blob: def84082a9a9b73deadc0d369c63a7ef981d8ea9 [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,
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400754 "%u clusters in bitmap, %u in gd",
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400755 free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500756 /*
757 * If we intent to continue, we consider group descritor
758 * corrupt and update bb_free using bitmap value
759 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500760 grp->bb_free = free;
761 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400762 mb_set_largest_free_order(sb, grp);
Alex Tomasc9de5602008-01-29 00:19:52 -0500763
764 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
765
766 period = get_cycles() - period;
767 spin_lock(&EXT4_SB(sb)->s_bal_lock);
768 EXT4_SB(sb)->s_mb_buddies_generated++;
769 EXT4_SB(sb)->s_mb_generation_time += period;
770 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
771}
772
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400773static void mb_regenerate_buddy(struct ext4_buddy *e4b)
774{
775 int count;
776 int order = 1;
777 void *buddy;
778
779 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
780 ext4_set_bits(buddy, 0, count);
781 }
782 e4b->bd_info->bb_fragments = 0;
783 memset(e4b->bd_info->bb_counters, 0,
784 sizeof(*e4b->bd_info->bb_counters) *
785 (e4b->bd_sb->s_blocksize_bits + 2));
786
787 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
788 e4b->bd_bitmap, e4b->bd_group);
789}
790
Alex Tomasc9de5602008-01-29 00:19:52 -0500791/* The buddy information is attached the buddy cache inode
792 * for convenience. The information regarding each group
793 * is loaded via ext4_mb_load_buddy. The information involve
794 * block bitmap and buddy information. The information are
795 * stored in the inode as
796 *
797 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500798 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500799 *
800 *
801 * one block each for bitmap and buddy information.
802 * So for each group we take up 2 blocks. A page can
803 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
804 * So it can have information regarding groups_per_page which
805 * is blocks_per_page/2
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400806 *
807 * Locking note: This routine takes the block group lock of all groups
808 * for this page; do not hold this lock when calling this routine!
Alex Tomasc9de5602008-01-29 00:19:52 -0500809 */
810
811static int ext4_mb_init_cache(struct page *page, char *incore)
812{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400813 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500814 int blocksize;
815 int blocks_per_page;
816 int groups_per_page;
817 int err = 0;
818 int i;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500819 ext4_group_t first_group, group;
Alex Tomasc9de5602008-01-29 00:19:52 -0500820 int first_block;
821 struct super_block *sb;
822 struct buffer_head *bhs;
Darrick J. Wongfa77dcf2012-04-29 18:35:10 -0400823 struct buffer_head **bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -0500824 struct inode *inode;
825 char *data;
826 char *bitmap;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400827 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -0500828
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400829 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500830
831 inode = page->mapping->host;
832 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400833 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500834 blocksize = 1 << inode->i_blkbits;
835 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
836
837 groups_per_page = blocks_per_page >> 1;
838 if (groups_per_page == 0)
839 groups_per_page = 1;
840
841 /* allocate buffer_heads to read bitmaps */
842 if (groups_per_page > 1) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500843 i = sizeof(struct buffer_head *) * groups_per_page;
844 bh = kzalloc(i, GFP_NOFS);
Theodore Ts'o813e5722012-02-20 17:52:46 -0500845 if (bh == NULL) {
846 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500847 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500848 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500849 } else
850 bh = &bhs;
851
852 first_group = page->index * blocks_per_page / 2;
853
854 /* read all groups the page covers into the cache */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500855 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
856 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500857 break;
858
Theodore Ts'o813e5722012-02-20 17:52:46 -0500859 grinfo = ext4_get_group_info(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400860 /*
861 * If page is uptodate then we came here after online resize
862 * which added some new uninitialized group info structs, so
863 * we must skip all initialized uptodate buddies on the page,
864 * which may be currently in use by an allocating task.
865 */
866 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
867 bh[i] = NULL;
868 continue;
869 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500870 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
871 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500872 goto out;
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500873 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500874 mb_debug(1, "read bitmap for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500875 }
876
877 /* wait for I/O completion */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500878 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
879 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
880 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -0500881 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500882 }
883 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500884
885 first_block = page->index * blocks_per_page;
886 for (i = 0; i < blocks_per_page; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500887 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400888 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500889 break;
890
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400891 if (!bh[group - first_group])
892 /* skip initialized uptodate buddy */
893 continue;
894
Alex Tomasc9de5602008-01-29 00:19:52 -0500895 /*
896 * data carry information regarding this
897 * particular group in the format specified
898 * above
899 *
900 */
901 data = page_address(page) + (i * blocksize);
902 bitmap = bh[group - first_group]->b_data;
903
904 /*
905 * We place the buddy block and bitmap block
906 * close together
907 */
908 if ((first_block + i) & 1) {
909 /* this is block of buddy */
910 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400911 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500912 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400913 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500914 grinfo = ext4_get_group_info(sb, group);
915 grinfo->bb_fragments = 0;
916 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400917 sizeof(*grinfo->bb_counters) *
918 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500919 /*
920 * incore got set to the group block bitmap below
921 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500922 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400923 /* init the buddy */
924 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500925 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500926 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500927 incore = NULL;
928 } else {
929 /* this is block of bitmap */
930 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400931 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500932 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400933 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500934
935 /* see comments in ext4_mb_put_pa() */
936 ext4_lock_group(sb, group);
937 memcpy(data, bitmap, blocksize);
938
939 /* mark all preallocated blks used in in-core bitmap */
940 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500941 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500942 ext4_unlock_group(sb, group);
943
944 /* set incore so that the buddy information can be
945 * generated using this
946 */
947 incore = data;
948 }
949 }
950 SetPageUptodate(page);
951
952out:
953 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400954 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500955 brelse(bh[i]);
956 if (bh != &bhs)
957 kfree(bh);
958 }
959 return err;
960}
961
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400962/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400963 * Lock the buddy and bitmap pages. This make sure other parallel init_group
964 * on the same buddy page doesn't happen whild holding the buddy page lock.
965 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
966 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400967 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400968static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
969 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400970{
Amir Goldstein2de88072011-05-09 21:48:13 -0400971 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
972 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400973 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400974 struct page *page;
975
976 e4b->bd_buddy_page = NULL;
977 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400978
979 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
980 /*
981 * the buddy cache inode stores the block bitmap
982 * and buddy information in consecutive blocks.
983 * So for each group we need two blocks.
984 */
985 block = group * 2;
986 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400987 poff = block % blocks_per_page;
988 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
989 if (!page)
990 return -EIO;
991 BUG_ON(page->mapping != inode->i_mapping);
992 e4b->bd_bitmap_page = page;
993 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400994
Amir Goldstein2de88072011-05-09 21:48:13 -0400995 if (blocks_per_page >= 2) {
996 /* buddy and bitmap are on the same page */
997 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400998 }
Amir Goldstein2de88072011-05-09 21:48:13 -0400999
1000 block++;
1001 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001002 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1003 if (!page)
1004 return -EIO;
1005 BUG_ON(page->mapping != inode->i_mapping);
1006 e4b->bd_buddy_page = page;
1007 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001008}
1009
Amir Goldstein2de88072011-05-09 21:48:13 -04001010static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001011{
Amir Goldstein2de88072011-05-09 21:48:13 -04001012 if (e4b->bd_bitmap_page) {
1013 unlock_page(e4b->bd_bitmap_page);
1014 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001015 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001016 if (e4b->bd_buddy_page) {
1017 unlock_page(e4b->bd_buddy_page);
1018 page_cache_release(e4b->bd_buddy_page);
1019 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001020}
1021
1022/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001023 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1024 * block group lock of all groups for this page; do not hold the BG lock when
1025 * calling this routine!
1026 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001027static noinline_for_stack
1028int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1029{
1030
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001031 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001032 struct ext4_buddy e4b;
1033 struct page *page;
1034 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001035
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001036 might_sleep();
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001037 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001038 this_grp = ext4_get_group_info(sb, group);
1039 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001040 * This ensures that we don't reinit the buddy cache
1041 * page which map to the group from which we are already
1042 * allocating. If we are looking at the buddy cache we would
1043 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001044 * would have pinned buddy page to page cache.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001045 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001046 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1047 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001048 /*
1049 * somebody initialized the group
1050 * return without doing anything
1051 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001052 goto err;
1053 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001054
1055 page = e4b.bd_bitmap_page;
1056 ret = ext4_mb_init_cache(page, NULL);
1057 if (ret)
1058 goto err;
1059 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001060 ret = -EIO;
1061 goto err;
1062 }
1063 mark_page_accessed(page);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001064
Amir Goldstein2de88072011-05-09 21:48:13 -04001065 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001066 /*
1067 * If both the bitmap and buddy are in
1068 * the same page we don't need to force
1069 * init the buddy
1070 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001071 ret = 0;
1072 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001073 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001074 /* init buddy cache */
1075 page = e4b.bd_buddy_page;
1076 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1077 if (ret)
1078 goto err;
1079 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001080 ret = -EIO;
1081 goto err;
1082 }
1083 mark_page_accessed(page);
1084err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001085 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001086 return ret;
1087}
1088
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001089/*
1090 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1091 * block group lock of all groups for this page; do not hold the BG lock when
1092 * calling this routine!
1093 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001094static noinline_for_stack int
1095ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1096 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001097{
Alex Tomasc9de5602008-01-29 00:19:52 -05001098 int blocks_per_page;
1099 int block;
1100 int pnum;
1101 int poff;
1102 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001103 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001104 struct ext4_group_info *grp;
1105 struct ext4_sb_info *sbi = EXT4_SB(sb);
1106 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001107
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001108 might_sleep();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001109 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001110
1111 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001112 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001113
1114 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001115 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001116 e4b->bd_sb = sb;
1117 e4b->bd_group = group;
1118 e4b->bd_buddy_page = NULL;
1119 e4b->bd_bitmap_page = NULL;
1120
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001121 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001122 /*
1123 * we need full data about the group
1124 * to make a good selection
1125 */
1126 ret = ext4_mb_init_group(sb, group);
1127 if (ret)
1128 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001129 }
1130
Alex Tomasc9de5602008-01-29 00:19:52 -05001131 /*
1132 * the buddy cache inode stores the block bitmap
1133 * and buddy information in consecutive blocks.
1134 * So for each group we need two blocks.
1135 */
1136 block = group * 2;
1137 pnum = block / blocks_per_page;
1138 poff = block % blocks_per_page;
1139
1140 /* we could use find_or_create_page(), but it locks page
1141 * what we'd like to avoid in fast path ... */
1142 page = find_get_page(inode->i_mapping, pnum);
1143 if (page == NULL || !PageUptodate(page)) {
1144 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001145 /*
1146 * drop the page reference and try
1147 * to get the page with lock. If we
1148 * are not uptodate that implies
1149 * somebody just created the page but
1150 * is yet to initialize the same. So
1151 * wait for it to initialize.
1152 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001153 page_cache_release(page);
1154 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1155 if (page) {
1156 BUG_ON(page->mapping != inode->i_mapping);
1157 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001158 ret = ext4_mb_init_cache(page, NULL);
1159 if (ret) {
1160 unlock_page(page);
1161 goto err;
1162 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001163 mb_cmp_bitmaps(e4b, page_address(page) +
1164 (poff * sb->s_blocksize));
1165 }
1166 unlock_page(page);
1167 }
1168 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001169 if (page == NULL || !PageUptodate(page)) {
1170 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001171 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001172 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001173 e4b->bd_bitmap_page = page;
1174 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1175 mark_page_accessed(page);
1176
1177 block++;
1178 pnum = block / blocks_per_page;
1179 poff = block % blocks_per_page;
1180
1181 page = find_get_page(inode->i_mapping, pnum);
1182 if (page == NULL || !PageUptodate(page)) {
1183 if (page)
1184 page_cache_release(page);
1185 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1186 if (page) {
1187 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001188 if (!PageUptodate(page)) {
1189 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1190 if (ret) {
1191 unlock_page(page);
1192 goto err;
1193 }
1194 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001195 unlock_page(page);
1196 }
1197 }
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001198 if (page == NULL || !PageUptodate(page)) {
1199 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001200 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001201 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001202 e4b->bd_buddy_page = page;
1203 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1204 mark_page_accessed(page);
1205
1206 BUG_ON(e4b->bd_bitmap_page == NULL);
1207 BUG_ON(e4b->bd_buddy_page == NULL);
1208
1209 return 0;
1210
1211err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001212 if (page)
1213 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001214 if (e4b->bd_bitmap_page)
1215 page_cache_release(e4b->bd_bitmap_page);
1216 if (e4b->bd_buddy_page)
1217 page_cache_release(e4b->bd_buddy_page);
1218 e4b->bd_buddy = NULL;
1219 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001220 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001221}
1222
Jing Zhange39e07f2010-05-14 00:00:00 -04001223static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001224{
1225 if (e4b->bd_bitmap_page)
1226 page_cache_release(e4b->bd_bitmap_page);
1227 if (e4b->bd_buddy_page)
1228 page_cache_release(e4b->bd_buddy_page);
1229}
1230
1231
1232static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1233{
1234 int order = 1;
1235 void *bb;
1236
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001237 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001238 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1239
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001240 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001241 while (order <= e4b->bd_blkbits + 1) {
1242 block = block >> 1;
1243 if (!mb_test_bit(block, bb)) {
1244 /* this block is part of buddy of order 'order' */
1245 return order;
1246 }
1247 bb += 1 << (e4b->bd_blkbits - order);
1248 order++;
1249 }
1250 return 0;
1251}
1252
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001253static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001254{
1255 __u32 *addr;
1256
1257 len = cur + len;
1258 while (cur < len) {
1259 if ((cur & 31) == 0 && (len - cur) >= 32) {
1260 /* fast path: clear whole word at once */
1261 addr = bm + (cur >> 3);
1262 *addr = 0;
1263 cur += 32;
1264 continue;
1265 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001266 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001267 cur++;
1268 }
1269}
1270
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001271/* clear bits in given range
1272 * will return first found zero bit if any, -1 otherwise
1273 */
1274static int mb_test_and_clear_bits(void *bm, int cur, int len)
1275{
1276 __u32 *addr;
1277 int zero_bit = -1;
1278
1279 len = cur + len;
1280 while (cur < len) {
1281 if ((cur & 31) == 0 && (len - cur) >= 32) {
1282 /* fast path: clear whole word at once */
1283 addr = bm + (cur >> 3);
1284 if (*addr != (__u32)(-1) && zero_bit == -1)
1285 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1286 *addr = 0;
1287 cur += 32;
1288 continue;
1289 }
1290 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1291 zero_bit = cur;
1292 cur++;
1293 }
1294
1295 return zero_bit;
1296}
1297
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001298void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001299{
1300 __u32 *addr;
1301
1302 len = cur + len;
1303 while (cur < len) {
1304 if ((cur & 31) == 0 && (len - cur) >= 32) {
1305 /* fast path: set whole word at once */
1306 addr = bm + (cur >> 3);
1307 *addr = 0xffffffff;
1308 cur += 32;
1309 continue;
1310 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001311 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001312 cur++;
1313 }
1314}
1315
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001316/*
1317 * _________________________________________________________________ */
1318
1319static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001320{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001321 if (mb_test_bit(*bit + side, bitmap)) {
1322 mb_clear_bit(*bit, bitmap);
1323 (*bit) -= side;
1324 return 1;
1325 }
1326 else {
1327 (*bit) += side;
1328 mb_set_bit(*bit, bitmap);
1329 return -1;
1330 }
1331}
1332
1333static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1334{
1335 int max;
1336 int order = 1;
1337 void *buddy = mb_find_buddy(e4b, order, &max);
1338
1339 while (buddy) {
1340 void *buddy2;
1341
1342 /* Bits in range [first; last] are known to be set since
1343 * corresponding blocks were allocated. Bits in range
1344 * (first; last) will stay set because they form buddies on
1345 * upper layer. We just deal with borders if they don't
1346 * align with upper layer and then go up.
