blob: 17a5a57c415a2dcdd36104c08473438c238f38ee [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090026#include <linux/slab.h>
Theodore Ts'o9bffad12009-06-17 11:48:11 -040027#include <trace/events/ext4.h>
28
Alex Tomasc9de5602008-01-29 00:19:52 -050029/*
30 * MUSTDO:
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
33 *
34 * TODO v4:
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
38 * - quota
39 * - reservation for superuser
40 *
41 * TODO v3:
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
46 * - error handling
47 */
48
49/*
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
52 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040053 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
Alex Tomasc9de5602008-01-29 00:19:52 -050062 *
63 * The main motivation for having small file use group preallocation is to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040064 * ensure that we have small files closer together on the disk.
Alex Tomasc9de5602008-01-29 00:19:52 -050065 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040066 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
69 * represented as:
Alex Tomasc9de5602008-01-29 00:19:52 -050070 *
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
Daniel Mack1537a362010-01-29 15:57:49 +080073 * pa_len -> length for this prealloc space
Alex Tomasc9de5602008-01-29 00:19:52 -050074 * pa_free -> free space available in this prealloc space
75 *
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
Tao Macaaf7a22011-07-11 18:42:42 -040078 * space we will consume the particular prealloc space. This makes sure that
79 * we have contiguous physical blocks representing the file blocks
Alex Tomasc9de5602008-01-29 00:19:52 -050080 *
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
83 * pa_free.
84 *
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
Tao Macaaf7a22011-07-11 18:42:42 -040087 * prealloc space. These are per CPU prealloc list represented as
Alex Tomasc9de5602008-01-29 00:19:52 -050088 *
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
90 *
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
93 *
94 * The locality group prealloc space is used looking at whether we have
Lucas De Marchi25985ed2011-03-30 22:57:33 -030095 * enough free space (pa_free) within the prealloc space.
Alex Tomasc9de5602008-01-29 00:19:52 -050096 *
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
105 * inode as:
106 *
107 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500109 *
110 *
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
115 *
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
118 *
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
122 *
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
Alex Tomasc9de5602008-01-29 00:19:52 -0500130 * 512 blocks. This can be tuned via
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400131 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
Alex Tomasc9de5602008-01-29 00:19:52 -0500132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400134 * the smallest multiple of the stripe value (sbi->s_stripe) which is
135 * greater than the default mb_group_prealloc.
Alex Tomasc9de5602008-01-29 00:19:52 -0500136 *
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400137 * The regular allocator (using the buddy cache) supports a few tunables.
Alex Tomasc9de5602008-01-29 00:19:52 -0500138 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400139 * /sys/fs/ext4/<partition>/mb_min_to_scan
140 * /sys/fs/ext4/<partition>/mb_max_to_scan
141 * /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -0500142 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400143 * The regular allocator uses buddy scan only if the request len is power of
Alex Tomasc9de5602008-01-29 00:19:52 -0500144 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
145 * value of s_mb_order2_reqs can be tuned via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400146 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200147 * stripe size (sbi->s_stripe), we try to search for contiguous block in
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400148 * stripe size. This should result in better allocation on RAID setups. If
149 * not, we search in the specific group using bitmap for best extents. The
150 * tunable min_to_scan and max_to_scan control the behaviour here.
Alex Tomasc9de5602008-01-29 00:19:52 -0500151 * min_to_scan indicate how long the mballoc __must__ look for a best
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400152 * extent and max_to_scan indicates how long the mballoc __can__ look for a
Alex Tomasc9de5602008-01-29 00:19:52 -0500153 * best extent in the found extents. Searching for the blocks starts with
154 * the group specified as the goal value in allocation context via
155 * ac_g_ex. Each group is first checked based on the criteria whether it
Tao Macaaf7a22011-07-11 18:42:42 -0400156 * can be used for allocation. ext4_mb_good_group explains how the groups are
Alex Tomasc9de5602008-01-29 00:19:52 -0500157 * checked.
158 *
159 * Both the prealloc space are getting populated as above. So for the first
160 * request we will hit the buddy cache which will result in this prealloc
161 * space getting filled. The prealloc space is then later used for the
162 * subsequent request.
163 */
164
165/*
166 * mballoc operates on the following data:
167 * - on-disk bitmap
168 * - in-core buddy (actually includes buddy and bitmap)
169 * - preallocation descriptors (PAs)
170 *
171 * there are two types of preallocations:
172 * - inode
173 * assiged to specific inode and can be used for this inode only.
174 * it describes part of inode's space preallocated to specific
175 * physical blocks. any block from that preallocated can be used
176 * independent. the descriptor just tracks number of blocks left
177 * unused. so, before taking some block from descriptor, one must
178 * make sure corresponded logical block isn't allocated yet. this
179 * also means that freeing any block within descriptor's range
180 * must discard all preallocated blocks.
181 * - locality group
182 * assigned to specific locality group which does not translate to
183 * permanent set of inodes: inode can join and leave group. space
184 * from this type of preallocation can be used for any inode. thus
185 * it's consumed from the beginning to the end.
186 *
187 * relation between them can be expressed as:
188 * in-core buddy = on-disk bitmap + preallocation descriptors
189 *
190 * this mean blocks mballoc considers used are:
191 * - allocated blocks (persistent)
192 * - preallocated blocks (non-persistent)
193 *
194 * consistency in mballoc world means that at any time a block is either
195 * free or used in ALL structures. notice: "any time" should not be read
196 * literally -- time is discrete and delimited by locks.
197 *
198 * to keep it simple, we don't use block numbers, instead we count number of
199 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 *
201 * all operations can be expressed as:
202 * - init buddy: buddy = on-disk + PAs
203 * - new PA: buddy += N; PA = N
204 * - use inode PA: on-disk += N; PA -= N
205 * - discard inode PA buddy -= on-disk - PA; PA = 0
206 * - use locality group PA on-disk += N; PA -= N
207 * - discard locality group PA buddy -= PA; PA = 0
208 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
209 * is used in real operation because we can't know actual used
210 * bits from PA, only from on-disk bitmap
211 *
212 * if we follow this strict logic, then all operations above should be atomic.
213 * given some of them can block, we'd have to use something like semaphores
214 * killing performance on high-end SMP hardware. let's try to relax it using
215 * the following knowledge:
216 * 1) if buddy is referenced, it's already initialized
217 * 2) while block is used in buddy and the buddy is referenced,
218 * nobody can re-allocate that block
219 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
220 * bit set and PA claims same block, it's OK. IOW, one can set bit in
221 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
222 * block
223 *
224 * so, now we're building a concurrency table:
225 * - init buddy vs.
226 * - new PA
227 * blocks for PA are allocated in the buddy, buddy must be referenced
228 * until PA is linked to allocation group to avoid concurrent buddy init
229 * - use inode PA
230 * we need to make sure that either on-disk bitmap or PA has uptodate data
231 * given (3) we care that PA-=N operation doesn't interfere with init
232 * - discard inode PA
233 * the simplest way would be to have buddy initialized by the discard
234 * - use locality group PA
235 * again PA-=N must be serialized with init
236 * - discard locality group PA
237 * the simplest way would be to have buddy initialized by the discard
238 * - new PA vs.
239 * - use inode PA
240 * i_data_sem serializes them
241 * - discard inode PA
242 * discard process must wait until PA isn't used by another process
243 * - use locality group PA
244 * some mutex should serialize them
245 * - discard locality group PA
246 * discard process must wait until PA isn't used by another process
247 * - use inode PA
248 * - use inode PA
249 * i_data_sem or another mutex should serializes them
250 * - discard inode PA
251 * discard process must wait until PA isn't used by another process
252 * - use locality group PA
253 * nothing wrong here -- they're different PAs covering different blocks
254 * - discard locality group PA
255 * discard process must wait until PA isn't used by another process
256 *
257 * now we're ready to make few consequences:
258 * - PA is referenced and while it is no discard is possible
259 * - PA is referenced until block isn't marked in on-disk bitmap
260 * - PA changes only after on-disk bitmap
261 * - discard must not compete with init. either init is done before
262 * any discard or they're serialized somehow
263 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 *
265 * a special case when we've used PA to emptiness. no need to modify buddy
266 * in this case, but we should care about concurrent init
267 *
268 */
269
270 /*
271 * Logic in few words:
272 *
273 * - allocation:
274 * load group
275 * find blocks
276 * mark bits in on-disk bitmap
277 * release group
278 *
279 * - use preallocation:
280 * find proper PA (per-inode or group)
281 * load group
282 * mark bits in on-disk bitmap
283 * release group
284 * release PA
285 *
286 * - free:
287 * load group
288 * mark bits in on-disk bitmap
289 * release group
290 *
291 * - discard preallocations in group:
292 * mark PAs deleted
293 * move them onto local list
294 * load on-disk bitmap
295 * load group
296 * remove PA from object (inode or locality group)
297 * mark free blocks in-core
298 *
299 * - discard inode's preallocations:
300 */
301
302/*
303 * Locking rules
304 *
305 * Locks:
306 * - bitlock on a group (group)
307 * - object (inode/locality) (object)
308 * - per-pa lock (pa)
309 *
310 * Paths:
311 * - new pa
312 * object
313 * group
314 *
315 * - find and use pa:
316 * pa
317 *
318 * - release consumed pa:
319 * pa
320 * group
321 * object
322 *
323 * - generate in-core bitmap:
324 * group
325 * pa
326 *
327 * - discard all for given object (inode, locality group):
328 * object
329 * pa
330 * group
331 *
332 * - discard all for given group:
333 * group
334 * pa
335 * group
336 * object
337 *
338 */
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500339static struct kmem_cache *ext4_pspace_cachep;
340static struct kmem_cache *ext4_ac_cachep;
341static struct kmem_cache *ext4_free_ext_cachep;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400342
343/* We create slab caches for groupinfo data structures based on the
344 * superblock block size. There will be one per mounted filesystem for
345 * each unique s_blocksize_bits */
Eric Sandeen2892c152011-02-12 08:12:18 -0500346#define NR_GRPINFO_CACHES 8
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400347static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
348
Eric Sandeen2892c152011-02-12 08:12:18 -0500349static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
350 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
351 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
352 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
353};
354
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500355static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
356 ext4_group_t group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500357static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
358 ext4_group_t group);
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500359static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
360
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500361static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
362{
Alex Tomasc9de5602008-01-29 00:19:52 -0500363#if BITS_PER_LONG == 64
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500364 *bit += ((unsigned long) addr & 7UL) << 3;
365 addr = (void *) ((unsigned long) addr & ~7UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500366#elif BITS_PER_LONG == 32
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500367 *bit += ((unsigned long) addr & 3UL) << 3;
368 addr = (void *) ((unsigned long) addr & ~3UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500369#else
370#error "how many bits you are?!"
371#endif
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500372 return addr;
373}
Alex Tomasc9de5602008-01-29 00:19:52 -0500374
375static inline int mb_test_bit(int bit, void *addr)
376{
377 /*
378 * ext4_test_bit on architecture like powerpc
379 * needs unsigned long aligned address
380 */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500381 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500382 return ext4_test_bit(bit, addr);
383}
384
385static inline void mb_set_bit(int bit, void *addr)
386{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500387 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500388 ext4_set_bit(bit, addr);
389}
390
Alex Tomasc9de5602008-01-29 00:19:52 -0500391static inline void mb_clear_bit(int bit, void *addr)
392{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500393 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500394 ext4_clear_bit(bit, addr);
395}
396
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500397static inline int mb_find_next_zero_bit(void *addr, int max, int start)
398{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400399 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500400 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400401 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500402 start += fix;
403
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400404 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
405 if (ret > max)
406 return max;
407 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500408}
409
410static inline int mb_find_next_bit(void *addr, int max, int start)
411{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400412 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500413 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400414 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500415 start += fix;
416
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400417 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
418 if (ret > max)
419 return max;
420 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500421}
422
Alex Tomasc9de5602008-01-29 00:19:52 -0500423static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
424{
425 char *bb;
426
Alex Tomasc9de5602008-01-29 00:19:52 -0500427 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
428 BUG_ON(max == NULL);
429
430 if (order > e4b->bd_blkbits + 1) {
431 *max = 0;
432 return NULL;
433 }
434
435 /* at order 0 we see each particular block */
Coly Li84b775a2011-02-24 12:51:59 -0500436 if (order == 0) {
437 *max = 1 << (e4b->bd_blkbits + 3);
Alex Tomasc9de5602008-01-29 00:19:52 -0500438 return EXT4_MB_BITMAP(e4b);
Coly Li84b775a2011-02-24 12:51:59 -0500439 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500440
441 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
442 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
443
444 return bb;
445}
446
447#ifdef DOUBLE_CHECK
448static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
449 int first, int count)
450{
451 int i;
452 struct super_block *sb = e4b->bd_sb;
453
454 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
455 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400456 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500457 for (i = 0; i < count; i++) {
458 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
459 ext4_fsblk_t blocknr;
Akinobu Mita5661bd62010-03-03 23:53:39 -0500460
461 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500462 blocknr += first + i;
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -0500463 ext4_grp_locked_error(sb, e4b->bd_group,
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400464 inode ? inode->i_ino : 0,
465 blocknr,
466 "freeing block already freed "
467 "(bit %u)",
468 first + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500469 }
470 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
471 }
472}
473
474static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
475{
476 int i;
477
478 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
479 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400480 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500481 for (i = 0; i < count; i++) {
482 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
483 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
484 }
485}
486
487static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
488{
489 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
490 unsigned char *b1, *b2;
491 int i;
492 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
493 b2 = (unsigned char *) bitmap;
494 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
495 if (b1[i] != b2[i]) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -0400496 ext4_msg(e4b->bd_sb, KERN_ERR,
497 "corruption in group %u "
498 "at byte %u(%u): %x in copy != %x "
499 "on disk/prealloc",
500 e4b->bd_group, i, i * 8, b1[i], b2[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500501 BUG();
502 }
503 }
504 }
505}
506
507#else
508static inline void mb_free_blocks_double(struct inode *inode,
509 struct ext4_buddy *e4b, int first, int count)
510{
511 return;
512}
513static inline void mb_mark_used_double(struct ext4_buddy *e4b,
514 int first, int count)
515{
516 return;
517}
518static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
519{
520 return;
521}
522#endif
523
524#ifdef AGGRESSIVE_CHECK
525
526#define MB_CHECK_ASSERT(assert) \
527do { \
528 if (!(assert)) { \
529 printk(KERN_EMERG \
530 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
531 function, file, line, # assert); \
532 BUG(); \
533 } \
534} while (0)
535
536static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
537 const char *function, int line)
538{
539 struct super_block *sb = e4b->bd_sb;
540 int order = e4b->bd_blkbits + 1;
541 int max;
542 int max2;
543 int i;
544 int j;
545 int k;
546 int count;
547 struct ext4_group_info *grp;
548 int fragments = 0;
549 int fstart;
550 struct list_head *cur;
551 void *buddy;
552 void *buddy2;
553
Alex Tomasc9de5602008-01-29 00:19:52 -0500554 {
555 static int mb_check_counter;
556 if (mb_check_counter++ % 100 != 0)
557 return 0;
558 }
559
560 while (order > 1) {
561 buddy = mb_find_buddy(e4b, order, &max);
562 MB_CHECK_ASSERT(buddy);
563 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
564 MB_CHECK_ASSERT(buddy2);
565 MB_CHECK_ASSERT(buddy != buddy2);
566 MB_CHECK_ASSERT(max * 2 == max2);
567
568 count = 0;
569 for (i = 0; i < max; i++) {
570
571 if (mb_test_bit(i, buddy)) {
572 /* only single bit in buddy2 may be 1 */
573 if (!mb_test_bit(i << 1, buddy2)) {
574 MB_CHECK_ASSERT(
575 mb_test_bit((i<<1)+1, buddy2));
576 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
577 MB_CHECK_ASSERT(
578 mb_test_bit(i << 1, buddy2));
579 }
580 continue;
581 }
582
583 /* both bits in buddy2 must be 0 */
584 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
585 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
586
587 for (j = 0; j < (1 << order); j++) {
588 k = (i * (1 << order)) + j;
589 MB_CHECK_ASSERT(
590 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
591 }
592 count++;
593 }
594 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
595 order--;
596 }
597
598 fstart = -1;
599 buddy = mb_find_buddy(e4b, 0, &max);
600 for (i = 0; i < max; i++) {
601 if (!mb_test_bit(i, buddy)) {
602 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
603 if (fstart == -1) {
604 fragments++;
605 fstart = i;
606 }
607 continue;
608 }
609 fstart = -1;
610 /* check used bits only */
611 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
612 buddy2 = mb_find_buddy(e4b, j, &max2);
613 k = i >> j;
614 MB_CHECK_ASSERT(k < max2);
615 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
616 }
617 }
618 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
619 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
620
621 grp = ext4_get_group_info(sb, e4b->bd_group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500622 list_for_each(cur, &grp->bb_prealloc_list) {
623 ext4_group_t groupnr;
624 struct ext4_prealloc_space *pa;
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400625 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
626 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
Alex Tomasc9de5602008-01-29 00:19:52 -0500627 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400628 for (i = 0; i < pa->pa_len; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500629 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
630 }
631 return 0;
632}
633#undef MB_CHECK_ASSERT
634#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
Harvey Harrison46e665e2008-04-17 10:38:59 -0400635 __FILE__, __func__, __LINE__)
Alex Tomasc9de5602008-01-29 00:19:52 -0500636#else
637#define mb_check_buddy(e4b)
638#endif
639
Coly Li7c786052011-02-24 13:24:25 -0500640/*
641 * Divide blocks started from @first with length @len into
642 * smaller chunks with power of 2 blocks.
643 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
644 * then increase bb_counters[] for corresponded chunk size.
645 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500646static void ext4_mb_mark_free_simple(struct super_block *sb,
Eric Sandeena36b4492009-08-25 22:36:45 -0400647 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
Alex Tomasc9de5602008-01-29 00:19:52 -0500648 struct ext4_group_info *grp)
649{
650 struct ext4_sb_info *sbi = EXT4_SB(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400651 ext4_grpblk_t min;
652 ext4_grpblk_t max;
653 ext4_grpblk_t chunk;
Alex Tomasc9de5602008-01-29 00:19:52 -0500654 unsigned short border;
655
Valerie Clementb73fce62008-02-15 13:48:51 -0500656 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
Alex Tomasc9de5602008-01-29 00:19:52 -0500657
658 border = 2 << sb->s_blocksize_bits;
659
660 while (len > 0) {
661 /* find how many blocks can be covered since this position */
662 max = ffs(first | border) - 1;
663
664 /* find how many blocks of power 2 we need to mark */
665 min = fls(len) - 1;
666
667 if (max < min)
668 min = max;
669 chunk = 1 << min;
670
671 /* mark multiblock chunks only */
672 grp->bb_counters[min]++;
673 if (min > 0)
674 mb_clear_bit(first >> min,
675 buddy + sbi->s_mb_offsets[min]);
676
677 len -= chunk;
678 first += chunk;
679 }
680}
681
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400682/*
683 * Cache the order of the largest free extent we have available in this block
684 * group.
