blob: 73ccbb3b973b8ff9029df4d48e99ebdd10b7176c [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)
Younger Liuc57ab392014-04-10 23:03:43 -0400992 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -0400993 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)
Younger Liuc57ab392014-04-10 23:03:43 -04001006 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001007 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 }
Younger Liuc57ab392014-04-10 23:03:43 -04001171 if (page == NULL) {
1172 ret = -ENOMEM;
1173 goto err;
1174 }
1175 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001176 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001177 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001178 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001179 e4b->bd_bitmap_page = page;
1180 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1181 mark_page_accessed(page);
1182
1183 block++;
1184 pnum = block / blocks_per_page;
1185 poff = block % blocks_per_page;
1186
1187 page = find_get_page(inode->i_mapping, pnum);
1188 if (page == NULL || !PageUptodate(page)) {
1189 if (page)
1190 page_cache_release(page);
1191 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1192 if (page) {
1193 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001194 if (!PageUptodate(page)) {
1195 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1196 if (ret) {
1197 unlock_page(page);
1198 goto err;
1199 }
1200 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001201 unlock_page(page);
1202 }
1203 }
Younger Liuc57ab392014-04-10 23:03:43 -04001204 if (page == NULL) {
1205 ret = -ENOMEM;
1206 goto err;
1207 }
1208 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001209 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001210 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001211 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001212 e4b->bd_buddy_page = page;
1213 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1214 mark_page_accessed(page);
1215
1216 BUG_ON(e4b->bd_bitmap_page == NULL);
1217 BUG_ON(e4b->bd_buddy_page == NULL);
1218
1219 return 0;
1220
1221err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001222 if (page)
1223 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001224 if (e4b->bd_bitmap_page)
1225 page_cache_release(e4b->bd_bitmap_page);
1226 if (e4b->bd_buddy_page)
1227 page_cache_release(e4b->bd_buddy_page);
1228 e4b->bd_buddy = NULL;
1229 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001230 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001231}
1232
Jing Zhange39e07f2010-05-14 00:00:00 -04001233static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001234{
1235 if (e4b->bd_bitmap_page)
1236 page_cache_release(e4b->bd_bitmap_page);
1237 if (e4b->bd_buddy_page)
1238 page_cache_release(e4b->bd_buddy_page);
1239}
1240
1241
1242static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1243{
1244 int order = 1;
1245 void *bb;
1246
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001247 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001248 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1249
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001250 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001251 while (order <= e4b->bd_blkbits + 1) {
1252 block = block >> 1;
1253 if (!mb_test_bit(block, bb)) {
1254 /* this block is part of buddy of order 'order' */
1255 return order;
1256 }
1257 bb += 1 << (e4b->bd_blkbits - order);
1258 order++;
1259 }
1260 return 0;
1261}
1262
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001263static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001264{
1265 __u32 *addr;
1266
1267 len = cur + len;
1268 while (cur < len) {
1269 if ((cur & 31) == 0 && (len - cur) >= 32) {
1270 /* fast path: clear whole word at once */
1271 addr = bm + (cur >> 3);
1272 *addr = 0;
1273 cur += 32;
1274 continue;
1275 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001276 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001277 cur++;
1278 }
1279}
1280
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001281/* clear bits in given range
1282 * will return first found zero bit if any, -1 otherwise
1283 */
1284static int mb_test_and_clear_bits(void *bm, int cur, int len)
1285{
1286 __u32 *addr;
1287 int zero_bit = -1;
1288
1289 len = cur + len;
1290 while (cur < len) {
1291 if ((cur & 31) == 0 && (len - cur) >= 32) {
1292 /* fast path: clear whole word at once */
1293 addr = bm + (cur >> 3);
1294 if (*addr != (__u32)(-1) && zero_bit == -1)
1295 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1296 *addr = 0;
1297 cur += 32;
1298 continue;
1299 }
1300 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1301 zero_bit = cur;
1302 cur++;
1303 }
1304
1305 return zero_bit;
1306}
1307
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001308void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001309{
1310 __u32 *addr;
1311
1312 len = cur + len;
1313 while (cur < len) {
1314 if ((cur & 31) == 0 && (len - cur) >= 32) {
1315 /* fast path: set whole word at once */
1316 addr = bm + (cur >> 3);
1317 *addr = 0xffffffff;
1318 cur += 32;
1319 continue;
1320 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001321 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001322 cur++;
1323 }
1324}
1325
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001326/*
1327 * _________________________________________________________________ */
1328
1329static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001330{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001331 if (mb_test_bit(*bit + side, bitmap)) {
1332 mb_clear_bit(*bit, bitmap);
1333 (*bit) -= side;
1334 return 1;
1335 }
1336 else {
1337 (*bit) += side;
1338 mb_set_bit(*bit, bitmap);
1339 return -1;
1340 }
1341}
1342
1343static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1344{
1345 int max;
1346 int order = 1;
1347 void *buddy = mb_find_buddy(e4b, order, &max);
1348
1349 while (buddy) {
1350 void *buddy2;
1351
1352 /* Bits in range [first; last] are known to be set since
1353 * corresponding blocks were allocated. Bits in range
1354 * (first; last) will stay set because they form buddies on
1355 * upper layer. We just deal with borders if they don't
1356 * align with upper layer and then go up.
1357 * Releasing entire group is all about clearing
1358 * single bit of highest order buddy.
1359 */
1360
1361 /* Example:
1362 * ---------------------------------
1363 * | 1 | 1 | 1 | 1 |
1364 * ---------------------------------
1365 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1366 * ---------------------------------
1367 * 0 1 2 3 4 5 6 7
1368 * \_____________________/
1369 *
1370 * Neither [1] nor [6] is aligned to above layer.
1371 * Left neighbour [0] is free, so mark it busy,
1372 * decrease bb_counters and extend range to
1373 * [0; 6]
1374 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1375 * mark [6] free, increase bb_counters and shrink range to
1376 * [0; 5].
1377 * Then shift range to [0; 2], go up and do the same.
1378 */
1379
1380
1381 if (first & 1)
1382 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1383 if (!(last & 1))
1384 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1385 if (first > last)
1386 break;
1387 order++;
1388
1389 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1390 mb_clear_bits(buddy, first, last - first + 1);
1391 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1392 break;
1393 }
1394 first >>= 1;
1395 last >>= 1;
1396 buddy = buddy2;
1397 }
1398}
1399
1400static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1401 int first, int count)
1402{
1403 int left_is_free = 0;
1404 int right_is_free = 0;
1405 int block;
1406 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001407 struct super_block *sb = e4b->bd_sb;
1408
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001409 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001410 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Darrick J. Wong163a2032013-08-28 17:35:51 -04001411 /* Don't bother if the block group is corrupt. */
1412 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1413 return;
1414
Alex Tomasc9de5602008-01-29 00:19:52 -05001415 mb_check_buddy(e4b);
1416 mb_free_blocks_double(inode, e4b, first, count);
1417
1418 e4b->bd_info->bb_free += count;
1419 if (first < e4b->bd_info->bb_first_free)
1420 e4b->bd_info->bb_first_free = first;
1421
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001422 /* access memory sequentially: check left neighbour,
1423 * clear range and then check right neighbour
1424 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001425 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001426 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1427 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1428 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1429 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1430
1431 if (unlikely(block != -1)) {
1432 ext4_fsblk_t blocknr;
1433
1434 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1435 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1436 ext4_grp_locked_error(sb, e4b->bd_group,
1437 inode ? inode->i_ino : 0,
1438 blocknr,
1439 "freeing already freed block "
Darrick J. Wong163a2032013-08-28 17:35:51 -04001440 "(bit %u); block bitmap corrupt.",
1441 block);
1442 /* Mark the block group as corrupt. */
1443 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1444 &e4b->bd_info->bb_state);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001445 mb_regenerate_buddy(e4b);
1446 goto done;
1447 }
1448
1449 /* let's maintain fragments counter */
1450 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001451 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001452 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001453 e4b->bd_info->bb_fragments++;
1454
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001455 /* buddy[0] == bd_bitmap is a special case, so handle
1456 * it right away and let mb_buddy_mark_free stay free of
1457 * zero order checks.
1458 * Check if neighbours are to be coaleasced,
1459 * adjust bitmap bb_counters and borders appropriately.
1460 */
1461 if (first & 1) {
1462 first += !left_is_free;
1463 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001464 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001465 if (!(last & 1)) {
1466 last -= !right_is_free;
1467 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1468 }
1469
1470 if (first <= last)
1471 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1472
1473done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001474 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001475 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001476}
1477
Robin Dong15c006a2012-08-17 10:02:17 -04001478static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001479 int needed, struct ext4_free_extent *ex)
1480{
1481 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001482 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001483 void *buddy;
1484
Vincent Minetbc8e6742009-05-15 08:33:18 -04001485 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001486 BUG_ON(ex == NULL);
1487
Robin Dong15c006a2012-08-17 10:02:17 -04001488 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001489 BUG_ON(buddy == NULL);
1490 BUG_ON(block >= max);
1491 if (mb_test_bit(block, buddy)) {
1492 ex->fe_len = 0;
1493 ex->fe_start = 0;
1494 ex->fe_group = 0;
1495 return 0;
1496 }
1497
Robin Dong15c006a2012-08-17 10:02:17 -04001498 /* find actual order */
1499 order = mb_find_order_for_block(e4b, block);
1500 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001501
1502 ex->fe_len = 1 << order;
1503 ex->fe_start = block << order;
1504 ex->fe_group = e4b->bd_group;
1505
1506 /* calc difference from given start */
1507 next = next - ex->fe_start;
1508 ex->fe_len -= next;
1509 ex->fe_start += next;
1510
1511 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001512 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001513
1514 if (block + 1 >= max)
1515 break;
1516
1517 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001518 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001519 break;
1520
Robin Dongb051d8d2011-10-26 05:30:30 -04001521 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001522
Alex Tomasc9de5602008-01-29 00:19:52 -05001523 block = next >> order;
1524 ex->fe_len += 1 << order;
1525 }
1526
1527 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1528 return ex->fe_len;
1529}
1530
1531static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1532{
1533 int ord;
1534 int mlen = 0;
1535 int max = 0;
1536 int cur;
1537 int start = ex->fe_start;
1538 int len = ex->fe_len;
1539 unsigned ret = 0;
1540 int len0 = len;
1541 void *buddy;
1542
1543 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1544 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001545 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001546 mb_check_buddy(e4b);
1547 mb_mark_used_double(e4b, start, len);
1548
1549 e4b->bd_info->bb_free -= len;
1550 if (e4b->bd_info->bb_first_free == start)
1551 e4b->bd_info->bb_first_free += len;
1552
1553 /* let's maintain fragments counter */
1554 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001555 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001556 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001557 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001558 if (mlen && max)
1559 e4b->bd_info->bb_fragments++;
1560 else if (!mlen && !max)
1561 e4b->bd_info->bb_fragments--;
1562
1563 /* let's maintain buddy itself */
1564 while (len) {
1565 ord = mb_find_order_for_block(e4b, start);
1566
1567 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1568 /* the whole chunk may be allocated at once! */
1569 mlen = 1 << ord;
1570 buddy = mb_find_buddy(e4b, ord, &max);
1571 BUG_ON((start >> ord) >= max);
1572 mb_set_bit(start >> ord, buddy);
1573 e4b->bd_info->bb_counters[ord]--;
1574 start += mlen;
1575 len -= mlen;
1576 BUG_ON(len < 0);
1577 continue;
1578 }
1579
1580 /* store for history */
1581 if (ret == 0)
1582 ret = len | (ord << 16);
1583
1584 /* we have to split large buddy */
1585 BUG_ON(ord <= 0);
1586 buddy = mb_find_buddy(e4b, ord, &max);
1587 mb_set_bit(start >> ord, buddy);
1588 e4b->bd_info->bb_counters[ord]--;
1589
1590 ord--;
1591 cur = (start >> ord) & ~1U;
1592 buddy = mb_find_buddy(e4b, ord, &max);
1593 mb_clear_bit(cur, buddy);
1594 mb_clear_bit(cur + 1, buddy);
1595 e4b->bd_info->bb_counters[ord]++;
1596 e4b->bd_info->bb_counters[ord]++;
1597 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001598 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001599
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001600 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001601 mb_check_buddy(e4b);
1602
1603 return ret;
1604}
1605
1606/*
1607 * Must be called under group lock!
1608 */
1609static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1610 struct ext4_buddy *e4b)
1611{
1612 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1613 int ret;
1614
1615 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1616 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1617
1618 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1619 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1620 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1621
1622 /* preallocation can change ac_b_ex, thus we store actually
1623 * allocated blocks for history */
1624 ac->ac_f_ex = ac->ac_b_ex;
1625
1626 ac->ac_status = AC_STATUS_FOUND;
1627 ac->ac_tail = ret & 0xffff;
1628 ac->ac_buddy = ret >> 16;
1629
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001630 /*
1631 * take the page reference. We want the page to be pinned
1632 * so that we don't get a ext4_mb_init_cache_call for this
1633 * group until we update the bitmap. That would mean we
1634 * double allocate blocks. The reference is dropped
1635 * in ext4_mb_release_context
1636 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001637 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1638 get_page(ac->ac_bitmap_page);
1639 ac->ac_buddy_page = e4b->bd_buddy_page;
1640 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001641 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001642 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001643 spin_lock(&sbi->s_md_lock);
1644 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1645 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1646 spin_unlock(&sbi->s_md_lock);
1647 }
1648}
1649
1650/*
1651 * regular allocator, for general purposes allocation
1652 */
1653
1654static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1655 struct ext4_buddy *e4b,
1656 int finish_group)
1657{
1658 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1659 struct ext4_free_extent *bex = &ac->ac_b_ex;
1660 struct ext4_free_extent *gex = &ac->ac_g_ex;
1661 struct ext4_free_extent ex;
1662 int max;
1663
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001664 if (ac->ac_status == AC_STATUS_FOUND)
1665 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001666 /*
1667 * We don't want to scan for a whole year
1668 */
1669 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1670 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1671 ac->ac_status = AC_STATUS_BREAK;
1672 return;
1673 }
1674
1675 /*
1676 * Haven't found good chunk so far, let's continue
1677 */
1678 if (bex->fe_len < gex->fe_len)
1679 return;
1680
1681 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1682 && bex->fe_group == e4b->bd_group) {
1683 /* recheck chunk's availability - we don't know
1684 * when it was found (within this lock-unlock
1685 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001686 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001687 if (max >= gex->fe_len) {
1688 ext4_mb_use_best_found(ac, e4b);
1689 return;
1690 }
1691 }
1692}
1693
1694/*
1695 * The routine checks whether found extent is good enough. If it is,
1696 * then the extent gets marked used and flag is set to the context
1697 * to stop scanning. Otherwise, the extent is compared with the
1698 * previous found extent and if new one is better, then it's stored
1699 * in the context. Later, the best found extent will be used, if
1700 * mballoc can't find good enough extent.
