~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ext4/mballoc.c

Version: ~ [ linux-5.4-rc3 ] ~ [ linux-5.3.6 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.79 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.149 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.196 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.196 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.75 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  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 
 24 #include "ext4_jbd2.h"
 25 #include "mballoc.h"
 26 #include <linux/log2.h>
 27 #include <linux/module.h>
 28 #include <linux/slab.h>
 29 #include <linux/backing-dev.h>
 30 #include <trace/events/ext4.h>
 31 
 32 #ifdef CONFIG_EXT4_DEBUG
 33 ushort ext4_mballoc_debug __read_mostly;
 34 
 35 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
 36 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
 37 #endif
 38 
 39 /*
 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  *
 63  * 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.
 72  *
 73  * The main motivation for having small file use group preallocation is to
 74  * ensure that we have small files closer together on the disk.
 75  *
 76  * 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:
 80  *
 81  * pa_lstart -> the logical start block for this prealloc space
 82  * pa_pstart -> the physical start block for this prealloc space
 83  * pa_len    -> length for this prealloc space (in clusters)
 84  * pa_free   ->  free space available in this prealloc space (in clusters)
 85  *
 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
 88  * space we will consume the particular prealloc space. This makes sure that
 89  * we have contiguous physical blocks representing the file blocks
 90  *
 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
 97  * prealloc space. These are per CPU prealloc list represented as
 98  *
 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
105  * enough free space (pa_free) within the prealloc space.
106  *
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                        }
118  *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
119  *
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
139  * 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
141  * 512 blocks. This can be tuned via
142  * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
143  * 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
145  * the smallest multiple of the stripe value (sbi->s_stripe) which is
146  * greater than the default mb_group_prealloc.
147  *
148  * The regular allocator (using the buddy cache) supports a few tunables.
149  *
150  * /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
153  *
154  * The regular allocator uses buddy scan only if the request len is power of
155  * 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
157  * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
158  * stripe size (sbi->s_stripe), we try to search for contiguous block in
159  * 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.
162  * min_to_scan indicate how long the mballoc __must__ look for a best
163  * extent and max_to_scan indicates how long the mballoc __can__ look for a
164  * 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
167  * can be used for allocation. ext4_mb_good_group explains how the groups are
168  * 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  */
350 static struct kmem_cache *ext4_pspace_cachep;
351 static struct kmem_cache *ext4_ac_cachep;
352 static struct kmem_cache *ext4_free_data_cachep;
353 
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 */
357 #define NR_GRPINFO_CACHES 8
358 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359 
360 static 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 
366 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367                                         ext4_group_t group);
368 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369                                                 ext4_group_t group);
370 static void ext4_free_data_callback(struct super_block *sb,
371                                 struct ext4_journal_cb_entry *jce, int rc);
372 
373 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374 {
375 #if BITS_PER_LONG == 64
376         *bit += ((unsigned long) addr & 7UL) << 3;
377         addr = (void *) ((unsigned long) addr & ~7UL);
378 #elif BITS_PER_LONG == 32
379         *bit += ((unsigned long) addr & 3UL) << 3;
380         addr = (void *) ((unsigned long) addr & ~3UL);
381 #else
382 #error "how many bits you are?!"
383 #endif
384         return addr;
385 }
386 
387 static 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          */
393         addr = mb_correct_addr_and_bit(&bit, addr);
394         return ext4_test_bit(bit, addr);
395 }
396 
397 static inline void mb_set_bit(int bit, void *addr)
398 {
399         addr = mb_correct_addr_and_bit(&bit, addr);
400         ext4_set_bit(bit, addr);
401 }
402 
403 static inline void mb_clear_bit(int bit, void *addr)
404 {
405         addr = mb_correct_addr_and_bit(&bit, addr);
406         ext4_clear_bit(bit, addr);
407 }
408 
409 static 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 
415 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416 {
417         int fix = 0, ret, tmpmax;
418         addr = mb_correct_addr_and_bit(&fix, addr);
419         tmpmax = max + fix;
420         start += fix;
421 
422         ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
423         if (ret > max)
424                 return max;
425         return ret;
426 }
427 
428 static inline int mb_find_next_bit(void *addr, int max, int start)
429 {
430         int fix = 0, ret, tmpmax;
431         addr = mb_correct_addr_and_bit(&fix, addr);
432         tmpmax = max + fix;
433         start += fix;
434 
435         ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
436         if (ret > max)
437                 return max;
438         return ret;
439 }
440 
441 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
442 {
443         char *bb;
444 
445         BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
446         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 */
454         if (order == 0) {
455                 *max = 1 << (e4b->bd_blkbits + 3);
456                 return e4b->bd_bitmap;
457         }
458 
459         bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
460         *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
461 
462         return bb;
463 }
464 
465 #ifdef DOUBLE_CHECK
466 static 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;
474         assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
475         for (i = 0; i < count; i++) {
476                 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
477                         ext4_fsblk_t blocknr;
478 
479                         blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
480                         blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
481                         ext4_grp_locked_error(sb, e4b->bd_group,
482                                               inode ? inode->i_ino : 0,
483                                               blocknr,
484                                               "freeing block already freed "
485                                               "(bit %u)",
486                                               first + i);
487                 }
488                 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
489         }
490 }
491 
492 static 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;
498         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
499         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 
505 static 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]) {
514                                 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]);
519                                 BUG();
520                         }
521                 }
522         }
523 }
524 
525 #else
526 static inline void mb_free_blocks_double(struct inode *inode,
527                                 struct ext4_buddy *e4b, int first, int count)
528 {
529         return;
530 }
531 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
532                                                 int first, int count)
533 {
534         return;
535 }
536 static 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)                                         \
545 do {                                                                    \
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 
554 static 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 
572         {
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 
601                         /* both bits in buddy2 must be 1 */
602                         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(
608                                         !mb_test_bit(k, e4b->bd_bitmap));
609                         }
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);
640         list_for_each(cur, &grp->bb_prealloc_list) {
641                 ext4_group_t groupnr;
642                 struct ext4_prealloc_space *pa;
643                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
644                 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
645                 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
646                 for (i = 0; i < pa->pa_len; i++)
647                         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,       \
653                                         __FILE__, __func__, __LINE__)
654 #else
655 #define mb_check_buddy(e4b)
656 #endif
657 
658 /*
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  */
664 static void ext4_mb_mark_free_simple(struct super_block *sb,
665                                 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
666                                         struct ext4_group_info *grp)
667 {
668         struct ext4_sb_info *sbi = EXT4_SB(sb);
669         ext4_grpblk_t min;
670         ext4_grpblk_t max;
671         ext4_grpblk_t chunk;
672         unsigned short border;
673 
674         BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
675 
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 
700 /*
701  * Cache the order of the largest free extent we have available in this block
702  * group.
703  */
704 static void
705 mb_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 
721 static noinline_for_stack
722 void ext4_mb_generate_buddy(struct super_block *sb,
723                                 void *buddy, void *bitmap, ext4_group_t group)
724 {
725         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
726         struct ext4_sb_info *sbi = EXT4_SB(sb);
727         ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
728         ext4_grpblk_t i = 0;
729         ext4_grpblk_t first;
730         ext4_grpblk_t len;
731         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 */
737         i = mb_find_next_zero_bit(bitmap, max, 0);
738         grp->bb_first_free = i;
739         while (i < max) {
740                 fragments++;
741                 first = i;
742                 i = mb_find_next_bit(bitmap, max, i);
743                 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)
750                         i = mb_find_next_zero_bit(bitmap, max, i);
751         }
752         grp->bb_fragments = fragments;
753 
754         if (free != grp->bb_free) {
755                 ext4_grp_locked_error(sb, group, 0, 0,
756                                       "block bitmap and bg descriptor "
757                                       "inconsistent: %u vs %u free clusters",
758                                       free, grp->bb_free);
759                 /*
760                  * If we intend to continue, we consider group descriptor
761                  * corrupt and update bb_free using bitmap value
762                  */
763                 grp->bb_free = free;
764                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
765                         percpu_counter_sub(&sbi->s_freeclusters_counter,
766                                            grp->bb_free);
767                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
768         }
769         mb_set_largest_free_order(sb, grp);
770 
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 
780 static 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 
798 /* 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                        }
805  * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
806  *
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
813  *
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!
