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 <linux/freezer.h>
20#include <trace/events/ext4.h>
21#include <kunit/static_stub.h>
22
23/*
24 * MUSTDO:
25 * - test ext4_ext_search_left() and ext4_ext_search_right()
26 * - search for metadata in few groups
27 *
28 * TODO v4:
29 * - normalization should take into account whether file is still open
30 * - discard preallocations if no free space left (policy?)
31 * - don't normalize tails
32 * - quota
33 * - reservation for superuser
34 *
35 * TODO v3:
36 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
37 * - track min/max extents in each group for better group selection
38 * - mb_mark_used() may allocate chunk right after splitting buddy
39 * - tree of groups sorted by number of free blocks
40 * - error handling
41 */
42
43/*
44 * The allocation request involve request for multiple number of blocks
45 * near to the goal(block) value specified.
46 *
47 * During initialization phase of the allocator we decide to use the
48 * group preallocation or inode preallocation depending on the size of
49 * the file. The size of the file could be the resulting file size we
50 * would have after allocation, or the current file size, which ever
51 * is larger. If the size is less than sbi->s_mb_stream_request we
52 * select to use the group preallocation. The default value of
53 * s_mb_stream_request is 16 blocks. This can also be tuned via
54 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
55 * terms of number of blocks.
56 *
57 * The main motivation for having small file use group preallocation is to
58 * ensure that we have small files closer together on the disk.
59 *
60 * First stage the allocator looks at the inode prealloc list,
61 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
62 * spaces for this particular inode. The inode prealloc space is
63 * represented as:
64 *
65 * pa_lstart -> the logical start block for this prealloc space
66 * pa_pstart -> the physical start block for this prealloc space
67 * pa_len -> length for this prealloc space (in clusters)
68 * pa_free -> free space available in this prealloc space (in clusters)
69 *
70 * The inode preallocation space is used looking at the _logical_ start
71 * block. If only the logical file block falls within the range of prealloc
72 * space we will consume the particular prealloc space. This makes sure that
73 * we have contiguous physical blocks representing the file blocks
74 *
75 * The important thing to be noted in case of inode prealloc space is that
76 * we don't modify the values associated to inode prealloc space except
77 * pa_free.
78 *
79 * If we are not able to find blocks in the inode prealloc space and if we
80 * have the group allocation flag set then we look at the locality group
81 * prealloc space. These are per CPU prealloc list represented as
82 *
83 * ext4_sb_info.s_locality_groups[smp_processor_id()]
84 *
85 * The reason for having a per cpu locality group is to reduce the contention
86 * between CPUs. It is possible to get scheduled at this point.
87 *
88 * The locality group prealloc space is used looking at whether we have
89 * enough free space (pa_free) within the prealloc space.
90 *
91 * If we can't allocate blocks via inode prealloc or/and locality group
92 * prealloc then we look at the buddy cache. The buddy cache is represented
93 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
94 * mapped to the buddy and bitmap information regarding different
95 * groups. The buddy information is attached to buddy cache inode so that
96 * we can access them through the page cache. The information regarding
97 * each group is loaded via ext4_mb_load_buddy. The information involve
98 * block bitmap and buddy information. The information are stored in the
99 * inode as:
100 *
101 * { page }
102 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
103 *
104 *
105 * one block each for bitmap and buddy information. So for each group we
106 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
107 * blocksize) blocks. So it can have information regarding groups_per_page
108 * which is blocks_per_page/2
109 *
110 * The buddy cache inode is not stored on disk. The inode is thrown
111 * away when the filesystem is unmounted.
112 *
113 * We look for count number of blocks in the buddy cache. If we were able
114 * to locate that many free blocks we return with additional information
115 * regarding rest of the contiguous physical block available
116 *
117 * Before allocating blocks via buddy cache we normalize the request
118 * blocks. This ensure we ask for more blocks that we needed. The extra
119 * blocks that we get after allocation is added to the respective prealloc
120 * list. In case of inode preallocation we follow a list of heuristics
121 * based on file size. This can be found in ext4_mb_normalize_request. If
122 * we are doing a group prealloc we try to normalize the request to
123 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
124 * dependent on the cluster size; for non-bigalloc file systems, it is
125 * 512 blocks. This can be tuned via
126 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * smallest multiple of the stripe value (sbi->s_stripe) which is
130 * greater than the default mb_group_prealloc.
131 *
132 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
133 * structures in two data structures:
134 *
135 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
136 *
137 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
138 *
139 * This is an array of lists where the index in the array represents the
140 * largest free order in the buddy bitmap of the participating group infos of
141 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
142 * number of buddy bitmap orders possible) number of lists. Group-infos are
143 * placed in appropriate lists.
144 *
145 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
146 *
147 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
148 *
149 * This is an array of lists where in the i-th list there are groups with
150 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
151 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
152 * Note that we don't bother with a special list for completely empty groups
153 * so we only have MB_NUM_ORDERS(sb) lists.
154 *
155 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
156 * structures to decide the order in which groups are to be traversed for
157 * fulfilling an allocation request.
158 *
159 * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order
160 * >= the order of the request. We directly look at the largest free order list
161 * in the data structure (1) above where largest_free_order = order of the
162 * request. If that list is empty, we look at remaining list in the increasing
163 * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED
164 * lookup in O(1) time.
165 *
166 * At CR_GOAL_LEN_FAST, we only consider groups where
167 * average fragment size > request size. So, we lookup a group which has average
168 * fragment size just above or equal to request size using our average fragment
169 * size group lists (data structure 2) in O(1) time.
170 *
171 * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied
172 * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in
173 * CR_GOAL_LEN_FAST suggests that there is no BG that has avg
174 * fragment size > goal length. So before falling to the slower
175 * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and
176 * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big
177 * enough average fragment size. This increases the chances of finding a
178 * suitable block group in O(1) time and results in faster allocation at the
179 * cost of reduced size of allocation.
180 *
181 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
182 * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and
183 * CR_GOAL_LEN_FAST phase.
184 *
185 * The regular allocator (using the buddy cache) supports a few tunables.
186 *
187 * /sys/fs/ext4/<partition>/mb_min_to_scan
188 * /sys/fs/ext4/<partition>/mb_max_to_scan
189 * /sys/fs/ext4/<partition>/mb_order2_req
190 * /sys/fs/ext4/<partition>/mb_linear_limit
191 *
192 * The regular allocator uses buddy scan only if the request len is power of
193 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
194 * value of s_mb_order2_reqs can be tuned via
195 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
196 * stripe size (sbi->s_stripe), we try to search for contiguous block in
197 * stripe size. This should result in better allocation on RAID setups. If
198 * not, we search in the specific group using bitmap for best extents. The
199 * tunable min_to_scan and max_to_scan control the behaviour here.
200 * min_to_scan indicate how long the mballoc __must__ look for a best
201 * extent and max_to_scan indicates how long the mballoc __can__ look for a
202 * best extent in the found extents. Searching for the blocks starts with
203 * the group specified as the goal value in allocation context via
204 * ac_g_ex. Each group is first checked based on the criteria whether it
205 * can be used for allocation. ext4_mb_good_group explains how the groups are
206 * checked.
207 *
208 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
209 * get traversed linearly. That may result in subsequent allocations being not
210 * close to each other. And so, the underlying device may get filled up in a
211 * non-linear fashion. While that may not matter on non-rotational devices, for
212 * rotational devices that may result in higher seek times. "mb_linear_limit"
213 * tells mballoc how many groups mballoc should search linearly before
214 * performing consulting above data structures for more efficient lookups. For
215 * non rotational devices, this value defaults to 0 and for rotational devices
216 * this is set to MB_DEFAULT_LINEAR_LIMIT.
217 *
218 * Both the prealloc space are getting populated as above. So for the first
219 * request we will hit the buddy cache which will result in this prealloc
220 * space getting filled. The prealloc space is then later used for the
221 * subsequent request.
222 */
223
224/*
225 * mballoc operates on the following data:
226 * - on-disk bitmap
227 * - in-core buddy (actually includes buddy and bitmap)
228 * - preallocation descriptors (PAs)
229 *
230 * there are two types of preallocations:
231 * - inode
232 * assiged to specific inode and can be used for this inode only.
233 * it describes part of inode's space preallocated to specific
234 * physical blocks. any block from that preallocated can be used
235 * independent. the descriptor just tracks number of blocks left
236 * unused. so, before taking some block from descriptor, one must
237 * make sure corresponded logical block isn't allocated yet. this
238 * also means that freeing any block within descriptor's range
239 * must discard all preallocated blocks.
240 * - locality group
241 * assigned to specific locality group which does not translate to
242 * permanent set of inodes: inode can join and leave group. space
243 * from this type of preallocation can be used for any inode. thus
244 * it's consumed from the beginning to the end.
245 *
246 * relation between them can be expressed as:
247 * in-core buddy = on-disk bitmap + preallocation descriptors
248 *
249 * this mean blocks mballoc considers used are:
250 * - allocated blocks (persistent)
251 * - preallocated blocks (non-persistent)
252 *
253 * consistency in mballoc world means that at any time a block is either
254 * free or used in ALL structures. notice: "any time" should not be read
255 * literally -- time is discrete and delimited by locks.
256 *
257 * to keep it simple, we don't use block numbers, instead we count number of
258 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
259 *
260 * all operations can be expressed as:
261 * - init buddy: buddy = on-disk + PAs
262 * - new PA: buddy += N; PA = N
263 * - use inode PA: on-disk += N; PA -= N
264 * - discard inode PA buddy -= on-disk - PA; PA = 0
265 * - use locality group PA on-disk += N; PA -= N
266 * - discard locality group PA buddy -= PA; PA = 0
267 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
268 * is used in real operation because we can't know actual used
269 * bits from PA, only from on-disk bitmap
270 *
271 * if we follow this strict logic, then all operations above should be atomic.
272 * given some of them can block, we'd have to use something like semaphores
273 * killing performance on high-end SMP hardware. let's try to relax it using
274 * the following knowledge:
275 * 1) if buddy is referenced, it's already initialized
276 * 2) while block is used in buddy and the buddy is referenced,
277 * nobody can re-allocate that block
278 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
279 * bit set and PA claims same block, it's OK. IOW, one can set bit in
280 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
281 * block
282 *
283 * so, now we're building a concurrency table:
284 * - init buddy vs.
285 * - new PA
286 * blocks for PA are allocated in the buddy, buddy must be referenced
287 * until PA is linked to allocation group to avoid concurrent buddy init
288 * - use inode PA
289 * we need to make sure that either on-disk bitmap or PA has uptodate data
290 * given (3) we care that PA-=N operation doesn't interfere with init
291 * - discard inode PA
292 * the simplest way would be to have buddy initialized by the discard
293 * - use locality group PA
294 * again PA-=N must be serialized with init
295 * - discard locality group PA
296 * the simplest way would be to have buddy initialized by the discard
297 * - new PA vs.
298 * - use inode PA
299 * i_data_sem serializes them
300 * - discard inode PA
301 * discard process must wait until PA isn't used by another process
302 * - use locality group PA
303 * some mutex should serialize them
304 * - discard locality group PA
305 * discard process must wait until PA isn't used by another process
306 * - use inode PA
307 * - use inode PA
308 * i_data_sem or another mutex should serializes them
309 * - discard inode PA
310 * discard process must wait until PA isn't used by another process
311 * - use locality group PA
312 * nothing wrong here -- they're different PAs covering different blocks
313 * - discard locality group PA
314 * discard process must wait until PA isn't used by another process
315 *
316 * now we're ready to make few consequences:
317 * - PA is referenced and while it is no discard is possible
318 * - PA is referenced until block isn't marked in on-disk bitmap
319 * - PA changes only after on-disk bitmap
320 * - discard must not compete with init. either init is done before
321 * any discard or they're serialized somehow
322 * - buddy init as sum of on-disk bitmap and PAs is done atomically
323 *
324 * a special case when we've used PA to emptiness. no need to modify buddy
325 * in this case, but we should care about concurrent init
326 *
327 */
328
329 /*
330 * Logic in few words:
331 *
332 * - allocation:
333 * load group
334 * find blocks
335 * mark bits in on-disk bitmap
336 * release group
337 *
338 * - use preallocation:
339 * find proper PA (per-inode or group)
340 * load group
341 * mark bits in on-disk bitmap
342 * release group
343 * release PA
344 *
345 * - free:
346 * load group
347 * mark bits in on-disk bitmap
348 * release group
349 *
350 * - discard preallocations in group:
351 * mark PAs deleted
352 * move them onto local list
353 * load on-disk bitmap
354 * load group
355 * remove PA from object (inode or locality group)
356 * mark free blocks in-core
357 *
358 * - discard inode's preallocations:
359 */
360
361/*
362 * Locking rules
363 *
364 * Locks:
365 * - bitlock on a group (group)
366 * - object (inode/locality) (object)
367 * - per-pa lock (pa)
368 * - cr_power2_aligned lists lock (cr_power2_aligned)
369 * - cr_goal_len_fast lists lock (cr_goal_len_fast)
370 *
371 * Paths:
372 * - new pa
373 * object
374 * group
375 *
376 * - find and use pa:
377 * pa
378 *
379 * - release consumed pa:
380 * pa
381 * group
382 * object
383 *
384 * - generate in-core bitmap:
385 * group
386 * pa
387 *
388 * - discard all for given object (inode, locality group):
389 * object
390 * pa
391 * group
392 *
393 * - discard all for given group:
394 * group
395 * pa
396 * group
397 * object
398 *
399 * - allocation path (ext4_mb_regular_allocator)
400 * group
401 * cr_power2_aligned/cr_goal_len_fast
402 */
403static struct kmem_cache *ext4_pspace_cachep;
404static struct kmem_cache *ext4_ac_cachep;
405static struct kmem_cache *ext4_free_data_cachep;
406
407/* We create slab caches for groupinfo data structures based on the
408 * superblock block size. There will be one per mounted filesystem for
409 * each unique s_blocksize_bits */
410#define NR_GRPINFO_CACHES 8
411static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
412
413static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
414 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
415 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
416 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
417};
418
419static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
420 ext4_group_t group);
421static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
422
423static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
424 ext4_group_t group, enum criteria cr);
425
426static int ext4_try_to_trim_range(struct super_block *sb,
427 struct ext4_buddy *e4b, ext4_grpblk_t start,
428 ext4_grpblk_t max, ext4_grpblk_t minblocks);
429
430/*
431 * The algorithm using this percpu seq counter goes below:
432 * 1. We sample the percpu discard_pa_seq counter before trying for block
433 * allocation in ext4_mb_new_blocks().
434 * 2. We increment this percpu discard_pa_seq counter when we either allocate
435 * or free these blocks i.e. while marking those blocks as used/free in
436 * mb_mark_used()/mb_free_blocks().
437 * 3. We also increment this percpu seq counter when we successfully identify
438 * that the bb_prealloc_list is not empty and hence proceed for discarding
439 * of those PAs inside ext4_mb_discard_group_preallocations().
440 *
441 * Now to make sure that the regular fast path of block allocation is not
442 * affected, as a small optimization we only sample the percpu seq counter
443 * on that cpu. Only when the block allocation fails and when freed blocks
444 * found were 0, that is when we sample percpu seq counter for all cpus using
445 * below function ext4_get_discard_pa_seq_sum(). This happens after making
446 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
447 */
448static DEFINE_PER_CPU(u64, discard_pa_seq);
449static inline u64 ext4_get_discard_pa_seq_sum(void)
450{
451 int __cpu;
452 u64 __seq = 0;
453
454 for_each_possible_cpu(__cpu)
455 __seq += per_cpu(discard_pa_seq, __cpu);
456 return __seq;
457}
458
459static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
460{
461#if BITS_PER_LONG == 64
462 *bit += ((unsigned long) addr & 7UL) << 3;
463 addr = (void *) ((unsigned long) addr & ~7UL);
464#elif BITS_PER_LONG == 32
465 *bit += ((unsigned long) addr & 3UL) << 3;
466 addr = (void *) ((unsigned long) addr & ~3UL);
467#else
468#error "how many bits you are?!"
469#endif
470 return addr;
471}
472
473static inline int mb_test_bit(int bit, void *addr)
474{
475 /*
476 * ext4_test_bit on architecture like powerpc
477 * needs unsigned long aligned address
478 */
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 return ext4_test_bit(bit, addr);
481}
482
483static inline void mb_set_bit(int bit, void *addr)
484{
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 ext4_set_bit(bit, addr);
487}
488
489static inline void mb_clear_bit(int bit, void *addr)
490{
491 addr = mb_correct_addr_and_bit(&bit, addr);
492 ext4_clear_bit(bit, addr);
493}
494
495static inline int mb_test_and_clear_bit(int bit, void *addr)
496{
497 addr = mb_correct_addr_and_bit(&bit, addr);
498 return ext4_test_and_clear_bit(bit, addr);
499}
500
501static inline int mb_find_next_zero_bit(void *addr, int max, int start)
502{
503 int fix = 0, ret, tmpmax;
504 addr = mb_correct_addr_and_bit(&fix, addr);
505 tmpmax = max + fix;
506 start += fix;
507
508 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
509 if (ret > max)
510 return max;
511 return ret;
512}
513
514static inline int mb_find_next_bit(void *addr, int max, int start)
515{
516 int fix = 0, ret, tmpmax;
517 addr = mb_correct_addr_and_bit(&fix, addr);
518 tmpmax = max + fix;
519 start += fix;
520
521 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
522 if (ret > max)
523 return max;
524 return ret;
525}
526
527static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
528{
529 char *bb;
530
531 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
532 BUG_ON(max == NULL);
533
534 if (order > e4b->bd_blkbits + 1) {
535 *max = 0;
536 return NULL;
537 }
538
539 /* at order 0 we see each particular block */
540 if (order == 0) {
541 *max = 1 << (e4b->bd_blkbits + 3);
542 return e4b->bd_bitmap;
543 }
544
545 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
546 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
547
548 return bb;
549}
550
551#ifdef DOUBLE_CHECK
552static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
553 int first, int count)
554{
555 int i;
556 struct super_block *sb = e4b->bd_sb;
557
558 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
559 return;
560 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
561 for (i = 0; i < count; i++) {
562 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
563 ext4_fsblk_t blocknr;
564
565 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
566 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
567 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
568 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
569 ext4_grp_locked_error(sb, e4b->bd_group,
570 inode ? inode->i_ino : 0,
571 blocknr,
572 "freeing block already freed "
573 "(bit %u)",
574 first + i);
575 }
576 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
577 }
578}
579
580static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
581{
582 int i;
583
584 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
585 return;
586 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
587 for (i = 0; i < count; i++) {
588 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
589 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
590 }
591}
592
593static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
594{
595 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
596 return;
597 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
598 unsigned char *b1, *b2;
599 int i;
600 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
601 b2 = (unsigned char *) bitmap;
602 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
603 if (b1[i] != b2[i]) {
604 ext4_msg(e4b->bd_sb, KERN_ERR,
605 "corruption in group %u "
606 "at byte %u(%u): %x in copy != %x "
607 "on disk/prealloc",
608 e4b->bd_group, i, i * 8, b1[i], b2[i]);
609 BUG();
610 }
611 }
612 }
613}
614
615static void mb_group_bb_bitmap_alloc(struct super_block *sb,
616 struct ext4_group_info *grp, ext4_group_t group)
617{
618 struct buffer_head *bh;
619
620 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
621 if (!grp->bb_bitmap)
622 return;
623
624 bh = ext4_read_block_bitmap(sb, group);
625 if (IS_ERR_OR_NULL(bh)) {
626 kfree(grp->bb_bitmap);
627 grp->bb_bitmap = NULL;
628 return;
629 }
630
631 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
632 put_bh(bh);
633}
634
635static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
636{
637 kfree(grp->bb_bitmap);
638}
639
640#else
641static inline void mb_free_blocks_double(struct inode *inode,
642 struct ext4_buddy *e4b, int first, int count)
643{
644 return;
645}
646static inline void mb_mark_used_double(struct ext4_buddy *e4b,
647 int first, int count)
648{
649 return;
650}
651static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
652{
653 return;
654}
655
656static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
657 struct ext4_group_info *grp, ext4_group_t group)
658{
659 return;
660}
661
662static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
663{
664 return;
665}
666#endif
667
668#ifdef AGGRESSIVE_CHECK
669
670#define MB_CHECK_ASSERT(assert) \
671do { \
672 if (!(assert)) { \
673 printk(KERN_EMERG \
674 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
675 function, file, line, # assert); \
676 BUG(); \
677 } \
678} while (0)
679
680static void __mb_check_buddy(struct ext4_buddy *e4b, char *file,
681 const char *function, int line)
682{
683 struct super_block *sb = e4b->bd_sb;
684 int order = e4b->bd_blkbits + 1;
685 int max;
686 int max2;
687 int i;
688 int j;
689 int k;
690 int count;
691 struct ext4_group_info *grp;
692 int fragments = 0;
693 int fstart;
694 struct list_head *cur;
695 void *buddy;
696 void *buddy2;
697
698 if (e4b->bd_info->bb_check_counter++ % 10)
699 return;
700
701 while (order > 1) {
702 buddy = mb_find_buddy(e4b, order, &max);
703 MB_CHECK_ASSERT(buddy);
704 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
705 MB_CHECK_ASSERT(buddy2);
706 MB_CHECK_ASSERT(buddy != buddy2);
707 MB_CHECK_ASSERT(max * 2 == max2);
708
709 count = 0;
710 for (i = 0; i < max; i++) {
711
712 if (mb_test_bit(i, buddy)) {
713 /* only single bit in buddy2 may be 0 */
714 if (!mb_test_bit(i << 1, buddy2)) {
715 MB_CHECK_ASSERT(
716 mb_test_bit((i<<1)+1, buddy2));
717 }
718 continue;
719 }
720
721 /* both bits in buddy2 must be 1 */
722 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
723 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
724
725 for (j = 0; j < (1 << order); j++) {
726 k = (i * (1 << order)) + j;
727 MB_CHECK_ASSERT(
728 !mb_test_bit(k, e4b->bd_bitmap));
729 }
730 count++;
731 }
732 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
733 order--;
734 }
735
736 fstart = -1;
737 buddy = mb_find_buddy(e4b, 0, &max);
738 for (i = 0; i < max; i++) {
739 if (!mb_test_bit(i, buddy)) {
740 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
741 if (fstart == -1) {
742 fragments++;
743 fstart = i;
744 }
745 continue;
746 }
747 fstart = -1;
748 /* check used bits only */
749 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
750 buddy2 = mb_find_buddy(e4b, j, &max2);
751 k = i >> j;
752 MB_CHECK_ASSERT(k < max2);
753 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
754 }
755 }
756 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
757 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
758
759 grp = ext4_get_group_info(sb, e4b->bd_group);
760 if (!grp)
761 return;
762 list_for_each(cur, &grp->bb_prealloc_list) {
763 ext4_group_t groupnr;
764 struct ext4_prealloc_space *pa;
765 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
766 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
767 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
768 for (i = 0; i < pa->pa_len; i++)
769 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
770 }
771}
772#undef MB_CHECK_ASSERT
773#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
774 __FILE__, __func__, __LINE__)
775#else
776#define mb_check_buddy(e4b)
777#endif
778
779/*
780 * Divide blocks started from @first with length @len into
781 * smaller chunks with power of 2 blocks.
782 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
783 * then increase bb_counters[] for corresponded chunk size.
784 */
785static void ext4_mb_mark_free_simple(struct super_block *sb,
786 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
787 struct ext4_group_info *grp)
788{
789 struct ext4_sb_info *sbi = EXT4_SB(sb);
790 ext4_grpblk_t min;
791 ext4_grpblk_t max;
792 ext4_grpblk_t chunk;
793 unsigned int border;
794
795 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
796
797 border = 2 << sb->s_blocksize_bits;
798
799 while (len > 0) {
800 /* find how many blocks can be covered since this position */
801 max = ffs(first | border) - 1;
802
803 /* find how many blocks of power 2 we need to mark */
804 min = fls(len) - 1;
805
806 if (max < min)
807 min = max;
808 chunk = 1 << min;
809
810 /* mark multiblock chunks only */
811 grp->bb_counters[min]++;
812 if (min > 0)
813 mb_clear_bit(first >> min,
814 buddy + sbi->s_mb_offsets[min]);
815
816 len -= chunk;
817 first += chunk;
818 }
819}
820
821static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
822{
823 int order;
824
825 /*
826 * We don't bother with a special lists groups with only 1 block free
827 * extents and for completely empty groups.
828 */
829 order = fls(len) - 2;
830 if (order < 0)
831 return 0;
832 if (order == MB_NUM_ORDERS(sb))
833 order--;
834 if (WARN_ON_ONCE(order > MB_NUM_ORDERS(sb)))
835 order = MB_NUM_ORDERS(sb) - 1;
836 return order;
837}
838
839/* Move group to appropriate avg_fragment_size list */
840static void
841mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
842{
843 struct ext4_sb_info *sbi = EXT4_SB(sb);
844 int new_order;
845
846 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 0)
847 return;
848
849 new_order = mb_avg_fragment_size_order(sb,
850 grp->bb_free / grp->bb_fragments);
851 if (new_order == grp->bb_avg_fragment_size_order)
852 return;
853
854 if (grp->bb_avg_fragment_size_order != -1) {
855 write_lock(&sbi->s_mb_avg_fragment_size_locks[
856 grp->bb_avg_fragment_size_order]);
857 list_del(&grp->bb_avg_fragment_size_node);
858 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
859 grp->bb_avg_fragment_size_order]);
860 }
861 grp->bb_avg_fragment_size_order = new_order;
862 write_lock(&sbi->s_mb_avg_fragment_size_locks[
863 grp->bb_avg_fragment_size_order]);
864 list_add_tail(&grp->bb_avg_fragment_size_node,
865 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
866 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
867 grp->bb_avg_fragment_size_order]);
868}
869
870/*
871 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
872 * cr level needs an update.
873 */
874static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac,
875 enum criteria *new_cr, ext4_group_t *group)
876{
877 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
878 struct ext4_group_info *iter;
879 int i;
880
881 if (ac->ac_status == AC_STATUS_FOUND)
882 return;
883
884 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED))
885 atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions);
886
887 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
888 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
889 continue;
890 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
891 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
892 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
893 continue;
894 }
895 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
896 bb_largest_free_order_node) {
897 if (sbi->s_mb_stats)
898 atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]);
899 if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) {
900 *group = iter->bb_group;
901 ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED;
902 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
903 return;
904 }
905 }
906 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
907 }
908
909 /* Increment cr and search again if no group is found */
910 *new_cr = CR_GOAL_LEN_FAST;
911}
912
913/*
914 * Find a suitable group of given order from the average fragments list.
915 */
916static struct ext4_group_info *
917ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order)
918{
919 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
920 struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order];
921 rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order];
922 struct ext4_group_info *grp = NULL, *iter;
923 enum criteria cr = ac->ac_criteria;
924
925 if (list_empty(frag_list))
926 return NULL;
927 read_lock(frag_list_lock);
928 if (list_empty(frag_list)) {
929 read_unlock(frag_list_lock);
930 return NULL;
931 }
932 list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) {
933 if (sbi->s_mb_stats)
934 atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]);
935 if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) {
936 grp = iter;
937 break;
938 }
939 }
940 read_unlock(frag_list_lock);
941 return grp;
942}
943
944/*
945 * Choose next group by traversing average fragment size list of suitable
946 * order. Updates *new_cr if cr level needs an update.