1701 *
1702 * FIXME: real allocation policy is to be designed yet!
1703 */
1704static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1705 struct ext4_free_extent *ex,
1706 struct ext4_buddy *e4b)
1707{
1708 struct ext4_free_extent *bex = &ac->ac_b_ex;
1709 struct ext4_free_extent *gex = &ac->ac_g_ex;
1710
1711 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001712 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1713 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001714 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1715
1716 ac->ac_found++;
1717
1718 /*
1719 * The special case - take what you catch first
1720 */
1721 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1722 *bex = *ex;
1723 ext4_mb_use_best_found(ac, e4b);
1724 return;
1725 }
1726
1727 /*
1728 * Let's check whether the chuck is good enough
1729 */
1730 if (ex->fe_len == gex->fe_len) {
1731 *bex = *ex;
1732 ext4_mb_use_best_found(ac, e4b);
1733 return;
1734 }
1735
1736 /*
1737 * If this is first found extent, just store it in the context
1738 */
1739 if (bex->fe_len == 0) {
1740 *bex = *ex;
1741 return;
1742 }
1743
1744 /*
1745 * If new found extent is better, store it in the context
1746 */
1747 if (bex->fe_len < gex->fe_len) {
1748 /* if the request isn't satisfied, any found extent
1749 * larger than previous best one is better */
1750 if (ex->fe_len > bex->fe_len)
1751 *bex = *ex;
1752 } else if (ex->fe_len > gex->fe_len) {
1753 /* if the request is satisfied, then we try to find
1754 * an extent that still satisfy the request, but is
1755 * smaller than previous one */
1756 if (ex->fe_len < bex->fe_len)
1757 *bex = *ex;
1758 }
1759
1760 ext4_mb_check_limits(ac, e4b, 0);
1761}
1762
Eric Sandeen089ceec2009-07-05 22:17:31 -04001763static noinline_for_stack
1764int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001765 struct ext4_buddy *e4b)
1766{
1767 struct ext4_free_extent ex = ac->ac_b_ex;
1768 ext4_group_t group = ex.fe_group;
1769 int max;
1770 int err;
1771
1772 BUG_ON(ex.fe_len <= 0);
1773 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1774 if (err)
1775 return err;
1776
1777 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001778 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001779
1780 if (max > 0) {
1781 ac->ac_b_ex = ex;
1782 ext4_mb_use_best_found(ac, e4b);
1783 }
1784
1785 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001786 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001787
1788 return 0;
1789}
1790
Eric Sandeen089ceec2009-07-05 22:17:31 -04001791static noinline_for_stack
1792int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001793 struct ext4_buddy *e4b)
1794{
1795 ext4_group_t group = ac->ac_g_ex.fe_group;
1796 int max;
1797 int err;
1798 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001799 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001800 struct ext4_free_extent ex;
1801
1802 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1803 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001804 if (grp->bb_free == 0)
1805 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001806
1807 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1808 if (err)
1809 return err;
1810
Darrick J. Wong163a2032013-08-28 17:35:51 -04001811 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1812 ext4_mb_unload_buddy(e4b);
1813 return 0;
1814 }
1815
Alex Tomasc9de5602008-01-29 00:19:52 -05001816 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001817 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001818 ac->ac_g_ex.fe_len, &ex);
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001819 ex.fe_logical = 0xDEADFA11; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001820
1821 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1822 ext4_fsblk_t start;
1823
Akinobu Mita5661bd62010-03-03 23:53:39 -05001824 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1825 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001826 /* use do_div to get remainder (would be 64-bit modulo) */
1827 if (do_div(start, sbi->s_stripe) == 0) {
1828 ac->ac_found++;
1829 ac->ac_b_ex = ex;
1830 ext4_mb_use_best_found(ac, e4b);
1831 }
1832 } else if (max >= ac->ac_g_ex.fe_len) {
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 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1840 /* Sometimes, caller may want to merge even small
1841 * number of blocks to an existing extent */
1842 BUG_ON(ex.fe_len <= 0);
1843 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1844 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1845 ac->ac_found++;
1846 ac->ac_b_ex = ex;
1847 ext4_mb_use_best_found(ac, e4b);
1848 }
1849 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001850 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001851
1852 return 0;
1853}
1854
1855/*
1856 * The routine scans buddy structures (not bitmap!) from given order
1857 * to max order and tries to find big enough chunk to satisfy the req
1858 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001859static noinline_for_stack
1860void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001861 struct ext4_buddy *e4b)
1862{
1863 struct super_block *sb = ac->ac_sb;
1864 struct ext4_group_info *grp = e4b->bd_info;
1865 void *buddy;
1866 int i;
1867 int k;
1868 int max;
1869
1870 BUG_ON(ac->ac_2order <= 0);
1871 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1872 if (grp->bb_counters[i] == 0)
1873 continue;
1874
1875 buddy = mb_find_buddy(e4b, i, &max);
1876 BUG_ON(buddy == NULL);
1877
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001878 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001879 BUG_ON(k >= max);
1880
1881 ac->ac_found++;
1882
1883 ac->ac_b_ex.fe_len = 1 << i;
1884 ac->ac_b_ex.fe_start = k << i;
1885 ac->ac_b_ex.fe_group = e4b->bd_group;
1886
1887 ext4_mb_use_best_found(ac, e4b);
1888
1889 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1890
1891 if (EXT4_SB(sb)->s_mb_stats)
1892 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1893
1894 break;
1895 }
1896}
1897
1898/*
1899 * The routine scans the group and measures all found extents.
1900 * In order to optimize scanning, caller must pass number of
1901 * free blocks in the group, so the routine can know upper limit.
1902 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001903static noinline_for_stack
1904void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001905 struct ext4_buddy *e4b)
1906{
1907 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001908 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001909 struct ext4_free_extent ex;
1910 int i;
1911 int free;
1912
1913 free = e4b->bd_info->bb_free;
1914 BUG_ON(free <= 0);
1915
1916 i = e4b->bd_info->bb_first_free;
1917
1918 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001919 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001920 EXT4_CLUSTERS_PER_GROUP(sb), i);
1921 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001922 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001923 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001924 * free blocks even though group info says we
1925 * we have free blocks
1926 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001927 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001928 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001929 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001930 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001931 break;
1932 }
1933
Robin Dong15c006a2012-08-17 10:02:17 -04001934 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001935 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001936 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001937 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001938 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001939 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001940 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001941 /*
1942 * The number of free blocks differs. This mostly
1943 * indicate that the bitmap is corrupt. So exit
1944 * without claiming the space.
1945 */
1946 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001947 }
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001948 ex.fe_logical = 0xDEADC0DE; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001949 ext4_mb_measure_extent(ac, &ex, e4b);
1950
1951 i += ex.fe_len;
1952 free -= ex.fe_len;
1953 }
1954
1955 ext4_mb_check_limits(ac, e4b, 1);
1956}
1957
1958/*
1959 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001960 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001961 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001962static noinline_for_stack
1963void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001964 struct ext4_buddy *e4b)
1965{
1966 struct super_block *sb = ac->ac_sb;
1967 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001968 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001969 struct ext4_free_extent ex;
1970 ext4_fsblk_t first_group_block;
1971 ext4_fsblk_t a;
1972 ext4_grpblk_t i;
1973 int max;
1974
1975 BUG_ON(sbi->s_stripe == 0);
1976
1977 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001978 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1979
Alex Tomasc9de5602008-01-29 00:19:52 -05001980 a = first_group_block + sbi->s_stripe - 1;
1981 do_div(a, sbi->s_stripe);
1982 i = (a * sbi->s_stripe) - first_group_block;
1983
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001984 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001985 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04001986 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001987 if (max >= sbi->s_stripe) {
1988 ac->ac_found++;
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001989 ex.fe_logical = 0xDEADF00D; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001990 ac->ac_b_ex = ex;
1991 ext4_mb_use_best_found(ac, e4b);
1992 break;
1993 }
1994 }
1995 i += sbi->s_stripe;
1996 }
1997}
1998
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001999/* This is now called BEFORE we load the buddy bitmap. */
Alex Tomasc9de5602008-01-29 00:19:52 -05002000static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2001 ext4_group_t group, int cr)
2002{
2003 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04002004 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05002005 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2006
2007 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002008
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04002009 free = grp->bb_free;
2010 if (free == 0)
2011 return 0;
2012 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2013 return 0;
2014
Darrick J. Wong163a2032013-08-28 17:35:51 -04002015 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2016 return 0;
2017
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002018 /* We only do this if the grp has never been initialized */
2019 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2020 int ret = ext4_mb_init_group(ac->ac_sb, group);
2021 if (ret)
2022 return 0;
2023 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002024
Alex Tomasc9de5602008-01-29 00:19:52 -05002025 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05002026 if (fragments == 0)
2027 return 0;
2028
2029 switch (cr) {
2030 case 0:
2031 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002032
Theodore Ts'oa4912122009-03-12 12:18:34 -04002033 /* Avoid using the first bg of a flexgroup for data files */
2034 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2035 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2036 ((group % flex_size) == 0))
2037 return 0;
2038
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002039 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2040 (free / fragments) >= ac->ac_g_ex.fe_len)
2041 return 1;
2042
2043 if (grp->bb_largest_free_order < ac->ac_2order)
2044 return 0;
2045
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002046 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002047 case 1:
2048 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2049 return 1;
2050 break;
2051 case 2:
2052 if (free >= ac->ac_g_ex.fe_len)
2053 return 1;
2054 break;
2055 case 3:
2056 return 1;
2057 default:
2058 BUG();
2059 }
2060
2061 return 0;
2062}
2063
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002064static noinline_for_stack int
2065ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002066{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002067 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002068 int cr;
2069 int err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002070 struct ext4_sb_info *sbi;
2071 struct super_block *sb;
2072 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002073
2074 sb = ac->ac_sb;
2075 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002076 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002077 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002078 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002079 ngroups = sbi->s_blockfile_groups;
2080
Alex Tomasc9de5602008-01-29 00:19:52 -05002081 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2082
2083 /* first, try the goal */
2084 err = ext4_mb_find_by_goal(ac, &e4b);
2085 if (err || ac->ac_status == AC_STATUS_FOUND)
2086 goto out;
2087
2088 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2089 goto out;
2090
2091 /*
2092 * ac->ac2_order is set only if the fe_len is a power of 2
2093 * if ac2_order is set we also set criteria to 0 so that we
2094 * try exact allocation using buddy.
2095 */
2096 i = fls(ac->ac_g_ex.fe_len);
2097 ac->ac_2order = 0;
2098 /*
2099 * We search using buddy data only if the order of the request
2100 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002101 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002102 */
2103 if (i >= sbi->s_mb_order2_reqs) {
2104 /*
2105 * This should tell if fe_len is exactly power of 2
2106 */
2107 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2108 ac->ac_2order = i - 1;
2109 }
2110
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002111 /* if stream allocation is enabled, use global goal */
2112 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002113 /* TBD: may be hot point */
2114 spin_lock(&sbi->s_md_lock);
2115 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2116 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2117 spin_unlock(&sbi->s_md_lock);
2118 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002119
Alex Tomasc9de5602008-01-29 00:19:52 -05002120 /* Let's just scan groups to find more-less suitable blocks */
2121 cr = ac->ac_2order ? 0 : 1;
2122 /*
2123 * cr == 0 try to get exact allocation,
2124 * cr == 3 try to get anything
2125 */
2126repeat:
2127 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2128 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002129 /*
2130 * searching for the right group start
2131 * from the goal value specified
2132 */
2133 group = ac->ac_g_ex.fe_group;
2134
Theodore Ts'o8df96752009-05-01 08:50:38 -04002135 for (i = 0; i < ngroups; group++, i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04002136 cond_resched();
Lachlan McIlroye6155732013-05-05 23:10:00 -04002137 /*
2138 * Artificially restricted ngroups for non-extent
2139 * files makes group > ngroups possible on first loop.
2140 */
2141 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002142 group = 0;
2143
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002144 /* This now checks without needing the buddy page */
2145 if (!ext4_mb_good_group(ac, group, cr))
Alex Tomasc9de5602008-01-29 00:19:52 -05002146 continue;
2147
Alex Tomasc9de5602008-01-29 00:19:52 -05002148 err = ext4_mb_load_buddy(sb, group, &e4b);
2149 if (err)
2150 goto out;
2151
2152 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002153
2154 /*
2155 * We need to check again after locking the
2156 * block group
2157 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002158 if (!ext4_mb_good_group(ac, group, cr)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002159 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002160 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002161 continue;
2162 }
2163
2164 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002165 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002166 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002167 else if (cr == 1 && sbi->s_stripe &&
2168 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002169 ext4_mb_scan_aligned(ac, &e4b);
2170 else
2171 ext4_mb_complex_scan_group(ac, &e4b);
2172
2173 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002174 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002175
2176 if (ac->ac_status != AC_STATUS_CONTINUE)
2177 break;
2178 }
2179 }
2180
2181 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2182 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2183 /*
2184 * We've been searching too long. Let's try to allocate
2185 * the best chunk we've found so far
2186 */
2187
2188 ext4_mb_try_best_found(ac, &e4b);
2189 if (ac->ac_status != AC_STATUS_FOUND) {
2190 /*
2191 * Someone more lucky has already allocated it.
