blob: 8b0f9ef517d690ea0ede0958c8aa726aebb86b35 [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);
Namjae Jeone43bb4e2014-06-26 10:11:53 -0400725 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -0400726 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400727 ext4_grpblk_t i = 0;
728 ext4_grpblk_t first;
729 ext4_grpblk_t len;
Alex Tomasc9de5602008-01-29 00:19:52 -0500730 unsigned free = 0;
731 unsigned fragments = 0;
732 unsigned long long period = get_cycles();
733
734 /* initialize buddy from bitmap which is aggregation
735 * of on-disk bitmap and preallocations */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500736 i = mb_find_next_zero_bit(bitmap, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -0500737 grp->bb_first_free = i;
738 while (i < max) {
739 fragments++;
740 first = i;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500741 i = mb_find_next_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500742 len = i - first;
743 free += len;
744 if (len > 1)
745 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
746 else
747 grp->bb_counters[0]++;
748 if (i < max)
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500749 i = mb_find_next_zero_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500750 }
751 grp->bb_fragments = fragments;
752
753 if (free != grp->bb_free) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400754 ext4_grp_locked_error(sb, group, 0, 0,
Theodore Ts'o94d4c062014-07-05 19:15:50 -0400755 "block bitmap and bg descriptor "
756 "inconsistent: %u vs %u free clusters",
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400757 free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500758 /*
Darrick J. Wong163a2032013-08-28 17:35:51 -0400759 * If we intend to continue, we consider group descriptor
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500760 * corrupt and update bb_free using bitmap value
761 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500762 grp->bb_free = free;
Namjae Jeone43bb4e2014-06-26 10:11:53 -0400763 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
764 percpu_counter_sub(&sbi->s_freeclusters_counter,
765 grp->bb_free);
Darrick J. Wong163a2032013-08-28 17:35:51 -0400766 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
Alex Tomasc9de5602008-01-29 00:19:52 -0500767 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400768 mb_set_largest_free_order(sb, grp);
Alex Tomasc9de5602008-01-29 00:19:52 -0500769
770 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
771
772 period = get_cycles() - period;
773 spin_lock(&EXT4_SB(sb)->s_bal_lock);
774 EXT4_SB(sb)->s_mb_buddies_generated++;
775 EXT4_SB(sb)->s_mb_generation_time += period;
776 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
777}
778
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400779static void mb_regenerate_buddy(struct ext4_buddy *e4b)
780{
781 int count;
782 int order = 1;
783 void *buddy;
784
785 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
786 ext4_set_bits(buddy, 0, count);
787 }
788 e4b->bd_info->bb_fragments = 0;
789 memset(e4b->bd_info->bb_counters, 0,
790 sizeof(*e4b->bd_info->bb_counters) *
791 (e4b->bd_sb->s_blocksize_bits + 2));
792
793 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
794 e4b->bd_bitmap, e4b->bd_group);
795}
796
Alex Tomasc9de5602008-01-29 00:19:52 -0500797/* The buddy information is attached the buddy cache inode
798 * for convenience. The information regarding each group
799 * is loaded via ext4_mb_load_buddy. The information involve
800 * block bitmap and buddy information. The information are
801 * stored in the inode as
802 *
803 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500804 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500805 *
806 *
807 * one block each for bitmap and buddy information.
808 * So for each group we take up 2 blocks. A page can
809 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
810 * So it can have information regarding groups_per_page which
811 * is blocks_per_page/2
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400812 *
813 * Locking note: This routine takes the block group lock of all groups
814 * for this page; do not hold this lock when calling this routine!
Alex Tomasc9de5602008-01-29 00:19:52 -0500815 */
816
817static int ext4_mb_init_cache(struct page *page, char *incore)
818{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400819 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500820 int blocksize;
821 int blocks_per_page;
822 int groups_per_page;
823 int err = 0;
824 int i;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500825 ext4_group_t first_group, group;
Alex Tomasc9de5602008-01-29 00:19:52 -0500826 int first_block;
827 struct super_block *sb;
828 struct buffer_head *bhs;
Darrick J. Wongfa77dcf2012-04-29 18:35:10 -0400829 struct buffer_head **bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -0500830 struct inode *inode;
831 char *data;
832 char *bitmap;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400833 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -0500834
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400835 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500836
837 inode = page->mapping->host;
838 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400839 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500840 blocksize = 1 << inode->i_blkbits;
841 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
842
843 groups_per_page = blocks_per_page >> 1;
844 if (groups_per_page == 0)
845 groups_per_page = 1;
846
847 /* allocate buffer_heads to read bitmaps */
848 if (groups_per_page > 1) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500849 i = sizeof(struct buffer_head *) * groups_per_page;
850 bh = kzalloc(i, GFP_NOFS);
Theodore Ts'o813e5722012-02-20 17:52:46 -0500851 if (bh == NULL) {
852 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500853 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500854 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500855 } else
856 bh = &bhs;
857
858 first_group = page->index * blocks_per_page / 2;
859
860 /* read all groups the page covers into the cache */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500861 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
862 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500863 break;
864
Theodore Ts'o813e5722012-02-20 17:52:46 -0500865 grinfo = ext4_get_group_info(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400866 /*
867 * If page is uptodate then we came here after online resize
868 * which added some new uninitialized group info structs, so
869 * we must skip all initialized uptodate buddies on the page,
870 * which may be currently in use by an allocating task.
871 */
872 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
873 bh[i] = NULL;
874 continue;
875 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500876 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
877 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500878 goto out;
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500879 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500880 mb_debug(1, "read bitmap for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500881 }
882
883 /* wait for I/O completion */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500884 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
885 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
886 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -0500887 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500888 }
889 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500890
891 first_block = page->index * blocks_per_page;
892 for (i = 0; i < blocks_per_page; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500893 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400894 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500895 break;
896
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400897 if (!bh[group - first_group])
898 /* skip initialized uptodate buddy */
899 continue;
900
Alex Tomasc9de5602008-01-29 00:19:52 -0500901 /*
902 * data carry information regarding this
903 * particular group in the format specified
904 * above
905 *
906 */
907 data = page_address(page) + (i * blocksize);
908 bitmap = bh[group - first_group]->b_data;
909
910 /*
911 * We place the buddy block and bitmap block
912 * close together
913 */
914 if ((first_block + i) & 1) {
915 /* this is block of buddy */
916 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400917 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500918 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400919 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500920 grinfo = ext4_get_group_info(sb, group);
921 grinfo->bb_fragments = 0;
922 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400923 sizeof(*grinfo->bb_counters) *
924 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500925 /*
926 * incore got set to the group block bitmap below
927 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500928 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400929 /* init the buddy */
930 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500931 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500932 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500933 incore = NULL;
934 } else {
935 /* this is block of bitmap */
936 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400937 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500938 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400939 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500940
941 /* see comments in ext4_mb_put_pa() */
942 ext4_lock_group(sb, group);
943 memcpy(data, bitmap, blocksize);
944
945 /* mark all preallocated blks used in in-core bitmap */
946 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500947 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500948 ext4_unlock_group(sb, group);
949
950 /* set incore so that the buddy information can be
951 * generated using this
952 */
953 incore = data;
954 }
955 }
956 SetPageUptodate(page);
957
958out:
959 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400960 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500961 brelse(bh[i]);
962 if (bh != &bhs)
963 kfree(bh);
964 }
965 return err;
966}
967
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400968/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400969 * Lock the buddy and bitmap pages. This make sure other parallel init_group
970 * on the same buddy page doesn't happen whild holding the buddy page lock.
971 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
972 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400973 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400974static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
975 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400976{
Amir Goldstein2de88072011-05-09 21:48:13 -0400977 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
978 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400979 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400980 struct page *page;
981
982 e4b->bd_buddy_page = NULL;
983 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400984
985 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
986 /*
987 * the buddy cache inode stores the block bitmap
988 * and buddy information in consecutive blocks.
989 * So for each group we need two blocks.
990 */
991 block = group * 2;
992 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400993 poff = block % blocks_per_page;
994 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
995 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -0400996 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -0400997 BUG_ON(page->mapping != inode->i_mapping);
998 e4b->bd_bitmap_page = page;
999 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001000
Amir Goldstein2de88072011-05-09 21:48:13 -04001001 if (blocks_per_page >= 2) {
1002 /* buddy and bitmap are on the same page */
1003 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001004 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001005
1006 block++;
1007 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001008 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1009 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -04001010 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001011 BUG_ON(page->mapping != inode->i_mapping);
1012 e4b->bd_buddy_page = page;
1013 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001014}
1015
Amir Goldstein2de88072011-05-09 21:48:13 -04001016static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001017{
Amir Goldstein2de88072011-05-09 21:48:13 -04001018 if (e4b->bd_bitmap_page) {
1019 unlock_page(e4b->bd_bitmap_page);
1020 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001021 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001022 if (e4b->bd_buddy_page) {
1023 unlock_page(e4b->bd_buddy_page);
1024 page_cache_release(e4b->bd_buddy_page);
1025 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001026}
1027
1028/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001029 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1030 * block group lock of all groups for this page; do not hold the BG lock when
1031 * calling this routine!
1032 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001033static noinline_for_stack
1034int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1035{
1036
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001037 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001038 struct ext4_buddy e4b;
1039 struct page *page;
1040 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001041
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001042 might_sleep();
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001043 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001044 this_grp = ext4_get_group_info(sb, group);
1045 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001046 * This ensures that we don't reinit the buddy cache
1047 * page which map to the group from which we are already
1048 * allocating. If we are looking at the buddy cache we would
1049 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001050 * would have pinned buddy page to page cache.
Mel Gorman2457aec2014-06-04 16:10:31 -07001051 * The call to ext4_mb_get_buddy_page_lock will mark the
1052 * page accessed.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001053 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001054 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1055 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001056 /*
1057 * somebody initialized the group
1058 * return without doing anything
1059 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001060 goto err;
1061 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001062
1063 page = e4b.bd_bitmap_page;
1064 ret = ext4_mb_init_cache(page, NULL);
1065 if (ret)
1066 goto err;
1067 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001068 ret = -EIO;
1069 goto err;
1070 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001071
Amir Goldstein2de88072011-05-09 21:48:13 -04001072 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001073 /*
1074 * If both the bitmap and buddy are in
1075 * the same page we don't need to force
1076 * init the buddy
1077 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001078 ret = 0;
1079 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001080 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001081 /* init buddy cache */
1082 page = e4b.bd_buddy_page;
1083 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1084 if (ret)
1085 goto err;
1086 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001087 ret = -EIO;
1088 goto err;
1089 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001090err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001091 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001092 return ret;
1093}
1094
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001095/*
1096 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1097 * block group lock of all groups for this page; do not hold the BG lock when
1098 * calling this routine!
1099 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001100static noinline_for_stack int
1101ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1102 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001103{
Alex Tomasc9de5602008-01-29 00:19:52 -05001104 int blocks_per_page;
1105 int block;
1106 int pnum;
1107 int poff;
1108 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001109 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001110 struct ext4_group_info *grp;
1111 struct ext4_sb_info *sbi = EXT4_SB(sb);
1112 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001113
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001114 might_sleep();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001115 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001116
1117 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001118 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001119
1120 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001121 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001122 e4b->bd_sb = sb;
1123 e4b->bd_group = group;
1124 e4b->bd_buddy_page = NULL;
1125 e4b->bd_bitmap_page = NULL;
1126
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001127 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001128 /*
1129 * we need full data about the group
1130 * to make a good selection
1131 */
1132 ret = ext4_mb_init_group(sb, group);
1133 if (ret)
1134 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001135 }
1136
Alex Tomasc9de5602008-01-29 00:19:52 -05001137 /*
1138 * the buddy cache inode stores the block bitmap
1139 * and buddy information in consecutive blocks.
1140 * So for each group we need two blocks.
1141 */
1142 block = group * 2;
1143 pnum = block / blocks_per_page;
1144 poff = block % blocks_per_page;
1145
1146 /* we could use find_or_create_page(), but it locks page
1147 * what we'd like to avoid in fast path ... */
Mel Gorman2457aec2014-06-04 16:10:31 -07001148 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001149 if (page == NULL || !PageUptodate(page)) {
1150 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001151 /*
1152 * drop the page reference and try
1153 * to get the page with lock. If we
1154 * are not uptodate that implies
1155 * somebody just created the page but
1156 * is yet to initialize the same. So
1157 * wait for it to initialize.
1158 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001159 page_cache_release(page);
1160 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1161 if (page) {
1162 BUG_ON(page->mapping != inode->i_mapping);
1163 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001164 ret = ext4_mb_init_cache(page, NULL);
1165 if (ret) {
1166 unlock_page(page);
1167 goto err;
1168 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001169 mb_cmp_bitmaps(e4b, page_address(page) +
1170 (poff * sb->s_blocksize));
1171 }
1172 unlock_page(page);
1173 }
1174 }
Younger Liuc57ab392014-04-10 23:03:43 -04001175 if (page == NULL) {
1176 ret = -ENOMEM;
1177 goto err;
1178 }
1179 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001180 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001181 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001182 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001183
1184 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001185 e4b->bd_bitmap_page = page;
1186 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001187
1188 block++;
1189 pnum = block / blocks_per_page;
1190 poff = block % blocks_per_page;
1191
Mel Gorman2457aec2014-06-04 16:10:31 -07001192 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001193 if (page == NULL || !PageUptodate(page)) {
1194 if (page)
1195 page_cache_release(page);
1196 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1197 if (page) {
1198 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001199 if (!PageUptodate(page)) {
1200 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1201 if (ret) {
1202 unlock_page(page);
1203 goto err;
1204 }
1205 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001206 unlock_page(page);
1207 }
1208 }
Younger Liuc57ab392014-04-10 23:03:43 -04001209 if (page == NULL) {
1210 ret = -ENOMEM;
1211 goto err;
1212 }
1213 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001214 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001215 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001216 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001217
1218 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001219 e4b->bd_buddy_page = page;
1220 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001221
1222 BUG_ON(e4b->bd_bitmap_page == NULL);
1223 BUG_ON(e4b->bd_buddy_page == NULL);
1224
1225 return 0;
1226
1227err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001228 if (page)
1229 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001230 if (e4b->bd_bitmap_page)
1231 page_cache_release(e4b->bd_bitmap_page);
1232 if (e4b->bd_buddy_page)
1233 page_cache_release(e4b->bd_buddy_page);
1234 e4b->bd_buddy = NULL;
1235 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001236 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001237}
1238
Jing Zhange39e07f2010-05-14 00:00:00 -04001239static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001240{
1241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1245}
1246
1247
1248static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1249{
1250 int order = 1;
1251 void *bb;
1252
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001253 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001254 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1255
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001256 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001257 while (order <= e4b->bd_blkbits + 1) {
1258 block = block >> 1;
1259 if (!mb_test_bit(block, bb)) {
1260 /* this block is part of buddy of order 'order' */
1261 return order;
1262 }
1263 bb += 1 << (e4b->bd_blkbits - order);
1264 order++;
1265 }
1266 return 0;
1267}
1268
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001269static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001270{
1271 __u32 *addr;
1272
1273 len = cur + len;
1274 while (cur < len) {
1275 if ((cur & 31) == 0 && (len - cur) >= 32) {
1276 /* fast path: clear whole word at once */
1277 addr = bm + (cur >> 3);
1278 *addr = 0;
1279 cur += 32;
1280 continue;
1281 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001282 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001283 cur++;
1284 }
1285}
1286
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001287/* clear bits in given range
1288 * will return first found zero bit if any, -1 otherwise
1289 */
1290static int mb_test_and_clear_bits(void *bm, int cur, int len)
1291{
1292 __u32 *addr;
1293 int zero_bit = -1;
1294
1295 len = cur + len;
1296 while (cur < len) {
1297 if ((cur & 31) == 0 && (len - cur) >= 32) {
1298 /* fast path: clear whole word at once */
1299 addr = bm + (cur >> 3);
1300 if (*addr != (__u32)(-1) && zero_bit == -1)
1301 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1302 *addr = 0;
1303 cur += 32;
1304 continue;
1305 }
1306 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1307 zero_bit = cur;
1308 cur++;
1309 }
1310
1311 return zero_bit;
1312}
1313
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001314void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001315{
1316 __u32 *addr;
1317
1318 len = cur + len;
1319 while (cur < len) {
1320 if ((cur & 31) == 0 && (len - cur) >= 32) {
1321 /* fast path: set whole word at once */
1322 addr = bm + (cur >> 3);
1323 *addr = 0xffffffff;
1324 cur += 32;
1325 continue;
1326 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001327 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001328 cur++;
1329 }
1330}
1331
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001332/*
1333 * _________________________________________________________________ */
1334
1335static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001336{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001337 if (mb_test_bit(*bit + side, bitmap)) {
1338 mb_clear_bit(*bit, bitmap);
1339 (*bit) -= side;
1340 return 1;
1341 }
1342 else {
1343 (*bit) += side;
1344 mb_set_bit(*bit, bitmap);
1345 return -1;
1346 }
1347}
1348
1349static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1350{
1351 int max;
1352 int order = 1;
1353 void *buddy = mb_find_buddy(e4b, order, &max);
1354
1355 while (buddy) {
1356 void *buddy2;
1357
1358 /* Bits in range [first; last] are known to be set since
1359 * corresponding blocks were allocated. Bits in range
1360 * (first; last) will stay set because they form buddies on
1361 * upper layer. We just deal with borders if they don't
1362 * align with upper layer and then go up.
