blob: f6aedf88da437ee324c314bb1020baa51db0423c [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>
Tejun Heo66114ca2015-05-22 17:13:32 -040029#include <linux/backing-dev.h>
Theodore Ts'o9bffad12009-06-17 11:48:11 -040030#include <trace/events/ext4.h>
31
Theodore Ts'oa0b30c12013-02-09 16:28:20 -050032#ifdef CONFIG_EXT4_DEBUG
33ushort ext4_mballoc_debug __read_mostly;
34
35module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
36MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
37#endif
38
Alex Tomasc9de5602008-01-29 00:19:52 -050039/*
40 * MUSTDO:
41 * - test ext4_ext_search_left() and ext4_ext_search_right()
42 * - search for metadata in few groups
43 *
44 * TODO v4:
45 * - normalization should take into account whether file is still open
46 * - discard preallocations if no free space left (policy?)
47 * - don't normalize tails
48 * - quota
49 * - reservation for superuser
50 *
51 * TODO v3:
52 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
53 * - track min/max extents in each group for better group selection
54 * - mb_mark_used() may allocate chunk right after splitting buddy
55 * - tree of groups sorted by number of free blocks
56 * - error handling
57 */
58
59/*
60 * The allocation request involve request for multiple number of blocks
61 * near to the goal(block) value specified.
62 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040063 * During initialization phase of the allocator we decide to use the
64 * group preallocation or inode preallocation depending on the size of
65 * the file. The size of the file could be the resulting file size we
66 * would have after allocation, or the current file size, which ever
67 * is larger. If the size is less than sbi->s_mb_stream_request we
68 * select to use the group preallocation. The default value of
69 * s_mb_stream_request is 16 blocks. This can also be tuned via
70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
71 * terms of number of blocks.
Alex Tomasc9de5602008-01-29 00:19:52 -050072 *
73 * The main motivation for having small file use group preallocation is to
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040074 * ensure that we have small files closer together on the disk.
Alex Tomasc9de5602008-01-29 00:19:52 -050075 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -040076 * First stage the allocator looks at the inode prealloc list,
77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
78 * spaces for this particular inode. The inode prealloc space is
79 * represented as:
Alex Tomasc9de5602008-01-29 00:19:52 -050080 *
81 * pa_lstart -> the logical start block for this prealloc space
82 * pa_pstart -> the physical start block for this prealloc space
Theodore Ts'o53accfa2011-09-09 18:48:51 -040083 * pa_len -> length for this prealloc space (in clusters)
84 * pa_free -> free space available in this prealloc space (in clusters)
Alex Tomasc9de5602008-01-29 00:19:52 -050085 *
86 * The inode preallocation space is used looking at the _logical_ start
87 * block. If only the logical file block falls within the range of prealloc
Tao Macaaf7a22011-07-11 18:42:42 -040088 * space we will consume the particular prealloc space. This makes sure that
89 * we have contiguous physical blocks representing the file blocks
Alex Tomasc9de5602008-01-29 00:19:52 -050090 *
91 * The important thing to be noted in case of inode prealloc space is that
92 * we don't modify the values associated to inode prealloc space except
93 * pa_free.
94 *
95 * If we are not able to find blocks in the inode prealloc space and if we
96 * have the group allocation flag set then we look at the locality group
Tao Macaaf7a22011-07-11 18:42:42 -040097 * prealloc space. These are per CPU prealloc list represented as
Alex Tomasc9de5602008-01-29 00:19:52 -050098 *
99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 *
101 * The reason for having a per cpu locality group is to reduce the contention
102 * between CPUs. It is possible to get scheduled at this point.
103 *
104 * The locality group prealloc space is used looking at whether we have
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300105 * enough free space (pa_free) within the prealloc space.
Alex Tomasc9de5602008-01-29 00:19:52 -0500106 *
107 * If we can't allocate blocks via inode prealloc or/and locality group
108 * prealloc then we look at the buddy cache. The buddy cache is represented
109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
110 * mapped to the buddy and bitmap information regarding different
111 * groups. The buddy information is attached to buddy cache inode so that
112 * we can access them through the page cache. The information regarding
113 * each group is loaded via ext4_mb_load_buddy. The information involve
114 * block bitmap and buddy information. The information are stored in the
115 * inode as:
116 *
117 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500118 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500119 *
120 *
121 * one block each for bitmap and buddy information. So for each group we
122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
123 * blocksize) blocks. So it can have information regarding groups_per_page
124 * which is blocks_per_page/2
125 *
126 * The buddy cache inode is not stored on disk. The inode is thrown
127 * away when the filesystem is unmounted.
128 *
129 * We look for count number of blocks in the buddy cache. If we were able
130 * to locate that many free blocks we return with additional information
131 * regarding rest of the contiguous physical block available
132 *
133 * Before allocating blocks via buddy cache we normalize the request
134 * blocks. This ensure we ask for more blocks that we needed. The extra
135 * blocks that we get after allocation is added to the respective prealloc
136 * list. In case of inode preallocation we follow a list of heuristics
137 * based on file size. This can be found in ext4_mb_normalize_request. If
138 * we are doing a group prealloc we try to normalize the request to
Theodore Ts'o27baebb2011-09-09 19:02:51 -0400139 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
140 * dependent on the cluster size; for non-bigalloc file systems, it is
Alex Tomasc9de5602008-01-29 00:19:52 -0500141 * 512 blocks. This can be tuned via
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
Alex Tomasc9de5602008-01-29 00:19:52 -0500143 * terms of number of blocks. If we have mounted the file system with -O
144 * stripe=<value> option the group prealloc request is normalized to the
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
146 * greater than the default mb_group_prealloc.
Alex Tomasc9de5602008-01-29 00:19:52 -0500147 *
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -0400148 * The regular allocator (using the buddy cache) supports a few tunables.
Alex Tomasc9de5602008-01-29 00:19:52 -0500149 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400150 * /sys/fs/ext4/<partition>/mb_min_to_scan
151 * /sys/fs/ext4/<partition>/mb_max_to_scan
152 * /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -0500153 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400154 * The regular allocator uses buddy scan only if the request len is power of
Alex Tomasc9de5602008-01-29 00:19:52 -0500155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
156 * value of s_mb_order2_reqs can be tuned via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400157 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400159 * stripe size. This should result in better allocation on RAID setups. If
160 * not, we search in the specific group using bitmap for best extents. The
161 * tunable min_to_scan and max_to_scan control the behaviour here.
Alex Tomasc9de5602008-01-29 00:19:52 -0500162 * min_to_scan indicate how long the mballoc __must__ look for a best
Theodore Ts'ob713a5e2009-03-31 09:11:14 -0400163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
Alex Tomasc9de5602008-01-29 00:19:52 -0500164 * best extent in the found extents. Searching for the blocks starts with
165 * the group specified as the goal value in allocation context via
166 * ac_g_ex. Each group is first checked based on the criteria whether it
Tao Macaaf7a22011-07-11 18:42:42 -0400167 * can be used for allocation. ext4_mb_good_group explains how the groups are
Alex Tomasc9de5602008-01-29 00:19:52 -0500168 * checked.
169 *
170 * Both the prealloc space are getting populated as above. So for the first
171 * request we will hit the buddy cache which will result in this prealloc
172 * space getting filled. The prealloc space is then later used for the
173 * subsequent request.
174 */
175
176/*
177 * mballoc operates on the following data:
178 * - on-disk bitmap
179 * - in-core buddy (actually includes buddy and bitmap)
180 * - preallocation descriptors (PAs)
181 *
182 * there are two types of preallocations:
183 * - inode
184 * assiged to specific inode and can be used for this inode only.
185 * it describes part of inode's space preallocated to specific
186 * physical blocks. any block from that preallocated can be used
187 * independent. the descriptor just tracks number of blocks left
188 * unused. so, before taking some block from descriptor, one must
189 * make sure corresponded logical block isn't allocated yet. this
190 * also means that freeing any block within descriptor's range
191 * must discard all preallocated blocks.
192 * - locality group
193 * assigned to specific locality group which does not translate to
194 * permanent set of inodes: inode can join and leave group. space
195 * from this type of preallocation can be used for any inode. thus
196 * it's consumed from the beginning to the end.
197 *
198 * relation between them can be expressed as:
199 * in-core buddy = on-disk bitmap + preallocation descriptors
200 *
201 * this mean blocks mballoc considers used are:
202 * - allocated blocks (persistent)
203 * - preallocated blocks (non-persistent)
204 *
205 * consistency in mballoc world means that at any time a block is either
206 * free or used in ALL structures. notice: "any time" should not be read
207 * literally -- time is discrete and delimited by locks.
208 *
209 * to keep it simple, we don't use block numbers, instead we count number of
210 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 *
212 * all operations can be expressed as:
213 * - init buddy: buddy = on-disk + PAs
214 * - new PA: buddy += N; PA = N
215 * - use inode PA: on-disk += N; PA -= N
216 * - discard inode PA buddy -= on-disk - PA; PA = 0
217 * - use locality group PA on-disk += N; PA -= N
218 * - discard locality group PA buddy -= PA; PA = 0
219 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
220 * is used in real operation because we can't know actual used
221 * bits from PA, only from on-disk bitmap
222 *
223 * if we follow this strict logic, then all operations above should be atomic.
224 * given some of them can block, we'd have to use something like semaphores
225 * killing performance on high-end SMP hardware. let's try to relax it using
226 * the following knowledge:
227 * 1) if buddy is referenced, it's already initialized
228 * 2) while block is used in buddy and the buddy is referenced,
229 * nobody can re-allocate that block
230 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
231 * bit set and PA claims same block, it's OK. IOW, one can set bit in
232 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
233 * block
234 *
235 * so, now we're building a concurrency table:
236 * - init buddy vs.
237 * - new PA
238 * blocks for PA are allocated in the buddy, buddy must be referenced
239 * until PA is linked to allocation group to avoid concurrent buddy init
240 * - use inode PA
241 * we need to make sure that either on-disk bitmap or PA has uptodate data
242 * given (3) we care that PA-=N operation doesn't interfere with init
243 * - discard inode PA
244 * the simplest way would be to have buddy initialized by the discard
245 * - use locality group PA
246 * again PA-=N must be serialized with init
247 * - discard locality group PA
248 * the simplest way would be to have buddy initialized by the discard
249 * - new PA vs.
250 * - use inode PA
251 * i_data_sem serializes them
252 * - discard inode PA
253 * discard process must wait until PA isn't used by another process
254 * - use locality group PA
255 * some mutex should serialize them
256 * - discard locality group PA
257 * discard process must wait until PA isn't used by another process
258 * - use inode PA
259 * - use inode PA
260 * i_data_sem or another mutex should serializes them
261 * - discard inode PA
262 * discard process must wait until PA isn't used by another process
263 * - use locality group PA
264 * nothing wrong here -- they're different PAs covering different blocks
265 * - discard locality group PA
266 * discard process must wait until PA isn't used by another process
267 *
268 * now we're ready to make few consequences:
269 * - PA is referenced and while it is no discard is possible
270 * - PA is referenced until block isn't marked in on-disk bitmap
271 * - PA changes only after on-disk bitmap
272 * - discard must not compete with init. either init is done before
273 * any discard or they're serialized somehow
274 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 *
276 * a special case when we've used PA to emptiness. no need to modify buddy
277 * in this case, but we should care about concurrent init
278 *
279 */
280
281 /*
282 * Logic in few words:
283 *
284 * - allocation:
285 * load group
286 * find blocks
287 * mark bits in on-disk bitmap
288 * release group
289 *
290 * - use preallocation:
291 * find proper PA (per-inode or group)
292 * load group
293 * mark bits in on-disk bitmap
294 * release group
295 * release PA
296 *
297 * - free:
298 * load group
299 * mark bits in on-disk bitmap
300 * release group
301 *
302 * - discard preallocations in group:
303 * mark PAs deleted
304 * move them onto local list
305 * load on-disk bitmap
306 * load group
307 * remove PA from object (inode or locality group)
308 * mark free blocks in-core
309 *
310 * - discard inode's preallocations:
311 */
312
313/*
314 * Locking rules
315 *
316 * Locks:
317 * - bitlock on a group (group)
318 * - object (inode/locality) (object)
319 * - per-pa lock (pa)
320 *
321 * Paths:
322 * - new pa
323 * object
324 * group
325 *
326 * - find and use pa:
327 * pa
328 *
329 * - release consumed pa:
330 * pa
331 * group
332 * object
333 *
334 * - generate in-core bitmap:
335 * group
336 * pa
337 *
338 * - discard all for given object (inode, locality group):
339 * object
340 * pa
341 * group
342 *
343 * - discard all for given group:
344 * group
345 * pa
346 * group
347 * object
348 *
349 */
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500350static struct kmem_cache *ext4_pspace_cachep;
351static struct kmem_cache *ext4_ac_cachep;
Bobi Jam18aadd42012-02-20 17:53:02 -0500352static struct kmem_cache *ext4_free_data_cachep;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400353
354/* We create slab caches for groupinfo data structures based on the
355 * superblock block size. There will be one per mounted filesystem for
356 * each unique s_blocksize_bits */
Eric Sandeen2892c152011-02-12 08:12:18 -0500357#define NR_GRPINFO_CACHES 8
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -0400358static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359
Eric Sandeen2892c152011-02-12 08:12:18 -0500360static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
361 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
362 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
363 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
364};
365
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500366static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367 ext4_group_t group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500368static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369 ext4_group_t group);
Bobi Jam18aadd42012-02-20 17:53:02 -0500370static void ext4_free_data_callback(struct super_block *sb,
371 struct ext4_journal_cb_entry *jce, int rc);
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500372
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500373static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374{
Alex Tomasc9de5602008-01-29 00:19:52 -0500375#if BITS_PER_LONG == 64
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500376 *bit += ((unsigned long) addr & 7UL) << 3;
377 addr = (void *) ((unsigned long) addr & ~7UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500378#elif BITS_PER_LONG == 32
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500379 *bit += ((unsigned long) addr & 3UL) << 3;
380 addr = (void *) ((unsigned long) addr & ~3UL);
Alex Tomasc9de5602008-01-29 00:19:52 -0500381#else
382#error "how many bits you are?!"
383#endif
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500384 return addr;
385}
Alex Tomasc9de5602008-01-29 00:19:52 -0500386
387static inline int mb_test_bit(int bit, void *addr)
388{
389 /*
390 * ext4_test_bit on architecture like powerpc
391 * needs unsigned long aligned address
392 */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500393 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500394 return ext4_test_bit(bit, addr);
395}
396
397static inline void mb_set_bit(int bit, void *addr)
398{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500399 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500400 ext4_set_bit(bit, addr);
401}
402
Alex Tomasc9de5602008-01-29 00:19:52 -0500403static inline void mb_clear_bit(int bit, void *addr)
404{
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500405 addr = mb_correct_addr_and_bit(&bit, addr);
Alex Tomasc9de5602008-01-29 00:19:52 -0500406 ext4_clear_bit(bit, addr);
407}
408
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400409static inline int mb_test_and_clear_bit(int bit, void *addr)
410{
411 addr = mb_correct_addr_and_bit(&bit, addr);
412 return ext4_test_and_clear_bit(bit, addr);
413}
414
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500415static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400417 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500418 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400419 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500420 start += fix;
421
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400422 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
423 if (ret > max)
424 return max;
425 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500426}
427
428static inline int mb_find_next_bit(void *addr, int max, int start)
429{
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400430 int fix = 0, ret, tmpmax;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500431 addr = mb_correct_addr_and_bit(&fix, addr);
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400432 tmpmax = max + fix;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500433 start += fix;
434
Aneesh Kumar K.Ve7dfb242008-07-11 19:27:31 -0400435 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
436 if (ret > max)
437 return max;
438 return ret;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500439}
440
Alex Tomasc9de5602008-01-29 00:19:52 -0500441static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
442{
443 char *bb;
444
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500445 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -0500446 BUG_ON(max == NULL);
447
448 if (order > e4b->bd_blkbits + 1) {
449 *max = 0;
450 return NULL;
451 }
452
453 /* at order 0 we see each particular block */
Coly Li84b775a2011-02-24 12:51:59 -0500454 if (order == 0) {
455 *max = 1 << (e4b->bd_blkbits + 3);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500456 return e4b->bd_bitmap;
Coly Li84b775a2011-02-24 12:51:59 -0500457 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500458
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500459 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
Alex Tomasc9de5602008-01-29 00:19:52 -0500460 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
461
462 return bb;
463}
464
465#ifdef DOUBLE_CHECK
466static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
467 int first, int count)
468{
469 int i;
470 struct super_block *sb = e4b->bd_sb;
471
472 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400474 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500475 for (i = 0; i < count; i++) {
476 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
477 ext4_fsblk_t blocknr;
Akinobu Mita5661bd62010-03-03 23:53:39 -0500478
479 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
Theodore Ts'o53accfa2011-09-09 18:48:51 -0400480 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -0500481 ext4_grp_locked_error(sb, e4b->bd_group,
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400482 inode ? inode->i_ino : 0,
483 blocknr,
484 "freeing block already freed "
485 "(bit %u)",
486 first + i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500487 }
488 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
489 }
490}
491
492static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
493{
494 int i;
495
496 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
497 return;
Vincent Minetbc8e6742009-05-15 08:33:18 -0400498 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -0500499 for (i = 0; i < count; i++) {
500 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
501 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
502 }
503}
504
505static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506{
507 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
508 unsigned char *b1, *b2;
509 int i;
510 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
511 b2 = (unsigned char *) bitmap;
512 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
513 if (b1[i] != b2[i]) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -0400514 ext4_msg(e4b->bd_sb, KERN_ERR,
515 "corruption in group %u "
516 "at byte %u(%u): %x in copy != %x "
517 "on disk/prealloc",
518 e4b->bd_group, i, i * 8, b1[i], b2[i]);
Alex Tomasc9de5602008-01-29 00:19:52 -0500519 BUG();
520 }
521 }
522 }
523}
524
525#else
526static inline void mb_free_blocks_double(struct inode *inode,
527 struct ext4_buddy *e4b, int first, int count)
528{
529 return;
530}
531static inline void mb_mark_used_double(struct ext4_buddy *e4b,
532 int first, int count)
533{
534 return;
535}
536static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
537{
538 return;
539}
540#endif
541
542#ifdef AGGRESSIVE_CHECK
543
544#define MB_CHECK_ASSERT(assert) \
545do { \
546 if (!(assert)) { \
547 printk(KERN_EMERG \
548 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
549 function, file, line, # assert); \
550 BUG(); \
551 } \
552} while (0)
553
554static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
555 const char *function, int line)
556{
557 struct super_block *sb = e4b->bd_sb;
558 int order = e4b->bd_blkbits + 1;
559 int max;
560 int max2;
561 int i;
562 int j;
563 int k;
564 int count;
565 struct ext4_group_info *grp;
566 int fragments = 0;
567 int fstart;
568 struct list_head *cur;
569 void *buddy;
570 void *buddy2;
571
Alex Tomasc9de5602008-01-29 00:19:52 -0500572 {
573 static int mb_check_counter;
574 if (mb_check_counter++ % 100 != 0)
575 return 0;
576 }
577
578 while (order > 1) {
579 buddy = mb_find_buddy(e4b, order, &max);
580 MB_CHECK_ASSERT(buddy);
581 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
582 MB_CHECK_ASSERT(buddy2);
583 MB_CHECK_ASSERT(buddy != buddy2);
584 MB_CHECK_ASSERT(max * 2 == max2);
585
586 count = 0;
587 for (i = 0; i < max; i++) {
588
589 if (mb_test_bit(i, buddy)) {
590 /* only single bit in buddy2 may be 1 */
591 if (!mb_test_bit(i << 1, buddy2)) {
592 MB_CHECK_ASSERT(
593 mb_test_bit((i<<1)+1, buddy2));
594 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
595 MB_CHECK_ASSERT(
596 mb_test_bit(i << 1, buddy2));
597 }
598 continue;
599 }
600
Robin Dong0a10da72011-10-26 08:48:54 -0400601 /* both bits in buddy2 must be 1 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500602 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
603 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
604
605 for (j = 0; j < (1 << order); j++) {
606 k = (i * (1 << order)) + j;
607 MB_CHECK_ASSERT(
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -0500608 !mb_test_bit(k, e4b->bd_bitmap));
Alex Tomasc9de5602008-01-29 00:19:52 -0500609 }
610 count++;
611 }
612 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
613 order--;
614 }
615
616 fstart = -1;
617 buddy = mb_find_buddy(e4b, 0, &max);
618 for (i = 0; i < max; i++) {
619 if (!mb_test_bit(i, buddy)) {
620 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
621 if (fstart == -1) {
622 fragments++;
623 fstart = i;
624 }
625 continue;
626 }
627 fstart = -1;
628 /* check used bits only */
629 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
630 buddy2 = mb_find_buddy(e4b, j, &max2);
631 k = i >> j;
632 MB_CHECK_ASSERT(k < max2);
633 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
634 }
635 }
636 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
637 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
638
639 grp = ext4_get_group_info(sb, e4b->bd_group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500640 list_for_each(cur, &grp->bb_prealloc_list) {
641 ext4_group_t groupnr;
642 struct ext4_prealloc_space *pa;
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400643 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
644 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
Alex Tomasc9de5602008-01-29 00:19:52 -0500645 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -0400646 for (i = 0; i < pa->pa_len; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500647 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
648 }
649 return 0;
650}
651#undef MB_CHECK_ASSERT
652#define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
Harvey Harrison46e665e2008-04-17 10:38:59 -0400653 __FILE__, __func__, __LINE__)
Alex Tomasc9de5602008-01-29 00:19:52 -0500654#else
655#define mb_check_buddy(e4b)
656#endif
657
Coly Li7c786052011-02-24 13:24:25 -0500658/*
659 * Divide blocks started from @first with length @len into
660 * smaller chunks with power of 2 blocks.
