]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blame - fs/ext4/mballoc.c
Merge tag 'mmc-4.1-rc1' of git://git.linaro.org/people/ulf.hansson/mmc
[mirror_ubuntu-zesty-kernel.git] / fs / ext4 / mballoc.c
CommitLineData
c9de560d
AT
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
18aadd47 24#include "ext4_jbd2.h"
8f6e39a7 25#include "mballoc.h"
28623c2f 26#include <linux/log2.h>
a0b30c12 27#include <linux/module.h>
5a0e3ad6 28#include <linux/slab.h>
9bffad1e
TT
29#include <trace/events/ext4.h>
30
a0b30c12
TT
31#ifdef CONFIG_EXT4_DEBUG
32ushort ext4_mballoc_debug __read_mostly;
33
34module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
36#endif
37
c9de560d
AT
38/*
39 * MUSTDO:
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
42 *
43 * TODO v4:
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
47 * - quota
48 * - reservation for superuser
49 *
50 * TODO v3:
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
55 * - error handling
56 */
57
58/*
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
61 *
b713a5ec
TT
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
c9de560d
AT
71 *
72 * The main motivation for having small file use group preallocation is to
b713a5ec 73 * ensure that we have small files closer together on the disk.
c9de560d 74 *
b713a5ec
TT
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
78 * represented as:
c9de560d
AT
79 *
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
53accfa9
TT
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
c9de560d
AT
84 *
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
caaf7a29
TM
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
c9de560d
AT
89 *
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
92 * pa_free.
93 *
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
caaf7a29 96 * prealloc space. These are per CPU prealloc list represented as
c9de560d
AT
97 *
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
99 *
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
102 *
103 * The locality group prealloc space is used looking at whether we have
25985edc 104 * enough free space (pa_free) within the prealloc space.
c9de560d
AT
105 *
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
114 * inode as:
115 *
116 * { page }
c3a326a6 117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
c9de560d
AT
118 *
119 *
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
124 *
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
127 *
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
131 *
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
27baebb8
TT
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
c9de560d 140 * 512 blocks. This can be tuned via
d7a1fee1 141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
c9de560d
AT
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
d7a1fee1
DE
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
c9de560d 146 *
d7a1fee1 147 * The regular allocator (using the buddy cache) supports a few tunables.
c9de560d 148 *
b713a5ec
TT
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
c9de560d 152 *
b713a5ec 153 * The regular allocator uses buddy scan only if the request len is power of
c9de560d
AT
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
b713a5ec 156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
af901ca1 157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
b713a5ec
TT
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
c9de560d 161 * min_to_scan indicate how long the mballoc __must__ look for a best
b713a5ec 162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
c9de560d
AT
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
caaf7a29 166 * can be used for allocation. ext4_mb_good_group explains how the groups are
c9de560d
AT
167 * checked.
168 *
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
173 */
174
175/*
176 * mballoc operates on the following data:
177 * - on-disk bitmap
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
180 *
181 * there are two types of preallocations:
182 * - inode
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
191 * - locality group
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
196 *
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
199 *
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
203 *
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
207 *
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
210 *
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
221 *
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
232 * block
233 *
234 * so, now we're building a concurrency table:
235 * - init buddy vs.
236 * - new PA
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
239 * - use inode PA
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
242 * - discard inode PA
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
248 * - new PA vs.
249 * - use inode PA
250 * i_data_sem serializes them
251 * - discard inode PA
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
257 * - use inode PA
258 * - use inode PA
259 * i_data_sem or another mutex should serializes them
260 * - discard inode PA
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
266 *
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
274 *
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
277 *
278 */
279
280 /*
281 * Logic in few words:
282 *
283 * - allocation:
284 * load group
285 * find blocks
286 * mark bits in on-disk bitmap
287 * release group
288 *
289 * - use preallocation:
290 * find proper PA (per-inode or group)
291 * load group
292 * mark bits in on-disk bitmap
293 * release group
294 * release PA
295 *
296 * - free:
297 * load group
298 * mark bits in on-disk bitmap
299 * release group
300 *
301 * - discard preallocations in group:
302 * mark PAs deleted
303 * move them onto local list
304 * load on-disk bitmap
305 * load group
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
308 *
309 * - discard inode's preallocations:
310 */
311
312/*
313 * Locking rules
314 *
315 * Locks:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
318 * - per-pa lock (pa)
319 *
320 * Paths:
321 * - new pa
322 * object
323 * group
324 *
325 * - find and use pa:
326 * pa
327 *
328 * - release consumed pa:
329 * pa
330 * group
331 * object
332 *
333 * - generate in-core bitmap:
334 * group
335 * pa
336 *
337 * - discard all for given object (inode, locality group):
338 * object
339 * pa
340 * group
341 *
342 * - discard all for given group:
343 * group
344 * pa
345 * group
346 * object
347 *
348 */
c3a326a6
AK
349static struct kmem_cache *ext4_pspace_cachep;
350static struct kmem_cache *ext4_ac_cachep;
18aadd47 351static struct kmem_cache *ext4_free_data_cachep;
fb1813f4
CW
352
353/* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
2892c15d 356#define NR_GRPINFO_CACHES 8
fb1813f4
CW
357static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
358
2892c15d
ES
359static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
363};
364
c3a326a6
AK
365static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
366 ext4_group_t group);
7a2fcbf7
AK
367static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
368 ext4_group_t group);
18aadd47
BJ
369static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
c3a326a6 371
ffad0a44
AK
372static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
373{
c9de560d 374#if BITS_PER_LONG == 64
ffad0a44
AK
375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
c9de560d 377#elif BITS_PER_LONG == 32
ffad0a44
AK
378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
c9de560d
AT
380#else
381#error "how many bits you are?!"
382#endif
ffad0a44
AK
383 return addr;
384}
c9de560d
AT
385
386static inline int mb_test_bit(int bit, void *addr)
387{
388 /*
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
391 */
ffad0a44 392 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT
393 return ext4_test_bit(bit, addr);
394}
395
396static inline void mb_set_bit(int bit, void *addr)
397{
ffad0a44 398 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT
399 ext4_set_bit(bit, addr);
400}
401
c9de560d
AT
402static inline void mb_clear_bit(int bit, void *addr)
403{
ffad0a44 404 addr = mb_correct_addr_and_bit(&bit, addr);
c9de560d
AT
405 ext4_clear_bit(bit, addr);
406}
407
eabe0444
AS
408static inline int mb_test_and_clear_bit(int bit, void *addr)
409{
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
412}
413
ffad0a44
AK
414static inline int mb_find_next_zero_bit(void *addr, int max, int start)
415{
e7dfb246 416 int fix = 0, ret, tmpmax;
ffad0a44 417 addr = mb_correct_addr_and_bit(&fix, addr);
e7dfb246 418 tmpmax = max + fix;
ffad0a44
AK
419 start += fix;
420
e7dfb246
AK
421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
422 if (ret > max)
423 return max;
424 return ret;
ffad0a44
AK
425}
426
427static inline int mb_find_next_bit(void *addr, int max, int start)
428{
e7dfb246 429 int fix = 0, ret, tmpmax;
ffad0a44 430 addr = mb_correct_addr_and_bit(&fix, addr);
e7dfb246 431 tmpmax = max + fix;
ffad0a44
AK
432 start += fix;
433
e7dfb246
AK
434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
435 if (ret > max)
436 return max;
437 return ret;
ffad0a44
AK
438}
439
c9de560d
AT
440static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
441{
442 char *bb;
443
c5e8f3f3 444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
c9de560d
AT
445 BUG_ON(max == NULL);
446
447 if (order > e4b->bd_blkbits + 1) {
448 *max = 0;
449 return NULL;
450 }
451
452 /* at order 0 we see each particular block */
84b775a3
CL
453 if (order == 0) {
454 *max = 1 << (e4b->bd_blkbits + 3);
c5e8f3f3 455 return e4b->bd_bitmap;
84b775a3 456 }
c9de560d 457
c5e8f3f3 458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
c9de560d
AT
459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
460
461 return bb;
462}
463
464#ifdef DOUBLE_CHECK
465static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
467{
468 int i;
469 struct super_block *sb = e4b->bd_sb;
470
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
472 return;
bc8e6740 473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
c9de560d
AT
474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
5661bd68
AM
477
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
53accfa9 479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
5d1b1b3f 480 ext4_grp_locked_error(sb, e4b->bd_group,
e29136f8
TT
481 inode ? inode->i_ino : 0,
482 blocknr,
483 "freeing block already freed "
484 "(bit %u)",
485 first + i);
c9de560d
AT
486 }
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
488 }
489}
490
491static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
492{
493 int i;
494
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
496 return;
bc8e6740 497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
c9de560d
AT
498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
501 }
502}
503
504static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
505{
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
508 int i;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
9d8b9ec4
TT
513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
516 "on disk/prealloc",
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
c9de560d
AT
518 BUG();
519 }
520 }
521 }
522}
523
524#else
525static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
527{
528 return;
529}
530static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
532{
533 return;
534}
535static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
536{
537 return;
538}
539#endif
540
541#ifdef AGGRESSIVE_CHECK
542
543#define MB_CHECK_ASSERT(assert) \
544do { \
545 if (!(assert)) { \
546 printk(KERN_EMERG \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
549 BUG(); \
550 } \
551} while (0)
552
553static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
555{
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
558 int max;
559 int max2;
560 int i;
561 int j;
562 int k;
563 int count;
564 struct ext4_group_info *grp;
565 int fragments = 0;
566 int fstart;
567 struct list_head *cur;
568 void *buddy;
569 void *buddy2;
570
c9de560d
AT
571 {
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
574 return 0;
575 }
576
577 while (order > 1) {
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
584
585 count = 0;
586 for (i = 0; i < max; i++) {
587
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
591 MB_CHECK_ASSERT(
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
594 MB_CHECK_ASSERT(
595 mb_test_bit(i << 1, buddy2));
596 }
597 continue;
598 }
599
0a10da73 600 /* both bits in buddy2 must be 1 */
c9de560d
AT
601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
603
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
606 MB_CHECK_ASSERT(
c5e8f3f3 607 !mb_test_bit(k, e4b->bd_bitmap));
c9de560d
AT
608 }
609 count++;
610 }
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
612 order--;
613 }
614
615 fstart = -1;
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
620 if (fstart == -1) {
621 fragments++;
622 fstart = i;
623 }
624 continue;
625 }
626 fstart = -1;
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
630 k = i >> j;
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
633 }
634 }
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
637
638 grp = ext4_get_group_info(sb, e4b->bd_group);
c9de560d
AT
639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
60bd63d1
SR
642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
c9de560d 644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
60bd63d1 645 for (i = 0; i < pa->pa_len; i++)
c9de560d
AT
646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
647 }
648 return 0;
649}
650#undef MB_CHECK_ASSERT
651#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
46e665e9 652 __FILE__, __func__, __LINE__)
c9de560d
AT
653#else
654#define mb_check_buddy(e4b)
655#endif
656
7c786059
CL
657/*
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
662 */
c9de560d 663static void ext4_mb_mark_free_simple(struct super_block *sb,
a36b4498 664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
c9de560d
AT
665 struct ext4_group_info *grp)
666{
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
a36b4498
ES
668 ext4_grpblk_t min;
669 ext4_grpblk_t max;
670 ext4_grpblk_t chunk;
c9de560d
AT
671 unsigned short border;
672
7137d7a4 673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
c9de560d
AT
674
675 border = 2 << sb->s_blocksize_bits;
676
677 while (len > 0) {
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
680
681 /* find how many blocks of power 2 we need to mark */
682 min = fls(len) - 1;
683
684 if (max < min)
685 min = max;
686 chunk = 1 << min;
687
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
690 if (min > 0)
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
693
694 len -= chunk;
695 first += chunk;
696 }
697}
698
8a57d9d6
CW
699/*
700 * Cache the order of the largest free extent we have available in this block
701 * group.
702 */
703static void
704mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
705{
706 int i;
707 int bits;
708
709 grp->bb_largest_free_order = -1; /* uninit */
710
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
715 break;
716 }
717 }
718}
719
089ceecc
ES
720static noinline_for_stack
721void ext4_mb_generate_buddy(struct super_block *sb,
c9de560d
AT
722 void *buddy, void *bitmap, ext4_group_t group)
723{
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
e43bb4e6 725 struct ext4_sb_info *sbi = EXT4_SB(sb);
7137d7a4 726 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
a36b4498
ES
727 ext4_grpblk_t i = 0;
728 ext4_grpblk_t first;
729 ext4_grpblk_t len;
c9de560d
AT
730 unsigned free = 0;
731 unsigned fragments = 0;
732 unsigned long long period = get_cycles();
733
734 /* initialize buddy from bitmap which is aggregation
735 * of on-disk bitmap and preallocations */
ffad0a44 736 i = mb_find_next_zero_bit(bitmap, max, 0);
c9de560d
AT
737 grp->bb_first_free = i;
738 while (i < max) {
739 fragments++;
740 first = i;
ffad0a44 741 i = mb_find_next_bit(bitmap, max, i);
c9de560d
AT
742 len = i - first;
743 free += len;
744 if (len > 1)
745 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
746 else
747 grp->bb_counters[0]++;
748 if (i < max)
ffad0a44 749 i = mb_find_next_zero_bit(bitmap, max, i);
c9de560d
AT
750 }
751 grp->bb_fragments = fragments;
752
753 if (free != grp->bb_free) {
e29136f8 754 ext4_grp_locked_error(sb, group, 0, 0,
94d4c066
TT
755 "block bitmap and bg descriptor "
756 "inconsistent: %u vs %u free clusters",
e29136f8 757 free, grp->bb_free);
e56eb659 758 /*
163a203d 759 * If we intend to continue, we consider group descriptor
e56eb659
AK
760 * corrupt and update bb_free using bitmap value
761 */
c9de560d 762 grp->bb_free = free;
e43bb4e6
NJ
763 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
764 percpu_counter_sub(&sbi->s_freeclusters_counter,
765 grp->bb_free);
163a203d 766 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
c9de560d 767 }
8a57d9d6 768 mb_set_largest_free_order(sb, grp);
c9de560d
AT
769
770 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
771
772 period = get_cycles() - period;
773 spin_lock(&EXT4_SB(sb)->s_bal_lock);
774 EXT4_SB(sb)->s_mb_buddies_generated++;
775 EXT4_SB(sb)->s_mb_generation_time += period;
776 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
777}
778
eabe0444
AS
779static void mb_regenerate_buddy(struct ext4_buddy *e4b)
780{
781 int count;
782 int order = 1;
783 void *buddy;
784
785 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
786 ext4_set_bits(buddy, 0, count);
787 }
788 e4b->bd_info->bb_fragments = 0;
789 memset(e4b->bd_info->bb_counters, 0,
790 sizeof(*e4b->bd_info->bb_counters) *
791 (e4b->bd_sb->s_blocksize_bits + 2));
792
793 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
794 e4b->bd_bitmap, e4b->bd_group);
795}
796
c9de560d
AT
797/* The buddy information is attached the buddy cache inode
798 * for convenience. The information regarding each group
799 * is loaded via ext4_mb_load_buddy. The information involve
800 * block bitmap and buddy information. The information are
801 * stored in the inode as
802 *
803 * { page }
c3a326a6 804 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
c9de560d
AT
805 *
806 *
807 * one block each for bitmap and buddy information.
808 * So for each group we take up 2 blocks. A page can
809 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
810 * So it can have information regarding groups_per_page which
811 * is blocks_per_page/2
8a57d9d6
CW
812 *
813 * Locking note: This routine takes the block group lock of all groups
814 * for this page; do not hold this lock when calling this routine!
c9de560d
AT
815 */
816
817static int ext4_mb_init_cache(struct page *page, char *incore)
818{
8df9675f 819 ext4_group_t ngroups;
c9de560d
AT
820 int blocksize;
821 int blocks_per_page;
822 int groups_per_page;
823 int err = 0;
824 int i;
813e5727 825 ext4_group_t first_group, group;
c9de560d
AT
826 int first_block;
827 struct super_block *sb;
828 struct buffer_head *bhs;
fa77dcfa 829 struct buffer_head **bh = NULL;
c9de560d
AT
830 struct inode *inode;
831 char *data;
832 char *bitmap;
9b8b7d35 833 struct ext4_group_info *grinfo;
c9de560d 834
6ba495e9 835 mb_debug(1, "init page %lu\n", page->index);
c9de560d
AT
836
837 inode = page->mapping->host;
838 sb = inode->i_sb;
8df9675f 839 ngroups = ext4_get_groups_count(sb);
c9de560d
AT
840 blocksize = 1 << inode->i_blkbits;
841 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
842
843 groups_per_page = blocks_per_page >> 1;
844 if (groups_per_page == 0)
845 groups_per_page = 1;
846
847 /* allocate buffer_heads to read bitmaps */
848 if (groups_per_page > 1) {
c9de560d
AT
849 i = sizeof(struct buffer_head *) * groups_per_page;
850 bh = kzalloc(i, GFP_NOFS);
813e5727
TT
851 if (bh == NULL) {
852 err = -ENOMEM;
c9de560d 853 goto out;
813e5727 854 }
c9de560d
AT
855 } else
856 bh = &bhs;
857
858 first_group = page->index * blocks_per_page / 2;
859
860 /* read all groups the page covers into the cache */
813e5727
TT
861 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
862 if (group >= ngroups)
c9de560d
AT
863 break;
864
813e5727 865 grinfo = ext4_get_group_info(sb, group);
9b8b7d35
AG
866 /*
867 * If page is uptodate then we came here after online resize
868 * which added some new uninitialized group info structs, so
869 * we must skip all initialized uptodate buddies on the page,
870 * which may be currently in use by an allocating task.
871 */
872 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
873 bh[i] = NULL;
874 continue;
875 }
813e5727
TT
876 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
877 err = -ENOMEM;
c9de560d 878 goto out;
2ccb5fb9 879 }
813e5727 880 mb_debug(1, "read bitmap for group %u\n", group);
c9de560d
AT
881 }
882
883 /* wait for I/O completion */
813e5727
TT
884 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
885 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
886 err = -EIO;
c9de560d 887 goto out;
813e5727
TT
888 }
889 }
c9de560d
AT
890
891 first_block = page->index * blocks_per_page;
892 for (i = 0; i < blocks_per_page; i++) {
c9de560d 893 group = (first_block + i) >> 1;
8df9675f 894 if (group >= ngroups)
c9de560d
AT
895 break;
896
9b8b7d35
AG
897 if (!bh[group - first_group])
898 /* skip initialized uptodate buddy */
899 continue;
900
c9de560d
AT
901 /*
902 * data carry information regarding this
903 * particular group in the format specified
904 * above
905 *
906 */
907 data = page_address(page) + (i * blocksize);
908 bitmap = bh[group - first_group]->b_data;
909
910 /*
911 * We place the buddy block and bitmap block
912 * close together
913 */
914 if ((first_block + i) & 1) {
915 /* this is block of buddy */
916 BUG_ON(incore == NULL);
6ba495e9 917 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
c9de560d 918 group, page->index, i * blocksize);
f307333e 919 trace_ext4_mb_buddy_bitmap_load(sb, group);
c9de560d
AT
920 grinfo = ext4_get_group_info(sb, group);
921 grinfo->bb_fragments = 0;
922 memset(grinfo->bb_counters, 0,
1927805e
ES
923 sizeof(*grinfo->bb_counters) *
924 (sb->s_blocksize_bits+2));
c9de560d
AT
925 /*
926 * incore got set to the group block bitmap below
927 */
7a2fcbf7 928 ext4_lock_group(sb, group);
9b8b7d35
AG
929 /* init the buddy */
930 memset(data, 0xff, blocksize);
c9de560d 931 ext4_mb_generate_buddy(sb, data, incore, group);
7a2fcbf7 932 ext4_unlock_group(sb, group);
c9de560d
AT
933 incore = NULL;
934 } else {
935 /* this is block of bitmap */
936 BUG_ON(incore != NULL);
6ba495e9 937 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
c9de560d 938 group, page->index, i * blocksize);
f307333e 939 trace_ext4_mb_bitmap_load(sb, group);
c9de560d
AT
940
941 /* see comments in ext4_mb_put_pa() */
942 ext4_lock_group(sb, group);
943 memcpy(data, bitmap, blocksize);
944
945 /* mark all preallocated blks used in in-core bitmap */
946 ext4_mb_generate_from_pa(sb, data, group);
7a2fcbf7 947 ext4_mb_generate_from_freelist(sb, data, group);
c9de560d
AT
948 ext4_unlock_group(sb, group);
949
950 /* set incore so that the buddy information can be
951 * generated using this
952 */
953 incore = data;
954 }
955 }
956 SetPageUptodate(page);
957
958out:
959 if (bh) {
9b8b7d35 960 for (i = 0; i < groups_per_page; i++)
c9de560d
AT
961 brelse(bh[i]);
962 if (bh != &bhs)
963 kfree(bh);
964 }
965 return err;
966}
967
eee4adc7 968/*
2de8807b
AG
969 * Lock the buddy and bitmap pages. This make sure other parallel init_group
970 * on the same buddy page doesn't happen whild holding the buddy page lock.
971 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
972 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
eee4adc7 973 */
2de8807b
AG
974static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
975 ext4_group_t group, struct ext4_buddy *e4b)
eee4adc7 976{
2de8807b
AG
977 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
978 int block, pnum, poff;
eee4adc7 979 int blocks_per_page;
2de8807b
AG
980 struct page *page;
981
982 e4b->bd_buddy_page = NULL;
983 e4b->bd_bitmap_page = NULL;
eee4adc7
ES
984
985 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
986 /*
987 * the buddy cache inode stores the block bitmap
988 * and buddy information in consecutive blocks.
989 * So for each group we need two blocks.