947 */
948static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac,
949 enum criteria *new_cr, ext4_group_t *group)
950{
951 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
952 struct ext4_group_info *grp = NULL;
953 int i;
954
955 if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) {
956 if (sbi->s_mb_stats)
957 atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions);
958 }
959
960 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
961 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
962 grp = ext4_mb_find_good_group_avg_frag_lists(ac, i);
963 if (grp) {
964 *group = grp->bb_group;
965 ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED;
966 return;
967 }
968 }
969
970 /*
971 * CR_BEST_AVAIL_LEN works based on the concept that we have
972 * a larger normalized goal len request which can be trimmed to
973 * a smaller goal len such that it can still satisfy original
974 * request len. However, allocation request for non-regular
975 * files never gets normalized.
976 * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA).
977 */
978 if (ac->ac_flags & EXT4_MB_HINT_DATA)
979 *new_cr = CR_BEST_AVAIL_LEN;
980 else
981 *new_cr = CR_GOAL_LEN_SLOW;
982}
983
984/*
985 * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment
986 * order we have and proactively trim the goal request length to that order to
987 * find a suitable group faster.
988 *
989 * This optimizes allocation speed at the cost of slightly reduced
990 * preallocations. However, we make sure that we don't trim the request too
991 * much and fall to CR_GOAL_LEN_SLOW in that case.
992 */
993static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac,
994 enum criteria *new_cr, ext4_group_t *group)
995{
996 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
997 struct ext4_group_info *grp = NULL;
998 int i, order, min_order;
999 unsigned long num_stripe_clusters = 0;
1000
1001 if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) {
1002 if (sbi->s_mb_stats)
1003 atomic_inc(&sbi->s_bal_best_avail_bad_suggestions);
1004 }
1005
1006 /*
1007 * mb_avg_fragment_size_order() returns order in a way that makes
1008 * retrieving back the length using (1 << order) inaccurate. Hence, use
1009 * fls() instead since we need to know the actual length while modifying
1010 * goal length.
1011 */
1012 order = fls(ac->ac_g_ex.fe_len) - 1;
1013 if (WARN_ON_ONCE(order - 1 > MB_NUM_ORDERS(ac->ac_sb)))
1014 order = MB_NUM_ORDERS(ac->ac_sb);
1015 min_order = order - sbi->s_mb_best_avail_max_trim_order;
1016 if (min_order < 0)
1017 min_order = 0;
1018
1019 if (sbi->s_stripe > 0) {
1020 /*
1021 * We are assuming that stripe size is always a multiple of
1022 * cluster ratio otherwise __ext4_fill_super exists early.
1023 */
1024 num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe);
1025 if (1 << min_order < num_stripe_clusters)
1026 /*
1027 * We consider 1 order less because later we round
1028 * up the goal len to num_stripe_clusters
1029 */
1030 min_order = fls(num_stripe_clusters) - 1;
1031 }
1032
1033 if (1 << min_order < ac->ac_o_ex.fe_len)
1034 min_order = fls(ac->ac_o_ex.fe_len);
1035
1036 for (i = order; i >= min_order; i--) {
1037 int frag_order;
1038 /*
1039 * Scale down goal len to make sure we find something
1040 * in the free fragments list. Basically, reduce
1041 * preallocations.
1042 */
1043 ac->ac_g_ex.fe_len = 1 << i;
1044
1045 if (num_stripe_clusters > 0) {
1046 /*
1047 * Try to round up the adjusted goal length to
1048 * stripe size (in cluster units) multiple for
1049 * efficiency.
1050 */
1051 ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len,
1052 num_stripe_clusters);
1053 }
1054
1055 frag_order = mb_avg_fragment_size_order(ac->ac_sb,
1056 ac->ac_g_ex.fe_len);
1057
1058 grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order);
1059 if (grp) {
1060 *group = grp->bb_group;
1061 ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED;
1062 return;
1063 }
1064 }
1065
1066 /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */
1067 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
1068 *new_cr = CR_GOAL_LEN_SLOW;
1069}
1070
1071static inline int should_optimize_scan(struct ext4_allocation_context *ac)
1072{
1073 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
1074 return 0;
1075 if (ac->ac_criteria >= CR_GOAL_LEN_SLOW)
1076 return 0;
1077 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1078 return 0;
1079 return 1;
1080}
1081
1082/*
1083 * Return next linear group for allocation.
1084 */
1085static ext4_group_t
1086next_linear_group(ext4_group_t group, ext4_group_t ngroups)
1087{
1088 /*
1089 * Artificially restricted ngroups for non-extent
1090 * files makes group > ngroups possible on first loop.
1091 */
1092 return group + 1 >= ngroups ? 0 : group + 1;
1093}
1094
1095/*
1096 * ext4_mb_choose_next_group: choose next group for allocation.
1097 *
1098 * @ac Allocation Context
1099 * @new_cr This is an output parameter. If the there is no good group
1100 * available at current CR level, this field is updated to indicate
1101 * the new cr level that should be used.
1102 * @group This is an input / output parameter. As an input it indicates the
1103 * next group that the allocator intends to use for allocation. As
1104 * output, this field indicates the next group that should be used as
1105 * determined by the optimization functions.
1106 * @ngroups Total number of groups
1107 */
1108static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1109 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1110{
1111 *new_cr = ac->ac_criteria;
1112
1113 if (!should_optimize_scan(ac)) {
1114 *group = next_linear_group(*group, ngroups);
1115 return;
1116 }
1117
1118 /*
1119 * Optimized scanning can return non adjacent groups which can cause
1120 * seek overhead for rotational disks. So try few linear groups before
1121 * trying optimized scan.
1122 */
1123 if (ac->ac_groups_linear_remaining) {
1124 *group = next_linear_group(*group, ngroups);
1125 ac->ac_groups_linear_remaining--;
1126 return;
1127 }
1128
1129 if (*new_cr == CR_POWER2_ALIGNED) {
1130 ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group);
1131 } else if (*new_cr == CR_GOAL_LEN_FAST) {
1132 ext4_mb_choose_next_group_goal_fast(ac, new_cr, group);
1133 } else if (*new_cr == CR_BEST_AVAIL_LEN) {
1134 ext4_mb_choose_next_group_best_avail(ac, new_cr, group);
1135 } else {
1136 /*
1137 * TODO: For CR_GOAL_LEN_SLOW, we can arrange groups in an
1138 * rb tree sorted by bb_free. But until that happens, we should
1139 * never come here.
1140 */
1141 WARN_ON(1);
1142 }
1143}
1144
1145/*
1146 * Cache the order of the largest free extent we have available in this block
1147 * group.
1148 */
1149static void
1150mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1151{
1152 struct ext4_sb_info *sbi = EXT4_SB(sb);
1153 int i;
1154
1155 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1156 if (grp->bb_counters[i] > 0)
1157 break;
1158 /* No need to move between order lists? */
1159 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1160 i == grp->bb_largest_free_order) {
1161 grp->bb_largest_free_order = i;
1162 return;
1163 }
1164
1165 if (grp->bb_largest_free_order >= 0) {
1166 write_lock(&sbi->s_mb_largest_free_orders_locks[
1167 grp->bb_largest_free_order]);
1168 list_del_init(&grp->bb_largest_free_order_node);
1169 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1170 grp->bb_largest_free_order]);
1171 }
1172 grp->bb_largest_free_order = i;
1173 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1174 write_lock(&sbi->s_mb_largest_free_orders_locks[
1175 grp->bb_largest_free_order]);
1176 list_add_tail(&grp->bb_largest_free_order_node,
1177 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1178 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1179 grp->bb_largest_free_order]);
1180 }
1181}
1182
1183static noinline_for_stack
1184void ext4_mb_generate_buddy(struct super_block *sb,
1185 void *buddy, void *bitmap, ext4_group_t group,
1186 struct ext4_group_info *grp)
1187{
1188 struct ext4_sb_info *sbi = EXT4_SB(sb);
1189 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1190 ext4_grpblk_t i = 0;
1191 ext4_grpblk_t first;
1192 ext4_grpblk_t len;
1193 unsigned free = 0;
1194 unsigned fragments = 0;
1195 unsigned long long period = get_cycles();
1196
1197 /* initialize buddy from bitmap which is aggregation
1198 * of on-disk bitmap and preallocations */
1199 i = mb_find_next_zero_bit(bitmap, max, 0);
1200 grp->bb_first_free = i;
1201 while (i < max) {
1202 fragments++;
1203 first = i;
1204 i = mb_find_next_bit(bitmap, max, i);
1205 len = i - first;
1206 free += len;
1207 if (len > 1)
1208 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1209 else
1210 grp->bb_counters[0]++;
1211 if (i < max)
1212 i = mb_find_next_zero_bit(bitmap, max, i);
1213 }
1214 grp->bb_fragments = fragments;
1215
1216 if (free != grp->bb_free) {
1217 ext4_grp_locked_error(sb, group, 0, 0,
1218 "block bitmap and bg descriptor "
1219 "inconsistent: %u vs %u free clusters",
1220 free, grp->bb_free);
1221 /*
1222 * If we intend to continue, we consider group descriptor
1223 * corrupt and update bb_free using bitmap value
1224 */
1225 grp->bb_free = free;
1226 ext4_mark_group_bitmap_corrupted(sb, group,
1227 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1228 }
1229 mb_set_largest_free_order(sb, grp);
1230 mb_update_avg_fragment_size(sb, grp);
1231
1232 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1233
1234 period = get_cycles() - period;
1235 atomic_inc(&sbi->s_mb_buddies_generated);
1236 atomic64_add(period, &sbi->s_mb_generation_time);
1237}
1238
1239static void mb_regenerate_buddy(struct ext4_buddy *e4b)
1240{
1241 int count;
1242 int order = 1;
1243 void *buddy;
1244
1245 while ((buddy = mb_find_buddy(e4b, order++, &count)))
1246 mb_set_bits(buddy, 0, count);
1247
1248 e4b->bd_info->bb_fragments = 0;
1249 memset(e4b->bd_info->bb_counters, 0,
1250 sizeof(*e4b->bd_info->bb_counters) *
1251 (e4b->bd_sb->s_blocksize_bits + 2));
1252
1253 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
1254 e4b->bd_bitmap, e4b->bd_group, e4b->bd_info);
1255}
1256
1257/* The buddy information is attached the buddy cache inode
1258 * for convenience. The information regarding each group
1259 * is loaded via ext4_mb_load_buddy. The information involve
1260 * block bitmap and buddy information. The information are
1261 * stored in the inode as
1262 *
1263 * { page }
1264 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1265 *
1266 *
1267 * one block each for bitmap and buddy information.
1268 * So for each group we take up 2 blocks. A page can
1269 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1270 * So it can have information regarding groups_per_page which
1271 * is blocks_per_page/2
1272 *
1273 * Locking note: This routine takes the block group lock of all groups
1274 * for this page; do not hold this lock when calling this routine!
1275 */
1276
1277static int ext4_mb_init_cache(struct folio *folio, char *incore, gfp_t gfp)
1278{
1279 ext4_group_t ngroups;
1280 unsigned int blocksize;
1281 int blocks_per_page;
1282 int groups_per_page;
1283 int err = 0;
1284 int i;
1285 ext4_group_t first_group, group;
1286 int first_block;
1287 struct super_block *sb;
1288 struct buffer_head *bhs;
1289 struct buffer_head **bh = NULL;
1290 struct inode *inode;
1291 char *data;
1292 char *bitmap;
1293 struct ext4_group_info *grinfo;
1294
1295 inode = folio->mapping->host;
1296 sb = inode->i_sb;
1297 ngroups = ext4_get_groups_count(sb);
1298 blocksize = i_blocksize(inode);
1299 blocks_per_page = PAGE_SIZE / blocksize;
1300
1301 mb_debug(sb, "init folio %lu\n", folio->index);
1302
1303 groups_per_page = blocks_per_page >> 1;
1304 if (groups_per_page == 0)
1305 groups_per_page = 1;
1306
1307 /* allocate buffer_heads to read bitmaps */
1308 if (groups_per_page > 1) {
1309 i = sizeof(struct buffer_head *) * groups_per_page;
1310 bh = kzalloc(i, gfp);
1311 if (bh == NULL)
1312 return -ENOMEM;
1313 } else
1314 bh = &bhs;
1315
1316 first_group = folio->index * blocks_per_page / 2;
1317
1318 /* read all groups the folio covers into the cache */
1319 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1320 if (group >= ngroups)
1321 break;
1322
1323 grinfo = ext4_get_group_info(sb, group);
1324 if (!grinfo)
1325 continue;
1326 /*
1327 * If page is uptodate then we came here after online resize
1328 * which added some new uninitialized group info structs, so
1329 * we must skip all initialized uptodate buddies on the folio,
1330 * which may be currently in use by an allocating task.
1331 */
1332 if (folio_test_uptodate(folio) &&
1333 !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1334 bh[i] = NULL;
1335 continue;
1336 }
1337 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1338 if (IS_ERR(bh[i])) {
1339 err = PTR_ERR(bh[i]);
1340 bh[i] = NULL;
1341 goto out;
1342 }
1343 mb_debug(sb, "read bitmap for group %u\n", group);
1344 }
1345
1346 /* wait for I/O completion */
1347 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1348 int err2;
1349
1350 if (!bh[i])
1351 continue;
1352 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1353 if (!err)
1354 err = err2;
1355 }
1356
1357 first_block = folio->index * blocks_per_page;
1358 for (i = 0; i < blocks_per_page; i++) {
1359 group = (first_block + i) >> 1;
1360 if (group >= ngroups)
1361 break;
1362
1363 if (!bh[group - first_group])
1364 /* skip initialized uptodate buddy */
1365 continue;
1366
1367 if (!buffer_verified(bh[group - first_group]))
1368 /* Skip faulty bitmaps */
1369 continue;
1370 err = 0;
1371
1372 /*
1373 * data carry information regarding this
1374 * particular group in the format specified
1375 * above
1376 *
1377 */
1378 data = folio_address(folio) + (i * blocksize);
1379 bitmap = bh[group - first_group]->b_data;
1380
1381 /*
1382 * We place the buddy block and bitmap block
1383 * close together
1384 */
1385 grinfo = ext4_get_group_info(sb, group);
1386 if (!grinfo) {
1387 err = -EFSCORRUPTED;
1388 goto out;
1389 }
1390 if ((first_block + i) & 1) {
1391 /* this is block of buddy */
1392 BUG_ON(incore == NULL);
1393 mb_debug(sb, "put buddy for group %u in folio %lu/%x\n",
1394 group, folio->index, i * blocksize);
1395 trace_ext4_mb_buddy_bitmap_load(sb, group);
1396 grinfo->bb_fragments = 0;
1397 memset(grinfo->bb_counters, 0,
1398 sizeof(*grinfo->bb_counters) *
1399 (MB_NUM_ORDERS(sb)));
1400 /*
1401 * incore got set to the group block bitmap below
1402 */
1403 ext4_lock_group(sb, group);
1404 /* init the buddy */
1405 memset(data, 0xff, blocksize);
1406 ext4_mb_generate_buddy(sb, data, incore, group, grinfo);
1407 ext4_unlock_group(sb, group);
1408 incore = NULL;
1409 } else {
1410 /* this is block of bitmap */
1411 BUG_ON(incore != NULL);
1412 mb_debug(sb, "put bitmap for group %u in folio %lu/%x\n",
1413 group, folio->index, i * blocksize);
1414 trace_ext4_mb_bitmap_load(sb, group);
1415
1416 /* see comments in ext4_mb_put_pa() */
1417 ext4_lock_group(sb, group);
1418 memcpy(data, bitmap, blocksize);
1419
1420 /* mark all preallocated blks used in in-core bitmap */
1421 ext4_mb_generate_from_pa(sb, data, group);
1422 WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root));
1423 ext4_unlock_group(sb, group);
1424
1425 /* set incore so that the buddy information can be
1426 * generated using this
1427 */
1428 incore = data;
1429 }
1430 }
1431 folio_mark_uptodate(folio);
1432
1433out:
1434 if (bh) {
1435 for (i = 0; i < groups_per_page; i++)
1436 brelse(bh[i]);
1437 if (bh != &bhs)
1438 kfree(bh);
1439 }
1440 return err;
1441}
1442
1443/*
1444 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1445 * on the same buddy page doesn't happen whild holding the buddy page lock.
1446 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1447 * are on the same page e4b->bd_buddy_folio is NULL and return value is 0.
1448 */
1449static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1450 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1451{
1452 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1453 int block, pnum, poff;
1454 int blocks_per_page;
1455 struct folio *folio;
1456
1457 e4b->bd_buddy_folio = NULL;
1458 e4b->bd_bitmap_folio = NULL;
1459
1460 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1461 /*
1462 * the buddy cache inode stores the block bitmap
1463 * and buddy information in consecutive blocks.
1464 * So for each group we need two blocks.
1465 */
1466 block = group * 2;
1467 pnum = block / blocks_per_page;
1468 poff = block % blocks_per_page;
1469 folio = __filemap_get_folio(inode->i_mapping, pnum,
1470 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1471 if (IS_ERR(folio))
1472 return PTR_ERR(folio);
1473 BUG_ON(folio->mapping != inode->i_mapping);
1474 e4b->bd_bitmap_folio = folio;
1475 e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize);
1476
1477 if (blocks_per_page >= 2) {
1478 /* buddy and bitmap are on the same page */
1479 return 0;
1480 }
1481
1482 /* blocks_per_page == 1, hence we need another page for the buddy */
1483 folio = __filemap_get_folio(inode->i_mapping, block + 1,
1484 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1485 if (IS_ERR(folio))
1486 return PTR_ERR(folio);
1487 BUG_ON(folio->mapping != inode->i_mapping);
1488 e4b->bd_buddy_folio = folio;
1489 return 0;
1490}
1491
1492static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1493{
1494 if (e4b->bd_bitmap_folio) {
1495 folio_unlock(e4b->bd_bitmap_folio);
1496 folio_put(e4b->bd_bitmap_folio);
1497 }
1498 if (e4b->bd_buddy_folio) {
1499 folio_unlock(e4b->bd_buddy_folio);
1500 folio_put(e4b->bd_buddy_folio);
1501 }
1502}
1503
1504/*
1505 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1506 * block group lock of all groups for this page; do not hold the BG lock when
1507 * calling this routine!
1508 */
1509static noinline_for_stack
1510int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1511{
1512
1513 struct ext4_group_info *this_grp;
1514 struct ext4_buddy e4b;
1515 struct folio *folio;
1516 int ret = 0;
1517
1518 might_sleep();
1519 mb_debug(sb, "init group %u\n", group);
1520 this_grp = ext4_get_group_info(sb, group);
1521 if (!this_grp)
1522 return -EFSCORRUPTED;
1523
1524 /*
1525 * This ensures that we don't reinit the buddy cache
1526 * page which map to the group from which we are already
1527 * allocating. If we are looking at the buddy cache we would
1528 * have taken a reference using ext4_mb_load_buddy and that
1529 * would have pinned buddy page to page cache.
1530 * The call to ext4_mb_get_buddy_page_lock will mark the
1531 * page accessed.
1532 */
1533 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1534 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1535 /*
1536 * somebody initialized the group
1537 * return without doing anything
1538 */
1539 goto err;
1540 }
1541
1542 folio = e4b.bd_bitmap_folio;
1543 ret = ext4_mb_init_cache(folio, NULL, gfp);
1544 if (ret)
1545 goto err;
1546 if (!folio_test_uptodate(folio)) {
1547 ret = -EIO;
1548 goto err;
1549 }
1550
1551 if (e4b.bd_buddy_folio == NULL) {
1552 /*
1553 * If both the bitmap and buddy are in
1554 * the same page we don't need to force
1555 * init the buddy
1556 */
1557 ret = 0;
1558 goto err;
1559 }
1560 /* init buddy cache */
1561 folio = e4b.bd_buddy_folio;
1562 ret = ext4_mb_init_cache(folio, e4b.bd_bitmap, gfp);
1563 if (ret)
1564 goto err;
1565 if (!folio_test_uptodate(folio)) {
1566 ret = -EIO;
1567 goto err;
1568 }
1569err:
1570 ext4_mb_put_buddy_page_lock(&e4b);
1571 return ret;
1572}
1573
1574/*
1575 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1576 * block group lock of all groups for this page; do not hold the BG lock when
1577 * calling this routine!
1578 */
1579static noinline_for_stack int
1580ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1581 struct ext4_buddy *e4b, gfp_t gfp)
1582{
1583 int blocks_per_page;
1584 int block;
1585 int pnum;
1586 int poff;
1587 struct folio *folio;
1588 int ret;
1589 struct ext4_group_info *grp;
1590 struct ext4_sb_info *sbi = EXT4_SB(sb);
1591 struct inode *inode = sbi->s_buddy_cache;
1592
1593 might_sleep();
1594 mb_debug(sb, "load group %u\n", group);
1595
1596 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1597 grp = ext4_get_group_info(sb, group);
1598 if (!grp)
1599 return -EFSCORRUPTED;
1600
1601 e4b->bd_blkbits = sb->s_blocksize_bits;
1602 e4b->bd_info = grp;
1603 e4b->bd_sb = sb;
1604 e4b->bd_group = group;
1605 e4b->bd_buddy_folio = NULL;
1606 e4b->bd_bitmap_folio = NULL;
1607
1608 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1609 /*
1610 * we need full data about the group
1611 * to make a good selection
1612 */
1613 ret = ext4_mb_init_group(sb, group, gfp);
1614 if (ret)
1615 return ret;
1616 }
1617
1618 /*
1619 * the buddy cache inode stores the block bitmap
1620 * and buddy information in consecutive blocks.
1621 * So for each group we need two blocks.
1622 */
1623 block = group * 2;
1624 pnum = block / blocks_per_page;
1625 poff = block % blocks_per_page;
1626
1627 /* Avoid locking the folio in the fast path ... */
1628 folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0);
1629 if (IS_ERR(folio) || !folio_test_uptodate(folio)) {
1630 if (!IS_ERR(folio))
1631 /*
1632 * drop the folio reference and try
1633 * to get the folio with lock. If we
1634 * are not uptodate that implies
1635 * somebody just created the folio but
1636 * is yet to initialize it. So
1637 * wait for it to initialize.
1638 */
1639 folio_put(folio);
1640 folio = __filemap_get_folio(inode->i_mapping, pnum,
1641 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1642 if (!IS_ERR(folio)) {
1643 if (WARN_RATELIMIT(folio->mapping != inode->i_mapping,
1644 "ext4: bitmap's mapping != inode->i_mapping\n")) {
1645 /* should never happen */
1646 folio_unlock(folio);
1647 ret = -EINVAL;
1648 goto err;
1649 }
1650 if (!folio_test_uptodate(folio)) {
1651 ret = ext4_mb_init_cache(folio, NULL, gfp);
1652 if (ret) {
1653 folio_unlock(folio);
1654 goto err;
1655 }
1656 mb_cmp_bitmaps(e4b, folio_address(folio) +
1657 (poff * sb->s_blocksize));
1658 }
1659 folio_unlock(folio);
1660 }
1661 }
1662 if (IS_ERR(folio)) {
1663 ret = PTR_ERR(folio);
1664 goto err;
1665 }
1666 if (!folio_test_uptodate(folio)) {
1667 ret = -EIO;
1668 goto err;
1669 }
1670
1671 /* Folios marked accessed already */
1672 e4b->bd_bitmap_folio = folio;
1673 e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize);
1674
1675 block++;
1676 pnum = block / blocks_per_page;
1677 poff = block % blocks_per_page;
1678
1679 folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0);
1680 if (IS_ERR(folio) || !folio_test_uptodate(folio)) {
1681 if (!IS_ERR(folio))
1682 folio_put(folio);
1683 folio = __filemap_get_folio(inode->i_mapping, pnum,
1684 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1685 if (!IS_ERR(folio)) {
1686 if (WARN_RATELIMIT(folio->mapping != inode->i_mapping,
1687 "ext4: buddy bitmap's mapping != inode->i_mapping\n")) {
1688 /* should never happen */
1689 folio_unlock(folio);
1690 ret = -EINVAL;
1691 goto err;
1692 }
1693 if (!folio_test_uptodate(folio)) {
1694 ret = ext4_mb_init_cache(folio, e4b->bd_bitmap,
1695 gfp);
1696 if (ret) {
1697 folio_unlock(folio);
1698 goto err;
1699 }
1700 }
1701 folio_unlock(folio);
1702 }
1703 }
1704 if (IS_ERR(folio)) {
1705 ret = PTR_ERR(folio);
1706 goto err;
1707 }
1708 if (!folio_test_uptodate(folio)) {
1709 ret = -EIO;
1710 goto err;
1711 }
1712
1713 /* Folios marked accessed already */
1714 e4b->bd_buddy_folio = folio;
1715 e4b->bd_buddy = folio_address(folio) + (poff * sb->s_blocksize);
1716
1717 return 0;
1718
1719err:
1720 if (!IS_ERR_OR_NULL(folio))
1721 folio_put(folio);
1722 if (e4b->bd_bitmap_folio)
1723 folio_put(e4b->bd_bitmap_folio);
1724
1725 e4b->bd_buddy = NULL;
1726 e4b->bd_bitmap = NULL;
1727 return ret;
1728}
1729
1730static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1731 struct ext4_buddy *e4b)
1732{
1733 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1734}
1735
1736static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1737{
1738 if (e4b->bd_bitmap_folio)
1739 folio_put(e4b->bd_bitmap_folio);
1740 if (e4b->bd_buddy_folio)
1741 folio_put(e4b->bd_buddy_folio);
1742}
1743
1744
1745static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1746{
1747 int order = 1, max;
1748 void *bb;
1749
1750 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1751 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1752
1753 while (order <= e4b->bd_blkbits + 1) {
1754 bb = mb_find_buddy(e4b, order, &max);
1755 if (!mb_test_bit(block >> order, bb)) {
1756 /* this block is part of buddy of order 'order' */
1757 return order;
1758 }
1759 order++;
1760 }
1761 return 0;
1762}
1763
1764static void mb_clear_bits(void *bm, int cur, int len)
1765{
1766 __u32 *addr;
1767
1768 len = cur + len;
1769 while (cur < len) {
1770 if ((cur & 31) == 0 && (len - cur) >= 32) {
1771 /* fast path: clear whole word at once */
1772 addr = bm + (cur >> 3);
1773 *addr = 0;
1774 cur += 32;
1775 continue;
1776 }
1777 mb_clear_bit(cur, bm);
1778 cur++;
1779 }
1780}
1781
1782/* clear bits in given range
1783 * will return first found zero bit if any, -1 otherwise
1784 */
1785static int mb_test_and_clear_bits(void *bm, int cur, int len)
1786{
1787 __u32 *addr;
1788 int zero_bit = -1;
1789
1790 len = cur + len;
1791 while (cur < len) {
1792 if ((cur & 31) == 0 && (len - cur) >= 32) {
1793 /* fast path: clear whole word at once */
1794 addr = bm + (cur >> 3);
1795 if (*addr != (__u32)(-1) && zero_bit == -1)
1796 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1797 *addr = 0;
1798 cur += 32;
1799 continue;
1800 }
1801 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1802 zero_bit = cur;
1803 cur++;
1804 }
1805
1806 return zero_bit;
1807}
1808
1809void mb_set_bits(void *bm, int cur, int len)
1810{
1811 __u32 *addr;
1812
1813 len = cur + len;
1814 while (cur < len) {
1815 if ((cur & 31) == 0 && (len - cur) >= 32) {
1816 /* fast path: set whole word at once */
1817 addr = bm + (cur >> 3);
1818 *addr = 0xffffffff;
1819 cur += 32;
1820 continue;
1821 }
1822 mb_set_bit(cur, bm);
1823 cur++;
1824 }
1825}
1826
1827static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1828{
1829 if (mb_test_bit(*bit + side, bitmap)) {
1830 mb_clear_bit(*bit, bitmap);
1831 (*bit) -= side;
1832 return 1;
1833 }
1834 else {
1835 (*bit) += side;
1836 mb_set_bit(*bit, bitmap);
1837 return -1;
1838 }
1839}
1840
1841static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1842{
1843 int max;
1844 int order = 1;
1845 void *buddy = mb_find_buddy(e4b, order, &max);
1846
1847 while (buddy) {
1848 void *buddy2;
1849
1850 /* Bits in range [first; last] are known to be set since
1851 * corresponding blocks were allocated. Bits in range
1852 * (first; last) will stay set because they form buddies on
1853 * upper layer. We just deal with borders if they don't
1854 * align with upper layer and then go up.