1363 * Releasing entire group is all about clearing
1364 * single bit of highest order buddy.
1365 */
1366
1367 /* Example:
1368 * ---------------------------------
1369 * | 1 | 1 | 1 | 1 |
1370 * ---------------------------------
1371 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1372 * ---------------------------------
1373 * 0 1 2 3 4 5 6 7
1374 * \_____________________/
1375 *
1376 * Neither [1] nor [6] is aligned to above layer.
1377 * Left neighbour [0] is free, so mark it busy,
1378 * decrease bb_counters and extend range to
1379 * [0; 6]
1380 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1381 * mark [6] free, increase bb_counters and shrink range to
1382 * [0; 5].
1383 * Then shift range to [0; 2], go up and do the same.
1384 */
1385
1386
1387 if (first & 1)
1388 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1389 if (!(last & 1))
1390 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1391 if (first > last)
1392 break;
1393 order++;
1394
1395 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1396 mb_clear_bits(buddy, first, last - first + 1);
1397 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1398 break;
1399 }
1400 first >>= 1;
1401 last >>= 1;
1402 buddy = buddy2;
1403 }
1404}
1405
1406static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1407 int first, int count)
1408{
1409 int left_is_free = 0;
1410 int right_is_free = 0;
1411 int block;
1412 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001413 struct super_block *sb = e4b->bd_sb;
1414
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04001415 if (WARN_ON(count == 0))
1416 return;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001417 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001418 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Darrick J. Wong163a2032013-08-28 17:35:51 -04001419 /* Don't bother if the block group is corrupt. */
1420 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1421 return;
1422
Alex Tomasc9de5602008-01-29 00:19:52 -05001423 mb_check_buddy(e4b);
1424 mb_free_blocks_double(inode, e4b, first, count);
1425
1426 e4b->bd_info->bb_free += count;
1427 if (first < e4b->bd_info->bb_first_free)
1428 e4b->bd_info->bb_first_free = first;
1429
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001430 /* access memory sequentially: check left neighbour,
1431 * clear range and then check right neighbour
1432 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001433 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001434 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1435 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1436 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1437 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1438
1439 if (unlikely(block != -1)) {
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001440 struct ext4_sb_info *sbi = EXT4_SB(sb);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001441 ext4_fsblk_t blocknr;
1442
1443 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1444 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1445 ext4_grp_locked_error(sb, e4b->bd_group,
1446 inode ? inode->i_ino : 0,
1447 blocknr,
1448 "freeing already freed block "
Darrick J. Wong163a2032013-08-28 17:35:51 -04001449 "(bit %u); block bitmap corrupt.",
1450 block);
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001451 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1452 percpu_counter_sub(&sbi->s_freeclusters_counter,
1453 e4b->bd_info->bb_free);
Darrick J. Wong163a2032013-08-28 17:35:51 -04001454 /* Mark the block group as corrupt. */
1455 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1456 &e4b->bd_info->bb_state);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001457 mb_regenerate_buddy(e4b);
1458 goto done;
1459 }
1460
1461 /* let's maintain fragments counter */
1462 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001463 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001464 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001465 e4b->bd_info->bb_fragments++;
1466
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001467 /* buddy[0] == bd_bitmap is a special case, so handle
1468 * it right away and let mb_buddy_mark_free stay free of
1469 * zero order checks.
1470 * Check if neighbours are to be coaleasced,
1471 * adjust bitmap bb_counters and borders appropriately.
1472 */
1473 if (first & 1) {
1474 first += !left_is_free;
1475 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001476 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001477 if (!(last & 1)) {
1478 last -= !right_is_free;
1479 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1480 }
1481
1482 if (first <= last)
1483 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1484
1485done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001486 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001487 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001488}
1489
Robin Dong15c006a2012-08-17 10:02:17 -04001490static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001491 int needed, struct ext4_free_extent *ex)
1492{
1493 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001494 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001495 void *buddy;
1496
Vincent Minetbc8e6742009-05-15 08:33:18 -04001497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001498 BUG_ON(ex == NULL);
1499
Robin Dong15c006a2012-08-17 10:02:17 -04001500 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001501 BUG_ON(buddy == NULL);
1502 BUG_ON(block >= max);
1503 if (mb_test_bit(block, buddy)) {
1504 ex->fe_len = 0;
1505 ex->fe_start = 0;
1506 ex->fe_group = 0;
1507 return 0;
1508 }
1509
Robin Dong15c006a2012-08-17 10:02:17 -04001510 /* find actual order */
1511 order = mb_find_order_for_block(e4b, block);
1512 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001513
1514 ex->fe_len = 1 << order;
1515 ex->fe_start = block << order;
1516 ex->fe_group = e4b->bd_group;
1517
1518 /* calc difference from given start */
1519 next = next - ex->fe_start;
1520 ex->fe_len -= next;
1521 ex->fe_start += next;
1522
1523 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001524 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001525
1526 if (block + 1 >= max)
1527 break;
1528
1529 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001530 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001531 break;
1532
Robin Dongb051d8d2011-10-26 05:30:30 -04001533 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001534
Alex Tomasc9de5602008-01-29 00:19:52 -05001535 block = next >> order;
1536 ex->fe_len += 1 << order;
1537 }
1538
1539 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1540 return ex->fe_len;
1541}
1542
1543static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1544{
1545 int ord;
1546 int mlen = 0;
1547 int max = 0;
1548 int cur;
1549 int start = ex->fe_start;
1550 int len = ex->fe_len;
1551 unsigned ret = 0;
1552 int len0 = len;
1553 void *buddy;
1554
1555 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1556 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001557 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001558 mb_check_buddy(e4b);
1559 mb_mark_used_double(e4b, start, len);
1560
1561 e4b->bd_info->bb_free -= len;
1562 if (e4b->bd_info->bb_first_free == start)
1563 e4b->bd_info->bb_first_free += len;
1564
1565 /* let's maintain fragments counter */
1566 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001567 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001568 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001569 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001570 if (mlen && max)
1571 e4b->bd_info->bb_fragments++;
1572 else if (!mlen && !max)
1573 e4b->bd_info->bb_fragments--;
1574
1575 /* let's maintain buddy itself */
1576 while (len) {
1577 ord = mb_find_order_for_block(e4b, start);
1578
1579 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1580 /* the whole chunk may be allocated at once! */
1581 mlen = 1 << ord;
1582 buddy = mb_find_buddy(e4b, ord, &max);
1583 BUG_ON((start >> ord) >= max);
1584 mb_set_bit(start >> ord, buddy);
1585 e4b->bd_info->bb_counters[ord]--;
1586 start += mlen;
1587 len -= mlen;
1588 BUG_ON(len < 0);
1589 continue;
1590 }
1591
1592 /* store for history */
1593 if (ret == 0)
1594 ret = len | (ord << 16);
1595
1596 /* we have to split large buddy */
1597 BUG_ON(ord <= 0);
1598 buddy = mb_find_buddy(e4b, ord, &max);
1599 mb_set_bit(start >> ord, buddy);
1600 e4b->bd_info->bb_counters[ord]--;
1601
1602 ord--;
1603 cur = (start >> ord) & ~1U;
1604 buddy = mb_find_buddy(e4b, ord, &max);
1605 mb_clear_bit(cur, buddy);
1606 mb_clear_bit(cur + 1, buddy);
1607 e4b->bd_info->bb_counters[ord]++;
1608 e4b->bd_info->bb_counters[ord]++;
1609 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001610 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001611
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001612 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001613 mb_check_buddy(e4b);
1614
1615 return ret;
1616}
1617
1618/*
1619 * Must be called under group lock!
1620 */
1621static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1622 struct ext4_buddy *e4b)
1623{
1624 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1625 int ret;
1626
1627 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1628 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1629
1630 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1631 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1632 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1633
1634 /* preallocation can change ac_b_ex, thus we store actually
1635 * allocated blocks for history */
1636 ac->ac_f_ex = ac->ac_b_ex;
1637
1638 ac->ac_status = AC_STATUS_FOUND;
1639 ac->ac_tail = ret & 0xffff;
1640 ac->ac_buddy = ret >> 16;
1641
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001642 /*
1643 * take the page reference. We want the page to be pinned
1644 * so that we don't get a ext4_mb_init_cache_call for this
1645 * group until we update the bitmap. That would mean we
1646 * double allocate blocks. The reference is dropped
1647 * in ext4_mb_release_context
1648 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001649 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1650 get_page(ac->ac_bitmap_page);
1651 ac->ac_buddy_page = e4b->bd_buddy_page;
1652 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001653 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001654 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001655 spin_lock(&sbi->s_md_lock);
1656 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1657 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1658 spin_unlock(&sbi->s_md_lock);
1659 }
1660}
1661
1662/*
1663 * regular allocator, for general purposes allocation
1664 */
1665
1666static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1667 struct ext4_buddy *e4b,
1668 int finish_group)
1669{
1670 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1671 struct ext4_free_extent *bex = &ac->ac_b_ex;
1672 struct ext4_free_extent *gex = &ac->ac_g_ex;
1673 struct ext4_free_extent ex;
1674 int max;
1675
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001676 if (ac->ac_status == AC_STATUS_FOUND)
1677 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001678 /*
1679 * We don't want to scan for a whole year
1680 */
1681 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1682 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1683 ac->ac_status = AC_STATUS_BREAK;
1684 return;
1685 }
1686
1687 /*
1688 * Haven't found good chunk so far, let's continue
1689 */
1690 if (bex->fe_len < gex->fe_len)
1691 return;
1692
1693 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1694 && bex->fe_group == e4b->bd_group) {
1695 /* recheck chunk's availability - we don't know
1696 * when it was found (within this lock-unlock
1697 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001698 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001699 if (max >= gex->fe_len) {
1700 ext4_mb_use_best_found(ac, e4b);
1701 return;
1702 }
1703 }
1704}
1705
1706/*
1707 * The routine checks whether found extent is good enough. If it is,
1708 * then the extent gets marked used and flag is set to the context
1709 * to stop scanning. Otherwise, the extent is compared with the
1710 * previous found extent and if new one is better, then it's stored
1711 * in the context. Later, the best found extent will be used, if
1712 * mballoc can't find good enough extent.
1713 *
1714 * FIXME: real allocation policy is to be designed yet!
1715 */
1716static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1717 struct ext4_free_extent *ex,
1718 struct ext4_buddy *e4b)
1719{
1720 struct ext4_free_extent *bex = &ac->ac_b_ex;
1721 struct ext4_free_extent *gex = &ac->ac_g_ex;
1722
1723 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001724 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1725 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001726 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1727
1728 ac->ac_found++;
1729
1730 /*
1731 * The special case - take what you catch first
1732 */
1733 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1734 *bex = *ex;
1735 ext4_mb_use_best_found(ac, e4b);
1736 return;
1737 }
1738
1739 /*
1740 * Let's check whether the chuck is good enough
1741 */
1742 if (ex->fe_len == gex->fe_len) {
1743 *bex = *ex;
1744 ext4_mb_use_best_found(ac, e4b);
1745 return;
1746 }
1747
1748 /*
1749 * If this is first found extent, just store it in the context
1750 */
1751 if (bex->fe_len == 0) {
1752 *bex = *ex;
1753 return;
1754 }
1755
1756 /*
1757 * If new found extent is better, store it in the context
1758 */
1759 if (bex->fe_len < gex->fe_len) {
1760 /* if the request isn't satisfied, any found extent
1761 * larger than previous best one is better */
1762 if (ex->fe_len > bex->fe_len)
1763 *bex = *ex;
1764 } else if (ex->fe_len > gex->fe_len) {
1765 /* if the request is satisfied, then we try to find
1766 * an extent that still satisfy the request, but is
1767 * smaller than previous one */
1768 if (ex->fe_len < bex->fe_len)
1769 *bex = *ex;
1770 }
1771
1772 ext4_mb_check_limits(ac, e4b, 0);
1773}
1774
Eric Sandeen089ceec2009-07-05 22:17:31 -04001775static noinline_for_stack
1776int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001777 struct ext4_buddy *e4b)
1778{
1779 struct ext4_free_extent ex = ac->ac_b_ex;
1780 ext4_group_t group = ex.fe_group;
1781 int max;
1782 int err;
1783
1784 BUG_ON(ex.fe_len <= 0);
1785 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1786 if (err)
1787 return err;
1788
1789 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001790 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001791
1792 if (max > 0) {
1793 ac->ac_b_ex = ex;
1794 ext4_mb_use_best_found(ac, e4b);
1795 }
1796
1797 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001798 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001799
1800 return 0;
1801}
1802
Eric Sandeen089ceec2009-07-05 22:17:31 -04001803static noinline_for_stack
1804int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001805 struct ext4_buddy *e4b)
1806{
1807 ext4_group_t group = ac->ac_g_ex.fe_group;
1808 int max;
1809 int err;
1810 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001811 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001812 struct ext4_free_extent ex;
1813
1814 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1815 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001816 if (grp->bb_free == 0)
1817 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001818
1819 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1820 if (err)
1821 return err;
1822
Darrick J. Wong163a2032013-08-28 17:35:51 -04001823 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1824 ext4_mb_unload_buddy(e4b);
1825 return 0;
1826 }
1827
Alex Tomasc9de5602008-01-29 00:19:52 -05001828 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001829 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001830 ac->ac_g_ex.fe_len, &ex);
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001831 ex.fe_logical = 0xDEADFA11; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001832
1833 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1834 ext4_fsblk_t start;
1835
Akinobu Mita5661bd62010-03-03 23:53:39 -05001836 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1837 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001838 /* use do_div to get remainder (would be 64-bit modulo) */
1839 if (do_div(start, sbi->s_stripe) == 0) {
1840 ac->ac_found++;
1841 ac->ac_b_ex = ex;
1842 ext4_mb_use_best_found(ac, e4b);
1843 }
1844 } else if (max >= ac->ac_g_ex.fe_len) {
1845 BUG_ON(ex.fe_len <= 0);
1846 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1847 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1848 ac->ac_found++;
1849 ac->ac_b_ex = ex;
1850 ext4_mb_use_best_found(ac, e4b);
1851 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1852 /* Sometimes, caller may want to merge even small
1853 * number of blocks to an existing extent */
1854 BUG_ON(ex.fe_len <= 0);
1855 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1856 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1857 ac->ac_found++;
1858 ac->ac_b_ex = ex;
1859 ext4_mb_use_best_found(ac, e4b);
1860 }
1861 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001862 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001863
1864 return 0;
1865}
1866
1867/*
1868 * The routine scans buddy structures (not bitmap!) from given order
1869 * to max order and tries to find big enough chunk to satisfy the req
1870 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001871static noinline_for_stack
1872void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001873 struct ext4_buddy *e4b)
1874{
1875 struct super_block *sb = ac->ac_sb;
1876 struct ext4_group_info *grp = e4b->bd_info;
1877 void *buddy;
1878 int i;
1879 int k;
1880 int max;
1881
1882 BUG_ON(ac->ac_2order <= 0);
1883 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1884 if (grp->bb_counters[i] == 0)
1885 continue;
1886
1887 buddy = mb_find_buddy(e4b, i, &max);
1888 BUG_ON(buddy == NULL);
1889
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001890 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001891 BUG_ON(k >= max);
1892
1893 ac->ac_found++;
1894
1895 ac->ac_b_ex.fe_len = 1 << i;
1896 ac->ac_b_ex.fe_start = k << i;
1897 ac->ac_b_ex.fe_group = e4b->bd_group;
1898
1899 ext4_mb_use_best_found(ac, e4b);
1900
1901 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1902
1903 if (EXT4_SB(sb)->s_mb_stats)
1904 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1905
1906 break;
1907 }
1908}
1909
1910/*
1911 * The routine scans the group and measures all found extents.