685 */
686static void
687mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
688{
689 int i;
690 int bits;
691
692 grp->bb_largest_free_order = -1; /* uninit */
693
694 bits = sb->s_blocksize_bits + 1;
695 for (i = bits; i >= 0; i--) {
696 if (grp->bb_counters[i] > 0) {
697 grp->bb_largest_free_order = i;
698 break;
699 }
700 }
701}
702
Eric Sandeen089ceec2009-07-05 22:17:31 -0400703static noinline_for_stack
704void ext4_mb_generate_buddy(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -0500705 void *buddy, void *bitmap, ext4_group_t group)
706{
707 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
Eric Sandeena36b4492009-08-25 22:36:45 -0400708 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
709 ext4_grpblk_t i = 0;
710 ext4_grpblk_t first;
711 ext4_grpblk_t len;
Alex Tomasc9de5602008-01-29 00:19:52 -0500712 unsigned free = 0;
713 unsigned fragments = 0;
714 unsigned long long period = get_cycles();
715
716 /* initialize buddy from bitmap which is aggregation
717 * of on-disk bitmap and preallocations */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500718 i = mb_find_next_zero_bit(bitmap, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -0500719 grp->bb_first_free = i;
720 while (i < max) {
721 fragments++;
722 first = i;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500723 i = mb_find_next_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500724 len = i - first;
725 free += len;
726 if (len > 1)
727 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
728 else
729 grp->bb_counters[0]++;
730 if (i < max)
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500731 i = mb_find_next_zero_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500732 }
733 grp->bb_fragments = fragments;
734
735 if (free != grp->bb_free) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400736 ext4_grp_locked_error(sb, group, 0, 0,
737 "%u blocks in bitmap, %u in gd",
738 free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500739 /*
740 * If we intent to continue, we consider group descritor
741 * corrupt and update bb_free using bitmap value
742 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500743 grp->bb_free = free;
744 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400745 mb_set_largest_free_order(sb, grp);
Alex Tomasc9de5602008-01-29 00:19:52 -0500746
747 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
748
749 period = get_cycles() - period;
750 spin_lock(&EXT4_SB(sb)->s_bal_lock);
751 EXT4_SB(sb)->s_mb_buddies_generated++;
752 EXT4_SB(sb)->s_mb_generation_time += period;
753 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
754}
755
756/* The buddy information is attached the buddy cache inode
757 * for convenience. The information regarding each group
758 * is loaded via ext4_mb_load_buddy. The information involve
759 * block bitmap and buddy information. The information are
760 * stored in the inode as
761 *
762 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500763 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500764 *
765 *
766 * one block each for bitmap and buddy information.
767 * So for each group we take up 2 blocks. A page can
768 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
769 * So it can have information regarding groups_per_page which
770 * is blocks_per_page/2
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400771 *
772 * Locking note: This routine takes the block group lock of all groups
773 * for this page; do not hold this lock when calling this routine!
Alex Tomasc9de5602008-01-29 00:19:52 -0500774 */
775
776static int ext4_mb_init_cache(struct page *page, char *incore)
777{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400778 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500779 int blocksize;
780 int blocks_per_page;
781 int groups_per_page;
782 int err = 0;
783 int i;
784 ext4_group_t first_group;
785 int first_block;
786 struct super_block *sb;
787 struct buffer_head *bhs;
788 struct buffer_head **bh;
789 struct inode *inode;
790 char *data;
791 char *bitmap;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400792 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -0500793
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400794 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500795
796 inode = page->mapping->host;
797 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400798 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500799 blocksize = 1 << inode->i_blkbits;
800 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
801
802 groups_per_page = blocks_per_page >> 1;
803 if (groups_per_page == 0)
804 groups_per_page = 1;
805
806 /* allocate buffer_heads to read bitmaps */
807 if (groups_per_page > 1) {
808 err = -ENOMEM;
809 i = sizeof(struct buffer_head *) * groups_per_page;
810 bh = kzalloc(i, GFP_NOFS);
811 if (bh == NULL)
812 goto out;
813 } else
814 bh = &bhs;
815
816 first_group = page->index * blocks_per_page / 2;
817
818 /* read all groups the page covers into the cache */
819 for (i = 0; i < groups_per_page; i++) {
820 struct ext4_group_desc *desc;
821
Theodore Ts'o8df96752009-05-01 08:50:38 -0400822 if (first_group + i >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500823 break;
824
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400825 grinfo = ext4_get_group_info(sb, first_group + i);
826 /*
827 * If page is uptodate then we came here after online resize
828 * which added some new uninitialized group info structs, so
829 * we must skip all initialized uptodate buddies on the page,
830 * which may be currently in use by an allocating task.
831 */
832 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
833 bh[i] = NULL;
834 continue;
835 }
836
Alex Tomasc9de5602008-01-29 00:19:52 -0500837 err = -EIO;
838 desc = ext4_get_group_desc(sb, first_group + i, NULL);
839 if (desc == NULL)
840 goto out;
841
842 err = -ENOMEM;
843 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
844 if (bh[i] == NULL)
845 goto out;
846
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500847 if (bitmap_uptodate(bh[i]))
Alex Tomasc9de5602008-01-29 00:19:52 -0500848 continue;
849
Frederic Bohec806e682008-10-10 08:09:18 -0400850 lock_buffer(bh[i]);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500851 if (bitmap_uptodate(bh[i])) {
852 unlock_buffer(bh[i]);
853 continue;
854 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400855 ext4_lock_group(sb, first_group + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500856 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
857 ext4_init_block_bitmap(sb, bh[i],
858 first_group + i, desc);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500859 set_bitmap_uptodate(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500860 set_buffer_uptodate(bh[i]);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400861 ext4_unlock_group(sb, first_group + i);
Aneesh Kumar K.V3300bed2009-01-03 22:33:39 -0500862 unlock_buffer(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500863 continue;
864 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -0400865 ext4_unlock_group(sb, first_group + i);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500866 if (buffer_uptodate(bh[i])) {
867 /*
868 * if not uninit if bh is uptodate,
869 * bitmap is also uptodate
870 */
871 set_bitmap_uptodate(bh[i]);
872 unlock_buffer(bh[i]);
873 continue;
874 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500875 get_bh(bh[i]);
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500876 /*
877 * submit the buffer_head for read. We can
878 * safely mark the bitmap as uptodate now.
879 * We do it here so the bitmap uptodate bit
880 * get set with buffer lock held.
881 */
882 set_bitmap_uptodate(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500883 bh[i]->b_end_io = end_buffer_read_sync;
884 submit_bh(READ, bh[i]);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400885 mb_debug(1, "read bitmap for group %u\n", first_group + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500886 }
887
888 /* wait for I/O completion */
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400889 for (i = 0; i < groups_per_page; i++)
890 if (bh[i])
891 wait_on_buffer(bh[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500892
893 err = -EIO;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400894 for (i = 0; i < groups_per_page; i++)
895 if (bh[i] && !buffer_uptodate(bh[i]))
Alex Tomasc9de5602008-01-29 00:19:52 -0500896 goto out;
897
Mingming Cao31b481d2008-07-11 19:27:31 -0400898 err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -0500899 first_block = page->index * blocks_per_page;
900 for (i = 0; i < blocks_per_page; i++) {
901 int group;
Alex Tomasc9de5602008-01-29 00:19:52 -0500902
903 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400904 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500905 break;
906
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400907 if (!bh[group - first_group])
908 /* skip initialized uptodate buddy */
909 continue;
910
Alex Tomasc9de5602008-01-29 00:19:52 -0500911 /*
912 * data carry information regarding this
913 * particular group in the format specified
914 * above
915 *
916 */
917 data = page_address(page) + (i * blocksize);
918 bitmap = bh[group - first_group]->b_data;
919
920 /*
921 * We place the buddy block and bitmap block
922 * close together
923 */
924 if ((first_block + i) & 1) {
925 /* this is block of buddy */
926 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400927 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500928 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400929 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500930 grinfo = ext4_get_group_info(sb, group);
931 grinfo->bb_fragments = 0;
932 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400933 sizeof(*grinfo->bb_counters) *
934 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500935 /*
936 * incore got set to the group block bitmap below
937 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500938 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400939 /* init the buddy */
940 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500941 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500942 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500943 incore = NULL;
944 } else {
945 /* this is block of bitmap */
946 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400947 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500948 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400949 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500950
951 /* see comments in ext4_mb_put_pa() */
952 ext4_lock_group(sb, group);
953 memcpy(data, bitmap, blocksize);
954
955 /* mark all preallocated blks used in in-core bitmap */
956 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500957 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500958 ext4_unlock_group(sb, group);
959
960 /* set incore so that the buddy information can be
961 * generated using this
962 */
963 incore = data;
964 }
965 }
966 SetPageUptodate(page);
967
968out:
969 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400970 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500971 brelse(bh[i]);
972 if (bh != &bhs)
973 kfree(bh);
974 }
975 return err;
976}
977
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400978/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400979 * Lock the buddy and bitmap pages. This make sure other parallel init_group
980 * on the same buddy page doesn't happen whild holding the buddy page lock.
981 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
982 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400983 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400984static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
985 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400986{
Amir Goldstein2de88072011-05-09 21:48:13 -0400987 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
988 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400989 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400990 struct page *page;
991
992 e4b->bd_buddy_page = NULL;
993 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400994
995 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
996 /*
997 * the buddy cache inode stores the block bitmap
998 * and buddy information in consecutive blocks.
999 * So for each group we need two blocks.
1000 */
1001 block = group * 2;
1002 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001003 poff = block % blocks_per_page;
1004 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1005 if (!page)
1006 return -EIO;
1007 BUG_ON(page->mapping != inode->i_mapping);
1008 e4b->bd_bitmap_page = page;
1009 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001010
Amir Goldstein2de88072011-05-09 21:48:13 -04001011 if (blocks_per_page >= 2) {
1012 /* buddy and bitmap are on the same page */
1013 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001014 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001015
1016 block++;
1017 pnum = block / blocks_per_page;
1018 poff = block % blocks_per_page;
1019 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1020 if (!page)
1021 return -EIO;
1022 BUG_ON(page->mapping != inode->i_mapping);
1023 e4b->bd_buddy_page = page;
1024 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001025}
1026
Amir Goldstein2de88072011-05-09 21:48:13 -04001027static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001028{
Amir Goldstein2de88072011-05-09 21:48:13 -04001029 if (e4b->bd_bitmap_page) {
1030 unlock_page(e4b->bd_bitmap_page);
1031 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001032 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 page_cache_release(e4b->bd_buddy_page);
1036 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001037}
1038
1039/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1043 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001044static noinline_for_stack
1045int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1046{
1047
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001048 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001049 struct ext4_buddy e4b;
1050 struct page *page;
1051 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001052
1053 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001054 this_grp = ext4_get_group_info(sb, group);
1055 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001056 * This ensures that we don't reinit the buddy cache
1057 * page which map to the group from which we are already
1058 * allocating. If we are looking at the buddy cache we would
1059 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001060 * would have pinned buddy page to page cache.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001061 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001062 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1063 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001064 /*
1065 * somebody initialized the group
1066 * return without doing anything
1067 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001068 goto err;
1069 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001070
1071 page = e4b.bd_bitmap_page;
1072 ret = ext4_mb_init_cache(page, NULL);
1073 if (ret)
1074 goto err;
1075 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001076 ret = -EIO;
1077 goto err;
1078 }
1079 mark_page_accessed(page);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001080
Amir Goldstein2de88072011-05-09 21:48:13 -04001081 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001082 /*
1083 * If both the bitmap and buddy are in
1084 * the same page we don't need to force
1085 * init the buddy
1086 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001087 ret = 0;
1088 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001089 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001090 /* init buddy cache */
1091 page = e4b.bd_buddy_page;
1092 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1093 if (ret)
1094 goto err;
1095 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001096 ret = -EIO;
1097 goto err;
1098 }
1099 mark_page_accessed(page);
1100err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001101 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001102 return ret;
1103}
1104
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001105/*
1106 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1107 * block group lock of all groups for this page; do not hold the BG lock when
1108 * calling this routine!
1109 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001110static noinline_for_stack int
1111ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1112 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001113{
Alex Tomasc9de5602008-01-29 00:19:52 -05001114 int blocks_per_page;
1115 int block;
1116 int pnum;
1117 int poff;
1118 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001119 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001120 struct ext4_group_info *grp;
1121 struct ext4_sb_info *sbi = EXT4_SB(sb);
1122 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001123
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001124 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001125
1126 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001127 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001128
1129 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001130 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001131 e4b->bd_sb = sb;
1132 e4b->bd_group = group;
1133 e4b->bd_buddy_page = NULL;
1134 e4b->bd_bitmap_page = NULL;
1135
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001136 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001137 /*
1138 * we need full data about the group
1139 * to make a good selection
1140 */
1141 ret = ext4_mb_init_group(sb, group);
1142 if (ret)
1143 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001144 }
1145
Alex Tomasc9de5602008-01-29 00:19:52 -05001146 /*
1147 * the buddy cache inode stores the block bitmap
1148 * and buddy information in consecutive blocks.
1149 * So for each group we need two blocks.
1150 */
1151 block = group * 2;
1152 pnum = block / blocks_per_page;
1153 poff = block % blocks_per_page;
1154
1155 /* we could use find_or_create_page(), but it locks page
1156 * what we'd like to avoid in fast path ... */
1157 page = find_get_page(inode->i_mapping, pnum);
1158 if (page == NULL || !PageUptodate(page)) {
1159 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001160 /*
1161 * drop the page reference and try
1162 * to get the page with lock. If we
1163 * are not uptodate that implies
1164 * somebody just created the page but
1165 * is yet to initialize the same. So
1166 * wait for it to initialize.