661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
662 * then increase bb_counters[] for corresponded chunk size.
663 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500664static void ext4_mb_mark_free_simple(struct super_block *sb,
Eric Sandeena36b4492009-08-25 22:36:45 -0400665 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
Alex Tomasc9de5602008-01-29 00:19:52 -0500666 struct ext4_group_info *grp)
667{
668 struct ext4_sb_info *sbi = EXT4_SB(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400669 ext4_grpblk_t min;
670 ext4_grpblk_t max;
671 ext4_grpblk_t chunk;
Alex Tomasc9de5602008-01-29 00:19:52 -0500672 unsigned short border;
673
Theodore Ts'o7137d7a2011-09-09 18:38:51 -0400674 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
Alex Tomasc9de5602008-01-29 00:19:52 -0500675
676 border = 2 << sb->s_blocksize_bits;
677
678 while (len > 0) {
679 /* find how many blocks can be covered since this position */
680 max = ffs(first | border) - 1;
681
682 /* find how many blocks of power 2 we need to mark */
683 min = fls(len) - 1;
684
685 if (max < min)
686 min = max;
687 chunk = 1 << min;
688
689 /* mark multiblock chunks only */
690 grp->bb_counters[min]++;
691 if (min > 0)
692 mb_clear_bit(first >> min,
693 buddy + sbi->s_mb_offsets[min]);
694
695 len -= chunk;
696 first += chunk;
697 }
698}
699
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400700/*
701 * Cache the order of the largest free extent we have available in this block
702 * group.
703 */
704static void
705mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
706{
707 int i;
708 int bits;
709
710 grp->bb_largest_free_order = -1; /* uninit */
711
712 bits = sb->s_blocksize_bits + 1;
713 for (i = bits; i >= 0; i--) {
714 if (grp->bb_counters[i] > 0) {
715 grp->bb_largest_free_order = i;
716 break;
717 }
718 }
719}
720
Eric Sandeen089ceec2009-07-05 22:17:31 -0400721static noinline_for_stack
722void ext4_mb_generate_buddy(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -0500723 void *buddy, void *bitmap, ext4_group_t group)
724{
725 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
Namjae Jeone43bb4e2014-06-26 10:11:53 -0400726 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -0400727 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
Eric Sandeena36b4492009-08-25 22:36:45 -0400728 ext4_grpblk_t i = 0;
729 ext4_grpblk_t first;
730 ext4_grpblk_t len;
Alex Tomasc9de5602008-01-29 00:19:52 -0500731 unsigned free = 0;
732 unsigned fragments = 0;
733 unsigned long long period = get_cycles();
734
735 /* initialize buddy from bitmap which is aggregation
736 * of on-disk bitmap and preallocations */
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500737 i = mb_find_next_zero_bit(bitmap, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -0500738 grp->bb_first_free = i;
739 while (i < max) {
740 fragments++;
741 first = i;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500742 i = mb_find_next_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500743 len = i - first;
744 free += len;
745 if (len > 1)
746 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
747 else
748 grp->bb_counters[0]++;
749 if (i < max)
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -0500750 i = mb_find_next_zero_bit(bitmap, max, i);
Alex Tomasc9de5602008-01-29 00:19:52 -0500751 }
752 grp->bb_fragments = fragments;
753
754 if (free != grp->bb_free) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400755 ext4_grp_locked_error(sb, group, 0, 0,
Theodore Ts'o94d4c062014-07-05 19:15:50 -0400756 "block bitmap and bg descriptor "
757 "inconsistent: %u vs %u free clusters",
Theodore Ts'oe29136f2010-06-29 12:54:28 -0400758 free, grp->bb_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500759 /*
Darrick J. Wong163a2032013-08-28 17:35:51 -0400760 * If we intend to continue, we consider group descriptor
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -0500761 * corrupt and update bb_free using bitmap value
762 */
Alex Tomasc9de5602008-01-29 00:19:52 -0500763 grp->bb_free = free;
Namjae Jeone43bb4e2014-06-26 10:11:53 -0400764 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
765 percpu_counter_sub(&sbi->s_freeclusters_counter,
766 grp->bb_free);
Darrick J. Wong163a2032013-08-28 17:35:51 -0400767 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
Alex Tomasc9de5602008-01-29 00:19:52 -0500768 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400769 mb_set_largest_free_order(sb, grp);
Alex Tomasc9de5602008-01-29 00:19:52 -0500770
771 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
772
773 period = get_cycles() - period;
774 spin_lock(&EXT4_SB(sb)->s_bal_lock);
775 EXT4_SB(sb)->s_mb_buddies_generated++;
776 EXT4_SB(sb)->s_mb_generation_time += period;
777 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
778}
779
Andrey Sidoroveabe0442013-04-09 12:22:29 -0400780static void mb_regenerate_buddy(struct ext4_buddy *e4b)
781{
782 int count;
783 int order = 1;
784 void *buddy;
785
786 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
787 ext4_set_bits(buddy, 0, count);
788 }
789 e4b->bd_info->bb_fragments = 0;
790 memset(e4b->bd_info->bb_counters, 0,
791 sizeof(*e4b->bd_info->bb_counters) *
792 (e4b->bd_sb->s_blocksize_bits + 2));
793
794 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
795 e4b->bd_bitmap, e4b->bd_group);
796}
797
Alex Tomasc9de5602008-01-29 00:19:52 -0500798/* The buddy information is attached the buddy cache inode
799 * for convenience. The information regarding each group
800 * is loaded via ext4_mb_load_buddy. The information involve
801 * block bitmap and buddy information. The information are
802 * stored in the inode as
803 *
804 * { page }
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -0500805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
Alex Tomasc9de5602008-01-29 00:19:52 -0500806 *
807 *
808 * one block each for bitmap and buddy information.
809 * So for each group we take up 2 blocks. A page can
810 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
811 * So it can have information regarding groups_per_page which
812 * is blocks_per_page/2
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400813 *
814 * Locking note: This routine takes the block group lock of all groups
815 * for this page; do not hold this lock when calling this routine!
Alex Tomasc9de5602008-01-29 00:19:52 -0500816 */
817
818static int ext4_mb_init_cache(struct page *page, char *incore)
819{
Theodore Ts'o8df96752009-05-01 08:50:38 -0400820 ext4_group_t ngroups;
Alex Tomasc9de5602008-01-29 00:19:52 -0500821 int blocksize;
822 int blocks_per_page;
823 int groups_per_page;
824 int err = 0;
825 int i;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500826 ext4_group_t first_group, group;
Alex Tomasc9de5602008-01-29 00:19:52 -0500827 int first_block;
828 struct super_block *sb;
829 struct buffer_head *bhs;
Darrick J. Wongfa77dcf2012-04-29 18:35:10 -0400830 struct buffer_head **bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -0500831 struct inode *inode;
832 char *data;
833 char *bitmap;
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400834 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -0500835
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400836 mb_debug(1, "init page %lu\n", page->index);
Alex Tomasc9de5602008-01-29 00:19:52 -0500837
838 inode = page->mapping->host;
839 sb = inode->i_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400840 ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -0500841 blocksize = 1 << inode->i_blkbits;
842 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
843
844 groups_per_page = blocks_per_page >> 1;
845 if (groups_per_page == 0)
846 groups_per_page = 1;
847
848 /* allocate buffer_heads to read bitmaps */
849 if (groups_per_page > 1) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500850 i = sizeof(struct buffer_head *) * groups_per_page;
851 bh = kzalloc(i, GFP_NOFS);
Theodore Ts'o813e5722012-02-20 17:52:46 -0500852 if (bh == NULL) {
853 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500854 goto out;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500855 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500856 } else
857 bh = &bhs;
858
859 first_group = page->index * blocks_per_page / 2;
860
861 /* read all groups the page covers into the cache */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500862 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
863 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500864 break;
865
Theodore Ts'o813e5722012-02-20 17:52:46 -0500866 grinfo = ext4_get_group_info(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400867 /*
868 * If page is uptodate then we came here after online resize
869 * which added some new uninitialized group info structs, so
870 * we must skip all initialized uptodate buddies on the page,
871 * which may be currently in use by an allocating task.
872 */
873 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
874 bh[i] = NULL;
875 continue;
876 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500877 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
878 err = -ENOMEM;
Alex Tomasc9de5602008-01-29 00:19:52 -0500879 goto out;
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500880 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500881 mb_debug(1, "read bitmap for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500882 }
883
884 /* wait for I/O completion */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500885 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
Lukas Czernerbbdc3222015-06-08 11:38:37 -0400886 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i]))
Theodore Ts'o813e5722012-02-20 17:52:46 -0500887 err = -EIO;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500888 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500889
890 first_block = page->index * blocks_per_page;
891 for (i = 0; i < blocks_per_page; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500892 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400893 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500894 break;
895
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400896 if (!bh[group - first_group])
897 /* skip initialized uptodate buddy */
898 continue;
899
Lukas Czernerbbdc3222015-06-08 11:38:37 -0400900 if (!buffer_verified(bh[group - first_group]))
901 /* Skip faulty bitmaps */
902 continue;
903 err = 0;
904
Alex Tomasc9de5602008-01-29 00:19:52 -0500905 /*
906 * data carry information regarding this
907 * particular group in the format specified
908 * above
909 *
910 */
911 data = page_address(page) + (i * blocksize);
912 bitmap = bh[group - first_group]->b_data;
913
914 /*
915 * We place the buddy block and bitmap block
916 * close together
917 */
918 if ((first_block + i) & 1) {
919 /* this is block of buddy */
920 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400921 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500922 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400923 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500924 grinfo = ext4_get_group_info(sb, group);
925 grinfo->bb_fragments = 0;
926 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400927 sizeof(*grinfo->bb_counters) *
928 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500929 /*
930 * incore got set to the group block bitmap below
931 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500932 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400933 /* init the buddy */
934 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500935 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500936 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500937 incore = NULL;
938 } else {
939 /* this is block of bitmap */
940 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400941 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500942 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400943 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500944
945 /* see comments in ext4_mb_put_pa() */
946 ext4_lock_group(sb, group);
947 memcpy(data, bitmap, blocksize);
948
949 /* mark all preallocated blks used in in-core bitmap */
950 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500951 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500952 ext4_unlock_group(sb, group);
953
954 /* set incore so that the buddy information can be
955 * generated using this
956 */
957 incore = data;
958 }
959 }
960 SetPageUptodate(page);
961
962out:
963 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400964 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500965 brelse(bh[i]);
966 if (bh != &bhs)
967 kfree(bh);
968 }
969 return err;
970}
971
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400972/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400973 * Lock the buddy and bitmap pages. This make sure other parallel init_group
974 * on the same buddy page doesn't happen whild holding the buddy page lock.
975 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
976 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400977 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400978static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
979 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400980{
Amir Goldstein2de88072011-05-09 21:48:13 -0400981 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
982 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400983 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400984 struct page *page;
985
986 e4b->bd_buddy_page = NULL;
987 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400988
989 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
990 /*
991 * the buddy cache inode stores the block bitmap
992 * and buddy information in consecutive blocks.
993 * So for each group we need two blocks.
994 */
995 block = group * 2;
996 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400997 poff = block % blocks_per_page;
998 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
999 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -04001000 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001001 BUG_ON(page->mapping != inode->i_mapping);
1002 e4b->bd_bitmap_page = page;
1003 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001004
Amir Goldstein2de88072011-05-09 21:48:13 -04001005 if (blocks_per_page >= 2) {
1006 /* buddy and bitmap are on the same page */
1007 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001008 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001009
1010 block++;
1011 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001012 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1013 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -04001014 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001015 BUG_ON(page->mapping != inode->i_mapping);
1016 e4b->bd_buddy_page = page;
1017 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001018}
1019
Amir Goldstein2de88072011-05-09 21:48:13 -04001020static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001021{
Amir Goldstein2de88072011-05-09 21:48:13 -04001022 if (e4b->bd_bitmap_page) {
1023 unlock_page(e4b->bd_bitmap_page);
1024 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001025 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001026 if (e4b->bd_buddy_page) {
1027 unlock_page(e4b->bd_buddy_page);
1028 page_cache_release(e4b->bd_buddy_page);
1029 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001030}
1031
1032/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001033 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1034 * block group lock of all groups for this page; do not hold the BG lock when
1035 * calling this routine!
1036 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001037static noinline_for_stack
1038int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1039{
1040
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001041 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001042 struct ext4_buddy e4b;
1043 struct page *page;
1044 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001045
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001046 might_sleep();
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001047 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001048 this_grp = ext4_get_group_info(sb, group);
1049 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001050 * This ensures that we don't reinit the buddy cache
1051 * page which map to the group from which we are already
1052 * allocating. If we are looking at the buddy cache we would
1053 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001054 * would have pinned buddy page to page cache.
Mel Gorman2457aec2014-06-04 16:10:31 -07001055 * The call to ext4_mb_get_buddy_page_lock will mark the
1056 * page accessed.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001057 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001058 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1059 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001060 /*
1061 * somebody initialized the group
1062 * return without doing anything
1063 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001064 goto err;
1065 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001066
1067 page = e4b.bd_bitmap_page;
1068 ret = ext4_mb_init_cache(page, NULL);
1069 if (ret)
1070 goto err;
1071 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001072 ret = -EIO;
1073 goto err;
1074 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001075
Amir Goldstein2de88072011-05-09 21:48:13 -04001076 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001077 /*
1078 * If both the bitmap and buddy are in
1079 * the same page we don't need to force
1080 * init the buddy
1081 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001082 ret = 0;
1083 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001084 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001085 /* init buddy cache */
1086 page = e4b.bd_buddy_page;
1087 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1088 if (ret)
1089 goto err;
1090 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001091 ret = -EIO;
1092 goto err;
1093 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001094err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001095 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001096 return ret;
1097}
1098
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001099/*
1100 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1101 * block group lock of all groups for this page; do not hold the BG lock when
1102 * calling this routine!
1103 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001104static noinline_for_stack int
1105ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1106 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001107{
Alex Tomasc9de5602008-01-29 00:19:52 -05001108 int blocks_per_page;
1109 int block;
1110 int pnum;
1111 int poff;
1112 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001113 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001114 struct ext4_group_info *grp;
1115 struct ext4_sb_info *sbi = EXT4_SB(sb);
1116 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001117
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001118 might_sleep();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001119 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001120
1121 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001122 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001123
1124 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001125 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001126 e4b->bd_sb = sb;
1127 e4b->bd_group = group;
1128 e4b->bd_buddy_page = NULL;
1129 e4b->bd_bitmap_page = NULL;
1130
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001131 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001132 /*
1133 * we need full data about the group
1134 * to make a good selection
1135 */
1136 ret = ext4_mb_init_group(sb, group);
1137 if (ret)
1138 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001139 }
1140
Alex Tomasc9de5602008-01-29 00:19:52 -05001141 /*
1142 * the buddy cache inode stores the block bitmap
1143 * and buddy information in consecutive blocks.
1144 * So for each group we need two blocks.
1145 */
1146 block = group * 2;
1147 pnum = block / blocks_per_page;
1148 poff = block % blocks_per_page;
1149
1150 /* we could use find_or_create_page(), but it locks page
1151 * what we'd like to avoid in fast path ... */
Mel Gorman2457aec2014-06-04 16:10:31 -07001152 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001153 if (page == NULL || !PageUptodate(page)) {
1154 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001155 /*
1156 * drop the page reference and try
1157 * to get the page with lock. If we
1158 * are not uptodate that implies
1159 * somebody just created the page but
1160 * is yet to initialize the same. So
1161 * wait for it to initialize.