1347 * Releasing entire group is all about clearing
1348 * single bit of highest order buddy.
1349 */
1350
1351 /* Example:
1352 * ---------------------------------
1353 * | 1 | 1 | 1 | 1 |
1354 * ---------------------------------
1355 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1356 * ---------------------------------
1357 * 0 1 2 3 4 5 6 7
1358 * \_____________________/
1359 *
1360 * Neither [1] nor [6] is aligned to above layer.
1361 * Left neighbour [0] is free, so mark it busy,
1362 * decrease bb_counters and extend range to
1363 * [0; 6]
1364 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1365 * mark [6] free, increase bb_counters and shrink range to
1366 * [0; 5].
1367 * Then shift range to [0; 2], go up and do the same.
1368 */
1369
1370
1371 if (first & 1)
1372 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1373 if (!(last & 1))
1374 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1375 if (first > last)
1376 break;
1377 order++;
1378
1379 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1380 mb_clear_bits(buddy, first, last - first + 1);
1381 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1382 break;
1383 }
1384 first >>= 1;
1385 last >>= 1;
1386 buddy = buddy2;
1387 }
1388}
1389
1390static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1391 int first, int count)
1392{
1393 int left_is_free = 0;
1394 int right_is_free = 0;
1395 int block;
1396 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001397 struct super_block *sb = e4b->bd_sb;
1398
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001399 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001400 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001401 mb_check_buddy(e4b);
1402 mb_free_blocks_double(inode, e4b, first, count);
1403
1404 e4b->bd_info->bb_free += count;
1405 if (first < e4b->bd_info->bb_first_free)
1406 e4b->bd_info->bb_first_free = first;
1407
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001408 /* access memory sequentially: check left neighbour,
1409 * clear range and then check right neighbour
1410 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001411 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001412 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1413 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1414 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1415 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1416
1417 if (unlikely(block != -1)) {
1418 ext4_fsblk_t blocknr;
1419
1420 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1421 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1422 ext4_grp_locked_error(sb, e4b->bd_group,
1423 inode ? inode->i_ino : 0,
1424 blocknr,
1425 "freeing already freed block "
1426 "(bit %u)", block);
1427 mb_regenerate_buddy(e4b);
1428 goto done;
1429 }
1430
1431 /* let's maintain fragments counter */
1432 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001433 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001434 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001435 e4b->bd_info->bb_fragments++;
1436
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001437 /* buddy[0] == bd_bitmap is a special case, so handle
1438 * it right away and let mb_buddy_mark_free stay free of
1439 * zero order checks.
1440 * Check if neighbours are to be coaleasced,
1441 * adjust bitmap bb_counters and borders appropriately.
1442 */
1443 if (first & 1) {
1444 first += !left_is_free;
1445 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001446 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001447 if (!(last & 1)) {
1448 last -= !right_is_free;
1449 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1450 }
1451
1452 if (first <= last)
1453 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1454
1455done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001456 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001457 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001458}
1459
Robin Dong15c006a2012-08-17 10:02:17 -04001460static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001461 int needed, struct ext4_free_extent *ex)
1462{
1463 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001464 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001465 void *buddy;
1466
Vincent Minetbc8e6742009-05-15 08:33:18 -04001467 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001468 BUG_ON(ex == NULL);
1469
Robin Dong15c006a2012-08-17 10:02:17 -04001470 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001471 BUG_ON(buddy == NULL);
1472 BUG_ON(block >= max);
1473 if (mb_test_bit(block, buddy)) {
1474 ex->fe_len = 0;
1475 ex->fe_start = 0;
1476 ex->fe_group = 0;
1477 return 0;
1478 }
1479
Robin Dong15c006a2012-08-17 10:02:17 -04001480 /* find actual order */
1481 order = mb_find_order_for_block(e4b, block);
1482 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001483
1484 ex->fe_len = 1 << order;
1485 ex->fe_start = block << order;
1486 ex->fe_group = e4b->bd_group;
1487
1488 /* calc difference from given start */
1489 next = next - ex->fe_start;
1490 ex->fe_len -= next;
1491 ex->fe_start += next;
1492
1493 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001494 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001495
1496 if (block + 1 >= max)
1497 break;
1498
1499 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001500 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001501 break;
1502
Robin Dongb051d8d2011-10-26 05:30:30 -04001503 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001504
Alex Tomasc9de5602008-01-29 00:19:52 -05001505 block = next >> order;
1506 ex->fe_len += 1 << order;
1507 }
1508
1509 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1510 return ex->fe_len;
1511}
1512
1513static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1514{
1515 int ord;
1516 int mlen = 0;
1517 int max = 0;
1518 int cur;
1519 int start = ex->fe_start;
1520 int len = ex->fe_len;
1521 unsigned ret = 0;
1522 int len0 = len;
1523 void *buddy;
1524
1525 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1526 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001527 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001528 mb_check_buddy(e4b);
1529 mb_mark_used_double(e4b, start, len);
1530
1531 e4b->bd_info->bb_free -= len;
1532 if (e4b->bd_info->bb_first_free == start)
1533 e4b->bd_info->bb_first_free += len;
1534
1535 /* let's maintain fragments counter */
1536 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001537 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001538 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001539 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001540 if (mlen && max)
1541 e4b->bd_info->bb_fragments++;
1542 else if (!mlen && !max)
1543 e4b->bd_info->bb_fragments--;
1544
1545 /* let's maintain buddy itself */
1546 while (len) {
1547 ord = mb_find_order_for_block(e4b, start);
1548
1549 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1550 /* the whole chunk may be allocated at once! */
1551 mlen = 1 << ord;
1552 buddy = mb_find_buddy(e4b, ord, &max);
1553 BUG_ON((start >> ord) >= max);
1554 mb_set_bit(start >> ord, buddy);
1555 e4b->bd_info->bb_counters[ord]--;
1556 start += mlen;
1557 len -= mlen;
1558 BUG_ON(len < 0);
1559 continue;
1560 }
1561
1562 /* store for history */
1563 if (ret == 0)
1564 ret = len | (ord << 16);
1565
1566 /* we have to split large buddy */
1567 BUG_ON(ord <= 0);
1568 buddy = mb_find_buddy(e4b, ord, &max);
1569 mb_set_bit(start >> ord, buddy);
1570 e4b->bd_info->bb_counters[ord]--;
1571
1572 ord--;
1573 cur = (start >> ord) & ~1U;
1574 buddy = mb_find_buddy(e4b, ord, &max);
1575 mb_clear_bit(cur, buddy);
1576 mb_clear_bit(cur + 1, buddy);
1577 e4b->bd_info->bb_counters[ord]++;
1578 e4b->bd_info->bb_counters[ord]++;
1579 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001580 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001581
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001582 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001583 mb_check_buddy(e4b);
1584
1585 return ret;
1586}
1587
1588/*
1589 * Must be called under group lock!
1590 */
1591static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1592 struct ext4_buddy *e4b)
1593{
1594 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1595 int ret;
1596
1597 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1598 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1599
1600 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1601 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1602 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1603
1604 /* preallocation can change ac_b_ex, thus we store actually
1605 * allocated blocks for history */
1606 ac->ac_f_ex = ac->ac_b_ex;
1607
1608 ac->ac_status = AC_STATUS_FOUND;
1609 ac->ac_tail = ret & 0xffff;
1610 ac->ac_buddy = ret >> 16;
1611
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001612 /*
1613 * take the page reference. We want the page to be pinned
1614 * so that we don't get a ext4_mb_init_cache_call for this
1615 * group until we update the bitmap. That would mean we
1616 * double allocate blocks. The reference is dropped
1617 * in ext4_mb_release_context
1618 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001619 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1620 get_page(ac->ac_bitmap_page);
1621 ac->ac_buddy_page = e4b->bd_buddy_page;
1622 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001623 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001624 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001625 spin_lock(&sbi->s_md_lock);
1626 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1627 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1628 spin_unlock(&sbi->s_md_lock);
1629 }
1630}
1631
1632/*
1633 * regular allocator, for general purposes allocation
1634 */
1635
1636static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1637 struct ext4_buddy *e4b,
1638 int finish_group)
1639{
1640 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1641 struct ext4_free_extent *bex = &ac->ac_b_ex;
1642 struct ext4_free_extent *gex = &ac->ac_g_ex;
1643 struct ext4_free_extent ex;
1644 int max;
1645
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001646 if (ac->ac_status == AC_STATUS_FOUND)
1647 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001648 /*
1649 * We don't want to scan for a whole year
1650 */
1651 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1652 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1653 ac->ac_status = AC_STATUS_BREAK;
1654 return;
1655 }
1656
1657 /*
1658 * Haven't found good chunk so far, let's continue
1659 */
1660 if (bex->fe_len < gex->fe_len)
1661 return;
1662
1663 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1664 && bex->fe_group == e4b->bd_group) {
1665 /* recheck chunk's availability - we don't know
1666 * when it was found (within this lock-unlock
1667 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001668 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001669 if (max >= gex->fe_len) {
1670 ext4_mb_use_best_found(ac, e4b);
1671 return;
1672 }
1673 }
1674}
1675
1676/*
1677 * The routine checks whether found extent is good enough. If it is,
1678 * then the extent gets marked used and flag is set to the context
1679 * to stop scanning. Otherwise, the extent is compared with the
1680 * previous found extent and if new one is better, then it's stored
1681 * in the context. Later, the best found extent will be used, if
1682 * mballoc can't find good enough extent.
1683 *
1684 * FIXME: real allocation policy is to be designed yet!
1685 */
1686static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1687 struct ext4_free_extent *ex,
1688 struct ext4_buddy *e4b)
1689{
1690 struct ext4_free_extent *bex = &ac->ac_b_ex;
1691 struct ext4_free_extent *gex = &ac->ac_g_ex;
1692
1693 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001694 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1695 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001696 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1697
1698 ac->ac_found++;
1699
1700 /*
1701 * The special case - take what you catch first
1702 */
1703 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1704 *bex = *ex;
1705 ext4_mb_use_best_found(ac, e4b);
1706 return;
1707 }
1708
1709 /*
1710 * Let's check whether the chuck is good enough
1711 */
1712 if (ex->fe_len == gex->fe_len) {
1713 *bex = *ex;
1714 ext4_mb_use_best_found(ac, e4b);
1715 return;
1716 }
1717
1718 /*
1719 * If this is first found extent, just store it in the context
1720 */
1721 if (bex->fe_len == 0) {
1722 *bex = *ex;
1723 return;
1724 }
1725
1726 /*
1727 * If new found extent is better, store it in the context
1728 */
1729 if (bex->fe_len < gex->fe_len) {
1730 /* if the request isn't satisfied, any found extent
1731 * larger than previous best one is better */
1732 if (ex->fe_len > bex->fe_len)
1733 *bex = *ex;
1734 } else if (ex->fe_len > gex->fe_len) {
1735 /* if the request is satisfied, then we try to find
1736 * an extent that still satisfy the request, but is
1737 * smaller than previous one */
1738 if (ex->fe_len < bex->fe_len)
1739 *bex = *ex;
1740 }
1741
1742 ext4_mb_check_limits(ac, e4b, 0);
1743}
1744
Eric Sandeen089ceec2009-07-05 22:17:31 -04001745static noinline_for_stack
1746int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001747 struct ext4_buddy *e4b)
1748{
1749 struct ext4_free_extent ex = ac->ac_b_ex;
1750 ext4_group_t group = ex.fe_group;
1751 int max;
1752 int err;
1753
1754 BUG_ON(ex.fe_len <= 0);
1755 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1756 if (err)
1757 return err;
1758
1759 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001760 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001761
1762 if (max > 0) {
1763 ac->ac_b_ex = ex;
1764 ext4_mb_use_best_found(ac, e4b);
1765 }
1766
1767 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001768 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001769
1770 return 0;
1771}
1772
Eric Sandeen089ceec2009-07-05 22:17:31 -04001773static noinline_for_stack
1774int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001775 struct ext4_buddy *e4b)
1776{
1777 ext4_group_t group = ac->ac_g_ex.fe_group;
1778 int max;
1779 int err;
1780 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001781 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001782 struct ext4_free_extent ex;
1783
1784 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1785 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001786 if (grp->bb_free == 0)
1787 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001788
1789 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1790 if (err)
1791 return err;
1792
1793 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001794 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001795 ac->ac_g_ex.fe_len, &ex);
1796
1797 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1798 ext4_fsblk_t start;
1799
Akinobu Mita5661bd62010-03-03 23:53:39 -05001800 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1801 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001802 /* use do_div to get remainder (would be 64-bit modulo) */
1803 if (do_div(start, sbi->s_stripe) == 0) {
1804 ac->ac_found++;
1805 ac->ac_b_ex = ex;
1806 ext4_mb_use_best_found(ac, e4b);
1807 }
1808 } else if (max >= ac->ac_g_ex.fe_len) {
1809 BUG_ON(ex.fe_len <= 0);
1810 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1811 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1812 ac->ac_found++;
1813 ac->ac_b_ex = ex;
1814 ext4_mb_use_best_found(ac, e4b);
1815 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1816 /* Sometimes, caller may want to merge even small
1817 * number of blocks to an existing extent */
1818 BUG_ON(ex.fe_len <= 0);
1819 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1820 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1821 ac->ac_found++;
1822 ac->ac_b_ex = ex;
1823 ext4_mb_use_best_found(ac, e4b);
1824 }
1825 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001826 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001827
1828 return 0;
1829}
1830
1831/*
1832 * The routine scans buddy structures (not bitmap!) from given order
1833 * to max order and tries to find big enough chunk to satisfy the req
1834 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001835static noinline_for_stack
1836void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001837 struct ext4_buddy *e4b)
1838{
1839 struct super_block *sb = ac->ac_sb;
1840 struct ext4_group_info *grp = e4b->bd_info;
1841 void *buddy;
1842 int i;
1843 int k;
1844 int max;
1845
1846 BUG_ON(ac->ac_2order <= 0);
1847 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1848 if (grp->bb_counters[i] == 0)
1849 continue;
1850
1851 buddy = mb_find_buddy(e4b, i, &max);
1852 BUG_ON(buddy == NULL);
1853
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001854 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001855 BUG_ON(k >= max);
1856
1857 ac->ac_found++;
1858
1859 ac->ac_b_ex.fe_len = 1 << i;
1860 ac->ac_b_ex.fe_start = k << i;
1861 ac->ac_b_ex.fe_group = e4b->bd_group;
1862
1863 ext4_mb_use_best_found(ac, e4b);
1864
1865 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1866
1867 if (EXT4_SB(sb)->s_mb_stats)
1868 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1869
1870 break;
1871 }
1872}
1873
1874/*
1875 * The routine scans the group and measures all found extents.
