blob: d34afad3e1371b89c3bfb600c51d55e714f00419 [file] [log] [blame]
Alex Tomasc9de5602008-01-29 00:19:52 -05001/*
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
Mingming Cao8f6e39a2008-04-29 22:01:31 -040024#include "mballoc.h"
Theodore Ts'o6ba495e2009-09-18 13:38:55 -040025#include <linux/debugfs.h>
Theodore Ts'o9bffad12009-06-17 11:48:11 -040026#include <trace/events/ext4.h>
27
Alex Tomasc9de5602008-01-29 00:19:52 -050028/*
29 * MUSTDO:
30 * - test ext4_ext_search_left() and ext4_ext_search_right()
31 * - search for metadata in few groups
32 *
33 * TODO v4:
34 * - normalization should take into account whether file is still open
35 * - discard preallocations if no free space left (policy?)
36 * - don't normalize tails
37 * - quota
38 * - reservation for superuser
39 *
40 * TODO v3:
41 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
42 * - track min/max extents in each group for better group selection
43 * - mb_mark_used() may allocate chunk right after splitting buddy
44 * - tree of groups sorted by number of free blocks
45 * - error handling
46 */
47
48/*
49 * The allocation request involve request for multiple number of blocks
50 * near to the goal(block) value specified.
51 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040052 * During initialization phase of the allocator we decide to use the
53 * group preallocation or inode preallocation depending on the size of
54 * the file. The size of the file could be the resulting file size we
55 * would have after allocation, or the current file size, which ever
56 * is larger. If the size is less than sbi->s_mb_stream_request we
57 * select to use the group preallocation. The default value of
58 * s_mb_stream_request is 16 blocks. This can also be tuned via
59 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
60 * terms of number of blocks.
Alex Tomasc9de5602008-01-29 00:19:52 -050061 *
62 * The main motivation for having small file use group preallocation is to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040063 * ensure that we have small files closer together on the disk.
Alex Tomasc9de5602008-01-29 00:19:52 -050064 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040065 * First stage the allocator looks at the inode prealloc list,
66 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
67 * spaces for this particular inode. The inode prealloc space is
68 * represented as:
Alex Tomasc9de5602008-01-29 00:19:52 -050069 *
70 * pa_lstart -> the logical start block for this prealloc space
71 * pa_pstart -> the physical start block for this prealloc space
72 * pa_len -> lenght for this prealloc space
73 * pa_free -> free space available in this prealloc space
74 *
75 * The inode preallocation space is used looking at the _logical_ start
76 * block. If only the logical file block falls within the range of prealloc
77 * space we will consume the particular prealloc space. This make sure that
78 * that the we have contiguous physical blocks representing the file blocks
79 *
80 * The important thing to be noted in case of inode prealloc space is that
81 * we don't modify the values associated to inode prealloc space except
82 * pa_free.
83 *
84 * If we are not able to find blocks in the inode prealloc space and if we
85 * have the group allocation flag set then we look at the locality group
86 * prealloc space. These are per CPU prealloc list repreasented as
87 *
88 * ext4_sb_info.s_locality_groups[smp_processor_id()]
89 *
90 * The reason for having a per cpu locality group is to reduce the contention
91 * between CPUs. It is possible to get scheduled at this point.
92 *
93 * The locality group prealloc space is used looking at whether we have
94 * enough free space (pa_free) withing the prealloc space.
95 *
96 * If we can't allocate blocks via inode prealloc or/and locality group
97 * prealloc then we look at the buddy cache. The buddy cache is represented
98 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
99 * mapped to the buddy and bitmap information regarding different
100 * groups. The buddy information is attached to buddy cache inode so that
101 * we can access them through the page cache. The information regarding
102 * each group is loaded via ext4_mb_load_buddy. The information involve
103 * block bitmap and buddy information. The information are stored in the
104 * inode as:
105 *
106 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500107 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500108 *
109 *
110 * one block each for bitmap and buddy information. So for each group we
111 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
112 * blocksize) blocks. So it can have information regarding groups_per_page
113 * which is blocks_per_page/2
114 *
115 * The buddy cache inode is not stored on disk. The inode is thrown
116 * away when the filesystem is unmounted.
117 *
118 * We look for count number of blocks in the buddy cache. If we were able
119 * to locate that many free blocks we return with additional information
120 * regarding rest of the contiguous physical block available
121 *
122 * Before allocating blocks via buddy cache we normalize the request
123 * blocks. This ensure we ask for more blocks that we needed. The extra
124 * blocks that we get after allocation is added to the respective prealloc
125 * list. In case of inode preallocation we follow a list of heuristics
126 * based on file size. This can be found in ext4_mb_normalize_request. If
127 * we are doing a group prealloc we try to normalize the request to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400128 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
Alex Tomasc9de5602008-01-29 00:19:52 -0500129 * 512 blocks. This can be tuned via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400130 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
Alex Tomasc9de5602008-01-29 00:19:52 -0500131 * terms of number of blocks. If we have mounted the file system with -O
132 * stripe=<value> option the group prealloc request is normalized to the
133 * stripe value (sbi->s_stripe)
134 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400135 * The regular allocator(using the buddy cache) supports few tunables.
Alex Tomasc9de5602008-01-29 00:19:52 -0500136 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400137 * /sys/fs/ext4/<partition>/mb_min_to_scan
138 * /sys/fs/ext4/<partition>/mb_max_to_scan
139 * /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -0500140 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400141 * The regular allocator uses buddy scan only if the request len is power of
Alex Tomasc9de5602008-01-29 00:19:52 -0500142 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
143 * value of s_mb_order2_reqs can be tuned via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400144 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200145 * stripe size (sbi->s_stripe), we try to search for contiguous block in
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400146 * stripe size. This should result in better allocation on RAID setups. If
147 * not, we search in the specific group using bitmap for best extents. The
148 * tunable min_to_scan and max_to_scan control the behaviour here.
Alex Tomasc9de5602008-01-29 00:19:52 -0500149 * min_to_scan indicate how long the mballoc __must__ look for a best
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400150 * extent and max_to_scan indicates how long the mballoc __can__ look for a
Alex Tomasc9de5602008-01-29 00:19:52 -0500151 * best extent in the found extents. Searching for the blocks starts with
152 * the group specified as the goal value in allocation context via
153 * ac_g_ex. Each group is first checked based on the criteria whether it
154 * can used for allocation. ext4_mb_good_group explains how the groups are
155 * checked.
156 *
157 * Both the prealloc space are getting populated as above. So for the first
158 * request we will hit the buddy cache which will result in this prealloc
159 * space getting filled. The prealloc space is then later used for the
160 * subsequent request.
161 */
162
163/*
164 * mballoc operates on the following data:
165 * - on-disk bitmap
166 * - in-core buddy (actually includes buddy and bitmap)
167 * - preallocation descriptors (PAs)
168 *
169 * there are two types of preallocations:
170 * - inode
171 * assiged to specific inode and can be used for this inode only.
172 * it describes part of inode's space preallocated to specific
173 * physical blocks. any block from that preallocated can be used
174 * independent. the descriptor just tracks number of blocks left
175 * unused. so, before taking some block from descriptor, one must
176 * make sure corresponded logical block isn't allocated yet. this
177 * also means that freeing any block within descriptor's range
178 * must discard all preallocated blocks.
179 * - locality group
180 * assigned to specific locality group which does not translate to
181 * permanent set of inodes: inode can join and leave group. space
182 * from this type of preallocation can be used for any inode. thus
183 * it's consumed from the beginning to the end.
184 *
185 * relation between them can be expressed as:
186 * in-core buddy = on-disk bitmap + preallocation descriptors
187 *
188 * this mean blocks mballoc considers used are:
189 * - allocated blocks (persistent)
190 * - preallocated blocks (non-persistent)
191 *
192 * consistency in mballoc world means that at any time a block is either
193 * free or used in ALL structures. notice: "any time" should not be read
194 * literally -- time is discrete and delimited by locks.
195 *
196 * to keep it simple, we don't use block numbers, instead we count number of
197 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
198 *
199 * all operations can be expressed as:
200 * - init buddy: buddy = on-disk + PAs
201 * - new PA: buddy += N; PA = N
202 * - use inode PA: on-disk += N; PA -= N
203 * - discard inode PA buddy -= on-disk - PA; PA = 0
204 * - use locality group PA on-disk += N; PA -= N
205 * - discard locality group PA buddy -= PA; PA = 0
206 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
207 * is used in real operation because we can't know actual used
208 * bits from PA, only from on-disk bitmap
209 *
210 * if we follow this strict logic, then all operations above should be atomic.
211 * given some of them can block, we'd have to use something like semaphores
212 * killing performance on high-end SMP hardware. let's try to relax it using
213 * the following knowledge:
214 * 1) if buddy is referenced, it's already initialized
215 * 2) while block is used in buddy and the buddy is referenced,
216 * nobody can re-allocate that block
217 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
218 * bit set and PA claims same block, it's OK. IOW, one can set bit in
219 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
220 * block
221 *
222 * so, now we're building a concurrency table:
223 * - init buddy vs.
224 * - new PA
225 * blocks for PA are allocated in the buddy, buddy must be referenced
226 * until PA is linked to allocation group to avoid concurrent buddy init
227 * - use inode PA
228 * we need to make sure that either on-disk bitmap or PA has uptodate data
229 * given (3) we care that PA-=N operation doesn't interfere with init
230 * - discard inode PA
231 * the simplest way would be to have buddy initialized by the discard
232 * - use locality group PA
233 * again PA-=N must be serialized with init
234 * - discard locality group PA
235 * the simplest way would be to have buddy initialized by the discard
236 * - new PA vs.
237 * - use inode PA
238 * i_data_sem serializes them
239 * - discard inode PA
240 * discard process must wait until PA isn't used by another process
241 * - use locality group PA
242 * some mutex should serialize them
243 * - discard locality group PA
244 * discard process must wait until PA isn't used by another process
245 * - use inode PA
246 * - use inode PA
247 * i_data_sem or another mutex should serializes them
248 * - discard inode PA
249 * discard process must wait until PA isn't used by another process
250 * - use locality group PA
251 * nothing wrong here -- they're different PAs covering different blocks
252 * - discard locality group PA
253 * discard process must wait until PA isn't used by another process
254 *
255 * now we're ready to make few consequences:
256 * - PA is referenced and while it is no discard is possible
257 * - PA is referenced until block isn't marked in on-disk bitmap
258 * - PA changes only after on-disk bitmap
259 * - discard must not compete with init. either init is done before
260 * any discard or they're serialized somehow
261 * - buddy init as sum of on-disk bitmap and PAs is done atomically
262 *
263 * a special case when we've used PA to emptiness. no need to modify buddy
264 * in this case, but we should care about concurrent init
265 *
266 */
267
268 /*
269 * Logic in few words:
270 *
271 * - allocation:
272 * load group
273 * find blocks
274 * mark bits in on-disk bitmap
275 * release group
276 *
277 * - use preallocation:
278 * find proper PA (per-inode or group)
279 * load group
280 * mark bits in on-disk bitmap
281 * release group
282 * release PA
283 *
284 * - free:
285 * load group
286 * mark bits in on-disk bitmap
287 * release group
288 *
289 * - discard preallocations in group:
290 * mark PAs deleted
291 * move them onto local list
292 * load on-disk bitmap
293 * load group
294 * remove PA from object (inode or locality group)
295 * mark free blocks in-core
296 *
297 * - discard inode's preallocations:
298 */
299
300/*
301 * Locking rules
302 *
303 * Locks:
304 * - bitlock on a group (group)
305 * - object (inode/locality) (object)
306 * - per-pa lock (pa)
307 *
308 * Paths:
309 * - new pa
310 * object
311 * group
312 *
313 * - find and use pa:
314 * pa
315 *
316 * - release consumed pa:
317 * pa
318 * group
319 * object
320 *
321 * - generate in-core bitmap:
322 * group
323 * pa
324 *
325 * - discard all for given object (inode, locality group):
326 * object
327 * pa
328 * group
329 *
330 * - discard all for given group:
331 * group
332 * pa
333 * group
334 * object
335 *
336 */
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500337static struct kmem_cache *ext4_pspace_cachep;
338static struct kmem_cache *ext4_ac_cachep;
339static struct kmem_cache *ext4_free_ext_cachep;
340static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
341 ext4_group_t group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500342static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
343 ext4_group_t group);
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500344static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
345
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500346static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
347{
Alex Tomasc9de5602008-01-29 00:19:52 -0500348#if BITS_PER_LONG == 64
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500349 *bit += ((unsigned long) addr & 7UL) << 3;
350 addr = (void *) ((unsigned long) addr & ~7UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500351#elif BITS_PER_LONG == 32
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500352 *bit += ((unsigned long) addr & 3UL) << 3;
353 addr = (void *) ((unsigned long) addr & ~3UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500354#else
355#error "how many bits you are?!"
356#endif
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500357 return addr;
358}
Alex Tomasc9de5602008-01-29 00:19:52 -0500359
360static inline int mb_test_bit(int bit, void *addr)
361{
362 /*
363 * ext4_test_bit on architecture like powerpc
364 * needs unsigned long aligned address
365 */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500366 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500367 return ext4_test_bit(bit, addr);
368}
369
370static inline void mb_set_bit(int bit, void *addr)
371{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500372 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500373 ext4_set_bit(bit, addr);
374}
375
Alex Tomasc9de5602008-01-29 00:19:52 -0500376static inline void mb_clear_bit(int bit, void *addr)
377{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500378 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500379 ext4_clear_bit(bit, addr);
380}
381
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500382static inline int mb_find_next_zero_bit(void *addr, int max, int start)
383{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400384 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500385 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400386 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500387 start += fix;
388
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400389 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
390 if (ret > max)
391 return max;
392 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500393}
394
395static inline int mb_find_next_bit(void *addr, int max, int start)
396{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400397 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500398 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400399 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500400 start += fix;
401
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400402 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
403 if (ret > max)
404 return max;
405 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500406}
407
Alex Tomasc9de5602008-01-29 00:19:52 -0500408static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
409{
410 char *bb;
411
Alex Tomasc9de5602008-01-29 00:19:52 -0500412 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
413 BUG_ON(max == NULL);
414
415 if (order > e4b->bd_blkbits + 1) {
416 *max = 0;
417 return NULL;
418 }
419
420 /* at order 0 we see each particular block */
421 *max = 1 << (e4b->bd_blkbits + 3);
422 if (order == 0)
423 return EXT4_MB_BITMAP(e4b);
424
425 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
426 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
427
428 return bb;
429}
430
431#ifdef DOUBLE_CHECK
432static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
433 int first, int count)
434{
435 int i;
436 struct super_block *sb = e4b->bd_sb;
437
438 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
439 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400440 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500441 for (i = 0; i < count; i++) {
442 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
443 ext4_fsblk_t blocknr;
444 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
445 blocknr += first + i;
446 blocknr +=
447 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -0500448 ext4_grp_locked_error(sb, e4b->bd_group,
449 __func__, "double-free of inode"
Theodore Ts'oa9df9a42009-01-05 22:18:16 -0500450 " %lu's block %llu(bit %u in group %u)",
Alex Tomasc9de5602008-01-29 00:19:52 -0500451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
453 }
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
455 }
456}
457
458static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
459{
460 int i;
461
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
463 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400464 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
468 }
469}
470
471static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
472{
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
475 int i;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
Theodore Ts'oa9df9a42009-01-05 22:18:16 -0500480 printk(KERN_ERR "corruption in group %u "
Theodore Ts'o47760042008-09-08 23:00:52 -0400481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b->bd_group, i, i * 8, b1[i], b2[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500484 BUG();
485 }
486 }
487 }
488}
489
490#else
491static inline void mb_free_blocks_double(struct inode *inode,
492 struct ext4_buddy *e4b, int first, int count)
493{
494 return;
495}
496static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497 int first, int count)
498{
499 return;
500}
501static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
502{
503 return;
504}
505#endif
506
507#ifdef AGGRESSIVE_CHECK
508
509#define MB_CHECK_ASSERT(assert) \
510do { \
511 if (!(assert)) { \
512 printk(KERN_EMERG \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
515 BUG(); \
516 } \
517} while (0)
518
519static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520 const char *function, int line)
521{
522 struct super_block *sb = e4b->bd_sb;
523 int order = e4b->bd_blkbits + 1;
524 int max;
525 int max2;
526 int i;
527 int j;
528 int k;
529 int count;
530 struct ext4_group_info *grp;
531 int fragments = 0;
532 int fstart;
533 struct list_head *cur;
534 void *buddy;
535 void *buddy2;
536
Alex Tomasc9de5602008-01-29 00:19:52 -0500537 {
538 static int mb_check_counter;
539 if (mb_check_counter++ % 100 != 0)
540 return 0;
541 }
542
543 while (order > 1) {
544 buddy = mb_find_buddy(e4b, order, &max);
545 MB_CHECK_ASSERT(buddy);
546 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547 MB_CHECK_ASSERT(buddy2);
548 MB_CHECK_ASSERT(buddy != buddy2);
549 MB_CHECK_ASSERT(max * 2 == max2);
550
551 count = 0;
552 for (i = 0; i < max; i++) {
553
554 if (mb_test_bit(i, buddy)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i << 1, buddy2)) {
557 MB_CHECK_ASSERT(
558 mb_test_bit((i<<1)+1, buddy2));
559 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
560 MB_CHECK_ASSERT(
561 mb_test_bit(i << 1, buddy2));
562 }
563 continue;
564 }
565
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
569
570 for (j = 0; j < (1 << order); j++) {
571 k = (i * (1 << order)) + j;
572 MB_CHECK_ASSERT(
573 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
574 }
575 count++;
576 }
577 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
578 order--;
579 }
580
581 fstart = -1;
582 buddy = mb_find_buddy(e4b, 0, &max);
583 for (i = 0; i < max; i++) {
584 if (!mb_test_bit(i, buddy)) {
585 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
586 if (fstart == -1) {
587 fragments++;
588 fstart = i;
589 }
590 continue;
591 }
592 fstart = -1;
593 /* check used bits only */
594 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595 buddy2 = mb_find_buddy(e4b, j, &max2);
596 k = i >> j;
597 MB_CHECK_ASSERT(k < max2);
598 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
599 }
600 }
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
603
604 grp = ext4_get_group_info(sb, e4b->bd_group);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 list_for_each(cur, &grp->bb_prealloc_list) {
607 ext4_group_t groupnr;
608 struct ext4_prealloc_space *pa;
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400609 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
Alex Tomasc9de5602008-01-29 00:19:52 -0500611 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400612 for (i = 0; i < pa->pa_len; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500613 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
614 }
615 return 0;
616}
617#undef MB_CHECK_ASSERT
618#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
Harvey Harrison46e665e2008-04-17 10:38:59 -0400619 __FILE__, __func__, __LINE__)
Alex Tomasc9de5602008-01-29 00:19:52 -0500620#else
621#define mb_check_buddy(e4b)
622#endif
623
624/* FIXME!! need more doc */
625static void ext4_mb_mark_free_simple(struct super_block *sb,
Eric Sandeena36b4492009-08-25 22:36:45 -0400626 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
Alex Tomasc9de5602008-01-29 00:19:52 -0500627 struct ext4_group_info *grp)
628{
629 struct ext4_sb_info *sbi = EXT4_SB(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400630 ext4_grpblk_t min;
631 ext4_grpblk_t max;
632 ext4_grpblk_t chunk;
Alex Tomasc9de5602008-01-29 00:19:52 -0500633 unsigned short border;
634
Valerie Clementb73fce62008-02-15 13:48:51 -0500635 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
Alex Tomasc9de5602008-01-29 00:19:52 -0500636
637 border = 2 << sb->s_blocksize_bits;
638
639 while (len > 0) {
640 /* find how many blocks can be covered since this position */
641 max = ffs(first | border) - 1;
642
643 /* find how many blocks of power 2 we need to mark */
644 min = fls(len) - 1;
645
646 if (max < min)
647 min = max;
648 chunk = 1 << min;
649
650 /* mark multiblock chunks only */
651 grp->bb_counters[min]++;
652 if (min > 0)
653 mb_clear_bit(first >> min,
654 buddy + sbi->s_mb_offsets[min]);
655
656 len -= chunk;
657 first += chunk;
658 }
659}
660
Eric Sandeen089ceec2009-07-05 22:17:31 -0400661static noinline_for_stack
662void ext4_mb_generate_buddy(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -0500663 void *buddy, void *bitmap, ext4_group_t group)
664{
665 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
Eric Sandeena36b4492009-08-25 22:36:45 -0400666 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
667 ext4_grpblk_t i = 0;
668 ext4_grpblk_t first;
669 ext4_grpblk_t len;
Alex Tomasc9de5602008-01-29 00:19:52 -0500670 unsigned free = 0;
671 unsigned fragments = 0;
672 unsigned long long period = get_cycles();
673
674 /* initialize buddy from bitmap which is aggregation
675 * of on-disk bitmap and preallocations */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500676 i = mb_find_next_zero_bit(bitmap, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -0500677 grp->bb_first_free = i;
678 while (i < max) {
679 fragments++;
680 first = i;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500681 i = mb_find_next_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500682 len = i - first;
683 free += len;
684 if (len > 1)
685 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
686 else
687 grp->bb_counters[0]++;
688 if (i < max)
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500689 i = mb_find_next_zero_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500690 }
691 grp->bb_fragments = fragments;
692
693 if (free != grp->bb_free) {
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -0500694 ext4_grp_locked_error(sb, group, __func__,
Theodore Ts'oa9df9a42009-01-05 22:18:16 -0500695 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
Alex Tomasc9de5602008-01-29 00:19:52 -0500696 group, free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500697 /*
698 * If we intent to continue, we consider group descritor
699 * corrupt and update bb_free using bitmap value
700 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500701 grp->bb_free = free;
702 }
703
704 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
705
706 period = get_cycles() - period;
707 spin_lock(&EXT4_SB(sb)->s_bal_lock);
708 EXT4_SB(sb)->s_mb_buddies_generated++;
709 EXT4_SB(sb)->s_mb_generation_time += period;
710 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
711}
712
713/* The buddy information is attached the buddy cache inode
714 * for convenience. The information regarding each group
715 * is loaded via ext4_mb_load_buddy. The information involve
716 * block bitmap and buddy information. The information are
717 * stored in the inode as
718 *
719 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500720 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500721 *
722 *
723 * one block each for bitmap and buddy information.
