(inode, block, count)) {
5943 ext4_error(sb, "Freeing blocks in system zone - "
5944 "Block = %llu, count = %lu", block, count);
5945 /* err = 0. ext4_std_error should be a no op */
5946 goto error_return;
5947 }
5948
5949 BUFFER_TRACE(bitmap_bh, "getting write access");
5950 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5951 EXT4_JTR_NONE);
5952 if (err)
5953 goto error_return;
5954
5955 /*
5956 * We are about to modify some metadata. Call the journal APIs
5957 * to unshare ->b_data if a currently-committing transaction is
5958 * using it
5959 */
5960 BUFFER_TRACE(gd_bh, "get_write_access");
5961 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
5962 if (err)
5963 goto error_return;
5964#ifdef AGGRESSIVE_CHECK
5965 {
5966 int i;
5967 for (i = 0; i < count_clusters; i++)
5968 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5969 }
5970#endif
5971 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5972
5973 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5974 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5975 GFP_NOFS|__GFP_NOFAIL);
5976 if (err)
5977 goto error_return;
5978
5979 /*
5980 * We need to make sure we don't reuse the freed block until after the
5981 * transaction is committed. We make an exception if the inode is to be
5982 * written in writeback mode since writeback mode has weak data
5983 * consistency guarantees.
5984 */
5985 if (ext4_handle_valid(handle) &&
5986 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5987 !ext4_should_writeback_data(inode))) {
5988 struct ext4_free_data *new_entry;
5989 /*
5990 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5991 * to fail.
5992 */
5993 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5994 GFP_NOFS|__GFP_NOFAIL);
5995 new_entry->efd_start_cluster = bit;
5996 new_entry->efd_group = block_group;
5997 new_entry->efd_count = count_clusters;
5998 new_entry->efd_tid = handle->h_transaction->t_tid;
5999
6000 ext4_lock_group(sb, block_group);
6001 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6002 ext4_mb_free_metadata(handle, &e4b, new_entry);
6003 } else {
6004 /* need to update group_info->bb_free and bitmap
6005 * with group lock held. generate_buddy look at
6006 * them with group lock_held
6007 */
6008 if (test_opt(sb, DISCARD)) {
6009 err = ext4_issue_discard(sb, block_group, bit, count,
6010 NULL);
6011 if (err && err != -EOPNOTSUPP)
6012 ext4_msg(sb, KERN_WARNING, "discard request in"
6013 " group:%u block:%d count:%lu failed"
6014 " with %d", block_group, bit, count,
6015 err);
6016 } else
6017 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6018
6019 ext4_lock_group(sb, block_group);
6020 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6021 mb_free_blocks(inode, &e4b, bit, count_clusters);
6022 }
6023
6024 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6025 ext4_free_group_clusters_set(sb, gdp, ret);
6026 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6027 ext4_group_desc_csum_set(sb, block_group, gdp);
6028 ext4_unlock_group(sb, block_group);
6029
6030 if (sbi->s_log_groups_per_flex) {
6031 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6032 atomic64_add(count_clusters,
6033 &sbi_array_rcu_deref(sbi, s_flex_groups,
6034 flex_group)->free_clusters);
6035 }
6036
6037 /*
6038 * on a bigalloc file system, defer the s_freeclusters_counter
6039 * update to the caller (ext4_remove_space and friends) so they
6040 * can determine if a cluster freed here should be rereserved
6041 */
6042 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6043 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6044 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6045 percpu_counter_add(&sbi->s_freeclusters_counter,
6046 count_clusters);
6047 }
6048
6049 ext4_mb_unload_buddy(&e4b);
6050
6051 /* We dirtied the bitmap block */
6052 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6053 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6054
6055 /* And the group descriptor block */
6056 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6057 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6058 if (!err)
6059 err = ret;
6060
6061 if (overflow && !err) {
6062 block += count;
6063 count = overflow;
6064 put_bh(bitmap_bh);
6065 /* The range changed so it's no longer validated */
6066 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6067 goto do_more;
6068 }