816  */
817 
818 static int ext4_mb_init_cache(struct page *page, char *incore)
819 {
820         ext4_group_t ngroups;
821         int blocksize;
822         int blocks_per_page;
823         int groups_per_page;
824         int err = 0;
825         int i;
826         ext4_group_t first_group, group;
827         int first_block;
828         struct super_block *sb;
829         struct buffer_head *bhs;
830         struct buffer_head **bh = NULL;
831         struct inode *inode;
832         char *data;
833         char *bitmap;
834         struct ext4_group_info *grinfo;
835 
836         mb_debug(1, "init page %lu\n", page->index);
837 
838         inode = page->mapping->host;
839         sb = inode->i_sb;
840         ngroups = ext4_get_groups_count(sb);
841         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) {
850                 i = sizeof(struct buffer_head *) * groups_per_page;
851                 bh = kzalloc(i, GFP_NOFS);
852                 if (bh == NULL) {
853                         err = -ENOMEM;
854                         goto out;
855                 }
856         } 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 */
862         for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
863                 if (group >= ngroups)
864                         break;
865 
866                 grinfo = ext4_get_group_info(sb, group);
867                 /*
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                 }
877                 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
878                         err = -ENOMEM;
879                         goto out;
880                 }
881                 mb_debug(1, "read bitmap for group %u\n", group);
882         }
883 
884         /* wait for I/O completion */
885         for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
886                 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i]))
887                         err = -EIO;
888         }
889 
890         first_block = page->index * blocks_per_page;
891         for (i = 0; i < blocks_per_page; i++) {
892                 group = (first_block + i) >> 1;
893                 if (group >= ngroups)
894                         break;
895 
896                 if (!bh[group - first_group])
897                         /* skip initialized uptodate buddy */
898                         continue;
899 
900                 if (!buffer_verified(bh[group - first_group]))
901                         /* Skip faulty bitmaps */
902                         continue;
903                 err = 0;
904 
905                 /*
906                  * data carry information regarding this
907                  * particular group in the format specified
908                  * above
909                  *
910                  */
911                 data = page_address(page) + (i * blocksize);
912                 bitmap = bh[group - first_group]->b_data;
913 
914                 /*
915                  * We place the buddy block and bitmap block
916                  * close together
917                  */
918                 if ((first_block + i) & 1) {
919                         /* this is block of buddy */
920                         BUG_ON(incore == NULL);
921                         mb_debug(1, "put buddy for group %u in page %lu/%x\n",
922                                 group, page->index, i * blocksize);
923                         trace_ext4_mb_buddy_bitmap_load(sb, group);
924                         grinfo = ext4_get_group_info(sb, group);
925                         grinfo->bb_fragments = 0;
926                         memset(grinfo->bb_counters, 0,
927                                sizeof(*grinfo->bb_counters) *
928                                 (sb->s_blocksize_bits+2));
929                         /*
930                          * incore got set to the group block bitmap below
931                          */
932                         ext4_lock_group(sb, group);
933                         /* init the buddy */
934                         memset(data, 0xff, blocksize);
935                         ext4_mb_generate_buddy(sb, data, incore, group);
936                         ext4_unlock_group(sb, group);
937                         incore = NULL;
938                 } else {
939                         /* this is block of bitmap */
940                         BUG_ON(incore != NULL);
941                         mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
942                                 group, page->index, i * blocksize);
943                         trace_ext4_mb_bitmap_load(sb, group);
944 
945                         /* see comments in ext4_mb_put_pa() */
946                         ext4_lock_group(sb, group);
947                         memcpy(data, bitmap, blocksize);
948 
949                         /* mark all preallocated blks used in in-core bitmap */
950                         ext4_mb_generate_from_pa(sb, data, group);
951                         ext4_mb_generate_from_freelist(sb, data, group);
952                         ext4_unlock_group(sb, group);
953 
954                         /* set incore so that the buddy information can be
955                          * generated using this
956                          */
957                         incore = data;
958                 }
959         }
960         SetPageUptodate(page);
961 
962 out:
963         if (bh) {
964                 for (i = 0; i < groups_per_page; i++)
965                         brelse(bh[i]);
966                 if (bh != &bhs)
967                         kfree(bh);
968         }
969         return err;
970 }
971 
972 /*
973  * Lock the buddy and bitmap pages. This make sure other parallel init_group
974  * on the same buddy page doesn't happen whild holding the buddy page lock.
975  * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
976  * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
977  */
978 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
979                 ext4_group_t group, struct ext4_buddy *e4b)
980 {
981         struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
982         int block, pnum, poff;
983         int blocks_per_page;
984         struct page *page;
985 
986         e4b->bd_buddy_page = NULL;
987         e4b->bd_bitmap_page = NULL;
988 
989         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
990         /*
991          * the buddy cache inode stores the block bitmap
992          * and buddy information in consecutive blocks.
993          * So for each group we need two blocks.
994          */
995         block = group * 2;
996         pnum = block / blocks_per_page;
997         poff = block % blocks_per_page;
998         page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
999         if (!page)
1000                 return -ENOMEM;
1001         BUG_ON(page->mapping != inode->i_mapping);
1002         e4b->bd_bitmap_page = page;
1003         e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1004 
1005         if (blocks_per_page >= 2) {
1006                 /* buddy and bitmap are on the same page */
1007                 return 0;
1008         }
1009 
1010         block++;
1011         pnum = block / blocks_per_page;
1012         page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1013         if (!page)
1014                 return -ENOMEM;
1015         BUG_ON(page->mapping != inode->i_mapping);
1016         e4b->bd_buddy_page = page;
1017         return 0;
1018 }
1019 
1020 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1021 {
1022         if (e4b->bd_bitmap_page) {
1023                 unlock_page(e4b->bd_bitmap_page);
1024                 page_cache_release(e4b->bd_bitmap_page);
1025         }
1026         if (e4b->bd_buddy_page) {
1027                 unlock_page(e4b->bd_buddy_page);
1028                 page_cache_release(e4b->bd_buddy_page);
1029         }
1030 }
1031 
1032 /*
1033  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1034  * block group lock of all groups for this page; do not hold the BG lock when
1035  * calling this routine!
1036  */
1037 static noinline_for_stack
1038 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1039 {
1040 
1041         struct ext4_group_info *this_grp;
1042         struct ext4_buddy e4b;
1043         struct page *page;
1044         int ret = 0;
1045 
1046         might_sleep();
1047         mb_debug(1, "init group %u\n", group);
1048         this_grp = ext4_get_group_info(sb, group);
1049         /*
1050          * This ensures that we don't reinit the buddy cache
1051          * page which map to the group from which we are already
1052          * allocating. If we are looking at the buddy cache we would
1053          * have taken a reference using ext4_mb_load_buddy and that
1054          * would have pinned buddy page to page cache.
1055          * The call to ext4_mb_get_buddy_page_lock will mark the
1056          * page accessed.
1057          */
1058         ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1059         if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1060                 /*
1061                  * somebody initialized the group
1062                  * return without doing anything
1063                  */
1064                 goto err;
1065         }
1066 
1067         page = e4b.bd_bitmap_page;
1068         ret = ext4_mb_init_cache(page, NULL);
1069         if (ret)
1070                 goto err;
1071         if (!PageUptodate(page)) {
1072                 ret = -EIO;
1073                 goto err;
1074         }
1075 
1076         if (e4b.bd_buddy_page == NULL) {
1077                 /*
1078                  * If both the bitmap and buddy are in
1079                  * the same page we don't need to force
1080                  * init the buddy
1081                  */
1082                 ret = 0;
1083                 goto err;
1084         }
1085         /* init buddy cache */
1086         page = e4b.bd_buddy_page;
1087         ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1088         if (ret)
1089                 goto err;
1090         if (!PageUptodate(page)) {
1091                 ret = -EIO;
1092                 goto err;
1093         }
1094 err:
1095         ext4_mb_put_buddy_page_lock(&e4b);
1096         return ret;
1097 }
1098 
1099 /*
1100  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1101  * block group lock of all groups for this page; do not hold the BG lock when
1102  * calling this routine!
1103  */
1104 static noinline_for_stack int
1105 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1106                                         struct ext4_buddy *e4b)
1107 {
1108         int blocks_per_page;
1109         int block;
1110         int pnum;
1111         int poff;
1112         struct page *page;
1113         int ret;
1114         struct ext4_group_info *grp;
1115         struct ext4_sb_info *sbi = EXT4_SB(sb);
1116         struct inode *inode = sbi->s_buddy_cache;
1117 
1118         might_sleep();
1119         mb_debug(1, "load group %u\n", group);
1120 
1121         blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1122         grp = ext4_get_group_info(sb, group);
1123 
1124         e4b->bd_blkbits = sb->s_blocksize_bits;
1125         e4b->bd_info = grp;
1126         e4b->bd_sb = sb;
1127         e4b->bd_group = group;
1128         e4b->bd_buddy_page = NULL;
1129         e4b->bd_bitmap_page = NULL;
1130 
1131         if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1132                 /*
1133                  * we need full data about the group
1134                  * to make a good selection
1135                  */
1136                 ret = ext4_mb_init_group(sb, group);
1137                 if (ret)
1138                         return ret;
1139         }
1140 
1141         /*
1142          * the buddy cache inode stores the block bitmap
1143          * and buddy information in consecutive blocks.
1144          * So for each group we need two blocks.
1145          */
1146         block = group * 2;
1147         pnum = block / blocks_per_page;
1148         poff = block % blocks_per_page;
1149 
1150         /* we could use find_or_create_page(), but it locks page
1151          * what we'd like to avoid in fast path ... */
1152         page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1153         if (page == NULL || !PageUptodate(page)) {
1154                 if (page)
1155                         /*
1156                          * drop the page reference and try
1157                          * to get the page with lock. If we
1158                          * are not uptodate that implies
1159                          * somebody just created the page but
1160                          * is yet to initialize the same. So
1161                          * wait for it to initialize.
