blob: 4424b7bf8ac64e431a11570e247a1d828634d29e [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 }
Darrick J. Wong9008a582015-10-17 21:33:24 -0400877 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
878 if (IS_ERR(bh[i])) {
879 err = PTR_ERR(bh[i]);
880 bh[i] = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -0500881 goto out;
Aneesh Kumar K.V2ccb5fb2009-01-05 21:49:55 -0500882 }
Theodore Ts'o813e5722012-02-20 17:52:46 -0500883 mb_debug(1, "read bitmap for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500884 }
885
886 /* wait for I/O completion */
Theodore Ts'o813e5722012-02-20 17:52:46 -0500887 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
Darrick J. Wong9008a582015-10-17 21:33:24 -0400888 int err2;
889
890 if (!bh[i])
891 continue;
892 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
893 if (!err)
894 err = err2;
Theodore Ts'o813e5722012-02-20 17:52:46 -0500895 }
Alex Tomasc9de5602008-01-29 00:19:52 -0500896
897 first_block = page->index * blocks_per_page;
898 for (i = 0; i < blocks_per_page; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -0500899 group = (first_block + i) >> 1;
Theodore Ts'o8df96752009-05-01 08:50:38 -0400900 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -0500901 break;
902
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400903 if (!bh[group - first_group])
904 /* skip initialized uptodate buddy */
905 continue;
906
Lukas Czernerbbdc3222015-06-08 11:38:37 -0400907 if (!buffer_verified(bh[group - first_group]))
908 /* Skip faulty bitmaps */
909 continue;
910 err = 0;
911
Alex Tomasc9de5602008-01-29 00:19:52 -0500912 /*
913 * data carry information regarding this
914 * particular group in the format specified
915 * above
916 *
917 */
918 data = page_address(page) + (i * blocksize);
919 bitmap = bh[group - first_group]->b_data;
920
921 /*
922 * We place the buddy block and bitmap block
923 * close together
924 */
925 if ((first_block + i) & 1) {
926 /* this is block of buddy */
927 BUG_ON(incore == NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400928 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500929 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400930 trace_ext4_mb_buddy_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500931 grinfo = ext4_get_group_info(sb, group);
932 grinfo->bb_fragments = 0;
933 memset(grinfo->bb_counters, 0,
Eric Sandeen19278052009-08-25 22:36:25 -0400934 sizeof(*grinfo->bb_counters) *
935 (sb->s_blocksize_bits+2));
Alex Tomasc9de5602008-01-29 00:19:52 -0500936 /*
937 * incore got set to the group block bitmap below
938 */
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500939 ext4_lock_group(sb, group);
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400940 /* init the buddy */
941 memset(data, 0xff, blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -0500942 ext4_mb_generate_buddy(sb, data, incore, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500943 ext4_unlock_group(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500944 incore = NULL;
945 } else {
946 /* this is block of bitmap */
947 BUG_ON(incore != NULL);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -0400948 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
Alex Tomasc9de5602008-01-29 00:19:52 -0500949 group, page->index, i * blocksize);
Theodore Ts'of3073332010-05-17 03:00:00 -0400950 trace_ext4_mb_bitmap_load(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500951
952 /* see comments in ext4_mb_put_pa() */
953 ext4_lock_group(sb, group);
954 memcpy(data, bitmap, blocksize);
955
956 /* mark all preallocated blks used in in-core bitmap */
957 ext4_mb_generate_from_pa(sb, data, group);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -0500958 ext4_mb_generate_from_freelist(sb, data, group);
Alex Tomasc9de5602008-01-29 00:19:52 -0500959 ext4_unlock_group(sb, group);
960
961 /* set incore so that the buddy information can be
962 * generated using this
963 */
964 incore = data;
965 }
966 }
967 SetPageUptodate(page);
968
969out:
970 if (bh) {
Amir Goldstein9b8b7d32011-05-09 21:49:42 -0400971 for (i = 0; i < groups_per_page; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -0500972 brelse(bh[i]);
973 if (bh != &bhs)
974 kfree(bh);
975 }
976 return err;
977}
978
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -0400979/*
Amir Goldstein2de88072011-05-09 21:48:13 -0400980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
981 * on the same buddy page doesn't happen whild holding the buddy page lock.
982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400984 */
Amir Goldstein2de88072011-05-09 21:48:13 -0400985static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
986 ext4_group_t group, struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400987{
Amir Goldstein2de88072011-05-09 21:48:13 -0400988 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
989 int block, pnum, poff;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400990 int blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -0400991 struct page *page;
992
993 e4b->bd_buddy_page = NULL;
994 e4b->bd_bitmap_page = NULL;
Eric Sandeeneee4adc2010-10-27 21:30:15 -0400995
996 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
997 /*
998 * the buddy cache inode stores the block bitmap
999 * and buddy information in consecutive blocks.
1000 * So for each group we need two blocks.
1001 */
1002 block = group * 2;
1003 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001004 poff = block % blocks_per_page;
1005 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1006 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -04001007 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001008 BUG_ON(page->mapping != inode->i_mapping);
1009 e4b->bd_bitmap_page = page;
1010 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001011
Amir Goldstein2de88072011-05-09 21:48:13 -04001012 if (blocks_per_page >= 2) {
1013 /* buddy and bitmap are on the same page */
1014 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001015 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001016
1017 block++;
1018 pnum = block / blocks_per_page;
Amir Goldstein2de88072011-05-09 21:48:13 -04001019 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1020 if (!page)
Younger Liuc57ab392014-04-10 23:03:43 -04001021 return -ENOMEM;
Amir Goldstein2de88072011-05-09 21:48:13 -04001022 BUG_ON(page->mapping != inode->i_mapping);
1023 e4b->bd_buddy_page = page;
1024 return 0;
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001025}
1026
Amir Goldstein2de88072011-05-09 21:48:13 -04001027static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001028{
Amir Goldstein2de88072011-05-09 21:48:13 -04001029 if (e4b->bd_bitmap_page) {
1030 unlock_page(e4b->bd_bitmap_page);
1031 page_cache_release(e4b->bd_bitmap_page);
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001032 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001033 if (e4b->bd_buddy_page) {
1034 unlock_page(e4b->bd_buddy_page);
1035 page_cache_release(e4b->bd_buddy_page);
1036 }
Eric Sandeeneee4adc2010-10-27 21:30:15 -04001037}
1038
1039/*
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001040 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1043 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001044static noinline_for_stack
1045int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1046{
1047
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001048 struct ext4_group_info *this_grp;
Amir Goldstein2de88072011-05-09 21:48:13 -04001049 struct ext4_buddy e4b;
1050 struct page *page;
1051 int ret = 0;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001052
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001053 might_sleep();
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001054 mb_debug(1, "init group %u\n", group);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001055 this_grp = ext4_get_group_info(sb, group);
1056 /*
Aneesh Kumar K.V08c3a812009-09-09 23:50:17 -04001057 * This ensures that we don't reinit the buddy cache
1058 * page which map to the group from which we are already
1059 * allocating. If we are looking at the buddy cache we would
1060 * have taken a reference using ext4_mb_load_buddy and that
Amir Goldstein2de88072011-05-09 21:48:13 -04001061 * would have pinned buddy page to page cache.
Mel Gorman2457aec2014-06-04 16:10:31 -07001062 * The call to ext4_mb_get_buddy_page_lock will mark the
1063 * page accessed.
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001064 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001065 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1066 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001067 /*
1068 * somebody initialized the group
1069 * return without doing anything
1070 */
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001071 goto err;
1072 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001073
1074 page = e4b.bd_bitmap_page;
1075 ret = ext4_mb_init_cache(page, NULL);
1076 if (ret)
1077 goto err;
1078 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001079 ret = -EIO;
1080 goto err;
1081 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001082
Amir Goldstein2de88072011-05-09 21:48:13 -04001083 if (e4b.bd_buddy_page == NULL) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001084 /*
1085 * If both the bitmap and buddy are in
1086 * the same page we don't need to force
1087 * init the buddy
1088 */
Amir Goldstein2de88072011-05-09 21:48:13 -04001089 ret = 0;
1090 goto err;
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001091 }
Amir Goldstein2de88072011-05-09 21:48:13 -04001092 /* init buddy cache */
1093 page = e4b.bd_buddy_page;
1094 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1095 if (ret)
1096 goto err;
1097 if (!PageUptodate(page)) {
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001098 ret = -EIO;
1099 goto err;
1100 }
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001101err:
Amir Goldstein2de88072011-05-09 21:48:13 -04001102 ext4_mb_put_buddy_page_lock(&e4b);
Aneesh Kumar K.Vb6a758e2009-09-09 23:47:46 -04001103 return ret;
1104}
1105
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001106/*
1107 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1110 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04001111static noinline_for_stack int
1112ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1113 struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001114{
Alex Tomasc9de5602008-01-29 00:19:52 -05001115 int blocks_per_page;
1116 int block;
1117 int pnum;
1118 int poff;
1119 struct page *page;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001120 int ret;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001121 struct ext4_group_info *grp;
1122 struct ext4_sb_info *sbi = EXT4_SB(sb);
1123 struct inode *inode = sbi->s_buddy_cache;
Alex Tomasc9de5602008-01-29 00:19:52 -05001124
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04001125 might_sleep();
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04001126 mb_debug(1, "load group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001127
1128 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001129 grp = ext4_get_group_info(sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001130
1131 e4b->bd_blkbits = sb->s_blocksize_bits;
Tao Ma529da702011-07-23 16:07:26 -04001132 e4b->bd_info = grp;
Alex Tomasc9de5602008-01-29 00:19:52 -05001133 e4b->bd_sb = sb;
1134 e4b->bd_group = group;
1135 e4b->bd_buddy_page = NULL;
1136 e4b->bd_bitmap_page = NULL;
1137
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001138 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001139 /*
1140 * we need full data about the group
1141 * to make a good selection
1142 */
1143 ret = ext4_mb_init_group(sb, group);
1144 if (ret)
1145 return ret;
Aneesh Kumar K.Vf41c0752009-09-09 23:34:50 -04001146 }
1147
Alex Tomasc9de5602008-01-29 00:19:52 -05001148 /*
1149 * the buddy cache inode stores the block bitmap
1150 * and buddy information in consecutive blocks.
1151 * So for each group we need two blocks.
1152 */
1153 block = group * 2;
1154 pnum = block / blocks_per_page;
1155 poff = block % blocks_per_page;
1156
1157 /* we could use find_or_create_page(), but it locks page
1158 * what we'd like to avoid in fast path ... */
Mel Gorman2457aec2014-06-04 16:10:31 -07001159 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001160 if (page == NULL || !PageUptodate(page)) {
1161 if (page)
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05001162 /*
1163 * drop the page reference and try
1164 * to get the page with lock. If we
1165 * are not uptodate that implies
1166 * somebody just created the page but
1167 * is yet to initialize the same. So
1168 * wait for it to initialize.
1169 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001170 page_cache_release(page);
1171 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1172 if (page) {
1173 BUG_ON(page->mapping != inode->i_mapping);
1174 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001175 ret = ext4_mb_init_cache(page, NULL);
1176 if (ret) {
1177 unlock_page(page);
1178 goto err;
1179 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001180 mb_cmp_bitmaps(e4b, page_address(page) +
1181 (poff * sb->s_blocksize));
1182 }
1183 unlock_page(page);
1184 }
1185 }
Younger Liuc57ab392014-04-10 23:03:43 -04001186 if (page == NULL) {
1187 ret = -ENOMEM;
1188 goto err;
1189 }
1190 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001191 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001192 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001193 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001194
1195 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001196 e4b->bd_bitmap_page = page;
1197 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001198
1199 block++;
1200 pnum = block / blocks_per_page;
1201 poff = block % blocks_per_page;
1202
Mel Gorman2457aec2014-06-04 16:10:31 -07001203 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
Alex Tomasc9de5602008-01-29 00:19:52 -05001204 if (page == NULL || !PageUptodate(page)) {
1205 if (page)
1206 page_cache_release(page);
1207 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1208 if (page) {
1209 BUG_ON(page->mapping != inode->i_mapping);
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001210 if (!PageUptodate(page)) {
1211 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1212 if (ret) {
1213 unlock_page(page);
1214 goto err;
1215 }
1216 }
Alex Tomasc9de5602008-01-29 00:19:52 -05001217 unlock_page(page);
1218 }
1219 }
Younger Liuc57ab392014-04-10 23:03:43 -04001220 if (page == NULL) {
1221 ret = -ENOMEM;
1222 goto err;
1223 }
1224 if (!PageUptodate(page)) {
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001225 ret = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05001226 goto err;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001227 }
Mel Gorman2457aec2014-06-04 16:10:31 -07001228
1229 /* Pages marked accessed already */
Alex Tomasc9de5602008-01-29 00:19:52 -05001230 e4b->bd_buddy_page = page;
1231 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
Alex Tomasc9de5602008-01-29 00:19:52 -05001232
1233 BUG_ON(e4b->bd_bitmap_page == NULL);
1234 BUG_ON(e4b->bd_buddy_page == NULL);
1235
1236 return 0;
1237
1238err:
Yang Ruirui26626f112011-04-16 19:17:48 -04001239 if (page)
1240 page_cache_release(page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001241 if (e4b->bd_bitmap_page)
1242 page_cache_release(e4b->bd_bitmap_page);
1243 if (e4b->bd_buddy_page)
1244 page_cache_release(e4b->bd_buddy_page);
1245 e4b->bd_buddy = NULL;
1246 e4b->bd_bitmap = NULL;
Shen Fengfdf6c7a2008-07-11 19:27:31 -04001247 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05001248}
1249
Jing Zhange39e07f2010-05-14 00:00:00 -04001250static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
Alex Tomasc9de5602008-01-29 00:19:52 -05001251{
1252 if (e4b->bd_bitmap_page)
1253 page_cache_release(e4b->bd_bitmap_page);
1254 if (e4b->bd_buddy_page)
1255 page_cache_release(e4b->bd_buddy_page);
1256}
1257
1258
1259static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1260{
1261 int order = 1;
1262 void *bb;
1263
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001264 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
Alex Tomasc9de5602008-01-29 00:19:52 -05001265 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1266
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001267 bb = e4b->bd_buddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05001268 while (order <= e4b->bd_blkbits + 1) {
1269 block = block >> 1;
1270 if (!mb_test_bit(block, bb)) {
1271 /* this block is part of buddy of order 'order' */
1272 return order;
1273 }
1274 bb += 1 << (e4b->bd_blkbits - order);
1275 order++;
1276 }
1277 return 0;
1278}
1279
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001280static void mb_clear_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001281{
1282 __u32 *addr;
1283
1284 len = cur + len;
1285 while (cur < len) {
1286 if ((cur & 31) == 0 && (len - cur) >= 32) {
1287 /* fast path: clear whole word at once */
1288 addr = bm + (cur >> 3);
1289 *addr = 0;
1290 cur += 32;
1291 continue;
1292 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001293 mb_clear_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001294 cur++;
1295 }
1296}
1297
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001298/* clear bits in given range
1299 * will return first found zero bit if any, -1 otherwise
1300 */
1301static int mb_test_and_clear_bits(void *bm, int cur, int len)
1302{
1303 __u32 *addr;
1304 int zero_bit = -1;
1305
1306 len = cur + len;
1307 while (cur < len) {
1308 if ((cur & 31) == 0 && (len - cur) >= 32) {
1309 /* fast path: clear whole word at once */
1310 addr = bm + (cur >> 3);
1311 if (*addr != (__u32)(-1) && zero_bit == -1)
1312 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1313 *addr = 0;
1314 cur += 32;
1315 continue;
1316 }
1317 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1318 zero_bit = cur;
1319 cur++;
1320 }
1321
1322 return zero_bit;
1323}
1324
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04001325void ext4_set_bits(void *bm, int cur, int len)
Alex Tomasc9de5602008-01-29 00:19:52 -05001326{
1327 __u32 *addr;
1328
1329 len = cur + len;
1330 while (cur < len) {
1331 if ((cur & 31) == 0 && (len - cur) >= 32) {
1332 /* fast path: set whole word at once */
1333 addr = bm + (cur >> 3);
1334 *addr = 0xffffffff;
1335 cur += 32;
1336 continue;
1337 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04001338 mb_set_bit(cur, bm);
Alex Tomasc9de5602008-01-29 00:19:52 -05001339 cur++;
1340 }
1341}
1342
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001343/*
1344 * _________________________________________________________________ */
1345
1346static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
Alex Tomasc9de5602008-01-29 00:19:52 -05001347{
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001348 if (mb_test_bit(*bit + side, bitmap)) {
1349 mb_clear_bit(*bit, bitmap);
1350 (*bit) -= side;
1351 return 1;
1352 }
1353 else {
1354 (*bit) += side;
1355 mb_set_bit(*bit, bitmap);
1356 return -1;
1357 }
1358}
1359
1360static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1361{
1362 int max;
1363 int order = 1;
1364 void *buddy = mb_find_buddy(e4b, order, &max);
1365
1366 while (buddy) {
1367 void *buddy2;
1368
1369 /* Bits in range [first; last] are known to be set since
1370 * corresponding blocks were allocated. Bits in range
1371 * (first; last) will stay set because they form buddies on
1372 * upper layer. We just deal with borders if they don't
1373 * align with upper layer and then go up.
1374 * Releasing entire group is all about clearing
1375 * single bit of highest order buddy.
1376 */
1377
1378 /* Example:
1379 * ---------------------------------
1380 * | 1 | 1 | 1 | 1 |
1381 * ---------------------------------
1382 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1383 * ---------------------------------
1384 * 0 1 2 3 4 5 6 7
1385 * \_____________________/
1386 *
1387 * Neither [1] nor [6] is aligned to above layer.
1388 * Left neighbour [0] is free, so mark it busy,
1389 * decrease bb_counters and extend range to
1390 * [0; 6]
1391 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1392 * mark [6] free, increase bb_counters and shrink range to
1393 * [0; 5].
1394 * Then shift range to [0; 2], go up and do the same.
1395 */
1396
1397
1398 if (first & 1)
1399 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1400 if (!(last & 1))
1401 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1402 if (first > last)
1403 break;
1404 order++;
1405
1406 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1407 mb_clear_bits(buddy, first, last - first + 1);
1408 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1409 break;
1410 }
1411 first >>= 1;
1412 last >>= 1;
1413 buddy = buddy2;
1414 }
1415}
1416
1417static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1418 int first, int count)
1419{
1420 int left_is_free = 0;
1421 int right_is_free = 0;
1422 int block;
1423 int last = first + count - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001424 struct super_block *sb = e4b->bd_sb;
1425
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04001426 if (WARN_ON(count == 0))
1427 return;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001428 BUG_ON(last >= (sb->s_blocksize << 3));
Vincent Minetbc8e6742009-05-15 08:33:18 -04001429 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
Darrick J. Wong163a2032013-08-28 17:35:51 -04001430 /* Don't bother if the block group is corrupt. */
1431 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1432 return;
1433
Alex Tomasc9de5602008-01-29 00:19:52 -05001434 mb_check_buddy(e4b);
1435 mb_free_blocks_double(inode, e4b, first, count);
1436
1437 e4b->bd_info->bb_free += count;
1438 if (first < e4b->bd_info->bb_first_free)
1439 e4b->bd_info->bb_first_free = first;
1440
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001441 /* access memory sequentially: check left neighbour,
1442 * clear range and then check right neighbour
1443 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001444 if (first != 0)
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001445 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1446 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1447 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1448 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1449
1450 if (unlikely(block != -1)) {
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001451 struct ext4_sb_info *sbi = EXT4_SB(sb);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001452 ext4_fsblk_t blocknr;
1453
1454 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1455 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1456 ext4_grp_locked_error(sb, e4b->bd_group,
1457 inode ? inode->i_ino : 0,
1458 blocknr,
1459 "freeing already freed block "
Darrick J. Wong163a2032013-08-28 17:35:51 -04001460 "(bit %u); block bitmap corrupt.",
1461 block);
Namjae Jeone43bb4e2014-06-26 10:11:53 -04001462 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1463 percpu_counter_sub(&sbi->s_freeclusters_counter,
1464 e4b->bd_info->bb_free);
Darrick J. Wong163a2032013-08-28 17:35:51 -04001465 /* Mark the block group as corrupt. */
1466 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1467 &e4b->bd_info->bb_state);
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001468 mb_regenerate_buddy(e4b);
1469 goto done;
1470 }
1471
1472 /* let's maintain fragments counter */
1473 if (left_is_free && right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001474 e4b->bd_info->bb_fragments--;
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001475 else if (!left_is_free && !right_is_free)
Alex Tomasc9de5602008-01-29 00:19:52 -05001476 e4b->bd_info->bb_fragments++;
1477
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001478 /* buddy[0] == bd_bitmap is a special case, so handle
1479 * it right away and let mb_buddy_mark_free stay free of
1480 * zero order checks.