6069error_return:
6070 brelse(bitmap_bh);
6071 ext4_std_error(sb, err);
6072 return;
6073}
6074
6075/**
6076 * ext4_free_blocks() -- Free given blocks and update quota
6077 * @handle: handle for this transaction
6078 * @inode: inode
6079 * @bh: optional buffer of the block to be freed
6080 * @block: starting physical block to be freed
6081 * @count: number of blocks to be freed
6082 * @flags: flags used by ext4_free_blocks
6083 */
6084void ext4_free_blocks(handle_t *handle, struct inode *inode,
6085 struct buffer_head *bh, ext4_fsblk_t block,
6086 unsigned long count, int flags)
6087{
6088 struct super_block *sb = inode->i_sb;
6089 unsigned int overflow;
6090 struct ext4_sb_info *sbi;
6091
6092 sbi = EXT4_SB(sb);
6093
6094 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6095 ext4_free_blocks_simple(inode, block, count);
6096 return;
6097 }
6098
6099 might_sleep();
6100 if (bh) {
6101 if (block)
6102 BUG_ON(block != bh->b_blocknr);
6103 else
6104 block = bh->b_blocknr;
6105 }
6106
6107 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6108 !ext4_inode_block_valid(inode, block, count)) {
6109 ext4_error(sb, "Freeing blocks not in datazone - "
6110 "block = %llu, count = %lu", block, count);
6111 return;
6112 }
6113 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6114
6115 ext4_debug("freeing block %llu\n", block);
6116 trace_ext4_free_blocks(inode, block, count, flags);
6117
6118 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6119 BUG_ON(count > 1);
6120
6121 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6122 inode, bh, block);
6123 }
6124
6125 /*
6126 * If the extent to be freed does not begin on a cluster
6127 * boundary, we need to deal with partial clusters at the
6128 * beginning and end of the extent. Normally we will free
6129 * blocks at the beginning or the end unless we are explicitly
6130 * requested to avoid doing so.
6131 */
6132 overflow = EXT4_PBLK_COFF(sbi, block);
6133 if (overflow) {
6134 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6135 overflow = sbi->s_cluster_ratio - overflow;
6136 block += overflow;
6137 if (count > overflow)
6138 count -= overflow;
6139 else
6140 return;
6141 } else {
6142 block -= overflow;
6143 count += overflow;
6144 }
6145 /* The range changed so it's no longer validated */
6146 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6147 }
6148 overflow = EXT4_LBLK_COFF(sbi, count);
6149 if (overflow) {
6150 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6151 if (count > overflow)
6152 count -= overflow;
6153 else
6154 return;
6155 } else
6156 count += sbi->s_cluster_ratio - overflow;
6157 /* The range changed so it's no longer validated */
6158 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6159 }
6160
6161 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6162 int i;
6163 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6164
6165 for (i = 0; i < count; i++) {
6166 cond_resched();
6167 if (is_metadata)
6168 bh = sb_find_get_block(inode->i_sb, block + i);
6169 ext4_forget(handle, is_metadata, inode, bh, block + i);
6170 }
6171 }
6172
6173 ext4_mb_clear_bb(handle, inode, block, count, flags);
6174 return;
6175}
6176
6177/**
6178 * ext4_group_add_blocks() -- Add given blocks to an existing group
6179 * @handle: handle to this transaction
6180 * @sb: super block
6181 * @block: start physical block to add to the block group
6182 * @count: number of blocks to free
6183 *
6184 * This marks the blocks as free in the bitmap and buddy.
6185 */
6186int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6187 ext4_fsblk_t block, unsigned long count)
6188{
6189 struct buffer_head *bitmap_bh = NULL;
6190 struct buffer_head *gd_bh;
6191 ext4_group_t block_group;
6192 ext4_grpblk_t bit;
6193 unsigned int i;
6194 struct ext4_group_desc *desc;
6195 struct ext4_sb_info *sbi = EXT4_SB(sb);
6196 struct ext4_buddy e4b;
6197 int err = 0, ret, free_clusters_count;
6198 ext4_grpblk_t clusters_freed;
6199 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6200 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6201 unsigned long cluster_count = last_cluster - first_cluster + 1;
6202
6203 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6204
6205 if (count == 0)
6206 return 0;
6207
6208 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6209 /*
6210 * Check to see if we are freeing blocks across a group
6211 * boundary.