1876 * In order to optimize scanning, caller must pass number of
1877 * free blocks in the group, so the routine can know upper limit.
1878 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001879static noinline_for_stack
1880void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001881 struct ext4_buddy *e4b)
1882{
1883 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001884 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001885 struct ext4_free_extent ex;
1886 int i;
1887 int free;
1888
1889 free = e4b->bd_info->bb_free;
1890 BUG_ON(free <= 0);
1891
1892 i = e4b->bd_info->bb_first_free;
1893
1894 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001895 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001896 EXT4_CLUSTERS_PER_GROUP(sb), i);
1897 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001898 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001899 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001900 * free blocks even though group info says we
1901 * we have free blocks
1902 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001903 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001904 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001905 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001906 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001907 break;
1908 }
1909
Robin Dong15c006a2012-08-17 10:02:17 -04001910 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001911 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001912 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001913 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001914 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001915 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001916 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001917 /*
1918 * The number of free blocks differs. This mostly
1919 * indicate that the bitmap is corrupt. So exit
1920 * without claiming the space.
1921 */
1922 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001923 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001924
1925 ext4_mb_measure_extent(ac, &ex, e4b);
1926
1927 i += ex.fe_len;
1928 free -= ex.fe_len;
1929 }
1930
1931 ext4_mb_check_limits(ac, e4b, 1);
1932}
1933
1934/*
1935 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001936 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001937 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001938static noinline_for_stack
1939void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001940 struct ext4_buddy *e4b)
1941{
1942 struct super_block *sb = ac->ac_sb;
1943 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001944 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001945 struct ext4_free_extent ex;
1946 ext4_fsblk_t first_group_block;
1947 ext4_fsblk_t a;
1948 ext4_grpblk_t i;
1949 int max;
1950
1951 BUG_ON(sbi->s_stripe == 0);
1952
1953 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001954 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1955
Alex Tomasc9de5602008-01-29 00:19:52 -05001956 a = first_group_block + sbi->s_stripe - 1;
1957 do_div(a, sbi->s_stripe);
1958 i = (a * sbi->s_stripe) - first_group_block;
1959
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001960 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001961 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04001962 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001963 if (max >= sbi->s_stripe) {
1964 ac->ac_found++;
1965 ac->ac_b_ex = ex;
1966 ext4_mb_use_best_found(ac, e4b);
1967 break;
1968 }
1969 }
1970 i += sbi->s_stripe;
1971 }
1972}
1973
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001974/* This is now called BEFORE we load the buddy bitmap. */
Alex Tomasc9de5602008-01-29 00:19:52 -05001975static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1976 ext4_group_t group, int cr)
1977{
1978 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04001979 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001980 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1981
1982 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001983
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04001984 free = grp->bb_free;
1985 if (free == 0)
1986 return 0;
1987 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
1988 return 0;
1989
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001990 /* We only do this if the grp has never been initialized */
1991 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1992 int ret = ext4_mb_init_group(ac->ac_sb, group);
1993 if (ret)
1994 return 0;
1995 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001996
Alex Tomasc9de5602008-01-29 00:19:52 -05001997 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05001998 if (fragments == 0)
1999 return 0;
2000
2001 switch (cr) {
2002 case 0:
2003 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002004
Theodore Ts'oa4912122009-03-12 12:18:34 -04002005 /* Avoid using the first bg of a flexgroup for data files */
2006 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2007 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2008 ((group % flex_size) == 0))
2009 return 0;
2010
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002011 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2012 (free / fragments) >= ac->ac_g_ex.fe_len)
2013 return 1;
2014
2015 if (grp->bb_largest_free_order < ac->ac_2order)
2016 return 0;
2017
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002018 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002019 case 1:
2020 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2021 return 1;
2022 break;
2023 case 2:
2024 if (free >= ac->ac_g_ex.fe_len)
2025 return 1;
2026 break;
2027 case 3:
2028 return 1;
2029 default:
2030 BUG();
2031 }
2032
2033 return 0;
2034}
2035
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002036static noinline_for_stack int
2037ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002038{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002039 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002040 int cr;
2041 int err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002042 struct ext4_sb_info *sbi;
2043 struct super_block *sb;
2044 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002045
2046 sb = ac->ac_sb;
2047 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002048 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002049 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002050 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002051 ngroups = sbi->s_blockfile_groups;
2052
Alex Tomasc9de5602008-01-29 00:19:52 -05002053 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2054
2055 /* first, try the goal */
2056 err = ext4_mb_find_by_goal(ac, &e4b);
2057 if (err || ac->ac_status == AC_STATUS_FOUND)
2058 goto out;
2059
2060 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2061 goto out;
2062
2063 /*
2064 * ac->ac2_order is set only if the fe_len is a power of 2
2065 * if ac2_order is set we also set criteria to 0 so that we
2066 * try exact allocation using buddy.
2067 */
2068 i = fls(ac->ac_g_ex.fe_len);
2069 ac->ac_2order = 0;
2070 /*
2071 * We search using buddy data only if the order of the request
2072 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002073 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002074 */
2075 if (i >= sbi->s_mb_order2_reqs) {
2076 /*
2077 * This should tell if fe_len is exactly power of 2
2078 */
2079 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2080 ac->ac_2order = i - 1;
2081 }
2082
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002083 /* if stream allocation is enabled, use global goal */
2084 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002085 /* TBD: may be hot point */
2086 spin_lock(&sbi->s_md_lock);
2087 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2088 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2089 spin_unlock(&sbi->s_md_lock);
2090 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002091
Alex Tomasc9de5602008-01-29 00:19:52 -05002092 /* Let's just scan groups to find more-less suitable blocks */
2093 cr = ac->ac_2order ? 0 : 1;
2094 /*
2095 * cr == 0 try to get exact allocation,
2096 * cr == 3 try to get anything
2097 */
2098repeat:
2099 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2100 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002101 /*
2102 * searching for the right group start
2103 * from the goal value specified
2104 */
2105 group = ac->ac_g_ex.fe_group;
2106
Theodore Ts'o8df96752009-05-01 08:50:38 -04002107 for (i = 0; i < ngroups; group++, i++) {
Lachlan McIlroye6155732013-05-05 23:10:00 -04002108 /*
2109 * Artificially restricted ngroups for non-extent
2110 * files makes group > ngroups possible on first loop.
2111 */
2112 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002113 group = 0;
2114
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002115 /* This now checks without needing the buddy page */
2116 if (!ext4_mb_good_group(ac, group, cr))
Alex Tomasc9de5602008-01-29 00:19:52 -05002117 continue;
2118
Alex Tomasc9de5602008-01-29 00:19:52 -05002119 err = ext4_mb_load_buddy(sb, group, &e4b);
2120 if (err)
2121 goto out;
2122
2123 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002124
2125 /*
2126 * We need to check again after locking the
2127 * block group
2128 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002129 if (!ext4_mb_good_group(ac, group, cr)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002130 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002131 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002132 continue;
2133 }
2134
2135 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002136 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002137 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002138 else if (cr == 1 && sbi->s_stripe &&
2139 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002140 ext4_mb_scan_aligned(ac, &e4b);
2141 else
2142 ext4_mb_complex_scan_group(ac, &e4b);
2143
2144 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002145 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002146
2147 if (ac->ac_status != AC_STATUS_CONTINUE)
2148 break;
2149 }
2150 }
2151
2152 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2153 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2154 /*
2155 * We've been searching too long. Let's try to allocate
2156 * the best chunk we've found so far
2157 */
2158
2159 ext4_mb_try_best_found(ac, &e4b);
2160 if (ac->ac_status != AC_STATUS_FOUND) {
2161 /*
2162 * Someone more lucky has already allocated it.
2163 * The only thing we can do is just take first
2164 * found block(s)
2165 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2166 */
2167 ac->ac_b_ex.fe_group = 0;
2168 ac->ac_b_ex.fe_start = 0;
2169 ac->ac_b_ex.fe_len = 0;
2170 ac->ac_status = AC_STATUS_CONTINUE;
2171 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2172 cr = 3;
2173 atomic_inc(&sbi->s_mb_lost_chunks);
2174 goto repeat;
2175 }
2176 }
2177out:
2178 return err;
2179}
2180
Alex Tomasc9de5602008-01-29 00:19:52 -05002181static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2182{
2183 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002184 ext4_group_t group;
2185
Theodore Ts'o8df96752009-05-01 08:50:38 -04002186 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002187 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002188 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002189 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002190}
2191
2192static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2193{
2194 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002195 ext4_group_t group;
2196
2197 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002198 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002199 return NULL;
2200 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002201 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002202}
2203
2204static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2205{
2206 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002207 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002208 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002209 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002210 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002211 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002212 struct sg {
2213 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002214 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002215 } sg;
2216
2217 group--;
2218 if (group == 0)
2219 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2220 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2221 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2222 "group", "free", "frags", "first",
2223 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2224 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2225
2226 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2227 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002228 grinfo = ext4_get_group_info(sb, group);
2229 /* Load the group info in memory only if not already loaded. */
2230 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2231 err = ext4_mb_load_buddy(sb, group, &e4b);
2232 if (err) {
2233 seq_printf(seq, "#%-5u: I/O error\n", group);
2234 return 0;
2235 }
2236 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002237 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002238
Alex Tomasc9de5602008-01-29 00:19:52 -05002239 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002240
2241 if (buddy_loaded)
2242 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002243
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002244 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002245 sg.info.bb_fragments, sg.info.bb_first_free);
2246 for (i = 0; i <= 13; i++)
2247 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2248 sg.info.bb_counters[i] : 0);
2249 seq_printf(seq, " ]\n");
2250
2251 return 0;
2252}
2253
2254static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2255{
2256}
2257
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002258static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002259 .start = ext4_mb_seq_groups_start,
2260 .next = ext4_mb_seq_groups_next,
2261 .stop = ext4_mb_seq_groups_stop,
2262 .show = ext4_mb_seq_groups_show,
2263};
2264
2265static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2266{
Al Virod9dda782013-03-31 18:16:14 -04002267 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002268 int rc;
2269
2270 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2271 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002272 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002273 m->private = sb;
2274 }
2275 return rc;
2276
2277}
2278
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002279static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002280 .owner = THIS_MODULE,
2281 .open = ext4_mb_seq_groups_open,
2282 .read = seq_read,
2283 .llseek = seq_lseek,
2284 .release = seq_release,
2285};
2286
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002287static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2288{
2289 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2290 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2291
2292 BUG_ON(!cachep);
2293 return cachep;
2294}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002295
Theodore Ts'o28623c22012-09-05 01:31:50 -04002296/*
2297 * Allocate the top-level s_group_info array for the specified number
2298 * of groups
2299 */
2300int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2301{
2302 struct ext4_sb_info *sbi = EXT4_SB(sb);
2303 unsigned size;
2304 struct ext4_group_info ***new_groupinfo;
2305
2306 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2307 EXT4_DESC_PER_BLOCK_BITS(sb);
2308 if (size <= sbi->s_group_info_size)
2309 return 0;
2310
2311 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2312 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2313 if (!new_groupinfo) {
2314 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2315 return -ENOMEM;
2316 }
2317 if (sbi->s_group_info) {
2318 memcpy(new_groupinfo, sbi->s_group_info,
2319 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2320 ext4_kvfree(sbi->s_group_info);
2321 }
2322 sbi->s_group_info = new_groupinfo;
2323 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2324 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2325 sbi->s_group_info_size);
2326 return 0;
2327}
2328
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002329/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002330int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002331 struct ext4_group_desc *desc)
2332{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002333 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002334 int metalen = 0;
2335 struct ext4_sb_info *sbi = EXT4_SB(sb);
2336 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002337 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002338
2339 /*
2340 * First check if this group is the first of a reserved block.