1162 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001163 page_cache_release(page);
1164 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1165 if (page) {
1166 BUG_ON(page->mapping != inode->i_mapping);
1167 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001168 ret = ext4_mb_init_cache(page, NULL);
1169 if (ret) {
1170 unlock_page(page);
1171 goto err;
1172 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001173 mb_cmp_bitmaps(e4b, page_address(page) +
1174 (poff * sb->s_blocksize));
1175 }
1176 unlock_page(page);
1177 }
1178 }
Younger Liuc57ab392014-04-10 23:03:43 -04001179 if (page == NULL) {
1180 ret = -ENOMEM;
1181 goto err;
1182 }
1183 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001184 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001185 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001186 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001187
1188 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001189 e4b->bd_bitmap_page = page;
1190 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001191
1192 block++;
1193 pnum = block / blocks_per_page;
1194 poff = block % blocks_per_page;
1195
Mel Gorman2457aec2014-06-04 16:10:31 -07001196 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001197 if (page == NULL || !PageUptodate(page)) {
1198 if (page)
1199 page_cache_release(page);
1200 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1201 if (page) {
1202 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001203 if (!PageUptodate(page)) {
1204 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1205 if (ret) {
1206 unlock_page(page);
1207 goto err;
1208 }
1209 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001210 unlock_page(page);
1211 }
1212 }
Younger Liuc57ab392014-04-10 23:03:43 -04001213 if (page == NULL) {
1214 ret = -ENOMEM;
1215 goto err;
1216 }
1217 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001218 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001219 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001220 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001221
1222 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001223 e4b->bd_buddy_page = page;
1224 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001225
1226 BUG_ON(e4b->bd_bitmap_page == NULL);
1227 BUG_ON(e4b->bd_buddy_page == NULL);
1228
1229 return 0;
1230
1231err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001232 if (page)
1233 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001234 if (e4b->bd_bitmap_page)
1235 page_cache_release(e4b->bd_bitmap_page);
1236 if (e4b->bd_buddy_page)
1237 page_cache_release(e4b->bd_buddy_page);
1238 e4b->bd_buddy = NULL;
1239 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001240 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001241}
1242
Jing Zhange39e07f2010-05-14 00:00:00 -04001243static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001244{
1245 if (e4b->bd_bitmap_page)
1246 page_cache_release(e4b->bd_bitmap_page);
1247 if (e4b->bd_buddy_page)
1248 page_cache_release(e4b->bd_buddy_page);
1249}
1250
1251
1252static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1253{
1254 int order = 1;
1255 void *bb;
1256
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001257 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001258 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1259
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001260 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001261 while (order <= e4b->bd_blkbits + 1) {
1262 block = block >> 1;
1263 if (!mb_test_bit(block, bb)) {
1264 /* this block is part of buddy of order 'order' */
1265 return order;
1266 }
1267 bb += 1 << (e4b->bd_blkbits - order);
1268 order++;
1269 }
1270 return 0;
1271}
1272
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001273static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001274{
1275 __u32 *addr;
1276
1277 len = cur + len;
1278 while (cur < len) {
1279 if ((cur & 31) == 0 && (len - cur) >= 32) {
1280 /* fast path: clear whole word at once */
1281 addr = bm + (cur >> 3);
1282 *addr = 0;
1283 cur += 32;
1284 continue;
1285 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001286 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001287 cur++;
1288 }
1289}
1290
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001291/* clear bits in given range
1292 * will return first found zero bit if any, -1 otherwise
1293 */
1294static int mb_test_and_clear_bits(void *bm, int cur, int len)
1295{
1296 __u32 *addr;
1297 int zero_bit = -1;
1298
1299 len = cur + len;
1300 while (cur < len) {
1301 if ((cur & 31) == 0 && (len - cur) >= 32) {
1302 /* fast path: clear whole word at once */
1303 addr = bm + (cur >> 3);
1304 if (*addr != (__u32)(-1) && zero_bit == -1)
1305 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1306 *addr = 0;
1307 cur += 32;
1308 continue;
1309 }
1310 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1311 zero_bit = cur;
1312 cur++;
1313 }
1314
1315 return zero_bit;
1316}
1317
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001318void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001319{
1320 __u32 *addr;
1321
1322 len = cur + len;
1323 while (cur < len) {
1324 if ((cur & 31) == 0 && (len - cur) >= 32) {
1325 /* fast path: set whole word at once */
1326 addr = bm + (cur >> 3);
1327 *addr = 0xffffffff;
1328 cur += 32;
1329 continue;
1330 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001331 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001332 cur++;
1333 }
1334}
1335
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001336/*
1337 * _________________________________________________________________ */
1338
1339static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001340{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001341 if (mb_test_bit(*bit + side, bitmap)) {
1342 mb_clear_bit(*bit, bitmap);
1343 (*bit) -= side;
1344 return 1;
1345 }
1346 else {
1347 (*bit) += side;
1348 mb_set_bit(*bit, bitmap);
1349 return -1;
1350 }
1351}
1352
1353static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1354{
1355 int max;
1356 int order = 1;
1357 void *buddy = mb_find_buddy(e4b, order, &max);
1358
1359 while (buddy) {
1360 void *buddy2;
1361
1362 /* Bits in range [first; last] are known to be set since
1363 * corresponding blocks were allocated. Bits in range
1364 * (first; last) will stay set because they form buddies on
1365 * upper layer. We just deal with borders if they don't
1366 * align with upper layer and then go up.
1367 * Releasing entire group is all about clearing
1368 * single bit of highest order buddy.
1369 */
1370
1371 /* Example:
1372 * ---------------------------------
1373 * | 1 | 1 | 1 | 1 |
1374 * ---------------------------------
1375 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1376 * ---------------------------------
1377 * 0 1 2 3 4 5 6 7
1378 * \_____________________/
1379 *
1380 * Neither [1] nor [6] is aligned to above layer.
1381 * Left neighbour [0] is free, so mark it busy,
1382 * decrease bb_counters and extend range to
1383 * [0; 6]
1384 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1385 * mark [6] free, increase bb_counters and shrink range to
1386 * [0; 5].
1387 * Then shift range to [0; 2], go up and do the same.
1388 */
1389
1390
1391 if (first & 1)
1392 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1393 if (!(last & 1))
1394 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1395 if (first > last)
1396 break;
1397 order++;
1398
1399 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1400 mb_clear_bits(buddy, first, last - first + 1);
1401 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1402 break;
1403 }
1404 first >>= 1;
1405 last >>= 1;
1406 buddy = buddy2;
1407 }
1408}
1409
1410static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1411 int first, int count)
1412{
1413 int left_is_free = 0;
1414 int right_is_free = 0;
1415 int block;
1416 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001417 struct super_block *sb = e4b->bd_sb;
1418
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04001419 if (WARN_ON(count == 0))
1420 return;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001421 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001422 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Darrick J. Wong163a2032013-08-28 17:35:51 -04001423 /* Don't bother if the block group is corrupt. */
1424 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1425 return;
1426
Alex Tomasc9de5602008-01-29 00:19:52 -05001427 mb_check_buddy(e4b);
1428 mb_free_blocks_double(inode, e4b, first, count);
1429
1430 e4b->bd_info->bb_free += count;
1431 if (first < e4b->bd_info->bb_first_free)
1432 e4b->bd_info->bb_first_free = first;
1433
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001434 /* access memory sequentially: check left neighbour,
1435 * clear range and then check right neighbour
1436 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001437 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001438 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1439 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1440 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1441 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1442
1443 if (unlikely(block != -1)) {
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001444 struct ext4_sb_info *sbi = EXT4_SB(sb);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001445 ext4_fsblk_t blocknr;
1446
1447 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1448 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1449 ext4_grp_locked_error(sb, e4b->bd_group,
1450 inode ? inode->i_ino : 0,
1451 blocknr,
1452 "freeing already freed block "
Darrick J. Wong163a2032013-08-28 17:35:51 -04001453 "(bit %u); block bitmap corrupt.",
1454 block);
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001455 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1456 percpu_counter_sub(&sbi->s_freeclusters_counter,
1457 e4b->bd_info->bb_free);
Darrick J. Wong163a2032013-08-28 17:35:51 -04001458 /* Mark the block group as corrupt. */
1459 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1460 &e4b->bd_info->bb_state);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001461 mb_regenerate_buddy(e4b);
1462 goto done;
1463 }
1464
1465 /* let's maintain fragments counter */
1466 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001467 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001468 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001469 e4b->bd_info->bb_fragments++;
1470
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001471 /* buddy[0] == bd_bitmap is a special case, so handle
1472 * it right away and let mb_buddy_mark_free stay free of
1473 * zero order checks.
1474 * Check if neighbours are to be coaleasced,
1475 * adjust bitmap bb_counters and borders appropriately.
1476 */
1477 if (first & 1) {
1478 first += !left_is_free;
1479 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001480 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001481 if (!(last & 1)) {
1482 last -= !right_is_free;
1483 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1484 }
1485
1486 if (first <= last)
1487 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1488
1489done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001490 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001491 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001492}
1493
Robin Dong15c006a2012-08-17 10:02:17 -04001494static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001495 int needed, struct ext4_free_extent *ex)
1496{
1497 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001498 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001499 void *buddy;
1500
Vincent Minetbc8e6742009-05-15 08:33:18 -04001501 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001502 BUG_ON(ex == NULL);
1503
Robin Dong15c006a2012-08-17 10:02:17 -04001504 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001505 BUG_ON(buddy == NULL);
1506 BUG_ON(block >= max);
1507 if (mb_test_bit(block, buddy)) {
1508 ex->fe_len = 0;
1509 ex->fe_start = 0;
1510 ex->fe_group = 0;
1511 return 0;
1512 }
1513
Robin Dong15c006a2012-08-17 10:02:17 -04001514 /* find actual order */
1515 order = mb_find_order_for_block(e4b, block);
1516 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001517
1518 ex->fe_len = 1 << order;
1519 ex->fe_start = block << order;
1520 ex->fe_group = e4b->bd_group;
1521
1522 /* calc difference from given start */
1523 next = next - ex->fe_start;
1524 ex->fe_len -= next;
1525 ex->fe_start += next;
1526
1527 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001528 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001529
1530 if (block + 1 >= max)
1531 break;
1532
1533 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001534 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001535 break;
1536
Robin Dongb051d8d2011-10-26 05:30:30 -04001537 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001538
Alex Tomasc9de5602008-01-29 00:19:52 -05001539 block = next >> order;
1540 ex->fe_len += 1 << order;
1541 }
1542
1543 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1544 return ex->fe_len;
1545}
1546
1547static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1548{
1549 int ord;
1550 int mlen = 0;
1551 int max = 0;
1552 int cur;
1553 int start = ex->fe_start;
1554 int len = ex->fe_len;
1555 unsigned ret = 0;
1556 int len0 = len;
1557 void *buddy;
1558
1559 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1560 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001561 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001562 mb_check_buddy(e4b);
1563 mb_mark_used_double(e4b, start, len);
1564
1565 e4b->bd_info->bb_free -= len;
1566 if (e4b->bd_info->bb_first_free == start)
1567 e4b->bd_info->bb_first_free += len;
1568
1569 /* let's maintain fragments counter */
1570 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001571 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001572 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001573 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001574 if (mlen && max)
1575 e4b->bd_info->bb_fragments++;
1576 else if (!mlen && !max)
1577 e4b->bd_info->bb_fragments--;
1578
1579 /* let's maintain buddy itself */
1580 while (len) {
1581 ord = mb_find_order_for_block(e4b, start);
1582
1583 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1584 /* the whole chunk may be allocated at once! */
1585 mlen = 1 << ord;
1586 buddy = mb_find_buddy(e4b, ord, &max);
1587 BUG_ON((start >> ord) >= max);
1588 mb_set_bit(start >> ord, buddy);
1589 e4b->bd_info->bb_counters[ord]--;
1590 start += mlen;
1591 len -= mlen;
1592 BUG_ON(len < 0);
1593 continue;
1594 }
1595
1596 /* store for history */
1597 if (ret == 0)
1598 ret = len | (ord << 16);
1599
1600 /* we have to split large buddy */
1601 BUG_ON(ord <= 0);
1602 buddy = mb_find_buddy(e4b, ord, &max);
1603 mb_set_bit(start >> ord, buddy);
1604 e4b->bd_info->bb_counters[ord]--;
1605
1606 ord--;
1607 cur = (start >> ord) & ~1U;
1608 buddy = mb_find_buddy(e4b, ord, &max);
1609 mb_clear_bit(cur, buddy);
1610 mb_clear_bit(cur + 1, buddy);
1611 e4b->bd_info->bb_counters[ord]++;
1612 e4b->bd_info->bb_counters[ord]++;
1613 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001614 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001615
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001616 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001617 mb_check_buddy(e4b);
1618
1619 return ret;
1620}
1621
1622/*
1623 * Must be called under group lock!
1624 */
1625static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1626 struct ext4_buddy *e4b)
1627{
1628 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1629 int ret;
1630
1631 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1632 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1633
1634 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1635 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1636 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1637
1638 /* preallocation can change ac_b_ex, thus we store actually
1639 * allocated blocks for history */
1640 ac->ac_f_ex = ac->ac_b_ex;
1641
1642 ac->ac_status = AC_STATUS_FOUND;
1643 ac->ac_tail = ret & 0xffff;
1644 ac->ac_buddy = ret >> 16;
1645
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001646 /*
1647 * take the page reference. We want the page to be pinned
1648 * so that we don't get a ext4_mb_init_cache_call for this
1649 * group until we update the bitmap. That would mean we
1650 * double allocate blocks. The reference is dropped
1651 * in ext4_mb_release_context
1652 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001653 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1654 get_page(ac->ac_bitmap_page);
1655 ac->ac_buddy_page = e4b->bd_buddy_page;
1656 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001657 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001658 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001659 spin_lock(&sbi->s_md_lock);
1660 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1661 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1662 spin_unlock(&sbi->s_md_lock);
1663 }
1664}
1665
1666/*
1667 * regular allocator, for general purposes allocation
1668 */
1669
1670static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1671 struct ext4_buddy *e4b,
1672 int finish_group)
1673{
1674 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1675 struct ext4_free_extent *bex = &ac->ac_b_ex;
1676 struct ext4_free_extent *gex = &ac->ac_g_ex;
1677 struct ext4_free_extent ex;
1678 int max;
1679
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001680 if (ac->ac_status == AC_STATUS_FOUND)
1681 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001682 /*
1683 * We don't want to scan for a whole year
1684 */
1685 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1686 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1687 ac->ac_status = AC_STATUS_BREAK;
1688 return;
1689 }
1690
1691 /*
1692 * Haven't found good chunk so far, let's continue
1693 */
1694 if (bex->fe_len < gex->fe_len)
1695 return;
1696
1697 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1698 && bex->fe_group == e4b->bd_group) {
1699 /* recheck chunk's availability - we don't know
1700 * when it was found (within this lock-unlock
1701 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001702 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001703 if (max >= gex->fe_len) {
1704 ext4_mb_use_best_found(ac, e4b);
1705 return;
1706 }
1707 }
1708}
1709
1710/*
1711 * The routine checks whether found extent is good enough. If it is,
1712 * then the extent gets marked used and flag is set to the context
1713 * to stop scanning. Otherwise, the extent is compared with the
1714 * previous found extent and if new one is better, then it's stored
1715 * in the context. Later, the best found extent will be used, if
1716 * mballoc can't find good enough extent.
1717 *
1718 * FIXME: real allocation policy is to be designed yet!
1719 */
1720static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1721 struct ext4_free_extent *ex,
1722 struct ext4_buddy *e4b)
1723{
1724 struct ext4_free_extent *bex = &ac->ac_b_ex;
1725 struct ext4_free_extent *gex = &ac->ac_g_ex;
1726
1727 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001728 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1729 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001730 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1731
1732 ac->ac_found++;
1733
1734 /*
1735 * The special case - take what you catch first
1736 */
1737 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1738 *bex = *ex;
1739 ext4_mb_use_best_found(ac, e4b);
1740 return;
1741 }
1742
1743 /*
1744 * Let's check whether the chuck is good enough
1745 */
1746 if (ex->fe_len == gex->fe_len) {
1747 *bex = *ex;
1748 ext4_mb_use_best_found(ac, e4b);
1749 return;
1750 }
1751
1752 /*
1753 * If this is first found extent, just store it in the context
1754 */
1755 if (bex->fe_len == 0) {
1756 *bex = *ex;
1757 return;
1758 }
1759
1760 /*
1761 * If new found extent is better, store it in the context
1762 */
1763 if (bex->fe_len < gex->fe_len) {
1764 /* if the request isn't satisfied, any found extent
1765 * larger than previous best one is better */
1766 if (ex->fe_len > bex->fe_len)
1767 *bex = *ex;
1768 } else if (ex->fe_len > gex->fe_len) {
1769 /* if the request is satisfied, then we try to find
1770 * an extent that still satisfy the request, but is
1771 * smaller than previous one */
1772 if (ex->fe_len < bex->fe_len)
1773 *bex = *ex;
1774 }
1775
1776 ext4_mb_check_limits(ac, e4b, 0);
1777}
1778
Eric Sandeen089ceec2009-07-05 22:17:31 -04001779static noinline_for_stack
1780int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001781 struct ext4_buddy *e4b)
1782{
1783 struct ext4_free_extent ex = ac->ac_b_ex;
1784 ext4_group_t group = ex.fe_group;
1785 int max;
1786 int err;
1787
1788 BUG_ON(ex.fe_len <= 0);
1789 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1790 if (err)
1791 return err;
1792
1793 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001794 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001795
1796 if (max > 0) {
1797 ac->ac_b_ex = ex;
1798 ext4_mb_use_best_found(ac, e4b);
1799 }
1800
1801 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001802 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001803
1804 return 0;
1805}
1806
Eric Sandeen089ceec2009-07-05 22:17:31 -04001807static noinline_for_stack
1808int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001809 struct ext4_buddy *e4b)
1810{
1811 ext4_group_t group = ac->ac_g_ex.fe_group;
1812 int max;
1813 int err;
1814 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001815 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001816 struct ext4_free_extent ex;
1817
1818 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1819 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001820 if (grp->bb_free == 0)
1821 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001822
1823 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1824 if (err)
1825 return err;
1826
Darrick J. Wong163a2032013-08-28 17:35:51 -04001827 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1828 ext4_mb_unload_buddy(e4b);
1829 return 0;
1830 }
1831
Alex Tomasc9de5602008-01-29 00:19:52 -05001832 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001833 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001834 ac->ac_g_ex.fe_len, &ex);
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001835 ex.fe_logical = 0xDEADFA11; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001836
1837 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1838 ext4_fsblk_t start;
1839
Akinobu Mita5661bd62010-03-03 23:53:39 -05001840 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1841 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001842 /* use do_div to get remainder (would be 64-bit modulo) */
1843 if (do_div(start, sbi->s_stripe) == 0) {
1844 ac->ac_found++;
1845 ac->ac_b_ex = ex;
1846 ext4_mb_use_best_found(ac, e4b);
1847 }
1848 } else if (max >= ac->ac_g_ex.fe_len) {
1849 BUG_ON(ex.fe_len <= 0);
1850 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1851 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1852 ac->ac_found++;
1853 ac->ac_b_ex = ex;
1854 ext4_mb_use_best_found(ac, e4b);
1855 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1856 /* Sometimes, caller may want to merge even small
1857 * number of blocks to an existing extent */
1858 BUG_ON(ex.fe_len <= 0);
1859 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1860 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1861 ac->ac_found++;
1862 ac->ac_b_ex = ex;
1863 ext4_mb_use_best_found(ac, e4b);
1864 }
1865 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001866 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001867
1868 return 0;
1869}
1870
1871/*
1872 * The routine scans buddy structures (not bitmap!) from given order
1873 * to max order and tries to find big enough chunk to satisfy the req
1874 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001875static noinline_for_stack
1876void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001877 struct ext4_buddy *e4b)
1878{
1879 struct super_block *sb = ac->ac_sb;
1880 struct ext4_group_info *grp = e4b->bd_info;
1881 void *buddy;
1882 int i;
1883 int k;
1884 int max;
1885
1886 BUG_ON(ac->ac_2order <= 0);
1887 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1888 if (grp->bb_counters[i] == 0)
1889 continue;
1890
1891 buddy = mb_find_buddy(e4b, i, &max);
1892 BUG_ON(buddy == NULL);
1893
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001894 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001895 BUG_ON(k >= max);
1896
1897 ac->ac_found++;
1898
1899 ac->ac_b_ex.fe_len = 1 << i;
1900 ac->ac_b_ex.fe_start = k << i;
1901 ac->ac_b_ex.fe_group = e4b->bd_group;
1902
1903 ext4_mb_use_best_found(ac, e4b);
1904
1905 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1906
1907 if (EXT4_SB(sb)->s_mb_stats)
1908 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1909
1910 break;
1911 }
1912}
1913
1914/*
1915 * The routine scans the group and measures all found extents.