6212 */
6213 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6214 ext4_warning(sb, "too many blocks added to group %u",
6215 block_group);
6216 err = -EINVAL;
6217 goto error_return;
6218 }
6219
6220 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6221 if (IS_ERR(bitmap_bh)) {
6222 err = PTR_ERR(bitmap_bh);
6223 bitmap_bh = NULL;
6224 goto error_return;
6225 }
6226
6227 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6228 if (!desc) {
6229 err = -EIO;
6230 goto error_return;
6231 }
6232
6233 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6234 ext4_error(sb, "Adding blocks in system zones - "
6235 "Block = %llu, count = %lu",
6236 block, count);
6237 err = -EINVAL;
6238 goto error_return;
6239 }
6240
6241 BUFFER_TRACE(bitmap_bh, "getting write access");
6242 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6243 EXT4_JTR_NONE);
6244 if (err)
6245 goto error_return;
6246
6247 /*
6248 * We are about to modify some metadata. Call the journal APIs
6249 * to unshare ->b_data if a currently-committing transaction is
6250 * using it
6251 */
6252 BUFFER_TRACE(gd_bh, "get_write_access");
6253 err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6254 if (err)
6255 goto error_return;
6256
6257 for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6258 BUFFER_TRACE(bitmap_bh, "clear bit");
6259 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6260 ext4_error(sb, "bit already cleared for block %llu",
6261 (ext4_fsblk_t)(block + i));
6262 BUFFER_TRACE(bitmap_bh, "bit already cleared");
6263 } else {
6264 clusters_freed++;
6265 }
6266 }
6267
6268 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6269 if (err)
6270 goto error_return;
6271
6272 /*
6273 * need to update group_info->bb_free and bitmap
6274 * with group lock held. generate_buddy look at
6275 * them with group lock_held
6276 */
6277 ext4_lock_group(sb, block_group);
6278 mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6279 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6280 free_clusters_count = clusters_freed +
6281 ext4_free_group_clusters(sb, desc);
6282 ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6283 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6284 ext4_group_desc_csum_set(sb, block_group, desc);
6285 ext4_unlock_group(sb, block_group);
6286 percpu_counter_add(&sbi->s_freeclusters_counter,
6287 clusters_freed);
6288
6289 if (sbi->s_log_groups_per_flex) {
6290 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6291 atomic64_add(clusters_freed,
6292 &sbi_array_rcu_deref(sbi, s_flex_groups,
6293 flex_group)->free_clusters);
6294 }
6295
6296 ext4_mb_unload_buddy(&e4b);
6297
6298 /* We dirtied the bitmap block */
6299 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6300 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6301
6302 /* And the group descriptor block */
6303 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6304 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6305 if (!err)
6306 err = ret;
6307
6308error_return:
6309 brelse(bitmap_bh);
6310 ext4_std_error(sb, err);
6311 return err;
6312}
6313
6314/**
6315 * ext4_trim_extent -- function to TRIM one single free extent in the group
6316 * @sb: super block for the file system
6317 * @start: starting block of the free extent in the alloc. group
6318 * @count: number of blocks to TRIM
6319 * @e4b: ext4 buddy for the group
6320 *
6321 * Trim "count" blocks starting at "start" in the "group". To assure that no
6322 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6323 * be called with under the group lock.
6324 */
6325static int ext4_trim_extent(struct super_block *sb,
6326 int start, int count, struct ext4_buddy *e4b)
6327__releases(bitlock)
6328__acquires(bitlock)
6329{
6330 struct ext4_free_extent ex;
6331 ext4_group_t group = e4b->bd_group;
6332 int ret = 0;
6333
6334 trace_ext4_trim_extent(sb, group, start, count);
6335
6336 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6337
6338 ex.fe_start = start;
6339 ex.fe_group = group;
6340 ex.fe_len = count;
6341
6342 /*
6343 * Mark blocks used, so no one can reuse them while
6344 * being trimmed.
