blob: 04a5c7504be9d6dbe55174d451f559064a7a05cf [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
Bobi Jam18aadd42012-02-20 17:53:02 -050024#include "ext4_jbd2.h"
Mingming Cao8f6e39a2008-04-29 22:01:31 -040025#include "mballoc.h"
Theodore Ts'o28623c22012-09-05 01:31:50 -040026#include <linux/log2.h>
Theodore Ts'oa0b30c12013-02-09 16:28:20 -050027#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090028#include <linux/slab.h>
Theodore Ts'o9bffad12009-06-17 11:48:11 -040029#include <trace/events/ext4.h>
30
Theodore Ts'oa0b30c12013-02-09 16:28:20 -050031#ifdef CONFIG_EXT4_DEBUG
32ushort ext4_mballoc_debug __read_mostly;
33
34module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
36#endif
37
Alex Tomasc9de5602008-01-29 00:19:52 -050038/*
39 * MUSTDO:
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
42 *
43 * TODO v4:
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
47 * - quota
48 * - reservation for superuser
49 *
50 * TODO v3:
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
55 * - error handling
56 */
57
58/*
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
61 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040062 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
Alex Tomasc9de5602008-01-29 00:19:52 -050071 *
72 * The main motivation for having small file use group preallocation is to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040073 * ensure that we have small files closer together on the disk.
Alex Tomasc9de5602008-01-29 00:19:52 -050074 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040075 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
78 * represented as:
Alex Tomasc9de5602008-01-29 00:19:52 -050079 *
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
Theodore Ts'o53accfa2011-09-09 18:48:51 -040082 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
Alex Tomasc9de5602008-01-29 00:19:52 -050084 *
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
Tao Macaaf7a22011-07-11 18:42:42 -040087 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
Alex Tomasc9de5602008-01-29 00:19:52 -050089 *
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
92 * pa_free.
93 *
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
Tao Macaaf7a22011-07-11 18:42:42 -040096 * prealloc space. These are per CPU prealloc list represented as
Alex Tomasc9de5602008-01-29 00:19:52 -050097 *
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
99 *
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
102 *
103 * The locality group prealloc space is used looking at whether we have
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300104 * enough free space (pa_free) within the prealloc space.
Alex Tomasc9de5602008-01-29 00:19:52 -0500105 *
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
114 * inode as:
115 *
116 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500118 *
119 *
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
124 *
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
127 *
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
131 *
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
Theodore Ts'o27baebb2011-09-09 19:02:51 -0400138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
Alex Tomasc9de5602008-01-29 00:19:52 -0500140 * 512 blocks. This can be tuned via
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
Alex Tomasc9de5602008-01-29 00:19:52 -0500142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
Alex Tomasc9de5602008-01-29 00:19:52 -0500146 *
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400147 * The regular allocator (using the buddy cache) supports a few tunables.
Alex Tomasc9de5602008-01-29 00:19:52 -0500148 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -0500152 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400153 * The regular allocator uses buddy scan only if the request len is power of
Alex Tomasc9de5602008-01-29 00:19:52 -0500154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
Alex Tomasc9de5602008-01-29 00:19:52 -0500161 * min_to_scan indicate how long the mballoc __must__ look for a best
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
Alex Tomasc9de5602008-01-29 00:19:52 -0500163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
Tao Macaaf7a22011-07-11 18:42:42 -0400166 * can be used for allocation. ext4_mb_good_group explains how the groups are
Alex Tomasc9de5602008-01-29 00:19:52 -0500167 * checked.
168 *
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
173 */
174
175/*
176 * mballoc operates on the following data:
177 * - on-disk bitmap
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
180 *
181 * there are two types of preallocations:
182 * - inode
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
191 * - locality group
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
196 *
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
199 *
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
203 *
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
207 *
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
210 *
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
221 *
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
232 * block
233 *
234 * so, now we're building a concurrency table:
235 * - init buddy vs.
236 * - new PA
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
239 * - use inode PA
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
242 * - discard inode PA
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
248 * - new PA vs.
249 * - use inode PA
250 * i_data_sem serializes them
251 * - discard inode PA
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
257 * - use inode PA
258 * - use inode PA
259 * i_data_sem or another mutex should serializes them
260 * - discard inode PA
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
266 *
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
274 *
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
277 *
278 */
279
280 /*
281 * Logic in few words:
282 *
283 * - allocation:
284 * load group
285 * find blocks
286 * mark bits in on-disk bitmap
287 * release group
288 *
289 * - use preallocation:
290 * find proper PA (per-inode or group)
291 * load group
292 * mark bits in on-disk bitmap
293 * release group
294 * release PA
295 *
296 * - free:
297 * load group
298 * mark bits in on-disk bitmap
299 * release group
300 *
301 * - discard preallocations in group:
302 * mark PAs deleted
303 * move them onto local list
304 * load on-disk bitmap
305 * load group
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
308 *
309 * - discard inode's preallocations:
310 */
311
312/*
313 * Locking rules
314 *
315 * Locks:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
318 * - per-pa lock (pa)
319 *
320 * Paths:
321 * - new pa
322 * object
323 * group
324 *
325 * - find and use pa:
326 * pa
327 *
328 * - release consumed pa:
329 * pa
330 * group
331 * object
332 *
333 * - generate in-core bitmap:
334 * group
335 * pa
336 *
337 * - discard all for given object (inode, locality group):
338 * object
339 * pa
340 * group
341 *
342 * - discard all for given group:
343 * group
344 * pa
345 * group
346 * object
347 *
348 */
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500349static struct kmem_cache *ext4_pspace_cachep;
350static struct kmem_cache *ext4_ac_cachep;
Bobi Jam18aadd42012-02-20 17:53:02 -0500351static struct kmem_cache *ext4_free_data_cachep;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400352
353/* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
Eric Sandeen2892c152011-02-12 08:12:18 -0500356#define NR_GRPINFO_CACHES 8
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400357static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
358
Eric Sandeen2892c152011-02-12 08:12:18 -0500359static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
363};
364
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500365static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
366 ext4_group_t group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500367static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
368 ext4_group_t group);
Bobi Jam18aadd42012-02-20 17:53:02 -0500369static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500371
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500372static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
373{
Alex Tomasc9de5602008-01-29 00:19:52 -0500374#if BITS_PER_LONG == 64
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500377#elif BITS_PER_LONG == 32
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500380#else
381#error "how many bits you are?!"
382#endif
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500383 return addr;
384}
Alex Tomasc9de5602008-01-29 00:19:52 -0500385
386static inline int mb_test_bit(int bit, void *addr)
387{
388 /*
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
391 */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500392 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500393 return ext4_test_bit(bit, addr);
394}
395
396static inline void mb_set_bit(int bit, void *addr)
397{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500398 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500399 ext4_set_bit(bit, addr);
400}
401
Alex Tomasc9de5602008-01-29 00:19:52 -0500402static inline void mb_clear_bit(int bit, void *addr)
403{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500404 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500405 ext4_clear_bit(bit, addr);
406}
407
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400408static inline int mb_test_and_clear_bit(int bit, void *addr)
409{
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
412}
413
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500414static inline int mb_find_next_zero_bit(void *addr, int max, int start)
415{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400416 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500417 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400418 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500419 start += fix;
420
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
422 if (ret > max)
423 return max;
424 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500425}
426
427static inline int mb_find_next_bit(void *addr, int max, int start)
428{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400429 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500430 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400431 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500432 start += fix;
433
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
435 if (ret > max)
436 return max;
437 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500438}
439
Alex Tomasc9de5602008-01-29 00:19:52 -0500440static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
441{
442 char *bb;
443
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -0500445 BUG_ON(max == NULL);
446
447 if (order > e4b->bd_blkbits + 1) {
448 *max = 0;
449 return NULL;
450 }
451
452 /* at order 0 we see each particular block */
Coly Li84b775a2011-02-24 12:51:59 -0500453 if (order == 0) {
454 *max = 1 << (e4b->bd_blkbits + 3);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500455 return e4b->bd_bitmap;
Coly Li84b775a2011-02-24 12:51:59 -0500456 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500457
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
Alex Tomasc9de5602008-01-29 00:19:52 -0500459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
460
461 return bb;
462}
463
464#ifdef DOUBLE_CHECK
465static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
467{
468 int i;
469 struct super_block *sb = e4b->bd_sb;
470
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
472 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
Akinobu Mita5661bd62010-03-03 23:53:39 -0500477
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -0500480 ext4_grp_locked_error(sb, e4b->bd_group,
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400481 inode ? inode->i_ino : 0,
482 blocknr,
483 "freeing block already freed "
484 "(bit %u)",
485 first + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500486 }
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
488 }
489}
490
491static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
492{
493 int i;
494
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
496 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
501 }
502}
503
504static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
505{
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
508 int i;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -0400513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
516 "on disk/prealloc",
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500518 BUG();
519 }
520 }
521 }
522}
523
524#else
525static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
527{
528 return;
529}
530static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
532{
533 return;
534}
535static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
536{
537 return;
538}
539#endif
540
541#ifdef AGGRESSIVE_CHECK
542
543#define MB_CHECK_ASSERT(assert) \
544do { \
545 if (!(assert)) { \
546 printk(KERN_EMERG \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
549 BUG(); \
550 } \
551} while (0)
552
553static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
555{
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
558 int max;
559 int max2;
560 int i;
561 int j;
562 int k;
563 int count;
564 struct ext4_group_info *grp;
565 int fragments = 0;
566 int fstart;
567 struct list_head *cur;
568 void *buddy;
569 void *buddy2;
570
Alex Tomasc9de5602008-01-29 00:19:52 -0500571 {
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
574 return 0;
575 }
576
577 while (order > 1) {
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
584
585 count = 0;
586 for (i = 0; i < max; i++) {
587
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
591 MB_CHECK_ASSERT(
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
594 MB_CHECK_ASSERT(
595 mb_test_bit(i << 1, buddy2));
596 }
597 continue;
598 }
599
Robin Dong0a10da72011-10-26 08:48:54 -0400600 /* both bits in buddy2 must be 1 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
603
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
606 MB_CHECK_ASSERT(
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500607 !mb_test_bit(k, e4b->bd_bitmap));
Alex Tomasc9de5602008-01-29 00:19:52 -0500608 }
609 count++;
610 }
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
612 order--;
613 }
614
615 fstart = -1;
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
620 if (fstart == -1) {
621 fragments++;
622 fstart = i;
623 }
624 continue;
625 }
626 fstart = -1;
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
630 k = i >> j;
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
633 }
634 }
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
637
638 grp = ext4_get_group_info(sb, e4b->bd_group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
Alex Tomasc9de5602008-01-29 00:19:52 -0500644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400645 for (i = 0; i < pa->pa_len; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
647 }
648 return 0;
649}
650#undef MB_CHECK_ASSERT
651#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
Harvey Harrison46e665e2008-04-17 10:38:59 -0400652 __FILE__, __func__, __LINE__)
Alex Tomasc9de5602008-01-29 00:19:52 -0500653#else
654#define mb_check_buddy(e4b)
655#endif
656
Coly Li7c786052011-02-24 13:24:25 -0500657/*
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
662 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500663static void ext4_mb_mark_free_simple(struct super_block *sb,
Eric Sandeena36b4492009-08-25 22:36:45 -0400664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
Alex Tomasc9de5602008-01-29 00:19:52 -0500665 struct ext4_group_info *grp)
666{
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400668 ext4_grpblk_t min;
669 ext4_grpblk_t max;
670 ext4_grpblk_t chunk;
Alex Tomasc9de5602008-01-29 00:19:52 -0500671 unsigned short border;
672
Theodore Ts'o7137d7a2011-09-09 18:38:51 -0400673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
Alex Tomasc9de5602008-01-29 00:19:52 -0500674
675 border = 2 << sb->s_blocksize_bits;
676
677 while (len > 0) {
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
680
681 /* find how many blocks of power 2 we need to mark */
682 min = fls(len) - 1;
683
684 if (max < min)
685 min = max;
686 chunk = 1 << min;
687
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
690 if (min > 0)
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
693
694 len -= chunk;
695 first += chunk;
696 }
697}
698
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400699/*
700 * Cache the order of the largest free extent we have available in this block
701 * group.
702 */
703static void
704mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
705{
706 int i;
707 int bits;
708
709 grp->bb_largest_free_order = -1; /* uninit */
710
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
715 break;
716 }
717 }
718}
719
Eric Sandeen089ceec2009-07-05 22:17:31 -0400720static noinline_for_stack
721void ext4_mb_generate_buddy(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -0500722 void *buddy, void *bitmap, ext4_group_t group)
723{
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -0400725 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400726 ext4_grpblk_t i = 0;
727 ext4_grpblk_t first;
728 ext4_grpblk_t len;
Alex Tomasc9de5602008-01-29 00:19:52 -0500729 unsigned free = 0;
730 unsigned fragments = 0;
731 unsigned long long period = get_cycles();
732
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500735 i = mb_find_next_zero_bit(bitmap, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -0500736 grp->bb_first_free = i;
737 while (i < max) {
738 fragments++;
739 first = i;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500740 i = mb_find_next_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500741 len = i - first;
742 free += len;
743 if (len > 1)
744 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
745 else
746 grp->bb_counters[0]++;
747 if (i < max)
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500748 i = mb_find_next_zero_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500749 }
750 grp->bb_fragments = fragments;
751
752 if (free != grp->bb_free) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400753 ext4_grp_locked_error(sb, group, 0, 0,
Darrick J. Wong163a2032013-08-28 17:35:51 -0400754 "%u clusters in bitmap, %u in gd; "
755 "block bitmap corrupt.",
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400756 free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500757 /*
Darrick J. Wong163a2032013-08-28 17:35:51 -0400758 * If we intend to continue, we consider group descriptor
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500759 * corrupt and update bb_free using bitmap value
760 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500761 grp->bb_free = free;
Darrick J. Wong163a2032013-08-28 17:35:51 -0400762 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
Alex Tomasc9de5602008-01-29 00:19:52 -0500763 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400764 mb_set_largest_free_order(sb, grp);
Alex Tomasc9de5602008-01-29 00:19:52 -0500765
766 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
767
768 period = get_cycles() - period;
769 spin_lock(&EXT4_SB(sb)->s_bal_lock);
770 EXT4_SB(sb)->s_mb_buddies_generated++;
771 EXT4_SB(sb)->s_mb_generation_time += period;
772 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
773}
774
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400775static void mb_regenerate_buddy(struct ext4_buddy *e4b)
776{
777 int count;
778 int order = 1;
779 void *buddy;
780
781 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
782 ext4_set_bits(buddy, 0, count);
783 }
784 e4b->bd_info->bb_fragments = 0;
785 memset(e4b->bd_info->bb_counters, 0,
786 sizeof(*e4b->bd_info->bb_counters) *
787 (e4b->bd_sb->s_blocksize_bits + 2));
788
789 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
790 e4b->bd_bitmap, e4b->bd_group);
791}
792
Alex Tomasc9de5602008-01-29 00:19:52 -0500793/* The buddy information is attached the buddy cache inode
794 * for convenience. The information regarding each group
795 * is loaded via ext4_mb_load_buddy. The information involve
796 * block bitmap and buddy information. The information are
797 * stored in the inode as
798 *
799 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500800 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500801 *
802 *
803 * one block each for bitmap and buddy information.
804 * So for each group we take up 2 blocks. A page can
805 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
806 * So it can have information regarding groups_per_page which
807 * is blocks_per_page/2
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400808 *
809 * Locking note: This routine takes the block group lock of all groups
810 * for this page; do not hold this lock when calling this routine!
Alex Tomasc9de5602008-01-29 00:19:52 -0500811 */
812
813static int ext4_mb_init_cache(struct page *page, char *incore)
814{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400815 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500816 int blocksize;
817 int blocks_per_page;
818 int groups_per_page;
819 int err = 0;
820 int i;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500821 ext4_group_t first_group, group;
Alex Tomasc9de5602008-01-29 00:19:52 -0500822 int first_block;
823 struct super_block *sb;
824 struct buffer_head *bhs;
Darrick J. Wongfa77dcf2012-04-29 18:35:10 -0400825 struct buffer_head **bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -0500826 struct inode *inode;
827 char *data;
828 char *bitmap;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400829 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -0500830
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400831 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500832
833 inode = page->mapping->host;
834 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400835 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500836 blocksize = 1 << inode->i_blkbits;
837 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
838
839 groups_per_page = blocks_per_page >> 1;
840 if (groups_per_page == 0)
841 groups_per_page = 1;
842
843 /* allocate buffer_heads to read bitmaps */
844 if (groups_per_page > 1) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500845 i = sizeof(struct buffer_head *) * groups_per_page;
846 bh = kzalloc(i, GFP_NOFS);
Theodore Ts'o813e5722012-02-20 17:52:46 -0500847 if (bh == NULL) {
848 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500849 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500850 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500851 } else
852 bh = &bhs;
853
854 first_group = page->index * blocks_per_page / 2;
855
856 /* read all groups the page covers into the cache */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500857 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
858 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500859 break;
860
Theodore Ts'o813e5722012-02-20 17:52:46 -0500861 grinfo = ext4_get_group_info(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400862 /*
863 * If page is uptodate then we came here after online resize
864 * which added some new uninitialized group info structs, so
865 * we must skip all initialized uptodate buddies on the page,
866 * which may be currently in use by an allocating task.