1912 * In order to optimize scanning, caller must pass number of
1913 * free blocks in the group, so the routine can know upper limit.
1914 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001915static noinline_for_stack
1916void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001917 struct ext4_buddy *e4b)
1918{
1919 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001920 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001921 struct ext4_free_extent ex;
1922 int i;
1923 int free;
1924
1925 free = e4b->bd_info->bb_free;
1926 BUG_ON(free <= 0);
1927
1928 i = e4b->bd_info->bb_first_free;
1929
1930 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001931 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001932 EXT4_CLUSTERS_PER_GROUP(sb), i);
1933 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001934 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001935 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001936 * free blocks even though group info says we
1937 * we have free blocks
1938 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001939 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001940 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001941 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001942 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001943 break;
1944 }
1945
Robin Dong15c006a2012-08-17 10:02:17 -04001946 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001947 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001948 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001949 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001950 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001951 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001952 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001953 /*
1954 * The number of free blocks differs. This mostly
1955 * indicate that the bitmap is corrupt. So exit
1956 * without claiming the space.
1957 */
1958 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001959 }
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001960 ex.fe_logical = 0xDEADC0DE; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001961 ext4_mb_measure_extent(ac, &ex, e4b);
1962
1963 i += ex.fe_len;
1964 free -= ex.fe_len;
1965 }
1966
1967 ext4_mb_check_limits(ac, e4b, 1);
1968}
1969
1970/*
1971 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001972 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001973 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001974static noinline_for_stack
1975void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001976 struct ext4_buddy *e4b)
1977{
1978 struct super_block *sb = ac->ac_sb;
1979 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001980 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001981 struct ext4_free_extent ex;
1982 ext4_fsblk_t first_group_block;
1983 ext4_fsblk_t a;
1984 ext4_grpblk_t i;
1985 int max;
1986
1987 BUG_ON(sbi->s_stripe == 0);
1988
1989 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001990 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1991
Alex Tomasc9de5602008-01-29 00:19:52 -05001992 a = first_group_block + sbi->s_stripe - 1;
1993 do_div(a, sbi->s_stripe);
1994 i = (a * sbi->s_stripe) - first_group_block;
1995
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001996 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001997 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04001998 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001999 if (max >= sbi->s_stripe) {
2000 ac->ac_found++;
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05002001 ex.fe_logical = 0xDEADF00D; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05002002 ac->ac_b_ex = ex;
2003 ext4_mb_use_best_found(ac, e4b);
2004 break;
2005 }
2006 }
2007 i += sbi->s_stripe;
2008 }
2009}
2010
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002011/* This is now called BEFORE we load the buddy bitmap. */
Alex Tomasc9de5602008-01-29 00:19:52 -05002012static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2013 ext4_group_t group, int cr)
2014{
2015 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04002016 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05002017 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2018
2019 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002020
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04002021 free = grp->bb_free;
2022 if (free == 0)
2023 return 0;
2024 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2025 return 0;
2026
Darrick J. Wong163a2032013-08-28 17:35:51 -04002027 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2028 return 0;
2029
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002030 /* We only do this if the grp has never been initialized */
2031 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2032 int ret = ext4_mb_init_group(ac->ac_sb, group);
2033 if (ret)
2034 return 0;
2035 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002036
Alex Tomasc9de5602008-01-29 00:19:52 -05002037 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05002038 if (fragments == 0)
2039 return 0;
2040
2041 switch (cr) {
2042 case 0:
2043 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002044
Theodore Ts'oa4912122009-03-12 12:18:34 -04002045 /* Avoid using the first bg of a flexgroup for data files */
2046 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2047 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2048 ((group % flex_size) == 0))
2049 return 0;
2050
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002051 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2052 (free / fragments) >= ac->ac_g_ex.fe_len)
2053 return 1;
2054
2055 if (grp->bb_largest_free_order < ac->ac_2order)
2056 return 0;
2057
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002058 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002059 case 1:
2060 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2061 return 1;
2062 break;
2063 case 2:
2064 if (free >= ac->ac_g_ex.fe_len)
2065 return 1;
2066 break;
2067 case 3:
2068 return 1;
2069 default:
2070 BUG();
2071 }
2072
2073 return 0;
2074}
2075
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002076static noinline_for_stack int
2077ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002078{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002079 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002080 int cr;
2081 int err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002082 struct ext4_sb_info *sbi;
2083 struct super_block *sb;
2084 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002085
2086 sb = ac->ac_sb;
2087 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002088 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002089 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002090 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002091 ngroups = sbi->s_blockfile_groups;
2092
Alex Tomasc9de5602008-01-29 00:19:52 -05002093 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2094
2095 /* first, try the goal */
2096 err = ext4_mb_find_by_goal(ac, &e4b);
2097 if (err || ac->ac_status == AC_STATUS_FOUND)
2098 goto out;
2099
2100 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2101 goto out;
2102
2103 /*
2104 * ac->ac2_order is set only if the fe_len is a power of 2
2105 * if ac2_order is set we also set criteria to 0 so that we
2106 * try exact allocation using buddy.
2107 */
2108 i = fls(ac->ac_g_ex.fe_len);
2109 ac->ac_2order = 0;
2110 /*
2111 * We search using buddy data only if the order of the request
2112 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002113 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002114 */
2115 if (i >= sbi->s_mb_order2_reqs) {
2116 /*
2117 * This should tell if fe_len is exactly power of 2
2118 */
2119 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2120 ac->ac_2order = i - 1;
2121 }
2122
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002123 /* if stream allocation is enabled, use global goal */
2124 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002125 /* TBD: may be hot point */
2126 spin_lock(&sbi->s_md_lock);
2127 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2128 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2129 spin_unlock(&sbi->s_md_lock);
2130 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002131
Alex Tomasc9de5602008-01-29 00:19:52 -05002132 /* Let's just scan groups to find more-less suitable blocks */
2133 cr = ac->ac_2order ? 0 : 1;
2134 /*
2135 * cr == 0 try to get exact allocation,
2136 * cr == 3 try to get anything
2137 */
2138repeat:
2139 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2140 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002141 /*
2142 * searching for the right group start
2143 * from the goal value specified
2144 */
2145 group = ac->ac_g_ex.fe_group;
2146
Theodore Ts'o8df96752009-05-01 08:50:38 -04002147 for (i = 0; i < ngroups; group++, i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04002148 cond_resched();
Lachlan McIlroye6155732013-05-05 23:10:00 -04002149 /*
2150 * Artificially restricted ngroups for non-extent
2151 * files makes group > ngroups possible on first loop.
2152 */
2153 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002154 group = 0;
2155
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002156 /* This now checks without needing the buddy page */
2157 if (!ext4_mb_good_group(ac, group, cr))
Alex Tomasc9de5602008-01-29 00:19:52 -05002158 continue;
2159
Alex Tomasc9de5602008-01-29 00:19:52 -05002160 err = ext4_mb_load_buddy(sb, group, &e4b);
2161 if (err)
2162 goto out;
2163
2164 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002165
2166 /*
2167 * We need to check again after locking the
2168 * block group
2169 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002170 if (!ext4_mb_good_group(ac, group, cr)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002171 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002172 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002173 continue;
2174 }
2175
2176 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002177 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002178 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002179 else if (cr == 1 && sbi->s_stripe &&
2180 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002181 ext4_mb_scan_aligned(ac, &e4b);
2182 else
2183 ext4_mb_complex_scan_group(ac, &e4b);
2184
2185 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002186 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002187
2188 if (ac->ac_status != AC_STATUS_CONTINUE)
2189 break;
2190 }
2191 }
2192
2193 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2194 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2195 /*
2196 * We've been searching too long. Let's try to allocate
2197 * the best chunk we've found so far
2198 */
2199
2200 ext4_mb_try_best_found(ac, &e4b);
2201 if (ac->ac_status != AC_STATUS_FOUND) {
2202 /*
2203 * Someone more lucky has already allocated it.
2204 * The only thing we can do is just take first
2205 * found block(s)
2206 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2207 */
2208 ac->ac_b_ex.fe_group = 0;
2209 ac->ac_b_ex.fe_start = 0;
2210 ac->ac_b_ex.fe_len = 0;
2211 ac->ac_status = AC_STATUS_CONTINUE;
2212 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2213 cr = 3;
2214 atomic_inc(&sbi->s_mb_lost_chunks);
2215 goto repeat;
2216 }
2217 }
2218out:
2219 return err;
2220}
2221
Alex Tomasc9de5602008-01-29 00:19:52 -05002222static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2223{
2224 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002225 ext4_group_t group;
2226
Theodore Ts'o8df96752009-05-01 08:50:38 -04002227 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002228 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002229 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002230 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002231}
2232
2233static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2234{
2235 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002236 ext4_group_t group;
2237
2238 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002239 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002240 return NULL;
2241 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002242 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002243}
2244
2245static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2246{
2247 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002248 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002249 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002250 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002251 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002252 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002253 struct sg {
2254 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002255 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002256 } sg;
2257
2258 group--;
2259 if (group == 0)
2260 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2261 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2262 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2263 "group", "free", "frags", "first",
2264 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2265 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2266
2267 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2268 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002269 grinfo = ext4_get_group_info(sb, group);
2270 /* Load the group info in memory only if not already loaded. */
2271 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2272 err = ext4_mb_load_buddy(sb, group, &e4b);
2273 if (err) {
2274 seq_printf(seq, "#%-5u: I/O error\n", group);
2275 return 0;
2276 }
2277 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002278 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002279
Alex Tomasc9de5602008-01-29 00:19:52 -05002280 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002281
2282 if (buddy_loaded)
2283 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002284
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002285 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002286 sg.info.bb_fragments, sg.info.bb_first_free);
2287 for (i = 0; i <= 13; i++)
2288 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2289 sg.info.bb_counters[i] : 0);
2290 seq_printf(seq, " ]\n");
2291
2292 return 0;
2293}
2294
2295static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2296{
2297}
2298
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002299static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002300 .start = ext4_mb_seq_groups_start,
2301 .next = ext4_mb_seq_groups_next,
2302 .stop = ext4_mb_seq_groups_stop,
2303 .show = ext4_mb_seq_groups_show,
2304};
2305
2306static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2307{
Al Virod9dda782013-03-31 18:16:14 -04002308 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002309 int rc;
2310
2311 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2312 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002313 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002314 m->private = sb;
2315 }
2316 return rc;
2317
2318}
2319
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002320static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002321 .owner = THIS_MODULE,
2322 .open = ext4_mb_seq_groups_open,
2323 .read = seq_read,
2324 .llseek = seq_lseek,
2325 .release = seq_release,
2326};
2327
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002328static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2329{
2330 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2331 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2332
2333 BUG_ON(!cachep);
2334 return cachep;
2335}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002336
Theodore Ts'o28623c22012-09-05 01:31:50 -04002337/*
2338 * Allocate the top-level s_group_info array for the specified number
2339 * of groups
2340 */
2341int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2342{
2343 struct ext4_sb_info *sbi = EXT4_SB(sb);
2344 unsigned size;
2345 struct ext4_group_info ***new_groupinfo;
2346
2347 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2348 EXT4_DESC_PER_BLOCK_BITS(sb);
2349 if (size <= sbi->s_group_info_size)
2350 return 0;
2351
2352 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2353 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2354 if (!new_groupinfo) {
2355 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2356 return -ENOMEM;
2357 }
2358 if (sbi->s_group_info) {
2359 memcpy(new_groupinfo, sbi->s_group_info,
2360 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2361 ext4_kvfree(sbi->s_group_info);
2362 }
2363 sbi->s_group_info = new_groupinfo;
2364 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2365 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2366 sbi->s_group_info_size);
2367 return 0;
2368}
2369
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002370/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002371int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002372 struct ext4_group_desc *desc)
2373{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002374 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002375 int metalen = 0;
2376 struct ext4_sb_info *sbi = EXT4_SB(sb);
2377 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002378 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002379
2380 /*
2381 * First check if this group is the first of a reserved block.