6345 */
6346 mb_mark_used(e4b, &ex);
6347 ext4_unlock_group(sb, group);
6348 ret = ext4_issue_discard(sb, group, start, count, NULL);
6349 ext4_lock_group(sb, group);
6350 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6351 return ret;
6352}
6353
6354static int ext4_try_to_trim_range(struct super_block *sb,
6355 struct ext4_buddy *e4b, ext4_grpblk_t start,
6356 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6357__acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6358__releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6359{
6360 ext4_grpblk_t next, count, free_count;
6361 void *bitmap;
6362
6363 bitmap = e4b->bd_bitmap;
6364 start = (e4b->bd_info->bb_first_free > start) ?
6365 e4b->bd_info->bb_first_free : start;
6366 count = 0;
6367 free_count = 0;
6368
6369 while (start <= max) {
6370 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6371 if (start > max)
6372 break;
6373 next = mb_find_next_bit(bitmap, max + 1, start);
6374
6375 if ((next - start) >= minblocks) {
6376 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6377
6378 if (ret && ret != -EOPNOTSUPP)
6379 break;
6380 count += next - start;
6381 }
6382 free_count += next - start;
6383 start = next + 1;
6384
6385 if (fatal_signal_pending(current)) {
6386 count = -ERESTARTSYS;
6387 break;
6388 }
6389
6390 if (need_resched()) {
6391 ext4_unlock_group(sb, e4b->bd_group);
6392 cond_resched();
6393 ext4_lock_group(sb, e4b->bd_group);
6394 }
6395
6396 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6397 break;
6398 }
6399
6400 return count;
6401}
6402
6403/**
6404 * ext4_trim_all_free -- function to trim all free space in alloc. group
6405 * @sb: super block for file system
6406 * @group: group to be trimmed
6407 * @start: first group block to examine
6408 * @max: last group block to examine
6409 * @minblocks: minimum extent block count
6410 * @set_trimmed: set the trimmed flag if at least one block is trimmed
6411 *
6412 * ext4_trim_all_free walks through group's block bitmap searching for free
6413 * extents. When the free extent is found, mark it as used in group buddy
6414 * bitmap. Then issue a TRIM command on this extent and free the extent in
6415 * the group buddy bitmap.
6416 */
6417static ext4_grpblk_t
6418ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6419 ext4_grpblk_t start, ext4_grpblk_t max,
6420 ext4_grpblk_t minblocks, bool set_trimmed)
6421{
6422 struct ext4_buddy e4b;
6423 int ret;
6424
6425 trace_ext4_trim_all_free(sb, group, start, max);
6426
6427 ret = ext4_mb_load_buddy(sb, group, &e4b);
6428 if (ret) {
6429 ext4_warning(sb, "Error %d loading buddy information for %u",
6430 ret, group);
6431 return ret;
6432 }
6433
6434 ext4_lock_group(sb, group);
6435
6436 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6437 minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6438 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6439 if (ret >= 0 && set_trimmed)
6440 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6441 } else {
6442 ret = 0;
6443 }
6444
6445 ext4_unlock_group(sb, group);
6446 ext4_mb_unload_buddy(&e4b);
6447
6448 ext4_debug("trimmed %d blocks in the group %d\n",
6449 ret, group);
6450
6451 return ret;
6452}
6453
6454/**
6455 * ext4_trim_fs() -- trim ioctl handle function
6456 * @sb: superblock for filesystem
6457 * @range: fstrim_range structure
6458 *
6459 * start: First Byte to trim
6460 * len: number of Bytes to trim from start
6461 * minlen: minimum extent length in Bytes
6462 * ext4_trim_fs goes through all allocation groups containing Bytes from
6463 * start to start+len. For each such a group ext4_trim_all_free function
6464 * is invoked to trim all free space.