724 * So for each group we take up 2 blocks. A page can
725 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
726 * So it can have information regarding groups_per_page which
727 * is blocks_per_page/2
728 */
729
730static int ext4_mb_init_cache(struct page *page, char *incore)
731{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400732 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500733 int blocksize;
734 int blocks_per_page;
735 int groups_per_page;
736 int err = 0;
737 int i;
738 ext4_group_t first_group;
739 int first_block;
740 struct super_block *sb;
741 struct buffer_head *bhs;
742 struct buffer_head **bh;
743 struct inode *inode;
744 char *data;
745 char *bitmap;
746
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400747 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500748
749 inode = page->mapping->host;
750 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400751 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500752 blocksize = 1 << inode->i_blkbits;
753 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
754
755 groups_per_page = blocks_per_page >> 1;
756 if (groups_per_page == 0)
757 groups_per_page = 1;
758
759 /* allocate buffer_heads to read bitmaps */
760 if (groups_per_page > 1) {
761 err = -ENOMEM;
762 i = sizeof(struct buffer_head *) * groups_per_page;
763 bh = kzalloc(i, GFP_NOFS);
764 if (bh == NULL)
765 goto out;
766 } else
767 bh = &bhs;
768
769 first_group = page->index * blocks_per_page / 2;
770
771 /* read all groups the page covers into the cache */
772 for (i = 0; i < groups_per_page; i++) {
773 struct ext4_group_desc *desc;
774
Theodore Ts'o8df96752009-05-01 08:50:38 -0400775 if (first_group + i >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500776 break;
777
778 err = -EIO;
779 desc = ext4_get_group_desc(sb, first_group + i, NULL);
780 if (desc == NULL)
781 goto out;
782
783 err = -ENOMEM;
784 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
785 if (bh[i] == NULL)
786 goto out;
787
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500788 if (bitmap_uptodate(bh[i]))
Alex Tomasc9de5602008-01-29 00:19:52 -0500789 continue;
790
Frederic Bohec806e682008-10-10 08:09:18 -0400791 lock_buffer(bh[i]);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500792 if (bitmap_uptodate(bh[i])) {
793 unlock_buffer(bh[i]);
794 continue;
795 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400796 ext4_lock_group(sb, first_group + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500797 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
798 ext4_init_block_bitmap(sb, bh[i],
799 first_group + i, desc);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500800 set_bitmap_uptodate(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500801 set_buffer_uptodate(bh[i]);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400802 ext4_unlock_group(sb, first_group + i);
Aneesh Kumar K.V3300bed2009-01-03 22:33:39 -0500803 unlock_buffer(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500804 continue;
805 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400806 ext4_unlock_group(sb, first_group + i);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500807 if (buffer_uptodate(bh[i])) {
808 /*
809 * if not uninit if bh is uptodate,
810 * bitmap is also uptodate
811 */
812 set_bitmap_uptodate(bh[i]);
813 unlock_buffer(bh[i]);
814 continue;
815 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500816 get_bh(bh[i]);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500817 /*
818 * submit the buffer_head for read. We can
819 * safely mark the bitmap as uptodate now.
820 * We do it here so the bitmap uptodate bit
821 * get set with buffer lock held.
822 */
823 set_bitmap_uptodate(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500824 bh[i]->b_end_io = end_buffer_read_sync;
825 submit_bh(READ, bh[i]);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400826 mb_debug(1, "read bitmap for group %u\n", first_group + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500827 }
828
829 /* wait for I/O completion */
830 for (i = 0; i < groups_per_page && bh[i]; i++)
831 wait_on_buffer(bh[i]);
832
833 err = -EIO;
834 for (i = 0; i < groups_per_page && bh[i]; i++)
835 if (!buffer_uptodate(bh[i]))
836 goto out;
837
Mingming Cao31b481d2008-07-11 19:27:31 -0400838 err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -0500839 first_block = page->index * blocks_per_page;
Aneesh Kumar K.V29eaf022009-01-05 21:48:56 -0500840 /* init the page */
841 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
Alex Tomasc9de5602008-01-29 00:19:52 -0500842 for (i = 0; i < blocks_per_page; i++) {
843 int group;
844 struct ext4_group_info *grinfo;
845
846 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400847 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500848 break;
849
850 /*
851 * data carry information regarding this
852 * particular group in the format specified
853 * above
854 *
855 */
856 data = page_address(page) + (i * blocksize);
857 bitmap = bh[group - first_group]->b_data;
858
859 /*
860 * We place the buddy block and bitmap block
861 * close together
862 */
863 if ((first_block + i) & 1) {
864 /* this is block of buddy */
865 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400866 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500867 group, page->index, i * blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500868 grinfo = ext4_get_group_info(sb, group);
869 grinfo->bb_fragments = 0;
870 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400871 sizeof(*grinfo->bb_counters) *
872 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500873 /*
874 * incore got set to the group block bitmap below
875 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500876 ext4_lock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500877 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500878 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500879 incore = NULL;
880 } else {
881 /* this is block of bitmap */
882 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400883 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500884 group, page->index, i * blocksize);
885
886 /* see comments in ext4_mb_put_pa() */
887 ext4_lock_group(sb, group);
888 memcpy(data, bitmap, blocksize);
889
890 /* mark all preallocated blks used in in-core bitmap */
891 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500892 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500893 ext4_unlock_group(sb, group);
894
895 /* set incore so that the buddy information can be
896 * generated using this
897 */
898 incore = data;
899 }
900 }
901 SetPageUptodate(page);
902
903out:
904 if (bh) {
905 for (i = 0; i < groups_per_page && bh[i]; i++)
906 brelse(bh[i]);
907 if (bh != &bhs)
908 kfree(bh);
909 }
910 return err;
911}
912
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -0400913static noinline_for_stack
914int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
915{
916
917 int ret = 0;
918 void *bitmap;
919 int blocks_per_page;
920 int block, pnum, poff;
921 int num_grp_locked = 0;
922 struct ext4_group_info *this_grp;
923 struct ext4_sb_info *sbi = EXT4_SB(sb);
924 struct inode *inode = sbi->s_buddy_cache;
925 struct page *page = NULL, *bitmap_page = NULL;
926
927 mb_debug(1, "init group %u\n", group);
928 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
929 this_grp = ext4_get_group_info(sb, group);
930 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -0400931 * This ensures that we don't reinit the buddy cache
932 * page which map to the group from which we are already
933 * allocating. If we are looking at the buddy cache we would
934 * have taken a reference using ext4_mb_load_buddy and that
935 * would have taken the alloc_sem lock.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -0400936 */
937 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
938 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
939 /*
940 * somebody initialized the group
941 * return without doing anything
942 */
943 ret = 0;
944 goto err;
945 }
946 /*
947 * the buddy cache inode stores the block bitmap
948 * and buddy information in consecutive blocks.
949 * So for each group we need two blocks.
950 */
951 block = group * 2;
952 pnum = block / blocks_per_page;
953 poff = block % blocks_per_page;
954 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
955 if (page) {
956 BUG_ON(page->mapping != inode->i_mapping);
957 ret = ext4_mb_init_cache(page, NULL);
958 if (ret) {
959 unlock_page(page);
960 goto err;
961 }
962 unlock_page(page);
963 }
964 if (page == NULL || !PageUptodate(page)) {
965 ret = -EIO;
966 goto err;
967 }
968 mark_page_accessed(page);
969 bitmap_page = page;
970 bitmap = page_address(page) + (poff * sb->s_blocksize);
971
972 /* init buddy cache */
973 block++;
974 pnum = block / blocks_per_page;
975 poff = block % blocks_per_page;
976 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
977 if (page == bitmap_page) {
978 /*
979 * If both the bitmap and buddy are in
980 * the same page we don't need to force
981 * init the buddy
982 */
983 unlock_page(page);
984 } else if (page) {
985 BUG_ON(page->mapping != inode->i_mapping);
986 ret = ext4_mb_init_cache(page, bitmap);
987 if (ret) {
988 unlock_page(page);
989 goto err;
990 }
991 unlock_page(page);
992 }
993 if (page == NULL || !PageUptodate(page)) {
994 ret = -EIO;
995 goto err;
996 }
997 mark_page_accessed(page);
998err:
999 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1000 if (bitmap_page)
1001 page_cache_release(bitmap_page);
1002 if (page)
1003 page_cache_release(page);
1004 return ret;
1005}
1006
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001007static noinline_for_stack int
1008ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1009 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001010{
Alex Tomasc9de5602008-01-29 00:19:52 -05001011 int blocks_per_page;
1012 int block;
1013 int pnum;
1014 int poff;
1015 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001016 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001017 struct ext4_group_info *grp;
1018 struct ext4_sb_info *sbi = EXT4_SB(sb);
1019 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001020
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001021 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001022
1023 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001024 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001025
1026 e4b->bd_blkbits = sb->s_blocksize_bits;
1027 e4b->bd_info = ext4_get_group_info(sb, group);
1028 e4b->bd_sb = sb;
1029 e4b->bd_group = group;
1030 e4b->bd_buddy_page = NULL;
1031 e4b->bd_bitmap_page = NULL;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001032 e4b->alloc_semp = &grp->alloc_sem;
1033
1034 /* Take the read lock on the group alloc
1035 * sem. This would make sure a parallel
1036 * ext4_mb_init_group happening on other
1037 * groups mapped by the page is blocked
1038 * till we are done with allocation
1039 */
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001040repeat_load_buddy:
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001041 down_read(e4b->alloc_semp);
Alex Tomasc9de5602008-01-29 00:19:52 -05001042
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001043 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1044 /* we need to check for group need init flag
1045 * with alloc_semp held so that we can be sure
1046 * that new blocks didn't get added to the group
1047 * when we are loading the buddy cache
1048 */
1049 up_read(e4b->alloc_semp);
1050 /*
1051 * we need full data about the group
1052 * to make a good selection
1053 */
1054 ret = ext4_mb_init_group(sb, group);
1055 if (ret)
1056 return ret;
1057 goto repeat_load_buddy;
1058 }
1059
Alex Tomasc9de5602008-01-29 00:19:52 -05001060 /*
1061 * the buddy cache inode stores the block bitmap
1062 * and buddy information in consecutive blocks.
1063 * So for each group we need two blocks.
1064 */
1065 block = group * 2;
1066 pnum = block / blocks_per_page;
1067 poff = block % blocks_per_page;
1068
1069 /* we could use find_or_create_page(), but it locks page
1070 * what we'd like to avoid in fast path ... */
1071 page = find_get_page(inode->i_mapping, pnum);
1072 if (page == NULL || !PageUptodate(page)) {
1073 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001074 /*
1075 * drop the page reference and try
1076 * to get the page with lock. If we
1077 * are not uptodate that implies
1078 * somebody just created the page but
1079 * is yet to initialize the same. So
1080 * wait for it to initialize.