2192 * The only thing we can do is just take first
2193 * found block(s)
2194 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2195 */
2196 ac->ac_b_ex.fe_group = 0;
2197 ac->ac_b_ex.fe_start = 0;
2198 ac->ac_b_ex.fe_len = 0;
2199 ac->ac_status = AC_STATUS_CONTINUE;
2200 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2201 cr = 3;
2202 atomic_inc(&sbi->s_mb_lost_chunks);
2203 goto repeat;
2204 }
2205 }
2206out:
2207 return err;
2208}
2209
Alex Tomasc9de5602008-01-29 00:19:52 -05002210static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2211{
2212 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002213 ext4_group_t group;
2214
Theodore Ts'o8df96752009-05-01 08:50:38 -04002215 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002216 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002217 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002218 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002219}
2220
2221static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2222{
2223 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002224 ext4_group_t group;
2225
2226 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002227 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002228 return NULL;
2229 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002230 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002231}
2232
2233static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2234{
2235 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002236 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002237 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002238 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002239 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002240 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002241 struct sg {
2242 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002243 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002244 } sg;
2245
2246 group--;
2247 if (group == 0)
2248 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2249 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2250 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2251 "group", "free", "frags", "first",
2252 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2253 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2254
2255 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2256 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002257 grinfo = ext4_get_group_info(sb, group);
2258 /* Load the group info in memory only if not already loaded. */
2259 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2260 err = ext4_mb_load_buddy(sb, group, &e4b);
2261 if (err) {
2262 seq_printf(seq, "#%-5u: I/O error\n", group);
2263 return 0;
2264 }
2265 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002266 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002267
Alex Tomasc9de5602008-01-29 00:19:52 -05002268 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002269
2270 if (buddy_loaded)
2271 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002272
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002273 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002274 sg.info.bb_fragments, sg.info.bb_first_free);
2275 for (i = 0; i <= 13; i++)
2276 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2277 sg.info.bb_counters[i] : 0);
2278 seq_printf(seq, " ]\n");
2279
2280 return 0;
2281}
2282
2283static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2284{
2285}
2286
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002287static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002288 .start = ext4_mb_seq_groups_start,
2289 .next = ext4_mb_seq_groups_next,
2290 .stop = ext4_mb_seq_groups_stop,
2291 .show = ext4_mb_seq_groups_show,
2292};
2293
2294static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2295{
Al Virod9dda782013-03-31 18:16:14 -04002296 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002297 int rc;
2298
2299 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2300 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002301 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002302 m->private = sb;
2303 }
2304 return rc;
2305
2306}
2307
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002308static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002309 .owner = THIS_MODULE,
2310 .open = ext4_mb_seq_groups_open,
2311 .read = seq_read,
2312 .llseek = seq_lseek,
2313 .release = seq_release,
2314};
2315
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002316static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2317{
2318 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2319 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2320
2321 BUG_ON(!cachep);
2322 return cachep;
2323}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002324
Theodore Ts'o28623c22012-09-05 01:31:50 -04002325/*
2326 * Allocate the top-level s_group_info array for the specified number
2327 * of groups
2328 */
2329int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2330{
2331 struct ext4_sb_info *sbi = EXT4_SB(sb);
2332 unsigned size;
2333 struct ext4_group_info ***new_groupinfo;
2334
2335 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2336 EXT4_DESC_PER_BLOCK_BITS(sb);
2337 if (size <= sbi->s_group_info_size)
2338 return 0;
2339
2340 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2341 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2342 if (!new_groupinfo) {
2343 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2344 return -ENOMEM;
2345 }
2346 if (sbi->s_group_info) {
2347 memcpy(new_groupinfo, sbi->s_group_info,
2348 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2349 ext4_kvfree(sbi->s_group_info);
2350 }
2351 sbi->s_group_info = new_groupinfo;
2352 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2353 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2354 sbi->s_group_info_size);
2355 return 0;
2356}
2357
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002358/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002359int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002360 struct ext4_group_desc *desc)
2361{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002362 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002363 int metalen = 0;
2364 struct ext4_sb_info *sbi = EXT4_SB(sb);
2365 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002366 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002367
2368 /*
2369 * First check if this group is the first of a reserved block.
2370 * If it's true, we have to allocate a new table of pointers
2371 * to ext4_group_info structures
2372 */
2373 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2374 metalen = sizeof(*meta_group_info) <<
2375 EXT4_DESC_PER_BLOCK_BITS(sb);
2376 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2377 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002378 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002379 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002380 goto exit_meta_group_info;
2381 }
2382 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2383 meta_group_info;
2384 }
2385
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002386 meta_group_info =
2387 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2388 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2389
Wei Yongjun85556c92012-09-26 20:43:37 -04002390 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002391 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002392 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002393 goto exit_group_info;
2394 }
2395 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2396 &(meta_group_info[i]->bb_state));
2397
2398 /*
2399 * initialize bb_free to be able to skip
2400 * empty groups without initialization
2401 */
2402 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2403 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002404 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002405 } else {
2406 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002407 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002408 }
2409
2410 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002411 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002412 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002413 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002414
2415#ifdef DOUBLE_CHECK
2416 {
2417 struct buffer_head *bh;
2418 meta_group_info[i]->bb_bitmap =
2419 kmalloc(sb->s_blocksize, GFP_KERNEL);
2420 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2421 bh = ext4_read_block_bitmap(sb, group);
2422 BUG_ON(bh == NULL);
2423 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2424 sb->s_blocksize);
2425 put_bh(bh);
2426 }
2427#endif
2428
2429 return 0;
2430
2431exit_group_info:
2432 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002433 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002434 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002435 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2436 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002437exit_meta_group_info:
2438 return -ENOMEM;
2439} /* ext4_mb_add_groupinfo */
2440
Alex Tomasc9de5602008-01-29 00:19:52 -05002441static int ext4_mb_init_backend(struct super_block *sb)
2442{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002443 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002444 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002445 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002446 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002447 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002448 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002449
Theodore Ts'o28623c22012-09-05 01:31:50 -04002450 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2451 if (err)
2452 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002453
Alex Tomasc9de5602008-01-29 00:19:52 -05002454 sbi->s_buddy_cache = new_inode(sb);
2455 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002456 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002457 goto err_freesgi;
2458 }
Yu Jian48e60612011-08-01 17:41:39 -04002459 /* To avoid potentially colliding with an valid on-disk inode number,
2460 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2461 * not in the inode hash, so it should never be found by iget(), but
2462 * this will avoid confusion if it ever shows up during debugging. */
2463 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002464 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002465 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002466 desc = ext4_get_group_desc(sb, i, NULL);
2467 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002468 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002469 goto err_freebuddy;
2470 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002471 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2472 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002473 }
2474
2475 return 0;
2476
2477err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002478 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002479 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002480 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002481 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002482 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002483 kfree(sbi->s_group_info[i]);
2484 iput(sbi->s_buddy_cache);
2485err_freesgi:
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002486 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002487 return -ENOMEM;
2488}
2489
Eric Sandeen2892c152011-02-12 08:12:18 -05002490static void ext4_groupinfo_destroy_slabs(void)
2491{
2492 int i;
2493
2494 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2495 if (ext4_groupinfo_caches[i])
2496 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2497 ext4_groupinfo_caches[i] = NULL;
2498 }
2499}
2500
2501static int ext4_groupinfo_create_slab(size_t size)
2502{
2503 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2504 int slab_size;
2505 int blocksize_bits = order_base_2(size);
2506 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2507 struct kmem_cache *cachep;
2508
2509 if (cache_index >= NR_GRPINFO_CACHES)
2510 return -EINVAL;
2511
2512 if (unlikely(cache_index < 0))
2513 cache_index = 0;
2514
2515 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2516 if (ext4_groupinfo_caches[cache_index]) {
2517 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2518 return 0; /* Already created */
2519 }
2520
2521 slab_size = offsetof(struct ext4_group_info,
2522 bb_counters[blocksize_bits + 2]);
2523
2524 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2525 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2526 NULL);
2527
Tao Ma823ba012011-07-11 18:26:01 -04002528 ext4_groupinfo_caches[cache_index] = cachep;
2529
Eric Sandeen2892c152011-02-12 08:12:18 -05002530 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2531 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002532 printk(KERN_EMERG
2533 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002534 return -ENOMEM;
2535 }
2536
Eric Sandeen2892c152011-02-12 08:12:18 -05002537 return 0;
2538}
2539
Akira Fujita9d990122012-05-28 14:19:25 -04002540int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002541{
2542 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002543 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002544 unsigned offset;
2545 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002546 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002547
Eric Sandeen19278052009-08-25 22:36:25 -04002548 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002549
2550 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2551 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002552 ret = -ENOMEM;
2553 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002554 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002555
Eric Sandeen19278052009-08-25 22:36:25 -04002556 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002557 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2558 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002559 ret = -ENOMEM;
2560 goto out;
2561 }
2562
Eric Sandeen2892c152011-02-12 08:12:18 -05002563 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2564 if (ret < 0)
2565 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002566
2567 /* order 0 is regular bitmap */
2568 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2569 sbi->s_mb_offsets[0] = 0;
2570
2571 i = 1;
2572 offset = 0;
2573 max = sb->s_blocksize << 2;
2574 do {
2575 sbi->s_mb_offsets[i] = offset;
2576 sbi->s_mb_maxs[i] = max;
2577 offset += 1 << (sb->s_blocksize_bits - i);
2578 max = max >> 1;
2579 i++;
2580 } while (i <= sb->s_blocksize_bits + 1);
2581
Alex Tomasc9de5602008-01-29 00:19:52 -05002582 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002583 spin_lock_init(&sbi->s_bal_lock);
2584
2585 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2586 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2587 sbi->s_mb_stats = MB_DEFAULT_STATS;
2588 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2589 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002590 /*
2591 * The default group preallocation is 512, which for 4k block
2592 * sizes translates to 2 megabytes. However for bigalloc file
2593 * systems, this is probably too big (i.e, if the cluster size
2594 * is 1 megabyte, then group preallocation size becomes half a
2595 * gigabyte!). As a default, we will keep a two megabyte
2596 * group pralloc size for cluster sizes up to 64k, and after
2597 * that, we will force a minimum group preallocation size of
2598 * 32 clusters. This translates to 8 megs when the cluster
2599 * size is 256k, and 32 megs when the cluster size is 1 meg,
2600 * which seems reasonable as a default.
2601 */
2602 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2603 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002604 /*
2605 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2606 * to the lowest multiple of s_stripe which is bigger than
2607 * the s_mb_group_prealloc as determined above. We want
2608 * the preallocation size to be an exact multiple of the
2609 * RAID stripe size so that preallocations don't fragment
2610 * the stripes.
2611 */
2612 if (sbi->s_stripe > 1) {
2613 sbi->s_mb_group_prealloc = roundup(
2614 sbi->s_mb_group_prealloc, sbi->s_stripe);
2615 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002616
Eric Sandeen730c2132008-09-13 15:23:29 -04002617 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002618 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002619 ret = -ENOMEM;
Tao Ma7aa0bae2011-10-06 10:22:28 -04002620 goto out_free_groupinfo_slab;
Alex Tomasc9de5602008-01-29 00:19:52 -05002621 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002622 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002623 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002624 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002625 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002626 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2627 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002628 spin_lock_init(&lg->lg_prealloc_lock);
2629 }
2630
Yu Jian79a77c52011-08-01 17:41:46 -04002631 /* init file for buddy data */
2632 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002633 if (ret != 0)
2634 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002635
Theodore Ts'o296c3552009-09-30 00:32:42 -04002636 if (sbi->s_proc)
2637 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2638 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002639
Tao Ma7aa0bae2011-10-06 10:22:28 -04002640 return 0;
2641
2642out_free_locality_groups:
2643 free_percpu(sbi->s_locality_groups);
2644 sbi->s_locality_groups = NULL;
2645out_free_groupinfo_slab:
2646 ext4_groupinfo_destroy_slabs();
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002647out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002648 kfree(sbi->s_mb_offsets);
2649 sbi->s_mb_offsets = NULL;
2650 kfree(sbi->s_mb_maxs);
2651 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002652 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002653}
2654
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002655/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002656static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2657{
2658 struct ext4_prealloc_space *pa;
2659 struct list_head *cur, *tmp;
2660 int count = 0;
2661
2662 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2663 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2664 list_del(&pa->pa_group_list);
2665 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002666 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002667 }
2668 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002669 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002670
2671}
2672
2673int ext4_mb_release(struct super_block *sb)
2674{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002675 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002676 ext4_group_t i;
2677 int num_meta_group_infos;
2678 struct ext4_group_info *grinfo;
2679 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002680 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002681
Salman Qazi95599962012-05-31 23:52:14 -04002682 if (sbi->s_proc)
2683 remove_proc_entry("mb_groups", sbi->s_proc);
2684
Alex Tomasc9de5602008-01-29 00:19:52 -05002685 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002686 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002687 grinfo = ext4_get_group_info(sb, i);
2688#ifdef DOUBLE_CHECK
2689 kfree(grinfo->bb_bitmap);
2690#endif
2691 ext4_lock_group(sb, i);
2692 ext4_mb_cleanup_pa(grinfo);
2693 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002694 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002695 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002696 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002697 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2698 EXT4_DESC_PER_BLOCK_BITS(sb);
2699 for (i = 0; i < num_meta_group_infos; i++)
2700 kfree(sbi->s_group_info[i]);
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002701 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002702 }
2703 kfree(sbi->s_mb_offsets);
2704 kfree(sbi->s_mb_maxs);
2705 if (sbi->s_buddy_cache)
2706 iput(sbi->s_buddy_cache);
2707 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002708 ext4_msg(sb, KERN_INFO,
2709 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002710 atomic_read(&sbi->s_bal_allocated),
2711 atomic_read(&sbi->s_bal_reqs),
2712 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002713 ext4_msg(sb, KERN_INFO,
2714 "mballoc: %u extents scanned, %u goal hits, "
2715 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002716 atomic_read(&sbi->s_bal_ex_scanned),
2717 atomic_read(&sbi->s_bal_goals),
2718 atomic_read(&sbi->s_bal_2orders),
2719 atomic_read(&sbi->s_bal_breaks),
2720 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002721 ext4_msg(sb, KERN_INFO,
2722 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002723 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002724 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002725 ext4_msg(sb, KERN_INFO,
2726 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002727 atomic_read(&sbi->s_mb_preallocated),
2728 atomic_read(&sbi->s_mb_discarded));
2729 }
2730
Eric Sandeen730c2132008-09-13 15:23:29 -04002731 free_percpu(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05002732
2733 return 0;
2734}
2735
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002736static inline int ext4_issue_discard(struct super_block *sb,
Theodore Ts'o84130192011-09-09 18:50:51 -04002737 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
Jiaying Zhang5c521832010-07-27 11:56:05 -04002738{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002739 ext4_fsblk_t discard_block;
2740
Theodore Ts'o84130192011-09-09 18:50:51 -04002741 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2742 ext4_group_first_block_no(sb, block_group));
2743 count = EXT4_C2B(EXT4_SB(sb), count);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002744 trace_ext4_discard_blocks(sb,
2745 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002746 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002747}
2748
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002749/*
2750 * This function is called by the jbd2 layer once the commit has finished,
2751 * so we know we can free the blocks that were released with that commit.