990 */
991 block = group * 2;
992 pnum = block / blocks_per_page;
2de8807b
AG
993 poff = block % blocks_per_page;
994 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
995 if (!page)
c57ab39b 996 return -ENOMEM;
2de8807b
AG
997 BUG_ON(page->mapping != inode->i_mapping);
998 e4b->bd_bitmap_page = page;
999 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1000
1001 if (blocks_per_page >= 2) {
1002 /* buddy and bitmap are on the same page */
1003 return 0;
eee4adc7 1004 }
2de8807b
AG
1005
1006 block++;
1007 pnum = block / blocks_per_page;
2de8807b
AG
1008 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1009 if (!page)
c57ab39b 1010 return -ENOMEM;
2de8807b
AG
1011 BUG_ON(page->mapping != inode->i_mapping);
1012 e4b->bd_buddy_page = page;
1013 return 0;
eee4adc7
ES
1014}
1015
2de8807b 1016static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
eee4adc7 1017{
2de8807b
AG
1018 if (e4b->bd_bitmap_page) {
1019 unlock_page(e4b->bd_bitmap_page);
1020 page_cache_release(e4b->bd_bitmap_page);
1021 }
1022 if (e4b->bd_buddy_page) {
1023 unlock_page(e4b->bd_buddy_page);
1024 page_cache_release(e4b->bd_buddy_page);
eee4adc7 1025 }
eee4adc7
ES
1026}
1027
8a57d9d6
CW
1028/*
1029 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1030 * block group lock of all groups for this page; do not hold the BG lock when
1031 * calling this routine!
1032 */
b6a758ec
AK
1033static noinline_for_stack
1034int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1035{
1036
b6a758ec 1037 struct ext4_group_info *this_grp;
2de8807b
AG
1038 struct ext4_buddy e4b;
1039 struct page *page;
1040 int ret = 0;
b6a758ec 1041
b10a44c3 1042 might_sleep();
b6a758ec 1043 mb_debug(1, "init group %u\n", group);
b6a758ec
AK
1044 this_grp = ext4_get_group_info(sb, group);
1045 /*
08c3a813
AK
1046 * This ensures that we don't reinit the buddy cache
1047 * page which map to the group from which we are already
1048 * allocating. If we are looking at the buddy cache we would
1049 * have taken a reference using ext4_mb_load_buddy and that
2de8807b 1050 * would have pinned buddy page to page cache.
2457aec6
MG
1051 * The call to ext4_mb_get_buddy_page_lock will mark the
1052 * page accessed.
b6a758ec 1053 */
2de8807b
AG
1054 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1055 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
b6a758ec
AK
1056 /*
1057 * somebody initialized the group
1058 * return without doing anything
1059 */
b6a758ec
AK
1060 goto err;
1061 }
2de8807b
AG
1062
1063 page = e4b.bd_bitmap_page;
1064 ret = ext4_mb_init_cache(page, NULL);
1065 if (ret)
1066 goto err;
1067 if (!PageUptodate(page)) {
b6a758ec
AK
1068 ret = -EIO;
1069 goto err;
1070 }
b6a758ec 1071
2de8807b 1072 if (e4b.bd_buddy_page == NULL) {
b6a758ec
AK
1073 /*
1074 * If both the bitmap and buddy are in
1075 * the same page we don't need to force
1076 * init the buddy
1077 */
2de8807b
AG
1078 ret = 0;
1079 goto err;
b6a758ec 1080 }
2de8807b
AG
1081 /* init buddy cache */
1082 page = e4b.bd_buddy_page;
1083 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1084 if (ret)
1085 goto err;
1086 if (!PageUptodate(page)) {
b6a758ec
AK
1087 ret = -EIO;
1088 goto err;
1089 }
b6a758ec 1090err:
2de8807b 1091 ext4_mb_put_buddy_page_lock(&e4b);
b6a758ec
AK
1092 return ret;
1093}
1094
8a57d9d6
CW
1095/*
1096 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1097 * block group lock of all groups for this page; do not hold the BG lock when
1098 * calling this routine!
1099 */
4ddfef7b
ES
1100static noinline_for_stack int
1101ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1102 struct ext4_buddy *e4b)
c9de560d 1103{
c9de560d
AT
1104 int blocks_per_page;
1105 int block;
1106 int pnum;
1107 int poff;
1108 struct page *page;
fdf6c7a7 1109 int ret;
920313a7
AK
1110 struct ext4_group_info *grp;
1111 struct ext4_sb_info *sbi = EXT4_SB(sb);
1112 struct inode *inode = sbi->s_buddy_cache;
c9de560d 1113
b10a44c3 1114 might_sleep();
6ba495e9 1115 mb_debug(1, "load group %u\n", group);
c9de560d
AT
1116
1117 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
920313a7 1118 grp = ext4_get_group_info(sb, group);
c9de560d
AT
1119
1120 e4b->bd_blkbits = sb->s_blocksize_bits;
529da704 1121 e4b->bd_info = grp;
c9de560d
AT
1122 e4b->bd_sb = sb;
1123 e4b->bd_group = group;
1124 e4b->bd_buddy_page = NULL;
1125 e4b->bd_bitmap_page = NULL;
1126
f41c0750 1127 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
f41c0750
AK
1128 /*
1129 * we need full data about the group
1130 * to make a good selection
1131 */
1132 ret = ext4_mb_init_group(sb, group);
1133 if (ret)
1134 return ret;
f41c0750
AK
1135 }
1136
c9de560d
AT
1137 /*
1138 * the buddy cache inode stores the block bitmap
1139 * and buddy information in consecutive blocks.
1140 * So for each group we need two blocks.
1141 */
1142 block = group * 2;
1143 pnum = block / blocks_per_page;
1144 poff = block % blocks_per_page;
1145
1146 /* we could use find_or_create_page(), but it locks page
1147 * what we'd like to avoid in fast path ... */
2457aec6 1148 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
c9de560d
AT
1149 if (page == NULL || !PageUptodate(page)) {
1150 if (page)
920313a7
AK
1151 /*
1152 * drop the page reference and try
1153 * to get the page with lock. If we
1154 * are not uptodate that implies
1155 * somebody just created the page but
1156 * is yet to initialize the same. So
1157 * wait for it to initialize.
1158 */
c9de560d
AT
1159 page_cache_release(page);
1160 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1161 if (page) {
1162 BUG_ON(page->mapping != inode->i_mapping);
1163 if (!PageUptodate(page)) {
fdf6c7a7
SF
1164 ret = ext4_mb_init_cache(page, NULL);
1165 if (ret) {
1166 unlock_page(page);
1167 goto err;
1168 }
c9de560d
AT
1169 mb_cmp_bitmaps(e4b, page_address(page) +
1170 (poff * sb->s_blocksize));
1171 }
1172 unlock_page(page);
1173 }
1174 }
c57ab39b
YL
1175 if (page == NULL) {
1176 ret = -ENOMEM;
1177 goto err;
1178 }
1179 if (!PageUptodate(page)) {
fdf6c7a7 1180 ret = -EIO;
c9de560d 1181 goto err;
fdf6c7a7 1182 }
2457aec6
MG
1183
1184 /* Pages marked accessed already */
c9de560d
AT
1185 e4b->bd_bitmap_page = page;
1186 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
c9de560d
AT
1187
1188 block++;
1189 pnum = block / blocks_per_page;
1190 poff = block % blocks_per_page;
1191
2457aec6 1192 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
c9de560d
AT
1193 if (page == NULL || !PageUptodate(page)) {
1194 if (page)
1195 page_cache_release(page);
1196 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1197 if (page) {
1198 BUG_ON(page->mapping != inode->i_mapping);
fdf6c7a7
SF
1199 if (!PageUptodate(page)) {
1200 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1201 if (ret) {
1202 unlock_page(page);
1203 goto err;
1204 }
1205 }
c9de560d
AT
1206 unlock_page(page);
1207 }
1208 }
c57ab39b
YL
1209 if (page == NULL) {
1210 ret = -ENOMEM;
1211 goto err;
1212 }
1213 if (!PageUptodate(page)) {
fdf6c7a7 1214 ret = -EIO;
c9de560d 1215 goto err;
fdf6c7a7 1216 }
2457aec6
MG
1217
1218 /* Pages marked accessed already */
c9de560d
AT
1219 e4b->bd_buddy_page = page;
1220 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
c9de560d
AT
1221
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1224
1225 return 0;
1226
1227err:
26626f11
YR
1228 if (page)
1229 page_cache_release(page);
c9de560d
AT
1230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
fdf6c7a7 1236 return ret;
c9de560d
AT
1237}
1238
e39e07fd 1239static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
c9de560d
AT
1240{
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1245}
1246
1247
1248static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1249{
1250 int order = 1;
1251 void *bb;
1252
c5e8f3f3 1253 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
c9de560d
AT
1254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1255
c5e8f3f3 1256 bb = e4b->bd_buddy;
c9de560d
AT
1257 while (order <= e4b->bd_blkbits + 1) {
1258 block = block >> 1;
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1261 return order;
1262 }
1263 bb += 1 << (e4b->bd_blkbits - order);
1264 order++;
1265 }
1266 return 0;
1267}
1268
955ce5f5 1269static void mb_clear_bits(void *bm, int cur, int len)
c9de560d
AT
1270{
1271 __u32 *addr;
1272
1273 len = cur + len;
1274 while (cur < len) {
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1278 *addr = 0;
1279 cur += 32;
1280 continue;
1281 }
955ce5f5 1282 mb_clear_bit(cur, bm);
c9de560d
AT
1283 cur++;
1284 }
1285}
1286
eabe0444
AS
1287/* clear bits in given range
1288 * will return first found zero bit if any, -1 otherwise
1289 */
1290static int mb_test_and_clear_bits(void *bm, int cur, int len)
1291{
1292 __u32 *addr;
1293 int zero_bit = -1;
1294
1295 len = cur + len;
1296 while (cur < len) {
1297 if ((cur & 31) == 0 && (len - cur) >= 32) {
1298 /* fast path: clear whole word at once */
1299 addr = bm + (cur >> 3);
1300 if (*addr != (__u32)(-1) && zero_bit == -1)
1301 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1302 *addr = 0;
1303 cur += 32;
1304 continue;
1305 }
1306 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1307 zero_bit = cur;
1308 cur++;
1309 }
1310
1311 return zero_bit;
1312}
1313
c3e94d1d 1314void ext4_set_bits(void *bm, int cur, int len)
c9de560d
AT
1315{
1316 __u32 *addr;
1317
1318 len = cur + len;
1319 while (cur < len) {
1320 if ((cur & 31) == 0 && (len - cur) >= 32) {
1321 /* fast path: set whole word at once */
1322 addr = bm + (cur >> 3);
1323 *addr = 0xffffffff;
1324 cur += 32;
1325 continue;
1326 }
955ce5f5 1327 mb_set_bit(cur, bm);
c9de560d
AT
1328 cur++;
1329 }
1330}
1331
eabe0444
AS
1332/*
1333 * _________________________________________________________________ */
1334
1335static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1336{
1337 if (mb_test_bit(*bit + side, bitmap)) {
1338 mb_clear_bit(*bit, bitmap);
1339 (*bit) -= side;
1340 return 1;
1341 }
1342 else {
1343 (*bit) += side;
1344 mb_set_bit(*bit, bitmap);
1345 return -1;
1346 }
1347}
1348
1349static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1350{
1351 int max;
1352 int order = 1;
1353 void *buddy = mb_find_buddy(e4b, order, &max);
1354
1355 while (buddy) {
1356 void *buddy2;
1357
1358 /* Bits in range [first; last] are known to be set since
1359 * corresponding blocks were allocated. Bits in range
1360 * (first; last) will stay set because they form buddies on
1361 * upper layer. We just deal with borders if they don't
1362 * align with upper layer and then go up.
1363 * Releasing entire group is all about clearing
1364 * single bit of highest order buddy.
1365 */
1366
1367 /* Example:
1368 * ---------------------------------
1369 * | 1 | 1 | 1 | 1 |
1370 * ---------------------------------
1371 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1372 * ---------------------------------
1373 * 0 1 2 3 4 5 6 7
1374 * \_____________________/
1375 *
1376 * Neither [1] nor [6] is aligned to above layer.
1377 * Left neighbour [0] is free, so mark it busy,
1378 * decrease bb_counters and extend range to
1379 * [0; 6]
1380 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1381 * mark [6] free, increase bb_counters and shrink range to
1382 * [0; 5].
1383 * Then shift range to [0; 2], go up and do the same.
1384 */
1385
1386
1387 if (first & 1)
1388 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1389 if (!(last & 1))
1390 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1391 if (first > last)
1392 break;
1393 order++;
1394
1395 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1396 mb_clear_bits(buddy, first, last - first + 1);
1397 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1398 break;
1399 }
1400 first >>= 1;
1401 last >>= 1;
1402 buddy = buddy2;
1403 }
1404}
1405
7e5a8cdd 1406static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
eabe0444 1407 int first, int count)
c9de560d 1408{
eabe0444
AS
1409 int left_is_free = 0;
1410 int right_is_free = 0;
1411 int block;
1412 int last = first + count - 1;
c9de560d
AT
1413 struct super_block *sb = e4b->bd_sb;
1414
c99d1e6e
TT
1415 if (WARN_ON(count == 0))
1416 return;
eabe0444 1417 BUG_ON(last >= (sb->s_blocksize << 3));
bc8e6740 1418 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
163a203d
DW
1419 /* Don't bother if the block group is corrupt. */
1420 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1421 return;
1422
c9de560d
AT
1423 mb_check_buddy(e4b);
1424 mb_free_blocks_double(inode, e4b, first, count);
1425
1426 e4b->bd_info->bb_free += count;
1427 if (first < e4b->bd_info->bb_first_free)
1428 e4b->bd_info->bb_first_free = first;
1429
eabe0444
AS
1430 /* access memory sequentially: check left neighbour,
1431 * clear range and then check right neighbour
1432 */
c9de560d 1433 if (first != 0)
eabe0444
AS
1434 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1435 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1436 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1437 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1438
1439 if (unlikely(block != -1)) {
e43bb4e6 1440 struct ext4_sb_info *sbi = EXT4_SB(sb);
eabe0444
AS
1441 ext4_fsblk_t blocknr;
1442
1443 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1444 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1445 ext4_grp_locked_error(sb, e4b->bd_group,
1446 inode ? inode->i_ino : 0,
1447 blocknr,
1448 "freeing already freed block "
163a203d
DW
1449 "(bit %u); block bitmap corrupt.",
1450 block);
e43bb4e6
NJ
1451 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1452 percpu_counter_sub(&sbi->s_freeclusters_counter,
1453 e4b->bd_info->bb_free);
163a203d
DW
1454 /* Mark the block group as corrupt. */
1455 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1456 &e4b->bd_info->bb_state);
eabe0444
AS
1457 mb_regenerate_buddy(e4b);
1458 goto done;
1459 }
1460
1461 /* let's maintain fragments counter */
1462 if (left_is_free && right_is_free)
c9de560d 1463 e4b->bd_info->bb_fragments--;
eabe0444 1464 else if (!left_is_free && !right_is_free)
c9de560d
AT
1465 e4b->bd_info->bb_fragments++;
1466
eabe0444
AS
1467 /* buddy[0] == bd_bitmap is a special case, so handle
1468 * it right away and let mb_buddy_mark_free stay free of
1469 * zero order checks.
1470 * Check if neighbours are to be coaleasced,
1471 * adjust bitmap bb_counters and borders appropriately.
1472 */
1473 if (first & 1) {
1474 first += !left_is_free;
1475 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1476 }
1477 if (!(last & 1)) {
1478 last -= !right_is_free;
1479 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1480 }
c9de560d 1481
eabe0444
AS
1482 if (first <= last)
1483 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
c9de560d 1484
eabe0444 1485done:
8a57d9d6 1486 mb_set_largest_free_order(sb, e4b->bd_info);
c9de560d 1487 mb_check_buddy(e4b);
c9de560d
AT
1488}
1489
15c006a2 1490static int mb_find_extent(struct ext4_buddy *e4b, int block,
c9de560d
AT
1491 int needed, struct ext4_free_extent *ex)
1492{
1493 int next = block;
15c006a2 1494 int max, order;
c9de560d
AT
1495 void *buddy;
1496
bc8e6740 1497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
c9de560d
AT
1498 BUG_ON(ex == NULL);
1499
15c006a2 1500 buddy = mb_find_buddy(e4b, 0, &max);
c9de560d
AT
1501 BUG_ON(buddy == NULL);
1502 BUG_ON(block >= max);
1503 if (mb_test_bit(block, buddy)) {
1504 ex->fe_len = 0;
1505 ex->fe_start = 0;
1506 ex->fe_group = 0;
1507 return 0;
1508 }
1509
15c006a2
RD
1510 /* find actual order */
1511 order = mb_find_order_for_block(e4b, block);
1512 block = block >> order;
c9de560d
AT
1513
1514 ex->fe_len = 1 << order;
1515 ex->fe_start = block << order;
1516 ex->fe_group = e4b->bd_group;
1517
1518 /* calc difference from given start */
1519 next = next - ex->fe_start;
1520 ex->fe_len -= next;
1521 ex->fe_start += next;
1522
1523 while (needed > ex->fe_len &&
d8ec0c39 1524 mb_find_buddy(e4b, order, &max)) {
c9de560d
AT
1525
1526 if (block + 1 >= max)
1527 break;
1528
1529 next = (block + 1) * (1 << order);
c5e8f3f3 1530 if (mb_test_bit(next, e4b->bd_bitmap))
c9de560d
AT
1531 break;
1532
b051d8dc 1533 order = mb_find_order_for_block(e4b, next);
c9de560d 1534
c9de560d
AT
1535 block = next >> order;
1536 ex->fe_len += 1 << order;
1537 }
1538
1539 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1540 return ex->fe_len;
1541}
1542
1543static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1544{
1545 int ord;
1546 int mlen = 0;
1547 int max = 0;
1548 int cur;
1549 int start = ex->fe_start;
1550 int len = ex->fe_len;
1551 unsigned ret = 0;
1552 int len0 = len;
1553 void *buddy;
1554
1555 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1556 BUG_ON(e4b->bd_group != ex->fe_group);
bc8e6740 1557 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
c9de560d
AT
1558 mb_check_buddy(e4b);
1559 mb_mark_used_double(e4b, start, len);
1560
1561 e4b->bd_info->bb_free -= len;
1562 if (e4b->bd_info->bb_first_free == start)
1563 e4b->bd_info->bb_first_free += len;
1564
1565 /* let's maintain fragments counter */
1566 if (start != 0)
c5e8f3f3 1567 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
c9de560d 1568 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
c5e8f3f3 1569 max = !mb_test_bit(start + len, e4b->bd_bitmap);
c9de560d
AT
1570 if (mlen && max)
1571 e4b->bd_info->bb_fragments++;
1572 else if (!mlen && !max)
1573 e4b->bd_info->bb_fragments--;
1574
1575 /* let's maintain buddy itself */
1576 while (len) {
1577 ord = mb_find_order_for_block(e4b, start);
1578
1579 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1580 /* the whole chunk may be allocated at once! */
1581 mlen = 1 << ord;
1582 buddy = mb_find_buddy(e4b, ord, &max);
1583 BUG_ON((start >> ord) >= max);
1584 mb_set_bit(start >> ord, buddy);
1585 e4b->bd_info->bb_counters[ord]--;
1586 start += mlen;
1587 len -= mlen;
1588 BUG_ON(len < 0);
1589 continue;
1590 }
1591
1592 /* store for history */
1593 if (ret == 0)
1594 ret = len | (ord << 16);
1595
1596 /* we have to split large buddy */
1597 BUG_ON(ord <= 0);
1598 buddy = mb_find_buddy(e4b, ord, &max);
1599 mb_set_bit(start >> ord, buddy);
1600 e4b->bd_info->bb_counters[ord]--;
1601
1602 ord--;
1603 cur = (start >> ord) & ~1U;
1604 buddy = mb_find_buddy(e4b, ord, &max);
1605 mb_clear_bit(cur, buddy);
1606 mb_clear_bit(cur + 1, buddy);
1607 e4b->bd_info->bb_counters[ord]++;
1608 e4b->bd_info->bb_counters[ord]++;
1609 }
8a57d9d6 1610 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
c9de560d 1611
c5e8f3f3 1612 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
c9de560d
AT
1613 mb_check_buddy(e4b);
1614
1615 return ret;
1616}
1617
1618/*
1619 * Must be called under group lock!
1620 */
1621static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1622 struct ext4_buddy *e4b)
1623{
1624 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1625 int ret;
1626
1627 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1628 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1629
1630 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1631 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1632 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1633
1634 /* preallocation can change ac_b_ex, thus we store actually
1635 * allocated blocks for history */
1636 ac->ac_f_ex = ac->ac_b_ex;
1637
1638 ac->ac_status = AC_STATUS_FOUND;
1639 ac->ac_tail = ret & 0xffff;
1640 ac->ac_buddy = ret >> 16;
1641
c3a326a6
AK
1642 /*
1643 * take the page reference. We want the page to be pinned
1644 * so that we don't get a ext4_mb_init_cache_call for this
1645 * group until we update the bitmap. That would mean we
1646 * double allocate blocks. The reference is dropped
1647 * in ext4_mb_release_context
1648 */
c9de560d
AT
1649 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1650 get_page(ac->ac_bitmap_page);
1651 ac->ac_buddy_page = e4b->bd_buddy_page;
1652 get_page(ac->ac_buddy_page);
c9de560d 1653 /* store last allocated for subsequent stream allocation */
4ba74d00 1654 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
c9de560d
AT
1655 spin_lock(&sbi->s_md_lock);
1656 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1657 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1658 spin_unlock(&sbi->s_md_lock);
1659 }
1660}
1661
1662/*
1663 * regular allocator, for general purposes allocation
1664 */
1665
1666static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1667 struct ext4_buddy *e4b,
1668 int finish_group)
1669{
1670 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1671 struct ext4_free_extent *bex = &ac->ac_b_ex;
1672 struct ext4_free_extent *gex = &ac->ac_g_ex;
1673 struct ext4_free_extent ex;
1674 int max;
1675
032115fc
AK
1676 if (ac->ac_status == AC_STATUS_FOUND)
1677 return;
c9de560d
AT
1678 /*
1679 * We don't want to scan for a whole year
1680 */
1681 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1682 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1683 ac->ac_status = AC_STATUS_BREAK;
1684 return;
1685 }
1686
1687 /*
1688 * Haven't found good chunk so far, let's continue
1689 */
1690 if (bex->fe_len < gex->fe_len)
1691 return;
1692
1693 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1694 && bex->fe_group == e4b->bd_group) {
1695 /* recheck chunk's availability - we don't know
1696 * when it was found (within this lock-unlock
1697 * period or not) */
15c006a2 1698 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
c9de560d
AT
1699 if (max >= gex->fe_len) {
1700 ext4_mb_use_best_found(ac, e4b);
1701 return;
1702 }
1703 }
1704}
1705
1706/*
1707 * The routine checks whether found extent is good enough. If it is,
1708 * then the extent gets marked used and flag is set to the context
1709 * to stop scanning. Otherwise, the extent is compared with the
1710 * previous found extent and if new one is better, then it's stored
1711 * in the context. Later, the best found extent will be used, if
1712 * mballoc can't find good enough extent.