1481 * Check if neighbours are to be coaleasced,
1482 * adjust bitmap bb_counters and borders appropriately.
1483 */
1484 if (first & 1) {
1485 first += !left_is_free;
1486 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05001487 }
Andrey Sidoroveabe0442013-04-09 12:22:29 -04001488 if (!(last & 1)) {
1489 last -= !right_is_free;
1490 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1491 }
1492
1493 if (first <= last)
1494 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1495
1496done:
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001497 mb_set_largest_free_order(sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001498 mb_check_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001499}
1500
Robin Dong15c006a2012-08-17 10:02:17 -04001501static int mb_find_extent(struct ext4_buddy *e4b, int block,
Alex Tomasc9de5602008-01-29 00:19:52 -05001502 int needed, struct ext4_free_extent *ex)
1503{
1504 int next = block;
Robin Dong15c006a2012-08-17 10:02:17 -04001505 int max, order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001506 void *buddy;
1507
Vincent Minetbc8e6742009-05-15 08:33:18 -04001508 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001509 BUG_ON(ex == NULL);
1510
Robin Dong15c006a2012-08-17 10:02:17 -04001511 buddy = mb_find_buddy(e4b, 0, &max);
Alex Tomasc9de5602008-01-29 00:19:52 -05001512 BUG_ON(buddy == NULL);
1513 BUG_ON(block >= max);
1514 if (mb_test_bit(block, buddy)) {
1515 ex->fe_len = 0;
1516 ex->fe_start = 0;
1517 ex->fe_group = 0;
1518 return 0;
1519 }
1520
Robin Dong15c006a2012-08-17 10:02:17 -04001521 /* find actual order */
1522 order = mb_find_order_for_block(e4b, block);
1523 block = block >> order;
Alex Tomasc9de5602008-01-29 00:19:52 -05001524
1525 ex->fe_len = 1 << order;
1526 ex->fe_start = block << order;
1527 ex->fe_group = e4b->bd_group;
1528
1529 /* calc difference from given start */
1530 next = next - ex->fe_start;
1531 ex->fe_len -= next;
1532 ex->fe_start += next;
1533
1534 while (needed > ex->fe_len &&
Alan Coxd8ec0c32012-11-08 12:19:58 -05001535 mb_find_buddy(e4b, order, &max)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001536
1537 if (block + 1 >= max)
1538 break;
1539
1540 next = (block + 1) * (1 << order);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001541 if (mb_test_bit(next, e4b->bd_bitmap))
Alex Tomasc9de5602008-01-29 00:19:52 -05001542 break;
1543
Robin Dongb051d8d2011-10-26 05:30:30 -04001544 order = mb_find_order_for_block(e4b, next);
Alex Tomasc9de5602008-01-29 00:19:52 -05001545
Alex Tomasc9de5602008-01-29 00:19:52 -05001546 block = next >> order;
1547 ex->fe_len += 1 << order;
1548 }
1549
1550 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1551 return ex->fe_len;
1552}
1553
1554static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1555{
1556 int ord;
1557 int mlen = 0;
1558 int max = 0;
1559 int cur;
1560 int start = ex->fe_start;
1561 int len = ex->fe_len;
1562 unsigned ret = 0;
1563 int len0 = len;
1564 void *buddy;
1565
1566 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1567 BUG_ON(e4b->bd_group != ex->fe_group);
Vincent Minetbc8e6742009-05-15 08:33:18 -04001568 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
Alex Tomasc9de5602008-01-29 00:19:52 -05001569 mb_check_buddy(e4b);
1570 mb_mark_used_double(e4b, start, len);
1571
1572 e4b->bd_info->bb_free -= len;
1573 if (e4b->bd_info->bb_first_free == start)
1574 e4b->bd_info->bb_first_free += len;
1575
1576 /* let's maintain fragments counter */
1577 if (start != 0)
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001578 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001579 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001580 max = !mb_test_bit(start + len, e4b->bd_bitmap);
Alex Tomasc9de5602008-01-29 00:19:52 -05001581 if (mlen && max)
1582 e4b->bd_info->bb_fragments++;
1583 else if (!mlen && !max)
1584 e4b->bd_info->bb_fragments--;
1585
1586 /* let's maintain buddy itself */
1587 while (len) {
1588 ord = mb_find_order_for_block(e4b, start);
1589
1590 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1591 /* the whole chunk may be allocated at once! */
1592 mlen = 1 << ord;
1593 buddy = mb_find_buddy(e4b, ord, &max);
1594 BUG_ON((start >> ord) >= max);
1595 mb_set_bit(start >> ord, buddy);
1596 e4b->bd_info->bb_counters[ord]--;
1597 start += mlen;
1598 len -= mlen;
1599 BUG_ON(len < 0);
1600 continue;
1601 }
1602
1603 /* store for history */
1604 if (ret == 0)
1605 ret = len | (ord << 16);
1606
1607 /* we have to split large buddy */
1608 BUG_ON(ord <= 0);
1609 buddy = mb_find_buddy(e4b, ord, &max);
1610 mb_set_bit(start >> ord, buddy);
1611 e4b->bd_info->bb_counters[ord]--;
1612
1613 ord--;
1614 cur = (start >> ord) & ~1U;
1615 buddy = mb_find_buddy(e4b, ord, &max);
1616 mb_clear_bit(cur, buddy);
1617 mb_clear_bit(cur + 1, buddy);
1618 e4b->bd_info->bb_counters[ord]++;
1619 e4b->bd_info->bb_counters[ord]++;
1620 }
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04001621 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05001622
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001623 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001624 mb_check_buddy(e4b);
1625
1626 return ret;
1627}
1628
1629/*
1630 * Must be called under group lock!
1631 */
1632static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1633 struct ext4_buddy *e4b)
1634{
1635 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1636 int ret;
1637
1638 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1639 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1640
1641 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1642 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1643 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1644
1645 /* preallocation can change ac_b_ex, thus we store actually
1646 * allocated blocks for history */
1647 ac->ac_f_ex = ac->ac_b_ex;
1648
1649 ac->ac_status = AC_STATUS_FOUND;
1650 ac->ac_tail = ret & 0xffff;
1651 ac->ac_buddy = ret >> 16;
1652
Aneesh Kumar K.Vc3a326a2008-11-25 15:11:52 -05001653 /*
1654 * take the page reference. We want the page to be pinned
1655 * so that we don't get a ext4_mb_init_cache_call for this
1656 * group until we update the bitmap. That would mean we
1657 * double allocate blocks. The reference is dropped
1658 * in ext4_mb_release_context
1659 */
Alex Tomasc9de5602008-01-29 00:19:52 -05001660 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1661 get_page(ac->ac_bitmap_page);
1662 ac->ac_buddy_page = e4b->bd_buddy_page;
1663 get_page(ac->ac_buddy_page);
Alex Tomasc9de5602008-01-29 00:19:52 -05001664 /* store last allocated for subsequent stream allocation */
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04001665 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05001666 spin_lock(&sbi->s_md_lock);
1667 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1668 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1669 spin_unlock(&sbi->s_md_lock);
1670 }
1671}
1672
1673/*
1674 * regular allocator, for general purposes allocation
1675 */
1676
1677static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1678 struct ext4_buddy *e4b,
1679 int finish_group)
1680{
1681 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1682 struct ext4_free_extent *bex = &ac->ac_b_ex;
1683 struct ext4_free_extent *gex = &ac->ac_g_ex;
1684 struct ext4_free_extent ex;
1685 int max;
1686
Aneesh Kumar K.V032115f2009-01-05 21:34:30 -05001687 if (ac->ac_status == AC_STATUS_FOUND)
1688 return;
Alex Tomasc9de5602008-01-29 00:19:52 -05001689 /*
1690 * We don't want to scan for a whole year
1691 */
1692 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1693 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1694 ac->ac_status = AC_STATUS_BREAK;
1695 return;
1696 }
1697
1698 /*
1699 * Haven't found good chunk so far, let's continue
1700 */
1701 if (bex->fe_len < gex->fe_len)
1702 return;
1703
1704 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1705 && bex->fe_group == e4b->bd_group) {
1706 /* recheck chunk's availability - we don't know
1707 * when it was found (within this lock-unlock
1708 * period or not) */
Robin Dong15c006a2012-08-17 10:02:17 -04001709 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001710 if (max >= gex->fe_len) {
1711 ext4_mb_use_best_found(ac, e4b);
1712 return;
1713 }
1714 }
1715}
1716
1717/*
1718 * The routine checks whether found extent is good enough. If it is,
1719 * then the extent gets marked used and flag is set to the context
1720 * to stop scanning. Otherwise, the extent is compared with the
1721 * previous found extent and if new one is better, then it's stored
1722 * in the context. Later, the best found extent will be used, if
1723 * mballoc can't find good enough extent.
1724 *
1725 * FIXME: real allocation policy is to be designed yet!
1726 */
1727static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1728 struct ext4_free_extent *ex,
1729 struct ext4_buddy *e4b)
1730{
1731 struct ext4_free_extent *bex = &ac->ac_b_ex;
1732 struct ext4_free_extent *gex = &ac->ac_g_ex;
1733
1734 BUG_ON(ex->fe_len <= 0);
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001735 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1736 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05001737 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1738
1739 ac->ac_found++;
1740
1741 /*
1742 * The special case - take what you catch first
1743 */
1744 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1745 *bex = *ex;
1746 ext4_mb_use_best_found(ac, e4b);
1747 return;
1748 }
1749
1750 /*
1751 * Let's check whether the chuck is good enough
1752 */
1753 if (ex->fe_len == gex->fe_len) {
1754 *bex = *ex;
1755 ext4_mb_use_best_found(ac, e4b);
1756 return;
1757 }
1758
1759 /*
1760 * If this is first found extent, just store it in the context
1761 */
1762 if (bex->fe_len == 0) {
1763 *bex = *ex;
1764 return;
1765 }
1766
1767 /*
1768 * If new found extent is better, store it in the context
1769 */
1770 if (bex->fe_len < gex->fe_len) {
1771 /* if the request isn't satisfied, any found extent
1772 * larger than previous best one is better */
1773 if (ex->fe_len > bex->fe_len)
1774 *bex = *ex;
1775 } else if (ex->fe_len > gex->fe_len) {
1776 /* if the request is satisfied, then we try to find
1777 * an extent that still satisfy the request, but is
1778 * smaller than previous one */
1779 if (ex->fe_len < bex->fe_len)
1780 *bex = *ex;
1781 }
1782
1783 ext4_mb_check_limits(ac, e4b, 0);
1784}
1785
Eric Sandeen089ceec2009-07-05 22:17:31 -04001786static noinline_for_stack
1787int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001788 struct ext4_buddy *e4b)
1789{
1790 struct ext4_free_extent ex = ac->ac_b_ex;
1791 ext4_group_t group = ex.fe_group;
1792 int max;
1793 int err;
1794
1795 BUG_ON(ex.fe_len <= 0);
1796 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1797 if (err)
1798 return err;
1799
1800 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001801 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001802
1803 if (max > 0) {
1804 ac->ac_b_ex = ex;
1805 ext4_mb_use_best_found(ac, e4b);
1806 }
1807
1808 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001809 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001810
1811 return 0;
1812}
1813
Eric Sandeen089ceec2009-07-05 22:17:31 -04001814static noinline_for_stack
1815int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001816 struct ext4_buddy *e4b)
1817{
1818 ext4_group_t group = ac->ac_g_ex.fe_group;
1819 int max;
1820 int err;
1821 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001822 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05001823 struct ext4_free_extent ex;
1824
1825 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1826 return 0;
Yongqiang Yang838cd0c2012-09-23 23:10:51 -04001827 if (grp->bb_free == 0)
1828 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05001829
1830 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1831 if (err)
1832 return err;
1833
Darrick J. Wong163a2032013-08-28 17:35:51 -04001834 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1835 ext4_mb_unload_buddy(e4b);
1836 return 0;
1837 }
1838
Alex Tomasc9de5602008-01-29 00:19:52 -05001839 ext4_lock_group(ac->ac_sb, group);
Robin Dong15c006a2012-08-17 10:02:17 -04001840 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
Alex Tomasc9de5602008-01-29 00:19:52 -05001841 ac->ac_g_ex.fe_len, &ex);
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001842 ex.fe_logical = 0xDEADFA11; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001843
1844 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1845 ext4_fsblk_t start;
1846
Akinobu Mita5661bd62010-03-03 23:53:39 -05001847 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1848 ex.fe_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05001849 /* use do_div to get remainder (would be 64-bit modulo) */
1850 if (do_div(start, sbi->s_stripe) == 0) {
1851 ac->ac_found++;
1852 ac->ac_b_ex = ex;
1853 ext4_mb_use_best_found(ac, e4b);
1854 }
1855 } else if (max >= ac->ac_g_ex.fe_len) {
1856 BUG_ON(ex.fe_len <= 0);
1857 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1858 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1859 ac->ac_found++;
1860 ac->ac_b_ex = ex;
1861 ext4_mb_use_best_found(ac, e4b);
1862 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1863 /* Sometimes, caller may want to merge even small
1864 * number of blocks to an existing extent */
1865 BUG_ON(ex.fe_len <= 0);
1866 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1867 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1868 ac->ac_found++;
1869 ac->ac_b_ex = ex;
1870 ext4_mb_use_best_found(ac, e4b);
1871 }
1872 ext4_unlock_group(ac->ac_sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04001873 ext4_mb_unload_buddy(e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05001874
1875 return 0;
1876}
1877
1878/*
1879 * The routine scans buddy structures (not bitmap!) from given order
1880 * to max order and tries to find big enough chunk to satisfy the req
1881 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001882static noinline_for_stack
1883void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001884 struct ext4_buddy *e4b)
1885{
1886 struct super_block *sb = ac->ac_sb;
1887 struct ext4_group_info *grp = e4b->bd_info;
1888 void *buddy;
1889 int i;
1890 int k;
1891 int max;
1892
1893 BUG_ON(ac->ac_2order <= 0);
1894 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1895 if (grp->bb_counters[i] == 0)
1896 continue;
1897
1898 buddy = mb_find_buddy(e4b, i, &max);
1899 BUG_ON(buddy == NULL);
1900
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001901 k = mb_find_next_zero_bit(buddy, max, 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05001902 BUG_ON(k >= max);
1903
1904 ac->ac_found++;
1905
1906 ac->ac_b_ex.fe_len = 1 << i;
1907 ac->ac_b_ex.fe_start = k << i;
1908 ac->ac_b_ex.fe_group = e4b->bd_group;
1909
1910 ext4_mb_use_best_found(ac, e4b);
1911
1912 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1913
1914 if (EXT4_SB(sb)->s_mb_stats)
1915 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1916
1917 break;
1918 }
1919}
1920
1921/*
1922 * The routine scans the group and measures all found extents.
1923 * In order to optimize scanning, caller must pass number of
1924 * free blocks in the group, so the routine can know upper limit.
1925 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001926static noinline_for_stack
1927void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001928 struct ext4_buddy *e4b)
1929{
1930 struct super_block *sb = ac->ac_sb;
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001931 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001932 struct ext4_free_extent ex;
1933 int i;
1934 int free;
1935
1936 free = e4b->bd_info->bb_free;
1937 BUG_ON(free <= 0);
1938
1939 i = e4b->bd_info->bb_first_free;
1940
1941 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05001942 i = mb_find_next_zero_bit(bitmap,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04001943 EXT4_CLUSTERS_PER_GROUP(sb), i);
1944 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001945 /*
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001946 * IF we have corrupt bitmap, we won't find any
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001947 * free blocks even though group info says we
1948 * we have free blocks
1949 */
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001950 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001951 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001952 "group info. But bitmap says 0",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001953 free);
Alex Tomasc9de5602008-01-29 00:19:52 -05001954 break;
1955 }
1956
Robin Dong15c006a2012-08-17 10:02:17 -04001957 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05001958 BUG_ON(ex.fe_len <= 0);
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001959 if (free < ex.fe_len) {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04001960 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04001961 "%d free clusters as per "
Theodore Ts'ofde4d952009-01-05 22:17:35 -05001962 "group info. But got %d blocks",
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001963 free, ex.fe_len);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05001964 /*
1965 * The number of free blocks differs. This mostly
1966 * indicate that the bitmap is corrupt. So exit
1967 * without claiming the space.
1968 */
1969 break;
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05001970 }
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05001971 ex.fe_logical = 0xDEADC0DE; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05001972 ext4_mb_measure_extent(ac, &ex, e4b);
1973
1974 i += ex.fe_len;
1975 free -= ex.fe_len;
1976 }
1977
1978 ext4_mb_check_limits(ac, e4b, 1);
1979}
1980
1981/*
1982 * This is a special case for storages like raid5
Eric Sandeen506bf2d2010-07-27 11:56:06 -04001983 * we try to find stripe-aligned chunks for stripe-size-multiple requests
Alex Tomasc9de5602008-01-29 00:19:52 -05001984 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04001985static noinline_for_stack
1986void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05001987 struct ext4_buddy *e4b)
1988{
1989 struct super_block *sb = ac->ac_sb;
1990 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'oc5e8f3f2012-02-20 17:54:06 -05001991 void *bitmap = e4b->bd_bitmap;
Alex Tomasc9de5602008-01-29 00:19:52 -05001992 struct ext4_free_extent ex;
1993 ext4_fsblk_t first_group_block;
1994 ext4_fsblk_t a;
1995 ext4_grpblk_t i;
1996 int max;
1997
1998 BUG_ON(sbi->s_stripe == 0);
1999
2000 /* find first stripe-aligned block in group */
Akinobu Mita5661bd62010-03-03 23:53:39 -05002001 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2002
Alex Tomasc9de5602008-01-29 00:19:52 -05002003 a = first_group_block + sbi->s_stripe - 1;
2004 do_div(a, sbi->s_stripe);
2005 i = (a * sbi->s_stripe) - first_group_block;
2006
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04002007 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002008 if (!mb_test_bit(i, bitmap)) {
Robin Dong15c006a2012-08-17 10:02:17 -04002009 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002010 if (max >= sbi->s_stripe) {
2011 ac->ac_found++;
Theodore Ts'oab0c00f2014-02-20 00:36:41 -05002012 ex.fe_logical = 0xDEADF00D; /* debug value */
Alex Tomasc9de5602008-01-29 00:19:52 -05002013 ac->ac_b_ex = ex;
2014 ext4_mb_use_best_found(ac, e4b);
2015 break;
2016 }
2017 }
2018 i += sbi->s_stripe;
2019 }
2020}
2021
Lukas Czerner42ac1842015-06-08 11:40:40 -04002022/*
2023 * This is now called BEFORE we load the buddy bitmap.
2024 * Returns either 1 or 0 indicating that the group is either suitable
2025 * for the allocation or not. In addition it can also return negative
2026 * error code when something goes wrong.
2027 */
Alex Tomasc9de5602008-01-29 00:19:52 -05002028static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2029 ext4_group_t group, int cr)
2030{
2031 unsigned free, fragments;
Theodore Ts'oa4912122009-03-12 12:18:34 -04002032 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05002033 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2034
2035 BUG_ON(cr < 0 || cr >= 4);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002036
Theodore Ts'o01fc48e2012-08-17 09:46:17 -04002037 free = grp->bb_free;
2038 if (free == 0)
2039 return 0;
2040 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2041 return 0;
2042
Darrick J. Wong163a2032013-08-28 17:35:51 -04002043 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2044 return 0;
2045
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002046 /* We only do this if the grp has never been initialized */
2047 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2048 int ret = ext4_mb_init_group(ac->ac_sb, group);
2049 if (ret)
Lukas Czerner42ac1842015-06-08 11:40:40 -04002050 return ret;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002051 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002052
Alex Tomasc9de5602008-01-29 00:19:52 -05002053 fragments = grp->bb_fragments;
Alex Tomasc9de5602008-01-29 00:19:52 -05002054 if (fragments == 0)
2055 return 0;
2056
2057 switch (cr) {
2058 case 0:
2059 BUG_ON(ac->ac_2order == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05002060
Theodore Ts'oa4912122009-03-12 12:18:34 -04002061 /* Avoid using the first bg of a flexgroup for data files */
2062 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2063 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2064 ((group % flex_size) == 0))
2065 return 0;
2066
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002067 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2068 (free / fragments) >= ac->ac_g_ex.fe_len)
2069 return 1;
2070
2071 if (grp->bb_largest_free_order < ac->ac_2order)
2072 return 0;
2073
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002074 return 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002075 case 1:
2076 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2077 return 1;
2078 break;
2079 case 2:
2080 if (free >= ac->ac_g_ex.fe_len)
2081 return 1;
2082 break;
2083 case 3:
2084 return 1;
2085 default:
2086 BUG();
2087 }
2088
2089 return 0;
2090}
2091
Eric Sandeen4ddfef72008-04-29 08:11:12 -04002092static noinline_for_stack int
2093ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05002094{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002095 ext4_group_t ngroups, group, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002096 int cr;
Lukas Czerner42ac1842015-06-08 11:40:40 -04002097 int err = 0, first_err = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002098 struct ext4_sb_info *sbi;
2099 struct super_block *sb;
2100 struct ext4_buddy e4b;
Alex Tomasc9de5602008-01-29 00:19:52 -05002101
2102 sb = ac->ac_sb;
2103 sbi = EXT4_SB(sb);
Theodore Ts'o8df96752009-05-01 08:50:38 -04002104 ngroups = ext4_get_groups_count(sb);
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002105 /* non-extent files are limited to low blocks/groups */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04002106 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04002107 ngroups = sbi->s_blockfile_groups;
2108
Alex Tomasc9de5602008-01-29 00:19:52 -05002109 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2110
2111 /* first, try the goal */
2112 err = ext4_mb_find_by_goal(ac, &e4b);
2113 if (err || ac->ac_status == AC_STATUS_FOUND)
2114 goto out;
2115
2116 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2117 goto out;
2118
2119 /*
2120 * ac->ac2_order is set only if the fe_len is a power of 2
2121 * if ac2_order is set we also set criteria to 0 so that we
2122 * try exact allocation using buddy.