1167 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001168 page_cache_release(page);
1169 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1170 if (page) {
1171 BUG_ON(page->mapping != inode->i_mapping);
1172 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001173 ret = ext4_mb_init_cache(page, NULL);
1174 if (ret) {
1175 unlock_page(page);
1176 goto err;
1177 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001178 mb_cmp_bitmaps(e4b, page_address(page) +
1179 (poff * sb->s_blocksize));
1180 }
1181 unlock_page(page);
1182 }
1183 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001184 if (page == NULL || !PageUptodate(page)) {
1185 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001186 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001187 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001188 e4b->bd_bitmap_page = page;
1189 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1190 mark_page_accessed(page);
1191
1192 block++;
1193 pnum = block / blocks_per_page;
1194 poff = block % blocks_per_page;
1195
1196 page = find_get_page(inode->i_mapping, pnum);
1197 if (page == NULL || !PageUptodate(page)) {
1198 if (page)
1199 page_cache_release(page);
1200 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1201 if (page) {
1202 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001203 if (!PageUptodate(page)) {
1204 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1205 if (ret) {
1206 unlock_page(page);
1207 goto err;
1208 }
1209 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001210 unlock_page(page);
1211 }
1212 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001213 if (page == NULL || !PageUptodate(page)) {
1214 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001215 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001216 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001217 e4b->bd_buddy_page = page;
1218 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1219 mark_page_accessed(page);
1220
1221 BUG_ON(e4b->bd_bitmap_page == NULL);
1222 BUG_ON(e4b->bd_buddy_page == NULL);
1223
1224 return 0;
1225
1226err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001227 if (page)
1228 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001229 if (e4b->bd_bitmap_page)
1230 page_cache_release(e4b->bd_bitmap_page);
1231 if (e4b->bd_buddy_page)
1232 page_cache_release(e4b->bd_buddy_page);
1233 e4b->bd_buddy = NULL;
1234 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001235 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001236}
1237
Jing Zhange39e07f2010-05-14 00:00:00 -04001238static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001239{
1240 if (e4b->bd_bitmap_page)
1241 page_cache_release(e4b->bd_bitmap_page);
1242 if (e4b->bd_buddy_page)
1243 page_cache_release(e4b->bd_buddy_page);
1244}
1245
1246
1247static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1248{
1249 int order = 1;
1250 void *bb;
1251
1252 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1253 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1254
1255 bb = EXT4_MB_BUDDY(e4b);
1256 while (order <= e4b->bd_blkbits + 1) {
1257 block = block >> 1;
1258 if (!mb_test_bit(block, bb)) {
1259 /* this block is part of buddy of order 'order' */
1260 return order;
1261 }
1262 bb += 1 << (e4b->bd_blkbits - order);
1263 order++;
1264 }
1265 return 0;
1266}
1267
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001268static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001269{
1270 __u32 *addr;
1271
1272 len = cur + len;
1273 while (cur < len) {
1274 if ((cur & 31) == 0 && (len - cur) >= 32) {
1275 /* fast path: clear whole word at once */
1276 addr = bm + (cur >> 3);
1277 *addr = 0;
1278 cur += 32;
1279 continue;
1280 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001281 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001282 cur++;
1283 }
1284}
1285
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001286void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001287{
1288 __u32 *addr;
1289
1290 len = cur + len;
1291 while (cur < len) {
1292 if ((cur & 31) == 0 && (len - cur) >= 32) {
1293 /* fast path: set whole word at once */
1294 addr = bm + (cur >> 3);
1295 *addr = 0xffffffff;
1296 cur += 32;
1297 continue;
1298 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001299 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001300 cur++;
1301 }
1302}
1303
Shen Feng7e5a8cd2008-07-13 21:03:31 -04001304static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
Alex Tomasc9de5602008-01-29 00:19:52 -05001305 int first, int count)
1306{
1307 int block = 0;
1308 int max = 0;
1309 int order;
1310 void *buddy;
1311 void *buddy2;
1312 struct super_block *sb = e4b->bd_sb;
1313
1314 BUG_ON(first + count > (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001315 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001316 mb_check_buddy(e4b);
1317 mb_free_blocks_double(inode, e4b, first, count);
1318
1319 e4b->bd_info->bb_free += count;
1320 if (first < e4b->bd_info->bb_first_free)
1321 e4b->bd_info->bb_first_free = first;
1322
1323 /* let's maintain fragments counter */
1324 if (first != 0)
1325 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1326 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1327 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1328 if (block && max)
1329 e4b->bd_info->bb_fragments--;
1330 else if (!block && !max)
1331 e4b->bd_info->bb_fragments++;
1332
1333 /* let's maintain buddy itself */
1334 while (count-- > 0) {
1335 block = first++;
1336 order = 0;
1337
1338 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1339 ext4_fsblk_t blocknr;
Akinobu Mita5661bd62010-03-03 23:53:39 -05001340
1341 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001342 blocknr += block;
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05001343 ext4_grp_locked_error(sb, e4b->bd_group,
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001344 inode ? inode->i_ino : 0,
1345 blocknr,
1346 "freeing already freed block "
1347 "(bit %u)", block);
Alex Tomasc9de5602008-01-29 00:19:52 -05001348 }
1349 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1350 e4b->bd_info->bb_counters[order]++;
1351
1352 /* start of the buddy */
1353 buddy = mb_find_buddy(e4b, order, &max);
1354
1355 do {
1356 block &= ~1UL;
1357 if (mb_test_bit(block, buddy) ||
1358 mb_test_bit(block + 1, buddy))
1359 break;
1360
1361 /* both the buddies are free, try to coalesce them */
1362 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1363
1364 if (!buddy2)
1365 break;
1366
1367 if (order > 0) {
1368 /* for special purposes, we don't set
1369 * free bits in bitmap */
1370 mb_set_bit(block, buddy);
1371 mb_set_bit(block + 1, buddy);
1372 }
1373 e4b->bd_info->bb_counters[order]--;
1374 e4b->bd_info->bb_counters[order]--;
1375
1376 block = block >> 1;
1377 order++;
1378 e4b->bd_info->bb_counters[order]++;
1379
1380 mb_clear_bit(block, buddy2);
1381 buddy = buddy2;
1382 } while (1);
1383 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001384 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001385 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001386}
1387
1388static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1389 int needed, struct ext4_free_extent *ex)
1390{
1391 int next = block;
1392 int max;
1393 int ord;
1394 void *buddy;
1395
Vincent Minetbc8e6742009-05-15 08:33:18 -04001396 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001397 BUG_ON(ex == NULL);
1398
1399 buddy = mb_find_buddy(e4b, order, &max);
1400 BUG_ON(buddy == NULL);
1401 BUG_ON(block >= max);
1402 if (mb_test_bit(block, buddy)) {
1403 ex->fe_len = 0;
1404 ex->fe_start = 0;
1405 ex->fe_group = 0;
1406 return 0;
1407 }
1408
1409 /* FIXME dorp order completely ? */
1410 if (likely(order == 0)) {
1411 /* find actual order */
1412 order = mb_find_order_for_block(e4b, block);
1413 block = block >> order;
1414 }
1415
1416 ex->fe_len = 1 << order;
1417 ex->fe_start = block << order;
1418 ex->fe_group = e4b->bd_group;
1419
1420 /* calc difference from given start */
1421 next = next - ex->fe_start;
1422 ex->fe_len -= next;
1423 ex->fe_start += next;
1424
1425 while (needed > ex->fe_len &&
1426 (buddy = mb_find_buddy(e4b, order, &max))) {
1427
1428 if (block + 1 >= max)
1429 break;
1430
1431 next = (block + 1) * (1 << order);
1432 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1433 break;
1434
1435 ord = mb_find_order_for_block(e4b, next);
1436
1437 order = ord;
1438 block = next >> order;
1439 ex->fe_len += 1 << order;
1440 }
1441
1442 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1443 return ex->fe_len;
1444}
1445
1446static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1447{
1448 int ord;
1449 int mlen = 0;
1450 int max = 0;
1451 int cur;
1452 int start = ex->fe_start;
1453 int len = ex->fe_len;
1454 unsigned ret = 0;
1455 int len0 = len;
1456 void *buddy;
1457
1458 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1459 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001460 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001461 mb_check_buddy(e4b);
1462 mb_mark_used_double(e4b, start, len);
1463
1464 e4b->bd_info->bb_free -= len;
1465 if (e4b->bd_info->bb_first_free == start)
1466 e4b->bd_info->bb_first_free += len;
1467
1468 /* let's maintain fragments counter */
1469 if (start != 0)
1470 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1471 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1472 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1473 if (mlen && max)
1474 e4b->bd_info->bb_fragments++;
1475 else if (!mlen && !max)
1476 e4b->bd_info->bb_fragments--;
1477
1478 /* let's maintain buddy itself */
1479 while (len) {
1480 ord = mb_find_order_for_block(e4b, start);
1481
1482 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1483 /* the whole chunk may be allocated at once! */
1484 mlen = 1 << ord;
1485 buddy = mb_find_buddy(e4b, ord, &max);
1486 BUG_ON((start >> ord) >= max);
1487 mb_set_bit(start >> ord, buddy);
1488 e4b->bd_info->bb_counters[ord]--;
1489 start += mlen;
1490 len -= mlen;
1491 BUG_ON(len < 0);
1492 continue;
1493 }
1494
1495 /* store for history */
1496 if (ret == 0)
1497 ret = len | (ord << 16);
1498
1499 /* we have to split large buddy */
1500 BUG_ON(ord <= 0);
1501 buddy = mb_find_buddy(e4b, ord, &max);
1502 mb_set_bit(start >> ord, buddy);
1503 e4b->bd_info->bb_counters[ord]--;
1504
1505 ord--;
1506 cur = (start >> ord) & ~1U;
1507 buddy = mb_find_buddy(e4b, ord, &max);
1508 mb_clear_bit(cur, buddy);
1509 mb_clear_bit(cur + 1, buddy);
1510 e4b->bd_info->bb_counters[ord]++;
1511 e4b->bd_info->bb_counters[ord]++;
1512 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001513 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001514
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001515 ext4_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001516 mb_check_buddy(e4b);
1517
1518 return ret;
1519}
1520
1521/*
1522 * Must be called under group lock!
1523 */
1524static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1525 struct ext4_buddy *e4b)
1526{
1527 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1528 int ret;
1529
1530 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1531 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1532
1533 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1534 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1535 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1536
1537 /* preallocation can change ac_b_ex, thus we store actually
1538 * allocated blocks for history */
1539 ac->ac_f_ex = ac->ac_b_ex;
1540
1541 ac->ac_status = AC_STATUS_FOUND;
1542 ac->ac_tail = ret & 0xffff;
1543 ac->ac_buddy = ret >> 16;
1544
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001545 /*
1546 * take the page reference. We want the page to be pinned
1547 * so that we don't get a ext4_mb_init_cache_call for this
1548 * group until we update the bitmap. That would mean we
1549 * double allocate blocks. The reference is dropped
1550 * in ext4_mb_release_context
1551 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001552 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1553 get_page(ac->ac_bitmap_page);
1554 ac->ac_buddy_page = e4b->bd_buddy_page;
1555 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001556 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001557 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001558 spin_lock(&sbi->s_md_lock);
1559 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1560 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1561 spin_unlock(&sbi->s_md_lock);
1562 }
1563}
1564
1565/*
1566 * regular allocator, for general purposes allocation
1567 */
1568
1569static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1570 struct ext4_buddy *e4b,
1571 int finish_group)
1572{
1573 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1574 struct ext4_free_extent *bex = &ac->ac_b_ex;
1575 struct ext4_free_extent *gex = &ac->ac_g_ex;
1576 struct ext4_free_extent ex;
1577 int max;
1578
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001579 if (ac->ac_status == AC_STATUS_FOUND)
1580 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001581 /*
1582 * We don't want to scan for a whole year
1583 */
1584 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1585 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1586 ac->ac_status = AC_STATUS_BREAK;
1587 return;
1588 }
1589
1590 /*
1591 * Haven't found good chunk so far, let's continue
1592 */
1593 if (bex->fe_len < gex->fe_len)
1594 return;
1595
1596 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1597 && bex->fe_group == e4b->bd_group) {
1598 /* recheck chunk's availability - we don't know
1599 * when it was found (within this lock-unlock
1600 * period or not) */
1601 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1602 if (max >= gex->fe_len) {
1603 ext4_mb_use_best_found(ac, e4b);
1604 return;
1605 }
1606 }
1607}
1608
1609/*
1610 * The routine checks whether found extent is good enough. If it is,
1611 * then the extent gets marked used and flag is set to the context
1612 * to stop scanning. Otherwise, the extent is compared with the
1613 * previous found extent and if new one is better, then it's stored
1614 * in the context. Later, the best found extent will be used, if
1615 * mballoc can't find good enough extent.
1616 *
1617 * FIXME: real allocation policy is to be designed yet!
1618 */
1619static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1620 struct ext4_free_extent *ex,
1621 struct ext4_buddy *e4b)
1622{
1623 struct ext4_free_extent *bex = &ac->ac_b_ex;
1624 struct ext4_free_extent *gex = &ac->ac_g_ex;
1625
1626 BUG_ON(ex->fe_len <= 0);
Eric Sandeen8d03c7a2009-03-14 11:51:46 -04001627 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001628 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1629 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1630
1631 ac->ac_found++;
1632
1633 /*
1634 * The special case - take what you catch first
1635 */
1636 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1637 *bex = *ex;
1638 ext4_mb_use_best_found(ac, e4b);
1639 return;
1640 }
1641
1642 /*
1643 * Let's check whether the chuck is good enough
1644 */
1645 if (ex->fe_len == gex->fe_len) {
1646 *bex = *ex;
1647 ext4_mb_use_best_found(ac, e4b);
1648 return;
1649 }
1650
1651 /*
1652 * If this is first found extent, just store it in the context
1653 */
1654 if (bex->fe_len == 0) {
1655 *bex = *ex;
1656 return;
1657 }
1658
1659 /*
1660 * If new found extent is better, store it in the context
1661 */
1662 if (bex->fe_len < gex->fe_len) {
1663 /* if the request isn't satisfied, any found extent
1664 * larger than previous best one is better */
1665 if (ex->fe_len > bex->fe_len)
1666 *bex = *ex;
1667 } else if (ex->fe_len > gex->fe_len) {
1668 /* if the request is satisfied, then we try to find
1669 * an extent that still satisfy the request, but is
1670 * smaller than previous one */
1671 if (ex->fe_len < bex->fe_len)
1672 *bex = *ex;
1673 }
1674
1675 ext4_mb_check_limits(ac, e4b, 0);
1676}
1677
Eric Sandeen089ceec2009-07-05 22:17:31 -04001678static noinline_for_stack
1679int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001680 struct ext4_buddy *e4b)
1681{
1682 struct ext4_free_extent ex = ac->ac_b_ex;
1683 ext4_group_t group = ex.fe_group;
1684 int max;
1685 int err;
1686
1687 BUG_ON(ex.fe_len <= 0);
1688 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1689 if (err)
1690 return err;
1691
1692 ext4_lock_group(ac->ac_sb, group);
1693 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1694
1695 if (max > 0) {
1696 ac->ac_b_ex = ex;
1697 ext4_mb_use_best_found(ac, e4b);
1698 }
1699
1700 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001701 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001702
1703 return 0;
1704}
1705
Eric Sandeen089ceec2009-07-05 22:17:31 -04001706static noinline_for_stack
1707int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001708 struct ext4_buddy *e4b)
1709{
1710 ext4_group_t group = ac->ac_g_ex.fe_group;
1711 int max;
1712 int err;
1713 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05001714 struct ext4_free_extent ex;
1715
1716 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1717 return 0;
1718
1719 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1720 if (err)
1721 return err;
1722
1723 ext4_lock_group(ac->ac_sb, group);
1724 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1725 ac->ac_g_ex.fe_len, &ex);
1726
1727 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1728 ext4_fsblk_t start;
1729
Akinobu Mita5661bd62010-03-03 23:53:39 -05001730 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1731 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001732 /* use do_div to get remainder (would be 64-bit modulo) */
1733 if (do_div(start, sbi->s_stripe) == 0) {
1734 ac->ac_found++;
1735 ac->ac_b_ex = ex;
1736 ext4_mb_use_best_found(ac, e4b);
1737 }
1738 } else if (max >= ac->ac_g_ex.fe_len) {
1739 BUG_ON(ex.fe_len <= 0);
1740 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1741 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1742 ac->ac_found++;
1743 ac->ac_b_ex = ex;
1744 ext4_mb_use_best_found(ac, e4b);
1745 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1746 /* Sometimes, caller may want to merge even small
1747 * number of blocks to an existing extent */
1748 BUG_ON(ex.fe_len <= 0);
1749 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1750 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1751 ac->ac_found++;
1752 ac->ac_b_ex = ex;
1753 ext4_mb_use_best_found(ac, e4b);
1754 }
1755 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001756 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001757
1758 return 0;
1759}
1760
1761/*
1762 * The routine scans buddy structures (not bitmap!) from given order
1763 * to max order and tries to find big enough chunk to satisfy the req
1764 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001765static noinline_for_stack
1766void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001767 struct ext4_buddy *e4b)
1768{
1769 struct super_block *sb = ac->ac_sb;
1770 struct ext4_group_info *grp = e4b->bd_info;
1771 void *buddy;
1772 int i;
1773 int k;
1774 int max;
1775
1776 BUG_ON(ac->ac_2order <= 0);
1777 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1778 if (grp->bb_counters[i] == 0)
1779 continue;
1780
1781 buddy = mb_find_buddy(e4b, i, &max);
1782 BUG_ON(buddy == NULL);
1783
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001784 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001785 BUG_ON(k >= max);
1786
1787 ac->ac_found++;
1788
1789 ac->ac_b_ex.fe_len = 1 << i;
1790 ac->ac_b_ex.fe_start = k << i;
1791 ac->ac_b_ex.fe_group = e4b->bd_group;
1792
1793 ext4_mb_use_best_found(ac, e4b);
1794
1795 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1796
1797 if (EXT4_SB(sb)->s_mb_stats)
1798 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1799
1800 break;
1801 }
1802}
1803
1804/*
1805 * The routine scans the group and measures all found extents.
1806 * In order to optimize scanning, caller must pass number of
1807 * free blocks in the group, so the routine can know upper limit.
1808 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001809static noinline_for_stack
1810void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001811 struct ext4_buddy *e4b)
1812{
1813 struct super_block *sb = ac->ac_sb;
1814 void *bitmap = EXT4_MB_BITMAP(e4b);
1815 struct ext4_free_extent ex;
1816 int i;
1817 int free;
1818
1819 free = e4b->bd_info->bb_free;
1820 BUG_ON(free <= 0);
1821
1822 i = e4b->bd_info->bb_first_free;
1823
1824 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001825 i = mb_find_next_zero_bit(bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05001826 EXT4_BLOCKS_PER_GROUP(sb), i);
1827 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001828 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001829 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001830 * free blocks even though group info says we
1831 * we have free blocks
1832 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001833 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1834 "%d free blocks as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001835 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001836 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001837 break;
1838 }
1839
1840 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1841 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001842 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001843 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1844 "%d free blocks as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001845 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001846 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001847 /*
1848 * The number of free blocks differs. This mostly
1849 * indicate that the bitmap is corrupt. So exit
1850 * without claiming the space.
1851 */
1852 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001853 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001854
1855 ext4_mb_measure_extent(ac, &ex, e4b);
1856
1857 i += ex.fe_len;
1858 free -= ex.fe_len;
1859 }
1860
1861 ext4_mb_check_limits(ac, e4b, 1);
1862}
1863
1864/*
1865 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001866 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001867 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001868static noinline_for_stack
1869void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001870 struct ext4_buddy *e4b)
1871{
1872 struct super_block *sb = ac->ac_sb;
1873 struct ext4_sb_info *sbi = EXT4_SB(sb);
1874 void *bitmap = EXT4_MB_BITMAP(e4b);
1875 struct ext4_free_extent ex;
1876 ext4_fsblk_t first_group_block;
1877 ext4_fsblk_t a;
1878 ext4_grpblk_t i;
1879 int max;
1880
1881 BUG_ON(sbi->s_stripe == 0);
1882
1883 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001884 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1885
Alex Tomasc9de5602008-01-29 00:19:52 -05001886 a = first_group_block + sbi->s_stripe - 1;
1887 do_div(a, sbi->s_stripe);
1888 i = (a * sbi->s_stripe) - first_group_block;
1889
1890 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1891 if (!mb_test_bit(i, bitmap)) {
1892 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1893 if (max >= sbi->s_stripe) {
1894 ac->ac_found++;
1895 ac->ac_b_ex = ex;
1896 ext4_mb_use_best_found(ac, e4b);
1897 break;
1898 }
1899 }
1900 i += sbi->s_stripe;
1901 }
1902}
1903
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001904/* This is now called BEFORE we load the buddy bitmap. */
Alex Tomasc9de5602008-01-29 00:19:52 -05001905static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1906 ext4_group_t group, int cr)
1907{
1908 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04001909 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001910 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1911
1912 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001913
1914 /* We only do this if the grp has never been initialized */
1915 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1916 int ret = ext4_mb_init_group(ac->ac_sb, group);
1917 if (ret)
1918 return 0;
1919 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001920
1921 free = grp->bb_free;
1922 fragments = grp->bb_fragments;
1923 if (free == 0)
1924 return 0;
1925 if (fragments == 0)
1926 return 0;
1927
1928 switch (cr) {
1929 case 0:
1930 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001931
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001932 if (grp->bb_largest_free_order < ac->ac_2order)
1933 return 0;
1934
Theodore Ts'oa4912122009-03-12 12:18:34 -04001935 /* Avoid using the first bg of a flexgroup for data files */
1936 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1937 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1938 ((group % flex_size) == 0))
1939 return 0;
1940
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001941 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001942 case 1:
1943 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1944 return 1;
1945 break;
1946 case 2:
1947 if (free >= ac->ac_g_ex.fe_len)
1948 return 1;
1949 break;
1950 case 3:
1951 return 1;
1952 default:
1953 BUG();
1954 }
1955
1956 return 0;
1957}
1958
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001959static noinline_for_stack int
1960ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05001961{
Theodore Ts'o8df96752009-05-01 08:50:38 -04001962 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05001963 int cr;
1964 int err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001965 struct ext4_sb_info *sbi;
1966 struct super_block *sb;
1967 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05001968
1969 sb = ac->ac_sb;
1970 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04001971 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04001972 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04001973 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04001974 ngroups = sbi->s_blockfile_groups;
1975
Alex Tomasc9de5602008-01-29 00:19:52 -05001976 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1977
1978 /* first, try the goal */
1979 err = ext4_mb_find_by_goal(ac, &e4b);
1980 if (err || ac->ac_status == AC_STATUS_FOUND)
1981 goto out;
1982
1983 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1984 goto out;
1985
1986 /*
1987 * ac->ac2_order is set only if the fe_len is a power of 2
1988 * if ac2_order is set we also set criteria to 0 so that we
1989 * try exact allocation using buddy.