1081 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001082 page_cache_release(page);
1083 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1084 if (page) {
1085 BUG_ON(page->mapping != inode->i_mapping);
1086 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001087 ret = ext4_mb_init_cache(page, NULL);
1088 if (ret) {
1089 unlock_page(page);
1090 goto err;
1091 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001092 mb_cmp_bitmaps(e4b, page_address(page) +
1093 (poff * sb->s_blocksize));
1094 }
1095 unlock_page(page);
1096 }
1097 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001098 if (page == NULL || !PageUptodate(page)) {
1099 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001100 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001101 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001102 e4b->bd_bitmap_page = page;
1103 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1104 mark_page_accessed(page);
1105
1106 block++;
1107 pnum = block / blocks_per_page;
1108 poff = block % blocks_per_page;
1109
1110 page = find_get_page(inode->i_mapping, pnum);
1111 if (page == NULL || !PageUptodate(page)) {
1112 if (page)
1113 page_cache_release(page);
1114 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1115 if (page) {
1116 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001117 if (!PageUptodate(page)) {
1118 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1119 if (ret) {
1120 unlock_page(page);
1121 goto err;
1122 }
1123 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001124 unlock_page(page);
1125 }
1126 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001127 if (page == NULL || !PageUptodate(page)) {
1128 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001129 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001130 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001131 e4b->bd_buddy_page = page;
1132 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1133 mark_page_accessed(page);
1134
1135 BUG_ON(e4b->bd_bitmap_page == NULL);
1136 BUG_ON(e4b->bd_buddy_page == NULL);
1137
1138 return 0;
1139
1140err:
1141 if (e4b->bd_bitmap_page)
1142 page_cache_release(e4b->bd_bitmap_page);
1143 if (e4b->bd_buddy_page)
1144 page_cache_release(e4b->bd_buddy_page);
1145 e4b->bd_buddy = NULL;
1146 e4b->bd_bitmap = NULL;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001147
1148 /* Done with the buddy cache */
1149 up_read(e4b->alloc_semp);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001150 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001151}
1152
1153static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1154{
1155 if (e4b->bd_bitmap_page)
1156 page_cache_release(e4b->bd_bitmap_page);
1157 if (e4b->bd_buddy_page)
1158 page_cache_release(e4b->bd_buddy_page);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001159 /* Done with the buddy cache */
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05001160 if (e4b->alloc_semp)
1161 up_read(e4b->alloc_semp);
Alex Tomasc9de5602008-01-29 00:19:52 -05001162}
1163
1164
1165static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1166{
1167 int order = 1;
1168 void *bb;
1169
1170 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1171 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1172
1173 bb = EXT4_MB_BUDDY(e4b);
1174 while (order <= e4b->bd_blkbits + 1) {
1175 block = block >> 1;
1176 if (!mb_test_bit(block, bb)) {
1177 /* this block is part of buddy of order 'order' */
1178 return order;
1179 }
1180 bb += 1 << (e4b->bd_blkbits - order);
1181 order++;
1182 }
1183 return 0;
1184}
1185
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001186static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001187{
1188 __u32 *addr;
1189
1190 len = cur + len;
1191 while (cur < len) {
1192 if ((cur & 31) == 0 && (len - cur) >= 32) {
1193 /* fast path: clear whole word at once */
1194 addr = bm + (cur >> 3);
1195 *addr = 0;
1196 cur += 32;
1197 continue;
1198 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001199 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001200 cur++;
1201 }
1202}
1203
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001204static void mb_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001205{
1206 __u32 *addr;
1207
1208 len = cur + len;
1209 while (cur < len) {
1210 if ((cur & 31) == 0 && (len - cur) >= 32) {
1211 /* fast path: set whole word at once */
1212 addr = bm + (cur >> 3);
1213 *addr = 0xffffffff;
1214 cur += 32;
1215 continue;
1216 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001217 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001218 cur++;
1219 }
1220}
1221
Shen Feng7e5a8cd2008-07-13 21:03:31 -04001222static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
Alex Tomasc9de5602008-01-29 00:19:52 -05001223 int first, int count)
1224{
1225 int block = 0;
1226 int max = 0;
1227 int order;
1228 void *buddy;
1229 void *buddy2;
1230 struct super_block *sb = e4b->bd_sb;
1231
1232 BUG_ON(first + count > (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001233 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001234 mb_check_buddy(e4b);
1235 mb_free_blocks_double(inode, e4b, first, count);
1236
1237 e4b->bd_info->bb_free += count;
1238 if (first < e4b->bd_info->bb_first_free)
1239 e4b->bd_info->bb_first_free = first;
1240
1241 /* let's maintain fragments counter */
1242 if (first != 0)
1243 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1244 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1245 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1246 if (block && max)
1247 e4b->bd_info->bb_fragments--;
1248 else if (!block && !max)
1249 e4b->bd_info->bb_fragments++;
1250
1251 /* let's maintain buddy itself */
1252 while (count-- > 0) {
1253 block = first++;
1254 order = 0;
1255
1256 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1257 ext4_fsblk_t blocknr;
1258 blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
1259 blocknr += block;
1260 blocknr +=
1261 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05001262 ext4_grp_locked_error(sb, e4b->bd_group,
1263 __func__, "double-free of inode"
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05001264 " %lu's block %llu(bit %u in group %u)",
Alex Tomasc9de5602008-01-29 00:19:52 -05001265 inode ? inode->i_ino : 0, blocknr, block,
1266 e4b->bd_group);
1267 }
1268 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1269 e4b->bd_info->bb_counters[order]++;
1270
1271 /* start of the buddy */
1272 buddy = mb_find_buddy(e4b, order, &max);
1273
1274 do {
1275 block &= ~1UL;
1276 if (mb_test_bit(block, buddy) ||
1277 mb_test_bit(block + 1, buddy))
1278 break;
1279
1280 /* both the buddies are free, try to coalesce them */
1281 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1282
1283 if (!buddy2)
1284 break;
1285
1286 if (order > 0) {
1287 /* for special purposes, we don't set
1288 * free bits in bitmap */
1289 mb_set_bit(block, buddy);
1290 mb_set_bit(block + 1, buddy);
1291 }
1292 e4b->bd_info->bb_counters[order]--;
1293 e4b->bd_info->bb_counters[order]--;
1294
1295 block = block >> 1;
1296 order++;
1297 e4b->bd_info->bb_counters[order]++;
1298
1299 mb_clear_bit(block, buddy2);
1300 buddy = buddy2;
1301 } while (1);
1302 }
1303 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001304}
1305
1306static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1307 int needed, struct ext4_free_extent *ex)
1308{
1309 int next = block;
1310 int max;
1311 int ord;
1312 void *buddy;
1313
Vincent Minetbc8e6742009-05-15 08:33:18 -04001314 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001315 BUG_ON(ex == NULL);
1316
1317 buddy = mb_find_buddy(e4b, order, &max);
1318 BUG_ON(buddy == NULL);
1319 BUG_ON(block >= max);
1320 if (mb_test_bit(block, buddy)) {
1321 ex->fe_len = 0;
1322 ex->fe_start = 0;
1323 ex->fe_group = 0;
1324 return 0;
1325 }
1326
1327 /* FIXME dorp order completely ? */
1328 if (likely(order == 0)) {
1329 /* find actual order */
1330 order = mb_find_order_for_block(e4b, block);
1331 block = block >> order;
1332 }
1333
1334 ex->fe_len = 1 << order;
1335 ex->fe_start = block << order;
1336 ex->fe_group = e4b->bd_group;
1337
1338 /* calc difference from given start */
1339 next = next - ex->fe_start;
1340 ex->fe_len -= next;
1341 ex->fe_start += next;
1342
1343 while (needed > ex->fe_len &&
1344 (buddy = mb_find_buddy(e4b, order, &max))) {
1345
1346 if (block + 1 >= max)
1347 break;
1348
1349 next = (block + 1) * (1 << order);
1350 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1351 break;
1352
1353 ord = mb_find_order_for_block(e4b, next);
1354
1355 order = ord;
1356 block = next >> order;
1357 ex->fe_len += 1 << order;
1358 }
1359
1360 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1361 return ex->fe_len;
1362}
1363
1364static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1365{
1366 int ord;
1367 int mlen = 0;
1368 int max = 0;
1369 int cur;
1370 int start = ex->fe_start;
1371 int len = ex->fe_len;
1372 unsigned ret = 0;
1373 int len0 = len;
1374 void *buddy;
1375
1376 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1377 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001378 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001379 mb_check_buddy(e4b);
1380 mb_mark_used_double(e4b, start, len);
1381
1382 e4b->bd_info->bb_free -= len;
1383 if (e4b->bd_info->bb_first_free == start)
1384 e4b->bd_info->bb_first_free += len;
1385
1386 /* let's maintain fragments counter */
1387 if (start != 0)
1388 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1389 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1390 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1391 if (mlen && max)
1392 e4b->bd_info->bb_fragments++;
1393 else if (!mlen && !max)
1394 e4b->bd_info->bb_fragments--;
1395
1396 /* let's maintain buddy itself */
1397 while (len) {
1398 ord = mb_find_order_for_block(e4b, start);
1399
1400 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1401 /* the whole chunk may be allocated at once! */
1402 mlen = 1 << ord;
1403 buddy = mb_find_buddy(e4b, ord, &max);
1404 BUG_ON((start >> ord) >= max);
1405 mb_set_bit(start >> ord, buddy);
1406 e4b->bd_info->bb_counters[ord]--;
1407 start += mlen;
1408 len -= mlen;
1409 BUG_ON(len < 0);
1410 continue;
1411 }
1412
1413 /* store for history */
1414 if (ret == 0)
1415 ret = len | (ord << 16);
1416
1417 /* we have to split large buddy */
1418 BUG_ON(ord <= 0);
1419 buddy = mb_find_buddy(e4b, ord, &max);
1420 mb_set_bit(start >> ord, buddy);
1421 e4b->bd_info->bb_counters[ord]--;
1422
1423 ord--;
1424 cur = (start >> ord) & ~1U;
1425 buddy = mb_find_buddy(e4b, ord, &max);
1426 mb_clear_bit(cur, buddy);
1427 mb_clear_bit(cur + 1, buddy);
1428 e4b->bd_info->bb_counters[ord]++;
1429 e4b->bd_info->bb_counters[ord]++;
1430 }
1431
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001432 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001433 mb_check_buddy(e4b);
1434
1435 return ret;
1436}
1437
1438/*
1439 * Must be called under group lock!
1440 */
1441static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1442 struct ext4_buddy *e4b)
1443{
1444 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1445 int ret;
1446
1447 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1448 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1449
1450 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1451 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1452 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1453
1454 /* preallocation can change ac_b_ex, thus we store actually
1455 * allocated blocks for history */
1456 ac->ac_f_ex = ac->ac_b_ex;
1457
1458 ac->ac_status = AC_STATUS_FOUND;
1459 ac->ac_tail = ret & 0xffff;
1460 ac->ac_buddy = ret >> 16;
1461
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001462 /*
1463 * take the page reference. We want the page to be pinned
1464 * so that we don't get a ext4_mb_init_cache_call for this
1465 * group until we update the bitmap. That would mean we
1466 * double allocate blocks. The reference is dropped
1467 * in ext4_mb_release_context
1468 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001469 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1470 get_page(ac->ac_bitmap_page);
1471 ac->ac_buddy_page = e4b->bd_buddy_page;
1472 get_page(ac->ac_buddy_page);
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05001473 /* on allocation we use ac to track the held semaphore */
1474 ac->alloc_semp = e4b->alloc_semp;
1475 e4b->alloc_semp = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05001476 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001477 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001478 spin_lock(&sbi->s_md_lock);
1479 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1480 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1481 spin_unlock(&sbi->s_md_lock);
1482 }
1483}
1484
1485/*
1486 * regular allocator, for general purposes allocation
1487 */
1488
1489static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1490 struct ext4_buddy *e4b,
1491 int finish_group)
1492{
1493 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1494 struct ext4_free_extent *bex = &ac->ac_b_ex;
1495 struct ext4_free_extent *gex = &ac->ac_g_ex;
1496 struct ext4_free_extent ex;
1497 int max;
1498
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001499 if (ac->ac_status == AC_STATUS_FOUND)
1500 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001501 /*
1502 * We don't want to scan for a whole year
1503 */
1504 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1505 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1506 ac->ac_status = AC_STATUS_BREAK;
1507 return;
1508 }
1509
1510 /*
1511 * Haven't found good chunk so far, let's continue
1512 */
1513 if (bex->fe_len < gex->fe_len)
1514 return;
1515
1516 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1517 && bex->fe_group == e4b->bd_group) {
1518 /* recheck chunk's availability - we don't know
1519 * when it was found (within this lock-unlock
1520 * period or not) */
1521 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1522 if (max >= gex->fe_len) {
1523 ext4_mb_use_best_found(ac, e4b);
1524 return;
1525 }
1526 }
1527}
1528
1529/*
1530 * The routine checks whether found extent is good enough. If it is,
1531 * then the extent gets marked used and flag is set to the context
1532 * to stop scanning. Otherwise, the extent is compared with the
1533 * previous found extent and if new one is better, then it's stored
1534 * in the context. Later, the best found extent will be used, if
1535 * mballoc can't find good enough extent.
1536 *
1537 * FIXME: real allocation policy is to be designed yet!
1538 */
1539static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1540 struct ext4_free_extent *ex,
1541 struct ext4_buddy *e4b)
1542{
1543 struct ext4_free_extent *bex = &ac->ac_b_ex;
1544 struct ext4_free_extent *gex = &ac->ac_g_ex;
1545
1546 BUG_ON(ex->fe_len <= 0);
Eric Sandeen8d03c7a2009-03-14 11:51:46 -04001547 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001548 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1549 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1550
1551 ac->ac_found++;
1552
1553 /*
1554 * The special case - take what you catch first
1555 */
1556 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1557 *bex = *ex;
1558 ext4_mb_use_best_found(ac, e4b);
1559 return;
1560 }
1561
1562 /*
1563 * Let's check whether the chuck is good enough
1564 */
1565 if (ex->fe_len == gex->fe_len) {
1566 *bex = *ex;
1567 ext4_mb_use_best_found(ac, e4b);
1568 return;
1569 }
1570
1571 /*
1572 * If this is first found extent, just store it in the context
1573 */
1574 if (bex->fe_len == 0) {
1575 *bex = *ex;
1576 return;
1577 }
1578
1579 /*
1580 * If new found extent is better, store it in the context
1581 */
1582 if (bex->fe_len < gex->fe_len) {
1583 /* if the request isn't satisfied, any found extent
1584 * larger than previous best one is better */
1585 if (ex->fe_len > bex->fe_len)
1586 *bex = *ex;
1587 } else if (ex->fe_len > gex->fe_len) {
1588 /* if the request is satisfied, then we try to find
1589 * an extent that still satisfy the request, but is
1590 * smaller than previous one */
1591 if (ex->fe_len < bex->fe_len)
1592 *bex = *ex;
1593 }
1594
1595 ext4_mb_check_limits(ac, e4b, 0);
1596}
1597
Eric Sandeen089ceec2009-07-05 22:17:31 -04001598static noinline_for_stack
1599int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001600 struct ext4_buddy *e4b)
1601{
1602 struct ext4_free_extent ex = ac->ac_b_ex;
1603 ext4_group_t group = ex.fe_group;
1604 int max;
1605 int err;
1606
1607 BUG_ON(ex.fe_len <= 0);
1608 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1609 if (err)
1610 return err;
1611
1612 ext4_lock_group(ac->ac_sb, group);
1613 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1614
1615 if (max > 0) {
1616 ac->ac_b_ex = ex;
1617 ext4_mb_use_best_found(ac, e4b);
1618 }
1619
1620 ext4_unlock_group(ac->ac_sb, group);
1621 ext4_mb_release_desc(e4b);
1622
1623 return 0;
1624}
1625
Eric Sandeen089ceec2009-07-05 22:17:31 -04001626static noinline_for_stack
1627int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001628 struct ext4_buddy *e4b)
1629{
1630 ext4_group_t group = ac->ac_g_ex.fe_group;
1631 int max;
1632 int err;
1633 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1634 struct ext4_super_block *es = sbi->s_es;
1635 struct ext4_free_extent ex;
1636
1637 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1638 return 0;
1639
1640 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1641 if (err)
1642 return err;
1643
1644 ext4_lock_group(ac->ac_sb, group);
1645 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1646 ac->ac_g_ex.fe_len, &ex);
1647
1648 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1649 ext4_fsblk_t start;
1650
1651 start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
1652 ex.fe_start + le32_to_cpu(es->s_first_data_block);
1653 /* use do_div to get remainder (would be 64-bit modulo) */
1654 if (do_div(start, sbi->s_stripe) == 0) {
1655 ac->ac_found++;
1656 ac->ac_b_ex = ex;
1657 ext4_mb_use_best_found(ac, e4b);
1658 }
1659 } else if (max >= ac->ac_g_ex.fe_len) {
1660 BUG_ON(ex.fe_len <= 0);
1661 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1662 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1663 ac->ac_found++;
1664 ac->ac_b_ex = ex;
1665 ext4_mb_use_best_found(ac, e4b);
1666 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1667 /* Sometimes, caller may want to merge even small
1668 * number of blocks to an existing extent */
1669 BUG_ON(ex.fe_len <= 0);
1670 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1671 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1672 ac->ac_found++;
1673 ac->ac_b_ex = ex;
1674 ext4_mb_use_best_found(ac, e4b);
1675 }
1676 ext4_unlock_group(ac->ac_sb, group);
1677 ext4_mb_release_desc(e4b);
1678
1679 return 0;
1680}
1681
1682/*
1683 * The routine scans buddy structures (not bitmap!) from given order
1684 * to max order and tries to find big enough chunk to satisfy the req
1685 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001686static noinline_for_stack
1687void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001688 struct ext4_buddy *e4b)
1689{
1690 struct super_block *sb = ac->ac_sb;
1691 struct ext4_group_info *grp = e4b->bd_info;
1692 void *buddy;
1693 int i;
1694 int k;
1695 int max;
1696
1697 BUG_ON(ac->ac_2order <= 0);
1698 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1699 if (grp->bb_counters[i] == 0)
1700 continue;
1701
1702 buddy = mb_find_buddy(e4b, i, &max);
1703 BUG_ON(buddy == NULL);
1704
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001705 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001706 BUG_ON(k >= max);
1707
1708 ac->ac_found++;
1709
1710 ac->ac_b_ex.fe_len = 1 << i;
1711 ac->ac_b_ex.fe_start = k << i;
1712 ac->ac_b_ex.fe_group = e4b->bd_group;
1713
1714 ext4_mb_use_best_found(ac, e4b);
1715
1716 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1717
1718 if (EXT4_SB(sb)->s_mb_stats)
1719 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1720
1721 break;
1722 }
1723}
1724
1725/*
1726 * The routine scans the group and measures all found extents.
1727 * In order to optimize scanning, caller must pass number of
1728 * free blocks in the group, so the routine can know upper limit.
1729 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001730static noinline_for_stack
1731void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001732 struct ext4_buddy *e4b)
1733{
1734 struct super_block *sb = ac->ac_sb;
1735 void *bitmap = EXT4_MB_BITMAP(e4b);
1736 struct ext4_free_extent ex;
1737 int i;
1738 int free;
1739
1740 free = e4b->bd_info->bb_free;
1741 BUG_ON(free <= 0);
1742
1743 i = e4b->bd_info->bb_first_free;
1744
1745 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001746 i = mb_find_next_zero_bit(bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05001747 EXT4_BLOCKS_PER_GROUP(sb), i);
1748 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001749 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001750 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001751 * free blocks even though group info says we
1752 * we have free blocks
1753 */
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05001754 ext4_grp_locked_error(sb, e4b->bd_group,
1755 __func__, "%d free blocks as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001756 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001757 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001758 break;
1759 }
1760
1761 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1762 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001763 if (free < ex.fe_len) {
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05001764 ext4_grp_locked_error(sb, e4b->bd_group,
1765 __func__, "%d free blocks as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001766 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001767 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001768 /*
1769 * The number of free blocks differs. This mostly
1770 * indicate that the bitmap is corrupt. So exit
1771 * without claiming the space.