1855 * Releasing entire group is all about clearing
1856 * single bit of highest order buddy.
1857 */
1858
1859 /* Example:
1860 * ---------------------------------
1861 * | 1 | 1 | 1 | 1 |
1862 * ---------------------------------
1863 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1864 * ---------------------------------
1865 * 0 1 2 3 4 5 6 7
1866 * \_____________________/
1867 *
1868 * Neither [1] nor [6] is aligned to above layer.
1869 * Left neighbour [0] is free, so mark it busy,
1870 * decrease bb_counters and extend range to
1871 * [0; 6]
1872 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1873 * mark [6] free, increase bb_counters and shrink range to
1874 * [0; 5].
1875 * Then shift range to [0; 2], go up and do the same.
1876 */
1877
1878
1879 if (first & 1)
1880 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1881 if (!(last & 1))
1882 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1883 if (first > last)
1884 break;
1885 order++;
1886
1887 buddy2 = mb_find_buddy(e4b, order, &max);
1888 if (!buddy2) {
1889 mb_clear_bits(buddy, first, last - first + 1);
1890 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1891 break;
1892 }
1893 first >>= 1;
1894 last >>= 1;
1895 buddy = buddy2;
1896 }
1897}
1898
1899static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1900 int first, int count)
1901{
1902 int left_is_free = 0;
1903 int right_is_free = 0;
1904 int block;
1905 int last = first + count - 1;
1906 struct super_block *sb = e4b->bd_sb;
1907
1908 if (WARN_ON(count == 0))
1909 return;
1910 BUG_ON(last >= (sb->s_blocksize << 3));
1911 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1912 /* Don't bother if the block group is corrupt. */
1913 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1914 return;
1915
1916 mb_check_buddy(e4b);
1917 mb_free_blocks_double(inode, e4b, first, count);
1918
1919 /* access memory sequentially: check left neighbour,
1920 * clear range and then check right neighbour
1921 */
1922 if (first != 0)
1923 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1924 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1925 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1926 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1927
1928 if (unlikely(block != -1)) {
1929 struct ext4_sb_info *sbi = EXT4_SB(sb);
1930 ext4_fsblk_t blocknr;
1931
1932 /*
1933 * Fastcommit replay can free already freed blocks which
1934 * corrupts allocation info. Regenerate it.
1935 */
1936 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
1937 mb_regenerate_buddy(e4b);
1938 goto check;
1939 }
1940
1941 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1942 blocknr += EXT4_C2B(sbi, block);
1943 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1944 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1945 ext4_grp_locked_error(sb, e4b->bd_group,
1946 inode ? inode->i_ino : 0, blocknr,
1947 "freeing already freed block (bit %u); block bitmap corrupt.",
1948 block);
1949 return;
1950 }
1951
1952 this_cpu_inc(discard_pa_seq);
1953 e4b->bd_info->bb_free += count;
1954 if (first < e4b->bd_info->bb_first_free)
1955 e4b->bd_info->bb_first_free = first;
1956
1957 /* let's maintain fragments counter */
1958 if (left_is_free && right_is_free)
1959 e4b->bd_info->bb_fragments--;
1960 else if (!left_is_free && !right_is_free)
1961 e4b->bd_info->bb_fragments++;
1962
1963 /* buddy[0] == bd_bitmap is a special case, so handle
1964 * it right away and let mb_buddy_mark_free stay free of
1965 * zero order checks.
1966 * Check if neighbours are to be coaleasced,
1967 * adjust bitmap bb_counters and borders appropriately.
1968 */
1969 if (first & 1) {
1970 first += !left_is_free;
1971 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1972 }
1973 if (!(last & 1)) {
1974 last -= !right_is_free;
1975 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1976 }
1977
1978 if (first <= last)
1979 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1980
1981 mb_set_largest_free_order(sb, e4b->bd_info);
1982 mb_update_avg_fragment_size(sb, e4b->bd_info);
1983check:
1984 mb_check_buddy(e4b);
1985}
1986
1987static int mb_find_extent(struct ext4_buddy *e4b, int block,
1988 int needed, struct ext4_free_extent *ex)
1989{
1990 int max, order, next;
1991 void *buddy;
1992
1993 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1994 BUG_ON(ex == NULL);
1995
1996 buddy = mb_find_buddy(e4b, 0, &max);
1997 BUG_ON(buddy == NULL);
1998 BUG_ON(block >= max);
1999 if (mb_test_bit(block, buddy)) {
2000 ex->fe_len = 0;
2001 ex->fe_start = 0;
2002 ex->fe_group = 0;
2003 return 0;
2004 }
2005
2006 /* find actual order */
2007 order = mb_find_order_for_block(e4b, block);
2008
2009 ex->fe_len = (1 << order) - (block & ((1 << order) - 1));
2010 ex->fe_start = block;
2011 ex->fe_group = e4b->bd_group;
2012
2013 block = block >> order;
2014
2015 while (needed > ex->fe_len &&
2016 mb_find_buddy(e4b, order, &max)) {
2017
2018 if (block + 1 >= max)
2019 break;
2020
2021 next = (block + 1) * (1 << order);
2022 if (mb_test_bit(next, e4b->bd_bitmap))
2023 break;
2024
2025 order = mb_find_order_for_block(e4b, next);
2026
2027 block = next >> order;
2028 ex->fe_len += 1 << order;
2029 }
2030
2031 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
2032 /* Should never happen! (but apparently sometimes does?!?) */
2033 WARN_ON(1);
2034 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
2035 "corruption or bug in mb_find_extent "
2036 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
2037 block, order, needed, ex->fe_group, ex->fe_start,
2038 ex->fe_len, ex->fe_logical);
2039 ex->fe_len = 0;
2040 ex->fe_start = 0;
2041 ex->fe_group = 0;
2042 }
2043 return ex->fe_len;
2044}
2045
2046static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
2047{
2048 int ord;
2049 int mlen = 0;
2050 int max = 0;
2051 int start = ex->fe_start;
2052 int len = ex->fe_len;
2053 unsigned ret = 0;
2054 int len0 = len;
2055 void *buddy;
2056 int ord_start, ord_end;
2057
2058 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
2059 BUG_ON(e4b->bd_group != ex->fe_group);
2060 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
2061 mb_check_buddy(e4b);
2062 mb_mark_used_double(e4b, start, len);
2063
2064 this_cpu_inc(discard_pa_seq);
2065 e4b->bd_info->bb_free -= len;
2066 if (e4b->bd_info->bb_first_free == start)
2067 e4b->bd_info->bb_first_free += len;
2068
2069 /* let's maintain fragments counter */
2070 if (start != 0)
2071 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
2072 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
2073 max = !mb_test_bit(start + len, e4b->bd_bitmap);
2074 if (mlen && max)
2075 e4b->bd_info->bb_fragments++;
2076 else if (!mlen && !max)
2077 e4b->bd_info->bb_fragments--;
2078
2079 /* let's maintain buddy itself */
2080 while (len) {
2081 ord = mb_find_order_for_block(e4b, start);
2082
2083 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
2084 /* the whole chunk may be allocated at once! */
2085 mlen = 1 << ord;
2086 buddy = mb_find_buddy(e4b, ord, &max);
2087 BUG_ON((start >> ord) >= max);
2088 mb_set_bit(start >> ord, buddy);
2089 e4b->bd_info->bb_counters[ord]--;
2090 start += mlen;
2091 len -= mlen;
2092 BUG_ON(len < 0);
2093 continue;
2094 }
2095
2096 /* store for history */
2097 if (ret == 0)
2098 ret = len | (ord << 16);
2099
2100 BUG_ON(ord <= 0);
2101 buddy = mb_find_buddy(e4b, ord, &max);
2102 mb_set_bit(start >> ord, buddy);
2103 e4b->bd_info->bb_counters[ord]--;
2104
2105 ord_start = (start >> ord) << ord;
2106 ord_end = ord_start + (1 << ord);
2107 /* first chunk */
2108 if (start > ord_start)
2109 ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy,
2110 ord_start, start - ord_start,
2111 e4b->bd_info);
2112
2113 /* last chunk */
2114 if (start + len < ord_end) {
2115 ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy,
2116 start + len,
2117 ord_end - (start + len),
2118 e4b->bd_info);
2119 break;
2120 }
2121 len = start + len - ord_end;
2122 start = ord_end;
2123 }
2124 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
2125
2126 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2127 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2128 mb_check_buddy(e4b);
2129
2130 return ret;
2131}
2132
2133/*
2134 * Must be called under group lock!
2135 */
2136static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2137 struct ext4_buddy *e4b)
2138{
2139 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2140 int ret;
2141
2142 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2143 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2144
2145 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2146 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2147 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2148
2149 /* preallocation can change ac_b_ex, thus we store actually
2150 * allocated blocks for history */
2151 ac->ac_f_ex = ac->ac_b_ex;
2152
2153 ac->ac_status = AC_STATUS_FOUND;
2154 ac->ac_tail = ret & 0xffff;
2155 ac->ac_buddy = ret >> 16;
2156
2157 /*
2158 * take the page reference. We want the page to be pinned
2159 * so that we don't get a ext4_mb_init_cache_call for this
2160 * group until we update the bitmap. That would mean we
2161 * double allocate blocks. The reference is dropped
2162 * in ext4_mb_release_context
2163 */
2164 ac->ac_bitmap_folio = e4b->bd_bitmap_folio;
2165 folio_get(ac->ac_bitmap_folio);
2166 ac->ac_buddy_folio = e4b->bd_buddy_folio;
2167 folio_get(ac->ac_buddy_folio);
2168 /* store last allocated for subsequent stream allocation */
2169 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2170 spin_lock(&sbi->s_md_lock);
2171 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2172 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2173 spin_unlock(&sbi->s_md_lock);
2174 }
2175 /*
2176 * As we've just preallocated more space than
2177 * user requested originally, we store allocated
2178 * space in a special descriptor.
2179 */
2180 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2181 ext4_mb_new_preallocation(ac);
2182
2183}
2184
2185static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2186 struct ext4_buddy *e4b,
2187 int finish_group)
2188{
2189 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2190 struct ext4_free_extent *bex = &ac->ac_b_ex;
2191 struct ext4_free_extent *gex = &ac->ac_g_ex;
2192
2193 if (ac->ac_status == AC_STATUS_FOUND)
2194 return;
2195 /*
2196 * We don't want to scan for a whole year
2197 */
2198 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2199 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2200 ac->ac_status = AC_STATUS_BREAK;
2201 return;
2202 }
2203
2204 /*
2205 * Haven't found good chunk so far, let's continue
2206 */
2207 if (bex->fe_len < gex->fe_len)
2208 return;
2209
2210 if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2211 ext4_mb_use_best_found(ac, e4b);
2212}
2213
2214/*
2215 * The routine checks whether found extent is good enough. If it is,
2216 * then the extent gets marked used and flag is set to the context
2217 * to stop scanning. Otherwise, the extent is compared with the
2218 * previous found extent and if new one is better, then it's stored
2219 * in the context. Later, the best found extent will be used, if
2220 * mballoc can't find good enough extent.
2221 *
2222 * The algorithm used is roughly as follows:
2223 *
2224 * * If free extent found is exactly as big as goal, then
2225 * stop the scan and use it immediately
2226 *
2227 * * If free extent found is smaller than goal, then keep retrying
2228 * upto a max of sbi->s_mb_max_to_scan times (default 200). After
2229 * that stop scanning and use whatever we have.
2230 *
2231 * * If free extent found is bigger than goal, then keep retrying
2232 * upto a max of sbi->s_mb_min_to_scan times (default 10) before
2233 * stopping the scan and using the extent.
2234 *
2235 *
2236 * FIXME: real allocation policy is to be designed yet!
2237 */
2238static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2239 struct ext4_free_extent *ex,
2240 struct ext4_buddy *e4b)
2241{
2242 struct ext4_free_extent *bex = &ac->ac_b_ex;
2243 struct ext4_free_extent *gex = &ac->ac_g_ex;
2244
2245 BUG_ON(ex->fe_len <= 0);
2246 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2247 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2248 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2249
2250 ac->ac_found++;
2251 ac->ac_cX_found[ac->ac_criteria]++;
2252
2253 /*
2254 * The special case - take what you catch first
2255 */
2256 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2257 *bex = *ex;
2258 ext4_mb_use_best_found(ac, e4b);
2259 return;
2260 }
2261
2262 /*
2263 * Let's check whether the chuck is good enough
2264 */
2265 if (ex->fe_len == gex->fe_len) {
2266 *bex = *ex;
2267 ext4_mb_use_best_found(ac, e4b);
2268 return;
2269 }
2270
2271 /*
2272 * If this is first found extent, just store it in the context
2273 */
2274 if (bex->fe_len == 0) {
2275 *bex = *ex;
2276 return;
2277 }
2278
2279 /*
2280 * If new found extent is better, store it in the context
2281 */
2282 if (bex->fe_len < gex->fe_len) {
2283 /* if the request isn't satisfied, any found extent
2284 * larger than previous best one is better */
2285 if (ex->fe_len > bex->fe_len)
2286 *bex = *ex;
2287 } else if (ex->fe_len > gex->fe_len) {
2288 /* if the request is satisfied, then we try to find
2289 * an extent that still satisfy the request, but is
2290 * smaller than previous one */
2291 if (ex->fe_len < bex->fe_len)
2292 *bex = *ex;
2293 }
2294
2295 ext4_mb_check_limits(ac, e4b, 0);
2296}
2297
2298static noinline_for_stack
2299void ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2300 struct ext4_buddy *e4b)
2301{
2302 struct ext4_free_extent ex = ac->ac_b_ex;
2303 ext4_group_t group = ex.fe_group;
2304 int max;
2305 int err;
2306
2307 BUG_ON(ex.fe_len <= 0);
2308 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2309 if (err)
2310 return;
2311
2312 ext4_lock_group(ac->ac_sb, group);
2313 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2314 goto out;
2315
2316 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2317
2318 if (max > 0) {
2319 ac->ac_b_ex = ex;
2320 ext4_mb_use_best_found(ac, e4b);
2321 }
2322
2323out:
2324 ext4_unlock_group(ac->ac_sb, group);
2325 ext4_mb_unload_buddy(e4b);
2326}
2327
2328static noinline_for_stack
2329int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2330 struct ext4_buddy *e4b)
2331{
2332 ext4_group_t group = ac->ac_g_ex.fe_group;
2333 int max;
2334 int err;
2335 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2336 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2337 struct ext4_free_extent ex;
2338
2339 if (!grp)
2340 return -EFSCORRUPTED;
2341 if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY)))
2342 return 0;
2343 if (grp->bb_free == 0)
2344 return 0;
2345
2346 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2347 if (err)
2348 return err;
2349
2350 ext4_lock_group(ac->ac_sb, group);
2351 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2352 goto out;
2353
2354 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2355 ac->ac_g_ex.fe_len, &ex);
2356 ex.fe_logical = 0xDEADFA11; /* debug value */
2357
2358 if (max >= ac->ac_g_ex.fe_len &&
2359 ac->ac_g_ex.fe_len == EXT4_NUM_B2C(sbi, sbi->s_stripe)) {
2360 ext4_fsblk_t start;
2361
2362 start = ext4_grp_offs_to_block(ac->ac_sb, &ex);
2363 /* use do_div to get remainder (would be 64-bit modulo) */
2364 if (do_div(start, sbi->s_stripe) == 0) {
2365 ac->ac_found++;
2366 ac->ac_b_ex = ex;
2367 ext4_mb_use_best_found(ac, e4b);
2368 }
2369 } else if (max >= ac->ac_g_ex.fe_len) {
2370 BUG_ON(ex.fe_len <= 0);
2371 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2372 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2373 ac->ac_found++;
2374 ac->ac_b_ex = ex;
2375 ext4_mb_use_best_found(ac, e4b);
2376 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2377 /* Sometimes, caller may want to merge even small
2378 * number of blocks to an existing extent */
2379 BUG_ON(ex.fe_len <= 0);
2380 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2381 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2382 ac->ac_found++;
2383 ac->ac_b_ex = ex;
2384 ext4_mb_use_best_found(ac, e4b);
2385 }
2386out:
2387 ext4_unlock_group(ac->ac_sb, group);
2388 ext4_mb_unload_buddy(e4b);
2389
2390 return 0;
2391}
2392
2393/*
2394 * The routine scans buddy structures (not bitmap!) from given order
2395 * to max order and tries to find big enough chunk to satisfy the req
2396 */
2397static noinline_for_stack
2398void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2399 struct ext4_buddy *e4b)
2400{
2401 struct super_block *sb = ac->ac_sb;
2402 struct ext4_group_info *grp = e4b->bd_info;
2403 void *buddy;
2404 int i;
2405 int k;
2406 int max;
2407
2408 BUG_ON(ac->ac_2order <= 0);
2409 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2410 if (grp->bb_counters[i] == 0)
2411 continue;
2412
2413 buddy = mb_find_buddy(e4b, i, &max);
2414 if (WARN_RATELIMIT(buddy == NULL,
2415 "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i))
2416 continue;
2417
2418 k = mb_find_next_zero_bit(buddy, max, 0);
2419 if (k >= max) {
2420 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2421 e4b->bd_group,
2422 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2423 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2424 "%d free clusters of order %d. But found 0",
2425 grp->bb_counters[i], i);
2426 break;
2427 }
2428 ac->ac_found++;
2429 ac->ac_cX_found[ac->ac_criteria]++;
2430
2431 ac->ac_b_ex.fe_len = 1 << i;
2432 ac->ac_b_ex.fe_start = k << i;
2433 ac->ac_b_ex.fe_group = e4b->bd_group;
2434
2435 ext4_mb_use_best_found(ac, e4b);
2436
2437 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2438
2439 if (EXT4_SB(sb)->s_mb_stats)
2440 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2441
2442 break;
2443 }
2444}
2445
2446/*
2447 * The routine scans the group and measures all found extents.
2448 * In order to optimize scanning, caller must pass number of
2449 * free blocks in the group, so the routine can know upper limit.
2450 */
2451static noinline_for_stack
2452void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2453 struct ext4_buddy *e4b)
2454{
2455 struct super_block *sb = ac->ac_sb;
2456 void *bitmap = e4b->bd_bitmap;
2457 struct ext4_free_extent ex;
2458 int i, j, freelen;
2459 int free;
2460
2461 free = e4b->bd_info->bb_free;
2462 if (WARN_ON(free <= 0))
2463 return;
2464
2465 i = e4b->bd_info->bb_first_free;
2466
2467 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2468 i = mb_find_next_zero_bit(bitmap,
2469 EXT4_CLUSTERS_PER_GROUP(sb), i);
2470 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2471 /*
2472 * IF we have corrupt bitmap, we won't find any
2473 * free blocks even though group info says we
2474 * have free blocks
2475 */
2476 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2477 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2478 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2479 "%d free clusters as per "
2480 "group info. But bitmap says 0",
2481 free);
2482 break;
2483 }
2484
2485 if (!ext4_mb_cr_expensive(ac->ac_criteria)) {
2486 /*
2487 * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are
2488 * sure that this group will have a large enough
2489 * continuous free extent, so skip over the smaller free
2490 * extents
2491 */
2492 j = mb_find_next_bit(bitmap,
2493 EXT4_CLUSTERS_PER_GROUP(sb), i);
2494 freelen = j - i;
2495
2496 if (freelen < ac->ac_g_ex.fe_len) {
2497 i = j;
2498 free -= freelen;
2499 continue;
2500 }
2501 }
2502
2503 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2504 if (WARN_ON(ex.fe_len <= 0))
2505 break;
2506 if (free < ex.fe_len) {
2507 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2508 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2509 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2510 "%d free clusters as per "
2511 "group info. But got %d blocks",
2512 free, ex.fe_len);
2513 /*
2514 * The number of free blocks differs. This mostly
2515 * indicate that the bitmap is corrupt. So exit
2516 * without claiming the space.
2517 */
2518 break;
2519 }
2520 ex.fe_logical = 0xDEADC0DE; /* debug value */
2521 ext4_mb_measure_extent(ac, &ex, e4b);
2522
2523 i += ex.fe_len;
2524 free -= ex.fe_len;
2525 }
2526
2527 ext4_mb_check_limits(ac, e4b, 1);
2528}
2529
2530/*
2531 * This is a special case for storages like raid5
2532 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2533 */
2534static noinline_for_stack
2535void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2536 struct ext4_buddy *e4b)
2537{
2538 struct super_block *sb = ac->ac_sb;
2539 struct ext4_sb_info *sbi = EXT4_SB(sb);
2540 void *bitmap = e4b->bd_bitmap;
2541 struct ext4_free_extent ex;
2542 ext4_fsblk_t first_group_block;
2543 ext4_fsblk_t a;
2544 ext4_grpblk_t i, stripe;
2545 int max;
2546
2547 BUG_ON(sbi->s_stripe == 0);
2548
2549 /* find first stripe-aligned block in group */
2550 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2551
2552 a = first_group_block + sbi->s_stripe - 1;
2553 do_div(a, sbi->s_stripe);
2554 i = (a * sbi->s_stripe) - first_group_block;
2555
2556 stripe = EXT4_NUM_B2C(sbi, sbi->s_stripe);
2557 i = EXT4_B2C(sbi, i);
2558 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2559 if (!mb_test_bit(i, bitmap)) {
2560 max = mb_find_extent(e4b, i, stripe, &ex);
2561 if (max >= stripe) {
2562 ac->ac_found++;
2563 ac->ac_cX_found[ac->ac_criteria]++;
2564 ex.fe_logical = 0xDEADF00D; /* debug value */
2565 ac->ac_b_ex = ex;
2566 ext4_mb_use_best_found(ac, e4b);
2567 break;
2568 }
2569 }
2570 i += stripe;
2571 }
2572}
2573
2574/*
2575 * This is also called BEFORE we load the buddy bitmap.
2576 * Returns either 1 or 0 indicating that the group is either suitable
2577 * for the allocation or not.
2578 */
2579static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2580 ext4_group_t group, enum criteria cr)
2581{
2582 ext4_grpblk_t free, fragments;
2583 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2584 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2585
2586 BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS);
2587
2588 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2589 return false;
2590
2591 free = grp->bb_free;
2592 if (free == 0)
2593 return false;
2594
2595 fragments = grp->bb_fragments;
2596 if (fragments == 0)
2597 return false;
2598
2599 switch (cr) {
2600 case CR_POWER2_ALIGNED:
2601 BUG_ON(ac->ac_2order == 0);
2602
2603 /* Avoid using the first bg of a flexgroup for data files */
2604 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2605 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2606 ((group % flex_size) == 0))
2607 return false;
2608
2609 if (free < ac->ac_g_ex.fe_len)
2610 return false;
2611
2612 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2613 return true;
2614
2615 if (grp->bb_largest_free_order < ac->ac_2order)
2616 return false;
2617
2618 return true;
2619 case CR_GOAL_LEN_FAST:
2620 case CR_BEST_AVAIL_LEN:
2621 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2622 return true;
2623 break;
2624 case CR_GOAL_LEN_SLOW:
2625 if (free >= ac->ac_g_ex.fe_len)
2626 return true;
2627 break;
2628 case CR_ANY_FREE:
2629 return true;
2630 default:
2631 BUG();
2632 }
2633
2634 return false;
2635}
2636
2637/*
2638 * This could return negative error code if something goes wrong
2639 * during ext4_mb_init_group(). This should not be called with
2640 * ext4_lock_group() held.
2641 *
2642 * Note: because we are conditionally operating with the group lock in
2643 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2644 * function using __acquire and __release. This means we need to be
2645 * super careful before messing with the error path handling via "goto
2646 * out"!
2647 */
2648static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2649 ext4_group_t group, enum criteria cr)
2650{
2651 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2652 struct super_block *sb = ac->ac_sb;
2653 struct ext4_sb_info *sbi = EXT4_SB(sb);
2654 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2655 ext4_grpblk_t free;
2656 int ret = 0;
2657
2658 if (!grp)
2659 return -EFSCORRUPTED;
2660 if (sbi->s_mb_stats)
2661 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2662 if (should_lock) {
2663 ext4_lock_group(sb, group);
2664 __release(ext4_group_lock_ptr(sb, group));
2665 }
2666 free = grp->bb_free;
2667 if (free == 0)
2668 goto out;
2669 /*
2670 * In all criterias except CR_ANY_FREE we try to avoid groups that
2671 * can't possibly satisfy the full goal request due to insufficient
2672 * free blocks.
2673 */
2674 if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len)
2675 goto out;
2676 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2677 goto out;
2678 if (should_lock) {
2679 __acquire(ext4_group_lock_ptr(sb, group));
2680 ext4_unlock_group(sb, group);
2681 }
2682
2683 /* We only do this if the grp has never been initialized */
2684 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2685 struct ext4_group_desc *gdp =
2686 ext4_get_group_desc(sb, group, NULL);
2687 int ret;
2688
2689 /*
2690 * CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic
2691 * search to find large good chunks almost for free. If buddy
2692 * data is not ready, then this optimization makes no sense. But
2693 * we never skip the first block group in a flex_bg, since this
2694 * gets used for metadata block allocation, and we want to make
2695 * sure we locate metadata blocks in the first block group in
2696 * the flex_bg if possible.
2697 */
2698 if (!ext4_mb_cr_expensive(cr) &&
2699 (!sbi->s_log_groups_per_flex ||
2700 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2701 !(ext4_has_group_desc_csum(sb) &&
2702 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2703 return 0;
2704 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2705 if (ret)
2706 return ret;
2707 }
2708
2709 if (should_lock) {
2710 ext4_lock_group(sb, group);
2711 __release(ext4_group_lock_ptr(sb, group));
2712 }
2713 ret = ext4_mb_good_group(ac, group, cr);
2714out:
2715 if (should_lock) {
2716 __acquire(ext4_group_lock_ptr(sb, group));
2717 ext4_unlock_group(sb, group);
2718 }
2719 return ret;
2720}
2721
2722/*
2723 * Start prefetching @nr block bitmaps starting at @group.
2724 * Return the next group which needs to be prefetched.