1713 *
1714 * FIXME: real allocation policy is to be designed yet!
1715 */
1716static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1717 struct ext4_free_extent *ex,
1718 struct ext4_buddy *e4b)
1719{
1720 struct ext4_free_extent *bex = &ac->ac_b_ex;
1721 struct ext4_free_extent *gex = &ac->ac_g_ex;
1722
1723 BUG_ON(ex->fe_len <= 0);
7137d7a4
TT
1724 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1725 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
c9de560d
AT
1726 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1727
1728 ac->ac_found++;
1729
1730 /*
1731 * The special case - take what you catch first
1732 */
1733 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1734 *bex = *ex;
1735 ext4_mb_use_best_found(ac, e4b);
1736 return;
1737 }
1738
1739 /*
1740 * Let's check whether the chuck is good enough
1741 */
1742 if (ex->fe_len == gex->fe_len) {
1743 *bex = *ex;
1744 ext4_mb_use_best_found(ac, e4b);
1745 return;
1746 }
1747
1748 /*
1749 * If this is first found extent, just store it in the context
1750 */
1751 if (bex->fe_len == 0) {
1752 *bex = *ex;
1753 return;
1754 }
1755
1756 /*
1757 * If new found extent is better, store it in the context
1758 */
1759 if (bex->fe_len < gex->fe_len) {
1760 /* if the request isn't satisfied, any found extent
1761 * larger than previous best one is better */
1762 if (ex->fe_len > bex->fe_len)
1763 *bex = *ex;
1764 } else if (ex->fe_len > gex->fe_len) {
1765 /* if the request is satisfied, then we try to find
1766 * an extent that still satisfy the request, but is
1767 * smaller than previous one */
1768 if (ex->fe_len < bex->fe_len)
1769 *bex = *ex;
1770 }
1771
1772 ext4_mb_check_limits(ac, e4b, 0);
1773}
1774
089ceecc
ES
1775static noinline_for_stack
1776int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
c9de560d
AT
1777 struct ext4_buddy *e4b)
1778{
1779 struct ext4_free_extent ex = ac->ac_b_ex;
1780 ext4_group_t group = ex.fe_group;
1781 int max;
1782 int err;
1783
1784 BUG_ON(ex.fe_len <= 0);
1785 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1786 if (err)
1787 return err;
1788
1789 ext4_lock_group(ac->ac_sb, group);
15c006a2 1790 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
c9de560d
AT
1791
1792 if (max > 0) {
1793 ac->ac_b_ex = ex;
1794 ext4_mb_use_best_found(ac, e4b);
1795 }
1796
1797 ext4_unlock_group(ac->ac_sb, group);
e39e07fd 1798 ext4_mb_unload_buddy(e4b);
c9de560d
AT
1799
1800 return 0;
1801}
1802
089ceecc
ES
1803static noinline_for_stack
1804int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
c9de560d
AT
1805 struct ext4_buddy *e4b)
1806{
1807 ext4_group_t group = ac->ac_g_ex.fe_group;
1808 int max;
1809 int err;
1810 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
838cd0cf 1811 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
c9de560d
AT
1812 struct ext4_free_extent ex;
1813
1814 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1815 return 0;
838cd0cf
YY
1816 if (grp->bb_free == 0)
1817 return 0;
c9de560d
AT
1818
1819 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1820 if (err)
1821 return err;
1822
163a203d
DW
1823 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1824 ext4_mb_unload_buddy(e4b);
1825 return 0;
1826 }
1827
c9de560d 1828 ext4_lock_group(ac->ac_sb, group);
15c006a2 1829 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
c9de560d 1830 ac->ac_g_ex.fe_len, &ex);
ab0c00fc 1831 ex.fe_logical = 0xDEADFA11; /* debug value */
c9de560d
AT
1832
1833 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1834 ext4_fsblk_t start;
1835
5661bd68
AM
1836 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1837 ex.fe_start;
c9de560d
AT
1838 /* use do_div to get remainder (would be 64-bit modulo) */
1839 if (do_div(start, sbi->s_stripe) == 0) {
1840 ac->ac_found++;
1841 ac->ac_b_ex = ex;
1842 ext4_mb_use_best_found(ac, e4b);
1843 }
1844 } else if (max >= ac->ac_g_ex.fe_len) {
1845 BUG_ON(ex.fe_len <= 0);
1846 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1847 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1848 ac->ac_found++;
1849 ac->ac_b_ex = ex;
1850 ext4_mb_use_best_found(ac, e4b);
1851 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1852 /* Sometimes, caller may want to merge even small
1853 * number of blocks to an existing extent */
1854 BUG_ON(ex.fe_len <= 0);
1855 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1856 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1857 ac->ac_found++;
1858 ac->ac_b_ex = ex;
1859 ext4_mb_use_best_found(ac, e4b);
1860 }
1861 ext4_unlock_group(ac->ac_sb, group);
e39e07fd 1862 ext4_mb_unload_buddy(e4b);
c9de560d
AT
1863
1864 return 0;
1865}
1866
1867/*
1868 * The routine scans buddy structures (not bitmap!) from given order
1869 * to max order and tries to find big enough chunk to satisfy the req
1870 */
089ceecc
ES
1871static noinline_for_stack
1872void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
c9de560d
AT
1873 struct ext4_buddy *e4b)
1874{
1875 struct super_block *sb = ac->ac_sb;
1876 struct ext4_group_info *grp = e4b->bd_info;
1877 void *buddy;
1878 int i;
1879 int k;
1880 int max;
1881
1882 BUG_ON(ac->ac_2order <= 0);
1883 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1884 if (grp->bb_counters[i] == 0)
1885 continue;
1886
1887 buddy = mb_find_buddy(e4b, i, &max);
1888 BUG_ON(buddy == NULL);
1889
ffad0a44 1890 k = mb_find_next_zero_bit(buddy, max, 0);
c9de560d
AT
1891 BUG_ON(k >= max);
1892
1893 ac->ac_found++;
1894
1895 ac->ac_b_ex.fe_len = 1 << i;
1896 ac->ac_b_ex.fe_start = k << i;
1897 ac->ac_b_ex.fe_group = e4b->bd_group;
1898
1899 ext4_mb_use_best_found(ac, e4b);
1900
1901 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1902
1903 if (EXT4_SB(sb)->s_mb_stats)
1904 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1905
1906 break;
1907 }
1908}
1909
1910/*
1911 * The routine scans the group and measures all found extents.
1912 * In order to optimize scanning, caller must pass number of
1913 * free blocks in the group, so the routine can know upper limit.
1914 */
089ceecc
ES
1915static noinline_for_stack
1916void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
c9de560d
AT
1917 struct ext4_buddy *e4b)
1918{
1919 struct super_block *sb = ac->ac_sb;
c5e8f3f3 1920 void *bitmap = e4b->bd_bitmap;
c9de560d
AT
1921 struct ext4_free_extent ex;
1922 int i;
1923 int free;
1924
1925 free = e4b->bd_info->bb_free;
1926 BUG_ON(free <= 0);
1927
1928 i = e4b->bd_info->bb_first_free;
1929
1930 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
ffad0a44 1931 i = mb_find_next_zero_bit(bitmap,
7137d7a4
TT
1932 EXT4_CLUSTERS_PER_GROUP(sb), i);
1933 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
26346ff6 1934 /*
e56eb659 1935 * IF we have corrupt bitmap, we won't find any
26346ff6
AK
1936 * free blocks even though group info says we
1937 * we have free blocks
1938 */
e29136f8 1939 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
53accfa9 1940 "%d free clusters as per "
fde4d95a 1941 "group info. But bitmap says 0",
26346ff6 1942 free);
c9de560d
AT
1943 break;
1944 }
1945
15c006a2 1946 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
c9de560d 1947 BUG_ON(ex.fe_len <= 0);
26346ff6 1948 if (free < ex.fe_len) {
e29136f8 1949 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
53accfa9 1950 "%d free clusters as per "
fde4d95a 1951 "group info. But got %d blocks",
26346ff6 1952 free, ex.fe_len);
e56eb659
AK
1953 /*
1954 * The number of free blocks differs. This mostly
1955 * indicate that the bitmap is corrupt. So exit
1956 * without claiming the space.
1957 */
1958 break;
26346ff6 1959 }
ab0c00fc 1960 ex.fe_logical = 0xDEADC0DE; /* debug value */
c9de560d
AT
1961 ext4_mb_measure_extent(ac, &ex, e4b);
1962
1963 i += ex.fe_len;
1964 free -= ex.fe_len;
1965 }
1966
1967 ext4_mb_check_limits(ac, e4b, 1);
1968}
1969
1970/*
1971 * This is a special case for storages like raid5
506bf2d8 1972 * we try to find stripe-aligned chunks for stripe-size-multiple requests
c9de560d 1973 */
089ceecc
ES
1974static noinline_for_stack
1975void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
c9de560d
AT
1976 struct ext4_buddy *e4b)
1977{
1978 struct super_block *sb = ac->ac_sb;
1979 struct ext4_sb_info *sbi = EXT4_SB(sb);
c5e8f3f3 1980 void *bitmap = e4b->bd_bitmap;
c9de560d
AT
1981 struct ext4_free_extent ex;
1982 ext4_fsblk_t first_group_block;
1983 ext4_fsblk_t a;
1984 ext4_grpblk_t i;
1985 int max;
1986
1987 BUG_ON(sbi->s_stripe == 0);
1988
1989 /* find first stripe-aligned block in group */
5661bd68
AM
1990 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1991
c9de560d
AT
1992 a = first_group_block + sbi->s_stripe - 1;
1993 do_div(a, sbi->s_stripe);
1994 i = (a * sbi->s_stripe) - first_group_block;
1995
7137d7a4 1996 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
c9de560d 1997 if (!mb_test_bit(i, bitmap)) {
15c006a2 1998 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
c9de560d
AT
1999 if (max >= sbi->s_stripe) {
2000 ac->ac_found++;
ab0c00fc 2001 ex.fe_logical = 0xDEADF00D; /* debug value */
c9de560d
AT
2002 ac->ac_b_ex = ex;
2003 ext4_mb_use_best_found(ac, e4b);
2004 break;
2005 }
2006 }
2007 i += sbi->s_stripe;
2008 }
2009}
2010
8a57d9d6 2011/* This is now called BEFORE we load the buddy bitmap. */
c9de560d
AT
2012static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2013 ext4_group_t group, int cr)
2014{
2015 unsigned free, fragments;
a4912123 2016 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
c9de560d
AT
2017 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2018
2019 BUG_ON(cr < 0 || cr >= 4);
8a57d9d6 2020
01fc48e8
TT
2021 free = grp->bb_free;
2022 if (free == 0)
2023 return 0;
2024 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2025 return 0;
2026
163a203d
DW
2027 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2028 return 0;
2029
8a57d9d6
CW
2030 /* We only do this if the grp has never been initialized */
2031 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2032 int ret = ext4_mb_init_group(ac->ac_sb, group);
2033 if (ret)
2034 return 0;
2035 }
c9de560d 2036
c9de560d 2037 fragments = grp->bb_fragments;
c9de560d
AT
2038 if (fragments == 0)
2039 return 0;
2040
2041 switch (cr) {
2042 case 0:
2043 BUG_ON(ac->ac_2order == 0);
c9de560d 2044
a4912123
TT
2045 /* Avoid using the first bg of a flexgroup for data files */
2046 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2047 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2048 ((group % flex_size) == 0))
2049 return 0;
2050
40ae3487
TT
2051 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2052 (free / fragments) >= ac->ac_g_ex.fe_len)
2053 return 1;
2054
2055 if (grp->bb_largest_free_order < ac->ac_2order)
2056 return 0;
2057
8a57d9d6 2058 return 1;
c9de560d
AT
2059 case 1:
2060 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2061 return 1;
2062 break;
2063 case 2:
2064 if (free >= ac->ac_g_ex.fe_len)
2065 return 1;
2066 break;
2067 case 3:
2068 return 1;
2069 default:
2070 BUG();
2071 }
2072
2073 return 0;
2074}
2075
4ddfef7b
ES
2076static noinline_for_stack int
2077ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
c9de560d 2078{
8df9675f 2079 ext4_group_t ngroups, group, i;
c9de560d
AT
2080 int cr;
2081 int err = 0;
c9de560d
AT
2082 struct ext4_sb_info *sbi;
2083 struct super_block *sb;
2084 struct ext4_buddy e4b;
c9de560d
AT
2085
2086 sb = ac->ac_sb;
2087 sbi = EXT4_SB(sb);
8df9675f 2088 ngroups = ext4_get_groups_count(sb);
fb0a387d 2089 /* non-extent files are limited to low blocks/groups */
12e9b892 2090 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
fb0a387d
ES
2091 ngroups = sbi->s_blockfile_groups;
2092
c9de560d
AT
2093 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2094
2095 /* first, try the goal */
2096 err = ext4_mb_find_by_goal(ac, &e4b);
2097 if (err || ac->ac_status == AC_STATUS_FOUND)
2098 goto out;
2099
2100 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2101 goto out;
2102
2103 /*
2104 * ac->ac2_order is set only if the fe_len is a power of 2
2105 * if ac2_order is set we also set criteria to 0 so that we
2106 * try exact allocation using buddy.
2107 */
2108 i = fls(ac->ac_g_ex.fe_len);
2109 ac->ac_2order = 0;
2110 /*
2111 * We search using buddy data only if the order of the request
2112 * is greater than equal to the sbi_s_mb_order2_reqs
b713a5ec 2113 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
c9de560d
AT
2114 */
2115 if (i >= sbi->s_mb_order2_reqs) {
2116 /*
2117 * This should tell if fe_len is exactly power of 2
2118 */
2119 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2120 ac->ac_2order = i - 1;
2121 }
2122
4ba74d00
TT
2123 /* if stream allocation is enabled, use global goal */
2124 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
c9de560d
AT
2125 /* TBD: may be hot point */
2126 spin_lock(&sbi->s_md_lock);
2127 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2128 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2129 spin_unlock(&sbi->s_md_lock);
2130 }
4ba74d00 2131
c9de560d
AT
2132 /* Let's just scan groups to find more-less suitable blocks */
2133 cr = ac->ac_2order ? 0 : 1;
2134 /*
2135 * cr == 0 try to get exact allocation,
2136 * cr == 3 try to get anything
2137 */
2138repeat:
2139 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2140 ac->ac_criteria = cr;
ed8f9c75
AK
2141 /*
2142 * searching for the right group start
2143 * from the goal value specified
2144 */
2145 group = ac->ac_g_ex.fe_group;
2146
8df9675f 2147 for (i = 0; i < ngroups; group++, i++) {
2ed5724d 2148 cond_resched();
e6155736
LM
2149 /*
2150 * Artificially restricted ngroups for non-extent
2151 * files makes group > ngroups possible on first loop.
2152 */
2153 if (group >= ngroups)
c9de560d
AT
2154 group = 0;
2155
8a57d9d6
CW
2156 /* This now checks without needing the buddy page */
2157 if (!ext4_mb_good_group(ac, group, cr))
c9de560d
AT
2158 continue;
2159
c9de560d
AT
2160 err = ext4_mb_load_buddy(sb, group, &e4b);
2161 if (err)
2162 goto out;
2163
2164 ext4_lock_group(sb, group);
8a57d9d6
CW
2165
2166 /*
2167 * We need to check again after locking the
2168 * block group
2169 */
c9de560d 2170 if (!ext4_mb_good_group(ac, group, cr)) {
c9de560d 2171 ext4_unlock_group(sb, group);
e39e07fd 2172 ext4_mb_unload_buddy(&e4b);
c9de560d
AT
2173 continue;
2174 }
2175
2176 ac->ac_groups_scanned++;
40ae3487 2177 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
c9de560d 2178 ext4_mb_simple_scan_group(ac, &e4b);
506bf2d8
ES
2179 else if (cr == 1 && sbi->s_stripe &&
2180 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
c9de560d
AT
2181 ext4_mb_scan_aligned(ac, &e4b);
2182 else
2183 ext4_mb_complex_scan_group(ac, &e4b);
2184
2185 ext4_unlock_group(sb, group);
e39e07fd 2186 ext4_mb_unload_buddy(&e4b);
c9de560d
AT
2187
2188 if (ac->ac_status != AC_STATUS_CONTINUE)
2189 break;
2190 }
2191 }
2192
2193 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2194 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2195 /*
2196 * We've been searching too long. Let's try to allocate
2197 * the best chunk we've found so far
2198 */
2199
2200 ext4_mb_try_best_found(ac, &e4b);
2201 if (ac->ac_status != AC_STATUS_FOUND) {
2202 /*
2203 * Someone more lucky has already allocated it.
2204 * The only thing we can do is just take first
2205 * found block(s)
2206 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2207 */
2208 ac->ac_b_ex.fe_group = 0;
2209 ac->ac_b_ex.fe_start = 0;
2210 ac->ac_b_ex.fe_len = 0;
2211 ac->ac_status = AC_STATUS_CONTINUE;
2212 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2213 cr = 3;
2214 atomic_inc(&sbi->s_mb_lost_chunks);
2215 goto repeat;
2216 }
2217 }
2218out:
2219 return err;
2220}
2221
c9de560d
AT
2222static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2223{
2224 struct super_block *sb = seq->private;
c9de560d
AT
2225 ext4_group_t group;
2226
8df9675f 2227 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
c9de560d 2228 return NULL;
c9de560d 2229 group = *pos + 1;
a9df9a49 2230 return (void *) ((unsigned long) group);
c9de560d
AT
2231}
2232
2233static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2234{
2235 struct super_block *sb = seq->private;
c9de560d
AT
2236 ext4_group_t group;
2237
2238 ++*pos;
8df9675f 2239 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
c9de560d
AT
2240 return NULL;
2241 group = *pos + 1;
a9df9a49 2242 return (void *) ((unsigned long) group);
c9de560d
AT
2243}
2244
2245static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2246{
2247 struct super_block *sb = seq->private;
a9df9a49 2248 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
c9de560d 2249 int i;
1c8457ca 2250 int err, buddy_loaded = 0;
c9de560d 2251 struct ext4_buddy e4b;
1c8457ca 2252 struct ext4_group_info *grinfo;
c9de560d
AT
2253 struct sg {
2254 struct ext4_group_info info;
a36b4498 2255 ext4_grpblk_t counters[16];
c9de560d
AT
2256 } sg;
2257
2258 group--;
2259 if (group == 0)
2260 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2261 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2262 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2263 "group", "free", "frags", "first",
2264 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2265 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2266
2267 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2268 sizeof(struct ext4_group_info);
1c8457ca
AK
2269 grinfo = ext4_get_group_info(sb, group);
2270 /* Load the group info in memory only if not already loaded. */
2271 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2272 err = ext4_mb_load_buddy(sb, group, &e4b);
2273 if (err) {
2274 seq_printf(seq, "#%-5u: I/O error\n", group);
2275 return 0;
2276 }
2277 buddy_loaded = 1;
c9de560d 2278 }
1c8457ca 2279
c9de560d 2280 memcpy(&sg, ext4_get_group_info(sb, group), i);
1c8457ca
AK
2281
2282 if (buddy_loaded)
2283 ext4_mb_unload_buddy(&e4b);
c9de560d 2284
a9df9a49 2285 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
c9de560d
AT
2286 sg.info.bb_fragments, sg.info.bb_first_free);
2287 for (i = 0; i <= 13; i++)
2288 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2289 sg.info.bb_counters[i] : 0);
2290 seq_printf(seq, " ]\n");
2291
2292 return 0;
2293}
2294
2295static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2296{
2297}
2298
7f1346a9 2299static const struct seq_operations ext4_mb_seq_groups_ops = {
c9de560d
AT
2300 .start = ext4_mb_seq_groups_start,
2301 .next = ext4_mb_seq_groups_next,
2302 .stop = ext4_mb_seq_groups_stop,
2303 .show = ext4_mb_seq_groups_show,
2304};
2305
2306static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2307{
d9dda78b 2308 struct super_block *sb = PDE_DATA(inode);
c9de560d
AT
2309 int rc;
2310
2311 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2312 if (rc == 0) {
a271fe85 2313 struct seq_file *m = file->private_data;
c9de560d
AT
2314 m->private = sb;
2315 }
2316 return rc;
2317
2318}
2319
7f1346a9 2320static const struct file_operations ext4_mb_seq_groups_fops = {
c9de560d
AT
2321 .owner = THIS_MODULE,
2322 .open = ext4_mb_seq_groups_open,
2323 .read = seq_read,
2324 .llseek = seq_lseek,
2325 .release = seq_release,
2326};
2327
fb1813f4
CW
2328static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2329{
2330 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2331 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2332
2333 BUG_ON(!cachep);
2334 return cachep;
2335}
5f21b0e6 2336
28623c2f
TT
2337/*
2338 * Allocate the top-level s_group_info array for the specified number
2339 * of groups
2340 */
2341int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2342{
2343 struct ext4_sb_info *sbi = EXT4_SB(sb);
2344 unsigned size;
2345 struct ext4_group_info ***new_groupinfo;
2346
2347 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2348 EXT4_DESC_PER_BLOCK_BITS(sb);
2349 if (size <= sbi->s_group_info_size)
2350 return 0;
2351
2352 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2353 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2354 if (!new_groupinfo) {
2355 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2356 return -ENOMEM;
2357 }
2358 if (sbi->s_group_info) {
2359 memcpy(new_groupinfo, sbi->s_group_info,
2360 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
b93b41d4 2361 kvfree(sbi->s_group_info);
28623c2f
TT
2362 }
2363 sbi->s_group_info = new_groupinfo;
2364 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2365 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2366 sbi->s_group_info_size);
2367 return 0;
2368}
2369
5f21b0e6 2370/* Create and initialize ext4_group_info data for the given group. */
920313a7 2371int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
5f21b0e6
FB
2372 struct ext4_group_desc *desc)
2373{
fb1813f4 2374 int i;
5f21b0e6
FB
2375 int metalen = 0;
2376 struct ext4_sb_info *sbi = EXT4_SB(sb);
2377 struct ext4_group_info **meta_group_info;
fb1813f4 2378 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
5f21b0e6
FB
2379
2380 /*
2381 * First check if this group is the first of a reserved block.