1162                          */
1163                         page_cache_release(page);
1164                 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1165                 if (page) {
1166                         BUG_ON(page->mapping != inode->i_mapping);
1167                         if (!PageUptodate(page)) {
1168                                 ret = ext4_mb_init_cache(page, NULL);
1169                                 if (ret) {
1170                                         unlock_page(page);
1171                                         goto err;
1172                                 }
1173                                 mb_cmp_bitmaps(e4b, page_address(page) +
1174                                                (poff * sb->s_blocksize));
1175                         }
1176                         unlock_page(page);
1177                 }
1178         }
1179         if (page == NULL) {
1180                 ret = -ENOMEM;
1181                 goto err;
1182         }
1183         if (!PageUptodate(page)) {
1184                 ret = -EIO;
1185                 goto err;
1186         }
1187 
1188         /* Pages marked accessed already */
1189         e4b->bd_bitmap_page = page;
1190         e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1191 
1192         block++;
1193         pnum = block / blocks_per_page;
1194         poff = block % blocks_per_page;
1195 
1196         page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1197         if (page == NULL || !PageUptodate(page)) {
1198                 if (page)
1199                         page_cache_release(page);
1200                 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1201                 if (page) {
1202                         BUG_ON(page->mapping != inode->i_mapping);
1203                         if (!PageUptodate(page)) {
1204                                 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1205                                 if (ret) {
1206                                         unlock_page(page);
1207                                         goto err;
1208                                 }
1209                         }
1210                         unlock_page(page);
1211                 }
1212         }
1213         if (page == NULL) {
1214                 ret = -ENOMEM;
1215                 goto err;
1216         }
1217         if (!PageUptodate(page)) {
1218                 ret = -EIO;
1219                 goto err;
1220         }
1221 
1222         /* Pages marked accessed already */
1223         e4b->bd_buddy_page = page;
1224         e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1225 
1226         BUG_ON(e4b->bd_bitmap_page == NULL);
1227         BUG_ON(e4b->bd_buddy_page == NULL);
1228 
1229         return 0;
1230 
1231 err:
1232         if (page)
1233                 page_cache_release(page);
1234         if (e4b->bd_bitmap_page)
1235                 page_cache_release(e4b->bd_bitmap_page);
1236         if (e4b->bd_buddy_page)
1237                 page_cache_release(e4b->bd_buddy_page);
1238         e4b->bd_buddy = NULL;
1239         e4b->bd_bitmap = NULL;
1240         return ret;
1241 }
1242 
1243 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1244 {
1245         if (e4b->bd_bitmap_page)
1246                 page_cache_release(e4b->bd_bitmap_page);
1247         if (e4b->bd_buddy_page)
1248                 page_cache_release(e4b->bd_buddy_page);
1249 }
1250 
1251 
1252 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1253 {
1254         int order = 1;
1255         void *bb;
1256 
1257         BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1258         BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1259 
1260         bb = e4b->bd_buddy;
1261         while (order <= e4b->bd_blkbits + 1) {
1262                 block = block >> 1;
1263                 if (!mb_test_bit(block, bb)) {
1264                         /* this block is part of buddy of order 'order' */
1265                         return order;
1266                 }
1267                 bb += 1 << (e4b->bd_blkbits - order);
1268                 order++;
1269         }
1270         return 0;
1271 }
1272 
1273 static void mb_clear_bits(void *bm, int cur, int len)
1274 {
1275         __u32 *addr;
1276 
1277         len = cur + len;
1278         while (cur < len) {
1279                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1280                         /* fast path: clear whole word at once */
1281                         addr = bm + (cur >> 3);
1282                         *addr = 0;
1283                         cur += 32;
1284                         continue;
1285                 }
1286                 mb_clear_bit(cur, bm);
1287                 cur++;
1288         }
1289 }
1290 
1291 /* clear bits in given range
1292  * will return first found zero bit if any, -1 otherwise
1293  */
1294 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1295 {
1296         __u32 *addr;
1297         int zero_bit = -1;
1298 
1299         len = cur + len;
1300         while (cur < len) {
1301                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1302                         /* fast path: clear whole word at once */
1303                         addr = bm + (cur >> 3);
1304                         if (*addr != (__u32)(-1) && zero_bit == -1)
1305                                 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1306                         *addr = 0;
1307                         cur += 32;
1308                         continue;
1309                 }
1310                 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1311                         zero_bit = cur;
1312                 cur++;
1313         }
1314 
1315         return zero_bit;
1316 }
1317 
1318 void ext4_set_bits(void *bm, int cur, int len)
1319 {
1320         __u32 *addr;
1321 
1322         len = cur + len;
1323         while (cur < len) {
1324                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1325                         /* fast path: set whole word at once */
1326                         addr = bm + (cur >> 3);
1327                         *addr = 0xffffffff;
1328                         cur += 32;
1329                         continue;
1330                 }
1331                 mb_set_bit(cur, bm);
1332                 cur++;
1333         }
1334 }
1335 
1336 /*
1337  * _________________________________________________________________ */
1338 
1339 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1340 {
1341         if (mb_test_bit(*bit + side, bitmap)) {
1342                 mb_clear_bit(*bit, bitmap);
1343                 (*bit) -= side;
1344                 return 1;
1345         }
1346         else {
1347                 (*bit) += side;
1348                 mb_set_bit(*bit, bitmap);
1349                 return -1;
1350         }
1351 }
1352 
1353 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1354 {
1355         int max;
1356         int order = 1;
1357         void *buddy = mb_find_buddy(e4b, order, &max);
1358 
1359         while (buddy) {
1360                 void *buddy2;
1361 
1362                 /* Bits in range [first; last] are known to be set since
1363                  * corresponding blocks were allocated. Bits in range
1364                  * (first; last) will stay set because they form buddies on
1365                  * upper layer. We just deal with borders if they don't
1366                  * align with upper layer and then go up.
1367                  * Releasing entire group is all about clearing
1368                  * single bit of highest order buddy.
1369                  */
1370 
1371                 /* Example:
1372                  * ---------------------------------
1373                  * |   1   |   1   |   1   |   1   |
1374                  * ---------------------------------
1375                  * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1376                  * ---------------------------------
1377                  *   0   1   2   3   4   5   6   7
1378                  *      \_____________________/
1379                  *
1380                  * Neither [1] nor [6] is aligned to above layer.
1381                  * Left neighbour [0] is free, so mark it busy,
1382                  * decrease bb_counters and extend range to
1383                  * [0; 6]
1384                  * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1385                  * mark [6] free, increase bb_counters and shrink range to
1386                  * [0; 5].
1387                  * Then shift range to [0; 2], go up and do the same.
1388                  */
1389 
1390 
1391                 if (first & 1)
1392                         e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1393                 if (!(last & 1))
1394                         e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1395                 if (first > last)
1396                         break;
1397                 order++;
1398 
1399                 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1400                         mb_clear_bits(buddy, first, last - first + 1);
1401                         e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1402                         break;
1403                 }
1404                 first >>= 1;
1405                 last >>= 1;
1406                 buddy = buddy2;
1407         }
1408 }
1409 
1410 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1411                            int first, int count)
1412 {
1413         int left_is_free = 0;
1414         int right_is_free = 0;
1415         int block;
1416         int last = first + count - 1;
1417         struct super_block *sb = e4b->bd_sb;
1418 
1419         if (WARN_ON(count == 0))
1420                 return;
1421         BUG_ON(last >= (sb->s_blocksize << 3));
1422         assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1423         /* Don't bother if the block group is corrupt. */
1424         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1425                 return;
1426 
1427         mb_check_buddy(e4b);
1428         mb_free_blocks_double(inode, e4b, first, count);
1429 
1430         e4b->bd_info->bb_free += count;
1431         if (first < e4b->bd_info->bb_first_free)
1432                 e4b->bd_info->bb_first_free = first;
1433 
1434         /* access memory sequentially: check left neighbour,
1435          * clear range and then check right neighbour
1436          */
1437         if (first != 0)
1438                 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1439         block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1440         if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1441                 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1442 
1443         if (unlikely(block != -1)) {
1444                 struct ext4_sb_info *sbi = EXT4_SB(sb);
1445                 ext4_fsblk_t blocknr;
1446 
1447                 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1448                 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1449                 ext4_grp_locked_error(sb, e4b->bd_group,
1450                                       inode ? inode->i_ino : 0,
1451                                       blocknr,
1452                                       "freeing already freed block "
1453                                       "(bit %u); block bitmap corrupt.",
1454                                       block);
1455                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1456                         percpu_counter_sub(&sbi->s_freeclusters_counter,
1457                                            e4b->bd_info->bb_free);
1458                 /* Mark the block group as corrupt. */
1459                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1460                         &e4b->bd_info->bb_state);
1461                 mb_regenerate_buddy(e4b);
1462                 goto done;
1463         }
1464 
1465         /* let's maintain fragments counter */
1466         if (left_is_free && right_is_free)
1467                 e4b->bd_info->bb_fragments--;
1468         else if (!left_is_free && !right_is_free)
1469                 e4b->bd_info->bb_fragments++;
1470 
1471         /* buddy[0] == bd_bitmap is a special case, so handle
1472          * it right away and let mb_buddy_mark_free stay free of
1473          * zero order checks.