2341 * If it's true, we have to allocate a new table of pointers
2342 * to ext4_group_info structures
2343 */
2344 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2345 metalen = sizeof(*meta_group_info) <<
2346 EXT4_DESC_PER_BLOCK_BITS(sb);
2347 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2348 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002349 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002350 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002351 goto exit_meta_group_info;
2352 }
2353 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2354 meta_group_info;
2355 }
2356
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002357 meta_group_info =
2358 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2359 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2360
Wei Yongjun85556c92012-09-26 20:43:37 -04002361 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002362 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002363 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002364 goto exit_group_info;
2365 }
2366 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2367 &(meta_group_info[i]->bb_state));
2368
2369 /*
2370 * initialize bb_free to be able to skip
2371 * empty groups without initialization
2372 */
2373 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2374 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002375 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002376 } else {
2377 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002378 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002379 }
2380
2381 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002382 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002383 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002384 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002385
2386#ifdef DOUBLE_CHECK
2387 {
2388 struct buffer_head *bh;
2389 meta_group_info[i]->bb_bitmap =
2390 kmalloc(sb->s_blocksize, GFP_KERNEL);
2391 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2392 bh = ext4_read_block_bitmap(sb, group);
2393 BUG_ON(bh == NULL);
2394 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2395 sb->s_blocksize);
2396 put_bh(bh);
2397 }
2398#endif
2399
2400 return 0;
2401
2402exit_group_info:
2403 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002404 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002405 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002406 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2407 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002408exit_meta_group_info:
2409 return -ENOMEM;
2410} /* ext4_mb_add_groupinfo */
2411
Alex Tomasc9de5602008-01-29 00:19:52 -05002412static int ext4_mb_init_backend(struct super_block *sb)
2413{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002414 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002415 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002416 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002417 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002418 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002419 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002420
Theodore Ts'o28623c22012-09-05 01:31:50 -04002421 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2422 if (err)
2423 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002424
Alex Tomasc9de5602008-01-29 00:19:52 -05002425 sbi->s_buddy_cache = new_inode(sb);
2426 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002427 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002428 goto err_freesgi;
2429 }
Yu Jian48e60612011-08-01 17:41:39 -04002430 /* To avoid potentially colliding with an valid on-disk inode number,
2431 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2432 * not in the inode hash, so it should never be found by iget(), but
2433 * this will avoid confusion if it ever shows up during debugging. */
2434 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002435 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002436 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002437 desc = ext4_get_group_desc(sb, i, NULL);
2438 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002439 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002440 goto err_freebuddy;
2441 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002442 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2443 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002444 }
2445
2446 return 0;
2447
2448err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002449 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002450 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002451 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002452 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002453 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002454 kfree(sbi->s_group_info[i]);
2455 iput(sbi->s_buddy_cache);
2456err_freesgi:
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002457 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002458 return -ENOMEM;
2459}
2460
Eric Sandeen2892c152011-02-12 08:12:18 -05002461static void ext4_groupinfo_destroy_slabs(void)
2462{
2463 int i;
2464
2465 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2466 if (ext4_groupinfo_caches[i])
2467 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2468 ext4_groupinfo_caches[i] = NULL;
2469 }
2470}
2471
2472static int ext4_groupinfo_create_slab(size_t size)
2473{
2474 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2475 int slab_size;
2476 int blocksize_bits = order_base_2(size);
2477 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2478 struct kmem_cache *cachep;
2479
2480 if (cache_index >= NR_GRPINFO_CACHES)
2481 return -EINVAL;
2482
2483 if (unlikely(cache_index < 0))
2484 cache_index = 0;
2485
2486 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2487 if (ext4_groupinfo_caches[cache_index]) {
2488 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2489 return 0; /* Already created */
2490 }
2491
2492 slab_size = offsetof(struct ext4_group_info,
2493 bb_counters[blocksize_bits + 2]);
2494
2495 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2496 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2497 NULL);
2498
Tao Ma823ba012011-07-11 18:26:01 -04002499 ext4_groupinfo_caches[cache_index] = cachep;
2500
Eric Sandeen2892c152011-02-12 08:12:18 -05002501 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2502 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002503 printk(KERN_EMERG
2504 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002505 return -ENOMEM;
2506 }
2507
Eric Sandeen2892c152011-02-12 08:12:18 -05002508 return 0;
2509}
2510
Akira Fujita9d990122012-05-28 14:19:25 -04002511int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002512{
2513 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002514 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002515 unsigned offset;
2516 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002517 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002518
Eric Sandeen19278052009-08-25 22:36:25 -04002519 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002520
2521 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2522 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002523 ret = -ENOMEM;
2524 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002525 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002526
Eric Sandeen19278052009-08-25 22:36:25 -04002527 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002528 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2529 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002530 ret = -ENOMEM;
2531 goto out;
2532 }
2533
Eric Sandeen2892c152011-02-12 08:12:18 -05002534 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2535 if (ret < 0)
2536 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002537
2538 /* order 0 is regular bitmap */
2539 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2540 sbi->s_mb_offsets[0] = 0;
2541
2542 i = 1;
2543 offset = 0;
2544 max = sb->s_blocksize << 2;
2545 do {
2546 sbi->s_mb_offsets[i] = offset;
2547 sbi->s_mb_maxs[i] = max;
2548 offset += 1 << (sb->s_blocksize_bits - i);
2549 max = max >> 1;
2550 i++;
2551 } while (i <= sb->s_blocksize_bits + 1);
2552
Alex Tomasc9de5602008-01-29 00:19:52 -05002553 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002554 spin_lock_init(&sbi->s_bal_lock);
2555
2556 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2557 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2558 sbi->s_mb_stats = MB_DEFAULT_STATS;
2559 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2560 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002561 /*
2562 * The default group preallocation is 512, which for 4k block
2563 * sizes translates to 2 megabytes. However for bigalloc file
2564 * systems, this is probably too big (i.e, if the cluster size
2565 * is 1 megabyte, then group preallocation size becomes half a
2566 * gigabyte!). As a default, we will keep a two megabyte
2567 * group pralloc size for cluster sizes up to 64k, and after
2568 * that, we will force a minimum group preallocation size of
2569 * 32 clusters. This translates to 8 megs when the cluster
2570 * size is 256k, and 32 megs when the cluster size is 1 meg,
2571 * which seems reasonable as a default.
2572 */
2573 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2574 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002575 /*
2576 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2577 * to the lowest multiple of s_stripe which is bigger than
2578 * the s_mb_group_prealloc as determined above. We want
2579 * the preallocation size to be an exact multiple of the
2580 * RAID stripe size so that preallocations don't fragment
2581 * the stripes.
2582 */
2583 if (sbi->s_stripe > 1) {
2584 sbi->s_mb_group_prealloc = roundup(
2585 sbi->s_mb_group_prealloc, sbi->s_stripe);
2586 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002587
Eric Sandeen730c2132008-09-13 15:23:29 -04002588 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002589 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002590 ret = -ENOMEM;
Tao Ma7aa0bae2011-10-06 10:22:28 -04002591 goto out_free_groupinfo_slab;
Alex Tomasc9de5602008-01-29 00:19:52 -05002592 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002593 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002594 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002595 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002596 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002597 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2598 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002599 spin_lock_init(&lg->lg_prealloc_lock);
2600 }
2601
Yu Jian79a77c52011-08-01 17:41:46 -04002602 /* init file for buddy data */
2603 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002604 if (ret != 0)
2605 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002606
Theodore Ts'o296c3552009-09-30 00:32:42 -04002607 if (sbi->s_proc)
2608 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2609 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002610
Tao Ma7aa0bae2011-10-06 10:22:28 -04002611 return 0;
2612
2613out_free_locality_groups:
2614 free_percpu(sbi->s_locality_groups);
2615 sbi->s_locality_groups = NULL;
2616out_free_groupinfo_slab:
2617 ext4_groupinfo_destroy_slabs();
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002618out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002619 kfree(sbi->s_mb_offsets);
2620 sbi->s_mb_offsets = NULL;
2621 kfree(sbi->s_mb_maxs);
2622 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002623 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002624}
2625
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002626/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002627static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2628{
2629 struct ext4_prealloc_space *pa;
2630 struct list_head *cur, *tmp;
2631 int count = 0;
2632
2633 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2634 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2635 list_del(&pa->pa_group_list);
2636 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002637 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002638 }
2639 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002640 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002641
2642}
2643
2644int ext4_mb_release(struct super_block *sb)
2645{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002646 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002647 ext4_group_t i;
2648 int num_meta_group_infos;
2649 struct ext4_group_info *grinfo;
2650 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002651 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002652
Salman Qazi95599962012-05-31 23:52:14 -04002653 if (sbi->s_proc)
2654 remove_proc_entry("mb_groups", sbi->s_proc);
2655
Alex Tomasc9de5602008-01-29 00:19:52 -05002656 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002657 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002658 grinfo = ext4_get_group_info(sb, i);
2659#ifdef DOUBLE_CHECK
2660 kfree(grinfo->bb_bitmap);
2661#endif
2662 ext4_lock_group(sb, i);
2663 ext4_mb_cleanup_pa(grinfo);
2664 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002665 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002666 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002667 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002668 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2669 EXT4_DESC_PER_BLOCK_BITS(sb);
2670 for (i = 0; i < num_meta_group_infos; i++)
2671 kfree(sbi->s_group_info[i]);
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002672 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002673 }
2674 kfree(sbi->s_mb_offsets);
2675 kfree(sbi->s_mb_maxs);
2676 if (sbi->s_buddy_cache)
2677 iput(sbi->s_buddy_cache);
2678 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002679 ext4_msg(sb, KERN_INFO,
2680 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002681 atomic_read(&sbi->s_bal_allocated),
2682 atomic_read(&sbi->s_bal_reqs),
2683 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002684 ext4_msg(sb, KERN_INFO,
2685 "mballoc: %u extents scanned, %u goal hits, "
2686 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002687 atomic_read(&sbi->s_bal_ex_scanned),
2688 atomic_read(&sbi->s_bal_goals),
2689 atomic_read(&sbi->s_bal_2orders),
2690 atomic_read(&sbi->s_bal_breaks),
2691 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002692 ext4_msg(sb, KERN_INFO,
2693 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002694 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002695 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002696 ext4_msg(sb, KERN_INFO,
2697 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002698 atomic_read(&sbi->s_mb_preallocated),
2699 atomic_read(&sbi->s_mb_discarded));
2700 }
2701
Eric Sandeen730c2132008-09-13 15:23:29 -04002702 free_percpu(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05002703
2704 return 0;
2705}
2706
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002707static inline int ext4_issue_discard(struct super_block *sb,
Theodore Ts'o84130192011-09-09 18:50:51 -04002708 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
Jiaying Zhang5c521832010-07-27 11:56:05 -04002709{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002710 ext4_fsblk_t discard_block;
2711
Theodore Ts'o84130192011-09-09 18:50:51 -04002712 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2713 ext4_group_first_block_no(sb, block_group));
2714 count = EXT4_C2B(EXT4_SB(sb), count);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002715 trace_ext4_discard_blocks(sb,
2716 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002717 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002718}
2719
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002720/*
2721 * This function is called by the jbd2 layer once the commit has finished,
2722 * so we know we can free the blocks that were released with that commit.
2723 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002724static void ext4_free_data_callback(struct super_block *sb,
2725 struct ext4_journal_cb_entry *jce,
2726 int rc)
Alex Tomasc9de5602008-01-29 00:19:52 -05002727{
Bobi Jam18aadd42012-02-20 17:53:02 -05002728 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
Alex Tomasc9de5602008-01-29 00:19:52 -05002729 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002730 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002731 int err, count = 0, count2 = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002732
Bobi Jam18aadd42012-02-20 17:53:02 -05002733 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2734 entry->efd_count, entry->efd_group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002735
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05002736 if (test_opt(sb, DISCARD)) {
2737 err = ext4_issue_discard(sb, entry->efd_group,
2738 entry->efd_start_cluster,
2739 entry->efd_count);
2740 if (err && err != -EOPNOTSUPP)
2741 ext4_msg(sb, KERN_WARNING, "discard request in"
2742 " group:%d block:%d count:%d failed"
2743 " with %d", entry->efd_group,
2744 entry->efd_start_cluster,
2745 entry->efd_count, err);
2746 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002747
Bobi Jam18aadd42012-02-20 17:53:02 -05002748 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2749 /* we expect to find existing buddy because it's pinned */
2750 BUG_ON(err != 0);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002751
Alex Tomasc9de5602008-01-29 00:19:52 -05002752
Bobi Jam18aadd42012-02-20 17:53:02 -05002753 db = e4b.bd_info;
2754 /* there are blocks to put in buddy to make them really free */
2755 count += entry->efd_count;
2756 count2++;
2757 ext4_lock_group(sb, entry->efd_group);
2758 /* Take it out of per group rb tree */
2759 rb_erase(&entry->efd_node, &(db->bb_free_root));
2760 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002761
Bobi Jam18aadd42012-02-20 17:53:02 -05002762 /*
2763 * Clear the trimmed flag for the group so that the next
2764 * ext4_trim_fs can trim it.
2765 * If the volume is mounted with -o discard, online discard
2766 * is supported and the free blocks will be trimmed online.
2767 */
2768 if (!test_opt(sb, DISCARD))
2769 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2770
2771 if (!db->bb_free_root.rb_node) {
2772 /* No more items in the per group rb tree
2773 * balance refcounts from ext4_mb_free_metadata()
Tao Ma3d56b8d2011-07-11 00:03:38 -04002774 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002775 page_cache_release(e4b.bd_buddy_page);
2776 page_cache_release(e4b.bd_bitmap_page);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002777 }
Bobi Jam18aadd42012-02-20 17:53:02 -05002778 ext4_unlock_group(sb, entry->efd_group);
2779 kmem_cache_free(ext4_free_data_cachep, entry);
2780 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002781
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002782 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002783}
2784
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002785int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002786{
Theodore Ts'o16828082010-10-27 21:30:09 -04002787 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2788 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002789 if (ext4_pspace_cachep == NULL)
2790 return -ENOMEM;
2791
Theodore Ts'o16828082010-10-27 21:30:09 -04002792 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2793 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002794 if (ext4_ac_cachep == NULL) {
2795 kmem_cache_destroy(ext4_pspace_cachep);
2796 return -ENOMEM;
2797 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002798
Bobi Jam18aadd42012-02-20 17:53:02 -05002799 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2800 SLAB_RECLAIM_ACCOUNT);
2801 if (ext4_free_data_cachep == NULL) {
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002802 kmem_cache_destroy(ext4_pspace_cachep);
2803 kmem_cache_destroy(ext4_ac_cachep);
2804 return -ENOMEM;
2805 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002806 return 0;
2807}
2808
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002809void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002810{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002811 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002812 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2813 * before destroying the slab cache.
2814 */
2815 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002816 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002817 kmem_cache_destroy(ext4_ac_cachep);
Bobi Jam18aadd42012-02-20 17:53:02 -05002818 kmem_cache_destroy(ext4_free_data_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002819 ext4_groupinfo_destroy_slabs();
Alex Tomasc9de5602008-01-29 00:19:52 -05002820}
2821
2822
2823/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002824 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002825 * Returns 0 if success or error code
2826 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002827static noinline_for_stack int
2828ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002829 handle_t *handle, unsigned int reserv_clstrs)
Alex Tomasc9de5602008-01-29 00:19:52 -05002830{
2831 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002832 struct ext4_group_desc *gdp;
2833 struct buffer_head *gdp_bh;
2834 struct ext4_sb_info *sbi;
2835 struct super_block *sb;
2836 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002837 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002838
2839 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2840 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2841
2842 sb = ac->ac_sb;
2843 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002844
2845 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002846 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002847 if (!bitmap_bh)
2848 goto out_err;
2849
2850 err = ext4_journal_get_write_access(handle, bitmap_bh);
2851 if (err)
2852 goto out_err;
2853
2854 err = -EIO;
2855 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2856 if (!gdp)
2857 goto out_err;
2858
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002859 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002860 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002861
Alex Tomasc9de5602008-01-29 00:19:52 -05002862 err = ext4_journal_get_write_access(handle, gdp_bh);
2863 if (err)
2864 goto out_err;
2865
Akinobu Mitabda00de2010-03-03 23:53:25 -05002866 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002867
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002868 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002869 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002870 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002871 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002872 /* File system mounted not to panic on error
2873 * Fix the bitmap and repeat the block allocation
2874 * We leak some of the blocks here.
2875 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002876 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002877 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2878 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002879 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002880 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002881 if (!err)
2882 err = -EAGAIN;
2883 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002884 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002885
2886 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002887#ifdef AGGRESSIVE_CHECK
2888 {
2889 int i;
2890 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2891 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2892 bitmap_bh->b_data));
2893 }
2894 }
2895#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002896 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2897 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002898 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2899 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002900 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002901 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002902 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002903 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002904 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2905 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002906 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002907 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002908
2909 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002910 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002911 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002912 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002913 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002914 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2915 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002916 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2917 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002918
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002919 if (sbi->s_log_groups_per_flex) {
2920 ext4_group_t flex_group = ext4_flex_group(sbi,
2921 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002922 atomic64_sub(ac->ac_b_ex.fe_len,
2923 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002924 }
2925
Frank Mayhar03901312009-01-07 00:06:22 -05002926 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002927 if (err)
2928 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002929 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002930
2931out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002932 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002933 return err;
2934}
2935
2936/*
2937 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002938 * Group request are normalized to s_mb_group_prealloc, which goes to
2939 * s_strip if we set the same via mount option.
2940 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002941 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002942 *
2943 * XXX: should we try to preallocate more than the group has now?