2123 */
2124 i = fls(ac->ac_g_ex.fe_len);
2125 ac->ac_2order = 0;
2126 /*
2127 * We search using buddy data only if the order of the request
2128 * is greater than equal to the sbi_s_mb_order2_reqs
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002129 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
Alex Tomasc9de5602008-01-29 00:19:52 -05002130 */
2131 if (i >= sbi->s_mb_order2_reqs) {
2132 /*
2133 * This should tell if fe_len is exactly power of 2
2134 */
2135 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2136 ac->ac_2order = i - 1;
2137 }
2138
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002139 /* if stream allocation is enabled, use global goal */
2140 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002141 /* TBD: may be hot point */
2142 spin_lock(&sbi->s_md_lock);
2143 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2144 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2145 spin_unlock(&sbi->s_md_lock);
2146 }
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04002147
Alex Tomasc9de5602008-01-29 00:19:52 -05002148 /* Let's just scan groups to find more-less suitable blocks */
2149 cr = ac->ac_2order ? 0 : 1;
2150 /*
2151 * cr == 0 try to get exact allocation,
2152 * cr == 3 try to get anything
2153 */
2154repeat:
2155 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2156 ac->ac_criteria = cr;
Aneesh Kumar K.Ved8f9c72008-07-11 19:27:31 -04002157 /*
2158 * searching for the right group start
2159 * from the goal value specified
2160 */
2161 group = ac->ac_g_ex.fe_group;
2162
Theodore Ts'o8df96752009-05-01 08:50:38 -04002163 for (i = 0; i < ngroups; group++, i++) {
Lukas Czerner42ac1842015-06-08 11:40:40 -04002164 int ret = 0;
Theodore Ts'o2ed57242013-06-12 11:43:02 -04002165 cond_resched();
Lachlan McIlroye6155732013-05-05 23:10:00 -04002166 /*
2167 * Artificially restricted ngroups for non-extent
2168 * files makes group > ngroups possible on first loop.
2169 */
2170 if (group >= ngroups)
Alex Tomasc9de5602008-01-29 00:19:52 -05002171 group = 0;
2172
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002173 /* This now checks without needing the buddy page */
Lukas Czerner42ac1842015-06-08 11:40:40 -04002174 ret = ext4_mb_good_group(ac, group, cr);
2175 if (ret <= 0) {
2176 if (!first_err)
2177 first_err = ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002178 continue;
Lukas Czerner42ac1842015-06-08 11:40:40 -04002179 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002180
Alex Tomasc9de5602008-01-29 00:19:52 -05002181 err = ext4_mb_load_buddy(sb, group, &e4b);
2182 if (err)
2183 goto out;
2184
2185 ext4_lock_group(sb, group);
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002186
2187 /*
2188 * We need to check again after locking the
2189 * block group
2190 */
Lukas Czerner42ac1842015-06-08 11:40:40 -04002191 ret = ext4_mb_good_group(ac, group, cr);
2192 if (ret <= 0) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002193 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002194 ext4_mb_unload_buddy(&e4b);
Lukas Czerner42ac1842015-06-08 11:40:40 -04002195 if (!first_err)
2196 first_err = ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002197 continue;
2198 }
2199
2200 ac->ac_groups_scanned++;
Theodore Ts'o40ae3482013-02-04 15:08:40 -05002201 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
Alex Tomasc9de5602008-01-29 00:19:52 -05002202 ext4_mb_simple_scan_group(ac, &e4b);
Eric Sandeen506bf2d2010-07-27 11:56:06 -04002203 else if (cr == 1 && sbi->s_stripe &&
2204 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
Alex Tomasc9de5602008-01-29 00:19:52 -05002205 ext4_mb_scan_aligned(ac, &e4b);
2206 else
2207 ext4_mb_complex_scan_group(ac, &e4b);
2208
2209 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04002210 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002211
2212 if (ac->ac_status != AC_STATUS_CONTINUE)
2213 break;
2214 }
2215 }
2216
2217 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2218 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2219 /*
2220 * We've been searching too long. Let's try to allocate
2221 * the best chunk we've found so far
2222 */
2223
2224 ext4_mb_try_best_found(ac, &e4b);
2225 if (ac->ac_status != AC_STATUS_FOUND) {
2226 /*
2227 * Someone more lucky has already allocated it.
2228 * The only thing we can do is just take first
2229 * found block(s)
2230 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2231 */
2232 ac->ac_b_ex.fe_group = 0;
2233 ac->ac_b_ex.fe_start = 0;
2234 ac->ac_b_ex.fe_len = 0;
2235 ac->ac_status = AC_STATUS_CONTINUE;
2236 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2237 cr = 3;
2238 atomic_inc(&sbi->s_mb_lost_chunks);
2239 goto repeat;
2240 }
2241 }
2242out:
Lukas Czerner42ac1842015-06-08 11:40:40 -04002243 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2244 err = first_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002245 return err;
2246}
2247
Alex Tomasc9de5602008-01-29 00:19:52 -05002248static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2249{
2250 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002251 ext4_group_t group;
2252
Theodore Ts'o8df96752009-05-01 08:50:38 -04002253 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002254 return NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002255 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002256 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002257}
2258
2259static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2260{
2261 struct super_block *sb = seq->private;
Alex Tomasc9de5602008-01-29 00:19:52 -05002262 ext4_group_t group;
2263
2264 ++*pos;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002265 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
Alex Tomasc9de5602008-01-29 00:19:52 -05002266 return NULL;
2267 group = *pos + 1;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002268 return (void *) ((unsigned long) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002269}
2270
2271static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2272{
2273 struct super_block *sb = seq->private;
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002274 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
Alex Tomasc9de5602008-01-29 00:19:52 -05002275 int i;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002276 int err, buddy_loaded = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05002277 struct ext4_buddy e4b;
Aditya Kali1c8457c2012-06-30 19:10:57 -04002278 struct ext4_group_info *grinfo;
Alex Tomasc9de5602008-01-29 00:19:52 -05002279 struct sg {
2280 struct ext4_group_info info;
Eric Sandeena36b4492009-08-25 22:36:45 -04002281 ext4_grpblk_t counters[16];
Alex Tomasc9de5602008-01-29 00:19:52 -05002282 } sg;
2283
2284 group--;
2285 if (group == 0)
Rasmus Villemoes97b4af22015-06-15 00:32:58 -04002286 seq_puts(seq, "#group: free frags first ["
2287 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
Huaitong Han802cf1f2016-02-12 00:17:16 -05002288 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
Alex Tomasc9de5602008-01-29 00:19:52 -05002289
2290 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2291 sizeof(struct ext4_group_info);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002292 grinfo = ext4_get_group_info(sb, group);
2293 /* Load the group info in memory only if not already loaded. */
2294 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2295 err = ext4_mb_load_buddy(sb, group, &e4b);
2296 if (err) {
2297 seq_printf(seq, "#%-5u: I/O error\n", group);
2298 return 0;
2299 }
2300 buddy_loaded = 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05002301 }
Aditya Kali1c8457c2012-06-30 19:10:57 -04002302
Alex Tomasc9de5602008-01-29 00:19:52 -05002303 memcpy(&sg, ext4_get_group_info(sb, group), i);
Aditya Kali1c8457c2012-06-30 19:10:57 -04002304
2305 if (buddy_loaded)
2306 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05002307
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002308 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
Alex Tomasc9de5602008-01-29 00:19:52 -05002309 sg.info.bb_fragments, sg.info.bb_first_free);
2310 for (i = 0; i <= 13; i++)
2311 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2312 sg.info.bb_counters[i] : 0);
2313 seq_printf(seq, " ]\n");
2314
2315 return 0;
2316}
2317
2318static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2319{
2320}
2321
Tobias Klauser7f1346a2009-09-05 09:28:54 -04002322static const struct seq_operations ext4_mb_seq_groups_ops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002323 .start = ext4_mb_seq_groups_start,
2324 .next = ext4_mb_seq_groups_next,
2325 .stop = ext4_mb_seq_groups_stop,
2326 .show = ext4_mb_seq_groups_show,
2327};
2328
2329static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2330{
Al Virod9dda782013-03-31 18:16:14 -04002331 struct super_block *sb = PDE_DATA(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05002332 int rc;
2333
2334 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2335 if (rc == 0) {
Joe Perchesa271fe82010-07-27 11:56:04 -04002336 struct seq_file *m = file->private_data;
Alex Tomasc9de5602008-01-29 00:19:52 -05002337 m->private = sb;
2338 }
2339 return rc;
2340
2341}
2342
Theodore Ts'oebd173b2015-09-23 12:46:17 -04002343const struct file_operations ext4_seq_mb_groups_fops = {
Alex Tomasc9de5602008-01-29 00:19:52 -05002344 .owner = THIS_MODULE,
2345 .open = ext4_mb_seq_groups_open,
2346 .read = seq_read,
2347 .llseek = seq_lseek,
2348 .release = seq_release,
2349};
2350
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002351static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2352{
2353 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2354 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2355
2356 BUG_ON(!cachep);
2357 return cachep;
2358}
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002359
Theodore Ts'o28623c22012-09-05 01:31:50 -04002360/*
2361 * Allocate the top-level s_group_info array for the specified number
2362 * of groups
2363 */
2364int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2365{
2366 struct ext4_sb_info *sbi = EXT4_SB(sb);
2367 unsigned size;
2368 struct ext4_group_info ***new_groupinfo;
2369
2370 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2371 EXT4_DESC_PER_BLOCK_BITS(sb);
2372 if (size <= sbi->s_group_info_size)
2373 return 0;
2374
2375 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2376 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2377 if (!new_groupinfo) {
2378 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2379 return -ENOMEM;
2380 }
2381 if (sbi->s_group_info) {
2382 memcpy(new_groupinfo, sbi->s_group_info,
2383 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
Al Virob93b41d2014-11-20 12:19:11 -05002384 kvfree(sbi->s_group_info);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002385 }
2386 sbi->s_group_info = new_groupinfo;
2387 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2388 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2389 sbi->s_group_info_size);
2390 return 0;
2391}
2392
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002393/* Create and initialize ext4_group_info data for the given group. */
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002394int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002395 struct ext4_group_desc *desc)
2396{
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002397 int i;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002398 int metalen = 0;
2399 struct ext4_sb_info *sbi = EXT4_SB(sb);
2400 struct ext4_group_info **meta_group_info;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002401 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002402
2403 /*
2404 * First check if this group is the first of a reserved block.
2405 * If it's true, we have to allocate a new table of pointers
2406 * to ext4_group_info structures
2407 */
2408 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2409 metalen = sizeof(*meta_group_info) <<
2410 EXT4_DESC_PER_BLOCK_BITS(sb);
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002411 meta_group_info = kmalloc(metalen, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002412 if (meta_group_info == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002413 ext4_msg(sb, KERN_ERR, "can't allocate mem "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002414 "for a buddy group");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002415 goto exit_meta_group_info;
2416 }
2417 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2418 meta_group_info;
2419 }
2420
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002421 meta_group_info =
2422 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2423 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2424
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002425 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002426 if (meta_group_info[i] == NULL) {
Joe Perches7f6a11e2012-03-19 23:09:43 -04002427 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002428 goto exit_group_info;
2429 }
2430 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2431 &(meta_group_info[i]->bb_state));
2432
2433 /*
2434 * initialize bb_free to be able to skip
2435 * empty groups without initialization
2436 */
2437 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2438 meta_group_info[i]->bb_free =
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002439 ext4_free_clusters_after_init(sb, group, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002440 } else {
2441 meta_group_info[i]->bb_free =
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002442 ext4_free_group_clusters(sb, desc);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002443 }
2444
2445 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05002446 init_rwsem(&meta_group_info[i]->alloc_sem);
Venkatesh Pallipadi64e290e2010-03-04 22:25:21 -05002447 meta_group_info[i]->bb_free_root = RB_ROOT;
Curt Wohlgemuth8a57d9d2010-05-16 15:00:00 -04002448 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002449
2450#ifdef DOUBLE_CHECK
2451 {
2452 struct buffer_head *bh;
2453 meta_group_info[i]->bb_bitmap =
Dmitry Monakhov4fdb5542014-11-25 13:08:04 -05002454 kmalloc(sb->s_blocksize, GFP_NOFS);
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002455 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2456 bh = ext4_read_block_bitmap(sb, group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04002457 BUG_ON(IS_ERR_OR_NULL(bh));
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002458 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2459 sb->s_blocksize);
2460 put_bh(bh);
2461 }
2462#endif
2463
2464 return 0;
2465
2466exit_group_info:
2467 /* If a meta_group_info table has been allocated, release it now */
Tao Macaaf7a22011-07-11 18:42:42 -04002468 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002469 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
Tao Macaaf7a22011-07-11 18:42:42 -04002470 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2471 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002472exit_meta_group_info:
2473 return -ENOMEM;
2474} /* ext4_mb_add_groupinfo */
2475
Alex Tomasc9de5602008-01-29 00:19:52 -05002476static int ext4_mb_init_backend(struct super_block *sb)
2477{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002478 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002479 ext4_group_t i;
Alex Tomasc9de5602008-01-29 00:19:52 -05002480 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o28623c22012-09-05 01:31:50 -04002481 int err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002482 struct ext4_group_desc *desc;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002483 struct kmem_cache *cachep;
Alex Tomasc9de5602008-01-29 00:19:52 -05002484
Theodore Ts'o28623c22012-09-05 01:31:50 -04002485 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2486 if (err)
2487 return err;
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002488
Alex Tomasc9de5602008-01-29 00:19:52 -05002489 sbi->s_buddy_cache = new_inode(sb);
2490 if (sbi->s_buddy_cache == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002491 ext4_msg(sb, KERN_ERR, "can't get new inode");
Alex Tomasc9de5602008-01-29 00:19:52 -05002492 goto err_freesgi;
2493 }
Yu Jian48e60612011-08-01 17:41:39 -04002494 /* To avoid potentially colliding with an valid on-disk inode number,
2495 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2496 * not in the inode hash, so it should never be found by iget(), but
2497 * this will avoid confusion if it ever shows up during debugging. */
2498 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
Alex Tomasc9de5602008-01-29 00:19:52 -05002499 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
Theodore Ts'o8df96752009-05-01 08:50:38 -04002500 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002501 desc = ext4_get_group_desc(sb, i, NULL);
2502 if (desc == NULL) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002503 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002504 goto err_freebuddy;
2505 }
Frederic Bohe5f21b0e2008-07-11 19:27:31 -04002506 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2507 goto err_freebuddy;
Alex Tomasc9de5602008-01-29 00:19:52 -05002508 }
2509
2510 return 0;
2511
2512err_freebuddy:
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002513 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Roel Kluinf1fa3342008-04-29 22:01:15 -04002514 while (i-- > 0)
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002515 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
Theodore Ts'o28623c22012-09-05 01:31:50 -04002516 i = sbi->s_group_info_size;
Roel Kluinf1fa3342008-04-29 22:01:15 -04002517 while (i-- > 0)
Alex Tomasc9de5602008-01-29 00:19:52 -05002518 kfree(sbi->s_group_info[i]);
2519 iput(sbi->s_buddy_cache);
2520err_freesgi:
Al Virob93b41d2014-11-20 12:19:11 -05002521 kvfree(sbi->s_group_info);
Alex Tomasc9de5602008-01-29 00:19:52 -05002522 return -ENOMEM;
2523}
2524
Eric Sandeen2892c152011-02-12 08:12:18 -05002525static void ext4_groupinfo_destroy_slabs(void)
2526{
2527 int i;
2528
2529 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2530 if (ext4_groupinfo_caches[i])
2531 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2532 ext4_groupinfo_caches[i] = NULL;
2533 }
2534}
2535
2536static int ext4_groupinfo_create_slab(size_t size)
2537{
2538 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2539 int slab_size;
2540 int blocksize_bits = order_base_2(size);
2541 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2542 struct kmem_cache *cachep;
2543
2544 if (cache_index >= NR_GRPINFO_CACHES)
2545 return -EINVAL;
2546
2547 if (unlikely(cache_index < 0))
2548 cache_index = 0;
2549
2550 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2551 if (ext4_groupinfo_caches[cache_index]) {
2552 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2553 return 0; /* Already created */
2554 }
2555
2556 slab_size = offsetof(struct ext4_group_info,
2557 bb_counters[blocksize_bits + 2]);
2558
2559 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2560 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2561 NULL);
2562
Tao Ma823ba012011-07-11 18:26:01 -04002563 ext4_groupinfo_caches[cache_index] = cachep;
2564
Eric Sandeen2892c152011-02-12 08:12:18 -05002565 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2566 if (!cachep) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04002567 printk(KERN_EMERG
2568 "EXT4-fs: no memory for groupinfo slab cache\n");
Eric Sandeen2892c152011-02-12 08:12:18 -05002569 return -ENOMEM;
2570 }
2571
Eric Sandeen2892c152011-02-12 08:12:18 -05002572 return 0;
2573}
2574
Akira Fujita9d990122012-05-28 14:19:25 -04002575int ext4_mb_init(struct super_block *sb)
Alex Tomasc9de5602008-01-29 00:19:52 -05002576{
2577 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002578 unsigned i, j;
Alex Tomasc9de5602008-01-29 00:19:52 -05002579 unsigned offset;
2580 unsigned max;
Shen Feng74767c52008-07-11 19:27:31 -04002581 int ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002582
Eric Sandeen19278052009-08-25 22:36:25 -04002583 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
Alex Tomasc9de5602008-01-29 00:19:52 -05002584
2585 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2586 if (sbi->s_mb_offsets == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002587 ret = -ENOMEM;
2588 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002589 }
Yasunori Gotoff7ef322008-12-17 00:48:39 -05002590
Eric Sandeen19278052009-08-25 22:36:25 -04002591 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002592 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2593 if (sbi->s_mb_maxs == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002594 ret = -ENOMEM;
2595 goto out;
2596 }
2597
Eric Sandeen2892c152011-02-12 08:12:18 -05002598 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2599 if (ret < 0)
2600 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002601
2602 /* order 0 is regular bitmap */
2603 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2604 sbi->s_mb_offsets[0] = 0;
2605
2606 i = 1;
2607 offset = 0;
2608 max = sb->s_blocksize << 2;
2609 do {
2610 sbi->s_mb_offsets[i] = offset;
2611 sbi->s_mb_maxs[i] = max;
2612 offset += 1 << (sb->s_blocksize_bits - i);
2613 max = max >> 1;
2614 i++;
2615 } while (i <= sb->s_blocksize_bits + 1);
2616
Alex Tomasc9de5602008-01-29 00:19:52 -05002617 spin_lock_init(&sbi->s_md_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05002618 spin_lock_init(&sbi->s_bal_lock);
2619
2620 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2621 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2622 sbi->s_mb_stats = MB_DEFAULT_STATS;
2623 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2624 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
Theodore Ts'o27baebb2011-09-09 19:02:51 -04002625 /*
2626 * The default group preallocation is 512, which for 4k block
2627 * sizes translates to 2 megabytes. However for bigalloc file
2628 * systems, this is probably too big (i.e, if the cluster size
2629 * is 1 megabyte, then group preallocation size becomes half a
2630 * gigabyte!). As a default, we will keep a two megabyte
2631 * group pralloc size for cluster sizes up to 64k, and after
2632 * that, we will force a minimum group preallocation size of
2633 * 32 clusters. This translates to 8 megs when the cluster
2634 * size is 256k, and 32 megs when the cluster size is 1 meg,
2635 * which seems reasonable as a default.
2636 */
2637 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2638 sbi->s_cluster_bits, 32);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002639 /*
2640 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2641 * to the lowest multiple of s_stripe which is bigger than
2642 * the s_mb_group_prealloc as determined above. We want
2643 * the preallocation size to be an exact multiple of the
2644 * RAID stripe size so that preallocations don't fragment
2645 * the stripes.