2382 * If it's true, we have to allocate a new table of pointers
2383 * to ext4_group_info structures
2384 */
2385 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2386 metalen = sizeof(*meta_group_info) <<
2387 EXT4_DESC_PER_BLOCK_BITS(sb);
4fdb5543 2388 meta_group_info = kmalloc(metalen, GFP_NOFS);
5f21b0e6 2389 if (meta_group_info == NULL) {
7f6a11e7 2390 ext4_msg(sb, KERN_ERR, "can't allocate mem "
9d8b9ec4 2391 "for a buddy group");
5f21b0e6
FB
2392 goto exit_meta_group_info;
2393 }
2394 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2395 meta_group_info;
2396 }
2397
5f21b0e6
FB
2398 meta_group_info =
2399 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2400 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2401
4fdb5543 2402 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
5f21b0e6 2403 if (meta_group_info[i] == NULL) {
7f6a11e7 2404 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
5f21b0e6
FB
2405 goto exit_group_info;
2406 }
2407 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2408 &(meta_group_info[i]->bb_state));
2409
2410 /*
2411 * initialize bb_free to be able to skip
2412 * empty groups without initialization
2413 */
2414 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2415 meta_group_info[i]->bb_free =
cff1dfd7 2416 ext4_free_clusters_after_init(sb, group, desc);
5f21b0e6
FB
2417 } else {
2418 meta_group_info[i]->bb_free =
021b65bb 2419 ext4_free_group_clusters(sb, desc);
5f21b0e6
FB
2420 }
2421
2422 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
920313a7 2423 init_rwsem(&meta_group_info[i]->alloc_sem);
64e290ec 2424 meta_group_info[i]->bb_free_root = RB_ROOT;
8a57d9d6 2425 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
5f21b0e6
FB
2426
2427#ifdef DOUBLE_CHECK
2428 {
2429 struct buffer_head *bh;
2430 meta_group_info[i]->bb_bitmap =
4fdb5543 2431 kmalloc(sb->s_blocksize, GFP_NOFS);
5f21b0e6
FB
2432 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2433 bh = ext4_read_block_bitmap(sb, group);
2434 BUG_ON(bh == NULL);
2435 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2436 sb->s_blocksize);
2437 put_bh(bh);
2438 }
2439#endif
2440
2441 return 0;
2442
2443exit_group_info:
2444 /* If a meta_group_info table has been allocated, release it now */
caaf7a29 2445 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
5f21b0e6 2446 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
caaf7a29
TM
2447 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2448 }
5f21b0e6
FB
2449exit_meta_group_info:
2450 return -ENOMEM;
2451} /* ext4_mb_add_groupinfo */
2452
c9de560d
AT
2453static int ext4_mb_init_backend(struct super_block *sb)
2454{
8df9675f 2455 ext4_group_t ngroups = ext4_get_groups_count(sb);
c9de560d 2456 ext4_group_t i;
c9de560d 2457 struct ext4_sb_info *sbi = EXT4_SB(sb);
28623c2f 2458 int err;
5f21b0e6 2459 struct ext4_group_desc *desc;
fb1813f4 2460 struct kmem_cache *cachep;
5f21b0e6 2461
28623c2f
TT
2462 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2463 if (err)
2464 return err;
c9de560d 2465
c9de560d
AT
2466 sbi->s_buddy_cache = new_inode(sb);
2467 if (sbi->s_buddy_cache == NULL) {
9d8b9ec4 2468 ext4_msg(sb, KERN_ERR, "can't get new inode");
c9de560d
AT
2469 goto err_freesgi;
2470 }
48e6061b
YJ
2471 /* To avoid potentially colliding with an valid on-disk inode number,
2472 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2473 * not in the inode hash, so it should never be found by iget(), but
2474 * this will avoid confusion if it ever shows up during debugging. */
2475 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
c9de560d 2476 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
8df9675f 2477 for (i = 0; i < ngroups; i++) {
c9de560d
AT
2478 desc = ext4_get_group_desc(sb, i, NULL);
2479 if (desc == NULL) {
9d8b9ec4 2480 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
c9de560d
AT
2481 goto err_freebuddy;
2482 }
5f21b0e6
FB
2483 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2484 goto err_freebuddy;
c9de560d
AT
2485 }
2486
2487 return 0;
2488
2489err_freebuddy:
fb1813f4 2490 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
f1fa3342 2491 while (i-- > 0)
fb1813f4 2492 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
28623c2f 2493 i = sbi->s_group_info_size;
f1fa3342 2494 while (i-- > 0)
c9de560d
AT
2495 kfree(sbi->s_group_info[i]);
2496 iput(sbi->s_buddy_cache);
2497err_freesgi:
b93b41d4 2498 kvfree(sbi->s_group_info);
c9de560d
AT
2499 return -ENOMEM;
2500}
2501
2892c15d
ES
2502static void ext4_groupinfo_destroy_slabs(void)
2503{
2504 int i;
2505
2506 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2507 if (ext4_groupinfo_caches[i])
2508 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2509 ext4_groupinfo_caches[i] = NULL;
2510 }
2511}
2512
2513static int ext4_groupinfo_create_slab(size_t size)
2514{
2515 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2516 int slab_size;
2517 int blocksize_bits = order_base_2(size);
2518 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2519 struct kmem_cache *cachep;
2520
2521 if (cache_index >= NR_GRPINFO_CACHES)
2522 return -EINVAL;
2523
2524 if (unlikely(cache_index < 0))
2525 cache_index = 0;
2526
2527 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2528 if (ext4_groupinfo_caches[cache_index]) {
2529 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2530 return 0; /* Already created */
2531 }
2532
2533 slab_size = offsetof(struct ext4_group_info,
2534 bb_counters[blocksize_bits + 2]);
2535
2536 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2537 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2538 NULL);
2539
823ba01f
TM
2540 ext4_groupinfo_caches[cache_index] = cachep;
2541
2892c15d
ES
2542 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2543 if (!cachep) {
9d8b9ec4
TT
2544 printk(KERN_EMERG
2545 "EXT4-fs: no memory for groupinfo slab cache\n");
2892c15d
ES
2546 return -ENOMEM;
2547 }
2548
2892c15d
ES
2549 return 0;
2550}
2551
9d99012f 2552int ext4_mb_init(struct super_block *sb)
c9de560d
AT
2553{
2554 struct ext4_sb_info *sbi = EXT4_SB(sb);
6be2ded1 2555 unsigned i, j;
c9de560d
AT
2556 unsigned offset;
2557 unsigned max;
74767c5a 2558 int ret;
c9de560d 2559
1927805e 2560 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
c9de560d
AT
2561
2562 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2563 if (sbi->s_mb_offsets == NULL) {
fb1813f4
CW
2564 ret = -ENOMEM;
2565 goto out;
c9de560d 2566 }
ff7ef329 2567
1927805e 2568 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
c9de560d
AT
2569 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2570 if (sbi->s_mb_maxs == NULL) {
fb1813f4
CW
2571 ret = -ENOMEM;
2572 goto out;
2573 }
2574
2892c15d
ES
2575 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2576 if (ret < 0)
2577 goto out;
c9de560d
AT
2578
2579 /* order 0 is regular bitmap */
2580 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2581 sbi->s_mb_offsets[0] = 0;
2582
2583 i = 1;
2584 offset = 0;
2585 max = sb->s_blocksize << 2;
2586 do {
2587 sbi->s_mb_offsets[i] = offset;
2588 sbi->s_mb_maxs[i] = max;
2589 offset += 1 << (sb->s_blocksize_bits - i);
2590 max = max >> 1;
2591 i++;
2592 } while (i <= sb->s_blocksize_bits + 1);
2593
c9de560d 2594 spin_lock_init(&sbi->s_md_lock);
c9de560d
AT
2595 spin_lock_init(&sbi->s_bal_lock);
2596
2597 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2598 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2599 sbi->s_mb_stats = MB_DEFAULT_STATS;
2600 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2601 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
27baebb8
TT
2602 /*
2603 * The default group preallocation is 512, which for 4k block
2604 * sizes translates to 2 megabytes. However for bigalloc file
2605 * systems, this is probably too big (i.e, if the cluster size
2606 * is 1 megabyte, then group preallocation size becomes half a
2607 * gigabyte!). As a default, we will keep a two megabyte
2608 * group pralloc size for cluster sizes up to 64k, and after
2609 * that, we will force a minimum group preallocation size of
2610 * 32 clusters. This translates to 8 megs when the cluster
2611 * size is 256k, and 32 megs when the cluster size is 1 meg,
2612 * which seems reasonable as a default.
2613 */
2614 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2615 sbi->s_cluster_bits, 32);
d7a1fee1
DE
2616 /*
2617 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2618 * to the lowest multiple of s_stripe which is bigger than
2619 * the s_mb_group_prealloc as determined above. We want
2620 * the preallocation size to be an exact multiple of the
2621 * RAID stripe size so that preallocations don't fragment
2622 * the stripes.
2623 */
2624 if (sbi->s_stripe > 1) {
2625 sbi->s_mb_group_prealloc = roundup(
2626 sbi->s_mb_group_prealloc, sbi->s_stripe);
2627 }
c9de560d 2628
730c213c 2629 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
c9de560d 2630 if (sbi->s_locality_groups == NULL) {
fb1813f4 2631 ret = -ENOMEM;
029b10c5 2632 goto out;
c9de560d 2633 }
730c213c 2634 for_each_possible_cpu(i) {
c9de560d 2635 struct ext4_locality_group *lg;
730c213c 2636 lg = per_cpu_ptr(sbi->s_locality_groups, i);
c9de560d 2637 mutex_init(&lg->lg_mutex);
6be2ded1
AK
2638 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2639 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
c9de560d
AT
2640 spin_lock_init(&lg->lg_prealloc_lock);
2641 }
2642
79a77c5a
YJ
2643 /* init file for buddy data */
2644 ret = ext4_mb_init_backend(sb);
7aa0baea
TM
2645 if (ret != 0)
2646 goto out_free_locality_groups;
79a77c5a 2647
296c355c
TT
2648 if (sbi->s_proc)
2649 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2650 &ext4_mb_seq_groups_fops, sb);
c9de560d 2651
7aa0baea
TM
2652 return 0;
2653
2654out_free_locality_groups:
2655 free_percpu(sbi->s_locality_groups);
2656 sbi->s_locality_groups = NULL;
fb1813f4 2657out:
7aa0baea
TM
2658 kfree(sbi->s_mb_offsets);
2659 sbi->s_mb_offsets = NULL;
2660 kfree(sbi->s_mb_maxs);
2661 sbi->s_mb_maxs = NULL;
fb1813f4 2662 return ret;
c9de560d
AT
2663}
2664
955ce5f5 2665/* need to called with the ext4 group lock held */
c9de560d
AT
2666static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2667{
2668 struct ext4_prealloc_space *pa;
2669 struct list_head *cur, *tmp;
2670 int count = 0;
2671
2672 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2673 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2674 list_del(&pa->pa_group_list);
2675 count++;
688f05a0 2676 kmem_cache_free(ext4_pspace_cachep, pa);
c9de560d
AT
2677 }
2678 if (count)
6ba495e9 2679 mb_debug(1, "mballoc: %u PAs left\n", count);
c9de560d
AT
2680
2681}
2682
2683int ext4_mb_release(struct super_block *sb)
2684{
8df9675f 2685 ext4_group_t ngroups = ext4_get_groups_count(sb);
c9de560d
AT
2686 ext4_group_t i;
2687 int num_meta_group_infos;
2688 struct ext4_group_info *grinfo;
2689 struct ext4_sb_info *sbi = EXT4_SB(sb);
fb1813f4 2690 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
c9de560d 2691
95599968
SQ
2692 if (sbi->s_proc)
2693 remove_proc_entry("mb_groups", sbi->s_proc);
2694
c9de560d 2695 if (sbi->s_group_info) {
8df9675f 2696 for (i = 0; i < ngroups; i++) {
c9de560d
AT
2697 grinfo = ext4_get_group_info(sb, i);
2698#ifdef DOUBLE_CHECK
2699 kfree(grinfo->bb_bitmap);
2700#endif
2701 ext4_lock_group(sb, i);
2702 ext4_mb_cleanup_pa(grinfo);
2703 ext4_unlock_group(sb, i);
fb1813f4 2704 kmem_cache_free(cachep, grinfo);
c9de560d 2705 }
8df9675f 2706 num_meta_group_infos = (ngroups +
c9de560d
AT
2707 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2708 EXT4_DESC_PER_BLOCK_BITS(sb);
2709 for (i = 0; i < num_meta_group_infos; i++)
2710 kfree(sbi->s_group_info[i]);
b93b41d4 2711 kvfree(sbi->s_group_info);
c9de560d
AT
2712 }
2713 kfree(sbi->s_mb_offsets);
2714 kfree(sbi->s_mb_maxs);
bfcba2d0 2715 iput(sbi->s_buddy_cache);
c9de560d 2716 if (sbi->s_mb_stats) {
9d8b9ec4
TT
2717 ext4_msg(sb, KERN_INFO,
2718 "mballoc: %u blocks %u reqs (%u success)",
c9de560d
AT
2719 atomic_read(&sbi->s_bal_allocated),
2720 atomic_read(&sbi->s_bal_reqs),
2721 atomic_read(&sbi->s_bal_success));
9d8b9ec4
TT
2722 ext4_msg(sb, KERN_INFO,
2723 "mballoc: %u extents scanned, %u goal hits, "
2724 "%u 2^N hits, %u breaks, %u lost",
c9de560d
AT
2725 atomic_read(&sbi->s_bal_ex_scanned),
2726 atomic_read(&sbi->s_bal_goals),
2727 atomic_read(&sbi->s_bal_2orders),
2728 atomic_read(&sbi->s_bal_breaks),
2729 atomic_read(&sbi->s_mb_lost_chunks));
9d8b9ec4
TT
2730 ext4_msg(sb, KERN_INFO,
2731 "mballoc: %lu generated and it took %Lu",
ced156e4 2732 sbi->s_mb_buddies_generated,
c9de560d 2733 sbi->s_mb_generation_time);
9d8b9ec4
TT
2734 ext4_msg(sb, KERN_INFO,
2735 "mballoc: %u preallocated, %u discarded",
c9de560d
AT
2736 atomic_read(&sbi->s_mb_preallocated),
2737 atomic_read(&sbi->s_mb_discarded));
2738 }
2739
730c213c 2740 free_percpu(sbi->s_locality_groups);
c9de560d
AT
2741
2742 return 0;
2743}
2744
77ca6cdf 2745static inline int ext4_issue_discard(struct super_block *sb,
84130193 2746 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
5c521830 2747{
5c521830
JZ
2748 ext4_fsblk_t discard_block;
2749
84130193
TT
2750 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2751 ext4_group_first_block_no(sb, block_group));
2752 count = EXT4_C2B(EXT4_SB(sb), count);
5c521830
JZ
2753 trace_ext4_discard_blocks(sb,
2754 (unsigned long long) discard_block, count);
93259636 2755 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
5c521830
JZ
2756}
2757
3e624fc7
TT
2758/*
2759 * This function is called by the jbd2 layer once the commit has finished,
2760 * so we know we can free the blocks that were released with that commit.
2761 */
18aadd47
BJ
2762static void ext4_free_data_callback(struct super_block *sb,
2763 struct ext4_journal_cb_entry *jce,
2764 int rc)
c9de560d 2765{
18aadd47 2766 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
c9de560d 2767 struct ext4_buddy e4b;
c894058d 2768 struct ext4_group_info *db;
d9f34504 2769 int err, count = 0, count2 = 0;
c9de560d 2770
18aadd47
BJ
2771 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2772 entry->efd_count, entry->efd_group, entry);
c9de560d 2773
d71c1ae2
LC
2774 if (test_opt(sb, DISCARD)) {
2775 err = ext4_issue_discard(sb, entry->efd_group,
2776 entry->efd_start_cluster,
2777 entry->efd_count);
2778 if (err && err != -EOPNOTSUPP)
2779 ext4_msg(sb, KERN_WARNING, "discard request in"
2780 " group:%d block:%d count:%d failed"
2781 " with %d", entry->efd_group,
2782 entry->efd_start_cluster,
2783 entry->efd_count, err);
2784 }
c9de560d 2785
18aadd47
BJ
2786 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2787 /* we expect to find existing buddy because it's pinned */
2788 BUG_ON(err != 0);
b90f6870 2789
c9de560d 2790
18aadd47
BJ
2791 db = e4b.bd_info;
2792 /* there are blocks to put in buddy to make them really free */
2793 count += entry->efd_count;
2794 count2++;
2795 ext4_lock_group(sb, entry->efd_group);
2796 /* Take it out of per group rb tree */
2797 rb_erase(&entry->efd_node, &(db->bb_free_root));
2798 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
c894058d 2799
18aadd47
BJ
2800 /*
2801 * Clear the trimmed flag for the group so that the next
2802 * ext4_trim_fs can trim it.
2803 * If the volume is mounted with -o discard, online discard
2804 * is supported and the free blocks will be trimmed online.
2805 */
2806 if (!test_opt(sb, DISCARD))
2807 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3d56b8d2 2808
18aadd47
BJ
2809 if (!db->bb_free_root.rb_node) {
2810 /* No more items in the per group rb tree
2811 * balance refcounts from ext4_mb_free_metadata()
2812 */
2813 page_cache_release(e4b.bd_buddy_page);
2814 page_cache_release(e4b.bd_bitmap_page);
3e624fc7 2815 }
18aadd47
BJ
2816 ext4_unlock_group(sb, entry->efd_group);
2817 kmem_cache_free(ext4_free_data_cachep, entry);
2818 ext4_mb_unload_buddy(&e4b);
c9de560d 2819
6ba495e9 2820 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
c9de560d
AT
2821}
2822
5dabfc78 2823int __init ext4_init_mballoc(void)
c9de560d 2824{
16828088
TT
2825 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2826 SLAB_RECLAIM_ACCOUNT);
c9de560d
AT
2827 if (ext4_pspace_cachep == NULL)
2828 return -ENOMEM;
2829
16828088
TT
2830 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2831 SLAB_RECLAIM_ACCOUNT);
256bdb49
ES
2832 if (ext4_ac_cachep == NULL) {
2833 kmem_cache_destroy(ext4_pspace_cachep);
2834 return -ENOMEM;
2835 }
c894058d 2836
18aadd47
BJ
2837 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2838 SLAB_RECLAIM_ACCOUNT);
2839 if (ext4_free_data_cachep == NULL) {
c894058d
AK
2840 kmem_cache_destroy(ext4_pspace_cachep);
2841 kmem_cache_destroy(ext4_ac_cachep);
2842 return -ENOMEM;
2843 }
c9de560d
AT
2844 return 0;
2845}
2846
5dabfc78 2847void ext4_exit_mballoc(void)
c9de560d 2848{
60e6679e 2849 /*
3e03f9ca
JDB
2850 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2851 * before destroying the slab cache.
2852 */
2853 rcu_barrier();
c9de560d 2854 kmem_cache_destroy(ext4_pspace_cachep);
256bdb49 2855 kmem_cache_destroy(ext4_ac_cachep);
18aadd47 2856 kmem_cache_destroy(ext4_free_data_cachep);
2892c15d 2857 ext4_groupinfo_destroy_slabs();
c9de560d
AT
2858}
2859
2860
2861/*
73b2c716 2862 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
c9de560d
AT
2863 * Returns 0 if success or error code
2864 */
4ddfef7b
ES
2865static noinline_for_stack int
2866ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
53accfa9 2867 handle_t *handle, unsigned int reserv_clstrs)
c9de560d
AT
2868{
2869 struct buffer_head *bitmap_bh = NULL;
c9de560d
AT
2870 struct ext4_group_desc *gdp;
2871 struct buffer_head *gdp_bh;
2872 struct ext4_sb_info *sbi;
2873 struct super_block *sb;
2874 ext4_fsblk_t block;
519deca0 2875 int err, len;
c9de560d
AT
2876
2877 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2878 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2879
2880 sb = ac->ac_sb;
2881 sbi = EXT4_SB(sb);
c9de560d
AT
2882
2883 err = -EIO;
574ca174 2884 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
c9de560d
AT
2885 if (!bitmap_bh)
2886 goto out_err;
2887
5d601255 2888 BUFFER_TRACE(bitmap_bh, "getting write access");
c9de560d
AT
2889 err = ext4_journal_get_write_access(handle, bitmap_bh);
2890 if (err)
2891 goto out_err;
2892
2893 err = -EIO;
2894 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2895 if (!gdp)
2896 goto out_err;
2897
a9df9a49 2898 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
021b65bb 2899 ext4_free_group_clusters(sb, gdp));
03cddb80 2900
5d601255 2901 BUFFER_TRACE(gdp_bh, "get_write_access");
c9de560d
AT
2902 err = ext4_journal_get_write_access(handle, gdp_bh);
2903 if (err)
2904 goto out_err;
2905
bda00de7 2906 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
c9de560d 2907
53accfa9 2908 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
6fd058f7 2909 if (!ext4_data_block_valid(sbi, block, len)) {
12062ddd 2910 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
1084f252 2911 "fs metadata", block, block+len);
519deca0
AK
2912 /* File system mounted not to panic on error
2913 * Fix the bitmap and repeat the block allocation
2914 * We leak some of the blocks here.