2382 * If it's true, we have to allocate a new table of pointers
2383 * to ext4_group_info structures
2384 */
2385 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2386 metalen = sizeof(*meta_group_info) <<
2387 EXT4_DESC_PER_BLOCK_BITS(sb);
2388 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2389 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002390 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002391 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002392 goto exit_meta_group_info;
2393 }
2394 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2395 meta_group_info;
2396 }
2397
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002398 meta_group_info =
2399 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2400 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2401
Wei Yongjun85556c92012-09-26 20:43:37 -04002402 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002403 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002404 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002405 goto exit_group_info;
2406 }
2407 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2408 &(meta_group_info[i]->bb_state));
2409
2410 /*
2411 * initialize bb_free to be able to skip
2412 * empty groups without initialization
2413 */
2414 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2415 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002416 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002417 } else {
2418 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002419 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002420 }
2421
2422 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002423 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002424 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002425 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002426
2427#ifdef DOUBLE_CHECK
2428 {
2429 struct buffer_head *bh;
2430 meta_group_info[i]->bb_bitmap =
2431 kmalloc(sb->s_blocksize, GFP_KERNEL);
2432 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2433 bh = ext4_read_block_bitmap(sb, group);
2434 BUG_ON(bh == NULL);
2435 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2436 sb->s_blocksize);
2437 put_bh(bh);
2438 }
2439#endif
2440
2441 return 0;
2442
2443exit_group_info:
2444 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002445 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002446 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002447 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2448 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002449exit_meta_group_info:
2450 return -ENOMEM;
2451} /* ext4_mb_add_groupinfo */
2452
Alex Tomasc9de5602008-01-29 00:19:52 -05002453static int ext4_mb_init_backend(struct super_block *sb)
2454{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002455 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002456 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002457 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002458 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002459 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002460 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002461
Theodore Ts'o28623c22012-09-05 01:31:50 -04002462 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2463 if (err)
2464 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002465
Alex Tomasc9de5602008-01-29 00:19:52 -05002466 sbi->s_buddy_cache = new_inode(sb);
2467 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002468 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002469 goto err_freesgi;
2470 }
Yu Jian48e60612011-08-01 17:41:39 -04002471 /* To avoid potentially colliding with an valid on-disk inode number,
2472 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2473 * not in the inode hash, so it should never be found by iget(), but
2474 * this will avoid confusion if it ever shows up during debugging. */
2475 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002476 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002477 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002478 desc = ext4_get_group_desc(sb, i, NULL);
2479 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002480 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002481 goto err_freebuddy;
2482 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002483 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2484 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002485 }
2486
2487 return 0;
2488
2489err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002490 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002491 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002492 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002493 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002494 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002495 kfree(sbi->s_group_info[i]);
2496 iput(sbi->s_buddy_cache);
2497err_freesgi:
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002498 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002499 return -ENOMEM;
2500}
2501
Eric Sandeen2892c152011-02-12 08:12:18 -05002502static void ext4_groupinfo_destroy_slabs(void)
2503{
2504 int i;
2505
2506 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2507 if (ext4_groupinfo_caches[i])
2508 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2509 ext4_groupinfo_caches[i] = NULL;
2510 }
2511}
2512
2513static int ext4_groupinfo_create_slab(size_t size)
2514{
2515 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2516 int slab_size;
2517 int blocksize_bits = order_base_2(size);
2518 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2519 struct kmem_cache *cachep;
2520
2521 if (cache_index >= NR_GRPINFO_CACHES)
2522 return -EINVAL;
2523
2524 if (unlikely(cache_index < 0))
2525 cache_index = 0;
2526
2527 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2528 if (ext4_groupinfo_caches[cache_index]) {
2529 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2530 return 0; /* Already created */
2531 }
2532
2533 slab_size = offsetof(struct ext4_group_info,
2534 bb_counters[blocksize_bits + 2]);
2535
2536 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2537 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2538 NULL);
2539
Tao Ma823ba012011-07-11 18:26:01 -04002540 ext4_groupinfo_caches[cache_index] = cachep;
2541
Eric Sandeen2892c152011-02-12 08:12:18 -05002542 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2543 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002544 printk(KERN_EMERG
2545 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002546 return -ENOMEM;
2547 }
2548
Eric Sandeen2892c152011-02-12 08:12:18 -05002549 return 0;
2550}
2551
Akira Fujita9d990122012-05-28 14:19:25 -04002552int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002553{
2554 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002555 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002556 unsigned offset;
2557 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002558 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002559
Eric Sandeen19278052009-08-25 22:36:25 -04002560 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002561
2562 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2563 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002564 ret = -ENOMEM;
2565 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002566 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002567
Eric Sandeen19278052009-08-25 22:36:25 -04002568 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002569 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2570 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002571 ret = -ENOMEM;
2572 goto out;
2573 }
2574
Eric Sandeen2892c152011-02-12 08:12:18 -05002575 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2576 if (ret < 0)
2577 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002578
2579 /* order 0 is regular bitmap */
2580 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2581 sbi->s_mb_offsets[0] = 0;
2582
2583 i = 1;
2584 offset = 0;
2585 max = sb->s_blocksize << 2;
2586 do {
2587 sbi->s_mb_offsets[i] = offset;
2588 sbi->s_mb_maxs[i] = max;
2589 offset += 1 << (sb->s_blocksize_bits - i);
2590 max = max >> 1;
2591 i++;
2592 } while (i <= sb->s_blocksize_bits + 1);
2593
Alex Tomasc9de5602008-01-29 00:19:52 -05002594 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002595 spin_lock_init(&sbi->s_bal_lock);
2596
2597 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2598 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2599 sbi->s_mb_stats = MB_DEFAULT_STATS;
2600 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2601 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002602 /*
2603 * The default group preallocation is 512, which for 4k block
2604 * sizes translates to 2 megabytes. However for bigalloc file
2605 * systems, this is probably too big (i.e, if the cluster size
2606 * is 1 megabyte, then group preallocation size becomes half a
2607 * gigabyte!). As a default, we will keep a two megabyte
2608 * group pralloc size for cluster sizes up to 64k, and after
2609 * that, we will force a minimum group preallocation size of
2610 * 32 clusters. This translates to 8 megs when the cluster
2611 * size is 256k, and 32 megs when the cluster size is 1 meg,
2612 * which seems reasonable as a default.
2613 */
2614 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2615 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002616 /*
2617 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2618 * to the lowest multiple of s_stripe which is bigger than
2619 * the s_mb_group_prealloc as determined above. We want
2620 * the preallocation size to be an exact multiple of the
2621 * RAID stripe size so that preallocations don't fragment
2622 * the stripes.
2623 */
2624 if (sbi->s_stripe > 1) {
2625 sbi->s_mb_group_prealloc = roundup(
2626 sbi->s_mb_group_prealloc, sbi->s_stripe);
2627 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002628
Eric Sandeen730c2132008-09-13 15:23:29 -04002629 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002630 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002631 ret = -ENOMEM;
Andrey Tsyvarev029b10c2014-05-12 12:34:21 -04002632 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002633 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002634 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002635 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002636 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002637 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002638 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2639 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002640 spin_lock_init(&lg->lg_prealloc_lock);
2641 }
2642
Yu Jian79a77c52011-08-01 17:41:46 -04002643 /* init file for buddy data */
2644 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002645 if (ret != 0)
2646 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002647
Theodore Ts'o296c3552009-09-30 00:32:42 -04002648 if (sbi->s_proc)
2649 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2650 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002651
Tao Ma7aa0bae2011-10-06 10:22:28 -04002652 return 0;
2653
2654out_free_locality_groups:
2655 free_percpu(sbi->s_locality_groups);
2656 sbi->s_locality_groups = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002657out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002658 kfree(sbi->s_mb_offsets);
2659 sbi->s_mb_offsets = NULL;
2660 kfree(sbi->s_mb_maxs);
2661 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002662 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002663}
2664
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002665/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002666static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2667{
2668 struct ext4_prealloc_space *pa;
2669 struct list_head *cur, *tmp;
2670 int count = 0;
2671
2672 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2673 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2674 list_del(&pa->pa_group_list);
2675 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002676 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002677 }
2678 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002679 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002680
2681}
2682
2683int ext4_mb_release(struct super_block *sb)
2684{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002685 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002686 ext4_group_t i;
2687 int num_meta_group_infos;
2688 struct ext4_group_info *grinfo;
2689 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002690 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002691
Salman Qazi95599962012-05-31 23:52:14 -04002692 if (sbi->s_proc)
2693 remove_proc_entry("mb_groups", sbi->s_proc);
2694
Alex Tomasc9de5602008-01-29 00:19:52 -05002695 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002696 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002697 grinfo = ext4_get_group_info(sb, i);
2698#ifdef DOUBLE_CHECK
2699 kfree(grinfo->bb_bitmap);
2700#endif
2701 ext4_lock_group(sb, i);
2702 ext4_mb_cleanup_pa(grinfo);
2703 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002704 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002705 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002706 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002707 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2708 EXT4_DESC_PER_BLOCK_BITS(sb);
2709 for (i = 0; i < num_meta_group_infos; i++)
2710 kfree(sbi->s_group_info[i]);
Theodore Ts'of18a5f22011-08-01 08:45:38 -04002711 ext4_kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002712 }
2713 kfree(sbi->s_mb_offsets);
2714 kfree(sbi->s_mb_maxs);
2715 if (sbi->s_buddy_cache)
2716 iput(sbi->s_buddy_cache);
2717 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002718 ext4_msg(sb, KERN_INFO,
2719 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002720 atomic_read(&sbi->s_bal_allocated),
2721 atomic_read(&sbi->s_bal_reqs),
2722 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002723 ext4_msg(sb, KERN_INFO,
2724 "mballoc: %u extents scanned, %u goal hits, "
2725 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002726 atomic_read(&sbi->s_bal_ex_scanned),
2727 atomic_read(&sbi->s_bal_goals),
2728 atomic_read(&sbi->s_bal_2orders),
2729 atomic_read(&sbi->s_bal_breaks),
2730 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002731 ext4_msg(sb, KERN_INFO,
2732 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002733 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002734 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002735 ext4_msg(sb, KERN_INFO,
2736 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002737 atomic_read(&sbi->s_mb_preallocated),
2738 atomic_read(&sbi->s_mb_discarded));
2739 }
2740
Eric Sandeen730c2132008-09-13 15:23:29 -04002741 free_percpu(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05002742
2743 return 0;
2744}
2745
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002746static inline int ext4_issue_discard(struct super_block *sb,
Theodore Ts'o84130192011-09-09 18:50:51 -04002747 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
Jiaying Zhang5c521832010-07-27 11:56:05 -04002748{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002749 ext4_fsblk_t discard_block;
2750
Theodore Ts'o84130192011-09-09 18:50:51 -04002751 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2752 ext4_group_first_block_no(sb, block_group));
2753 count = EXT4_C2B(EXT4_SB(sb), count);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002754 trace_ext4_discard_blocks(sb,
2755 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002756 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002757}
2758
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002759/*
2760 * This function is called by the jbd2 layer once the commit has finished,
2761 * so we know we can free the blocks that were released with that commit.
2762 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002763static void ext4_free_data_callback(struct super_block *sb,
2764 struct ext4_journal_cb_entry *jce,
2765 int rc)
Alex Tomasc9de5602008-01-29 00:19:52 -05002766{
Bobi Jam18aadd42012-02-20 17:53:02 -05002767 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
Alex Tomasc9de5602008-01-29 00:19:52 -05002768 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002769 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002770 int err, count = 0, count2 = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002771
Bobi Jam18aadd42012-02-20 17:53:02 -05002772 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2773 entry->efd_count, entry->efd_group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002774
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05002775 if (test_opt(sb, DISCARD)) {
2776 err = ext4_issue_discard(sb, entry->efd_group,
2777 entry->efd_start_cluster,
2778 entry->efd_count);
2779 if (err && err != -EOPNOTSUPP)
2780 ext4_msg(sb, KERN_WARNING, "discard request in"
2781 " group:%d block:%d count:%d failed"
2782 " with %d", entry->efd_group,
2783 entry->efd_start_cluster,
2784 entry->efd_count, err);
2785 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002786
Bobi Jam18aadd42012-02-20 17:53:02 -05002787 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2788 /* we expect to find existing buddy because it's pinned */
2789 BUG_ON(err != 0);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002790
Alex Tomasc9de5602008-01-29 00:19:52 -05002791
Bobi Jam18aadd42012-02-20 17:53:02 -05002792 db = e4b.bd_info;
2793 /* there are blocks to put in buddy to make them really free */
2794 count += entry->efd_count;
2795 count2++;
2796 ext4_lock_group(sb, entry->efd_group);
2797 /* Take it out of per group rb tree */
2798 rb_erase(&entry->efd_node, &(db->bb_free_root));
2799 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002800
Bobi Jam18aadd42012-02-20 17:53:02 -05002801 /*
2802 * Clear the trimmed flag for the group so that the next
2803 * ext4_trim_fs can trim it.
2804 * If the volume is mounted with -o discard, online discard
2805 * is supported and the free blocks will be trimmed online.
2806 */
2807 if (!test_opt(sb, DISCARD))
2808 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2809
2810 if (!db->bb_free_root.rb_node) {
2811 /* No more items in the per group rb tree
2812 * balance refcounts from ext4_mb_free_metadata()
Tao Ma3d56b8d2011-07-11 00:03:38 -04002813 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002814 page_cache_release(e4b.bd_buddy_page);
2815 page_cache_release(e4b.bd_bitmap_page);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002816 }
Bobi Jam18aadd42012-02-20 17:53:02 -05002817 ext4_unlock_group(sb, entry->efd_group);
2818 kmem_cache_free(ext4_free_data_cachep, entry);
2819 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002820
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002821 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002822}
2823
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002824int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002825{
Theodore Ts'o16828082010-10-27 21:30:09 -04002826 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2827 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002828 if (ext4_pspace_cachep == NULL)
2829 return -ENOMEM;
2830
Theodore Ts'o16828082010-10-27 21:30:09 -04002831 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2832 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002833 if (ext4_ac_cachep == NULL) {
2834 kmem_cache_destroy(ext4_pspace_cachep);
2835 return -ENOMEM;
2836 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002837
Bobi Jam18aadd42012-02-20 17:53:02 -05002838 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2839 SLAB_RECLAIM_ACCOUNT);
2840 if (ext4_free_data_cachep == NULL) {
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002841 kmem_cache_destroy(ext4_pspace_cachep);
2842 kmem_cache_destroy(ext4_ac_cachep);
2843 return -ENOMEM;
2844 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002845 return 0;
2846}
2847
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002848void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002849{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002850 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002851 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2852 * before destroying the slab cache.
2853 */
2854 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002855 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002856 kmem_cache_destroy(ext4_ac_cachep);
Bobi Jam18aadd42012-02-20 17:53:02 -05002857 kmem_cache_destroy(ext4_free_data_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002858 ext4_groupinfo_destroy_slabs();
Alex Tomasc9de5602008-01-29 00:19:52 -05002859}
2860
2861
2862/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002863 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002864 * Returns 0 if success or error code
2865 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002866static noinline_for_stack int
2867ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002868 handle_t *handle, unsigned int reserv_clstrs)
Alex Tomasc9de5602008-01-29 00:19:52 -05002869{
2870 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002871 struct ext4_group_desc *gdp;
2872 struct buffer_head *gdp_bh;
2873 struct ext4_sb_info *sbi;
2874 struct super_block *sb;
2875 ext4_fsblk_t block;
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002876 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002877
2878 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2879 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2880
2881 sb = ac->ac_sb;
2882 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002883
2884 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002885 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002886 if (!bitmap_bh)
2887 goto out_err;
2888
liang xie5d601252014-05-12 22:06:43 -04002889 BUFFER_TRACE(bitmap_bh, "getting write access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002890 err = ext4_journal_get_write_access(handle, bitmap_bh);
2891 if (err)
2892 goto out_err;
2893
2894 err = -EIO;
2895 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2896 if (!gdp)
2897 goto out_err;
2898
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002899 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002900 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002901
liang xie5d601252014-05-12 22:06:43 -04002902 BUFFER_TRACE(gdp_bh, "get_write_access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002903 err = ext4_journal_get_write_access(handle, gdp_bh);
2904 if (err)
2905 goto out_err;
2906
Akinobu Mitabda00de2010-03-03 23:53:25 -05002907 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002908
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002909 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002910 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002911 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002912 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002913 /* File system mounted not to panic on error
2914 * Fix the bitmap and repeat the block allocation
2915 * We leak some of the blocks here.
2916 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002917 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002918 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2919 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002920 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002921 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002922 if (!err)
2923 err = -EAGAIN;
2924 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002925 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002926
2927 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002928#ifdef AGGRESSIVE_CHECK
2929 {
2930 int i;
2931 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2932 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2933 bitmap_bh->b_data));
2934 }
2935 }
2936#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002937 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2938 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002939 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2940 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002941 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002942 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002943 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002944 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002945 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2946 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002947 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002948 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002949
2950 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002951 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002952 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002953 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002954 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002955 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2956 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002957 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2958 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002959
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002960 if (sbi->s_log_groups_per_flex) {
2961 ext4_group_t flex_group = ext4_flex_group(sbi,
2962 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002963 atomic64_sub(ac->ac_b_ex.fe_len,
2964 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002965 }
2966
Frank Mayhar03901312009-01-07 00:06:22 -05002967 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002968 if (err)
2969 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002970 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002971
2972out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002973 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002974 return err;
2975}
2976
2977/*
2978 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002979 * Group request are normalized to s_mb_group_prealloc, which goes to
2980 * s_strip if we set the same via mount option.