2725 */
2726ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2727 unsigned int nr, int *cnt)
2728{
2729 ext4_group_t ngroups = ext4_get_groups_count(sb);
2730 struct buffer_head *bh;
2731 struct blk_plug plug;
2732
2733 blk_start_plug(&plug);
2734 while (nr-- > 0) {
2735 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2736 NULL);
2737 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2738
2739 /*
2740 * Prefetch block groups with free blocks; but don't
2741 * bother if it is marked uninitialized on disk, since
2742 * it won't require I/O to read. Also only try to
2743 * prefetch once, so we avoid getblk() call, which can
2744 * be expensive.
2745 */
2746 if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2747 EXT4_MB_GRP_NEED_INIT(grp) &&
2748 ext4_free_group_clusters(sb, gdp) > 0 ) {
2749 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2750 if (bh && !IS_ERR(bh)) {
2751 if (!buffer_uptodate(bh) && cnt)
2752 (*cnt)++;
2753 brelse(bh);
2754 }
2755 }
2756 if (++group >= ngroups)
2757 group = 0;
2758 }
2759 blk_finish_plug(&plug);
2760 return group;
2761}
2762
2763/*
2764 * Prefetching reads the block bitmap into the buffer cache; but we
2765 * need to make sure that the buddy bitmap in the page cache has been
2766 * initialized. Note that ext4_mb_init_group() will block if the I/O
2767 * is not yet completed, or indeed if it was not initiated by
2768 * ext4_mb_prefetch did not start the I/O.
2769 *
2770 * TODO: We should actually kick off the buddy bitmap setup in a work
2771 * queue when the buffer I/O is completed, so that we don't block
2772 * waiting for the block allocation bitmap read to finish when
2773 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2774 */
2775void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2776 unsigned int nr)
2777{
2778 struct ext4_group_desc *gdp;
2779 struct ext4_group_info *grp;
2780
2781 while (nr-- > 0) {
2782 if (!group)
2783 group = ext4_get_groups_count(sb);
2784 group--;
2785 gdp = ext4_get_group_desc(sb, group, NULL);
2786 grp = ext4_get_group_info(sb, group);
2787
2788 if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) &&
2789 ext4_free_group_clusters(sb, gdp) > 0) {
2790 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2791 break;
2792 }
2793 }
2794}
2795
2796static noinline_for_stack int
2797ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2798{
2799 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2800 enum criteria new_cr, cr = CR_GOAL_LEN_FAST;
2801 int err = 0, first_err = 0;
2802 unsigned int nr = 0, prefetch_ios = 0;
2803 struct ext4_sb_info *sbi;
2804 struct super_block *sb;
2805 struct ext4_buddy e4b;
2806 int lost;
2807
2808 sb = ac->ac_sb;
2809 sbi = EXT4_SB(sb);
2810 ngroups = ext4_get_groups_count(sb);
2811 /* non-extent files are limited to low blocks/groups */
2812 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2813 ngroups = sbi->s_blockfile_groups;
2814
2815 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2816
2817 /* first, try the goal */
2818 err = ext4_mb_find_by_goal(ac, &e4b);
2819 if (err || ac->ac_status == AC_STATUS_FOUND)
2820 goto out;
2821
2822 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2823 goto out;
2824
2825 /*
2826 * ac->ac_2order is set only if the fe_len is a power of 2
2827 * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED
2828 * so that we try exact allocation using buddy.
2829 */
2830 i = fls(ac->ac_g_ex.fe_len);
2831 ac->ac_2order = 0;
2832 /*
2833 * We search using buddy data only if the order of the request
2834 * is greater than equal to the sbi_s_mb_order2_reqs
2835 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2836 * We also support searching for power-of-two requests only for
2837 * requests upto maximum buddy size we have constructed.
2838 */
2839 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2840 if (is_power_of_2(ac->ac_g_ex.fe_len))
2841 ac->ac_2order = array_index_nospec(i - 1,
2842 MB_NUM_ORDERS(sb));
2843 }
2844
2845 /* if stream allocation is enabled, use global goal */
2846 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2847 /* TBD: may be hot point */
2848 spin_lock(&sbi->s_md_lock);
2849 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2850 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2851 spin_unlock(&sbi->s_md_lock);
2852 }
2853
2854 /*
2855 * Let's just scan groups to find more-less suitable blocks We
2856 * start with CR_GOAL_LEN_FAST, unless it is power of 2
2857 * aligned, in which case let's do that faster approach first.
2858 */
2859 if (ac->ac_2order)
2860 cr = CR_POWER2_ALIGNED;
2861repeat:
2862 for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2863 ac->ac_criteria = cr;
2864 /*
2865 * searching for the right group start
2866 * from the goal value specified
2867 */
2868 group = ac->ac_g_ex.fe_group;
2869 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2870 prefetch_grp = group;
2871 nr = 0;
2872
2873 for (i = 0, new_cr = cr; i < ngroups; i++,
2874 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2875 int ret = 0;
2876
2877 cond_resched();
2878 if (new_cr != cr) {
2879 cr = new_cr;
2880 goto repeat;
2881 }
2882
2883 /*
2884 * Batch reads of the block allocation bitmaps
2885 * to get multiple READs in flight; limit
2886 * prefetching at inexpensive CR, otherwise mballoc
2887 * can spend a lot of time loading imperfect groups
2888 */
2889 if ((prefetch_grp == group) &&
2890 (ext4_mb_cr_expensive(cr) ||
2891 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2892 nr = sbi->s_mb_prefetch;
2893 if (ext4_has_feature_flex_bg(sb)) {
2894 nr = 1 << sbi->s_log_groups_per_flex;
2895 nr -= group & (nr - 1);
2896 nr = min(nr, sbi->s_mb_prefetch);
2897 }
2898 prefetch_grp = ext4_mb_prefetch(sb, group,
2899 nr, &prefetch_ios);
2900 }
2901
2902 /* This now checks without needing the buddy page */
2903 ret = ext4_mb_good_group_nolock(ac, group, cr);
2904 if (ret <= 0) {
2905 if (!first_err)
2906 first_err = ret;
2907 continue;
2908 }
2909
2910 err = ext4_mb_load_buddy(sb, group, &e4b);
2911 if (err)
2912 goto out;
2913
2914 ext4_lock_group(sb, group);
2915
2916 /*
2917 * We need to check again after locking the
2918 * block group
2919 */
2920 ret = ext4_mb_good_group(ac, group, cr);
2921 if (ret == 0) {
2922 ext4_unlock_group(sb, group);
2923 ext4_mb_unload_buddy(&e4b);
2924 continue;
2925 }
2926
2927 ac->ac_groups_scanned++;
2928 if (cr == CR_POWER2_ALIGNED)
2929 ext4_mb_simple_scan_group(ac, &e4b);
2930 else {
2931 bool is_stripe_aligned =
2932 (sbi->s_stripe >=
2933 sbi->s_cluster_ratio) &&
2934 !(ac->ac_g_ex.fe_len %
2935 EXT4_NUM_B2C(sbi, sbi->s_stripe));
2936
2937 if ((cr == CR_GOAL_LEN_FAST ||
2938 cr == CR_BEST_AVAIL_LEN) &&
2939 is_stripe_aligned)
2940 ext4_mb_scan_aligned(ac, &e4b);
2941
2942 if (ac->ac_status == AC_STATUS_CONTINUE)
2943 ext4_mb_complex_scan_group(ac, &e4b);
2944 }
2945
2946 ext4_unlock_group(sb, group);
2947 ext4_mb_unload_buddy(&e4b);
2948
2949 if (ac->ac_status != AC_STATUS_CONTINUE)
2950 break;
2951 }
2952 /* Processed all groups and haven't found blocks */
2953 if (sbi->s_mb_stats && i == ngroups)
2954 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2955
2956 if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN)
2957 /* Reset goal length to original goal length before
2958 * falling into CR_GOAL_LEN_SLOW */
2959 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
2960 }
2961
2962 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2963 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2964 /*
2965 * We've been searching too long. Let's try to allocate
2966 * the best chunk we've found so far
2967 */
2968 ext4_mb_try_best_found(ac, &e4b);
2969 if (ac->ac_status != AC_STATUS_FOUND) {
2970 /*
2971 * Someone more lucky has already allocated it.
2972 * The only thing we can do is just take first
2973 * found block(s)
2974 */
2975 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2976 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2977 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2978 ac->ac_b_ex.fe_len, lost);
2979
2980 ac->ac_b_ex.fe_group = 0;
2981 ac->ac_b_ex.fe_start = 0;
2982 ac->ac_b_ex.fe_len = 0;
2983 ac->ac_status = AC_STATUS_CONTINUE;
2984 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2985 cr = CR_ANY_FREE;
2986 goto repeat;
2987 }
2988 }
2989
2990 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2991 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2992out:
2993 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2994 err = first_err;
2995
2996 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2997 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2998 ac->ac_flags, cr, err);
2999
3000 if (nr)
3001 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
3002
3003 return err;
3004}
3005
3006static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
3007{
3008 struct super_block *sb = pde_data(file_inode(seq->file));
3009 ext4_group_t group;
3010
3011 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3012 return NULL;
3013 group = *pos + 1;
3014 return (void *) ((unsigned long) group);
3015}
3016
3017static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
3018{
3019 struct super_block *sb = pde_data(file_inode(seq->file));
3020 ext4_group_t group;
3021
3022 ++*pos;
3023 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3024 return NULL;
3025 group = *pos + 1;
3026 return (void *) ((unsigned long) group);
3027}
3028
3029static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
3030{
3031 struct super_block *sb = pde_data(file_inode(seq->file));
3032 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
3033 int i, err;
3034 char nbuf[16];
3035 struct ext4_buddy e4b;
3036 struct ext4_group_info *grinfo;
3037 unsigned char blocksize_bits = min_t(unsigned char,
3038 sb->s_blocksize_bits,
3039 EXT4_MAX_BLOCK_LOG_SIZE);
3040 struct sg {
3041 struct ext4_group_info info;
3042 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
3043 } sg;
3044
3045 group--;
3046 if (group == 0)
3047 seq_puts(seq, "#group: free frags first ["
3048 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
3049 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
3050
3051 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
3052 sizeof(struct ext4_group_info);
3053
3054 grinfo = ext4_get_group_info(sb, group);
3055 if (!grinfo)
3056 return 0;
3057 /* Load the group info in memory only if not already loaded. */
3058 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
3059 err = ext4_mb_load_buddy(sb, group, &e4b);
3060 if (err) {
3061 seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf));
3062 return 0;
3063 }
3064 ext4_mb_unload_buddy(&e4b);
3065 }
3066
3067 /*
3068 * We care only about free space counters in the group info and
3069 * these are safe to access even after the buddy has been unloaded
3070 */
3071 memcpy(&sg, grinfo, i);
3072 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
3073 sg.info.bb_fragments, sg.info.bb_first_free);
3074 for (i = 0; i <= 13; i++)
3075 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
3076 sg.info.bb_counters[i] : 0);
3077 seq_puts(seq, " ]");
3078 if (EXT4_MB_GRP_BBITMAP_CORRUPT(&sg.info))
3079 seq_puts(seq, " Block bitmap corrupted!");
3080 seq_putc(seq, '\n');
3081 return 0;
3082}
3083
3084static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
3085{
3086}
3087
3088const struct seq_operations ext4_mb_seq_groups_ops = {
3089 .start = ext4_mb_seq_groups_start,
3090 .next = ext4_mb_seq_groups_next,
3091 .stop = ext4_mb_seq_groups_stop,
3092 .show = ext4_mb_seq_groups_show,
3093};
3094
3095int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
3096{
3097 struct super_block *sb = seq->private;
3098 struct ext4_sb_info *sbi = EXT4_SB(sb);
3099
3100 seq_puts(seq, "mballoc:\n");
3101 if (!sbi->s_mb_stats) {
3102 seq_puts(seq, "\tmb stats collection turned off.\n");
3103 seq_puts(
3104 seq,
3105 "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
3106 return 0;
3107 }
3108 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
3109 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
3110
3111 seq_printf(seq, "\tgroups_scanned: %u\n",
3112 atomic_read(&sbi->s_bal_groups_scanned));
3113
3114 /* CR_POWER2_ALIGNED stats */
3115 seq_puts(seq, "\tcr_p2_aligned_stats:\n");
3116 seq_printf(seq, "\t\thits: %llu\n",
3117 atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED]));
3118 seq_printf(
3119 seq, "\t\tgroups_considered: %llu\n",
3120 atomic64_read(
3121 &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]));
3122 seq_printf(seq, "\t\textents_scanned: %u\n",
3123 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED]));
3124 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3125 atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED]));
3126 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3127 atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions));
3128
3129 /* CR_GOAL_LEN_FAST stats */
3130 seq_puts(seq, "\tcr_goal_fast_stats:\n");
3131 seq_printf(seq, "\t\thits: %llu\n",
3132 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST]));
3133 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3134 atomic64_read(
3135 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST]));
3136 seq_printf(seq, "\t\textents_scanned: %u\n",
3137 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST]));
3138 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3139 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST]));
3140 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3141 atomic_read(&sbi->s_bal_goal_fast_bad_suggestions));
3142
3143 /* CR_BEST_AVAIL_LEN stats */
3144 seq_puts(seq, "\tcr_best_avail_stats:\n");
3145 seq_printf(seq, "\t\thits: %llu\n",
3146 atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN]));
3147 seq_printf(
3148 seq, "\t\tgroups_considered: %llu\n",
3149 atomic64_read(
3150 &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN]));
3151 seq_printf(seq, "\t\textents_scanned: %u\n",
3152 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN]));
3153 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3154 atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN]));
3155 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3156 atomic_read(&sbi->s_bal_best_avail_bad_suggestions));
3157
3158 /* CR_GOAL_LEN_SLOW stats */
3159 seq_puts(seq, "\tcr_goal_slow_stats:\n");
3160 seq_printf(seq, "\t\thits: %llu\n",
3161 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW]));
3162 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3163 atomic64_read(
3164 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW]));
3165 seq_printf(seq, "\t\textents_scanned: %u\n",
3166 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW]));
3167 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3168 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW]));
3169
3170 /* CR_ANY_FREE stats */
3171 seq_puts(seq, "\tcr_any_free_stats:\n");
3172 seq_printf(seq, "\t\thits: %llu\n",
3173 atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE]));
3174 seq_printf(
3175 seq, "\t\tgroups_considered: %llu\n",
3176 atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE]));
3177 seq_printf(seq, "\t\textents_scanned: %u\n",
3178 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE]));
3179 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3180 atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE]));
3181
3182 /* Aggregates */
3183 seq_printf(seq, "\textents_scanned: %u\n",
3184 atomic_read(&sbi->s_bal_ex_scanned));
3185 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
3186 seq_printf(seq, "\t\tlen_goal_hits: %u\n",
3187 atomic_read(&sbi->s_bal_len_goals));
3188 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
3189 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
3190 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
3191 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
3192 atomic_read(&sbi->s_mb_buddies_generated),
3193 ext4_get_groups_count(sb));
3194 seq_printf(seq, "\tbuddies_time_used: %llu\n",
3195 atomic64_read(&sbi->s_mb_generation_time));
3196 seq_printf(seq, "\tpreallocated: %u\n",
3197 atomic_read(&sbi->s_mb_preallocated));
3198 seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded));
3199 return 0;
3200}
3201
3202static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
3203{
3204 struct super_block *sb = pde_data(file_inode(seq->file));
3205 unsigned long position;
3206
3207 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3208 return NULL;
3209 position = *pos + 1;
3210 return (void *) ((unsigned long) position);
3211}
3212
3213static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3214{
3215 struct super_block *sb = pde_data(file_inode(seq->file));
3216 unsigned long position;
3217
3218 ++*pos;
3219 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3220 return NULL;
3221 position = *pos + 1;
3222 return (void *) ((unsigned long) position);
3223}
3224
3225static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3226{
3227 struct super_block *sb = pde_data(file_inode(seq->file));
3228 struct ext4_sb_info *sbi = EXT4_SB(sb);
3229 unsigned long position = ((unsigned long) v);
3230 struct ext4_group_info *grp;
3231 unsigned int count;
3232
3233 position--;
3234 if (position >= MB_NUM_ORDERS(sb)) {
3235 position -= MB_NUM_ORDERS(sb);
3236 if (position == 0)
3237 seq_puts(seq, "avg_fragment_size_lists:\n");
3238
3239 count = 0;
3240 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
3241 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
3242 bb_avg_fragment_size_node)
3243 count++;
3244 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
3245 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3246 (unsigned int)position, count);
3247 return 0;
3248 }
3249
3250 if (position == 0) {
3251 seq_printf(seq, "optimize_scan: %d\n",
3252 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3253 seq_puts(seq, "max_free_order_lists:\n");
3254 }
3255 count = 0;
3256 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
3257 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3258 bb_largest_free_order_node)
3259 count++;
3260 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
3261 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3262 (unsigned int)position, count);
3263
3264 return 0;
3265}
3266
3267static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3268{
3269}
3270
3271const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3272 .start = ext4_mb_seq_structs_summary_start,
3273 .next = ext4_mb_seq_structs_summary_next,
3274 .stop = ext4_mb_seq_structs_summary_stop,
3275 .show = ext4_mb_seq_structs_summary_show,
3276};
3277
3278static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3279{
3280 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3281 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3282
3283 BUG_ON(!cachep);
3284 return cachep;
3285}
3286
3287/*
3288 * Allocate the top-level s_group_info array for the specified number
3289 * of groups
3290 */
3291int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3292{
3293 struct ext4_sb_info *sbi = EXT4_SB(sb);
3294 unsigned size;
3295 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3296
3297 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3298 EXT4_DESC_PER_BLOCK_BITS(sb);
3299 if (size <= sbi->s_group_info_size)
3300 return 0;
3301
3302 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3303 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3304 if (!new_groupinfo) {
3305 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3306 return -ENOMEM;
3307 }
3308 rcu_read_lock();
3309 old_groupinfo = rcu_dereference(sbi->s_group_info);
3310 if (old_groupinfo)
3311 memcpy(new_groupinfo, old_groupinfo,
3312 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3313 rcu_read_unlock();
3314 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3315 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3316 if (old_groupinfo)
3317 ext4_kvfree_array_rcu(old_groupinfo);
3318 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3319 sbi->s_group_info_size);
3320 return 0;
3321}
3322
3323/* Create and initialize ext4_group_info data for the given group. */
3324int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3325 struct ext4_group_desc *desc)
3326{
3327 int i;
3328 int metalen = 0;
3329 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3330 struct ext4_sb_info *sbi = EXT4_SB(sb);
3331 struct ext4_group_info **meta_group_info;
3332 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3333
3334 /*
3335 * First check if this group is the first of a reserved block.
3336 * If it's true, we have to allocate a new table of pointers
3337 * to ext4_group_info structures
3338 */
3339 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3340 metalen = sizeof(*meta_group_info) <<
3341 EXT4_DESC_PER_BLOCK_BITS(sb);
3342 meta_group_info = kmalloc(metalen, GFP_NOFS);
3343 if (meta_group_info == NULL) {
3344 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3345 "for a buddy group");
3346 return -ENOMEM;
3347 }
3348 rcu_read_lock();
3349 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3350 rcu_read_unlock();
3351 }
3352
3353 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3354 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3355
3356 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3357 if (meta_group_info[i] == NULL) {
3358 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3359 goto exit_group_info;
3360 }
3361 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3362 &(meta_group_info[i]->bb_state));
3363
3364 /*
3365 * initialize bb_free to be able to skip
3366 * empty groups without initialization
3367 */
3368 if (ext4_has_group_desc_csum(sb) &&
3369 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3370 meta_group_info[i]->bb_free =
3371 ext4_free_clusters_after_init(sb, group, desc);
3372 } else {
3373 meta_group_info[i]->bb_free =
3374 ext4_free_group_clusters(sb, desc);
3375 }
3376
3377 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3378 init_rwsem(&meta_group_info[i]->alloc_sem);
3379 meta_group_info[i]->bb_free_root = RB_ROOT;
3380 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3381 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3382 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3383 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3384 meta_group_info[i]->bb_group = group;
3385
3386 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3387 return 0;
3388
3389exit_group_info:
3390 /* If a meta_group_info table has been allocated, release it now */
3391 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3392 struct ext4_group_info ***group_info;
3393
3394 rcu_read_lock();
3395 group_info = rcu_dereference(sbi->s_group_info);
3396 kfree(group_info[idx]);
3397 group_info[idx] = NULL;
3398 rcu_read_unlock();
3399 }
3400 return -ENOMEM;
3401} /* ext4_mb_add_groupinfo */
3402
3403static int ext4_mb_init_backend(struct super_block *sb)
3404{
3405 ext4_group_t ngroups = ext4_get_groups_count(sb);
3406 ext4_group_t i;
3407 struct ext4_sb_info *sbi = EXT4_SB(sb);
3408 int err;
3409 struct ext4_group_desc *desc;
3410 struct ext4_group_info ***group_info;
3411 struct kmem_cache *cachep;
3412
3413 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3414 if (err)
3415 return err;
3416
3417 sbi->s_buddy_cache = new_inode(sb);
3418 if (sbi->s_buddy_cache == NULL) {
3419 ext4_msg(sb, KERN_ERR, "can't get new inode");
3420 goto err_freesgi;
3421 }
3422 /* To avoid potentially colliding with an valid on-disk inode number,
3423 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3424 * not in the inode hash, so it should never be found by iget(), but
3425 * this will avoid confusion if it ever shows up during debugging. */
3426 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3427 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3428 for (i = 0; i < ngroups; i++) {
3429 cond_resched();
3430 desc = ext4_get_group_desc(sb, i, NULL);
3431 if (desc == NULL) {
3432 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3433 goto err_freebuddy;
3434 }
3435 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3436 goto err_freebuddy;
3437 }
3438
3439 if (ext4_has_feature_flex_bg(sb)) {
3440 /* a single flex group is supposed to be read by a single IO.
3441 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3442 * unsigned integer, so the maximum shift is 32.
3443 */
3444 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3445 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3446 goto err_freebuddy;
3447 }
3448 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3449 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3450 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3451 } else {
3452 sbi->s_mb_prefetch = 32;
3453 }
3454 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3455 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3456 /*
3457 * now many real IOs to prefetch within a single allocation at
3458 * CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related
3459 * optimization we shouldn't try to load too many groups, at some point
3460 * we should start to use what we've got in memory.
3461 * with an average random access time 5ms, it'd take a second to get
3462 * 200 groups (* N with flex_bg), so let's make this limit 4
3463 */
3464 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3465 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3466 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3467
3468 return 0;
3469
3470err_freebuddy:
3471 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3472 while (i-- > 0) {
3473 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3474
3475 if (grp)
3476 kmem_cache_free(cachep, grp);
3477 }
3478 i = sbi->s_group_info_size;
3479 rcu_read_lock();
3480 group_info = rcu_dereference(sbi->s_group_info);
3481 while (i-- > 0)
3482 kfree(group_info[i]);
3483 rcu_read_unlock();
3484 iput(sbi->s_buddy_cache);
3485err_freesgi:
3486 rcu_read_lock();
3487 kvfree(rcu_dereference(sbi->s_group_info));
3488 rcu_read_unlock();
3489 return -ENOMEM;
3490}
3491
3492static void ext4_groupinfo_destroy_slabs(void)
3493{
3494 int i;
3495
3496 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3497 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3498 ext4_groupinfo_caches[i] = NULL;
3499 }
3500}
3501
3502static int ext4_groupinfo_create_slab(size_t size)
3503{
3504 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3505 int slab_size;
3506 int blocksize_bits = order_base_2(size);
3507 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3508 struct kmem_cache *cachep;
3509
3510 if (cache_index >= NR_GRPINFO_CACHES)
3511 return -EINVAL;
3512
3513 if (unlikely(cache_index < 0))
3514 cache_index = 0;
3515
3516 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3517 if (ext4_groupinfo_caches[cache_index]) {
3518 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3519 return 0; /* Already created */
3520 }
3521
3522 slab_size = offsetof(struct ext4_group_info,
3523 bb_counters[blocksize_bits + 2]);
3524
3525 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3526 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3527 NULL);
3528
3529 ext4_groupinfo_caches[cache_index] = cachep;
3530
3531 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3532 if (!cachep) {
3533 printk(KERN_EMERG
3534 "EXT4-fs: no memory for groupinfo slab cache\n");
3535 return -ENOMEM;
3536 }
3537
3538 return 0;
3539}
3540
3541static void ext4_discard_work(struct work_struct *work)
3542{
3543 struct ext4_sb_info *sbi = container_of(work,
3544 struct ext4_sb_info, s_discard_work);
3545 struct super_block *sb = sbi->s_sb;
3546 struct ext4_free_data *fd, *nfd;
3547 struct ext4_buddy e4b;
3548 LIST_HEAD(discard_list);
3549 ext4_group_t grp, load_grp;
3550 int err = 0;
3551
3552 spin_lock(&sbi->s_md_lock);
3553 list_splice_init(&sbi->s_discard_list, &discard_list);
3554 spin_unlock(&sbi->s_md_lock);
3555
3556 load_grp = UINT_MAX;
3557 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3558 /*
3559 * If filesystem is umounting or no memory or suffering
3560 * from no space, give up the discard
3561 */
3562 if ((sb->s_flags & SB_ACTIVE) && !err &&
3563 !atomic_read(&sbi->s_retry_alloc_pending)) {
3564 grp = fd->efd_group;
3565 if (grp != load_grp) {
3566 if (load_grp != UINT_MAX)
3567 ext4_mb_unload_buddy(&e4b);
3568
3569 err = ext4_mb_load_buddy(sb, grp, &e4b);
3570 if (err) {
3571 kmem_cache_free(ext4_free_data_cachep, fd);
3572 load_grp = UINT_MAX;
3573 continue;
3574 } else {
3575 load_grp = grp;
3576 }
3577 }
3578
3579 ext4_lock_group(sb, grp);
3580 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3581 fd->efd_start_cluster + fd->efd_count - 1, 1);
3582 ext4_unlock_group(sb, grp);
3583 }
3584 kmem_cache_free(ext4_free_data_cachep, fd);
3585 }
3586
3587 if (load_grp != UINT_MAX)
3588 ext4_mb_unload_buddy(&e4b);
3589}
3590
3591int ext4_mb_init(struct super_block *sb)
3592{
3593 struct ext4_sb_info *sbi = EXT4_SB(sb);
3594 unsigned i, j;
3595 unsigned offset, offset_incr;
3596 unsigned max;
3597 int ret;
3598
3599 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3600
3601 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3602 if (sbi->s_mb_offsets == NULL) {
3603 ret = -ENOMEM;
3604 goto out;
3605 }
3606
3607 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3608 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3609 if (sbi->s_mb_maxs == NULL) {
3610 ret = -ENOMEM;
3611 goto out;
3612 }
3613
3614 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3615 if (ret < 0)
3616 goto out;
3617
3618 /* order 0 is regular bitmap */
3619 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3620 sbi->s_mb_offsets[0] = 0;
3621
3622 i = 1;
3623 offset = 0;
3624 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3625 max = sb->s_blocksize << 2;
3626 do {
3627 sbi->s_mb_offsets[i] = offset;
3628 sbi->s_mb_maxs[i] = max;
3629 offset += offset_incr;
3630 offset_incr = offset_incr >> 1;
3631 max = max >> 1;
3632 i++;
3633 } while (i < MB_NUM_ORDERS(sb));
3634
3635 sbi->s_mb_avg_fragment_size =
3636 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3637 GFP_KERNEL);
3638 if (!sbi->s_mb_avg_fragment_size) {
3639 ret = -ENOMEM;
3640 goto out;
3641 }
3642 sbi->s_mb_avg_fragment_size_locks =
3643 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3644 GFP_KERNEL);
3645 if (!sbi->s_mb_avg_fragment_size_locks) {
3646 ret = -ENOMEM;
3647 goto out;
3648 }
3649 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3650 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3651 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3652 }
3653 sbi->s_mb_largest_free_orders =
3654 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3655 GFP_KERNEL);
3656 if (!sbi->s_mb_largest_free_orders) {
3657 ret = -ENOMEM;
3658 goto out;
3659 }
3660 sbi->s_mb_largest_free_orders_locks =
3661 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3662 GFP_KERNEL);
3663 if (!sbi->s_mb_largest_free_orders_locks) {
3664 ret = -ENOMEM;
3665 goto out;
3666 }
3667 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3668 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3669 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3670 }
3671
3672 spin_lock_init(&sbi->s_md_lock);
3673 sbi->s_mb_free_pending = 0;
3674 INIT_LIST_HEAD(&sbi->s_freed_data_list[0]);
3675 INIT_LIST_HEAD(&sbi->s_freed_data_list[1]);
3676 INIT_LIST_HEAD(&sbi->s_discard_list);
3677 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3678 atomic_set(&sbi->s_retry_alloc_pending, 0);
3679
3680 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3681 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3682 sbi->s_mb_stats = MB_DEFAULT_STATS;
3683 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3684 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3685 sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER;
3686
3687 /*
3688 * The default group preallocation is 512, which for 4k block
3689 * sizes translates to 2 megabytes. However for bigalloc file
3690 * systems, this is probably too big (i.e, if the cluster size
3691 * is 1 megabyte, then group preallocation size becomes half a
3692 * gigabyte!). As a default, we will keep a two megabyte
3693 * group pralloc size for cluster sizes up to 64k, and after
3694 * that, we will force a minimum group preallocation size of
3695 * 32 clusters. This translates to 8 megs when the cluster
3696 * size is 256k, and 32 megs when the cluster size is 1 meg,
3697 * which seems reasonable as a default.