1772 */
1773 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001774 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001775
1776 ext4_mb_measure_extent(ac, &ex, e4b);
1777
1778 i += ex.fe_len;
1779 free -= ex.fe_len;
1780 }
1781
1782 ext4_mb_check_limits(ac, e4b, 1);
1783}
1784
1785/*
1786 * This is a special case for storages like raid5
1787 * we try to find stripe-aligned chunks for stripe-size requests
1788 * XXX should do so at least for multiples of stripe size as well
1789 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001790static noinline_for_stack
1791void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001792 struct ext4_buddy *e4b)
1793{
1794 struct super_block *sb = ac->ac_sb;
1795 struct ext4_sb_info *sbi = EXT4_SB(sb);
1796 void *bitmap = EXT4_MB_BITMAP(e4b);
1797 struct ext4_free_extent ex;
1798 ext4_fsblk_t first_group_block;
1799 ext4_fsblk_t a;
1800 ext4_grpblk_t i;
1801 int max;
1802
1803 BUG_ON(sbi->s_stripe == 0);
1804
1805 /* find first stripe-aligned block in group */
1806 first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
1807 + le32_to_cpu(sbi->s_es->s_first_data_block);
1808 a = first_group_block + sbi->s_stripe - 1;
1809 do_div(a, sbi->s_stripe);
1810 i = (a * sbi->s_stripe) - first_group_block;
1811
1812 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1813 if (!mb_test_bit(i, bitmap)) {
1814 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1815 if (max >= sbi->s_stripe) {
1816 ac->ac_found++;
1817 ac->ac_b_ex = ex;
1818 ext4_mb_use_best_found(ac, e4b);
1819 break;
1820 }
1821 }
1822 i += sbi->s_stripe;
1823 }
1824}
1825
1826static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1827 ext4_group_t group, int cr)
1828{
1829 unsigned free, fragments;
1830 unsigned i, bits;
Theodore Ts'oa4912122009-03-12 12:18:34 -04001831 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001832 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1833
1834 BUG_ON(cr < 0 || cr >= 4);
1835 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1836
1837 free = grp->bb_free;
1838 fragments = grp->bb_fragments;
1839 if (free == 0)
1840 return 0;
1841 if (fragments == 0)
1842 return 0;
1843
1844 switch (cr) {
1845 case 0:
1846 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001847
Theodore Ts'oa4912122009-03-12 12:18:34 -04001848 /* Avoid using the first bg of a flexgroup for data files */
1849 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1850 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1851 ((group % flex_size) == 0))
1852 return 0;
1853
Alex Tomasc9de5602008-01-29 00:19:52 -05001854 bits = ac->ac_sb->s_blocksize_bits + 1;
1855 for (i = ac->ac_2order; i <= bits; i++)
1856 if (grp->bb_counters[i] > 0)
1857 return 1;
1858 break;
1859 case 1:
1860 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1861 return 1;
1862 break;
1863 case 2:
1864 if (free >= ac->ac_g_ex.fe_len)
1865 return 1;
1866 break;
1867 case 3:
1868 return 1;
1869 default:
1870 BUG();
1871 }
1872
1873 return 0;
1874}
1875
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001876/*
1877 * lock the group_info alloc_sem of all the groups
1878 * belonging to the same buddy cache page. This
1879 * make sure other parallel operation on the buddy
1880 * cache doesn't happen whild holding the buddy cache
1881 * lock
1882 */
1883int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1884{
1885 int i;
1886 int block, pnum;
1887 int blocks_per_page;
1888 int groups_per_page;
Theodore Ts'o8df96752009-05-01 08:50:38 -04001889 ext4_group_t ngroups = ext4_get_groups_count(sb);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001890 ext4_group_t first_group;
1891 struct ext4_group_info *grp;
1892
1893 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1894 /*
1895 * the buddy cache inode stores the block bitmap
1896 * and buddy information in consecutive blocks.
1897 * So for each group we need two blocks.
1898 */
1899 block = group * 2;
1900 pnum = block / blocks_per_page;
1901 first_group = pnum * blocks_per_page / 2;
1902
1903 groups_per_page = blocks_per_page >> 1;
1904 if (groups_per_page == 0)
1905 groups_per_page = 1;
1906 /* read all groups the page covers into the cache */
1907 for (i = 0; i < groups_per_page; i++) {
1908
Theodore Ts'o8df96752009-05-01 08:50:38 -04001909 if ((first_group + i) >= ngroups)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001910 break;
1911 grp = ext4_get_group_info(sb, first_group + i);
1912 /* take all groups write allocation
1913 * semaphore. This make sure there is
1914 * no block allocation going on in any
1915 * of that groups
1916 */
Aneesh Kumar K.Vb7be0192008-11-23 23:51:53 -05001917 down_write_nested(&grp->alloc_sem, i);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001918 }
1919 return i;
1920}
1921
1922void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1923 ext4_group_t group, int locked_group)
1924{
1925 int i;
1926 int block, pnum;
1927 int blocks_per_page;
1928 ext4_group_t first_group;
1929 struct ext4_group_info *grp;
1930
1931 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1932 /*
1933 * the buddy cache inode stores the block bitmap
1934 * and buddy information in consecutive blocks.
1935 * So for each group we need two blocks.
1936 */
1937 block = group * 2;
1938 pnum = block / blocks_per_page;
1939 first_group = pnum * blocks_per_page / 2;
1940 /* release locks on all the groups */
1941 for (i = 0; i < locked_group; i++) {
1942
1943 grp = ext4_get_group_info(sb, first_group + i);
1944 /* take all groups write allocation
1945 * semaphore. This make sure there is
1946 * no block allocation going on in any
1947 * of that groups
1948 */
1949 up_write(&grp->alloc_sem);
1950 }
1951
1952}
1953
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001954static noinline_for_stack int
1955ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05001956{
Theodore Ts'o8df96752009-05-01 08:50:38 -04001957 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05001958 int cr;
1959 int err = 0;
1960 int bsbits;
1961 struct ext4_sb_info *sbi;
1962 struct super_block *sb;
1963 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05001964
1965 sb = ac->ac_sb;
1966 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04001967 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04001968 /* non-extent files are limited to low blocks/groups */
1969 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL))
1970 ngroups = sbi->s_blockfile_groups;
1971
Alex Tomasc9de5602008-01-29 00:19:52 -05001972 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1973
1974 /* first, try the goal */
1975 err = ext4_mb_find_by_goal(ac, &e4b);
1976 if (err || ac->ac_status == AC_STATUS_FOUND)
1977 goto out;
1978
1979 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1980 goto out;
1981
1982 /*
1983 * ac->ac2_order is set only if the fe_len is a power of 2
1984 * if ac2_order is set we also set criteria to 0 so that we
1985 * try exact allocation using buddy.
1986 */
1987 i = fls(ac->ac_g_ex.fe_len);
1988 ac->ac_2order = 0;
1989 /*
1990 * We search using buddy data only if the order of the request
1991 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04001992 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05001993 */
1994 if (i >= sbi->s_mb_order2_reqs) {
1995 /*
1996 * This should tell if fe_len is exactly power of 2
1997 */
1998 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1999 ac->ac_2order = i - 1;
2000 }
2001
2002 bsbits = ac->ac_sb->s_blocksize_bits;
Alex Tomasc9de5602008-01-29 00:19:52 -05002003
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002004 /* if stream allocation is enabled, use global goal */
2005 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002006 /* TBD: may be hot point */
2007 spin_lock(&sbi->s_md_lock);
2008 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2009 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2010 spin_unlock(&sbi->s_md_lock);
2011 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002012
Alex Tomasc9de5602008-01-29 00:19:52 -05002013 /* Let's just scan groups to find more-less suitable blocks */
2014 cr = ac->ac_2order ? 0 : 1;
2015 /*
2016 * cr == 0 try to get exact allocation,
2017 * cr == 3 try to get anything
2018 */
2019repeat:
2020 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2021 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002022 /*
2023 * searching for the right group start
2024 * from the goal value specified
2025 */
2026 group = ac->ac_g_ex.fe_group;
2027
Theodore Ts'o8df96752009-05-01 08:50:38 -04002028 for (i = 0; i < ngroups; group++, i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002029 struct ext4_group_info *grp;
2030 struct ext4_group_desc *desc;
2031
Theodore Ts'o8df96752009-05-01 08:50:38 -04002032 if (group == ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002033 group = 0;
2034
2035 /* quick check to skip empty groups */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002036 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002037 if (grp->bb_free == 0)
2038 continue;
2039
Alex Tomasc9de5602008-01-29 00:19:52 -05002040 err = ext4_mb_load_buddy(sb, group, &e4b);
2041 if (err)
2042 goto out;
2043
2044 ext4_lock_group(sb, group);
2045 if (!ext4_mb_good_group(ac, group, cr)) {
2046 /* someone did allocation from this group */
2047 ext4_unlock_group(sb, group);
2048 ext4_mb_release_desc(&e4b);
2049 continue;
2050 }
2051
2052 ac->ac_groups_scanned++;
2053 desc = ext4_get_group_desc(sb, group, NULL);
Theodore Ts'o75507ef2009-05-01 12:58:36 -04002054 if (cr == 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002055 ext4_mb_simple_scan_group(ac, &e4b);
2056 else if (cr == 1 &&
2057 ac->ac_g_ex.fe_len == sbi->s_stripe)
2058 ext4_mb_scan_aligned(ac, &e4b);
2059 else
2060 ext4_mb_complex_scan_group(ac, &e4b);
2061
2062 ext4_unlock_group(sb, group);
2063 ext4_mb_release_desc(&e4b);
2064
2065 if (ac->ac_status != AC_STATUS_CONTINUE)
2066 break;
2067 }
2068 }
2069
2070 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2071 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2072 /*
2073 * We've been searching too long. Let's try to allocate
2074 * the best chunk we've found so far
2075 */
2076
2077 ext4_mb_try_best_found(ac, &e4b);
2078 if (ac->ac_status != AC_STATUS_FOUND) {
2079 /*
2080 * Someone more lucky has already allocated it.
2081 * The only thing we can do is just take first
2082 * found block(s)
2083 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2084 */
2085 ac->ac_b_ex.fe_group = 0;
2086 ac->ac_b_ex.fe_start = 0;
2087 ac->ac_b_ex.fe_len = 0;
2088 ac->ac_status = AC_STATUS_CONTINUE;
2089 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2090 cr = 3;
2091 atomic_inc(&sbi->s_mb_lost_chunks);
2092 goto repeat;
2093 }
2094 }
2095out:
2096 return err;
2097}
2098
Alex Tomasc9de5602008-01-29 00:19:52 -05002099static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2100{
2101 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002102 ext4_group_t group;
2103
Theodore Ts'o8df96752009-05-01 08:50:38 -04002104 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002105 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002106 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002107 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002108}
2109
2110static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2111{
2112 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002113 ext4_group_t group;
2114
2115 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002116 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002117 return NULL;
2118 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002119 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002120}
2121
2122static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2123{
2124 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002125 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002126 int i;
2127 int err;
2128 struct ext4_buddy e4b;
2129 struct sg {
2130 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002131 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002132 } sg;
2133
2134 group--;
2135 if (group == 0)
2136 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2137 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2138 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2139 "group", "free", "frags", "first",
2140 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2141 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2142
2143 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2144 sizeof(struct ext4_group_info);
2145 err = ext4_mb_load_buddy(sb, group, &e4b);
2146 if (err) {
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002147 seq_printf(seq, "#%-5u: I/O error\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002148 return 0;
2149 }
2150 ext4_lock_group(sb, group);
2151 memcpy(&sg, ext4_get_group_info(sb, group), i);
2152 ext4_unlock_group(sb, group);
2153 ext4_mb_release_desc(&e4b);
2154
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002155 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002156 sg.info.bb_fragments, sg.info.bb_first_free);
2157 for (i = 0; i <= 13; i++)
2158 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2159 sg.info.bb_counters[i] : 0);
2160 seq_printf(seq, " ]\n");
2161
2162 return 0;
2163}
2164
2165static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2166{
2167}
2168
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002169static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002170 .start = ext4_mb_seq_groups_start,
2171 .next = ext4_mb_seq_groups_next,
2172 .stop = ext4_mb_seq_groups_stop,
2173 .show = ext4_mb_seq_groups_show,
2174};
2175
2176static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2177{
2178 struct super_block *sb = PDE(inode)->data;
2179 int rc;
2180
2181 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2182 if (rc == 0) {
2183 struct seq_file *m = (struct seq_file *)file->private_data;
2184 m->private = sb;
2185 }
2186 return rc;
2187
2188}
2189
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002190static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002191 .owner = THIS_MODULE,
2192 .open = ext4_mb_seq_groups_open,
2193 .read = seq_read,
2194 .llseek = seq_lseek,
2195 .release = seq_release,
2196};
2197
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002198
2199/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002200int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002201 struct ext4_group_desc *desc)
2202{
2203 int i, len;
2204 int metalen = 0;
2205 struct ext4_sb_info *sbi = EXT4_SB(sb);
2206 struct ext4_group_info **meta_group_info;
2207
2208 /*
2209 * First check if this group is the first of a reserved block.
2210 * If it's true, we have to allocate a new table of pointers
2211 * to ext4_group_info structures
2212 */
2213 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2214 metalen = sizeof(*meta_group_info) <<
2215 EXT4_DESC_PER_BLOCK_BITS(sb);
2216 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2217 if (meta_group_info == NULL) {
2218 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2219 "buddy group\n");
2220 goto exit_meta_group_info;
2221 }
2222 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2223 meta_group_info;
2224 }
2225
2226 /*
2227 * calculate needed size. if change bb_counters size,
2228 * don't forget about ext4_mb_generate_buddy()
2229 */
2230 len = offsetof(typeof(**meta_group_info),
2231 bb_counters[sb->s_blocksize_bits + 2]);
2232
2233 meta_group_info =
2234 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2235 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2236
2237 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2238 if (meta_group_info[i] == NULL) {
2239 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2240 goto exit_group_info;
2241 }
2242 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2243 &(meta_group_info[i]->bb_state));
2244
2245 /*
2246 * initialize bb_free to be able to skip
2247 * empty groups without initialization
2248 */
2249 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2250 meta_group_info[i]->bb_free =
2251 ext4_free_blocks_after_init(sb, group, desc);
2252 } else {
2253 meta_group_info[i]->bb_free =
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002254 ext4_free_blks_count(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002255 }
2256
2257 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002258 init_rwsem(&meta_group_info[i]->alloc_sem);
Joe Perches5a4a7982009-07-05 22:33:08 -04002259 meta_group_info[i]->bb_free_root.rb_node = NULL;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002260
2261#ifdef DOUBLE_CHECK
2262 {
2263 struct buffer_head *bh;
2264 meta_group_info[i]->bb_bitmap =
2265 kmalloc(sb->s_blocksize, GFP_KERNEL);
2266 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2267 bh = ext4_read_block_bitmap(sb, group);
2268 BUG_ON(bh == NULL);
2269 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2270 sb->s_blocksize);
2271 put_bh(bh);
2272 }
2273#endif
2274
2275 return 0;
2276
2277exit_group_info:
2278 /* If a meta_group_info table has been allocated, release it now */
2279 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2280 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2281exit_meta_group_info:
2282 return -ENOMEM;
2283} /* ext4_mb_add_groupinfo */
2284
Alex Tomasc9de5602008-01-29 00:19:52 -05002285static int ext4_mb_init_backend(struct super_block *sb)
2286{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002287 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002288 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002289 struct ext4_sb_info *sbi = EXT4_SB(sb);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002290 struct ext4_super_block *es = sbi->s_es;
2291 int num_meta_group_infos;
2292 int num_meta_group_infos_max;
2293 int array_size;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002294 struct ext4_group_desc *desc;
Alex Tomasc9de5602008-01-29 00:19:52 -05002295
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002296 /* This is the number of blocks used by GDT */
Theodore Ts'o8df96752009-05-01 08:50:38 -04002297 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002298 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2299
2300 /*
2301 * This is the total number of blocks used by GDT including
2302 * the number of reserved blocks for GDT.
2303 * The s_group_info array is allocated with this value
2304 * to allow a clean online resize without a complex
2305 * manipulation of pointer.
2306 * The drawback is the unused memory when no resize
2307 * occurs but it's very low in terms of pages
2308 * (see comments below)
2309 * Need to handle this properly when META_BG resizing is allowed
2310 */
2311 num_meta_group_infos_max = num_meta_group_infos +
2312 le16_to_cpu(es->s_reserved_gdt_blocks);
2313
2314 /*
2315 * array_size is the size of s_group_info array. We round it
2316 * to the next power of two because this approximation is done
2317 * internally by kmalloc so we can have some more memory
2318 * for free here (e.g. may be used for META_BG resize).