2981 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002982 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002983 *
2984 * XXX: should we try to preallocate more than the group has now?
2985 */
2986static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2987{
2988 struct super_block *sb = ac->ac_sb;
2989 struct ext4_locality_group *lg = ac->ac_lg;
2990
2991 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002992 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002993 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05002994 current->pid, ac->ac_g_ex.fe_len);
2995}
2996
2997/*
2998 * Normalization means making request better in terms of
2999 * size and alignment
3000 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003001static noinline_for_stack void
3002ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05003003 struct ext4_allocation_request *ar)
3004{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003005 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003006 int bsbits, max;
3007 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05003008 loff_t size, start_off;
3009 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04003010 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003011 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003012 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05003013
3014 /* do normalize only data requests, metadata requests
3015 do not need preallocation */
3016 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3017 return;
3018
3019 /* sometime caller may want exact blocks */
3020 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3021 return;
3022
3023 /* caller may indicate that preallocation isn't
3024 * required (it's a tail, for example) */
3025 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3026 return;
3027
3028 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3029 ext4_mb_normalize_group_request(ac);
3030 return ;
3031 }
3032
3033 bsbits = ac->ac_sb->s_blocksize_bits;
3034
3035 /* first, let's learn actual file size
3036 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003037 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003038 size = size << bsbits;
3039 if (size < i_size_read(ac->ac_inode))
3040 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003041 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003042
Valerie Clement19304792008-05-13 19:31:14 -04003043 /* max size of free chunks */
3044 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003045
Valerie Clement19304792008-05-13 19:31:14 -04003046#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3047 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003048
3049 /* first, try to predict filesize */
3050 /* XXX: should this table be tunable? */
3051 start_off = 0;
3052 if (size <= 16 * 1024) {
3053 size = 16 * 1024;
3054 } else if (size <= 32 * 1024) {
3055 size = 32 * 1024;
3056 } else if (size <= 64 * 1024) {
3057 size = 64 * 1024;
3058 } else if (size <= 128 * 1024) {
3059 size = 128 * 1024;
3060 } else if (size <= 256 * 1024) {
3061 size = 256 * 1024;
3062 } else if (size <= 512 * 1024) {
3063 size = 512 * 1024;
3064 } else if (size <= 1024 * 1024) {
3065 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003066 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003067 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003068 (21 - bsbits)) << 21;
3069 size = 2 * 1024 * 1024;
3070 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003071 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3072 (22 - bsbits)) << 22;
3073 size = 4 * 1024 * 1024;
3074 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003075 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003076 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3077 (23 - bsbits)) << 23;
3078 size = 8 * 1024 * 1024;
3079 } else {
Xiaoguang Wangb27b1532014-07-27 22:26:36 -04003080 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3081 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3082 ac->ac_o_ex.fe_len) << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003083 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003084 size = size >> bsbits;
3085 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003086
3087 /* don't cover already allocated blocks in selected range */
3088 if (ar->pleft && start <= ar->lleft) {
3089 size -= ar->lleft + 1 - start;
3090 start = ar->lleft + 1;
3091 }
3092 if (ar->pright && start + size - 1 >= ar->lright)
3093 size -= start + size - ar->lright;
3094
3095 end = start + size;
3096
3097 /* check we don't cross already preallocated blocks */
3098 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003099 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003100 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003101
Alex Tomasc9de5602008-01-29 00:19:52 -05003102 if (pa->pa_deleted)
3103 continue;
3104 spin_lock(&pa->pa_lock);
3105 if (pa->pa_deleted) {
3106 spin_unlock(&pa->pa_lock);
3107 continue;
3108 }
3109
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003110 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3111 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003112
3113 /* PA must not overlap original request */
3114 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3115 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3116
Eric Sandeen38877f42009-08-17 23:55:24 -04003117 /* skip PAs this normalized request doesn't overlap with */
3118 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003119 spin_unlock(&pa->pa_lock);
3120 continue;
3121 }
3122 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3123
Eric Sandeen38877f42009-08-17 23:55:24 -04003124 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003125 if (pa_end <= ac->ac_o_ex.fe_logical) {
3126 BUG_ON(pa_end < start);
3127 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003128 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003129 BUG_ON(pa->pa_lstart > end);
3130 end = pa->pa_lstart;
3131 }
3132 spin_unlock(&pa->pa_lock);
3133 }
3134 rcu_read_unlock();
3135 size = end - start;
3136
3137 /* XXX: extra loop to check we really don't overlap preallocations */
3138 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003139 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003140 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003141
Alex Tomasc9de5602008-01-29 00:19:52 -05003142 spin_lock(&pa->pa_lock);
3143 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003144 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3145 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003146 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3147 }
3148 spin_unlock(&pa->pa_lock);
3149 }
3150 rcu_read_unlock();
3151
3152 if (start + size <= ac->ac_o_ex.fe_logical &&
3153 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003154 ext4_msg(ac->ac_sb, KERN_ERR,
3155 "start %lu, size %lu, fe_logical %lu",
3156 (unsigned long) start, (unsigned long) size,
3157 (unsigned long) ac->ac_o_ex.fe_logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05003158 }
3159 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3160 start > ac->ac_o_ex.fe_logical);
Maurizio Lombardib5b60772014-05-27 12:48:56 -04003161 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003162
3163 /* now prepare goal request */
3164
3165 /* XXX: is it better to align blocks WRT to logical
3166 * placement or satisfy big request as is */
3167 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003168 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003169
3170 /* define goal start in order to merge */
3171 if (ar->pright && (ar->lright == (start + size))) {
3172 /* merge to the right */
3173 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3174 &ac->ac_f_ex.fe_group,
3175 &ac->ac_f_ex.fe_start);
3176 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3177 }
3178 if (ar->pleft && (ar->lleft + 1 == start)) {
3179 /* merge to the left */
3180 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3181 &ac->ac_f_ex.fe_group,
3182 &ac->ac_f_ex.fe_start);
3183 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3184 }
3185
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003186 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003187 (unsigned) orig_size, (unsigned) start);
3188}
3189
3190static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3191{
3192 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3193
3194 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3195 atomic_inc(&sbi->s_bal_reqs);
3196 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003197 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003198 atomic_inc(&sbi->s_bal_success);
3199 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3200 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3201 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3202 atomic_inc(&sbi->s_bal_goals);
3203 if (ac->ac_found > sbi->s_mb_max_to_scan)
3204 atomic_inc(&sbi->s_bal_breaks);
3205 }
3206
Theodore Ts'o296c3552009-09-30 00:32:42 -04003207 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3208 trace_ext4_mballoc_alloc(ac);
3209 else
3210 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003211}
3212
3213/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003214 * Called on failure; free up any blocks from the inode PA for this
3215 * context. We don't need this for MB_GROUP_PA because we only change
3216 * pa_free in ext4_mb_release_context(), but on failure, we've already
3217 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3218 */
3219static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3220{
3221 struct ext4_prealloc_space *pa = ac->ac_pa;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003222 struct ext4_buddy e4b;
3223 int err;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003224
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003225 if (pa == NULL) {
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003226 if (ac->ac_f_ex.fe_len == 0)
3227 return;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003228 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3229 if (err) {
3230 /*
3231 * This should never happen since we pin the
3232 * pages in the ext4_allocation_context so
3233 * ext4_mb_load_buddy() should never fail.
3234 */
3235 WARN(1, "mb_load_buddy failed (%d)", err);
3236 return;
3237 }
3238 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3239 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3240 ac->ac_f_ex.fe_len);
3241 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003242 ext4_mb_unload_buddy(&e4b);
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003243 return;
3244 }
3245 if (pa->pa_type == MB_INODE_PA)
Zheng Liu400db9d2012-05-28 17:53:53 -04003246 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003247}
3248
3249/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003250 * use blocks preallocated to inode
3251 */
3252static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3253 struct ext4_prealloc_space *pa)
3254{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003255 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003256 ext4_fsblk_t start;
3257 ext4_fsblk_t end;
3258 int len;
3259
3260 /* found preallocated blocks, use them */
3261 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003262 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3263 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3264 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003265 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3266 &ac->ac_b_ex.fe_start);
3267 ac->ac_b_ex.fe_len = len;
3268 ac->ac_status = AC_STATUS_FOUND;
3269 ac->ac_pa = pa;
3270
3271 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003272 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003273 BUG_ON(pa->pa_free < len);
3274 pa->pa_free -= len;
3275
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003276 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003277}
3278
3279/*
3280 * use blocks preallocated to locality group
3281 */
3282static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3283 struct ext4_prealloc_space *pa)
3284{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003285 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003286
Alex Tomasc9de5602008-01-29 00:19:52 -05003287 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3288 &ac->ac_b_ex.fe_group,
3289 &ac->ac_b_ex.fe_start);
3290 ac->ac_b_ex.fe_len = len;
3291 ac->ac_status = AC_STATUS_FOUND;
3292 ac->ac_pa = pa;
3293
3294 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003295 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003296 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003297 * in on-disk bitmap -- see ext4_mb_release_context()
3298 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003299 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003300 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003301}
3302
3303/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003304 * Return the prealloc space that have minimal distance
3305 * from the goal block. @cpa is the prealloc
3306 * space that is having currently known minimal distance
3307 * from the goal block.
3308 */
3309static struct ext4_prealloc_space *
3310ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3311 struct ext4_prealloc_space *pa,
3312 struct ext4_prealloc_space *cpa)
3313{
3314 ext4_fsblk_t cur_distance, new_distance;
3315
3316 if (cpa == NULL) {
3317 atomic_inc(&pa->pa_count);
3318 return pa;
3319 }
3320 cur_distance = abs(goal_block - cpa->pa_pstart);
3321 new_distance = abs(goal_block - pa->pa_pstart);
3322
Coly Li5a54b2f2011-02-24 14:10:05 -05003323 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003324 return cpa;
3325
3326 /* drop the previous reference */
3327 atomic_dec(&cpa->pa_count);
3328 atomic_inc(&pa->pa_count);
3329 return pa;
3330}
3331
3332/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003333 * search goal blocks in preallocated space
3334 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003335static noinline_for_stack int
3336ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003337{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003338 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003339 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003340 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3341 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003342 struct ext4_prealloc_space *pa, *cpa = NULL;
3343 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003344
3345 /* only data can be preallocated */
3346 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3347 return 0;
3348
3349 /* first, try per-file preallocation */
3350 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003351 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003352
3353 /* all fields in this condition don't change,
3354 * so we can skip locking for them */
3355 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003356 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3357 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003358 continue;
3359
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003360 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003361 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003362 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3363 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003364 continue;
3365
Alex Tomasc9de5602008-01-29 00:19:52 -05003366 /* found preallocated blocks, use them */
3367 spin_lock(&pa->pa_lock);
3368 if (pa->pa_deleted == 0 && pa->pa_free) {
3369 atomic_inc(&pa->pa_count);
3370 ext4_mb_use_inode_pa(ac, pa);
3371 spin_unlock(&pa->pa_lock);
3372 ac->ac_criteria = 10;
3373 rcu_read_unlock();
3374 return 1;
3375 }
3376 spin_unlock(&pa->pa_lock);
3377 }
3378 rcu_read_unlock();
3379
3380 /* can we use group allocation? */
3381 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3382 return 0;
3383
3384 /* inode may have no locality group for some reason */
3385 lg = ac->ac_lg;
3386 if (lg == NULL)
3387 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003388 order = fls(ac->ac_o_ex.fe_len) - 1;
3389 if (order > PREALLOC_TB_SIZE - 1)
3390 /* The max size of hash table is PREALLOC_TB_SIZE */
3391 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003392
Akinobu Mitabda00de2010-03-03 23:53:25 -05003393 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003394 /*
3395 * search for the prealloc space that is having
3396 * minimal distance from the goal block.
3397 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003398 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3399 rcu_read_lock();
3400 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3401 pa_inode_list) {
3402 spin_lock(&pa->pa_lock);
3403 if (pa->pa_deleted == 0 &&
3404 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003405
3406 cpa = ext4_mb_check_group_pa(goal_block,
3407 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003408 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003409 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003410 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003411 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003412 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003413 if (cpa) {
3414 ext4_mb_use_group_pa(ac, cpa);
3415 ac->ac_criteria = 20;
3416 return 1;
3417 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003418 return 0;
3419}
3420
3421/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003422 * the function goes through all block freed in the group
3423 * but not yet committed and marks them used in in-core bitmap.
3424 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003425 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003426 */
3427static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3428 ext4_group_t group)
3429{
3430 struct rb_node *n;
3431 struct ext4_group_info *grp;
3432 struct ext4_free_data *entry;
3433
3434 grp = ext4_get_group_info(sb, group);
3435 n = rb_first(&(grp->bb_free_root));
3436
3437 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003438 entry = rb_entry(n, struct ext4_free_data, efd_node);
3439 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003440 n = rb_next(n);
3441 }
3442 return;
3443}
3444
3445/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003446 * the function goes through all preallocation in this group and marks them
3447 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003448 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003449 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003450static noinline_for_stack
3451void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003452 ext4_group_t group)
3453{
3454 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3455 struct ext4_prealloc_space *pa;
3456 struct list_head *cur;
3457 ext4_group_t groupnr;
3458 ext4_grpblk_t start;
3459 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003460 int len;
3461
3462 /* all form of preallocation discards first load group,
3463 * so the only competing code is preallocation use.
3464 * we don't need any locking here
3465 * notice we do NOT ignore preallocations with pa_deleted
3466 * otherwise we could leave used blocks available for
3467 * allocation in buddy when concurrent ext4_mb_put_pa()
3468 * is dropping preallocation
3469 */
3470 list_for_each(cur, &grp->bb_prealloc_list) {
3471 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3472 spin_lock(&pa->pa_lock);
3473 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3474 &groupnr, &start);
3475 len = pa->pa_len;
3476 spin_unlock(&pa->pa_lock);
3477 if (unlikely(len == 0))
3478 continue;
3479 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003480 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003481 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003482 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003483 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003484}
3485
3486static void ext4_mb_pa_callback(struct rcu_head *head)
3487{
3488 struct ext4_prealloc_space *pa;
3489 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003490
3491 BUG_ON(atomic_read(&pa->pa_count));
3492 BUG_ON(pa->pa_deleted == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05003493 kmem_cache_free(ext4_pspace_cachep, pa);
3494}
3495
3496/*
3497 * drops a reference to preallocated space descriptor
3498 * if this was the last reference and the space is consumed
3499 */
3500static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3501 struct super_block *sb, struct ext4_prealloc_space *pa)
3502{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003503 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003504 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003505
Alex Tomasc9de5602008-01-29 00:19:52 -05003506 /* in this short window concurrent discard can set pa_deleted */
3507 spin_lock(&pa->pa_lock);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003508 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3509 spin_unlock(&pa->pa_lock);
3510 return;
3511 }
3512
Alex Tomasc9de5602008-01-29 00:19:52 -05003513 if (pa->pa_deleted == 1) {
3514 spin_unlock(&pa->pa_lock);
3515 return;
3516 }
3517
3518 pa->pa_deleted = 1;
3519 spin_unlock(&pa->pa_lock);
3520
Eric Sandeend33a1972009-03-16 23:25:40 -04003521 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003522 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003523 * If doing group-based preallocation, pa_pstart may be in the
3524 * next group when pa is used up
3525 */
3526 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003527 grp_blk--;
3528
Lukas Czernerbd862982013-04-03 23:32:34 -04003529 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003530
3531 /*
3532 * possible race:
3533 *
3534 * P1 (buddy init) P2 (regular allocation)
3535 * find block B in PA
3536 * copy on-disk bitmap to buddy
3537 * mark B in on-disk bitmap
3538 * drop PA from group
3539 * mark all PAs in buddy
3540 *
3541 * thus, P1 initializes buddy with B available. to prevent this
3542 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3543 * against that pair
3544 */
3545 ext4_lock_group(sb, grp);
3546 list_del(&pa->pa_group_list);
3547 ext4_unlock_group(sb, grp);
3548
3549 spin_lock(pa->pa_obj_lock);
3550 list_del_rcu(&pa->pa_inode_list);
3551 spin_unlock(pa->pa_obj_lock);
3552
3553 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3554}
3555
3556/*
3557 * creates new preallocated space for given inode
3558 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003559static noinline_for_stack int
3560ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003561{
3562 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003563 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003564 struct ext4_prealloc_space *pa;
3565 struct ext4_group_info *grp;
3566 struct ext4_inode_info *ei;
3567
3568 /* preallocate only when found space is larger then requested */
3569 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3570 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3571 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3572
3573 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3574 if (pa == NULL)
3575 return -ENOMEM;
3576
3577 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3578 int winl;
3579 int wins;
3580 int win;
3581 int offs;
3582
3583 /* we can't allocate as much as normalizer wants.