2646 */
2647 if (sbi->s_stripe > 1) {
2648 sbi->s_mb_group_prealloc = roundup(
2649 sbi->s_mb_group_prealloc, sbi->s_stripe);
2650 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002651
Eric Sandeen730c2132008-09-13 15:23:29 -04002652 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002653 if (sbi->s_locality_groups == NULL) {
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002654 ret = -ENOMEM;
Andrey Tsyvarev029b10c2014-05-12 12:34:21 -04002655 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05002656 }
Eric Sandeen730c2132008-09-13 15:23:29 -04002657 for_each_possible_cpu(i) {
Alex Tomasc9de5602008-01-29 00:19:52 -05002658 struct ext4_locality_group *lg;
Eric Sandeen730c2132008-09-13 15:23:29 -04002659 lg = per_cpu_ptr(sbi->s_locality_groups, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05002660 mutex_init(&lg->lg_mutex);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04002661 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2662 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
Alex Tomasc9de5602008-01-29 00:19:52 -05002663 spin_lock_init(&lg->lg_prealloc_lock);
2664 }
2665
Yu Jian79a77c52011-08-01 17:41:46 -04002666 /* init file for buddy data */
2667 ret = ext4_mb_init_backend(sb);
Tao Ma7aa0bae2011-10-06 10:22:28 -04002668 if (ret != 0)
2669 goto out_free_locality_groups;
Yu Jian79a77c52011-08-01 17:41:46 -04002670
Tao Ma7aa0bae2011-10-06 10:22:28 -04002671 return 0;
2672
2673out_free_locality_groups:
2674 free_percpu(sbi->s_locality_groups);
2675 sbi->s_locality_groups = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002676out:
Tao Ma7aa0bae2011-10-06 10:22:28 -04002677 kfree(sbi->s_mb_offsets);
2678 sbi->s_mb_offsets = NULL;
2679 kfree(sbi->s_mb_maxs);
2680 sbi->s_mb_maxs = NULL;
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002681 return ret;
Alex Tomasc9de5602008-01-29 00:19:52 -05002682}
2683
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002684/* need to called with the ext4 group lock held */
Alex Tomasc9de5602008-01-29 00:19:52 -05002685static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2686{
2687 struct ext4_prealloc_space *pa;
2688 struct list_head *cur, *tmp;
2689 int count = 0;
2690
2691 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2692 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2693 list_del(&pa->pa_group_list);
2694 count++;
Aneesh Kumar K.V688f05a2008-10-13 12:14:14 -04002695 kmem_cache_free(ext4_pspace_cachep, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05002696 }
2697 if (count)
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04002698 mb_debug(1, "mballoc: %u PAs left\n", count);
Alex Tomasc9de5602008-01-29 00:19:52 -05002699
2700}
2701
2702int ext4_mb_release(struct super_block *sb)
2703{
Theodore Ts'o8df96752009-05-01 08:50:38 -04002704 ext4_group_t ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05002705 ext4_group_t i;
2706 int num_meta_group_infos;
2707 struct ext4_group_info *grinfo;
2708 struct ext4_sb_info *sbi = EXT4_SB(sb);
Curt Wohlgemuthfb1813f2010-10-27 21:29:12 -04002709 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
Alex Tomasc9de5602008-01-29 00:19:52 -05002710
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.V519deca2008-05-15 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
Theodore Ts'o574ca172008-07-11 19:27:31 -04002899 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04002900 if (IS_ERR(bitmap_bh)) {
2901 err = PTR_ERR(bitmap_bh);
2902 bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05002903 goto out_err;
Darrick J. Wong9008a582015-10-17 21:33:24 -04002904 }
Alex Tomasc9de5602008-01-29 00:19:52 -05002905
liang xie5d601252014-05-12 22:06:43 -04002906 BUFFER_TRACE(bitmap_bh, "getting write access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002907 err = ext4_journal_get_write_access(handle, bitmap_bh);
2908 if (err)
2909 goto out_err;
2910
2911 err = -EIO;
2912 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2913 if (!gdp)
2914 goto out_err;
2915
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05002916 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002917 ext4_free_group_clusters(sb, gdp));
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04002918
liang xie5d601252014-05-12 22:06:43 -04002919 BUFFER_TRACE(gdp_bh, "get_write_access");
Alex Tomasc9de5602008-01-29 00:19:52 -05002920 err = ext4_journal_get_write_access(handle, gdp_bh);
2921 if (err)
2922 goto out_err;
2923
Akinobu Mitabda00de2010-03-03 23:53:25 -05002924 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
Alex Tomasc9de5602008-01-29 00:19:52 -05002925
Theodore Ts'o53accfa2011-09-09 18:48:51 -04002926 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Theodore Ts'o6fd058f2009-05-17 15:38:01 -04002927 if (!ext4_data_block_valid(sbi, block, len)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05002928 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
Theodore Ts'o1084f252012-03-19 23:13:43 -04002929 "fs metadata", block, block+len);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002930 /* File system mounted not to panic on error
2931 * Fix the bitmap and repeat the block allocation
2932 * We leak some of the blocks here.
2933 */
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002934 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002935 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2936 ac->ac_b_ex.fe_len);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002937 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Frank Mayhar03901312009-01-07 00:06:22 -05002938 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04002939 if (!err)
2940 err = -EAGAIN;
2941 goto out_err;
Alex Tomasc9de5602008-01-29 00:19:52 -05002942 }
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002943
2944 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05002945#ifdef AGGRESSIVE_CHECK
2946 {
2947 int i;
2948 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2949 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2950 bitmap_bh->b_data));
2951 }
2952 }
2953#endif
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04002954 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2955 ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05002956 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2957 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002958 ext4_free_group_clusters_set(sb, gdp,
Theodore Ts'ocff1dfd72011-09-09 19:12:51 -04002959 ext4_free_clusters_after_init(sb,
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002960 ac->ac_b_ex.fe_group, gdp));
Alex Tomasc9de5602008-01-29 00:19:52 -05002961 }
Theodore Ts'o021b65b2011-09-09 19:08:51 -04002962 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2963 ext4_free_group_clusters_set(sb, gdp, len);
Tao Ma79f1ba42012-10-22 00:34:32 -04002964 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04002965 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04002966
2967 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04002968 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
Mingming Caod2a17632008-07-14 17:52:37 -04002969 /*
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002970 * Now reduce the dirty block count also. Should not go negative
Mingming Caod2a17632008-07-14 17:52:37 -04002971 */
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04002972 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2973 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04002974 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2975 reserv_clstrs);
Alex Tomasc9de5602008-01-29 00:19:52 -05002976
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002977 if (sbi->s_log_groups_per_flex) {
2978 ext4_group_t flex_group = ext4_flex_group(sbi,
2979 ac->ac_b_ex.fe_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04002980 atomic64_sub(ac->ac_b_ex.fe_len,
2981 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04002982 }
2983
Frank Mayhar03901312009-01-07 00:06:22 -05002984 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002985 if (err)
2986 goto out_err;
Frank Mayhar03901312009-01-07 00:06:22 -05002987 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002988
2989out_err:
Aneesh Kumar K.V42a10ad2008-02-10 01:07:28 -05002990 brelse(bitmap_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05002991 return err;
2992}
2993
2994/*
2995 * here we normalize request for locality group
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04002996 * Group request are normalized to s_mb_group_prealloc, which goes to
2997 * s_strip if we set the same via mount option.
2998 * s_mb_group_prealloc can be configured via
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04002999 * /sys/fs/ext4/<partition>/mb_group_prealloc
Alex Tomasc9de5602008-01-29 00:19:52 -05003000 *
3001 * XXX: should we try to preallocate more than the group has now?
3002 */
3003static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3004{
3005 struct super_block *sb = ac->ac_sb;
3006 struct ext4_locality_group *lg = ac->ac_lg;
3007
3008 BUG_ON(lg == NULL);
Dan Ehrenbergd7a1fee2011-07-17 21:11:30 -04003009 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003010 mb_debug(1, "#%u: goal %u blocks for locality group\n",
Alex Tomasc9de5602008-01-29 00:19:52 -05003011 current->pid, ac->ac_g_ex.fe_len);
3012}
3013
3014/*
3015 * Normalization means making request better in terms of
3016 * size and alignment
3017 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003018static noinline_for_stack void
3019ext4_mb_normalize_request(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05003020 struct ext4_allocation_request *ar)
3021{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003022 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003023 int bsbits, max;
3024 ext4_lblk_t end;
Curt Wohlgemuth1592d2c2012-02-20 17:53:03 -05003025 loff_t size, start_off;
3026 loff_t orig_size __maybe_unused;
Andi Kleen5a0790c2010-06-14 13:28:03 -04003027 ext4_lblk_t start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003028 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003029 struct ext4_prealloc_space *pa;
Alex Tomasc9de5602008-01-29 00:19:52 -05003030
3031 /* do normalize only data requests, metadata requests
3032 do not need preallocation */
3033 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3034 return;
3035
3036 /* sometime caller may want exact blocks */
3037 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3038 return;
3039
3040 /* caller may indicate that preallocation isn't
3041 * required (it's a tail, for example) */
3042 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3043 return;
3044
3045 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3046 ext4_mb_normalize_group_request(ac);
3047 return ;
3048 }
3049
3050 bsbits = ac->ac_sb->s_blocksize_bits;
3051
3052 /* first, let's learn actual file size
3053 * given current request is allocated */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003054 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003055 size = size << bsbits;
3056 if (size < i_size_read(ac->ac_inode))
3057 size = i_size_read(ac->ac_inode);
Andi Kleen5a0790c2010-06-14 13:28:03 -04003058 orig_size = size;
Alex Tomasc9de5602008-01-29 00:19:52 -05003059
Valerie Clement19304792008-05-13 19:31:14 -04003060 /* max size of free chunks */
3061 max = 2 << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003062
Valerie Clement19304792008-05-13 19:31:14 -04003063#define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3064 (req <= (size) || max <= (chunk_size))
Alex Tomasc9de5602008-01-29 00:19:52 -05003065
3066 /* first, try to predict filesize */
3067 /* XXX: should this table be tunable? */
3068 start_off = 0;
3069 if (size <= 16 * 1024) {
3070 size = 16 * 1024;
3071 } else if (size <= 32 * 1024) {
3072 size = 32 * 1024;
3073 } else if (size <= 64 * 1024) {
3074 size = 64 * 1024;
3075 } else if (size <= 128 * 1024) {
3076 size = 128 * 1024;
3077 } else if (size <= 256 * 1024) {
3078 size = 256 * 1024;
3079 } else if (size <= 512 * 1024) {
3080 size = 512 * 1024;
3081 } else if (size <= 1024 * 1024) {
3082 size = 1024 * 1024;
Valerie Clement19304792008-05-13 19:31:14 -04003083 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003084 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
Valerie Clement19304792008-05-13 19:31:14 -04003085 (21 - bsbits)) << 21;
3086 size = 2 * 1024 * 1024;
3087 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003088 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3089 (22 - bsbits)) << 22;
3090 size = 4 * 1024 * 1024;
3091 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
Valerie Clement19304792008-05-13 19:31:14 -04003092 (8<<20)>>bsbits, max, 8 * 1024)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003093 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3094 (23 - bsbits)) << 23;
3095 size = 8 * 1024 * 1024;
3096 } else {
Xiaoguang Wangb27b1532014-07-27 22:26:36 -04003097 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3098 size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3099 ac->ac_o_ex.fe_len) << bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003100 }
Andi Kleen5a0790c2010-06-14 13:28:03 -04003101 size = size >> bsbits;
3102 start = start_off >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05003103
3104 /* don't cover already allocated blocks in selected range */
3105 if (ar->pleft && start <= ar->lleft) {
3106 size -= ar->lleft + 1 - start;
3107 start = ar->lleft + 1;
3108 }
3109 if (ar->pright && start + size - 1 >= ar->lright)
3110 size -= start + size - ar->lright;
3111
3112 end = start + size;
3113
3114 /* check we don't cross already preallocated blocks */
3115 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003116 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003117 ext4_lblk_t pa_end;
Alex Tomasc9de5602008-01-29 00:19:52 -05003118
Alex Tomasc9de5602008-01-29 00:19:52 -05003119 if (pa->pa_deleted)
3120 continue;
3121 spin_lock(&pa->pa_lock);
3122 if (pa->pa_deleted) {
3123 spin_unlock(&pa->pa_lock);
3124 continue;
3125 }
3126
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003127 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3128 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003129
3130 /* PA must not overlap original request */
3131 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3132 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3133
Eric Sandeen38877f42009-08-17 23:55:24 -04003134 /* skip PAs this normalized request doesn't overlap with */
3135 if (pa->pa_lstart >= end || pa_end <= start) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003136 spin_unlock(&pa->pa_lock);
3137 continue;
3138 }
3139 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3140
Eric Sandeen38877f42009-08-17 23:55:24 -04003141 /* adjust start or end to be adjacent to this pa */
Alex Tomasc9de5602008-01-29 00:19:52 -05003142 if (pa_end <= ac->ac_o_ex.fe_logical) {
3143 BUG_ON(pa_end < start);
3144 start = pa_end;
Eric Sandeen38877f42009-08-17 23:55:24 -04003145 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003146 BUG_ON(pa->pa_lstart > end);
3147 end = pa->pa_lstart;
3148 }
3149 spin_unlock(&pa->pa_lock);
3150 }
3151 rcu_read_unlock();
3152 size = end - start;
3153
3154 /* XXX: extra loop to check we really don't overlap preallocations */
3155 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003156 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003157 ext4_lblk_t pa_end;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003158
Alex Tomasc9de5602008-01-29 00:19:52 -05003159 spin_lock(&pa->pa_lock);
3160 if (pa->pa_deleted == 0) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003161 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3162 pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003163 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3164 }
3165 spin_unlock(&pa->pa_lock);
3166 }
3167 rcu_read_unlock();
3168
3169 if (start + size <= ac->ac_o_ex.fe_logical &&
3170 start > ac->ac_o_ex.fe_logical) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003171 ext4_msg(ac->ac_sb, KERN_ERR,
3172 "start %lu, size %lu, fe_logical %lu",
3173 (unsigned long) start, (unsigned long) size,
3174 (unsigned long) ac->ac_o_ex.fe_logical);
Dmitry Monakhovdfe076c2014-10-01 22:26:17 -04003175 BUG();
Alex Tomasc9de5602008-01-29 00:19:52 -05003176 }
Maurizio Lombardib5b60772014-05-27 12:48:56 -04003177 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
Alex Tomasc9de5602008-01-29 00:19:52 -05003178
3179 /* now prepare goal request */
3180
3181 /* XXX: is it better to align blocks WRT to logical
3182 * placement or satisfy big request as is */
3183 ac->ac_g_ex.fe_logical = start;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003184 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
Alex Tomasc9de5602008-01-29 00:19:52 -05003185
3186 /* define goal start in order to merge */
3187 if (ar->pright && (ar->lright == (start + size))) {
3188 /* merge to the right */
3189 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3190 &ac->ac_f_ex.fe_group,
3191 &ac->ac_f_ex.fe_start);
3192 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3193 }
3194 if (ar->pleft && (ar->lleft + 1 == start)) {
3195 /* merge to the left */
3196 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3197 &ac->ac_f_ex.fe_group,
3198 &ac->ac_f_ex.fe_start);
3199 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3200 }
3201
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003202 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
Alex Tomasc9de5602008-01-29 00:19:52 -05003203 (unsigned) orig_size, (unsigned) start);
3204}
3205
3206static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3207{
3208 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3209
3210 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3211 atomic_inc(&sbi->s_bal_reqs);
3212 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
Curt Wohlgemuth291dae42010-05-16 16:00:00 -04003213 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
Alex Tomasc9de5602008-01-29 00:19:52 -05003214 atomic_inc(&sbi->s_bal_success);
3215 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3216 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3217 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3218 atomic_inc(&sbi->s_bal_goals);
3219 if (ac->ac_found > sbi->s_mb_max_to_scan)
3220 atomic_inc(&sbi->s_bal_breaks);
3221 }
3222
Theodore Ts'o296c3552009-09-30 00:32:42 -04003223 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3224 trace_ext4_mballoc_alloc(ac);
3225 else
3226 trace_ext4_mballoc_prealloc(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05003227}
3228
3229/*
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003230 * Called on failure; free up any blocks from the inode PA for this
3231 * context. We don't need this for MB_GROUP_PA because we only change
3232 * pa_free in ext4_mb_release_context(), but on failure, we've already
3233 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3234 */
3235static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3236{
3237 struct ext4_prealloc_space *pa = ac->ac_pa;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003238 struct ext4_buddy e4b;
3239 int err;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003240
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003241 if (pa == NULL) {
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003242 if (ac->ac_f_ex.fe_len == 0)
3243 return;
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003244 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3245 if (err) {
3246 /*
3247 * This should never happen since we pin the
3248 * pages in the ext4_allocation_context so
3249 * ext4_mb_load_buddy() should never fail.
3250 */
3251 WARN(1, "mb_load_buddy failed (%d)", err);
3252 return;
3253 }
3254 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3255 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3256 ac->ac_f_ex.fe_len);
3257 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
Theodore Ts'oc99d1e62014-08-23 17:47:28 -04003258 ext4_mb_unload_buddy(&e4b);
Theodore Ts'o86f0afd2014-07-30 22:17:17 -04003259 return;
3260 }
3261 if (pa->pa_type == MB_INODE_PA)
Zheng Liu400db9d2012-05-28 17:53:53 -04003262 pa->pa_free += ac->ac_b_ex.fe_len;
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05003263}
3264
3265/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003266 * use blocks preallocated to inode
3267 */
3268static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3269 struct ext4_prealloc_space *pa)
3270{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003271 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003272 ext4_fsblk_t start;
3273 ext4_fsblk_t end;
3274 int len;
3275
3276 /* found preallocated blocks, use them */
3277 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003278 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3279 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3280 len = EXT4_NUM_B2C(sbi, end - start);
Alex Tomasc9de5602008-01-29 00:19:52 -05003281 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3282 &ac->ac_b_ex.fe_start);
3283 ac->ac_b_ex.fe_len = len;
3284 ac->ac_status = AC_STATUS_FOUND;
3285 ac->ac_pa = pa;
3286
3287 BUG_ON(start < pa->pa_pstart);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003288 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
Alex Tomasc9de5602008-01-29 00:19:52 -05003289 BUG_ON(pa->pa_free < len);
3290 pa->pa_free -= len;
3291
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003292 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003293}
3294
3295/*
3296 * use blocks preallocated to locality group
3297 */
3298static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3299 struct ext4_prealloc_space *pa)
3300{
Aneesh Kumar K.V03cddb82008-06-05 20:59:29 -04003301 unsigned int len = ac->ac_o_ex.fe_len;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003302
Alex Tomasc9de5602008-01-29 00:19:52 -05003303 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3304 &ac->ac_b_ex.fe_group,
3305 &ac->ac_b_ex.fe_start);
3306 ac->ac_b_ex.fe_len = len;
3307 ac->ac_status = AC_STATUS_FOUND;
3308 ac->ac_pa = pa;
3309
3310 /* we don't correct pa_pstart or pa_plen here to avoid
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003311 * possible race when the group is being loaded concurrently
Alex Tomasc9de5602008-01-29 00:19:52 -05003312 * instead we correct pa later, after blocks are marked
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05003313 * in on-disk bitmap -- see ext4_mb_release_context()
3314 * Other CPUs are prevented from allocating from this pa by lg_mutex
Alex Tomasc9de5602008-01-29 00:19:52 -05003315 */
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003316 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003317}
3318
3319/*
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003320 * Return the prealloc space that have minimal distance
3321 * from the goal block. @cpa is the prealloc
3322 * space that is having currently known minimal distance
3323 * from the goal block.