1474          * Check if neighbours are to be coaleasced,
1475          * adjust bitmap bb_counters and borders appropriately.
1476          */
1477         if (first & 1) {
1478                 first += !left_is_free;
1479                 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1480         }
1481         if (!(last & 1)) {
1482                 last -= !right_is_free;
1483                 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1484         }
1485 
1486         if (first <= last)
1487                 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1488 
1489 done:
1490         mb_set_largest_free_order(sb, e4b->bd_info);
1491         mb_check_buddy(e4b);
1492 }
1493 
1494 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1495                                 int needed, struct ext4_free_extent *ex)
1496 {
1497         int next = block;
1498         int max, order;
1499         void *buddy;
1500 
1501         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1502         BUG_ON(ex == NULL);
1503 
1504         buddy = mb_find_buddy(e4b, 0, &max);
1505         BUG_ON(buddy == NULL);
1506         BUG_ON(block >= max);
1507         if (mb_test_bit(block, buddy)) {
1508                 ex->fe_len = 0;
1509                 ex->fe_start = 0;
1510                 ex->fe_group = 0;
1511                 return 0;
1512         }
1513 
1514         /* find actual order */
1515         order = mb_find_order_for_block(e4b, block);
1516         block = block >> order;
1517 
1518         ex->fe_len = 1 << order;
1519         ex->fe_start = block << order;
1520         ex->fe_group = e4b->bd_group;
1521 
1522         /* calc difference from given start */
1523         next = next - ex->fe_start;
1524         ex->fe_len -= next;
1525         ex->fe_start += next;
1526 
1527         while (needed > ex->fe_len &&
1528                mb_find_buddy(e4b, order, &max)) {
1529 
1530                 if (block + 1 >= max)
1531                         break;
1532 
1533                 next = (block + 1) * (1 << order);
1534                 if (mb_test_bit(next, e4b->bd_bitmap))
1535                         break;
1536 
1537                 order = mb_find_order_for_block(e4b, next);
1538 
1539                 block = next >> order;
1540                 ex->fe_len += 1 << order;
1541         }
1542 
1543         BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1544         return ex->fe_len;
1545 }
1546 
1547 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1548 {
1549         int ord;
1550         int mlen = 0;
1551         int max = 0;
1552         int cur;
1553         int start = ex->fe_start;
1554         int len = ex->fe_len;
1555         unsigned ret = 0;
1556         int len0 = len;
1557         void *buddy;
1558 
1559         BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1560         BUG_ON(e4b->bd_group != ex->fe_group);
1561         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1562         mb_check_buddy(e4b);
1563         mb_mark_used_double(e4b, start, len);
1564 
1565         e4b->bd_info->bb_free -= len;
1566         if (e4b->bd_info->bb_first_free == start)
1567                 e4b->bd_info->bb_first_free += len;
1568 
1569         /* let's maintain fragments counter */
1570         if (start != 0)
1571                 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1572         if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1573                 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1574         if (mlen && max)
1575                 e4b->bd_info->bb_fragments++;
1576         else if (!mlen && !max)
1577                 e4b->bd_info->bb_fragments--;
1578 
1579         /* let's maintain buddy itself */
1580         while (len) {
1581                 ord = mb_find_order_for_block(e4b, start);
1582 
1583                 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1584                         /* the whole chunk may be allocated at once! */
1585                         mlen = 1 << ord;
1586                         buddy = mb_find_buddy(e4b, ord, &max);
1587                         BUG_ON((start >> ord) >= max);
1588                         mb_set_bit(start >> ord, buddy);
1589                         e4b->bd_info->bb_counters[ord]--;
1590                         start += mlen;
1591                         len -= mlen;
1592                         BUG_ON(len < 0);
1593                         continue;
1594                 }
1595 
1596                 /* store for history */
1597                 if (ret == 0)
1598                         ret = len | (ord << 16);
1599 
1600                 /* we have to split large buddy */
1601                 BUG_ON(ord <= 0);
1602                 buddy = mb_find_buddy(e4b, ord, &max);
1603                 mb_set_bit(start >> ord, buddy);
1604                 e4b->bd_info->bb_counters[ord]--;
1605 
1606                 ord--;
1607                 cur = (start >> ord) & ~1U;
1608                 buddy = mb_find_buddy(e4b, ord, &max);
1609                 mb_clear_bit(cur, buddy);
1610                 mb_clear_bit(cur + 1, buddy);
1611                 e4b->bd_info->bb_counters[ord]++;
1612                 e4b->bd_info->bb_counters[ord]++;
1613         }
1614         mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1615 
1616         ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1617         mb_check_buddy(e4b);
1618 
1619         return ret;
1620 }
1621 
1622 /*
1623  * Must be called under group lock!
1624  */
1625 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1626                                         struct ext4_buddy *e4b)
1627 {
1628         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1629         int ret;
1630 
1631         BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1632         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1633 
1634         ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1635         ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1636         ret = mb_mark_used(e4b, &ac->ac_b_ex);
1637 
1638         /* preallocation can change ac_b_ex, thus we store actually
1639          * allocated blocks for history */
1640         ac->ac_f_ex = ac->ac_b_ex;
1641 
1642         ac->ac_status = AC_STATUS_FOUND;
1643         ac->ac_tail = ret & 0xffff;
1644         ac->ac_buddy = ret >> 16;
1645 
1646         /*
1647          * take the page reference. We want the page to be pinned
1648          * so that we don't get a ext4_mb_init_cache_call for this
1649          * group until we update the bitmap. That would mean we
1650          * double allocate blocks. The reference is dropped
1651          * in ext4_mb_release_context
1652          */
1653         ac->ac_bitmap_page = e4b->bd_bitmap_page;
1654         get_page(ac->ac_bitmap_page);
1655         ac->ac_buddy_page = e4b->bd_buddy_page;
1656         get_page(ac->ac_buddy_page);
1657         /* store last allocated for subsequent stream allocation */
1658         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1659                 spin_lock(&sbi->s_md_lock);
1660                 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1661                 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1662                 spin_unlock(&sbi->s_md_lock);
1663         }
1664 }
1665 
1666 /*
1667  * regular allocator, for general purposes allocation
1668  */
1669 
1670 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1671                                         struct ext4_buddy *e4b,
1672                                         int finish_group)
1673 {
1674         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1675         struct ext4_free_extent *bex = &ac->ac_b_ex;
1676         struct ext4_free_extent *gex = &ac->ac_g_ex;
1677         struct ext4_free_extent ex;
1678         int max;
1679 
1680         if (ac->ac_status == AC_STATUS_FOUND)
1681                 return;
1682         /*
1683          * We don't want to scan for a whole year
1684          */
1685         if (ac->ac_found > sbi->s_mb_max_to_scan &&
1686                         !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1687                 ac->ac_status = AC_STATUS_BREAK;
1688                 return;
1689         }
1690 
1691         /*
1692          * Haven't found good chunk so far, let's continue
1693          */
1694         if (bex->fe_len < gex->fe_len)
1695                 return;
1696 
1697         if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1698                         && bex->fe_group == e4b->bd_group) {
1699                 /* recheck chunk's availability - we don't know
1700                  * when it was found (within this lock-unlock
1701                  * period or not) */
1702                 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1703                 if (max >= gex->fe_len) {
1704                         ext4_mb_use_best_found(ac, e4b);
1705                         return;
1706                 }
1707         }
1708 }
1709 
1710 /*
1711  * The routine checks whether found extent is good enough. If it is,
1712  * then the extent gets marked used and flag is set to the context
1713  * to stop scanning. Otherwise, the extent is compared with the
1714  * previous found extent and if new one is better, then it's stored
1715  * in the context. Later, the best found extent will be used, if
1716  * mballoc can't find good enough extent.
1717  *
1718  * FIXME: real allocation policy is to be designed yet!