2319 */
2320 array_size = 1;
2321 while (array_size < sizeof(*sbi->s_group_info) *
2322 num_meta_group_infos_max)
2323 array_size = array_size << 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002324 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2325 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2326 * So a two level scheme suffices for now. */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002327 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
Alex Tomasc9de5602008-01-29 00:19:52 -05002328 if (sbi->s_group_info == NULL) {
2329 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2330 return -ENOMEM;
2331 }
2332 sbi->s_buddy_cache = new_inode(sb);
2333 if (sbi->s_buddy_cache == NULL) {
2334 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2335 goto err_freesgi;
2336 }
2337 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002338 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002339 desc = ext4_get_group_desc(sb, i, NULL);
2340 if (desc == NULL) {
2341 printk(KERN_ERR
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002342 "EXT4-fs: can't read descriptor %u\n", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002343 goto err_freebuddy;
2344 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002345 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2346 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002347 }
2348
2349 return 0;
2350
2351err_freebuddy:
Roel Kluinf1fa3342008-04-29 22:01:15 -04002352 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002353 kfree(ext4_get_group_info(sb, i));
Alex Tomasc9de5602008-01-29 00:19:52 -05002354 i = num_meta_group_infos;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002355 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002356 kfree(sbi->s_group_info[i]);
2357 iput(sbi->s_buddy_cache);
2358err_freesgi:
2359 kfree(sbi->s_group_info);
2360 return -ENOMEM;
2361}
2362
2363int ext4_mb_init(struct super_block *sb, int needs_recovery)
2364{
2365 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002366 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002367 unsigned offset;
2368 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002369 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002370
Eric Sandeen19278052009-08-25 22:36:25 -04002371 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002372
2373 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2374 if (sbi->s_mb_offsets == NULL) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002375 return -ENOMEM;
2376 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002377
Eric Sandeen19278052009-08-25 22:36:25 -04002378 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002379 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2380 if (sbi->s_mb_maxs == NULL) {
Dan Carpentera7b19442009-03-27 19:42:54 -04002381 kfree(sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002382 return -ENOMEM;
2383 }
2384
2385 /* order 0 is regular bitmap */
2386 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2387 sbi->s_mb_offsets[0] = 0;
2388
2389 i = 1;
2390 offset = 0;
2391 max = sb->s_blocksize << 2;
2392 do {
2393 sbi->s_mb_offsets[i] = offset;
2394 sbi->s_mb_maxs[i] = max;
2395 offset += 1 << (sb->s_blocksize_bits - i);
2396 max = max >> 1;
2397 i++;
2398 } while (i <= sb->s_blocksize_bits + 1);
2399
2400 /* init file for buddy data */
Shen Feng74767c52008-07-11 19:27:31 -04002401 ret = ext4_mb_init_backend(sb);
2402 if (ret != 0) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002403 kfree(sbi->s_mb_offsets);
2404 kfree(sbi->s_mb_maxs);
Shen Feng74767c52008-07-11 19:27:31 -04002405 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002406 }
2407
2408 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002409 spin_lock_init(&sbi->s_bal_lock);
2410
2411 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2412 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2413 sbi->s_mb_stats = MB_DEFAULT_STATS;
2414 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2415 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Alex Tomasc9de5602008-01-29 00:19:52 -05002416 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2417
Eric Sandeen730c2132008-09-13 15:23:29 -04002418 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002419 if (sbi->s_locality_groups == NULL) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002420 kfree(sbi->s_mb_offsets);
2421 kfree(sbi->s_mb_maxs);
2422 return -ENOMEM;
2423 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002424 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002425 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002426 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002427 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002428 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2429 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002430 spin_lock_init(&lg->lg_prealloc_lock);
2431 }
2432
Theodore Ts'o296c3552009-09-30 00:32:42 -04002433 if (sbi->s_proc)
2434 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2435 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002436
Frank Mayhar03901312009-01-07 00:06:22 -05002437 if (sbi->s_journal)
2438 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
Alex Tomasc9de5602008-01-29 00:19:52 -05002439 return 0;
2440}
2441
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002442/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002443static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2444{
2445 struct ext4_prealloc_space *pa;
2446 struct list_head *cur, *tmp;
2447 int count = 0;
2448
2449 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2450 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2451 list_del(&pa->pa_group_list);
2452 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002453 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002454 }
2455 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002456 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002457
2458}
2459
2460int ext4_mb_release(struct super_block *sb)
2461{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002462 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002463 ext4_group_t i;
2464 int num_meta_group_infos;
2465 struct ext4_group_info *grinfo;
2466 struct ext4_sb_info *sbi = EXT4_SB(sb);
2467
Alex Tomasc9de5602008-01-29 00:19:52 -05002468 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002469 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002470 grinfo = ext4_get_group_info(sb, i);
2471#ifdef DOUBLE_CHECK
2472 kfree(grinfo->bb_bitmap);
2473#endif
2474 ext4_lock_group(sb, i);
2475 ext4_mb_cleanup_pa(grinfo);
2476 ext4_unlock_group(sb, i);
2477 kfree(grinfo);
2478 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002479 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002480 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2481 EXT4_DESC_PER_BLOCK_BITS(sb);
2482 for (i = 0; i < num_meta_group_infos; i++)
2483 kfree(sbi->s_group_info[i]);
2484 kfree(sbi->s_group_info);
2485 }
2486 kfree(sbi->s_mb_offsets);
2487 kfree(sbi->s_mb_maxs);
2488 if (sbi->s_buddy_cache)
2489 iput(sbi->s_buddy_cache);
2490 if (sbi->s_mb_stats) {
2491 printk(KERN_INFO
2492 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2493 atomic_read(&sbi->s_bal_allocated),
2494 atomic_read(&sbi->s_bal_reqs),
2495 atomic_read(&sbi->s_bal_success));
2496 printk(KERN_INFO
2497 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2498 "%u 2^N hits, %u breaks, %u lost\n",
2499 atomic_read(&sbi->s_bal_ex_scanned),
2500 atomic_read(&sbi->s_bal_goals),
2501 atomic_read(&sbi->s_bal_2orders),
2502 atomic_read(&sbi->s_bal_breaks),
2503 atomic_read(&sbi->s_mb_lost_chunks));
2504 printk(KERN_INFO
2505 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2506 sbi->s_mb_buddies_generated++,
2507 sbi->s_mb_generation_time);
2508 printk(KERN_INFO
2509 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2510 atomic_read(&sbi->s_mb_preallocated),
2511 atomic_read(&sbi->s_mb_discarded));
2512 }
2513
Eric Sandeen730c2132008-09-13 15:23:29 -04002514 free_percpu(sbi->s_locality_groups);
Theodore Ts'o296c3552009-09-30 00:32:42 -04002515 if (sbi->s_proc)
2516 remove_proc_entry("mb_groups", sbi->s_proc);
Alex Tomasc9de5602008-01-29 00:19:52 -05002517
2518 return 0;
2519}
2520
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002521/*
2522 * This function is called by the jbd2 layer once the commit has finished,
2523 * so we know we can free the blocks that were released with that commit.
2524 */
2525static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
Alex Tomasc9de5602008-01-29 00:19:52 -05002526{
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002527 struct super_block *sb = journal->j_private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002528 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002529 struct ext4_group_info *db;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002530 int err, count = 0, count2 = 0;
2531 struct ext4_free_data *entry;
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002532 struct list_head *l, *ltmp;
Alex Tomasc9de5602008-01-29 00:19:52 -05002533
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002534 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2535 entry = list_entry(l, struct ext4_free_data, list);
Alex Tomasc9de5602008-01-29 00:19:52 -05002536
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002537 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002538 entry->count, entry->group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002539
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002540 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002541 /* we expect to find existing buddy because it's pinned */
2542 BUG_ON(err != 0);
2543
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002544 db = e4b.bd_info;
Alex Tomasc9de5602008-01-29 00:19:52 -05002545 /* there are blocks to put in buddy to make them really free */
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002546 count += entry->count;
Alex Tomasc9de5602008-01-29 00:19:52 -05002547 count2++;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002548 ext4_lock_group(sb, entry->group);
2549 /* Take it out of per group rb tree */
2550 rb_erase(&entry->node, &(db->bb_free_root));
2551 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2552
2553 if (!db->bb_free_root.rb_node) {
2554 /* No more items in the per group rb tree
2555 * balance refcounts from ext4_mb_free_metadata()
2556 */
2557 page_cache_release(e4b.bd_buddy_page);
2558 page_cache_release(e4b.bd_bitmap_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05002559 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002560 ext4_unlock_group(sb, entry->group);
Eric Sandeen5328e632009-11-19 14:25:42 -05002561 if (test_opt(sb, DISCARD)) {
2562 ext4_fsblk_t discard_block;
2563 struct ext4_super_block *es = EXT4_SB(sb)->s_es;
Alex Tomasc9de5602008-01-29 00:19:52 -05002564
Eric Sandeen5328e632009-11-19 14:25:42 -05002565 discard_block = (ext4_fsblk_t)entry->group *
2566 EXT4_BLOCKS_PER_GROUP(sb)
2567 + entry->start_blk
2568 + le32_to_cpu(es->s_first_data_block);
2569 trace_ext4_discard_blocks(sb,
2570 (unsigned long long)discard_block,
2571 entry->count);
2572 sb_issue_discard(sb, discard_block, entry->count);
2573 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002574 kmem_cache_free(ext4_free_ext_cachep, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002575 ext4_mb_release_desc(&e4b);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002576 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002577
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002578 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002579}
2580
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002581#ifdef CONFIG_EXT4_DEBUG
2582u8 mb_enable_debug __read_mostly;
2583
2584static struct dentry *debugfs_dir;
2585static struct dentry *debugfs_debug;
2586
2587static void __init ext4_create_debugfs_entry(void)
2588{
2589 debugfs_dir = debugfs_create_dir("ext4", NULL);
2590 if (debugfs_dir)
2591 debugfs_debug = debugfs_create_u8("mballoc-debug",
2592 S_IRUGO | S_IWUSR,
2593 debugfs_dir,
2594 &mb_enable_debug);
2595}
2596
2597static void ext4_remove_debugfs_entry(void)
2598{
2599 debugfs_remove(debugfs_debug);
2600 debugfs_remove(debugfs_dir);
2601}
2602
2603#else
2604
2605static void __init ext4_create_debugfs_entry(void)
2606{
2607}
2608
2609static void ext4_remove_debugfs_entry(void)
2610{
2611}
2612
2613#endif
2614
Alex Tomasc9de5602008-01-29 00:19:52 -05002615int __init init_ext4_mballoc(void)
2616{
2617 ext4_pspace_cachep =
2618 kmem_cache_create("ext4_prealloc_space",
2619 sizeof(struct ext4_prealloc_space),
2620 0, SLAB_RECLAIM_ACCOUNT, NULL);
2621 if (ext4_pspace_cachep == NULL)
2622 return -ENOMEM;
2623
Eric Sandeen256bdb42008-02-10 01:13:33 -05002624 ext4_ac_cachep =
2625 kmem_cache_create("ext4_alloc_context",
2626 sizeof(struct ext4_allocation_context),
2627 0, SLAB_RECLAIM_ACCOUNT, NULL);
2628 if (ext4_ac_cachep == NULL) {
2629 kmem_cache_destroy(ext4_pspace_cachep);
2630 return -ENOMEM;
2631 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002632
2633 ext4_free_ext_cachep =
2634 kmem_cache_create("ext4_free_block_extents",
2635 sizeof(struct ext4_free_data),
2636 0, SLAB_RECLAIM_ACCOUNT, NULL);
2637 if (ext4_free_ext_cachep == NULL) {
2638 kmem_cache_destroy(ext4_pspace_cachep);
2639 kmem_cache_destroy(ext4_ac_cachep);
2640 return -ENOMEM;
2641 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002642 ext4_create_debugfs_entry();
Alex Tomasc9de5602008-01-29 00:19:52 -05002643 return 0;
2644}
2645
2646void exit_ext4_mballoc(void)
2647{
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002648 /*
2649 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2650 * before destroying the slab cache.
2651 */
2652 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002653 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002654 kmem_cache_destroy(ext4_ac_cachep);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002655 kmem_cache_destroy(ext4_free_ext_cachep);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002656 ext4_remove_debugfs_entry();
Alex Tomasc9de5602008-01-29 00:19:52 -05002657}
2658
2659
2660/*
2661 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2662 * Returns 0 if success or error code
2663 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002664static noinline_for_stack int
2665ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o498e5f22008-11-05 00:14:04 -05002666 handle_t *handle, unsigned int reserv_blks)
Alex Tomasc9de5602008-01-29 00:19:52 -05002667{
2668 struct buffer_head *bitmap_bh = NULL;
2669 struct ext4_super_block *es;
2670 struct ext4_group_desc *gdp;
2671 struct buffer_head *gdp_bh;
2672 struct ext4_sb_info *sbi;
2673 struct super_block *sb;
2674 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002675 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002676
2677 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2678 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2679
2680 sb = ac->ac_sb;
2681 sbi = EXT4_SB(sb);
2682 es = sbi->s_es;
2683
Alex Tomasc9de5602008-01-29 00:19:52 -05002684
2685 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002686 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002687 if (!bitmap_bh)
2688 goto out_err;
2689
2690 err = ext4_journal_get_write_access(handle, bitmap_bh);
2691 if (err)
2692 goto out_err;
2693
2694 err = -EIO;
2695 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2696 if (!gdp)
2697 goto out_err;
2698
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002699 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Thadeu Lima de Souza Cascardo9fd97842009-01-26 19:26:26 -05002700 ext4_free_blks_count(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002701
Alex Tomasc9de5602008-01-29 00:19:52 -05002702 err = ext4_journal_get_write_access(handle, gdp_bh);
2703 if (err)
2704 goto out_err;
2705
2706 block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
2707 + ac->ac_b_ex.fe_start
2708 + le32_to_cpu(es->s_first_data_block);
2709
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002710 len = ac->ac_b_ex.fe_len;
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002711 if (!ext4_data_block_valid(sbi, block, len)) {
Harvey Harrison46e665e2008-04-17 10:38:59 -04002712 ext4_error(sb, __func__,
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002713 "Allocating blocks %llu-%llu which overlap "
2714 "fs metadata\n", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002715 /* File system mounted not to panic on error
2716 * Fix the bitmap and repeat the block allocation
2717 * We leak some of the blocks here.
2718 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002719 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2720 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2721 ac->ac_b_ex.fe_len);
2722 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002723 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002724 if (!err)
2725 err = -EAGAIN;
2726 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002727 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002728
2729 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002730#ifdef AGGRESSIVE_CHECK
2731 {
2732 int i;
2733 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2734 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2735 bitmap_bh->b_data));
2736 }
2737 }
2738#endif
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002739 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002740 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2741 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002742 ext4_free_blks_set(sb, gdp,
2743 ext4_free_blocks_after_init(sb,
2744 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002745 }
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002746 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2747 ext4_free_blks_set(sb, gdp, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002748 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002749
2750 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002751 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002752 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002753 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002754 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002755 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2756 /* release all the reserved blocks if non delalloc */
2757 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
Alex Tomasc9de5602008-01-29 00:19:52 -05002758
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002759 if (sbi->s_log_groups_per_flex) {
2760 ext4_group_t flex_group = ext4_flex_group(sbi,
2761 ac->ac_b_ex.fe_group);
Theodore Ts'o9f24e422009-03-04 19:09:10 -05002762 atomic_sub(ac->ac_b_ex.fe_len,
2763 &sbi->s_flex_groups[flex_group].free_blocks);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002764 }
2765
Frank Mayhar03901312009-01-07 00:06:22 -05002766 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002767 if (err)
2768 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002769 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002770
2771out_err:
2772 sb->s_dirt = 1;
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002773 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002774 return err;
2775}
2776
2777/*
2778 * here we normalize request for locality group
2779 * Group request are normalized to s_strip size if we set the same via mount
2780 * option. If not we set it to s_mb_group_prealloc which can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002781 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002782 *
2783 * XXX: should we try to preallocate more than the group has now?
2784 */
2785static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2786{
2787 struct super_block *sb = ac->ac_sb;
2788 struct ext4_locality_group *lg = ac->ac_lg;
2789
2790 BUG_ON(lg == NULL);
2791 if (EXT4_SB(sb)->s_stripe)
2792 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2793 else
2794 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002795 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002796 current->pid, ac->ac_g_ex.fe_len);
2797}
2798
2799/*
2800 * Normalization means making request better in terms of
2801 * size and alignment
2802 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002803static noinline_for_stack void
2804ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05002805 struct ext4_allocation_request *ar)
2806{
2807 int bsbits, max;
2808 ext4_lblk_t end;
Alex Tomasc9de5602008-01-29 00:19:52 -05002809 loff_t size, orig_size, start_off;
2810 ext4_lblk_t start, orig_start;
2811 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002812 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05002813
2814 /* do normalize only data requests, metadata requests
2815 do not need preallocation */
2816 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2817 return;
2818
2819 /* sometime caller may want exact blocks */
2820 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2821 return;
2822
2823 /* caller may indicate that preallocation isn't
2824 * required (it's a tail, for example) */
2825 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2826 return;
2827
2828 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2829 ext4_mb_normalize_group_request(ac);
2830 return ;
2831 }
2832
2833 bsbits = ac->ac_sb->s_blocksize_bits;
2834
2835 /* first, let's learn actual file size
2836 * given current request is allocated */
2837 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2838 size = size << bsbits;
2839 if (size < i_size_read(ac->ac_inode))
2840 size = i_size_read(ac->ac_inode);
2841
Valerie Clement19304792008-05-13 19:31:14 -04002842 /* max size of free chunks */
2843 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05002844
Valerie Clement19304792008-05-13 19:31:14 -04002845#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2846 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05002847
2848 /* first, try to predict filesize */
2849 /* XXX: should this table be tunable? */
2850 start_off = 0;
2851 if (size <= 16 * 1024) {
2852 size = 16 * 1024;
2853 } else if (size <= 32 * 1024) {
2854 size = 32 * 1024;
2855 } else if (size <= 64 * 1024) {
2856 size = 64 * 1024;
2857 } else if (size <= 128 * 1024) {
2858 size = 128 * 1024;
2859 } else if (size <= 256 * 1024) {
2860 size = 256 * 1024;
2861 } else if (size <= 512 * 1024) {
2862 size = 512 * 1024;
2863 } else if (size <= 1024 * 1024) {
2864 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04002865 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002866 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04002867 (21 - bsbits)) << 21;
2868 size = 2 * 1024 * 1024;
2869 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002870 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2871 (22 - bsbits)) << 22;
2872 size = 4 * 1024 * 1024;
2873 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04002874 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002875 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2876 (23 - bsbits)) << 23;
2877 size = 8 * 1024 * 1024;
2878 } else {
2879 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2880 size = ac->ac_o_ex.fe_len << bsbits;
2881 }
2882 orig_size = size = size >> bsbits;
2883 orig_start = start = start_off >> bsbits;
2884
2885 /* don't cover already allocated blocks in selected range */
2886 if (ar->pleft && start <= ar->lleft) {
2887 size -= ar->lleft + 1 - start;
2888 start = ar->lleft + 1;
2889 }
2890 if (ar->pright && start + size - 1 >= ar->lright)
2891 size -= start + size - ar->lright;
2892
2893 end = start + size;
2894
2895 /* check we don't cross already preallocated blocks */
2896 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002897 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05002898 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05002899
Alex Tomasc9de5602008-01-29 00:19:52 -05002900 if (pa->pa_deleted)
2901 continue;
2902 spin_lock(&pa->pa_lock);
2903 if (pa->pa_deleted) {
2904 spin_unlock(&pa->pa_lock);
2905 continue;
2906 }
2907
2908 pa_end = pa->pa_lstart + pa->pa_len;
2909
2910 /* PA must not overlap original request */
2911 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2912 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2913
Eric Sandeen38877f42009-08-17 23:55:24 -04002914 /* skip PAs this normalized request doesn't overlap with */
2915 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002916 spin_unlock(&pa->pa_lock);
2917 continue;
2918 }
2919 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2920
Eric Sandeen38877f42009-08-17 23:55:24 -04002921 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05002922 if (pa_end <= ac->ac_o_ex.fe_logical) {
2923 BUG_ON(pa_end < start);
2924 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04002925 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002926 BUG_ON(pa->pa_lstart > end);
2927 end = pa->pa_lstart;
2928 }
2929 spin_unlock(&pa->pa_lock);
2930 }
2931 rcu_read_unlock();
2932 size = end - start;
2933
2934 /* XXX: extra loop to check we really don't overlap preallocations */
2935 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002936 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05002937 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05002938 spin_lock(&pa->pa_lock);
2939 if (pa->pa_deleted == 0) {
2940 pa_end = pa->pa_lstart + pa->pa_len;
2941 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
2942 }
2943 spin_unlock(&pa->pa_lock);
2944 }
2945 rcu_read_unlock();
2946
2947 if (start + size <= ac->ac_o_ex.fe_logical &&
2948 start > ac->ac_o_ex.fe_logical) {
2949 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
2950 (unsigned long) start, (unsigned long) size,
2951 (unsigned long) ac->ac_o_ex.fe_logical);
2952 }
2953 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
2954 start > ac->ac_o_ex.fe_logical);
Eric Sandeen8d03c7a2009-03-14 11:51:46 -04002955 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05002956
2957 /* now prepare goal request */
2958
2959 /* XXX: is it better to align blocks WRT to logical
2960 * placement or satisfy big request as is */
2961 ac->ac_g_ex.fe_logical = start;
2962 ac->ac_g_ex.fe_len = size;
2963
2964 /* define goal start in order to merge */
2965 if (ar->pright && (ar->lright == (start + size))) {
2966 /* merge to the right */
2967 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
2968 &ac->ac_f_ex.fe_group,
2969 &ac->ac_f_ex.fe_start);
2970 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
2971 }
2972 if (ar->pleft && (ar->lleft + 1 == start)) {
2973 /* merge to the left */
2974 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
2975 &ac->ac_f_ex.fe_group,
2976 &ac->ac_f_ex.fe_start);
2977 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
2978 }
2979
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002980 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05002981 (unsigned) orig_size, (unsigned) start);
2982}
2983
2984static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
2985{
2986 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2987
2988 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
2989 atomic_inc(&sbi->s_bal_reqs);
2990 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
2991 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
2992 atomic_inc(&sbi->s_bal_success);
2993 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
2994 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2995 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2996 atomic_inc(&sbi->s_bal_goals);
2997 if (ac->ac_found > sbi->s_mb_max_to_scan)
2998 atomic_inc(&sbi->s_bal_breaks);
2999 }
3000
Theodore Ts'o296c3552009-09-30 00:32:42 -04003001 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3002 trace_ext4_mballoc_alloc(ac);
3003 else
3004 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003005}
3006
3007/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003008 * Called on failure; free up any blocks from the inode PA for this
3009 * context. We don't need this for MB_GROUP_PA because we only change
3010 * pa_free in ext4_mb_release_context(), but on failure, we've already
3011 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3012 */
3013static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3014{
3015 struct ext4_prealloc_space *pa = ac->ac_pa;
3016 int len;
3017
3018 if (pa && pa->pa_type == MB_INODE_PA) {
3019 len = ac->ac_b_ex.fe_len;
3020 pa->pa_free += len;
3021 }
3022
3023}
3024
3025/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003026 * use blocks preallocated to inode
3027 */
3028static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3029 struct ext4_prealloc_space *pa)
3030{
3031 ext4_fsblk_t start;
3032 ext4_fsblk_t end;
3033 int len;
3034
3035 /* found preallocated blocks, use them */
3036 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3037 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3038 len = end - start;
3039 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3040 &ac->ac_b_ex.fe_start);
3041 ac->ac_b_ex.fe_len = len;
3042 ac->ac_status = AC_STATUS_FOUND;
3043 ac->ac_pa = pa;
3044
3045 BUG_ON(start < pa->pa_pstart);
3046 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3047 BUG_ON(pa->pa_free < len);
3048 pa->pa_free -= len;
3049
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003050 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003051}
3052
3053/*
3054 * use blocks preallocated to locality group
3055 */
3056static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3057 struct ext4_prealloc_space *pa)
3058{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003059 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003060
Alex Tomasc9de5602008-01-29 00:19:52 -05003061 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3062 &ac->ac_b_ex.fe_group,
3063 &ac->ac_b_ex.fe_start);
3064 ac->ac_b_ex.fe_len = len;
3065 ac->ac_status = AC_STATUS_FOUND;
3066 ac->ac_pa = pa;
3067
3068 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003069 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003070 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003071 * in on-disk bitmap -- see ext4_mb_release_context()
3072 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003073 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003074 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003075}
3076
3077/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003078 * Return the prealloc space that have minimal distance
3079 * from the goal block. @cpa is the prealloc
3080 * space that is having currently known minimal distance
3081 * from the goal block.