867 */
868 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
869 bh[i] = NULL;
870 continue;
871 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500872 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
873 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500874 goto out;
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500875 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500876 mb_debug(1, "read bitmap for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500877 }
878
879 /* wait for I/O completion */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500880 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
881 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
882 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -0500883 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500884 }
885 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500886
887 first_block = page->index * blocks_per_page;
888 for (i = 0; i < blocks_per_page; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500889 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400890 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500891 break;
892
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400893 if (!bh[group - first_group])
894 /* skip initialized uptodate buddy */
895 continue;
896
Alex Tomasc9de5602008-01-29 00:19:52 -0500897 /*
898 * data carry information regarding this
899 * particular group in the format specified
900 * above
901 *
902 */
903 data = page_address(page) + (i * blocksize);
904 bitmap = bh[group - first_group]->b_data;
905
906 /*
907 * We place the buddy block and bitmap block
908 * close together
909 */
910 if ((first_block + i) & 1) {
911 /* this is block of buddy */
912 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400913 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500914 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400915 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500916 grinfo = ext4_get_group_info(sb, group);
917 grinfo->bb_fragments = 0;
918 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400919 sizeof(*grinfo->bb_counters) *
920 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500921 /*
922 * incore got set to the group block bitmap below
923 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500924 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400925 /* init the buddy */
926 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500927 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500928 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500929 incore = NULL;
930 } else {
931 /* this is block of bitmap */
932 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400933 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500934 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400935 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500936
937 /* see comments in ext4_mb_put_pa() */
938 ext4_lock_group(sb, group);
939 memcpy(data, bitmap, blocksize);
940
941 /* mark all preallocated blks used in in-core bitmap */
942 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500943 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500944 ext4_unlock_group(sb, group);
945
946 /* set incore so that the buddy information can be
947 * generated using this
948 */
949 incore = data;
950 }
951 }
952 SetPageUptodate(page);
953
954out:
955 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400956 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500957 brelse(bh[i]);
958 if (bh != &bhs)
959 kfree(bh);
960 }
961 return err;
962}
963
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400964/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400965 * Lock the buddy and bitmap pages. This make sure other parallel init_group
966 * on the same buddy page doesn't happen whild holding the buddy page lock.
967 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
968 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400969 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400970static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
971 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400972{
Amir Goldstein2de88072011-05-09 21:48:13 -0400973 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
974 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400975 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400976 struct page *page;
977
978 e4b->bd_buddy_page = NULL;
979 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400980
981 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
982 /*
983 * the buddy cache inode stores the block bitmap
984 * and buddy information in consecutive blocks.
985 * So for each group we need two blocks.
986 */
987 block = group * 2;
988 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400989 poff = block % blocks_per_page;
990 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
991 if (!page)
992 return -EIO;
993 BUG_ON(page->mapping != inode->i_mapping);
994 e4b->bd_bitmap_page = page;
995 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400996
Amir Goldstein2de88072011-05-09 21:48:13 -0400997 if (blocks_per_page >= 2) {
998 /* buddy and bitmap are on the same page */
999 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001000 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001001
1002 block++;
1003 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001004 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_buddy_page = page;
1009 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001010}
1011
Amir Goldstein2de88072011-05-09 21:48:13 -04001012static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001013{
Amir Goldstein2de88072011-05-09 21:48:13 -04001014 if (e4b->bd_bitmap_page) {
1015 unlock_page(e4b->bd_bitmap_page);
1016 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001017 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001018 if (e4b->bd_buddy_page) {
1019 unlock_page(e4b->bd_buddy_page);
1020 page_cache_release(e4b->bd_buddy_page);
1021 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001022}
1023
1024/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001025 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1026 * block group lock of all groups for this page; do not hold the BG lock when
1027 * calling this routine!
1028 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001029static noinline_for_stack
1030int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1031{
1032
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001033 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001034 struct ext4_buddy e4b;
1035 struct page *page;
1036 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001037
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001038 might_sleep();
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001039 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001040 this_grp = ext4_get_group_info(sb, group);
1041 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001042 * This ensures that we don't reinit the buddy cache
1043 * page which map to the group from which we are already
1044 * allocating. If we are looking at the buddy cache we would
1045 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001046 * would have pinned buddy page to page cache.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001047 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001048 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1049 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001050 /*
1051 * somebody initialized the group
1052 * return without doing anything
1053 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001054 goto err;
1055 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001056
1057 page = e4b.bd_bitmap_page;
1058 ret = ext4_mb_init_cache(page, NULL);
1059 if (ret)
1060 goto err;
1061 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001062 ret = -EIO;
1063 goto err;
1064 }
1065 mark_page_accessed(page);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001066
Amir Goldstein2de88072011-05-09 21:48:13 -04001067 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001068 /*
1069 * If both the bitmap and buddy are in
1070 * the same page we don't need to force
1071 * init the buddy
1072 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001073 ret = 0;
1074 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001075 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001076 /* init buddy cache */
1077 page = e4b.bd_buddy_page;
1078 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1079 if (ret)
1080 goto err;
1081 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001082 ret = -EIO;
1083 goto err;
1084 }
1085 mark_page_accessed(page);
1086err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001087 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001088 return ret;
1089}
1090
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001091/*
1092 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1093 * block group lock of all groups for this page; do not hold the BG lock when
1094 * calling this routine!
1095 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001096static noinline_for_stack int
1097ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1098 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001099{
Alex Tomasc9de5602008-01-29 00:19:52 -05001100 int blocks_per_page;
1101 int block;
1102 int pnum;
1103 int poff;
1104 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001105 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001106 struct ext4_group_info *grp;
1107 struct ext4_sb_info *sbi = EXT4_SB(sb);
1108 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001109
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001110 might_sleep();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001111 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001112
1113 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001114 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001115
1116 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001117 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001118 e4b->bd_sb = sb;
1119 e4b->bd_group = group;
1120 e4b->bd_buddy_page = NULL;
1121 e4b->bd_bitmap_page = NULL;
1122
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001123 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001124 /*
1125 * we need full data about the group
1126 * to make a good selection
1127 */
1128 ret = ext4_mb_init_group(sb, group);
1129 if (ret)
1130 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001131 }
1132
Alex Tomasc9de5602008-01-29 00:19:52 -05001133 /*
1134 * the buddy cache inode stores the block bitmap
1135 * and buddy information in consecutive blocks.
1136 * So for each group we need two blocks.
1137 */
1138 block = group * 2;
1139 pnum = block / blocks_per_page;
1140 poff = block % blocks_per_page;
1141
1142 /* we could use find_or_create_page(), but it locks page
1143 * what we'd like to avoid in fast path ... */
1144 page = find_get_page(inode->i_mapping, pnum);
1145 if (page == NULL || !PageUptodate(page)) {
1146 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001147 /*
1148 * drop the page reference and try
1149 * to get the page with lock. If we
1150 * are not uptodate that implies
1151 * somebody just created the page but
1152 * is yet to initialize the same. So
1153 * wait for it to initialize.
1154 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001155 page_cache_release(page);
1156 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1157 if (page) {
1158 BUG_ON(page->mapping != inode->i_mapping);
1159 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001160 ret = ext4_mb_init_cache(page, NULL);
1161 if (ret) {
1162 unlock_page(page);
1163 goto err;
1164 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001165 mb_cmp_bitmaps(e4b, page_address(page) +
1166 (poff * sb->s_blocksize));
1167 }
1168 unlock_page(page);
1169 }
1170 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001171 if (page == NULL || !PageUptodate(page)) {
1172 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001173 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001174 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001175 e4b->bd_bitmap_page = page;
1176 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1177 mark_page_accessed(page);
1178
1179 block++;
1180 pnum = block / blocks_per_page;
1181 poff = block % blocks_per_page;
1182
1183 page = find_get_page(inode->i_mapping, pnum);
1184 if (page == NULL || !PageUptodate(page)) {
1185 if (page)
1186 page_cache_release(page);
1187 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1188 if (page) {
1189 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001190 if (!PageUptodate(page)) {
1191 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1192 if (ret) {
1193 unlock_page(page);
1194 goto err;
1195 }
1196 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001197 unlock_page(page);
1198 }
1199 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001200 if (page == NULL || !PageUptodate(page)) {
1201 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001202 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001203 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001204 e4b->bd_buddy_page = page;
1205 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1206 mark_page_accessed(page);
1207
1208 BUG_ON(e4b->bd_bitmap_page == NULL);
1209 BUG_ON(e4b->bd_buddy_page == NULL);
1210
1211 return 0;
1212
1213err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001214 if (page)
1215 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001216 if (e4b->bd_bitmap_page)
1217 page_cache_release(e4b->bd_bitmap_page);
1218 if (e4b->bd_buddy_page)
1219 page_cache_release(e4b->bd_buddy_page);
1220 e4b->bd_buddy = NULL;
1221 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001222 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001223}
1224
Jing Zhange39e07f2010-05-14 00:00:00 -04001225static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001226{
1227 if (e4b->bd_bitmap_page)
1228 page_cache_release(e4b->bd_bitmap_page);
1229 if (e4b->bd_buddy_page)
1230 page_cache_release(e4b->bd_buddy_page);
1231}
1232
1233
1234static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1235{
1236 int order = 1;
1237 void *bb;
1238
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001239 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001240 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1241
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001242 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001243 while (order <= e4b->bd_blkbits + 1) {
1244 block = block >> 1;
1245 if (!mb_test_bit(block, bb)) {
1246 /* this block is part of buddy of order 'order' */
1247 return order;
1248 }
1249 bb += 1 << (e4b->bd_blkbits - order);
1250 order++;
1251 }
1252 return 0;
1253}
1254
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001255static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001256{
1257 __u32 *addr;
1258
1259 len = cur + len;
1260 while (cur < len) {
1261 if ((cur & 31) == 0 && (len - cur) >= 32) {
1262 /* fast path: clear whole word at once */
1263 addr = bm + (cur >> 3);
1264 *addr = 0;
1265 cur += 32;
1266 continue;
1267 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001268 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001269 cur++;
1270 }
1271}
1272
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001273/* clear bits in given range
1274 * will return first found zero bit if any, -1 otherwise
1275 */
1276static int mb_test_and_clear_bits(void *bm, int cur, int len)
1277{
1278 __u32 *addr;
1279 int zero_bit = -1;
1280
1281 len = cur + len;
1282 while (cur < len) {
1283 if ((cur & 31) == 0 && (len - cur) >= 32) {
1284 /* fast path: clear whole word at once */
1285 addr = bm + (cur >> 3);
1286 if (*addr != (__u32)(-1) && zero_bit == -1)
1287 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1288 *addr = 0;
1289 cur += 32;
1290 continue;
1291 }
1292 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1293 zero_bit = cur;
1294 cur++;
1295 }
1296
1297 return zero_bit;
1298}
1299
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001300void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001301{
1302 __u32 *addr;
1303
1304 len = cur + len;
1305 while (cur < len) {
1306 if ((cur & 31) == 0 && (len - cur) >= 32) {
1307 /* fast path: set whole word at once */
1308 addr = bm + (cur >> 3);
1309 *addr = 0xffffffff;
1310 cur += 32;
1311 continue;
1312 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001313 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001314 cur++;
1315 }
1316}
1317
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001318/*
1319 * _________________________________________________________________ */
1320
1321static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001322{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001323 if (mb_test_bit(*bit + side, bitmap)) {
1324 mb_clear_bit(*bit, bitmap);
1325 (*bit) -= side;
1326 return 1;
1327 }
1328 else {
1329 (*bit) += side;
1330 mb_set_bit(*bit, bitmap);
1331 return -1;
1332 }
1333}
1334
1335static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1336{
1337 int max;
1338 int order = 1;
1339 void *buddy = mb_find_buddy(e4b, order, &max);
1340
1341 while (buddy) {
1342 void *buddy2;
1343
1344 /* Bits in range [first; last] are known to be set since
1345 * corresponding blocks were allocated. Bits in range
1346 * (first; last) will stay set because they form buddies on
1347 * upper layer. We just deal with borders if they don't
1348 * align with upper layer and then go up.
1349 * Releasing entire group is all about clearing
1350 * single bit of highest order buddy.
1351 */
1352
1353 /* Example:
1354 * ---------------------------------
1355 * | 1 | 1 | 1 | 1 |
1356 * ---------------------------------
1357 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1358 * ---------------------------------
1359 * 0 1 2 3 4 5 6 7
1360 * \_____________________/
1361 *
1362 * Neither [1] nor [6] is aligned to above layer.
1363 * Left neighbour [0] is free, so mark it busy,
1364 * decrease bb_counters and extend range to
1365 * [0; 6]
1366 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1367 * mark [6] free, increase bb_counters and shrink range to
1368 * [0; 5].
1369 * Then shift range to [0; 2], go up and do the same.
1370 */
1371
1372
1373 if (first & 1)
1374 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1375 if (!(last & 1))
1376 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1377 if (first > last)
1378 break;
1379 order++;
1380
1381 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1382 mb_clear_bits(buddy, first, last - first + 1);
1383 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1384 break;
1385 }
1386 first >>= 1;
1387 last >>= 1;
1388 buddy = buddy2;
1389 }
1390}
1391
1392static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1393 int first, int count)
1394{
1395 int left_is_free = 0;
1396 int right_is_free = 0;
1397 int block;
1398 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001399 struct super_block *sb = e4b->bd_sb;
1400
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001401 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001402 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Darrick J. Wong163a2032013-08-28 17:35:51 -04001403 /* Don't bother if the block group is corrupt. */
1404 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1405 return;
1406
Alex Tomasc9de5602008-01-29 00:19:52 -05001407 mb_check_buddy(e4b);
1408 mb_free_blocks_double(inode, e4b, first, count);
1409
1410 e4b->bd_info->bb_free += count;
1411 if (first < e4b->bd_info->bb_first_free)
1412 e4b->bd_info->bb_first_free = first;
1413
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001414 /* access memory sequentially: check left neighbour,
1415 * clear range and then check right neighbour
1416 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001417 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001418 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1419 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1420 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1421 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1422
1423 if (unlikely(block != -1)) {
1424 ext4_fsblk_t blocknr;
1425
1426 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1427 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1428 ext4_grp_locked_error(sb, e4b->bd_group,
1429 inode ? inode->i_ino : 0,
1430 blocknr,
1431 "freeing already freed block "
Darrick J. Wong163a2032013-08-28 17:35:51 -04001432 "(bit %u); block bitmap corrupt.",
1433 block);
1434 /* Mark the block group as corrupt. */
1435 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1436 &e4b->bd_info->bb_state);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001437 mb_regenerate_buddy(e4b);
1438 goto done;
1439 }
1440
1441 /* let's maintain fragments counter */
1442 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001443 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001444 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001445 e4b->bd_info->bb_fragments++;
1446
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001447 /* buddy[0] == bd_bitmap is a special case, so handle
1448 * it right away and let mb_buddy_mark_free stay free of
1449 * zero order checks.
1450 * Check if neighbours are to be coaleasced,
1451 * adjust bitmap bb_counters and borders appropriately.