2944 */
2945static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2946{
2947 struct super_block *sb = ac->ac_sb;
2948 struct ext4_locality_group *lg = ac->ac_lg;
2949
2950 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002951 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002952 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002953 current->pid, ac->ac_g_ex.fe_len);
2954}
2955
2956/*
2957 * Normalization means making request better in terms of
2958 * size and alignment
2959 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002960static noinline_for_stack void
2961ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05002962 struct ext4_allocation_request *ar)
2963{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002964 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002965 int bsbits, max;
2966 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05002967 loff_t size, start_off;
2968 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04002969 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05002970 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04002971 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05002972
2973 /* do normalize only data requests, metadata requests
2974 do not need preallocation */
2975 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2976 return;
2977
2978 /* sometime caller may want exact blocks */
2979 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2980 return;
2981
2982 /* caller may indicate that preallocation isn't
2983 * required (it's a tail, for example) */
2984 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2985 return;
2986
2987 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2988 ext4_mb_normalize_group_request(ac);
2989 return ;
2990 }
2991
2992 bsbits = ac->ac_sb->s_blocksize_bits;
2993
2994 /* first, let's learn actual file size
2995 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002996 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002997 size = size << bsbits;
2998 if (size < i_size_read(ac->ac_inode))
2999 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003000 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003001
Valerie Clement19304792008-05-13 19:31:14 -04003002 /* max size of free chunks */
3003 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003004
Valerie Clement19304792008-05-13 19:31:14 -04003005#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3006 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003007
3008 /* first, try to predict filesize */
3009 /* XXX: should this table be tunable? */
3010 start_off = 0;
3011 if (size <= 16 * 1024) {
3012 size = 16 * 1024;
3013 } else if (size <= 32 * 1024) {
3014 size = 32 * 1024;
3015 } else if (size <= 64 * 1024) {
3016 size = 64 * 1024;
3017 } else if (size <= 128 * 1024) {
3018 size = 128 * 1024;
3019 } else if (size <= 256 * 1024) {
3020 size = 256 * 1024;
3021 } else if (size <= 512 * 1024) {
3022 size = 512 * 1024;
3023 } else if (size <= 1024 * 1024) {
3024 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003025 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003026 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003027 (21 - bsbits)) << 21;
3028 size = 2 * 1024 * 1024;
3029 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003030 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3031 (22 - bsbits)) << 22;
3032 size = 4 * 1024 * 1024;
3033 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003034 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003035 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3036 (23 - bsbits)) << 23;
3037 size = 8 * 1024 * 1024;
3038 } else {
3039 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3040 size = ac->ac_o_ex.fe_len << bsbits;
3041 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003042 size = size >> bsbits;
3043 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003044
3045 /* don't cover already allocated blocks in selected range */
3046 if (ar->pleft && start <= ar->lleft) {
3047 size -= ar->lleft + 1 - start;
3048 start = ar->lleft + 1;
3049 }
3050 if (ar->pright && start + size - 1 >= ar->lright)
3051 size -= start + size - ar->lright;
3052
3053 end = start + size;
3054
3055 /* check we don't cross already preallocated blocks */
3056 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003057 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003058 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003059
Alex Tomasc9de5602008-01-29 00:19:52 -05003060 if (pa->pa_deleted)
3061 continue;
3062 spin_lock(&pa->pa_lock);
3063 if (pa->pa_deleted) {
3064 spin_unlock(&pa->pa_lock);
3065 continue;
3066 }
3067
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003068 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3069 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003070
3071 /* PA must not overlap original request */
3072 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3073 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3074
Eric Sandeen38877f42009-08-17 23:55:24 -04003075 /* skip PAs this normalized request doesn't overlap with */
3076 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003077 spin_unlock(&pa->pa_lock);
3078 continue;
3079 }
3080 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3081
Eric Sandeen38877f42009-08-17 23:55:24 -04003082 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003083 if (pa_end <= ac->ac_o_ex.fe_logical) {
3084 BUG_ON(pa_end < start);
3085 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003086 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003087 BUG_ON(pa->pa_lstart > end);
3088 end = pa->pa_lstart;
3089 }
3090 spin_unlock(&pa->pa_lock);
3091 }
3092 rcu_read_unlock();
3093 size = end - start;
3094
3095 /* XXX: extra loop to check we really don't overlap preallocations */
3096 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003097 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003098 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003099
Alex Tomasc9de5602008-01-29 00:19:52 -05003100 spin_lock(&pa->pa_lock);
3101 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003102 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3103 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003104 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3105 }
3106 spin_unlock(&pa->pa_lock);
3107 }
3108 rcu_read_unlock();
3109
3110 if (start + size <= ac->ac_o_ex.fe_logical &&
3111 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003112 ext4_msg(ac->ac_sb, KERN_ERR,
3113 "start %lu, size %lu, fe_logical %lu",
3114 (unsigned long) start, (unsigned long) size,
3115 (unsigned long) ac->ac_o_ex.fe_logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05003116 }
3117 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3118 start > ac->ac_o_ex.fe_logical);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003119 BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003120
3121 /* now prepare goal request */
3122
3123 /* XXX: is it better to align blocks WRT to logical
3124 * placement or satisfy big request as is */
3125 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003126 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003127
3128 /* define goal start in order to merge */
3129 if (ar->pright && (ar->lright == (start + size))) {
3130 /* merge to the right */
3131 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3132 &ac->ac_f_ex.fe_group,
3133 &ac->ac_f_ex.fe_start);
3134 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3135 }
3136 if (ar->pleft && (ar->lleft + 1 == start)) {
3137 /* merge to the left */
3138 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3139 &ac->ac_f_ex.fe_group,
3140 &ac->ac_f_ex.fe_start);
3141 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3142 }
3143
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003144 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003145 (unsigned) orig_size, (unsigned) start);
3146}
3147
3148static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3149{
3150 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3151
3152 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3153 atomic_inc(&sbi->s_bal_reqs);
3154 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003155 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003156 atomic_inc(&sbi->s_bal_success);
3157 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3158 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3159 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3160 atomic_inc(&sbi->s_bal_goals);
3161 if (ac->ac_found > sbi->s_mb_max_to_scan)
3162 atomic_inc(&sbi->s_bal_breaks);
3163 }
3164
Theodore Ts'o296c3552009-09-30 00:32:42 -04003165 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3166 trace_ext4_mballoc_alloc(ac);
3167 else
3168 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003169}
3170
3171/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003172 * Called on failure; free up any blocks from the inode PA for this
3173 * context. We don't need this for MB_GROUP_PA because we only change
3174 * pa_free in ext4_mb_release_context(), but on failure, we've already
3175 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3176 */
3177static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3178{
3179 struct ext4_prealloc_space *pa = ac->ac_pa;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003180
Zheng Liu400db9d2012-05-28 17:53:53 -04003181 if (pa && pa->pa_type == MB_INODE_PA)
3182 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003183}
3184
3185/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003186 * use blocks preallocated to inode
3187 */
3188static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3189 struct ext4_prealloc_space *pa)
3190{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003191 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003192 ext4_fsblk_t start;
3193 ext4_fsblk_t end;
3194 int len;
3195
3196 /* found preallocated blocks, use them */
3197 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003198 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3199 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3200 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003201 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3202 &ac->ac_b_ex.fe_start);
3203 ac->ac_b_ex.fe_len = len;
3204 ac->ac_status = AC_STATUS_FOUND;
3205 ac->ac_pa = pa;
3206
3207 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003208 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003209 BUG_ON(pa->pa_free < len);
3210 pa->pa_free -= len;
3211
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003212 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003213}
3214
3215/*
3216 * use blocks preallocated to locality group
3217 */
3218static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3219 struct ext4_prealloc_space *pa)
3220{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003221 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003222
Alex Tomasc9de5602008-01-29 00:19:52 -05003223 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3224 &ac->ac_b_ex.fe_group,
3225 &ac->ac_b_ex.fe_start);
3226 ac->ac_b_ex.fe_len = len;
3227 ac->ac_status = AC_STATUS_FOUND;
3228 ac->ac_pa = pa;
3229
3230 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003231 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003232 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003233 * in on-disk bitmap -- see ext4_mb_release_context()
3234 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003235 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003236 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003237}
3238
3239/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003240 * Return the prealloc space that have minimal distance
3241 * from the goal block. @cpa is the prealloc
3242 * space that is having currently known minimal distance
3243 * from the goal block.
3244 */
3245static struct ext4_prealloc_space *
3246ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3247 struct ext4_prealloc_space *pa,
3248 struct ext4_prealloc_space *cpa)
3249{
3250 ext4_fsblk_t cur_distance, new_distance;
3251
3252 if (cpa == NULL) {
3253 atomic_inc(&pa->pa_count);
3254 return pa;
3255 }
3256 cur_distance = abs(goal_block - cpa->pa_pstart);
3257 new_distance = abs(goal_block - pa->pa_pstart);
3258
Coly Li5a54b2f2011-02-24 14:10:05 -05003259 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003260 return cpa;
3261
3262 /* drop the previous reference */
3263 atomic_dec(&cpa->pa_count);
3264 atomic_inc(&pa->pa_count);
3265 return pa;
3266}
3267
3268/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003269 * search goal blocks in preallocated space
3270 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003271static noinline_for_stack int
3272ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003273{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003274 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003275 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003276 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3277 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003278 struct ext4_prealloc_space *pa, *cpa = NULL;
3279 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003280
3281 /* only data can be preallocated */
3282 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3283 return 0;
3284
3285 /* first, try per-file preallocation */
3286 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003287 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003288
3289 /* all fields in this condition don't change,
3290 * so we can skip locking for them */
3291 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003292 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3293 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003294 continue;
3295
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003296 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003297 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003298 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3299 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003300 continue;
3301
Alex Tomasc9de5602008-01-29 00:19:52 -05003302 /* found preallocated blocks, use them */
3303 spin_lock(&pa->pa_lock);
3304 if (pa->pa_deleted == 0 && pa->pa_free) {
3305 atomic_inc(&pa->pa_count);
3306 ext4_mb_use_inode_pa(ac, pa);
3307 spin_unlock(&pa->pa_lock);
3308 ac->ac_criteria = 10;
3309 rcu_read_unlock();
3310 return 1;
3311 }
3312 spin_unlock(&pa->pa_lock);
3313 }
3314 rcu_read_unlock();
3315
3316 /* can we use group allocation? */
3317 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3318 return 0;
3319
3320 /* inode may have no locality group for some reason */
3321 lg = ac->ac_lg;
3322 if (lg == NULL)
3323 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003324 order = fls(ac->ac_o_ex.fe_len) - 1;
3325 if (order > PREALLOC_TB_SIZE - 1)
3326 /* The max size of hash table is PREALLOC_TB_SIZE */
3327 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003328
Akinobu Mitabda00de2010-03-03 23:53:25 -05003329 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003330 /*
3331 * search for the prealloc space that is having
3332 * minimal distance from the goal block.
3333 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003334 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3335 rcu_read_lock();
3336 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3337 pa_inode_list) {
3338 spin_lock(&pa->pa_lock);
3339 if (pa->pa_deleted == 0 &&
3340 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003341
3342 cpa = ext4_mb_check_group_pa(goal_block,
3343 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003344 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003345 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003346 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003347 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003348 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003349 if (cpa) {
3350 ext4_mb_use_group_pa(ac, cpa);
3351 ac->ac_criteria = 20;
3352 return 1;
3353 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003354 return 0;
3355}
3356
3357/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003358 * the function goes through all block freed in the group
3359 * but not yet committed and marks them used in in-core bitmap.
3360 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003361 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003362 */
3363static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3364 ext4_group_t group)
3365{
3366 struct rb_node *n;
3367 struct ext4_group_info *grp;
3368 struct ext4_free_data *entry;
3369
3370 grp = ext4_get_group_info(sb, group);
3371 n = rb_first(&(grp->bb_free_root));
3372
3373 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003374 entry = rb_entry(n, struct ext4_free_data, efd_node);
3375 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003376 n = rb_next(n);
3377 }
3378 return;
3379}
3380
3381/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003382 * the function goes through all preallocation in this group and marks them
3383 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003384 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003385 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003386static noinline_for_stack
3387void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003388 ext4_group_t group)
3389{
3390 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3391 struct ext4_prealloc_space *pa;
3392 struct list_head *cur;
3393 ext4_group_t groupnr;
3394 ext4_grpblk_t start;
3395 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003396 int len;
3397
3398 /* all form of preallocation discards first load group,
3399 * so the only competing code is preallocation use.
3400 * we don't need any locking here
3401 * notice we do NOT ignore preallocations with pa_deleted
3402 * otherwise we could leave used blocks available for
3403 * allocation in buddy when concurrent ext4_mb_put_pa()
3404 * is dropping preallocation
3405 */
3406 list_for_each(cur, &grp->bb_prealloc_list) {
3407 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3408 spin_lock(&pa->pa_lock);
3409 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3410 &groupnr, &start);
3411 len = pa->pa_len;
3412 spin_unlock(&pa->pa_lock);
3413 if (unlikely(len == 0))
3414 continue;
3415 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003416 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003417 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003418 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003419 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003420}
3421
3422static void ext4_mb_pa_callback(struct rcu_head *head)
3423{
3424 struct ext4_prealloc_space *pa;
3425 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3426 kmem_cache_free(ext4_pspace_cachep, pa);
3427}
3428
3429/*
3430 * drops a reference to preallocated space descriptor
3431 * if this was the last reference and the space is consumed
3432 */
3433static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3434 struct super_block *sb, struct ext4_prealloc_space *pa)
3435{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003436 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003437 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003438
3439 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3440 return;
3441
3442 /* in this short window concurrent discard can set pa_deleted */
3443 spin_lock(&pa->pa_lock);
3444 if (pa->pa_deleted == 1) {
3445 spin_unlock(&pa->pa_lock);
3446 return;
3447 }
3448
3449 pa->pa_deleted = 1;
3450 spin_unlock(&pa->pa_lock);
3451
Eric Sandeend33a1972009-03-16 23:25:40 -04003452 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003453 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003454 * If doing group-based preallocation, pa_pstart may be in the
3455 * next group when pa is used up
3456 */
3457 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003458 grp_blk--;
3459
Lukas Czernerbd862982013-04-03 23:32:34 -04003460 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003461
3462 /*
3463 * possible race:
3464 *
3465 * P1 (buddy init) P2 (regular allocation)
3466 * find block B in PA
3467 * copy on-disk bitmap to buddy
3468 * mark B in on-disk bitmap
3469 * drop PA from group
3470 * mark all PAs in buddy
3471 *
3472 * thus, P1 initializes buddy with B available. to prevent this
3473 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3474 * against that pair
3475 */
3476 ext4_lock_group(sb, grp);
3477 list_del(&pa->pa_group_list);
3478 ext4_unlock_group(sb, grp);
3479
3480 spin_lock(pa->pa_obj_lock);
3481 list_del_rcu(&pa->pa_inode_list);
3482 spin_unlock(pa->pa_obj_lock);
3483
3484 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3485}
3486
3487/*
3488 * creates new preallocated space for given inode
3489 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003490static noinline_for_stack int
3491ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003492{
3493 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003494 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003495 struct ext4_prealloc_space *pa;
3496 struct ext4_group_info *grp;
3497 struct ext4_inode_info *ei;
3498
3499 /* preallocate only when found space is larger then requested */
3500 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3501 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3502 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3503
3504 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3505 if (pa == NULL)
3506 return -ENOMEM;
3507
3508 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3509 int winl;
3510 int wins;
3511 int win;
3512 int offs;
3513
3514 /* we can't allocate as much as normalizer wants.