3082 */
3083static struct ext4_prealloc_space *
3084ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3085 struct ext4_prealloc_space *pa,
3086 struct ext4_prealloc_space *cpa)
3087{
3088 ext4_fsblk_t cur_distance, new_distance;
3089
3090 if (cpa == NULL) {
3091 atomic_inc(&pa->pa_count);
3092 return pa;
3093 }
3094 cur_distance = abs(goal_block - cpa->pa_pstart);
3095 new_distance = abs(goal_block - pa->pa_pstart);
3096
3097 if (cur_distance < new_distance)
3098 return cpa;
3099
3100 /* drop the previous reference */
3101 atomic_dec(&cpa->pa_count);
3102 atomic_inc(&pa->pa_count);
3103 return pa;
3104}
3105
3106/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003107 * search goal blocks in preallocated space
3108 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003109static noinline_for_stack int
3110ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003111{
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003112 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003113 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3114 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003115 struct ext4_prealloc_space *pa, *cpa = NULL;
3116 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003117
3118 /* only data can be preallocated */
3119 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3120 return 0;
3121
3122 /* first, try per-file preallocation */
3123 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003124 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003125
3126 /* all fields in this condition don't change,
3127 * so we can skip locking for them */
3128 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3129 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3130 continue;
3131
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003132 /* non-extent files can't have physical blocks past 2^32 */
3133 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) &&
3134 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3135 continue;
3136
Alex Tomasc9de5602008-01-29 00:19:52 -05003137 /* found preallocated blocks, use them */
3138 spin_lock(&pa->pa_lock);
3139 if (pa->pa_deleted == 0 && pa->pa_free) {
3140 atomic_inc(&pa->pa_count);
3141 ext4_mb_use_inode_pa(ac, pa);
3142 spin_unlock(&pa->pa_lock);
3143 ac->ac_criteria = 10;
3144 rcu_read_unlock();
3145 return 1;
3146 }
3147 spin_unlock(&pa->pa_lock);
3148 }
3149 rcu_read_unlock();
3150
3151 /* can we use group allocation? */
3152 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3153 return 0;
3154
3155 /* inode may have no locality group for some reason */
3156 lg = ac->ac_lg;
3157 if (lg == NULL)
3158 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003159 order = fls(ac->ac_o_ex.fe_len) - 1;
3160 if (order > PREALLOC_TB_SIZE - 1)
3161 /* The max size of hash table is PREALLOC_TB_SIZE */
3162 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003163
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003164 goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) +
3165 ac->ac_g_ex.fe_start +
3166 le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block);
3167 /*
3168 * search for the prealloc space that is having
3169 * minimal distance from the goal block.
3170 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003171 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3172 rcu_read_lock();
3173 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3174 pa_inode_list) {
3175 spin_lock(&pa->pa_lock);
3176 if (pa->pa_deleted == 0 &&
3177 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003178
3179 cpa = ext4_mb_check_group_pa(goal_block,
3180 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003181 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003182 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003183 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003184 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003185 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003186 if (cpa) {
3187 ext4_mb_use_group_pa(ac, cpa);
3188 ac->ac_criteria = 20;
3189 return 1;
3190 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003191 return 0;
3192}
3193
3194/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003195 * the function goes through all block freed in the group
3196 * but not yet committed and marks them used in in-core bitmap.
3197 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003198 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003199 */
3200static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3201 ext4_group_t group)
3202{
3203 struct rb_node *n;
3204 struct ext4_group_info *grp;
3205 struct ext4_free_data *entry;
3206
3207 grp = ext4_get_group_info(sb, group);
3208 n = rb_first(&(grp->bb_free_root));
3209
3210 while (n) {
3211 entry = rb_entry(n, struct ext4_free_data, node);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003212 mb_set_bits(bitmap, entry->start_blk, entry->count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003213 n = rb_next(n);
3214 }
3215 return;
3216}
3217
3218/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003219 * the function goes through all preallocation in this group and marks them
3220 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003221 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003222 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003223static noinline_for_stack
3224void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003225 ext4_group_t group)
3226{
3227 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3228 struct ext4_prealloc_space *pa;
3229 struct list_head *cur;
3230 ext4_group_t groupnr;
3231 ext4_grpblk_t start;
3232 int preallocated = 0;
3233 int count = 0;
3234 int len;
3235
3236 /* all form of preallocation discards first load group,
3237 * so the only competing code is preallocation use.
3238 * we don't need any locking here
3239 * notice we do NOT ignore preallocations with pa_deleted
3240 * otherwise we could leave used blocks available for
3241 * allocation in buddy when concurrent ext4_mb_put_pa()
3242 * is dropping preallocation
3243 */
3244 list_for_each(cur, &grp->bb_prealloc_list) {
3245 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3246 spin_lock(&pa->pa_lock);
3247 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3248 &groupnr, &start);
3249 len = pa->pa_len;
3250 spin_unlock(&pa->pa_lock);
3251 if (unlikely(len == 0))
3252 continue;
3253 BUG_ON(groupnr != group);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003254 mb_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003255 preallocated += len;
3256 count++;
3257 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003258 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003259}
3260
3261static void ext4_mb_pa_callback(struct rcu_head *head)
3262{
3263 struct ext4_prealloc_space *pa;
3264 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3265 kmem_cache_free(ext4_pspace_cachep, pa);
3266}
3267
3268/*
3269 * drops a reference to preallocated space descriptor
3270 * if this was the last reference and the space is consumed
3271 */
3272static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3273 struct super_block *sb, struct ext4_prealloc_space *pa)
3274{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003275 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003276 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003277
3278 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3279 return;
3280
3281 /* in this short window concurrent discard can set pa_deleted */
3282 spin_lock(&pa->pa_lock);
3283 if (pa->pa_deleted == 1) {
3284 spin_unlock(&pa->pa_lock);
3285 return;
3286 }
3287
3288 pa->pa_deleted = 1;
3289 spin_unlock(&pa->pa_lock);
3290
Eric Sandeend33a1972009-03-16 23:25:40 -04003291 grp_blk = pa->pa_pstart;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003292 /*
3293 * If doing group-based preallocation, pa_pstart may be in the
3294 * next group when pa is used up
3295 */
3296 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003297 grp_blk--;
3298
3299 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
Alex Tomasc9de5602008-01-29 00:19:52 -05003300
3301 /*
3302 * possible race:
3303 *
3304 * P1 (buddy init) P2 (regular allocation)
3305 * find block B in PA
3306 * copy on-disk bitmap to buddy
3307 * mark B in on-disk bitmap
3308 * drop PA from group
3309 * mark all PAs in buddy
3310 *
3311 * thus, P1 initializes buddy with B available. to prevent this
3312 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3313 * against that pair
3314 */
3315 ext4_lock_group(sb, grp);
3316 list_del(&pa->pa_group_list);
3317 ext4_unlock_group(sb, grp);
3318
3319 spin_lock(pa->pa_obj_lock);
3320 list_del_rcu(&pa->pa_inode_list);
3321 spin_unlock(pa->pa_obj_lock);
3322
3323 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3324}
3325
3326/*
3327 * creates new preallocated space for given inode
3328 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003329static noinline_for_stack int
3330ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003331{
3332 struct super_block *sb = ac->ac_sb;
3333 struct ext4_prealloc_space *pa;
3334 struct ext4_group_info *grp;
3335 struct ext4_inode_info *ei;
3336
3337 /* preallocate only when found space is larger then requested */
3338 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3339 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3340 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3341
3342 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3343 if (pa == NULL)
3344 return -ENOMEM;
3345
3346 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3347 int winl;
3348 int wins;
3349 int win;
3350 int offs;
3351
3352 /* we can't allocate as much as normalizer wants.
3353 * so, found space must get proper lstart
3354 * to cover original request */
3355 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3356 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3357
3358 /* we're limited by original request in that
3359 * logical block must be covered any way
3360 * winl is window we can move our chunk within */
3361 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3362
3363 /* also, we should cover whole original request */
3364 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3365
3366 /* the smallest one defines real window */
3367 win = min(winl, wins);
3368
3369 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3370 if (offs && offs < win)
3371 win = offs;
3372
3373 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3374 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3375 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3376 }
3377
3378 /* preallocation can change ac_b_ex, thus we store actually
3379 * allocated blocks for history */
3380 ac->ac_f_ex = ac->ac_b_ex;
3381
3382 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3383 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3384 pa->pa_len = ac->ac_b_ex.fe_len;
3385 pa->pa_free = pa->pa_len;
3386 atomic_set(&pa->pa_count, 1);
3387 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003388 INIT_LIST_HEAD(&pa->pa_inode_list);
3389 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003390 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003391 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003392
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003393 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003394 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003395 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003396
3397 ext4_mb_use_inode_pa(ac, pa);
3398 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3399
3400 ei = EXT4_I(ac->ac_inode);
3401 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3402
3403 pa->pa_obj_lock = &ei->i_prealloc_lock;
3404 pa->pa_inode = ac->ac_inode;
3405
3406 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3407 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3408 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3409
3410 spin_lock(pa->pa_obj_lock);
3411 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3412 spin_unlock(pa->pa_obj_lock);
3413
3414 return 0;
3415}
3416
3417/*
3418 * creates new preallocated space for locality group inodes belongs to
3419 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003420static noinline_for_stack int
3421ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003422{
3423 struct super_block *sb = ac->ac_sb;
3424 struct ext4_locality_group *lg;
3425 struct ext4_prealloc_space *pa;
3426 struct ext4_group_info *grp;
3427
3428 /* preallocate only when found space is larger then requested */
3429 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3430 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3431 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3432
3433 BUG_ON(ext4_pspace_cachep == NULL);
3434 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3435 if (pa == NULL)
3436 return -ENOMEM;
3437
3438 /* preallocation can change ac_b_ex, thus we store actually
3439 * allocated blocks for history */
3440 ac->ac_f_ex = ac->ac_b_ex;
3441
3442 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3443 pa->pa_lstart = pa->pa_pstart;
3444 pa->pa_len = ac->ac_b_ex.fe_len;
3445 pa->pa_free = pa->pa_len;
3446 atomic_set(&pa->pa_count, 1);
3447 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003448 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003449 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003450 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003451 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003452
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003453 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003454 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3455 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003456
3457 ext4_mb_use_group_pa(ac, pa);
3458 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3459
3460 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3461 lg = ac->ac_lg;
3462 BUG_ON(lg == NULL);
3463
3464 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3465 pa->pa_inode = NULL;
3466
3467 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3468 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3469 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3470
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003471 /*
3472 * We will later add the new pa to the right bucket
3473 * after updating the pa_free in ext4_mb_release_context
3474 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003475 return 0;
3476}
3477
3478static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3479{
3480 int err;
3481
3482 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3483 err = ext4_mb_new_group_pa(ac);
3484 else
3485 err = ext4_mb_new_inode_pa(ac);
3486 return err;
3487}
3488
3489/*
3490 * finds all unused blocks in on-disk bitmap, frees them in
3491 * in-core bitmap and buddy.
3492 * @pa must be unlinked from inode and group lists, so that
3493 * nobody else can find/use it.
3494 * the caller MUST hold group/inode locks.
3495 * TODO: optimize the case when there are no in-core structures yet
3496 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003497static noinline_for_stack int
3498ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003499 struct ext4_prealloc_space *pa,
3500 struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003501{
Alex Tomasc9de5602008-01-29 00:19:52 -05003502 struct super_block *sb = e4b->bd_sb;
3503 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003504 unsigned int end;
3505 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003506 ext4_group_t group;
3507 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003508 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003509 sector_t start;
3510 int err = 0;
3511 int free = 0;
3512
3513 BUG_ON(pa->pa_deleted == 0);
3514 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'oba80b102009-01-03 20:03:21 -05003515 grp_blk_start = pa->pa_pstart - bit;
Alex Tomasc9de5602008-01-29 00:19:52 -05003516 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3517 end = bit + pa->pa_len;
3518
Eric Sandeen256bdb42008-02-10 01:13:33 -05003519 if (ac) {
3520 ac->ac_sb = sb;
3521 ac->ac_inode = pa->pa_inode;
Eric Sandeen256bdb42008-02-10 01:13:33 -05003522 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003523
3524 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003525 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003526 if (bit >= end)
3527 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003528 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003529 start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
3530 le32_to_cpu(sbi->s_es->s_first_data_block);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003531 mb_debug(1, " free preallocated %u/%u in group %u\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003532 (unsigned) start, (unsigned) next - bit,
3533 (unsigned) group);
3534 free += next - bit;
3535
Eric Sandeen256bdb42008-02-10 01:13:33 -05003536 if (ac) {
3537 ac->ac_b_ex.fe_group = group;
3538 ac->ac_b_ex.fe_start = bit;
3539 ac->ac_b_ex.fe_len = next - bit;
3540 ac->ac_b_ex.fe_logical = 0;
Theodore Ts'o296c3552009-09-30 00:32:42 -04003541 trace_ext4_mballoc_discard(ac);
Eric Sandeen256bdb42008-02-10 01:13:33 -05003542 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003543
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003544 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3545 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003546 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3547 bit = next + 1;
3548 }
3549 if (free != pa->pa_free) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003550 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003551 pa, (unsigned long) pa->pa_lstart,
3552 (unsigned long) pa->pa_pstart,
3553 (unsigned long) pa->pa_len);
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003554 ext4_grp_locked_error(sb, group,
3555 __func__, "free %u, pa_free %u",
3556 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003557 /*
3558 * pa is already deleted so we use the value obtained
3559 * from the bitmap and continue.