1916 * In order to optimize scanning, caller must pass number of
1917 * free blocks in the group, so the routine can know upper limit.
1918 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001919static noinline_for_stack
1920void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001921 struct ext4_buddy *e4b)
1922{
1923 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001924 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001925 struct ext4_free_extent ex;
1926 int i;
1927 int free;
1928
1929 free = e4b->bd_info->bb_free;
1930 BUG_ON(free <= 0);
1931
1932 i = e4b->bd_info->bb_first_free;
1933
1934 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001935 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001936 EXT4_CLUSTERS_PER_GROUP(sb), i);
1937 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001938 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001939 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001940 * free blocks even though group info says we
1941 * we have free blocks
1942 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001943 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001944 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001945 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001946 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001947 break;
1948 }
1949
Robin Dong15c006a2012-08-17 10:02:17 -04001950 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001951 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001952 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001953 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001954 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001955 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001956 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001957 /*
1958 * The number of free blocks differs. This mostly
1959 * indicate that the bitmap is corrupt. So exit
1960 * without claiming the space.
1961 */
1962 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001963 }
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001964 ex.fe_logical = 0xDEADC0DE; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001965 ext4_mb_measure_extent(ac, &ex, e4b);
1966
1967 i += ex.fe_len;
1968 free -= ex.fe_len;
1969 }
1970
1971 ext4_mb_check_limits(ac, e4b, 1);
1972}
1973
1974/*
1975 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001976 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001977 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001978static noinline_for_stack
1979void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001980 struct ext4_buddy *e4b)
1981{
1982 struct super_block *sb = ac->ac_sb;
1983 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001984 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001985 struct ext4_free_extent ex;
1986 ext4_fsblk_t first_group_block;
1987 ext4_fsblk_t a;
1988 ext4_grpblk_t i;
1989 int max;
1990
1991 BUG_ON(sbi->s_stripe == 0);
1992
1993 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05001994 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1995
Alex Tomasc9de5602008-01-29 00:19:52 -05001996 a = first_group_block + sbi->s_stripe - 1;
1997 do_div(a, sbi->s_stripe);
1998 i = (a * sbi->s_stripe) - first_group_block;
1999
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04002000 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002001 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04002002 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002003 if (max >= sbi->s_stripe) {
2004 ac->ac_found++;
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05002005 ex.fe_logical = 0xDEADF00D; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05002006 ac->ac_b_ex = ex;
2007 ext4_mb_use_best_found(ac, e4b);
2008 break;
2009 }
2010 }
2011 i += sbi->s_stripe;
2012 }
2013}
2014
Lukas Czerner42ac1842015-06-08 11:40:40 -04002015/*
2016 * This is now called BEFORE we load the buddy bitmap.
2017 * Returns either 1 or 0 indicating that the group is either suitable
2018 * for the allocation or not. In addition it can also return negative
2019 * error code when something goes wrong.
2020 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002021static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2022 ext4_group_t group, int cr)
2023{
2024 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04002025 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05002026 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2027
2028 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002029
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04002030 free = grp->bb_free;
2031 if (free == 0)
2032 return 0;
2033 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2034 return 0;
2035
Darrick J. Wong163a2032013-08-28 17:35:51 -04002036 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2037 return 0;
2038
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002039 /* We only do this if the grp has never been initialized */
2040 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2041 int ret = ext4_mb_init_group(ac->ac_sb, group);
2042 if (ret)
Lukas Czerner42ac1842015-06-08 11:40:40 -04002043 return ret;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002044 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002045
Alex Tomasc9de5602008-01-29 00:19:52 -05002046 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05002047 if (fragments == 0)
2048 return 0;
2049
2050 switch (cr) {
2051 case 0:
2052 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002053
Theodore Ts'oa4912122009-03-12 12:18:34 -04002054 /* Avoid using the first bg of a flexgroup for data files */
2055 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2056 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2057 ((group % flex_size) == 0))
2058 return 0;
2059
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002060 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2061 (free / fragments) >= ac->ac_g_ex.fe_len)
2062 return 1;
2063
2064 if (grp->bb_largest_free_order < ac->ac_2order)
2065 return 0;
2066
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002067 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002068 case 1:
2069 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2070 return 1;
2071 break;
2072 case 2:
2073 if (free >= ac->ac_g_ex.fe_len)
2074 return 1;
2075 break;
2076 case 3:
2077 return 1;
2078 default:
2079 BUG();
2080 }
2081
2082 return 0;
2083}
2084
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002085static noinline_for_stack int
2086ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002087{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002088 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002089 int cr;
Lukas Czerner42ac1842015-06-08 11:40:40 -04002090 int err = 0, first_err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002091 struct ext4_sb_info *sbi;
2092 struct super_block *sb;
2093 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002094
2095 sb = ac->ac_sb;
2096 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002097 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002098 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002099 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002100 ngroups = sbi->s_blockfile_groups;
2101
Alex Tomasc9de5602008-01-29 00:19:52 -05002102 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2103
2104 /* first, try the goal */
2105 err = ext4_mb_find_by_goal(ac, &e4b);
2106 if (err || ac->ac_status == AC_STATUS_FOUND)
2107 goto out;
2108
2109 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2110 goto out;
2111
2112 /*
2113 * ac->ac2_order is set only if the fe_len is a power of 2
2114 * if ac2_order is set we also set criteria to 0 so that we
2115 * try exact allocation using buddy.
2116 */
2117 i = fls(ac->ac_g_ex.fe_len);
2118 ac->ac_2order = 0;
2119 /*
2120 * We search using buddy data only if the order of the request
2121 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002122 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002123 */
2124 if (i >= sbi->s_mb_order2_reqs) {
2125 /*
2126 * This should tell if fe_len is exactly power of 2
2127 */
2128 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2129 ac->ac_2order = i - 1;
2130 }
2131
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002132 /* if stream allocation is enabled, use global goal */
2133 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002134 /* TBD: may be hot point */
2135 spin_lock(&sbi->s_md_lock);
2136 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2137 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2138 spin_unlock(&sbi->s_md_lock);
2139 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002140
Alex Tomasc9de5602008-01-29 00:19:52 -05002141 /* Let's just scan groups to find more-less suitable blocks */
2142 cr = ac->ac_2order ? 0 : 1;
2143 /*
2144 * cr == 0 try to get exact allocation,
2145 * cr == 3 try to get anything
2146 */
2147repeat:
2148 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2149 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002150 /*
2151 * searching for the right group start
2152 * from the goal value specified
2153 */
2154 group = ac->ac_g_ex.fe_group;
2155
Theodore Ts'o8df96752009-05-01 08:50:38 -04002156 for (i = 0; i < ngroups; group++, i++) {
Lukas Czerner42ac1842015-06-08 11:40:40 -04002157 int ret = 0;
Theodore Ts'o2ed57242013-06-12 11:43:02 -04002158 cond_resched();
Lachlan McIlroye6155732013-05-05 23:10:00 -04002159 /*
2160 * Artificially restricted ngroups for non-extent
2161 * files makes group > ngroups possible on first loop.
2162 */
2163 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002164 group = 0;
2165
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002166 /* This now checks without needing the buddy page */
Lukas Czerner42ac1842015-06-08 11:40:40 -04002167 ret = ext4_mb_good_group(ac, group, cr);
2168 if (ret <= 0) {
2169 if (!first_err)
2170 first_err = ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002171 continue;
Lukas Czerner42ac1842015-06-08 11:40:40 -04002172 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002173
Alex Tomasc9de5602008-01-29 00:19:52 -05002174 err = ext4_mb_load_buddy(sb, group, &e4b);
2175 if (err)
2176 goto out;
2177
2178 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002179
2180 /*
2181 * We need to check again after locking the
2182 * block group
2183 */
Lukas Czerner42ac1842015-06-08 11:40:40 -04002184 ret = ext4_mb_good_group(ac, group, cr);
2185 if (ret <= 0) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002186 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002187 ext4_mb_unload_buddy(&e4b);
Lukas Czerner42ac1842015-06-08 11:40:40 -04002188 if (!first_err)
2189 first_err = ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002190 continue;
2191 }
2192
2193 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002194 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002195 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002196 else if (cr == 1 && sbi->s_stripe &&
2197 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002198 ext4_mb_scan_aligned(ac, &e4b);
2199 else
2200 ext4_mb_complex_scan_group(ac, &e4b);
2201
2202 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002203 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002204
2205 if (ac->ac_status != AC_STATUS_CONTINUE)
2206 break;
2207 }
2208 }
2209
2210 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2211 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2212 /*
2213 * We've been searching too long. Let's try to allocate
2214 * the best chunk we've found so far
2215 */
2216
2217 ext4_mb_try_best_found(ac, &e4b);
2218 if (ac->ac_status != AC_STATUS_FOUND) {
2219 /*
2220 * Someone more lucky has already allocated it.
2221 * The only thing we can do is just take first
2222 * found block(s)
2223 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2224 */
2225 ac->ac_b_ex.fe_group = 0;
2226 ac->ac_b_ex.fe_start = 0;
2227 ac->ac_b_ex.fe_len = 0;
2228 ac->ac_status = AC_STATUS_CONTINUE;
2229 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2230 cr = 3;
2231 atomic_inc(&sbi->s_mb_lost_chunks);
2232 goto repeat;
2233 }
2234 }
2235out:
Lukas Czerner42ac1842015-06-08 11:40:40 -04002236 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2237 err = first_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002238 return err;
2239}
2240
Alex Tomasc9de5602008-01-29 00:19:52 -05002241static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2242{
2243 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002244 ext4_group_t group;
2245
Theodore Ts'o8df96752009-05-01 08:50:38 -04002246 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002247 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002248 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002249 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002250}
2251
2252static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2253{
2254 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002255 ext4_group_t group;
2256
2257 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002258 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002259 return NULL;
2260 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002261 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002262}
2263
2264static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2265{
2266 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002267 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002268 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002269 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002270 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002271 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002272 struct sg {
2273 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002274 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002275 } sg;
2276
2277 group--;
2278 if (group == 0)
Rasmus Villemoes97b4af22015-06-15 00:32:58 -04002279 seq_puts(seq, "#group: free frags first ["
2280 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2281 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]");
Alex Tomasc9de5602008-01-29 00:19:52 -05002282
2283 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2284 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002285 grinfo = ext4_get_group_info(sb, group);
2286 /* Load the group info in memory only if not already loaded. */
2287 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2288 err = ext4_mb_load_buddy(sb, group, &e4b);
2289 if (err) {
2290 seq_printf(seq, "#%-5u: I/O error\n", group);
2291 return 0;
2292 }
2293 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002294 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002295
Alex Tomasc9de5602008-01-29 00:19:52 -05002296 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002297
2298 if (buddy_loaded)
2299 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002300
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002301 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002302 sg.info.bb_fragments, sg.info.bb_first_free);
2303 for (i = 0; i <= 13; i++)
2304 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2305 sg.info.bb_counters[i] : 0);
2306 seq_printf(seq, " ]\n");
2307
2308 return 0;
2309}
2310
2311static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2312{
2313}
2314
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002315static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002316 .start = ext4_mb_seq_groups_start,
2317 .next = ext4_mb_seq_groups_next,
2318 .stop = ext4_mb_seq_groups_stop,
2319 .show = ext4_mb_seq_groups_show,
2320};
2321
2322static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2323{
Al Virod9dda782013-03-31 18:16:14 -04002324 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002325 int rc;
2326
2327 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2328 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002329 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002330 m->private = sb;
2331 }
2332 return rc;
2333
2334}
2335
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002336static const struct file_operations ext4_mb_seq_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002337 .owner = THIS_MODULE,
2338 .open = ext4_mb_seq_groups_open,
2339 .read = seq_read,
2340 .llseek = seq_lseek,
2341 .release = seq_release,
2342};
2343
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002344static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2345{
2346 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2347 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2348
2349 BUG_ON(!cachep);
2350 return cachep;
2351}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002352
Theodore Ts'o28623c22012-09-05 01:31:50 -04002353/*
2354 * Allocate the top-level s_group_info array for the specified number
2355 * of groups
2356 */
2357int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2358{
2359 struct ext4_sb_info *sbi = EXT4_SB(sb);
2360 unsigned size;
2361 struct ext4_group_info ***new_groupinfo;
2362
2363 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2364 EXT4_DESC_PER_BLOCK_BITS(sb);
2365 if (size <= sbi->s_group_info_size)
2366 return 0;
2367
2368 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2369 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2370 if (!new_groupinfo) {
2371 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2372 return -ENOMEM;
2373 }
2374 if (sbi->s_group_info) {
2375 memcpy(new_groupinfo, sbi->s_group_info,
2376 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
Al Virob93b41d2014-11-20 12:19:11 -05002377 kvfree(sbi->s_group_info);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002378 }
2379 sbi->s_group_info = new_groupinfo;
2380 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2381 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2382 sbi->s_group_info_size);
2383 return 0;
2384}
2385
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002386/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002387int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002388 struct ext4_group_desc *desc)
2389{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002390 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002391 int metalen = 0;
2392 struct ext4_sb_info *sbi = EXT4_SB(sb);
2393 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002394 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002395
2396 /*
2397 * First check if this group is the first of a reserved block.
2398 * If it's true, we have to allocate a new table of pointers
2399 * to ext4_group_info structures
2400 */
2401 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2402 metalen = sizeof(*meta_group_info) <<
2403 EXT4_DESC_PER_BLOCK_BITS(sb);
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002404 meta_group_info = kmalloc(metalen, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002405 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002406 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002407 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002408 goto exit_meta_group_info;
2409 }
2410 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2411 meta_group_info;
2412 }
2413
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002414 meta_group_info =
2415 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2416 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2417
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002418 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002419 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002420 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002421 goto exit_group_info;
2422 }
2423 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2424 &(meta_group_info[i]->bb_state));
2425
2426 /*
2427 * initialize bb_free to be able to skip
2428 * empty groups without initialization
2429 */
2430 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2431 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002432 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002433 } else {
2434 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002435 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002436 }
2437
2438 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002439 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002440 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002441 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002442
2443#ifdef DOUBLE_CHECK
2444 {
2445 struct buffer_head *bh;
2446 meta_group_info[i]->bb_bitmap =
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002447 kmalloc(sb->s_blocksize, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002448 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2449 bh = ext4_read_block_bitmap(sb, group);
2450 BUG_ON(bh == NULL);
2451 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2452 sb->s_blocksize);
2453 put_bh(bh);
2454 }
2455#endif
2456
2457 return 0;
2458
2459exit_group_info:
2460 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002461 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002462 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002463 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2464 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002465exit_meta_group_info:
2466 return -ENOMEM;
2467} /* ext4_mb_add_groupinfo */
2468
Alex Tomasc9de5602008-01-29 00:19:52 -05002469static int ext4_mb_init_backend(struct super_block *sb)
2470{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002471 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002472 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002473 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002474 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002475 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002476 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002477
Theodore Ts'o28623c22012-09-05 01:31:50 -04002478 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2479 if (err)
2480 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002481
Alex Tomasc9de5602008-01-29 00:19:52 -05002482 sbi->s_buddy_cache = new_inode(sb);
2483 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002484 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002485 goto err_freesgi;
2486 }
Yu Jian48e60612011-08-01 17:41:39 -04002487 /* To avoid potentially colliding with an valid on-disk inode number,
2488 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2489 * not in the inode hash, so it should never be found by iget(), but
2490 * this will avoid confusion if it ever shows up during debugging. */
2491 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002492 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002493 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002494 desc = ext4_get_group_desc(sb, i, NULL);
2495 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002496 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002497 goto err_freebuddy;
2498 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002499 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2500 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002501 }
2502
2503 return 0;
2504
2505err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002506 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002507 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002508 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002509 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002510 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002511 kfree(sbi->s_group_info[i]);
2512 iput(sbi->s_buddy_cache);
2513err_freesgi:
Al Virob93b41d2014-11-20 12:19:11 -05002514 kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002515 return -ENOMEM;
2516}
2517
Eric Sandeen2892c152011-02-12 08:12:18 -05002518static void ext4_groupinfo_destroy_slabs(void)
2519{
2520 int i;
2521
2522 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2523 if (ext4_groupinfo_caches[i])
2524 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2525 ext4_groupinfo_caches[i] = NULL;
2526 }
2527}
2528
2529static int ext4_groupinfo_create_slab(size_t size)
2530{
2531 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2532 int slab_size;
2533 int blocksize_bits = order_base_2(size);
2534 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2535 struct kmem_cache *cachep;
2536
2537 if (cache_index >= NR_GRPINFO_CACHES)
2538 return -EINVAL;
2539
2540 if (unlikely(cache_index < 0))
2541 cache_index = 0;
2542
2543 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2544 if (ext4_groupinfo_caches[cache_index]) {
2545 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2546 return 0; /* Already created */
2547 }
2548
2549 slab_size = offsetof(struct ext4_group_info,
2550 bb_counters[blocksize_bits + 2]);
2551
2552 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2553 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2554 NULL);
2555
Tao Ma823ba012011-07-11 18:26:01 -04002556 ext4_groupinfo_caches[cache_index] = cachep;
2557
Eric Sandeen2892c152011-02-12 08:12:18 -05002558 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2559 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002560 printk(KERN_EMERG
2561 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002562 return -ENOMEM;
2563 }
2564
Eric Sandeen2892c152011-02-12 08:12:18 -05002565 return 0;
2566}
2567
Akira Fujita9d990122012-05-28 14:19:25 -04002568int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002569{
2570 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002571 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002572 unsigned offset;
2573 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002574 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002575
Eric Sandeen19278052009-08-25 22:36:25 -04002576 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002577
2578 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2579 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002580 ret = -ENOMEM;
2581 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002582 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002583
Eric Sandeen19278052009-08-25 22:36:25 -04002584 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002585 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2586 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002587 ret = -ENOMEM;
2588 goto out;
2589 }
2590
Eric Sandeen2892c152011-02-12 08:12:18 -05002591 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2592 if (ret < 0)
2593 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002594
2595 /* order 0 is regular bitmap */
2596 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2597 sbi->s_mb_offsets[0] = 0;
2598
2599 i = 1;
2600 offset = 0;
2601 max = sb->s_blocksize << 2;
2602 do {
2603 sbi->s_mb_offsets[i] = offset;
2604 sbi->s_mb_maxs[i] = max;
2605 offset += 1 << (sb->s_blocksize_bits - i);
2606 max = max >> 1;
2607 i++;
2608 } while (i <= sb->s_blocksize_bits + 1);
2609
Alex Tomasc9de5602008-01-29 00:19:52 -05002610 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002611 spin_lock_init(&sbi->s_bal_lock);
2612
2613 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2614 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2615 sbi->s_mb_stats = MB_DEFAULT_STATS;
2616 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2617 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002618 /*
2619 * The default group preallocation is 512, which for 4k block
2620 * sizes translates to 2 megabytes. However for bigalloc file
2621 * systems, this is probably too big (i.e, if the cluster size
2622 * is 1 megabyte, then group preallocation size becomes half a
2623 * gigabyte!). As a default, we will keep a two megabyte
2624 * group pralloc size for cluster sizes up to 64k, and after
2625 * that, we will force a minimum group preallocation size of
2626 * 32 clusters. This translates to 8 megs when the cluster
2627 * size is 256k, and 32 megs when the cluster size is 1 meg,
2628 * which seems reasonable as a default.