3698 */
3699 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3700 sbi->s_cluster_bits, 32);
3701 /*
3702 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3703 * to the lowest multiple of s_stripe which is bigger than
3704 * the s_mb_group_prealloc as determined above. We want
3705 * the preallocation size to be an exact multiple of the
3706 * RAID stripe size so that preallocations don't fragment
3707 * the stripes.
3708 */
3709 if (sbi->s_stripe > 1) {
3710 sbi->s_mb_group_prealloc = roundup(
3711 sbi->s_mb_group_prealloc, EXT4_NUM_B2C(sbi, sbi->s_stripe));
3712 }
3713
3714 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3715 if (sbi->s_locality_groups == NULL) {
3716 ret = -ENOMEM;
3717 goto out;
3718 }
3719 for_each_possible_cpu(i) {
3720 struct ext4_locality_group *lg;
3721 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3722 mutex_init(&lg->lg_mutex);
3723 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3724 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3725 spin_lock_init(&lg->lg_prealloc_lock);
3726 }
3727
3728 if (bdev_nonrot(sb->s_bdev))
3729 sbi->s_mb_max_linear_groups = 0;
3730 else
3731 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3732 /* init file for buddy data */
3733 ret = ext4_mb_init_backend(sb);
3734 if (ret != 0)
3735 goto out_free_locality_groups;
3736
3737 return 0;
3738
3739out_free_locality_groups:
3740 free_percpu(sbi->s_locality_groups);
3741 sbi->s_locality_groups = NULL;
3742out:
3743 kfree(sbi->s_mb_avg_fragment_size);
3744 kfree(sbi->s_mb_avg_fragment_size_locks);
3745 kfree(sbi->s_mb_largest_free_orders);
3746 kfree(sbi->s_mb_largest_free_orders_locks);
3747 kfree(sbi->s_mb_offsets);
3748 sbi->s_mb_offsets = NULL;
3749 kfree(sbi->s_mb_maxs);
3750 sbi->s_mb_maxs = NULL;
3751 return ret;
3752}
3753
3754/* need to called with the ext4 group lock held */
3755static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3756{
3757 struct ext4_prealloc_space *pa;
3758 struct list_head *cur, *tmp;
3759 int count = 0;
3760
3761 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3762 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3763 list_del(&pa->pa_group_list);
3764 count++;
3765 kmem_cache_free(ext4_pspace_cachep, pa);
3766 }
3767 return count;
3768}
3769
3770void ext4_mb_release(struct super_block *sb)
3771{
3772 ext4_group_t ngroups = ext4_get_groups_count(sb);
3773 ext4_group_t i;
3774 int num_meta_group_infos;
3775 struct ext4_group_info *grinfo, ***group_info;
3776 struct ext4_sb_info *sbi = EXT4_SB(sb);
3777 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3778 int count;
3779
3780 if (test_opt(sb, DISCARD)) {
3781 /*
3782 * wait the discard work to drain all of ext4_free_data
3783 */
3784 flush_work(&sbi->s_discard_work);
3785 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3786 }
3787
3788 if (sbi->s_group_info) {
3789 for (i = 0; i < ngroups; i++) {
3790 cond_resched();
3791 grinfo = ext4_get_group_info(sb, i);
3792 if (!grinfo)
3793 continue;
3794 mb_group_bb_bitmap_free(grinfo);
3795 ext4_lock_group(sb, i);
3796 count = ext4_mb_cleanup_pa(grinfo);
3797 if (count)
3798 mb_debug(sb, "mballoc: %d PAs left\n",
3799 count);
3800 ext4_unlock_group(sb, i);
3801 kmem_cache_free(cachep, grinfo);
3802 }
3803 num_meta_group_infos = (ngroups +
3804 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3805 EXT4_DESC_PER_BLOCK_BITS(sb);
3806 rcu_read_lock();
3807 group_info = rcu_dereference(sbi->s_group_info);
3808 for (i = 0; i < num_meta_group_infos; i++)
3809 kfree(group_info[i]);
3810 kvfree(group_info);
3811 rcu_read_unlock();
3812 }
3813 kfree(sbi->s_mb_avg_fragment_size);
3814 kfree(sbi->s_mb_avg_fragment_size_locks);
3815 kfree(sbi->s_mb_largest_free_orders);
3816 kfree(sbi->s_mb_largest_free_orders_locks);
3817 kfree(sbi->s_mb_offsets);
3818 kfree(sbi->s_mb_maxs);
3819 iput(sbi->s_buddy_cache);
3820 if (sbi->s_mb_stats) {
3821 ext4_msg(sb, KERN_INFO,
3822 "mballoc: %u blocks %u reqs (%u success)",
3823 atomic_read(&sbi->s_bal_allocated),
3824 atomic_read(&sbi->s_bal_reqs),
3825 atomic_read(&sbi->s_bal_success));
3826 ext4_msg(sb, KERN_INFO,
3827 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3828 "%u 2^N hits, %u breaks, %u lost",
3829 atomic_read(&sbi->s_bal_ex_scanned),
3830 atomic_read(&sbi->s_bal_groups_scanned),
3831 atomic_read(&sbi->s_bal_goals),
3832 atomic_read(&sbi->s_bal_2orders),
3833 atomic_read(&sbi->s_bal_breaks),
3834 atomic_read(&sbi->s_mb_lost_chunks));
3835 ext4_msg(sb, KERN_INFO,
3836 "mballoc: %u generated and it took %llu",
3837 atomic_read(&sbi->s_mb_buddies_generated),
3838 atomic64_read(&sbi->s_mb_generation_time));
3839 ext4_msg(sb, KERN_INFO,
3840 "mballoc: %u preallocated, %u discarded",
3841 atomic_read(&sbi->s_mb_preallocated),
3842 atomic_read(&sbi->s_mb_discarded));
3843 }
3844
3845 free_percpu(sbi->s_locality_groups);
3846}
3847
3848static inline int ext4_issue_discard(struct super_block *sb,
3849 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
3850{
3851 ext4_fsblk_t discard_block;
3852
3853 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3854 ext4_group_first_block_no(sb, block_group));
3855 count = EXT4_C2B(EXT4_SB(sb), count);
3856 trace_ext4_discard_blocks(sb,
3857 (unsigned long long) discard_block, count);
3858
3859 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3860}
3861
3862static void ext4_free_data_in_buddy(struct super_block *sb,
3863 struct ext4_free_data *entry)
3864{
3865 struct ext4_buddy e4b;
3866 struct ext4_group_info *db;
3867 int err, count = 0;
3868
3869 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3870 entry->efd_count, entry->efd_group, entry);
3871
3872 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3873 /* we expect to find existing buddy because it's pinned */
3874 BUG_ON(err != 0);
3875
3876 spin_lock(&EXT4_SB(sb)->s_md_lock);
3877 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3878 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3879
3880 db = e4b.bd_info;
3881 /* there are blocks to put in buddy to make them really free */
3882 count += entry->efd_count;
3883 ext4_lock_group(sb, entry->efd_group);
3884 /* Take it out of per group rb tree */
3885 rb_erase(&entry->efd_node, &(db->bb_free_root));
3886 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3887
3888 /*
3889 * Clear the trimmed flag for the group so that the next
3890 * ext4_trim_fs can trim it.
3891 */
3892 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3893
3894 if (!db->bb_free_root.rb_node) {
3895 /* No more items in the per group rb tree
3896 * balance refcounts from ext4_mb_free_metadata()
3897 */
3898 folio_put(e4b.bd_buddy_folio);
3899 folio_put(e4b.bd_bitmap_folio);
3900 }
3901 ext4_unlock_group(sb, entry->efd_group);
3902 ext4_mb_unload_buddy(&e4b);
3903
3904 mb_debug(sb, "freed %d blocks in 1 structures\n", count);
3905}
3906
3907/*
3908 * This function is called by the jbd2 layer once the commit has finished,
3909 * so we know we can free the blocks that were released with that commit.
3910 */
3911void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3912{
3913 struct ext4_sb_info *sbi = EXT4_SB(sb);
3914 struct ext4_free_data *entry, *tmp;
3915 LIST_HEAD(freed_data_list);
3916 struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1];
3917 bool wake;
3918
3919 list_replace_init(s_freed_head, &freed_data_list);
3920
3921 list_for_each_entry(entry, &freed_data_list, efd_list)
3922 ext4_free_data_in_buddy(sb, entry);
3923
3924 if (test_opt(sb, DISCARD)) {
3925 spin_lock(&sbi->s_md_lock);
3926 wake = list_empty(&sbi->s_discard_list);
3927 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3928 spin_unlock(&sbi->s_md_lock);
3929 if (wake)
3930 queue_work(system_unbound_wq, &sbi->s_discard_work);
3931 } else {
3932 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3933 kmem_cache_free(ext4_free_data_cachep, entry);
3934 }
3935}
3936
3937int __init ext4_init_mballoc(void)
3938{
3939 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3940 SLAB_RECLAIM_ACCOUNT);
3941 if (ext4_pspace_cachep == NULL)
3942 goto out;
3943
3944 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3945 SLAB_RECLAIM_ACCOUNT);
3946 if (ext4_ac_cachep == NULL)
3947 goto out_pa_free;
3948
3949 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3950 SLAB_RECLAIM_ACCOUNT);
3951 if (ext4_free_data_cachep == NULL)
3952 goto out_ac_free;
3953
3954 return 0;
3955
3956out_ac_free:
3957 kmem_cache_destroy(ext4_ac_cachep);
3958out_pa_free:
3959 kmem_cache_destroy(ext4_pspace_cachep);
3960out:
3961 return -ENOMEM;
3962}
3963
3964void ext4_exit_mballoc(void)
3965{
3966 /*
3967 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3968 * before destroying the slab cache.
3969 */
3970 rcu_barrier();
3971 kmem_cache_destroy(ext4_pspace_cachep);
3972 kmem_cache_destroy(ext4_ac_cachep);
3973 kmem_cache_destroy(ext4_free_data_cachep);
3974 ext4_groupinfo_destroy_slabs();
3975}
3976
3977#define EXT4_MB_BITMAP_MARKED_CHECK 0x0001
3978#define EXT4_MB_SYNC_UPDATE 0x0002
3979static int
3980ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state,
3981 ext4_group_t group, ext4_grpblk_t blkoff,
3982 ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed)
3983{
3984 struct ext4_sb_info *sbi = EXT4_SB(sb);
3985 struct buffer_head *bitmap_bh = NULL;
3986 struct ext4_group_desc *gdp;
3987 struct buffer_head *gdp_bh;
3988 int err;
3989 unsigned int i, already, changed = len;
3990
3991 KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context,
3992 handle, sb, state, group, blkoff, len,
3993 flags, ret_changed);
3994
3995 if (ret_changed)
3996 *ret_changed = 0;
3997 bitmap_bh = ext4_read_block_bitmap(sb, group);
3998 if (IS_ERR(bitmap_bh))
3999 return PTR_ERR(bitmap_bh);
4000
4001 if (handle) {
4002 BUFFER_TRACE(bitmap_bh, "getting write access");
4003 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
4004 EXT4_JTR_NONE);
4005 if (err)
4006 goto out_err;
4007 }
4008
4009 err = -EIO;
4010 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
4011 if (!gdp)
4012 goto out_err;
4013
4014 if (handle) {
4015 BUFFER_TRACE(gdp_bh, "get_write_access");
4016 err = ext4_journal_get_write_access(handle, sb, gdp_bh,
4017 EXT4_JTR_NONE);
4018 if (err)
4019 goto out_err;
4020 }
4021
4022 ext4_lock_group(sb, group);
4023 if (ext4_has_group_desc_csum(sb) &&
4024 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
4025 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
4026 ext4_free_group_clusters_set(sb, gdp,
4027 ext4_free_clusters_after_init(sb, group, gdp));
4028 }
4029
4030 if (flags & EXT4_MB_BITMAP_MARKED_CHECK) {
4031 already = 0;
4032 for (i = 0; i < len; i++)
4033 if (mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
4034 state)
4035 already++;
4036 changed = len - already;
4037 }
4038
4039 if (state) {
4040 mb_set_bits(bitmap_bh->b_data, blkoff, len);
4041 ext4_free_group_clusters_set(sb, gdp,
4042 ext4_free_group_clusters(sb, gdp) - changed);
4043 } else {
4044 mb_clear_bits(bitmap_bh->b_data, blkoff, len);
4045 ext4_free_group_clusters_set(sb, gdp,
4046 ext4_free_group_clusters(sb, gdp) + changed);
4047 }
4048
4049 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh);
4050 ext4_group_desc_csum_set(sb, group, gdp);
4051 ext4_unlock_group(sb, group);
4052 if (ret_changed)
4053 *ret_changed = changed;
4054
4055 if (sbi->s_log_groups_per_flex) {
4056 ext4_group_t flex_group = ext4_flex_group(sbi, group);
4057 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
4058 s_flex_groups, flex_group);
4059
4060 if (state)
4061 atomic64_sub(changed, &fg->free_clusters);
4062 else
4063 atomic64_add(changed, &fg->free_clusters);
4064 }
4065
4066 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4067 if (err)
4068 goto out_err;
4069 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
4070 if (err)
4071 goto out_err;
4072
4073 if (flags & EXT4_MB_SYNC_UPDATE) {
4074 sync_dirty_buffer(bitmap_bh);
4075 sync_dirty_buffer(gdp_bh);
4076 }
4077
4078out_err:
4079 brelse(bitmap_bh);
4080 return err;
4081}
4082
4083/*
4084 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
4085 * Returns 0 if success or error code
4086 */
4087static noinline_for_stack int
4088ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
4089 handle_t *handle, unsigned int reserv_clstrs)
4090{
4091 struct ext4_group_desc *gdp;
4092 struct ext4_sb_info *sbi;
4093 struct super_block *sb;
4094 ext4_fsblk_t block;
4095 int err, len;
4096 int flags = 0;
4097 ext4_grpblk_t changed;
4098
4099 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4100 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4101
4102 sb = ac->ac_sb;
4103 sbi = EXT4_SB(sb);
4104
4105 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL);
4106 if (!gdp)
4107 return -EIO;
4108 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
4109 ext4_free_group_clusters(sb, gdp));
4110
4111 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4112 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4113 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
4114 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
4115 "fs metadata", block, block+len);
4116 /* File system mounted not to panic on error
4117 * Fix the bitmap and return EFSCORRUPTED
4118 * We leak some of the blocks here.
4119 */
4120 err = ext4_mb_mark_context(handle, sb, true,
4121 ac->ac_b_ex.fe_group,
4122 ac->ac_b_ex.fe_start,
4123 ac->ac_b_ex.fe_len,
4124 0, NULL);
4125 if (!err)
4126 err = -EFSCORRUPTED;
4127 return err;
4128 }
4129
4130#ifdef AGGRESSIVE_CHECK
4131 flags |= EXT4_MB_BITMAP_MARKED_CHECK;
4132#endif
4133 err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group,
4134 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len,
4135 flags, &changed);
4136
4137 if (err && changed == 0)
4138 return err;
4139
4140#ifdef AGGRESSIVE_CHECK
4141 BUG_ON(changed != ac->ac_b_ex.fe_len);
4142#endif
4143 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
4144 /*
4145 * Now reduce the dirty block count also. Should not go negative
4146 */
4147 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
4148 /* release all the reserved blocks if non delalloc */
4149 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4150 reserv_clstrs);
4151
4152 return err;
4153}
4154
4155/*
4156 * Idempotent helper for Ext4 fast commit replay path to set the state of
4157 * blocks in bitmaps and update counters.
4158 */
4159void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
4160 int len, bool state)
4161{
4162 struct ext4_sb_info *sbi = EXT4_SB(sb);
4163 ext4_group_t group;
4164 ext4_grpblk_t blkoff;
4165 int err = 0;
4166 unsigned int clen, thisgrp_len;
4167
4168 while (len > 0) {
4169 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
4170
4171 /*
4172 * Check to see if we are freeing blocks across a group
4173 * boundary.
4174 * In case of flex_bg, this can happen that (block, len) may
4175 * span across more than one group. In that case we need to
4176 * get the corresponding group metadata to work with.
4177 * For this we have goto again loop.
4178 */
4179 thisgrp_len = min_t(unsigned int, (unsigned int)len,
4180 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
4181 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
4182
4183 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
4184 ext4_error(sb, "Marking blocks in system zone - "
4185 "Block = %llu, len = %u",
4186 block, thisgrp_len);
4187 break;
4188 }
4189
4190 err = ext4_mb_mark_context(NULL, sb, state,
4191 group, blkoff, clen,
4192 EXT4_MB_BITMAP_MARKED_CHECK |
4193 EXT4_MB_SYNC_UPDATE,
4194 NULL);
4195 if (err)
4196 break;
4197
4198 block += thisgrp_len;
4199 len -= thisgrp_len;
4200 BUG_ON(len < 0);
4201 }
4202}
4203
4204/*
4205 * here we normalize request for locality group
4206 * Group request are normalized to s_mb_group_prealloc, which goes to
4207 * s_strip if we set the same via mount option.
4208 * s_mb_group_prealloc can be configured via
4209 * /sys/fs/ext4/<partition>/mb_group_prealloc
4210 *
4211 * XXX: should we try to preallocate more than the group has now?
4212 */
4213static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4214{
4215 struct super_block *sb = ac->ac_sb;
4216 struct ext4_locality_group *lg = ac->ac_lg;
4217
4218 BUG_ON(lg == NULL);
4219 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4220 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4221}
4222
4223/*
4224 * This function returns the next element to look at during inode
4225 * PA rbtree walk. We assume that we have held the inode PA rbtree lock
4226 * (ei->i_prealloc_lock)
4227 *
4228 * new_start The start of the range we want to compare
4229 * cur_start The existing start that we are comparing against
4230 * node The node of the rb_tree
4231 */
4232static inline struct rb_node*
4233ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node)
4234{
4235 if (new_start < cur_start)
4236 return node->rb_left;
4237 else
4238 return node->rb_right;
4239}
4240
4241static inline void
4242ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
4243 ext4_lblk_t start, loff_t end)
4244{
4245 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4246 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4247 struct ext4_prealloc_space *tmp_pa;
4248 ext4_lblk_t tmp_pa_start;
4249 loff_t tmp_pa_end;
4250 struct rb_node *iter;
4251
4252 read_lock(&ei->i_prealloc_lock);
4253 for (iter = ei->i_prealloc_node.rb_node; iter;
4254 iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) {
4255 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4256 pa_node.inode_node);
4257 tmp_pa_start = tmp_pa->pa_lstart;
4258 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4259
4260 spin_lock(&tmp_pa->pa_lock);
4261 if (tmp_pa->pa_deleted == 0)
4262 BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
4263 spin_unlock(&tmp_pa->pa_lock);
4264 }
4265 read_unlock(&ei->i_prealloc_lock);
4266}
4267
4268/*
4269 * Given an allocation context "ac" and a range "start", "end", check
4270 * and adjust boundaries if the range overlaps with any of the existing
4271 * preallocatoins stored in the corresponding inode of the allocation context.
4272 *
4273 * Parameters:
4274 * ac allocation context
4275 * start start of the new range
4276 * end end of the new range
4277 */
4278static inline void
4279ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
4280 ext4_lblk_t *start, loff_t *end)
4281{
4282 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4283 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4284 struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL;
4285 struct rb_node *iter;
4286 ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1;
4287 loff_t new_end, tmp_pa_end, left_pa_end = -1;
4288
4289 new_start = *start;
4290 new_end = *end;
4291
4292 /*
4293 * Adjust the normalized range so that it doesn't overlap with any
4294 * existing preallocated blocks(PAs). Make sure to hold the rbtree lock
4295 * so it doesn't change underneath us.
4296 */
4297 read_lock(&ei->i_prealloc_lock);
4298
4299 /* Step 1: find any one immediate neighboring PA of the normalized range */
4300 for (iter = ei->i_prealloc_node.rb_node; iter;
4301 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4302 tmp_pa_start, iter)) {
4303 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4304 pa_node.inode_node);
4305 tmp_pa_start = tmp_pa->pa_lstart;
4306 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4307
4308 /* PA must not overlap original request */
4309 spin_lock(&tmp_pa->pa_lock);
4310 if (tmp_pa->pa_deleted == 0)
4311 BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
4312 ac->ac_o_ex.fe_logical < tmp_pa_start));
4313 spin_unlock(&tmp_pa->pa_lock);
4314 }
4315
4316 /*
4317 * Step 2: check if the found PA is left or right neighbor and
4318 * get the other neighbor
4319 */
4320 if (tmp_pa) {
4321 if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) {
4322 struct rb_node *tmp;
4323
4324 left_pa = tmp_pa;
4325 tmp = rb_next(&left_pa->pa_node.inode_node);
4326 if (tmp) {
4327 right_pa = rb_entry(tmp,
4328 struct ext4_prealloc_space,
4329 pa_node.inode_node);
4330 }
4331 } else {
4332 struct rb_node *tmp;
4333
4334 right_pa = tmp_pa;
4335 tmp = rb_prev(&right_pa->pa_node.inode_node);
4336 if (tmp) {
4337 left_pa = rb_entry(tmp,
4338 struct ext4_prealloc_space,
4339 pa_node.inode_node);
4340 }
4341 }
4342 }
4343
4344 /* Step 3: get the non deleted neighbors */
4345 if (left_pa) {
4346 for (iter = &left_pa->pa_node.inode_node;;
4347 iter = rb_prev(iter)) {
4348 if (!iter) {
4349 left_pa = NULL;
4350 break;
4351 }
4352
4353 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4354 pa_node.inode_node);
4355 left_pa = tmp_pa;
4356 spin_lock(&tmp_pa->pa_lock);
4357 if (tmp_pa->pa_deleted == 0) {
4358 spin_unlock(&tmp_pa->pa_lock);
4359 break;
4360 }
4361 spin_unlock(&tmp_pa->pa_lock);
4362 }
4363 }
4364
4365 if (right_pa) {
4366 for (iter = &right_pa->pa_node.inode_node;;
4367 iter = rb_next(iter)) {
4368 if (!iter) {
4369 right_pa = NULL;
4370 break;
4371 }
4372
4373 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4374 pa_node.inode_node);
4375 right_pa = tmp_pa;
4376 spin_lock(&tmp_pa->pa_lock);
4377 if (tmp_pa->pa_deleted == 0) {
4378 spin_unlock(&tmp_pa->pa_lock);
4379 break;
4380 }
4381 spin_unlock(&tmp_pa->pa_lock);
4382 }
4383 }
4384
4385 if (left_pa) {
4386 left_pa_end = pa_logical_end(sbi, left_pa);
4387 BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical);
4388 }
4389
4390 if (right_pa) {
4391 right_pa_start = right_pa->pa_lstart;
4392 BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical);
4393 }
4394
4395 /* Step 4: trim our normalized range to not overlap with the neighbors */
4396 if (left_pa) {
4397 if (left_pa_end > new_start)
4398 new_start = left_pa_end;
4399 }
4400
4401 if (right_pa) {
4402 if (right_pa_start < new_end)
4403 new_end = right_pa_start;
4404 }
4405 read_unlock(&ei->i_prealloc_lock);
4406
4407 /* XXX: extra loop to check we really don't overlap preallocations */
4408 ext4_mb_pa_assert_overlap(ac, new_start, new_end);
4409
4410 *start = new_start;
4411 *end = new_end;
4412}
4413
4414/*
4415 * Normalization means making request better in terms of
4416 * size and alignment
4417 */
4418static noinline_for_stack void
4419ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4420 struct ext4_allocation_request *ar)
4421{
4422 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4423 struct ext4_super_block *es = sbi->s_es;
4424 int bsbits, max;
4425 loff_t size, start_off, end;
4426 loff_t orig_size __maybe_unused;
4427 ext4_lblk_t start;
4428
4429 /* do normalize only data requests, metadata requests
4430 do not need preallocation */
4431 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4432 return;
4433
4434 /* sometime caller may want exact blocks */
4435 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4436 return;
4437
4438 /* caller may indicate that preallocation isn't
4439 * required (it's a tail, for example) */
4440 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4441 return;
4442
4443 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4444 ext4_mb_normalize_group_request(ac);
4445 return ;
4446 }
4447
4448 bsbits = ac->ac_sb->s_blocksize_bits;
4449
4450 /* first, let's learn actual file size
4451 * given current request is allocated */
4452 size = extent_logical_end(sbi, &ac->ac_o_ex);
4453 size = size << bsbits;
4454 if (size < i_size_read(ac->ac_inode))
4455 size = i_size_read(ac->ac_inode);
4456 orig_size = size;
4457
4458 /* max size of free chunks */
4459 max = 2 << bsbits;
4460
4461#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4462 (req <= (size) || max <= (chunk_size))
4463
4464 /* first, try to predict filesize */
4465 /* XXX: should this table be tunable? */
4466 start_off = 0;
4467 if (size <= 16 * 1024) {
4468 size = 16 * 1024;
4469 } else if (size <= 32 * 1024) {
4470 size = 32 * 1024;
4471 } else if (size <= 64 * 1024) {
4472 size = 64 * 1024;
4473 } else if (size <= 128 * 1024) {
4474 size = 128 * 1024;
4475 } else if (size <= 256 * 1024) {
4476 size = 256 * 1024;
4477 } else if (size <= 512 * 1024) {
4478 size = 512 * 1024;
4479 } else if (size <= 1024 * 1024) {
4480 size = 1024 * 1024;
4481 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4482 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4483 (21 - bsbits)) << 21;
4484 size = 2 * 1024 * 1024;
4485 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4486 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4487 (22 - bsbits)) << 22;
4488 size = 4 * 1024 * 1024;
4489 } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len),
4490 (8<<20)>>bsbits, max, 8 * 1024)) {
4491 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4492 (23 - bsbits)) << 23;
4493 size = 8 * 1024 * 1024;
4494 } else {
4495 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4496 size = (loff_t) EXT4_C2B(sbi,
4497 ac->ac_o_ex.fe_len) << bsbits;
4498 }
4499 size = size >> bsbits;
4500 start = start_off >> bsbits;
4501
4502 /*
4503 * For tiny groups (smaller than 8MB) the chosen allocation
4504 * alignment may be larger than group size. Make sure the
4505 * alignment does not move allocation to a different group which
4506 * makes mballoc fail assertions later.