3560 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003561 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003562 atomic_add(free, &sbi->s_mb_discarded);
3563
3564 return err;
3565}
3566
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003567static noinline_for_stack int
3568ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003569 struct ext4_prealloc_space *pa,
3570 struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003571{
Alex Tomasc9de5602008-01-29 00:19:52 -05003572 struct super_block *sb = e4b->bd_sb;
3573 ext4_group_t group;
3574 ext4_grpblk_t bit;
3575
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003576 trace_ext4_mb_release_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003577 BUG_ON(pa->pa_deleted == 0);
3578 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3579 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3580 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3581 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3582
Eric Sandeen256bdb42008-02-10 01:13:33 -05003583 if (ac) {
3584 ac->ac_sb = sb;
3585 ac->ac_inode = NULL;
3586 ac->ac_b_ex.fe_group = group;
3587 ac->ac_b_ex.fe_start = bit;
3588 ac->ac_b_ex.fe_len = pa->pa_len;
3589 ac->ac_b_ex.fe_logical = 0;
Theodore Ts'o296c3552009-09-30 00:32:42 -04003590 trace_ext4_mballoc_discard(ac);
Eric Sandeen256bdb42008-02-10 01:13:33 -05003591 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003592
3593 return 0;
3594}
3595
3596/*
3597 * releases all preallocations in given group
3598 *
3599 * first, we need to decide discard policy:
3600 * - when do we discard
3601 * 1) ENOSPC
3602 * - how many do we discard
3603 * 1) how many requested
3604 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003605static noinline_for_stack int
3606ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003607 ext4_group_t group, int needed)
3608{
3609 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3610 struct buffer_head *bitmap_bh = NULL;
3611 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003612 struct ext4_allocation_context *ac;
Alex Tomasc9de5602008-01-29 00:19:52 -05003613 struct list_head list;
3614 struct ext4_buddy e4b;
3615 int err;
3616 int busy = 0;
3617 int free = 0;
3618
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003619 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003620
3621 if (list_empty(&grp->bb_prealloc_list))
3622 return 0;
3623
Theodore Ts'o574ca172008-07-11 19:27:31 -04003624 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003625 if (bitmap_bh == NULL) {
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003626 ext4_error(sb, __func__, "Error in reading block "
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003627 "bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003628 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003629 }
3630
3631 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003632 if (err) {
3633 ext4_error(sb, __func__, "Error in loading buddy "
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003634 "information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003635 put_bh(bitmap_bh);
3636 return 0;
3637 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003638
3639 if (needed == 0)
3640 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3641
Alex Tomasc9de5602008-01-29 00:19:52 -05003642 INIT_LIST_HEAD(&list);
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003643 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003644 if (ac)
3645 ac->ac_sb = sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05003646repeat:
3647 ext4_lock_group(sb, group);
3648 list_for_each_entry_safe(pa, tmp,
3649 &grp->bb_prealloc_list, pa_group_list) {
3650 spin_lock(&pa->pa_lock);
3651 if (atomic_read(&pa->pa_count)) {
3652 spin_unlock(&pa->pa_lock);
3653 busy = 1;
3654 continue;
3655 }
3656 if (pa->pa_deleted) {
3657 spin_unlock(&pa->pa_lock);
3658 continue;
3659 }
3660
3661 /* seems this one can be freed ... */
3662 pa->pa_deleted = 1;
3663
3664 /* we can trust pa_free ... */
3665 free += pa->pa_free;
3666
3667 spin_unlock(&pa->pa_lock);
3668
3669 list_del(&pa->pa_group_list);
3670 list_add(&pa->u.pa_tmp_list, &list);
3671 }
3672
3673 /* if we still need more blocks and some PAs were used, try again */
3674 if (free < needed && busy) {
3675 busy = 0;
3676 ext4_unlock_group(sb, group);
3677 /*
3678 * Yield the CPU here so that we don't get soft lockup
3679 * in non preempt case.
3680 */
3681 yield();
3682 goto repeat;
3683 }
3684
3685 /* found anything to free? */
3686 if (list_empty(&list)) {
3687 BUG_ON(free != 0);
3688 goto out;
3689 }
3690
3691 /* now free all selected PAs */
3692 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3693
3694 /* remove from object (inode or locality group) */
3695 spin_lock(pa->pa_obj_lock);
3696 list_del_rcu(&pa->pa_inode_list);
3697 spin_unlock(pa->pa_obj_lock);
3698
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003699 if (pa->pa_type == MB_GROUP_PA)
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003700 ext4_mb_release_group_pa(&e4b, pa, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003701 else
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003702 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003703
3704 list_del(&pa->u.pa_tmp_list);
3705 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3706 }
3707
3708out:
3709 ext4_unlock_group(sb, group);
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003710 if (ac)
3711 kmem_cache_free(ext4_ac_cachep, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003712 ext4_mb_release_desc(&e4b);
3713 put_bh(bitmap_bh);
3714 return free;
3715}
3716
3717/*
3718 * releases all non-used preallocated blocks for given inode
3719 *
3720 * It's important to discard preallocations under i_data_sem
3721 * We don't want another block to be served from the prealloc
3722 * space when we are discarding the inode prealloc space.
3723 *
3724 * FIXME!! Make sure it is valid at all the call sites
3725 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003726void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003727{
3728 struct ext4_inode_info *ei = EXT4_I(inode);
3729 struct super_block *sb = inode->i_sb;
3730 struct buffer_head *bitmap_bh = NULL;
3731 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003732 struct ext4_allocation_context *ac;
Alex Tomasc9de5602008-01-29 00:19:52 -05003733 ext4_group_t group = 0;
3734 struct list_head list;
3735 struct ext4_buddy e4b;
3736 int err;
3737
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003738 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003739 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3740 return;
3741 }
3742
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003743 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003744 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003745
3746 INIT_LIST_HEAD(&list);
3747
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003748 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003749 if (ac) {
3750 ac->ac_sb = sb;
3751 ac->ac_inode = inode;
3752 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003753repeat:
3754 /* first, collect all pa's in the inode */
3755 spin_lock(&ei->i_prealloc_lock);
3756 while (!list_empty(&ei->i_prealloc_list)) {
3757 pa = list_entry(ei->i_prealloc_list.next,
3758 struct ext4_prealloc_space, pa_inode_list);
3759 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3760 spin_lock(&pa->pa_lock);
3761 if (atomic_read(&pa->pa_count)) {
3762 /* this shouldn't happen often - nobody should
3763 * use preallocation while we're discarding it */
3764 spin_unlock(&pa->pa_lock);
3765 spin_unlock(&ei->i_prealloc_lock);
3766 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3767 WARN_ON(1);
3768 schedule_timeout_uninterruptible(HZ);
3769 goto repeat;
3770
3771 }
3772 if (pa->pa_deleted == 0) {
3773 pa->pa_deleted = 1;
3774 spin_unlock(&pa->pa_lock);
3775 list_del_rcu(&pa->pa_inode_list);
3776 list_add(&pa->u.pa_tmp_list, &list);
3777 continue;
3778 }
3779
3780 /* someone is deleting pa right now */
3781 spin_unlock(&pa->pa_lock);
3782 spin_unlock(&ei->i_prealloc_lock);
3783
3784 /* we have to wait here because pa_deleted
3785 * doesn't mean pa is already unlinked from
3786 * the list. as we might be called from
3787 * ->clear_inode() the inode will get freed
3788 * and concurrent thread which is unlinking
3789 * pa from inode's list may access already
3790 * freed memory, bad-bad-bad */
3791
3792 /* XXX: if this happens too often, we can
3793 * add a flag to force wait only in case
3794 * of ->clear_inode(), but not in case of
3795 * regular truncate */
3796 schedule_timeout_uninterruptible(HZ);
3797 goto repeat;
3798 }
3799 spin_unlock(&ei->i_prealloc_lock);
3800
3801 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003802 BUG_ON(pa->pa_type != MB_INODE_PA);
Alex Tomasc9de5602008-01-29 00:19:52 -05003803 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3804
3805 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003806 if (err) {
3807 ext4_error(sb, __func__, "Error in loading buddy "
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003808 "information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003809 continue;
3810 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003811
Theodore Ts'o574ca172008-07-11 19:27:31 -04003812 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003813 if (bitmap_bh == NULL) {
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003814 ext4_error(sb, __func__, "Error in reading block "
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003815 "bitmap for %u", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003816 ext4_mb_release_desc(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003817 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05003818 }
3819
3820 ext4_lock_group(sb, group);
3821 list_del(&pa->pa_group_list);
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003822 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003823 ext4_unlock_group(sb, group);
3824
3825 ext4_mb_release_desc(&e4b);
3826 put_bh(bitmap_bh);
3827
3828 list_del(&pa->u.pa_tmp_list);
3829 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3830 }
Aneesh Kumar K.Vc83617d2008-04-29 22:00:47 -04003831 if (ac)
3832 kmem_cache_free(ext4_ac_cachep, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003833}
3834
3835/*
3836 * finds all preallocated spaces and return blocks being freed to them
3837 * if preallocated space becomes full (no block is used from the space)
3838 * then the function frees space in buddy
3839 * XXX: at the moment, truncate (which is the only way to free blocks)
3840 * discards all preallocations
3841 */
3842static void ext4_mb_return_to_preallocation(struct inode *inode,
3843 struct ext4_buddy *e4b,
3844 sector_t block, int count)
3845{
3846 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3847}
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003848#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05003849static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3850{
3851 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04003852 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003853
3854 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3855 " Allocation context details:\n");
3856 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3857 ac->ac_status, ac->ac_flags);
3858 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3859 "best %lu/%lu/%lu@%lu cr %d\n",
3860 (unsigned long)ac->ac_o_ex.fe_group,
3861 (unsigned long)ac->ac_o_ex.fe_start,
3862 (unsigned long)ac->ac_o_ex.fe_len,
3863 (unsigned long)ac->ac_o_ex.fe_logical,
3864 (unsigned long)ac->ac_g_ex.fe_group,
3865 (unsigned long)ac->ac_g_ex.fe_start,
3866 (unsigned long)ac->ac_g_ex.fe_len,
3867 (unsigned long)ac->ac_g_ex.fe_logical,
3868 (unsigned long)ac->ac_b_ex.fe_group,
3869 (unsigned long)ac->ac_b_ex.fe_start,
3870 (unsigned long)ac->ac_b_ex.fe_len,
3871 (unsigned long)ac->ac_b_ex.fe_logical,
3872 (int)ac->ac_criteria);
3873 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3874 ac->ac_found);
3875 printk(KERN_ERR "EXT4-fs: groups: \n");
Theodore Ts'o8df96752009-05-01 08:50:38 -04003876 ngroups = ext4_get_groups_count(sb);
3877 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003878 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3879 struct ext4_prealloc_space *pa;
3880 ext4_grpblk_t start;
3881 struct list_head *cur;
3882 ext4_lock_group(sb, i);
3883 list_for_each(cur, &grp->bb_prealloc_list) {
3884 pa = list_entry(cur, struct ext4_prealloc_space,
3885 pa_group_list);
3886 spin_lock(&pa->pa_lock);
3887 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3888 NULL, &start);
3889 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04003890 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3891 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003892 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04003893 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05003894
3895 if (grp->bb_free == 0)
3896 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04003897 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003898 i, grp->bb_free, grp->bb_fragments);
3899 }
3900 printk(KERN_ERR "\n");
3901}
3902#else
3903static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3904{
3905 return;
3906}
3907#endif
3908
3909/*
3910 * We use locality group preallocation for small size file. The size of the
3911 * file is determined by the current size or the resulting size after
3912 * allocation which ever is larger
3913 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04003914 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05003915 */
3916static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3917{
3918 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3919 int bsbits = ac->ac_sb->s_blocksize_bits;
3920 loff_t size, isize;
3921
3922 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3923 return;
3924
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003925 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3926 return;
3927
Alex Tomasc9de5602008-01-29 00:19:52 -05003928 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
Theodore Ts'o50797482009-09-18 13:34:02 -04003929 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3930 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003931
Theodore Ts'o50797482009-09-18 13:34:02 -04003932 if ((size == isize) &&
3933 !ext4_fs_is_busy(sbi) &&
3934 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3935 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3936 return;
3937 }
3938
Alex Tomasc9de5602008-01-29 00:19:52 -05003939 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04003940 size = max(size, isize);
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003941 if (size >= sbi->s_mb_stream_request) {
3942 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05003943 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003944 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003945
3946 BUG_ON(ac->ac_lg != NULL);
3947 /*
3948 * locality group prealloc space are per cpu. The reason for having
3949 * per cpu locality group is to reduce the contention between block
3950 * request from multiple CPUs.
3951 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09003952 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05003953
3954 /* we're going to use group allocation */
3955 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3956
3957 /* serialize all allocations in the group */
3958 mutex_lock(&ac->ac_lg->lg_mutex);
3959}
3960
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003961static noinline_for_stack int
3962ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05003963 struct ext4_allocation_request *ar)
3964{
3965 struct super_block *sb = ar->inode->i_sb;
3966 struct ext4_sb_info *sbi = EXT4_SB(sb);
3967 struct ext4_super_block *es = sbi->s_es;
3968 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003969 unsigned int len;
3970 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05003971 ext4_grpblk_t block;
3972
3973 /* we can't allocate > group size */
3974 len = ar->len;
3975
3976 /* just a dirty hack to filter too big requests */
3977 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
3978 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
3979
3980 /* start searching from the goal */
3981 goal = ar->goal;
3982 if (goal < le32_to_cpu(es->s_first_data_block) ||
3983 goal >= ext4_blocks_count(es))
3984 goal = le32_to_cpu(es->s_first_data_block);
3985 ext4_get_group_no_and_offset(sb, goal, &group, &block);
3986
3987 /* set up allocation goals */
Theodore Ts'o833576b2009-07-13 09:45:52 -04003988 memset(ac, 0, sizeof(struct ext4_allocation_context));
Alex Tomasc9de5602008-01-29 00:19:52 -05003989 ac->ac_b_ex.fe_logical = ar->logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05003990 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05003991 ac->ac_sb = sb;
3992 ac->ac_inode = ar->inode;
3993 ac->ac_o_ex.fe_logical = ar->logical;
3994 ac->ac_o_ex.fe_group = group;
3995 ac->ac_o_ex.fe_start = block;
3996 ac->ac_o_ex.fe_len = len;
3997 ac->ac_g_ex.fe_logical = ar->logical;
3998 ac->ac_g_ex.fe_group = group;
3999 ac->ac_g_ex.fe_start = block;
4000 ac->ac_g_ex.fe_len = len;
Alex Tomasc9de5602008-01-29 00:19:52 -05004001 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004002
4003 /* we have to define context: we'll we work with a file or
4004 * locality group. this is a policy, actually */
4005 ext4_mb_group_or_file(ac);
4006
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004007 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004008 "left: %u/%u, right %u/%u to %swritable\n",
4009 (unsigned) ar->len, (unsigned) ar->logical,
4010 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4011 (unsigned) ar->lleft, (unsigned) ar->pleft,
4012 (unsigned) ar->lright, (unsigned) ar->pright,
4013 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4014 return 0;
4015
4016}
4017
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004018static noinline_for_stack void
4019ext4_mb_discard_lg_preallocations(struct super_block *sb,
4020 struct ext4_locality_group *lg,
4021 int order, int total_entries)
4022{
4023 ext4_group_t group = 0;
4024 struct ext4_buddy e4b;
4025 struct list_head discard_list;
4026 struct ext4_prealloc_space *pa, *tmp;
4027 struct ext4_allocation_context *ac;
4028
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004029 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004030
4031 INIT_LIST_HEAD(&discard_list);
4032 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004033 if (ac)
4034 ac->ac_sb = sb;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004035
4036 spin_lock(&lg->lg_prealloc_lock);
4037 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4038 pa_inode_list) {
4039 spin_lock(&pa->pa_lock);
4040 if (atomic_read(&pa->pa_count)) {
4041 /*
4042 * This is the pa that we just used
4043 * for block allocation. So don't
4044 * free that
4045 */
4046 spin_unlock(&pa->pa_lock);
4047 continue;
4048 }
4049 if (pa->pa_deleted) {
4050 spin_unlock(&pa->pa_lock);
4051 continue;
4052 }
4053 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004054 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004055
4056 /* seems this one can be freed ... */
4057 pa->pa_deleted = 1;
4058 spin_unlock(&pa->pa_lock);
4059
4060 list_del_rcu(&pa->pa_inode_list);
4061 list_add(&pa->u.pa_tmp_list, &discard_list);
4062
4063 total_entries--;
4064 if (total_entries <= 5) {
4065 /*
4066 * we want to keep only 5 entries
4067 * allowing it to grow to 8. This
4068 * mak sure we don't call discard
4069 * soon for this list.
4070 */
4071 break;
4072 }
4073 }
4074 spin_unlock(&lg->lg_prealloc_lock);
4075
4076 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4077
4078 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4079 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4080 ext4_error(sb, __func__, "Error in loading buddy "
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05004081 "information for %u", group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004082 continue;
4083 }
4084 ext4_lock_group(sb, group);
4085 list_del(&pa->pa_group_list);
4086 ext4_mb_release_group_pa(&e4b, pa, ac);
4087 ext4_unlock_group(sb, group);
4088
4089 ext4_mb_release_desc(&e4b);
4090 list_del(&pa->u.pa_tmp_list);
4091 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4092 }
4093 if (ac)
4094 kmem_cache_free(ext4_ac_cachep, ac);
4095}
4096
4097/*
4098 * We have incremented pa_count. So it cannot be freed at this
4099 * point. Also we hold lg_mutex. So no parallel allocation is
4100 * possible from this lg. That means pa_free cannot be updated.
4101 *
4102 * A parallel ext4_mb_discard_group_preallocations is possible.
4103 * which can cause the lg_prealloc_list to be updated.