2629 */
2630 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2631 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002632 /*
2633 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2634 * to the lowest multiple of s_stripe which is bigger than
2635 * the s_mb_group_prealloc as determined above. We want
2636 * the preallocation size to be an exact multiple of the
2637 * RAID stripe size so that preallocations don't fragment
2638 * the stripes.
2639 */
2640 if (sbi->s_stripe > 1) {
2641 sbi->s_mb_group_prealloc = roundup(
2642 sbi->s_mb_group_prealloc, sbi->s_stripe);
2643 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002644
Eric Sandeen730c2132008-09-13 15:23:29 -04002645 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002646 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002647 ret = -ENOMEM;
Andrey Tsyvarev029b10c2014-05-12 12:34:21 -04002648 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002649 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002650 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002651 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002652 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002653 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002654 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2655 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002656 spin_lock_init(&lg->lg_prealloc_lock);
2657 }
2658
Yu Jian79a77c52011-08-01 17:41:46 -04002659 /* init file for buddy data */
2660 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002661 if (ret != 0)
2662 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002663
Theodore Ts'o296c3552009-09-30 00:32:42 -04002664 if (sbi->s_proc)
2665 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2666 &ext4_mb_seq_groups_fops, sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002667
Tao Ma7aa0bae2011-10-06 10:22:28 -04002668 return 0;
2669
2670out_free_locality_groups:
2671 free_percpu(sbi->s_locality_groups);
2672 sbi->s_locality_groups = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002673out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002674 kfree(sbi->s_mb_offsets);
2675 sbi->s_mb_offsets = NULL;
2676 kfree(sbi->s_mb_maxs);
2677 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002678 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002679}
2680
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002681/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002682static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2683{
2684 struct ext4_prealloc_space *pa;
2685 struct list_head *cur, *tmp;
2686 int count = 0;
2687
2688 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2689 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2690 list_del(&pa->pa_group_list);
2691 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002692 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002693 }
2694 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002695 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002696
2697}
2698
2699int ext4_mb_release(struct super_block *sb)
2700{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002701 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002702 ext4_group_t i;
2703 int num_meta_group_infos;
2704 struct ext4_group_info *grinfo;
2705 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002706 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002707
Salman Qazi95599962012-05-31 23:52:14 -04002708 if (sbi->s_proc)
2709 remove_proc_entry("mb_groups", sbi->s_proc);
2710
Alex Tomasc9de5602008-01-29 00:19:52 -05002711 if (sbi->s_group_info) {
Theodore Ts'o8df96752009-05-01 08:50:38 -04002712 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002713 grinfo = ext4_get_group_info(sb, i);
2714#ifdef DOUBLE_CHECK
2715 kfree(grinfo->bb_bitmap);
2716#endif
2717 ext4_lock_group(sb, i);
2718 ext4_mb_cleanup_pa(grinfo);
2719 ext4_unlock_group(sb, i);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002720 kmem_cache_free(cachep, grinfo);
Alex Tomasc9de5602008-01-29 00:19:52 -05002721 }
Theodore Ts'o8df96752009-05-01 08:50:38 -04002722 num_meta_group_infos = (ngroups +
Alex Tomasc9de5602008-01-29 00:19:52 -05002723 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2724 EXT4_DESC_PER_BLOCK_BITS(sb);
2725 for (i = 0; i < num_meta_group_infos; i++)
2726 kfree(sbi->s_group_info[i]);
Al Virob93b41d2014-11-20 12:19:11 -05002727 kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002728 }
2729 kfree(sbi->s_mb_offsets);
2730 kfree(sbi->s_mb_maxs);
Markus Elfringbfcba2d2014-11-25 20:01:37 -05002731 iput(sbi->s_buddy_cache);
Alex Tomasc9de5602008-01-29 00:19:52 -05002732 if (sbi->s_mb_stats) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002733 ext4_msg(sb, KERN_INFO,
2734 "mballoc: %u blocks %u reqs (%u success)",
Alex Tomasc9de5602008-01-29 00:19:52 -05002735 atomic_read(&sbi->s_bal_allocated),
2736 atomic_read(&sbi->s_bal_reqs),
2737 atomic_read(&sbi->s_bal_success));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002738 ext4_msg(sb, KERN_INFO,
2739 "mballoc: %u extents scanned, %u goal hits, "
2740 "%u 2^N hits, %u breaks, %u lost",
Alex Tomasc9de5602008-01-29 00:19:52 -05002741 atomic_read(&sbi->s_bal_ex_scanned),
2742 atomic_read(&sbi->s_bal_goals),
2743 atomic_read(&sbi->s_bal_2orders),
2744 atomic_read(&sbi->s_bal_breaks),
2745 atomic_read(&sbi->s_mb_lost_chunks));
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002746 ext4_msg(sb, KERN_INFO,
2747 "mballoc: %lu generated and it took %Lu",
Tao Maced156e2011-07-23 16:18:05 -04002748 sbi->s_mb_buddies_generated,
Alex Tomasc9de5602008-01-29 00:19:52 -05002749 sbi->s_mb_generation_time);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002750 ext4_msg(sb, KERN_INFO,
2751 "mballoc: %u preallocated, %u discarded",
Alex Tomasc9de5602008-01-29 00:19:52 -05002752 atomic_read(&sbi->s_mb_preallocated),
2753 atomic_read(&sbi->s_mb_discarded));
2754 }
2755
Eric Sandeen730c2132008-09-13 15:23:29 -04002756 free_percpu(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05002757
2758 return 0;
2759}
2760
Lukas Czerner77ca6cd2010-10-27 21:30:11 -04002761static inline int ext4_issue_discard(struct super_block *sb,
Theodore Ts'o84130192011-09-09 18:50:51 -04002762 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
Jiaying Zhang5c521832010-07-27 11:56:05 -04002763{
Jiaying Zhang5c521832010-07-27 11:56:05 -04002764 ext4_fsblk_t discard_block;
2765
Theodore Ts'o84130192011-09-09 18:50:51 -04002766 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2767 ext4_group_first_block_no(sb, block_group));
2768 count = EXT4_C2B(EXT4_SB(sb), count);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002769 trace_ext4_discard_blocks(sb,
2770 (unsigned long long) discard_block, count);
Lukas Czerner93259632011-01-10 12:09:59 -05002771 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
Jiaying Zhang5c521832010-07-27 11:56:05 -04002772}
2773
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002774/*
2775 * This function is called by the jbd2 layer once the commit has finished,
2776 * so we know we can free the blocks that were released with that commit.
2777 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002778static void ext4_free_data_callback(struct super_block *sb,
2779 struct ext4_journal_cb_entry *jce,
2780 int rc)
Alex Tomasc9de5602008-01-29 00:19:52 -05002781{
Bobi Jam18aadd42012-02-20 17:53:02 -05002782 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
Alex Tomasc9de5602008-01-29 00:19:52 -05002783 struct ext4_buddy e4b;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002784 struct ext4_group_info *db;
Theodore Ts'od9f34502011-04-30 13:47:24 -04002785 int err, count = 0, count2 = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002786
Bobi Jam18aadd42012-02-20 17:53:02 -05002787 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2788 entry->efd_count, entry->efd_group, entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05002789
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05002790 if (test_opt(sb, DISCARD)) {
2791 err = ext4_issue_discard(sb, entry->efd_group,
2792 entry->efd_start_cluster,
2793 entry->efd_count);
2794 if (err && err != -EOPNOTSUPP)
2795 ext4_msg(sb, KERN_WARNING, "discard request in"
2796 " group:%d block:%d count:%d failed"
2797 " with %d", entry->efd_group,
2798 entry->efd_start_cluster,
2799 entry->efd_count, err);
2800 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002801
Bobi Jam18aadd42012-02-20 17:53:02 -05002802 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2803 /* we expect to find existing buddy because it's pinned */
2804 BUG_ON(err != 0);
Theodore Ts'ob90f6872010-04-20 16:51:59 -04002805
Alex Tomasc9de5602008-01-29 00:19:52 -05002806
Bobi Jam18aadd42012-02-20 17:53:02 -05002807 db = e4b.bd_info;
2808 /* there are blocks to put in buddy to make them really free */
2809 count += entry->efd_count;
2810 count2++;
2811 ext4_lock_group(sb, entry->efd_group);
2812 /* Take it out of per group rb tree */
2813 rb_erase(&entry->efd_node, &(db->bb_free_root));
2814 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002815
Bobi Jam18aadd42012-02-20 17:53:02 -05002816 /*
2817 * Clear the trimmed flag for the group so that the next
2818 * ext4_trim_fs can trim it.
2819 * If the volume is mounted with -o discard, online discard
2820 * is supported and the free blocks will be trimmed online.
2821 */
2822 if (!test_opt(sb, DISCARD))
2823 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2824
2825 if (!db->bb_free_root.rb_node) {
2826 /* No more items in the per group rb tree
2827 * balance refcounts from ext4_mb_free_metadata()
Tao Ma3d56b8d2011-07-11 00:03:38 -04002828 */
Bobi Jam18aadd42012-02-20 17:53:02 -05002829 page_cache_release(e4b.bd_buddy_page);
2830 page_cache_release(e4b.bd_bitmap_page);
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04002831 }
Bobi Jam18aadd42012-02-20 17:53:02 -05002832 ext4_unlock_group(sb, entry->efd_group);
2833 kmem_cache_free(ext4_free_data_cachep, entry);
2834 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002835
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002836 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
Alex Tomasc9de5602008-01-29 00:19:52 -05002837}
2838
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002839int __init ext4_init_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002840{
Theodore Ts'o16828082010-10-27 21:30:09 -04002841 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2842 SLAB_RECLAIM_ACCOUNT);
Alex Tomasc9de5602008-01-29 00:19:52 -05002843 if (ext4_pspace_cachep == NULL)
2844 return -ENOMEM;
2845
Theodore Ts'o16828082010-10-27 21:30:09 -04002846 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2847 SLAB_RECLAIM_ACCOUNT);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002848 if (ext4_ac_cachep == NULL) {
2849 kmem_cache_destroy(ext4_pspace_cachep);
2850 return -ENOMEM;
2851 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002852
Bobi Jam18aadd42012-02-20 17:53:02 -05002853 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2854 SLAB_RECLAIM_ACCOUNT);
2855 if (ext4_free_data_cachep == NULL) {
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04002856 kmem_cache_destroy(ext4_pspace_cachep);
2857 kmem_cache_destroy(ext4_ac_cachep);
2858 return -ENOMEM;
2859 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002860 return 0;
2861}
2862
Theodore Ts'o5dabfc72010-10-27 21:30:14 -04002863void ext4_exit_mballoc(void)
Alex Tomasc9de5602008-01-29 00:19:52 -05002864{
Theodore Ts'o60e66792010-05-17 07:00:00 -04002865 /*
Jesper Dangaard Brouer3e03f9c2009-07-05 22:29:27 -04002866 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2867 * before destroying the slab cache.
2868 */
2869 rcu_barrier();
Alex Tomasc9de5602008-01-29 00:19:52 -05002870 kmem_cache_destroy(ext4_pspace_cachep);
Eric Sandeen256bdb42008-02-10 01:13:33 -05002871 kmem_cache_destroy(ext4_ac_cachep);
Bobi Jam18aadd42012-02-20 17:53:02 -05002872 kmem_cache_destroy(ext4_free_data_cachep);
Eric Sandeen2892c152011-02-12 08:12:18 -05002873 ext4_groupinfo_destroy_slabs();
Alex Tomasc9de5602008-01-29 00:19:52 -05002874}
2875
2876
2877/*
Uwe Kleine-König73b2c712010-07-30 21:02:47 +02002878 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
Alex Tomasc9de5602008-01-29 00:19:52 -05002879 * Returns 0 if success or error code
2880 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002881static noinline_for_stack int
2882ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002883 handle_t *handle, unsigned int reserv_clstrs)
Alex Tomasc9de5602008-01-29 00:19:52 -05002884{
2885 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002886 struct ext4_group_desc *gdp;
2887 struct buffer_head *gdp_bh;
2888 struct ext4_sb_info *sbi;
2889 struct super_block *sb;
2890 ext4_fsblk_t block;
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04002891 int err, len;
Alex Tomasc9de5602008-01-29 00:19:52 -05002892
2893 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2894 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2895
2896 sb = ac->ac_sb;
2897 sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002898
2899 err = -EIO;
Theodore Ts'o574ca172008-07-11 19:27:31 -04002900 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002901 if (!bitmap_bh)
2902 goto out_err;
2903
liang xie5d601252014-05-12 22:06:43 -04002904 BUFFER_TRACE(bitmap_bh, "getting write access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002905 err = ext4_journal_get_write_access(handle, bitmap_bh);
2906 if (err)
2907 goto out_err;
2908
2909 err = -EIO;
2910 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2911 if (!gdp)
2912 goto out_err;
2913
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002914 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002915 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002916
liang xie5d601252014-05-12 22:06:43 -04002917 BUFFER_TRACE(gdp_bh, "get_write_access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002918 err = ext4_journal_get_write_access(handle, gdp_bh);
2919 if (err)
2920 goto out_err;
2921
Akinobu Mitabda00de2010-03-03 23:53:25 -05002922 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002923
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002924 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002925 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002926 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002927 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04002928 /* File system mounted not to panic on error
2929 * Fix the bitmap and repeat the block allocation
2930 * We leak some of the blocks here.
2931 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002932 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002933 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2934 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002935 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002936 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04002937 if (!err)
2938 err = -EAGAIN;
2939 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002940 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002941
2942 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002943#ifdef AGGRESSIVE_CHECK
2944 {
2945 int i;
2946 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2947 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2948 bitmap_bh->b_data));
2949 }
2950 }
2951#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002952 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2953 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002954 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2955 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002956 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002957 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002958 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002959 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002960 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2961 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002962 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002963 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002964
2965 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002966 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002967 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002968 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002969 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002970 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2971 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002972 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2973 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002974
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002975 if (sbi->s_log_groups_per_flex) {
2976 ext4_group_t flex_group = ext4_flex_group(sbi,
2977 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002978 atomic64_sub(ac->ac_b_ex.fe_len,
2979 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002980 }
2981
Frank Mayhar03901312009-01-07 00:06:22 -05002982 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002983 if (err)
2984 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002985 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002986
2987out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002988 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002989 return err;
2990}
2991
2992/*
2993 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002994 * Group request are normalized to s_mb_group_prealloc, which goes to
2995 * s_strip if we set the same via mount option.
2996 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002997 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05002998 *
2999 * XXX: should we try to preallocate more than the group has now?
3000 */
3001static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3002{
3003 struct super_block *sb = ac->ac_sb;
3004 struct ext4_locality_group *lg = ac->ac_lg;
3005
3006 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04003007 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003008 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003009 current->pid, ac->ac_g_ex.fe_len);
3010}
3011
3012/*
3013 * Normalization means making request better in terms of
3014 * size and alignment
3015 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003016static noinline_for_stack void
3017ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05003018 struct ext4_allocation_request *ar)
3019{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003020 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003021 int bsbits, max;
3022 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05003023 loff_t size, start_off;
3024 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04003025 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003026 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003027 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05003028
3029 /* do normalize only data requests, metadata requests
3030 do not need preallocation */
3031 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3032 return;
3033
3034 /* sometime caller may want exact blocks */
3035 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3036 return;
3037
3038 /* caller may indicate that preallocation isn't
3039 * required (it's a tail, for example) */
3040 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3041 return;
3042
3043 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3044 ext4_mb_normalize_group_request(ac);
3045 return ;
3046 }
3047
3048 bsbits = ac->ac_sb->s_blocksize_bits;
3049
3050 /* first, let's learn actual file size
3051 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003052 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003053 size = size << bsbits;
3054 if (size < i_size_read(ac->ac_inode))
3055 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003056 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003057
Valerie Clement19304792008-05-13 19:31:14 -04003058 /* max size of free chunks */
3059 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003060
Valerie Clement19304792008-05-13 19:31:14 -04003061#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3062 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003063
3064 /* first, try to predict filesize */
3065 /* XXX: should this table be tunable? */
3066 start_off = 0;
3067 if (size <= 16 * 1024) {
3068 size = 16 * 1024;
3069 } else if (size <= 32 * 1024) {
3070 size = 32 * 1024;
3071 } else if (size <= 64 * 1024) {
3072 size = 64 * 1024;
3073 } else if (size <= 128 * 1024) {
3074 size = 128 * 1024;
3075 } else if (size <= 256 * 1024) {
3076 size = 256 * 1024;
3077 } else if (size <= 512 * 1024) {
3078 size = 512 * 1024;
3079 } else if (size <= 1024 * 1024) {
3080 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003081 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003082 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003083 (21 - bsbits)) << 21;
3084 size = 2 * 1024 * 1024;
3085 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003086 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3087 (22 - bsbits)) << 22;
3088 size = 4 * 1024 * 1024;
3089 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003090 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003091 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3092 (23 - bsbits)) << 23;
3093 size = 8 * 1024 * 1024;
3094 } else {
Xiaoguang Wangb27b1532014-07-27 22:26:36 -04003095 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3096 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3097 ac->ac_o_ex.fe_len) << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003098 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003099 size = size >> bsbits;
3100 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003101
3102 /* don't cover already allocated blocks in selected range */
3103 if (ar->pleft && start <= ar->lleft) {
3104 size -= ar->lleft + 1 - start;
3105 start = ar->lleft + 1;
3106 }
3107 if (ar->pright && start + size - 1 >= ar->lright)
3108 size -= start + size - ar->lright;
3109
3110 end = start + size;
3111
3112 /* check we don't cross already preallocated blocks */
3113 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003114 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003115 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003116
Alex Tomasc9de5602008-01-29 00:19:52 -05003117 if (pa->pa_deleted)
3118 continue;
3119 spin_lock(&pa->pa_lock);
3120 if (pa->pa_deleted) {
3121 spin_unlock(&pa->pa_lock);
3122 continue;
3123 }
3124
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003125 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3126 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003127
3128 /* PA must not overlap original request */
3129 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3130 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3131
Eric Sandeen38877f42009-08-17 23:55:24 -04003132 /* skip PAs this normalized request doesn't overlap with */
3133 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003134 spin_unlock(&pa->pa_lock);
3135 continue;
3136 }
3137 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3138
Eric Sandeen38877f42009-08-17 23:55:24 -04003139 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003140 if (pa_end <= ac->ac_o_ex.fe_logical) {
3141 BUG_ON(pa_end < start);
3142 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003143 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003144 BUG_ON(pa->pa_lstart > end);
3145 end = pa->pa_lstart;
3146 }
3147 spin_unlock(&pa->pa_lock);
3148 }
3149 rcu_read_unlock();
3150 size = end - start;
3151
3152 /* XXX: extra loop to check we really don't overlap preallocations */
3153 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003154 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003155 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003156
Alex Tomasc9de5602008-01-29 00:19:52 -05003157 spin_lock(&pa->pa_lock);
3158 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003159 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3160 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003161 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3162 }
3163 spin_unlock(&pa->pa_lock);
3164 }
3165 rcu_read_unlock();
3166
3167 if (start + size <= ac->ac_o_ex.fe_logical &&
3168 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003169 ext4_msg(ac->ac_sb, KERN_ERR,
3170 "start %lu, size %lu, fe_logical %lu",
3171 (unsigned long) start, (unsigned long) size,
3172 (unsigned long) ac->ac_o_ex.fe_logical);
Dmitry Monakhovdfe076c2014-10-01 22:26:17 -04003173 BUG();
Alex Tomasc9de5602008-01-29 00:19:52 -05003174 }
Maurizio Lombardib5b60772014-05-27 12:48:56 -04003175 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003176
3177 /* now prepare goal request */
3178
3179 /* XXX: is it better to align blocks WRT to logical
3180 * placement or satisfy big request as is */
3181 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003182 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003183
3184 /* define goal start in order to merge */
3185 if (ar->pright && (ar->lright == (start + size))) {
3186 /* merge to the right */
3187 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3188 &ac->ac_f_ex.fe_group,
3189 &ac->ac_f_ex.fe_start);
3190 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3191 }
3192 if (ar->pleft && (ar->lleft + 1 == start)) {
3193 /* merge to the left */
3194 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3195 &ac->ac_f_ex.fe_group,
3196 &ac->ac_f_ex.fe_start);
3197 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3198 }
3199
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003200 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003201 (unsigned) orig_size, (unsigned) start);
3202}
3203
3204static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3205{
3206 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3207
3208 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3209 atomic_inc(&sbi->s_bal_reqs);
3210 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003211 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003212 atomic_inc(&sbi->s_bal_success);
3213 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3214 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3215 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3216 atomic_inc(&sbi->s_bal_goals);
3217 if (ac->ac_found > sbi->s_mb_max_to_scan)
3218 atomic_inc(&sbi->s_bal_breaks);
3219 }
3220
Theodore Ts'o296c3552009-09-30 00:32:42 -04003221 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3222 trace_ext4_mballoc_alloc(ac);
3223 else
3224 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003225}
3226
3227/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003228 * Called on failure; free up any blocks from the inode PA for this
3229 * context. We don't need this for MB_GROUP_PA because we only change
3230 * pa_free in ext4_mb_release_context(), but on failure, we've already
3231 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3232 */
3233static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3234{
3235 struct ext4_prealloc_space *pa = ac->ac_pa;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003236 struct ext4_buddy e4b;
3237 int err;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003238
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003239 if (pa == NULL) {
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003240 if (ac->ac_f_ex.fe_len == 0)
3241 return;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003242 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3243 if (err) {
3244 /*
3245 * This should never happen since we pin the
3246 * pages in the ext4_allocation_context so
3247 * ext4_mb_load_buddy() should never fail.