3324 */
3325static struct ext4_prealloc_space *
3326ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3327 struct ext4_prealloc_space *pa,
3328 struct ext4_prealloc_space *cpa)
3329{
3330 ext4_fsblk_t cur_distance, new_distance;
3331
3332 if (cpa == NULL) {
3333 atomic_inc(&pa->pa_count);
3334 return pa;
3335 }
Andrew Morton79211c82015-11-09 14:58:13 -08003336 cur_distance = abs(goal_block - cpa->pa_pstart);
3337 new_distance = abs(goal_block - pa->pa_pstart);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003338
Coly Li5a54b2f2011-02-24 14:10:05 -05003339 if (cur_distance <= new_distance)
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003340 return cpa;
3341
3342 /* drop the previous reference */
3343 atomic_dec(&cpa->pa_count);
3344 atomic_inc(&pa->pa_count);
3345 return pa;
3346}
3347
3348/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003349 * search goal blocks in preallocated space
3350 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003351static noinline_for_stack int
3352ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003353{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003354 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003355 int order, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05003356 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3357 struct ext4_locality_group *lg;
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003358 struct ext4_prealloc_space *pa, *cpa = NULL;
3359 ext4_fsblk_t goal_block;
Alex Tomasc9de5602008-01-29 00:19:52 -05003360
3361 /* only data can be preallocated */
3362 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3363 return 0;
3364
3365 /* first, try per-file preallocation */
3366 rcu_read_lock();
Aneesh Kumar K.V9a0762c2008-04-17 10:38:59 -04003367 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003368
3369 /* all fields in this condition don't change,
3370 * so we can skip locking for them */
3371 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003372 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3373 EXT4_C2B(sbi, pa->pa_len)))
Alex Tomasc9de5602008-01-29 00:19:52 -05003374 continue;
3375
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003376 /* non-extent files can't have physical blocks past 2^32 */
Dmitry Monakhov12e9b892010-05-16 22:00:00 -04003377 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003378 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3379 EXT4_MAX_BLOCK_FILE_PHYS))
Eric Sandeenfb0a3872009-09-16 14:45:10 -04003380 continue;
3381
Alex Tomasc9de5602008-01-29 00:19:52 -05003382 /* found preallocated blocks, use them */
3383 spin_lock(&pa->pa_lock);
3384 if (pa->pa_deleted == 0 && pa->pa_free) {
3385 atomic_inc(&pa->pa_count);
3386 ext4_mb_use_inode_pa(ac, pa);
3387 spin_unlock(&pa->pa_lock);
3388 ac->ac_criteria = 10;
3389 rcu_read_unlock();
3390 return 1;
3391 }
3392 spin_unlock(&pa->pa_lock);
3393 }
3394 rcu_read_unlock();
3395
3396 /* can we use group allocation? */
3397 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3398 return 0;
3399
3400 /* inode may have no locality group for some reason */
3401 lg = ac->ac_lg;
3402 if (lg == NULL)
3403 return 0;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003404 order = fls(ac->ac_o_ex.fe_len) - 1;
3405 if (order > PREALLOC_TB_SIZE - 1)
3406 /* The max size of hash table is PREALLOC_TB_SIZE */
3407 order = PREALLOC_TB_SIZE - 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003408
Akinobu Mitabda00de2010-03-03 23:53:25 -05003409 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003410 /*
3411 * search for the prealloc space that is having
3412 * minimal distance from the goal block.
3413 */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003414 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3415 rcu_read_lock();
3416 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3417 pa_inode_list) {
3418 spin_lock(&pa->pa_lock);
3419 if (pa->pa_deleted == 0 &&
3420 pa->pa_free >= ac->ac_o_ex.fe_len) {
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003421
3422 cpa = ext4_mb_check_group_pa(goal_block,
3423 pa, cpa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003424 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003425 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05003426 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003427 rcu_read_unlock();
Alex Tomasc9de5602008-01-29 00:19:52 -05003428 }
Aneesh Kumar K.V5e745b02008-08-18 18:00:57 -04003429 if (cpa) {
3430 ext4_mb_use_group_pa(ac, cpa);
3431 ac->ac_criteria = 20;
3432 return 1;
3433 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003434 return 0;
3435}
3436
3437/*
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003438 * the function goes through all block freed in the group
3439 * but not yet committed and marks them used in in-core bitmap.
3440 * buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003441 * Need to be called with the ext4 group lock held
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003442 */
3443static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3444 ext4_group_t group)
3445{
3446 struct rb_node *n;
3447 struct ext4_group_info *grp;
3448 struct ext4_free_data *entry;
3449
3450 grp = ext4_get_group_info(sb, group);
3451 n = rb_first(&(grp->bb_free_root));
3452
3453 while (n) {
Bobi Jam18aadd42012-02-20 17:53:02 -05003454 entry = rb_entry(n, struct ext4_free_data, efd_node);
3455 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05003456 n = rb_next(n);
3457 }
3458 return;
3459}
3460
3461/*
Alex Tomasc9de5602008-01-29 00:19:52 -05003462 * the function goes through all preallocation in this group and marks them
3463 * used in in-core bitmap. buddy must be generated from this bitmap
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04003464 * Need to be called with ext4 group lock held
Alex Tomasc9de5602008-01-29 00:19:52 -05003465 */
Eric Sandeen089ceec2009-07-05 22:17:31 -04003466static noinline_for_stack
3467void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
Alex Tomasc9de5602008-01-29 00:19:52 -05003468 ext4_group_t group)
3469{
3470 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3471 struct ext4_prealloc_space *pa;
3472 struct list_head *cur;
3473 ext4_group_t groupnr;
3474 ext4_grpblk_t start;
3475 int preallocated = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003476 int len;
3477
3478 /* all form of preallocation discards first load group,
3479 * so the only competing code is preallocation use.
3480 * we don't need any locking here
3481 * notice we do NOT ignore preallocations with pa_deleted
3482 * otherwise we could leave used blocks available for
3483 * allocation in buddy when concurrent ext4_mb_put_pa()
3484 * is dropping preallocation
3485 */
3486 list_for_each(cur, &grp->bb_prealloc_list) {
3487 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3488 spin_lock(&pa->pa_lock);
3489 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3490 &groupnr, &start);
3491 len = pa->pa_len;
3492 spin_unlock(&pa->pa_lock);
3493 if (unlikely(len == 0))
3494 continue;
3495 BUG_ON(groupnr != group);
Yongqiang Yangc3e94d12011-07-26 22:05:53 -04003496 ext4_set_bits(bitmap, start, len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003497 preallocated += len;
Alex Tomasc9de5602008-01-29 00:19:52 -05003498 }
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003499 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003500}
3501
3502static void ext4_mb_pa_callback(struct rcu_head *head)
3503{
3504 struct ext4_prealloc_space *pa;
3505 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003506
3507 BUG_ON(atomic_read(&pa->pa_count));
3508 BUG_ON(pa->pa_deleted == 0);
Alex Tomasc9de5602008-01-29 00:19:52 -05003509 kmem_cache_free(ext4_pspace_cachep, pa);
3510}
3511
3512/*
3513 * drops a reference to preallocated space descriptor
3514 * if this was the last reference and the space is consumed
3515 */
3516static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3517 struct super_block *sb, struct ext4_prealloc_space *pa)
3518{
Theodore Ts'oa9df9a42009-01-05 22:18:16 -05003519 ext4_group_t grp;
Eric Sandeend33a1972009-03-16 23:25:40 -04003520 ext4_fsblk_t grp_blk;
Alex Tomasc9de5602008-01-29 00:19:52 -05003521
Alex Tomasc9de5602008-01-29 00:19:52 -05003522 /* in this short window concurrent discard can set pa_deleted */
3523 spin_lock(&pa->pa_lock);
Junho Ryu4e8d2132013-12-03 18:10:28 -05003524 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3525 spin_unlock(&pa->pa_lock);
3526 return;
3527 }
3528
Alex Tomasc9de5602008-01-29 00:19:52 -05003529 if (pa->pa_deleted == 1) {
3530 spin_unlock(&pa->pa_lock);
3531 return;
3532 }
3533
3534 pa->pa_deleted = 1;
3535 spin_unlock(&pa->pa_lock);
3536
Eric Sandeend33a1972009-03-16 23:25:40 -04003537 grp_blk = pa->pa_pstart;
Theodore Ts'o60e66792010-05-17 07:00:00 -04003538 /*
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003539 * If doing group-based preallocation, pa_pstart may be in the
3540 * next group when pa is used up
3541 */
3542 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeend33a1972009-03-16 23:25:40 -04003543 grp_blk--;
3544
Lukas Czernerbd862982013-04-03 23:32:34 -04003545 grp = ext4_get_group_number(sb, grp_blk);
Alex Tomasc9de5602008-01-29 00:19:52 -05003546
3547 /*
3548 * possible race:
3549 *
3550 * P1 (buddy init) P2 (regular allocation)
3551 * find block B in PA
3552 * copy on-disk bitmap to buddy
3553 * mark B in on-disk bitmap
3554 * drop PA from group
3555 * mark all PAs in buddy
3556 *
3557 * thus, P1 initializes buddy with B available. to prevent this
3558 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3559 * against that pair
3560 */
3561 ext4_lock_group(sb, grp);
3562 list_del(&pa->pa_group_list);
3563 ext4_unlock_group(sb, grp);
3564
3565 spin_lock(pa->pa_obj_lock);
3566 list_del_rcu(&pa->pa_inode_list);
3567 spin_unlock(pa->pa_obj_lock);
3568
3569 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3570}
3571
3572/*
3573 * creates new preallocated space for given inode
3574 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003575static noinline_for_stack int
3576ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003577{
3578 struct super_block *sb = ac->ac_sb;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003579 struct ext4_sb_info *sbi = EXT4_SB(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05003580 struct ext4_prealloc_space *pa;
3581 struct ext4_group_info *grp;
3582 struct ext4_inode_info *ei;
3583
3584 /* preallocate only when found space is larger then requested */
3585 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3586 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3587 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3588
3589 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3590 if (pa == NULL)
3591 return -ENOMEM;
3592
3593 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3594 int winl;
3595 int wins;
3596 int win;
3597 int offs;
3598
3599 /* we can't allocate as much as normalizer wants.
3600 * so, found space must get proper lstart
3601 * to cover original request */
3602 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3603 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3604
3605 /* we're limited by original request in that
3606 * logical block must be covered any way
3607 * winl is window we can move our chunk within */
3608 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3609
3610 /* also, we should cover whole original request */
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003611 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003612
3613 /* the smallest one defines real window */
3614 win = min(winl, wins);
3615
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003616 offs = ac->ac_o_ex.fe_logical %
3617 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003618 if (offs && offs < win)
3619 win = offs;
3620
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003621 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
Lukas Czerner810da242013-03-02 17:18:58 -05003622 EXT4_NUM_B2C(sbi, win);
Alex Tomasc9de5602008-01-29 00:19:52 -05003623 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3624 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3625 }
3626
3627 /* preallocation can change ac_b_ex, thus we store actually
3628 * allocated blocks for history */
3629 ac->ac_f_ex = ac->ac_b_ex;
3630
3631 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3632 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3633 pa->pa_len = ac->ac_b_ex.fe_len;
3634 pa->pa_free = pa->pa_len;
3635 atomic_set(&pa->pa_count, 1);
3636 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003637 INIT_LIST_HEAD(&pa->pa_inode_list);
3638 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003639 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003640 pa->pa_type = MB_INODE_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003641
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003642 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
Alex Tomasc9de5602008-01-29 00:19:52 -05003643 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003644 trace_ext4_mb_new_inode_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003645
3646 ext4_mb_use_inode_pa(ac, pa);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003647 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
Alex Tomasc9de5602008-01-29 00:19:52 -05003648
3649 ei = EXT4_I(ac->ac_inode);
3650 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3651
3652 pa->pa_obj_lock = &ei->i_prealloc_lock;
3653 pa->pa_inode = ac->ac_inode;
3654
3655 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3656 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3657 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3658
3659 spin_lock(pa->pa_obj_lock);
3660 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3661 spin_unlock(pa->pa_obj_lock);
3662
3663 return 0;
3664}
3665
3666/*
3667 * creates new preallocated space for locality group inodes belongs to
3668 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003669static noinline_for_stack int
3670ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
Alex Tomasc9de5602008-01-29 00:19:52 -05003671{
3672 struct super_block *sb = ac->ac_sb;
3673 struct ext4_locality_group *lg;
3674 struct ext4_prealloc_space *pa;
3675 struct ext4_group_info *grp;
3676
3677 /* preallocate only when found space is larger then requested */
3678 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3679 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3680 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3681
3682 BUG_ON(ext4_pspace_cachep == NULL);
3683 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3684 if (pa == NULL)
3685 return -ENOMEM;
3686
3687 /* preallocation can change ac_b_ex, thus we store actually
3688 * allocated blocks for history */
3689 ac->ac_f_ex = ac->ac_b_ex;
3690
3691 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3692 pa->pa_lstart = pa->pa_pstart;
3693 pa->pa_len = ac->ac_b_ex.fe_len;
3694 pa->pa_free = pa->pa_len;
3695 atomic_set(&pa->pa_count, 1);
3696 spin_lock_init(&pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003697 INIT_LIST_HEAD(&pa->pa_inode_list);
Aneesh Kumar K.Vd794bf82009-02-14 10:31:16 -05003698 INIT_LIST_HEAD(&pa->pa_group_list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003699 pa->pa_deleted = 0;
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003700 pa->pa_type = MB_GROUP_PA;
Alex Tomasc9de5602008-01-29 00:19:52 -05003701
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003702 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003703 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3704 trace_ext4_mb_new_group_pa(ac, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003705
3706 ext4_mb_use_group_pa(ac, pa);
3707 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3708
3709 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3710 lg = ac->ac_lg;
3711 BUG_ON(lg == NULL);
3712
3713 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3714 pa->pa_inode = NULL;
3715
3716 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3717 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3718 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3719
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04003720 /*
3721 * We will later add the new pa to the right bucket
3722 * after updating the pa_free in ext4_mb_release_context
3723 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003724 return 0;
3725}
3726
3727static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3728{
3729 int err;
3730
3731 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3732 err = ext4_mb_new_group_pa(ac);
3733 else
3734 err = ext4_mb_new_inode_pa(ac);
3735 return err;
3736}
3737
3738/*
3739 * finds all unused blocks in on-disk bitmap, frees them in
3740 * in-core bitmap and buddy.
3741 * @pa must be unlinked from inode and group lists, so that
3742 * nobody else can find/use it.
3743 * the caller MUST hold group/inode locks.
3744 * TODO: optimize the case when there are no in-core structures yet
3745 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003746static noinline_for_stack int
3747ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003748 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003749{
Alex Tomasc9de5602008-01-29 00:19:52 -05003750 struct super_block *sb = e4b->bd_sb;
3751 struct ext4_sb_info *sbi = EXT4_SB(sb);
Theodore Ts'o498e5f22008-11-05 00:14:04 -05003752 unsigned int end;
3753 unsigned int next;
Alex Tomasc9de5602008-01-29 00:19:52 -05003754 ext4_group_t group;
3755 ext4_grpblk_t bit;
Theodore Ts'oba80b102009-01-03 20:03:21 -05003756 unsigned long long grp_blk_start;
Alex Tomasc9de5602008-01-29 00:19:52 -05003757 int err = 0;
3758 int free = 0;
3759
3760 BUG_ON(pa->pa_deleted == 0);
3761 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003762 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003763 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3764 end = bit + pa->pa_len;
3765
Alex Tomasc9de5602008-01-29 00:19:52 -05003766 while (bit < end) {
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003767 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003768 if (bit >= end)
3769 break;
Aneesh Kumar K.Vffad0a42008-02-23 01:38:34 -05003770 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003771 mb_debug(1, " free preallocated %u/%u in group %u\n",
Andi Kleen5a0790c2010-06-14 13:28:03 -04003772 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3773 (unsigned) next - bit, (unsigned) group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003774 free += next - bit;
3775
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003776 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04003777 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3778 EXT4_C2B(sbi, bit)),
Lukas Czernera9c667f2011-06-06 09:51:52 -04003779 next - bit);
Alex Tomasc9de5602008-01-29 00:19:52 -05003780 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3781 bit = next + 1;
3782 }
3783 if (free != pa->pa_free) {
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003784 ext4_msg(e4b->bd_sb, KERN_CRIT,
3785 "pa %p: logic %lu, phys. %lu, len %lu",
3786 pa, (unsigned long) pa->pa_lstart,
3787 (unsigned long) pa->pa_pstart,
3788 (unsigned long) pa->pa_len);
Theodore Ts'oe29136f2010-06-29 12:54:28 -04003789 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
Aneesh Kumar K.V5d1b1b32009-01-05 22:19:52 -05003790 free, pa->pa_free);
Aneesh Kumar K.Ve56eb652008-02-15 13:48:21 -05003791 /*
3792 * pa is already deleted so we use the value obtained
3793 * from the bitmap and continue.
3794 */
Alex Tomasc9de5602008-01-29 00:19:52 -05003795 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003796 atomic_add(free, &sbi->s_mb_discarded);
3797
3798 return err;
3799}
3800
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003801static noinline_for_stack int
3802ext4_mb_release_group_pa(struct ext4_buddy *e4b,
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003803 struct ext4_prealloc_space *pa)
Alex Tomasc9de5602008-01-29 00:19:52 -05003804{
Alex Tomasc9de5602008-01-29 00:19:52 -05003805 struct super_block *sb = e4b->bd_sb;
3806 ext4_group_t group;
3807 ext4_grpblk_t bit;
3808
Yongqiang Yang60e07cf2011-12-18 15:49:54 -05003809 trace_ext4_mb_release_group_pa(sb, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003810 BUG_ON(pa->pa_deleted == 0);
3811 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3812 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3813 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3814 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003815 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05003816
3817 return 0;
3818}
3819
3820/*
3821 * releases all preallocations in given group
3822 *
3823 * first, we need to decide discard policy:
3824 * - when do we discard
3825 * 1) ENOSPC
3826 * - how many do we discard
3827 * 1) how many requested
3828 */
Eric Sandeen4ddfef72008-04-29 08:11:12 -04003829static noinline_for_stack int
3830ext4_mb_discard_group_preallocations(struct super_block *sb,
Alex Tomasc9de5602008-01-29 00:19:52 -05003831 ext4_group_t group, int needed)
3832{
3833 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3834 struct buffer_head *bitmap_bh = NULL;
3835 struct ext4_prealloc_space *pa, *tmp;
3836 struct list_head list;
3837 struct ext4_buddy e4b;
3838 int err;
3839 int busy = 0;
3840 int free = 0;
3841
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003842 mb_debug(1, "discard preallocation for group %u\n", group);
Alex Tomasc9de5602008-01-29 00:19:52 -05003843
3844 if (list_empty(&grp->bb_prealloc_list))
3845 return 0;
3846
Theodore Ts'o574ca172008-07-11 19:27:31 -04003847 bitmap_bh = ext4_read_block_bitmap(sb, group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04003848 if (IS_ERR(bitmap_bh)) {
3849 err = PTR_ERR(bitmap_bh);
3850 ext4_error(sb, "Error %d reading block bitmap for %u",
3851 err, group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003852 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05003853 }
3854
3855 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003856 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05003857 ext4_error(sb, "Error loading buddy information for %u", group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04003858 put_bh(bitmap_bh);
3859 return 0;
3860 }
Alex Tomasc9de5602008-01-29 00:19:52 -05003861
3862 if (needed == 0)
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04003863 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
Alex Tomasc9de5602008-01-29 00:19:52 -05003864
Alex Tomasc9de5602008-01-29 00:19:52 -05003865 INIT_LIST_HEAD(&list);
Alex Tomasc9de5602008-01-29 00:19:52 -05003866repeat:
3867 ext4_lock_group(sb, group);
3868 list_for_each_entry_safe(pa, tmp,
3869 &grp->bb_prealloc_list, pa_group_list) {
3870 spin_lock(&pa->pa_lock);
3871 if (atomic_read(&pa->pa_count)) {
3872 spin_unlock(&pa->pa_lock);
3873 busy = 1;
3874 continue;
3875 }
3876 if (pa->pa_deleted) {
3877 spin_unlock(&pa->pa_lock);
3878 continue;
3879 }
3880
3881 /* seems this one can be freed ... */
3882 pa->pa_deleted = 1;
3883
3884 /* we can trust pa_free ... */
3885 free += pa->pa_free;
3886
3887 spin_unlock(&pa->pa_lock);
3888
3889 list_del(&pa->pa_group_list);
3890 list_add(&pa->u.pa_tmp_list, &list);
3891 }
3892
3893 /* if we still need more blocks and some PAs were used, try again */
3894 if (free < needed && busy) {
3895 busy = 0;
3896 ext4_unlock_group(sb, group);
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04003897 cond_resched();
Alex Tomasc9de5602008-01-29 00:19:52 -05003898 goto repeat;
3899 }
3900
3901 /* found anything to free? */
3902 if (list_empty(&list)) {
3903 BUG_ON(free != 0);
3904 goto out;
3905 }
3906
3907 /* now free all selected PAs */
3908 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3909
3910 /* remove from object (inode or locality group) */
3911 spin_lock(pa->pa_obj_lock);
3912 list_del_rcu(&pa->pa_inode_list);
3913 spin_unlock(pa->pa_obj_lock);
3914
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04003915 if (pa->pa_type == MB_GROUP_PA)
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003916 ext4_mb_release_group_pa(&e4b, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003917 else
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04003918 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05003919
3920 list_del(&pa->u.pa_tmp_list);
3921 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3922 }
3923
3924out:
3925 ext4_unlock_group(sb, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04003926 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05003927 put_bh(bitmap_bh);
3928 return free;
3929}
3930
3931/*
3932 * releases all non-used preallocated blocks for given inode
3933 *
3934 * It's important to discard preallocations under i_data_sem
3935 * We don't want another block to be served from the prealloc
3936 * space when we are discarding the inode prealloc space.