1990 */
1991 i = fls(ac->ac_g_ex.fe_len);
1992 ac->ac_2order = 0;
1993 /*
1994 * We search using buddy data only if the order of the request
1995 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04001996 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05001997 */
1998 if (i >= sbi->s_mb_order2_reqs) {
1999 /*
2000 * This should tell if fe_len is exactly power of 2
2001 */
2002 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2003 ac->ac_2order = i - 1;
2004 }
2005
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002006 /* if stream allocation is enabled, use global goal */
2007 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002008 /* TBD: may be hot point */
2009 spin_lock(&sbi->s_md_lock);
2010 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2011 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2012 spin_unlock(&sbi->s_md_lock);
2013 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002014
Alex Tomasc9de5602008-01-29 00:19:52 -05002015 /* Let's just scan groups to find more-less suitable blocks */
2016 cr = ac->ac_2order ? 0 : 1;
2017 /*
2018 * cr == 0 try to get exact allocation,
2019 * cr == 3 try to get anything
2020 */
2021repeat:
2022 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2023 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002024 /*
2025 * searching for the right group start
2026 * from the goal value specified
2027 */
2028 group = ac->ac_g_ex.fe_group;
2029
Theodore Ts'o8df96752009-05-01 08:50:38 -04002030 for (i = 0; i < ngroups; group++, i++) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002031 if (group == ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002032 group = 0;
2033
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002034 /* This now checks without needing the buddy page */
2035 if (!ext4_mb_good_group(ac, group, cr))
Alex Tomasc9de5602008-01-29 00:19:52 -05002036 continue;
2037
Alex Tomasc9de5602008-01-29 00:19:52 -05002038 err = ext4_mb_load_buddy(sb, group, &e4b);
2039 if (err)
2040 goto out;
2041
2042 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002043
2044 /*
2045 * We need to check again after locking the
2046 * block group
2047 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002048 if (!ext4_mb_good_group(ac, group, cr)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002049 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002050 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002051 continue;
2052 }
2053
2054 ac->ac_groups_scanned++;
Theodore Ts'o75507ef2009-05-01 12:58:36 -04002055 if (cr == 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002056 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002057 else if (cr == 1 && sbi->s_stripe &&
2058 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002059 ext4_mb_scan_aligned(ac, &e4b);
2060 else
2061 ext4_mb_complex_scan_group(ac, &e4b);
2062
2063 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002064 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002065
2066 if (ac->ac_status != AC_STATUS_CONTINUE)
2067 break;
2068 }
2069 }
2070
2071 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2072 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2073 /*
2074 * We've been searching too long. Let's try to allocate
2075 * the best chunk we've found so far
2076 */
2077
2078 ext4_mb_try_best_found(ac, &e4b);
2079 if (ac->ac_status != AC_STATUS_FOUND) {
2080 /*
2081 * Someone more lucky has already allocated it.
2082 * The only thing we can do is just take first
2083 * found block(s)
2084 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2085 */
2086 ac->ac_b_ex.fe_group = 0;
2087 ac->ac_b_ex.fe_start = 0;
2088 ac->ac_b_ex.fe_len = 0;
2089 ac->ac_status = AC_STATUS_CONTINUE;
2090 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2091 cr = 3;
2092 atomic_inc(&sbi->s_mb_lost_chunks);
2093 goto repeat;
2094 }
2095 }
2096out:
2097 return err;
2098}
2099
Alex Tomasc9de5602008-01-29 00:19:52 -05002100static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2101{
2102 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002103 ext4_group_t group;
2104
Theodore Ts'o8df96752009-05-01 08:50:38 -04002105 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002106 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002107 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002108 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002109}
2110
2111static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2112{
2113 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002114 ext4_group_t group;
2115
2116 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002117 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002118 return NULL;
2119 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002120 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002121}
2122
2123static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2124{
2125 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002126 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002127 int i;
2128 int err;
2129 struct ext4_buddy e4b;
2130 struct sg {
2131 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002132 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002133 } sg;
2134
2135 group--;
2136 if (group == 0)
2137 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2138 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2139 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2140 "group", "free", "frags", "first",
2141 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2142 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2143
2144 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2145 sizeof(struct ext4_group_info);
2146 err = ext4_mb_load_buddy(sb, group, &e4b);
2147 if (err) {
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002148 seq_printf(seq, "#%-5u: I/O error\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002149 return 0;
2150 }
2151 ext4_lock_group(sb, group);
2152 memcpy(&sg, ext4_get_group_info(sb, group), i);
2153 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002154 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002155
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002156 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002157 sg.info.bb_fragments, sg.info.bb_first_free);
2158 for (i = 0; i <= 13; i++)
2159 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2160 sg.info.bb_counters[i] : 0);
2161 seq_printf(seq, " ]\n");
2162
2163 return 0;
2164}
2165
2166static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2167{
2168}
2169
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002170static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002171 .start = ext4_mb_seq_groups_start,
2172 .next = ext4_mb_seq_groups_next,
2173 .stop = ext4_mb_seq_groups_stop,
2174 .show = ext4_mb_seq_groups_show,
2175};
2176
2177static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2178{
2179 struct super_block *sb = PDE(inode)->data;
2180 int rc;
2181
2182 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2183 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002184 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002185 m->private = sb;
2186 }
2187 return rc;
2188
2189}
2190
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002191static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002192 .owner = THIS_MODULE,
2193 .open = ext4_mb_seq_groups_open,
2194 .read = seq_read,
2195 .llseek = seq_lseek,
2196 .release = seq_release,
2197};
2198
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002199static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2200{
2201 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2202 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2203
2204 BUG_ON(!cachep);
2205 return cachep;
2206}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002207
2208/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002209int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002210 struct ext4_group_desc *desc)
2211{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002212 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002213 int metalen = 0;
2214 struct ext4_sb_info *sbi = EXT4_SB(sb);
2215 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002216 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002217
2218 /*
2219 * First check if this group is the first of a reserved block.
2220 * If it's true, we have to allocate a new table of pointers
2221 * to ext4_group_info structures
2222 */
2223 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2224 metalen = sizeof(*meta_group_info) <<
2225 EXT4_DESC_PER_BLOCK_BITS(sb);
2226 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2227 if (meta_group_info == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002228 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate mem "
2229 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002230 goto exit_meta_group_info;
2231 }
2232 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2233 meta_group_info;
2234 }
2235
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002236 meta_group_info =
2237 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2238 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2239
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002240 meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002241 if (meta_group_info[i] == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002242 ext4_msg(sb, KERN_ERR, "EXT4-fs: can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002243 goto exit_group_info;
2244 }
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002245 memset(meta_group_info[i], 0, kmem_cache_size(cachep));
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002246 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2247 &(meta_group_info[i]->bb_state));
2248
2249 /*
2250 * initialize bb_free to be able to skip
2251 * empty groups without initialization
2252 */
2253 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2254 meta_group_info[i]->bb_free =
2255 ext4_free_blocks_after_init(sb, group, desc);
2256 } else {
2257 meta_group_info[i]->bb_free =
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002258 ext4_free_blks_count(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002259 }
2260
2261 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002262 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002263 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002264 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002265
2266#ifdef DOUBLE_CHECK
2267 {
2268 struct buffer_head *bh;
2269 meta_group_info[i]->bb_bitmap =
2270 kmalloc(sb->s_blocksize, GFP_KERNEL);
2271 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2272 bh = ext4_read_block_bitmap(sb, group);
2273 BUG_ON(bh == NULL);
2274 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2275 sb->s_blocksize);
2276 put_bh(bh);
2277 }
2278#endif
2279
2280 return 0;
2281
2282exit_group_info:
2283 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002284 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002285 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002286 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2287 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002288exit_meta_group_info:
2289 return -ENOMEM;
2290} /* ext4_mb_add_groupinfo */
2291
Alex Tomasc9de5602008-01-29 00:19:52 -05002292static int ext4_mb_init_backend(struct super_block *sb)
2293{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002294 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002295 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002296 struct ext4_sb_info *sbi = EXT4_SB(sb);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002297 struct ext4_super_block *es = sbi->s_es;
2298 int num_meta_group_infos;
2299 int num_meta_group_infos_max;
2300 int array_size;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002301 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002302 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002303
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002304 /* This is the number of blocks used by GDT */
Theodore Ts'o8df96752009-05-01 08:50:38 -04002305 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002306 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2307
2308 /*
2309 * This is the total number of blocks used by GDT including
2310 * the number of reserved blocks for GDT.
2311 * The s_group_info array is allocated with this value
2312 * to allow a clean online resize without a complex
2313 * manipulation of pointer.
2314 * The drawback is the unused memory when no resize
2315 * occurs but it's very low in terms of pages
2316 * (see comments below)
2317 * Need to handle this properly when META_BG resizing is allowed
2318 */
2319 num_meta_group_infos_max = num_meta_group_infos +
2320 le16_to_cpu(es->s_reserved_gdt_blocks);
2321
2322 /*
2323 * array_size is the size of s_group_info array. We round it
2324 * to the next power of two because this approximation is done
2325 * internally by kmalloc so we can have some more memory
2326 * for free here (e.g. may be used for META_BG resize).
2327 */
2328 array_size = 1;
2329 while (array_size < sizeof(*sbi->s_group_info) *
2330 num_meta_group_infos_max)
2331 array_size = array_size << 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002332 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2333 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2334 * So a two level scheme suffices for now. */
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002335 sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
Alex Tomasc9de5602008-01-29 00:19:52 -05002336 if (sbi->s_group_info == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002337 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
Alex Tomasc9de5602008-01-29 00:19:52 -05002338 return -ENOMEM;
2339 }
2340 sbi->s_buddy_cache = new_inode(sb);
2341 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002342 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002343 goto err_freesgi;
2344 }
Yu Jian48e60612011-08-01 17:41:39 -04002345 /* To avoid potentially colliding with an valid on-disk inode number,
2346 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2347 * not in the inode hash, so it should never be found by iget(), but
2348 * this will avoid confusion if it ever shows up during debugging. */
2349 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002350 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002351 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002352 desc = ext4_get_group_desc(sb, i, NULL);
2353 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002354 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002355 goto err_freebuddy;
2356 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002357 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2358 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002359 }
2360
2361 return 0;
2362
2363err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002364 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002365 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002366 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Alex Tomasc9de5602008-01-29 00:19:52 -05002367 i = num_meta_group_infos;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002368 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002369 kfree(sbi->s_group_info[i]);
2370 iput(sbi->s_buddy_cache);
2371err_freesgi:
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002372 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002373 return -ENOMEM;
2374}
2375
Eric Sandeen2892c152011-02-12 08:12:18 -05002376static void ext4_groupinfo_destroy_slabs(void)
2377{
2378 int i;
2379
2380 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2381 if (ext4_groupinfo_caches[i])
2382 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2383 ext4_groupinfo_caches[i] = NULL;
2384 }
2385}
2386
2387static int ext4_groupinfo_create_slab(size_t size)
2388{
2389 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2390 int slab_size;
2391 int blocksize_bits = order_base_2(size);
2392 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2393 struct kmem_cache *cachep;
2394
2395 if (cache_index >= NR_GRPINFO_CACHES)
2396 return -EINVAL;
2397
2398 if (unlikely(cache_index < 0))
2399 cache_index = 0;
2400
2401 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2402 if (ext4_groupinfo_caches[cache_index]) {
2403 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2404 return 0; /* Already created */
2405 }
2406
2407 slab_size = offsetof(struct ext4_group_info,
2408 bb_counters[blocksize_bits + 2]);
2409
2410 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2411 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2412 NULL);
2413
Tao Ma823ba012011-07-11 18:26:01 -04002414 ext4_groupinfo_caches[cache_index] = cachep;
2415
Eric Sandeen2892c152011-02-12 08:12:18 -05002416 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2417 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002418 printk(KERN_EMERG
2419 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002420 return -ENOMEM;
2421 }
2422
Eric Sandeen2892c152011-02-12 08:12:18 -05002423 return 0;
2424}
2425
Alex Tomasc9de5602008-01-29 00:19:52 -05002426int ext4_mb_init(struct super_block *sb, int needs_recovery)
2427{
2428 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002429 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002430 unsigned offset;
2431 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002432 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002433
Eric Sandeen19278052009-08-25 22:36:25 -04002434 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002435
2436 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2437 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002438 ret = -ENOMEM;
2439 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002440 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002441
Eric Sandeen19278052009-08-25 22:36:25 -04002442 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002443 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2444 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002445 ret = -ENOMEM;
2446 goto out;
2447 }
2448
Eric Sandeen2892c152011-02-12 08:12:18 -05002449 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2450 if (ret < 0)
2451 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002452
2453 /* order 0 is regular bitmap */
2454 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2455 sbi->s_mb_offsets[0] = 0;
2456
2457 i = 1;
2458 offset = 0;
2459 max = sb->s_blocksize << 2;
2460 do {
2461 sbi->s_mb_offsets[i] = offset;
2462 sbi->s_mb_maxs[i] = max;
2463 offset += 1 << (sb->s_blocksize_bits - i);
2464 max = max >> 1;
2465 i++;
2466 } while (i <= sb->s_blocksize_bits + 1);
2467
Alex Tomasc9de5602008-01-29 00:19:52 -05002468 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002469 spin_lock_init(&sbi->s_bal_lock);
2470
2471 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2472 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2473 sbi->s_mb_stats = MB_DEFAULT_STATS;
2474 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2475 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Alex Tomasc9de5602008-01-29 00:19:52 -05002476 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002477 /*
2478 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2479 * to the lowest multiple of s_stripe which is bigger than
2480 * the s_mb_group_prealloc as determined above. We want
2481 * the preallocation size to be an exact multiple of the
2482 * RAID stripe size so that preallocations don't fragment
2483 * the stripes.
2484 */
2485 if (sbi->s_stripe > 1) {
2486 sbi->s_mb_group_prealloc = roundup(
2487 sbi->s_mb_group_prealloc, sbi->s_stripe);
2488 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002489
Eric Sandeen730c2132008-09-13 15:23:29 -04002490 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002491 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002492 ret = -ENOMEM;
2493 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002494 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002495 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002496 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002497 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002498 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002499 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2500 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002501 spin_lock_init(&lg->lg_prealloc_lock);
2502 }
2503
Yu Jian79a77c52011-08-01 17:41:46 -04002504 /* init file for buddy data */
2505 ret = ext4_mb_init_backend(sb);
2506 if (ret != 0) {
2507 goto out;
2508 }
2509
Theodore Ts'o296c3552009-09-30 00:32:42 -04002510 if (sbi->s_proc)
2511 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2512 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002513
Frank Mayhar03901312009-01-07 00:06:22 -05002514 if (sbi->s_journal)
2515 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002516out:
2517 if (ret) {
2518 kfree(sbi->s_mb_offsets);
2519 kfree(sbi->s_mb_maxs);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002520 }
2521 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002522}
2523
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002524/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002525static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2526{
2527 struct ext4_prealloc_space *pa;
2528 struct list_head *cur, *tmp;
2529 int count = 0;
2530
2531 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2532 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2533 list_del(&pa->pa_group_list);
2534 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002535 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002536 }
2537 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002538 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002539
2540}
2541
2542int ext4_mb_release(struct super_block *sb)
2543{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002544 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002545 ext4_group_t i;
2546 int num_meta_group_infos;
2547 struct ext4_group_info *grinfo;
2548 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002549 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002550
Alex Tomasc9de5602008-01-29 00:19:52 -05002551 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002552 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002553 grinfo = ext4_get_group_info(sb, i);
2554#ifdef DOUBLE_CHECK
2555 kfree(grinfo->bb_bitmap);
2556#endif
2557 ext4_lock_group(sb, i);
2558 ext4_mb_cleanup_pa(grinfo);
2559 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002560 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002561 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002562 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002563 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2564 EXT4_DESC_PER_BLOCK_BITS(sb);
2565 for (i = 0; i < num_meta_group_infos; i++)
2566 kfree(sbi->s_group_info[i]);
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002567 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002568 }
2569 kfree(sbi->s_mb_offsets);
2570 kfree(sbi->s_mb_maxs);
2571 if (sbi->s_buddy_cache)
2572 iput(sbi->s_buddy_cache);
2573 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002574 ext4_msg(sb, KERN_INFO,
2575 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002576 atomic_read(&sbi->s_bal_allocated),
2577 atomic_read(&sbi->s_bal_reqs),
2578 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002579 ext4_msg(sb, KERN_INFO,
2580 "mballoc: %u extents scanned, %u goal hits, "
2581 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002582 atomic_read(&sbi->s_bal_ex_scanned),
2583 atomic_read(&sbi->s_bal_goals),
2584 atomic_read(&sbi->s_bal_2orders),
2585 atomic_read(&sbi->s_bal_breaks),
2586 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002587 ext4_msg(sb, KERN_INFO,
2588 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002589 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002590 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002591 ext4_msg(sb, KERN_INFO,
2592 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002593 atomic_read(&sbi->s_mb_preallocated),
2594 atomic_read(&sbi->s_mb_discarded));
2595 }
2596
Eric Sandeen730c2132008-09-13 15:23:29 -04002597 free_percpu(sbi->s_locality_groups);
Theodore Ts'o296c3552009-09-30 00:32:42 -04002598 if (sbi->s_proc)
2599 remove_proc_entry("mb_groups", sbi->s_proc);
Alex Tomasc9de5602008-01-29 00:19:52 -05002600
2601 return 0;
2602}
2603
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002604static inline int ext4_issue_discard(struct super_block *sb,
Jiaying Zhang5c521832010-07-27 11:56:05 -04002605 ext4_group_t block_group, ext4_grpblk_t block, int count)
2606{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002607 ext4_fsblk_t discard_block;
2608
2609 discard_block = block + ext4_group_first_block_no(sb, block_group);
2610 trace_ext4_discard_blocks(sb,
2611 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002612 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002613}
2614
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002615/*
2616 * This function is called by the jbd2 layer once the commit has finished,
2617 * so we know we can free the blocks that were released with that commit.