3248 */
3249 WARN(1, "mb_load_buddy failed (%d)", err);
3250 return;
3251 }
3252 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3253 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3254 ac->ac_f_ex.fe_len);
3255 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003256 ext4_mb_unload_buddy(&e4b);
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003257 return;
3258 }
3259 if (pa->pa_type == MB_INODE_PA)
Zheng Liu400db9d2012-05-28 17:53:53 -04003260 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003261}
3262
3263/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003264 * use blocks preallocated to inode
3265 */
3266static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3267 struct ext4_prealloc_space *pa)
3268{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003269 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003270 ext4_fsblk_t start;
3271 ext4_fsblk_t end;
3272 int len;
3273
3274 /* found preallocated blocks, use them */
3275 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003276 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3277 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3278 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003279 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3280 &ac->ac_b_ex.fe_start);
3281 ac->ac_b_ex.fe_len = len;
3282 ac->ac_status = AC_STATUS_FOUND;
3283 ac->ac_pa = pa;
3284
3285 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003286 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003287 BUG_ON(pa->pa_free < len);
3288 pa->pa_free -= len;
3289
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003290 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003291}
3292
3293/*
3294 * use blocks preallocated to locality group
3295 */
3296static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3297 struct ext4_prealloc_space *pa)
3298{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003299 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003300
Alex Tomasc9de5602008-01-29 00:19:52 -05003301 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3302 &ac->ac_b_ex.fe_group,
3303 &ac->ac_b_ex.fe_start);
3304 ac->ac_b_ex.fe_len = len;
3305 ac->ac_status = AC_STATUS_FOUND;
3306 ac->ac_pa = pa;
3307
3308 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003309 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003310 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003311 * in on-disk bitmap -- see ext4_mb_release_context()
3312 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003313 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003314 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003315}
3316
3317/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003318 * Return the prealloc space that have minimal distance
3319 * from the goal block. @cpa is the prealloc
3320 * space that is having currently known minimal distance
3321 * from the goal block.
3322 */
3323static struct ext4_prealloc_space *
3324ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3325 struct ext4_prealloc_space *pa,
3326 struct ext4_prealloc_space *cpa)
3327{
3328 ext4_fsblk_t cur_distance, new_distance;
3329
3330 if (cpa == NULL) {
3331 atomic_inc(&pa->pa_count);
3332 return pa;
3333 }
3334 cur_distance = abs(goal_block - cpa->pa_pstart);
3335 new_distance = abs(goal_block - pa->pa_pstart);
3336
Coly Li5a54b2f2011-02-24 14:10:05 -05003337 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003338 return cpa;
3339
3340 /* drop the previous reference */
3341 atomic_dec(&cpa->pa_count);
3342 atomic_inc(&pa->pa_count);
3343 return pa;
3344}
3345
3346/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003347 * search goal blocks in preallocated space
3348 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003349static noinline_for_stack int
3350ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003351{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003352 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003353 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003354 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3355 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003356 struct ext4_prealloc_space *pa, *cpa = NULL;
3357 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003358
3359 /* only data can be preallocated */
3360 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3361 return 0;
3362
3363 /* first, try per-file preallocation */
3364 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003365 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003366
3367 /* all fields in this condition don't change,
3368 * so we can skip locking for them */
3369 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003370 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3371 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003372 continue;
3373
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003374 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003375 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003376 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3377 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003378 continue;
3379
Alex Tomasc9de5602008-01-29 00:19:52 -05003380 /* found preallocated blocks, use them */
3381 spin_lock(&pa->pa_lock);
3382 if (pa->pa_deleted == 0 && pa->pa_free) {
3383 atomic_inc(&pa->pa_count);
3384 ext4_mb_use_inode_pa(ac, pa);
3385 spin_unlock(&pa->pa_lock);
3386 ac->ac_criteria = 10;
3387 rcu_read_unlock();
3388 return 1;
3389 }
3390 spin_unlock(&pa->pa_lock);
3391 }
3392 rcu_read_unlock();
3393
3394 /* can we use group allocation? */
3395 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3396 return 0;
3397
3398 /* inode may have no locality group for some reason */
3399 lg = ac->ac_lg;
3400 if (lg == NULL)
3401 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003402 order = fls(ac->ac_o_ex.fe_len) - 1;
3403 if (order > PREALLOC_TB_SIZE - 1)
3404 /* The max size of hash table is PREALLOC_TB_SIZE */
3405 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003406
Akinobu Mitabda00de2010-03-03 23:53:25 -05003407 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003408 /*
3409 * search for the prealloc space that is having
3410 * minimal distance from the goal block.
3411 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003412 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3413 rcu_read_lock();
3414 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3415 pa_inode_list) {
3416 spin_lock(&pa->pa_lock);
3417 if (pa->pa_deleted == 0 &&
3418 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003419
3420 cpa = ext4_mb_check_group_pa(goal_block,
3421 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003422 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003423 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003424 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003425 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003426 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003427 if (cpa) {
3428 ext4_mb_use_group_pa(ac, cpa);
3429 ac->ac_criteria = 20;
3430 return 1;
3431 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003432 return 0;
3433}
3434
3435/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003436 * the function goes through all block freed in the group
3437 * but not yet committed and marks them used in in-core bitmap.
3438 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003439 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003440 */
3441static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3442 ext4_group_t group)
3443{
3444 struct rb_node *n;
3445 struct ext4_group_info *grp;
3446 struct ext4_free_data *entry;
3447
3448 grp = ext4_get_group_info(sb, group);
3449 n = rb_first(&(grp->bb_free_root));
3450
3451 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003452 entry = rb_entry(n, struct ext4_free_data, efd_node);
3453 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003454 n = rb_next(n);
3455 }
3456 return;
3457}
3458
3459/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003460 * the function goes through all preallocation in this group and marks them
3461 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003462 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003463 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003464static noinline_for_stack
3465void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003466 ext4_group_t group)
3467{
3468 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3469 struct ext4_prealloc_space *pa;
3470 struct list_head *cur;
3471 ext4_group_t groupnr;
3472 ext4_grpblk_t start;
3473 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003474 int len;
3475
3476 /* all form of preallocation discards first load group,
3477 * so the only competing code is preallocation use.
3478 * we don't need any locking here
3479 * notice we do NOT ignore preallocations with pa_deleted
3480 * otherwise we could leave used blocks available for
3481 * allocation in buddy when concurrent ext4_mb_put_pa()
3482 * is dropping preallocation
3483 */
3484 list_for_each(cur, &grp->bb_prealloc_list) {
3485 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3486 spin_lock(&pa->pa_lock);
3487 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3488 &groupnr, &start);
3489 len = pa->pa_len;
3490 spin_unlock(&pa->pa_lock);
3491 if (unlikely(len == 0))
3492 continue;
3493 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003494 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003495 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003496 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003497 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003498}
3499
3500static void ext4_mb_pa_callback(struct rcu_head *head)
3501{
3502 struct ext4_prealloc_space *pa;
3503 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003504
3505 BUG_ON(atomic_read(&pa->pa_count));
3506 BUG_ON(pa->pa_deleted == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05003507 kmem_cache_free(ext4_pspace_cachep, pa);
3508}
3509
3510/*
3511 * drops a reference to preallocated space descriptor
3512 * if this was the last reference and the space is consumed
3513 */
3514static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3515 struct super_block *sb, struct ext4_prealloc_space *pa)
3516{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003517 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003518 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003519
Alex Tomasc9de5602008-01-29 00:19:52 -05003520 /* in this short window concurrent discard can set pa_deleted */
3521 spin_lock(&pa->pa_lock);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003522 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3523 spin_unlock(&pa->pa_lock);
3524 return;
3525 }
3526
Alex Tomasc9de5602008-01-29 00:19:52 -05003527 if (pa->pa_deleted == 1) {
3528 spin_unlock(&pa->pa_lock);
3529 return;
3530 }
3531
3532 pa->pa_deleted = 1;
3533 spin_unlock(&pa->pa_lock);
3534
Eric Sandeend33a1972009-03-16 23:25:40 -04003535 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003536 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003537 * If doing group-based preallocation, pa_pstart may be in the
3538 * next group when pa is used up
3539 */
3540 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003541 grp_blk--;
3542
Lukas Czernerbd862982013-04-03 23:32:34 -04003543 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003544
3545 /*
3546 * possible race:
3547 *
3548 * P1 (buddy init) P2 (regular allocation)
3549 * find block B in PA
3550 * copy on-disk bitmap to buddy
3551 * mark B in on-disk bitmap
3552 * drop PA from group
3553 * mark all PAs in buddy
3554 *
3555 * thus, P1 initializes buddy with B available. to prevent this
3556 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3557 * against that pair
3558 */
3559 ext4_lock_group(sb, grp);
3560 list_del(&pa->pa_group_list);
3561 ext4_unlock_group(sb, grp);
3562
3563 spin_lock(pa->pa_obj_lock);
3564 list_del_rcu(&pa->pa_inode_list);
3565 spin_unlock(pa->pa_obj_lock);
3566
3567 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3568}
3569
3570/*
3571 * creates new preallocated space for given inode
3572 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003573static noinline_for_stack int
3574ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003575{
3576 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003577 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003578 struct ext4_prealloc_space *pa;
3579 struct ext4_group_info *grp;
3580 struct ext4_inode_info *ei;
3581
3582 /* preallocate only when found space is larger then requested */
3583 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3584 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3585 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3586
3587 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3588 if (pa == NULL)
3589 return -ENOMEM;
3590
3591 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3592 int winl;
3593 int wins;
3594 int win;
3595 int offs;
3596
3597 /* we can't allocate as much as normalizer wants.
3598 * so, found space must get proper lstart
3599 * to cover original request */
3600 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3601 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3602
3603 /* we're limited by original request in that
3604 * logical block must be covered any way
3605 * winl is window we can move our chunk within */
3606 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3607
3608 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003609 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003610
3611 /* the smallest one defines real window */
3612 win = min(winl, wins);
3613
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003614 offs = ac->ac_o_ex.fe_logical %
3615 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003616 if (offs && offs < win)
3617 win = offs;
3618
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003619 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003620 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003621 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3622 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3623 }
3624
3625 /* preallocation can change ac_b_ex, thus we store actually
3626 * allocated blocks for history */
3627 ac->ac_f_ex = ac->ac_b_ex;
3628
3629 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3630 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3631 pa->pa_len = ac->ac_b_ex.fe_len;
3632 pa->pa_free = pa->pa_len;
3633 atomic_set(&pa->pa_count, 1);
3634 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003635 INIT_LIST_HEAD(&pa->pa_inode_list);
3636 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003637 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003638 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003639
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003640 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003641 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003642 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003643
3644 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003645 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003646
3647 ei = EXT4_I(ac->ac_inode);
3648 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3649
3650 pa->pa_obj_lock = &ei->i_prealloc_lock;
3651 pa->pa_inode = ac->ac_inode;
3652
3653 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3654 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3655 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3656
3657 spin_lock(pa->pa_obj_lock);
3658 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3659 spin_unlock(pa->pa_obj_lock);
3660
3661 return 0;
3662}
3663
3664/*
3665 * creates new preallocated space for locality group inodes belongs to
3666 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003667static noinline_for_stack int
3668ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003669{
3670 struct super_block *sb = ac->ac_sb;
3671 struct ext4_locality_group *lg;
3672 struct ext4_prealloc_space *pa;
3673 struct ext4_group_info *grp;
3674
3675 /* preallocate only when found space is larger then requested */
3676 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3677 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3678 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3679
3680 BUG_ON(ext4_pspace_cachep == NULL);
3681 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3682 if (pa == NULL)
3683 return -ENOMEM;
3684
3685 /* preallocation can change ac_b_ex, thus we store actually
3686 * allocated blocks for history */
3687 ac->ac_f_ex = ac->ac_b_ex;
3688
3689 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3690 pa->pa_lstart = pa->pa_pstart;
3691 pa->pa_len = ac->ac_b_ex.fe_len;
3692 pa->pa_free = pa->pa_len;
3693 atomic_set(&pa->pa_count, 1);
3694 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003695 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003696 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003697 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003698 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003699
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003700 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003701 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3702 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003703
3704 ext4_mb_use_group_pa(ac, pa);
3705 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3706
3707 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3708 lg = ac->ac_lg;
3709 BUG_ON(lg == NULL);
3710
3711 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3712 pa->pa_inode = NULL;
3713
3714 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3715 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3716 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3717
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003718 /*
3719 * We will later add the new pa to the right bucket
3720 * after updating the pa_free in ext4_mb_release_context
3721 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003722 return 0;
3723}
3724
3725static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3726{
3727 int err;
3728
3729 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3730 err = ext4_mb_new_group_pa(ac);
3731 else
3732 err = ext4_mb_new_inode_pa(ac);
3733 return err;
3734}
3735
3736/*
3737 * finds all unused blocks in on-disk bitmap, frees them in
3738 * in-core bitmap and buddy.
3739 * @pa must be unlinked from inode and group lists, so that
3740 * nobody else can find/use it.
3741 * the caller MUST hold group/inode locks.
3742 * TODO: optimize the case when there are no in-core structures yet
3743 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003744static noinline_for_stack int
3745ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003746 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003747{
Alex Tomasc9de5602008-01-29 00:19:52 -05003748 struct super_block *sb = e4b->bd_sb;
3749 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003750 unsigned int end;
3751 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003752 ext4_group_t group;
3753 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003754 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003755 int err = 0;
3756 int free = 0;
3757
3758 BUG_ON(pa->pa_deleted == 0);
3759 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003760 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003761 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3762 end = bit + pa->pa_len;
3763
Alex Tomasc9de5602008-01-29 00:19:52 -05003764 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003765 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003766 if (bit >= end)
3767 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003768 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003769 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003770 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3771 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003772 free += next - bit;
3773
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003774 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003775 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3776 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003777 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003778 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3779 bit = next + 1;
3780 }
3781 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003782 ext4_msg(e4b->bd_sb, KERN_CRIT,
3783 "pa %p: logic %lu, phys. %lu, len %lu",
3784 pa, (unsigned long) pa->pa_lstart,
3785 (unsigned long) pa->pa_pstart,
3786 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003787 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003788 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003789 /*
3790 * pa is already deleted so we use the value obtained
3791 * from the bitmap and continue.
3792 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003793 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003794 atomic_add(free, &sbi->s_mb_discarded);
3795
3796 return err;
3797}
3798
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003799static noinline_for_stack int
3800ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003801 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003802{
Alex Tomasc9de5602008-01-29 00:19:52 -05003803 struct super_block *sb = e4b->bd_sb;
3804 ext4_group_t group;
3805 ext4_grpblk_t bit;
3806
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003807 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003808 BUG_ON(pa->pa_deleted == 0);
3809 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3810 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3811 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3812 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003813 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003814
3815 return 0;
3816}
3817
3818/*
3819 * releases all preallocations in given group
3820 *
3821 * first, we need to decide discard policy:
3822 * - when do we discard
3823 * 1) ENOSPC
3824 * - how many do we discard
3825 * 1) how many requested
3826 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003827static noinline_for_stack int
3828ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003829 ext4_group_t group, int needed)
3830{
3831 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3832 struct buffer_head *bitmap_bh = NULL;
3833 struct ext4_prealloc_space *pa, *tmp;
3834 struct list_head list;
3835 struct ext4_buddy e4b;
3836 int err;
3837 int busy = 0;
3838 int free = 0;
3839
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003840 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003841
3842 if (list_empty(&grp->bb_prealloc_list))
3843 return 0;
3844
Theodore Ts'o574ca172008-07-11 19:27:31 -04003845 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003846 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003847 ext4_error(sb, "Error reading block bitmap for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003848 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003849 }
3850
3851 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003852 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003853 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003854 put_bh(bitmap_bh);
3855 return 0;
3856 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003857
3858 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003859 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003860
Alex Tomasc9de5602008-01-29 00:19:52 -05003861 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003862repeat:
3863 ext4_lock_group(sb, group);
3864 list_for_each_entry_safe(pa, tmp,
3865 &grp->bb_prealloc_list, pa_group_list) {
3866 spin_lock(&pa->pa_lock);
3867 if (atomic_read(&pa->pa_count)) {
3868 spin_unlock(&pa->pa_lock);
3869 busy = 1;
3870 continue;
3871 }
3872 if (pa->pa_deleted) {
3873 spin_unlock(&pa->pa_lock);
3874 continue;
3875 }
3876
3877 /* seems this one can be freed ... */
3878 pa->pa_deleted = 1;
3879
3880 /* we can trust pa_free ... */
3881 free += pa->pa_free;
3882
3883 spin_unlock(&pa->pa_lock);
3884
3885 list_del(&pa->pa_group_list);
3886 list_add(&pa->u.pa_tmp_list, &list);
3887 }
3888
3889 /* if we still need more blocks and some PAs were used, try again */
3890 if (free < needed && busy) {
3891 busy = 0;
3892 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003893 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003894 goto repeat;
3895 }
3896
3897 /* found anything to free? */
3898 if (list_empty(&list)) {
3899 BUG_ON(free != 0);
3900 goto out;
3901 }
3902
3903 /* now free all selected PAs */
3904 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3905
3906 /* remove from object (inode or locality group) */
3907 spin_lock(pa->pa_obj_lock);
3908 list_del_rcu(&pa->pa_inode_list);
3909 spin_unlock(pa->pa_obj_lock);
3910
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003911 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003912 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003913 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003914 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003915
3916 list_del(&pa->u.pa_tmp_list);
3917 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3918 }
3919
3920out:
3921 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003922 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003923 put_bh(bitmap_bh);
3924 return free;
3925}
3926
3927/*
3928 * releases all non-used preallocated blocks for given inode
3929 *
3930 * It's important to discard preallocations under i_data_sem
3931 * We don't want another block to be served from the prealloc
3932 * space when we are discarding the inode prealloc space.