4507 */
4508 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4509 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4510
4511 /* avoid unnecessary preallocation that may trigger assertions */
4512 if (start + size > EXT_MAX_BLOCKS)
4513 size = EXT_MAX_BLOCKS - start;
4514
4515 /* don't cover already allocated blocks in selected range */
4516 if (ar->pleft && start <= ar->lleft) {
4517 size -= ar->lleft + 1 - start;
4518 start = ar->lleft + 1;
4519 }
4520 if (ar->pright && start + size - 1 >= ar->lright)
4521 size -= start + size - ar->lright;
4522
4523 /*
4524 * Trim allocation request for filesystems with artificially small
4525 * groups.
4526 */
4527 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4528 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4529
4530 end = start + size;
4531
4532 ext4_mb_pa_adjust_overlap(ac, &start, &end);
4533
4534 size = end - start;
4535
4536 /*
4537 * In this function "start" and "size" are normalized for better
4538 * alignment and length such that we could preallocate more blocks.
4539 * This normalization is done such that original request of
4540 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4541 * "size" boundaries.
4542 * (Note fe_len can be relaxed since FS block allocation API does not
4543 * provide gurantee on number of contiguous blocks allocation since that
4544 * depends upon free space left, etc).
4545 * In case of inode pa, later we use the allocated blocks
4546 * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated
4547 * range of goal/best blocks [start, size] to put it at the
4548 * ac_o_ex.fe_logical extent of this inode.
4549 * (See ext4_mb_use_inode_pa() for more details)
4550 */
4551 if (start + size <= ac->ac_o_ex.fe_logical ||
4552 start > ac->ac_o_ex.fe_logical) {
4553 ext4_msg(ac->ac_sb, KERN_ERR,
4554 "start %lu, size %lu, fe_logical %lu",
4555 (unsigned long) start, (unsigned long) size,
4556 (unsigned long) ac->ac_o_ex.fe_logical);
4557 BUG();
4558 }
4559 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4560
4561 /* now prepare goal request */
4562
4563 /* XXX: is it better to align blocks WRT to logical
4564 * placement or satisfy big request as is */
4565 ac->ac_g_ex.fe_logical = start;
4566 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4567 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
4568
4569 /* define goal start in order to merge */
4570 if (ar->pright && (ar->lright == (start + size)) &&
4571 ar->pright >= size &&
4572 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
4573 /* merge to the right */
4574 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4575 &ac->ac_g_ex.fe_group,
4576 &ac->ac_g_ex.fe_start);
4577 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4578 }
4579 if (ar->pleft && (ar->lleft + 1 == start) &&
4580 ar->pleft + 1 < ext4_blocks_count(es)) {
4581 /* merge to the left */
4582 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4583 &ac->ac_g_ex.fe_group,
4584 &ac->ac_g_ex.fe_start);
4585 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4586 }
4587
4588 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4589 orig_size, start);
4590}
4591
4592static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4593{
4594 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4595
4596 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4597 atomic_inc(&sbi->s_bal_reqs);
4598 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4599 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4600 atomic_inc(&sbi->s_bal_success);
4601
4602 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4603 for (int i=0; i<EXT4_MB_NUM_CRS; i++) {
4604 atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]);
4605 }
4606
4607 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4608 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4609 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4610 atomic_inc(&sbi->s_bal_goals);
4611 /* did we allocate as much as normalizer originally wanted? */
4612 if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len)
4613 atomic_inc(&sbi->s_bal_len_goals);
4614
4615 if (ac->ac_found > sbi->s_mb_max_to_scan)
4616 atomic_inc(&sbi->s_bal_breaks);
4617 }
4618
4619 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4620 trace_ext4_mballoc_alloc(ac);
4621 else
4622 trace_ext4_mballoc_prealloc(ac);
4623}
4624
4625/*
4626 * Called on failure; free up any blocks from the inode PA for this
4627 * context. We don't need this for MB_GROUP_PA because we only change
4628 * pa_free in ext4_mb_release_context(), but on failure, we've already
4629 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4630 */
4631static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4632{
4633 struct ext4_prealloc_space *pa = ac->ac_pa;
4634 struct ext4_buddy e4b;
4635 int err;
4636
4637 if (pa == NULL) {
4638 if (ac->ac_f_ex.fe_len == 0)
4639 return;
4640 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4641 if (WARN_RATELIMIT(err,
4642 "ext4: mb_load_buddy failed (%d)", err))
4643 /*
4644 * This should never happen since we pin the
4645 * pages in the ext4_allocation_context so
4646 * ext4_mb_load_buddy() should never fail.
4647 */
4648 return;
4649 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4650 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4651 ac->ac_f_ex.fe_len);
4652 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4653 ext4_mb_unload_buddy(&e4b);
4654 return;
4655 }
4656 if (pa->pa_type == MB_INODE_PA) {
4657 spin_lock(&pa->pa_lock);
4658 pa->pa_free += ac->ac_b_ex.fe_len;
4659 spin_unlock(&pa->pa_lock);
4660 }
4661}
4662
4663/*
4664 * use blocks preallocated to inode
4665 */
4666static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4667 struct ext4_prealloc_space *pa)
4668{
4669 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4670 ext4_fsblk_t start;
4671 ext4_fsblk_t end;
4672 int len;
4673
4674 /* found preallocated blocks, use them */
4675 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4676 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4677 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4678 len = EXT4_NUM_B2C(sbi, end - start);
4679 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4680 &ac->ac_b_ex.fe_start);
4681 ac->ac_b_ex.fe_len = len;
4682 ac->ac_status = AC_STATUS_FOUND;
4683 ac->ac_pa = pa;
4684
4685 BUG_ON(start < pa->pa_pstart);
4686 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4687 BUG_ON(pa->pa_free < len);
4688 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4689 pa->pa_free -= len;
4690
4691 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4692}
4693
4694/*
4695 * use blocks preallocated to locality group
4696 */
4697static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4698 struct ext4_prealloc_space *pa)
4699{
4700 unsigned int len = ac->ac_o_ex.fe_len;
4701
4702 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4703 &ac->ac_b_ex.fe_group,
4704 &ac->ac_b_ex.fe_start);
4705 ac->ac_b_ex.fe_len = len;
4706 ac->ac_status = AC_STATUS_FOUND;
4707 ac->ac_pa = pa;
4708
4709 /* we don't correct pa_pstart or pa_len here to avoid
4710 * possible race when the group is being loaded concurrently
4711 * instead we correct pa later, after blocks are marked
4712 * in on-disk bitmap -- see ext4_mb_release_context()
4713 * Other CPUs are prevented from allocating from this pa by lg_mutex
4714 */
4715 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4716 pa->pa_lstart, len, pa);
4717}
4718
4719/*
4720 * Return the prealloc space that have minimal distance
4721 * from the goal block. @cpa is the prealloc
4722 * space that is having currently known minimal distance
4723 * from the goal block.
4724 */
4725static struct ext4_prealloc_space *
4726ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4727 struct ext4_prealloc_space *pa,
4728 struct ext4_prealloc_space *cpa)
4729{
4730 ext4_fsblk_t cur_distance, new_distance;
4731
4732 if (cpa == NULL) {
4733 atomic_inc(&pa->pa_count);
4734 return pa;
4735 }
4736 cur_distance = abs(goal_block - cpa->pa_pstart);
4737 new_distance = abs(goal_block - pa->pa_pstart);
4738
4739 if (cur_distance <= new_distance)
4740 return cpa;
4741
4742 /* drop the previous reference */
4743 atomic_dec(&cpa->pa_count);
4744 atomic_inc(&pa->pa_count);
4745 return pa;
4746}
4747
4748/*
4749 * check if found pa meets EXT4_MB_HINT_GOAL_ONLY
4750 */
4751static bool
4752ext4_mb_pa_goal_check(struct ext4_allocation_context *ac,
4753 struct ext4_prealloc_space *pa)
4754{
4755 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4756 ext4_fsblk_t start;
4757
4758 if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)))
4759 return true;
4760
4761 /*
4762 * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted
4763 * in ext4_mb_normalize_request and will keep same with ac_o_ex
4764 * from ext4_mb_initialize_context. Choose ac_g_ex here to keep
4765 * consistent with ext4_mb_find_by_goal.
4766 */
4767 start = pa->pa_pstart +
4768 (ac->ac_g_ex.fe_logical - pa->pa_lstart);
4769 if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start)
4770 return false;
4771
4772 if (ac->ac_g_ex.fe_len > pa->pa_len -
4773 EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart))
4774 return false;
4775
4776 return true;
4777}
4778
4779/*
4780 * search goal blocks in preallocated space
4781 */
4782static noinline_for_stack bool
4783ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4784{
4785 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4786 int order, i;
4787 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4788 struct ext4_locality_group *lg;
4789 struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL;
4790 struct rb_node *iter;
4791 ext4_fsblk_t goal_block;
4792
4793 /* only data can be preallocated */
4794 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4795 return false;
4796
4797 /*
4798 * first, try per-file preallocation by searching the inode pa rbtree.
4799 *
4800 * Here, we can't do a direct traversal of the tree because
4801 * ext4_mb_discard_group_preallocation() can paralelly mark the pa
4802 * deleted and that can cause direct traversal to skip some entries.
4803 */
4804 read_lock(&ei->i_prealloc_lock);
4805
4806 if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) {
4807 goto try_group_pa;
4808 }
4809
4810 /*
4811 * Step 1: Find a pa with logical start immediately adjacent to the
4812 * original logical start. This could be on the left or right.
4813 *
4814 * (tmp_pa->pa_lstart never changes so we can skip locking for it).
4815 */
4816 for (iter = ei->i_prealloc_node.rb_node; iter;
4817 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4818 tmp_pa->pa_lstart, iter)) {
4819 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4820 pa_node.inode_node);
4821 }
4822
4823 /*
4824 * Step 2: The adjacent pa might be to the right of logical start, find
4825 * the left adjacent pa. After this step we'd have a valid tmp_pa whose
4826 * logical start is towards the left of original request's logical start
4827 */
4828 if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4829 struct rb_node *tmp;
4830 tmp = rb_prev(&tmp_pa->pa_node.inode_node);
4831
4832 if (tmp) {
4833 tmp_pa = rb_entry(tmp, struct ext4_prealloc_space,
4834 pa_node.inode_node);
4835 } else {
4836 /*
4837 * If there is no adjacent pa to the left then finding
4838 * an overlapping pa is not possible hence stop searching
4839 * inode pa tree
4840 */
4841 goto try_group_pa;
4842 }
4843 }
4844
4845 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4846
4847 /*
4848 * Step 3: If the left adjacent pa is deleted, keep moving left to find
4849 * the first non deleted adjacent pa. After this step we should have a
4850 * valid tmp_pa which is guaranteed to be non deleted.
4851 */
4852 for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) {
4853 if (!iter) {
4854 /*
4855 * no non deleted left adjacent pa, so stop searching
4856 * inode pa tree
4857 */
4858 goto try_group_pa;
4859 }
4860 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4861 pa_node.inode_node);
4862 spin_lock(&tmp_pa->pa_lock);
4863 if (tmp_pa->pa_deleted == 0) {
4864 /*
4865 * We will keep holding the pa_lock from
4866 * this point on because we don't want group discard
4867 * to delete this pa underneath us. Since group
4868 * discard is anyways an ENOSPC operation it
4869 * should be okay for it to wait a few more cycles.
4870 */
4871 break;
4872 } else {
4873 spin_unlock(&tmp_pa->pa_lock);
4874 }
4875 }
4876
4877 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4878 BUG_ON(tmp_pa->pa_deleted == 1);
4879
4880 /*
4881 * Step 4: We now have the non deleted left adjacent pa. Only this
4882 * pa can possibly satisfy the request hence check if it overlaps
4883 * original logical start and stop searching if it doesn't.
4884 */
4885 if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) {
4886 spin_unlock(&tmp_pa->pa_lock);
4887 goto try_group_pa;
4888 }
4889
4890 /* non-extent files can't have physical blocks past 2^32 */
4891 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4892 (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
4893 EXT4_MAX_BLOCK_FILE_PHYS)) {
4894 /*
4895 * Since PAs don't overlap, we won't find any other PA to
4896 * satisfy this.
4897 */
4898 spin_unlock(&tmp_pa->pa_lock);
4899 goto try_group_pa;
4900 }
4901
4902 if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) {
4903 atomic_inc(&tmp_pa->pa_count);
4904 ext4_mb_use_inode_pa(ac, tmp_pa);
4905 spin_unlock(&tmp_pa->pa_lock);
4906 read_unlock(&ei->i_prealloc_lock);
4907 return true;
4908 } else {
4909 /*
4910 * We found a valid overlapping pa but couldn't use it because
4911 * it had no free blocks. This should ideally never happen
4912 * because:
4913 *
4914 * 1. When a new inode pa is added to rbtree it must have
4915 * pa_free > 0 since otherwise we won't actually need
4916 * preallocation.
4917 *
4918 * 2. An inode pa that is in the rbtree can only have it's
4919 * pa_free become zero when another thread calls:
4920 * ext4_mb_new_blocks
4921 * ext4_mb_use_preallocated
4922 * ext4_mb_use_inode_pa
4923 *
4924 * 3. Further, after the above calls make pa_free == 0, we will
4925 * immediately remove it from the rbtree in:
4926 * ext4_mb_new_blocks
4927 * ext4_mb_release_context
4928 * ext4_mb_put_pa
4929 *
4930 * 4. Since the pa_free becoming 0 and pa_free getting removed
4931 * from tree both happen in ext4_mb_new_blocks, which is always
4932 * called with i_data_sem held for data allocations, we can be
4933 * sure that another process will never see a pa in rbtree with
4934 * pa_free == 0.
4935 */
4936 WARN_ON_ONCE(tmp_pa->pa_free == 0);
4937 }
4938 spin_unlock(&tmp_pa->pa_lock);
4939try_group_pa:
4940 read_unlock(&ei->i_prealloc_lock);
4941
4942 /* can we use group allocation? */
4943 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4944 return false;
4945
4946 /* inode may have no locality group for some reason */
4947 lg = ac->ac_lg;
4948 if (lg == NULL)
4949 return false;
4950 order = fls(ac->ac_o_ex.fe_len) - 1;
4951 if (order > PREALLOC_TB_SIZE - 1)
4952 /* The max size of hash table is PREALLOC_TB_SIZE */
4953 order = PREALLOC_TB_SIZE - 1;
4954
4955 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4956 /*
4957 * search for the prealloc space that is having
4958 * minimal distance from the goal block.
4959 */
4960 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4961 rcu_read_lock();
4962 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
4963 pa_node.lg_list) {
4964 spin_lock(&tmp_pa->pa_lock);
4965 if (tmp_pa->pa_deleted == 0 &&
4966 tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
4967
4968 cpa = ext4_mb_check_group_pa(goal_block,
4969 tmp_pa, cpa);
4970 }
4971 spin_unlock(&tmp_pa->pa_lock);
4972 }
4973 rcu_read_unlock();
4974 }
4975 if (cpa) {
4976 ext4_mb_use_group_pa(ac, cpa);
4977 return true;
4978 }
4979 return false;
4980}
4981
4982/*
4983 * the function goes through all preallocation in this group and marks them
4984 * used in in-core bitmap. buddy must be generated from this bitmap
4985 * Need to be called with ext4 group lock held
4986 */
4987static noinline_for_stack
4988void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4989 ext4_group_t group)
4990{
4991 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4992 struct ext4_prealloc_space *pa;
4993 struct list_head *cur;
4994 ext4_group_t groupnr;
4995 ext4_grpblk_t start;
4996 int preallocated = 0;
4997 int len;
4998
4999 if (!grp)
5000 return;
5001
5002 /* all form of preallocation discards first load group,
5003 * so the only competing code is preallocation use.
5004 * we don't need any locking here
5005 * notice we do NOT ignore preallocations with pa_deleted
5006 * otherwise we could leave used blocks available for
5007 * allocation in buddy when concurrent ext4_mb_put_pa()
5008 * is dropping preallocation
5009 */
5010 list_for_each(cur, &grp->bb_prealloc_list) {
5011 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
5012 spin_lock(&pa->pa_lock);
5013 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5014 &groupnr, &start);
5015 len = pa->pa_len;
5016 spin_unlock(&pa->pa_lock);
5017 if (unlikely(len == 0))
5018 continue;
5019 BUG_ON(groupnr != group);
5020 mb_set_bits(bitmap, start, len);
5021 preallocated += len;
5022 }
5023 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
5024}
5025
5026static void ext4_mb_mark_pa_deleted(struct super_block *sb,
5027 struct ext4_prealloc_space *pa)
5028{
5029 struct ext4_inode_info *ei;
5030
5031 if (pa->pa_deleted) {
5032 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
5033 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
5034 pa->pa_len);
5035 return;
5036 }
5037
5038 pa->pa_deleted = 1;
5039
5040 if (pa->pa_type == MB_INODE_PA) {
5041 ei = EXT4_I(pa->pa_inode);
5042 atomic_dec(&ei->i_prealloc_active);
5043 }
5044}
5045
5046static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
5047{
5048 BUG_ON(!pa);
5049 BUG_ON(atomic_read(&pa->pa_count));
5050 BUG_ON(pa->pa_deleted == 0);
5051 kmem_cache_free(ext4_pspace_cachep, pa);
5052}
5053
5054static void ext4_mb_pa_callback(struct rcu_head *head)
5055{
5056 struct ext4_prealloc_space *pa;
5057
5058 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
5059 ext4_mb_pa_free(pa);
5060}
5061
5062/*
5063 * drops a reference to preallocated space descriptor
5064 * if this was the last reference and the space is consumed
5065 */
5066static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
5067 struct super_block *sb, struct ext4_prealloc_space *pa)
5068{
5069 ext4_group_t grp;
5070 ext4_fsblk_t grp_blk;
5071 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
5072
5073 /* in this short window concurrent discard can set pa_deleted */
5074 spin_lock(&pa->pa_lock);
5075 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
5076 spin_unlock(&pa->pa_lock);
5077 return;
5078 }
5079
5080 if (pa->pa_deleted == 1) {
5081 spin_unlock(&pa->pa_lock);
5082 return;
5083 }
5084
5085 ext4_mb_mark_pa_deleted(sb, pa);
5086 spin_unlock(&pa->pa_lock);
5087
5088 grp_blk = pa->pa_pstart;
5089 /*
5090 * If doing group-based preallocation, pa_pstart may be in the
5091 * next group when pa is used up
5092 */
5093 if (pa->pa_type == MB_GROUP_PA)
5094 grp_blk--;
5095
5096 grp = ext4_get_group_number(sb, grp_blk);
5097
5098 /*
5099 * possible race:
5100 *
5101 * P1 (buddy init) P2 (regular allocation)
5102 * find block B in PA
5103 * copy on-disk bitmap to buddy
5104 * mark B in on-disk bitmap
5105 * drop PA from group
5106 * mark all PAs in buddy
5107 *
5108 * thus, P1 initializes buddy with B available. to prevent this
5109 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
5110 * against that pair
5111 */
5112 ext4_lock_group(sb, grp);
5113 list_del(&pa->pa_group_list);
5114 ext4_unlock_group(sb, grp);
5115
5116 if (pa->pa_type == MB_INODE_PA) {
5117 write_lock(pa->pa_node_lock.inode_lock);
5118 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5119 write_unlock(pa->pa_node_lock.inode_lock);
5120 ext4_mb_pa_free(pa);
5121 } else {
5122 spin_lock(pa->pa_node_lock.lg_lock);
5123 list_del_rcu(&pa->pa_node.lg_list);
5124 spin_unlock(pa->pa_node_lock.lg_lock);
5125 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5126 }
5127}
5128
5129static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new)
5130{
5131 struct rb_node **iter = &root->rb_node, *parent = NULL;
5132 struct ext4_prealloc_space *iter_pa, *new_pa;
5133 ext4_lblk_t iter_start, new_start;
5134
5135 while (*iter) {
5136 iter_pa = rb_entry(*iter, struct ext4_prealloc_space,
5137 pa_node.inode_node);
5138 new_pa = rb_entry(new, struct ext4_prealloc_space,
5139 pa_node.inode_node);
5140 iter_start = iter_pa->pa_lstart;
5141 new_start = new_pa->pa_lstart;
5142
5143 parent = *iter;
5144 if (new_start < iter_start)
5145 iter = &((*iter)->rb_left);
5146 else
5147 iter = &((*iter)->rb_right);
5148 }
5149
5150 rb_link_node(new, parent, iter);
5151 rb_insert_color(new, root);
5152}
5153
5154/*
5155 * creates new preallocated space for given inode
5156 */
5157static noinline_for_stack void
5158ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
5159{
5160 struct super_block *sb = ac->ac_sb;
5161 struct ext4_sb_info *sbi = EXT4_SB(sb);
5162 struct ext4_prealloc_space *pa;
5163 struct ext4_group_info *grp;
5164 struct ext4_inode_info *ei;
5165
5166 /* preallocate only when found space is larger then requested */
5167 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5168 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5169 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5170 BUG_ON(ac->ac_pa == NULL);
5171
5172 pa = ac->ac_pa;
5173
5174 if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) {
5175 struct ext4_free_extent ex = {
5176 .fe_logical = ac->ac_g_ex.fe_logical,
5177 .fe_len = ac->ac_orig_goal_len,
5178 };
5179 loff_t orig_goal_end = extent_logical_end(sbi, &ex);
5180 loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex);
5181
5182 /*
5183 * We can't allocate as much as normalizer wants, so we try
5184 * to get proper lstart to cover the original request, except
5185 * when the goal doesn't cover the original request as below:
5186 *
5187 * orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048
5188 * best_ex:0/200(200) -> adjusted: 1848/2048(200)
5189 */
5190 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
5191 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
5192
5193 /*
5194 * Use the below logic for adjusting best extent as it keeps
5195 * fragmentation in check while ensuring logical range of best
5196 * extent doesn't overflow out of goal extent:
5197 *
5198 * 1. Check if best ex can be kept at end of goal (before
5199 * cr_best_avail trimmed it) and still cover original start
5200 * 2. Else, check if best ex can be kept at start of goal and
5201 * still cover original end
5202 * 3. Else, keep the best ex at start of original request.
5203 */
5204 ex.fe_len = ac->ac_b_ex.fe_len;
5205
5206 ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len);
5207 if (ac->ac_o_ex.fe_logical >= ex.fe_logical)
5208 goto adjust_bex;
5209
5210 ex.fe_logical = ac->ac_g_ex.fe_logical;
5211 if (o_ex_end <= extent_logical_end(sbi, &ex))
5212 goto adjust_bex;
5213
5214 ex.fe_logical = ac->ac_o_ex.fe_logical;
5215adjust_bex:
5216 ac->ac_b_ex.fe_logical = ex.fe_logical;
5217
5218 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
5219 BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end);
5220 }
5221
5222 pa->pa_lstart = ac->ac_b_ex.fe_logical;
5223 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5224 pa->pa_len = ac->ac_b_ex.fe_len;
5225 pa->pa_free = pa->pa_len;
5226 spin_lock_init(&pa->pa_lock);
5227 INIT_LIST_HEAD(&pa->pa_group_list);
5228 pa->pa_deleted = 0;
5229 pa->pa_type = MB_INODE_PA;
5230
5231 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5232 pa->pa_len, pa->pa_lstart);
5233 trace_ext4_mb_new_inode_pa(ac, pa);
5234
5235 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
5236 ext4_mb_use_inode_pa(ac, pa);
5237
5238 ei = EXT4_I(ac->ac_inode);
5239 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5240 if (!grp)
5241 return;
5242
5243 pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
5244 pa->pa_inode = ac->ac_inode;
5245
5246 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5247
5248 write_lock(pa->pa_node_lock.inode_lock);
5249 ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node);
5250 write_unlock(pa->pa_node_lock.inode_lock);
5251 atomic_inc(&ei->i_prealloc_active);
5252}
5253
5254/*
5255 * creates new preallocated space for locality group inodes belongs to
5256 */
5257static noinline_for_stack void
5258ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
5259{
5260 struct super_block *sb = ac->ac_sb;
5261 struct ext4_locality_group *lg;
5262 struct ext4_prealloc_space *pa;
5263 struct ext4_group_info *grp;
5264
5265 /* preallocate only when found space is larger then requested */
5266 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5267 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5268 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5269 BUG_ON(ac->ac_pa == NULL);
5270
5271 pa = ac->ac_pa;
5272
5273 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5274 pa->pa_lstart = pa->pa_pstart;
5275 pa->pa_len = ac->ac_b_ex.fe_len;
5276 pa->pa_free = pa->pa_len;
5277 spin_lock_init(&pa->pa_lock);
5278 INIT_LIST_HEAD(&pa->pa_node.lg_list);
5279 INIT_LIST_HEAD(&pa->pa_group_list);
5280 pa->pa_deleted = 0;
5281 pa->pa_type = MB_GROUP_PA;
5282
5283 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5284 pa->pa_len, pa->pa_lstart);
5285 trace_ext4_mb_new_group_pa(ac, pa);
5286
5287 ext4_mb_use_group_pa(ac, pa);
5288 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
5289
5290 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5291 if (!grp)
5292 return;
5293 lg = ac->ac_lg;
5294 BUG_ON(lg == NULL);
5295
5296 pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
5297 pa->pa_inode = NULL;
5298
5299 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5300
5301 /*
5302 * We will later add the new pa to the right bucket
5303 * after updating the pa_free in ext4_mb_release_context
5304 */
5305}
5306
5307static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
5308{
5309 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5310 ext4_mb_new_group_pa(ac);
5311 else
5312 ext4_mb_new_inode_pa(ac);
5313}
5314
5315/*
5316 * finds all unused blocks in on-disk bitmap, frees them in
5317 * in-core bitmap and buddy.
5318 * @pa must be unlinked from inode and group lists, so that
5319 * nobody else can find/use it.
5320 * the caller MUST hold group/inode locks.