2915 */
955ce5f5 2916 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
c3e94d1d
YY
2917 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2918 ac->ac_b_ex.fe_len);
955ce5f5 2919 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
0390131b 2920 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
519deca0
AK
2921 if (!err)
2922 err = -EAGAIN;
2923 goto out_err;
c9de560d 2924 }
955ce5f5
AK
2925
2926 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
c9de560d
AT
2927#ifdef AGGRESSIVE_CHECK
2928 {
2929 int i;
2930 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2931 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2932 bitmap_bh->b_data));
2933 }
2934 }
2935#endif
c3e94d1d
YY
2936 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2937 ac->ac_b_ex.fe_len);
c9de560d
AT
2938 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2939 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
021b65bb 2940 ext4_free_group_clusters_set(sb, gdp,
cff1dfd7 2941 ext4_free_clusters_after_init(sb,
021b65bb 2942 ac->ac_b_ex.fe_group, gdp));
c9de560d 2943 }
021b65bb
TT
2944 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2945 ext4_free_group_clusters_set(sb, gdp, len);
79f1ba49 2946 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
feb0ab32 2947 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
955ce5f5
AK
2948
2949 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
57042651 2950 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
d2a17637 2951 /*
6bc6e63f 2952 * Now reduce the dirty block count also. Should not go negative
d2a17637 2953 */
6bc6e63f
AK
2954 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2955 /* release all the reserved blocks if non delalloc */
57042651
TT
2956 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2957 reserv_clstrs);
c9de560d 2958
772cb7c8
JS
2959 if (sbi->s_log_groups_per_flex) {
2960 ext4_group_t flex_group = ext4_flex_group(sbi,
2961 ac->ac_b_ex.fe_group);
90ba983f
TT
2962 atomic64_sub(ac->ac_b_ex.fe_len,
2963 &sbi->s_flex_groups[flex_group].free_clusters);
772cb7c8
JS
2964 }
2965
0390131b 2966 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
c9de560d
AT
2967 if (err)
2968 goto out_err;
0390131b 2969 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
c9de560d
AT
2970
2971out_err:
42a10add 2972 brelse(bitmap_bh);
c9de560d
AT
2973 return err;
2974}
2975
2976/*
2977 * here we normalize request for locality group
d7a1fee1
DE
2978 * Group request are normalized to s_mb_group_prealloc, which goes to
2979 * s_strip if we set the same via mount option.
2980 * s_mb_group_prealloc can be configured via
b713a5ec 2981 * /sys/fs/ext4/<partition>/mb_group_prealloc
c9de560d
AT
2982 *
2983 * XXX: should we try to preallocate more than the group has now?
2984 */
2985static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2986{
2987 struct super_block *sb = ac->ac_sb;
2988 struct ext4_locality_group *lg = ac->ac_lg;
2989
2990 BUG_ON(lg == NULL);
d7a1fee1 2991 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
6ba495e9 2992 mb_debug(1, "#%u: goal %u blocks for locality group\n",
c9de560d
AT
2993 current->pid, ac->ac_g_ex.fe_len);
2994}
2995
2996/*
2997 * Normalization means making request better in terms of
2998 * size and alignment
2999 */
4ddfef7b
ES
3000static noinline_for_stack void
3001ext4_mb_normalize_request(struct ext4_allocation_context *ac,
c9de560d
AT
3002 struct ext4_allocation_request *ar)
3003{
53accfa9 3004 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
c9de560d
AT
3005 int bsbits, max;
3006 ext4_lblk_t end;
1592d2c5
CW
3007 loff_t size, start_off;
3008 loff_t orig_size __maybe_unused;
5a0790c2 3009 ext4_lblk_t start;
c9de560d 3010 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
9a0762c5 3011 struct ext4_prealloc_space *pa;
c9de560d
AT
3012
3013 /* do normalize only data requests, metadata requests
3014 do not need preallocation */
3015 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3016 return;
3017
3018 /* sometime caller may want exact blocks */
3019 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3020 return;
3021
3022 /* caller may indicate that preallocation isn't
3023 * required (it's a tail, for example) */
3024 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3025 return;
3026
3027 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3028 ext4_mb_normalize_group_request(ac);
3029 return ;
3030 }
3031
3032 bsbits = ac->ac_sb->s_blocksize_bits;
3033
3034 /* first, let's learn actual file size
3035 * given current request is allocated */
53accfa9 3036 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
c9de560d
AT
3037 size = size << bsbits;
3038 if (size < i_size_read(ac->ac_inode))
3039 size = i_size_read(ac->ac_inode);
5a0790c2 3040 orig_size = size;
c9de560d 3041
1930479c
VC
3042 /* max size of free chunks */
3043 max = 2 << bsbits;
c9de560d 3044
1930479c
VC
3045#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3046 (req <= (size) || max <= (chunk_size))
c9de560d
AT
3047
3048 /* first, try to predict filesize */
3049 /* XXX: should this table be tunable? */
3050 start_off = 0;
3051 if (size <= 16 * 1024) {
3052 size = 16 * 1024;
3053 } else if (size <= 32 * 1024) {
3054 size = 32 * 1024;
3055 } else if (size <= 64 * 1024) {
3056 size = 64 * 1024;
3057 } else if (size <= 128 * 1024) {
3058 size = 128 * 1024;
3059 } else if (size <= 256 * 1024) {
3060 size = 256 * 1024;
3061 } else if (size <= 512 * 1024) {
3062 size = 512 * 1024;
3063 } else if (size <= 1024 * 1024) {
3064 size = 1024 * 1024;
1930479c 3065 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
c9de560d 3066 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
1930479c
VC
3067 (21 - bsbits)) << 21;
3068 size = 2 * 1024 * 1024;
3069 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
c9de560d
AT
3070 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3071 (22 - bsbits)) << 22;
3072 size = 4 * 1024 * 1024;
3073 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
1930479c 3074 (8<<20)>>bsbits, max, 8 * 1024)) {
c9de560d
AT
3075 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3076 (23 - bsbits)) << 23;
3077 size = 8 * 1024 * 1024;
3078 } else {
b27b1535
XW
3079 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3080 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3081 ac->ac_o_ex.fe_len) << bsbits;
c9de560d 3082 }
5a0790c2
AK
3083 size = size >> bsbits;
3084 start = start_off >> bsbits;
c9de560d
AT
3085
3086 /* don't cover already allocated blocks in selected range */
3087 if (ar->pleft && start <= ar->lleft) {
3088 size -= ar->lleft + 1 - start;
3089 start = ar->lleft + 1;
3090 }
3091 if (ar->pright && start + size - 1 >= ar->lright)
3092 size -= start + size - ar->lright;
3093
3094 end = start + size;
3095
3096 /* check we don't cross already preallocated blocks */
3097 rcu_read_lock();
9a0762c5 3098 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
498e5f24 3099 ext4_lblk_t pa_end;
c9de560d 3100
c9de560d
AT
3101 if (pa->pa_deleted)
3102 continue;
3103 spin_lock(&pa->pa_lock);
3104 if (pa->pa_deleted) {
3105 spin_unlock(&pa->pa_lock);
3106 continue;
3107 }
3108
53accfa9
TT
3109 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3110 pa->pa_len);
c9de560d
AT
3111
3112 /* PA must not overlap original request */
3113 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3114 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3115
38877f4e
ES
3116 /* skip PAs this normalized request doesn't overlap with */
3117 if (pa->pa_lstart >= end || pa_end <= start) {
c9de560d
AT
3118 spin_unlock(&pa->pa_lock);
3119 continue;
3120 }
3121 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3122
38877f4e 3123 /* adjust start or end to be adjacent to this pa */
c9de560d
AT
3124 if (pa_end <= ac->ac_o_ex.fe_logical) {
3125 BUG_ON(pa_end < start);
3126 start = pa_end;
38877f4e 3127 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
c9de560d
AT
3128 BUG_ON(pa->pa_lstart > end);
3129 end = pa->pa_lstart;
3130 }
3131 spin_unlock(&pa->pa_lock);
3132 }
3133 rcu_read_unlock();
3134 size = end - start;
3135
3136 /* XXX: extra loop to check we really don't overlap preallocations */
3137 rcu_read_lock();
9a0762c5 3138 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
498e5f24 3139 ext4_lblk_t pa_end;
53accfa9 3140
c9de560d
AT
3141 spin_lock(&pa->pa_lock);
3142 if (pa->pa_deleted == 0) {
53accfa9
TT
3143 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3144 pa->pa_len);
c9de560d
AT
3145 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3146 }
3147 spin_unlock(&pa->pa_lock);
3148 }
3149 rcu_read_unlock();
3150
3151 if (start + size <= ac->ac_o_ex.fe_logical &&
3152 start > ac->ac_o_ex.fe_logical) {
9d8b9ec4
TT
3153 ext4_msg(ac->ac_sb, KERN_ERR,
3154 "start %lu, size %lu, fe_logical %lu",
3155 (unsigned long) start, (unsigned long) size,
3156 (unsigned long) ac->ac_o_ex.fe_logical);
dfe076c1 3157 BUG();
c9de560d 3158 }
b5b60778 3159 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
c9de560d
AT
3160
3161 /* now prepare goal request */
3162
3163 /* XXX: is it better to align blocks WRT to logical
3164 * placement or satisfy big request as is */
3165 ac->ac_g_ex.fe_logical = start;
53accfa9 3166 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
c9de560d
AT
3167
3168 /* define goal start in order to merge */
3169 if (ar->pright && (ar->lright == (start + size))) {
3170 /* merge to the right */
3171 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3172 &ac->ac_f_ex.fe_group,
3173 &ac->ac_f_ex.fe_start);
3174 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3175 }
3176 if (ar->pleft && (ar->lleft + 1 == start)) {
3177 /* merge to the left */
3178 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3179 &ac->ac_f_ex.fe_group,
3180 &ac->ac_f_ex.fe_start);
3181 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3182 }
3183
6ba495e9 3184 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
c9de560d
AT
3185 (unsigned) orig_size, (unsigned) start);
3186}
3187
3188static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3189{
3190 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3191
3192 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3193 atomic_inc(&sbi->s_bal_reqs);
3194 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
291dae47 3195 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
c9de560d
AT
3196 atomic_inc(&sbi->s_bal_success);
3197 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3198 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3199 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3200 atomic_inc(&sbi->s_bal_goals);
3201 if (ac->ac_found > sbi->s_mb_max_to_scan)
3202 atomic_inc(&sbi->s_bal_breaks);
3203 }
3204
296c355c
TT
3205 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3206 trace_ext4_mballoc_alloc(ac);
3207 else
3208 trace_ext4_mballoc_prealloc(ac);
c9de560d
AT
3209}
3210
b844167e
CW
3211/*
3212 * Called on failure; free up any blocks from the inode PA for this
3213 * context. We don't need this for MB_GROUP_PA because we only change
3214 * pa_free in ext4_mb_release_context(), but on failure, we've already
3215 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3216 */
3217static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3218{
3219 struct ext4_prealloc_space *pa = ac->ac_pa;
86f0afd4
TT
3220 struct ext4_buddy e4b;
3221 int err;
b844167e 3222
86f0afd4 3223 if (pa == NULL) {
c99d1e6e
TT
3224 if (ac->ac_f_ex.fe_len == 0)
3225 return;
86f0afd4
TT
3226 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3227 if (err) {
3228 /*
3229 * This should never happen since we pin the
3230 * pages in the ext4_allocation_context so
3231 * ext4_mb_load_buddy() should never fail.
3232 */
3233 WARN(1, "mb_load_buddy failed (%d)", err);
3234 return;
3235 }
3236 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3237 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3238 ac->ac_f_ex.fe_len);
3239 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
c99d1e6e 3240 ext4_mb_unload_buddy(&e4b);
86f0afd4
TT
3241 return;
3242 }
3243 if (pa->pa_type == MB_INODE_PA)
400db9d3 3244 pa->pa_free += ac->ac_b_ex.fe_len;
b844167e
CW
3245}
3246
c9de560d
AT
3247/*
3248 * use blocks preallocated to inode
3249 */
3250static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3251 struct ext4_prealloc_space *pa)
3252{
53accfa9 3253 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
c9de560d
AT
3254 ext4_fsblk_t start;
3255 ext4_fsblk_t end;
3256 int len;
3257
3258 /* found preallocated blocks, use them */
3259 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
53accfa9
TT
3260 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3261 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3262 len = EXT4_NUM_B2C(sbi, end - start);
c9de560d
AT
3263 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3264 &ac->ac_b_ex.fe_start);
3265 ac->ac_b_ex.fe_len = len;
3266 ac->ac_status = AC_STATUS_FOUND;
3267 ac->ac_pa = pa;
3268
3269 BUG_ON(start < pa->pa_pstart);
53accfa9 3270 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
c9de560d
AT
3271 BUG_ON(pa->pa_free < len);
3272 pa->pa_free -= len;
3273
6ba495e9 3274 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
c9de560d
AT
3275}
3276
3277/*
3278 * use blocks preallocated to locality group
3279 */
3280static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3281 struct ext4_prealloc_space *pa)
3282{
03cddb80 3283 unsigned int len = ac->ac_o_ex.fe_len;
6be2ded1 3284
c9de560d
AT
3285 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3286 &ac->ac_b_ex.fe_group,
3287 &ac->ac_b_ex.fe_start);
3288 ac->ac_b_ex.fe_len = len;
3289 ac->ac_status = AC_STATUS_FOUND;
3290 ac->ac_pa = pa;
3291
3292 /* we don't correct pa_pstart or pa_plen here to avoid
26346ff6 3293 * possible race when the group is being loaded concurrently
c9de560d 3294 * instead we correct pa later, after blocks are marked
26346ff6
AK
3295 * in on-disk bitmap -- see ext4_mb_release_context()
3296 * Other CPUs are prevented from allocating from this pa by lg_mutex
c9de560d 3297 */
6ba495e9 3298 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
c9de560d
AT
3299}
3300
5e745b04
AK
3301/*
3302 * Return the prealloc space that have minimal distance
3303 * from the goal block. @cpa is the prealloc
3304 * space that is having currently known minimal distance
3305 * from the goal block.
3306 */
3307static struct ext4_prealloc_space *
3308ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3309 struct ext4_prealloc_space *pa,
3310 struct ext4_prealloc_space *cpa)
3311{
3312 ext4_fsblk_t cur_distance, new_distance;
3313
3314 if (cpa == NULL) {
3315 atomic_inc(&pa->pa_count);
3316 return pa;
3317 }
3318 cur_distance = abs(goal_block - cpa->pa_pstart);
3319 new_distance = abs(goal_block - pa->pa_pstart);
3320
5a54b2f1 3321 if (cur_distance <= new_distance)
5e745b04
AK
3322 return cpa;
3323
3324 /* drop the previous reference */
3325 atomic_dec(&cpa->pa_count);
3326 atomic_inc(&pa->pa_count);
3327 return pa;
3328}
3329
c9de560d
AT
3330/*
3331 * search goal blocks in preallocated space
3332 */
4ddfef7b
ES
3333static noinline_for_stack int
3334ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
c9de560d 3335{
53accfa9 3336 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
6be2ded1 3337 int order, i;
c9de560d
AT
3338 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3339 struct ext4_locality_group *lg;
5e745b04
AK
3340 struct ext4_prealloc_space *pa, *cpa = NULL;
3341 ext4_fsblk_t goal_block;
c9de560d
AT
3342
3343 /* only data can be preallocated */
3344 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3345 return 0;
3346
3347 /* first, try per-file preallocation */
3348 rcu_read_lock();
9a0762c5 3349 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
c9de560d
AT
3350
3351 /* all fields in this condition don't change,
3352 * so we can skip locking for them */
3353 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
53accfa9
TT
3354 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3355 EXT4_C2B(sbi, pa->pa_len)))
c9de560d
AT
3356 continue;
3357
fb0a387d 3358 /* non-extent files can't have physical blocks past 2^32 */
12e9b892 3359 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
53accfa9
TT
3360 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3361 EXT4_MAX_BLOCK_FILE_PHYS))
fb0a387d
ES
3362 continue;
3363
c9de560d
AT
3364 /* found preallocated blocks, use them */
3365 spin_lock(&pa->pa_lock);
3366 if (pa->pa_deleted == 0 && pa->pa_free) {
3367 atomic_inc(&pa->pa_count);
3368 ext4_mb_use_inode_pa(ac, pa);
3369 spin_unlock(&pa->pa_lock);
3370 ac->ac_criteria = 10;
3371 rcu_read_unlock();
3372 return 1;
3373 }
3374 spin_unlock(&pa->pa_lock);
3375 }
3376 rcu_read_unlock();
3377
3378 /* can we use group allocation? */
3379 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3380 return 0;
3381
3382 /* inode may have no locality group for some reason */
3383 lg = ac->ac_lg;
3384 if (lg == NULL)
3385 return 0;
6be2ded1
AK
3386 order = fls(ac->ac_o_ex.fe_len) - 1;
3387 if (order > PREALLOC_TB_SIZE - 1)
3388 /* The max size of hash table is PREALLOC_TB_SIZE */
3389 order = PREALLOC_TB_SIZE - 1;
3390
bda00de7 3391 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
5e745b04
AK
3392 /*
3393 * search for the prealloc space that is having
3394 * minimal distance from the goal block.
3395 */
6be2ded1
AK
3396 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3397 rcu_read_lock();
3398 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3399 pa_inode_list) {
3400 spin_lock(&pa->pa_lock);
3401 if (pa->pa_deleted == 0 &&
3402 pa->pa_free >= ac->ac_o_ex.fe_len) {
5e745b04
AK
3403
3404 cpa = ext4_mb_check_group_pa(goal_block,
3405 pa, cpa);
6be2ded1 3406 }
c9de560d 3407 spin_unlock(&pa->pa_lock);
c9de560d 3408 }
6be2ded1 3409 rcu_read_unlock();
c9de560d 3410 }
5e745b04
AK
3411 if (cpa) {
3412 ext4_mb_use_group_pa(ac, cpa);
3413 ac->ac_criteria = 20;
3414 return 1;
3415 }
c9de560d
AT
3416 return 0;
3417}
3418
7a2fcbf7
AK
3419/*
3420 * the function goes through all block freed in the group
3421 * but not yet committed and marks them used in in-core bitmap.
3422 * buddy must be generated from this bitmap
955ce5f5 3423 * Need to be called with the ext4 group lock held
7a2fcbf7
AK
3424 */
3425static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3426 ext4_group_t group)
3427{
3428 struct rb_node *n;
3429 struct ext4_group_info *grp;
3430 struct ext4_free_data *entry;
3431
3432 grp = ext4_get_group_info(sb, group);
3433 n = rb_first(&(grp->bb_free_root));
3434
3435 while (n) {
18aadd47
BJ
3436 entry = rb_entry(n, struct ext4_free_data, efd_node);
3437 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
7a2fcbf7
AK
3438 n = rb_next(n);
3439 }
3440 return;
3441}
3442
c9de560d
AT
3443/*
3444 * the function goes through all preallocation in this group and marks them
3445 * used in in-core bitmap. buddy must be generated from this bitmap
955ce5f5 3446 * Need to be called with ext4 group lock held
c9de560d 3447 */
089ceecc
ES
3448static noinline_for_stack
3449void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
c9de560d
AT
3450 ext4_group_t group)
3451{
3452 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3453 struct ext4_prealloc_space *pa;
3454 struct list_head *cur;
3455 ext4_group_t groupnr;
3456 ext4_grpblk_t start;
3457 int preallocated = 0;
c9de560d
AT
3458 int len;
3459
3460 /* all form of preallocation discards first load group,
3461 * so the only competing code is preallocation use.