1719  */
1720 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1721                                         struct ext4_free_extent *ex,
1722                                         struct ext4_buddy *e4b)
1723 {
1724         struct ext4_free_extent *bex = &ac->ac_b_ex;
1725         struct ext4_free_extent *gex = &ac->ac_g_ex;
1726 
1727         BUG_ON(ex->fe_len <= 0);
1728         BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1729         BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1730         BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1731 
1732         ac->ac_found++;
1733 
1734         /*
1735          * The special case - take what you catch first
1736          */
1737         if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1738                 *bex = *ex;
1739                 ext4_mb_use_best_found(ac, e4b);
1740                 return;
1741         }
1742 
1743         /*
1744          * Let's check whether the chuck is good enough
1745          */
1746         if (ex->fe_len == gex->fe_len) {
1747                 *bex = *ex;
1748                 ext4_mb_use_best_found(ac, e4b);
1749                 return;
1750         }
1751 
1752         /*
1753          * If this is first found extent, just store it in the context
1754          */
1755         if (bex->fe_len == 0) {
1756                 *bex = *ex;
1757                 return;
1758         }
1759 
1760         /*
1761          * If new found extent is better, store it in the context
1762          */
1763         if (bex->fe_len < gex->fe_len) {
1764                 /* if the request isn't satisfied, any found extent
1765                  * larger than previous best one is better */
1766                 if (ex->fe_len > bex->fe_len)
1767                         *bex = *ex;
1768         } else if (ex->fe_len > gex->fe_len) {
1769                 /* if the request is satisfied, then we try to find
1770                  * an extent that still satisfy the request, but is
1771                  * smaller than previous one */
1772                 if (ex->fe_len < bex->fe_len)
1773                         *bex = *ex;
1774         }
1775 
1776         ext4_mb_check_limits(ac, e4b, 0);
1777 }
1778 
1779 static noinline_for_stack
1780 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1781                                         struct ext4_buddy *e4b)
1782 {
1783         struct ext4_free_extent ex = ac->ac_b_ex;
1784         ext4_group_t group = ex.fe_group;
1785         int max;
1786         int err;
1787 
1788         BUG_ON(ex.fe_len <= 0);
1789         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1790         if (err)
1791                 return err;
1792 
1793         ext4_lock_group(ac->ac_sb, group);
1794         max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1795 
1796         if (max > 0) {
1797                 ac->ac_b_ex = ex;
1798                 ext4_mb_use_best_found(ac, e4b);
1799         }
1800 
1801         ext4_unlock_group(ac->ac_sb, group);
1802         ext4_mb_unload_buddy(e4b);
1803 
1804         return 0;
1805 }
1806 
1807 static noinline_for_stack
1808 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1809                                 struct ext4_buddy *e4b)
1810 {
1811         ext4_group_t group = ac->ac_g_ex.fe_group;
1812         int max;
1813         int err;
1814         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1815         struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1816         struct ext4_free_extent ex;
1817 
1818         if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1819                 return 0;
1820         if (grp->bb_free == 0)
1821                 return 0;
1822 
1823         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1824         if (err)
1825                 return err;
1826 
1827         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1828                 ext4_mb_unload_buddy(e4b);
1829                 return 0;
1830         }
1831 
1832         ext4_lock_group(ac->ac_sb, group);
1833         max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1834                              ac->ac_g_ex.fe_len, &ex);
1835         ex.fe_logical = 0xDEADFA11; /* debug value */
1836 
1837         if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1838                 ext4_fsblk_t start;
1839 
1840                 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1841                         ex.fe_start;
1842                 /* use do_div to get remainder (would be 64-bit modulo) */
1843                 if (do_div(start, sbi->s_stripe) == 0) {
1844                         ac->ac_found++;
1845                         ac->ac_b_ex = ex;
1846                         ext4_mb_use_best_found(ac, e4b);
1847                 }
1848         } else if (max >= ac->ac_g_ex.fe_len) {
1849                 BUG_ON(ex.fe_len <= 0);
1850                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1851                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1852                 ac->ac_found++;
1853                 ac->ac_b_ex = ex;
1854                 ext4_mb_use_best_found(ac, e4b);
1855         } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1856                 /* Sometimes, caller may want to merge even small
1857                  * number of blocks to an existing extent */
1858                 BUG_ON(ex.fe_len <= 0);
1859                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1860                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1861                 ac->ac_found++;
1862                 ac->ac_b_ex = ex;
1863                 ext4_mb_use_best_found(ac, e4b);
1864         }
1865         ext4_unlock_group(ac->ac_sb, group);
1866         ext4_mb_unload_buddy(e4b);
1867 
1868         return 0;
1869 }
1870 
1871 /*
1872  * The routine scans buddy structures (not bitmap!) from given order
1873  * to max order and tries to find big enough chunk to satisfy the req
1874  */
1875 static noinline_for_stack
1876 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1877                                         struct ext4_buddy *e4b)
1878 {
1879         struct super_block *sb = ac->ac_sb;
1880         struct ext4_group_info *grp = e4b->bd_info;
1881         void *buddy;
1882         int i;
1883         int k;
1884         int max;
1885 
1886         BUG_ON(ac->ac_2order <= 0);
1887         for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1888                 if (grp->bb_counters[i] == 0)
1889                         continue;
1890 
1891                 buddy = mb_find_buddy(e4b, i, &max);
1892                 BUG_ON(buddy == NULL);
1893 
1894                 k = mb_find_next_zero_bit(buddy, max, 0);
1895                 BUG_ON(k >= max);
1896 
1897                 ac->ac_found++;
1898 
1899                 ac->ac_b_ex.fe_len = 1 << i;
1900                 ac->ac_b_ex.fe_start = k << i;
1901                 ac->ac_b_ex.fe_group = e4b->bd_group;
1902 
1903                 ext4_mb_use_best_found(ac, e4b);
1904 
1905                 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1906 
1907                 if (EXT4_SB(sb)->s_mb_stats)
1908                         atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1909 
1910                 break;
1911         }
1912 }
1913 
1914 /*
1915  * The routine scans the group and measures all found extents.
1916  * In order to optimize scanning, caller must pass number of
1917  * free blocks in the group, so the routine can know upper limit.
1918  */
1919 static noinline_for_stack
1920 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1921                                         struct ext4_buddy *e4b)
1922 {
1923         struct super_block *sb = ac->ac_sb;
1924         void *bitmap = e4b->bd_bitmap;
1925         struct ext4_free_extent ex;
1926         int i;
1927         int free;
1928 
1929         free = e4b->bd_info->bb_free;
1930         BUG_ON(free <= 0);
1931 
1932         i = e4b->bd_info->bb_first_free;
1933 
1934         while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1935                 i = mb_find_next_zero_bit(bitmap,
1936                                                 EXT4_CLUSTERS_PER_GROUP(sb), i);
1937                 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1938                         /*
1939                          * IF we have corrupt bitmap, we won't find any
1940                          * free blocks even though group info says we
1941                          * we have free blocks
1942                          */
1943                         ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1944                                         "%d free clusters as per "
1945                                         "group info. But bitmap says 0",
1946                                         free);
1947                         break;
1948                 }
1949 
1950                 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1951                 BUG_ON(ex.fe_len <= 0);
1952                 if (free < ex.fe_len) {
1953                         ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1954                                         "%d free clusters as per "
1955                                         "group info. But got %d blocks",
1956                                         free, ex.fe_len);
1957                         /*
1958                          * The number of free blocks differs. This mostly
1959                          * indicate that the bitmap is corrupt. So exit
1960                          * without claiming the space.
1961                          */
1962                         break;
1963                 }
1964                 ex.fe_logical = 0xDEADC0DE; /* debug value */
1965                 ext4_mb_measure_extent(ac, &ex, e4b);
1966 
1967                 i += ex.fe_len;
1968                 free -= ex.fe_len;
1969         }
1970 
1971         ext4_mb_check_limits(ac, e4b, 1);
1972 }
1973 
1974 /*
1975  * This is a special case for storages like raid5
1976  * we try to find stripe-aligned chunks for stripe-size-multiple requests
1977  */
1978 static noinline_for_stack
1979 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1980                                  struct ext4_buddy *e4b)
1981 {
1982         struct super_block *sb = ac->ac_sb;
1983         struct ext4_sb_info *sbi = EXT4_SB(sb);
1984         void *bitmap = e4b->bd_bitmap;
1985         struct ext4_free_extent ex;
1986         ext4_fsblk_t first_group_block;
1987         ext4_fsblk_t a;
1988         ext4_grpblk_t i;
1989         int max;
1990 
1991         BUG_ON(sbi->s_stripe == 0);
1992 
1993         /* find first stripe-aligned block in group */
1994         first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1995 
1996         a = first_group_block + sbi->s_stripe - 1;
1997         do_div(a, sbi->s_stripe);
1998         i = (a * sbi->s_stripe) - first_group_block;
1999 
2000         while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2001                 if (!mb_test_bit(i, bitmap)) {
2002                         max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2003                         if (max >= sbi->s_stripe) {
2004                                 ac->ac_found++;
2005                                 ex.fe_logical = 0xDEADF00D; /* debug value */
2006                                 ac->ac_b_ex = ex;
2007                                 ext4_mb_use_best_found(ac, e4b);
2008                                 break;
2009                         }
2010                 }
2011                 i += sbi->s_stripe;
2012         }
2013 }
2014 
2015 /*
2016  * This is now called BEFORE we load the buddy bitmap.
2017  * Returns either 1 or 0 indicating that the group is either suitable
2018  * for the allocation or not. In addition it can also return negative
2019  * error code when something goes wrong.