2752 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002753static void ext4_free_data_callback(struct super_block *sb,
2754 struct ext4_journal_cb_entry *jce,
2755 int rc)
Alex Tomasc9de5602008-01-29 00:19:52 -05002756{
Bobi Jam18aadd42012-02-20 17:53:02 -05002757 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
Alex Tomasc9de5602008-01-29 00:19:52 -05002758 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002759 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002760 int err, count = 0, count2 = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002761
Bobi Jam18aadd42012-02-20 17:53:02 -05002762 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2763 entry->efd_count, entry->efd_group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002764
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05002765 if (test_opt(sb, DISCARD)) {
2766 err = ext4_issue_discard(sb, entry->efd_group,
2767 entry->efd_start_cluster,
2768 entry->efd_count);
2769 if (err && err != -EOPNOTSUPP)
2770 ext4_msg(sb, KERN_WARNING, "discard request in"
2771 " group:%d block:%d count:%d failed"
2772 " with %d", entry->efd_group,
2773 entry->efd_start_cluster,
2774 entry->efd_count, err);
2775 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002776
Bobi Jam18aadd42012-02-20 17:53:02 -05002777 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2778 /* we expect to find existing buddy because it's pinned */
2779 BUG_ON(err != 0);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002780
Alex Tomasc9de5602008-01-29 00:19:52 -05002781
Bobi Jam18aadd42012-02-20 17:53:02 -05002782 db = e4b.bd_info;
2783 /* there are blocks to put in buddy to make them really free */
2784 count += entry->efd_count;
2785 count2++;
2786 ext4_lock_group(sb, entry->efd_group);
2787 /* Take it out of per group rb tree */
2788 rb_erase(&entry->efd_node, &(db->bb_free_root));
2789 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002790
Bobi Jam18aadd42012-02-20 17:53:02 -05002791 /*
2792 * Clear the trimmed flag for the group so that the next
2793 * ext4_trim_fs can trim it.
2794 * If the volume is mounted with -o discard, online discard
2795 * is supported and the free blocks will be trimmed online.
2796 */
2797 if (!test_opt(sb, DISCARD))
2798 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2799
2800 if (!db->bb_free_root.rb_node) {
2801 /* No more items in the per group rb tree
2802 * balance refcounts from ext4_mb_free_metadata()
Tao Ma3d56b8d2011-07-11 00:03:38 -04002803 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002804 page_cache_release(e4b.bd_buddy_page);
2805 page_cache_release(e4b.bd_bitmap_page);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002806 }
Bobi Jam18aadd42012-02-20 17:53:02 -05002807 ext4_unlock_group(sb, entry->efd_group);
2808 kmem_cache_free(ext4_free_data_cachep, entry);
2809 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002810
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002811 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002812}
2813
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002814int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002815{
Theodore Ts'o16828082010-10-27 21:30:09 -04002816 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2817 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002818 if (ext4_pspace_cachep == NULL)
2819 return -ENOMEM;
2820
Theodore Ts'o16828082010-10-27 21:30:09 -04002821 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2822 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002823 if (ext4_ac_cachep == NULL) {
2824 kmem_cache_destroy(ext4_pspace_cachep);
2825 return -ENOMEM;
2826 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002827
Bobi Jam18aadd42012-02-20 17:53:02 -05002828 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2829 SLAB_RECLAIM_ACCOUNT);
2830 if (ext4_free_data_cachep == NULL) {
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002831 kmem_cache_destroy(ext4_pspace_cachep);
2832 kmem_cache_destroy(ext4_ac_cachep);
2833 return -ENOMEM;
2834 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002835 return 0;
2836}
2837
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002838void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002839{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002840 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002841 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2842 * before destroying the slab cache.
2843 */
2844 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002845 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002846 kmem_cache_destroy(ext4_ac_cachep);
Bobi Jam18aadd42012-02-20 17:53:02 -05002847 kmem_cache_destroy(ext4_free_data_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002848 ext4_groupinfo_destroy_slabs();
Alex Tomasc9de5602008-01-29 00:19:52 -05002849}
2850
2851
2852/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002853 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002854 * Returns 0 if success or error code
2855 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002856static noinline_for_stack int
2857ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002858 handle_t *handle, unsigned int reserv_clstrs)
Alex Tomasc9de5602008-01-29 00:19:52 -05002859{
2860 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002861 struct ext4_group_desc *gdp;
2862 struct buffer_head *gdp_bh;
2863 struct ext4_sb_info *sbi;
2864 struct super_block *sb;
2865 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002866 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002867
2868 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2869 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2870
2871 sb = ac->ac_sb;
2872 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002873
2874 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002875 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002876 if (!bitmap_bh)
2877 goto out_err;
2878
2879 err = ext4_journal_get_write_access(handle, bitmap_bh);
2880 if (err)
2881 goto out_err;
2882
2883 err = -EIO;
2884 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2885 if (!gdp)
2886 goto out_err;
2887
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002888 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002889 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002890
Alex Tomasc9de5602008-01-29 00:19:52 -05002891 err = ext4_journal_get_write_access(handle, gdp_bh);
2892 if (err)
2893 goto out_err;
2894
Akinobu Mitabda00de2010-03-03 23:53:25 -05002895 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002896
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002897 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002898 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002899 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002900 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002901 /* File system mounted not to panic on error
2902 * Fix the bitmap and repeat the block allocation
2903 * We leak some of the blocks here.
2904 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002905 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002906 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2907 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002908 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002909 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002910 if (!err)
2911 err = -EAGAIN;
2912 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002913 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002914
2915 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002916#ifdef AGGRESSIVE_CHECK
2917 {
2918 int i;
2919 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2920 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2921 bitmap_bh->b_data));
2922 }
2923 }
2924#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002925 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2926 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002927 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2928 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002929 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002930 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002931 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002932 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002933 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2934 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002935 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002936 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002937
2938 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002939 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002940 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002941 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002942 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002943 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2944 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002945 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2946 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002947
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002948 if (sbi->s_log_groups_per_flex) {
2949 ext4_group_t flex_group = ext4_flex_group(sbi,
2950 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002951 atomic64_sub(ac->ac_b_ex.fe_len,
2952 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002953 }
2954
Frank Mayhar03901312009-01-07 00:06:22 -05002955 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002956 if (err)
2957 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002958 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002959
2960out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002961 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002962 return err;
2963}
2964
2965/*
2966 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002967 * Group request are normalized to s_mb_group_prealloc, which goes to
2968 * s_strip if we set the same via mount option.
2969 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002970 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002971 *
2972 * XXX: should we try to preallocate more than the group has now?
2973 */
2974static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2975{
2976 struct super_block *sb = ac->ac_sb;
2977 struct ext4_locality_group *lg = ac->ac_lg;
2978
2979 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002980 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002981 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002982 current->pid, ac->ac_g_ex.fe_len);
2983}
2984
2985/*
2986 * Normalization means making request better in terms of
2987 * size and alignment
2988 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002989static noinline_for_stack void
2990ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05002991 struct ext4_allocation_request *ar)
2992{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002993 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002994 int bsbits, max;
2995 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05002996 loff_t size, start_off;
2997 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04002998 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05002999 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003000 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05003001
3002 /* do normalize only data requests, metadata requests
3003 do not need preallocation */
3004 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3005 return;
3006
3007 /* sometime caller may want exact blocks */
3008 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3009 return;
3010
3011 /* caller may indicate that preallocation isn't
3012 * required (it's a tail, for example) */
3013 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3014 return;
3015
3016 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3017 ext4_mb_normalize_group_request(ac);
3018 return ;
3019 }
3020
3021 bsbits = ac->ac_sb->s_blocksize_bits;
3022
3023 /* first, let's learn actual file size
3024 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003025 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003026 size = size << bsbits;
3027 if (size < i_size_read(ac->ac_inode))
3028 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003029 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003030
Valerie Clement19304792008-05-13 19:31:14 -04003031 /* max size of free chunks */
3032 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003033
Valerie Clement19304792008-05-13 19:31:14 -04003034#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3035 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003036
3037 /* first, try to predict filesize */
3038 /* XXX: should this table be tunable? */
3039 start_off = 0;
3040 if (size <= 16 * 1024) {
3041 size = 16 * 1024;
3042 } else if (size <= 32 * 1024) {
3043 size = 32 * 1024;
3044 } else if (size <= 64 * 1024) {
3045 size = 64 * 1024;
3046 } else if (size <= 128 * 1024) {
3047 size = 128 * 1024;
3048 } else if (size <= 256 * 1024) {
3049 size = 256 * 1024;
3050 } else if (size <= 512 * 1024) {
3051 size = 512 * 1024;
3052 } else if (size <= 1024 * 1024) {
3053 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003054 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003055 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003056 (21 - bsbits)) << 21;
3057 size = 2 * 1024 * 1024;
3058 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003059 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3060 (22 - bsbits)) << 22;
3061 size = 4 * 1024 * 1024;
3062 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003063 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003064 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3065 (23 - bsbits)) << 23;
3066 size = 8 * 1024 * 1024;
3067 } else {
3068 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3069 size = ac->ac_o_ex.fe_len << bsbits;
3070 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003071 size = size >> bsbits;
3072 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003073
3074 /* don't cover already allocated blocks in selected range */
3075 if (ar->pleft && start <= ar->lleft) {
3076 size -= ar->lleft + 1 - start;
3077 start = ar->lleft + 1;
3078 }
3079 if (ar->pright && start + size - 1 >= ar->lright)
3080 size -= start + size - ar->lright;
3081
3082 end = start + size;
3083
3084 /* check we don't cross already preallocated blocks */
3085 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003086 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003087 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003088
Alex Tomasc9de5602008-01-29 00:19:52 -05003089 if (pa->pa_deleted)
3090 continue;
3091 spin_lock(&pa->pa_lock);
3092 if (pa->pa_deleted) {
3093 spin_unlock(&pa->pa_lock);
3094 continue;
3095 }
3096
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003097 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3098 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003099
3100 /* PA must not overlap original request */
3101 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3102 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3103
Eric Sandeen38877f42009-08-17 23:55:24 -04003104 /* skip PAs this normalized request doesn't overlap with */
3105 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003106 spin_unlock(&pa->pa_lock);
3107 continue;
3108 }
3109 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3110
Eric Sandeen38877f42009-08-17 23:55:24 -04003111 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003112 if (pa_end <= ac->ac_o_ex.fe_logical) {
3113 BUG_ON(pa_end < start);
3114 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003115 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003116 BUG_ON(pa->pa_lstart > end);
3117 end = pa->pa_lstart;
3118 }
3119 spin_unlock(&pa->pa_lock);
3120 }
3121 rcu_read_unlock();
3122 size = end - start;
3123
3124 /* XXX: extra loop to check we really don't overlap preallocations */
3125 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003126 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003127 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003128
Alex Tomasc9de5602008-01-29 00:19:52 -05003129 spin_lock(&pa->pa_lock);
3130 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003131 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3132 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003133 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3134 }
3135 spin_unlock(&pa->pa_lock);
3136 }
3137 rcu_read_unlock();
3138
3139 if (start + size <= ac->ac_o_ex.fe_logical &&
3140 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003141 ext4_msg(ac->ac_sb, KERN_ERR,
3142 "start %lu, size %lu, fe_logical %lu",
3143 (unsigned long) start, (unsigned long) size,
3144 (unsigned long) ac->ac_o_ex.fe_logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05003145 }
3146 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3147 start > ac->ac_o_ex.fe_logical);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003148 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003149
3150 /* now prepare goal request */
3151
3152 /* XXX: is it better to align blocks WRT to logical
3153 * placement or satisfy big request as is */
3154 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003155 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003156
3157 /* define goal start in order to merge */
3158 if (ar->pright && (ar->lright == (start + size))) {
3159 /* merge to the right */
3160 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3161 &ac->ac_f_ex.fe_group,
3162 &ac->ac_f_ex.fe_start);
3163 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3164 }
3165 if (ar->pleft && (ar->lleft + 1 == start)) {
3166 /* merge to the left */
3167 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3168 &ac->ac_f_ex.fe_group,
3169 &ac->ac_f_ex.fe_start);
3170 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3171 }
3172
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003173 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003174 (unsigned) orig_size, (unsigned) start);
3175}
3176
3177static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3178{
3179 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3180
3181 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3182 atomic_inc(&sbi->s_bal_reqs);
3183 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003184 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003185 atomic_inc(&sbi->s_bal_success);
3186 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3187 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3188 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3189 atomic_inc(&sbi->s_bal_goals);
3190 if (ac->ac_found > sbi->s_mb_max_to_scan)
3191 atomic_inc(&sbi->s_bal_breaks);
3192 }
3193
Theodore Ts'o296c3552009-09-30 00:32:42 -04003194 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3195 trace_ext4_mballoc_alloc(ac);
3196 else
3197 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003198}
3199
3200/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003201 * Called on failure; free up any blocks from the inode PA for this
3202 * context. We don't need this for MB_GROUP_PA because we only change
3203 * pa_free in ext4_mb_release_context(), but on failure, we've already
3204 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3205 */
3206static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3207{
3208 struct ext4_prealloc_space *pa = ac->ac_pa;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003209
Zheng Liu400db9d2012-05-28 17:53:53 -04003210 if (pa && pa->pa_type == MB_INODE_PA)
3211 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003212}
3213
3214/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003215 * use blocks preallocated to inode
3216 */
3217static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3218 struct ext4_prealloc_space *pa)
3219{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003220 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003221 ext4_fsblk_t start;
3222 ext4_fsblk_t end;
3223 int len;
3224
3225 /* found preallocated blocks, use them */
3226 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003227 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3228 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3229 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003230 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3231 &ac->ac_b_ex.fe_start);
3232 ac->ac_b_ex.fe_len = len;
3233 ac->ac_status = AC_STATUS_FOUND;
3234 ac->ac_pa = pa;
3235
3236 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003237 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003238 BUG_ON(pa->pa_free < len);
3239 pa->pa_free -= len;
3240
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003241 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003242}
3243
3244/*
3245 * use blocks preallocated to locality group
3246 */
3247static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3248 struct ext4_prealloc_space *pa)
3249{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003250 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003251
Alex Tomasc9de5602008-01-29 00:19:52 -05003252 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3253 &ac->ac_b_ex.fe_group,
3254 &ac->ac_b_ex.fe_start);
3255 ac->ac_b_ex.fe_len = len;
3256 ac->ac_status = AC_STATUS_FOUND;
3257 ac->ac_pa = pa;
3258
3259 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003260 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003261 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003262 * in on-disk bitmap -- see ext4_mb_release_context()
3263 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003264 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003265 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003266}
3267
3268/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003269 * Return the prealloc space that have minimal distance
3270 * from the goal block. @cpa is the prealloc
3271 * space that is having currently known minimal distance
3272 * from the goal block.