3937 *
3938 * FIXME!! Make sure it is valid at all the call sites
3939 */
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003940void ext4_discard_preallocations(struct inode *inode)
Alex Tomasc9de5602008-01-29 00:19:52 -05003941{
3942 struct ext4_inode_info *ei = EXT4_I(inode);
3943 struct super_block *sb = inode->i_sb;
3944 struct buffer_head *bitmap_bh = NULL;
3945 struct ext4_prealloc_space *pa, *tmp;
3946 ext4_group_t group = 0;
3947 struct list_head list;
3948 struct ext4_buddy e4b;
3949 int err;
3950
Theodore Ts'oc2ea3fd2008-10-10 09:40:52 -04003951 if (!S_ISREG(inode->i_mode)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05003952 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3953 return;
3954 }
3955
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04003956 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
Theodore Ts'o9bffad12009-06-17 11:48:11 -04003957 trace_ext4_discard_preallocations(inode);
Alex Tomasc9de5602008-01-29 00:19:52 -05003958
3959 INIT_LIST_HEAD(&list);
3960
3961repeat:
3962 /* first, collect all pa's in the inode */
3963 spin_lock(&ei->i_prealloc_lock);
3964 while (!list_empty(&ei->i_prealloc_list)) {
3965 pa = list_entry(ei->i_prealloc_list.next,
3966 struct ext4_prealloc_space, pa_inode_list);
3967 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3968 spin_lock(&pa->pa_lock);
3969 if (atomic_read(&pa->pa_count)) {
3970 /* this shouldn't happen often - nobody should
3971 * use preallocation while we're discarding it */
3972 spin_unlock(&pa->pa_lock);
3973 spin_unlock(&ei->i_prealloc_lock);
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04003974 ext4_msg(sb, KERN_ERR,
3975 "uh-oh! used pa while discarding");
Alex Tomasc9de5602008-01-29 00:19:52 -05003976 WARN_ON(1);
3977 schedule_timeout_uninterruptible(HZ);
3978 goto repeat;
3979
3980 }
3981 if (pa->pa_deleted == 0) {
3982 pa->pa_deleted = 1;
3983 spin_unlock(&pa->pa_lock);
3984 list_del_rcu(&pa->pa_inode_list);
3985 list_add(&pa->u.pa_tmp_list, &list);
3986 continue;
3987 }
3988
3989 /* someone is deleting pa right now */
3990 spin_unlock(&pa->pa_lock);
3991 spin_unlock(&ei->i_prealloc_lock);
3992
3993 /* we have to wait here because pa_deleted
3994 * doesn't mean pa is already unlinked from
3995 * the list. as we might be called from
3996 * ->clear_inode() the inode will get freed
3997 * and concurrent thread which is unlinking
3998 * pa from inode's list may access already
3999 * freed memory, bad-bad-bad */
4000
4001 /* XXX: if this happens too often, we can
4002 * add a flag to force wait only in case
4003 * of ->clear_inode(), but not in case of
4004 * regular truncate */
4005 schedule_timeout_uninterruptible(HZ);
4006 goto repeat;
4007 }
4008 spin_unlock(&ei->i_prealloc_lock);
4009
4010 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004011 BUG_ON(pa->pa_type != MB_INODE_PA);
Lukas Czernerbd862982013-04-03 23:32:34 -04004012 group = ext4_get_group_number(sb, pa->pa_pstart);
Alex Tomasc9de5602008-01-29 00:19:52 -05004013
4014 err = ext4_mb_load_buddy(sb, group, &e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004015 if (err) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004016 ext4_error(sb, "Error loading buddy information for %u",
4017 group);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004018 continue;
4019 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004020
Theodore Ts'o574ca172008-07-11 19:27:31 -04004021 bitmap_bh = ext4_read_block_bitmap(sb, group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04004022 if (IS_ERR(bitmap_bh)) {
4023 err = PTR_ERR(bitmap_bh);
4024 ext4_error(sb, "Error %d reading block bitmap for %u",
4025 err, group);
Jing Zhange39e07f2010-05-14 00:00:00 -04004026 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004027 continue;
Alex Tomasc9de5602008-01-29 00:19:52 -05004028 }
4029
4030 ext4_lock_group(sb, group);
4031 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004032 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
Alex Tomasc9de5602008-01-29 00:19:52 -05004033 ext4_unlock_group(sb, group);
4034
Jing Zhange39e07f2010-05-14 00:00:00 -04004035 ext4_mb_unload_buddy(&e4b);
Alex Tomasc9de5602008-01-29 00:19:52 -05004036 put_bh(bitmap_bh);
4037
4038 list_del(&pa->u.pa_tmp_list);
4039 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4040 }
4041}
4042
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004043#ifdef CONFIG_EXT4_DEBUG
Alex Tomasc9de5602008-01-29 00:19:52 -05004044static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4045{
4046 struct super_block *sb = ac->ac_sb;
Theodore Ts'o8df96752009-05-01 08:50:38 -04004047 ext4_group_t ngroups, i;
Alex Tomasc9de5602008-01-29 00:19:52 -05004048
Theodore Ts'oa0b30c12013-02-09 16:28:20 -05004049 if (!ext4_mballoc_debug ||
Theodore Ts'o4dd89fc2011-02-27 17:23:47 -05004050 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
Eric Sandeene3570632010-07-27 11:56:08 -04004051 return;
4052
Joe Perches7f6a11e2012-03-19 23:09:43 -04004053 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004054 " Allocation context details:");
Joe Perches7f6a11e2012-03-19 23:09:43 -04004055 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004056 ac->ac_status, ac->ac_flags);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004057 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
Theodore Ts'o9d8b9ec2011-08-01 17:41:35 -04004058 "goal %lu/%lu/%lu@%lu, "
4059 "best %lu/%lu/%lu@%lu cr %d",
Alex Tomasc9de5602008-01-29 00:19:52 -05004060 (unsigned long)ac->ac_o_ex.fe_group,
4061 (unsigned long)ac->ac_o_ex.fe_start,
4062 (unsigned long)ac->ac_o_ex.fe_len,
4063 (unsigned long)ac->ac_o_ex.fe_logical,
4064 (unsigned long)ac->ac_g_ex.fe_group,
4065 (unsigned long)ac->ac_g_ex.fe_start,
4066 (unsigned long)ac->ac_g_ex.fe_len,
4067 (unsigned long)ac->ac_g_ex.fe_logical,
4068 (unsigned long)ac->ac_b_ex.fe_group,
4069 (unsigned long)ac->ac_b_ex.fe_start,
4070 (unsigned long)ac->ac_b_ex.fe_len,
4071 (unsigned long)ac->ac_b_ex.fe_logical,
4072 (int)ac->ac_criteria);
Eric Sandeendc9ddd92014-02-20 13:32:10 -05004073 ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
Joe Perches7f6a11e2012-03-19 23:09:43 -04004074 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
Theodore Ts'o8df96752009-05-01 08:50:38 -04004075 ngroups = ext4_get_groups_count(sb);
4076 for (i = 0; i < ngroups; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004077 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4078 struct ext4_prealloc_space *pa;
4079 ext4_grpblk_t start;
4080 struct list_head *cur;
4081 ext4_lock_group(sb, i);
4082 list_for_each(cur, &grp->bb_prealloc_list) {
4083 pa = list_entry(cur, struct ext4_prealloc_space,
4084 pa_group_list);
4085 spin_lock(&pa->pa_lock);
4086 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4087 NULL, &start);
4088 spin_unlock(&pa->pa_lock);
Akira Fujita1c718502009-07-05 23:04:36 -04004089 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4090 start, pa->pa_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004091 }
Solofo Ramangalahy60bd63d2008-04-29 21:59:59 -04004092 ext4_unlock_group(sb, i);
Alex Tomasc9de5602008-01-29 00:19:52 -05004093
4094 if (grp->bb_free == 0)
4095 continue;
Akira Fujita1c718502009-07-05 23:04:36 -04004096 printk(KERN_ERR "%u: %d/%d \n",
Alex Tomasc9de5602008-01-29 00:19:52 -05004097 i, grp->bb_free, grp->bb_fragments);
4098 }
4099 printk(KERN_ERR "\n");
4100}
4101#else
4102static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4103{
4104 return;
4105}
4106#endif
4107
4108/*
4109 * We use locality group preallocation for small size file. The size of the
4110 * file is determined by the current size or the resulting size after
4111 * allocation which ever is larger
4112 *
Theodore Ts'ob713a5e2009-03-31 09:11:14 -04004113 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
Alex Tomasc9de5602008-01-29 00:19:52 -05004114 */
4115static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4116{
4117 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4118 int bsbits = ac->ac_sb->s_blocksize_bits;
4119 loff_t size, isize;
4120
4121 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4122 return;
4123
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004124 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4125 return;
4126
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004127 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
Theodore Ts'o50797482009-09-18 13:34:02 -04004128 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4129 >> bsbits;
Alex Tomasc9de5602008-01-29 00:19:52 -05004130
Theodore Ts'o50797482009-09-18 13:34:02 -04004131 if ((size == isize) &&
4132 !ext4_fs_is_busy(sbi) &&
4133 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4134 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4135 return;
4136 }
4137
Robin Dongebbe0272011-10-26 05:14:27 -04004138 if (sbi->s_mb_group_prealloc <= 0) {
4139 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4140 return;
4141 }
4142
Alex Tomasc9de5602008-01-29 00:19:52 -05004143 /* don't use group allocation for large files */
Theodore Ts'o71780572009-09-28 00:06:20 -04004144 size = max(size, isize);
Tao Macc483f12010-03-01 19:06:35 -05004145 if (size > sbi->s_mb_stream_request) {
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004146 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
Alex Tomasc9de5602008-01-29 00:19:52 -05004147 return;
Theodore Ts'o4ba74d02009-08-09 22:01:13 -04004148 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004149
4150 BUG_ON(ac->ac_lg != NULL);
4151 /*
4152 * locality group prealloc space are per cpu. The reason for having
4153 * per cpu locality group is to reduce the contention between block
4154 * request from multiple CPUs.
4155 */
Christoph Lametera0b6bc62014-08-17 12:30:28 -05004156 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
Alex Tomasc9de5602008-01-29 00:19:52 -05004157
4158 /* we're going to use group allocation */
4159 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4160
4161 /* serialize all allocations in the group */
4162 mutex_lock(&ac->ac_lg->lg_mutex);
4163}
4164
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004165static noinline_for_stack int
4166ext4_mb_initialize_context(struct ext4_allocation_context *ac,
Alex Tomasc9de5602008-01-29 00:19:52 -05004167 struct ext4_allocation_request *ar)
4168{
4169 struct super_block *sb = ar->inode->i_sb;
4170 struct ext4_sb_info *sbi = EXT4_SB(sb);
4171 struct ext4_super_block *es = sbi->s_es;
4172 ext4_group_t group;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004173 unsigned int len;
4174 ext4_fsblk_t goal;
Alex Tomasc9de5602008-01-29 00:19:52 -05004175 ext4_grpblk_t block;
4176
4177 /* we can't allocate > group size */
4178 len = ar->len;
4179
4180 /* just a dirty hack to filter too big requests */
Theodore Ts'o40ae3482013-02-04 15:08:40 -05004181 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4182 len = EXT4_CLUSTERS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004183
4184 /* start searching from the goal */
4185 goal = ar->goal;
4186 if (goal < le32_to_cpu(es->s_first_data_block) ||
4187 goal >= ext4_blocks_count(es))
4188 goal = le32_to_cpu(es->s_first_data_block);
4189 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4190
4191 /* set up allocation goals */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004192 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
Alex Tomasc9de5602008-01-29 00:19:52 -05004193 ac->ac_status = AC_STATUS_CONTINUE;
Alex Tomasc9de5602008-01-29 00:19:52 -05004194 ac->ac_sb = sb;
4195 ac->ac_inode = ar->inode;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004196 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
Alex Tomasc9de5602008-01-29 00:19:52 -05004197 ac->ac_o_ex.fe_group = group;
4198 ac->ac_o_ex.fe_start = block;
4199 ac->ac_o_ex.fe_len = len;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004200 ac->ac_g_ex = ac->ac_o_ex;
Alex Tomasc9de5602008-01-29 00:19:52 -05004201 ac->ac_flags = ar->flags;
Alex Tomasc9de5602008-01-29 00:19:52 -05004202
4203 /* we have to define context: we'll we work with a file or
4204 * locality group. this is a policy, actually */
4205 ext4_mb_group_or_file(ac);
4206
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004207 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
Alex Tomasc9de5602008-01-29 00:19:52 -05004208 "left: %u/%u, right %u/%u to %swritable\n",
4209 (unsigned) ar->len, (unsigned) ar->logical,
4210 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4211 (unsigned) ar->lleft, (unsigned) ar->pleft,
4212 (unsigned) ar->lright, (unsigned) ar->pright,
4213 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4214 return 0;
4215
4216}
4217
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004218static noinline_for_stack void
4219ext4_mb_discard_lg_preallocations(struct super_block *sb,
4220 struct ext4_locality_group *lg,
4221 int order, int total_entries)
4222{
4223 ext4_group_t group = 0;
4224 struct ext4_buddy e4b;
4225 struct list_head discard_list;
4226 struct ext4_prealloc_space *pa, *tmp;
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004227
Theodore Ts'o6ba495e2009-09-18 13:38:55 -04004228 mb_debug(1, "discard locality group preallocation\n");
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004229
4230 INIT_LIST_HEAD(&discard_list);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004231
4232 spin_lock(&lg->lg_prealloc_lock);
4233 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4234 pa_inode_list) {
4235 spin_lock(&pa->pa_lock);
4236 if (atomic_read(&pa->pa_count)) {
4237 /*
4238 * This is the pa that we just used
4239 * for block allocation. So don't
4240 * free that
4241 */
4242 spin_unlock(&pa->pa_lock);
4243 continue;
4244 }
4245 if (pa->pa_deleted) {
4246 spin_unlock(&pa->pa_lock);
4247 continue;
4248 }
4249 /* only lg prealloc space */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004250 BUG_ON(pa->pa_type != MB_GROUP_PA);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004251
4252 /* seems this one can be freed ... */
4253 pa->pa_deleted = 1;
4254 spin_unlock(&pa->pa_lock);
4255
4256 list_del_rcu(&pa->pa_inode_list);
4257 list_add(&pa->u.pa_tmp_list, &discard_list);
4258
4259 total_entries--;
4260 if (total_entries <= 5) {
4261 /*
4262 * we want to keep only 5 entries
4263 * allowing it to grow to 8. This
4264 * mak sure we don't call discard
4265 * soon for this list.
4266 */
4267 break;
4268 }
4269 }
4270 spin_unlock(&lg->lg_prealloc_lock);
4271
4272 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4273
Lukas Czernerbd862982013-04-03 23:32:34 -04004274 group = ext4_get_group_number(sb, pa->pa_pstart);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004275 if (ext4_mb_load_buddy(sb, group, &e4b)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004276 ext4_error(sb, "Error loading buddy information for %u",
4277 group);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004278 continue;
4279 }
4280 ext4_lock_group(sb, group);
4281 list_del(&pa->pa_group_list);
Eric Sandeen3e1e5f52010-10-27 21:30:07 -04004282 ext4_mb_release_group_pa(&e4b, pa);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004283 ext4_unlock_group(sb, group);
4284
Jing Zhange39e07f2010-05-14 00:00:00 -04004285 ext4_mb_unload_buddy(&e4b);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004286 list_del(&pa->u.pa_tmp_list);
4287 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4288 }
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004289}
4290
4291/*
4292 * We have incremented pa_count. So it cannot be freed at this
4293 * point. Also we hold lg_mutex. So no parallel allocation is
4294 * possible from this lg. That means pa_free cannot be updated.
4295 *
4296 * A parallel ext4_mb_discard_group_preallocations is possible.
4297 * which can cause the lg_prealloc_list to be updated.
4298 */
4299
4300static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4301{
4302 int order, added = 0, lg_prealloc_count = 1;
4303 struct super_block *sb = ac->ac_sb;
4304 struct ext4_locality_group *lg = ac->ac_lg;
4305 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4306
4307 order = fls(pa->pa_free) - 1;
4308 if (order > PREALLOC_TB_SIZE - 1)
4309 /* The max size of hash table is PREALLOC_TB_SIZE */
4310 order = PREALLOC_TB_SIZE - 1;
4311 /* Add the prealloc space to lg */
Niu Yaweif1167002013-02-01 21:31:27 -05004312 spin_lock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004313 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4314 pa_inode_list) {
4315 spin_lock(&tmp_pa->pa_lock);
4316 if (tmp_pa->pa_deleted) {
Theodore Ts'oe7c9e3e2009-03-27 19:43:21 -04004317 spin_unlock(&tmp_pa->pa_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004318 continue;
4319 }
4320 if (!added && pa->pa_free < tmp_pa->pa_free) {
4321 /* Add to the tail of the previous entry */
4322 list_add_tail_rcu(&pa->pa_inode_list,
4323 &tmp_pa->pa_inode_list);
4324 added = 1;
4325 /*
4326 * we want to count the total
4327 * number of entries in the list
4328 */
4329 }
4330 spin_unlock(&tmp_pa->pa_lock);
4331 lg_prealloc_count++;
4332 }
4333 if (!added)
4334 list_add_tail_rcu(&pa->pa_inode_list,
4335 &lg->lg_prealloc_list[order]);
Niu Yaweif1167002013-02-01 21:31:27 -05004336 spin_unlock(&lg->lg_prealloc_lock);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004337
4338 /* Now trim the list to be not more than 8 elements */
4339 if (lg_prealloc_count > 8) {
4340 ext4_mb_discard_lg_preallocations(sb, lg,
Niu Yaweif1167002013-02-01 21:31:27 -05004341 order, lg_prealloc_count);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004342 return;
4343 }
4344 return ;
4345}
4346
Alex Tomasc9de5602008-01-29 00:19:52 -05004347/*
4348 * release all resource we used in allocation
4349 */
4350static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4351{
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004352 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004353 struct ext4_prealloc_space *pa = ac->ac_pa;
4354 if (pa) {
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004355 if (pa->pa_type == MB_GROUP_PA) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004356 /* see comment in ext4_mb_use_group_pa() */
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004357 spin_lock(&pa->pa_lock);
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004358 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4359 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
Aneesh Kumar K.V6be2ded2008-07-23 14:14:05 -04004360 pa->pa_free -= ac->ac_b_ex.fe_len;
4361 pa->pa_len -= ac->ac_b_ex.fe_len;
4362 spin_unlock(&pa->pa_lock);
Alex Tomasc9de5602008-01-29 00:19:52 -05004363 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004364 }
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004365 if (pa) {
4366 /*
4367 * We want to add the pa to the right bucket.
4368 * Remove it from the list and while adding
4369 * make sure the list to which we are adding
Amir Goldstein44183d42011-05-09 21:52:36 -04004370 * doesn't grow big.
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004371 */
Aneesh Kumar K.Vcc0fb9a2009-03-27 17:16:58 -04004372 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
Aneesh Kumar K.Vba443912009-02-10 11:14:34 -05004373 spin_lock(pa->pa_obj_lock);
4374 list_del_rcu(&pa->pa_inode_list);
4375 spin_unlock(pa->pa_obj_lock);
4376 ext4_mb_add_n_trim(ac);
4377 }
4378 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4379 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004380 if (ac->ac_bitmap_page)
4381 page_cache_release(ac->ac_bitmap_page);
4382 if (ac->ac_buddy_page)
4383 page_cache_release(ac->ac_buddy_page);
4384 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4385 mutex_unlock(&ac->ac_lg->lg_mutex);
4386 ext4_mb_collect_stats(ac);
4387 return 0;
4388}
4389
4390static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4391{
Theodore Ts'o8df96752009-05-01 08:50:38 -04004392 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004393 int ret;
4394 int freed = 0;
4395
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004396 trace_ext4_mb_discard_preallocations(sb, needed);
Theodore Ts'o8df96752009-05-01 08:50:38 -04004397 for (i = 0; i < ngroups && needed > 0; i++) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004398 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4399 freed += ret;
4400 needed -= ret;
4401 }
4402
4403 return freed;
4404}
4405
4406/*
4407 * Main entry point into mballoc to allocate blocks
4408 * it tries to use preallocation first, then falls back
4409 * to usual allocation
4410 */
4411ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
Aditya Kali6c7a1202010-08-05 16:22:24 -04004412 struct ext4_allocation_request *ar, int *errp)
Alex Tomasc9de5602008-01-29 00:19:52 -05004413{
Aneesh Kumar K.V6bc6e632008-10-10 09:39:00 -04004414 int freed;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004415 struct ext4_allocation_context *ac = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004416 struct ext4_sb_info *sbi;
4417 struct super_block *sb;
4418 ext4_fsblk_t block = 0;
Mingming Cao60e58e02009-01-22 18:13:05 +01004419 unsigned int inquota = 0;
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004420 unsigned int reserv_clstrs = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004421
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004422 might_sleep();
Alex Tomasc9de5602008-01-29 00:19:52 -05004423 sb = ar->inode->i_sb;
4424 sbi = EXT4_SB(sb);
4425
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004426 trace_ext4_request_blocks(ar);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004427
Dmitry Monakhov45dc63e2011-10-20 20:07:23 -04004428 /* Allow to use superuser reservation for quota file */
4429 if (IS_NOQUOTA(ar->inode))
4430 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4431
Theodore Ts'oe3cf5d52014-09-04 18:07:25 -04004432 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
Mingming Cao60e58e02009-01-22 18:13:05 +01004433 /* Without delayed allocation we need to verify
4434 * there is enough free blocks to do block allocation
4435 * and verify allocation doesn't exceed the quota limits.