2618 */
2619static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
Alex Tomasc9de5602008-01-29 00:19:52 -05002620{
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002621 struct super_block *sb = journal->j_private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002622 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002623 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002624 int err, count = 0, count2 = 0;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002625 struct ext4_free_data *entry;
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002626 struct list_head *l, *ltmp;
Alex Tomasc9de5602008-01-29 00:19:52 -05002627
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002628 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2629 entry = list_entry(l, struct ext4_free_data, list);
Alex Tomasc9de5602008-01-29 00:19:52 -05002630
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002631 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002632 entry->count, entry->group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002633
Theodore Ts'od9f34502011-04-30 13:47:24 -04002634 if (test_opt(sb, DISCARD))
2635 ext4_issue_discard(sb, entry->group,
2636 entry->start_blk, entry->count);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002637
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002638 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002639 /* we expect to find existing buddy because it's pinned */
2640 BUG_ON(err != 0);
2641
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002642 db = e4b.bd_info;
Alex Tomasc9de5602008-01-29 00:19:52 -05002643 /* there are blocks to put in buddy to make them really free */
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002644 count += entry->count;
Alex Tomasc9de5602008-01-29 00:19:52 -05002645 count2++;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002646 ext4_lock_group(sb, entry->group);
2647 /* Take it out of per group rb tree */
2648 rb_erase(&entry->node, &(db->bb_free_root));
2649 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2650
Tao Ma3d56b8d2011-07-11 00:03:38 -04002651 /*
2652 * Clear the trimmed flag for the group so that the next
2653 * ext4_trim_fs can trim it.
2654 * If the volume is mounted with -o discard, online discard
2655 * is supported and the free blocks will be trimmed online.
2656 */
2657 if (!test_opt(sb, DISCARD))
2658 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2659
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002660 if (!db->bb_free_root.rb_node) {
2661 /* No more items in the per group rb tree
2662 * balance refcounts from ext4_mb_free_metadata()
2663 */
2664 page_cache_release(e4b.bd_buddy_page);
2665 page_cache_release(e4b.bd_bitmap_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05002666 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002667 ext4_unlock_group(sb, entry->group);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002668 kmem_cache_free(ext4_free_ext_cachep, entry);
Jing Zhange39e07f2010-05-14 00:00:00 -04002669 ext4_mb_unload_buddy(&e4b);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002670 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002671
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002672 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002673}
2674
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002675#ifdef CONFIG_EXT4_DEBUG
2676u8 mb_enable_debug __read_mostly;
2677
2678static struct dentry *debugfs_dir;
2679static struct dentry *debugfs_debug;
2680
2681static void __init ext4_create_debugfs_entry(void)
2682{
2683 debugfs_dir = debugfs_create_dir("ext4", NULL);
2684 if (debugfs_dir)
2685 debugfs_debug = debugfs_create_u8("mballoc-debug",
2686 S_IRUGO | S_IWUSR,
2687 debugfs_dir,
2688 &mb_enable_debug);
2689}
2690
2691static void ext4_remove_debugfs_entry(void)
2692{
2693 debugfs_remove(debugfs_debug);
2694 debugfs_remove(debugfs_dir);
2695}
2696
2697#else
2698
2699static void __init ext4_create_debugfs_entry(void)
2700{
2701}
2702
2703static void ext4_remove_debugfs_entry(void)
2704{
2705}
2706
2707#endif
2708
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002709int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002710{
Theodore Ts'o16828082010-10-27 21:30:09 -04002711 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2712 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002713 if (ext4_pspace_cachep == NULL)
2714 return -ENOMEM;
2715
Theodore Ts'o16828082010-10-27 21:30:09 -04002716 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2717 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002718 if (ext4_ac_cachep == NULL) {
2719 kmem_cache_destroy(ext4_pspace_cachep);
2720 return -ENOMEM;
2721 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002722
Theodore Ts'o16828082010-10-27 21:30:09 -04002723 ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data,
2724 SLAB_RECLAIM_ACCOUNT);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002725 if (ext4_free_ext_cachep == NULL) {
2726 kmem_cache_destroy(ext4_pspace_cachep);
2727 kmem_cache_destroy(ext4_ac_cachep);
2728 return -ENOMEM;
2729 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002730 ext4_create_debugfs_entry();
Alex Tomasc9de5602008-01-29 00:19:52 -05002731 return 0;
2732}
2733
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002734void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002735{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002736 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002737 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2738 * before destroying the slab cache.
2739 */
2740 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002741 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002742 kmem_cache_destroy(ext4_ac_cachep);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002743 kmem_cache_destroy(ext4_free_ext_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002744 ext4_groupinfo_destroy_slabs();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002745 ext4_remove_debugfs_entry();
Alex Tomasc9de5602008-01-29 00:19:52 -05002746}
2747
2748
2749/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002750 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002751 * Returns 0 if success or error code
2752 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002753static noinline_for_stack int
2754ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o498e5f22008-11-05 00:14:04 -05002755 handle_t *handle, unsigned int reserv_blks)
Alex Tomasc9de5602008-01-29 00:19:52 -05002756{
2757 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002758 struct ext4_group_desc *gdp;
2759 struct buffer_head *gdp_bh;
2760 struct ext4_sb_info *sbi;
2761 struct super_block *sb;
2762 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002763 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002764
2765 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2766 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2767
2768 sb = ac->ac_sb;
2769 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002770
2771 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002772 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002773 if (!bitmap_bh)
2774 goto out_err;
2775
2776 err = ext4_journal_get_write_access(handle, bitmap_bh);
2777 if (err)
2778 goto out_err;
2779
2780 err = -EIO;
2781 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2782 if (!gdp)
2783 goto out_err;
2784
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002785 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Thadeu Lima de Souza Cascardo9fd97842009-01-26 19:26:26 -05002786 ext4_free_blks_count(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002787
Alex Tomasc9de5602008-01-29 00:19:52 -05002788 err = ext4_journal_get_write_access(handle, gdp_bh);
2789 if (err)
2790 goto out_err;
2791
Akinobu Mitabda00de2010-03-03 23:53:25 -05002792 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002793
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002794 len = ac->ac_b_ex.fe_len;
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002795 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002796 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002797 "fs metadata\n", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002798 /* File system mounted not to panic on error
2799 * Fix the bitmap and repeat the block allocation
2800 * We leak some of the blocks here.
2801 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002802 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002803 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2804 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002805 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002806 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002807 if (!err)
2808 err = -EAGAIN;
2809 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002810 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002811
2812 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002813#ifdef AGGRESSIVE_CHECK
2814 {
2815 int i;
2816 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2817 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2818 bitmap_bh->b_data));
2819 }
2820 }
2821#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002822 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2823 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002824 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2825 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002826 ext4_free_blks_set(sb, gdp,
2827 ext4_free_blocks_after_init(sb,
2828 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002829 }
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05002830 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2831 ext4_free_blks_set(sb, gdp, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002832 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002833
2834 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002835 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002836 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002837 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002838 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002839 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2840 /* release all the reserved blocks if non delalloc */
2841 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
Alex Tomasc9de5602008-01-29 00:19:52 -05002842
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002843 if (sbi->s_log_groups_per_flex) {
2844 ext4_group_t flex_group = ext4_flex_group(sbi,
2845 ac->ac_b_ex.fe_group);
Theodore Ts'o9f24e422009-03-04 19:09:10 -05002846 atomic_sub(ac->ac_b_ex.fe_len,
2847 &sbi->s_flex_groups[flex_group].free_blocks);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002848 }
2849
Frank Mayhar03901312009-01-07 00:06:22 -05002850 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002851 if (err)
2852 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002853 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002854
2855out_err:
Theodore Ts'oa0375152010-06-11 23:14:04 -04002856 ext4_mark_super_dirty(sb);
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002857 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002858 return err;
2859}
2860
2861/*
2862 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002863 * Group request are normalized to s_mb_group_prealloc, which goes to
2864 * s_strip if we set the same via mount option.
2865 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002866 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002867 *
2868 * XXX: should we try to preallocate more than the group has now?
2869 */
2870static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2871{
2872 struct super_block *sb = ac->ac_sb;
2873 struct ext4_locality_group *lg = ac->ac_lg;
2874
2875 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002876 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002877 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002878 current->pid, ac->ac_g_ex.fe_len);
2879}
2880
2881/*
2882 * Normalization means making request better in terms of
2883 * size and alignment
2884 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002885static noinline_for_stack void
2886ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05002887 struct ext4_allocation_request *ar)
2888{
2889 int bsbits, max;
2890 ext4_lblk_t end;
Alex Tomasc9de5602008-01-29 00:19:52 -05002891 loff_t size, orig_size, start_off;
Andi Kleen5a0790c2010-06-14 13:28:03 -04002892 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05002893 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002894 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05002895
2896 /* do normalize only data requests, metadata requests
2897 do not need preallocation */
2898 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2899 return;
2900
2901 /* sometime caller may want exact blocks */
2902 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2903 return;
2904
2905 /* caller may indicate that preallocation isn't
2906 * required (it's a tail, for example) */
2907 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2908 return;
2909
2910 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2911 ext4_mb_normalize_group_request(ac);
2912 return ;
2913 }
2914
2915 bsbits = ac->ac_sb->s_blocksize_bits;
2916
2917 /* first, let's learn actual file size
2918 * given current request is allocated */
2919 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2920 size = size << bsbits;
2921 if (size < i_size_read(ac->ac_inode))
2922 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04002923 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05002924
Valerie Clement19304792008-05-13 19:31:14 -04002925 /* max size of free chunks */
2926 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05002927
Valerie Clement19304792008-05-13 19:31:14 -04002928#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2929 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05002930
2931 /* first, try to predict filesize */
2932 /* XXX: should this table be tunable? */
2933 start_off = 0;
2934 if (size <= 16 * 1024) {
2935 size = 16 * 1024;
2936 } else if (size <= 32 * 1024) {
2937 size = 32 * 1024;
2938 } else if (size <= 64 * 1024) {
2939 size = 64 * 1024;
2940 } else if (size <= 128 * 1024) {
2941 size = 128 * 1024;
2942 } else if (size <= 256 * 1024) {
2943 size = 256 * 1024;
2944 } else if (size <= 512 * 1024) {
2945 size = 512 * 1024;
2946 } else if (size <= 1024 * 1024) {
2947 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04002948 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002949 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04002950 (21 - bsbits)) << 21;
2951 size = 2 * 1024 * 1024;
2952 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002953 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2954 (22 - bsbits)) << 22;
2955 size = 4 * 1024 * 1024;
2956 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04002957 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002958 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2959 (23 - bsbits)) << 23;
2960 size = 8 * 1024 * 1024;
2961 } else {
2962 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2963 size = ac->ac_o_ex.fe_len << bsbits;
2964 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04002965 size = size >> bsbits;
2966 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05002967
2968 /* don't cover already allocated blocks in selected range */
2969 if (ar->pleft && start <= ar->lleft) {
2970 size -= ar->lleft + 1 - start;
2971 start = ar->lleft + 1;
2972 }
2973 if (ar->pright && start + size - 1 >= ar->lright)
2974 size -= start + size - ar->lright;
2975
2976 end = start + size;
2977
2978 /* check we don't cross already preallocated blocks */
2979 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002980 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05002981 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05002982
Alex Tomasc9de5602008-01-29 00:19:52 -05002983 if (pa->pa_deleted)
2984 continue;
2985 spin_lock(&pa->pa_lock);
2986 if (pa->pa_deleted) {
2987 spin_unlock(&pa->pa_lock);
2988 continue;
2989 }
2990
2991 pa_end = pa->pa_lstart + pa->pa_len;
2992
2993 /* PA must not overlap original request */
2994 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2995 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2996
Eric Sandeen38877f42009-08-17 23:55:24 -04002997 /* skip PAs this normalized request doesn't overlap with */
2998 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002999 spin_unlock(&pa->pa_lock);
3000 continue;
3001 }
3002 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3003
Eric Sandeen38877f42009-08-17 23:55:24 -04003004 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003005 if (pa_end <= ac->ac_o_ex.fe_logical) {
3006 BUG_ON(pa_end < start);
3007 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003008 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003009 BUG_ON(pa->pa_lstart > end);
3010 end = pa->pa_lstart;
3011 }
3012 spin_unlock(&pa->pa_lock);
3013 }
3014 rcu_read_unlock();
3015 size = end - start;
3016
3017 /* XXX: extra loop to check we really don't overlap preallocations */
3018 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003019 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003020 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003021 spin_lock(&pa->pa_lock);
3022 if (pa->pa_deleted == 0) {
3023 pa_end = pa->pa_lstart + pa->pa_len;
3024 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3025 }
3026 spin_unlock(&pa->pa_lock);
3027 }
3028 rcu_read_unlock();
3029
3030 if (start + size <= ac->ac_o_ex.fe_logical &&
3031 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003032 ext4_msg(ac->ac_sb, KERN_ERR,
3033 "start %lu, size %lu, fe_logical %lu",
3034 (unsigned long) start, (unsigned long) size,
3035 (unsigned long) ac->ac_o_ex.fe_logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05003036 }
3037 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3038 start > ac->ac_o_ex.fe_logical);
Eric Sandeen8d03c7a2009-03-14 11:51:46 -04003039 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003040
3041 /* now prepare goal request */
3042
3043 /* XXX: is it better to align blocks WRT to logical
3044 * placement or satisfy big request as is */
3045 ac->ac_g_ex.fe_logical = start;
3046 ac->ac_g_ex.fe_len = size;
3047
3048 /* define goal start in order to merge */
3049 if (ar->pright && (ar->lright == (start + size))) {
3050 /* merge to the right */
3051 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3052 &ac->ac_f_ex.fe_group,
3053 &ac->ac_f_ex.fe_start);
3054 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3055 }
3056 if (ar->pleft && (ar->lleft + 1 == start)) {
3057 /* merge to the left */
3058 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3059 &ac->ac_f_ex.fe_group,
3060 &ac->ac_f_ex.fe_start);
3061 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3062 }
3063
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003064 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003065 (unsigned) orig_size, (unsigned) start);
3066}
3067
3068static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3069{
3070 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3071
3072 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3073 atomic_inc(&sbi->s_bal_reqs);
3074 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003075 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003076 atomic_inc(&sbi->s_bal_success);
3077 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3078 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3079 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3080 atomic_inc(&sbi->s_bal_goals);
3081 if (ac->ac_found > sbi->s_mb_max_to_scan)
3082 atomic_inc(&sbi->s_bal_breaks);
3083 }
3084
Theodore Ts'o296c3552009-09-30 00:32:42 -04003085 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3086 trace_ext4_mballoc_alloc(ac);
3087 else
3088 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003089}
3090
3091/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003092 * Called on failure; free up any blocks from the inode PA for this
3093 * context. We don't need this for MB_GROUP_PA because we only change
3094 * pa_free in ext4_mb_release_context(), but on failure, we've already
3095 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3096 */
3097static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3098{
3099 struct ext4_prealloc_space *pa = ac->ac_pa;
3100 int len;
3101
3102 if (pa && pa->pa_type == MB_INODE_PA) {
3103 len = ac->ac_b_ex.fe_len;
3104 pa->pa_free += len;
3105 }
3106
3107}
3108
3109/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003110 * use blocks preallocated to inode
3111 */
3112static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3113 struct ext4_prealloc_space *pa)
3114{
3115 ext4_fsblk_t start;
3116 ext4_fsblk_t end;
3117 int len;
3118
3119 /* found preallocated blocks, use them */
3120 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3121 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3122 len = end - start;
3123 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3124 &ac->ac_b_ex.fe_start);
3125 ac->ac_b_ex.fe_len = len;
3126 ac->ac_status = AC_STATUS_FOUND;
3127 ac->ac_pa = pa;
3128
3129 BUG_ON(start < pa->pa_pstart);
3130 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3131 BUG_ON(pa->pa_free < len);
3132 pa->pa_free -= len;
3133
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003134 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003135}
3136
3137/*
3138 * use blocks preallocated to locality group
3139 */
3140static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3141 struct ext4_prealloc_space *pa)
3142{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003143 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003144
Alex Tomasc9de5602008-01-29 00:19:52 -05003145 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3146 &ac->ac_b_ex.fe_group,
3147 &ac->ac_b_ex.fe_start);
3148 ac->ac_b_ex.fe_len = len;
3149 ac->ac_status = AC_STATUS_FOUND;
3150 ac->ac_pa = pa;
3151
3152 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003153 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003154 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003155 * in on-disk bitmap -- see ext4_mb_release_context()
3156 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003157 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003158 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003159}
3160
3161/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003162 * Return the prealloc space that have minimal distance
3163 * from the goal block. @cpa is the prealloc
3164 * space that is having currently known minimal distance
3165 * from the goal block.
3166 */
3167static struct ext4_prealloc_space *
3168ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3169 struct ext4_prealloc_space *pa,
3170 struct ext4_prealloc_space *cpa)
3171{
3172 ext4_fsblk_t cur_distance, new_distance;
3173
3174 if (cpa == NULL) {
3175 atomic_inc(&pa->pa_count);
3176 return pa;
3177 }
3178 cur_distance = abs(goal_block - cpa->pa_pstart);
3179 new_distance = abs(goal_block - pa->pa_pstart);
3180
Coly Li5a54b2f2011-02-24 14:10:05 -05003181 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003182 return cpa;
3183
3184 /* drop the previous reference */
3185 atomic_dec(&cpa->pa_count);
3186 atomic_inc(&pa->pa_count);
3187 return pa;
3188}
3189
3190/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003191 * search goal blocks in preallocated space
3192 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003193static noinline_for_stack int
3194ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003195{
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003196 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003197 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3198 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003199 struct ext4_prealloc_space *pa, *cpa = NULL;
3200 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003201
3202 /* only data can be preallocated */
3203 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3204 return 0;
3205
3206 /* first, try per-file preallocation */
3207 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003208 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003209
3210 /* all fields in this condition don't change,
3211 * so we can skip locking for them */
3212 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3213 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3214 continue;
3215
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003216 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003217 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003218 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3219 continue;
3220
Alex Tomasc9de5602008-01-29 00:19:52 -05003221 /* found preallocated blocks, use them */
3222 spin_lock(&pa->pa_lock);
3223 if (pa->pa_deleted == 0 && pa->pa_free) {
3224 atomic_inc(&pa->pa_count);
3225 ext4_mb_use_inode_pa(ac, pa);
3226 spin_unlock(&pa->pa_lock);
3227 ac->ac_criteria = 10;
3228 rcu_read_unlock();
3229 return 1;
3230 }
3231 spin_unlock(&pa->pa_lock);
3232 }
3233 rcu_read_unlock();
3234
3235 /* can we use group allocation? */
3236 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3237 return 0;
3238
3239 /* inode may have no locality group for some reason */
3240 lg = ac->ac_lg;
3241 if (lg == NULL)
3242 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003243 order = fls(ac->ac_o_ex.fe_len) - 1;
3244 if (order > PREALLOC_TB_SIZE - 1)
3245 /* The max size of hash table is PREALLOC_TB_SIZE */
3246 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003247
Akinobu Mitabda00de2010-03-03 23:53:25 -05003248 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003249 /*
3250 * search for the prealloc space that is having
3251 * minimal distance from the goal block.