3273 */
3274static struct ext4_prealloc_space *
3275ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3276 struct ext4_prealloc_space *pa,
3277 struct ext4_prealloc_space *cpa)
3278{
3279 ext4_fsblk_t cur_distance, new_distance;
3280
3281 if (cpa == NULL) {
3282 atomic_inc(&pa->pa_count);
3283 return pa;
3284 }
3285 cur_distance = abs(goal_block - cpa->pa_pstart);
3286 new_distance = abs(goal_block - pa->pa_pstart);
3287
Coly Li5a54b2f2011-02-24 14:10:05 -05003288 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003289 return cpa;
3290
3291 /* drop the previous reference */
3292 atomic_dec(&cpa->pa_count);
3293 atomic_inc(&pa->pa_count);
3294 return pa;
3295}
3296
3297/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003298 * search goal blocks in preallocated space
3299 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003300static noinline_for_stack int
3301ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003302{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003303 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003304 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003305 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3306 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003307 struct ext4_prealloc_space *pa, *cpa = NULL;
3308 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003309
3310 /* only data can be preallocated */
3311 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3312 return 0;
3313
3314 /* first, try per-file preallocation */
3315 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003316 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003317
3318 /* all fields in this condition don't change,
3319 * so we can skip locking for them */
3320 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003321 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3322 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003323 continue;
3324
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003325 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003326 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003327 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3328 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003329 continue;
3330
Alex Tomasc9de5602008-01-29 00:19:52 -05003331 /* found preallocated blocks, use them */
3332 spin_lock(&pa->pa_lock);
3333 if (pa->pa_deleted == 0 && pa->pa_free) {
3334 atomic_inc(&pa->pa_count);
3335 ext4_mb_use_inode_pa(ac, pa);
3336 spin_unlock(&pa->pa_lock);
3337 ac->ac_criteria = 10;
3338 rcu_read_unlock();
3339 return 1;
3340 }
3341 spin_unlock(&pa->pa_lock);
3342 }
3343 rcu_read_unlock();
3344
3345 /* can we use group allocation? */
3346 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3347 return 0;
3348
3349 /* inode may have no locality group for some reason */
3350 lg = ac->ac_lg;
3351 if (lg == NULL)
3352 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003353 order = fls(ac->ac_o_ex.fe_len) - 1;
3354 if (order > PREALLOC_TB_SIZE - 1)
3355 /* The max size of hash table is PREALLOC_TB_SIZE */
3356 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003357
Akinobu Mitabda00de2010-03-03 23:53:25 -05003358 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003359 /*
3360 * search for the prealloc space that is having
3361 * minimal distance from the goal block.
3362 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003363 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3364 rcu_read_lock();
3365 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3366 pa_inode_list) {
3367 spin_lock(&pa->pa_lock);
3368 if (pa->pa_deleted == 0 &&
3369 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003370
3371 cpa = ext4_mb_check_group_pa(goal_block,
3372 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003373 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003374 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003375 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003376 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003377 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003378 if (cpa) {
3379 ext4_mb_use_group_pa(ac, cpa);
3380 ac->ac_criteria = 20;
3381 return 1;
3382 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003383 return 0;
3384}
3385
3386/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003387 * the function goes through all block freed in the group
3388 * but not yet committed and marks them used in in-core bitmap.
3389 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003390 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003391 */
3392static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3393 ext4_group_t group)
3394{
3395 struct rb_node *n;
3396 struct ext4_group_info *grp;
3397 struct ext4_free_data *entry;
3398
3399 grp = ext4_get_group_info(sb, group);
3400 n = rb_first(&(grp->bb_free_root));
3401
3402 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003403 entry = rb_entry(n, struct ext4_free_data, efd_node);
3404 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003405 n = rb_next(n);
3406 }
3407 return;
3408}
3409
3410/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003411 * the function goes through all preallocation in this group and marks them
3412 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003413 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003414 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003415static noinline_for_stack
3416void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003417 ext4_group_t group)
3418{
3419 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3420 struct ext4_prealloc_space *pa;
3421 struct list_head *cur;
3422 ext4_group_t groupnr;
3423 ext4_grpblk_t start;
3424 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003425 int len;
3426
3427 /* all form of preallocation discards first load group,
3428 * so the only competing code is preallocation use.
3429 * we don't need any locking here
3430 * notice we do NOT ignore preallocations with pa_deleted
3431 * otherwise we could leave used blocks available for
3432 * allocation in buddy when concurrent ext4_mb_put_pa()
3433 * is dropping preallocation
3434 */
3435 list_for_each(cur, &grp->bb_prealloc_list) {
3436 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3437 spin_lock(&pa->pa_lock);
3438 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3439 &groupnr, &start);
3440 len = pa->pa_len;
3441 spin_unlock(&pa->pa_lock);
3442 if (unlikely(len == 0))
3443 continue;
3444 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003445 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003446 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003447 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003448 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003449}
3450
3451static void ext4_mb_pa_callback(struct rcu_head *head)
3452{
3453 struct ext4_prealloc_space *pa;
3454 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003455
3456 BUG_ON(atomic_read(&pa->pa_count));
3457 BUG_ON(pa->pa_deleted == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05003458 kmem_cache_free(ext4_pspace_cachep, pa);
3459}
3460
3461/*
3462 * drops a reference to preallocated space descriptor
3463 * if this was the last reference and the space is consumed
3464 */
3465static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3466 struct super_block *sb, struct ext4_prealloc_space *pa)
3467{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003468 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003469 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003470
Alex Tomasc9de5602008-01-29 00:19:52 -05003471 /* in this short window concurrent discard can set pa_deleted */
3472 spin_lock(&pa->pa_lock);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003473 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3474 spin_unlock(&pa->pa_lock);
3475 return;
3476 }
3477
Alex Tomasc9de5602008-01-29 00:19:52 -05003478 if (pa->pa_deleted == 1) {
3479 spin_unlock(&pa->pa_lock);
3480 return;
3481 }
3482
3483 pa->pa_deleted = 1;
3484 spin_unlock(&pa->pa_lock);
3485
Eric Sandeend33a1972009-03-16 23:25:40 -04003486 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003487 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003488 * If doing group-based preallocation, pa_pstart may be in the
3489 * next group when pa is used up
3490 */
3491 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003492 grp_blk--;
3493
Lukas Czernerbd862982013-04-03 23:32:34 -04003494 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003495
3496 /*
3497 * possible race:
3498 *
3499 * P1 (buddy init) P2 (regular allocation)
3500 * find block B in PA
3501 * copy on-disk bitmap to buddy
3502 * mark B in on-disk bitmap
3503 * drop PA from group
3504 * mark all PAs in buddy
3505 *
3506 * thus, P1 initializes buddy with B available. to prevent this
3507 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3508 * against that pair
3509 */
3510 ext4_lock_group(sb, grp);
3511 list_del(&pa->pa_group_list);
3512 ext4_unlock_group(sb, grp);
3513
3514 spin_lock(pa->pa_obj_lock);
3515 list_del_rcu(&pa->pa_inode_list);
3516 spin_unlock(pa->pa_obj_lock);
3517
3518 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3519}
3520
3521/*
3522 * creates new preallocated space for given inode
3523 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003524static noinline_for_stack int
3525ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003526{
3527 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003528 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003529 struct ext4_prealloc_space *pa;
3530 struct ext4_group_info *grp;
3531 struct ext4_inode_info *ei;
3532
3533 /* preallocate only when found space is larger then requested */
3534 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3535 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3536 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3537
3538 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3539 if (pa == NULL)
3540 return -ENOMEM;
3541
3542 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3543 int winl;
3544 int wins;
3545 int win;
3546 int offs;
3547
3548 /* we can't allocate as much as normalizer wants.
3549 * so, found space must get proper lstart
3550 * to cover original request */
3551 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3552 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3553
3554 /* we're limited by original request in that
3555 * logical block must be covered any way
3556 * winl is window we can move our chunk within */
3557 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3558
3559 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003560 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003561
3562 /* the smallest one defines real window */
3563 win = min(winl, wins);
3564
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003565 offs = ac->ac_o_ex.fe_logical %
3566 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003567 if (offs && offs < win)
3568 win = offs;
3569
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003570 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003571 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003572 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3573 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3574 }
3575
3576 /* preallocation can change ac_b_ex, thus we store actually
3577 * allocated blocks for history */
3578 ac->ac_f_ex = ac->ac_b_ex;
3579
3580 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3581 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3582 pa->pa_len = ac->ac_b_ex.fe_len;
3583 pa->pa_free = pa->pa_len;
3584 atomic_set(&pa->pa_count, 1);
3585 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003586 INIT_LIST_HEAD(&pa->pa_inode_list);
3587 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003588 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003589 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003590
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003591 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003592 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003593 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003594
3595 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003596 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003597
3598 ei = EXT4_I(ac->ac_inode);
3599 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3600
3601 pa->pa_obj_lock = &ei->i_prealloc_lock;
3602 pa->pa_inode = ac->ac_inode;
3603
3604 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3605 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3606 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3607
3608 spin_lock(pa->pa_obj_lock);
3609 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3610 spin_unlock(pa->pa_obj_lock);
3611
3612 return 0;
3613}
3614
3615/*
3616 * creates new preallocated space for locality group inodes belongs to
3617 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003618static noinline_for_stack int
3619ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003620{
3621 struct super_block *sb = ac->ac_sb;
3622 struct ext4_locality_group *lg;
3623 struct ext4_prealloc_space *pa;
3624 struct ext4_group_info *grp;
3625
3626 /* preallocate only when found space is larger then requested */
3627 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3628 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3629 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3630
3631 BUG_ON(ext4_pspace_cachep == NULL);
3632 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3633 if (pa == NULL)
3634 return -ENOMEM;
3635
3636 /* preallocation can change ac_b_ex, thus we store actually
3637 * allocated blocks for history */
3638 ac->ac_f_ex = ac->ac_b_ex;
3639
3640 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3641 pa->pa_lstart = pa->pa_pstart;
3642 pa->pa_len = ac->ac_b_ex.fe_len;
3643 pa->pa_free = pa->pa_len;
3644 atomic_set(&pa->pa_count, 1);
3645 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003646 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003647 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003648 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003649 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003650
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003651 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003652 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3653 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003654
3655 ext4_mb_use_group_pa(ac, pa);
3656 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3657
3658 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3659 lg = ac->ac_lg;
3660 BUG_ON(lg == NULL);
3661
3662 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3663 pa->pa_inode = NULL;
3664
3665 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3666 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3667 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3668
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003669 /*
3670 * We will later add the new pa to the right bucket
3671 * after updating the pa_free in ext4_mb_release_context
3672 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003673 return 0;
3674}
3675
3676static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3677{
3678 int err;
3679
3680 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3681 err = ext4_mb_new_group_pa(ac);
3682 else
3683 err = ext4_mb_new_inode_pa(ac);
3684 return err;
3685}
3686
3687/*
3688 * finds all unused blocks in on-disk bitmap, frees them in
3689 * in-core bitmap and buddy.
3690 * @pa must be unlinked from inode and group lists, so that
3691 * nobody else can find/use it.
3692 * the caller MUST hold group/inode locks.