3462 * we don't need any locking here
3463 * notice we do NOT ignore preallocations with pa_deleted
3464 * otherwise we could leave used blocks available for
3465 * allocation in buddy when concurrent ext4_mb_put_pa()
3466 * is dropping preallocation
3467 */
3468 list_for_each(cur, &grp->bb_prealloc_list) {
3469 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3470 spin_lock(&pa->pa_lock);
3471 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3472 &groupnr, &start);
3473 len = pa->pa_len;
3474 spin_unlock(&pa->pa_lock);
3475 if (unlikely(len == 0))
3476 continue;
3477 BUG_ON(groupnr != group);
c3e94d1d 3478 ext4_set_bits(bitmap, start, len);
c9de560d 3479 preallocated += len;
c9de560d 3480 }
6ba495e9 3481 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
c9de560d
AT
3482}
3483
3484static void ext4_mb_pa_callback(struct rcu_head *head)
3485{
3486 struct ext4_prealloc_space *pa;
3487 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4e8d2139
JR
3488
3489 BUG_ON(atomic_read(&pa->pa_count));
3490 BUG_ON(pa->pa_deleted == 0);
c9de560d
AT
3491 kmem_cache_free(ext4_pspace_cachep, pa);
3492}
3493
3494/*
3495 * drops a reference to preallocated space descriptor
3496 * if this was the last reference and the space is consumed
3497 */
3498static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3499 struct super_block *sb, struct ext4_prealloc_space *pa)
3500{
a9df9a49 3501 ext4_group_t grp;
d33a1976 3502 ext4_fsblk_t grp_blk;
c9de560d 3503
c9de560d
AT
3504 /* in this short window concurrent discard can set pa_deleted */
3505 spin_lock(&pa->pa_lock);
4e8d2139
JR
3506 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3507 spin_unlock(&pa->pa_lock);
3508 return;
3509 }
3510
c9de560d
AT
3511 if (pa->pa_deleted == 1) {
3512 spin_unlock(&pa->pa_lock);
3513 return;
3514 }
3515
3516 pa->pa_deleted = 1;
3517 spin_unlock(&pa->pa_lock);
3518
d33a1976 3519 grp_blk = pa->pa_pstart;
60e6679e 3520 /*
cc0fb9ad
AK
3521 * If doing group-based preallocation, pa_pstart may be in the
3522 * next group when pa is used up
3523 */
3524 if (pa->pa_type == MB_GROUP_PA)
d33a1976
ES
3525 grp_blk--;
3526
bd86298e 3527 grp = ext4_get_group_number(sb, grp_blk);
c9de560d
AT
3528
3529 /*
3530 * possible race:
3531 *
3532 * P1 (buddy init) P2 (regular allocation)
3533 * find block B in PA
3534 * copy on-disk bitmap to buddy
3535 * mark B in on-disk bitmap
3536 * drop PA from group
3537 * mark all PAs in buddy
3538 *
3539 * thus, P1 initializes buddy with B available. to prevent this
3540 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3541 * against that pair
3542 */
3543 ext4_lock_group(sb, grp);
3544 list_del(&pa->pa_group_list);
3545 ext4_unlock_group(sb, grp);
3546
3547 spin_lock(pa->pa_obj_lock);
3548 list_del_rcu(&pa->pa_inode_list);
3549 spin_unlock(pa->pa_obj_lock);
3550
3551 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3552}
3553
3554/*
3555 * creates new preallocated space for given inode
3556 */
4ddfef7b
ES
3557static noinline_for_stack int
3558ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
c9de560d
AT
3559{
3560 struct super_block *sb = ac->ac_sb;
53accfa9 3561 struct ext4_sb_info *sbi = EXT4_SB(sb);
c9de560d
AT
3562 struct ext4_prealloc_space *pa;
3563 struct ext4_group_info *grp;
3564 struct ext4_inode_info *ei;
3565
3566 /* preallocate only when found space is larger then requested */
3567 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3568 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3569 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3570
3571 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3572 if (pa == NULL)
3573 return -ENOMEM;
3574
3575 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3576 int winl;
3577 int wins;
3578 int win;
3579 int offs;
3580
3581 /* we can't allocate as much as normalizer wants.
3582 * so, found space must get proper lstart
3583 * to cover original request */
3584 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3585 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3586
3587 /* we're limited by original request in that
3588 * logical block must be covered any way
3589 * winl is window we can move our chunk within */
3590 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3591
3592 /* also, we should cover whole original request */
53accfa9 3593 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
c9de560d
AT
3594
3595 /* the smallest one defines real window */
3596 win = min(winl, wins);
3597
53accfa9
TT
3598 offs = ac->ac_o_ex.fe_logical %
3599 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
c9de560d
AT
3600 if (offs && offs < win)
3601 win = offs;
3602
53accfa9 3603 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
810da240 3604 EXT4_NUM_B2C(sbi, win);
c9de560d
AT
3605 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3606 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3607 }
3608
3609 /* preallocation can change ac_b_ex, thus we store actually
3610 * allocated blocks for history */
3611 ac->ac_f_ex = ac->ac_b_ex;
3612
3613 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3614 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3615 pa->pa_len = ac->ac_b_ex.fe_len;
3616 pa->pa_free = pa->pa_len;
3617 atomic_set(&pa->pa_count, 1);
3618 spin_lock_init(&pa->pa_lock);
d794bf8e
AK
3619 INIT_LIST_HEAD(&pa->pa_inode_list);
3620 INIT_LIST_HEAD(&pa->pa_group_list);
c9de560d 3621 pa->pa_deleted = 0;
cc0fb9ad 3622 pa->pa_type = MB_INODE_PA;
c9de560d 3623
6ba495e9 3624 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
c9de560d 3625 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
9bffad1e 3626 trace_ext4_mb_new_inode_pa(ac, pa);
c9de560d
AT
3627
3628 ext4_mb_use_inode_pa(ac, pa);
53accfa9 3629 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
c9de560d
AT
3630
3631 ei = EXT4_I(ac->ac_inode);
3632 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3633
3634 pa->pa_obj_lock = &ei->i_prealloc_lock;
3635 pa->pa_inode = ac->ac_inode;
3636
3637 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3638 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3639 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3640
3641 spin_lock(pa->pa_obj_lock);
3642 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3643 spin_unlock(pa->pa_obj_lock);
3644
3645 return 0;
3646}
3647
3648/*
3649 * creates new preallocated space for locality group inodes belongs to
3650 */
4ddfef7b
ES
3651static noinline_for_stack int
3652ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
c9de560d
AT
3653{
3654 struct super_block *sb = ac->ac_sb;
3655 struct ext4_locality_group *lg;
3656 struct ext4_prealloc_space *pa;
3657 struct ext4_group_info *grp;
3658
3659 /* preallocate only when found space is larger then requested */
3660 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3661 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3662 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3663
3664 BUG_ON(ext4_pspace_cachep == NULL);
3665 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3666 if (pa == NULL)
3667 return -ENOMEM;
3668
3669 /* preallocation can change ac_b_ex, thus we store actually
3670 * allocated blocks for history */
3671 ac->ac_f_ex = ac->ac_b_ex;
3672
3673 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3674 pa->pa_lstart = pa->pa_pstart;
3675 pa->pa_len = ac->ac_b_ex.fe_len;
3676 pa->pa_free = pa->pa_len;
3677 atomic_set(&pa->pa_count, 1);
3678 spin_lock_init(&pa->pa_lock);
6be2ded1 3679 INIT_LIST_HEAD(&pa->pa_inode_list);
d794bf8e 3680 INIT_LIST_HEAD(&pa->pa_group_list);
c9de560d 3681 pa->pa_deleted = 0;
cc0fb9ad 3682 pa->pa_type = MB_GROUP_PA;
c9de560d 3683
6ba495e9 3684 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
9bffad1e
TT
3685 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3686 trace_ext4_mb_new_group_pa(ac, pa);
c9de560d
AT
3687
3688 ext4_mb_use_group_pa(ac, pa);
3689 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3690
3691 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3692 lg = ac->ac_lg;
3693 BUG_ON(lg == NULL);
3694
3695 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3696 pa->pa_inode = NULL;
3697
3698 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3699 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3700 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3701
6be2ded1
AK
3702 /*
3703 * We will later add the new pa to the right bucket
3704 * after updating the pa_free in ext4_mb_release_context
3705 */
c9de560d
AT
3706 return 0;
3707}
3708
3709static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3710{
3711 int err;
3712
3713 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3714 err = ext4_mb_new_group_pa(ac);
3715 else
3716 err = ext4_mb_new_inode_pa(ac);
3717 return err;
3718}
3719
3720/*
3721 * finds all unused blocks in on-disk bitmap, frees them in
3722 * in-core bitmap and buddy.
3723 * @pa must be unlinked from inode and group lists, so that
3724 * nobody else can find/use it.
3725 * the caller MUST hold group/inode locks.
3726 * TODO: optimize the case when there are no in-core structures yet
3727 */
4ddfef7b
ES
3728static noinline_for_stack int
3729ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3e1e5f50 3730 struct ext4_prealloc_space *pa)
c9de560d 3731{
c9de560d
AT
3732 struct super_block *sb = e4b->bd_sb;
3733 struct ext4_sb_info *sbi = EXT4_SB(sb);
498e5f24
TT
3734 unsigned int end;
3735 unsigned int next;
c9de560d
AT
3736 ext4_group_t group;
3737 ext4_grpblk_t bit;
ba80b101 3738 unsigned long long grp_blk_start;
c9de560d
AT
3739 int err = 0;
3740 int free = 0;
3741
3742 BUG_ON(pa->pa_deleted == 0);
3743 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
53accfa9 3744 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
c9de560d
AT
3745 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3746 end = bit + pa->pa_len;
3747
c9de560d 3748 while (bit < end) {
ffad0a44 3749 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
c9de560d
AT
3750 if (bit >= end)
3751 break;
ffad0a44 3752 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
6ba495e9 3753 mb_debug(1, " free preallocated %u/%u in group %u\n",
5a0790c2
AK
3754 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3755 (unsigned) next - bit, (unsigned) group);
c9de560d
AT
3756 free += next - bit;
3757
3e1e5f50 3758 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
53accfa9
TT
3759 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3760 EXT4_C2B(sbi, bit)),
a9c667f8 3761 next - bit);
c9de560d
AT
3762 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3763 bit = next + 1;
3764 }
3765 if (free != pa->pa_free) {
9d8b9ec4
TT
3766 ext4_msg(e4b->bd_sb, KERN_CRIT,
3767 "pa %p: logic %lu, phys. %lu, len %lu",
3768 pa, (unsigned long) pa->pa_lstart,
3769 (unsigned long) pa->pa_pstart,
3770 (unsigned long) pa->pa_len);
e29136f8 3771 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
5d1b1b3f 3772 free, pa->pa_free);
e56eb659
AK
3773 /*
3774 * pa is already deleted so we use the value obtained
3775 * from the bitmap and continue.
3776 */
c9de560d 3777 }
c9de560d
AT
3778 atomic_add(free, &sbi->s_mb_discarded);
3779
3780 return err;
3781}
3782
4ddfef7b
ES
3783static noinline_for_stack int
3784ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3e1e5f50 3785 struct ext4_prealloc_space *pa)
c9de560d 3786{
c9de560d
AT
3787 struct super_block *sb = e4b->bd_sb;
3788 ext4_group_t group;
3789 ext4_grpblk_t bit;
3790
60e07cf5 3791 trace_ext4_mb_release_group_pa(sb, pa);
c9de560d
AT
3792 BUG_ON(pa->pa_deleted == 0);
3793 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3794 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3795 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3796 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3e1e5f50 3797 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
c9de560d
AT
3798
3799 return 0;
3800}
3801
3802/*
3803 * releases all preallocations in given group
3804 *
3805 * first, we need to decide discard policy:
3806 * - when do we discard
3807 * 1) ENOSPC
3808 * - how many do we discard
3809 * 1) how many requested
3810 */
4ddfef7b
ES
3811static noinline_for_stack int
3812ext4_mb_discard_group_preallocations(struct super_block *sb,
c9de560d
AT
3813 ext4_group_t group, int needed)
3814{
3815 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3816 struct buffer_head *bitmap_bh = NULL;
3817 struct ext4_prealloc_space *pa, *tmp;
3818 struct list_head list;
3819 struct ext4_buddy e4b;
3820 int err;
3821 int busy = 0;
3822 int free = 0;
3823
6ba495e9 3824 mb_debug(1, "discard preallocation for group %u\n", group);
c9de560d
AT
3825
3826 if (list_empty(&grp->bb_prealloc_list))
3827 return 0;
3828
574ca174 3829 bitmap_bh = ext4_read_block_bitmap(sb, group);
c9de560d 3830 if (bitmap_bh == NULL) {
12062ddd 3831 ext4_error(sb, "Error reading block bitmap for %u", group);
ce89f46c 3832 return 0;
c9de560d
AT
3833 }
3834
3835 err = ext4_mb_load_buddy(sb, group, &e4b);
ce89f46c 3836 if (err) {
12062ddd 3837 ext4_error(sb, "Error loading buddy information for %u", group);
ce89f46c
AK
3838 put_bh(bitmap_bh);
3839 return 0;
3840 }
c9de560d
AT
3841
3842 if (needed == 0)
7137d7a4 3843 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
c9de560d 3844
c9de560d 3845 INIT_LIST_HEAD(&list);
c9de560d
AT
3846repeat:
3847 ext4_lock_group(sb, group);
3848 list_for_each_entry_safe(pa, tmp,
3849 &grp->bb_prealloc_list, pa_group_list) {
3850 spin_lock(&pa->pa_lock);
3851 if (atomic_read(&pa->pa_count)) {
3852 spin_unlock(&pa->pa_lock);
3853 busy = 1;
3854 continue;
3855 }
3856 if (pa->pa_deleted) {
3857 spin_unlock(&pa->pa_lock);
3858 continue;
3859 }
3860
3861 /* seems this one can be freed ... */
3862 pa->pa_deleted = 1;
3863
3864 /* we can trust pa_free ... */
3865 free += pa->pa_free;
3866
3867 spin_unlock(&pa->pa_lock);
3868
3869 list_del(&pa->pa_group_list);
3870 list_add(&pa->u.pa_tmp_list, &list);
3871 }
3872
3873 /* if we still need more blocks and some PAs were used, try again */
3874 if (free < needed && busy) {
3875 busy = 0;
3876 ext4_unlock_group(sb, group);
bb8b20ed 3877 cond_resched();
c9de560d
AT
3878 goto repeat;
3879 }
3880
3881 /* found anything to free? */
3882 if (list_empty(&list)) {
3883 BUG_ON(free != 0);
3884 goto out;
3885 }
3886
3887 /* now free all selected PAs */
3888 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3889
3890 /* remove from object (inode or locality group) */
3891 spin_lock(pa->pa_obj_lock);
3892 list_del_rcu(&pa->pa_inode_list);
3893 spin_unlock(pa->pa_obj_lock);
3894
cc0fb9ad 3895 if (pa->pa_type == MB_GROUP_PA)
3e1e5f50 3896 ext4_mb_release_group_pa(&e4b, pa);
c9de560d 3897 else
3e1e5f50 3898 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
c9de560d
AT
3899
3900 list_del(&pa->u.pa_tmp_list);
3901 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3902 }
3903
3904out:
3905 ext4_unlock_group(sb, group);
e39e07fd 3906 ext4_mb_unload_buddy(&e4b);
c9de560d
AT
3907 put_bh(bitmap_bh);
3908 return free;
3909}
3910
3911/*
3912 * releases all non-used preallocated blocks for given inode
3913 *
3914 * It's important to discard preallocations under i_data_sem
3915 * We don't want another block to be served from the prealloc
3916 * space when we are discarding the inode prealloc space.
3917 *
3918 * FIXME!! Make sure it is valid at all the call sites
3919 */
c2ea3fde 3920void ext4_discard_preallocations(struct inode *inode)
c9de560d
AT
3921{
3922 struct ext4_inode_info *ei = EXT4_I(inode);
3923 struct super_block *sb = inode->i_sb;
3924 struct buffer_head *bitmap_bh = NULL;
3925 struct ext4_prealloc_space *pa, *tmp;
3926 ext4_group_t group = 0;
3927 struct list_head list;
3928 struct ext4_buddy e4b;
3929 int err;
3930
c2ea3fde 3931 if (!S_ISREG(inode->i_mode)) {
c9de560d
AT
3932 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3933 return;
3934 }
3935
6ba495e9 3936 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
9bffad1e 3937 trace_ext4_discard_preallocations(inode);
c9de560d
AT
3938
3939 INIT_LIST_HEAD(&list);
3940
3941repeat:
3942 /* first, collect all pa's in the inode */
3943 spin_lock(&ei->i_prealloc_lock);
3944 while (!list_empty(&ei->i_prealloc_list)) {
3945 pa = list_entry(ei->i_prealloc_list.next,
3946 struct ext4_prealloc_space, pa_inode_list);
3947 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3948 spin_lock(&pa->pa_lock);
3949 if (atomic_read(&pa->pa_count)) {
3950 /* this shouldn't happen often - nobody should
3951 * use preallocation while we're discarding it */
3952 spin_unlock(&pa->pa_lock);
3953 spin_unlock(&ei->i_prealloc_lock);
9d8b9ec4
TT
3954 ext4_msg(sb, KERN_ERR,
3955 "uh-oh! used pa while discarding");
c9de560d
AT
3956 WARN_ON(1);
3957 schedule_timeout_uninterruptible(HZ);
3958 goto repeat;
3959
3960 }
3961 if (pa->pa_deleted == 0) {
3962 pa->pa_deleted = 1;
3963 spin_unlock(&pa->pa_lock);
3964 list_del_rcu(&pa->pa_inode_list);
3965 list_add(&pa->u.pa_tmp_list, &list);
3966 continue;
3967 }
3968
3969 /* someone is deleting pa right now */
3970 spin_unlock(&pa->pa_lock);
3971 spin_unlock(&ei->i_prealloc_lock);
3972
3973 /* we have to wait here because pa_deleted
3974 * doesn't mean pa is already unlinked from
3975 * the list. as we might be called from
3976 * ->clear_inode() the inode will get freed
3977 * and concurrent thread which is unlinking
3978 * pa from inode's list may access already
3979 * freed memory, bad-bad-bad */
3980
3981 /* XXX: if this happens too often, we can
3982 * add a flag to force wait only in case
3983 * of ->clear_inode(), but not in case of
3984 * regular truncate */
3985 schedule_timeout_uninterruptible(HZ);
3986 goto repeat;
3987 }
3988 spin_unlock(&ei->i_prealloc_lock);
3989
3990 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
cc0fb9ad 3991 BUG_ON(pa->pa_type != MB_INODE_PA);
bd86298e 3992 group = ext4_get_group_number(sb, pa->pa_pstart);
c9de560d
AT
3993
3994 err = ext4_mb_load_buddy(sb, group, &e4b);
ce89f46c 3995 if (err) {
12062ddd
ES
3996 ext4_error(sb, "Error loading buddy information for %u",
3997 group);
ce89f46c
AK
3998 continue;
3999 }
c9de560d 4000
574ca174 4001 bitmap_bh = ext4_read_block_bitmap(sb, group);
c9de560d 4002 if (bitmap_bh == NULL) {
12062ddd
ES
4003 ext4_error(sb, "Error reading block bitmap for %u",
4004 group);
e39e07fd 4005 ext4_mb_unload_buddy(&e4b);
ce89f46c 4006 continue;
c9de560d
AT
4007 }
4008
4009 ext4_lock_group(sb, group);
4010 list_del(&pa->pa_group_list);
3e1e5f50 4011 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
c9de560d
AT
4012 ext4_unlock_group(sb, group);
4013
e39e07fd 4014 ext4_mb_unload_buddy(&e4b);
c9de560d
AT
4015 put_bh(bitmap_bh);
4016
4017 list_del(&pa->u.pa_tmp_list);
4018 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4019 }
4020}
4021
6ba495e9 4022#ifdef CONFIG_EXT4_DEBUG
c9de560d
AT
4023static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4024{
4025 struct super_block *sb = ac->ac_sb;
8df9675f 4026 ext4_group_t ngroups, i;
c9de560d 4027
a0b30c12 4028 if (!ext4_mballoc_debug ||
4dd89fc6 4029 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
e3570639
ES
4030 return;
4031
7f6a11e7 4032 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
9d8b9ec4 4033 " Allocation context details:");
7f6a11e7 4034 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
c9de560d 4035 ac->ac_status, ac->ac_flags);
7f6a11e7 4036 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
9d8b9ec4
TT
4037 "goal %lu/%lu/%lu@%lu, "
4038 "best %lu/%lu/%lu@%lu cr %d",
c9de560d
AT
4039 (unsigned long)ac->ac_o_ex.fe_group,
4040 (unsigned long)ac->ac_o_ex.fe_start,
4041 (unsigned long)ac->ac_o_ex.fe_len,
4042 (unsigned long)ac->ac_o_ex.fe_logical,
4043 (unsigned long)ac->ac_g_ex.fe_group,
4044 (unsigned long)ac->ac_g_ex.fe_start,
4045 (unsigned long)ac->ac_g_ex.fe_len,
4046 (unsigned long)ac->ac_g_ex.fe_logical,
4047 (unsigned long)ac->ac_b_ex.fe_group,
4048 (unsigned long)ac->ac_b_ex.fe_start,
4049 (unsigned long)ac->ac_b_ex.fe_len,
4050 (unsigned long)ac->ac_b_ex.fe_logical,
4051 (int)ac->ac_criteria);
dc9ddd98 4052 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
7f6a11e7 4053 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
8df9675f
TT
4054 ngroups = ext4_get_groups_count(sb);
4055 for (i = 0; i < ngroups; i++) {
c9de560d
AT
4056 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4057 struct ext4_prealloc_space *pa;
4058 ext4_grpblk_t start;
4059 struct list_head *cur;
4060 ext4_lock_group(sb, i);
4061 list_for_each(cur, &grp->bb_prealloc_list) {
4062 pa = list_entry(cur, struct ext4_prealloc_space,
4063 pa_group_list);
4064 spin_lock(&pa->pa_lock);
4065 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4066 NULL, &start);
4067 spin_unlock(&pa->pa_lock);
1c718505
AF
4068 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4069 start, pa->pa_len);
c9de560d 4070 }
60bd63d1 4071 ext4_unlock_group(sb, i);
c9de560d
AT
4072
4073 if (grp->bb_free == 0)
4074 continue;
1c718505 4075 printk(KERN_ERR "%u: %d/%d \n",
c9de560d
AT
4076 i, grp->bb_free, grp->bb_fragments);
4077 }
4078 printk(KERN_ERR "\n");
4079}
4080#else
4081static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4082{
4083 return;
4084}
4085#endif
4086
4087/*
4088 * We use locality group preallocation for small size file. The size of the
4089 * file is determined by the current size or the resulting size after
4090 * allocation which ever is larger
4091 *
b713a5ec 4092 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
c9de560d
AT
4093 */
4094static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4095{
4096 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4097 int bsbits = ac->ac_sb->s_blocksize_bits;
4098 loff_t size, isize;
4099
4100 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4101 return;
4102
4ba74d00
TT
4103 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4104 return;
4105
53accfa9 4106 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
50797481
TT
4107 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4108 >> bsbits;
c9de560d 4109
50797481
TT
4110 if ((size == isize) &&
4111 !ext4_fs_is_busy(sbi) &&
4112 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4113 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4114 return;
4115 }
4116
ebbe0277
RD
4117 if (sbi->s_mb_group_prealloc <= 0) {
4118 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4119 return;
4120 }
4121
c9de560d 4122 /* don't use group allocation for large files */
71780577 4123 size = max(size, isize);
cc483f10 4124 if (size > sbi->s_mb_stream_request) {
4ba74d00 4125 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
c9de560d 4126 return;
4ba74d00 4127 }
c9de560d
AT
4128
4129 BUG_ON(ac->ac_lg != NULL);
4130 /*
4131 * locality group prealloc space are per cpu. The reason for having
4132 * per cpu locality group is to reduce the contention between block
4133 * request from multiple CPUs.