6465 */
6466int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6467{
6468 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6469 struct ext4_group_info *grp;
6470 ext4_group_t group, first_group, last_group;
6471 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6472 uint64_t start, end, minlen, trimmed = 0;
6473 ext4_fsblk_t first_data_blk =
6474 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6475 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6476 bool whole_group, eof = false;
6477 int ret = 0;
6478
6479 start = range->start >> sb->s_blocksize_bits;
6480 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6481 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6482 range->minlen >> sb->s_blocksize_bits);
6483
6484 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6485 start >= max_blks ||
6486 range->len < sb->s_blocksize)
6487 return -EINVAL;
6488 /* No point to try to trim less than discard granularity */
6489 if (range->minlen < discard_granularity) {
6490 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6491 discard_granularity >> sb->s_blocksize_bits);
6492 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6493 goto out;
6494 }
6495 if (end >= max_blks - 1) {
6496 end = max_blks - 1;
6497 eof = true;
6498 }
6499 if (end <= first_data_blk)
6500 goto out;
6501 if (start < first_data_blk)
6502 start = first_data_blk;
6503
6504 /* Determine first and last group to examine based on start and end */
6505 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6506 &first_group, &first_cluster);
6507 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6508 &last_group, &last_cluster);
6509
6510 /* end now represents the last cluster to discard in this group */
6511 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6512 whole_group = true;
6513
6514 for (group = first_group; group <= last_group; group++) {
6515 grp = ext4_get_group_info(sb, group);
6516 /* We only do this if the grp has never been initialized */
6517 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6518 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6519 if (ret)
6520 break;
6521 }
6522
6523 /*
6524 * For all the groups except the last one, last cluster will
6525 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6526 * change it for the last group, note that last_cluster is
6527 * already computed earlier by ext4_get_group_no_and_offset()
6528 */
6529 if (group == last_group) {
6530 end = last_cluster;
6531 whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6532 }
6533 if (grp->bb_free >= minlen) {
6534 cnt = ext4_trim_all_free(sb, group, first_cluster,
6535 end, minlen, whole_group);
6536 if (cnt < 0) {
6537 ret = cnt;
6538 break;
6539 }
6540 trimmed += cnt;
6541 }
6542
6543 /*
6544 * For every group except the first one, we are sure
6545 * that the first cluster to discard will be cluster #0.
6546 */
6547 first_cluster = 0;
6548 }
6549
6550 if (!ret)
6551 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6552
6553out:
6554 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6555 return ret;
6556}
6557
6558/* Iterate all the free extents in the group. */
6559int
6560ext4_mballoc_query_range(
6561 struct super_block *sb,
6562 ext4_group_t group,
6563 ext4_grpblk_t start,
6564 ext4_grpblk_t end,
6565 ext4_mballoc_query_range_fn formatter,
6566 void *priv)
6567{
6568 void *bitmap;
6569 ext4_grpblk_t next;
6570 struct ext4_buddy e4b;
6571 int error;
6572
6573 error = ext4_mb_load_buddy(sb, group, &e4b);
6574 if (error)
6575 return error;
6576 bitmap = e4b.bd_bitmap;
6577
6578 ext4_lock_group(sb, group);
6579
6580 start = (e4b.bd_info->bb_first_free > start) ?
6581 e4b.bd_info->bb_first_free : start;
6582 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6583 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6584
6585 while (start <= end) {
6586 start = mb_find_next_zero_bit(bitmap, end + 1, start);
6587 if (start > end)
6588 break;
6589 next = mb_find_next_bit(bitmap, end + 1, start);
6590
6591 ext4_unlock_group(sb, group);
6592 error = formatter(sb, group, start, next - start, priv);
6593 if (error)
6594 goto out_unload;
6595 ext4_lock_group(sb, group);
6596
6597 start = next + 1;
6598 }
6599
6600 ext4_unlock_group(sb, group);
6601out_unload:
6602 ext4_mb_unload_buddy(&e4b);
6603
6604 return error;
6605}