3933 *
3934 * FIXME!! Make sure it is valid at all the call sites
3935 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003936void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003937{
3938 struct ext4_inode_info *ei = EXT4_I(inode);
3939 struct super_block *sb = inode->i_sb;
3940 struct buffer_head *bitmap_bh = NULL;
3941 struct ext4_prealloc_space *pa, *tmp;
3942 ext4_group_t group = 0;
3943 struct list_head list;
3944 struct ext4_buddy e4b;
3945 int err;
3946
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003947 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003948 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3949 return;
3950 }
3951
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003952 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003953 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003954
3955 INIT_LIST_HEAD(&list);
3956
3957repeat:
3958 /* first, collect all pa's in the inode */
3959 spin_lock(&ei->i_prealloc_lock);
3960 while (!list_empty(&ei->i_prealloc_list)) {
3961 pa = list_entry(ei->i_prealloc_list.next,
3962 struct ext4_prealloc_space, pa_inode_list);
3963 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3964 spin_lock(&pa->pa_lock);
3965 if (atomic_read(&pa->pa_count)) {
3966 /* this shouldn't happen often - nobody should
3967 * use preallocation while we're discarding it */
3968 spin_unlock(&pa->pa_lock);
3969 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003970 ext4_msg(sb, KERN_ERR,
3971 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003972 WARN_ON(1);
3973 schedule_timeout_uninterruptible(HZ);
3974 goto repeat;
3975
3976 }
3977 if (pa->pa_deleted == 0) {
3978 pa->pa_deleted = 1;
3979 spin_unlock(&pa->pa_lock);
3980 list_del_rcu(&pa->pa_inode_list);
3981 list_add(&pa->u.pa_tmp_list, &list);
3982 continue;
3983 }
3984
3985 /* someone is deleting pa right now */
3986 spin_unlock(&pa->pa_lock);
3987 spin_unlock(&ei->i_prealloc_lock);
3988
3989 /* we have to wait here because pa_deleted
3990 * doesn't mean pa is already unlinked from
3991 * the list. as we might be called from
3992 * ->clear_inode() the inode will get freed
3993 * and concurrent thread which is unlinking
3994 * pa from inode's list may access already
3995 * freed memory, bad-bad-bad */
3996
3997 /* XXX: if this happens too often, we can
3998 * add a flag to force wait only in case
3999 * of ->clear_inode(), but not in case of
4000 * regular truncate */
4001 schedule_timeout_uninterruptible(HZ);
4002 goto repeat;
4003 }
4004 spin_unlock(&ei->i_prealloc_lock);
4005
4006 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004007 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04004008 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05004009
4010 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004011 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004012 ext4_error(sb, "Error loading buddy information for %u",
4013 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004014 continue;
4015 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004016
Theodore Ts'o574ca172008-07-11 19:27:31 -04004017 bitmap_bh = ext4_read_block_bitmap(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004018 if (bitmap_bh == NULL) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004019 ext4_error(sb, "Error reading block bitmap for %u",
4020 group);
Jing Zhange39e07f2010-05-14 00:00:00 -04004021 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004022 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05004023 }
4024
4025 ext4_lock_group(sb, group);
4026 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004027 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05004028 ext4_unlock_group(sb, group);
4029
Jing Zhange39e07f2010-05-14 00:00:00 -04004030 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05004031 put_bh(bitmap_bh);
4032
4033 list_del(&pa->u.pa_tmp_list);
4034 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4035 }
4036}
4037
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004038#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05004039static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4040{
4041 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04004042 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05004043
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05004044 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05004045 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04004046 return;
4047
Joe Perches7f6a11e2012-03-19 23:09:43 -04004048 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004049 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04004050 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004051 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004052 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004053 "goal %lu/%lu/%lu@%lu, "
4054 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004055 (unsigned long)ac->ac_o_ex.fe_group,
4056 (unsigned long)ac->ac_o_ex.fe_start,
4057 (unsigned long)ac->ac_o_ex.fe_len,
4058 (unsigned long)ac->ac_o_ex.fe_logical,
4059 (unsigned long)ac->ac_g_ex.fe_group,
4060 (unsigned long)ac->ac_g_ex.fe_start,
4061 (unsigned long)ac->ac_g_ex.fe_len,
4062 (unsigned long)ac->ac_g_ex.fe_logical,
4063 (unsigned long)ac->ac_b_ex.fe_group,
4064 (unsigned long)ac->ac_b_ex.fe_start,
4065 (unsigned long)ac->ac_b_ex.fe_len,
4066 (unsigned long)ac->ac_b_ex.fe_logical,
4067 (int)ac->ac_criteria);
Eric Sandeendc9ddd92014-02-20 13:32:10 -05004068 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004069 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04004070 ngroups = ext4_get_groups_count(sb);
4071 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004072 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4073 struct ext4_prealloc_space *pa;
4074 ext4_grpblk_t start;
4075 struct list_head *cur;
4076 ext4_lock_group(sb, i);
4077 list_for_each(cur, &grp->bb_prealloc_list) {
4078 pa = list_entry(cur, struct ext4_prealloc_space,
4079 pa_group_list);
4080 spin_lock(&pa->pa_lock);
4081 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4082 NULL, &start);
4083 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004084 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4085 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004086 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004087 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004088
4089 if (grp->bb_free == 0)
4090 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004091 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004092 i, grp->bb_free, grp->bb_fragments);
4093 }
4094 printk(KERN_ERR "\n");
4095}
4096#else
4097static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4098{
4099 return;
4100}
4101#endif
4102
4103/*
4104 * We use locality group preallocation for small size file. The size of the
4105 * file is determined by the current size or the resulting size after
4106 * allocation which ever is larger
4107 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004108 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004109 */
4110static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4111{
4112 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4113 int bsbits = ac->ac_sb->s_blocksize_bits;
4114 loff_t size, isize;
4115
4116 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4117 return;
4118
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004119 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4120 return;
4121
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004122 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004123 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4124 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004125
Theodore Ts'o50797482009-09-18 13:34:02 -04004126 if ((size == isize) &&
4127 !ext4_fs_is_busy(sbi) &&
4128 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4129 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4130 return;
4131 }
4132
Robin Dongebbe0272011-10-26 05:14:27 -04004133 if (sbi->s_mb_group_prealloc <= 0) {
4134 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4135 return;
4136 }
4137
Alex Tomasc9de5602008-01-29 00:19:52 -05004138 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004139 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004140 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004141 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004142 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004143 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004144
4145 BUG_ON(ac->ac_lg != NULL);
4146 /*
4147 * locality group prealloc space are per cpu. The reason for having
4148 * per cpu locality group is to reduce the contention between block
4149 * request from multiple CPUs.
4150 */
Christoph Lametera0b6bc62014-08-17 12:30:28 -05004151 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004152
4153 /* we're going to use group allocation */
4154 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4155
4156 /* serialize all allocations in the group */
4157 mutex_lock(&ac->ac_lg->lg_mutex);
4158}
4159
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004160static noinline_for_stack int
4161ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004162 struct ext4_allocation_request *ar)
4163{
4164 struct super_block *sb = ar->inode->i_sb;
4165 struct ext4_sb_info *sbi = EXT4_SB(sb);
4166 struct ext4_super_block *es = sbi->s_es;
4167 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004168 unsigned int len;
4169 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004170 ext4_grpblk_t block;
4171
4172 /* we can't allocate > group size */
4173 len = ar->len;
4174
4175 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004176 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4177 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004178
4179 /* start searching from the goal */
4180 goal = ar->goal;
4181 if (goal < le32_to_cpu(es->s_first_data_block) ||
4182 goal >= ext4_blocks_count(es))
4183 goal = le32_to_cpu(es->s_first_data_block);
4184 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4185
4186 /* set up allocation goals */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004187 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05004188 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004189 ac->ac_sb = sb;
4190 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004191 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004192 ac->ac_o_ex.fe_group = group;
4193 ac->ac_o_ex.fe_start = block;
4194 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004195 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004196 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004197
4198 /* we have to define context: we'll we work with a file or
4199 * locality group. this is a policy, actually */
4200 ext4_mb_group_or_file(ac);
4201
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004202 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004203 "left: %u/%u, right %u/%u to %swritable\n",
4204 (unsigned) ar->len, (unsigned) ar->logical,
4205 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4206 (unsigned) ar->lleft, (unsigned) ar->pleft,
4207 (unsigned) ar->lright, (unsigned) ar->pright,
4208 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4209 return 0;
4210
4211}
4212
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004213static noinline_for_stack void
4214ext4_mb_discard_lg_preallocations(struct super_block *sb,
4215 struct ext4_locality_group *lg,
4216 int order, int total_entries)
4217{
4218 ext4_group_t group = 0;
4219 struct ext4_buddy e4b;
4220 struct list_head discard_list;
4221 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004222
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004223 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004224
4225 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004226
4227 spin_lock(&lg->lg_prealloc_lock);
4228 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4229 pa_inode_list) {
4230 spin_lock(&pa->pa_lock);
4231 if (atomic_read(&pa->pa_count)) {
4232 /*
4233 * This is the pa that we just used
4234 * for block allocation. So don't
4235 * free that
4236 */
4237 spin_unlock(&pa->pa_lock);
4238 continue;
4239 }
4240 if (pa->pa_deleted) {
4241 spin_unlock(&pa->pa_lock);
4242 continue;
4243 }
4244 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004245 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004246
4247 /* seems this one can be freed ... */
4248 pa->pa_deleted = 1;
4249 spin_unlock(&pa->pa_lock);
4250
4251 list_del_rcu(&pa->pa_inode_list);
4252 list_add(&pa->u.pa_tmp_list, &discard_list);
4253
4254 total_entries--;
4255 if (total_entries <= 5) {
4256 /*
4257 * we want to keep only 5 entries
4258 * allowing it to grow to 8. This
4259 * mak sure we don't call discard
4260 * soon for this list.
4261 */
4262 break;
4263 }
4264 }
4265 spin_unlock(&lg->lg_prealloc_lock);
4266
4267 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4268
Lukas Czernerbd862982013-04-03 23:32:34 -04004269 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004270 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004271 ext4_error(sb, "Error loading buddy information for %u",
4272 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004273 continue;
4274 }
4275 ext4_lock_group(sb, group);
4276 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004277 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004278 ext4_unlock_group(sb, group);
4279
Jing Zhange39e07f2010-05-14 00:00:00 -04004280 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004281 list_del(&pa->u.pa_tmp_list);
4282 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4283 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004284}
4285
4286/*
4287 * We have incremented pa_count. So it cannot be freed at this
4288 * point. Also we hold lg_mutex. So no parallel allocation is
4289 * possible from this lg. That means pa_free cannot be updated.
4290 *
4291 * A parallel ext4_mb_discard_group_preallocations is possible.
4292 * which can cause the lg_prealloc_list to be updated.
4293 */
4294
4295static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4296{
4297 int order, added = 0, lg_prealloc_count = 1;
4298 struct super_block *sb = ac->ac_sb;
4299 struct ext4_locality_group *lg = ac->ac_lg;
4300 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4301
4302 order = fls(pa->pa_free) - 1;
4303 if (order > PREALLOC_TB_SIZE - 1)
4304 /* The max size of hash table is PREALLOC_TB_SIZE */
4305 order = PREALLOC_TB_SIZE - 1;
4306 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004307 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004308 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4309 pa_inode_list) {
4310 spin_lock(&tmp_pa->pa_lock);
4311 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004312 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004313 continue;
4314 }
4315 if (!added && pa->pa_free < tmp_pa->pa_free) {
4316 /* Add to the tail of the previous entry */
4317 list_add_tail_rcu(&pa->pa_inode_list,
4318 &tmp_pa->pa_inode_list);
4319 added = 1;
4320 /*
4321 * we want to count the total
4322 * number of entries in the list
4323 */
4324 }
4325 spin_unlock(&tmp_pa->pa_lock);
4326 lg_prealloc_count++;
4327 }
4328 if (!added)
4329 list_add_tail_rcu(&pa->pa_inode_list,
4330 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004331 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004332
4333 /* Now trim the list to be not more than 8 elements */
4334 if (lg_prealloc_count > 8) {
4335 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004336 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004337 return;
4338 }
4339 return ;
4340}
4341
Alex Tomasc9de5602008-01-29 00:19:52 -05004342/*
4343 * release all resource we used in allocation
4344 */
4345static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4346{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004347 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004348 struct ext4_prealloc_space *pa = ac->ac_pa;
4349 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004350 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004351 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004352 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004353 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4354 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004355 pa->pa_free -= ac->ac_b_ex.fe_len;
4356 pa->pa_len -= ac->ac_b_ex.fe_len;
4357 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004358 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004359 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004360 if (pa) {
4361 /*
4362 * We want to add the pa to the right bucket.
4363 * Remove it from the list and while adding
4364 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004365 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004366 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004367 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004368 spin_lock(pa->pa_obj_lock);
4369 list_del_rcu(&pa->pa_inode_list);
4370 spin_unlock(pa->pa_obj_lock);
4371 ext4_mb_add_n_trim(ac);
4372 }
4373 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4374 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004375 if (ac->ac_bitmap_page)
4376 page_cache_release(ac->ac_bitmap_page);
4377 if (ac->ac_buddy_page)
4378 page_cache_release(ac->ac_buddy_page);
4379 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4380 mutex_unlock(&ac->ac_lg->lg_mutex);
4381 ext4_mb_collect_stats(ac);
4382 return 0;
4383}
4384
4385static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4386{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004387 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004388 int ret;
4389 int freed = 0;
4390
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004391 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004392 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004393 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4394 freed += ret;
4395 needed -= ret;
4396 }
4397
4398 return freed;
4399}
4400
4401/*
4402 * Main entry point into mballoc to allocate blocks
4403 * it tries to use preallocation first, then falls back
4404 * to usual allocation
4405 */
4406ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004407 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004408{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004409 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004410 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004411 struct ext4_sb_info *sbi;
4412 struct super_block *sb;
4413 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004414 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004415 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004416
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004417 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004418 sb = ar->inode->i_sb;
4419 sbi = EXT4_SB(sb);
4420
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004421 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004422
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004423 /* Allow to use superuser reservation for quota file */
4424 if (IS_NOQUOTA(ar->inode))
4425 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4426
Theodore Ts'oe3cf5d52014-09-04 18:07:25 -04004427 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
Mingming Cao60e58e02009-01-22 18:13:05 +01004428 /* Without delayed allocation we need to verify
4429 * there is enough free blocks to do block allocation
4430 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004431 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004432 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004433 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004434
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004435 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004436 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004437 ar->len = ar->len >> 1;
4438 }
4439 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004440 *errp = -ENOSPC;
4441 return 0;
4442 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004443 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004444 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004445 dquot_alloc_block_nofail(ar->inode,
4446 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004447 } else {
4448 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004449 dquot_alloc_block(ar->inode,
4450 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004451
4452 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4453 ar->len--;
4454 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004455 }
4456 inquota = ar->len;
4457 if (ar->len == 0) {
4458 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004459 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004460 }
Mingming Caod2a17632008-07-14 17:52:37 -04004461 }
Mingming Caod2a17632008-07-14 17:52:37 -04004462
Wei Yongjun85556c92012-09-26 20:43:37 -04004463 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004464 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004465 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004466 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004467 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004468 }
4469
Eric Sandeen256bdb42008-02-10 01:13:33 -05004470 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004471 if (*errp) {
4472 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004473 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004474 }
4475
Eric Sandeen256bdb42008-02-10 01:13:33 -05004476 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4477 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004478 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4479 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004480repeat:
4481 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004482 *errp = ext4_mb_regular_allocator(ac);
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004483 if (*errp)
4484 goto discard_and_exit;
4485
4486 /* as we've just preallocated more space than
4487 * user requested originally, we store allocated
4488 * space in a special descriptor */
4489 if (ac->ac_status == AC_STATUS_FOUND &&
4490 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4491 *errp = ext4_mb_new_preallocation(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004492 if (*errp) {
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004493 discard_and_exit:
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004494 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004495 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004496 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004497 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004498 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004499 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004500 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004501 /*
4502 * drop the reference that we took
4503 * in ext4_mb_use_best_found
4504 */
4505 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04004506 ac->ac_b_ex.fe_group = 0;
4507 ac->ac_b_ex.fe_start = 0;
4508 ac->ac_b_ex.fe_len = 0;
4509 ac->ac_status = AC_STATUS_CONTINUE;
4510 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004511 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004512 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004513 goto errout;
4514 } else {
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04004515 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4516 ar->len = ac->ac_b_ex.fe_len;
4517 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004518 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004519 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004520 if (freed)
4521 goto repeat;
4522 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004523 }
4524
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004525errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004526 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004527 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004528 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004529 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004530 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004531 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004532out:
4533 if (ac)
4534 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004535 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004536 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004537 if (!ar->len) {
Theodore Ts'oe3cf5d52014-09-04 18:07:25 -04004538 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004539 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004540 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004541 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004542 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004543
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004544 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004545
Alex Tomasc9de5602008-01-29 00:19:52 -05004546 return block;
4547}
Alex Tomasc9de5602008-01-29 00:19:52 -05004548
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004549/*
4550 * We can merge two free data extents only if the physical blocks
4551 * are contiguous, AND the extents were freed by the same transaction,
4552 * AND the blocks are associated with the same group.