3584 * so, found space must get proper lstart
3585 * to cover original request */
3586 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3587 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3588
3589 /* we're limited by original request in that
3590 * logical block must be covered any way
3591 * winl is window we can move our chunk within */
3592 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3593
3594 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003595 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003596
3597 /* the smallest one defines real window */
3598 win = min(winl, wins);
3599
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003600 offs = ac->ac_o_ex.fe_logical %
3601 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003602 if (offs && offs < win)
3603 win = offs;
3604
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003605 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003606 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003607 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3608 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3609 }
3610
3611 /* preallocation can change ac_b_ex, thus we store actually
3612 * allocated blocks for history */
3613 ac->ac_f_ex = ac->ac_b_ex;
3614
3615 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3616 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3617 pa->pa_len = ac->ac_b_ex.fe_len;
3618 pa->pa_free = pa->pa_len;
3619 atomic_set(&pa->pa_count, 1);
3620 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003621 INIT_LIST_HEAD(&pa->pa_inode_list);
3622 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003623 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003624 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003625
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003626 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003627 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003628 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003629
3630 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003631 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003632
3633 ei = EXT4_I(ac->ac_inode);
3634 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3635
3636 pa->pa_obj_lock = &ei->i_prealloc_lock;
3637 pa->pa_inode = ac->ac_inode;
3638
3639 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3640 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3641 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3642
3643 spin_lock(pa->pa_obj_lock);
3644 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3645 spin_unlock(pa->pa_obj_lock);
3646
3647 return 0;
3648}
3649
3650/*
3651 * creates new preallocated space for locality group inodes belongs to
3652 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003653static noinline_for_stack int
3654ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003655{
3656 struct super_block *sb = ac->ac_sb;
3657 struct ext4_locality_group *lg;
3658 struct ext4_prealloc_space *pa;
3659 struct ext4_group_info *grp;
3660
3661 /* preallocate only when found space is larger then requested */
3662 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3663 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3664 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3665
3666 BUG_ON(ext4_pspace_cachep == NULL);
3667 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3668 if (pa == NULL)
3669 return -ENOMEM;
3670
3671 /* preallocation can change ac_b_ex, thus we store actually
3672 * allocated blocks for history */
3673 ac->ac_f_ex = ac->ac_b_ex;
3674
3675 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3676 pa->pa_lstart = pa->pa_pstart;
3677 pa->pa_len = ac->ac_b_ex.fe_len;
3678 pa->pa_free = pa->pa_len;
3679 atomic_set(&pa->pa_count, 1);
3680 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003681 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003682 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003683 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003684 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003685
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003686 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003687 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3688 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003689
3690 ext4_mb_use_group_pa(ac, pa);
3691 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3692
3693 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3694 lg = ac->ac_lg;
3695 BUG_ON(lg == NULL);
3696
3697 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3698 pa->pa_inode = NULL;
3699
3700 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3701 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3702 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3703
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003704 /*
3705 * We will later add the new pa to the right bucket
3706 * after updating the pa_free in ext4_mb_release_context
3707 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003708 return 0;
3709}
3710
3711static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3712{
3713 int err;
3714
3715 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3716 err = ext4_mb_new_group_pa(ac);
3717 else
3718 err = ext4_mb_new_inode_pa(ac);
3719 return err;
3720}
3721
3722/*
3723 * finds all unused blocks in on-disk bitmap, frees them in
3724 * in-core bitmap and buddy.
3725 * @pa must be unlinked from inode and group lists, so that
3726 * nobody else can find/use it.
3727 * the caller MUST hold group/inode locks.
3728 * TODO: optimize the case when there are no in-core structures yet
3729 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003730static noinline_for_stack int
3731ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003732 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003733{
Alex Tomasc9de5602008-01-29 00:19:52 -05003734 struct super_block *sb = e4b->bd_sb;
3735 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003736 unsigned int end;
3737 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003738 ext4_group_t group;
3739 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003740 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003741 int err = 0;
3742 int free = 0;
3743
3744 BUG_ON(pa->pa_deleted == 0);
3745 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003746 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003747 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3748 end = bit + pa->pa_len;
3749
Alex Tomasc9de5602008-01-29 00:19:52 -05003750 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003751 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003752 if (bit >= end)
3753 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003754 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003755 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003756 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3757 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003758 free += next - bit;
3759
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003760 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003761 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3762 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003763 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003764 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3765 bit = next + 1;
3766 }
3767 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003768 ext4_msg(e4b->bd_sb, KERN_CRIT,
3769 "pa %p: logic %lu, phys. %lu, len %lu",
3770 pa, (unsigned long) pa->pa_lstart,
3771 (unsigned long) pa->pa_pstart,
3772 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003773 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003774 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003775 /*
3776 * pa is already deleted so we use the value obtained
3777 * from the bitmap and continue.
3778 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003779 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003780 atomic_add(free, &sbi->s_mb_discarded);
3781
3782 return err;
3783}
3784
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003785static noinline_for_stack int
3786ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003787 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003788{
Alex Tomasc9de5602008-01-29 00:19:52 -05003789 struct super_block *sb = e4b->bd_sb;
3790 ext4_group_t group;
3791 ext4_grpblk_t bit;
3792
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003793 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003794 BUG_ON(pa->pa_deleted == 0);
3795 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3796 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3797 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3798 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003799 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003800
3801 return 0;
3802}
3803
3804/*
3805 * releases all preallocations in given group
3806 *
3807 * first, we need to decide discard policy:
3808 * - when do we discard
3809 * 1) ENOSPC
3810 * - how many do we discard
3811 * 1) how many requested
3812 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003813static noinline_for_stack int
3814ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003815 ext4_group_t group, int needed)
3816{
3817 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3818 struct buffer_head *bitmap_bh = NULL;
3819 struct ext4_prealloc_space *pa, *tmp;
3820 struct list_head list;
3821 struct ext4_buddy e4b;
3822 int err;
3823 int busy = 0;
3824 int free = 0;
3825
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003826 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003827
3828 if (list_empty(&grp->bb_prealloc_list))
3829 return 0;
3830
Theodore Ts'o574ca172008-07-11 19:27:31 -04003831 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003832 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003833 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003834 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003835 }
3836
3837 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003838 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003839 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003840 put_bh(bitmap_bh);
3841 return 0;
3842 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003843
3844 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003845 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003846
Alex Tomasc9de5602008-01-29 00:19:52 -05003847 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003848repeat:
3849 ext4_lock_group(sb, group);
3850 list_for_each_entry_safe(pa, tmp,
3851 &grp->bb_prealloc_list, pa_group_list) {
3852 spin_lock(&pa->pa_lock);
3853 if (atomic_read(&pa->pa_count)) {
3854 spin_unlock(&pa->pa_lock);
3855 busy = 1;
3856 continue;
3857 }
3858 if (pa->pa_deleted) {
3859 spin_unlock(&pa->pa_lock);
3860 continue;
3861 }
3862
3863 /* seems this one can be freed ... */
3864 pa->pa_deleted = 1;
3865
3866 /* we can trust pa_free ... */
3867 free += pa->pa_free;
3868
3869 spin_unlock(&pa->pa_lock);
3870
3871 list_del(&pa->pa_group_list);
3872 list_add(&pa->u.pa_tmp_list, &list);
3873 }
3874
3875 /* if we still need more blocks and some PAs were used, try again */
3876 if (free < needed && busy) {
3877 busy = 0;
3878 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003879 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003880 goto repeat;
3881 }
3882
3883 /* found anything to free? */
3884 if (list_empty(&list)) {
3885 BUG_ON(free != 0);
3886 goto out;
3887 }
3888
3889 /* now free all selected PAs */
3890 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3891
3892 /* remove from object (inode or locality group) */
3893 spin_lock(pa->pa_obj_lock);
3894 list_del_rcu(&pa->pa_inode_list);
3895 spin_unlock(pa->pa_obj_lock);
3896
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003897 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003898 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003899 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003900 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003901
3902 list_del(&pa->u.pa_tmp_list);
3903 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3904 }
3905
3906out:
3907 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003908 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003909 put_bh(bitmap_bh);
3910 return free;
3911}
3912
3913/*
3914 * releases all non-used preallocated blocks for given inode
3915 *
3916 * It's important to discard preallocations under i_data_sem
3917 * We don't want another block to be served from the prealloc
3918 * space when we are discarding the inode prealloc space.
3919 *
3920 * FIXME!! Make sure it is valid at all the call sites
3921 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003922void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003923{
3924 struct ext4_inode_info *ei = EXT4_I(inode);
3925 struct super_block *sb = inode->i_sb;
3926 struct buffer_head *bitmap_bh = NULL;
3927 struct ext4_prealloc_space *pa, *tmp;
3928 ext4_group_t group = 0;
3929 struct list_head list;
3930 struct ext4_buddy e4b;
3931 int err;
3932
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003933 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003934 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3935 return;
3936 }
3937
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003938 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003939 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003940
3941 INIT_LIST_HEAD(&list);
3942
3943repeat:
3944 /* first, collect all pa's in the inode */
3945 spin_lock(&ei->i_prealloc_lock);
3946 while (!list_empty(&ei->i_prealloc_list)) {
3947 pa = list_entry(ei->i_prealloc_list.next,
3948 struct ext4_prealloc_space, pa_inode_list);
3949 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3950 spin_lock(&pa->pa_lock);
3951 if (atomic_read(&pa->pa_count)) {
3952 /* this shouldn't happen often - nobody should
3953 * use preallocation while we're discarding it */
3954 spin_unlock(&pa->pa_lock);
3955 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003956 ext4_msg(sb, KERN_ERR,
3957 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003958 WARN_ON(1);
3959 schedule_timeout_uninterruptible(HZ);
3960 goto repeat;
3961
3962 }
3963 if (pa->pa_deleted == 0) {
3964 pa->pa_deleted = 1;
3965 spin_unlock(&pa->pa_lock);
3966 list_del_rcu(&pa->pa_inode_list);
3967 list_add(&pa->u.pa_tmp_list, &list);
3968 continue;
3969 }
3970
3971 /* someone is deleting pa right now */
3972 spin_unlock(&pa->pa_lock);
3973 spin_unlock(&ei->i_prealloc_lock);
3974
3975 /* we have to wait here because pa_deleted
3976 * doesn't mean pa is already unlinked from
3977 * the list. as we might be called from
3978 * ->clear_inode() the inode will get freed
3979 * and concurrent thread which is unlinking
3980 * pa from inode's list may access already
3981 * freed memory, bad-bad-bad */
3982
3983 /* XXX: if this happens too often, we can
3984 * add a flag to force wait only in case
3985 * of ->clear_inode(), but not in case of
3986 * regular truncate */
3987 schedule_timeout_uninterruptible(HZ);
3988 goto repeat;
3989 }
3990 spin_unlock(&ei->i_prealloc_lock);
3991
3992 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003993 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04003994 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05003995
3996 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003997 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003998 ext4_error(sb, "Error loading buddy information for %u",
3999 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004000 continue;
4001 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004002
Theodore Ts'o574ca172008-07-11 19:27:31 -04004003 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004004 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004005 ext4_error(sb, "Error reading block bitmap for %u",
4006 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04004007 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004008 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05004009 }
4010
4011 ext4_lock_group(sb, group);
4012 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004013 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05004014 ext4_unlock_group(sb, group);
4015
Jing Zhange39e07f2010-05-14 00:00:00 -04004016 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05004017 put_bh(bitmap_bh);
4018
4019 list_del(&pa->u.pa_tmp_list);
4020 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4021 }
4022}
4023
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004024#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05004025static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4026{
4027 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04004028 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05004029
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05004030 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05004031 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04004032 return;
4033
Joe Perches7f6a11e2012-03-19 23:09:43 -04004034 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004035 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04004036 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004037 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004038 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004039 "goal %lu/%lu/%lu@%lu, "
4040 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004041 (unsigned long)ac->ac_o_ex.fe_group,
4042 (unsigned long)ac->ac_o_ex.fe_start,
4043 (unsigned long)ac->ac_o_ex.fe_len,
4044 (unsigned long)ac->ac_o_ex.fe_logical,
4045 (unsigned long)ac->ac_g_ex.fe_group,
4046 (unsigned long)ac->ac_g_ex.fe_start,
4047 (unsigned long)ac->ac_g_ex.fe_len,
4048 (unsigned long)ac->ac_g_ex.fe_logical,
4049 (unsigned long)ac->ac_b_ex.fe_group,
4050 (unsigned long)ac->ac_b_ex.fe_start,
4051 (unsigned long)ac->ac_b_ex.fe_len,
4052 (unsigned long)ac->ac_b_ex.fe_logical,
4053 (int)ac->ac_criteria);
Eric Sandeendc9ddd92014-02-20 13:32:10 -05004054 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004055 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04004056 ngroups = ext4_get_groups_count(sb);
4057 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004058 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4059 struct ext4_prealloc_space *pa;
4060 ext4_grpblk_t start;
4061 struct list_head *cur;
4062 ext4_lock_group(sb, i);
4063 list_for_each(cur, &grp->bb_prealloc_list) {
4064 pa = list_entry(cur, struct ext4_prealloc_space,
4065 pa_group_list);
4066 spin_lock(&pa->pa_lock);
4067 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4068 NULL, &start);
4069 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004070 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4071 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004072 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004073 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004074
4075 if (grp->bb_free == 0)
4076 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004077 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004078 i, grp->bb_free, grp->bb_fragments);
4079 }
4080 printk(KERN_ERR "\n");
4081}
4082#else
4083static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4084{
4085 return;
4086}
4087#endif
4088
4089/*
4090 * We use locality group preallocation for small size file. The size of the
4091 * file is determined by the current size or the resulting size after
4092 * allocation which ever is larger
4093 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004094 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004095 */
4096static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4097{
4098 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4099 int bsbits = ac->ac_sb->s_blocksize_bits;
4100 loff_t size, isize;
4101
4102 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4103 return;
4104
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004105 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4106 return;
4107
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004108 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004109 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4110 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004111
Theodore Ts'o50797482009-09-18 13:34:02 -04004112 if ((size == isize) &&
4113 !ext4_fs_is_busy(sbi) &&
4114 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4115 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4116 return;
4117 }
4118
Robin Dongebbe0272011-10-26 05:14:27 -04004119 if (sbi->s_mb_group_prealloc <= 0) {
4120 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4121 return;
4122 }
4123
Alex Tomasc9de5602008-01-29 00:19:52 -05004124 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004125 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004126 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004127 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004128 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004129 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004130
4131 BUG_ON(ac->ac_lg != NULL);
4132 /*
4133 * locality group prealloc space are per cpu. The reason for having
4134 * per cpu locality group is to reduce the contention between block
4135 * request from multiple CPUs.