3515 * so, found space must get proper lstart
3516 * to cover original request */
3517 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3518 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3519
3520 /* we're limited by original request in that
3521 * logical block must be covered any way
3522 * winl is window we can move our chunk within */
3523 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3524
3525 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003526 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003527
3528 /* the smallest one defines real window */
3529 win = min(winl, wins);
3530
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003531 offs = ac->ac_o_ex.fe_logical %
3532 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003533 if (offs && offs < win)
3534 win = offs;
3535
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003536 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003537 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003538 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3539 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3540 }
3541
3542 /* preallocation can change ac_b_ex, thus we store actually
3543 * allocated blocks for history */
3544 ac->ac_f_ex = ac->ac_b_ex;
3545
3546 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3547 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3548 pa->pa_len = ac->ac_b_ex.fe_len;
3549 pa->pa_free = pa->pa_len;
3550 atomic_set(&pa->pa_count, 1);
3551 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003552 INIT_LIST_HEAD(&pa->pa_inode_list);
3553 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003554 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003555 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003556
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003557 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003558 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003559 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003560
3561 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003562 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003563
3564 ei = EXT4_I(ac->ac_inode);
3565 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3566
3567 pa->pa_obj_lock = &ei->i_prealloc_lock;
3568 pa->pa_inode = ac->ac_inode;
3569
3570 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3571 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3572 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3573
3574 spin_lock(pa->pa_obj_lock);
3575 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3576 spin_unlock(pa->pa_obj_lock);
3577
3578 return 0;
3579}
3580
3581/*
3582 * creates new preallocated space for locality group inodes belongs to
3583 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003584static noinline_for_stack int
3585ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003586{
3587 struct super_block *sb = ac->ac_sb;
3588 struct ext4_locality_group *lg;
3589 struct ext4_prealloc_space *pa;
3590 struct ext4_group_info *grp;
3591
3592 /* preallocate only when found space is larger then requested */
3593 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3594 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3595 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3596
3597 BUG_ON(ext4_pspace_cachep == NULL);
3598 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3599 if (pa == NULL)
3600 return -ENOMEM;
3601
3602 /* preallocation can change ac_b_ex, thus we store actually
3603 * allocated blocks for history */
3604 ac->ac_f_ex = ac->ac_b_ex;
3605
3606 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3607 pa->pa_lstart = pa->pa_pstart;
3608 pa->pa_len = ac->ac_b_ex.fe_len;
3609 pa->pa_free = pa->pa_len;
3610 atomic_set(&pa->pa_count, 1);
3611 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003612 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003613 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003614 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003615 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003616
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003617 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003618 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3619 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003620
3621 ext4_mb_use_group_pa(ac, pa);
3622 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3623
3624 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3625 lg = ac->ac_lg;
3626 BUG_ON(lg == NULL);
3627
3628 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3629 pa->pa_inode = NULL;
3630
3631 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3632 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3633 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3634
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003635 /*
3636 * We will later add the new pa to the right bucket
3637 * after updating the pa_free in ext4_mb_release_context
3638 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003639 return 0;
3640}
3641
3642static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3643{
3644 int err;
3645
3646 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3647 err = ext4_mb_new_group_pa(ac);
3648 else
3649 err = ext4_mb_new_inode_pa(ac);
3650 return err;
3651}
3652
3653/*
3654 * finds all unused blocks in on-disk bitmap, frees them in
3655 * in-core bitmap and buddy.
3656 * @pa must be unlinked from inode and group lists, so that
3657 * nobody else can find/use it.
3658 * the caller MUST hold group/inode locks.
3659 * TODO: optimize the case when there are no in-core structures yet
3660 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003661static noinline_for_stack int
3662ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003663 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003664{
Alex Tomasc9de5602008-01-29 00:19:52 -05003665 struct super_block *sb = e4b->bd_sb;
3666 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003667 unsigned int end;
3668 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003669 ext4_group_t group;
3670 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003671 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003672 int err = 0;
3673 int free = 0;
3674
3675 BUG_ON(pa->pa_deleted == 0);
3676 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003677 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003678 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3679 end = bit + pa->pa_len;
3680
Alex Tomasc9de5602008-01-29 00:19:52 -05003681 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003682 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003683 if (bit >= end)
3684 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003685 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003686 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003687 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3688 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003689 free += next - bit;
3690
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003691 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003692 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3693 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003694 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003695 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3696 bit = next + 1;
3697 }
3698 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003699 ext4_msg(e4b->bd_sb, KERN_CRIT,
3700 "pa %p: logic %lu, phys. %lu, len %lu",
3701 pa, (unsigned long) pa->pa_lstart,
3702 (unsigned long) pa->pa_pstart,
3703 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003704 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003705 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003706 /*
3707 * pa is already deleted so we use the value obtained
3708 * from the bitmap and continue.
3709 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003710 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003711 atomic_add(free, &sbi->s_mb_discarded);
3712
3713 return err;
3714}
3715
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003716static noinline_for_stack int
3717ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003718 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003719{
Alex Tomasc9de5602008-01-29 00:19:52 -05003720 struct super_block *sb = e4b->bd_sb;
3721 ext4_group_t group;
3722 ext4_grpblk_t bit;
3723
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003724 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003725 BUG_ON(pa->pa_deleted == 0);
3726 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3727 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3728 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3729 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003730 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003731
3732 return 0;
3733}
3734
3735/*
3736 * releases all preallocations in given group
3737 *
3738 * first, we need to decide discard policy:
3739 * - when do we discard
3740 * 1) ENOSPC
3741 * - how many do we discard
3742 * 1) how many requested
3743 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003744static noinline_for_stack int
3745ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003746 ext4_group_t group, int needed)
3747{
3748 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3749 struct buffer_head *bitmap_bh = NULL;
3750 struct ext4_prealloc_space *pa, *tmp;
3751 struct list_head list;
3752 struct ext4_buddy e4b;
3753 int err;
3754 int busy = 0;
3755 int free = 0;
3756
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003757 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003758
3759 if (list_empty(&grp->bb_prealloc_list))
3760 return 0;
3761
Theodore Ts'o574ca172008-07-11 19:27:31 -04003762 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003763 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003764 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003765 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003766 }
3767
3768 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003769 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003770 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003771 put_bh(bitmap_bh);
3772 return 0;
3773 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003774
3775 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003776 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003777
Alex Tomasc9de5602008-01-29 00:19:52 -05003778 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003779repeat:
3780 ext4_lock_group(sb, group);
3781 list_for_each_entry_safe(pa, tmp,
3782 &grp->bb_prealloc_list, pa_group_list) {
3783 spin_lock(&pa->pa_lock);
3784 if (atomic_read(&pa->pa_count)) {
3785 spin_unlock(&pa->pa_lock);
3786 busy = 1;
3787 continue;
3788 }
3789 if (pa->pa_deleted) {
3790 spin_unlock(&pa->pa_lock);
3791 continue;
3792 }
3793
3794 /* seems this one can be freed ... */
3795 pa->pa_deleted = 1;
3796
3797 /* we can trust pa_free ... */
3798 free += pa->pa_free;
3799
3800 spin_unlock(&pa->pa_lock);
3801
3802 list_del(&pa->pa_group_list);
3803 list_add(&pa->u.pa_tmp_list, &list);
3804 }
3805
3806 /* if we still need more blocks and some PAs were used, try again */
3807 if (free < needed && busy) {
3808 busy = 0;
3809 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003810 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003811 goto repeat;
3812 }
3813
3814 /* found anything to free? */
3815 if (list_empty(&list)) {
3816 BUG_ON(free != 0);
3817 goto out;
3818 }
3819
3820 /* now free all selected PAs */
3821 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3822
3823 /* remove from object (inode or locality group) */
3824 spin_lock(pa->pa_obj_lock);
3825 list_del_rcu(&pa->pa_inode_list);
3826 spin_unlock(pa->pa_obj_lock);
3827
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003828 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003829 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003830 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003831 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003832
3833 list_del(&pa->u.pa_tmp_list);
3834 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3835 }
3836
3837out:
3838 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003839 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003840 put_bh(bitmap_bh);
3841 return free;
3842}
3843
3844/*
3845 * releases all non-used preallocated blocks for given inode
3846 *
3847 * It's important to discard preallocations under i_data_sem
3848 * We don't want another block to be served from the prealloc
3849 * space when we are discarding the inode prealloc space.
3850 *
3851 * FIXME!! Make sure it is valid at all the call sites
3852 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003853void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003854{
3855 struct ext4_inode_info *ei = EXT4_I(inode);
3856 struct super_block *sb = inode->i_sb;
3857 struct buffer_head *bitmap_bh = NULL;
3858 struct ext4_prealloc_space *pa, *tmp;
3859 ext4_group_t group = 0;
3860 struct list_head list;
3861 struct ext4_buddy e4b;
3862 int err;
3863
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003864 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003865 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3866 return;
3867 }
3868
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003869 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003870 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003871
3872 INIT_LIST_HEAD(&list);
3873
3874repeat:
3875 /* first, collect all pa's in the inode */
3876 spin_lock(&ei->i_prealloc_lock);
3877 while (!list_empty(&ei->i_prealloc_list)) {
3878 pa = list_entry(ei->i_prealloc_list.next,
3879 struct ext4_prealloc_space, pa_inode_list);
3880 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3881 spin_lock(&pa->pa_lock);
3882 if (atomic_read(&pa->pa_count)) {
3883 /* this shouldn't happen often - nobody should
3884 * use preallocation while we're discarding it */
3885 spin_unlock(&pa->pa_lock);
3886 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003887 ext4_msg(sb, KERN_ERR,
3888 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003889 WARN_ON(1);
3890 schedule_timeout_uninterruptible(HZ);
3891 goto repeat;
3892
3893 }
3894 if (pa->pa_deleted == 0) {
3895 pa->pa_deleted = 1;
3896 spin_unlock(&pa->pa_lock);
3897 list_del_rcu(&pa->pa_inode_list);
3898 list_add(&pa->u.pa_tmp_list, &list);
3899 continue;
3900 }
3901
3902 /* someone is deleting pa right now */
3903 spin_unlock(&pa->pa_lock);
3904 spin_unlock(&ei->i_prealloc_lock);
3905
3906 /* we have to wait here because pa_deleted
3907 * doesn't mean pa is already unlinked from
3908 * the list. as we might be called from
3909 * ->clear_inode() the inode will get freed
3910 * and concurrent thread which is unlinking
3911 * pa from inode's list may access already
3912 * freed memory, bad-bad-bad */
3913
3914 /* XXX: if this happens too often, we can
3915 * add a flag to force wait only in case
3916 * of ->clear_inode(), but not in case of
3917 * regular truncate */
3918 schedule_timeout_uninterruptible(HZ);
3919 goto repeat;
3920 }
3921 spin_unlock(&ei->i_prealloc_lock);
3922
3923 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003924 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04003925 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05003926
3927 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003928 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003929 ext4_error(sb, "Error loading buddy information for %u",
3930 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003931 continue;
3932 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003933
Theodore Ts'o574ca172008-07-11 19:27:31 -04003934 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003935 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003936 ext4_error(sb, "Error reading block bitmap for %u",
3937 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003938 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003939 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05003940 }
3941
3942 ext4_lock_group(sb, group);
3943 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003944 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003945 ext4_unlock_group(sb, group);
3946
Jing Zhange39e07f2010-05-14 00:00:00 -04003947 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003948 put_bh(bitmap_bh);
3949
3950 list_del(&pa->u.pa_tmp_list);
3951 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3952 }
3953}
3954
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003955#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05003956static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3957{
3958 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04003959 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003960
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05003961 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05003962 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04003963 return;
3964
Joe Perches7f6a11e2012-03-19 23:09:43 -04003965 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003966 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04003967 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003968 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04003969 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003970 "goal %lu/%lu/%lu@%lu, "
3971 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05003972 (unsigned long)ac->ac_o_ex.fe_group,
3973 (unsigned long)ac->ac_o_ex.fe_start,
3974 (unsigned long)ac->ac_o_ex.fe_len,
3975 (unsigned long)ac->ac_o_ex.fe_logical,
3976 (unsigned long)ac->ac_g_ex.fe_group,
3977 (unsigned long)ac->ac_g_ex.fe_start,
3978 (unsigned long)ac->ac_g_ex.fe_len,
3979 (unsigned long)ac->ac_g_ex.fe_logical,
3980 (unsigned long)ac->ac_b_ex.fe_group,
3981 (unsigned long)ac->ac_b_ex.fe_start,
3982 (unsigned long)ac->ac_b_ex.fe_len,
3983 (unsigned long)ac->ac_b_ex.fe_logical,
3984 (int)ac->ac_criteria);
Joe Perches7f6a11e2012-03-19 23:09:43 -04003985 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003986 ac->ac_ex_scanned, ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04003987 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04003988 ngroups = ext4_get_groups_count(sb);
3989 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003990 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3991 struct ext4_prealloc_space *pa;
3992 ext4_grpblk_t start;
3993 struct list_head *cur;
3994 ext4_lock_group(sb, i);
3995 list_for_each(cur, &grp->bb_prealloc_list) {
3996 pa = list_entry(cur, struct ext4_prealloc_space,
3997 pa_group_list);
3998 spin_lock(&pa->pa_lock);
3999 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4000 NULL, &start);
4001 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004002 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4003 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004004 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004005 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004006
4007 if (grp->bb_free == 0)
4008 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004009 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004010 i, grp->bb_free, grp->bb_fragments);
4011 }
4012 printk(KERN_ERR "\n");
4013}
4014#else
4015static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4016{
4017 return;
4018}
4019#endif
4020
4021/*
4022 * We use locality group preallocation for small size file. The size of the
4023 * file is determined by the current size or the resulting size after
4024 * allocation which ever is larger
4025 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004026 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004027 */
4028static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4029{
4030 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4031 int bsbits = ac->ac_sb->s_blocksize_bits;
4032 loff_t size, isize;
4033
4034 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4035 return;
4036
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004037 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4038 return;
4039
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004040 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004041 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4042 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004043
Theodore Ts'o50797482009-09-18 13:34:02 -04004044 if ((size == isize) &&
4045 !ext4_fs_is_busy(sbi) &&
4046 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4047 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4048 return;
4049 }
4050
Robin Dongebbe0272011-10-26 05:14:27 -04004051 if (sbi->s_mb_group_prealloc <= 0) {
4052 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4053 return;
4054 }
4055
Alex Tomasc9de5602008-01-29 00:19:52 -05004056 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004057 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004058 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004059 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004060 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004061 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004062
4063 BUG_ON(ac->ac_lg != NULL);
4064 /*
4065 * locality group prealloc space are per cpu. The reason for having
4066 * per cpu locality group is to reduce the contention between block
4067 * request from multiple CPUs.