2020  */
2021 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2022                                 ext4_group_t group, int cr)
2023 {
2024         unsigned free, fragments;
2025         int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2026         struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2027 
2028         BUG_ON(cr < 0 || cr >= 4);
2029 
2030         free = grp->bb_free;
2031         if (free == 0)
2032                 return 0;
2033         if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2034                 return 0;
2035 
2036         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2037                 return 0;
2038 
2039         /* We only do this if the grp has never been initialized */
2040         if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2041                 int ret = ext4_mb_init_group(ac->ac_sb, group);
2042                 if (ret)
2043                         return ret;
2044         }
2045 
2046         fragments = grp->bb_fragments;
2047         if (fragments == 0)
2048                 return 0;
2049 
2050         switch (cr) {
2051         case 0:
2052                 BUG_ON(ac->ac_2order == 0);
2053 
2054                 /* Avoid using the first bg of a flexgroup for data files */
2055                 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2056                     (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2057                     ((group % flex_size) == 0))
2058                         return 0;
2059 
2060                 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2061                     (free / fragments) >= ac->ac_g_ex.fe_len)
2062                         return 1;
2063 
2064                 if (grp->bb_largest_free_order < ac->ac_2order)
2065                         return 0;
2066 
2067                 return 1;
2068         case 1:
2069                 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2070                         return 1;
2071                 break;
2072         case 2:
2073                 if (free >= ac->ac_g_ex.fe_len)
2074                         return 1;
2075                 break;
2076         case 3:
2077                 return 1;
2078         default:
2079                 BUG();
2080         }
2081 
2082         return 0;
2083 }
2084 
2085 static noinline_for_stack int
2086 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2087 {
2088         ext4_group_t ngroups, group, i;
2089         int cr;
2090         int err = 0, first_err = 0;
2091         struct ext4_sb_info *sbi;
2092         struct super_block *sb;
2093         struct ext4_buddy e4b;
2094 
2095         sb = ac->ac_sb;
2096         sbi = EXT4_SB(sb);
2097         ngroups = ext4_get_groups_count(sb);
2098         /* non-extent files are limited to low blocks/groups */
2099         if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2100                 ngroups = sbi->s_blockfile_groups;
2101 
2102         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2103 
2104         /* first, try the goal */
2105         err = ext4_mb_find_by_goal(ac, &e4b);
2106         if (err || ac->ac_status == AC_STATUS_FOUND)
2107                 goto out;
2108 
2109         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2110                 goto out;
2111 
2112         /*
2113          * ac->ac2_order is set only if the fe_len is a power of 2
2114          * if ac2_order is set we also set criteria to 0 so that we
2115          * try exact allocation using buddy.
2116          */
2117         i = fls(ac->ac_g_ex.fe_len);
2118         ac->ac_2order = 0;
2119         /*
2120          * We search using buddy data only if the order of the request
2121          * is greater than equal to the sbi_s_mb_order2_reqs
2122          * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2123          */
2124         if (i >= sbi->s_mb_order2_reqs) {
2125                 /*
2126                  * This should tell if fe_len is exactly power of 2
2127                  */
2128                 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2129                         ac->ac_2order = i - 1;
2130         }
2131 
2132         /* if stream allocation is enabled, use global goal */
2133         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2134                 /* TBD: may be hot point */
2135                 spin_lock(&sbi->s_md_lock);
2136                 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2137                 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2138                 spin_unlock(&sbi->s_md_lock);
2139         }
2140 
2141         /* Let's just scan groups to find more-less suitable blocks */
2142         cr = ac->ac_2order ? 0 : 1;
2143         /*
2144          * cr == 0 try to get exact allocation,
2145          * cr == 3  try to get anything
2146          */
2147 repeat:
2148         for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2149                 ac->ac_criteria = cr;
2150                 /*
2151                  * searching for the right group start
2152                  * from the goal value specified
2153                  */
2154                 group = ac->ac_g_ex.fe_group;
2155 
2156                 for (i = 0; i < ngroups; group++, i++) {
2157                         int ret = 0;
2158                         cond_resched();
2159                         /*
2160                          * Artificially restricted ngroups for non-extent
2161                          * files makes group > ngroups possible on first loop.
2162                          */
2163                         if (group >= ngroups)
2164                                 group = 0;
2165 
2166                         /* This now checks without needing the buddy page */
2167                         ret = ext4_mb_good_group(ac, group, cr);
2168                         if (ret <= 0) {
2169                                 if (!first_err)
2170                                         first_err = ret;
2171                                 continue;
2172                         }
2173 
2174                         err = ext4_mb_load_buddy(sb, group, &e4b);
2175                         if (err)
2176                                 goto out;
2177 
2178                         ext4_lock_group(sb, group);
2179 
2180                         /*
2181                          * We need to check again after locking the
2182                          * block group
2183                          */
2184                         ret = ext4_mb_good_group(ac, group, cr);
2185                         if (ret <= 0) {
2186                                 ext4_unlock_group(sb, group);
2187                                 ext4_mb_unload_buddy(&e4b);
2188                                 if (!first_err)
2189                                         first_err = ret;
2190                                 continue;
2191                         }
2192 
2193                         ac->ac_groups_scanned++;
2194                         if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2195                                 ext4_mb_simple_scan_group(ac, &e4b);
2196                         else if (cr == 1 && sbi->s_stripe &&
2197                                         !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2198                                 ext4_mb_scan_aligned(ac, &e4b);
2199                         else
2200                                 ext4_mb_complex_scan_group(ac, &e4b);
2201 
2202                         ext4_unlock_group(sb, group);
2203                         ext4_mb_unload_buddy(&e4b);
2204 
2205                         if (ac->ac_status != AC_STATUS_CONTINUE)
2206                                 break;
2207                 }
2208         }
2209 
2210         if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2211             !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2212                 /*
2213                  * We've been searching too long. Let's try to allocate
2214                  * the best chunk we've found so far
2215                  */
2216 
2217                 ext4_mb_try_best_found(ac, &e4b);
2218                 if (ac->ac_status != AC_STATUS_FOUND) {
2219                         /*
2220                          * Someone more lucky has already allocated it.
2221                          * The only thing we can do is just take first
2222                          * found block(s)
2223                         printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2224                          */
2225                         ac->ac_b_ex.fe_group = 0;
2226                         ac->ac_b_ex.fe_start = 0;
2227                         ac->ac_b_ex.fe_len = 0;
2228                         ac->ac_status = AC_STATUS_CONTINUE;
2229                         ac->ac_flags |= EXT4_MB_HINT_FIRST;
2230                         cr = 3;
2231                         atomic_inc(&sbi->s_mb_lost_chunks);
2232                         goto repeat;
2233                 }
2234         }
2235 out:
2236         if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2237                 err = first_err;
2238         return err;
2239 }
2240 
2241 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2242 {
2243         struct super_block *sb = seq->private;
2244         ext4_group_t group;
2245 
2246         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2247                 return NULL;
2248         group = *pos + 1;
2249         return (void *) ((unsigned long) group);
2250 }
2251 
2252 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2253 {
2254         struct super_block *sb = seq->private;
2255         ext4_group_t group;
2256 
2257         ++*pos;
2258         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2259                 return NULL;
2260         group = *pos + 1;
2261         return (void *) ((unsigned long) group);
2262 }
2263 
2264 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2265 {
2266         struct super_block *sb = seq->private;
2267         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2268         int i;
2269         int err, buddy_loaded = 0;
2270         struct ext4_buddy e4b;
2271         struct ext4_group_info *grinfo;
2272         struct sg {
2273                 struct ext4_group_info info;
2274                 ext4_grpblk_t counters[16];
2275         } sg;
2276 
2277         group--;
2278         if (group == 0)
2279                 seq_puts(seq, "#group: free  frags first ["
2280                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
2281                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]");
2282 
2283         i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2284                 sizeof(struct ext4_group_info);
2285         grinfo = ext4_get_group_info(sb, group);
2286         /* Load the group info in memory only if not already loaded. */
2287         if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2288                 err = ext4_mb_load_buddy(sb, group, &e4b);
2289                 if (err) {
2290                         seq_printf(seq, "#%-5u: I/O error\n", group);
2291                         return 0;
2292                 }
2293                 buddy_loaded = 1;
2294         }
2295 
2296         memcpy(&sg, ext4_get_group_info(sb, group), i);
2297 
2298         if (buddy_loaded)
2299                 ext4_mb_unload_buddy(&e4b);
2300 
2301         seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2302                         sg.info.bb_fragments, sg.info.bb_first_free);
2303         for (i = 0; i <= 13; i++)
2304                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2305                                 sg.info.bb_counters[i] : 0);
2306         seq_printf(seq, " ]\n");
2307 
2308         return 0;
2309 }
2310 
2311 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2312 {
2313 }
2314 
2315 static const struct seq_operations ext4_mb_seq_groups_ops = {
2316         .start  = ext4_mb_seq_groups_start,
2317         .next   = ext4_mb_seq_groups_next,
2318         .stop   = ext4_mb_seq_groups_stop,
2319         .show   = ext4_mb_seq_groups_show,
2320 };
2321 
2322 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2323 {
2324         struct super_block *sb = PDE_DATA(inode);
2325         int rc;
2326 
2327         rc = seq_open(file, &ext4_mb_seq_groups_ops);
2328         if (rc == 0) {
2329                 struct seq_file *m = file->private_data;
2330                 m->private = sb;
2331         }
2332         return rc;
2333 
2334 }
2335 
2336 static const struct file_operations ext4_mb_seq_groups_fops = {
2337         .owner          = THIS_MODULE,
2338         .open           = ext4_mb_seq_groups_open,
2339         .read           = seq_read,
2340         .llseek         = seq_lseek,
2341         .release        = seq_release,
2342 };
2343 
2344 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2345 {
2346         int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2347         struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2348 
2349         BUG_ON(!cachep);
2350         return cachep;
2351 }
2352 
2353 /*
2354  * Allocate the top-level s_group_info array for the specified number
2355  * of groups
2356  */
2357 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2358 {
2359         struct ext4_sb_info *sbi = EXT4_SB(sb);
2360         unsigned size;
2361         struct ext4_group_info ***new_groupinfo;
2362 
2363         size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2364                 EXT4_DESC_PER_BLOCK_BITS(sb);
2365         if (size <= sbi->s_group_info_size)
2366                 return 0;
2367 
2368         size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2369         new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2370         if (!new_groupinfo) {
2371                 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2372                 return -ENOMEM;
2373         }
2374         if (sbi->s_group_info) {
2375                 memcpy(new_groupinfo, sbi->s_group_info,
2376                        sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2377                 kvfree(sbi->s_group_info);
2378         }
2379         sbi->s_group_info = new_groupinfo;
2380         sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2381         ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 
2382                    sbi->s_group_info_size);
2383         return 0;
2384 }
2385 
2386 /* Create and initialize ext4_group_info data for the given group. */
2387 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2388                           struct ext4_group_desc *desc)
2389 {
2390         int i;
2391         int metalen = 0;
2392         struct ext4_sb_info *sbi = EXT4_SB(sb);
2393         struct ext4_group_info **meta_group_info;
2394         struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2395 
2396         /*
2397          * First check if this group is the first of a reserved block.