4136 */
Christoph Lameterca0c9582009-10-03 19:48:22 +09004137 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004138
4139 /* we're going to use group allocation */
4140 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4141
4142 /* serialize all allocations in the group */
4143 mutex_lock(&ac->ac_lg->lg_mutex);
4144}
4145
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004146static noinline_for_stack int
4147ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004148 struct ext4_allocation_request *ar)
4149{
4150 struct super_block *sb = ar->inode->i_sb;
4151 struct ext4_sb_info *sbi = EXT4_SB(sb);
4152 struct ext4_super_block *es = sbi->s_es;
4153 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004154 unsigned int len;
4155 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004156 ext4_grpblk_t block;
4157
4158 /* we can't allocate > group size */
4159 len = ar->len;
4160
4161 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004162 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4163 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004164
4165 /* start searching from the goal */
4166 goal = ar->goal;
4167 if (goal < le32_to_cpu(es->s_first_data_block) ||
4168 goal >= ext4_blocks_count(es))
4169 goal = le32_to_cpu(es->s_first_data_block);
4170 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4171
4172 /* set up allocation goals */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004173 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05004174 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004175 ac->ac_sb = sb;
4176 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004177 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004178 ac->ac_o_ex.fe_group = group;
4179 ac->ac_o_ex.fe_start = block;
4180 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004181 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004182 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004183
4184 /* we have to define context: we'll we work with a file or
4185 * locality group. this is a policy, actually */
4186 ext4_mb_group_or_file(ac);
4187
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004188 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004189 "left: %u/%u, right %u/%u to %swritable\n",
4190 (unsigned) ar->len, (unsigned) ar->logical,
4191 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4192 (unsigned) ar->lleft, (unsigned) ar->pleft,
4193 (unsigned) ar->lright, (unsigned) ar->pright,
4194 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4195 return 0;
4196
4197}
4198
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004199static noinline_for_stack void
4200ext4_mb_discard_lg_preallocations(struct super_block *sb,
4201 struct ext4_locality_group *lg,
4202 int order, int total_entries)
4203{
4204 ext4_group_t group = 0;
4205 struct ext4_buddy e4b;
4206 struct list_head discard_list;
4207 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004208
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004209 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004210
4211 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004212
4213 spin_lock(&lg->lg_prealloc_lock);
4214 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4215 pa_inode_list) {
4216 spin_lock(&pa->pa_lock);
4217 if (atomic_read(&pa->pa_count)) {
4218 /*
4219 * This is the pa that we just used
4220 * for block allocation. So don't
4221 * free that
4222 */
4223 spin_unlock(&pa->pa_lock);
4224 continue;
4225 }
4226 if (pa->pa_deleted) {
4227 spin_unlock(&pa->pa_lock);
4228 continue;
4229 }
4230 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004231 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004232
4233 /* seems this one can be freed ... */
4234 pa->pa_deleted = 1;
4235 spin_unlock(&pa->pa_lock);
4236
4237 list_del_rcu(&pa->pa_inode_list);
4238 list_add(&pa->u.pa_tmp_list, &discard_list);
4239
4240 total_entries--;
4241 if (total_entries <= 5) {
4242 /*
4243 * we want to keep only 5 entries
4244 * allowing it to grow to 8. This
4245 * mak sure we don't call discard
4246 * soon for this list.
4247 */
4248 break;
4249 }
4250 }
4251 spin_unlock(&lg->lg_prealloc_lock);
4252
4253 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4254
Lukas Czernerbd862982013-04-03 23:32:34 -04004255 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004256 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004257 ext4_error(sb, "Error loading buddy information for %u",
4258 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004259 continue;
4260 }
4261 ext4_lock_group(sb, group);
4262 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004263 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004264 ext4_unlock_group(sb, group);
4265
Jing Zhange39e07f2010-05-14 00:00:00 -04004266 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004267 list_del(&pa->u.pa_tmp_list);
4268 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4269 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004270}
4271
4272/*
4273 * We have incremented pa_count. So it cannot be freed at this
4274 * point. Also we hold lg_mutex. So no parallel allocation is
4275 * possible from this lg. That means pa_free cannot be updated.
4276 *
4277 * A parallel ext4_mb_discard_group_preallocations is possible.
4278 * which can cause the lg_prealloc_list to be updated.
4279 */
4280
4281static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4282{
4283 int order, added = 0, lg_prealloc_count = 1;
4284 struct super_block *sb = ac->ac_sb;
4285 struct ext4_locality_group *lg = ac->ac_lg;
4286 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4287
4288 order = fls(pa->pa_free) - 1;
4289 if (order > PREALLOC_TB_SIZE - 1)
4290 /* The max size of hash table is PREALLOC_TB_SIZE */
4291 order = PREALLOC_TB_SIZE - 1;
4292 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004293 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004294 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4295 pa_inode_list) {
4296 spin_lock(&tmp_pa->pa_lock);
4297 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004298 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004299 continue;
4300 }
4301 if (!added && pa->pa_free < tmp_pa->pa_free) {
4302 /* Add to the tail of the previous entry */
4303 list_add_tail_rcu(&pa->pa_inode_list,
4304 &tmp_pa->pa_inode_list);
4305 added = 1;
4306 /*
4307 * we want to count the total
4308 * number of entries in the list
4309 */
4310 }
4311 spin_unlock(&tmp_pa->pa_lock);
4312 lg_prealloc_count++;
4313 }
4314 if (!added)
4315 list_add_tail_rcu(&pa->pa_inode_list,
4316 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004317 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004318
4319 /* Now trim the list to be not more than 8 elements */
4320 if (lg_prealloc_count > 8) {
4321 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004322 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004323 return;
4324 }
4325 return ;
4326}
4327
Alex Tomasc9de5602008-01-29 00:19:52 -05004328/*
4329 * release all resource we used in allocation
4330 */
4331static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4332{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004333 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004334 struct ext4_prealloc_space *pa = ac->ac_pa;
4335 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004336 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004337 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004338 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004339 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4340 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004341 pa->pa_free -= ac->ac_b_ex.fe_len;
4342 pa->pa_len -= ac->ac_b_ex.fe_len;
4343 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004344 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004345 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004346 if (pa) {
4347 /*
4348 * We want to add the pa to the right bucket.
4349 * Remove it from the list and while adding
4350 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004351 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004352 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004353 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004354 spin_lock(pa->pa_obj_lock);
4355 list_del_rcu(&pa->pa_inode_list);
4356 spin_unlock(pa->pa_obj_lock);
4357 ext4_mb_add_n_trim(ac);
4358 }
4359 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4360 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004361 if (ac->ac_bitmap_page)
4362 page_cache_release(ac->ac_bitmap_page);
4363 if (ac->ac_buddy_page)
4364 page_cache_release(ac->ac_buddy_page);
4365 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4366 mutex_unlock(&ac->ac_lg->lg_mutex);
4367 ext4_mb_collect_stats(ac);
4368 return 0;
4369}
4370
4371static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4372{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004373 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004374 int ret;
4375 int freed = 0;
4376
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004377 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004378 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004379 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4380 freed += ret;
4381 needed -= ret;
4382 }
4383
4384 return freed;
4385}
4386
4387/*
4388 * Main entry point into mballoc to allocate blocks
4389 * it tries to use preallocation first, then falls back
4390 * to usual allocation
4391 */
4392ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004393 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004394{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004395 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004396 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004397 struct ext4_sb_info *sbi;
4398 struct super_block *sb;
4399 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004400 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004401 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004402
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004403 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004404 sb = ar->inode->i_sb;
4405 sbi = EXT4_SB(sb);
4406
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004407 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004408
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004409 /* Allow to use superuser reservation for quota file */
4410 if (IS_NOQUOTA(ar->inode))
4411 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4412
Mingming Cao60e58e02009-01-22 18:13:05 +01004413 /*
4414 * For delayed allocation, we could skip the ENOSPC and
4415 * EDQUOT check, as blocks and quotas have been already
4416 * reserved when data being copied into pagecache.
4417 */
Theodore Ts'of2321092011-01-10 12:12:36 -05004418 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
Mingming Cao60e58e02009-01-22 18:13:05 +01004419 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4420 else {
4421 /* Without delayed allocation we need to verify
4422 * there is enough free blocks to do block allocation
4423 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004424 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004425 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004426 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004427
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004428 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004429 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004430 ar->len = ar->len >> 1;
4431 }
4432 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004433 *errp = -ENOSPC;
4434 return 0;
4435 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004436 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004437 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004438 dquot_alloc_block_nofail(ar->inode,
4439 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004440 } else {
4441 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004442 dquot_alloc_block(ar->inode,
4443 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004444
4445 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4446 ar->len--;
4447 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004448 }
4449 inquota = ar->len;
4450 if (ar->len == 0) {
4451 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004452 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004453 }
Mingming Caod2a17632008-07-14 17:52:37 -04004454 }
Mingming Caod2a17632008-07-14 17:52:37 -04004455
Wei Yongjun85556c92012-09-26 20:43:37 -04004456 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004457 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004458 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004459 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004460 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004461 }
4462
Eric Sandeen256bdb42008-02-10 01:13:33 -05004463 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004464 if (*errp) {
4465 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004466 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004467 }
4468
Eric Sandeen256bdb42008-02-10 01:13:33 -05004469 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4470 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004471 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4472 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004473repeat:
4474 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004475 *errp = ext4_mb_regular_allocator(ac);
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004476 if (*errp)
4477 goto discard_and_exit;
4478
4479 /* as we've just preallocated more space than
4480 * user requested originally, we store allocated
4481 * space in a special descriptor */
4482 if (ac->ac_status == AC_STATUS_FOUND &&
4483 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4484 *errp = ext4_mb_new_preallocation(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004485 if (*errp) {
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004486 discard_and_exit:
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004487 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004488 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004489 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004490 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004491 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004492 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004493 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004494 /*
4495 * drop the reference that we took
4496 * in ext4_mb_use_best_found
4497 */
4498 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004499 ac->ac_b_ex.fe_group = 0;
4500 ac->ac_b_ex.fe_start = 0;
4501 ac->ac_b_ex.fe_len = 0;
4502 ac->ac_status = AC_STATUS_CONTINUE;
4503 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004504 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004505 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004506 goto errout;
4507 } else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004508 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4509 ar->len = ac->ac_b_ex.fe_len;
4510 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004511 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004512 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004513 if (freed)
4514 goto repeat;
4515 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004516 }
4517
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004518errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004519 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004520 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004521 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004522 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004523 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004524 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004525out:
4526 if (ac)
4527 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004528 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004529 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004530 if (!ar->len) {
Theodore Ts'of2321092011-01-10 12:12:36 -05004531 if (!ext4_test_inode_state(ar->inode,
4532 EXT4_STATE_DELALLOC_RESERVED))
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004533 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004534 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004535 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004536 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004537
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004538 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004539
Alex Tomasc9de5602008-01-29 00:19:52 -05004540 return block;
4541}
Alex Tomasc9de5602008-01-29 00:19:52 -05004542
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004543/*
4544 * We can merge two free data extents only if the physical blocks
4545 * are contiguous, AND the extents were freed by the same transaction,
4546 * AND the blocks are associated with the same group.
4547 */
4548static int can_merge(struct ext4_free_data *entry1,
4549 struct ext4_free_data *entry2)
4550{
Bobi Jam18aadd42012-02-20 17:53:02 -05004551 if ((entry1->efd_tid == entry2->efd_tid) &&
4552 (entry1->efd_group == entry2->efd_group) &&
4553 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004554 return 1;
4555 return 0;
4556}
4557
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004558static noinline_for_stack int
4559ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004560 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004561{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004562 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004563 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004564 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004565 struct ext4_group_info *db = e4b->bd_info;
4566 struct super_block *sb = e4b->bd_sb;
4567 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004568 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4569 struct rb_node *parent = NULL, *new_node;
4570
Frank Mayhar03901312009-01-07 00:06:22 -05004571 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004572 BUG_ON(e4b->bd_bitmap_page == NULL);
4573 BUG_ON(e4b->bd_buddy_page == NULL);
4574
Bobi Jam18aadd42012-02-20 17:53:02 -05004575 new_node = &new_entry->efd_node;
4576 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004577
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004578 if (!*n) {
4579 /* first free block exent. We need to
4580 protect buddy cache from being freed,
4581 * otherwise we'll refresh it from
4582 * on-disk bitmap and lose not-yet-available
4583 * blocks */
4584 page_cache_get(e4b->bd_buddy_page);
4585 page_cache_get(e4b->bd_bitmap_page);
4586 }
4587 while (*n) {
4588 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004589 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4590 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004591 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004592 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004593 n = &(*n)->rb_right;
4594 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004595 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004596 ext4_group_first_block_no(sb, group) +
4597 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004598 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004599 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004600 }
4601 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004602
4603 rb_link_node(new_node, parent, n);
4604 rb_insert_color(new_node, &db->bb_free_root);
4605
4606 /* Now try to see the extent can be merged to left and right */
4607 node = rb_prev(new_node);
4608 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004609 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004610 if (can_merge(entry, new_entry) &&
4611 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004612 new_entry->efd_start_cluster = entry->efd_start_cluster;
4613 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004614 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004615 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004616 }
4617 }
4618
4619 node = rb_next(new_node);
4620 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004621 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004622 if (can_merge(new_entry, entry) &&
4623 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004624 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004625 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004626 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004627 }
4628 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004629 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004630 ext4_journal_callback_add(handle, ext4_free_data_callback,
4631 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004632 return 0;
4633}
4634
Theodore Ts'o44338712009-11-22 07:44:56 -05004635/**
4636 * ext4_free_blocks() -- Free given blocks and update quota
4637 * @handle: handle for this transaction
4638 * @inode: inode
4639 * @block: start physical block to free
4640 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004641 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004642 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004643void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004644 struct buffer_head *bh, ext4_fsblk_t block,
4645 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004646{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004647 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004648 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004649 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004650 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004651 ext4_grpblk_t bit;
4652 struct buffer_head *gd_bh;
4653 ext4_group_t block_group;
4654 struct ext4_sb_info *sbi;
4655 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004656 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004657 int err = 0;
4658 int ret;
4659
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004660 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004661 if (bh) {
4662 if (block)
4663 BUG_ON(block != bh->b_blocknr);
4664 else
4665 block = bh->b_blocknr;
4666 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004667
Alex Tomasc9de5602008-01-29 00:19:52 -05004668 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004669 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4670 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004671 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004672 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004673 goto error_return;
4674 }
4675
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004676 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004677 trace_ext4_free_blocks(inode, block, count, flags);
4678
4679 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4680 struct buffer_head *tbh = bh;
4681 int i;
4682
4683 BUG_ON(bh && (count > 1));
4684
4685 for (i = 0; i < count; i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004686 cond_resched();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004687 if (!bh)
4688 tbh = sb_find_get_block(inode->i_sb,
4689 block + i);
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004690 if (!tbh)
Namhyung Kim87783692010-10-27 21:30:11 -04004691 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004692 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004693 inode, tbh, block + i);
4694 }
4695 }
4696
Theodore Ts'o60e66792010-05-17 07:00:00 -04004697 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004698 * We need to make sure we don't reuse the freed block until
4699 * after the transaction is committed, which we can do by
4700 * treating the block as metadata, below. We make an
4701 * exception if the inode is to be written in writeback mode
4702 * since writeback mode has weak data consistency guarantees.