3693 * TODO: optimize the case when there are no in-core structures yet
3694 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003695static noinline_for_stack int
3696ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003697 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003698{
Alex Tomasc9de5602008-01-29 00:19:52 -05003699 struct super_block *sb = e4b->bd_sb;
3700 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003701 unsigned int end;
3702 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003703 ext4_group_t group;
3704 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003705 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003706 int err = 0;
3707 int free = 0;
3708
3709 BUG_ON(pa->pa_deleted == 0);
3710 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003711 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003712 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3713 end = bit + pa->pa_len;
3714
Alex Tomasc9de5602008-01-29 00:19:52 -05003715 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003716 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003717 if (bit >= end)
3718 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003719 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003720 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003721 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3722 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003723 free += next - bit;
3724
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003725 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003726 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3727 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003728 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003729 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3730 bit = next + 1;
3731 }
3732 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003733 ext4_msg(e4b->bd_sb, KERN_CRIT,
3734 "pa %p: logic %lu, phys. %lu, len %lu",
3735 pa, (unsigned long) pa->pa_lstart,
3736 (unsigned long) pa->pa_pstart,
3737 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003738 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003739 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003740 /*
3741 * pa is already deleted so we use the value obtained
3742 * from the bitmap and continue.
3743 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003744 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003745 atomic_add(free, &sbi->s_mb_discarded);
3746
3747 return err;
3748}
3749
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003750static noinline_for_stack int
3751ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003752 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003753{
Alex Tomasc9de5602008-01-29 00:19:52 -05003754 struct super_block *sb = e4b->bd_sb;
3755 ext4_group_t group;
3756 ext4_grpblk_t bit;
3757
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003758 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003759 BUG_ON(pa->pa_deleted == 0);
3760 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3761 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3762 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3763 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003764 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003765
3766 return 0;
3767}
3768
3769/*
3770 * releases all preallocations in given group
3771 *
3772 * first, we need to decide discard policy:
3773 * - when do we discard
3774 * 1) ENOSPC
3775 * - how many do we discard
3776 * 1) how many requested
3777 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003778static noinline_for_stack int
3779ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003780 ext4_group_t group, int needed)
3781{
3782 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3783 struct buffer_head *bitmap_bh = NULL;
3784 struct ext4_prealloc_space *pa, *tmp;
3785 struct list_head list;
3786 struct ext4_buddy e4b;
3787 int err;
3788 int busy = 0;
3789 int free = 0;
3790
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003791 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003792
3793 if (list_empty(&grp->bb_prealloc_list))
3794 return 0;
3795
Theodore Ts'o574ca172008-07-11 19:27:31 -04003796 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003797 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003798 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003799 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003800 }
3801
3802 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003803 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003804 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003805 put_bh(bitmap_bh);
3806 return 0;
3807 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003808
3809 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003810 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003811
Alex Tomasc9de5602008-01-29 00:19:52 -05003812 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003813repeat:
3814 ext4_lock_group(sb, group);
3815 list_for_each_entry_safe(pa, tmp,
3816 &grp->bb_prealloc_list, pa_group_list) {
3817 spin_lock(&pa->pa_lock);
3818 if (atomic_read(&pa->pa_count)) {
3819 spin_unlock(&pa->pa_lock);
3820 busy = 1;
3821 continue;
3822 }
3823 if (pa->pa_deleted) {
3824 spin_unlock(&pa->pa_lock);
3825 continue;
3826 }
3827
3828 /* seems this one can be freed ... */
3829 pa->pa_deleted = 1;
3830
3831 /* we can trust pa_free ... */
3832 free += pa->pa_free;
3833
3834 spin_unlock(&pa->pa_lock);
3835
3836 list_del(&pa->pa_group_list);
3837 list_add(&pa->u.pa_tmp_list, &list);
3838 }
3839
3840 /* if we still need more blocks and some PAs were used, try again */
3841 if (free < needed && busy) {
3842 busy = 0;
3843 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003844 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003845 goto repeat;
3846 }
3847
3848 /* found anything to free? */
3849 if (list_empty(&list)) {
3850 BUG_ON(free != 0);
3851 goto out;
3852 }
3853
3854 /* now free all selected PAs */
3855 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3856
3857 /* remove from object (inode or locality group) */
3858 spin_lock(pa->pa_obj_lock);
3859 list_del_rcu(&pa->pa_inode_list);
3860 spin_unlock(pa->pa_obj_lock);
3861
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003862 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003863 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003864 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003865 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003866
3867 list_del(&pa->u.pa_tmp_list);
3868 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3869 }
3870
3871out:
3872 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003873 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003874 put_bh(bitmap_bh);
3875 return free;
3876}
3877
3878/*
3879 * releases all non-used preallocated blocks for given inode
3880 *
3881 * It's important to discard preallocations under i_data_sem
3882 * We don't want another block to be served from the prealloc
3883 * space when we are discarding the inode prealloc space.
3884 *
3885 * FIXME!! Make sure it is valid at all the call sites
3886 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003887void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003888{
3889 struct ext4_inode_info *ei = EXT4_I(inode);
3890 struct super_block *sb = inode->i_sb;
3891 struct buffer_head *bitmap_bh = NULL;
3892 struct ext4_prealloc_space *pa, *tmp;
3893 ext4_group_t group = 0;
3894 struct list_head list;
3895 struct ext4_buddy e4b;
3896 int err;
3897
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003898 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003899 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3900 return;
3901 }
3902
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003903 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003904 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003905
3906 INIT_LIST_HEAD(&list);
3907
3908repeat:
3909 /* first, collect all pa's in the inode */
3910 spin_lock(&ei->i_prealloc_lock);
3911 while (!list_empty(&ei->i_prealloc_list)) {
3912 pa = list_entry(ei->i_prealloc_list.next,
3913 struct ext4_prealloc_space, pa_inode_list);
3914 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3915 spin_lock(&pa->pa_lock);
3916 if (atomic_read(&pa->pa_count)) {
3917 /* this shouldn't happen often - nobody should
3918 * use preallocation while we're discarding it */
3919 spin_unlock(&pa->pa_lock);
3920 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003921 ext4_msg(sb, KERN_ERR,
3922 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003923 WARN_ON(1);
3924 schedule_timeout_uninterruptible(HZ);
3925 goto repeat;
3926
3927 }
3928 if (pa->pa_deleted == 0) {
3929 pa->pa_deleted = 1;
3930 spin_unlock(&pa->pa_lock);
3931 list_del_rcu(&pa->pa_inode_list);
3932 list_add(&pa->u.pa_tmp_list, &list);
3933 continue;
3934 }
3935
3936 /* someone is deleting pa right now */
3937 spin_unlock(&pa->pa_lock);
3938 spin_unlock(&ei->i_prealloc_lock);
3939
3940 /* we have to wait here because pa_deleted
3941 * doesn't mean pa is already unlinked from
3942 * the list. as we might be called from
3943 * ->clear_inode() the inode will get freed
3944 * and concurrent thread which is unlinking
3945 * pa from inode's list may access already
3946 * freed memory, bad-bad-bad */
3947
3948 /* XXX: if this happens too often, we can
3949 * add a flag to force wait only in case
3950 * of ->clear_inode(), but not in case of
3951 * regular truncate */
3952 schedule_timeout_uninterruptible(HZ);
3953 goto repeat;
3954 }
3955 spin_unlock(&ei->i_prealloc_lock);
3956
3957 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003958 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04003959 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05003960
3961 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003962 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003963 ext4_error(sb, "Error loading buddy information for %u",
3964 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003965 continue;
3966 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003967
Theodore Ts'o574ca172008-07-11 19:27:31 -04003968 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003969 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003970 ext4_error(sb, "Error reading block bitmap for %u",
3971 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003972 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003973 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05003974 }
3975
3976 ext4_lock_group(sb, group);
3977 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003978 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003979 ext4_unlock_group(sb, group);
3980
Jing Zhange39e07f2010-05-14 00:00:00 -04003981 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003982 put_bh(bitmap_bh);
3983
3984 list_del(&pa->u.pa_tmp_list);
3985 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3986 }
3987}
3988
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003989#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05003990static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3991{
3992 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04003993 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003994
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05003995 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05003996 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04003997 return;
3998
Joe Perches7f6a11e2012-03-19 23:09:43 -04003999 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004000 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04004001 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004002 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004003 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004004 "goal %lu/%lu/%lu@%lu, "
4005 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004006 (unsigned long)ac->ac_o_ex.fe_group,
4007 (unsigned long)ac->ac_o_ex.fe_start,
4008 (unsigned long)ac->ac_o_ex.fe_len,
4009 (unsigned long)ac->ac_o_ex.fe_logical,
4010 (unsigned long)ac->ac_g_ex.fe_group,
4011 (unsigned long)ac->ac_g_ex.fe_start,
4012 (unsigned long)ac->ac_g_ex.fe_len,
4013 (unsigned long)ac->ac_g_ex.fe_logical,
4014 (unsigned long)ac->ac_b_ex.fe_group,
4015 (unsigned long)ac->ac_b_ex.fe_start,
4016 (unsigned long)ac->ac_b_ex.fe_len,
4017 (unsigned long)ac->ac_b_ex.fe_logical,
4018 (int)ac->ac_criteria);
Eric Sandeendc9ddd92014-02-20 13:32:10 -05004019 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004020 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04004021 ngroups = ext4_get_groups_count(sb);
4022 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004023 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4024 struct ext4_prealloc_space *pa;
4025 ext4_grpblk_t start;
4026 struct list_head *cur;
4027 ext4_lock_group(sb, i);
4028 list_for_each(cur, &grp->bb_prealloc_list) {
4029 pa = list_entry(cur, struct ext4_prealloc_space,
4030 pa_group_list);
4031 spin_lock(&pa->pa_lock);
4032 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4033 NULL, &start);
4034 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004035 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4036 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004037 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004038 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004039
4040 if (grp->bb_free == 0)
4041 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004042 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004043 i, grp->bb_free, grp->bb_fragments);
4044 }
4045 printk(KERN_ERR "\n");
4046}
4047#else
4048static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4049{
4050 return;
4051}
4052#endif
4053
4054/*
4055 * We use locality group preallocation for small size file. The size of the
4056 * file is determined by the current size or the resulting size after
4057 * allocation which ever is larger
4058 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004059 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004060 */
4061static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4062{
4063 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4064 int bsbits = ac->ac_sb->s_blocksize_bits;
4065 loff_t size, isize;
4066
4067 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4068 return;
4069
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004070 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4071 return;
4072
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004073 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004074 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4075 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004076
Theodore Ts'o50797482009-09-18 13:34:02 -04004077 if ((size == isize) &&
4078 !ext4_fs_is_busy(sbi) &&
4079 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4080 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4081 return;
4082 }
4083
Robin Dongebbe0272011-10-26 05:14:27 -04004084 if (sbi->s_mb_group_prealloc <= 0) {
4085 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4086 return;
4087 }
4088
Alex Tomasc9de5602008-01-29 00:19:52 -05004089 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004090 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004091 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004092 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004093 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004094 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004095
4096 BUG_ON(ac->ac_lg != NULL);
4097 /*
4098 * locality group prealloc space are per cpu. The reason for having
4099 * per cpu locality group is to reduce the contention between block
4100 * request from multiple CPUs.
4101 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09004102 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004103
4104 /* we're going to use group allocation */
4105 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4106
4107 /* serialize all allocations in the group */
4108 mutex_lock(&ac->ac_lg->lg_mutex);
4109}
4110
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004111static noinline_for_stack int
4112ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004113 struct ext4_allocation_request *ar)
4114{
4115 struct super_block *sb = ar->inode->i_sb;
4116 struct ext4_sb_info *sbi = EXT4_SB(sb);
4117 struct ext4_super_block *es = sbi->s_es;
4118 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004119 unsigned int len;
4120 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004121 ext4_grpblk_t block;
4122
4123 /* we can't allocate > group size */
4124 len = ar->len;
4125
4126 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004127 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4128 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004129
4130 /* start searching from the goal */
4131 goal = ar->goal;
4132 if (goal < le32_to_cpu(es->s_first_data_block) ||
4133 goal >= ext4_blocks_count(es))
4134 goal = le32_to_cpu(es->s_first_data_block);
4135 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4136
4137 /* set up allocation goals */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004138 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05004139 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004140 ac->ac_sb = sb;
4141 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004142 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004143 ac->ac_o_ex.fe_group = group;
4144 ac->ac_o_ex.fe_start = block;
4145 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004146 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004147 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004148
4149 /* we have to define context: we'll we work with a file or
4150 * locality group. this is a policy, actually */
4151 ext4_mb_group_or_file(ac);
4152
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004153 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004154 "left: %u/%u, right %u/%u to %swritable\n",
4155 (unsigned) ar->len, (unsigned) ar->logical,
4156 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4157 (unsigned) ar->lleft, (unsigned) ar->pleft,
4158 (unsigned) ar->lright, (unsigned) ar->pright,
4159 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4160 return 0;
4161
4162}
4163
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004164static noinline_for_stack void
4165ext4_mb_discard_lg_preallocations(struct super_block *sb,
4166 struct ext4_locality_group *lg,
4167 int order, int total_entries)
4168{
4169 ext4_group_t group = 0;
4170 struct ext4_buddy e4b;
4171 struct list_head discard_list;
4172 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004173
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004174 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004175
4176 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004177
4178 spin_lock(&lg->lg_prealloc_lock);
4179 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4180 pa_inode_list) {
4181 spin_lock(&pa->pa_lock);
4182 if (atomic_read(&pa->pa_count)) {
4183 /*
4184 * This is the pa that we just used
4185 * for block allocation. So don't
4186 * free that
4187 */
4188 spin_unlock(&pa->pa_lock);
4189 continue;
4190 }
4191 if (pa->pa_deleted) {
4192 spin_unlock(&pa->pa_lock);
4193 continue;
4194 }
4195 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004196 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004197
4198 /* seems this one can be freed ... */
4199 pa->pa_deleted = 1;
4200 spin_unlock(&pa->pa_lock);
4201
4202 list_del_rcu(&pa->pa_inode_list);
4203 list_add(&pa->u.pa_tmp_list, &discard_list);
4204
4205 total_entries--;
4206 if (total_entries <= 5) {
4207 /*
4208 * we want to keep only 5 entries
4209 * allowing it to grow to 8. This
4210 * mak sure we don't call discard
4211 * soon for this list.