1452 */
1453 if (first & 1) {
1454 first += !left_is_free;
1455 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001456 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001457 if (!(last & 1)) {
1458 last -= !right_is_free;
1459 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1460 }
1461
1462 if (first <= last)
1463 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1464
1465done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001466 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001467 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001468}
1469
Robin Dong15c006a2012-08-17 10:02:17 -04001470static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001471 int needed, struct ext4_free_extent *ex)
1472{
1473 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001474 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001475 void *buddy;
1476
Vincent Minetbc8e6742009-05-15 08:33:18 -04001477 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001478 BUG_ON(ex == NULL);
1479
Robin Dong15c006a2012-08-17 10:02:17 -04001480 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001481 BUG_ON(buddy == NULL);
1482 BUG_ON(block >= max);
1483 if (mb_test_bit(block, buddy)) {
1484 ex->fe_len = 0;
1485 ex->fe_start = 0;
1486 ex->fe_group = 0;
1487 return 0;
1488 }
1489
Robin Dong15c006a2012-08-17 10:02:17 -04001490 /* find actual order */
1491 order = mb_find_order_for_block(e4b, block);
1492 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001493
1494 ex->fe_len = 1 << order;
1495 ex->fe_start = block << order;
1496 ex->fe_group = e4b->bd_group;
1497
1498 /* calc difference from given start */
1499 next = next - ex->fe_start;
1500 ex->fe_len -= next;
1501 ex->fe_start += next;
1502
1503 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001504 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001505
1506 if (block + 1 >= max)
1507 break;
1508
1509 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001510 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001511 break;
1512
Robin Dongb051d8d2011-10-26 05:30:30 -04001513 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001514
Alex Tomasc9de5602008-01-29 00:19:52 -05001515 block = next >> order;
1516 ex->fe_len += 1 << order;
1517 }
1518
1519 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1520 return ex->fe_len;
1521}
1522
1523static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1524{
1525 int ord;
1526 int mlen = 0;
1527 int max = 0;
1528 int cur;
1529 int start = ex->fe_start;
1530 int len = ex->fe_len;
1531 unsigned ret = 0;
1532 int len0 = len;
1533 void *buddy;
1534
1535 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1536 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001537 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001538 mb_check_buddy(e4b);
1539 mb_mark_used_double(e4b, start, len);
1540
1541 e4b->bd_info->bb_free -= len;
1542 if (e4b->bd_info->bb_first_free == start)
1543 e4b->bd_info->bb_first_free += len;
1544
1545 /* let's maintain fragments counter */
1546 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001547 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001548 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001549 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001550 if (mlen && max)
1551 e4b->bd_info->bb_fragments++;
1552 else if (!mlen && !max)
1553 e4b->bd_info->bb_fragments--;
1554
1555 /* let's maintain buddy itself */
1556 while (len) {
1557 ord = mb_find_order_for_block(e4b, start);
1558
1559 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1560 /* the whole chunk may be allocated at once! */
1561 mlen = 1 << ord;
1562 buddy = mb_find_buddy(e4b, ord, &max);
1563 BUG_ON((start >> ord) >= max);
1564 mb_set_bit(start >> ord, buddy);
1565 e4b->bd_info->bb_counters[ord]--;
1566 start += mlen;
1567 len -= mlen;
1568 BUG_ON(len < 0);
1569 continue;
1570 }
1571
1572 /* store for history */
1573 if (ret == 0)
1574 ret = len | (ord << 16);
1575
1576 /* we have to split large buddy */
1577 BUG_ON(ord <= 0);
1578 buddy = mb_find_buddy(e4b, ord, &max);
1579 mb_set_bit(start >> ord, buddy);
1580 e4b->bd_info->bb_counters[ord]--;
1581
1582 ord--;
1583 cur = (start >> ord) & ~1U;
1584 buddy = mb_find_buddy(e4b, ord, &max);
1585 mb_clear_bit(cur, buddy);
1586 mb_clear_bit(cur + 1, buddy);
1587 e4b->bd_info->bb_counters[ord]++;
1588 e4b->bd_info->bb_counters[ord]++;
1589 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001590 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001591
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001592 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001593 mb_check_buddy(e4b);
1594
1595 return ret;
1596}
1597
1598/*
1599 * Must be called under group lock!
1600 */
1601static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1602 struct ext4_buddy *e4b)
1603{
1604 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1605 int ret;
1606
1607 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1608 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1609
1610 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1611 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1612 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1613
1614 /* preallocation can change ac_b_ex, thus we store actually
1615 * allocated blocks for history */
1616 ac->ac_f_ex = ac->ac_b_ex;
1617
1618 ac->ac_status = AC_STATUS_FOUND;
1619 ac->ac_tail = ret & 0xffff;
1620 ac->ac_buddy = ret >> 16;
1621
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001622 /*
1623 * take the page reference. We want the page to be pinned
1624 * so that we don't get a ext4_mb_init_cache_call for this
1625 * group until we update the bitmap. That would mean we
1626 * double allocate blocks. The reference is dropped
1627 * in ext4_mb_release_context
1628 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001629 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1630 get_page(ac->ac_bitmap_page);
1631 ac->ac_buddy_page = e4b->bd_buddy_page;
1632 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001633 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001634 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001635 spin_lock(&sbi->s_md_lock);
1636 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1637 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1638 spin_unlock(&sbi->s_md_lock);
1639 }
1640}
1641
1642/*
1643 * regular allocator, for general purposes allocation
1644 */
1645
1646static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1647 struct ext4_buddy *e4b,
1648 int finish_group)
1649{
1650 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1651 struct ext4_free_extent *bex = &ac->ac_b_ex;
1652 struct ext4_free_extent *gex = &ac->ac_g_ex;
1653 struct ext4_free_extent ex;
1654 int max;
1655
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001656 if (ac->ac_status == AC_STATUS_FOUND)
1657 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001658 /*
1659 * We don't want to scan for a whole year
1660 */
1661 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1662 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1663 ac->ac_status = AC_STATUS_BREAK;
1664 return;
1665 }
1666
1667 /*
1668 * Haven't found good chunk so far, let's continue
1669 */
1670 if (bex->fe_len < gex->fe_len)
1671 return;
1672
1673 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1674 && bex->fe_group == e4b->bd_group) {
1675 /* recheck chunk's availability - we don't know
1676 * when it was found (within this lock-unlock
1677 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001678 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001679 if (max >= gex->fe_len) {
1680 ext4_mb_use_best_found(ac, e4b);
1681 return;
1682 }
1683 }
1684}
1685
1686/*
1687 * The routine checks whether found extent is good enough. If it is,
1688 * then the extent gets marked used and flag is set to the context
1689 * to stop scanning. Otherwise, the extent is compared with the
1690 * previous found extent and if new one is better, then it's stored
1691 * in the context. Later, the best found extent will be used, if
1692 * mballoc can't find good enough extent.
1693 *
1694 * FIXME: real allocation policy is to be designed yet!
1695 */
1696static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1697 struct ext4_free_extent *ex,
1698 struct ext4_buddy *e4b)
1699{
1700 struct ext4_free_extent *bex = &ac->ac_b_ex;
1701 struct ext4_free_extent *gex = &ac->ac_g_ex;
1702
1703 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001704 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1705 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001706 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1707
1708 ac->ac_found++;
1709
1710 /*
1711 * The special case - take what you catch first
1712 */
1713 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1714 *bex = *ex;
1715 ext4_mb_use_best_found(ac, e4b);
1716 return;
1717 }
1718
1719 /*
1720 * Let's check whether the chuck is good enough
1721 */
1722 if (ex->fe_len == gex->fe_len) {
1723 *bex = *ex;
1724 ext4_mb_use_best_found(ac, e4b);
1725 return;
1726 }
1727
1728 /*
1729 * If this is first found extent, just store it in the context
1730 */
1731 if (bex->fe_len == 0) {
1732 *bex = *ex;
1733 return;
1734 }
1735
1736 /*
1737 * If new found extent is better, store it in the context
1738 */
1739 if (bex->fe_len < gex->fe_len) {
1740 /* if the request isn't satisfied, any found extent
1741 * larger than previous best one is better */
1742 if (ex->fe_len > bex->fe_len)
1743 *bex = *ex;
1744 } else if (ex->fe_len > gex->fe_len) {
1745 /* if the request is satisfied, then we try to find
1746 * an extent that still satisfy the request, but is
1747 * smaller than previous one */
1748 if (ex->fe_len < bex->fe_len)
1749 *bex = *ex;
1750 }
1751
1752 ext4_mb_check_limits(ac, e4b, 0);
1753}
1754
Eric Sandeen089ceec2009-07-05 22:17:31 -04001755static noinline_for_stack
1756int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001757 struct ext4_buddy *e4b)
1758{
1759 struct ext4_free_extent ex = ac->ac_b_ex;
1760 ext4_group_t group = ex.fe_group;
1761 int max;
1762 int err;
1763
1764 BUG_ON(ex.fe_len <= 0);
1765 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1766 if (err)
1767 return err;
1768
1769 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001770 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001771
1772 if (max > 0) {
1773 ac->ac_b_ex = ex;
1774 ext4_mb_use_best_found(ac, e4b);
1775 }
1776
1777 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001778 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001779
1780 return 0;
1781}
1782
Eric Sandeen089ceec2009-07-05 22:17:31 -04001783static noinline_for_stack
1784int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001785 struct ext4_buddy *e4b)
1786{
1787 ext4_group_t group = ac->ac_g_ex.fe_group;
1788 int max;
1789 int err;
1790 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001791 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001792 struct ext4_free_extent ex;
1793
1794 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1795 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001796 if (grp->bb_free == 0)
1797 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001798
1799 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1800 if (err)
1801 return err;
1802
Darrick J. Wong163a2032013-08-28 17:35:51 -04001803 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1804 ext4_mb_unload_buddy(e4b);
1805 return 0;
1806 }
1807
Alex Tomasc9de5602008-01-29 00:19:52 -05001808 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001809 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001810 ac->ac_g_ex.fe_len, &ex);
1811
1812 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1813 ext4_fsblk_t start;
1814
Akinobu Mita5661bd62010-03-03 23:53:39 -05001815 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1816 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001817 /* use do_div to get remainder (would be 64-bit modulo) */
1818 if (do_div(start, sbi->s_stripe) == 0) {
1819 ac->ac_found++;
1820 ac->ac_b_ex = ex;
1821 ext4_mb_use_best_found(ac, e4b);
1822 }
1823 } else if (max >= ac->ac_g_ex.fe_len) {
1824 BUG_ON(ex.fe_len <= 0);
1825 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1826 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1827 ac->ac_found++;
1828 ac->ac_b_ex = ex;
1829 ext4_mb_use_best_found(ac, e4b);
1830 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1831 /* Sometimes, caller may want to merge even small
1832 * number of blocks to an existing extent */
1833 BUG_ON(ex.fe_len <= 0);
1834 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1835 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1836 ac->ac_found++;
1837 ac->ac_b_ex = ex;
1838 ext4_mb_use_best_found(ac, e4b);
1839 }
1840 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001841 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001842
1843 return 0;
1844}
1845
1846/*
1847 * The routine scans buddy structures (not bitmap!) from given order
1848 * to max order and tries to find big enough chunk to satisfy the req
1849 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001850static noinline_for_stack
1851void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001852 struct ext4_buddy *e4b)
1853{
1854 struct super_block *sb = ac->ac_sb;
1855 struct ext4_group_info *grp = e4b->bd_info;
1856 void *buddy;
1857 int i;
1858 int k;
1859 int max;
1860
1861 BUG_ON(ac->ac_2order <= 0);
1862 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1863 if (grp->bb_counters[i] == 0)
1864 continue;
1865
1866 buddy = mb_find_buddy(e4b, i, &max);
1867 BUG_ON(buddy == NULL);
1868
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001869 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001870 BUG_ON(k >= max);
1871
1872 ac->ac_found++;
1873
1874 ac->ac_b_ex.fe_len = 1 << i;
1875 ac->ac_b_ex.fe_start = k << i;
1876 ac->ac_b_ex.fe_group = e4b->bd_group;
1877
1878 ext4_mb_use_best_found(ac, e4b);
1879
1880 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1881
1882 if (EXT4_SB(sb)->s_mb_stats)
1883 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1884
1885 break;
1886 }
1887}
1888
1889/*
1890 * The routine scans the group and measures all found extents.
1891 * In order to optimize scanning, caller must pass number of
1892 * free blocks in the group, so the routine can know upper limit.
1893 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001894static noinline_for_stack
1895void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001896 struct ext4_buddy *e4b)
1897{
1898 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001899 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001900 struct ext4_free_extent ex;
1901 int i;
1902 int free;
1903
1904 free = e4b->bd_info->bb_free;
1905 BUG_ON(free <= 0);
1906
1907 i = e4b->bd_info->bb_first_free;
1908
1909 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001910 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001911 EXT4_CLUSTERS_PER_GROUP(sb), i);
1912 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001913 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001914 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001915 * free blocks even though group info says we
1916 * we have free blocks
1917 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001918 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001919 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001920 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001921 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001922 break;
1923 }
1924
Robin Dong15c006a2012-08-17 10:02:17 -04001925 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001926 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001927 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001928 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001929 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001930 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001931 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001932 /*
1933 * The number of free blocks differs. This mostly
1934 * indicate that the bitmap is corrupt. So exit
1935 * without claiming the space.
1936 */
1937 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001938 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001939
1940 ext4_mb_measure_extent(ac, &ex, e4b);
1941
1942 i += ex.fe_len;
1943 free -= ex.fe_len;
1944 }
1945
1946 ext4_mb_check_limits(ac, e4b, 1);
1947}
1948
1949/*
1950 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001951 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001952 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001953static noinline_for_stack
1954void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001955 struct ext4_buddy *e4b)
1956{
1957 struct super_block *sb = ac->ac_sb;
1958 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001959 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001960 struct ext4_free_extent ex;
1961 ext4_fsblk_t first_group_block;
1962 ext4_fsblk_t a;
1963 ext4_grpblk_t i;
1964 int max;
1965
1966 BUG_ON(sbi->s_stripe == 0);
1967
1968 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001969 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1970
Alex Tomasc9de5602008-01-29 00:19:52 -05001971 a = first_group_block + sbi->s_stripe - 1;
1972 do_div(a, sbi->s_stripe);
1973 i = (a * sbi->s_stripe) - first_group_block;
1974
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001975 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001976 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04001977 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001978 if (max >= sbi->s_stripe) {
1979 ac->ac_found++;
1980 ac->ac_b_ex = ex;
1981 ext4_mb_use_best_found(ac, e4b);
1982 break;
1983 }
1984 }
1985 i += sbi->s_stripe;
1986 }
1987}
1988
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001989/* This is now called BEFORE we load the buddy bitmap. */
Alex Tomasc9de5602008-01-29 00:19:52 -05001990static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1991 ext4_group_t group, int cr)
1992{
1993 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04001994 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001995 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1996
1997 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001998
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04001999 free = grp->bb_free;
2000 if (free == 0)
2001 return 0;
2002 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2003 return 0;
2004
Darrick J. Wong163a2032013-08-28 17:35:51 -04002005 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2006 return 0;
2007
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002008 /* We only do this if the grp has never been initialized */
2009 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2010 int ret = ext4_mb_init_group(ac->ac_sb, group);
2011 if (ret)
2012 return 0;
2013 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002014
Alex Tomasc9de5602008-01-29 00:19:52 -05002015 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05002016 if (fragments == 0)
2017 return 0;
2018
2019 switch (cr) {
2020 case 0:
2021 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002022
Theodore Ts'oa4912122009-03-12 12:18:34 -04002023 /* Avoid using the first bg of a flexgroup for data files */
2024 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2025 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2026 ((group % flex_size) == 0))
2027 return 0;
2028
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002029 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2030 (free / fragments) >= ac->ac_g_ex.fe_len)
2031 return 1;
2032
2033 if (grp->bb_largest_free_order < ac->ac_2order)
2034 return 0;
2035
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002036 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002037 case 1:
2038 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2039 return 1;
2040 break;
2041 case 2:
2042 if (free >= ac->ac_g_ex.fe_len)
2043 return 1;
2044 break;
2045 case 3:
2046 return 1;
2047 default:
2048 BUG();
2049 }
2050
2051 return 0;
2052}
2053
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002054static noinline_for_stack int
2055ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002056{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002057 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002058 int cr;
2059 int err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002060 struct ext4_sb_info *sbi;
2061 struct super_block *sb;
2062 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002063
2064 sb = ac->ac_sb;
2065 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002066 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002067 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002068 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002069 ngroups = sbi->s_blockfile_groups;
2070
Alex Tomasc9de5602008-01-29 00:19:52 -05002071 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2072
2073 /* first, try the goal */
2074 err = ext4_mb_find_by_goal(ac, &e4b);
2075 if (err || ac->ac_status == AC_STATUS_FOUND)
2076 goto out;
2077
2078 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2079 goto out;
2080
2081 /*
2082 * ac->ac2_order is set only if the fe_len is a power of 2
2083 * if ac2_order is set we also set criteria to 0 so that we
2084 * try exact allocation using buddy.
2085 */
2086 i = fls(ac->ac_g_ex.fe_len);
2087 ac->ac_2order = 0;
2088 /*
2089 * We search using buddy data only if the order of the request
2090 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002091 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002092 */
2093 if (i >= sbi->s_mb_order2_reqs) {
2094 /*
2095 * This should tell if fe_len is exactly power of 2
2096 */
2097 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2098 ac->ac_2order = i - 1;
2099 }
2100
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002101 /* if stream allocation is enabled, use global goal */
2102 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002103 /* TBD: may be hot point */
2104 spin_lock(&sbi->s_md_lock);
2105 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2106 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2107 spin_unlock(&sbi->s_md_lock);
2108 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002109
Alex Tomasc9de5602008-01-29 00:19:52 -05002110 /* Let's just scan groups to find more-less suitable blocks */
2111 cr = ac->ac_2order ? 0 : 1;
2112 /*
2113 * cr == 0 try to get exact allocation,
2114 * cr == 3 try to get anything
2115 */
2116repeat:
2117 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2118 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002119 /*
2120 * searching for the right group start
2121 * from the goal value specified
2122 */
2123 group = ac->ac_g_ex.fe_group;
2124
Theodore Ts'o8df96752009-05-01 08:50:38 -04002125 for (i = 0; i < ngroups; group++, i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04002126 cond_resched();
Lachlan McIlroye6155732013-05-05 23:10:00 -04002127 /*
2128 * Artificially restricted ngroups for non-extent
2129 * files makes group > ngroups possible on first loop.