Mingming Caod2a17632008-07-14 17:52:37 -04004436 */
Allison Henderson55f020d2011-05-25 07:41:26 -04004437 while (ar->len &&
Theodore Ts'oe7d5f312011-09-09 19:14:51 -04004438 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004439
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004440 /* let others to free the space */
Lukas Czernerbb8b20e2013-03-10 22:28:09 -04004441 cond_resched();
Aneesh Kumar K.V030ba6b2008-09-08 23:14:50 -04004442 ar->len = ar->len >> 1;
4443 }
4444 if (!ar->len) {
Aneesh Kumar K.Va30d542a2008-10-09 10:56:23 -04004445 *errp = -ENOSPC;
4446 return 0;
4447 }
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004448 reserv_clstrs = ar->len;
Allison Henderson55f020d2011-05-25 07:41:26 -04004449 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004450 dquot_alloc_block_nofail(ar->inode,
4451 EXT4_C2B(sbi, ar->len));
Allison Henderson55f020d2011-05-25 07:41:26 -04004452 } else {
4453 while (ar->len &&
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004454 dquot_alloc_block(ar->inode,
4455 EXT4_C2B(sbi, ar->len))) {
Allison Henderson55f020d2011-05-25 07:41:26 -04004456
4457 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4458 ar->len--;
4459 }
Mingming Cao60e58e02009-01-22 18:13:05 +01004460 }
4461 inquota = ar->len;
4462 if (ar->len == 0) {
4463 *errp = -EDQUOT;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004464 goto out;
Mingming Cao60e58e02009-01-22 18:13:05 +01004465 }
Mingming Caod2a17632008-07-14 17:52:37 -04004466 }
Mingming Caod2a17632008-07-14 17:52:37 -04004467
Wei Yongjun85556c92012-09-26 20:43:37 -04004468 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
Theodore Ts'o833576b2009-07-13 09:45:52 -04004469 if (!ac) {
Shen Feng363d4252008-07-11 19:27:31 -04004470 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004471 *errp = -ENOMEM;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004472 goto out;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004473 }
4474
Eric Sandeen256bdb42008-02-10 01:13:33 -05004475 *errp = ext4_mb_initialize_context(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004476 if (*errp) {
4477 ar->len = 0;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004478 goto out;
Alex Tomasc9de5602008-01-29 00:19:52 -05004479 }
4480
Eric Sandeen256bdb42008-02-10 01:13:33 -05004481 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4482 if (!ext4_mb_use_preallocated(ac)) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004483 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4484 ext4_mb_normalize_request(ac, ar);
Alex Tomasc9de5602008-01-29 00:19:52 -05004485repeat:
4486 /* allocate space in core */
Aditya Kali6c7a1202010-08-05 16:22:24 -04004487 *errp = ext4_mb_regular_allocator(ac);
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004488 if (*errp)
4489 goto discard_and_exit;
4490
4491 /* as we've just preallocated more space than
4492 * user requested originally, we store allocated
4493 * space in a special descriptor */
4494 if (ac->ac_status == AC_STATUS_FOUND &&
4495 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4496 *errp = ext4_mb_new_preallocation(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004497 if (*errp) {
Alexey Khoroshilov2c00ef32013-07-01 08:12:36 -04004498 discard_and_exit:
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004499 ext4_discard_allocated_blocks(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004500 goto errout;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004501 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004502 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004503 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004504 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004505 if (*errp == -EAGAIN) {
Aneesh Kumar K.V8556e8f2009-01-05 21:46:55 -05004506 /*
4507 * drop the reference that we took
4508 * in ext4_mb_use_best_found
4509 */
4510 ext4_mb_release_context(ac);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004511 ac->ac_b_ex.fe_group = 0;
4512 ac->ac_b_ex.fe_start = 0;
4513 ac->ac_b_ex.fe_len = 0;
4514 ac->ac_status = AC_STATUS_CONTINUE;
4515 goto repeat;
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004516 } else if (*errp) {
Curt Wohlgemuthb8441672009-12-08 22:18:25 -05004517 ext4_discard_allocated_blocks(ac);
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004518 goto errout;
4519 } else {
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004520 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4521 ar->len = ac->ac_b_ex.fe_len;
4522 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004523 } else {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004524 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
Alex Tomasc9de5602008-01-29 00:19:52 -05004525 if (freed)
4526 goto repeat;
4527 *errp = -ENOSPC;
Aditya Kali6c7a1202010-08-05 16:22:24 -04004528 }
4529
Eric Sandeen6d138ce2012-11-08 11:11:59 -05004530errout:
Aditya Kali6c7a1202010-08-05 16:22:24 -04004531 if (*errp) {
Eric Sandeen256bdb42008-02-10 01:13:33 -05004532 ac->ac_b_ex.fe_len = 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004533 ar->len = 0;
Eric Sandeen256bdb42008-02-10 01:13:33 -05004534 ext4_mb_show_ac(ac);
Alex Tomasc9de5602008-01-29 00:19:52 -05004535 }
Eric Sandeen256bdb42008-02-10 01:13:33 -05004536 ext4_mb_release_context(ac);
Aditya Kali6c7a1202010-08-05 16:22:24 -04004537out:
4538 if (ac)
4539 kmem_cache_free(ext4_ac_cachep, ac);
Mingming Cao60e58e02009-01-22 18:13:05 +01004540 if (inquota && ar->len < inquota)
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004541 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004542 if (!ar->len) {
Theodore Ts'oe3cf5d52014-09-04 18:07:25 -04004543 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004544 /* release all the reserved blocks if non delalloc */
Theodore Ts'o57042652011-09-09 18:56:51 -04004545 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
Theodore Ts'o53accfa2011-09-09 18:48:51 -04004546 reserv_clstrs);
Aneesh Kumar K.V0087d9f2009-01-05 21:49:12 -05004547 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004548
Theodore Ts'o9bffad12009-06-17 11:48:11 -04004549 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
Theodore Ts'oba80b102009-01-03 20:03:21 -05004550
Alex Tomasc9de5602008-01-29 00:19:52 -05004551 return block;
4552}
Alex Tomasc9de5602008-01-29 00:19:52 -05004553
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004554/*
4555 * We can merge two free data extents only if the physical blocks
4556 * are contiguous, AND the extents were freed by the same transaction,
4557 * AND the blocks are associated with the same group.
4558 */
4559static int can_merge(struct ext4_free_data *entry1,
4560 struct ext4_free_data *entry2)
4561{
Bobi Jam18aadd42012-02-20 17:53:02 -05004562 if ((entry1->efd_tid == entry2->efd_tid) &&
4563 (entry1->efd_group == entry2->efd_group) &&
4564 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004565 return 1;
4566 return 0;
4567}
4568
Eric Sandeen4ddfef72008-04-29 08:11:12 -04004569static noinline_for_stack int
4570ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004571 struct ext4_free_data *new_entry)
Alex Tomasc9de5602008-01-29 00:19:52 -05004572{
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004573 ext4_group_t group = e4b->bd_group;
Theodore Ts'o84130192011-09-09 18:50:51 -04004574 ext4_grpblk_t cluster;
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004575 struct ext4_free_data *entry;
Alex Tomasc9de5602008-01-29 00:19:52 -05004576 struct ext4_group_info *db = e4b->bd_info;
4577 struct super_block *sb = e4b->bd_sb;
4578 struct ext4_sb_info *sbi = EXT4_SB(sb);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004579 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4580 struct rb_node *parent = NULL, *new_node;
4581
Frank Mayhar03901312009-01-07 00:06:22 -05004582 BUG_ON(!ext4_handle_valid(handle));
Alex Tomasc9de5602008-01-29 00:19:52 -05004583 BUG_ON(e4b->bd_bitmap_page == NULL);
4584 BUG_ON(e4b->bd_buddy_page == NULL);
4585
Bobi Jam18aadd42012-02-20 17:53:02 -05004586 new_node = &new_entry->efd_node;
4587 cluster = new_entry->efd_start_cluster;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004588
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004589 if (!*n) {
4590 /* first free block exent. We need to
4591 protect buddy cache from being freed,
4592 * otherwise we'll refresh it from
4593 * on-disk bitmap and lose not-yet-available
4594 * blocks */
4595 page_cache_get(e4b->bd_buddy_page);
4596 page_cache_get(e4b->bd_bitmap_page);
4597 }
4598 while (*n) {
4599 parent = *n;
Bobi Jam18aadd42012-02-20 17:53:02 -05004600 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4601 if (cluster < entry->efd_start_cluster)
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004602 n = &(*n)->rb_left;
Bobi Jam18aadd42012-02-20 17:53:02 -05004603 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004604 n = &(*n)->rb_right;
4605 else {
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004606 ext4_grp_locked_error(sb, group, 0,
Theodore Ts'o84130192011-09-09 18:50:51 -04004607 ext4_group_first_block_no(sb, group) +
4608 EXT4_C2B(sbi, cluster),
Theodore Ts'oe29136f2010-06-29 12:54:28 -04004609 "Block already on to-be-freed list");
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004610 return 0;
Alex Tomasc9de5602008-01-29 00:19:52 -05004611 }
4612 }
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004613
4614 rb_link_node(new_node, parent, n);
4615 rb_insert_color(new_node, &db->bb_free_root);
4616
4617 /* Now try to see the extent can be merged to left and right */
4618 node = rb_prev(new_node);
4619 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004620 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004621 if (can_merge(entry, new_entry) &&
4622 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004623 new_entry->efd_start_cluster = entry->efd_start_cluster;
4624 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004625 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004626 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004627 }
4628 }
4629
4630 node = rb_next(new_node);
4631 if (node) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004632 entry = rb_entry(node, struct ext4_free_data, efd_node);
Dmitry Monakhov5d3ee202013-04-03 22:08:52 -04004633 if (can_merge(new_entry, entry) &&
4634 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
Bobi Jam18aadd42012-02-20 17:53:02 -05004635 new_entry->efd_count += entry->efd_count;
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004636 rb_erase(node, &(db->bb_free_root));
Bobi Jam18aadd42012-02-20 17:53:02 -05004637 kmem_cache_free(ext4_free_data_cachep, entry);
Aneesh Kumar K.Vc8940582008-10-16 10:14:27 -04004638 }
4639 }
Theodore Ts'o3e624fc2008-10-16 20:00:24 -04004640 /* Add the extent to transaction's private list */
Bobi Jam18aadd42012-02-20 17:53:02 -05004641 ext4_journal_callback_add(handle, ext4_free_data_callback,
4642 &new_entry->efd_jce);
Alex Tomasc9de5602008-01-29 00:19:52 -05004643 return 0;
4644}
4645
Theodore Ts'o44338712009-11-22 07:44:56 -05004646/**
4647 * ext4_free_blocks() -- Free given blocks and update quota
4648 * @handle: handle for this transaction
4649 * @inode: inode
4650 * @block: start physical block to free
4651 * @count: number of blocks to count
Yongqiang Yang5def1362011-06-05 23:26:40 -04004652 * @flags: flags used by ext4_free_blocks
Alex Tomasc9de5602008-01-29 00:19:52 -05004653 */
Theodore Ts'o44338712009-11-22 07:44:56 -05004654void ext4_free_blocks(handle_t *handle, struct inode *inode,
Theodore Ts'oe6362602009-11-23 07:17:05 -05004655 struct buffer_head *bh, ext4_fsblk_t block,
4656 unsigned long count, int flags)
Alex Tomasc9de5602008-01-29 00:19:52 -05004657{
Aneesh Kumar K.V26346ff2008-02-10 01:10:04 -05004658 struct buffer_head *bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004659 struct super_block *sb = inode->i_sb;
Alex Tomasc9de5602008-01-29 00:19:52 -05004660 struct ext4_group_desc *gdp;
Theodore Ts'o498e5f22008-11-05 00:14:04 -05004661 unsigned int overflow;
Alex Tomasc9de5602008-01-29 00:19:52 -05004662 ext4_grpblk_t bit;
4663 struct buffer_head *gd_bh;
4664 ext4_group_t block_group;
4665 struct ext4_sb_info *sbi;
4666 struct ext4_buddy e4b;
Theodore Ts'o84130192011-09-09 18:50:51 -04004667 unsigned int count_clusters;
Alex Tomasc9de5602008-01-29 00:19:52 -05004668 int err = 0;
4669 int ret;
4670
Theodore Ts'ob10a44c2013-04-03 22:00:52 -04004671 might_sleep();
Theodore Ts'oe6362602009-11-23 07:17:05 -05004672 if (bh) {
4673 if (block)
4674 BUG_ON(block != bh->b_blocknr);
4675 else
4676 block = bh->b_blocknr;
4677 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004678
Alex Tomasc9de5602008-01-29 00:19:52 -05004679 sbi = EXT4_SB(sb);
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004680 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4681 !ext4_data_block_valid(sbi, block, count)) {
Eric Sandeen12062dd2010-02-15 14:19:27 -05004682 ext4_error(sb, "Freeing blocks not in datazone - "
Theodore Ts'o1f2acb62010-01-22 17:40:42 -05004683 "block = %llu, count = %lu", block, count);
Alex Tomasc9de5602008-01-29 00:19:52 -05004684 goto error_return;
4685 }
4686
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004687 ext4_debug("freeing block %llu\n", block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004688 trace_ext4_free_blocks(inode, block, count, flags);
4689
Daeho Jeong9c02ac92015-10-17 22:28:21 -04004690 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4691 BUG_ON(count > 1);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004692
Daeho Jeong9c02ac92015-10-17 22:28:21 -04004693 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4694 inode, bh, block);
Theodore Ts'oe6362602009-11-23 07:17:05 -05004695 }
4696
Theodore Ts'o60e66792010-05-17 07:00:00 -04004697 /*
Theodore Ts'oe6362602009-11-23 07:17:05 -05004698 * We need to make sure we don't reuse the freed block until
4699 * after the transaction is committed, which we can do by
4700 * treating the block as metadata, below. We make an
4701 * exception if the inode is to be written in writeback mode
4702 * since writeback mode has weak data consistency guarantees.
4703 */
4704 if (!ext4_should_writeback_data(inode))
4705 flags |= EXT4_FREE_BLOCKS_METADATA;
Alex Tomasc9de5602008-01-29 00:19:52 -05004706
Theodore Ts'o84130192011-09-09 18:50:51 -04004707 /*
4708 * If the extent to be freed does not begin on a cluster
4709 * boundary, we need to deal with partial clusters at the
4710 * beginning and end of the extent. Normally we will free
4711 * blocks at the beginning or the end unless we are explicitly
4712 * requested to avoid doing so.
4713 */
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004714 overflow = EXT4_PBLK_COFF(sbi, block);
Theodore Ts'o84130192011-09-09 18:50:51 -04004715 if (overflow) {
4716 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4717 overflow = sbi->s_cluster_ratio - overflow;
4718 block += overflow;
4719 if (count > overflow)
4720 count -= overflow;
4721 else
4722 return;
4723 } else {
4724 block -= overflow;
4725 count += overflow;
4726 }
4727 }
Theodore Ts'of5a44db2013-12-20 09:29:35 -05004728 overflow = EXT4_LBLK_COFF(sbi, count);
Theodore Ts'o84130192011-09-09 18:50:51 -04004729 if (overflow) {
4730 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4731 if (count > overflow)
4732 count -= overflow;
4733 else
4734 return;
4735 } else
4736 count += sbi->s_cluster_ratio - overflow;
4737 }
4738
Daeho Jeong9c02ac92015-10-17 22:28:21 -04004739 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4740 int i;
4741
4742 for (i = 0; i < count; i++) {
4743 cond_resched();
4744 bh = sb_find_get_block(inode->i_sb, block + i);
4745 if (!bh)
4746 continue;
4747 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4748 inode, bh, block + i);
4749 }
4750 }
4751
Alex Tomasc9de5602008-01-29 00:19:52 -05004752do_more:
4753 overflow = 0;
4754 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4755
Darrick J. Wong163a2032013-08-28 17:35:51 -04004756 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4757 ext4_get_group_info(sb, block_group))))
4758 return;
4759
Alex Tomasc9de5602008-01-29 00:19:52 -05004760 /*
4761 * Check to see if we are freeing blocks across a group
4762 * boundary.
4763 */
Theodore Ts'o84130192011-09-09 18:50:51 -04004764 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4765 overflow = EXT4_C2B(sbi, bit) + count -
4766 EXT4_BLOCKS_PER_GROUP(sb);
Alex Tomasc9de5602008-01-29 00:19:52 -05004767 count -= overflow;
4768 }
Lukas Czerner810da242013-03-02 17:18:58 -05004769 count_clusters = EXT4_NUM_B2C(sbi, count);
Theodore Ts'o574ca172008-07-11 19:27:31 -04004770 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04004771 if (IS_ERR(bitmap_bh)) {
4772 err = PTR_ERR(bitmap_bh);
4773 bitmap_bh = NULL;
Alex Tomasc9de5602008-01-29 00:19:52 -05004774 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004775 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004776 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004777 if (!gdp) {
4778 err = -EIO;
Alex Tomasc9de5602008-01-29 00:19:52 -05004779 goto error_return;
Aneesh Kumar K.Vce89f462008-07-23 14:09:29 -04004780 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004781
4782 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4783 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4784 in_range(block, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004785 EXT4_SB(sb)->s_itb_per_group) ||
Alex Tomasc9de5602008-01-29 00:19:52 -05004786 in_range(block + count - 1, ext4_inode_table(sb, gdp),
Theodore Ts'o84130192011-09-09 18:50:51 -04004787 EXT4_SB(sb)->s_itb_per_group)) {
Alex Tomasc9de5602008-01-29 00:19:52 -05004788
Eric Sandeen12062dd2010-02-15 14:19:27 -05004789 ext4_error(sb, "Freeing blocks in system zone - "
Theodore Ts'o0610b6e2009-06-15 03:45:05 -04004790 "Block = %llu, count = %lu", block, count);
Aneesh Kumar K.V519deca2008-05-15 14:43:20 -04004791 /* err = 0. ext4_std_error should be a no op */
4792 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004793 }
4794
4795 BUFFER_TRACE(bitmap_bh, "getting write access");
4796 err = ext4_journal_get_write_access(handle, bitmap_bh);
4797 if (err)
4798 goto error_return;
4799
4800 /*
4801 * We are about to modify some metadata. Call the journal APIs
4802 * to unshare ->b_data if a currently-committing transaction is
4803 * using it
4804 */
4805 BUFFER_TRACE(gd_bh, "get_write_access");
4806 err = ext4_journal_get_write_access(handle, gd_bh);
4807 if (err)
4808 goto error_return;
Alex Tomasc9de5602008-01-29 00:19:52 -05004809#ifdef AGGRESSIVE_CHECK
4810 {
4811 int i;
Theodore Ts'o84130192011-09-09 18:50:51 -04004812 for (i = 0; i < count_clusters; i++)
Alex Tomasc9de5602008-01-29 00:19:52 -05004813 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4814 }
4815#endif
Theodore Ts'o84130192011-09-09 18:50:51 -04004816 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004817
Aneesh Kumar K.V920313a2009-01-05 21:36:19 -05004818 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4819 if (err)
4820 goto error_return;
Theodore Ts'oe6362602009-11-23 07:17:05 -05004821
4822 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004823 struct ext4_free_data *new_entry;
4824 /*
4825 * blocks being freed are metadata. these blocks shouldn't
4826 * be used until this transaction is committed
Michal Hocko7444a072015-07-05 12:33:44 -04004827 *
4828 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4829 * to fail.