4553 */
4554static int can_merge(struct ext4_free_data *entry1,
4555 struct ext4_free_data *entry2)
4556{
Bobi Jam18aadd42012-02-20 17:53:02 -05004557 if ((entry1->efd_tid == entry2->efd_tid) &&
4558 (entry1->efd_group == entry2->efd_group) &&
4559 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004560 return 1;
4561 return 0;
4562}
4563
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004564static noinline_for_stack int
4565ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004566 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004567{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004568 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004569 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004570 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004571 struct ext4_group_info *db = e4b->bd_info;
4572 struct super_block *sb = e4b->bd_sb;
4573 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004574 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4575 struct rb_node *parent = NULL, *new_node;
4576
Frank Mayhar03901312009-01-07 00:06:22 -05004577 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004578 BUG_ON(e4b->bd_bitmap_page == NULL);
4579 BUG_ON(e4b->bd_buddy_page == NULL);
4580
Bobi Jam18aadd42012-02-20 17:53:02 -05004581 new_node = &new_entry->efd_node;
4582 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004583
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004584 if (!*n) {
4585 /* first free block exent. We need to
4586 protect buddy cache from being freed,
4587 * otherwise we'll refresh it from
4588 * on-disk bitmap and lose not-yet-available
4589 * blocks */
4590 page_cache_get(e4b->bd_buddy_page);
4591 page_cache_get(e4b->bd_bitmap_page);
4592 }
4593 while (*n) {
4594 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004595 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4596 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004597 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004598 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004599 n = &(*n)->rb_right;
4600 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004601 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004602 ext4_group_first_block_no(sb, group) +
4603 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004604 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004605 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004606 }
4607 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004608
4609 rb_link_node(new_node, parent, n);
4610 rb_insert_color(new_node, &db->bb_free_root);
4611
4612 /* Now try to see the extent can be merged to left and right */
4613 node = rb_prev(new_node);
4614 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004615 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004616 if (can_merge(entry, new_entry) &&
4617 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004618 new_entry->efd_start_cluster = entry->efd_start_cluster;
4619 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004620 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004621 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004622 }
4623 }
4624
4625 node = rb_next(new_node);
4626 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004627 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004628 if (can_merge(new_entry, entry) &&
4629 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004630 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004631 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004632 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004633 }
4634 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004635 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004636 ext4_journal_callback_add(handle, ext4_free_data_callback,
4637 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004638 return 0;
4639}
4640
Theodore Ts'o44338712009-11-22 07:44:56 -05004641/**
4642 * ext4_free_blocks() -- Free given blocks and update quota
4643 * @handle: handle for this transaction
4644 * @inode: inode
4645 * @block: start physical block to free
4646 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004647 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004648 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004649void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004650 struct buffer_head *bh, ext4_fsblk_t block,
4651 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004652{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004653 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004654 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004655 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004656 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004657 ext4_grpblk_t bit;
4658 struct buffer_head *gd_bh;
4659 ext4_group_t block_group;
4660 struct ext4_sb_info *sbi;
4661 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004662 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004663 int err = 0;
4664 int ret;
4665
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004666 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004667 if (bh) {
4668 if (block)
4669 BUG_ON(block != bh->b_blocknr);
4670 else
4671 block = bh->b_blocknr;
4672 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004673
Alex Tomasc9de5602008-01-29 00:19:52 -05004674 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004675 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4676 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004677 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004678 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004679 goto error_return;
4680 }
4681
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004682 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004683 trace_ext4_free_blocks(inode, block, count, flags);
4684
4685 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4686 struct buffer_head *tbh = bh;
4687 int i;
4688
4689 BUG_ON(bh && (count > 1));
4690
4691 for (i = 0; i < count; i++) {
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004692 cond_resched();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004693 if (!bh)
4694 tbh = sb_find_get_block(inode->i_sb,
4695 block + i);
Theodore Ts'o2ed57242013-06-12 11:43:02 -04004696 if (!tbh)
Namhyung Kim87783692010-10-27 21:30:11 -04004697 continue;
Theodore Ts'o60e66792010-05-17 07:00:00 -04004698 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004699 inode, tbh, block + i);
4700 }
4701 }
4702
Theodore Ts'o60e66792010-05-17 07:00:00 -04004703 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004704 * We need to make sure we don't reuse the freed block until
4705 * after the transaction is committed, which we can do by
4706 * treating the block as metadata, below. We make an
4707 * exception if the inode is to be written in writeback mode
4708 * since writeback mode has weak data consistency guarantees.
4709 */
4710 if (!ext4_should_writeback_data(inode))
4711 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004712
Theodore Ts'o84130192011-09-09 18:50:51 -04004713 /*
4714 * If the extent to be freed does not begin on a cluster
4715 * boundary, we need to deal with partial clusters at the
4716 * beginning and end of the extent. Normally we will free
4717 * blocks at the beginning or the end unless we are explicitly
4718 * requested to avoid doing so.
4719 */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004720 overflow = EXT4_PBLK_COFF(sbi, block);
Theodore Ts'o84130192011-09-09 18:50:51 -04004721 if (overflow) {
4722 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4723 overflow = sbi->s_cluster_ratio - overflow;
4724 block += overflow;
4725 if (count > overflow)
4726 count -= overflow;
4727 else
4728 return;
4729 } else {
4730 block -= overflow;
4731 count += overflow;
4732 }
4733 }
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004734 overflow = EXT4_LBLK_COFF(sbi, count);
Theodore Ts'o84130192011-09-09 18:50:51 -04004735 if (overflow) {
4736 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4737 if (count > overflow)
4738 count -= overflow;
4739 else
4740 return;
4741 } else
4742 count += sbi->s_cluster_ratio - overflow;
4743 }
4744
Alex Tomasc9de5602008-01-29 00:19:52 -05004745do_more:
4746 overflow = 0;
4747 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4748
Darrick J. Wong163a2032013-08-28 17:35:51 -04004749 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4750 ext4_get_group_info(sb, block_group))))
4751 return;
4752
Alex Tomasc9de5602008-01-29 00:19:52 -05004753 /*
4754 * Check to see if we are freeing blocks across a group
4755 * boundary.
4756 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004757 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4758 overflow = EXT4_C2B(sbi, bit) + count -
4759 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004760 count -= overflow;
4761 }
Lukas Czerner810da242013-03-02 17:18:58 -05004762 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004763 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004764 if (!bitmap_bh) {
4765 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004766 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004767 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004768 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004769 if (!gdp) {
4770 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004771 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004772 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004773
4774 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4775 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4776 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004777 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004778 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004779 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004780
Eric Sandeen12062dd2010-02-15 14:19:27 -05004781 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004782 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca02008-05-15 14:43:20 -04004783 /* err = 0. ext4_std_error should be a no op */
4784 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004785 }
4786
4787 BUFFER_TRACE(bitmap_bh, "getting write access");
4788 err = ext4_journal_get_write_access(handle, bitmap_bh);
4789 if (err)
4790 goto error_return;
4791
4792 /*
4793 * We are about to modify some metadata. Call the journal APIs
4794 * to unshare ->b_data if a currently-committing transaction is
4795 * using it
4796 */
4797 BUFFER_TRACE(gd_bh, "get_write_access");
4798 err = ext4_journal_get_write_access(handle, gd_bh);
4799 if (err)
4800 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004801#ifdef AGGRESSIVE_CHECK
4802 {
4803 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004804 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004805 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4806 }
4807#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004808 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004809
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004810 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4811 if (err)
4812 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004813
4814 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004815 struct ext4_free_data *new_entry;
4816 /*
4817 * blocks being freed are metadata. these blocks shouldn't
4818 * be used until this transaction is committed
4819 */
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004820 retry:
Bobi Jam18aadd42012-02-20 17:53:02 -05004821 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004822 if (!new_entry) {
Theodore Ts'oe7676a72013-07-13 00:40:35 -04004823 /*
4824 * We use a retry loop because
4825 * ext4_free_blocks() is not allowed to fail.
4826 */
4827 cond_resched();
4828 congestion_wait(BLK_RW_ASYNC, HZ/50);
4829 goto retry;
Theodore Ts'ob72143a2010-12-20 07:26:59 -05004830 }
Bobi Jam18aadd42012-02-20 17:53:02 -05004831 new_entry->efd_start_cluster = bit;
4832 new_entry->efd_group = block_group;
4833 new_entry->efd_count = count_clusters;
4834 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004835
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004836 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004837 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004838 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004839 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004840 /* need to update group_info->bb_free and bitmap
4841 * with group lock held. generate_buddy look at
4842 * them with group lock_held
4843 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004844 if (test_opt(sb, DISCARD)) {
4845 err = ext4_issue_discard(sb, block_group, bit, count);
4846 if (err && err != -EOPNOTSUPP)
4847 ext4_msg(sb, KERN_WARNING, "discard request in"
4848 " group:%d block:%d count:%lu failed"
4849 " with %d", block_group, bit, count,
4850 err);
Lukas Czerner8f9ff182013-10-30 11:10:52 -04004851 } else
4852 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004853
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004854 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004855 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4856 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004857 }
4858
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004859 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4860 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004861 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004862 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004863 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004864
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004865 if (sbi->s_log_groups_per_flex) {
4866 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004867 atomic64_add(count_clusters,
4868 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004869 }
4870
Theodore Ts'o71d4f7d2014-07-15 06:02:38 -04004871 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Aditya Kali7b415bf2011-09-09 19:04:51 -04004872 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
Jan Kara7d734532013-08-17 09:36:54 -04004873 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4874
4875 ext4_mb_unload_buddy(&e4b);
Aditya Kali7b415bf2011-09-09 19:04:51 -04004876
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004877 /* We dirtied the bitmap block */
4878 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4879 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4880
Alex Tomasc9de5602008-01-29 00:19:52 -05004881 /* And the group descriptor block */
4882 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004883 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004884 if (!err)
4885 err = ret;
4886
4887 if (overflow && !err) {
4888 block += count;
4889 count = overflow;
4890 put_bh(bitmap_bh);
4891 goto do_more;
4892 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004893error_return:
4894 brelse(bitmap_bh);
4895 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004896 return;
4897}
Lukas Czerner7360d172010-10-27 21:30:12 -04004898
4899/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004900 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004901 * @handle: handle to this transaction
4902 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004903 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004904 * @count: number of blocks to free
4905 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004906 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004907 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004908int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004909 ext4_fsblk_t block, unsigned long count)
4910{
4911 struct buffer_head *bitmap_bh = NULL;
4912 struct buffer_head *gd_bh;
4913 ext4_group_t block_group;
4914 ext4_grpblk_t bit;
4915 unsigned int i;
4916 struct ext4_group_desc *desc;
4917 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004918 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004919 int err = 0, ret, blk_free_count;
4920 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004921
4922 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4923
Yongqiang Yang4740b832011-07-26 21:51:08 -04004924 if (count == 0)
4925 return 0;
4926
Amir Goldstein2846e822011-05-09 10:46:41 -04004927 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004928 /*
4929 * Check to see if we are freeing blocks across a group
4930 * boundary.
4931 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004932 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4933 ext4_warning(sb, "too much blocks added to group %u\n",
4934 block_group);
4935 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004936 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004937 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004938
Amir Goldstein2846e822011-05-09 10:46:41 -04004939 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004940 if (!bitmap_bh) {
4941 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004942 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004943 }
4944
Amir Goldstein2846e822011-05-09 10:46:41 -04004945 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004946 if (!desc) {
4947 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004948 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004949 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004950
4951 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4952 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4953 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4954 in_range(block + count - 1, ext4_inode_table(sb, desc),
4955 sbi->s_itb_per_group)) {
4956 ext4_error(sb, "Adding blocks in system zones - "
4957 "Block = %llu, count = %lu",
4958 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004959 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004960 goto error_return;
4961 }
4962
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004963 BUFFER_TRACE(bitmap_bh, "getting write access");
4964 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004965 if (err)
4966 goto error_return;
4967
4968 /*
4969 * We are about to modify some metadata. Call the journal APIs
4970 * to unshare ->b_data if a currently-committing transaction is
4971 * using it
4972 */
4973 BUFFER_TRACE(gd_bh, "get_write_access");
4974 err = ext4_journal_get_write_access(handle, gd_bh);
4975 if (err)
4976 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004977
Amir Goldstein2846e822011-05-09 10:46:41 -04004978 for (i = 0, blocks_freed = 0; i < count; i++) {
4979 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004980 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004981 ext4_error(sb, "bit already cleared for block %llu",
4982 (ext4_fsblk_t)(block + i));
4983 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4984 } else {
4985 blocks_freed++;
4986 }
4987 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004988
4989 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4990 if (err)
4991 goto error_return;
4992
4993 /*
4994 * need to update group_info->bb_free and bitmap
4995 * with group lock held. generate_buddy look at
4996 * them with group lock_held
4997 */
Amir Goldstein2846e822011-05-09 10:46:41 -04004998 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004999 mb_clear_bits(bitmap_bh->b_data, bit, count);
5000 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04005001 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5002 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04005003 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04005004 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04005005 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04005006 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05005007 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04005008
5009 if (sbi->s_log_groups_per_flex) {
5010 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04005011 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5012 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04005013 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04005014
5015 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04005016
5017 /* We dirtied the bitmap block */
5018 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5019 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5020
5021 /* And the group descriptor block */
5022 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5023 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5024 if (!err)
5025 err = ret;
5026
5027error_return:
5028 brelse(bitmap_bh);
5029 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04005030 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04005031}
5032
5033/**
Lukas Czerner7360d172010-10-27 21:30:12 -04005034 * ext4_trim_extent -- function to TRIM one single free extent in the group
5035 * @sb: super block for the file system
5036 * @start: starting block of the free extent in the alloc. group
5037 * @count: number of blocks to TRIM
5038 * @group: alloc. group we are working with
5039 * @e4b: ext4 buddy for the group
5040 *
5041 * Trim "count" blocks starting at "start" in the "group". To assure that no
5042 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5043 * be called with under the group lock.
5044 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005045static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04005046 ext4_group_t group, struct ext4_buddy *e4b)
jon ernste2cbd582014-04-12 23:01:28 -04005047__releases(bitlock)
5048__acquires(bitlock)
Lukas Czerner7360d172010-10-27 21:30:12 -04005049{
5050 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005051 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005052
Tao Mab3d4c2b2011-07-11 00:01:52 -04005053 trace_ext4_trim_extent(sb, group, start, count);
5054
Lukas Czerner7360d172010-10-27 21:30:12 -04005055 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5056
5057 ex.fe_start = start;
5058 ex.fe_group = group;
5059 ex.fe_len = count;
5060
5061 /*
5062 * Mark blocks used, so no one can reuse them while
5063 * being trimmed.
5064 */
5065 mb_mark_used(e4b, &ex);
5066 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005067 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04005068 ext4_lock_group(sb, group);
5069 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005070 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005071}
5072
5073/**
5074 * ext4_trim_all_free -- function to trim all free space in alloc. group
5075 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04005076 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04005077 * @start: first group block to examine
5078 * @max: last group block to examine
5079 * @minblocks: minimum extent block count
5080 *
5081 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5082 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5083 * the extent.
5084 *
5085 *
5086 * ext4_trim_all_free walks through group's block bitmap searching for free
5087 * extents. When the free extent is found, mark it as used in group buddy
5088 * bitmap. Then issue a TRIM command on this extent and free the extent in
5089 * the group buddy bitmap. This is done until whole group is scanned.
5090 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005091static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005092ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5093 ext4_grpblk_t start, ext4_grpblk_t max,
5094 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005095{
5096 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005097 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005098 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005099 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005100
Tao Mab3d4c2b2011-07-11 00:01:52 -04005101 trace_ext4_trim_all_free(sb, group, start, max);
5102
Lukas Czerner78944082011-05-24 18:16:27 -04005103 ret = ext4_mb_load_buddy(sb, group, &e4b);
5104 if (ret) {
5105 ext4_error(sb, "Error in loading buddy "
5106 "information for %u", group);
5107 return ret;
5108 }
Lukas Czerner78944082011-05-24 18:16:27 -04005109 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005110
5111 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005112 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5113 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5114 goto out;
5115
Lukas Czerner78944082011-05-24 18:16:27 -04005116 start = (e4b.bd_info->bb_first_free > start) ?
5117 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005118
Lukas Czerner913eed832012-03-21 21:22:22 -04005119 while (start <= max) {
5120 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5121 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005122 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005123 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005124
5125 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005126 ret = ext4_trim_extent(sb, start,
5127 next - start, group, &e4b);
5128 if (ret && ret != -EOPNOTSUPP)
5129 break;
5130 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005131 count += next - start;
5132 }
Tao Ma169ddc32011-07-11 00:00:07 -04005133 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005134 start = next + 1;
5135
5136 if (fatal_signal_pending(current)) {
5137 count = -ERESTARTSYS;
5138 break;
5139 }
5140
5141 if (need_resched()) {
5142 ext4_unlock_group(sb, group);
5143 cond_resched();
5144 ext4_lock_group(sb, group);
5145 }
5146
Tao Ma169ddc32011-07-11 00:00:07 -04005147 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005148 break;
5149 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005150
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005151 if (!ret) {
5152 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005153 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005154 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005155out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005156 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005157 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005158
5159 ext4_debug("trimmed %d blocks in the group %d\n",
5160 count, group);
5161
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005162 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005163}
5164
5165/**
5166 * ext4_trim_fs() -- trim ioctl handle function
5167 * @sb: superblock for filesystem
5168 * @range: fstrim_range structure
5169 *
5170 * start: First Byte to trim
5171 * len: number of Bytes to trim from start
5172 * minlen: minimum extent length in Bytes
5173 * ext4_trim_fs goes through all allocation groups containing Bytes from
5174 * start to start+len. For each such a group ext4_trim_all_free function
5175 * is invoked to trim all free space.
5176 */
5177int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5178{
Lukas Czerner78944082011-05-24 18:16:27 -04005179 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005180 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005181 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005182 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005183 ext4_fsblk_t first_data_blk =
5184 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005185 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005186 int ret = 0;
5187
5188 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005189 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005190 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5191 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005192
Lukas Czerner5de35e82012-10-22 18:01:19 -04005193 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5194 start >= max_blks ||
5195 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005196 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005197 if (end >= max_blks)
5198 end = max_blks - 1;
5199 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005200 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005201 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005202 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005203
Lukas Czerner913eed832012-03-21 21:22:22 -04005204 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005205 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005206 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005207 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005208 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005209
Lukas Czerner913eed832012-03-21 21:22:22 -04005210 /* end now represents the last cluster to discard in this group */
5211 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005212
5213 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005214 grp = ext4_get_group_info(sb, group);
5215 /* We only do this if the grp has never been initialized */
5216 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5217 ret = ext4_mb_init_group(sb, group);
5218 if (ret)
5219 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005220 }
5221
Tao Ma0ba08512011-03-23 15:48:11 -04005222 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005223 * For all the groups except the last one, last cluster will
5224 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5225 * change it for the last group, note that last_cluster is
5226 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005227 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005228 if (group == last_group)
5229 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005230
Lukas Czerner78944082011-05-24 18:16:27 -04005231 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005232 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005233 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005234 if (cnt < 0) {
5235 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005236 break;
5237 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005238 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005239 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005240
5241 /*
5242 * For every group except the first one, we are sure
5243 * that the first cluster to discard will be cluster #0.
5244 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005245 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005246 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005247
Tao Ma3d56b8d2011-07-11 00:03:38 -04005248 if (!ret)
5249 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5250
Tao Ma22f10452011-07-10 23:52:37 -04005251out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005252 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005253 return ret;
5254}