2130 */
2131 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002132 group = 0;
2133
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002134 /* This now checks without needing the buddy page */
2135 if (!ext4_mb_good_group(ac, group, cr))
Alex Tomasc9de5602008-01-29 00:19:52 -05002136 continue;
2137
Alex Tomasc9de5602008-01-29 00:19:52 -05002138 err = ext4_mb_load_buddy(sb, group, &e4b);
2139 if (err)
2140 goto out;
2141
2142 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002143
2144 /*
2145 * We need to check again after locking the
2146 * block group
2147 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002148 if (!ext4_mb_good_group(ac, group, cr)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002149 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002150 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002151 continue;
2152 }
2153
2154 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002155 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002156 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002157 else if (cr == 1 && sbi->s_stripe &&
2158 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002159 ext4_mb_scan_aligned(ac, &e4b);
2160 else
2161 ext4_mb_complex_scan_group(ac, &e4b);
2162
2163 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002164 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002165
2166 if (ac->ac_status != AC_STATUS_CONTINUE)
2167 break;
2168 }
2169 }
2170
2171 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2172 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2173 /*
2174 * We've been searching too long. Let's try to allocate
2175 * the best chunk we've found so far
2176 */
2177
2178 ext4_mb_try_best_found(ac, &e4b);
2179 if (ac->ac_status != AC_STATUS_FOUND) {
2180 /*
2181 * Someone more lucky has already allocated it.
2182 * The only thing we can do is just take first
2183 * found block(s)
2184 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2185 */
2186 ac->ac_b_ex.fe_group = 0;
2187 ac->ac_b_ex.fe_start = 0;
2188 ac->ac_b_ex.fe_len = 0;
2189 ac->ac_status = AC_STATUS_CONTINUE;
2190 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2191 cr = 3;
2192 atomic_inc(&sbi->s_mb_lost_chunks);
2193 goto repeat;
2194 }
2195 }
2196out:
2197 return err;
2198}
2199
Alex Tomasc9de5602008-01-29 00:19:52 -05002200static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2201{
2202 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002203 ext4_group_t group;
2204
Theodore Ts'o8df96752009-05-01 08:50:38 -04002205 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002206 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002207 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002208 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002209}
2210
2211static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2212{
2213 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002214 ext4_group_t group;
2215
2216 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002217 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002218 return NULL;
2219 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002220 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002221}
2222
2223static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2224{
2225 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002226 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002227 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002228 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002229 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002230 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002231 struct sg {
2232 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002233 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002234 } sg;
2235
2236 group--;
2237 if (group == 0)
2238 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2239 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2240 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2241 "group", "free", "frags", "first",
2242 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2243 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2244
2245 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2246 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002247 grinfo = ext4_get_group_info(sb, group);
2248 /* Load the group info in memory only if not already loaded. */
2249 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2250 err = ext4_mb_load_buddy(sb, group, &e4b);
2251 if (err) {
2252 seq_printf(seq, "#%-5u: I/O error\n", group);
2253 return 0;
2254 }
2255 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002256 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002257
Alex Tomasc9de5602008-01-29 00:19:52 -05002258 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002259
2260 if (buddy_loaded)
2261 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002262
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002263 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002264 sg.info.bb_fragments, sg.info.bb_first_free);
2265 for (i = 0; i <= 13; i++)
2266 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2267 sg.info.bb_counters[i] : 0);
2268 seq_printf(seq, " ]\n");
2269
2270 return 0;
2271}
2272
2273static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2274{
2275}
2276
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002277static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002278 .start = ext4_mb_seq_groups_start,
2279 .next = ext4_mb_seq_groups_next,
2280 .stop = ext4_mb_seq_groups_stop,
2281 .show = ext4_mb_seq_groups_show,
2282};
2283
2284static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2285{
Al Virod9dda782013-03-31 18:16:14 -04002286 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002287 int rc;
2288
2289 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2290 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002291 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002292 m->private = sb;
2293 }
2294 return rc;
2295
2296}
2297
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002298static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002299 .owner = THIS_MODULE,
2300 .open = ext4_mb_seq_groups_open,
2301 .read = seq_read,
2302 .llseek = seq_lseek,
2303 .release = seq_release,
2304};
2305
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002306static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2307{
2308 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2309 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2310
2311 BUG_ON(!cachep);
2312 return cachep;
2313}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002314
Theodore Ts'o28623c22012-09-05 01:31:50 -04002315/*
2316 * Allocate the top-level s_group_info array for the specified number
2317 * of groups
2318 */
2319int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2320{
2321 struct ext4_sb_info *sbi = EXT4_SB(sb);
2322 unsigned size;
2323 struct ext4_group_info ***new_groupinfo;
2324
2325 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2326 EXT4_DESC_PER_BLOCK_BITS(sb);
2327 if (size <= sbi->s_group_info_size)
2328 return 0;
2329
2330 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2331 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2332 if (!new_groupinfo) {
2333 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2334 return -ENOMEM;
2335 }
2336 if (sbi->s_group_info) {
2337 memcpy(new_groupinfo, sbi->s_group_info,
2338 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2339 ext4_kvfree(sbi->s_group_info);
2340 }
2341 sbi->s_group_info = new_groupinfo;
2342 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2343 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2344 sbi->s_group_info_size);
2345 return 0;
2346}
2347
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002348/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002349int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002350 struct ext4_group_desc *desc)
2351{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002352 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002353 int metalen = 0;
2354 struct ext4_sb_info *sbi = EXT4_SB(sb);
2355 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002356 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002357
2358 /*
2359 * First check if this group is the first of a reserved block.
2360 * If it's true, we have to allocate a new table of pointers
2361 * to ext4_group_info structures
2362 */
2363 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2364 metalen = sizeof(*meta_group_info) <<
2365 EXT4_DESC_PER_BLOCK_BITS(sb);
2366 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2367 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002368 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002369 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002370 goto exit_meta_group_info;
2371 }
2372 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2373 meta_group_info;
2374 }
2375
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002376 meta_group_info =
2377 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2378 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2379
Wei Yongjun85556c92012-09-26 20:43:37 -04002380 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002381 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002382 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002383 goto exit_group_info;
2384 }
2385 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2386 &(meta_group_info[i]->bb_state));
2387
2388 /*
2389 * initialize bb_free to be able to skip
2390 * empty groups without initialization
2391 */
2392 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2393 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002394 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002395 } else {
2396 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002397 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002398 }
2399
2400 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002401 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002402 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002403 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002404
2405#ifdef DOUBLE_CHECK
2406 {
2407 struct buffer_head *bh;
2408 meta_group_info[i]->bb_bitmap =
2409 kmalloc(sb->s_blocksize, GFP_KERNEL);
2410 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2411 bh = ext4_read_block_bitmap(sb, group);
2412 BUG_ON(bh == NULL);
2413 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2414 sb->s_blocksize);
2415 put_bh(bh);
2416 }
2417#endif
2418
2419 return 0;
2420
2421exit_group_info:
2422 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002423 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002424 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002425 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2426 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002427exit_meta_group_info:
2428 return -ENOMEM;
2429} /* ext4_mb_add_groupinfo */
2430
Alex Tomasc9de5602008-01-29 00:19:52 -05002431static int ext4_mb_init_backend(struct super_block *sb)
2432{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002433 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002434 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002435 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002436 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002437 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002438 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002439
Theodore Ts'o28623c22012-09-05 01:31:50 -04002440 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2441 if (err)
2442 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002443
Alex Tomasc9de5602008-01-29 00:19:52 -05002444 sbi->s_buddy_cache = new_inode(sb);
2445 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002446 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002447 goto err_freesgi;
2448 }
Yu Jian48e60612011-08-01 17:41:39 -04002449 /* To avoid potentially colliding with an valid on-disk inode number,
2450 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2451 * not in the inode hash, so it should never be found by iget(), but
2452 * this will avoid confusion if it ever shows up during debugging. */
2453 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002454 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002455 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002456 desc = ext4_get_group_desc(sb, i, NULL);
2457 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002458 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002459 goto err_freebuddy;
2460 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002461 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2462 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002463 }
2464
2465 return 0;
2466
2467err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002468 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002469 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002470 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002471 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002472 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002473 kfree(sbi->s_group_info[i]);
2474 iput(sbi->s_buddy_cache);
2475err_freesgi:
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002476 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002477 return -ENOMEM;
2478}
2479
Eric Sandeen2892c152011-02-12 08:12:18 -05002480static void ext4_groupinfo_destroy_slabs(void)
2481{
2482 int i;
2483
2484 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2485 if (ext4_groupinfo_caches[i])
2486 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2487 ext4_groupinfo_caches[i] = NULL;
2488 }
2489}
2490
2491static int ext4_groupinfo_create_slab(size_t size)
2492{
2493 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2494 int slab_size;
2495 int blocksize_bits = order_base_2(size);
2496 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2497 struct kmem_cache *cachep;
2498
2499 if (cache_index >= NR_GRPINFO_CACHES)
2500 return -EINVAL;
2501
2502 if (unlikely(cache_index < 0))
2503 cache_index = 0;
2504
2505 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2506 if (ext4_groupinfo_caches[cache_index]) {
2507 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2508 return 0; /* Already created */
2509 }
2510
2511 slab_size = offsetof(struct ext4_group_info,
2512 bb_counters[blocksize_bits + 2]);
2513
2514 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2515 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2516 NULL);
2517
Tao Ma823ba012011-07-11 18:26:01 -04002518 ext4_groupinfo_caches[cache_index] = cachep;
2519
Eric Sandeen2892c152011-02-12 08:12:18 -05002520 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2521 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002522 printk(KERN_EMERG
2523 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002524 return -ENOMEM;
2525 }
2526
Eric Sandeen2892c152011-02-12 08:12:18 -05002527 return 0;
2528}
2529
Akira Fujita9d990122012-05-28 14:19:25 -04002530int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002531{
2532 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002533 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002534 unsigned offset;
2535 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002536 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002537
Eric Sandeen19278052009-08-25 22:36:25 -04002538 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002539
2540 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2541 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002542 ret = -ENOMEM;
2543 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002544 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002545
Eric Sandeen19278052009-08-25 22:36:25 -04002546 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002547 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2548 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002549 ret = -ENOMEM;
2550 goto out;
2551 }
2552
Eric Sandeen2892c152011-02-12 08:12:18 -05002553 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2554 if (ret < 0)
2555 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002556
2557 /* order 0 is regular bitmap */
2558 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2559 sbi->s_mb_offsets[0] = 0;
2560
2561 i = 1;
2562 offset = 0;
2563 max = sb->s_blocksize << 2;
2564 do {
2565 sbi->s_mb_offsets[i] = offset;
2566 sbi->s_mb_maxs[i] = max;
2567 offset += 1 << (sb->s_blocksize_bits - i);
2568 max = max >> 1;
2569 i++;
2570 } while (i <= sb->s_blocksize_bits + 1);
2571
Alex Tomasc9de5602008-01-29 00:19:52 -05002572 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002573 spin_lock_init(&sbi->s_bal_lock);
2574
2575 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2576 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2577 sbi->s_mb_stats = MB_DEFAULT_STATS;
2578 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2579 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002580 /*
2581 * The default group preallocation is 512, which for 4k block
2582 * sizes translates to 2 megabytes. However for bigalloc file
2583 * systems, this is probably too big (i.e, if the cluster size
2584 * is 1 megabyte, then group preallocation size becomes half a
2585 * gigabyte!). As a default, we will keep a two megabyte
2586 * group pralloc size for cluster sizes up to 64k, and after
2587 * that, we will force a minimum group preallocation size of
2588 * 32 clusters. This translates to 8 megs when the cluster
2589 * size is 256k, and 32 megs when the cluster size is 1 meg,
2590 * which seems reasonable as a default.
2591 */
2592 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2593 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002594 /*
2595 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2596 * to the lowest multiple of s_stripe which is bigger than
2597 * the s_mb_group_prealloc as determined above. We want
2598 * the preallocation size to be an exact multiple of the
2599 * RAID stripe size so that preallocations don't fragment
2600 * the stripes.
2601 */
2602 if (sbi->s_stripe > 1) {
2603 sbi->s_mb_group_prealloc = roundup(
2604 sbi->s_mb_group_prealloc, sbi->s_stripe);
2605 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002606
Eric Sandeen730c2132008-09-13 15:23:29 -04002607 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002608 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002609 ret = -ENOMEM;
Tao Ma7aa0bae2011-10-06 10:22:28 -04002610 goto out_free_groupinfo_slab;
Alex Tomasc9de5602008-01-29 00:19:52 -05002611 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002612 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002613 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002614 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002615 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002616 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2617 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002618 spin_lock_init(&lg->lg_prealloc_lock);
2619 }
2620
Yu Jian79a77c52011-08-01 17:41:46 -04002621 /* init file for buddy data */
2622 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002623 if (ret != 0)
2624 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002625
Theodore Ts'o296c3552009-09-30 00:32:42 -04002626 if (sbi->s_proc)
2627 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2628 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002629
Tao Ma7aa0bae2011-10-06 10:22:28 -04002630 return 0;
2631
2632out_free_locality_groups:
2633 free_percpu(sbi->s_locality_groups);
2634 sbi->s_locality_groups = NULL;
2635out_free_groupinfo_slab:
2636 ext4_groupinfo_destroy_slabs();
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002637out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002638 kfree(sbi->s_mb_offsets);
2639 sbi->s_mb_offsets = NULL;
2640 kfree(sbi->s_mb_maxs);
2641 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002642 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002643}
2644
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002645/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002646static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2647{
2648 struct ext4_prealloc_space *pa;
2649 struct list_head *cur, *tmp;
2650 int count = 0;
2651
2652 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2653 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2654 list_del(&pa->pa_group_list);
2655 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002656 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002657 }
2658 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002659 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002660
2661}
2662
2663int ext4_mb_release(struct super_block *sb)
2664{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002665 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002666 ext4_group_t i;
2667 int num_meta_group_infos;
2668 struct ext4_group_info *grinfo;
2669 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002670 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002671
Salman Qazi95599962012-05-31 23:52:14 -04002672 if (sbi->s_proc)
2673 remove_proc_entry("mb_groups", sbi->s_proc);
2674
Alex Tomasc9de5602008-01-29 00:19:52 -05002675 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002676 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002677 grinfo = ext4_get_group_info(sb, i);
2678#ifdef DOUBLE_CHECK
2679 kfree(grinfo->bb_bitmap);
2680#endif
2681 ext4_lock_group(sb, i);
2682 ext4_mb_cleanup_pa(grinfo);
2683 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002684 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002685 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002686 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002687 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2688 EXT4_DESC_PER_BLOCK_BITS(sb);
2689 for (i = 0; i < num_meta_group_infos; i++)
2690 kfree(sbi->s_group_info[i]);
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002691 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002692 }
2693 kfree(sbi->s_mb_offsets);
2694 kfree(sbi->s_mb_maxs);
2695 if (sbi->s_buddy_cache)
2696 iput(sbi->s_buddy_cache);
2697 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002698 ext4_msg(sb, KERN_INFO,
2699 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002700 atomic_read(&sbi->s_bal_allocated),
2701 atomic_read(&sbi->s_bal_reqs),
2702 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002703 ext4_msg(sb, KERN_INFO,
2704 "mballoc: %u extents scanned, %u goal hits, "
2705 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002706 atomic_read(&sbi->s_bal_ex_scanned),
2707 atomic_read(&sbi->s_bal_goals),
2708 atomic_read(&sbi->s_bal_2orders),
2709 atomic_read(&sbi->s_bal_breaks),
2710 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002711 ext4_msg(sb, KERN_INFO,
2712 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002713 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002714 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002715 ext4_msg(sb, KERN_INFO,
2716 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002717 atomic_read(&sbi->s_mb_preallocated),
2718 atomic_read(&sbi->s_mb_discarded));
2719 }
2720
Eric Sandeen730c2132008-09-13 15:23:29 -04002721 free_percpu(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05002722
2723 return 0;
2724}
2725
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002726static inline int ext4_issue_discard(struct super_block *sb,
Theodore Ts'o84130192011-09-09 18:50:51 -04002727 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
Jiaying Zhang5c521832010-07-27 11:56:05 -04002728{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002729 ext4_fsblk_t discard_block;
2730
Theodore Ts'o84130192011-09-09 18:50:51 -04002731 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2732 ext4_group_first_block_no(sb, block_group));
2733 count = EXT4_C2B(EXT4_SB(sb), count);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002734 trace_ext4_discard_blocks(sb,
2735 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002736 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002737}
2738
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002739/*
2740 * This function is called by the jbd2 layer once the commit has finished,
2741 * so we know we can free the blocks that were released with that commit.