3252 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003253 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3254 rcu_read_lock();
3255 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3256 pa_inode_list) {
3257 spin_lock(&pa->pa_lock);
3258 if (pa->pa_deleted == 0 &&
3259 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003260
3261 cpa = ext4_mb_check_group_pa(goal_block,
3262 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003263 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003264 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003265 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003266 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003267 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003268 if (cpa) {
3269 ext4_mb_use_group_pa(ac, cpa);
3270 ac->ac_criteria = 20;
3271 return 1;
3272 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003273 return 0;
3274}
3275
3276/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003277 * the function goes through all block freed in the group
3278 * but not yet committed and marks them used in in-core bitmap.
3279 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003280 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003281 */
3282static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3283 ext4_group_t group)
3284{
3285 struct rb_node *n;
3286 struct ext4_group_info *grp;
3287 struct ext4_free_data *entry;
3288
3289 grp = ext4_get_group_info(sb, group);
3290 n = rb_first(&(grp->bb_free_root));
3291
3292 while (n) {
3293 entry = rb_entry(n, struct ext4_free_data, node);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003294 ext4_set_bits(bitmap, entry->start_blk, entry->count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003295 n = rb_next(n);
3296 }
3297 return;
3298}
3299
3300/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003301 * the function goes through all preallocation in this group and marks them
3302 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003303 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003304 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003305static noinline_for_stack
3306void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003307 ext4_group_t group)
3308{
3309 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3310 struct ext4_prealloc_space *pa;
3311 struct list_head *cur;
3312 ext4_group_t groupnr;
3313 ext4_grpblk_t start;
3314 int preallocated = 0;
3315 int count = 0;
3316 int len;
3317
3318 /* all form of preallocation discards first load group,
3319 * so the only competing code is preallocation use.
3320 * we don't need any locking here
3321 * notice we do NOT ignore preallocations with pa_deleted
3322 * otherwise we could leave used blocks available for
3323 * allocation in buddy when concurrent ext4_mb_put_pa()
3324 * is dropping preallocation
3325 */
3326 list_for_each(cur, &grp->bb_prealloc_list) {
3327 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3328 spin_lock(&pa->pa_lock);
3329 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3330 &groupnr, &start);
3331 len = pa->pa_len;
3332 spin_unlock(&pa->pa_lock);
3333 if (unlikely(len == 0))
3334 continue;
3335 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003336 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003337 preallocated += len;
3338 count++;
3339 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003340 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003341}
3342
3343static void ext4_mb_pa_callback(struct rcu_head *head)
3344{
3345 struct ext4_prealloc_space *pa;
3346 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3347 kmem_cache_free(ext4_pspace_cachep, pa);
3348}
3349
3350/*
3351 * drops a reference to preallocated space descriptor
3352 * if this was the last reference and the space is consumed
3353 */
3354static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3355 struct super_block *sb, struct ext4_prealloc_space *pa)
3356{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003357 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003358 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003359
3360 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3361 return;
3362
3363 /* in this short window concurrent discard can set pa_deleted */
3364 spin_lock(&pa->pa_lock);
3365 if (pa->pa_deleted == 1) {
3366 spin_unlock(&pa->pa_lock);
3367 return;
3368 }
3369
3370 pa->pa_deleted = 1;
3371 spin_unlock(&pa->pa_lock);
3372
Eric Sandeend33a1972009-03-16 23:25:40 -04003373 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003374 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003375 * If doing group-based preallocation, pa_pstart may be in the
3376 * next group when pa is used up
3377 */
3378 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003379 grp_blk--;
3380
3381 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
Alex Tomasc9de5602008-01-29 00:19:52 -05003382
3383 /*
3384 * possible race:
3385 *
3386 * P1 (buddy init) P2 (regular allocation)
3387 * find block B in PA
3388 * copy on-disk bitmap to buddy
3389 * mark B in on-disk bitmap
3390 * drop PA from group
3391 * mark all PAs in buddy
3392 *
3393 * thus, P1 initializes buddy with B available. to prevent this
3394 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3395 * against that pair
3396 */
3397 ext4_lock_group(sb, grp);
3398 list_del(&pa->pa_group_list);
3399 ext4_unlock_group(sb, grp);
3400
3401 spin_lock(pa->pa_obj_lock);
3402 list_del_rcu(&pa->pa_inode_list);
3403 spin_unlock(pa->pa_obj_lock);
3404
3405 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3406}
3407
3408/*
3409 * creates new preallocated space for given inode
3410 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003411static noinline_for_stack int
3412ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003413{
3414 struct super_block *sb = ac->ac_sb;
3415 struct ext4_prealloc_space *pa;
3416 struct ext4_group_info *grp;
3417 struct ext4_inode_info *ei;
3418
3419 /* preallocate only when found space is larger then requested */
3420 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3421 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3422 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3423
3424 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3425 if (pa == NULL)
3426 return -ENOMEM;
3427
3428 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3429 int winl;
3430 int wins;
3431 int win;
3432 int offs;
3433
3434 /* we can't allocate as much as normalizer wants.
3435 * so, found space must get proper lstart
3436 * to cover original request */
3437 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3438 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3439
3440 /* we're limited by original request in that
3441 * logical block must be covered any way
3442 * winl is window we can move our chunk within */
3443 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3444
3445 /* also, we should cover whole original request */
3446 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3447
3448 /* the smallest one defines real window */
3449 win = min(winl, wins);
3450
3451 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3452 if (offs && offs < win)
3453 win = offs;
3454
3455 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3456 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3457 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3458 }
3459
3460 /* preallocation can change ac_b_ex, thus we store actually
3461 * allocated blocks for history */
3462 ac->ac_f_ex = ac->ac_b_ex;
3463
3464 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3465 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3466 pa->pa_len = ac->ac_b_ex.fe_len;
3467 pa->pa_free = pa->pa_len;
3468 atomic_set(&pa->pa_count, 1);
3469 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003470 INIT_LIST_HEAD(&pa->pa_inode_list);
3471 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003472 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003473 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003474
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003475 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003476 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003477 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003478
3479 ext4_mb_use_inode_pa(ac, pa);
3480 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3481
3482 ei = EXT4_I(ac->ac_inode);
3483 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3484
3485 pa->pa_obj_lock = &ei->i_prealloc_lock;
3486 pa->pa_inode = ac->ac_inode;
3487
3488 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3489 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3490 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3491
3492 spin_lock(pa->pa_obj_lock);
3493 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3494 spin_unlock(pa->pa_obj_lock);
3495
3496 return 0;
3497}
3498
3499/*
3500 * creates new preallocated space for locality group inodes belongs to
3501 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003502static noinline_for_stack int
3503ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003504{
3505 struct super_block *sb = ac->ac_sb;
3506 struct ext4_locality_group *lg;
3507 struct ext4_prealloc_space *pa;
3508 struct ext4_group_info *grp;
3509
3510 /* preallocate only when found space is larger then requested */
3511 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3512 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3513 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3514
3515 BUG_ON(ext4_pspace_cachep == NULL);
3516 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3517 if (pa == NULL)
3518 return -ENOMEM;
3519
3520 /* preallocation can change ac_b_ex, thus we store actually
3521 * allocated blocks for history */
3522 ac->ac_f_ex = ac->ac_b_ex;
3523
3524 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3525 pa->pa_lstart = pa->pa_pstart;
3526 pa->pa_len = ac->ac_b_ex.fe_len;
3527 pa->pa_free = pa->pa_len;
3528 atomic_set(&pa->pa_count, 1);
3529 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003530 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003531 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003532 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003533 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003534
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003535 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003536 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3537 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003538
3539 ext4_mb_use_group_pa(ac, pa);
3540 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3541
3542 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3543 lg = ac->ac_lg;
3544 BUG_ON(lg == NULL);
3545
3546 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3547 pa->pa_inode = NULL;
3548
3549 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3550 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3551 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3552
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003553 /*
3554 * We will later add the new pa to the right bucket
3555 * after updating the pa_free in ext4_mb_release_context
3556 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003557 return 0;
3558}
3559
3560static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3561{
3562 int err;
3563
3564 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3565 err = ext4_mb_new_group_pa(ac);
3566 else
3567 err = ext4_mb_new_inode_pa(ac);
3568 return err;
3569}
3570
3571/*
3572 * finds all unused blocks in on-disk bitmap, frees them in
3573 * in-core bitmap and buddy.
3574 * @pa must be unlinked from inode and group lists, so that
3575 * nobody else can find/use it.
3576 * the caller MUST hold group/inode locks.
3577 * TODO: optimize the case when there are no in-core structures yet
3578 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003579static noinline_for_stack int
3580ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003581 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003582{
Alex Tomasc9de5602008-01-29 00:19:52 -05003583 struct super_block *sb = e4b->bd_sb;
3584 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003585 unsigned int end;
3586 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003587 ext4_group_t group;
3588 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003589 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003590 int err = 0;
3591 int free = 0;
3592
3593 BUG_ON(pa->pa_deleted == 0);
3594 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'oba80b102009-01-03 20:03:21 -05003595 grp_blk_start = pa->pa_pstart - bit;
Alex Tomasc9de5602008-01-29 00:19:52 -05003596 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3597 end = bit + pa->pa_len;
3598
Alex Tomasc9de5602008-01-29 00:19:52 -05003599 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003600 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003601 if (bit >= end)
3602 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003603 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003604 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003605 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3606 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003607 free += next - bit;
3608
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003609 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Lukas Czernera9c667f2011-06-06 09:51:52 -04003610 trace_ext4_mb_release_inode_pa(pa, grp_blk_start + bit,
3611 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003612 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3613 bit = next + 1;
3614 }
3615 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003616 ext4_msg(e4b->bd_sb, KERN_CRIT,
3617 "pa %p: logic %lu, phys. %lu, len %lu",
3618 pa, (unsigned long) pa->pa_lstart,
3619 (unsigned long) pa->pa_pstart,
3620 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003621 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003622 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003623 /*
3624 * pa is already deleted so we use the value obtained
3625 * from the bitmap and continue.
3626 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003627 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003628 atomic_add(free, &sbi->s_mb_discarded);
3629
3630 return err;
3631}
3632
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003633static noinline_for_stack int
3634ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003635 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003636{
Alex Tomasc9de5602008-01-29 00:19:52 -05003637 struct super_block *sb = e4b->bd_sb;
3638 ext4_group_t group;
3639 ext4_grpblk_t bit;
3640
Lukas Czernera9c667f2011-06-06 09:51:52 -04003641 trace_ext4_mb_release_group_pa(pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003642 BUG_ON(pa->pa_deleted == 0);
3643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3644 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3645 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3646 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003647 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003648
3649 return 0;
3650}
3651
3652/*
3653 * releases all preallocations in given group
3654 *
3655 * first, we need to decide discard policy:
3656 * - when do we discard
3657 * 1) ENOSPC
3658 * - how many do we discard
3659 * 1) how many requested
3660 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003661static noinline_for_stack int
3662ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003663 ext4_group_t group, int needed)
3664{
3665 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3666 struct buffer_head *bitmap_bh = NULL;
3667 struct ext4_prealloc_space *pa, *tmp;
3668 struct list_head list;
3669 struct ext4_buddy e4b;
3670 int err;
3671 int busy = 0;
3672 int free = 0;
3673
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003674 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003675
3676 if (list_empty(&grp->bb_prealloc_list))
3677 return 0;
3678
Theodore Ts'o574ca172008-07-11 19:27:31 -04003679 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003680 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003681 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003682 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003683 }
3684
3685 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003686 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003687 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003688 put_bh(bitmap_bh);
3689 return 0;
3690 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003691
3692 if (needed == 0)
3693 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3694
Alex Tomasc9de5602008-01-29 00:19:52 -05003695 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003696repeat:
3697 ext4_lock_group(sb, group);
3698 list_for_each_entry_safe(pa, tmp,
3699 &grp->bb_prealloc_list, pa_group_list) {
3700 spin_lock(&pa->pa_lock);
3701 if (atomic_read(&pa->pa_count)) {
3702 spin_unlock(&pa->pa_lock);
3703 busy = 1;
3704 continue;
3705 }
3706 if (pa->pa_deleted) {
3707 spin_unlock(&pa->pa_lock);
3708 continue;
3709 }
3710
3711 /* seems this one can be freed ... */
3712 pa->pa_deleted = 1;
3713
3714 /* we can trust pa_free ... */
3715 free += pa->pa_free;
3716
3717 spin_unlock(&pa->pa_lock);
3718
3719 list_del(&pa->pa_group_list);
3720 list_add(&pa->u.pa_tmp_list, &list);
3721 }
3722
3723 /* if we still need more blocks and some PAs were used, try again */
3724 if (free < needed && busy) {
3725 busy = 0;
3726 ext4_unlock_group(sb, group);
3727 /*
3728 * Yield the CPU here so that we don't get soft lockup
3729 * in non preempt case.
3730 */
3731 yield();
3732 goto repeat;
3733 }
3734
3735 /* found anything to free? */
3736 if (list_empty(&list)) {
3737 BUG_ON(free != 0);
3738 goto out;
3739 }
3740
3741 /* now free all selected PAs */
3742 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3743
3744 /* remove from object (inode or locality group) */
3745 spin_lock(pa->pa_obj_lock);
3746 list_del_rcu(&pa->pa_inode_list);
3747 spin_unlock(pa->pa_obj_lock);
3748
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003749 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003750 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003751 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003752 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003753
3754 list_del(&pa->u.pa_tmp_list);
3755 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3756 }
3757
3758out:
3759 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003760 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003761 put_bh(bitmap_bh);
3762 return free;
3763}
3764
3765/*
3766 * releases all non-used preallocated blocks for given inode
3767 *
3768 * It's important to discard preallocations under i_data_sem
3769 * We don't want another block to be served from the prealloc
3770 * space when we are discarding the inode prealloc space.
3771 *
3772 * FIXME!! Make sure it is valid at all the call sites
3773 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003774void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003775{
3776 struct ext4_inode_info *ei = EXT4_I(inode);
3777 struct super_block *sb = inode->i_sb;
3778 struct buffer_head *bitmap_bh = NULL;
3779 struct ext4_prealloc_space *pa, *tmp;
3780 ext4_group_t group = 0;
3781 struct list_head list;
3782 struct ext4_buddy e4b;
3783 int err;
3784
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003785 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003786 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3787 return;
3788 }
3789
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003790 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003791 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003792
3793 INIT_LIST_HEAD(&list);
3794
3795repeat:
3796 /* first, collect all pa's in the inode */
3797 spin_lock(&ei->i_prealloc_lock);
3798 while (!list_empty(&ei->i_prealloc_list)) {
3799 pa = list_entry(ei->i_prealloc_list.next,
3800 struct ext4_prealloc_space, pa_inode_list);
3801 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3802 spin_lock(&pa->pa_lock);
3803 if (atomic_read(&pa->pa_count)) {
3804 /* this shouldn't happen often - nobody should
3805 * use preallocation while we're discarding it */
3806 spin_unlock(&pa->pa_lock);
3807 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003808 ext4_msg(sb, KERN_ERR,
3809 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003810 WARN_ON(1);
3811 schedule_timeout_uninterruptible(HZ);
3812 goto repeat;
3813
3814 }
3815 if (pa->pa_deleted == 0) {
3816 pa->pa_deleted = 1;
3817 spin_unlock(&pa->pa_lock);
3818 list_del_rcu(&pa->pa_inode_list);
3819 list_add(&pa->u.pa_tmp_list, &list);
3820 continue;
3821 }
3822
3823 /* someone is deleting pa right now */
3824 spin_unlock(&pa->pa_lock);
3825 spin_unlock(&ei->i_prealloc_lock);
3826
3827 /* we have to wait here because pa_deleted
3828 * doesn't mean pa is already unlinked from
3829 * the list. as we might be called from
3830 * ->clear_inode() the inode will get freed
3831 * and concurrent thread which is unlinking
3832 * pa from inode's list may access already
3833 * freed memory, bad-bad-bad */
3834
3835 /* XXX: if this happens too often, we can
3836 * add a flag to force wait only in case
3837 * of ->clear_inode(), but not in case of
3838 * regular truncate */
3839 schedule_timeout_uninterruptible(HZ);
3840 goto repeat;
3841 }
3842 spin_unlock(&ei->i_prealloc_lock);
3843
3844 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003845 BUG_ON(pa->pa_type != MB_INODE_PA);
Alex Tomasc9de5602008-01-29 00:19:52 -05003846 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3847
3848 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003849 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003850 ext4_error(sb, "Error loading buddy information for %u",
3851 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003852 continue;
3853 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003854
Theodore Ts'o574ca172008-07-11 19:27:31 -04003855 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003856 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003857 ext4_error(sb, "Error reading block bitmap for %u",
3858 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003859 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003860 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05003861 }
3862
3863 ext4_lock_group(sb, group);
3864 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003865 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003866 ext4_unlock_group(sb, group);
3867
Jing Zhange39e07f2010-05-14 00:00:00 -04003868 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003869 put_bh(bitmap_bh);
3870
3871 list_del(&pa->u.pa_tmp_list);
3872 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3873 }
3874}
3875
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003876#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05003877static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3878{
3879 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04003880 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003881
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05003882 if (!mb_enable_debug ||
3883 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04003884 return;
3885
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003886 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: Can't allocate:"
3887 " Allocation context details:");
3888 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003889 ac->ac_status, ac->ac_flags);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003890 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: orig %lu/%lu/%lu@%lu, "
3891 "goal %lu/%lu/%lu@%lu, "
3892 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003893 (unsigned long)ac->ac_o_ex.fe_group,
3894 (unsigned long)ac->ac_o_ex.fe_start,
3895 (unsigned long)ac->ac_o_ex.fe_len,
3896 (unsigned long)ac->ac_o_ex.fe_logical,
3897 (unsigned long)ac->ac_g_ex.fe_group,
3898 (unsigned long)ac->ac_g_ex.fe_start,
3899 (unsigned long)ac->ac_g_ex.fe_len,
3900 (unsigned long)ac->ac_g_ex.fe_logical,
3901 (unsigned long)ac->ac_b_ex.fe_group,
3902 (unsigned long)ac->ac_b_ex.fe_start,
3903 (unsigned long)ac->ac_b_ex.fe_len,
3904 (unsigned long)ac->ac_b_ex.fe_logical,
3905 (int)ac->ac_criteria);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003906 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: %lu scanned, %d found",
3907 ac->ac_ex_scanned, ac->ac_found);
3908 ext4_msg(ac->ac_sb, KERN_ERR, "EXT4-fs: groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04003909 ngroups = ext4_get_groups_count(sb);
3910 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003911 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3912 struct ext4_prealloc_space *pa;
3913 ext4_grpblk_t start;
3914 struct list_head *cur;
3915 ext4_lock_group(sb, i);
3916 list_for_each(cur, &grp->bb_prealloc_list) {
3917 pa = list_entry(cur, struct ext4_prealloc_space,
3918 pa_group_list);
3919 spin_lock(&pa->pa_lock);
3920 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3921 NULL, &start);
3922 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04003923 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3924 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003925 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04003926 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05003927
3928 if (grp->bb_free == 0)
3929 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04003930 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003931 i, grp->bb_free, grp->bb_fragments);
3932 }
3933 printk(KERN_ERR "\n");
3934}
3935#else
3936static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3937{
3938 return;
3939}
3940#endif
3941
3942/*
3943 * We use locality group preallocation for small size file. The size of the
3944 * file is determined by the current size or the resulting size after
3945 * allocation which ever is larger
3946 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04003947 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05003948 */
3949static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3950{
3951 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3952 int bsbits = ac->ac_sb->s_blocksize_bits;
3953 loff_t size, isize;
3954
3955 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3956 return;
3957
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003958 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3959 return;
3960
Alex Tomasc9de5602008-01-29 00:19:52 -05003961 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
Theodore Ts'o50797482009-09-18 13:34:02 -04003962 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3963 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003964
Theodore Ts'o50797482009-09-18 13:34:02 -04003965 if ((size == isize) &&
3966 !ext4_fs_is_busy(sbi) &&
3967 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3968 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3969 return;
3970 }
3971
Alex Tomasc9de5602008-01-29 00:19:52 -05003972 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04003973 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05003974 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003975 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05003976 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04003977 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003978
3979 BUG_ON(ac->ac_lg != NULL);
3980 /*
3981 * locality group prealloc space are per cpu. The reason for having
3982 * per cpu locality group is to reduce the contention between block
3983 * request from multiple CPUs.