4104 */
4105
4106static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4107{
4108 int order, added = 0, lg_prealloc_count = 1;
4109 struct super_block *sb = ac->ac_sb;
4110 struct ext4_locality_group *lg = ac->ac_lg;
4111 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4112
4113 order = fls(pa->pa_free) - 1;
4114 if (order > PREALLOC_TB_SIZE - 1)
4115 /* The max size of hash table is PREALLOC_TB_SIZE */
4116 order = PREALLOC_TB_SIZE - 1;
4117 /* Add the prealloc space to lg */
4118 rcu_read_lock();
4119 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4120 pa_inode_list) {
4121 spin_lock(&tmp_pa->pa_lock);
4122 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004123 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004124 continue;
4125 }
4126 if (!added && pa->pa_free < tmp_pa->pa_free) {
4127 /* Add to the tail of the previous entry */
4128 list_add_tail_rcu(&pa->pa_inode_list,
4129 &tmp_pa->pa_inode_list);
4130 added = 1;
4131 /*
4132 * we want to count the total
4133 * number of entries in the list
4134 */
4135 }
4136 spin_unlock(&tmp_pa->pa_lock);
4137 lg_prealloc_count++;
4138 }
4139 if (!added)
4140 list_add_tail_rcu(&pa->pa_inode_list,
4141 &lg->lg_prealloc_list[order]);
4142 rcu_read_unlock();
4143
4144 /* Now trim the list to be not more than 8 elements */
4145 if (lg_prealloc_count > 8) {
4146 ext4_mb_discard_lg_preallocations(sb, lg,
4147 order, lg_prealloc_count);
4148 return;
4149 }
4150 return ;
4151}
4152
Alex Tomasc9de5602008-01-29 00:19:52 -05004153/*
4154 * release all resource we used in allocation
4155 */
4156static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4157{
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004158 struct ext4_prealloc_space *pa = ac->ac_pa;
4159 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004160 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004161 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004162 spin_lock(&pa->pa_lock);
4163 pa->pa_pstart += ac->ac_b_ex.fe_len;
4164 pa->pa_lstart += ac->ac_b_ex.fe_len;
4165 pa->pa_free -= ac->ac_b_ex.fe_len;
4166 pa->pa_len -= ac->ac_b_ex.fe_len;
4167 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004168 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004169 }
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004170 if (ac->alloc_semp)
4171 up_read(ac->alloc_semp);
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004172 if (pa) {
4173 /*
4174 * We want to add the pa to the right bucket.
4175 * Remove it from the list and while adding
4176 * make sure the list to which we are adding
4177 * doesn't grow big. We need to release
4178 * alloc_semp before calling ext4_mb_add_n_trim()
4179 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004180 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004181 spin_lock(pa->pa_obj_lock);
4182 list_del_rcu(&pa->pa_inode_list);
4183 spin_unlock(pa->pa_obj_lock);
4184 ext4_mb_add_n_trim(ac);
4185 }
4186 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4187 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004188 if (ac->ac_bitmap_page)
4189 page_cache_release(ac->ac_bitmap_page);
4190 if (ac->ac_buddy_page)
4191 page_cache_release(ac->ac_buddy_page);
4192 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4193 mutex_unlock(&ac->ac_lg->lg_mutex);
4194 ext4_mb_collect_stats(ac);
4195 return 0;
4196}
4197
4198static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4199{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004200 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004201 int ret;
4202 int freed = 0;
4203
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004204 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004205 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004206 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4207 freed += ret;
4208 needed -= ret;
4209 }
4210
4211 return freed;
4212}
4213
4214/*
4215 * Main entry point into mballoc to allocate blocks
4216 * it tries to use preallocation first, then falls back
4217 * to usual allocation
4218 */
4219ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4220 struct ext4_allocation_request *ar, int *errp)
4221{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004222 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004223 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004224 struct ext4_sb_info *sbi;
4225 struct super_block *sb;
4226 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004227 unsigned int inquota = 0;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004228 unsigned int reserv_blks = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004229
4230 sb = ar->inode->i_sb;
4231 sbi = EXT4_SB(sb);
4232
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004233 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004234
Mingming Cao60e58e02009-01-22 18:13:05 +01004235 /*
4236 * For delayed allocation, we could skip the ENOSPC and
4237 * EDQUOT check, as blocks and quotas have been already
4238 * reserved when data being copied into pagecache.
4239 */
4240 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4241 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4242 else {
4243 /* Without delayed allocation we need to verify
4244 * there is enough free blocks to do block allocation
4245 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004246 */
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004247 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4248 /* let others to free the space */
4249 yield();
4250 ar->len = ar->len >> 1;
4251 }
4252 if (!ar->len) {
Aneesh Kumar K.Va30d5422008-10-09 10:56:23 -04004253 *errp = -ENOSPC;
4254 return 0;
4255 }
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004256 reserv_blks = ar->len;
Jan Karaa269eb12009-01-26 17:04:39 +01004257 while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) {
Mingming Cao60e58e02009-01-22 18:13:05 +01004258 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4259 ar->len--;
4260 }
4261 inquota = ar->len;
4262 if (ar->len == 0) {
4263 *errp = -EDQUOT;
4264 goto out3;
4265 }
Mingming Caod2a17632008-07-14 17:52:37 -04004266 }
Mingming Caod2a17632008-07-14 17:52:37 -04004267
Eric Sandeen256bdb42008-02-10 01:13:33 -05004268 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004269 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004270 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004271 *errp = -ENOMEM;
Shen Feng363d4252008-07-11 19:27:31 -04004272 goto out1;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004273 }
4274
Eric Sandeen256bdb42008-02-10 01:13:33 -05004275 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004276 if (*errp) {
4277 ar->len = 0;
Shen Feng363d4252008-07-11 19:27:31 -04004278 goto out2;
Alex Tomasc9de5602008-01-29 00:19:52 -05004279 }
4280
Eric Sandeen256bdb42008-02-10 01:13:33 -05004281 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4282 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004283 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4284 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004285repeat:
4286 /* allocate space in core */
Eric Sandeen256bdb42008-02-10 01:13:33 -05004287 ext4_mb_regular_allocator(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004288
4289 /* as we've just preallocated more space than
4290 * user requested orinally, we store allocated
4291 * space in a special descriptor */
Eric Sandeen256bdb42008-02-10 01:13:33 -05004292 if (ac->ac_status == AC_STATUS_FOUND &&
4293 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4294 ext4_mb_new_preallocation(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004295 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004296 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004297 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004298 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004299 /*
4300 * drop the reference that we took
4301 * in ext4_mb_use_best_found
4302 */
4303 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004304 ac->ac_b_ex.fe_group = 0;
4305 ac->ac_b_ex.fe_start = 0;
4306 ac->ac_b_ex.fe_len = 0;
4307 ac->ac_status = AC_STATUS_CONTINUE;
4308 goto repeat;
4309 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004310 ext4_discard_allocated_blocks(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004311 ac->ac_b_ex.fe_len = 0;
4312 ar->len = 0;
4313 ext4_mb_show_ac(ac);
4314 } else {
4315 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4316 ar->len = ac->ac_b_ex.fe_len;
4317 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004318 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004319 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004320 if (freed)
4321 goto repeat;
4322 *errp = -ENOSPC;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004323 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004324 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004325 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004326 }
4327
Eric Sandeen256bdb42008-02-10 01:13:33 -05004328 ext4_mb_release_context(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004329
Shen Feng363d4252008-07-11 19:27:31 -04004330out2:
4331 kmem_cache_free(ext4_ac_cachep, ac);
4332out1:
Mingming Cao60e58e02009-01-22 18:13:05 +01004333 if (inquota && ar->len < inquota)
Jan Karaa269eb12009-01-26 17:04:39 +01004334 vfs_dq_free_block(ar->inode, inquota - ar->len);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004335out3:
4336 if (!ar->len) {
4337 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4338 /* release all the reserved blocks if non delalloc */
4339 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4340 reserv_blks);
4341 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004342
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004343 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004344
Alex Tomasc9de5602008-01-29 00:19:52 -05004345 return block;
4346}
Alex Tomasc9de5602008-01-29 00:19:52 -05004347
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004348/*
4349 * We can merge two free data extents only if the physical blocks
4350 * are contiguous, AND the extents were freed by the same transaction,
4351 * AND the blocks are associated with the same group.
4352 */
4353static int can_merge(struct ext4_free_data *entry1,
4354 struct ext4_free_data *entry2)
4355{
4356 if ((entry1->t_tid == entry2->t_tid) &&
4357 (entry1->group == entry2->group) &&
4358 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4359 return 1;
4360 return 0;
4361}
4362
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004363static noinline_for_stack int
4364ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004365 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004366{
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004367 ext4_grpblk_t block;
4368 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004369 struct ext4_group_info *db = e4b->bd_info;
4370 struct super_block *sb = e4b->bd_sb;
4371 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004372 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4373 struct rb_node *parent = NULL, *new_node;
4374
Frank Mayhar03901312009-01-07 00:06:22 -05004375 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004376 BUG_ON(e4b->bd_bitmap_page == NULL);
4377 BUG_ON(e4b->bd_buddy_page == NULL);
4378
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004379 new_node = &new_entry->node;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004380 block = new_entry->start_blk;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004381
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004382 if (!*n) {
4383 /* first free block exent. We need to
4384 protect buddy cache from being freed,
4385 * otherwise we'll refresh it from
4386 * on-disk bitmap and lose not-yet-available
4387 * blocks */
4388 page_cache_get(e4b->bd_buddy_page);
4389 page_cache_get(e4b->bd_bitmap_page);
4390 }
4391 while (*n) {
4392 parent = *n;
4393 entry = rb_entry(parent, struct ext4_free_data, node);
4394 if (block < entry->start_blk)
4395 n = &(*n)->rb_left;
4396 else if (block >= (entry->start_blk + entry->count))
4397 n = &(*n)->rb_right;
4398 else {
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05004399 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4400 "Double free of blocks %d (%d %d)",
4401 block, entry->start_blk, entry->count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004402 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004403 }
4404 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004405
4406 rb_link_node(new_node, parent, n);
4407 rb_insert_color(new_node, &db->bb_free_root);
4408
4409 /* Now try to see the extent can be merged to left and right */
4410 node = rb_prev(new_node);
4411 if (node) {
4412 entry = rb_entry(node, struct ext4_free_data, node);
4413 if (can_merge(entry, new_entry)) {
4414 new_entry->start_blk = entry->start_blk;
4415 new_entry->count += entry->count;
4416 rb_erase(node, &(db->bb_free_root));
4417 spin_lock(&sbi->s_md_lock);
4418 list_del(&entry->list);
4419 spin_unlock(&sbi->s_md_lock);
4420 kmem_cache_free(ext4_free_ext_cachep, entry);
4421 }
4422 }
4423
4424 node = rb_next(new_node);
4425 if (node) {
4426 entry = rb_entry(node, struct ext4_free_data, node);
4427 if (can_merge(new_entry, entry)) {
4428 new_entry->count += entry->count;
4429 rb_erase(node, &(db->bb_free_root));
4430 spin_lock(&sbi->s_md_lock);
4431 list_del(&entry->list);
4432 spin_unlock(&sbi->s_md_lock);
4433 kmem_cache_free(ext4_free_ext_cachep, entry);
4434 }
4435 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004436 /* Add the extent to transaction's private list */
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004437 spin_lock(&sbi->s_md_lock);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004438 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004439 spin_unlock(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004440 return 0;
4441}
4442
Theodore Ts'o44338712009-11-22 07:44:56 -05004443/**
4444 * ext4_free_blocks() -- Free given blocks and update quota
4445 * @handle: handle for this transaction
4446 * @inode: inode
4447 * @block: start physical block to free
4448 * @count: number of blocks to count
4449 * @metadata: Are these metadata blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004450 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004451void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004452 struct buffer_head *bh, ext4_fsblk_t block,
4453 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004454{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004455 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004456 struct super_block *sb = inode->i_sb;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004457 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004458 struct ext4_group_desc *gdp;
4459 struct ext4_super_block *es;
Theodore Ts'o44338712009-11-22 07:44:56 -05004460 unsigned long freed = 0;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004461 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004462 ext4_grpblk_t bit;
4463 struct buffer_head *gd_bh;
4464 ext4_group_t block_group;
4465 struct ext4_sb_info *sbi;
4466 struct ext4_buddy e4b;
4467 int err = 0;
4468 int ret;
4469
Theodore Ts'oe6362602009-11-23 07:17:05 -05004470 if (bh) {
4471 if (block)
4472 BUG_ON(block != bh->b_blocknr);
4473 else
4474 block = bh->b_blocknr;
4475 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004476
Alex Tomasc9de5602008-01-29 00:19:52 -05004477 sbi = EXT4_SB(sb);
4478 es = EXT4_SB(sb)->s_es;
Theodore Ts'o9084d472009-11-22 20:48:42 -05004479 if (!ext4_data_block_valid(sbi, block, count)) {
Harvey Harrison46e665e2008-04-17 10:38:59 -04004480 ext4_error(sb, __func__,
Alex Tomasc9de5602008-01-29 00:19:52 -05004481 "Freeing blocks not in datazone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004482 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004483 goto error_return;
4484 }
4485
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004486 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004487 trace_ext4_free_blocks(inode, block, count, flags);
4488
4489 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4490 struct buffer_head *tbh = bh;
4491 int i;
4492
4493 BUG_ON(bh && (count > 1));
4494
4495 for (i = 0; i < count; i++) {
4496 if (!bh)
4497 tbh = sb_find_get_block(inode->i_sb,
4498 block + i);
4499 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4500 inode, tbh, block + i);
4501 }
4502 }
4503
4504 /*
4505 * We need to make sure we don't reuse the freed block until
4506 * after the transaction is committed, which we can do by
4507 * treating the block as metadata, below. We make an
4508 * exception if the inode is to be written in writeback mode
4509 * since writeback mode has weak data consistency guarantees.
4510 */
4511 if (!ext4_should_writeback_data(inode))
4512 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004513
Eric Sandeen256bdb42008-02-10 01:13:33 -05004514 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4515 if (ac) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004516 ac->ac_inode = inode;
4517 ac->ac_sb = sb;
4518 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004519
4520do_more:
4521 overflow = 0;
4522 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4523
4524 /*
4525 * Check to see if we are freeing blocks across a group
4526 * boundary.
4527 */
4528 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4529 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4530 count -= overflow;
4531 }
Theodore Ts'o574ca172008-07-11 19:27:31 -04004532 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004533 if (!bitmap_bh) {
4534 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004535 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004536 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004537 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004538 if (!gdp) {
4539 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004540 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004541 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004542
4543 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4544 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4545 in_range(block, ext4_inode_table(sb, gdp),
4546 EXT4_SB(sb)->s_itb_per_group) ||
4547 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4548 EXT4_SB(sb)->s_itb_per_group)) {
4549
Harvey Harrison46e665e2008-04-17 10:38:59 -04004550 ext4_error(sb, __func__,
Alex Tomasc9de5602008-01-29 00:19:52 -05004551 "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004552 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004553 /* err = 0. ext4_std_error should be a no op */
4554 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004555 }
4556
4557 BUFFER_TRACE(bitmap_bh, "getting write access");
4558 err = ext4_journal_get_write_access(handle, bitmap_bh);
4559 if (err)
4560 goto error_return;
4561
4562 /*
4563 * We are about to modify some metadata. Call the journal APIs
4564 * to unshare ->b_data if a currently-committing transaction is
4565 * using it
4566 */
4567 BUFFER_TRACE(gd_bh, "get_write_access");
4568 err = ext4_journal_get_write_access(handle, gd_bh);
4569 if (err)
4570 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004571#ifdef AGGRESSIVE_CHECK
4572 {
4573 int i;
4574 for (i = 0; i < count; i++)
4575 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4576 }
4577#endif
Eric Sandeen256bdb42008-02-10 01:13:33 -05004578 if (ac) {
4579 ac->ac_b_ex.fe_group = block_group;
4580 ac->ac_b_ex.fe_start = bit;
4581 ac->ac_b_ex.fe_len = count;
Theodore Ts'o296c3552009-09-30 00:32:42 -04004582 trace_ext4_mballoc_free(ac);
Eric Sandeen256bdb42008-02-10 01:13:33 -05004583 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004584
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004585 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4586 if (err)
4587 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004588
4589 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004590 struct ext4_free_data *new_entry;
4591 /*
4592 * blocks being freed are metadata. these blocks shouldn't
4593 * be used until this transaction is committed
4594 */
4595 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4596 new_entry->start_blk = bit;
4597 new_entry->group = block_group;
4598 new_entry->count = count;
4599 new_entry->t_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004600
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004601 ext4_lock_group(sb, block_group);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004602 mb_clear_bits(bitmap_bh->b_data, bit, count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004603 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004604 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004605 /* need to update group_info->bb_free and bitmap
4606 * with group lock held. generate_buddy look at
4607 * them with group lock_held
4608 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004609 ext4_lock_group(sb, block_group);
4610 mb_clear_bits(bitmap_bh->b_data, bit, count);
Shen Feng7e5a8cd2008-07-13 21:03:31 -04004611 mb_free_blocks(inode, &e4b, bit, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004612 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004613 }
4614
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05004615 ret = ext4_free_blks_count(sb, gdp) + count;
4616 ext4_free_blks_set(sb, gdp, ret);
Alex Tomasc9de5602008-01-29 00:19:52 -05004617 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004618 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004619 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4620
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004621 if (sbi->s_log_groups_per_flex) {
4622 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o9f24e422009-03-04 19:09:10 -05004623 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004624 }
4625
Alex Tomasc9de5602008-01-29 00:19:52 -05004626 ext4_mb_release_desc(&e4b);
4627
Theodore Ts'o44338712009-11-22 07:44:56 -05004628 freed += count;
Alex Tomasc9de5602008-01-29 00:19:52 -05004629
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004630 /* We dirtied the bitmap block */
4631 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4632 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4633
Alex Tomasc9de5602008-01-29 00:19:52 -05004634 /* And the group descriptor block */
4635 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004636 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004637 if (!err)
4638 err = ret;
4639
4640 if (overflow && !err) {
4641 block += count;
4642 count = overflow;
4643 put_bh(bitmap_bh);
4644 goto do_more;
4645 }
4646 sb->s_dirt = 1;
4647error_return:
Theodore Ts'o44338712009-11-22 07:44:56 -05004648 if (freed)
4649 vfs_dq_free_block(inode, freed);
Alex Tomasc9de5602008-01-29 00:19:52 -05004650 brelse(bitmap_bh);
4651 ext4_std_error(sb, err);
Eric Sandeen256bdb42008-02-10 01:13:33 -05004652 if (ac)
4653 kmem_cache_free(ext4_ac_cachep, ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004654 return;
4655}