2742 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002743static void ext4_free_data_callback(struct super_block *sb,
2744 struct ext4_journal_cb_entry *jce,
2745 int rc)
Alex Tomasc9de5602008-01-29 00:19:52 -05002746{
Bobi Jam18aadd42012-02-20 17:53:02 -05002747 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
Alex Tomasc9de5602008-01-29 00:19:52 -05002748 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002749 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002750 int err, count = 0, count2 = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002751
Bobi Jam18aadd42012-02-20 17:53:02 -05002752 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2753 entry->efd_count, entry->efd_group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002754
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05002755 if (test_opt(sb, DISCARD)) {
2756 err = ext4_issue_discard(sb, entry->efd_group,
2757 entry->efd_start_cluster,
2758 entry->efd_count);
2759 if (err && err != -EOPNOTSUPP)
2760 ext4_msg(sb, KERN_WARNING, "discard request in"
2761 " group:%d block:%d count:%d failed"
2762 " with %d", entry->efd_group,
2763 entry->efd_start_cluster,
2764 entry->efd_count, err);
2765 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002766
Bobi Jam18aadd42012-02-20 17:53:02 -05002767 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2768 /* we expect to find existing buddy because it's pinned */
2769 BUG_ON(err != 0);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002770
Alex Tomasc9de5602008-01-29 00:19:52 -05002771
Bobi Jam18aadd42012-02-20 17:53:02 -05002772 db = e4b.bd_info;
2773 /* there are blocks to put in buddy to make them really free */
2774 count += entry->efd_count;
2775 count2++;
2776 ext4_lock_group(sb, entry->efd_group);
2777 /* Take it out of per group rb tree */
2778 rb_erase(&entry->efd_node, &(db->bb_free_root));
2779 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002780
Bobi Jam18aadd42012-02-20 17:53:02 -05002781 /*
2782 * Clear the trimmed flag for the group so that the next
2783 * ext4_trim_fs can trim it.
2784 * If the volume is mounted with -o discard, online discard
2785 * is supported and the free blocks will be trimmed online.
2786 */
2787 if (!test_opt(sb, DISCARD))
2788 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2789
2790 if (!db->bb_free_root.rb_node) {
2791 /* No more items in the per group rb tree
2792 * balance refcounts from ext4_mb_free_metadata()
Tao Ma3d56b8d2011-07-11 00:03:38 -04002793 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002794 page_cache_release(e4b.bd_buddy_page);
2795 page_cache_release(e4b.bd_bitmap_page);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002796 }
Bobi Jam18aadd42012-02-20 17:53:02 -05002797 ext4_unlock_group(sb, entry->efd_group);
2798 kmem_cache_free(ext4_free_data_cachep, entry);
2799 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002800
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002801 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002802}
2803
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002804int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002805{
Theodore Ts'o16828082010-10-27 21:30:09 -04002806 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2807 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002808 if (ext4_pspace_cachep == NULL)
2809 return -ENOMEM;
2810
Theodore Ts'o16828082010-10-27 21:30:09 -04002811 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2812 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002813 if (ext4_ac_cachep == NULL) {
2814 kmem_cache_destroy(ext4_pspace_cachep);
2815 return -ENOMEM;
2816 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002817
Bobi Jam18aadd42012-02-20 17:53:02 -05002818 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2819 SLAB_RECLAIM_ACCOUNT);
2820 if (ext4_free_data_cachep == NULL) {
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002821 kmem_cache_destroy(ext4_pspace_cachep);
2822 kmem_cache_destroy(ext4_ac_cachep);
2823 return -ENOMEM;
2824 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002825 return 0;
2826}
2827
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002828void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002829{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002830 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002831 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2832 * before destroying the slab cache.
2833 */
2834 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002835 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002836 kmem_cache_destroy(ext4_ac_cachep);
Bobi Jam18aadd42012-02-20 17:53:02 -05002837 kmem_cache_destroy(ext4_free_data_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002838 ext4_groupinfo_destroy_slabs();
Alex Tomasc9de5602008-01-29 00:19:52 -05002839}
2840
2841
2842/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002843 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002844 * Returns 0 if success or error code
2845 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002846static noinline_for_stack int
2847ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002848 handle_t *handle, unsigned int reserv_clstrs)
Alex Tomasc9de5602008-01-29 00:19:52 -05002849{
2850 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002851 struct ext4_group_desc *gdp;
2852 struct buffer_head *gdp_bh;
2853 struct ext4_sb_info *sbi;
2854 struct super_block *sb;
2855 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002856 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002857
2858 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2859 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2860
2861 sb = ac->ac_sb;
2862 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002863
2864 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002865 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002866 if (!bitmap_bh)
2867 goto out_err;
2868
2869 err = ext4_journal_get_write_access(handle, bitmap_bh);
2870 if (err)
2871 goto out_err;
2872
2873 err = -EIO;
2874 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2875 if (!gdp)
2876 goto out_err;
2877
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002878 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002879 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002880
Alex Tomasc9de5602008-01-29 00:19:52 -05002881 err = ext4_journal_get_write_access(handle, gdp_bh);
2882 if (err)
2883 goto out_err;
2884
Akinobu Mitabda00de2010-03-03 23:53:25 -05002885 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002886
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002887 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002888 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002889 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002890 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002891 /* File system mounted not to panic on error
2892 * Fix the bitmap and repeat the block allocation
2893 * We leak some of the blocks here.
2894 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002895 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002896 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2897 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002898 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002899 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002900 if (!err)
2901 err = -EAGAIN;
2902 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002903 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002904
2905 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002906#ifdef AGGRESSIVE_CHECK
2907 {
2908 int i;
2909 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2910 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2911 bitmap_bh->b_data));
2912 }
2913 }
2914#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002915 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2916 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002917 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2918 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002919 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002920 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002921 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002922 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002923 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2924 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002925 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002926 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002927
2928 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002929 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002930 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002931 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002932 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002933 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2934 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002935 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2936 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002937
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002938 if (sbi->s_log_groups_per_flex) {
2939 ext4_group_t flex_group = ext4_flex_group(sbi,
2940 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002941 atomic64_sub(ac->ac_b_ex.fe_len,
2942 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002943 }
2944
Frank Mayhar03901312009-01-07 00:06:22 -05002945 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002946 if (err)
2947 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002948 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002949
2950out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002951 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002952 return err;
2953}
2954
2955/*
2956 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002957 * Group request are normalized to s_mb_group_prealloc, which goes to
2958 * s_strip if we set the same via mount option.
2959 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002960 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002961 *
2962 * XXX: should we try to preallocate more than the group has now?
2963 */
2964static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2965{
2966 struct super_block *sb = ac->ac_sb;
2967 struct ext4_locality_group *lg = ac->ac_lg;
2968
2969 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002970 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002971 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002972 current->pid, ac->ac_g_ex.fe_len);
2973}
2974
2975/*
2976 * Normalization means making request better in terms of
2977 * size and alignment
2978 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002979static noinline_for_stack void
2980ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05002981 struct ext4_allocation_request *ar)
2982{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002983 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002984 int bsbits, max;
2985 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05002986 loff_t size, start_off;
2987 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04002988 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05002989 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002990 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05002991
2992 /* do normalize only data requests, metadata requests
2993 do not need preallocation */
2994 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2995 return;
2996
2997 /* sometime caller may want exact blocks */
2998 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2999 return;
3000
3001 /* caller may indicate that preallocation isn't
3002 * required (it's a tail, for example) */
3003 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3004 return;
3005
3006 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3007 ext4_mb_normalize_group_request(ac);
3008 return ;
3009 }
3010
3011 bsbits = ac->ac_sb->s_blocksize_bits;
3012
3013 /* first, let's learn actual file size
3014 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003015 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003016 size = size << bsbits;
3017 if (size < i_size_read(ac->ac_inode))
3018 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003019 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003020
Valerie Clement19304792008-05-13 19:31:14 -04003021 /* max size of free chunks */
3022 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003023
Valerie Clement19304792008-05-13 19:31:14 -04003024#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3025 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003026
3027 /* first, try to predict filesize */
3028 /* XXX: should this table be tunable? */
3029 start_off = 0;
3030 if (size <= 16 * 1024) {
3031 size = 16 * 1024;
3032 } else if (size <= 32 * 1024) {
3033 size = 32 * 1024;
3034 } else if (size <= 64 * 1024) {
3035 size = 64 * 1024;
3036 } else if (size <= 128 * 1024) {
3037 size = 128 * 1024;
3038 } else if (size <= 256 * 1024) {
3039 size = 256 * 1024;
3040 } else if (size <= 512 * 1024) {
3041 size = 512 * 1024;
3042 } else if (size <= 1024 * 1024) {
3043 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003044 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003045 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003046 (21 - bsbits)) << 21;
3047 size = 2 * 1024 * 1024;
3048 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003049 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3050 (22 - bsbits)) << 22;
3051 size = 4 * 1024 * 1024;
3052 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003053 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003054 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3055 (23 - bsbits)) << 23;
3056 size = 8 * 1024 * 1024;
3057 } else {
3058 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3059 size = ac->ac_o_ex.fe_len << bsbits;
3060 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003061 size = size >> bsbits;
3062 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003063
3064 /* don't cover already allocated blocks in selected range */
3065 if (ar->pleft && start <= ar->lleft) {
3066 size -= ar->lleft + 1 - start;
3067 start = ar->lleft + 1;
3068 }
3069 if (ar->pright && start + size - 1 >= ar->lright)
3070 size -= start + size - ar->lright;
3071
3072 end = start + size;
3073
3074 /* check we don't cross already preallocated blocks */
3075 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003076 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003077 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003078
Alex Tomasc9de5602008-01-29 00:19:52 -05003079 if (pa->pa_deleted)
3080 continue;
3081 spin_lock(&pa->pa_lock);
3082 if (pa->pa_deleted) {
3083 spin_unlock(&pa->pa_lock);
3084 continue;
3085 }
3086
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003087 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3088 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003089
3090 /* PA must not overlap original request */
3091 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3092 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3093
Eric Sandeen38877f42009-08-17 23:55:24 -04003094 /* skip PAs this normalized request doesn't overlap with */
3095 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003096 spin_unlock(&pa->pa_lock);
3097 continue;
3098 }
3099 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3100
Eric Sandeen38877f42009-08-17 23:55:24 -04003101 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003102 if (pa_end <= ac->ac_o_ex.fe_logical) {
3103 BUG_ON(pa_end < start);
3104 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003105 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003106 BUG_ON(pa->pa_lstart > end);
3107 end = pa->pa_lstart;
3108 }
3109 spin_unlock(&pa->pa_lock);
3110 }
3111 rcu_read_unlock();
3112 size = end - start;
3113
3114 /* XXX: extra loop to check we really don't overlap preallocations */
3115 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003116 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003117 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003118
Alex Tomasc9de5602008-01-29 00:19:52 -05003119 spin_lock(&pa->pa_lock);
3120 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003121 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3122 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003123 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3124 }
3125 spin_unlock(&pa->pa_lock);
3126 }
3127 rcu_read_unlock();
3128
3129 if (start + size <= ac->ac_o_ex.fe_logical &&
3130 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003131 ext4_msg(ac->ac_sb, KERN_ERR,
3132 "start %lu, size %lu, fe_logical %lu",
3133 (unsigned long) start, (unsigned long) size,
3134 (unsigned long) ac->ac_o_ex.fe_logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05003135 }
3136 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3137 start > ac->ac_o_ex.fe_logical);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003138 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003139
3140 /* now prepare goal request */
3141
3142 /* XXX: is it better to align blocks WRT to logical
3143 * placement or satisfy big request as is */
3144 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003145 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003146
3147 /* define goal start in order to merge */
3148 if (ar->pright && (ar->lright == (start + size))) {
3149 /* merge to the right */
3150 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3151 &ac->ac_f_ex.fe_group,
3152 &ac->ac_f_ex.fe_start);
3153 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3154 }
3155 if (ar->pleft && (ar->lleft + 1 == start)) {
3156 /* merge to the left */
3157 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3158 &ac->ac_f_ex.fe_group,
3159 &ac->ac_f_ex.fe_start);
3160 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3161 }
3162
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003163 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003164 (unsigned) orig_size, (unsigned) start);
3165}
3166
3167static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3168{
3169 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3170
3171 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3172 atomic_inc(&sbi->s_bal_reqs);
3173 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003174 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003175 atomic_inc(&sbi->s_bal_success);
3176 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3177 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3178 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3179 atomic_inc(&sbi->s_bal_goals);
3180 if (ac->ac_found > sbi->s_mb_max_to_scan)
3181 atomic_inc(&sbi->s_bal_breaks);
3182 }
3183
Theodore Ts'o296c3552009-09-30 00:32:42 -04003184 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3185 trace_ext4_mballoc_alloc(ac);
3186 else
3187 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003188}
3189
3190/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003191 * Called on failure; free up any blocks from the inode PA for this
3192 * context. We don't need this for MB_GROUP_PA because we only change
3193 * pa_free in ext4_mb_release_context(), but on failure, we've already
3194 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3195 */
3196static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3197{
3198 struct ext4_prealloc_space *pa = ac->ac_pa;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003199
Zheng Liu400db9d2012-05-28 17:53:53 -04003200 if (pa && pa->pa_type == MB_INODE_PA)
3201 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003202}
3203
3204/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003205 * use blocks preallocated to inode
3206 */
3207static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3208 struct ext4_prealloc_space *pa)
3209{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003210 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003211 ext4_fsblk_t start;
3212 ext4_fsblk_t end;
3213 int len;
3214
3215 /* found preallocated blocks, use them */
3216 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003217 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3218 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3219 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003220 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3221 &ac->ac_b_ex.fe_start);
3222 ac->ac_b_ex.fe_len = len;
3223 ac->ac_status = AC_STATUS_FOUND;
3224 ac->ac_pa = pa;
3225
3226 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003227 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003228 BUG_ON(pa->pa_free < len);
3229 pa->pa_free -= len;
3230
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003231 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003232}
3233
3234/*
3235 * use blocks preallocated to locality group
3236 */
3237static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3238 struct ext4_prealloc_space *pa)
3239{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003240 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003241
Alex Tomasc9de5602008-01-29 00:19:52 -05003242 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3243 &ac->ac_b_ex.fe_group,
3244 &ac->ac_b_ex.fe_start);
3245 ac->ac_b_ex.fe_len = len;
3246 ac->ac_status = AC_STATUS_FOUND;
3247 ac->ac_pa = pa;
3248
3249 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003250 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003251 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003252 * in on-disk bitmap -- see ext4_mb_release_context()
3253 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003254 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003255 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003256}
3257
3258/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003259 * Return the prealloc space that have minimal distance
3260 * from the goal block. @cpa is the prealloc
3261 * space that is having currently known minimal distance
3262 * from the goal block.
3263 */
3264static struct ext4_prealloc_space *
3265ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3266 struct ext4_prealloc_space *pa,
3267 struct ext4_prealloc_space *cpa)
3268{
3269 ext4_fsblk_t cur_distance, new_distance;
3270
3271 if (cpa == NULL) {
3272 atomic_inc(&pa->pa_count);
3273 return pa;
3274 }
3275 cur_distance = abs(goal_block - cpa->pa_pstart);
3276 new_distance = abs(goal_block - pa->pa_pstart);
3277
Coly Li5a54b2f2011-02-24 14:10:05 -05003278 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003279 return cpa;
3280
3281 /* drop the previous reference */
3282 atomic_dec(&cpa->pa_count);
3283 atomic_inc(&pa->pa_count);
3284 return pa;
3285}
3286
3287/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003288 * search goal blocks in preallocated space
3289 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003290static noinline_for_stack int
3291ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003292{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003293 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003294 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003295 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3296 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003297 struct ext4_prealloc_space *pa, *cpa = NULL;
3298 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003299
3300 /* only data can be preallocated */
3301 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3302 return 0;
3303
3304 /* first, try per-file preallocation */
3305 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003306 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003307
3308 /* all fields in this condition don't change,
3309 * so we can skip locking for them */
3310 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003311 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3312 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003313 continue;
3314
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003315 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003316 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003317 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3318 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003319 continue;
3320
Alex Tomasc9de5602008-01-29 00:19:52 -05003321 /* found preallocated blocks, use them */
3322 spin_lock(&pa->pa_lock);
3323 if (pa->pa_deleted == 0 && pa->pa_free) {
3324 atomic_inc(&pa->pa_count);
3325 ext4_mb_use_inode_pa(ac, pa);
3326 spin_unlock(&pa->pa_lock);
3327 ac->ac_criteria = 10;
3328 rcu_read_unlock();
3329 return 1;
3330 }
3331 spin_unlock(&pa->pa_lock);
3332 }
3333 rcu_read_unlock();
3334
3335 /* can we use group allocation? */
3336 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3337 return 0;
3338
3339 /* inode may have no locality group for some reason */
3340 lg = ac->ac_lg;
3341 if (lg == NULL)
3342 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003343 order = fls(ac->ac_o_ex.fe_len) - 1;
3344 if (order > PREALLOC_TB_SIZE - 1)
3345 /* The max size of hash table is PREALLOC_TB_SIZE */
3346 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003347
Akinobu Mitabda00de2010-03-03 23:53:25 -05003348 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003349 /*
3350 * search for the prealloc space that is having
3351 * minimal distance from the goal block.