3984 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09003985 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05003986
3987 /* we're going to use group allocation */
3988 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3989
3990 /* serialize all allocations in the group */
3991 mutex_lock(&ac->ac_lg->lg_mutex);
3992}
3993
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003994static noinline_for_stack int
3995ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05003996 struct ext4_allocation_request *ar)
3997{
3998 struct super_block *sb = ar->inode->i_sb;
3999 struct ext4_sb_info *sbi = EXT4_SB(sb);
4000 struct ext4_super_block *es = sbi->s_es;
4001 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004002 unsigned int len;
4003 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004004 ext4_grpblk_t block;
4005
4006 /* we can't allocate > group size */
4007 len = ar->len;
4008
4009 /* just a dirty hack to filter too big requests */
4010 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
4011 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
4012
4013 /* start searching from the goal */
4014 goal = ar->goal;
4015 if (goal < le32_to_cpu(es->s_first_data_block) ||
4016 goal >= ext4_blocks_count(es))
4017 goal = le32_to_cpu(es->s_first_data_block);
4018 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4019
4020 /* set up allocation goals */
Theodore Ts'o833576b2009-07-13 09:45:52 -04004021 memset(ac, 0, sizeof(struct ext4_allocation_context));
Alex Tomasc9de5602008-01-29 00:19:52 -05004022 ac->ac_b_ex.fe_logical = ar->logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004023 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004024 ac->ac_sb = sb;
4025 ac->ac_inode = ar->inode;
4026 ac->ac_o_ex.fe_logical = ar->logical;
4027 ac->ac_o_ex.fe_group = group;
4028 ac->ac_o_ex.fe_start = block;
4029 ac->ac_o_ex.fe_len = len;
4030 ac->ac_g_ex.fe_logical = ar->logical;
4031 ac->ac_g_ex.fe_group = group;
4032 ac->ac_g_ex.fe_start = block;
4033 ac->ac_g_ex.fe_len = len;
Alex Tomasc9de5602008-01-29 00:19:52 -05004034 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004035
4036 /* we have to define context: we'll we work with a file or
4037 * locality group. this is a policy, actually */
4038 ext4_mb_group_or_file(ac);
4039
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004040 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004041 "left: %u/%u, right %u/%u to %swritable\n",
4042 (unsigned) ar->len, (unsigned) ar->logical,
4043 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4044 (unsigned) ar->lleft, (unsigned) ar->pleft,
4045 (unsigned) ar->lright, (unsigned) ar->pright,
4046 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4047 return 0;
4048
4049}
4050
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004051static noinline_for_stack void
4052ext4_mb_discard_lg_preallocations(struct super_block *sb,
4053 struct ext4_locality_group *lg,
4054 int order, int total_entries)
4055{
4056 ext4_group_t group = 0;
4057 struct ext4_buddy e4b;
4058 struct list_head discard_list;
4059 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004060
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004061 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004062
4063 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004064
4065 spin_lock(&lg->lg_prealloc_lock);
4066 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4067 pa_inode_list) {
4068 spin_lock(&pa->pa_lock);
4069 if (atomic_read(&pa->pa_count)) {
4070 /*
4071 * This is the pa that we just used
4072 * for block allocation. So don't
4073 * free that
4074 */
4075 spin_unlock(&pa->pa_lock);
4076 continue;
4077 }
4078 if (pa->pa_deleted) {
4079 spin_unlock(&pa->pa_lock);
4080 continue;
4081 }
4082 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004083 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004084
4085 /* seems this one can be freed ... */
4086 pa->pa_deleted = 1;
4087 spin_unlock(&pa->pa_lock);
4088
4089 list_del_rcu(&pa->pa_inode_list);
4090 list_add(&pa->u.pa_tmp_list, &discard_list);
4091
4092 total_entries--;
4093 if (total_entries <= 5) {
4094 /*
4095 * we want to keep only 5 entries
4096 * allowing it to grow to 8. This
4097 * mak sure we don't call discard
4098 * soon for this list.
4099 */
4100 break;
4101 }
4102 }
4103 spin_unlock(&lg->lg_prealloc_lock);
4104
4105 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4106
4107 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4108 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004109 ext4_error(sb, "Error loading buddy information for %u",
4110 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004111 continue;
4112 }
4113 ext4_lock_group(sb, group);
4114 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004115 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004116 ext4_unlock_group(sb, group);
4117
Jing Zhange39e07f2010-05-14 00:00:00 -04004118 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004119 list_del(&pa->u.pa_tmp_list);
4120 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4121 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004122}
4123
4124/*
4125 * We have incremented pa_count. So it cannot be freed at this
4126 * point. Also we hold lg_mutex. So no parallel allocation is
4127 * possible from this lg. That means pa_free cannot be updated.
4128 *
4129 * A parallel ext4_mb_discard_group_preallocations is possible.
4130 * which can cause the lg_prealloc_list to be updated.
4131 */
4132
4133static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4134{
4135 int order, added = 0, lg_prealloc_count = 1;
4136 struct super_block *sb = ac->ac_sb;
4137 struct ext4_locality_group *lg = ac->ac_lg;
4138 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4139
4140 order = fls(pa->pa_free) - 1;
4141 if (order > PREALLOC_TB_SIZE - 1)
4142 /* The max size of hash table is PREALLOC_TB_SIZE */
4143 order = PREALLOC_TB_SIZE - 1;
4144 /* Add the prealloc space to lg */
4145 rcu_read_lock();
4146 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4147 pa_inode_list) {
4148 spin_lock(&tmp_pa->pa_lock);
4149 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004150 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004151 continue;
4152 }
4153 if (!added && pa->pa_free < tmp_pa->pa_free) {
4154 /* Add to the tail of the previous entry */
4155 list_add_tail_rcu(&pa->pa_inode_list,
4156 &tmp_pa->pa_inode_list);
4157 added = 1;
4158 /*
4159 * we want to count the total
4160 * number of entries in the list
4161 */
4162 }
4163 spin_unlock(&tmp_pa->pa_lock);
4164 lg_prealloc_count++;
4165 }
4166 if (!added)
4167 list_add_tail_rcu(&pa->pa_inode_list,
4168 &lg->lg_prealloc_list[order]);
4169 rcu_read_unlock();
4170
4171 /* Now trim the list to be not more than 8 elements */
4172 if (lg_prealloc_count > 8) {
4173 ext4_mb_discard_lg_preallocations(sb, lg,
4174 order, lg_prealloc_count);
4175 return;
4176 }
4177 return ;
4178}
4179
Alex Tomasc9de5602008-01-29 00:19:52 -05004180/*
4181 * release all resource we used in allocation
4182 */
4183static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4184{
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004185 struct ext4_prealloc_space *pa = ac->ac_pa;
4186 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004187 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004188 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004189 spin_lock(&pa->pa_lock);
4190 pa->pa_pstart += ac->ac_b_ex.fe_len;
4191 pa->pa_lstart += ac->ac_b_ex.fe_len;
4192 pa->pa_free -= ac->ac_b_ex.fe_len;
4193 pa->pa_len -= ac->ac_b_ex.fe_len;
4194 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004195 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004196 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004197 if (pa) {
4198 /*
4199 * We want to add the pa to the right bucket.
4200 * Remove it from the list and while adding
4201 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004202 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004203 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004204 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004205 spin_lock(pa->pa_obj_lock);
4206 list_del_rcu(&pa->pa_inode_list);
4207 spin_unlock(pa->pa_obj_lock);
4208 ext4_mb_add_n_trim(ac);
4209 }
4210 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4211 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004212 if (ac->ac_bitmap_page)
4213 page_cache_release(ac->ac_bitmap_page);
4214 if (ac->ac_buddy_page)
4215 page_cache_release(ac->ac_buddy_page);
4216 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4217 mutex_unlock(&ac->ac_lg->lg_mutex);
4218 ext4_mb_collect_stats(ac);
4219 return 0;
4220}
4221
4222static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4223{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004224 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004225 int ret;
4226 int freed = 0;
4227
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004228 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004229 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004230 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4231 freed += ret;
4232 needed -= ret;
4233 }
4234
4235 return freed;
4236}
4237
4238/*
4239 * Main entry point into mballoc to allocate blocks
4240 * it tries to use preallocation first, then falls back
4241 * to usual allocation
4242 */
4243ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004244 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004245{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004246 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004247 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004248 struct ext4_sb_info *sbi;
4249 struct super_block *sb;
4250 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004251 unsigned int inquota = 0;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004252 unsigned int reserv_blks = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004253
4254 sb = ar->inode->i_sb;
4255 sbi = EXT4_SB(sb);
4256
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004257 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004258
Mingming Cao60e58e02009-01-22 18:13:05 +01004259 /*
4260 * For delayed allocation, we could skip the ENOSPC and
4261 * EDQUOT check, as blocks and quotas have been already
4262 * reserved when data being copied into pagecache.
4263 */
Theodore Ts'of2321092011-01-10 12:12:36 -05004264 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
Mingming Cao60e58e02009-01-22 18:13:05 +01004265 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4266 else {
4267 /* Without delayed allocation we need to verify
4268 * there is enough free blocks to do block allocation
4269 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004270 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004271 while (ar->len &&
4272 ext4_claim_free_blocks(sbi, ar->len, ar->flags)) {
4273
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004274 /* let others to free the space */
4275 yield();
4276 ar->len = ar->len >> 1;
4277 }
4278 if (!ar->len) {
Aneesh Kumar K.Va30d5422008-10-09 10:56:23 -04004279 *errp = -ENOSPC;
4280 return 0;
4281 }
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004282 reserv_blks = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004283 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4284 dquot_alloc_block_nofail(ar->inode, ar->len);
4285 } else {
4286 while (ar->len &&
4287 dquot_alloc_block(ar->inode, ar->len)) {
4288
4289 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4290 ar->len--;
4291 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004292 }
4293 inquota = ar->len;
4294 if (ar->len == 0) {
4295 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004296 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004297 }
Mingming Caod2a17632008-07-14 17:52:37 -04004298 }
Mingming Caod2a17632008-07-14 17:52:37 -04004299
Eric Sandeen256bdb42008-02-10 01:13:33 -05004300 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004301 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004302 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004303 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004304 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004305 }
4306
Eric Sandeen256bdb42008-02-10 01:13:33 -05004307 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004308 if (*errp) {
4309 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004310 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004311 }
4312
Eric Sandeen256bdb42008-02-10 01:13:33 -05004313 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4314 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004315 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4316 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004317repeat:
4318 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004319 *errp = ext4_mb_regular_allocator(ac);
4320 if (*errp)
4321 goto errout;
Alex Tomasc9de5602008-01-29 00:19:52 -05004322
4323 /* as we've just preallocated more space than
4324 * user requested orinally, we store allocated
4325 * space in a special descriptor */
Eric Sandeen256bdb42008-02-10 01:13:33 -05004326 if (ac->ac_status == AC_STATUS_FOUND &&
4327 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4328 ext4_mb_new_preallocation(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004329 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004330 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004331 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004332 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004333 /*
4334 * drop the reference that we took
4335 * in ext4_mb_use_best_found
4336 */
4337 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004338 ac->ac_b_ex.fe_group = 0;
4339 ac->ac_b_ex.fe_start = 0;
4340 ac->ac_b_ex.fe_len = 0;
4341 ac->ac_status = AC_STATUS_CONTINUE;
4342 goto repeat;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004343 } else if (*errp)
4344 errout:
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004345 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004346 else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004347 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4348 ar->len = ac->ac_b_ex.fe_len;
4349 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004350 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004351 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004352 if (freed)
4353 goto repeat;
4354 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004355 }
4356
4357 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004358 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004359 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004360 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004361 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004362 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004363out:
4364 if (ac)
4365 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004366 if (inquota && ar->len < inquota)
Christoph Hellwig5dd40562010-03-03 09:05:00 -05004367 dquot_free_block(ar->inode, inquota - ar->len);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004368 if (!ar->len) {
Theodore Ts'of2321092011-01-10 12:12:36 -05004369 if (!ext4_test_inode_state(ar->inode,
4370 EXT4_STATE_DELALLOC_RESERVED))
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004371 /* release all the reserved blocks if non delalloc */
4372 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4373 reserv_blks);
4374 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004375
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004376 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004377
Alex Tomasc9de5602008-01-29 00:19:52 -05004378 return block;
4379}
Alex Tomasc9de5602008-01-29 00:19:52 -05004380
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004381/*
4382 * We can merge two free data extents only if the physical blocks
4383 * are contiguous, AND the extents were freed by the same transaction,
4384 * AND the blocks are associated with the same group.
4385 */
4386static int can_merge(struct ext4_free_data *entry1,
4387 struct ext4_free_data *entry2)
4388{
4389 if ((entry1->t_tid == entry2->t_tid) &&
4390 (entry1->group == entry2->group) &&
4391 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4392 return 1;
4393 return 0;
4394}
4395
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004396static noinline_for_stack int
4397ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004398 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004399{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004400 ext4_group_t group = e4b->bd_group;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004401 ext4_grpblk_t block;
4402 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004403 struct ext4_group_info *db = e4b->bd_info;
4404 struct super_block *sb = e4b->bd_sb;
4405 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004406 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4407 struct rb_node *parent = NULL, *new_node;
4408
Frank Mayhar03901312009-01-07 00:06:22 -05004409 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004410 BUG_ON(e4b->bd_bitmap_page == NULL);
4411 BUG_ON(e4b->bd_buddy_page == NULL);
4412
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004413 new_node = &new_entry->node;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004414 block = new_entry->start_blk;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004415
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004416 if (!*n) {
4417 /* first free block exent. We need to
4418 protect buddy cache from being freed,
4419 * otherwise we'll refresh it from
4420 * on-disk bitmap and lose not-yet-available
4421 * blocks */
4422 page_cache_get(e4b->bd_buddy_page);
4423 page_cache_get(e4b->bd_bitmap_page);
4424 }
4425 while (*n) {
4426 parent = *n;
4427 entry = rb_entry(parent, struct ext4_free_data, node);
4428 if (block < entry->start_blk)
4429 n = &(*n)->rb_left;
4430 else if (block >= (entry->start_blk + entry->count))
4431 n = &(*n)->rb_right;
4432 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004433 ext4_grp_locked_error(sb, group, 0,
4434 ext4_group_first_block_no(sb, group) + block,
4435 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004436 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004437 }
4438 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004439
4440 rb_link_node(new_node, parent, n);
4441 rb_insert_color(new_node, &db->bb_free_root);
4442
4443 /* Now try to see the extent can be merged to left and right */
4444 node = rb_prev(new_node);
4445 if (node) {
4446 entry = rb_entry(node, struct ext4_free_data, node);
4447 if (can_merge(entry, new_entry)) {
4448 new_entry->start_blk = entry->start_blk;
4449 new_entry->count += entry->count;
4450 rb_erase(node, &(db->bb_free_root));
4451 spin_lock(&sbi->s_md_lock);
4452 list_del(&entry->list);
4453 spin_unlock(&sbi->s_md_lock);
4454 kmem_cache_free(ext4_free_ext_cachep, entry);
4455 }
4456 }
4457
4458 node = rb_next(new_node);
4459 if (node) {
4460 entry = rb_entry(node, struct ext4_free_data, node);
4461 if (can_merge(new_entry, entry)) {
4462 new_entry->count += entry->count;
4463 rb_erase(node, &(db->bb_free_root));
4464 spin_lock(&sbi->s_md_lock);
4465 list_del(&entry->list);
4466 spin_unlock(&sbi->s_md_lock);
4467 kmem_cache_free(ext4_free_ext_cachep, entry);
4468 }
4469 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004470 /* Add the extent to transaction's private list */
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004471 spin_lock(&sbi->s_md_lock);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004472 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004473 spin_unlock(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004474 return 0;
4475}
4476
Theodore Ts'o44338712009-11-22 07:44:56 -05004477/**
4478 * ext4_free_blocks() -- Free given blocks and update quota
4479 * @handle: handle for this transaction
4480 * @inode: inode
4481 * @block: start physical block to free
4482 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004483 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004484 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004485void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004486 struct buffer_head *bh, ext4_fsblk_t block,
4487 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004488{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004489 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004490 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004491 struct ext4_group_desc *gdp;
Theodore Ts'o44338712009-11-22 07:44:56 -05004492 unsigned long freed = 0;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004493 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004494 ext4_grpblk_t bit;
4495 struct buffer_head *gd_bh;
4496 ext4_group_t block_group;
4497 struct ext4_sb_info *sbi;
4498 struct ext4_buddy e4b;
4499 int err = 0;
4500 int ret;
4501
Theodore Ts'oe6362602009-11-23 07:17:05 -05004502 if (bh) {
4503 if (block)
4504 BUG_ON(block != bh->b_blocknr);
4505 else
4506 block = bh->b_blocknr;
4507 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004508
Alex Tomasc9de5602008-01-29 00:19:52 -05004509 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004510 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4511 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004512 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004513 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004514 goto error_return;
4515 }
4516
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004517 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004518 trace_ext4_free_blocks(inode, block, count, flags);
4519
4520 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4521 struct buffer_head *tbh = bh;
4522 int i;
4523
4524 BUG_ON(bh && (count > 1));
4525
4526 for (i = 0; i < count; i++) {
4527 if (!bh)
4528 tbh = sb_find_get_block(inode->i_sb,
4529 block + i);
Namhyung Kim87783692010-10-27 21:30:11 -04004530 if (unlikely(!tbh))
4531 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004532 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004533 inode, tbh, block + i);
4534 }
4535 }
4536
Theodore Ts'o60e66792010-05-17 07:00:00 -04004537 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004538 * We need to make sure we don't reuse the freed block until
4539 * after the transaction is committed, which we can do by
4540 * treating the block as metadata, below. We make an
4541 * exception if the inode is to be written in writeback mode
4542 * since writeback mode has weak data consistency guarantees.