4703 */
4704 if (!ext4_should_writeback_data(inode))
4705 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004706
Theodore Ts'o84130192011-09-09 18:50:51 -04004707 /*
4708 * If the extent to be freed does not begin on a cluster
4709 * boundary, we need to deal with partial clusters at the
4710 * beginning and end of the extent. Normally we will free
4711 * blocks at the beginning or the end unless we are explicitly
4712 * requested to avoid doing so.
4713 */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004714 overflow = EXT4_PBLK_COFF(sbi, block);
Theodore Ts'o84130192011-09-09 18:50:51 -04004715 if (overflow) {
4716 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4717 overflow = sbi->s_cluster_ratio - overflow;
4718 block += overflow;
4719 if (count > overflow)
4720 count -= overflow;
4721 else
4722 return;
4723 } else {
4724 block -= overflow;
4725 count += overflow;
4726 }
4727 }
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004728 overflow = EXT4_LBLK_COFF(sbi, count);
Theodore Ts'o84130192011-09-09 18:50:51 -04004729 if (overflow) {
4730 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4731 if (count > overflow)
4732 count -= overflow;
4733 else
4734 return;
4735 } else
4736 count += sbi->s_cluster_ratio - overflow;
4737 }
4738
Alex Tomasc9de5602008-01-29 00:19:52 -05004739do_more:
4740 overflow = 0;
4741 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4742
Darrick J. Wong163a2032013-08-28 17:35:51 -04004743 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4744 ext4_get_group_info(sb, block_group))))
4745 return;
4746
Alex Tomasc9de5602008-01-29 00:19:52 -05004747 /*
4748 * Check to see if we are freeing blocks across a group
4749 * boundary.
4750 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004751 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4752 overflow = EXT4_C2B(sbi, bit) + count -
4753 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004754 count -= overflow;
4755 }
Lukas Czerner810da242013-03-02 17:18:58 -05004756 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004757 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004758 if (!bitmap_bh) {
4759 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004760 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004761 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004762 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004763 if (!gdp) {
4764 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004765 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004766 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004767
4768 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4769 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4770 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004771 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004772 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004773 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004774
Eric Sandeen12062dd2010-02-15 14:19:27 -05004775 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004776 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004777 /* err = 0. ext4_std_error should be a no op */
4778 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004779 }
4780
4781 BUFFER_TRACE(bitmap_bh, "getting write access");
4782 err = ext4_journal_get_write_access(handle, bitmap_bh);
4783 if (err)
4784 goto error_return;
4785
4786 /*
4787 * We are about to modify some metadata. Call the journal APIs
4788 * to unshare ->b_data if a currently-committing transaction is
4789 * using it
4790 */
4791 BUFFER_TRACE(gd_bh, "get_write_access");
4792 err = ext4_journal_get_write_access(handle, gd_bh);
4793 if (err)
4794 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004795#ifdef AGGRESSIVE_CHECK
4796 {
4797 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004798 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004799 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4800 }
4801#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004802 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004803
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004804 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4805 if (err)
4806 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004807
4808 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004809 struct ext4_free_data *new_entry;
4810 /*
4811 * blocks being freed are metadata. these blocks shouldn't
4812 * be used until this transaction is committed
4813 */
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004814 retry:
Bobi Jam18aadd42012-02-20 17:53:02 -05004815 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004816 if (!new_entry) {
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004817 /*
4818 * We use a retry loop because
4819 * ext4_free_blocks() is not allowed to fail.
4820 */
4821 cond_resched();
4822 congestion_wait(BLK_RW_ASYNC, HZ/50);
4823 goto retry;
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004824 }
Bobi Jam18aadd42012-02-20 17:53:02 -05004825 new_entry->efd_start_cluster = bit;
4826 new_entry->efd_group = block_group;
4827 new_entry->efd_count = count_clusters;
4828 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004829
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004830 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004831 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004832 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004833 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004834 /* need to update group_info->bb_free and bitmap
4835 * with group lock held. generate_buddy look at
4836 * them with group lock_held
4837 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004838 if (test_opt(sb, DISCARD)) {
4839 err = ext4_issue_discard(sb, block_group, bit, count);
4840 if (err && err != -EOPNOTSUPP)
4841 ext4_msg(sb, KERN_WARNING, "discard request in"
4842 " group:%d block:%d count:%lu failed"
4843 " with %d", block_group, bit, count,
4844 err);
Lukas Czerner8f9ff182013-10-30 11:10:52 -04004845 } else
4846 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004847
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004848 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004849 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4850 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004851 }
4852
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004853 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4854 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004855 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004856 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004857 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004858
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004859 if (sbi->s_log_groups_per_flex) {
4860 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004861 atomic64_add(count_clusters,
4862 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004863 }
4864
Theodore Ts'o71d4f7d2014-07-15 06:02:38 -04004865 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Aditya Kali7b415bf2011-09-09 19:04:51 -04004866 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
Jan Kara7d734532013-08-17 09:36:54 -04004867 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4868
4869 ext4_mb_unload_buddy(&e4b);
Aditya Kali7b415bf2011-09-09 19:04:51 -04004870
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004871 /* We dirtied the bitmap block */
4872 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4873 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4874
Alex Tomasc9de5602008-01-29 00:19:52 -05004875 /* And the group descriptor block */
4876 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004877 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004878 if (!err)
4879 err = ret;
4880
4881 if (overflow && !err) {
4882 block += count;
4883 count = overflow;
4884 put_bh(bitmap_bh);
4885 goto do_more;
4886 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004887error_return:
4888 brelse(bitmap_bh);
4889 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004890 return;
4891}
Lukas Czerner7360d172010-10-27 21:30:12 -04004892
4893/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004894 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004895 * @handle: handle to this transaction
4896 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004897 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004898 * @count: number of blocks to free
4899 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004900 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004901 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004902int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004903 ext4_fsblk_t block, unsigned long count)
4904{
4905 struct buffer_head *bitmap_bh = NULL;
4906 struct buffer_head *gd_bh;
4907 ext4_group_t block_group;
4908 ext4_grpblk_t bit;
4909 unsigned int i;
4910 struct ext4_group_desc *desc;
4911 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004912 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004913 int err = 0, ret, blk_free_count;
4914 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004915
4916 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4917
Yongqiang Yang4740b832011-07-26 21:51:08 -04004918 if (count == 0)
4919 return 0;
4920
Amir Goldstein2846e822011-05-09 10:46:41 -04004921 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004922 /*
4923 * Check to see if we are freeing blocks across a group
4924 * boundary.
4925 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004926 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4927 ext4_warning(sb, "too much blocks added to group %u\n",
4928 block_group);
4929 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004930 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004931 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004932
Amir Goldstein2846e822011-05-09 10:46:41 -04004933 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004934 if (!bitmap_bh) {
4935 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004936 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004937 }
4938
Amir Goldstein2846e822011-05-09 10:46:41 -04004939 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004940 if (!desc) {
4941 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004942 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004943 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004944
4945 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4946 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4947 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4948 in_range(block + count - 1, ext4_inode_table(sb, desc),
4949 sbi->s_itb_per_group)) {
4950 ext4_error(sb, "Adding blocks in system zones - "
4951 "Block = %llu, count = %lu",
4952 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004953 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004954 goto error_return;
4955 }
4956
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004957 BUFFER_TRACE(bitmap_bh, "getting write access");
4958 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004959 if (err)
4960 goto error_return;
4961
4962 /*
4963 * We are about to modify some metadata. Call the journal APIs
4964 * to unshare ->b_data if a currently-committing transaction is
4965 * using it
4966 */
4967 BUFFER_TRACE(gd_bh, "get_write_access");
4968 err = ext4_journal_get_write_access(handle, gd_bh);
4969 if (err)
4970 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004971
Amir Goldstein2846e822011-05-09 10:46:41 -04004972 for (i = 0, blocks_freed = 0; i < count; i++) {
4973 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004974 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004975 ext4_error(sb, "bit already cleared for block %llu",
4976 (ext4_fsblk_t)(block + i));
4977 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4978 } else {
4979 blocks_freed++;
4980 }
4981 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004982
4983 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4984 if (err)
4985 goto error_return;
4986
4987 /*
4988 * need to update group_info->bb_free and bitmap
4989 * with group lock held. generate_buddy look at
4990 * them with group lock_held
4991 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004992 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004993 mb_clear_bits(bitmap_bh->b_data, bit, count);
4994 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004995 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4996 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04004997 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004998 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04004999 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04005000 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05005001 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04005002
5003 if (sbi->s_log_groups_per_flex) {
5004 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04005005 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5006 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04005007 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04005008
5009 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04005010
5011 /* We dirtied the bitmap block */
5012 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5013 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5014
5015 /* And the group descriptor block */
5016 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5017 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5018 if (!err)
5019 err = ret;
5020
5021error_return:
5022 brelse(bitmap_bh);
5023 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04005024 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04005025}
5026
5027/**
Lukas Czerner7360d172010-10-27 21:30:12 -04005028 * ext4_trim_extent -- function to TRIM one single free extent in the group
5029 * @sb: super block for the file system
5030 * @start: starting block of the free extent in the alloc. group
5031 * @count: number of blocks to TRIM
5032 * @group: alloc. group we are working with
5033 * @e4b: ext4 buddy for the group
5034 *
5035 * Trim "count" blocks starting at "start" in the "group". To assure that no
5036 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5037 * be called with under the group lock.
5038 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005039static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04005040 ext4_group_t group, struct ext4_buddy *e4b)
jon ernste2cbd582014-04-12 23:01:28 -04005041__releases(bitlock)
5042__acquires(bitlock)
Lukas Czerner7360d172010-10-27 21:30:12 -04005043{
5044 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005045 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005046
Tao Mab3d4c2b2011-07-11 00:01:52 -04005047 trace_ext4_trim_extent(sb, group, start, count);
5048
Lukas Czerner7360d172010-10-27 21:30:12 -04005049 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5050
5051 ex.fe_start = start;
5052 ex.fe_group = group;
5053 ex.fe_len = count;
5054
5055 /*
5056 * Mark blocks used, so no one can reuse them while
5057 * being trimmed.
5058 */
5059 mb_mark_used(e4b, &ex);
5060 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005061 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04005062 ext4_lock_group(sb, group);
5063 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005064 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005065}
5066
5067/**
5068 * ext4_trim_all_free -- function to trim all free space in alloc. group
5069 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04005070 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04005071 * @start: first group block to examine
5072 * @max: last group block to examine
5073 * @minblocks: minimum extent block count
5074 *
5075 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5076 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5077 * the extent.
5078 *
5079 *
5080 * ext4_trim_all_free walks through group's block bitmap searching for free
5081 * extents. When the free extent is found, mark it as used in group buddy
5082 * bitmap. Then issue a TRIM command on this extent and free the extent in
5083 * the group buddy bitmap. This is done until whole group is scanned.
5084 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005085static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005086ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5087 ext4_grpblk_t start, ext4_grpblk_t max,
5088 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005089{
5090 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005091 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005092 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005093 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005094
Tao Mab3d4c2b2011-07-11 00:01:52 -04005095 trace_ext4_trim_all_free(sb, group, start, max);
5096
Lukas Czerner78944082011-05-24 18:16:27 -04005097 ret = ext4_mb_load_buddy(sb, group, &e4b);
5098 if (ret) {
5099 ext4_error(sb, "Error in loading buddy "
5100 "information for %u", group);
5101 return ret;
5102 }
Lukas Czerner78944082011-05-24 18:16:27 -04005103 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005104
5105 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005106 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5107 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5108 goto out;
5109
Lukas Czerner78944082011-05-24 18:16:27 -04005110 start = (e4b.bd_info->bb_first_free > start) ?
5111 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005112
Lukas Czerner913eed832012-03-21 21:22:22 -04005113 while (start <= max) {
5114 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5115 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005116 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005117 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005118
5119 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005120 ret = ext4_trim_extent(sb, start,
5121 next - start, group, &e4b);
5122 if (ret && ret != -EOPNOTSUPP)
5123 break;
5124 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005125 count += next - start;
5126 }
Tao Ma169ddc32011-07-11 00:00:07 -04005127 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005128 start = next + 1;
5129
5130 if (fatal_signal_pending(current)) {
5131 count = -ERESTARTSYS;
5132 break;
5133 }
5134
5135 if (need_resched()) {
5136 ext4_unlock_group(sb, group);
5137 cond_resched();
5138 ext4_lock_group(sb, group);
5139 }
5140
Tao Ma169ddc32011-07-11 00:00:07 -04005141 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005142 break;
5143 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005144
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005145 if (!ret) {
5146 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005147 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005148 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005149out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005150 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005151 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005152
5153 ext4_debug("trimmed %d blocks in the group %d\n",
5154 count, group);
5155
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005156 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005157}
5158
5159/**
5160 * ext4_trim_fs() -- trim ioctl handle function
5161 * @sb: superblock for filesystem
5162 * @range: fstrim_range structure
5163 *
5164 * start: First Byte to trim
5165 * len: number of Bytes to trim from start
5166 * minlen: minimum extent length in Bytes
5167 * ext4_trim_fs goes through all allocation groups containing Bytes from
5168 * start to start+len. For each such a group ext4_trim_all_free function
5169 * is invoked to trim all free space.
5170 */
5171int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5172{
Lukas Czerner78944082011-05-24 18:16:27 -04005173 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005174 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005175 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005176 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005177 ext4_fsblk_t first_data_blk =
5178 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005179 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005180 int ret = 0;
5181
5182 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005183 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005184 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5185 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005186
Lukas Czerner5de35e82012-10-22 18:01:19 -04005187 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5188 start >= max_blks ||
5189 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005190 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005191 if (end >= max_blks)
5192 end = max_blks - 1;
5193 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005194 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005195 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005196 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005197
Lukas Czerner913eed832012-03-21 21:22:22 -04005198 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005199 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005200 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005201 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005202 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005203
Lukas Czerner913eed832012-03-21 21:22:22 -04005204 /* end now represents the last cluster to discard in this group */
5205 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005206
5207 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005208 grp = ext4_get_group_info(sb, group);
5209 /* We only do this if the grp has never been initialized */
5210 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5211 ret = ext4_mb_init_group(sb, group);
5212 if (ret)
5213 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005214 }
5215
Tao Ma0ba08512011-03-23 15:48:11 -04005216 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005217 * For all the groups except the last one, last cluster will
5218 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5219 * change it for the last group, note that last_cluster is
5220 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005221 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005222 if (group == last_group)
5223 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005224
Lukas Czerner78944082011-05-24 18:16:27 -04005225 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005226 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005227 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005228 if (cnt < 0) {
5229 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005230 break;
5231 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005232 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005233 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005234
5235 /*
5236 * For every group except the first one, we are sure
5237 * that the first cluster to discard will be cluster #0.
5238 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005239 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005240 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005241
Tao Ma3d56b8d2011-07-11 00:03:38 -04005242 if (!ret)
5243 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5244
Tao Ma22f10452011-07-10 23:52:37 -04005245out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005246 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005247 return ret;
5248}