3352 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003353 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3354 rcu_read_lock();
3355 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3356 pa_inode_list) {
3357 spin_lock(&pa->pa_lock);
3358 if (pa->pa_deleted == 0 &&
3359 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003360
3361 cpa = ext4_mb_check_group_pa(goal_block,
3362 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003363 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003364 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003365 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003366 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003367 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003368 if (cpa) {
3369 ext4_mb_use_group_pa(ac, cpa);
3370 ac->ac_criteria = 20;
3371 return 1;
3372 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003373 return 0;
3374}
3375
3376/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003377 * the function goes through all block freed in the group
3378 * but not yet committed and marks them used in in-core bitmap.
3379 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003380 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003381 */
3382static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3383 ext4_group_t group)
3384{
3385 struct rb_node *n;
3386 struct ext4_group_info *grp;
3387 struct ext4_free_data *entry;
3388
3389 grp = ext4_get_group_info(sb, group);
3390 n = rb_first(&(grp->bb_free_root));
3391
3392 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003393 entry = rb_entry(n, struct ext4_free_data, efd_node);
3394 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003395 n = rb_next(n);
3396 }
3397 return;
3398}
3399
3400/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003401 * the function goes through all preallocation in this group and marks them
3402 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003403 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003404 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003405static noinline_for_stack
3406void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003407 ext4_group_t group)
3408{
3409 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3410 struct ext4_prealloc_space *pa;
3411 struct list_head *cur;
3412 ext4_group_t groupnr;
3413 ext4_grpblk_t start;
3414 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003415 int len;
3416
3417 /* all form of preallocation discards first load group,
3418 * so the only competing code is preallocation use.
3419 * we don't need any locking here
3420 * notice we do NOT ignore preallocations with pa_deleted
3421 * otherwise we could leave used blocks available for
3422 * allocation in buddy when concurrent ext4_mb_put_pa()
3423 * is dropping preallocation
3424 */
3425 list_for_each(cur, &grp->bb_prealloc_list) {
3426 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3427 spin_lock(&pa->pa_lock);
3428 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3429 &groupnr, &start);
3430 len = pa->pa_len;
3431 spin_unlock(&pa->pa_lock);
3432 if (unlikely(len == 0))
3433 continue;
3434 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003435 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003436 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003437 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003438 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003439}
3440
3441static void ext4_mb_pa_callback(struct rcu_head *head)
3442{
3443 struct ext4_prealloc_space *pa;
3444 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003445
3446 BUG_ON(atomic_read(&pa->pa_count));
3447 BUG_ON(pa->pa_deleted == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05003448 kmem_cache_free(ext4_pspace_cachep, pa);
3449}
3450
3451/*
3452 * drops a reference to preallocated space descriptor
3453 * if this was the last reference and the space is consumed
3454 */
3455static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3456 struct super_block *sb, struct ext4_prealloc_space *pa)
3457{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003458 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003459 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003460
Alex Tomasc9de5602008-01-29 00:19:52 -05003461 /* in this short window concurrent discard can set pa_deleted */
3462 spin_lock(&pa->pa_lock);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003463 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3464 spin_unlock(&pa->pa_lock);
3465 return;
3466 }
3467
Alex Tomasc9de5602008-01-29 00:19:52 -05003468 if (pa->pa_deleted == 1) {
3469 spin_unlock(&pa->pa_lock);
3470 return;
3471 }
3472
3473 pa->pa_deleted = 1;
3474 spin_unlock(&pa->pa_lock);
3475
Eric Sandeend33a1972009-03-16 23:25:40 -04003476 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003477 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003478 * If doing group-based preallocation, pa_pstart may be in the
3479 * next group when pa is used up
3480 */
3481 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003482 grp_blk--;
3483
Lukas Czernerbd862982013-04-03 23:32:34 -04003484 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003485
3486 /*
3487 * possible race:
3488 *
3489 * P1 (buddy init) P2 (regular allocation)
3490 * find block B in PA
3491 * copy on-disk bitmap to buddy
3492 * mark B in on-disk bitmap
3493 * drop PA from group
3494 * mark all PAs in buddy
3495 *
3496 * thus, P1 initializes buddy with B available. to prevent this
3497 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3498 * against that pair
3499 */
3500 ext4_lock_group(sb, grp);
3501 list_del(&pa->pa_group_list);
3502 ext4_unlock_group(sb, grp);
3503
3504 spin_lock(pa->pa_obj_lock);
3505 list_del_rcu(&pa->pa_inode_list);
3506 spin_unlock(pa->pa_obj_lock);
3507
3508 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3509}
3510
3511/*
3512 * creates new preallocated space for given inode
3513 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003514static noinline_for_stack int
3515ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003516{
3517 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003518 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003519 struct ext4_prealloc_space *pa;
3520 struct ext4_group_info *grp;
3521 struct ext4_inode_info *ei;
3522
3523 /* preallocate only when found space is larger then requested */
3524 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3525 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3526 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3527
3528 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3529 if (pa == NULL)
3530 return -ENOMEM;
3531
3532 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3533 int winl;
3534 int wins;
3535 int win;
3536 int offs;
3537
3538 /* we can't allocate as much as normalizer wants.
3539 * so, found space must get proper lstart
3540 * to cover original request */
3541 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3542 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3543
3544 /* we're limited by original request in that
3545 * logical block must be covered any way
3546 * winl is window we can move our chunk within */
3547 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3548
3549 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003550 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003551
3552 /* the smallest one defines real window */
3553 win = min(winl, wins);
3554
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003555 offs = ac->ac_o_ex.fe_logical %
3556 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003557 if (offs && offs < win)
3558 win = offs;
3559
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003560 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003561 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003562 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3563 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3564 }
3565
3566 /* preallocation can change ac_b_ex, thus we store actually
3567 * allocated blocks for history */
3568 ac->ac_f_ex = ac->ac_b_ex;
3569
3570 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3571 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3572 pa->pa_len = ac->ac_b_ex.fe_len;
3573 pa->pa_free = pa->pa_len;
3574 atomic_set(&pa->pa_count, 1);
3575 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003576 INIT_LIST_HEAD(&pa->pa_inode_list);
3577 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003578 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003579 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003580
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003581 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003582 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003583 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003584
3585 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003586 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003587
3588 ei = EXT4_I(ac->ac_inode);
3589 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3590
3591 pa->pa_obj_lock = &ei->i_prealloc_lock;
3592 pa->pa_inode = ac->ac_inode;
3593
3594 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3595 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3596 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3597
3598 spin_lock(pa->pa_obj_lock);
3599 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3600 spin_unlock(pa->pa_obj_lock);
3601
3602 return 0;
3603}
3604
3605/*
3606 * creates new preallocated space for locality group inodes belongs to
3607 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003608static noinline_for_stack int
3609ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003610{
3611 struct super_block *sb = ac->ac_sb;
3612 struct ext4_locality_group *lg;
3613 struct ext4_prealloc_space *pa;
3614 struct ext4_group_info *grp;
3615
3616 /* preallocate only when found space is larger then requested */
3617 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3618 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3619 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3620
3621 BUG_ON(ext4_pspace_cachep == NULL);
3622 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3623 if (pa == NULL)
3624 return -ENOMEM;
3625
3626 /* preallocation can change ac_b_ex, thus we store actually
3627 * allocated blocks for history */
3628 ac->ac_f_ex = ac->ac_b_ex;
3629
3630 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3631 pa->pa_lstart = pa->pa_pstart;
3632 pa->pa_len = ac->ac_b_ex.fe_len;
3633 pa->pa_free = pa->pa_len;
3634 atomic_set(&pa->pa_count, 1);
3635 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003636 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003637 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003638 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003639 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003640
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003641 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003642 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3643 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003644
3645 ext4_mb_use_group_pa(ac, pa);
3646 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3647
3648 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3649 lg = ac->ac_lg;
3650 BUG_ON(lg == NULL);
3651
3652 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3653 pa->pa_inode = NULL;
3654
3655 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3656 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3657 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3658
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003659 /*
3660 * We will later add the new pa to the right bucket
3661 * after updating the pa_free in ext4_mb_release_context
3662 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003663 return 0;
3664}
3665
3666static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3667{
3668 int err;
3669
3670 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3671 err = ext4_mb_new_group_pa(ac);
3672 else
3673 err = ext4_mb_new_inode_pa(ac);
3674 return err;
3675}
3676
3677/*
3678 * finds all unused blocks in on-disk bitmap, frees them in
3679 * in-core bitmap and buddy.
3680 * @pa must be unlinked from inode and group lists, so that
3681 * nobody else can find/use it.
3682 * the caller MUST hold group/inode locks.
3683 * TODO: optimize the case when there are no in-core structures yet
3684 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003685static noinline_for_stack int
3686ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003687 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003688{
Alex Tomasc9de5602008-01-29 00:19:52 -05003689 struct super_block *sb = e4b->bd_sb;
3690 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003691 unsigned int end;
3692 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003693 ext4_group_t group;
3694 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003695 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003696 int err = 0;
3697 int free = 0;
3698
3699 BUG_ON(pa->pa_deleted == 0);
3700 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003701 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003702 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3703 end = bit + pa->pa_len;
3704
Alex Tomasc9de5602008-01-29 00:19:52 -05003705 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003706 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003707 if (bit >= end)
3708 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003709 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003710 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003711 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3712 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003713 free += next - bit;
3714
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003715 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003716 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3717 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003718 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003719 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3720 bit = next + 1;
3721 }
3722 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003723 ext4_msg(e4b->bd_sb, KERN_CRIT,
3724 "pa %p: logic %lu, phys. %lu, len %lu",
3725 pa, (unsigned long) pa->pa_lstart,
3726 (unsigned long) pa->pa_pstart,
3727 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003728 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003729 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003730 /*
3731 * pa is already deleted so we use the value obtained
3732 * from the bitmap and continue.
3733 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003734 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003735 atomic_add(free, &sbi->s_mb_discarded);
3736
3737 return err;
3738}
3739
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003740static noinline_for_stack int
3741ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003742 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003743{
Alex Tomasc9de5602008-01-29 00:19:52 -05003744 struct super_block *sb = e4b->bd_sb;
3745 ext4_group_t group;
3746 ext4_grpblk_t bit;
3747
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003748 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003749 BUG_ON(pa->pa_deleted == 0);
3750 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3751 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3752 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3753 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003754 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003755
3756 return 0;
3757}
3758
3759/*
3760 * releases all preallocations in given group
3761 *
3762 * first, we need to decide discard policy:
3763 * - when do we discard
3764 * 1) ENOSPC
3765 * - how many do we discard
3766 * 1) how many requested
3767 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003768static noinline_for_stack int
3769ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003770 ext4_group_t group, int needed)
3771{
3772 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3773 struct buffer_head *bitmap_bh = NULL;
3774 struct ext4_prealloc_space *pa, *tmp;
3775 struct list_head list;
3776 struct ext4_buddy e4b;
3777 int err;
3778 int busy = 0;
3779 int free = 0;
3780
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003781 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003782
3783 if (list_empty(&grp->bb_prealloc_list))
3784 return 0;
3785
Theodore Ts'o574ca172008-07-11 19:27:31 -04003786 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003787 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003788 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003789 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003790 }
3791
3792 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003793 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003794 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003795 put_bh(bitmap_bh);
3796 return 0;
3797 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003798
3799 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003800 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003801
Alex Tomasc9de5602008-01-29 00:19:52 -05003802 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003803repeat:
3804 ext4_lock_group(sb, group);
3805 list_for_each_entry_safe(pa, tmp,
3806 &grp->bb_prealloc_list, pa_group_list) {
3807 spin_lock(&pa->pa_lock);
3808 if (atomic_read(&pa->pa_count)) {
3809 spin_unlock(&pa->pa_lock);
3810 busy = 1;
3811 continue;
3812 }
3813 if (pa->pa_deleted) {
3814 spin_unlock(&pa->pa_lock);
3815 continue;
3816 }
3817
3818 /* seems this one can be freed ... */
3819 pa->pa_deleted = 1;
3820
3821 /* we can trust pa_free ... */
3822 free += pa->pa_free;
3823
3824 spin_unlock(&pa->pa_lock);
3825
3826 list_del(&pa->pa_group_list);
3827 list_add(&pa->u.pa_tmp_list, &list);
3828 }
3829
3830 /* if we still need more blocks and some PAs were used, try again */
3831 if (free < needed && busy) {
3832 busy = 0;
3833 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003834 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003835 goto repeat;
3836 }
3837
3838 /* found anything to free? */
3839 if (list_empty(&list)) {
3840 BUG_ON(free != 0);
3841 goto out;
3842 }
3843
3844 /* now free all selected PAs */
3845 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3846
3847 /* remove from object (inode or locality group) */
3848 spin_lock(pa->pa_obj_lock);
3849 list_del_rcu(&pa->pa_inode_list);
3850 spin_unlock(pa->pa_obj_lock);
3851
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003852 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003853 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003854 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003855 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003856
3857 list_del(&pa->u.pa_tmp_list);
3858 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3859 }
3860
3861out:
3862 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003863 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003864 put_bh(bitmap_bh);
3865 return free;
3866}
3867
3868/*
3869 * releases all non-used preallocated blocks for given inode
3870 *
3871 * It's important to discard preallocations under i_data_sem
3872 * We don't want another block to be served from the prealloc
3873 * space when we are discarding the inode prealloc space.
3874 *
3875 * FIXME!! Make sure it is valid at all the call sites
3876 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003877void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003878{
3879 struct ext4_inode_info *ei = EXT4_I(inode);
3880 struct super_block *sb = inode->i_sb;
3881 struct buffer_head *bitmap_bh = NULL;
3882 struct ext4_prealloc_space *pa, *tmp;
3883 ext4_group_t group = 0;
3884 struct list_head list;
3885 struct ext4_buddy e4b;
3886 int err;
3887
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003888 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003889 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3890 return;
3891 }
3892
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003893 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003894 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003895
3896 INIT_LIST_HEAD(&list);
3897
3898repeat:
3899 /* first, collect all pa's in the inode */
3900 spin_lock(&ei->i_prealloc_lock);
3901 while (!list_empty(&ei->i_prealloc_list)) {
3902 pa = list_entry(ei->i_prealloc_list.next,
3903 struct ext4_prealloc_space, pa_inode_list);
3904 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3905 spin_lock(&pa->pa_lock);
3906 if (atomic_read(&pa->pa_count)) {
3907 /* this shouldn't happen often - nobody should
3908 * use preallocation while we're discarding it */
3909 spin_unlock(&pa->pa_lock);
3910 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003911 ext4_msg(sb, KERN_ERR,
3912 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003913 WARN_ON(1);
3914 schedule_timeout_uninterruptible(HZ);
3915 goto repeat;
3916
3917 }
3918 if (pa->pa_deleted == 0) {
3919 pa->pa_deleted = 1;
3920 spin_unlock(&pa->pa_lock);
3921 list_del_rcu(&pa->pa_inode_list);
3922 list_add(&pa->u.pa_tmp_list, &list);
3923 continue;
3924 }
3925
3926 /* someone is deleting pa right now */
3927 spin_unlock(&pa->pa_lock);
3928 spin_unlock(&ei->i_prealloc_lock);
3929
3930 /* we have to wait here because pa_deleted
3931 * doesn't mean pa is already unlinked from
3932 * the list. as we might be called from
3933 * ->clear_inode() the inode will get freed
3934 * and concurrent thread which is unlinking
3935 * pa from inode's list may access already
3936 * freed memory, bad-bad-bad */
3937
3938 /* XXX: if this happens too often, we can
3939 * add a flag to force wait only in case
3940 * of ->clear_inode(), but not in case of
3941 * regular truncate */
3942 schedule_timeout_uninterruptible(HZ);
3943 goto repeat;
3944 }
3945 spin_unlock(&ei->i_prealloc_lock);
3946
3947 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003948 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04003949 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05003950
3951 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003952 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003953 ext4_error(sb, "Error loading buddy information for %u",
3954 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003955 continue;
3956 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003957
Theodore Ts'o574ca172008-07-11 19:27:31 -04003958 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003959 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003960 ext4_error(sb, "Error reading block bitmap for %u",
3961 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003962 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003963 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05003964 }
3965
3966 ext4_lock_group(sb, group);
3967 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003968 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003969 ext4_unlock_group(sb, group);
3970
Jing Zhange39e07f2010-05-14 00:00:00 -04003971 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003972 put_bh(bitmap_bh);
3973
3974 list_del(&pa->u.pa_tmp_list);
3975 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3976 }
3977}
3978
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003979#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05003980static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3981{
3982 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04003983 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003984
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05003985 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05003986 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04003987 return;
3988
Joe Perches7f6a11e2012-03-19 23:09:43 -04003989 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003990 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04003991 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003992 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04003993 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003994 "goal %lu/%lu/%lu@%lu, "
3995 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003996 (unsigned long)ac->ac_o_ex.fe_group,
3997 (unsigned long)ac->ac_o_ex.fe_start,
3998 (unsigned long)ac->ac_o_ex.fe_len,
3999 (unsigned long)ac->ac_o_ex.fe_logical,
4000 (unsigned long)ac->ac_g_ex.fe_group,
4001 (unsigned long)ac->ac_g_ex.fe_start,
4002 (unsigned long)ac->ac_g_ex.fe_len,
4003 (unsigned long)ac->ac_g_ex.fe_logical,
4004 (unsigned long)ac->ac_b_ex.fe_group,
4005 (unsigned long)ac->ac_b_ex.fe_start,
4006 (unsigned long)ac->ac_b_ex.fe_len,
4007 (unsigned long)ac->ac_b_ex.fe_logical,
4008 (int)ac->ac_criteria);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004009 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004010 ac->ac_ex_scanned, ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004011 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04004012 ngroups = ext4_get_groups_count(sb);
4013 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004014 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4015 struct ext4_prealloc_space *pa;
4016 ext4_grpblk_t start;
4017 struct list_head *cur;
4018 ext4_lock_group(sb, i);
4019 list_for_each(cur, &grp->bb_prealloc_list) {
4020 pa = list_entry(cur, struct ext4_prealloc_space,
4021 pa_group_list);
4022 spin_lock(&pa->pa_lock);
4023 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4024 NULL, &start);
4025 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004026 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4027 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004028 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004029 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004030
4031 if (grp->bb_free == 0)
4032 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004033 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004034 i, grp->bb_free, grp->bb_fragments);
4035 }
4036 printk(KERN_ERR "\n");
4037}
4038#else
4039static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4040{
4041 return;
4042}
4043#endif
4044
4045/*
4046 * We use locality group preallocation for small size file. The size of the
4047 * file is determined by the current size or the resulting size after
4048 * allocation which ever is larger
4049 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004050 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004051 */
4052static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4053{
4054 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4055 int bsbits = ac->ac_sb->s_blocksize_bits;
4056 loff_t size, isize;
4057
4058 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4059 return;
4060
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004061 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4062 return;
4063
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004064 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004065 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4066 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004067
Theodore Ts'o50797482009-09-18 13:34:02 -04004068 if ((size == isize) &&
4069 !ext4_fs_is_busy(sbi) &&
4070 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4071 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4072 return;
4073 }
4074
Robin Dongebbe0272011-10-26 05:14:27 -04004075 if (sbi->s_mb_group_prealloc <= 0) {
4076 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4077 return;
4078 }
4079
Alex Tomasc9de5602008-01-29 00:19:52 -05004080 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004081 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004082 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004083 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004084 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004085 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004086
4087 BUG_ON(ac->ac_lg != NULL);
4088 /*
4089 * locality group prealloc space are per cpu. The reason for having
4090 * per cpu locality group is to reduce the contention between block
4091 * request from multiple CPUs.