5321 * TODO: optimize the case when there are no in-core structures yet
5322 */
5323static noinline_for_stack void
5324ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
5325 struct ext4_prealloc_space *pa)
5326{
5327 struct super_block *sb = e4b->bd_sb;
5328 struct ext4_sb_info *sbi = EXT4_SB(sb);
5329 unsigned int end;
5330 unsigned int next;
5331 ext4_group_t group;
5332 ext4_grpblk_t bit;
5333 unsigned long long grp_blk_start;
5334 int free = 0;
5335
5336 BUG_ON(pa->pa_deleted == 0);
5337 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5338 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
5339 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
5340 end = bit + pa->pa_len;
5341
5342 while (bit < end) {
5343 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
5344 if (bit >= end)
5345 break;
5346 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
5347 mb_debug(sb, "free preallocated %u/%u in group %u\n",
5348 (unsigned) ext4_group_first_block_no(sb, group) + bit,
5349 (unsigned) next - bit, (unsigned) group);
5350 free += next - bit;
5351
5352 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
5353 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
5354 EXT4_C2B(sbi, bit)),
5355 next - bit);
5356 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
5357 bit = next + 1;
5358 }
5359 if (free != pa->pa_free) {
5360 ext4_msg(e4b->bd_sb, KERN_CRIT,
5361 "pa %p: logic %lu, phys. %lu, len %d",
5362 pa, (unsigned long) pa->pa_lstart,
5363 (unsigned long) pa->pa_pstart,
5364 pa->pa_len);
5365 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
5366 free, pa->pa_free);
5367 /*
5368 * pa is already deleted so we use the value obtained
5369 * from the bitmap and continue.
5370 */
5371 }
5372 atomic_add(free, &sbi->s_mb_discarded);
5373}
5374
5375static noinline_for_stack void
5376ext4_mb_release_group_pa(struct ext4_buddy *e4b,
5377 struct ext4_prealloc_space *pa)
5378{
5379 struct super_block *sb = e4b->bd_sb;
5380 ext4_group_t group;
5381 ext4_grpblk_t bit;
5382
5383 trace_ext4_mb_release_group_pa(sb, pa);
5384 BUG_ON(pa->pa_deleted == 0);
5385 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5386 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
5387 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
5388 e4b->bd_group, group, pa->pa_pstart);
5389 return;
5390 }
5391 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
5392 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
5393 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
5394}
5395
5396/*
5397 * releases all preallocations in given group
5398 *
5399 * first, we need to decide discard policy:
5400 * - when do we discard
5401 * 1) ENOSPC
5402 * - how many do we discard
5403 * 1) how many requested
5404 */
5405static noinline_for_stack int
5406ext4_mb_discard_group_preallocations(struct super_block *sb,
5407 ext4_group_t group, int *busy)
5408{
5409 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
5410 struct buffer_head *bitmap_bh = NULL;
5411 struct ext4_prealloc_space *pa, *tmp;
5412 LIST_HEAD(list);
5413 struct ext4_buddy e4b;
5414 struct ext4_inode_info *ei;
5415 int err;
5416 int free = 0;
5417
5418 if (!grp)
5419 return 0;
5420 mb_debug(sb, "discard preallocation for group %u\n", group);
5421 if (list_empty(&grp->bb_prealloc_list))
5422 goto out_dbg;
5423
5424 bitmap_bh = ext4_read_block_bitmap(sb, group);
5425 if (IS_ERR(bitmap_bh)) {
5426 err = PTR_ERR(bitmap_bh);
5427 ext4_error_err(sb, -err,
5428 "Error %d reading block bitmap for %u",
5429 err, group);
5430 goto out_dbg;
5431 }
5432
5433 err = ext4_mb_load_buddy(sb, group, &e4b);
5434 if (err) {
5435 ext4_warning(sb, "Error %d loading buddy information for %u",
5436 err, group);
5437 put_bh(bitmap_bh);
5438 goto out_dbg;
5439 }
5440
5441 ext4_lock_group(sb, group);
5442 list_for_each_entry_safe(pa, tmp,
5443 &grp->bb_prealloc_list, pa_group_list) {
5444 spin_lock(&pa->pa_lock);
5445 if (atomic_read(&pa->pa_count)) {
5446 spin_unlock(&pa->pa_lock);
5447 *busy = 1;
5448 continue;
5449 }
5450 if (pa->pa_deleted) {
5451 spin_unlock(&pa->pa_lock);
5452 continue;
5453 }
5454
5455 /* seems this one can be freed ... */
5456 ext4_mb_mark_pa_deleted(sb, pa);
5457
5458 if (!free)
5459 this_cpu_inc(discard_pa_seq);
5460
5461 /* we can trust pa_free ... */
5462 free += pa->pa_free;
5463
5464 spin_unlock(&pa->pa_lock);
5465
5466 list_del(&pa->pa_group_list);
5467 list_add(&pa->u.pa_tmp_list, &list);
5468 }
5469
5470 /* now free all selected PAs */
5471 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5472
5473 /* remove from object (inode or locality group) */
5474 if (pa->pa_type == MB_GROUP_PA) {
5475 spin_lock(pa->pa_node_lock.lg_lock);
5476 list_del_rcu(&pa->pa_node.lg_list);
5477 spin_unlock(pa->pa_node_lock.lg_lock);
5478 } else {
5479 write_lock(pa->pa_node_lock.inode_lock);
5480 ei = EXT4_I(pa->pa_inode);
5481 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5482 write_unlock(pa->pa_node_lock.inode_lock);
5483 }
5484
5485 list_del(&pa->u.pa_tmp_list);
5486
5487 if (pa->pa_type == MB_GROUP_PA) {
5488 ext4_mb_release_group_pa(&e4b, pa);
5489 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5490 } else {
5491 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5492 ext4_mb_pa_free(pa);
5493 }
5494 }
5495
5496 ext4_unlock_group(sb, group);
5497 ext4_mb_unload_buddy(&e4b);
5498 put_bh(bitmap_bh);
5499out_dbg:
5500 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
5501 free, group, grp->bb_free);
5502 return free;
5503}
5504
5505/*
5506 * releases all non-used preallocated blocks for given inode
5507 *
5508 * It's important to discard preallocations under i_data_sem
5509 * We don't want another block to be served from the prealloc
5510 * space when we are discarding the inode prealloc space.
5511 *
5512 * FIXME!! Make sure it is valid at all the call sites
5513 */
5514void ext4_discard_preallocations(struct inode *inode)
5515{
5516 struct ext4_inode_info *ei = EXT4_I(inode);
5517 struct super_block *sb = inode->i_sb;
5518 struct buffer_head *bitmap_bh = NULL;
5519 struct ext4_prealloc_space *pa, *tmp;
5520 ext4_group_t group = 0;
5521 LIST_HEAD(list);
5522 struct ext4_buddy e4b;
5523 struct rb_node *iter;
5524 int err;
5525
5526 if (!S_ISREG(inode->i_mode))
5527 return;
5528
5529 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
5530 return;
5531
5532 mb_debug(sb, "discard preallocation for inode %lu\n",
5533 inode->i_ino);
5534 trace_ext4_discard_preallocations(inode,
5535 atomic_read(&ei->i_prealloc_active));
5536
5537repeat:
5538 /* first, collect all pa's in the inode */
5539 write_lock(&ei->i_prealloc_lock);
5540 for (iter = rb_first(&ei->i_prealloc_node); iter;
5541 iter = rb_next(iter)) {
5542 pa = rb_entry(iter, struct ext4_prealloc_space,
5543 pa_node.inode_node);
5544 BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
5545
5546 spin_lock(&pa->pa_lock);
5547 if (atomic_read(&pa->pa_count)) {
5548 /* this shouldn't happen often - nobody should
5549 * use preallocation while we're discarding it */
5550 spin_unlock(&pa->pa_lock);
5551 write_unlock(&ei->i_prealloc_lock);
5552 ext4_msg(sb, KERN_ERR,
5553 "uh-oh! used pa while discarding");
5554 WARN_ON(1);
5555 schedule_timeout_uninterruptible(HZ);
5556 goto repeat;
5557
5558 }
5559 if (pa->pa_deleted == 0) {
5560 ext4_mb_mark_pa_deleted(sb, pa);
5561 spin_unlock(&pa->pa_lock);
5562 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5563 list_add(&pa->u.pa_tmp_list, &list);
5564 continue;
5565 }
5566
5567 /* someone is deleting pa right now */
5568 spin_unlock(&pa->pa_lock);
5569 write_unlock(&ei->i_prealloc_lock);
5570
5571 /* we have to wait here because pa_deleted
5572 * doesn't mean pa is already unlinked from
5573 * the list. as we might be called from
5574 * ->clear_inode() the inode will get freed
5575 * and concurrent thread which is unlinking
5576 * pa from inode's list may access already
5577 * freed memory, bad-bad-bad */
5578
5579 /* XXX: if this happens too often, we can
5580 * add a flag to force wait only in case
5581 * of ->clear_inode(), but not in case of
5582 * regular truncate */
5583 schedule_timeout_uninterruptible(HZ);
5584 goto repeat;
5585 }
5586 write_unlock(&ei->i_prealloc_lock);
5587
5588 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5589 BUG_ON(pa->pa_type != MB_INODE_PA);
5590 group = ext4_get_group_number(sb, pa->pa_pstart);
5591
5592 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5593 GFP_NOFS|__GFP_NOFAIL);
5594 if (err) {
5595 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5596 err, group);
5597 continue;
5598 }
5599
5600 bitmap_bh = ext4_read_block_bitmap(sb, group);
5601 if (IS_ERR(bitmap_bh)) {
5602 err = PTR_ERR(bitmap_bh);
5603 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5604 err, group);
5605 ext4_mb_unload_buddy(&e4b);
5606 continue;
5607 }
5608
5609 ext4_lock_group(sb, group);
5610 list_del(&pa->pa_group_list);
5611 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5612 ext4_unlock_group(sb, group);
5613
5614 ext4_mb_unload_buddy(&e4b);
5615 put_bh(bitmap_bh);
5616
5617 list_del(&pa->u.pa_tmp_list);
5618 ext4_mb_pa_free(pa);
5619 }
5620}
5621
5622static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5623{
5624 struct ext4_prealloc_space *pa;
5625
5626 BUG_ON(ext4_pspace_cachep == NULL);
5627 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5628 if (!pa)
5629 return -ENOMEM;
5630 atomic_set(&pa->pa_count, 1);
5631 ac->ac_pa = pa;
5632 return 0;
5633}
5634
5635static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
5636{
5637 struct ext4_prealloc_space *pa = ac->ac_pa;
5638
5639 BUG_ON(!pa);
5640 ac->ac_pa = NULL;
5641 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5642 /*
5643 * current function is only called due to an error or due to
5644 * len of found blocks < len of requested blocks hence the PA has not
5645 * been added to grp->bb_prealloc_list. So we don't need to lock it
5646 */
5647 pa->pa_deleted = 1;
5648 ext4_mb_pa_free(pa);
5649}
5650
5651#ifdef CONFIG_EXT4_DEBUG
5652static inline void ext4_mb_show_pa(struct super_block *sb)
5653{
5654 ext4_group_t i, ngroups;
5655
5656 if (ext4_forced_shutdown(sb))
5657 return;
5658
5659 ngroups = ext4_get_groups_count(sb);
5660 mb_debug(sb, "groups: ");
5661 for (i = 0; i < ngroups; i++) {
5662 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5663 struct ext4_prealloc_space *pa;
5664 ext4_grpblk_t start;
5665 struct list_head *cur;
5666
5667 if (!grp)
5668 continue;
5669 ext4_lock_group(sb, i);
5670 list_for_each(cur, &grp->bb_prealloc_list) {
5671 pa = list_entry(cur, struct ext4_prealloc_space,
5672 pa_group_list);
5673 spin_lock(&pa->pa_lock);
5674 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5675 NULL, &start);
5676 spin_unlock(&pa->pa_lock);
5677 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5678 pa->pa_len);
5679 }
5680 ext4_unlock_group(sb, i);
5681 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5682 grp->bb_fragments);
5683 }
5684}
5685
5686static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5687{
5688 struct super_block *sb = ac->ac_sb;
5689
5690 if (ext4_forced_shutdown(sb))
5691 return;
5692
5693 mb_debug(sb, "Can't allocate:"
5694 " Allocation context details:");
5695 mb_debug(sb, "status %u flags 0x%x",
5696 ac->ac_status, ac->ac_flags);
5697 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5698 "goal %lu/%lu/%lu@%lu, "
5699 "best %lu/%lu/%lu@%lu cr %d",
5700 (unsigned long)ac->ac_o_ex.fe_group,
5701 (unsigned long)ac->ac_o_ex.fe_start,
5702 (unsigned long)ac->ac_o_ex.fe_len,
5703 (unsigned long)ac->ac_o_ex.fe_logical,
5704 (unsigned long)ac->ac_g_ex.fe_group,
5705 (unsigned long)ac->ac_g_ex.fe_start,
5706 (unsigned long)ac->ac_g_ex.fe_len,
5707 (unsigned long)ac->ac_g_ex.fe_logical,
5708 (unsigned long)ac->ac_b_ex.fe_group,
5709 (unsigned long)ac->ac_b_ex.fe_start,
5710 (unsigned long)ac->ac_b_ex.fe_len,
5711 (unsigned long)ac->ac_b_ex.fe_logical,
5712 (int)ac->ac_criteria);
5713 mb_debug(sb, "%u found", ac->ac_found);
5714 mb_debug(sb, "used pa: %s, ", str_yes_no(ac->ac_pa));
5715 if (ac->ac_pa)
5716 mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ?
5717 "group pa" : "inode pa");
5718 ext4_mb_show_pa(sb);
5719}
5720#else
5721static inline void ext4_mb_show_pa(struct super_block *sb)
5722{
5723}
5724static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5725{
5726 ext4_mb_show_pa(ac->ac_sb);
5727}
5728#endif
5729
5730/*
5731 * We use locality group preallocation for small size file. The size of the
5732 * file is determined by the current size or the resulting size after
5733 * allocation which ever is larger
5734 *
5735 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5736 */
5737static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5738{
5739 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5740 int bsbits = ac->ac_sb->s_blocksize_bits;
5741 loff_t size, isize;
5742 bool inode_pa_eligible, group_pa_eligible;
5743
5744 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5745 return;
5746
5747 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5748 return;
5749
5750 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5751 inode_pa_eligible = true;
5752 size = extent_logical_end(sbi, &ac->ac_o_ex);
5753 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5754 >> bsbits;
5755
5756 /* No point in using inode preallocation for closed files */
5757 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5758 !inode_is_open_for_write(ac->ac_inode))
5759 inode_pa_eligible = false;
5760
5761 size = max(size, isize);
5762 /* Don't use group allocation for large files */
5763 if (size > sbi->s_mb_stream_request)
5764 group_pa_eligible = false;
5765
5766 if (!group_pa_eligible) {
5767 if (inode_pa_eligible)
5768 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5769 else
5770 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5771 return;
5772 }
5773
5774 BUG_ON(ac->ac_lg != NULL);
5775 /*
5776 * locality group prealloc space are per cpu. The reason for having
5777 * per cpu locality group is to reduce the contention between block
5778 * request from multiple CPUs.
5779 */
5780 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5781
5782 /* we're going to use group allocation */
5783 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5784
5785 /* serialize all allocations in the group */
5786 mutex_lock(&ac->ac_lg->lg_mutex);
5787}
5788
5789static noinline_for_stack void
5790ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5791 struct ext4_allocation_request *ar)
5792{
5793 struct super_block *sb = ar->inode->i_sb;
5794 struct ext4_sb_info *sbi = EXT4_SB(sb);
5795 struct ext4_super_block *es = sbi->s_es;
5796 ext4_group_t group;
5797 unsigned int len;
5798 ext4_fsblk_t goal;
5799 ext4_grpblk_t block;
5800
5801 /* we can't allocate > group size */
5802 len = ar->len;
5803
5804 /* just a dirty hack to filter too big requests */
5805 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5806 len = EXT4_CLUSTERS_PER_GROUP(sb);
5807
5808 /* start searching from the goal */
5809 goal = ar->goal;
5810 if (goal < le32_to_cpu(es->s_first_data_block) ||
5811 goal >= ext4_blocks_count(es))
5812 goal = le32_to_cpu(es->s_first_data_block);
5813 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5814
5815 /* set up allocation goals */
5816 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5817 ac->ac_status = AC_STATUS_CONTINUE;
5818 ac->ac_sb = sb;
5819 ac->ac_inode = ar->inode;
5820 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5821 ac->ac_o_ex.fe_group = group;
5822 ac->ac_o_ex.fe_start = block;
5823 ac->ac_o_ex.fe_len = len;
5824 ac->ac_g_ex = ac->ac_o_ex;
5825 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
5826 ac->ac_flags = ar->flags;
5827
5828 /* we have to define context: we'll work with a file or
5829 * locality group. this is a policy, actually */
5830 ext4_mb_group_or_file(ac);
5831
5832 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5833 "left: %u/%u, right %u/%u to %swritable\n",
5834 (unsigned) ar->len, (unsigned) ar->logical,
5835 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5836 (unsigned) ar->lleft, (unsigned) ar->pleft,
5837 (unsigned) ar->lright, (unsigned) ar->pright,
5838 inode_is_open_for_write(ar->inode) ? "" : "non-");
5839}
5840
5841static noinline_for_stack void
5842ext4_mb_discard_lg_preallocations(struct super_block *sb,
5843 struct ext4_locality_group *lg,
5844 int order, int total_entries)
5845{
5846 ext4_group_t group = 0;
5847 struct ext4_buddy e4b;
5848 LIST_HEAD(discard_list);
5849 struct ext4_prealloc_space *pa, *tmp;
5850
5851 mb_debug(sb, "discard locality group preallocation\n");
5852
5853 spin_lock(&lg->lg_prealloc_lock);
5854 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5855 pa_node.lg_list,
5856 lockdep_is_held(&lg->lg_prealloc_lock)) {
5857 spin_lock(&pa->pa_lock);
5858 if (atomic_read(&pa->pa_count)) {
5859 /*
5860 * This is the pa that we just used
5861 * for block allocation. So don't
5862 * free that
5863 */
5864 spin_unlock(&pa->pa_lock);
5865 continue;
5866 }
5867 if (pa->pa_deleted) {
5868 spin_unlock(&pa->pa_lock);
5869 continue;
5870 }
5871 /* only lg prealloc space */
5872 BUG_ON(pa->pa_type != MB_GROUP_PA);
5873
5874 /* seems this one can be freed ... */
5875 ext4_mb_mark_pa_deleted(sb, pa);
5876 spin_unlock(&pa->pa_lock);
5877
5878 list_del_rcu(&pa->pa_node.lg_list);
5879 list_add(&pa->u.pa_tmp_list, &discard_list);
5880
5881 total_entries--;
5882 if (total_entries <= 5) {
5883 /*
5884 * we want to keep only 5 entries
5885 * allowing it to grow to 8. This
5886 * mak sure we don't call discard
5887 * soon for this list.
5888 */
5889 break;
5890 }
5891 }
5892 spin_unlock(&lg->lg_prealloc_lock);
5893
5894 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5895 int err;
5896
5897 group = ext4_get_group_number(sb, pa->pa_pstart);
5898 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5899 GFP_NOFS|__GFP_NOFAIL);
5900 if (err) {
5901 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5902 err, group);
5903 continue;
5904 }
5905 ext4_lock_group(sb, group);
5906 list_del(&pa->pa_group_list);
5907 ext4_mb_release_group_pa(&e4b, pa);
5908 ext4_unlock_group(sb, group);
5909
5910 ext4_mb_unload_buddy(&e4b);
5911 list_del(&pa->u.pa_tmp_list);
5912 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5913 }
5914}
5915
5916/*
5917 * We have incremented pa_count. So it cannot be freed at this
5918 * point. Also we hold lg_mutex. So no parallel allocation is
5919 * possible from this lg. That means pa_free cannot be updated.
5920 *
5921 * A parallel ext4_mb_discard_group_preallocations is possible.
5922 * which can cause the lg_prealloc_list to be updated.
5923 */
5924
5925static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5926{
5927 int order, added = 0, lg_prealloc_count = 1;
5928 struct super_block *sb = ac->ac_sb;
5929 struct ext4_locality_group *lg = ac->ac_lg;
5930 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5931
5932 order = fls(pa->pa_free) - 1;
5933 if (order > PREALLOC_TB_SIZE - 1)
5934 /* The max size of hash table is PREALLOC_TB_SIZE */
5935 order = PREALLOC_TB_SIZE - 1;
5936 /* Add the prealloc space to lg */
5937 spin_lock(&lg->lg_prealloc_lock);
5938 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5939 pa_node.lg_list,
5940 lockdep_is_held(&lg->lg_prealloc_lock)) {
5941 spin_lock(&tmp_pa->pa_lock);
5942 if (tmp_pa->pa_deleted) {
5943 spin_unlock(&tmp_pa->pa_lock);
5944 continue;
5945 }
5946 if (!added && pa->pa_free < tmp_pa->pa_free) {
5947 /* Add to the tail of the previous entry */
5948 list_add_tail_rcu(&pa->pa_node.lg_list,
5949 &tmp_pa->pa_node.lg_list);
5950 added = 1;
5951 /*
5952 * we want to count the total
5953 * number of entries in the list
5954 */
5955 }
5956 spin_unlock(&tmp_pa->pa_lock);
5957 lg_prealloc_count++;
5958 }
5959 if (!added)
5960 list_add_tail_rcu(&pa->pa_node.lg_list,
5961 &lg->lg_prealloc_list[order]);
5962 spin_unlock(&lg->lg_prealloc_lock);
5963
5964 /* Now trim the list to be not more than 8 elements */
5965 if (lg_prealloc_count > 8)
5966 ext4_mb_discard_lg_preallocations(sb, lg,
5967 order, lg_prealloc_count);
5968}
5969
5970/*
5971 * release all resource we used in allocation
5972 */
5973static void ext4_mb_release_context(struct ext4_allocation_context *ac)
5974{
5975 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5976 struct ext4_prealloc_space *pa = ac->ac_pa;
5977 if (pa) {
5978 if (pa->pa_type == MB_GROUP_PA) {
5979 /* see comment in ext4_mb_use_group_pa() */
5980 spin_lock(&pa->pa_lock);
5981 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5982 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5983 pa->pa_free -= ac->ac_b_ex.fe_len;
5984 pa->pa_len -= ac->ac_b_ex.fe_len;
5985 spin_unlock(&pa->pa_lock);
5986
5987 /*
5988 * We want to add the pa to the right bucket.
5989 * Remove it from the list and while adding
5990 * make sure the list to which we are adding
5991 * doesn't grow big.
5992 */
5993 if (likely(pa->pa_free)) {
5994 spin_lock(pa->pa_node_lock.lg_lock);
5995 list_del_rcu(&pa->pa_node.lg_list);
5996 spin_unlock(pa->pa_node_lock.lg_lock);
5997 ext4_mb_add_n_trim(ac);
5998 }
5999 }
6000
6001 ext4_mb_put_pa(ac, ac->ac_sb, pa);
6002 }
6003 if (ac->ac_bitmap_folio)
6004 folio_put(ac->ac_bitmap_folio);
6005 if (ac->ac_buddy_folio)
6006 folio_put(ac->ac_buddy_folio);
6007 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
6008 mutex_unlock(&ac->ac_lg->lg_mutex);
6009 ext4_mb_collect_stats(ac);
6010}
6011
6012static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
6013{
6014 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
6015 int ret;
6016 int freed = 0, busy = 0;
6017 int retry = 0;
6018
6019 trace_ext4_mb_discard_preallocations(sb, needed);
6020
6021 if (needed == 0)
6022 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
6023 repeat:
6024 for (i = 0; i < ngroups && needed > 0; i++) {
6025 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
6026 freed += ret;
6027 needed -= ret;
6028 cond_resched();
6029 }
6030
6031 if (needed > 0 && busy && ++retry < 3) {
6032 busy = 0;
6033 goto repeat;
6034 }
6035
6036 return freed;
6037}
6038
6039static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
6040 struct ext4_allocation_context *ac, u64 *seq)
6041{
6042 int freed;
6043 u64 seq_retry = 0;
6044 bool ret = false;
6045
6046 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
6047 if (freed) {
6048 ret = true;
6049 goto out_dbg;
6050 }
6051 seq_retry = ext4_get_discard_pa_seq_sum();
6052 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
6053 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
6054 *seq = seq_retry;
6055 ret = true;
6056 }
6057
6058out_dbg:
6059 mb_debug(sb, "freed %d, retry ? %s\n", freed, str_yes_no(ret));
6060 return ret;
6061}
6062
6063/*
6064 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
6065 * linearly starting at the goal block and also excludes the blocks which
6066 * are going to be in use after fast commit replay.
6067 */
6068static ext4_fsblk_t
6069ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp)
6070{
6071 struct buffer_head *bitmap_bh;
6072 struct super_block *sb = ar->inode->i_sb;
6073 struct ext4_sb_info *sbi = EXT4_SB(sb);
6074 ext4_group_t group, nr;
6075 ext4_grpblk_t blkoff;
6076 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
6077 ext4_grpblk_t i = 0;
6078 ext4_fsblk_t goal, block;
6079 struct ext4_super_block *es = sbi->s_es;
6080
6081 goal = ar->goal;
6082 if (goal < le32_to_cpu(es->s_first_data_block) ||
6083 goal >= ext4_blocks_count(es))
6084 goal = le32_to_cpu(es->s_first_data_block);
6085
6086 ar->len = 0;
6087 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
6088 for (nr = ext4_get_groups_count(sb); nr > 0; nr--) {
6089 bitmap_bh = ext4_read_block_bitmap(sb, group);
6090 if (IS_ERR(bitmap_bh)) {
6091 *errp = PTR_ERR(bitmap_bh);
6092 pr_warn("Failed to read block bitmap\n");
6093 return 0;
6094 }
6095
6096 while (1) {
6097 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
6098 blkoff);
6099 if (i >= max)
6100 break;
6101 if (ext4_fc_replay_check_excluded(sb,
6102 ext4_group_first_block_no(sb, group) +
6103 EXT4_C2B(sbi, i))) {
6104 blkoff = i + 1;
6105 } else
6106 break;
6107 }
6108 brelse(bitmap_bh);
6109 if (i < max)
6110 break;
6111
6112 if (++group >= ext4_get_groups_count(sb))
6113 group = 0;
6114
6115 blkoff = 0;
6116 }
6117
6118 if (i >= max) {
6119 *errp = -ENOSPC;
6120 return 0;
6121 }
6122
6123 block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i);
6124 ext4_mb_mark_bb(sb, block, 1, true);
6125 ar->len = 1;
6126
6127 *errp = 0;
6128 return block;
6129}
6130
6131/*
6132 * Main entry point into mballoc to allocate blocks
6133 * it tries to use preallocation first, then falls back
6134 * to usual allocation
6135 */
6136ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
6137 struct ext4_allocation_request *ar, int *errp)
6138{
6139 struct ext4_allocation_context *ac = NULL;
6140 struct ext4_sb_info *sbi;
6141 struct super_block *sb;
6142 ext4_fsblk_t block = 0;
6143 unsigned int inquota = 0;
6144 unsigned int reserv_clstrs = 0;
6145 int retries = 0;
6146 u64 seq;
6147
6148 might_sleep();
6149 sb = ar->inode->i_sb;
6150 sbi = EXT4_SB(sb);
6151
6152 trace_ext4_request_blocks(ar);
6153 if (sbi->s_mount_state & EXT4_FC_REPLAY)
6154 return ext4_mb_new_blocks_simple(ar, errp);
6155
6156 /* Allow to use superuser reservation for quota file */
6157 if (ext4_is_quota_file(ar->inode))
6158 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
6159
6160 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
6161 /* Without delayed allocation we need to verify
6162 * there is enough free blocks to do block allocation
6163 * and verify allocation doesn't exceed the quota limits.