4134 */
a0b6bc63 4135 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
c9de560d
AT
4136
4137 /* we're going to use group allocation */
4138 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4139
4140 /* serialize all allocations in the group */
4141 mutex_lock(&ac->ac_lg->lg_mutex);
4142}
4143
4ddfef7b
ES
4144static noinline_for_stack int
4145ext4_mb_initialize_context(struct ext4_allocation_context *ac,
c9de560d
AT
4146 struct ext4_allocation_request *ar)
4147{
4148 struct super_block *sb = ar->inode->i_sb;
4149 struct ext4_sb_info *sbi = EXT4_SB(sb);
4150 struct ext4_super_block *es = sbi->s_es;
4151 ext4_group_t group;
498e5f24
TT
4152 unsigned int len;
4153 ext4_fsblk_t goal;
c9de560d
AT
4154 ext4_grpblk_t block;
4155
4156 /* we can't allocate > group size */
4157 len = ar->len;
4158
4159 /* just a dirty hack to filter too big requests */
40ae3487
TT
4160 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4161 len = EXT4_CLUSTERS_PER_GROUP(sb);
c9de560d
AT
4162
4163 /* start searching from the goal */
4164 goal = ar->goal;
4165 if (goal < le32_to_cpu(es->s_first_data_block) ||
4166 goal >= ext4_blocks_count(es))
4167 goal = le32_to_cpu(es->s_first_data_block);
4168 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4169
4170 /* set up allocation goals */
f5a44db5 4171 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
c9de560d 4172 ac->ac_status = AC_STATUS_CONTINUE;
c9de560d
AT
4173 ac->ac_sb = sb;
4174 ac->ac_inode = ar->inode;
53accfa9 4175 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
c9de560d
AT
4176 ac->ac_o_ex.fe_group = group;
4177 ac->ac_o_ex.fe_start = block;
4178 ac->ac_o_ex.fe_len = len;
53accfa9 4179 ac->ac_g_ex = ac->ac_o_ex;
c9de560d 4180 ac->ac_flags = ar->flags;
c9de560d
AT
4181
4182 /* we have to define context: we'll we work with a file or
4183 * locality group. this is a policy, actually */
4184 ext4_mb_group_or_file(ac);
4185
6ba495e9 4186 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
c9de560d
AT
4187 "left: %u/%u, right %u/%u to %swritable\n",
4188 (unsigned) ar->len, (unsigned) ar->logical,
4189 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4190 (unsigned) ar->lleft, (unsigned) ar->pleft,
4191 (unsigned) ar->lright, (unsigned) ar->pright,
4192 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4193 return 0;
4194
4195}
4196
6be2ded1
AK
4197static noinline_for_stack void
4198ext4_mb_discard_lg_preallocations(struct super_block *sb,
4199 struct ext4_locality_group *lg,
4200 int order, int total_entries)
4201{
4202 ext4_group_t group = 0;
4203 struct ext4_buddy e4b;
4204 struct list_head discard_list;
4205 struct ext4_prealloc_space *pa, *tmp;
6be2ded1 4206
6ba495e9 4207 mb_debug(1, "discard locality group preallocation\n");
6be2ded1
AK
4208
4209 INIT_LIST_HEAD(&discard_list);
6be2ded1
AK
4210
4211 spin_lock(&lg->lg_prealloc_lock);
4212 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4213 pa_inode_list) {
4214 spin_lock(&pa->pa_lock);
4215 if (atomic_read(&pa->pa_count)) {
4216 /*
4217 * This is the pa that we just used
4218 * for block allocation. So don't
4219 * free that
4220 */
4221 spin_unlock(&pa->pa_lock);
4222 continue;
4223 }
4224 if (pa->pa_deleted) {
4225 spin_unlock(&pa->pa_lock);
4226 continue;
4227 }
4228 /* only lg prealloc space */
cc0fb9ad 4229 BUG_ON(pa->pa_type != MB_GROUP_PA);
6be2ded1
AK
4230
4231 /* seems this one can be freed ... */
4232 pa->pa_deleted = 1;
4233 spin_unlock(&pa->pa_lock);
4234
4235 list_del_rcu(&pa->pa_inode_list);
4236 list_add(&pa->u.pa_tmp_list, &discard_list);
4237
4238 total_entries--;
4239 if (total_entries <= 5) {
4240 /*
4241 * we want to keep only 5 entries
4242 * allowing it to grow to 8. This
4243 * mak sure we don't call discard
4244 * soon for this list.
4245 */
4246 break;
4247 }
4248 }
4249 spin_unlock(&lg->lg_prealloc_lock);
4250
4251 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4252
bd86298e 4253 group = ext4_get_group_number(sb, pa->pa_pstart);
6be2ded1 4254 if (ext4_mb_load_buddy(sb, group, &e4b)) {
12062ddd
ES
4255 ext4_error(sb, "Error loading buddy information for %u",
4256 group);
6be2ded1
AK
4257 continue;
4258 }
4259 ext4_lock_group(sb, group);
4260 list_del(&pa->pa_group_list);
3e1e5f50 4261 ext4_mb_release_group_pa(&e4b, pa);
6be2ded1
AK
4262 ext4_unlock_group(sb, group);
4263
e39e07fd 4264 ext4_mb_unload_buddy(&e4b);
6be2ded1
AK
4265 list_del(&pa->u.pa_tmp_list);
4266 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4267 }
6be2ded1
AK
4268}
4269
4270/*
4271 * We have incremented pa_count. So it cannot be freed at this
4272 * point. Also we hold lg_mutex. So no parallel allocation is
4273 * possible from this lg. That means pa_free cannot be updated.
4274 *
4275 * A parallel ext4_mb_discard_group_preallocations is possible.
4276 * which can cause the lg_prealloc_list to be updated.
4277 */
4278
4279static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4280{
4281 int order, added = 0, lg_prealloc_count = 1;
4282 struct super_block *sb = ac->ac_sb;
4283 struct ext4_locality_group *lg = ac->ac_lg;
4284 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4285
4286 order = fls(pa->pa_free) - 1;
4287 if (order > PREALLOC_TB_SIZE - 1)
4288 /* The max size of hash table is PREALLOC_TB_SIZE */
4289 order = PREALLOC_TB_SIZE - 1;
4290 /* Add the prealloc space to lg */
f1167009 4291 spin_lock(&lg->lg_prealloc_lock);
6be2ded1
AK
4292 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4293 pa_inode_list) {
4294 spin_lock(&tmp_pa->pa_lock);
4295 if (tmp_pa->pa_deleted) {
e7c9e3e9 4296 spin_unlock(&tmp_pa->pa_lock);
6be2ded1
AK
4297 continue;
4298 }
4299 if (!added && pa->pa_free < tmp_pa->pa_free) {
4300 /* Add to the tail of the previous entry */
4301 list_add_tail_rcu(&pa->pa_inode_list,
4302 &tmp_pa->pa_inode_list);
4303 added = 1;
4304 /*
4305 * we want to count the total
4306 * number of entries in the list
4307 */
4308 }
4309 spin_unlock(&tmp_pa->pa_lock);
4310 lg_prealloc_count++;
4311 }
4312 if (!added)
4313 list_add_tail_rcu(&pa->pa_inode_list,
4314 &lg->lg_prealloc_list[order]);
f1167009 4315 spin_unlock(&lg->lg_prealloc_lock);
6be2ded1
AK
4316
4317 /* Now trim the list to be not more than 8 elements */
4318 if (lg_prealloc_count > 8) {
4319 ext4_mb_discard_lg_preallocations(sb, lg,
f1167009 4320 order, lg_prealloc_count);
6be2ded1
AK
4321 return;
4322 }
4323 return ;
4324}
4325
c9de560d
AT
4326/*
4327 * release all resource we used in allocation
4328 */
4329static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4330{
53accfa9 4331 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
6be2ded1
AK
4332 struct ext4_prealloc_space *pa = ac->ac_pa;
4333 if (pa) {
cc0fb9ad 4334 if (pa->pa_type == MB_GROUP_PA) {
c9de560d 4335 /* see comment in ext4_mb_use_group_pa() */
6be2ded1 4336 spin_lock(&pa->pa_lock);
53accfa9
TT
4337 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4338 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
6be2ded1
AK
4339 pa->pa_free -= ac->ac_b_ex.fe_len;
4340 pa->pa_len -= ac->ac_b_ex.fe_len;
4341 spin_unlock(&pa->pa_lock);
c9de560d 4342 }
c9de560d 4343 }
ba443916
AK
4344 if (pa) {
4345 /*
4346 * We want to add the pa to the right bucket.
4347 * Remove it from the list and while adding
4348 * make sure the list to which we are adding
44183d42 4349 * doesn't grow big.
ba443916 4350 */
cc0fb9ad 4351 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
ba443916
AK
4352 spin_lock(pa->pa_obj_lock);
4353 list_del_rcu(&pa->pa_inode_list);
4354 spin_unlock(pa->pa_obj_lock);
4355 ext4_mb_add_n_trim(ac);
4356 }
4357 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4358 }
c9de560d
AT
4359 if (ac->ac_bitmap_page)
4360 page_cache_release(ac->ac_bitmap_page);
4361 if (ac->ac_buddy_page)
4362 page_cache_release(ac->ac_buddy_page);
4363 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4364 mutex_unlock(&ac->ac_lg->lg_mutex);
4365 ext4_mb_collect_stats(ac);
4366 return 0;
4367}
4368
4369static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4370{
8df9675f 4371 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
c9de560d
AT
4372 int ret;
4373 int freed = 0;
4374
9bffad1e 4375 trace_ext4_mb_discard_preallocations(sb, needed);
8df9675f 4376 for (i = 0; i < ngroups && needed > 0; i++) {
c9de560d
AT
4377 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4378 freed += ret;
4379 needed -= ret;
4380 }
4381
4382 return freed;
4383}
4384
4385/*
4386 * Main entry point into mballoc to allocate blocks
4387 * it tries to use preallocation first, then falls back
4388 * to usual allocation
4389 */
4390ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
6c7a120a 4391 struct ext4_allocation_request *ar, int *errp)
c9de560d 4392{
6bc6e63f 4393 int freed;
256bdb49 4394 struct ext4_allocation_context *ac = NULL;
c9de560d
AT
4395 struct ext4_sb_info *sbi;
4396 struct super_block *sb;
4397 ext4_fsblk_t block = 0;
60e58e0f 4398 unsigned int inquota = 0;
53accfa9 4399 unsigned int reserv_clstrs = 0;
c9de560d 4400
b10a44c3 4401 might_sleep();
c9de560d
AT
4402 sb = ar->inode->i_sb;
4403 sbi = EXT4_SB(sb);
4404
9bffad1e 4405 trace_ext4_request_blocks(ar);
ba80b101 4406
45dc63e7
DM
4407 /* Allow to use superuser reservation for quota file */
4408 if (IS_NOQUOTA(ar->inode))
4409 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4410
e3cf5d5d 4411 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
60e58e0f
MC
4412 /* Without delayed allocation we need to verify
4413 * there is enough free blocks to do block allocation
4414 * and verify allocation doesn't exceed the quota limits.
d2a17637 4415 */
55f020db 4416 while (ar->len &&
e7d5f315 4417 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
55f020db 4418
030ba6bc 4419 /* let others to free the space */
bb8b20ed 4420 cond_resched();
030ba6bc
AK
4421 ar->len = ar->len >> 1;
4422 }
4423 if (!ar->len) {
a30d542a
AK
4424 *errp = -ENOSPC;
4425 return 0;
4426 }
53accfa9 4427 reserv_clstrs = ar->len;
55f020db 4428 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
53accfa9
TT
4429 dquot_alloc_block_nofail(ar->inode,
4430 EXT4_C2B(sbi, ar->len));
55f020db
AH
4431 } else {
4432 while (ar->len &&
53accfa9
TT
4433 dquot_alloc_block(ar->inode,
4434 EXT4_C2B(sbi, ar->len))) {
55f020db
AH
4435
4436 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4437 ar->len--;
4438 }
60e58e0f
MC
4439 }
4440 inquota = ar->len;
4441 if (ar->len == 0) {
4442 *errp = -EDQUOT;
6c7a120a 4443 goto out;
60e58e0f 4444 }
07031431 4445 }
d2a17637 4446
85556c9a 4447 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
833576b3 4448 if (!ac) {
363d4251 4449 ar->len = 0;
256bdb49 4450 *errp = -ENOMEM;
6c7a120a 4451 goto out;
256bdb49
ES
4452 }
4453
256bdb49 4454 *errp = ext4_mb_initialize_context(ac, ar);
c9de560d
AT
4455 if (*errp) {
4456 ar->len = 0;
6c7a120a 4457 goto out;
c9de560d
AT
4458 }
4459
256bdb49
ES
4460 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4461 if (!ext4_mb_use_preallocated(ac)) {
256bdb49
ES
4462 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4463 ext4_mb_normalize_request(ac, ar);
c9de560d
AT
4464repeat:
4465 /* allocate space in core */
6c7a120a 4466 *errp = ext4_mb_regular_allocator(ac);
2c00ef3e
AK
4467 if (*errp)
4468 goto discard_and_exit;
c9de560d
AT
4469
4470 /* as we've just preallocated more space than
2c00ef3e 4471 * user requested originally, we store allocated
c9de560d 4472 * space in a special descriptor */
256bdb49 4473 if (ac->ac_status == AC_STATUS_FOUND &&
2c00ef3e
AK
4474 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4475 *errp = ext4_mb_new_preallocation(ac);
4476 if (*errp) {
4477 discard_and_exit:
4478 ext4_discard_allocated_blocks(ac);
4479 goto errout;
4480 }
c9de560d 4481 }
256bdb49 4482 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
53accfa9 4483 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
6c7a120a 4484 if (*errp == -EAGAIN) {
8556e8f3
AK
4485 /*
4486 * drop the reference that we took
4487 * in ext4_mb_use_best_found
4488 */
4489 ext4_mb_release_context(ac);
519deca0
AK
4490 ac->ac_b_ex.fe_group = 0;
4491 ac->ac_b_ex.fe_start = 0;
4492 ac->ac_b_ex.fe_len = 0;
4493 ac->ac_status = AC_STATUS_CONTINUE;
4494 goto repeat;
6d138ced 4495 } else if (*errp) {
b844167e 4496 ext4_discard_allocated_blocks(ac);
6d138ced
ES
4497 goto errout;
4498 } else {
519deca0
AK
4499 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4500 ar->len = ac->ac_b_ex.fe_len;
4501 }
c9de560d 4502 } else {
256bdb49 4503 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
c9de560d
AT
4504 if (freed)
4505 goto repeat;
4506 *errp = -ENOSPC;
6c7a120a
AK
4507 }
4508
6d138ced 4509errout:
6c7a120a 4510 if (*errp) {
256bdb49 4511 ac->ac_b_ex.fe_len = 0;
c9de560d 4512 ar->len = 0;
256bdb49 4513 ext4_mb_show_ac(ac);
c9de560d 4514 }
256bdb49 4515 ext4_mb_release_context(ac);
6c7a120a
AK
4516out:
4517 if (ac)
4518 kmem_cache_free(ext4_ac_cachep, ac);
60e58e0f 4519 if (inquota && ar->len < inquota)
53accfa9 4520 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
0087d9fb 4521 if (!ar->len) {
e3cf5d5d 4522 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
0087d9fb 4523 /* release all the reserved blocks if non delalloc */
57042651 4524 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
53accfa9 4525 reserv_clstrs);
0087d9fb 4526 }
c9de560d 4527
9bffad1e 4528 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
ba80b101 4529
c9de560d
AT
4530 return block;
4531}
c9de560d 4532
c894058d
AK
4533/*
4534 * We can merge two free data extents only if the physical blocks
4535 * are contiguous, AND the extents were freed by the same transaction,
4536 * AND the blocks are associated with the same group.
4537 */
4538static int can_merge(struct ext4_free_data *entry1,
4539 struct ext4_free_data *entry2)
4540{
18aadd47
BJ
4541 if ((entry1->efd_tid == entry2->efd_tid) &&
4542 (entry1->efd_group == entry2->efd_group) &&
4543 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
c894058d
AK
4544 return 1;
4545 return 0;
4546}
4547
4ddfef7b
ES
4548static noinline_for_stack int
4549ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
7a2fcbf7 4550 struct ext4_free_data *new_entry)
c9de560d 4551{
e29136f8 4552 ext4_group_t group = e4b->bd_group;
84130193 4553 ext4_grpblk_t cluster;
7a2fcbf7 4554 struct ext4_free_data *entry;
c9de560d
AT
4555 struct ext4_group_info *db = e4b->bd_info;
4556 struct super_block *sb = e4b->bd_sb;
4557 struct ext4_sb_info *sbi = EXT4_SB(sb);
c894058d
AK
4558 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4559 struct rb_node *parent = NULL, *new_node;
4560
0390131b 4561 BUG_ON(!ext4_handle_valid(handle));
c9de560d
AT
4562 BUG_ON(e4b->bd_bitmap_page == NULL);
4563 BUG_ON(e4b->bd_buddy_page == NULL);
4564
18aadd47
BJ
4565 new_node = &new_entry->efd_node;
4566 cluster = new_entry->efd_start_cluster;
c894058d 4567
c894058d
AK
4568 if (!*n) {
4569 /* first free block exent. We need to
4570 protect buddy cache from being freed,
4571 * otherwise we'll refresh it from
4572 * on-disk bitmap and lose not-yet-available
4573 * blocks */
4574 page_cache_get(e4b->bd_buddy_page);
4575 page_cache_get(e4b->bd_bitmap_page);
4576 }
4577 while (*n) {
4578 parent = *n;
18aadd47
BJ
4579 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4580 if (cluster < entry->efd_start_cluster)
c894058d 4581 n = &(*n)->rb_left;
18aadd47 4582 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
c894058d
AK
4583 n = &(*n)->rb_right;
4584 else {
e29136f8 4585 ext4_grp_locked_error(sb, group, 0,
84130193
TT
4586 ext4_group_first_block_no(sb, group) +
4587 EXT4_C2B(sbi, cluster),
e29136f8 4588 "Block already on to-be-freed list");
c894058d 4589 return 0;
c9de560d 4590 }
c894058d 4591 }
c9de560d 4592
c894058d
AK
4593 rb_link_node(new_node, parent, n);
4594 rb_insert_color(new_node, &db->bb_free_root);
4595
4596 /* Now try to see the extent can be merged to left and right */
4597 node = rb_prev(new_node);
4598 if (node) {
18aadd47 4599 entry = rb_entry(node, struct ext4_free_data, efd_node);
5d3ee208
DM
4600 if (can_merge(entry, new_entry) &&
4601 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
18aadd47
BJ
4602 new_entry->efd_start_cluster = entry->efd_start_cluster;
4603 new_entry->efd_count += entry->efd_count;
c894058d 4604 rb_erase(node, &(db->bb_free_root));
18aadd47 4605 kmem_cache_free(ext4_free_data_cachep, entry);
c9de560d 4606 }
c894058d 4607 }
c9de560d 4608
c894058d
AK
4609 node = rb_next(new_node);
4610 if (node) {
18aadd47 4611 entry = rb_entry(node, struct ext4_free_data, efd_node);
5d3ee208
DM
4612 if (can_merge(new_entry, entry) &&
4613 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
18aadd47 4614 new_entry->efd_count += entry->efd_count;
c894058d 4615 rb_erase(node, &(db->bb_free_root));
18aadd47 4616 kmem_cache_free(ext4_free_data_cachep, entry);
c9de560d
AT
4617 }
4618 }
3e624fc7 4619 /* Add the extent to transaction's private list */
18aadd47
BJ
4620 ext4_journal_callback_add(handle, ext4_free_data_callback,
4621 &new_entry->efd_jce);
c9de560d
AT
4622 return 0;
4623}
4624
44338711
TT
4625/**
4626 * ext4_free_blocks() -- Free given blocks and update quota
4627 * @handle: handle for this transaction
4628 * @inode: inode
4629 * @block: start physical block to free
4630 * @count: number of blocks to count
5def1360 4631 * @flags: flags used by ext4_free_blocks
c9de560d 4632 */
44338711 4633void ext4_free_blocks(handle_t *handle, struct inode *inode,
e6362609
TT
4634 struct buffer_head *bh, ext4_fsblk_t block,
4635 unsigned long count, int flags)
c9de560d 4636{
26346ff6 4637 struct buffer_head *bitmap_bh = NULL;
c9de560d 4638 struct super_block *sb = inode->i_sb;
c9de560d 4639 struct ext4_group_desc *gdp;
498e5f24 4640 unsigned int overflow;
c9de560d
AT
4641 ext4_grpblk_t bit;
4642 struct buffer_head *gd_bh;
4643 ext4_group_t block_group;
4644 struct ext4_sb_info *sbi;
4645 struct ext4_buddy e4b;
84130193 4646 unsigned int count_clusters;
c9de560d
AT
4647 int err = 0;
4648 int ret;
4649
b10a44c3 4650 might_sleep();
e6362609
TT
4651 if (bh) {
4652 if (block)
4653 BUG_ON(block != bh->b_blocknr);
4654 else
4655 block = bh->b_blocknr;
4656 }
c9de560d 4657
c9de560d 4658 sbi = EXT4_SB(sb);
1f2acb60
TT
4659 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4660 !ext4_data_block_valid(sbi, block, count)) {
12062ddd 4661 ext4_error(sb, "Freeing blocks not in datazone - "
1f2acb60 4662 "block = %llu, count = %lu", block, count);
c9de560d
AT
4663 goto error_return;
4664 }
4665
0610b6e9 4666 ext4_debug("freeing block %llu\n", block);
e6362609
TT
4667 trace_ext4_free_blocks(inode, block, count, flags);
4668
4669 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4670 struct buffer_head *tbh = bh;
4671 int i;
4672
4673 BUG_ON(bh && (count > 1));
4674
4675 for (i = 0; i < count; i++) {
2ed5724d 4676 cond_resched();
e6362609
TT
4677 if (!bh)
4678 tbh = sb_find_get_block(inode->i_sb,
4679 block + i);
2ed5724d 4680 if (!tbh)
87783690 4681 continue;
60e6679e 4682 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
e6362609
TT
4683 inode, tbh, block + i);
4684 }
4685 }
4686
60e6679e 4687 /*
e6362609
TT
4688 * We need to make sure we don't reuse the freed block until
4689 * after the transaction is committed, which we can do by
4690 * treating the block as metadata, below. We make an
4691 * exception if the inode is to be written in writeback mode
4692 * since writeback mode has weak data consistency guarantees.