4092 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09004093 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004094
4095 /* we're going to use group allocation */
4096 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4097
4098 /* serialize all allocations in the group */
4099 mutex_lock(&ac->ac_lg->lg_mutex);
4100}
4101
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004102static noinline_for_stack int
4103ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004104 struct ext4_allocation_request *ar)
4105{
4106 struct super_block *sb = ar->inode->i_sb;
4107 struct ext4_sb_info *sbi = EXT4_SB(sb);
4108 struct ext4_super_block *es = sbi->s_es;
4109 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004110 unsigned int len;
4111 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004112 ext4_grpblk_t block;
4113
4114 /* we can't allocate > group size */
4115 len = ar->len;
4116
4117 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004118 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4119 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004120
4121 /* start searching from the goal */
4122 goal = ar->goal;
4123 if (goal < le32_to_cpu(es->s_first_data_block) ||
4124 goal >= ext4_blocks_count(es))
4125 goal = le32_to_cpu(es->s_first_data_block);
4126 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4127
4128 /* set up allocation goals */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004129 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05004130 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004131 ac->ac_sb = sb;
4132 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004133 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004134 ac->ac_o_ex.fe_group = group;
4135 ac->ac_o_ex.fe_start = block;
4136 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004137 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004138 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004139
4140 /* we have to define context: we'll we work with a file or
4141 * locality group. this is a policy, actually */
4142 ext4_mb_group_or_file(ac);
4143
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004144 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004145 "left: %u/%u, right %u/%u to %swritable\n",
4146 (unsigned) ar->len, (unsigned) ar->logical,
4147 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4148 (unsigned) ar->lleft, (unsigned) ar->pleft,
4149 (unsigned) ar->lright, (unsigned) ar->pright,
4150 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4151 return 0;
4152
4153}
4154
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004155static noinline_for_stack void
4156ext4_mb_discard_lg_preallocations(struct super_block *sb,
4157 struct ext4_locality_group *lg,
4158 int order, int total_entries)
4159{
4160 ext4_group_t group = 0;
4161 struct ext4_buddy e4b;
4162 struct list_head discard_list;
4163 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004164
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004165 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004166
4167 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004168
4169 spin_lock(&lg->lg_prealloc_lock);
4170 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4171 pa_inode_list) {
4172 spin_lock(&pa->pa_lock);
4173 if (atomic_read(&pa->pa_count)) {
4174 /*
4175 * This is the pa that we just used
4176 * for block allocation. So don't
4177 * free that
4178 */
4179 spin_unlock(&pa->pa_lock);
4180 continue;
4181 }
4182 if (pa->pa_deleted) {
4183 spin_unlock(&pa->pa_lock);
4184 continue;
4185 }
4186 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004187 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004188
4189 /* seems this one can be freed ... */
4190 pa->pa_deleted = 1;
4191 spin_unlock(&pa->pa_lock);
4192
4193 list_del_rcu(&pa->pa_inode_list);
4194 list_add(&pa->u.pa_tmp_list, &discard_list);
4195
4196 total_entries--;
4197 if (total_entries <= 5) {
4198 /*
4199 * we want to keep only 5 entries
4200 * allowing it to grow to 8. This
4201 * mak sure we don't call discard
4202 * soon for this list.
4203 */
4204 break;
4205 }
4206 }
4207 spin_unlock(&lg->lg_prealloc_lock);
4208
4209 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4210
Lukas Czernerbd862982013-04-03 23:32:34 -04004211 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004212 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004213 ext4_error(sb, "Error loading buddy information for %u",
4214 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004215 continue;
4216 }
4217 ext4_lock_group(sb, group);
4218 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004219 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004220 ext4_unlock_group(sb, group);
4221
Jing Zhange39e07f2010-05-14 00:00:00 -04004222 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004223 list_del(&pa->u.pa_tmp_list);
4224 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4225 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004226}
4227
4228/*
4229 * We have incremented pa_count. So it cannot be freed at this
4230 * point. Also we hold lg_mutex. So no parallel allocation is
4231 * possible from this lg. That means pa_free cannot be updated.
4232 *
4233 * A parallel ext4_mb_discard_group_preallocations is possible.
4234 * which can cause the lg_prealloc_list to be updated.
4235 */
4236
4237static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4238{
4239 int order, added = 0, lg_prealloc_count = 1;
4240 struct super_block *sb = ac->ac_sb;
4241 struct ext4_locality_group *lg = ac->ac_lg;
4242 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4243
4244 order = fls(pa->pa_free) - 1;
4245 if (order > PREALLOC_TB_SIZE - 1)
4246 /* The max size of hash table is PREALLOC_TB_SIZE */
4247 order = PREALLOC_TB_SIZE - 1;
4248 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004249 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004250 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4251 pa_inode_list) {
4252 spin_lock(&tmp_pa->pa_lock);
4253 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004254 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004255 continue;
4256 }
4257 if (!added && pa->pa_free < tmp_pa->pa_free) {
4258 /* Add to the tail of the previous entry */
4259 list_add_tail_rcu(&pa->pa_inode_list,
4260 &tmp_pa->pa_inode_list);
4261 added = 1;
4262 /*
4263 * we want to count the total
4264 * number of entries in the list
4265 */
4266 }
4267 spin_unlock(&tmp_pa->pa_lock);
4268 lg_prealloc_count++;
4269 }
4270 if (!added)
4271 list_add_tail_rcu(&pa->pa_inode_list,
4272 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004273 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004274
4275 /* Now trim the list to be not more than 8 elements */
4276 if (lg_prealloc_count > 8) {
4277 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004278 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004279 return;
4280 }
4281 return ;
4282}
4283
Alex Tomasc9de5602008-01-29 00:19:52 -05004284/*
4285 * release all resource we used in allocation
4286 */
4287static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4288{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004289 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004290 struct ext4_prealloc_space *pa = ac->ac_pa;
4291 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004292 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004293 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004294 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004295 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4296 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004297 pa->pa_free -= ac->ac_b_ex.fe_len;
4298 pa->pa_len -= ac->ac_b_ex.fe_len;
4299 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004300 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004301 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004302 if (pa) {
4303 /*
4304 * We want to add the pa to the right bucket.
4305 * Remove it from the list and while adding
4306 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004307 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004308 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004309 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004310 spin_lock(pa->pa_obj_lock);
4311 list_del_rcu(&pa->pa_inode_list);
4312 spin_unlock(pa->pa_obj_lock);
4313 ext4_mb_add_n_trim(ac);
4314 }
4315 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4316 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004317 if (ac->ac_bitmap_page)
4318 page_cache_release(ac->ac_bitmap_page);
4319 if (ac->ac_buddy_page)
4320 page_cache_release(ac->ac_buddy_page);
4321 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4322 mutex_unlock(&ac->ac_lg->lg_mutex);
4323 ext4_mb_collect_stats(ac);
4324 return 0;
4325}
4326
4327static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4328{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004329 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004330 int ret;
4331 int freed = 0;
4332
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004333 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004334 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004335 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4336 freed += ret;
4337 needed -= ret;
4338 }
4339
4340 return freed;
4341}
4342
4343/*
4344 * Main entry point into mballoc to allocate blocks
4345 * it tries to use preallocation first, then falls back
4346 * to usual allocation
4347 */
4348ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004349 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004350{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004351 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004352 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004353 struct ext4_sb_info *sbi;
4354 struct super_block *sb;
4355 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004356 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004357 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004358
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004359 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004360 sb = ar->inode->i_sb;
4361 sbi = EXT4_SB(sb);
4362
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004363 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004364
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004365 /* Allow to use superuser reservation for quota file */
4366 if (IS_NOQUOTA(ar->inode))
4367 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4368
Mingming Cao60e58e02009-01-22 18:13:05 +01004369 /*
4370 * For delayed allocation, we could skip the ENOSPC and
4371 * EDQUOT check, as blocks and quotas have been already
4372 * reserved when data being copied into pagecache.
4373 */
Theodore Ts'of2321092011-01-10 12:12:36 -05004374 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
Mingming Cao60e58e02009-01-22 18:13:05 +01004375 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4376 else {
4377 /* Without delayed allocation we need to verify
4378 * there is enough free blocks to do block allocation
4379 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004380 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004381 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004382 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004383
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004384 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004385 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004386 ar->len = ar->len >> 1;
4387 }
4388 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004389 *errp = -ENOSPC;
4390 return 0;
4391 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004392 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004393 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004394 dquot_alloc_block_nofail(ar->inode,
4395 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004396 } else {
4397 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004398 dquot_alloc_block(ar->inode,
4399 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004400
4401 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4402 ar->len--;
4403 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004404 }
4405 inquota = ar->len;
4406 if (ar->len == 0) {
4407 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004408 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004409 }
Mingming Caod2a17632008-07-14 17:52:37 -04004410 }
Mingming Caod2a17632008-07-14 17:52:37 -04004411
Wei Yongjun85556c92012-09-26 20:43:37 -04004412 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004413 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004414 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004415 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004416 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004417 }
4418
Eric Sandeen256bdb42008-02-10 01:13:33 -05004419 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004420 if (*errp) {
4421 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004422 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004423 }
4424
Eric Sandeen256bdb42008-02-10 01:13:33 -05004425 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4426 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004427 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4428 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004429repeat:
4430 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004431 *errp = ext4_mb_regular_allocator(ac);
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004432 if (*errp)
4433 goto discard_and_exit;
4434
4435 /* as we've just preallocated more space than
4436 * user requested originally, we store allocated
4437 * space in a special descriptor */
4438 if (ac->ac_status == AC_STATUS_FOUND &&
4439 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4440 *errp = ext4_mb_new_preallocation(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004441 if (*errp) {
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004442 discard_and_exit:
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004443 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004444 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004445 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004446 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004447 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004448 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004449 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004450 /*
4451 * drop the reference that we took
4452 * in ext4_mb_use_best_found
4453 */
4454 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004455 ac->ac_b_ex.fe_group = 0;
4456 ac->ac_b_ex.fe_start = 0;
4457 ac->ac_b_ex.fe_len = 0;
4458 ac->ac_status = AC_STATUS_CONTINUE;
4459 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004460 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004461 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004462 goto errout;
4463 } else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004464 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4465 ar->len = ac->ac_b_ex.fe_len;
4466 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004467 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004468 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004469 if (freed)
4470 goto repeat;
4471 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004472 }
4473
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004474errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004475 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004476 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004477 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004478 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004479 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004480 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004481out:
4482 if (ac)
4483 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004484 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004485 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004486 if (!ar->len) {
Theodore Ts'of2321092011-01-10 12:12:36 -05004487 if (!ext4_test_inode_state(ar->inode,
4488 EXT4_STATE_DELALLOC_RESERVED))
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004489 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004490 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004491 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004492 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004493
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004494 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004495
Alex Tomasc9de5602008-01-29 00:19:52 -05004496 return block;
4497}
Alex Tomasc9de5602008-01-29 00:19:52 -05004498
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004499/*
4500 * We can merge two free data extents only if the physical blocks
4501 * are contiguous, AND the extents were freed by the same transaction,
4502 * AND the blocks are associated with the same group.
4503 */
4504static int can_merge(struct ext4_free_data *entry1,
4505 struct ext4_free_data *entry2)
4506{
Bobi Jam18aadd42012-02-20 17:53:02 -05004507 if ((entry1->efd_tid == entry2->efd_tid) &&
4508 (entry1->efd_group == entry2->efd_group) &&
4509 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004510 return 1;
4511 return 0;
4512}
4513
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004514static noinline_for_stack int
4515ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004516 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004517{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004518 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004519 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004520 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004521 struct ext4_group_info *db = e4b->bd_info;
4522 struct super_block *sb = e4b->bd_sb;
4523 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004524 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4525 struct rb_node *parent = NULL, *new_node;
4526
Frank Mayhar03901312009-01-07 00:06:22 -05004527 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004528 BUG_ON(e4b->bd_bitmap_page == NULL);
4529 BUG_ON(e4b->bd_buddy_page == NULL);
4530
Bobi Jam18aadd42012-02-20 17:53:02 -05004531 new_node = &new_entry->efd_node;
4532 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004533
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004534 if (!*n) {
4535 /* first free block exent. We need to
4536 protect buddy cache from being freed,
4537 * otherwise we'll refresh it from
4538 * on-disk bitmap and lose not-yet-available
4539 * blocks */
4540 page_cache_get(e4b->bd_buddy_page);
4541 page_cache_get(e4b->bd_bitmap_page);
4542 }
4543 while (*n) {
4544 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004545 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4546 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004547 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004548 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004549 n = &(*n)->rb_right;
4550 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004551 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004552 ext4_group_first_block_no(sb, group) +
4553 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004554 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004555 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004556 }
4557 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004558
4559 rb_link_node(new_node, parent, n);
4560 rb_insert_color(new_node, &db->bb_free_root);
4561
4562 /* Now try to see the extent can be merged to left and right */
4563 node = rb_prev(new_node);
4564 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004565 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004566 if (can_merge(entry, new_entry) &&
4567 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004568 new_entry->efd_start_cluster = entry->efd_start_cluster;
4569 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004570 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004571 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004572 }
4573 }
4574
4575 node = rb_next(new_node);
4576 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004577 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004578 if (can_merge(new_entry, entry) &&
4579 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004580 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004581 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004582 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004583 }
4584 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004585 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004586 ext4_journal_callback_add(handle, ext4_free_data_callback,
4587 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004588 return 0;
4589}
4590
Theodore Ts'o44338712009-11-22 07:44:56 -05004591/**
4592 * ext4_free_blocks() -- Free given blocks and update quota
4593 * @handle: handle for this transaction
4594 * @inode: inode
4595 * @block: start physical block to free
4596 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004597 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004598 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004599void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004600 struct buffer_head *bh, ext4_fsblk_t block,
4601 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004602{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004603 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004604 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004605 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004606 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004607 ext4_grpblk_t bit;
4608 struct buffer_head *gd_bh;
4609 ext4_group_t block_group;
4610 struct ext4_sb_info *sbi;
Jan Kara7d734532013-08-17 09:36:54 -04004611 struct ext4_inode_info *ei = EXT4_I(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05004612 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004613 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004614 int err = 0;
4615 int ret;
4616
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004617 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004618 if (bh) {
4619 if (block)
4620 BUG_ON(block != bh->b_blocknr);
4621 else
4622 block = bh->b_blocknr;
4623 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004624
Alex Tomasc9de5602008-01-29 00:19:52 -05004625 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004626 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4627 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004628 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004629 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004630 goto error_return;
4631 }
4632
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004633 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004634 trace_ext4_free_blocks(inode, block, count, flags);
4635
4636 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4637 struct buffer_head *tbh = bh;
4638 int i;
4639
4640 BUG_ON(bh && (count > 1));
4641
4642 for (i = 0; i < count; i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004643 cond_resched();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004644 if (!bh)
4645 tbh = sb_find_get_block(inode->i_sb,
4646 block + i);
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004647 if (!tbh)
Namhyung Kim87783692010-10-27 21:30:11 -04004648 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004649 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004650 inode, tbh, block + i);
4651 }
4652 }
4653
Theodore Ts'o60e66792010-05-17 07:00:00 -04004654 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004655 * We need to make sure we don't reuse the freed block until
4656 * after the transaction is committed, which we can do by
4657 * treating the block as metadata, below. We make an
4658 * exception if the inode is to be written in writeback mode
4659 * since writeback mode has weak data consistency guarantees.