4212 */
4213 break;
4214 }
4215 }
4216 spin_unlock(&lg->lg_prealloc_lock);
4217
4218 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4219
Lukas Czernerbd862982013-04-03 23:32:34 -04004220 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004221 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004222 ext4_error(sb, "Error loading buddy information for %u",
4223 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004224 continue;
4225 }
4226 ext4_lock_group(sb, group);
4227 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004228 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004229 ext4_unlock_group(sb, group);
4230
Jing Zhange39e07f2010-05-14 00:00:00 -04004231 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004232 list_del(&pa->u.pa_tmp_list);
4233 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4234 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004235}
4236
4237/*
4238 * We have incremented pa_count. So it cannot be freed at this
4239 * point. Also we hold lg_mutex. So no parallel allocation is
4240 * possible from this lg. That means pa_free cannot be updated.
4241 *
4242 * A parallel ext4_mb_discard_group_preallocations is possible.
4243 * which can cause the lg_prealloc_list to be updated.
4244 */
4245
4246static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4247{
4248 int order, added = 0, lg_prealloc_count = 1;
4249 struct super_block *sb = ac->ac_sb;
4250 struct ext4_locality_group *lg = ac->ac_lg;
4251 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4252
4253 order = fls(pa->pa_free) - 1;
4254 if (order > PREALLOC_TB_SIZE - 1)
4255 /* The max size of hash table is PREALLOC_TB_SIZE */
4256 order = PREALLOC_TB_SIZE - 1;
4257 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004258 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004259 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4260 pa_inode_list) {
4261 spin_lock(&tmp_pa->pa_lock);
4262 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004263 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004264 continue;
4265 }
4266 if (!added && pa->pa_free < tmp_pa->pa_free) {
4267 /* Add to the tail of the previous entry */
4268 list_add_tail_rcu(&pa->pa_inode_list,
4269 &tmp_pa->pa_inode_list);
4270 added = 1;
4271 /*
4272 * we want to count the total
4273 * number of entries in the list
4274 */
4275 }
4276 spin_unlock(&tmp_pa->pa_lock);
4277 lg_prealloc_count++;
4278 }
4279 if (!added)
4280 list_add_tail_rcu(&pa->pa_inode_list,
4281 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004282 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004283
4284 /* Now trim the list to be not more than 8 elements */
4285 if (lg_prealloc_count > 8) {
4286 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004287 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004288 return;
4289 }
4290 return ;
4291}
4292
Alex Tomasc9de5602008-01-29 00:19:52 -05004293/*
4294 * release all resource we used in allocation
4295 */
4296static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4297{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004298 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004299 struct ext4_prealloc_space *pa = ac->ac_pa;
4300 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004301 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004302 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004303 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004304 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4305 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004306 pa->pa_free -= ac->ac_b_ex.fe_len;
4307 pa->pa_len -= ac->ac_b_ex.fe_len;
4308 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004309 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004310 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004311 if (pa) {
4312 /*
4313 * We want to add the pa to the right bucket.
4314 * Remove it from the list and while adding
4315 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004316 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004317 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004318 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004319 spin_lock(pa->pa_obj_lock);
4320 list_del_rcu(&pa->pa_inode_list);
4321 spin_unlock(pa->pa_obj_lock);
4322 ext4_mb_add_n_trim(ac);
4323 }
4324 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4325 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004326 if (ac->ac_bitmap_page)
4327 page_cache_release(ac->ac_bitmap_page);
4328 if (ac->ac_buddy_page)
4329 page_cache_release(ac->ac_buddy_page);
4330 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4331 mutex_unlock(&ac->ac_lg->lg_mutex);
4332 ext4_mb_collect_stats(ac);
4333 return 0;
4334}
4335
4336static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4337{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004338 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004339 int ret;
4340 int freed = 0;
4341
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004342 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004343 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004344 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4345 freed += ret;
4346 needed -= ret;
4347 }
4348
4349 return freed;
4350}
4351
4352/*
4353 * Main entry point into mballoc to allocate blocks
4354 * it tries to use preallocation first, then falls back
4355 * to usual allocation
4356 */
4357ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004358 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004359{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004360 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004361 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004362 struct ext4_sb_info *sbi;
4363 struct super_block *sb;
4364 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004365 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004366 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004367
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004368 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004369 sb = ar->inode->i_sb;
4370 sbi = EXT4_SB(sb);
4371
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004372 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004373
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004374 /* Allow to use superuser reservation for quota file */
4375 if (IS_NOQUOTA(ar->inode))
4376 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4377
Mingming Cao60e58e02009-01-22 18:13:05 +01004378 /*
4379 * For delayed allocation, we could skip the ENOSPC and
4380 * EDQUOT check, as blocks and quotas have been already
4381 * reserved when data being copied into pagecache.
4382 */
Theodore Ts'of2321092011-01-10 12:12:36 -05004383 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
Mingming Cao60e58e02009-01-22 18:13:05 +01004384 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4385 else {
4386 /* Without delayed allocation we need to verify
4387 * there is enough free blocks to do block allocation
4388 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004389 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004390 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004391 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004392
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004393 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004394 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004395 ar->len = ar->len >> 1;
4396 }
4397 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004398 *errp = -ENOSPC;
4399 return 0;
4400 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004401 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004402 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004403 dquot_alloc_block_nofail(ar->inode,
4404 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004405 } else {
4406 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004407 dquot_alloc_block(ar->inode,
4408 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004409
4410 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4411 ar->len--;
4412 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004413 }
4414 inquota = ar->len;
4415 if (ar->len == 0) {
4416 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004417 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004418 }
Mingming Caod2a17632008-07-14 17:52:37 -04004419 }
Mingming Caod2a17632008-07-14 17:52:37 -04004420
Wei Yongjun85556c92012-09-26 20:43:37 -04004421 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004422 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004423 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004424 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004425 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004426 }
4427
Eric Sandeen256bdb42008-02-10 01:13:33 -05004428 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004429 if (*errp) {
4430 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004431 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004432 }
4433
Eric Sandeen256bdb42008-02-10 01:13:33 -05004434 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4435 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004436 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4437 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004438repeat:
4439 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004440 *errp = ext4_mb_regular_allocator(ac);
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004441 if (*errp)
4442 goto discard_and_exit;
4443
4444 /* as we've just preallocated more space than
4445 * user requested originally, we store allocated
4446 * space in a special descriptor */
4447 if (ac->ac_status == AC_STATUS_FOUND &&
4448 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4449 *errp = ext4_mb_new_preallocation(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004450 if (*errp) {
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004451 discard_and_exit:
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004452 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004453 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004454 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004455 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004456 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004457 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004458 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004459 /*
4460 * drop the reference that we took
4461 * in ext4_mb_use_best_found
4462 */
4463 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004464 ac->ac_b_ex.fe_group = 0;
4465 ac->ac_b_ex.fe_start = 0;
4466 ac->ac_b_ex.fe_len = 0;
4467 ac->ac_status = AC_STATUS_CONTINUE;
4468 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004469 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004470 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004471 goto errout;
4472 } else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004473 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4474 ar->len = ac->ac_b_ex.fe_len;
4475 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004476 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004477 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004478 if (freed)
4479 goto repeat;
4480 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004481 }
4482
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004483errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004484 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004485 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004486 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004487 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004488 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004489 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004490out:
4491 if (ac)
4492 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004493 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004494 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004495 if (!ar->len) {
Theodore Ts'of2321092011-01-10 12:12:36 -05004496 if (!ext4_test_inode_state(ar->inode,
4497 EXT4_STATE_DELALLOC_RESERVED))
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004498 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004499 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004500 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004501 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004502
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004503 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004504
Alex Tomasc9de5602008-01-29 00:19:52 -05004505 return block;
4506}
Alex Tomasc9de5602008-01-29 00:19:52 -05004507
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004508/*
4509 * We can merge two free data extents only if the physical blocks
4510 * are contiguous, AND the extents were freed by the same transaction,
4511 * AND the blocks are associated with the same group.
4512 */
4513static int can_merge(struct ext4_free_data *entry1,
4514 struct ext4_free_data *entry2)
4515{
Bobi Jam18aadd42012-02-20 17:53:02 -05004516 if ((entry1->efd_tid == entry2->efd_tid) &&
4517 (entry1->efd_group == entry2->efd_group) &&
4518 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004519 return 1;
4520 return 0;
4521}
4522
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004523static noinline_for_stack int
4524ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004525 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004526{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004527 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004528 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004529 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004530 struct ext4_group_info *db = e4b->bd_info;
4531 struct super_block *sb = e4b->bd_sb;
4532 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004533 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4534 struct rb_node *parent = NULL, *new_node;
4535
Frank Mayhar03901312009-01-07 00:06:22 -05004536 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004537 BUG_ON(e4b->bd_bitmap_page == NULL);
4538 BUG_ON(e4b->bd_buddy_page == NULL);
4539
Bobi Jam18aadd42012-02-20 17:53:02 -05004540 new_node = &new_entry->efd_node;
4541 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004542
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004543 if (!*n) {
4544 /* first free block exent. We need to
4545 protect buddy cache from being freed,
4546 * otherwise we'll refresh it from
4547 * on-disk bitmap and lose not-yet-available
4548 * blocks */
4549 page_cache_get(e4b->bd_buddy_page);
4550 page_cache_get(e4b->bd_bitmap_page);
4551 }
4552 while (*n) {
4553 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004554 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4555 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004556 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004557 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004558 n = &(*n)->rb_right;
4559 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004560 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004561 ext4_group_first_block_no(sb, group) +
4562 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004563 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004564 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004565 }
4566 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004567
4568 rb_link_node(new_node, parent, n);
4569 rb_insert_color(new_node, &db->bb_free_root);
4570
4571 /* Now try to see the extent can be merged to left and right */
4572 node = rb_prev(new_node);
4573 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004574 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004575 if (can_merge(entry, new_entry) &&
4576 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004577 new_entry->efd_start_cluster = entry->efd_start_cluster;
4578 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004579 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004580 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004581 }
4582 }
4583
4584 node = rb_next(new_node);
4585 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004586 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004587 if (can_merge(new_entry, entry) &&
4588 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004589 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004590 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004591 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004592 }
4593 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004594 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004595 ext4_journal_callback_add(handle, ext4_free_data_callback,
4596 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004597 return 0;
4598}
4599
Theodore Ts'o44338712009-11-22 07:44:56 -05004600/**
4601 * ext4_free_blocks() -- Free given blocks and update quota
4602 * @handle: handle for this transaction
4603 * @inode: inode
4604 * @block: start physical block to free
4605 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004606 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004607 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004608void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004609 struct buffer_head *bh, ext4_fsblk_t block,
4610 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004611{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004612 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004613 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004614 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004615 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004616 ext4_grpblk_t bit;
4617 struct buffer_head *gd_bh;
4618 ext4_group_t block_group;
4619 struct ext4_sb_info *sbi;
Jan Kara7d734532013-08-17 09:36:54 -04004620 struct ext4_inode_info *ei = EXT4_I(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05004621 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004622 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004623 int err = 0;
4624 int ret;
4625
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004626 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004627 if (bh) {
4628 if (block)
4629 BUG_ON(block != bh->b_blocknr);
4630 else
4631 block = bh->b_blocknr;
4632 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004633
Alex Tomasc9de5602008-01-29 00:19:52 -05004634 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004635 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4636 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004637 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004638 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004639 goto error_return;
4640 }
4641
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004642 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004643 trace_ext4_free_blocks(inode, block, count, flags);
4644
4645 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4646 struct buffer_head *tbh = bh;
4647 int i;
4648
4649 BUG_ON(bh && (count > 1));
4650
4651 for (i = 0; i < count; i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004652 cond_resched();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004653 if (!bh)
4654 tbh = sb_find_get_block(inode->i_sb,
4655 block + i);
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004656 if (!tbh)
Namhyung Kim87783692010-10-27 21:30:11 -04004657 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004658 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004659 inode, tbh, block + i);
4660 }
4661 }
4662
Theodore Ts'o60e66792010-05-17 07:00:00 -04004663 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004664 * We need to make sure we don't reuse the freed block until
4665 * after the transaction is committed, which we can do by
4666 * treating the block as metadata, below. We make an
4667 * exception if the inode is to be written in writeback mode
4668 * since writeback mode has weak data consistency guarantees.
4669 */
4670 if (!ext4_should_writeback_data(inode))
4671 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004672
Theodore Ts'o84130192011-09-09 18:50:51 -04004673 /*
4674 * If the extent to be freed does not begin on a cluster
4675 * boundary, we need to deal with partial clusters at the
4676 * beginning and end of the extent. Normally we will free
4677 * blocks at the beginning or the end unless we are explicitly
4678 * requested to avoid doing so.
4679 */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004680 overflow = EXT4_PBLK_COFF(sbi, block);
Theodore Ts'o84130192011-09-09 18:50:51 -04004681 if (overflow) {
4682 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4683 overflow = sbi->s_cluster_ratio - overflow;
4684 block += overflow;
4685 if (count > overflow)
4686 count -= overflow;
4687 else
4688 return;
4689 } else {
4690 block -= overflow;
4691 count += overflow;
4692 }
4693 }
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004694 overflow = EXT4_LBLK_COFF(sbi, count);
Theodore Ts'o84130192011-09-09 18:50:51 -04004695 if (overflow) {
4696 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4697 if (count > overflow)
4698 count -= overflow;
4699 else
4700 return;
4701 } else
4702 count += sbi->s_cluster_ratio - overflow;
4703 }
4704
Alex Tomasc9de5602008-01-29 00:19:52 -05004705do_more:
4706 overflow = 0;
4707 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4708
Darrick J. Wong163a2032013-08-28 17:35:51 -04004709 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4710 ext4_get_group_info(sb, block_group))))
4711 return;
4712
Alex Tomasc9de5602008-01-29 00:19:52 -05004713 /*
4714 * Check to see if we are freeing blocks across a group
4715 * boundary.