2398          * If it's true, we have to allocate a new table of pointers
2399          * to ext4_group_info structures
2400          */
2401         if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2402                 metalen = sizeof(*meta_group_info) <<
2403                         EXT4_DESC_PER_BLOCK_BITS(sb);
2404                 meta_group_info = kmalloc(metalen, GFP_NOFS);
2405                 if (meta_group_info == NULL) {
2406                         ext4_msg(sb, KERN_ERR, "can't allocate mem "
2407                                  "for a buddy group");
2408                         goto exit_meta_group_info;
2409                 }
2410                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2411                         meta_group_info;
2412         }
2413 
2414         meta_group_info =
2415                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2416         i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2417 
2418         meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2419         if (meta_group_info[i] == NULL) {
2420                 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2421                 goto exit_group_info;
2422         }
2423         set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2424                 &(meta_group_info[i]->bb_state));
2425 
2426         /*
2427          * initialize bb_free to be able to skip
2428          * empty groups without initialization
2429          */
2430         if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2431                 meta_group_info[i]->bb_free =
2432                         ext4_free_clusters_after_init(sb, group, desc);
2433         } else {
2434                 meta_group_info[i]->bb_free =
2435                         ext4_free_group_clusters(sb, desc);
2436         }
2437 
2438         INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2439         init_rwsem(&meta_group_info[i]->alloc_sem);
2440         meta_group_info[i]->bb_free_root = RB_ROOT;
2441         meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2442 
2443 #ifdef DOUBLE_CHECK
2444         {
2445                 struct buffer_head *bh;
2446                 meta_group_info[i]->bb_bitmap =
2447                         kmalloc(sb->s_blocksize, GFP_NOFS);
2448                 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2449                 bh = ext4_read_block_bitmap(sb, group);
2450                 BUG_ON(bh == NULL);
2451                 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2452                         sb->s_blocksize);
2453                 put_bh(bh);
2454         }
2455 #endif
2456 
2457         return 0;
2458 
2459 exit_group_info:
2460         /* If a meta_group_info table has been allocated, release it now */
2461         if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2462                 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2463                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2464         }
2465 exit_meta_group_info:
2466         return -ENOMEM;
2467 } /* ext4_mb_add_groupinfo */
2468 
2469 static int ext4_mb_init_backend(struct super_block *sb)
2470 {
2471         ext4_group_t ngroups = ext4_get_groups_count(sb);
2472         ext4_group_t i;
2473         struct ext4_sb_info *sbi = EXT4_SB(sb);
2474         int err;
2475         struct ext4_group_desc *desc;
2476         struct kmem_cache *cachep;
2477 
2478         err = ext4_mb_alloc_groupinfo(sb, ngroups);
2479         if (err)
2480                 return err;
2481 
2482         sbi->s_buddy_cache = new_inode(sb);
2483         if (sbi->s_buddy_cache == NULL) {
2484                 ext4_msg(sb, KERN_ERR, "can't get new inode");
2485                 goto err_freesgi;
2486         }
2487         /* To avoid potentially colliding with an valid on-disk inode number,
2488          * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2489          * not in the inode hash, so it should never be found by iget(), but
2490          * this will avoid confusion if it ever shows up during debugging. */
2491         sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2492         EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2493         for (i = 0; i < ngroups; i++) {
2494                 desc = ext4_get_group_desc(sb, i, NULL);
2495                 if (desc == NULL) {
2496                         ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2497                         goto err_freebuddy;
2498                 }
2499                 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2500                         goto err_freebuddy;
2501         }
2502 
2503         return 0;
2504 
2505 err_freebuddy:
2506         cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2507         while (i-- > 0)
2508                 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2509         i = sbi->s_group_info_size;
2510         while (i-- > 0)
2511                 kfree(sbi->s_group_info[i]);
2512         iput(sbi->s_buddy_cache);
2513 err_freesgi:
2514         kvfree(sbi->s_group_info);
2515         return -ENOMEM;
2516 }
2517 
2518 static void ext4_groupinfo_destroy_slabs(void)
2519 {
2520         int i;
2521 
2522         for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2523                 if (ext4_groupinfo_caches[i])
2524                         kmem_cache_destroy(ext4_groupinfo_caches[i]);
2525                 ext4_groupinfo_caches[i] = NULL;
2526         }
2527 }
2528 
2529 static int ext4_groupinfo_create_slab(size_t size)
2530 {
2531         static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2532         int slab_size;
2533         int blocksize_bits = order_base_2(size);
2534         int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2535         struct kmem_cache *cachep;
2536 
2537         if (cache_index >= NR_GRPINFO_CACHES)
2538                 return -EINVAL;
2539 
2540         if (unlikely(cache_index < 0))
2541                 cache_index = 0;
2542 
2543         mutex_lock(&ext4_grpinfo_slab_create_mutex);
2544         if (ext4_groupinfo_caches[cache_index]) {
2545                 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2546                 return 0;       /* Already created */
2547         }
2548 
2549         slab_size = offsetof(struct ext4_group_info,
2550                                 bb_counters[blocksize_bits + 2]);
2551 
2552         cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2553                                         slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2554                                         NULL);
2555 
2556         ext4_groupinfo_caches[cache_index] = cachep;
2557 
2558         mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2559         if (!cachep) {
2560                 printk(KERN_EMERG
2561                        "EXT4-fs: no memory for groupinfo slab cache\n");
2562                 return -ENOMEM;
2563         }
2564 
2565         return 0;
2566 }
2567 
2568 int ext4_mb_init(struct super_block *sb)
2569 {
2570         struct ext4_sb_info *sbi = EXT4_SB(sb);
2571         unsigned i, j;
2572         unsigned offset;
2573         unsigned max;
2574         int ret;
2575 
2576         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2577 
2578         sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2579         if (sbi->s_mb_offsets == NULL) {
2580                 ret = -ENOMEM;
2581                 goto out;
2582         }
2583 
2584         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2585         sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2586         if (sbi->s_mb_maxs == NULL) {
2587                 ret = -ENOMEM;
2588                 goto out;
2589         }
2590 
2591         ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2592         if (ret < 0)
2593                 goto out;
2594 
2595         /* order 0 is regular bitmap */
2596         sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2597         sbi->s_mb_offsets[0] = 0;
2598 
2599         i = 1;
2600         offset = 0;
2601         max = sb->s_blocksize << 2;
2602         do {
2603                 sbi->s_mb_offsets[i] = offset;
2604                 sbi->s_mb_maxs[i] = max;
2605                 offset += 1 << (sb->s_blocksize_bits - i);
2606                 max = max >> 1;
2607                 i++;
2608         } while (i <= sb->s_blocksize_bits + 1);
2609 
2610         spin_lock_init(&sbi->s_md_lock);
2611         spin_lock_init(&sbi->s_bal_lock);
2612 
2613         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2614         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2615         sbi->s_mb_stats = MB_DEFAULT_STATS;
2616         sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2617         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2618         /*
2619          * The default group preallocation is 512, which for 4k block
2620          * sizes translates to 2 megabytes.  However for bigalloc file
2621          * systems, this is probably too big (i.e, if the cluster size
2622          * is 1 megabyte, then group preallocation size becomes half a
2623          * gigabyte!).  As a default, we will keep a two megabyte
2624          * group pralloc size for cluster sizes up to 64k, and after
2625          * that, we will force a minimum group preallocation size of
2626          * 32 clusters.  This translates to 8 megs when the cluster
2627          * size is 256k, and 32 megs when the cluster size is 1 meg,
2628          * which seems reasonable as a default.
2629          */
2630         sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2631                                        sbi->s_cluster_bits, 32);
2632         /*
2633          * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2634          * to the lowest multiple of s_stripe which is bigger than
2635          * the s_mb_group_prealloc as determined above. We want
2636          * the preallocation size to be an exact multiple of the
2637          * RAID stripe size so that preallocations don't fragment
2638          * the stripes.