4660 */
4661 if (!ext4_should_writeback_data(inode))
4662 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004663
Theodore Ts'o84130192011-09-09 18:50:51 -04004664 /*
4665 * If the extent to be freed does not begin on a cluster
4666 * boundary, we need to deal with partial clusters at the
4667 * beginning and end of the extent. Normally we will free
4668 * blocks at the beginning or the end unless we are explicitly
4669 * requested to avoid doing so.
4670 */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004671 overflow = EXT4_PBLK_COFF(sbi, block);
Theodore Ts'o84130192011-09-09 18:50:51 -04004672 if (overflow) {
4673 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4674 overflow = sbi->s_cluster_ratio - overflow;
4675 block += overflow;
4676 if (count > overflow)
4677 count -= overflow;
4678 else
4679 return;
4680 } else {
4681 block -= overflow;
4682 count += overflow;
4683 }
4684 }
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004685 overflow = EXT4_LBLK_COFF(sbi, count);
Theodore Ts'o84130192011-09-09 18:50:51 -04004686 if (overflow) {
4687 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4688 if (count > overflow)
4689 count -= overflow;
4690 else
4691 return;
4692 } else
4693 count += sbi->s_cluster_ratio - overflow;
4694 }
4695
Alex Tomasc9de5602008-01-29 00:19:52 -05004696do_more:
4697 overflow = 0;
4698 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4699
Darrick J. Wong163a2032013-08-28 17:35:51 -04004700 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4701 ext4_get_group_info(sb, block_group))))
4702 return;
4703
Alex Tomasc9de5602008-01-29 00:19:52 -05004704 /*
4705 * Check to see if we are freeing blocks across a group
4706 * boundary.
4707 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004708 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4709 overflow = EXT4_C2B(sbi, bit) + count -
4710 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004711 count -= overflow;
4712 }
Lukas Czerner810da242013-03-02 17:18:58 -05004713 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004714 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004715 if (!bitmap_bh) {
4716 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004717 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004718 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004719 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004720 if (!gdp) {
4721 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004722 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004723 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004724
4725 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4726 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4727 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004728 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004729 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004730 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004731
Eric Sandeen12062dd2010-02-15 14:19:27 -05004732 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004733 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004734 /* err = 0. ext4_std_error should be a no op */
4735 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004736 }
4737
4738 BUFFER_TRACE(bitmap_bh, "getting write access");
4739 err = ext4_journal_get_write_access(handle, bitmap_bh);
4740 if (err)
4741 goto error_return;
4742
4743 /*
4744 * We are about to modify some metadata. Call the journal APIs
4745 * to unshare ->b_data if a currently-committing transaction is
4746 * using it
4747 */
4748 BUFFER_TRACE(gd_bh, "get_write_access");
4749 err = ext4_journal_get_write_access(handle, gd_bh);
4750 if (err)
4751 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004752#ifdef AGGRESSIVE_CHECK
4753 {
4754 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004755 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004756 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4757 }
4758#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004759 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004760
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004761 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4762 if (err)
4763 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004764
4765 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004766 struct ext4_free_data *new_entry;
4767 /*
4768 * blocks being freed are metadata. these blocks shouldn't
4769 * be used until this transaction is committed
4770 */
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004771 retry:
Bobi Jam18aadd42012-02-20 17:53:02 -05004772 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004773 if (!new_entry) {
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004774 /*
4775 * We use a retry loop because
4776 * ext4_free_blocks() is not allowed to fail.
4777 */
4778 cond_resched();
4779 congestion_wait(BLK_RW_ASYNC, HZ/50);
4780 goto retry;
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004781 }
Bobi Jam18aadd42012-02-20 17:53:02 -05004782 new_entry->efd_start_cluster = bit;
4783 new_entry->efd_group = block_group;
4784 new_entry->efd_count = count_clusters;
4785 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004786
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004787 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004788 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004789 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004790 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004791 /* need to update group_info->bb_free and bitmap
4792 * with group lock held. generate_buddy look at
4793 * them with group lock_held
4794 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004795 if (test_opt(sb, DISCARD)) {
4796 err = ext4_issue_discard(sb, block_group, bit, count);
4797 if (err && err != -EOPNOTSUPP)
4798 ext4_msg(sb, KERN_WARNING, "discard request in"
4799 " group:%d block:%d count:%lu failed"
4800 " with %d", block_group, bit, count,
4801 err);
Lukas Czerner8f9ff182013-10-30 11:10:52 -04004802 } else
4803 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004804
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004805 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004806 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4807 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004808 }
4809
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004810 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4811 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004812 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004813 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004814 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004815
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004816 if (sbi->s_log_groups_per_flex) {
4817 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004818 atomic64_add(count_clusters,
4819 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004820 }
4821
Jan Kara7d734532013-08-17 09:36:54 -04004822 if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) {
4823 percpu_counter_add(&sbi->s_dirtyclusters_counter,
4824 count_clusters);
4825 spin_lock(&ei->i_block_reservation_lock);
4826 if (flags & EXT4_FREE_BLOCKS_METADATA)
4827 ei->i_reserved_meta_blocks += count_clusters;
4828 else
4829 ei->i_reserved_data_blocks += count_clusters;
4830 spin_unlock(&ei->i_block_reservation_lock);
4831 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4832 dquot_reclaim_block(inode,
4833 EXT4_C2B(sbi, count_clusters));
4834 } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Aditya Kali7b415bf2011-09-09 19:04:51 -04004835 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
Jan Kara7d734532013-08-17 09:36:54 -04004836 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4837
4838 ext4_mb_unload_buddy(&e4b);
Aditya Kali7b415bf2011-09-09 19:04:51 -04004839
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004840 /* We dirtied the bitmap block */
4841 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4842 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4843
Alex Tomasc9de5602008-01-29 00:19:52 -05004844 /* And the group descriptor block */
4845 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004846 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004847 if (!err)
4848 err = ret;
4849
4850 if (overflow && !err) {
4851 block += count;
4852 count = overflow;
4853 put_bh(bitmap_bh);
4854 goto do_more;
4855 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004856error_return:
4857 brelse(bitmap_bh);
4858 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004859 return;
4860}
Lukas Czerner7360d172010-10-27 21:30:12 -04004861
4862/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004863 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004864 * @handle: handle to this transaction
4865 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004866 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004867 * @count: number of blocks to free
4868 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004869 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004870 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004871int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004872 ext4_fsblk_t block, unsigned long count)
4873{
4874 struct buffer_head *bitmap_bh = NULL;
4875 struct buffer_head *gd_bh;
4876 ext4_group_t block_group;
4877 ext4_grpblk_t bit;
4878 unsigned int i;
4879 struct ext4_group_desc *desc;
4880 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004881 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004882 int err = 0, ret, blk_free_count;
4883 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004884
4885 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4886
Yongqiang Yang4740b832011-07-26 21:51:08 -04004887 if (count == 0)
4888 return 0;
4889
Amir Goldstein2846e822011-05-09 10:46:41 -04004890 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004891 /*
4892 * Check to see if we are freeing blocks across a group
4893 * boundary.
4894 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004895 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4896 ext4_warning(sb, "too much blocks added to group %u\n",
4897 block_group);
4898 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004899 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004900 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004901
Amir Goldstein2846e822011-05-09 10:46:41 -04004902 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004903 if (!bitmap_bh) {
4904 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004905 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004906 }
4907
Amir Goldstein2846e822011-05-09 10:46:41 -04004908 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004909 if (!desc) {
4910 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004911 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004912 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004913
4914 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4915 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4916 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4917 in_range(block + count - 1, ext4_inode_table(sb, desc),
4918 sbi->s_itb_per_group)) {
4919 ext4_error(sb, "Adding blocks in system zones - "
4920 "Block = %llu, count = %lu",
4921 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004922 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004923 goto error_return;
4924 }
4925
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004926 BUFFER_TRACE(bitmap_bh, "getting write access");
4927 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004928 if (err)
4929 goto error_return;
4930
4931 /*
4932 * We are about to modify some metadata. Call the journal APIs
4933 * to unshare ->b_data if a currently-committing transaction is
4934 * using it
4935 */
4936 BUFFER_TRACE(gd_bh, "get_write_access");
4937 err = ext4_journal_get_write_access(handle, gd_bh);
4938 if (err)
4939 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004940
Amir Goldstein2846e822011-05-09 10:46:41 -04004941 for (i = 0, blocks_freed = 0; i < count; i++) {
4942 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004943 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004944 ext4_error(sb, "bit already cleared for block %llu",
4945 (ext4_fsblk_t)(block + i));
4946 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4947 } else {
4948 blocks_freed++;
4949 }
4950 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004951
4952 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4953 if (err)
4954 goto error_return;
4955
4956 /*
4957 * need to update group_info->bb_free and bitmap
4958 * with group lock held. generate_buddy look at
4959 * them with group lock_held
4960 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004961 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004962 mb_clear_bits(bitmap_bh->b_data, bit, count);
4963 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004964 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4965 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04004966 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004967 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04004968 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04004969 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05004970 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04004971
4972 if (sbi->s_log_groups_per_flex) {
4973 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004974 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4975 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04004976 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004977
4978 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04004979
4980 /* We dirtied the bitmap block */
4981 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4982 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4983
4984 /* And the group descriptor block */
4985 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4986 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4987 if (!err)
4988 err = ret;
4989
4990error_return:
4991 brelse(bitmap_bh);
4992 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004993 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04004994}
4995
4996/**
Lukas Czerner7360d172010-10-27 21:30:12 -04004997 * ext4_trim_extent -- function to TRIM one single free extent in the group
4998 * @sb: super block for the file system
4999 * @start: starting block of the free extent in the alloc. group
5000 * @count: number of blocks to TRIM
5001 * @group: alloc. group we are working with
5002 * @e4b: ext4 buddy for the group
5003 *
5004 * Trim "count" blocks starting at "start" in the "group". To assure that no
5005 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5006 * be called with under the group lock.
5007 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005008static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04005009 ext4_group_t group, struct ext4_buddy *e4b)
Lukas Czerner7360d172010-10-27 21:30:12 -04005010{
5011 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005012 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005013
Tao Mab3d4c2b2011-07-11 00:01:52 -04005014 trace_ext4_trim_extent(sb, group, start, count);
5015
Lukas Czerner7360d172010-10-27 21:30:12 -04005016 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5017
5018 ex.fe_start = start;
5019 ex.fe_group = group;
5020 ex.fe_len = count;
5021
5022 /*
5023 * Mark blocks used, so no one can reuse them while
5024 * being trimmed.
5025 */
5026 mb_mark_used(e4b, &ex);
5027 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005028 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04005029 ext4_lock_group(sb, group);
5030 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005031 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005032}
5033
5034/**
5035 * ext4_trim_all_free -- function to trim all free space in alloc. group
5036 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04005037 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04005038 * @start: first group block to examine
5039 * @max: last group block to examine
5040 * @minblocks: minimum extent block count
5041 *
5042 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5043 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5044 * the extent.
5045 *
5046 *
5047 * ext4_trim_all_free walks through group's block bitmap searching for free
5048 * extents. When the free extent is found, mark it as used in group buddy
5049 * bitmap. Then issue a TRIM command on this extent and free the extent in
5050 * the group buddy bitmap. This is done until whole group is scanned.
5051 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005052static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005053ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5054 ext4_grpblk_t start, ext4_grpblk_t max,
5055 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005056{
5057 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005058 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005059 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005060 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005061
Tao Mab3d4c2b2011-07-11 00:01:52 -04005062 trace_ext4_trim_all_free(sb, group, start, max);
5063
Lukas Czerner78944082011-05-24 18:16:27 -04005064 ret = ext4_mb_load_buddy(sb, group, &e4b);
5065 if (ret) {
5066 ext4_error(sb, "Error in loading buddy "
5067 "information for %u", group);
5068 return ret;
5069 }
Lukas Czerner78944082011-05-24 18:16:27 -04005070 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005071
5072 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005073 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5074 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5075 goto out;
5076
Lukas Czerner78944082011-05-24 18:16:27 -04005077 start = (e4b.bd_info->bb_first_free > start) ?
5078 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005079
Lukas Czerner913eed832012-03-21 21:22:22 -04005080 while (start <= max) {
5081 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5082 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005083 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005084 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005085
5086 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005087 ret = ext4_trim_extent(sb, start,
5088 next - start, group, &e4b);
5089 if (ret && ret != -EOPNOTSUPP)
5090 break;
5091 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005092 count += next - start;
5093 }
Tao Ma169ddc32011-07-11 00:00:07 -04005094 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005095 start = next + 1;
5096
5097 if (fatal_signal_pending(current)) {
5098 count = -ERESTARTSYS;
5099 break;
5100 }
5101
5102 if (need_resched()) {
5103 ext4_unlock_group(sb, group);
5104 cond_resched();
5105 ext4_lock_group(sb, group);
5106 }
5107
Tao Ma169ddc32011-07-11 00:00:07 -04005108 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005109 break;
5110 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005111
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005112 if (!ret) {
5113 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005114 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005115 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005116out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005117 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005118 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005119
5120 ext4_debug("trimmed %d blocks in the group %d\n",
5121 count, group);
5122
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005123 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005124}
5125
5126/**
5127 * ext4_trim_fs() -- trim ioctl handle function
5128 * @sb: superblock for filesystem
5129 * @range: fstrim_range structure
5130 *
5131 * start: First Byte to trim
5132 * len: number of Bytes to trim from start
5133 * minlen: minimum extent length in Bytes
5134 * ext4_trim_fs goes through all allocation groups containing Bytes from
5135 * start to start+len. For each such a group ext4_trim_all_free function
5136 * is invoked to trim all free space.
5137 */
5138int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5139{
Lukas Czerner78944082011-05-24 18:16:27 -04005140 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005141 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005142 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005143 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005144 ext4_fsblk_t first_data_blk =
5145 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005146 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005147 int ret = 0;
5148
5149 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005150 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005151 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5152 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005153
Lukas Czerner5de35e82012-10-22 18:01:19 -04005154 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5155 start >= max_blks ||
5156 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005157 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005158 if (end >= max_blks)
5159 end = max_blks - 1;
5160 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005161 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005162 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005163 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005164
Lukas Czerner913eed832012-03-21 21:22:22 -04005165 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005166 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005167 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005168 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005169 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005170
Lukas Czerner913eed832012-03-21 21:22:22 -04005171 /* end now represents the last cluster to discard in this group */
5172 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005173
5174 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005175 grp = ext4_get_group_info(sb, group);
5176 /* We only do this if the grp has never been initialized */
5177 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5178 ret = ext4_mb_init_group(sb, group);
5179 if (ret)
5180 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005181 }
5182
Tao Ma0ba08512011-03-23 15:48:11 -04005183 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005184 * For all the groups except the last one, last cluster will
5185 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5186 * change it for the last group, note that last_cluster is
5187 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005188 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005189 if (group == last_group)
5190 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005191
Lukas Czerner78944082011-05-24 18:16:27 -04005192 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005193 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005194 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005195 if (cnt < 0) {
5196 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005197 break;
5198 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005199 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005200 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005201
5202 /*
5203 * For every group except the first one, we are sure
5204 * that the first cluster to discard will be cluster #0.
5205 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005206 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005207 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005208
Tao Ma3d56b8d2011-07-11 00:03:38 -04005209 if (!ret)
5210 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5211
Tao Ma22f10452011-07-10 23:52:37 -04005212out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005213 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005214 return ret;
5215}