4068 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09004069 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004070
4071 /* we're going to use group allocation */
4072 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4073
4074 /* serialize all allocations in the group */
4075 mutex_lock(&ac->ac_lg->lg_mutex);
4076}
4077
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004078static noinline_for_stack int
4079ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004080 struct ext4_allocation_request *ar)
4081{
4082 struct super_block *sb = ar->inode->i_sb;
4083 struct ext4_sb_info *sbi = EXT4_SB(sb);
4084 struct ext4_super_block *es = sbi->s_es;
4085 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004086 unsigned int len;
4087 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004088 ext4_grpblk_t block;
4089
4090 /* we can't allocate > group size */
4091 len = ar->len;
4092
4093 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004094 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4095 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004096
4097 /* start searching from the goal */
4098 goal = ar->goal;
4099 if (goal < le32_to_cpu(es->s_first_data_block) ||
4100 goal >= ext4_blocks_count(es))
4101 goal = le32_to_cpu(es->s_first_data_block);
4102 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4103
4104 /* set up allocation goals */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004105 ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
Alex Tomasc9de5602008-01-29 00:19:52 -05004106 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004107 ac->ac_sb = sb;
4108 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004109 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004110 ac->ac_o_ex.fe_group = group;
4111 ac->ac_o_ex.fe_start = block;
4112 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004113 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004114 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004115
4116 /* we have to define context: we'll we work with a file or
4117 * locality group. this is a policy, actually */
4118 ext4_mb_group_or_file(ac);
4119
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004120 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004121 "left: %u/%u, right %u/%u to %swritable\n",
4122 (unsigned) ar->len, (unsigned) ar->logical,
4123 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4124 (unsigned) ar->lleft, (unsigned) ar->pleft,
4125 (unsigned) ar->lright, (unsigned) ar->pright,
4126 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4127 return 0;
4128
4129}
4130
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004131static noinline_for_stack void
4132ext4_mb_discard_lg_preallocations(struct super_block *sb,
4133 struct ext4_locality_group *lg,
4134 int order, int total_entries)
4135{
4136 ext4_group_t group = 0;
4137 struct ext4_buddy e4b;
4138 struct list_head discard_list;
4139 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004140
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004141 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004142
4143 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004144
4145 spin_lock(&lg->lg_prealloc_lock);
4146 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4147 pa_inode_list) {
4148 spin_lock(&pa->pa_lock);
4149 if (atomic_read(&pa->pa_count)) {
4150 /*
4151 * This is the pa that we just used
4152 * for block allocation. So don't
4153 * free that
4154 */
4155 spin_unlock(&pa->pa_lock);
4156 continue;
4157 }
4158 if (pa->pa_deleted) {
4159 spin_unlock(&pa->pa_lock);
4160 continue;
4161 }
4162 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004163 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004164
4165 /* seems this one can be freed ... */
4166 pa->pa_deleted = 1;
4167 spin_unlock(&pa->pa_lock);
4168
4169 list_del_rcu(&pa->pa_inode_list);
4170 list_add(&pa->u.pa_tmp_list, &discard_list);
4171
4172 total_entries--;
4173 if (total_entries <= 5) {
4174 /*
4175 * we want to keep only 5 entries
4176 * allowing it to grow to 8. This
4177 * mak sure we don't call discard
4178 * soon for this list.
4179 */
4180 break;
4181 }
4182 }
4183 spin_unlock(&lg->lg_prealloc_lock);
4184
4185 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4186
Lukas Czernerbd862982013-04-03 23:32:34 -04004187 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004188 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004189 ext4_error(sb, "Error loading buddy information for %u",
4190 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004191 continue;
4192 }
4193 ext4_lock_group(sb, group);
4194 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004195 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004196 ext4_unlock_group(sb, group);
4197
Jing Zhange39e07f2010-05-14 00:00:00 -04004198 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004199 list_del(&pa->u.pa_tmp_list);
4200 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4201 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004202}
4203
4204/*
4205 * We have incremented pa_count. So it cannot be freed at this
4206 * point. Also we hold lg_mutex. So no parallel allocation is
4207 * possible from this lg. That means pa_free cannot be updated.
4208 *
4209 * A parallel ext4_mb_discard_group_preallocations is possible.
4210 * which can cause the lg_prealloc_list to be updated.
4211 */
4212
4213static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4214{
4215 int order, added = 0, lg_prealloc_count = 1;
4216 struct super_block *sb = ac->ac_sb;
4217 struct ext4_locality_group *lg = ac->ac_lg;
4218 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4219
4220 order = fls(pa->pa_free) - 1;
4221 if (order > PREALLOC_TB_SIZE - 1)
4222 /* The max size of hash table is PREALLOC_TB_SIZE */
4223 order = PREALLOC_TB_SIZE - 1;
4224 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004225 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004226 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4227 pa_inode_list) {
4228 spin_lock(&tmp_pa->pa_lock);
4229 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004230 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004231 continue;
4232 }
4233 if (!added && pa->pa_free < tmp_pa->pa_free) {
4234 /* Add to the tail of the previous entry */
4235 list_add_tail_rcu(&pa->pa_inode_list,
4236 &tmp_pa->pa_inode_list);
4237 added = 1;
4238 /*
4239 * we want to count the total
4240 * number of entries in the list
4241 */
4242 }
4243 spin_unlock(&tmp_pa->pa_lock);
4244 lg_prealloc_count++;
4245 }
4246 if (!added)
4247 list_add_tail_rcu(&pa->pa_inode_list,
4248 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004249 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004250
4251 /* Now trim the list to be not more than 8 elements */
4252 if (lg_prealloc_count > 8) {
4253 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004254 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004255 return;
4256 }
4257 return ;
4258}
4259
Alex Tomasc9de5602008-01-29 00:19:52 -05004260/*
4261 * release all resource we used in allocation
4262 */
4263static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4264{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004265 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004266 struct ext4_prealloc_space *pa = ac->ac_pa;
4267 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004268 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004269 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004270 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004271 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4272 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004273 pa->pa_free -= ac->ac_b_ex.fe_len;
4274 pa->pa_len -= ac->ac_b_ex.fe_len;
4275 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004276 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004277 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004278 if (pa) {
4279 /*
4280 * We want to add the pa to the right bucket.
4281 * Remove it from the list and while adding
4282 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004283 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004284 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004285 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004286 spin_lock(pa->pa_obj_lock);
4287 list_del_rcu(&pa->pa_inode_list);
4288 spin_unlock(pa->pa_obj_lock);
4289 ext4_mb_add_n_trim(ac);
4290 }
4291 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4292 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004293 if (ac->ac_bitmap_page)
4294 page_cache_release(ac->ac_bitmap_page);
4295 if (ac->ac_buddy_page)
4296 page_cache_release(ac->ac_buddy_page);
4297 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4298 mutex_unlock(&ac->ac_lg->lg_mutex);
4299 ext4_mb_collect_stats(ac);
4300 return 0;
4301}
4302
4303static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4304{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004305 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004306 int ret;
4307 int freed = 0;
4308
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004309 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004310 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004311 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4312 freed += ret;
4313 needed -= ret;
4314 }
4315
4316 return freed;
4317}
4318
4319/*
4320 * Main entry point into mballoc to allocate blocks
4321 * it tries to use preallocation first, then falls back
4322 * to usual allocation
4323 */
4324ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004325 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004326{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004327 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004328 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004329 struct ext4_sb_info *sbi;
4330 struct super_block *sb;
4331 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004332 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004333 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004334
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004335 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004336 sb = ar->inode->i_sb;
4337 sbi = EXT4_SB(sb);
4338
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004339 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004340
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004341 /* Allow to use superuser reservation for quota file */
4342 if (IS_NOQUOTA(ar->inode))
4343 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4344
Mingming Cao60e58e02009-01-22 18:13:05 +01004345 /*
4346 * For delayed allocation, we could skip the ENOSPC and
4347 * EDQUOT check, as blocks and quotas have been already
4348 * reserved when data being copied into pagecache.
4349 */
Theodore Ts'of2321092011-01-10 12:12:36 -05004350 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
Mingming Cao60e58e02009-01-22 18:13:05 +01004351 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4352 else {
4353 /* Without delayed allocation we need to verify
4354 * there is enough free blocks to do block allocation
4355 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004356 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004357 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004358 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004359
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004360 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004361 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004362 ar->len = ar->len >> 1;
4363 }
4364 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004365 *errp = -ENOSPC;
4366 return 0;
4367 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004368 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004369 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004370 dquot_alloc_block_nofail(ar->inode,
4371 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004372 } else {
4373 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004374 dquot_alloc_block(ar->inode,
4375 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004376
4377 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4378 ar->len--;
4379 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004380 }
4381 inquota = ar->len;
4382 if (ar->len == 0) {
4383 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004384 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004385 }
Mingming Caod2a17632008-07-14 17:52:37 -04004386 }
Mingming Caod2a17632008-07-14 17:52:37 -04004387
Wei Yongjun85556c92012-09-26 20:43:37 -04004388 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004389 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004390 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004391 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004392 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004393 }
4394
Eric Sandeen256bdb42008-02-10 01:13:33 -05004395 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004396 if (*errp) {
4397 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004398 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004399 }
4400
Eric Sandeen256bdb42008-02-10 01:13:33 -05004401 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4402 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004403 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4404 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004405repeat:
4406 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004407 *errp = ext4_mb_regular_allocator(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004408 if (*errp) {
4409 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004410 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004411 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004412
4413 /* as we've just preallocated more space than
4414 * user requested orinally, we store allocated
4415 * space in a special descriptor */
Eric Sandeen256bdb42008-02-10 01:13:33 -05004416 if (ac->ac_status == AC_STATUS_FOUND &&
4417 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4418 ext4_mb_new_preallocation(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004419 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004420 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004421 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004422 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004423 /*
4424 * drop the reference that we took
4425 * in ext4_mb_use_best_found
4426 */
4427 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004428 ac->ac_b_ex.fe_group = 0;
4429 ac->ac_b_ex.fe_start = 0;
4430 ac->ac_b_ex.fe_len = 0;
4431 ac->ac_status = AC_STATUS_CONTINUE;
4432 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004433 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004434 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004435 goto errout;
4436 } else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004437 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4438 ar->len = ac->ac_b_ex.fe_len;
4439 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004440 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004441 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004442 if (freed)
4443 goto repeat;
4444 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004445 }
4446
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004447errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004448 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004449 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004450 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004451 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004452 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004453 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004454out:
4455 if (ac)
4456 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004457 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004458 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004459 if (!ar->len) {
Theodore Ts'of2321092011-01-10 12:12:36 -05004460 if (!ext4_test_inode_state(ar->inode,
4461 EXT4_STATE_DELALLOC_RESERVED))
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004462 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004463 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004464 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004465 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004466
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004467 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004468
Alex Tomasc9de5602008-01-29 00:19:52 -05004469 return block;
4470}
Alex Tomasc9de5602008-01-29 00:19:52 -05004471
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004472/*
4473 * We can merge two free data extents only if the physical blocks
4474 * are contiguous, AND the extents were freed by the same transaction,
4475 * AND the blocks are associated with the same group.
4476 */
4477static int can_merge(struct ext4_free_data *entry1,
4478 struct ext4_free_data *entry2)
4479{
Bobi Jam18aadd42012-02-20 17:53:02 -05004480 if ((entry1->efd_tid == entry2->efd_tid) &&
4481 (entry1->efd_group == entry2->efd_group) &&
4482 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004483 return 1;
4484 return 0;
4485}
4486
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004487static noinline_for_stack int
4488ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004489 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004490{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004491 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004492 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004493 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004494 struct ext4_group_info *db = e4b->bd_info;
4495 struct super_block *sb = e4b->bd_sb;
4496 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004497 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4498 struct rb_node *parent = NULL, *new_node;
4499
Frank Mayhar03901312009-01-07 00:06:22 -05004500 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004501 BUG_ON(e4b->bd_bitmap_page == NULL);
4502 BUG_ON(e4b->bd_buddy_page == NULL);
4503
Bobi Jam18aadd42012-02-20 17:53:02 -05004504 new_node = &new_entry->efd_node;
4505 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004506
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004507 if (!*n) {
4508 /* first free block exent. We need to
4509 protect buddy cache from being freed,
4510 * otherwise we'll refresh it from
4511 * on-disk bitmap and lose not-yet-available
4512 * blocks */
4513 page_cache_get(e4b->bd_buddy_page);
4514 page_cache_get(e4b->bd_bitmap_page);
4515 }
4516 while (*n) {
4517 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004518 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4519 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004520 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004521 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004522 n = &(*n)->rb_right;
4523 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004524 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004525 ext4_group_first_block_no(sb, group) +
4526 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004527 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004528 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004529 }
4530 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004531
4532 rb_link_node(new_node, parent, n);
4533 rb_insert_color(new_node, &db->bb_free_root);
4534
4535 /* Now try to see the extent can be merged to left and right */
4536 node = rb_prev(new_node);
4537 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004538 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004539 if (can_merge(entry, new_entry) &&
4540 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004541 new_entry->efd_start_cluster = entry->efd_start_cluster;
4542 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004543 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004544 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004545 }
4546 }
4547
4548 node = rb_next(new_node);
4549 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004550 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004551 if (can_merge(new_entry, entry) &&
4552 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004553 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004554 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004555 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004556 }
4557 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004558 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004559 ext4_journal_callback_add(handle, ext4_free_data_callback,
4560 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004561 return 0;
4562}
4563
Theodore Ts'o44338712009-11-22 07:44:56 -05004564/**
4565 * ext4_free_blocks() -- Free given blocks and update quota
4566 * @handle: handle for this transaction
4567 * @inode: inode
4568 * @block: start physical block to free
4569 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004570 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004571 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004572void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004573 struct buffer_head *bh, ext4_fsblk_t block,
4574 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004575{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004576 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004577 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004578 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004579 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004580 ext4_grpblk_t bit;
4581 struct buffer_head *gd_bh;
4582 ext4_group_t block_group;
4583 struct ext4_sb_info *sbi;
4584 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004585 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004586 int err = 0;
4587 int ret;
4588
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004589 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004590 if (bh) {
4591 if (block)
4592 BUG_ON(block != bh->b_blocknr);
4593 else
4594 block = bh->b_blocknr;
4595 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004596
Alex Tomasc9de5602008-01-29 00:19:52 -05004597 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004598 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4599 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004600 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004601 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004602 goto error_return;
4603 }
4604
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004605 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004606 trace_ext4_free_blocks(inode, block, count, flags);
4607
4608 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4609 struct buffer_head *tbh = bh;
4610 int i;
4611
4612 BUG_ON(bh && (count > 1));
4613
4614 for (i = 0; i < count; i++) {
4615 if (!bh)
4616 tbh = sb_find_get_block(inode->i_sb,
4617 block + i);
Namhyung Kim87783692010-10-27 21:30:11 -04004618 if (unlikely(!tbh))
4619 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004620 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004621 inode, tbh, block + i);
4622 }
4623 }
4624
Theodore Ts'o60e66792010-05-17 07:00:00 -04004625 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004626 * We need to make sure we don't reuse the freed block until
4627 * after the transaction is committed, which we can do by
4628 * treating the block as metadata, below. We make an
4629 * exception if the inode is to be written in writeback mode
4630 * since writeback mode has weak data consistency guarantees.