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004830 */
Michal Hocko7444a072015-07-05 12:33:44 -04004831 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4832 GFP_NOFS|__GFP_NOFAIL);
Bobi Jam18aadd42012-02-20 17:53:02 -05004833 new_entry->efd_start_cluster = bit;
4834 new_entry->efd_group = block_group;
4835 new_entry->efd_count = count_clusters;
4836 new_entry->efd_tid = handle->h_transaction->t_tid;
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004837
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004838 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004839 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004840 ext4_mb_free_metadata(handle, &e4b, new_entry);
Alex Tomasc9de5602008-01-29 00:19:52 -05004841 } else {
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004842 /* need to update group_info->bb_free and bitmap
4843 * with group lock held. generate_buddy look at
4844 * them with group lock_held
4845 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004846 if (test_opt(sb, DISCARD)) {
4847 err = ext4_issue_discard(sb, block_group, bit, count);
4848 if (err && err != -EOPNOTSUPP)
4849 ext4_msg(sb, KERN_WARNING, "discard request in"
4850 " group:%d block:%d count:%lu failed"
4851 " with %d", block_group, bit, count,
4852 err);
Lukas Czerner8f9ff182013-10-30 11:10:52 -04004853 } else
4854 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05004855
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004856 ext4_lock_group(sb, block_group);
Theodore Ts'o84130192011-09-09 18:50:51 -04004857 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4858 mb_free_blocks(inode, &e4b, bit, count_clusters);
Alex Tomasc9de5602008-01-29 00:19:52 -05004859 }
4860
Theodore Ts'o021b65b2011-09-09 19:08:51 -04004861 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4862 ext4_free_group_clusters_set(sb, gdp, ret);
Tao Ma79f1ba42012-10-22 00:34:32 -04004863 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04004864 ext4_group_desc_csum_set(sb, block_group, gdp);
Aneesh Kumar K.V955ce5f2009-05-02 20:35:09 -04004865 ext4_unlock_group(sb, block_group);
Alex Tomasc9de5602008-01-29 00:19:52 -05004866
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004867 if (sbi->s_log_groups_per_flex) {
4868 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04004869 atomic64_add(count_clusters,
4870 &sbi->s_flex_groups[flex_group].free_clusters);
Jose R. Santos772cb7c2008-07-11 19:27:31 -04004871 }
4872
Theodore Ts'o71d4f7d2014-07-15 06:02:38 -04004873 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
Aditya Kali7b415bf2011-09-09 19:04:51 -04004874 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
Jan Kara7d734532013-08-17 09:36:54 -04004875 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4876
4877 ext4_mb_unload_buddy(&e4b);
Aditya Kali7b415bf2011-09-09 19:04:51 -04004878
Aneesh Kumar K.V7a2fcbf2009-01-05 21:36:55 -05004879 /* We dirtied the bitmap block */
4880 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4881 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4882
Alex Tomasc9de5602008-01-29 00:19:52 -05004883 /* And the group descriptor block */
4884 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
Frank Mayhar03901312009-01-07 00:06:22 -05004885 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
Alex Tomasc9de5602008-01-29 00:19:52 -05004886 if (!err)
4887 err = ret;
4888
4889 if (overflow && !err) {
4890 block += count;
4891 count = overflow;
4892 put_bh(bitmap_bh);
4893 goto do_more;
4894 }
Alex Tomasc9de5602008-01-29 00:19:52 -05004895error_return:
4896 brelse(bitmap_bh);
4897 ext4_std_error(sb, err);
Alex Tomasc9de5602008-01-29 00:19:52 -05004898 return;
4899}
Lukas Czerner7360d172010-10-27 21:30:12 -04004900
4901/**
Yongqiang Yang05291552011-07-26 21:43:56 -04004902 * ext4_group_add_blocks() -- Add given blocks to an existing group
Amir Goldstein2846e822011-05-09 10:46:41 -04004903 * @handle: handle to this transaction
4904 * @sb: super block
Anatol Pomozov4907cb72012-09-01 10:31:09 -07004905 * @block: start physical block to add to the block group
Amir Goldstein2846e822011-05-09 10:46:41 -04004906 * @count: number of blocks to free
4907 *
Amir Goldsteine73a3472011-05-09 21:40:01 -04004908 * This marks the blocks as free in the bitmap and buddy.
Amir Goldstein2846e822011-05-09 10:46:41 -04004909 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004910int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
Amir Goldstein2846e822011-05-09 10:46:41 -04004911 ext4_fsblk_t block, unsigned long count)
4912{
4913 struct buffer_head *bitmap_bh = NULL;
4914 struct buffer_head *gd_bh;
4915 ext4_group_t block_group;
4916 ext4_grpblk_t bit;
4917 unsigned int i;
4918 struct ext4_group_desc *desc;
4919 struct ext4_sb_info *sbi = EXT4_SB(sb);
Amir Goldsteine73a3472011-05-09 21:40:01 -04004920 struct ext4_buddy e4b;
Amir Goldstein2846e822011-05-09 10:46:41 -04004921 int err = 0, ret, blk_free_count;
4922 ext4_grpblk_t blocks_freed;
Amir Goldstein2846e822011-05-09 10:46:41 -04004923
4924 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4925
Yongqiang Yang4740b832011-07-26 21:51:08 -04004926 if (count == 0)
4927 return 0;
4928
Amir Goldstein2846e822011-05-09 10:46:41 -04004929 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
Amir Goldstein2846e822011-05-09 10:46:41 -04004930 /*
4931 * Check to see if we are freeing blocks across a group
4932 * boundary.
4933 */
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004934 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4935 ext4_warning(sb, "too much blocks added to group %u\n",
4936 block_group);
4937 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004938 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004939 }
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004940
Amir Goldstein2846e822011-05-09 10:46:41 -04004941 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
Darrick J. Wong9008a582015-10-17 21:33:24 -04004942 if (IS_ERR(bitmap_bh)) {
4943 err = PTR_ERR(bitmap_bh);
4944 bitmap_bh = NULL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004945 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004946 }
4947
Amir Goldstein2846e822011-05-09 10:46:41 -04004948 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004949 if (!desc) {
4950 err = -EIO;
Amir Goldstein2846e822011-05-09 10:46:41 -04004951 goto error_return;
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004952 }
Amir Goldstein2846e822011-05-09 10:46:41 -04004953
4954 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4955 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4956 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4957 in_range(block + count - 1, ext4_inode_table(sb, desc),
4958 sbi->s_itb_per_group)) {
4959 ext4_error(sb, "Adding blocks in system zones - "
4960 "Block = %llu, count = %lu",
4961 block, count);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04004962 err = -EINVAL;
Amir Goldstein2846e822011-05-09 10:46:41 -04004963 goto error_return;
4964 }
4965
Theodore Ts'o2cd05cc2011-05-09 10:58:45 -04004966 BUFFER_TRACE(bitmap_bh, "getting write access");
4967 err = ext4_journal_get_write_access(handle, bitmap_bh);
Amir Goldstein2846e822011-05-09 10:46:41 -04004968 if (err)
4969 goto error_return;
4970
4971 /*
4972 * We are about to modify some metadata. Call the journal APIs
4973 * to unshare ->b_data if a currently-committing transaction is
4974 * using it
4975 */
4976 BUFFER_TRACE(gd_bh, "get_write_access");
4977 err = ext4_journal_get_write_access(handle, gd_bh);
4978 if (err)
4979 goto error_return;
Amir Goldsteine73a3472011-05-09 21:40:01 -04004980
Amir Goldstein2846e822011-05-09 10:46:41 -04004981 for (i = 0, blocks_freed = 0; i < count; i++) {
4982 BUFFER_TRACE(bitmap_bh, "clear bit");
Amir Goldsteine73a3472011-05-09 21:40:01 -04004983 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
Amir Goldstein2846e822011-05-09 10:46:41 -04004984 ext4_error(sb, "bit already cleared for block %llu",
4985 (ext4_fsblk_t)(block + i));
4986 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4987 } else {
4988 blocks_freed++;
4989 }
4990 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04004991
4992 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4993 if (err)
4994 goto error_return;
4995
4996 /*
4997 * need to update group_info->bb_free and bitmap
4998 * with group lock held. generate_buddy look at
4999 * them with group lock_held
5000 */
Amir Goldstein2846e822011-05-09 10:46:41 -04005001 ext4_lock_group(sb, block_group);
Amir Goldsteine73a3472011-05-09 21:40:01 -04005002 mb_clear_bits(bitmap_bh->b_data, bit, count);
5003 mb_free_blocks(NULL, &e4b, bit, count);
Theodore Ts'o021b65b2011-09-09 19:08:51 -04005004 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5005 ext4_free_group_clusters_set(sb, desc, blk_free_count);
Tao Ma79f1ba42012-10-22 00:34:32 -04005006 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
Darrick J. Wongfeb0ab32012-04-29 18:45:10 -04005007 ext4_group_desc_csum_set(sb, block_group, desc);
Amir Goldstein2846e822011-05-09 10:46:41 -04005008 ext4_unlock_group(sb, block_group);
Theodore Ts'o57042652011-09-09 18:56:51 -04005009 percpu_counter_add(&sbi->s_freeclusters_counter,
Lukas Czerner810da242013-03-02 17:18:58 -05005010 EXT4_NUM_B2C(sbi, blocks_freed));
Amir Goldstein2846e822011-05-09 10:46:41 -04005011
5012 if (sbi->s_log_groups_per_flex) {
5013 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
Theodore Ts'o90ba9832013-03-11 23:39:59 -04005014 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5015 &sbi->s_flex_groups[flex_group].free_clusters);
Amir Goldstein2846e822011-05-09 10:46:41 -04005016 }
Amir Goldsteine73a3472011-05-09 21:40:01 -04005017
5018 ext4_mb_unload_buddy(&e4b);
Amir Goldstein2846e822011-05-09 10:46:41 -04005019
5020 /* We dirtied the bitmap block */
5021 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5022 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5023
5024 /* And the group descriptor block */
5025 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5026 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5027 if (!err)
5028 err = ret;
5029
5030error_return:
5031 brelse(bitmap_bh);
5032 ext4_std_error(sb, err);
Yongqiang Yangcc7365d2011-07-26 21:46:07 -04005033 return err;
Amir Goldstein2846e822011-05-09 10:46:41 -04005034}
5035
5036/**
Lukas Czerner7360d172010-10-27 21:30:12 -04005037 * ext4_trim_extent -- function to TRIM one single free extent in the group
5038 * @sb: super block for the file system
5039 * @start: starting block of the free extent in the alloc. group
5040 * @count: number of blocks to TRIM
5041 * @group: alloc. group we are working with
5042 * @e4b: ext4 buddy for the group
5043 *
5044 * Trim "count" blocks starting at "start" in the "group". To assure that no
5045 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5046 * be called with under the group lock.
5047 */
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005048static int ext4_trim_extent(struct super_block *sb, int start, int count,
Theodore Ts'od9f34502011-04-30 13:47:24 -04005049 ext4_group_t group, struct ext4_buddy *e4b)
jon ernste2cbd582014-04-12 23:01:28 -04005050__releases(bitlock)
5051__acquires(bitlock)
Lukas Czerner7360d172010-10-27 21:30:12 -04005052{
5053 struct ext4_free_extent ex;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005054 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005055
Tao Mab3d4c2b2011-07-11 00:01:52 -04005056 trace_ext4_trim_extent(sb, group, start, count);
5057
Lukas Czerner7360d172010-10-27 21:30:12 -04005058 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5059
5060 ex.fe_start = start;
5061 ex.fe_group = group;
5062 ex.fe_len = count;
5063
5064 /*
5065 * Mark blocks used, so no one can reuse them while
5066 * being trimmed.
5067 */
5068 mb_mark_used(e4b, &ex);
5069 ext4_unlock_group(sb, group);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005070 ret = ext4_issue_discard(sb, group, start, count);
Lukas Czerner7360d172010-10-27 21:30:12 -04005071 ext4_lock_group(sb, group);
5072 mb_free_blocks(NULL, e4b, start, ex.fe_len);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005073 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005074}
5075
5076/**
5077 * ext4_trim_all_free -- function to trim all free space in alloc. group
5078 * @sb: super block for file system
Tao Ma22612282011-07-11 00:04:34 -04005079 * @group: group to be trimmed
Lukas Czerner7360d172010-10-27 21:30:12 -04005080 * @start: first group block to examine
5081 * @max: last group block to examine
5082 * @minblocks: minimum extent block count
5083 *
5084 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5085 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5086 * the extent.
5087 *
5088 *
5089 * ext4_trim_all_free walks through group's block bitmap searching for free
5090 * extents. When the free extent is found, mark it as used in group buddy
5091 * bitmap. Then issue a TRIM command on this extent and free the extent in
5092 * the group buddy bitmap. This is done until whole group is scanned.
5093 */
Lukas Czerner0b75a842011-02-23 12:22:49 -05005094static ext4_grpblk_t
Lukas Czerner78944082011-05-24 18:16:27 -04005095ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5096 ext4_grpblk_t start, ext4_grpblk_t max,
5097 ext4_grpblk_t minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005098{
5099 void *bitmap;
Tao Ma169ddc32011-07-11 00:00:07 -04005100 ext4_grpblk_t next, count = 0, free_count = 0;
Lukas Czerner78944082011-05-24 18:16:27 -04005101 struct ext4_buddy e4b;
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005102 int ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005103
Tao Mab3d4c2b2011-07-11 00:01:52 -04005104 trace_ext4_trim_all_free(sb, group, start, max);
5105
Lukas Czerner78944082011-05-24 18:16:27 -04005106 ret = ext4_mb_load_buddy(sb, group, &e4b);
5107 if (ret) {
5108 ext4_error(sb, "Error in loading buddy "
5109 "information for %u", group);
5110 return ret;
5111 }
Lukas Czerner78944082011-05-24 18:16:27 -04005112 bitmap = e4b.bd_bitmap;
Lukas Czerner28739ee2011-05-24 18:28:07 -04005113
5114 ext4_lock_group(sb, group);
Tao Ma3d56b8d2011-07-11 00:03:38 -04005115 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5116 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5117 goto out;
5118
Lukas Czerner78944082011-05-24 18:16:27 -04005119 start = (e4b.bd_info->bb_first_free > start) ?
5120 e4b.bd_info->bb_first_free : start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005121
Lukas Czerner913eed832012-03-21 21:22:22 -04005122 while (start <= max) {
5123 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5124 if (start > max)
Lukas Czerner7360d172010-10-27 21:30:12 -04005125 break;
Lukas Czerner913eed832012-03-21 21:22:22 -04005126 next = mb_find_next_bit(bitmap, max + 1, start);
Lukas Czerner7360d172010-10-27 21:30:12 -04005127
5128 if ((next - start) >= minblocks) {
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005129 ret = ext4_trim_extent(sb, start,
5130 next - start, group, &e4b);
5131 if (ret && ret != -EOPNOTSUPP)
5132 break;
5133 ret = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005134 count += next - start;
5135 }
Tao Ma169ddc32011-07-11 00:00:07 -04005136 free_count += next - start;
Lukas Czerner7360d172010-10-27 21:30:12 -04005137 start = next + 1;
5138
5139 if (fatal_signal_pending(current)) {
5140 count = -ERESTARTSYS;
5141 break;
5142 }
5143
5144 if (need_resched()) {
5145 ext4_unlock_group(sb, group);
5146 cond_resched();
5147 ext4_lock_group(sb, group);
5148 }
5149
Tao Ma169ddc32011-07-11 00:00:07 -04005150 if ((e4b.bd_info->bb_free - free_count) < minblocks)
Lukas Czerner7360d172010-10-27 21:30:12 -04005151 break;
5152 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005153
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005154 if (!ret) {
5155 ret = count;
Tao Ma3d56b8d2011-07-11 00:03:38 -04005156 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005157 }
Tao Ma3d56b8d2011-07-11 00:03:38 -04005158out:
Lukas Czerner7360d172010-10-27 21:30:12 -04005159 ext4_unlock_group(sb, group);
Lukas Czerner78944082011-05-24 18:16:27 -04005160 ext4_mb_unload_buddy(&e4b);
Lukas Czerner7360d172010-10-27 21:30:12 -04005161
5162 ext4_debug("trimmed %d blocks in the group %d\n",
5163 count, group);
5164
Lukas Czernerd71c1ae2012-11-08 14:04:52 -05005165 return ret;
Lukas Czerner7360d172010-10-27 21:30:12 -04005166}
5167
5168/**
5169 * ext4_trim_fs() -- trim ioctl handle function
5170 * @sb: superblock for filesystem
5171 * @range: fstrim_range structure
5172 *
5173 * start: First Byte to trim
5174 * len: number of Bytes to trim from start
5175 * minlen: minimum extent length in Bytes
5176 * ext4_trim_fs goes through all allocation groups containing Bytes from
5177 * start to start+len. For each such a group ext4_trim_all_free function
5178 * is invoked to trim all free space.
5179 */
5180int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5181{
Lukas Czerner78944082011-05-24 18:16:27 -04005182 struct ext4_group_info *grp;
Lukas Czerner913eed832012-03-21 21:22:22 -04005183 ext4_group_t group, first_group, last_group;
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005184 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
Lukas Czerner913eed832012-03-21 21:22:22 -04005185 uint64_t start, end, minlen, trimmed = 0;
Jan Kara0f0a25b2011-01-11 15:16:31 -05005186 ext4_fsblk_t first_data_blk =
5187 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
Lukas Czerner913eed832012-03-21 21:22:22 -04005188 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
Lukas Czerner7360d172010-10-27 21:30:12 -04005189 int ret = 0;
5190
5191 start = range->start >> sb->s_blocksize_bits;
Lukas Czerner913eed832012-03-21 21:22:22 -04005192 end = start + (range->len >> sb->s_blocksize_bits) - 1;
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005193 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5194 range->minlen >> sb->s_blocksize_bits);
Lukas Czerner7360d172010-10-27 21:30:12 -04005195
Lukas Czerner5de35e82012-10-22 18:01:19 -04005196 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5197 start >= max_blks ||
5198 range->len < sb->s_blocksize)
Lukas Czerner7360d172010-10-27 21:30:12 -04005199 return -EINVAL;
Lukas Czerner913eed832012-03-21 21:22:22 -04005200 if (end >= max_blks)
5201 end = max_blks - 1;
5202 if (end <= first_data_blk)
Tao Ma22f10452011-07-10 23:52:37 -04005203 goto out;
Lukas Czerner913eed832012-03-21 21:22:22 -04005204 if (start < first_data_blk)
Jan Kara0f0a25b2011-01-11 15:16:31 -05005205 start = first_data_blk;
Lukas Czerner7360d172010-10-27 21:30:12 -04005206
Lukas Czerner913eed832012-03-21 21:22:22 -04005207 /* Determine first and last group to examine based on start and end */
Lukas Czerner7360d172010-10-27 21:30:12 -04005208 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005209 &first_group, &first_cluster);
Lukas Czerner913eed832012-03-21 21:22:22 -04005210 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005211 &last_group, &last_cluster);
Lukas Czerner7360d172010-10-27 21:30:12 -04005212
Lukas Czerner913eed832012-03-21 21:22:22 -04005213 /* end now represents the last cluster to discard in this group */
5214 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
Lukas Czerner7360d172010-10-27 21:30:12 -04005215
5216 for (group = first_group; group <= last_group; group++) {
Lukas Czerner78944082011-05-24 18:16:27 -04005217 grp = ext4_get_group_info(sb, group);
5218 /* We only do this if the grp has never been initialized */
5219 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5220 ret = ext4_mb_init_group(sb, group);
5221 if (ret)
5222 break;
Lukas Czerner7360d172010-10-27 21:30:12 -04005223 }
5224
Tao Ma0ba08512011-03-23 15:48:11 -04005225 /*
Lukas Czerner913eed832012-03-21 21:22:22 -04005226 * For all the groups except the last one, last cluster will
5227 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5228 * change it for the last group, note that last_cluster is
5229 * already computed earlier by ext4_get_group_no_and_offset()
Tao Ma0ba08512011-03-23 15:48:11 -04005230 */
Lukas Czerner913eed832012-03-21 21:22:22 -04005231 if (group == last_group)
5232 end = last_cluster;
Lukas Czerner7360d172010-10-27 21:30:12 -04005233
Lukas Czerner78944082011-05-24 18:16:27 -04005234 if (grp->bb_free >= minlen) {
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005235 cnt = ext4_trim_all_free(sb, group, first_cluster,
Lukas Czerner913eed832012-03-21 21:22:22 -04005236 end, minlen);
Lukas Czerner7360d172010-10-27 21:30:12 -04005237 if (cnt < 0) {
5238 ret = cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005239 break;
5240 }
Lukas Czerner21e7fd22012-03-21 21:24:22 -04005241 trimmed += cnt;
Lukas Czerner7360d172010-10-27 21:30:12 -04005242 }
Lukas Czerner913eed832012-03-21 21:22:22 -04005243
5244 /*
5245 * For every group except the first one, we are sure
5246 * that the first cluster to discard will be cluster #0.
5247 */
Theodore Ts'o7137d7a2011-09-09 18:38:51 -04005248 first_cluster = 0;
Lukas Czerner7360d172010-10-27 21:30:12 -04005249 }
Lukas Czerner7360d172010-10-27 21:30:12 -04005250
Tao Ma3d56b8d2011-07-11 00:03:38 -04005251 if (!ret)
5252 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5253
Tao Ma22f10452011-07-10 23:52:37 -04005254out:
Lukas Czerneraaf7d732012-09-26 22:21:21 -04005255 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
Lukas Czerner7360d172010-10-27 21:30:12 -04005256 return ret;
5257}