4693 */
4694 if (!ext4_should_writeback_data(inode))
4695 flags |= EXT4_FREE_BLOCKS_METADATA;
c9de560d 4696
84130193
TT
4697 /*
4698 * If the extent to be freed does not begin on a cluster
4699 * boundary, we need to deal with partial clusters at the
4700 * beginning and end of the extent. Normally we will free
4701 * blocks at the beginning or the end unless we are explicitly
4702 * requested to avoid doing so.
4703 */
f5a44db5 4704 overflow = EXT4_PBLK_COFF(sbi, block);
84130193
TT
4705 if (overflow) {
4706 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4707 overflow = sbi->s_cluster_ratio - overflow;
4708 block += overflow;
4709 if (count > overflow)
4710 count -= overflow;
4711 else
4712 return;
4713 } else {
4714 block -= overflow;
4715 count += overflow;
4716 }
4717 }
f5a44db5 4718 overflow = EXT4_LBLK_COFF(sbi, count);
84130193
TT
4719 if (overflow) {
4720 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4721 if (count > overflow)
4722 count -= overflow;
4723 else
4724 return;
4725 } else
4726 count += sbi->s_cluster_ratio - overflow;
4727 }
4728
c9de560d
AT
4729do_more:
4730 overflow = 0;
4731 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4732
163a203d
DW
4733 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4734 ext4_get_group_info(sb, block_group))))
4735 return;
4736
c9de560d
AT
4737 /*
4738 * Check to see if we are freeing blocks across a group
4739 * boundary.
4740 */
84130193
TT
4741 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4742 overflow = EXT4_C2B(sbi, bit) + count -
4743 EXT4_BLOCKS_PER_GROUP(sb);
c9de560d
AT
4744 count -= overflow;
4745 }
810da240 4746 count_clusters = EXT4_NUM_B2C(sbi, count);
574ca174 4747 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
ce89f46c
AK
4748 if (!bitmap_bh) {
4749 err = -EIO;
c9de560d 4750 goto error_return;
ce89f46c 4751 }
c9de560d 4752 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
ce89f46c
AK
4753 if (!gdp) {
4754 err = -EIO;
c9de560d 4755 goto error_return;
ce89f46c 4756 }
c9de560d
AT
4757
4758 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4759 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4760 in_range(block, ext4_inode_table(sb, gdp),
84130193 4761 EXT4_SB(sb)->s_itb_per_group) ||
c9de560d 4762 in_range(block + count - 1, ext4_inode_table(sb, gdp),
84130193 4763 EXT4_SB(sb)->s_itb_per_group)) {
c9de560d 4764
12062ddd 4765 ext4_error(sb, "Freeing blocks in system zone - "
0610b6e9 4766 "Block = %llu, count = %lu", block, count);
519deca0
AK
4767 /* err = 0. ext4_std_error should be a no op */
4768 goto error_return;
c9de560d
AT
4769 }
4770
4771 BUFFER_TRACE(bitmap_bh, "getting write access");
4772 err = ext4_journal_get_write_access(handle, bitmap_bh);
4773 if (err)
4774 goto error_return;
4775
4776 /*
4777 * We are about to modify some metadata. Call the journal APIs
4778 * to unshare ->b_data if a currently-committing transaction is
4779 * using it
4780 */
4781 BUFFER_TRACE(gd_bh, "get_write_access");
4782 err = ext4_journal_get_write_access(handle, gd_bh);
4783 if (err)
4784 goto error_return;
c9de560d
AT
4785#ifdef AGGRESSIVE_CHECK
4786 {
4787 int i;
84130193 4788 for (i = 0; i < count_clusters; i++)
c9de560d
AT
4789 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4790 }
4791#endif
84130193 4792 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
c9de560d 4793
920313a7
AK
4794 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4795 if (err)
4796 goto error_return;
e6362609
TT
4797
4798 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
7a2fcbf7
AK
4799 struct ext4_free_data *new_entry;
4800 /*
4801 * blocks being freed are metadata. these blocks shouldn't
4802 * be used until this transaction is committed
4803 */
e7676a70 4804 retry:
18aadd47 4805 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
b72143ab 4806 if (!new_entry) {
e7676a70
TT
4807 /*
4808 * We use a retry loop because
4809 * ext4_free_blocks() is not allowed to fail.
4810 */
4811 cond_resched();
4812 congestion_wait(BLK_RW_ASYNC, HZ/50);
4813 goto retry;
b72143ab 4814 }
18aadd47
BJ
4815 new_entry->efd_start_cluster = bit;
4816 new_entry->efd_group = block_group;
4817 new_entry->efd_count = count_clusters;
4818 new_entry->efd_tid = handle->h_transaction->t_tid;
955ce5f5 4819
7a2fcbf7 4820 ext4_lock_group(sb, block_group);
84130193 4821 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
7a2fcbf7 4822 ext4_mb_free_metadata(handle, &e4b, new_entry);
c9de560d 4823 } else {
7a2fcbf7
AK
4824 /* need to update group_info->bb_free and bitmap
4825 * with group lock held. generate_buddy look at
4826 * them with group lock_held
4827 */
d71c1ae2
LC
4828 if (test_opt(sb, DISCARD)) {
4829 err = ext4_issue_discard(sb, block_group, bit, count);
4830 if (err && err != -EOPNOTSUPP)
4831 ext4_msg(sb, KERN_WARNING, "discard request in"
4832 " group:%d block:%d count:%lu failed"
4833 " with %d", block_group, bit, count,
4834 err);
8f9ff189
LC
4835 } else
4836 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
d71c1ae2 4837
955ce5f5 4838 ext4_lock_group(sb, block_group);
84130193
TT
4839 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4840 mb_free_blocks(inode, &e4b, bit, count_clusters);
c9de560d
AT
4841 }
4842
021b65bb
TT
4843 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4844 ext4_free_group_clusters_set(sb, gdp, ret);
79f1ba49 4845 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
feb0ab32 4846 ext4_group_desc_csum_set(sb, block_group, gdp);
955ce5f5 4847 ext4_unlock_group(sb, block_group);
c9de560d 4848
772cb7c8
JS
4849 if (sbi->s_log_groups_per_flex) {
4850 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
90ba983f
TT
4851 atomic64_add(count_clusters,
4852 &sbi->s_flex_groups[flex_group].free_clusters);
772cb7c8
JS
4853 }
4854
71d4f7d0 4855 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
7b415bf6 4856 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
7d734532
JK
4857 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4858
4859 ext4_mb_unload_buddy(&e4b);
7b415bf6 4860
7a2fcbf7
AK
4861 /* We dirtied the bitmap block */
4862 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4863 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4864
c9de560d
AT
4865 /* And the group descriptor block */
4866 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
0390131b 4867 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
c9de560d
AT
4868 if (!err)
4869 err = ret;
4870
4871 if (overflow && !err) {
4872 block += count;
4873 count = overflow;
4874 put_bh(bitmap_bh);
4875 goto do_more;
4876 }
c9de560d
AT
4877error_return:
4878 brelse(bitmap_bh);
4879 ext4_std_error(sb, err);
4880 return;
4881}
7360d173 4882
2846e820 4883/**
0529155e 4884 * ext4_group_add_blocks() -- Add given blocks to an existing group
2846e820
AG
4885 * @handle: handle to this transaction
4886 * @sb: super block
4907cb7b 4887 * @block: start physical block to add to the block group
2846e820
AG
4888 * @count: number of blocks to free
4889 *
e73a347b 4890 * This marks the blocks as free in the bitmap and buddy.
2846e820 4891 */
cc7365df 4892int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
2846e820
AG
4893 ext4_fsblk_t block, unsigned long count)
4894{
4895 struct buffer_head *bitmap_bh = NULL;
4896 struct buffer_head *gd_bh;
4897 ext4_group_t block_group;
4898 ext4_grpblk_t bit;
4899 unsigned int i;
4900 struct ext4_group_desc *desc;
4901 struct ext4_sb_info *sbi = EXT4_SB(sb);
e73a347b 4902 struct ext4_buddy e4b;
2846e820
AG
4903 int err = 0, ret, blk_free_count;
4904 ext4_grpblk_t blocks_freed;
2846e820
AG
4905
4906 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4907
4740b830
YY
4908 if (count == 0)
4909 return 0;
4910
2846e820 4911 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
2846e820
AG
4912 /*
4913 * Check to see if we are freeing blocks across a group
4914 * boundary.
4915 */
cc7365df
YY
4916 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4917 ext4_warning(sb, "too much blocks added to group %u\n",
4918 block_group);
4919 err = -EINVAL;
2846e820 4920 goto error_return;
cc7365df 4921 }
2cd05cc3 4922
2846e820 4923 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
cc7365df
YY
4924 if (!bitmap_bh) {
4925 err = -EIO;
2846e820 4926 goto error_return;
cc7365df
YY
4927 }
4928
2846e820 4929 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
cc7365df
YY
4930 if (!desc) {
4931 err = -EIO;
2846e820 4932 goto error_return;
cc7365df 4933 }
2846e820
AG
4934
4935 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4936 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4937 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4938 in_range(block + count - 1, ext4_inode_table(sb, desc),
4939 sbi->s_itb_per_group)) {
4940 ext4_error(sb, "Adding blocks in system zones - "
4941 "Block = %llu, count = %lu",
4942 block, count);
cc7365df 4943 err = -EINVAL;
2846e820
AG
4944 goto error_return;
4945 }
4946
2cd05cc3
TT
4947 BUFFER_TRACE(bitmap_bh, "getting write access");
4948 err = ext4_journal_get_write_access(handle, bitmap_bh);
2846e820
AG
4949 if (err)
4950 goto error_return;
4951
4952 /*
4953 * We are about to modify some metadata. Call the journal APIs
4954 * to unshare ->b_data if a currently-committing transaction is
4955 * using it
4956 */
4957 BUFFER_TRACE(gd_bh, "get_write_access");
4958 err = ext4_journal_get_write_access(handle, gd_bh);
4959 if (err)
4960 goto error_return;
e73a347b 4961
2846e820
AG
4962 for (i = 0, blocks_freed = 0; i < count; i++) {
4963 BUFFER_TRACE(bitmap_bh, "clear bit");
e73a347b 4964 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
2846e820
AG
4965 ext4_error(sb, "bit already cleared for block %llu",
4966 (ext4_fsblk_t)(block + i));
4967 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4968 } else {
4969 blocks_freed++;
4970 }
4971 }
e73a347b
AG
4972
4973 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4974 if (err)
4975 goto error_return;
4976
4977 /*
4978 * need to update group_info->bb_free and bitmap
4979 * with group lock held. generate_buddy look at
4980 * them with group lock_held
4981 */
2846e820 4982 ext4_lock_group(sb, block_group);
e73a347b
AG
4983 mb_clear_bits(bitmap_bh->b_data, bit, count);
4984 mb_free_blocks(NULL, &e4b, bit, count);
021b65bb
TT
4985 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4986 ext4_free_group_clusters_set(sb, desc, blk_free_count);
79f1ba49 4987 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
feb0ab32 4988 ext4_group_desc_csum_set(sb, block_group, desc);
2846e820 4989 ext4_unlock_group(sb, block_group);
57042651 4990 percpu_counter_add(&sbi->s_freeclusters_counter,
810da240 4991 EXT4_NUM_B2C(sbi, blocks_freed));
2846e820
AG
4992
4993 if (sbi->s_log_groups_per_flex) {
4994 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
90ba983f
TT
4995 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4996 &sbi->s_flex_groups[flex_group].free_clusters);
2846e820 4997 }
e73a347b
AG
4998
4999 ext4_mb_unload_buddy(&e4b);
2846e820
AG
5000
5001 /* We dirtied the bitmap block */
5002 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5003 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5004
5005 /* And the group descriptor block */
5006 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5007 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5008 if (!err)
5009 err = ret;
5010
5011error_return:
5012 brelse(bitmap_bh);
5013 ext4_std_error(sb, err);
cc7365df 5014 return err;
2846e820
AG
5015}
5016
7360d173
LC
5017/**
5018 * ext4_trim_extent -- function to TRIM one single free extent in the group
5019 * @sb: super block for the file system
5020 * @start: starting block of the free extent in the alloc. group
5021 * @count: number of blocks to TRIM
5022 * @group: alloc. group we are working with
5023 * @e4b: ext4 buddy for the group
5024 *
5025 * Trim "count" blocks starting at "start" in the "group". To assure that no
5026 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5027 * be called with under the group lock.
5028 */
d71c1ae2 5029static int ext4_trim_extent(struct super_block *sb, int start, int count,
d9f34504 5030 ext4_group_t group, struct ext4_buddy *e4b)
e2cbd587 5031__releases(bitlock)
5032__acquires(bitlock)
7360d173
LC
5033{
5034 struct ext4_free_extent ex;
d71c1ae2 5035 int ret = 0;
7360d173 5036
b3d4c2b1
TM
5037 trace_ext4_trim_extent(sb, group, start, count);
5038
7360d173
LC
5039 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5040
5041 ex.fe_start = start;
5042 ex.fe_group = group;
5043 ex.fe_len = count;
5044
5045 /*
5046 * Mark blocks used, so no one can reuse them while
5047 * being trimmed.
5048 */
5049 mb_mark_used(e4b, &ex);
5050 ext4_unlock_group(sb, group);
d71c1ae2 5051 ret = ext4_issue_discard(sb, group, start, count);
7360d173
LC
5052 ext4_lock_group(sb, group);
5053 mb_free_blocks(NULL, e4b, start, ex.fe_len);
d71c1ae2 5054 return ret;
7360d173
LC
5055}
5056
5057/**
5058 * ext4_trim_all_free -- function to trim all free space in alloc. group
5059 * @sb: super block for file system
22612283 5060 * @group: group to be trimmed
7360d173
LC
5061 * @start: first group block to examine
5062 * @max: last group block to examine
5063 * @minblocks: minimum extent block count
5064 *
5065 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5066 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5067 * the extent.
5068 *
5069 *
5070 * ext4_trim_all_free walks through group's block bitmap searching for free
5071 * extents. When the free extent is found, mark it as used in group buddy
5072 * bitmap. Then issue a TRIM command on this extent and free the extent in
5073 * the group buddy bitmap. This is done until whole group is scanned.
5074 */
0b75a840 5075static ext4_grpblk_t
78944086
LC
5076ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5077 ext4_grpblk_t start, ext4_grpblk_t max,
5078 ext4_grpblk_t minblocks)
7360d173
LC
5079{
5080 void *bitmap;
169ddc3e 5081 ext4_grpblk_t next, count = 0, free_count = 0;
78944086 5082 struct ext4_buddy e4b;
d71c1ae2 5083 int ret = 0;
7360d173 5084
b3d4c2b1
TM
5085 trace_ext4_trim_all_free(sb, group, start, max);
5086
78944086
LC
5087 ret = ext4_mb_load_buddy(sb, group, &e4b);
5088 if (ret) {
5089 ext4_error(sb, "Error in loading buddy "
5090 "information for %u", group);
5091 return ret;
5092 }
78944086 5093 bitmap = e4b.bd_bitmap;
28739eea
LC
5094
5095 ext4_lock_group(sb, group);
3d56b8d2
TM
5096 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5097 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5098 goto out;
5099
78944086
LC
5100 start = (e4b.bd_info->bb_first_free > start) ?
5101 e4b.bd_info->bb_first_free : start;
7360d173 5102
913eed83
LC
5103 while (start <= max) {
5104 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5105 if (start > max)
7360d173 5106 break;
913eed83 5107 next = mb_find_next_bit(bitmap, max + 1, start);
7360d173
LC
5108
5109 if ((next - start) >= minblocks) {
d71c1ae2
LC
5110 ret = ext4_trim_extent(sb, start,
5111 next - start, group, &e4b);
5112 if (ret && ret != -EOPNOTSUPP)
5113 break;
5114 ret = 0;
7360d173
LC
5115 count += next - start;
5116 }
169ddc3e 5117 free_count += next - start;
7360d173
LC
5118 start = next + 1;
5119
5120 if (fatal_signal_pending(current)) {
5121 count = -ERESTARTSYS;
5122 break;
5123 }
5124
5125 if (need_resched()) {
5126 ext4_unlock_group(sb, group);
5127 cond_resched();
5128 ext4_lock_group(sb, group);
5129 }
5130
169ddc3e 5131 if ((e4b.bd_info->bb_free - free_count) < minblocks)
7360d173
LC
5132 break;
5133 }
3d56b8d2 5134
d71c1ae2
LC
5135 if (!ret) {
5136 ret = count;
3d56b8d2 5137 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
d71c1ae2 5138 }
3d56b8d2 5139out:
7360d173 5140 ext4_unlock_group(sb, group);
78944086 5141 ext4_mb_unload_buddy(&e4b);
7360d173
LC
5142
5143 ext4_debug("trimmed %d blocks in the group %d\n",
5144 count, group);
5145
d71c1ae2 5146 return ret;
7360d173
LC
5147}
5148
5149/**
5150 * ext4_trim_fs() -- trim ioctl handle function
5151 * @sb: superblock for filesystem
5152 * @range: fstrim_range structure
5153 *
5154 * start: First Byte to trim
5155 * len: number of Bytes to trim from start
5156 * minlen: minimum extent length in Bytes
5157 * ext4_trim_fs goes through all allocation groups containing Bytes from
5158 * start to start+len. For each such a group ext4_trim_all_free function
5159 * is invoked to trim all free space.
5160 */
5161int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5162{
78944086 5163 struct ext4_group_info *grp;
913eed83 5164 ext4_group_t group, first_group, last_group;
7137d7a4 5165 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
913eed83 5166 uint64_t start, end, minlen, trimmed = 0;
0f0a25bf
JK
5167 ext4_fsblk_t first_data_blk =
5168 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
913eed83 5169 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
7360d173
LC
5170 int ret = 0;
5171
5172 start = range->start >> sb->s_blocksize_bits;
913eed83 5173 end = start + (range->len >> sb->s_blocksize_bits) - 1;
aaf7d73e
LC
5174 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5175 range->minlen >> sb->s_blocksize_bits);
7360d173 5176
5de35e8d
LC
5177 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5178 start >= max_blks ||
5179 range->len < sb->s_blocksize)
7360d173 5180 return -EINVAL;
913eed83
LC
5181 if (end >= max_blks)
5182 end = max_blks - 1;
5183 if (end <= first_data_blk)
22f10457 5184 goto out;
913eed83 5185 if (start < first_data_blk)
0f0a25bf 5186 start = first_data_blk;
7360d173 5187
913eed83 5188 /* Determine first and last group to examine based on start and end */
7360d173 5189 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
7137d7a4 5190 &first_group, &first_cluster);
913eed83 5191 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
7137d7a4 5192 &last_group, &last_cluster);
7360d173 5193
913eed83
LC
5194 /* end now represents the last cluster to discard in this group */
5195 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
7360d173
LC
5196
5197 for (group = first_group; group <= last_group; group++) {
78944086
LC
5198 grp = ext4_get_group_info(sb, group);
5199 /* We only do this if the grp has never been initialized */
5200 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5201 ret = ext4_mb_init_group(sb, group);
5202 if (ret)
5203 break;
7360d173
LC
5204 }
5205
0ba08517 5206 /*
913eed83
LC
5207 * For all the groups except the last one, last cluster will
5208 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5209 * change it for the last group, note that last_cluster is
5210 * already computed earlier by ext4_get_group_no_and_offset()
0ba08517 5211 */
913eed83
LC
5212 if (group == last_group)
5213 end = last_cluster;
7360d173 5214
78944086 5215 if (grp->bb_free >= minlen) {
7137d7a4 5216 cnt = ext4_trim_all_free(sb, group, first_cluster,
913eed83 5217 end, minlen);
7360d173
LC
5218 if (cnt < 0) {
5219 ret = cnt;
7360d173
LC
5220 break;
5221 }
21e7fd22 5222 trimmed += cnt;
7360d173 5223 }
913eed83
LC
5224
5225 /*
5226 * For every group except the first one, we are sure
5227 * that the first cluster to discard will be cluster #0.
5228 */
7137d7a4 5229 first_cluster = 0;
7360d173 5230 }
7360d173 5231
3d56b8d2
TM
5232 if (!ret)
5233 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5234
22f10457 5235out:
aaf7d73e 5236 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
7360d173
LC
5237 return ret;
5238}