4631 */
4632 if (!ext4_should_writeback_data(inode))
4633 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004634
Theodore Ts'o84130192011-09-09 18:50:51 -04004635 /*
4636 * If the extent to be freed does not begin on a cluster
4637 * boundary, we need to deal with partial clusters at the
4638 * beginning and end of the extent. Normally we will free
4639 * blocks at the beginning or the end unless we are explicitly
4640 * requested to avoid doing so.
4641 */
4642 overflow = block & (sbi->s_cluster_ratio - 1);
4643 if (overflow) {
4644 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4645 overflow = sbi->s_cluster_ratio - overflow;
4646 block += overflow;
4647 if (count > overflow)
4648 count -= overflow;
4649 else
4650 return;
4651 } else {
4652 block -= overflow;
4653 count += overflow;
4654 }
4655 }
4656 overflow = count & (sbi->s_cluster_ratio - 1);
4657 if (overflow) {
4658 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4659 if (count > overflow)
4660 count -= overflow;
4661 else
4662 return;
4663 } else
4664 count += sbi->s_cluster_ratio - overflow;
4665 }
4666
Alex Tomasc9de5602008-01-29 00:19:52 -05004667do_more:
4668 overflow = 0;
4669 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4670
4671 /*
4672 * Check to see if we are freeing blocks across a group
4673 * boundary.
4674 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004675 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4676 overflow = EXT4_C2B(sbi, bit) + count -
4677 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004678 count -= overflow;
4679 }
Lukas Czerner810da242013-03-02 17:18:58 -05004680 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004681 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004682 if (!bitmap_bh) {
4683 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004684 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004685 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004686 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004687 if (!gdp) {
4688 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004689 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004690 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004691
4692 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4693 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4694 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004695 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004696 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004697 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004698
Eric Sandeen12062dd2010-02-15 14:19:27 -05004699 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004700 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004701 /* err = 0. ext4_std_error should be a no op */
4702 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004703 }
4704
4705 BUFFER_TRACE(bitmap_bh, "getting write access");
4706 err = ext4_journal_get_write_access(handle, bitmap_bh);
4707 if (err)
4708 goto error_return;
4709
4710 /*
4711 * We are about to modify some metadata. Call the journal APIs
4712 * to unshare ->b_data if a currently-committing transaction is
4713 * using it
4714 */
4715 BUFFER_TRACE(gd_bh, "get_write_access");
4716 err = ext4_journal_get_write_access(handle, gd_bh);
4717 if (err)
4718 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004719#ifdef AGGRESSIVE_CHECK
4720 {
4721 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004722 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004723 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4724 }
4725#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004726 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004727
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004728 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4729 if (err)
4730 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004731
4732 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004733 struct ext4_free_data *new_entry;
4734 /*
4735 * blocks being freed are metadata. these blocks shouldn't
4736 * be used until this transaction is committed
4737 */
Bobi Jam18aadd42012-02-20 17:53:02 -05004738 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004739 if (!new_entry) {
Salman Qazi02b78312012-05-31 23:51:27 -04004740 ext4_mb_unload_buddy(&e4b);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004741 err = -ENOMEM;
4742 goto error_return;
4743 }
Bobi Jam18aadd42012-02-20 17:53:02 -05004744 new_entry->efd_start_cluster = bit;
4745 new_entry->efd_group = block_group;
4746 new_entry->efd_count = count_clusters;
4747 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004748
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004749 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004750 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004751 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004752 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004753 /* need to update group_info->bb_free and bitmap
4754 * with group lock held. generate_buddy look at
4755 * them with group lock_held
4756 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004757 if (test_opt(sb, DISCARD)) {
4758 err = ext4_issue_discard(sb, block_group, bit, count);
4759 if (err && err != -EOPNOTSUPP)
4760 ext4_msg(sb, KERN_WARNING, "discard request in"
4761 " group:%d block:%d count:%lu failed"
4762 " with %d", block_group, bit, count,
4763 err);
4764 }
4765
4766
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004767 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004768 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4769 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004770 }
4771
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004772 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4773 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004774 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004775 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004776 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04004777 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004778
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004779 if (sbi->s_log_groups_per_flex) {
4780 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004781 atomic64_add(count_clusters,
4782 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004783 }
4784
Jing Zhange39e07f2010-05-14 00:00:00 -04004785 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05004786
Aditya Kali7b415bf2011-09-09 19:04:51 -04004787 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4788 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4789
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004790 /* We dirtied the bitmap block */
4791 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4792 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4793
Alex Tomasc9de5602008-01-29 00:19:52 -05004794 /* And the group descriptor block */
4795 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004796 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004797 if (!err)
4798 err = ret;
4799
4800 if (overflow && !err) {
4801 block += count;
4802 count = overflow;
4803 put_bh(bitmap_bh);
4804 goto do_more;
4805 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004806error_return:
4807 brelse(bitmap_bh);
4808 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004809 return;
4810}
Lukas Czerner7360d172010-10-27 21:30:12 -04004811
4812/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004813 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004814 * @handle: handle to this transaction
4815 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004816 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004817 * @count: number of blocks to free
4818 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004819 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004820 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004821int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004822 ext4_fsblk_t block, unsigned long count)
4823{
4824 struct buffer_head *bitmap_bh = NULL;
4825 struct buffer_head *gd_bh;
4826 ext4_group_t block_group;
4827 ext4_grpblk_t bit;
4828 unsigned int i;
4829 struct ext4_group_desc *desc;
4830 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004831 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004832 int err = 0, ret, blk_free_count;
4833 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004834
4835 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4836
Yongqiang Yang4740b832011-07-26 21:51:08 -04004837 if (count == 0)
4838 return 0;
4839
Amir Goldstein2846e822011-05-09 10:46:41 -04004840 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004841 /*
4842 * Check to see if we are freeing blocks across a group
4843 * boundary.
4844 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004845 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4846 ext4_warning(sb, "too much blocks added to group %u\n",
4847 block_group);
4848 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004849 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004850 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004851
Amir Goldstein2846e822011-05-09 10:46:41 -04004852 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004853 if (!bitmap_bh) {
4854 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004855 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004856 }
4857
Amir Goldstein2846e822011-05-09 10:46:41 -04004858 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004859 if (!desc) {
4860 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004861 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004862 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004863
4864 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4865 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4866 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4867 in_range(block + count - 1, ext4_inode_table(sb, desc),
4868 sbi->s_itb_per_group)) {
4869 ext4_error(sb, "Adding blocks in system zones - "
4870 "Block = %llu, count = %lu",
4871 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004872 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004873 goto error_return;
4874 }
4875
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004876 BUFFER_TRACE(bitmap_bh, "getting write access");
4877 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004878 if (err)
4879 goto error_return;
4880
4881 /*
4882 * We are about to modify some metadata. Call the journal APIs
4883 * to unshare ->b_data if a currently-committing transaction is
4884 * using it
4885 */
4886 BUFFER_TRACE(gd_bh, "get_write_access");
4887 err = ext4_journal_get_write_access(handle, gd_bh);
4888 if (err)
4889 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004890
Amir Goldstein2846e822011-05-09 10:46:41 -04004891 for (i = 0, blocks_freed = 0; i < count; i++) {
4892 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004893 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004894 ext4_error(sb, "bit already cleared for block %llu",
4895 (ext4_fsblk_t)(block + i));
4896 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4897 } else {
4898 blocks_freed++;
4899 }
4900 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004901
4902 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4903 if (err)
4904 goto error_return;
4905
4906 /*
4907 * need to update group_info->bb_free and bitmap
4908 * with group lock held. generate_buddy look at
4909 * them with group lock_held
4910 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004911 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004912 mb_clear_bits(bitmap_bh->b_data, bit, count);
4913 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004914 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4915 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04004916 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004917 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04004918 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04004919 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05004920 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04004921
4922 if (sbi->s_log_groups_per_flex) {
4923 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004924 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4925 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04004926 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004927
4928 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04004929
4930 /* We dirtied the bitmap block */
4931 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4932 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4933
4934 /* And the group descriptor block */
4935 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4936 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4937 if (!err)
4938 err = ret;
4939
4940error_return:
4941 brelse(bitmap_bh);
4942 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004943 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04004944}
4945
4946/**
Lukas Czerner7360d172010-10-27 21:30:12 -04004947 * ext4_trim_extent -- function to TRIM one single free extent in the group
4948 * @sb: super block for the file system
4949 * @start: starting block of the free extent in the alloc. group
4950 * @count: number of blocks to TRIM
4951 * @group: alloc. group we are working with
4952 * @e4b: ext4 buddy for the group
4953 *
4954 * Trim "count" blocks starting at "start" in the "group". To assure that no
4955 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4956 * be called with under the group lock.
4957 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004958static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04004959 ext4_group_t group, struct ext4_buddy *e4b)
Lukas Czerner7360d172010-10-27 21:30:12 -04004960{
4961 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004962 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04004963
Tao Mab3d4c2b2011-07-11 00:01:52 -04004964 trace_ext4_trim_extent(sb, group, start, count);
4965
Lukas Czerner7360d172010-10-27 21:30:12 -04004966 assert_spin_locked(ext4_group_lock_ptr(sb, group));
4967
4968 ex.fe_start = start;
4969 ex.fe_group = group;
4970 ex.fe_len = count;
4971
4972 /*
4973 * Mark blocks used, so no one can reuse them while
4974 * being trimmed.
4975 */
4976 mb_mark_used(e4b, &ex);
4977 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004978 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04004979 ext4_lock_group(sb, group);
4980 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004981 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04004982}
4983
4984/**
4985 * ext4_trim_all_free -- function to trim all free space in alloc. group
4986 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04004987 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04004988 * @start: first group block to examine
4989 * @max: last group block to examine
4990 * @minblocks: minimum extent block count
4991 *
4992 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4993 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4994 * the extent.
4995 *
4996 *
4997 * ext4_trim_all_free walks through group's block bitmap searching for free
4998 * extents. When the free extent is found, mark it as used in group buddy
4999 * bitmap. Then issue a TRIM command on this extent and free the extent in
5000 * the group buddy bitmap. This is done until whole group is scanned.
5001 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005002static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005003ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5004 ext4_grpblk_t start, ext4_grpblk_t max,
5005 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005006{
5007 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005008 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005009 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005010 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005011
Tao Mab3d4c2b2011-07-11 00:01:52 -04005012 trace_ext4_trim_all_free(sb, group, start, max);
5013
Lukas Czerner78944082011-05-24 18:16:27 -04005014 ret = ext4_mb_load_buddy(sb, group, &e4b);
5015 if (ret) {
5016 ext4_error(sb, "Error in loading buddy "
5017 "information for %u", group);
5018 return ret;
5019 }
Lukas Czerner78944082011-05-24 18:16:27 -04005020 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005021
5022 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005023 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5024 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5025 goto out;
5026
Lukas Czerner78944082011-05-24 18:16:27 -04005027 start = (e4b.bd_info->bb_first_free > start) ?
5028 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005029
Lukas Czerner913eed832012-03-21 21:22:22 -04005030 while (start <= max) {
5031 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5032 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005033 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005034 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005035
5036 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005037 ret = ext4_trim_extent(sb, start,
5038 next - start, group, &e4b);
5039 if (ret && ret != -EOPNOTSUPP)
5040 break;
5041 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005042 count += next - start;
5043 }
Tao Ma169ddc32011-07-11 00:00:07 -04005044 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005045 start = next + 1;
5046
5047 if (fatal_signal_pending(current)) {
5048 count = -ERESTARTSYS;
5049 break;
5050 }
5051
5052 if (need_resched()) {
5053 ext4_unlock_group(sb, group);
5054 cond_resched();
5055 ext4_lock_group(sb, group);
5056 }
5057
Tao Ma169ddc32011-07-11 00:00:07 -04005058 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005059 break;
5060 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005061
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005062 if (!ret) {
5063 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005064 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005065 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005066out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005067 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005068 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005069
5070 ext4_debug("trimmed %d blocks in the group %d\n",
5071 count, group);
5072
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005073 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005074}
5075
5076/**
5077 * ext4_trim_fs() -- trim ioctl handle function
5078 * @sb: superblock for filesystem
5079 * @range: fstrim_range structure
5080 *
5081 * start: First Byte to trim
5082 * len: number of Bytes to trim from start
5083 * minlen: minimum extent length in Bytes
5084 * ext4_trim_fs goes through all allocation groups containing Bytes from
5085 * start to start+len. For each such a group ext4_trim_all_free function
5086 * is invoked to trim all free space.
5087 */
5088int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5089{
Lukas Czerner78944082011-05-24 18:16:27 -04005090 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005091 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005092 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005093 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005094 ext4_fsblk_t first_data_blk =
5095 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005096 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005097 int ret = 0;
5098
5099 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005100 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005101 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5102 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005103
Lukas Czerner5de35e82012-10-22 18:01:19 -04005104 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5105 start >= max_blks ||
5106 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005107 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005108 if (end >= max_blks)
5109 end = max_blks - 1;
5110 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005111 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005112 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005113 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005114
Lukas Czerner913eed832012-03-21 21:22:22 -04005115 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005116 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005117 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005118 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005119 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005120
Lukas Czerner913eed832012-03-21 21:22:22 -04005121 /* end now represents the last cluster to discard in this group */
5122 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005123
5124 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005125 grp = ext4_get_group_info(sb, group);
5126 /* We only do this if the grp has never been initialized */
5127 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5128 ret = ext4_mb_init_group(sb, group);
5129 if (ret)
5130 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005131 }
5132
Tao Ma0ba08512011-03-23 15:48:11 -04005133 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005134 * For all the groups except the last one, last cluster will
5135 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5136 * change it for the last group, note that last_cluster is
5137 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005138 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005139 if (group == last_group)
5140 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005141
Lukas Czerner78944082011-05-24 18:16:27 -04005142 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005143 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005144 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005145 if (cnt < 0) {
5146 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005147 break;
5148 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005149 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005150 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005151
5152 /*
5153 * For every group except the first one, we are sure
5154 * that the first cluster to discard will be cluster #0.
5155 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005156 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005157 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005158
Tao Ma3d56b8d2011-07-11 00:03:38 -04005159 if (!ret)
5160 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5161
Tao Ma22f10452011-07-10 23:52:37 -04005162out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005163 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005164 return ret;
5165}