4543 */
4544 if (!ext4_should_writeback_data(inode))
4545 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004546
Alex Tomasc9de5602008-01-29 00:19:52 -05004547do_more:
4548 overflow = 0;
4549 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4550
4551 /*
4552 * Check to see if we are freeing blocks across a group
4553 * boundary.
4554 */
4555 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4556 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4557 count -= overflow;
4558 }
Theodore Ts'o574ca172008-07-11 19:27:31 -04004559 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004560 if (!bitmap_bh) {
4561 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004562 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004563 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004564 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004565 if (!gdp) {
4566 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004567 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004568 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004569
4570 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4571 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4572 in_range(block, ext4_inode_table(sb, gdp),
4573 EXT4_SB(sb)->s_itb_per_group) ||
4574 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4575 EXT4_SB(sb)->s_itb_per_group)) {
4576
Eric Sandeen12062dd2010-02-15 14:19:27 -05004577 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004578 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004579 /* err = 0. ext4_std_error should be a no op */
4580 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004581 }
4582
4583 BUFFER_TRACE(bitmap_bh, "getting write access");
4584 err = ext4_journal_get_write_access(handle, bitmap_bh);
4585 if (err)
4586 goto error_return;
4587
4588 /*
4589 * We are about to modify some metadata. Call the journal APIs
4590 * to unshare ->b_data if a currently-committing transaction is
4591 * using it
4592 */
4593 BUFFER_TRACE(gd_bh, "get_write_access");
4594 err = ext4_journal_get_write_access(handle, gd_bh);
4595 if (err)
4596 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004597#ifdef AGGRESSIVE_CHECK
4598 {
4599 int i;
4600 for (i = 0; i < count; i++)
4601 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4602 }
4603#endif
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004604 trace_ext4_mballoc_free(sb, inode, block_group, bit, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004605
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004606 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4607 if (err)
4608 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004609
4610 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004611 struct ext4_free_data *new_entry;
4612 /*
4613 * blocks being freed are metadata. these blocks shouldn't
4614 * be used until this transaction is committed
4615 */
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004616 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4617 if (!new_entry) {
4618 err = -ENOMEM;
4619 goto error_return;
4620 }
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004621 new_entry->start_blk = bit;
4622 new_entry->group = block_group;
4623 new_entry->count = count;
4624 new_entry->t_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004625
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004626 ext4_lock_group(sb, block_group);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004627 mb_clear_bits(bitmap_bh->b_data, bit, count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004628 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004629 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004630 /* need to update group_info->bb_free and bitmap
4631 * with group lock held. generate_buddy look at
4632 * them with group lock_held
4633 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004634 ext4_lock_group(sb, block_group);
4635 mb_clear_bits(bitmap_bh->b_data, bit, count);
Shen Feng7e5a8cd2008-07-13 21:03:31 -04004636 mb_free_blocks(inode, &e4b, bit, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004637 }
4638
Aneesh Kumar K.V560671a2009-01-05 22:20:24 -05004639 ret = ext4_free_blks_count(sb, gdp) + count;
4640 ext4_free_blks_set(sb, gdp, ret);
Alex Tomasc9de5602008-01-29 00:19:52 -05004641 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004642 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004643 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4644
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004645 if (sbi->s_log_groups_per_flex) {
4646 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o9f24e422009-03-04 19:09:10 -05004647 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004648 }
4649
Jing Zhange39e07f2010-05-14 00:00:00 -04004650 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05004651
Theodore Ts'o44338712009-11-22 07:44:56 -05004652 freed += count;
Alex Tomasc9de5602008-01-29 00:19:52 -05004653
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004654 /* We dirtied the bitmap block */
4655 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4656 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4657
Alex Tomasc9de5602008-01-29 00:19:52 -05004658 /* And the group descriptor block */
4659 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004660 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004661 if (!err)
4662 err = ret;
4663
4664 if (overflow && !err) {
4665 block += count;
4666 count = overflow;
4667 put_bh(bitmap_bh);
4668 goto do_more;
4669 }
Theodore Ts'oa0375152010-06-11 23:14:04 -04004670 ext4_mark_super_dirty(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004671error_return:
Maxim Patlasov7132de72011-07-10 19:37:48 -04004672 if (freed && !(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Christoph Hellwig5dd40562010-03-03 09:05:00 -05004673 dquot_free_block(inode, freed);
Alex Tomasc9de5602008-01-29 00:19:52 -05004674 brelse(bitmap_bh);
4675 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004676 return;
4677}
Lukas Czerner7360d172010-10-27 21:30:12 -04004678
4679/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004680 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004681 * @handle: handle to this transaction
4682 * @sb: super block
4683 * @block: start physcial block to add to the block group
4684 * @count: number of blocks to free
4685 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004686 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004687 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004688int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004689 ext4_fsblk_t block, unsigned long count)
4690{
4691 struct buffer_head *bitmap_bh = NULL;
4692 struct buffer_head *gd_bh;
4693 ext4_group_t block_group;
4694 ext4_grpblk_t bit;
4695 unsigned int i;
4696 struct ext4_group_desc *desc;
4697 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004698 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004699 int err = 0, ret, blk_free_count;
4700 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004701
4702 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4703
Yongqiang Yang4740b832011-07-26 21:51:08 -04004704 if (count == 0)
4705 return 0;
4706
Amir Goldstein2846e822011-05-09 10:46:41 -04004707 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004708 /*
4709 * Check to see if we are freeing blocks across a group
4710 * boundary.
4711 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004712 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4713 ext4_warning(sb, "too much blocks added to group %u\n",
4714 block_group);
4715 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004716 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004717 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004718
Amir Goldstein2846e822011-05-09 10:46:41 -04004719 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004720 if (!bitmap_bh) {
4721 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004722 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004723 }
4724
Amir Goldstein2846e822011-05-09 10:46:41 -04004725 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004726 if (!desc) {
4727 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004728 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004729 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004730
4731 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4732 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4733 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4734 in_range(block + count - 1, ext4_inode_table(sb, desc),
4735 sbi->s_itb_per_group)) {
4736 ext4_error(sb, "Adding blocks in system zones - "
4737 "Block = %llu, count = %lu",
4738 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004739 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004740 goto error_return;
4741 }
4742
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004743 BUFFER_TRACE(bitmap_bh, "getting write access");
4744 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004745 if (err)
4746 goto error_return;
4747
4748 /*
4749 * We are about to modify some metadata. Call the journal APIs
4750 * to unshare ->b_data if a currently-committing transaction is
4751 * using it
4752 */
4753 BUFFER_TRACE(gd_bh, "get_write_access");
4754 err = ext4_journal_get_write_access(handle, gd_bh);
4755 if (err)
4756 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004757
Amir Goldstein2846e822011-05-09 10:46:41 -04004758 for (i = 0, blocks_freed = 0; i < count; i++) {
4759 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004760 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004761 ext4_error(sb, "bit already cleared for block %llu",
4762 (ext4_fsblk_t)(block + i));
4763 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4764 } else {
4765 blocks_freed++;
4766 }
4767 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004768
4769 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4770 if (err)
4771 goto error_return;
4772
4773 /*
4774 * need to update group_info->bb_free and bitmap
4775 * with group lock held. generate_buddy look at
4776 * them with group lock_held
4777 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004778 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004779 mb_clear_bits(bitmap_bh->b_data, bit, count);
4780 mb_free_blocks(NULL, &e4b, bit, count);
Amir Goldstein2846e822011-05-09 10:46:41 -04004781 blk_free_count = blocks_freed + ext4_free_blks_count(sb, desc);
4782 ext4_free_blks_set(sb, desc, blk_free_count);
4783 desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc);
4784 ext4_unlock_group(sb, block_group);
4785 percpu_counter_add(&sbi->s_freeblocks_counter, blocks_freed);
4786
4787 if (sbi->s_log_groups_per_flex) {
4788 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4789 atomic_add(blocks_freed,
4790 &sbi->s_flex_groups[flex_group].free_blocks);
4791 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004792
4793 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04004794
4795 /* We dirtied the bitmap block */
4796 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4797 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4798
4799 /* And the group descriptor block */
4800 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4801 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4802 if (!err)
4803 err = ret;
4804
4805error_return:
4806 brelse(bitmap_bh);
4807 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004808 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04004809}
4810
4811/**
Lukas Czerner7360d172010-10-27 21:30:12 -04004812 * ext4_trim_extent -- function to TRIM one single free extent in the group
4813 * @sb: super block for the file system
4814 * @start: starting block of the free extent in the alloc. group
4815 * @count: number of blocks to TRIM
4816 * @group: alloc. group we are working with
4817 * @e4b: ext4 buddy for the group
4818 *
4819 * Trim "count" blocks starting at "start" in the "group". To assure that no
4820 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4821 * be called with under the group lock.
4822 */
Theodore Ts'od9f34502011-04-30 13:47:24 -04004823static void ext4_trim_extent(struct super_block *sb, int start, int count,
4824 ext4_group_t group, struct ext4_buddy *e4b)
Lukas Czerner7360d172010-10-27 21:30:12 -04004825{
4826 struct ext4_free_extent ex;
Lukas Czerner7360d172010-10-27 21:30:12 -04004827
Tao Mab3d4c2b2011-07-11 00:01:52 -04004828 trace_ext4_trim_extent(sb, group, start, count);
4829
Lukas Czerner7360d172010-10-27 21:30:12 -04004830 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4831
4832 ex.fe_start = start;
4833 ex.fe_group = group;
4834 ex.fe_len = count;
4835
4836 /*
4837 * Mark blocks used, so no one can reuse them while
4838 * being trimmed.
4839 */
4840 mb_mark_used(e4b, &ex);
4841 ext4_unlock_group(sb, group);
Theodore Ts'od9f34502011-04-30 13:47:24 -04004842 ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04004843 ext4_lock_group(sb, group);
4844 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czerner7360d172010-10-27 21:30:12 -04004845}
4846
4847/**
4848 * ext4_trim_all_free -- function to trim all free space in alloc. group
4849 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04004850 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04004851 * @start: first group block to examine
4852 * @max: last group block to examine
4853 * @minblocks: minimum extent block count
4854 *
4855 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4856 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4857 * the extent.
4858 *
4859 *
4860 * ext4_trim_all_free walks through group's block bitmap searching for free
4861 * extents. When the free extent is found, mark it as used in group buddy
4862 * bitmap. Then issue a TRIM command on this extent and free the extent in
4863 * the group buddy bitmap. This is done until whole group is scanned.
4864 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05004865static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04004866ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4867 ext4_grpblk_t start, ext4_grpblk_t max,
4868 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04004869{
4870 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04004871 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04004872 struct ext4_buddy e4b;
4873 int ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04004874
Tao Mab3d4c2b2011-07-11 00:01:52 -04004875 trace_ext4_trim_all_free(sb, group, start, max);
4876
Lukas Czerner78944082011-05-24 18:16:27 -04004877 ret = ext4_mb_load_buddy(sb, group, &e4b);
4878 if (ret) {
4879 ext4_error(sb, "Error in loading buddy "
4880 "information for %u", group);
4881 return ret;
4882 }
Lukas Czerner78944082011-05-24 18:16:27 -04004883 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04004884
4885 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04004886 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4887 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4888 goto out;
4889
Lukas Czerner78944082011-05-24 18:16:27 -04004890 start = (e4b.bd_info->bb_first_free > start) ?
4891 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04004892
4893 while (start < max) {
4894 start = mb_find_next_zero_bit(bitmap, max, start);
4895 if (start >= max)
4896 break;
4897 next = mb_find_next_bit(bitmap, max, start);
4898
4899 if ((next - start) >= minblocks) {
Theodore Ts'od9f34502011-04-30 13:47:24 -04004900 ext4_trim_extent(sb, start,
Lukas Czerner78944082011-05-24 18:16:27 -04004901 next - start, group, &e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04004902 count += next - start;
4903 }
Tao Ma169ddc32011-07-11 00:00:07 -04004904 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04004905 start = next + 1;
4906
4907 if (fatal_signal_pending(current)) {
4908 count = -ERESTARTSYS;
4909 break;
4910 }
4911
4912 if (need_resched()) {
4913 ext4_unlock_group(sb, group);
4914 cond_resched();
4915 ext4_lock_group(sb, group);
4916 }
4917
Tao Ma169ddc32011-07-11 00:00:07 -04004918 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04004919 break;
4920 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04004921
4922 if (!ret)
4923 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4924out:
Lukas Czerner7360d172010-10-27 21:30:12 -04004925 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04004926 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04004927
4928 ext4_debug("trimmed %d blocks in the group %d\n",
4929 count, group);
4930
Lukas Czerner7360d172010-10-27 21:30:12 -04004931 return count;
4932}
4933
4934/**
4935 * ext4_trim_fs() -- trim ioctl handle function
4936 * @sb: superblock for filesystem
4937 * @range: fstrim_range structure
4938 *
4939 * start: First Byte to trim
4940 * len: number of Bytes to trim from start
4941 * minlen: minimum extent length in Bytes
4942 * ext4_trim_fs goes through all allocation groups containing Bytes from
4943 * start to start+len. For each such a group ext4_trim_all_free function
4944 * is invoked to trim all free space.
4945 */
4946int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4947{
Lukas Czerner78944082011-05-24 18:16:27 -04004948 struct ext4_group_info *grp;
Lukas Czerner7360d172010-10-27 21:30:12 -04004949 ext4_group_t first_group, last_group;
4950 ext4_group_t group, ngroups = ext4_get_groups_count(sb);
4951 ext4_grpblk_t cnt = 0, first_block, last_block;
Lukas Czerner78944082011-05-24 18:16:27 -04004952 uint64_t start, len, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05004953 ext4_fsblk_t first_data_blk =
4954 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner7360d172010-10-27 21:30:12 -04004955 int ret = 0;
4956
4957 start = range->start >> sb->s_blocksize_bits;
4958 len = range->len >> sb->s_blocksize_bits;
4959 minlen = range->minlen >> sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04004960
4961 if (unlikely(minlen > EXT4_BLOCKS_PER_GROUP(sb)))
4962 return -EINVAL;
Tao Ma22f10452011-07-10 23:52:37 -04004963 if (start + len <= first_data_blk)
4964 goto out;
Jan Kara0f0a25b2011-01-11 15:16:31 -05004965 if (start < first_data_blk) {
4966 len -= first_data_blk - start;
4967 start = first_data_blk;
4968 }
Lukas Czerner7360d172010-10-27 21:30:12 -04004969
4970 /* Determine first and last group to examine based on start and len */
4971 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
4972 &first_group, &first_block);
4973 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len),
4974 &last_group, &last_block);
4975 last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group;
4976 last_block = EXT4_BLOCKS_PER_GROUP(sb);
4977
4978 if (first_group > last_group)
4979 return -EINVAL;
4980
4981 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04004982 grp = ext4_get_group_info(sb, group);
4983 /* We only do this if the grp has never been initialized */
4984 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
4985 ret = ext4_mb_init_group(sb, group);
4986 if (ret)
4987 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04004988 }
4989
Tao Ma0ba08512011-03-23 15:48:11 -04004990 /*
4991 * For all the groups except the last one, last block will
4992 * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to
4993 * change it for the last group in which case start +
4994 * len < EXT4_BLOCKS_PER_GROUP(sb).
4995 */
4996 if (first_block + len < EXT4_BLOCKS_PER_GROUP(sb))
Jan Karaca6e9092011-01-10 12:30:39 -05004997 last_block = first_block + len;
Tao Ma0ba08512011-03-23 15:48:11 -04004998 len -= last_block - first_block;
Lukas Czerner7360d172010-10-27 21:30:12 -04004999
Lukas Czerner78944082011-05-24 18:16:27 -04005000 if (grp->bb_free >= minlen) {
5001 cnt = ext4_trim_all_free(sb, group, first_block,
Lukas Czerner7360d172010-10-27 21:30:12 -04005002 last_block, minlen);
5003 if (cnt < 0) {
5004 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005005 break;
5006 }
5007 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005008 trimmed += cnt;
5009 first_block = 0;
5010 }
5011 range->len = trimmed * sb->s_blocksize;
5012
Tao Ma3d56b8d2011-07-11 00:03:38 -04005013 if (!ret)
5014 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5015
Tao Ma22f10452011-07-10 23:52:37 -04005016out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005017 return ret;
5018}