4716 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004717 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4718 overflow = EXT4_C2B(sbi, bit) + count -
4719 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004720 count -= overflow;
4721 }
Lukas Czerner810da242013-03-02 17:18:58 -05004722 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004723 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004724 if (!bitmap_bh) {
4725 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004726 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004727 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004728 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004729 if (!gdp) {
4730 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004731 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004732 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004733
4734 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4735 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4736 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004737 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004738 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004739 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004740
Eric Sandeen12062dd2010-02-15 14:19:27 -05004741 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004742 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004743 /* err = 0. ext4_std_error should be a no op */
4744 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004745 }
4746
4747 BUFFER_TRACE(bitmap_bh, "getting write access");
4748 err = ext4_journal_get_write_access(handle, bitmap_bh);
4749 if (err)
4750 goto error_return;
4751
4752 /*
4753 * We are about to modify some metadata. Call the journal APIs
4754 * to unshare ->b_data if a currently-committing transaction is
4755 * using it
4756 */
4757 BUFFER_TRACE(gd_bh, "get_write_access");
4758 err = ext4_journal_get_write_access(handle, gd_bh);
4759 if (err)
4760 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004761#ifdef AGGRESSIVE_CHECK
4762 {
4763 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004764 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004765 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4766 }
4767#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004768 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004769
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004770 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4771 if (err)
4772 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004773
4774 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004775 struct ext4_free_data *new_entry;
4776 /*
4777 * blocks being freed are metadata. these blocks shouldn't
4778 * be used until this transaction is committed
4779 */
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004780 retry:
Bobi Jam18aadd42012-02-20 17:53:02 -05004781 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004782 if (!new_entry) {
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004783 /*
4784 * We use a retry loop because
4785 * ext4_free_blocks() is not allowed to fail.
4786 */
4787 cond_resched();
4788 congestion_wait(BLK_RW_ASYNC, HZ/50);
4789 goto retry;
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004790 }
Bobi Jam18aadd42012-02-20 17:53:02 -05004791 new_entry->efd_start_cluster = bit;
4792 new_entry->efd_group = block_group;
4793 new_entry->efd_count = count_clusters;
4794 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004795
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004796 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004797 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004798 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004799 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004800 /* need to update group_info->bb_free and bitmap
4801 * with group lock held. generate_buddy look at
4802 * them with group lock_held
4803 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004804 if (test_opt(sb, DISCARD)) {
4805 err = ext4_issue_discard(sb, block_group, bit, count);
4806 if (err && err != -EOPNOTSUPP)
4807 ext4_msg(sb, KERN_WARNING, "discard request in"
4808 " group:%d block:%d count:%lu failed"
4809 " with %d", block_group, bit, count,
4810 err);
Lukas Czerner8f9ff182013-10-30 11:10:52 -04004811 } else
4812 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004813
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004814 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004815 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4816 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004817 }
4818
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004819 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4820 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004821 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004822 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004823 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004824
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004825 if (sbi->s_log_groups_per_flex) {
4826 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004827 atomic64_add(count_clusters,
4828 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004829 }
4830
Jan Kara7d734532013-08-17 09:36:54 -04004831 if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) {
4832 percpu_counter_add(&sbi->s_dirtyclusters_counter,
4833 count_clusters);
4834 spin_lock(&ei->i_block_reservation_lock);
4835 if (flags & EXT4_FREE_BLOCKS_METADATA)
4836 ei->i_reserved_meta_blocks += count_clusters;
4837 else
4838 ei->i_reserved_data_blocks += count_clusters;
4839 spin_unlock(&ei->i_block_reservation_lock);
4840 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4841 dquot_reclaim_block(inode,
4842 EXT4_C2B(sbi, count_clusters));
4843 } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Aditya Kali7b415bf2011-09-09 19:04:51 -04004844 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
Jan Kara7d734532013-08-17 09:36:54 -04004845 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4846
4847 ext4_mb_unload_buddy(&e4b);
Aditya Kali7b415bf2011-09-09 19:04:51 -04004848
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004849 /* We dirtied the bitmap block */
4850 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4851 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4852
Alex Tomasc9de5602008-01-29 00:19:52 -05004853 /* And the group descriptor block */
4854 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004855 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004856 if (!err)
4857 err = ret;
4858
4859 if (overflow && !err) {
4860 block += count;
4861 count = overflow;
4862 put_bh(bitmap_bh);
4863 goto do_more;
4864 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004865error_return:
4866 brelse(bitmap_bh);
4867 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004868 return;
4869}
Lukas Czerner7360d172010-10-27 21:30:12 -04004870
4871/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004872 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004873 * @handle: handle to this transaction
4874 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004875 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004876 * @count: number of blocks to free
4877 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004878 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004879 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004880int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004881 ext4_fsblk_t block, unsigned long count)
4882{
4883 struct buffer_head *bitmap_bh = NULL;
4884 struct buffer_head *gd_bh;
4885 ext4_group_t block_group;
4886 ext4_grpblk_t bit;
4887 unsigned int i;
4888 struct ext4_group_desc *desc;
4889 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004890 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004891 int err = 0, ret, blk_free_count;
4892 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004893
4894 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4895
Yongqiang Yang4740b832011-07-26 21:51:08 -04004896 if (count == 0)
4897 return 0;
4898
Amir Goldstein2846e822011-05-09 10:46:41 -04004899 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004900 /*
4901 * Check to see if we are freeing blocks across a group
4902 * boundary.
4903 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004904 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4905 ext4_warning(sb, "too much blocks added to group %u\n",
4906 block_group);
4907 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004908 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004909 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004910
Amir Goldstein2846e822011-05-09 10:46:41 -04004911 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004912 if (!bitmap_bh) {
4913 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004914 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004915 }
4916
Amir Goldstein2846e822011-05-09 10:46:41 -04004917 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004918 if (!desc) {
4919 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004920 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004921 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004922
4923 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4924 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4925 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4926 in_range(block + count - 1, ext4_inode_table(sb, desc),
4927 sbi->s_itb_per_group)) {
4928 ext4_error(sb, "Adding blocks in system zones - "
4929 "Block = %llu, count = %lu",
4930 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004931 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004932 goto error_return;
4933 }
4934
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004935 BUFFER_TRACE(bitmap_bh, "getting write access");
4936 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004937 if (err)
4938 goto error_return;
4939
4940 /*
4941 * We are about to modify some metadata. Call the journal APIs
4942 * to unshare ->b_data if a currently-committing transaction is
4943 * using it
4944 */
4945 BUFFER_TRACE(gd_bh, "get_write_access");
4946 err = ext4_journal_get_write_access(handle, gd_bh);
4947 if (err)
4948 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004949
Amir Goldstein2846e822011-05-09 10:46:41 -04004950 for (i = 0, blocks_freed = 0; i < count; i++) {
4951 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004952 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004953 ext4_error(sb, "bit already cleared for block %llu",
4954 (ext4_fsblk_t)(block + i));
4955 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4956 } else {
4957 blocks_freed++;
4958 }
4959 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004960
4961 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4962 if (err)
4963 goto error_return;
4964
4965 /*
4966 * need to update group_info->bb_free and bitmap
4967 * with group lock held. generate_buddy look at
4968 * them with group lock_held
4969 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004970 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004971 mb_clear_bits(bitmap_bh->b_data, bit, count);
4972 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004973 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4974 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04004975 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004976 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04004977 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04004978 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05004979 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04004980
4981 if (sbi->s_log_groups_per_flex) {
4982 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004983 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4984 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04004985 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004986
4987 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04004988
4989 /* We dirtied the bitmap block */
4990 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4991 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4992
4993 /* And the group descriptor block */
4994 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4995 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4996 if (!err)
4997 err = ret;
4998
4999error_return:
5000 brelse(bitmap_bh);
5001 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04005002 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04005003}
5004
5005/**
Lukas Czerner7360d172010-10-27 21:30:12 -04005006 * ext4_trim_extent -- function to TRIM one single free extent in the group
5007 * @sb: super block for the file system
5008 * @start: starting block of the free extent in the alloc. group
5009 * @count: number of blocks to TRIM
5010 * @group: alloc. group we are working with
5011 * @e4b: ext4 buddy for the group
5012 *
5013 * Trim "count" blocks starting at "start" in the "group". To assure that no
5014 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5015 * be called with under the group lock.
5016 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005017static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04005018 ext4_group_t group, struct ext4_buddy *e4b)
Lukas Czerner7360d172010-10-27 21:30:12 -04005019{
5020 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005021 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005022
Tao Mab3d4c2b2011-07-11 00:01:52 -04005023 trace_ext4_trim_extent(sb, group, start, count);
5024
Lukas Czerner7360d172010-10-27 21:30:12 -04005025 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5026
5027 ex.fe_start = start;
5028 ex.fe_group = group;
5029 ex.fe_len = count;
5030
5031 /*
5032 * Mark blocks used, so no one can reuse them while
5033 * being trimmed.
5034 */
5035 mb_mark_used(e4b, &ex);
5036 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005037 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04005038 ext4_lock_group(sb, group);
5039 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005040 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005041}
5042
5043/**
5044 * ext4_trim_all_free -- function to trim all free space in alloc. group
5045 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04005046 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04005047 * @start: first group block to examine
5048 * @max: last group block to examine
5049 * @minblocks: minimum extent block count
5050 *
5051 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5052 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5053 * the extent.
5054 *
5055 *
5056 * ext4_trim_all_free walks through group's block bitmap searching for free
5057 * extents. When the free extent is found, mark it as used in group buddy
5058 * bitmap. Then issue a TRIM command on this extent and free the extent in
5059 * the group buddy bitmap. This is done until whole group is scanned.
5060 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005061static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005062ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5063 ext4_grpblk_t start, ext4_grpblk_t max,
5064 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005065{
5066 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005067 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005068 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005069 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005070
Tao Mab3d4c2b2011-07-11 00:01:52 -04005071 trace_ext4_trim_all_free(sb, group, start, max);
5072
Lukas Czerner78944082011-05-24 18:16:27 -04005073 ret = ext4_mb_load_buddy(sb, group, &e4b);
5074 if (ret) {
5075 ext4_error(sb, "Error in loading buddy "
5076 "information for %u", group);
5077 return ret;
5078 }
Lukas Czerner78944082011-05-24 18:16:27 -04005079 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005080
5081 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005082 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5083 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5084 goto out;
5085
Lukas Czerner78944082011-05-24 18:16:27 -04005086 start = (e4b.bd_info->bb_first_free > start) ?
5087 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005088
Lukas Czerner913eed832012-03-21 21:22:22 -04005089 while (start <= max) {
5090 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5091 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005092 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005093 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005094
5095 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005096 ret = ext4_trim_extent(sb, start,
5097 next - start, group, &e4b);
5098 if (ret && ret != -EOPNOTSUPP)
5099 break;
5100 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005101 count += next - start;
5102 }
Tao Ma169ddc32011-07-11 00:00:07 -04005103 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005104 start = next + 1;
5105
5106 if (fatal_signal_pending(current)) {
5107 count = -ERESTARTSYS;
5108 break;
5109 }
5110
5111 if (need_resched()) {
5112 ext4_unlock_group(sb, group);
5113 cond_resched();
5114 ext4_lock_group(sb, group);
5115 }
5116
Tao Ma169ddc32011-07-11 00:00:07 -04005117 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005118 break;
5119 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005120
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005121 if (!ret) {
5122 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005123 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005124 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005125out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005126 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005127 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005128
5129 ext4_debug("trimmed %d blocks in the group %d\n",
5130 count, group);
5131
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005132 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005133}
5134
5135/**
5136 * ext4_trim_fs() -- trim ioctl handle function
5137 * @sb: superblock for filesystem
5138 * @range: fstrim_range structure
5139 *
5140 * start: First Byte to trim
5141 * len: number of Bytes to trim from start
5142 * minlen: minimum extent length in Bytes
5143 * ext4_trim_fs goes through all allocation groups containing Bytes from
5144 * start to start+len. For each such a group ext4_trim_all_free function
5145 * is invoked to trim all free space.
5146 */
5147int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5148{
Lukas Czerner78944082011-05-24 18:16:27 -04005149 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005150 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005151 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005152 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005153 ext4_fsblk_t first_data_blk =
5154 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005155 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005156 int ret = 0;
5157
5158 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005159 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005160 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5161 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005162
Lukas Czerner5de35e82012-10-22 18:01:19 -04005163 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5164 start >= max_blks ||
5165 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005166 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005167 if (end >= max_blks)
5168 end = max_blks - 1;
5169 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005170 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005171 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005172 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005173
Lukas Czerner913eed832012-03-21 21:22:22 -04005174 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005175 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005176 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005177 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005178 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005179
Lukas Czerner913eed832012-03-21 21:22:22 -04005180 /* end now represents the last cluster to discard in this group */
5181 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005182
5183 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005184 grp = ext4_get_group_info(sb, group);
5185 /* We only do this if the grp has never been initialized */
5186 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5187 ret = ext4_mb_init_group(sb, group);
5188 if (ret)
5189 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005190 }
5191
Tao Ma0ba08512011-03-23 15:48:11 -04005192 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005193 * For all the groups except the last one, last cluster will
5194 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5195 * change it for the last group, note that last_cluster is
5196 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005197 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005198 if (group == last_group)
5199 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005200
Lukas Czerner78944082011-05-24 18:16:27 -04005201 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005202 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005203 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005204 if (cnt < 0) {
5205 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005206 break;
5207 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005208 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005209 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005210
5211 /*
5212 * For every group except the first one, we are sure
5213 * that the first cluster to discard will be cluster #0.
5214 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005215 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005216 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005217
Tao Ma3d56b8d2011-07-11 00:03:38 -04005218 if (!ret)
5219 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5220
Tao Ma22f10452011-07-10 23:52:37 -04005221out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005222 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005223 return ret;
5224}