6164 */
6165 while (ar->len &&
6166 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
6167
6168 /* let others to free the space */
6169 cond_resched();
6170 ar->len = ar->len >> 1;
6171 }
6172 if (!ar->len) {
6173 ext4_mb_show_pa(sb);
6174 *errp = -ENOSPC;
6175 return 0;
6176 }
6177 reserv_clstrs = ar->len;
6178 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
6179 dquot_alloc_block_nofail(ar->inode,
6180 EXT4_C2B(sbi, ar->len));
6181 } else {
6182 while (ar->len &&
6183 dquot_alloc_block(ar->inode,
6184 EXT4_C2B(sbi, ar->len))) {
6185
6186 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
6187 ar->len--;
6188 }
6189 }
6190 inquota = ar->len;
6191 if (ar->len == 0) {
6192 *errp = -EDQUOT;
6193 goto out;
6194 }
6195 }
6196
6197 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
6198 if (!ac) {
6199 ar->len = 0;
6200 *errp = -ENOMEM;
6201 goto out;
6202 }
6203
6204 ext4_mb_initialize_context(ac, ar);
6205
6206 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
6207 seq = this_cpu_read(discard_pa_seq);
6208 if (!ext4_mb_use_preallocated(ac)) {
6209 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
6210 ext4_mb_normalize_request(ac, ar);
6211
6212 *errp = ext4_mb_pa_alloc(ac);
6213 if (*errp)
6214 goto errout;
6215repeat:
6216 /* allocate space in core */
6217 *errp = ext4_mb_regular_allocator(ac);
6218 /*
6219 * pa allocated above is added to grp->bb_prealloc_list only
6220 * when we were able to allocate some block i.e. when
6221 * ac->ac_status == AC_STATUS_FOUND.
6222 * And error from above mean ac->ac_status != AC_STATUS_FOUND
6223 * So we have to free this pa here itself.
6224 */
6225 if (*errp) {
6226 ext4_mb_pa_put_free(ac);
6227 ext4_discard_allocated_blocks(ac);
6228 goto errout;
6229 }
6230 if (ac->ac_status == AC_STATUS_FOUND &&
6231 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
6232 ext4_mb_pa_put_free(ac);
6233 }
6234 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
6235 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
6236 if (*errp) {
6237 ext4_discard_allocated_blocks(ac);
6238 goto errout;
6239 } else {
6240 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
6241 ar->len = ac->ac_b_ex.fe_len;
6242 }
6243 } else {
6244 if (++retries < 3 &&
6245 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
6246 goto repeat;
6247 /*
6248 * If block allocation fails then the pa allocated above
6249 * needs to be freed here itself.
6250 */
6251 ext4_mb_pa_put_free(ac);
6252 *errp = -ENOSPC;
6253 }
6254
6255 if (*errp) {
6256errout:
6257 ac->ac_b_ex.fe_len = 0;
6258 ar->len = 0;
6259 ext4_mb_show_ac(ac);
6260 }
6261 ext4_mb_release_context(ac);
6262 kmem_cache_free(ext4_ac_cachep, ac);
6263out:
6264 if (inquota && ar->len < inquota)
6265 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
6266 if (!ar->len) {
6267 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
6268 /* release all the reserved blocks if non delalloc */
6269 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
6270 reserv_clstrs);
6271 }
6272
6273 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
6274
6275 return block;
6276}
6277
6278/*
6279 * We can merge two free data extents only if the physical blocks
6280 * are contiguous, AND the extents were freed by the same transaction,
6281 * AND the blocks are associated with the same group.
6282 */
6283static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
6284 struct ext4_free_data *entry,
6285 struct ext4_free_data *new_entry,
6286 struct rb_root *entry_rb_root)
6287{
6288 if ((entry->efd_tid != new_entry->efd_tid) ||
6289 (entry->efd_group != new_entry->efd_group))
6290 return;
6291 if (entry->efd_start_cluster + entry->efd_count ==
6292 new_entry->efd_start_cluster) {
6293 new_entry->efd_start_cluster = entry->efd_start_cluster;
6294 new_entry->efd_count += entry->efd_count;
6295 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
6296 entry->efd_start_cluster) {
6297 new_entry->efd_count += entry->efd_count;
6298 } else
6299 return;
6300 spin_lock(&sbi->s_md_lock);
6301 list_del(&entry->efd_list);
6302 spin_unlock(&sbi->s_md_lock);
6303 rb_erase(&entry->efd_node, entry_rb_root);
6304 kmem_cache_free(ext4_free_data_cachep, entry);
6305}
6306
6307static noinline_for_stack void
6308ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
6309 struct ext4_free_data *new_entry)
6310{
6311 ext4_group_t group = e4b->bd_group;
6312 ext4_grpblk_t cluster;
6313 ext4_grpblk_t clusters = new_entry->efd_count;
6314 struct ext4_free_data *entry;
6315 struct ext4_group_info *db = e4b->bd_info;
6316 struct super_block *sb = e4b->bd_sb;
6317 struct ext4_sb_info *sbi = EXT4_SB(sb);
6318 struct rb_node **n = &db->bb_free_root.rb_node, *node;
6319 struct rb_node *parent = NULL, *new_node;
6320
6321 BUG_ON(!ext4_handle_valid(handle));
6322 BUG_ON(e4b->bd_bitmap_folio == NULL);
6323 BUG_ON(e4b->bd_buddy_folio == NULL);
6324
6325 new_node = &new_entry->efd_node;
6326 cluster = new_entry->efd_start_cluster;
6327
6328 if (!*n) {
6329 /* first free block exent. We need to
6330 protect buddy cache from being freed,
6331 * otherwise we'll refresh it from
6332 * on-disk bitmap and lose not-yet-available
6333 * blocks */
6334 folio_get(e4b->bd_buddy_folio);
6335 folio_get(e4b->bd_bitmap_folio);
6336 }
6337 while (*n) {
6338 parent = *n;
6339 entry = rb_entry(parent, struct ext4_free_data, efd_node);
6340 if (cluster < entry->efd_start_cluster)
6341 n = &(*n)->rb_left;
6342 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
6343 n = &(*n)->rb_right;
6344 else {
6345 ext4_grp_locked_error(sb, group, 0,
6346 ext4_group_first_block_no(sb, group) +
6347 EXT4_C2B(sbi, cluster),
6348 "Block already on to-be-freed list");
6349 kmem_cache_free(ext4_free_data_cachep, new_entry);
6350 return;
6351 }
6352 }
6353
6354 rb_link_node(new_node, parent, n);
6355 rb_insert_color(new_node, &db->bb_free_root);
6356
6357 /* Now try to see the extent can be merged to left and right */
6358 node = rb_prev(new_node);
6359 if (node) {
6360 entry = rb_entry(node, struct ext4_free_data, efd_node);
6361 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6362 &(db->bb_free_root));
6363 }
6364
6365 node = rb_next(new_node);
6366 if (node) {
6367 entry = rb_entry(node, struct ext4_free_data, efd_node);
6368 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6369 &(db->bb_free_root));
6370 }
6371
6372 spin_lock(&sbi->s_md_lock);
6373 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]);
6374 sbi->s_mb_free_pending += clusters;
6375 spin_unlock(&sbi->s_md_lock);
6376}
6377
6378static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
6379 unsigned long count)
6380{
6381 struct super_block *sb = inode->i_sb;
6382 ext4_group_t group;
6383 ext4_grpblk_t blkoff;
6384
6385 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
6386 ext4_mb_mark_context(NULL, sb, false, group, blkoff, count,
6387 EXT4_MB_BITMAP_MARKED_CHECK |
6388 EXT4_MB_SYNC_UPDATE,
6389 NULL);
6390}
6391
6392/**
6393 * ext4_mb_clear_bb() -- helper function for freeing blocks.
6394 * Used by ext4_free_blocks()
6395 * @handle: handle for this transaction
6396 * @inode: inode
6397 * @block: starting physical block to be freed
6398 * @count: number of blocks to be freed
6399 * @flags: flags used by ext4_free_blocks
6400 */
6401static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
6402 ext4_fsblk_t block, unsigned long count,
6403 int flags)
6404{
6405 struct super_block *sb = inode->i_sb;
6406 struct ext4_group_info *grp;
6407 unsigned int overflow;
6408 ext4_grpblk_t bit;
6409 ext4_group_t block_group;
6410 struct ext4_sb_info *sbi;
6411 struct ext4_buddy e4b;
6412 unsigned int count_clusters;
6413 int err = 0;
6414 int mark_flags = 0;
6415 ext4_grpblk_t changed;
6416
6417 sbi = EXT4_SB(sb);
6418
6419 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6420 !ext4_inode_block_valid(inode, block, count)) {
6421 ext4_error(sb, "Freeing blocks in system zone - "
6422 "Block = %llu, count = %lu", block, count);
6423 /* err = 0. ext4_std_error should be a no op */
6424 goto error_out;
6425 }
6426 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6427
6428do_more:
6429 overflow = 0;
6430 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6431
6432 grp = ext4_get_group_info(sb, block_group);
6433 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
6434 return;
6435
6436 /*
6437 * Check to see if we are freeing blocks across a group
6438 * boundary.
6439 */
6440 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
6441 overflow = EXT4_C2B(sbi, bit) + count -
6442 EXT4_BLOCKS_PER_GROUP(sb);
6443 count -= overflow;
6444 /* The range changed so it's no longer validated */
6445 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6446 }
6447 count_clusters = EXT4_NUM_B2C(sbi, count);
6448 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6449
6450 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6451 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6452 GFP_NOFS|__GFP_NOFAIL);
6453 if (err)
6454 goto error_out;
6455
6456 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6457 !ext4_inode_block_valid(inode, block, count)) {
6458 ext4_error(sb, "Freeing blocks in system zone - "
6459 "Block = %llu, count = %lu", block, count);
6460 /* err = 0. ext4_std_error should be a no op */
6461 goto error_clean;
6462 }
6463
6464#ifdef AGGRESSIVE_CHECK
6465 mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK;
6466#endif
6467 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6468 count_clusters, mark_flags, &changed);
6469
6470
6471 if (err && changed == 0)
6472 goto error_clean;
6473
6474#ifdef AGGRESSIVE_CHECK
6475 BUG_ON(changed != count_clusters);
6476#endif
6477
6478 /*
6479 * We need to make sure we don't reuse the freed block until after the
6480 * transaction is committed. We make an exception if the inode is to be
6481 * written in writeback mode since writeback mode has weak data
6482 * consistency guarantees.
6483 */
6484 if (ext4_handle_valid(handle) &&
6485 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6486 !ext4_should_writeback_data(inode))) {
6487 struct ext4_free_data *new_entry;
6488 /*
6489 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6490 * to fail.
6491 */
6492 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6493 GFP_NOFS|__GFP_NOFAIL);
6494 new_entry->efd_start_cluster = bit;
6495 new_entry->efd_group = block_group;
6496 new_entry->efd_count = count_clusters;
6497 new_entry->efd_tid = handle->h_transaction->t_tid;
6498
6499 ext4_lock_group(sb, block_group);
6500 ext4_mb_free_metadata(handle, &e4b, new_entry);
6501 } else {
6502 if (test_opt(sb, DISCARD)) {
6503 err = ext4_issue_discard(sb, block_group, bit,
6504 count_clusters);
6505 /*
6506 * Ignore EOPNOTSUPP error. This is consistent with
6507 * what happens when using journal.
6508 */
6509 if (err == -EOPNOTSUPP)
6510 err = 0;
6511 if (err)
6512 ext4_msg(sb, KERN_WARNING, "discard request in"
6513 " group:%u block:%d count:%lu failed"
6514 " with %d", block_group, bit, count,
6515 err);
6516 }
6517
6518 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6519
6520 ext4_lock_group(sb, block_group);
6521 mb_free_blocks(inode, &e4b, bit, count_clusters);
6522 }
6523
6524 ext4_unlock_group(sb, block_group);
6525
6526 /*
6527 * on a bigalloc file system, defer the s_freeclusters_counter
6528 * update to the caller (ext4_remove_space and friends) so they
6529 * can determine if a cluster freed here should be rereserved
6530 */
6531 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6532 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6533 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6534 percpu_counter_add(&sbi->s_freeclusters_counter,
6535 count_clusters);
6536 }
6537
6538 if (overflow && !err) {
6539 block += count;
6540 count = overflow;
6541 ext4_mb_unload_buddy(&e4b);
6542 /* The range changed so it's no longer validated */
6543 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6544 goto do_more;
6545 }
6546
6547error_clean:
6548 ext4_mb_unload_buddy(&e4b);
6549error_out:
6550 ext4_std_error(sb, err);
6551}
6552
6553/**
6554 * ext4_free_blocks() -- Free given blocks and update quota
6555 * @handle: handle for this transaction
6556 * @inode: inode
6557 * @bh: optional buffer of the block to be freed
6558 * @block: starting physical block to be freed
6559 * @count: number of blocks to be freed
6560 * @flags: flags used by ext4_free_blocks
6561 */
6562void ext4_free_blocks(handle_t *handle, struct inode *inode,
6563 struct buffer_head *bh, ext4_fsblk_t block,
6564 unsigned long count, int flags)
6565{
6566 struct super_block *sb = inode->i_sb;
6567 unsigned int overflow;
6568 struct ext4_sb_info *sbi;
6569
6570 sbi = EXT4_SB(sb);
6571
6572 if (bh) {
6573 if (block)
6574 BUG_ON(block != bh->b_blocknr);
6575 else
6576 block = bh->b_blocknr;
6577 }
6578
6579 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6580 ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count));
6581 return;
6582 }
6583
6584 might_sleep();
6585
6586 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6587 !ext4_inode_block_valid(inode, block, count)) {
6588 ext4_error(sb, "Freeing blocks not in datazone - "
6589 "block = %llu, count = %lu", block, count);
6590 return;
6591 }
6592 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6593
6594 ext4_debug("freeing block %llu\n", block);
6595 trace_ext4_free_blocks(inode, block, count, flags);
6596
6597 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6598 BUG_ON(count > 1);
6599
6600 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6601 inode, bh, block);
6602 }
6603
6604 /*
6605 * If the extent to be freed does not begin on a cluster
6606 * boundary, we need to deal with partial clusters at the
6607 * beginning and end of the extent. Normally we will free
6608 * blocks at the beginning or the end unless we are explicitly
6609 * requested to avoid doing so.
6610 */
6611 overflow = EXT4_PBLK_COFF(sbi, block);
6612 if (overflow) {
6613 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6614 overflow = sbi->s_cluster_ratio - overflow;
6615 block += overflow;
6616 if (count > overflow)
6617 count -= overflow;
6618 else
6619 return;
6620 } else {
6621 block -= overflow;
6622 count += overflow;
6623 }
6624 /* The range changed so it's no longer validated */
6625 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6626 }
6627 overflow = EXT4_LBLK_COFF(sbi, count);
6628 if (overflow) {
6629 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6630 if (count > overflow)
6631 count -= overflow;
6632 else
6633 return;
6634 } else
6635 count += sbi->s_cluster_ratio - overflow;
6636 /* The range changed so it's no longer validated */
6637 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6638 }
6639
6640 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6641 int i;
6642 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6643
6644 for (i = 0; i < count; i++) {
6645 cond_resched();
6646 if (is_metadata)
6647 bh = sb_find_get_block(inode->i_sb, block + i);
6648 ext4_forget(handle, is_metadata, inode, bh, block + i);
6649 }
6650 }
6651
6652 ext4_mb_clear_bb(handle, inode, block, count, flags);
6653}
6654
6655/**
6656 * ext4_group_add_blocks() -- Add given blocks to an existing group
6657 * @handle: handle to this transaction
6658 * @sb: super block
6659 * @block: start physical block to add to the block group
6660 * @count: number of blocks to free
6661 *
6662 * This marks the blocks as free in the bitmap and buddy.
6663 */
6664int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6665 ext4_fsblk_t block, unsigned long count)
6666{
6667 ext4_group_t block_group;
6668 ext4_grpblk_t bit;
6669 struct ext4_sb_info *sbi = EXT4_SB(sb);
6670 struct ext4_buddy e4b;
6671 int err = 0;
6672 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6673 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6674 unsigned long cluster_count = last_cluster - first_cluster + 1;
6675 ext4_grpblk_t changed;
6676
6677 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6678
6679 if (cluster_count == 0)
6680 return 0;
6681
6682 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6683 /*
6684 * Check to see if we are freeing blocks across a group
6685 * boundary.
6686 */
6687 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6688 ext4_warning(sb, "too many blocks added to group %u",
6689 block_group);
6690 err = -EINVAL;
6691 goto error_out;
6692 }
6693
6694 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6695 if (err)
6696 goto error_out;
6697
6698 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6699 ext4_error(sb, "Adding blocks in system zones - "
6700 "Block = %llu, count = %lu",
6701 block, count);
6702 err = -EINVAL;
6703 goto error_clean;
6704 }
6705
6706 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6707 cluster_count, EXT4_MB_BITMAP_MARKED_CHECK,
6708 &changed);
6709 if (err && changed == 0)
6710 goto error_clean;
6711
6712 if (changed != cluster_count)
6713 ext4_error(sb, "bit already cleared in group %u", block_group);
6714
6715 ext4_lock_group(sb, block_group);
6716 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6717 ext4_unlock_group(sb, block_group);
6718 percpu_counter_add(&sbi->s_freeclusters_counter,
6719 changed);
6720
6721error_clean:
6722 ext4_mb_unload_buddy(&e4b);
6723error_out:
6724 ext4_std_error(sb, err);
6725 return err;
6726}
6727
6728/**
6729 * ext4_trim_extent -- function to TRIM one single free extent in the group
6730 * @sb: super block for the file system
6731 * @start: starting block of the free extent in the alloc. group
6732 * @count: number of blocks to TRIM
6733 * @e4b: ext4 buddy for the group
6734 *
6735 * Trim "count" blocks starting at "start" in the "group". To assure that no
6736 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6737 * be called with under the group lock.
6738 */
6739static int ext4_trim_extent(struct super_block *sb,
6740 int start, int count, struct ext4_buddy *e4b)
6741__releases(bitlock)
6742__acquires(bitlock)
6743{
6744 struct ext4_free_extent ex;
6745 ext4_group_t group = e4b->bd_group;
6746 int ret = 0;
6747
6748 trace_ext4_trim_extent(sb, group, start, count);
6749
6750 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6751
6752 ex.fe_start = start;
6753 ex.fe_group = group;
6754 ex.fe_len = count;
6755
6756 /*
6757 * Mark blocks used, so no one can reuse them while
6758 * being trimmed.
6759 */
6760 mb_mark_used(e4b, &ex);
6761 ext4_unlock_group(sb, group);
6762 ret = ext4_issue_discard(sb, group, start, count);
6763 ext4_lock_group(sb, group);
6764 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6765 return ret;
6766}
6767
6768static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb,
6769 ext4_group_t grp)
6770{
6771 unsigned long nr_clusters_in_group;
6772
6773 if (grp < (ext4_get_groups_count(sb) - 1))
6774 nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb);
6775 else
6776 nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) -
6777 ext4_group_first_block_no(sb, grp))
6778 >> EXT4_CLUSTER_BITS(sb);
6779
6780 return nr_clusters_in_group - 1;
6781}
6782
6783static bool ext4_trim_interrupted(void)
6784{
6785 return fatal_signal_pending(current) || freezing(current);
6786}
6787
6788static int ext4_try_to_trim_range(struct super_block *sb,
6789 struct ext4_buddy *e4b, ext4_grpblk_t start,
6790 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6791__acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6792__releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6793{
6794 ext4_grpblk_t next, count, free_count, last, origin_start;
6795 bool set_trimmed = false;
6796 void *bitmap;
6797
6798 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
6799 return 0;
6800
6801 last = ext4_last_grp_cluster(sb, e4b->bd_group);
6802 bitmap = e4b->bd_bitmap;
6803 if (start == 0 && max >= last)
6804 set_trimmed = true;
6805 origin_start = start;
6806 start = max(e4b->bd_info->bb_first_free, start);
6807 count = 0;
6808 free_count = 0;
6809
6810 while (start <= max) {
6811 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6812 if (start > max)
6813 break;
6814
6815 next = mb_find_next_bit(bitmap, last + 1, start);
6816 if (origin_start == 0 && next >= last)
6817 set_trimmed = true;
6818
6819 if ((next - start) >= minblocks) {
6820 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6821
6822 if (ret && ret != -EOPNOTSUPP)
6823 return count;
6824 count += next - start;
6825 }
6826 free_count += next - start;
6827 start = next + 1;
6828
6829 if (ext4_trim_interrupted())
6830 return count;
6831
6832 if (need_resched()) {
6833 ext4_unlock_group(sb, e4b->bd_group);
6834 cond_resched();
6835 ext4_lock_group(sb, e4b->bd_group);
6836 }
6837
6838 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6839 break;
6840 }
6841
6842 if (set_trimmed)
6843 EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info);
6844
6845 return count;
6846}
6847
6848/**
6849 * ext4_trim_all_free -- function to trim all free space in alloc. group
6850 * @sb: super block for file system
6851 * @group: group to be trimmed
6852 * @start: first group block to examine
6853 * @max: last group block to examine
6854 * @minblocks: minimum extent block count
6855 *
6856 * ext4_trim_all_free walks through group's block bitmap searching for free
6857 * extents. When the free extent is found, mark it as used in group buddy
6858 * bitmap. Then issue a TRIM command on this extent and free the extent in
6859 * the group buddy bitmap.
6860 */
6861static ext4_grpblk_t
6862ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6863 ext4_grpblk_t start, ext4_grpblk_t max,
6864 ext4_grpblk_t minblocks)
6865{
6866 struct ext4_buddy e4b;
6867 int ret;
6868
6869 trace_ext4_trim_all_free(sb, group, start, max);
6870
6871 ret = ext4_mb_load_buddy(sb, group, &e4b);
6872 if (ret) {
6873 ext4_warning(sb, "Error %d loading buddy information for %u",
6874 ret, group);
6875 return ret;
6876 }
6877
6878 ext4_lock_group(sb, group);
6879
6880 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6881 minblocks < EXT4_SB(sb)->s_last_trim_minblks)
6882 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6883 else
6884 ret = 0;
6885
6886 ext4_unlock_group(sb, group);
6887 ext4_mb_unload_buddy(&e4b);
6888
6889 ext4_debug("trimmed %d blocks in the group %d\n",
6890 ret, group);
6891
6892 return ret;
6893}
6894
6895/**
6896 * ext4_trim_fs() -- trim ioctl handle function
6897 * @sb: superblock for filesystem
6898 * @range: fstrim_range structure
6899 *
6900 * start: First Byte to trim
6901 * len: number of Bytes to trim from start
6902 * minlen: minimum extent length in Bytes
6903 * ext4_trim_fs goes through all allocation groups containing Bytes from
6904 * start to start+len. For each such a group ext4_trim_all_free function
6905 * is invoked to trim all free space.
6906 */
6907int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6908{
6909 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6910 struct ext4_group_info *grp;
6911 ext4_group_t group, first_group, last_group;
6912 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6913 uint64_t start, end, minlen, trimmed = 0;
6914 ext4_fsblk_t first_data_blk =
6915 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6916 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6917 int ret = 0;
6918
6919 start = range->start >> sb->s_blocksize_bits;
6920 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6921 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6922 range->minlen >> sb->s_blocksize_bits);
6923
6924 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6925 start >= max_blks ||
6926 range->len < sb->s_blocksize)
6927 return -EINVAL;
6928 /* No point to try to trim less than discard granularity */
6929 if (range->minlen < discard_granularity) {
6930 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6931 discard_granularity >> sb->s_blocksize_bits);
6932 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6933 goto out;
6934 }
6935 if (end >= max_blks - 1)
6936 end = max_blks - 1;
6937 if (end <= first_data_blk)
6938 goto out;
6939 if (start < first_data_blk)
6940 start = first_data_blk;
6941
6942 /* Determine first and last group to examine based on start and end */
6943 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6944 &first_group, &first_cluster);
6945 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6946 &last_group, &last_cluster);
6947
6948 /* end now represents the last cluster to discard in this group */
6949 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6950
6951 for (group = first_group; group <= last_group; group++) {
6952 if (ext4_trim_interrupted())
6953 break;
6954 grp = ext4_get_group_info(sb, group);
6955 if (!grp)
6956 continue;
6957 /* We only do this if the grp has never been initialized */
6958 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6959 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6960 if (ret)
6961 break;
6962 }
6963
6964 /*
6965 * For all the groups except the last one, last cluster will
6966 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6967 * change it for the last group, note that last_cluster is
6968 * already computed earlier by ext4_get_group_no_and_offset()
6969 */
6970 if (group == last_group)
6971 end = last_cluster;
6972 if (grp->bb_free >= minlen) {
6973 cnt = ext4_trim_all_free(sb, group, first_cluster,
6974 end, minlen);
6975 if (cnt < 0) {
6976 ret = cnt;
6977 break;
6978 }
6979 trimmed += cnt;
6980 }
6981
6982 /*
6983 * For every group except the first one, we are sure
6984 * that the first cluster to discard will be cluster #0.
6985 */
6986 first_cluster = 0;
6987 }
6988
6989 if (!ret)
6990 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6991
6992out:
6993 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6994 return ret;
6995}
6996
6997/* Iterate all the free extents in the group. */
6998int
6999ext4_mballoc_query_range(
7000 struct super_block *sb,
7001 ext4_group_t group,
7002 ext4_grpblk_t first,
7003 ext4_grpblk_t end,
7004 ext4_mballoc_query_range_fn meta_formatter,
7005 ext4_mballoc_query_range_fn formatter,
7006 void *priv)
7007{
7008 void *bitmap;
7009 ext4_grpblk_t start, next;
7010 struct ext4_buddy e4b;
7011 int error;
7012
7013 error = ext4_mb_load_buddy(sb, group, &e4b);
7014 if (error)
7015 return error;
7016 bitmap = e4b.bd_bitmap;
7017
7018 ext4_lock_group(sb, group);
7019
7020 start = max(e4b.bd_info->bb_first_free, first);
7021 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
7022 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
7023 if (meta_formatter && start != first) {
7024 if (start > end)
7025 start = end;
7026 ext4_unlock_group(sb, group);
7027 error = meta_formatter(sb, group, first, start - first,
7028 priv);
7029 if (error)
7030 goto out_unload;
7031 ext4_lock_group(sb, group);
7032 }
7033 while (start <= end) {
7034 start = mb_find_next_zero_bit(bitmap, end + 1, start);
7035 if (start > end)
7036 break;
7037 next = mb_find_next_bit(bitmap, end + 1, start);
7038
7039 ext4_unlock_group(sb, group);
7040 error = formatter(sb, group, start, next - start, priv);
7041 if (error)
7042 goto out_unload;
7043 ext4_lock_group(sb, group);
7044
7045 start = next + 1;
7046 }
7047
7048 ext4_unlock_group(sb, group);
7049out_unload:
7050 ext4_mb_unload_buddy(&e4b);
7051
7052 return error;
7053}
7054
7055#ifdef CONFIG_EXT4_KUNIT_TESTS
7056#include "mballoc-test.c"
7057#endif