2639          */
2640         if (sbi->s_stripe > 1) {
2641                 sbi->s_mb_group_prealloc = roundup(
2642                         sbi->s_mb_group_prealloc, sbi->s_stripe);
2643         }
2644 
2645         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2646         if (sbi->s_locality_groups == NULL) {
2647                 ret = -ENOMEM;
2648                 goto out;
2649         }
2650         for_each_possible_cpu(i) {
2651                 struct ext4_locality_group *lg;
2652                 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2653                 mutex_init(&lg->lg_mutex);
2654                 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2655                         INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2656                 spin_lock_init(&lg->lg_prealloc_lock);
2657         }
2658 
2659         /* init file for buddy data */
2660         ret = ext4_mb_init_backend(sb);
2661         if (ret != 0)
2662                 goto out_free_locality_groups;
2663 
2664         if (sbi->s_proc)
2665                 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2666                                  &ext4_mb_seq_groups_fops, sb);
2667 
2668         return 0;
2669 
2670 out_free_locality_groups:
2671         free_percpu(sbi->s_locality_groups);
2672         sbi->s_locality_groups = NULL;
2673 out:
2674         kfree(sbi->s_mb_offsets);
2675         sbi->s_mb_offsets = NULL;
2676         kfree(sbi->s_mb_maxs);
2677         sbi->s_mb_maxs = NULL;
2678         return ret;
2679 }
2680 
2681 /* need to called with the ext4 group lock held */
2682 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2683 {
2684         struct ext4_prealloc_space *pa;
2685         struct list_head *cur, *tmp;
2686         int count = 0;
2687 
2688         list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2689                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2690                 list_del(&pa->pa_group_list);
2691                 count++;
2692                 kmem_cache_free(ext4_pspace_cachep, pa);
2693         }
2694         if (count)
2695                 mb_debug(1, "mballoc: %u PAs left\n", count);
2696 
2697 }
2698 
2699 int ext4_mb_release(struct super_block *sb)
2700 {
2701         ext4_group_t ngroups = ext4_get_groups_count(sb);
2702         ext4_group_t i;
2703         int num_meta_group_infos;
2704         struct ext4_group_info *grinfo;
2705         struct ext4_sb_info *sbi = EXT4_SB(sb);
2706         struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2707 
2708         if (sbi->s_proc)
2709                 remove_proc_entry("mb_groups", sbi->s_proc);
2710 
2711         if (sbi->s_group_info) {
2712                 for (i = 0; i < ngroups; i++) {
2713                         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);
2720                         kmem_cache_free(cachep, grinfo);
2721                 }
2722                 num_meta_group_infos = (ngroups +
2723                                 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]);
2727                 kvfree(sbi->s_group_info);
2728         }
2729         kfree(sbi->s_mb_offsets);
2730         kfree(sbi->s_mb_maxs);
2731         iput(sbi->s_buddy_cache);
2732         if (sbi->s_mb_stats) {
2733                 ext4_msg(sb, KERN_INFO,
2734                        "mballoc: %u blocks %u reqs (%u success)",
2735                                 atomic_read(&sbi->s_bal_allocated),
2736                                 atomic_read(&sbi->s_bal_reqs),
2737                                 atomic_read(&sbi->s_bal_success));
2738                 ext4_msg(sb, KERN_INFO,
2739                       "mballoc: %u extents scanned, %u goal hits, "
2740                                 "%u 2^N hits, %u breaks, %u lost",
2741                                 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));
2746                 ext4_msg(sb, KERN_INFO,
2747                        "mballoc: %lu generated and it took %Lu",
2748                                 sbi->s_mb_buddies_generated,
2749                                 sbi->s_mb_generation_time);
2750                 ext4_msg(sb, KERN_INFO,
2751                        "mballoc: %u preallocated, %u discarded",
2752                                 atomic_read(&sbi->s_mb_preallocated),
2753                                 atomic_read(&sbi->s_mb_discarded));
2754         }
2755 
2756         free_percpu(sbi->s_locality_groups);
2757 
2758         return 0;
2759 }
2760 
2761 static inline int ext4_issue_discard(struct super_block *sb,
2762                 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2763 {
2764         ext4_fsblk_t discard_block;
2765 
2766         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);
2769         trace_ext4_discard_blocks(sb,
2770                         (unsigned long long) discard_block, count);
2771         return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2772 }
2773 
2774 /*
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  */
2778 static void ext4_free_data_callback(struct super_block *sb,
2779                                     struct ext4_journal_cb_entry *jce,
2780                                     int rc)
2781 {
2782         struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2783         struct ext4_buddy e4b;
2784         struct ext4_group_info *db;
2785         int err, count = 0, count2 = 0;
2786 
2787         mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2788                  entry->efd_count, entry->efd_group, entry);
2789 
2790         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         }
2801 
2802         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);
2805 
2806 
2807         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);
2815 
2816         /*
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()
2828                  */
2829                 page_cache_release(e4b.bd_buddy_page);
2830                 page_cache_release(e4b.bd_bitmap_page);
2831         }
2832         ext4_unlock_group(sb, entry->efd_group);
2833         kmem_cache_free(ext4_free_data_cachep, entry);
2834         ext4_mb_unload_buddy(&e4b);
2835 
2836         mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2837 }
2838 
2839 int __init ext4_init_mballoc(void)
2840 {
2841         ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2842                                         SLAB_RECLAIM_ACCOUNT);
2843         if (ext4_pspace_cachep == NULL)
2844                 return -ENOMEM;
2845 
2846         ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2847                                     SLAB_RECLAIM_ACCOUNT);
2848         if (ext4_ac_cachep == NULL) {
2849                 kmem_cache_destroy(ext4_pspace_cachep);
2850                 return -ENOMEM;
2851         }
2852 
2853         ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2854                                            SLAB_RECLAIM_ACCOUNT);
2855         if (ext4_free_data_cachep == NULL) {
2856                 kmem_cache_destroy(ext4_pspace_cachep);
2857                 kmem_cache_destroy(ext4_ac_cachep);
2858                 return -ENOMEM;
2859         }
2860         return 0;
2861 }
2862 
2863 void ext4_exit_mballoc(void)
2864 {
2865         /*
2866          * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2867          * before destroying the slab cache.
2868          */
2869         rcu_barrier();
2870         kmem_cache_destroy(ext4_pspace_cachep);
2871         kmem_cache_destroy(ext4_ac_cachep);
2872         kmem_cache_destroy(ext4_free_data_cachep);
2873         ext4_groupinfo_destroy_slabs();
2874 }
2875 
2876 
2877 /*
2878  * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2879  * Returns 0 if success or error code
2880  */
2881 static noinline_for_stack int
2882 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2883                                 handle_t *handle, unsigned int reserv_clstrs)
2884 {
2885         struct buffer_head *bitmap_bh = NULL;
2886         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;
2891         int err, len;
2892 
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);
2898 
2899         err = -EIO;
2900         bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2901         if (!bitmap_bh)
2902                 goto out_err;
2903 
2904         BUFFER_TRACE(bitmap_bh, "getting write access");
2905         err = ext4_journal_get_write_access(handle, bitmap_bh);
2906         if (err)
2907                 goto out_err;
2908 
2909         err = -EIO;
2910         gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2911         if (!gdp)
2912                 goto out_err;
2913 
2914         ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2915                         ext4_free_group_clusters(sb, gdp));
2916 
2917         BUFFER_TRACE(gdp_bh, "get_write_access");
2918         err = ext4_journal_get_write_access(handle, gdp_bh);
2919         if (err)
2920                 goto out_err;
2921 
2922         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2923 
2924         len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2925         if (!ext4_data_block_valid(sbi, block, len)) {
2926                 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2927                            "fs metadata", block, block+len);
2928                 /* File system mounted not to panic on error
2929                  * Fix the bitmap and repeat the block allocation
2930                  * We leak some of the blocks here.
2931                  */
2932                 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2933                 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2934                               ac->ac_b_ex.fe_len);
2935                 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2936                 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2937                 if (!err)
2938                         err = -EAGAIN;
2939                 goto out_err;
2940         }
2941 
2942         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2943 #ifdef AGGRESSIVE_CHECK
2944         {
2945                 int i;
2946                 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2947                         BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2948                                                 bitmap_bh->b_data));
2949                 }
2950         }
2951 #endif
2952         ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2953                       ac->ac_b_ex.fe_len);
2954         if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2955                 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2956                 ext4_free_group_clusters_set(sb, gdp,
2957                                              ext4_free_clusters_after_init(sb,
2958                                                 ac->ac_b_ex.fe_group, gdp));
2959         }
2960         len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2961         ext4_free_group_clusters_set(sb, gdp, len);
2962         ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2963         ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2964 
2965         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2966         percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2967         /*
2968          * Now reduce the dirty block count also. Should not go negative
2969          */
2970         if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2971                 /* release all the reserved blocks if non delalloc */
2972                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2973                                    reserv_clstrs);
2974 
2975         if (sbi->s_log_groups_per_flex) {
2976                 ext4_group_t flex_group = ext4_flex_group(sbi,
2977                                                           ac->ac_b_ex.fe_group);
2978                 atomic64_sub(ac->ac_b_ex.fe_len,
2979                              &sbi->s_flex_groups[flex_group].free_clusters);
2980         }
2981 
2982         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2983         if (err)
2984                 goto out_err;
2985         err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2986 
2987 out_err:
2988         brelse(bitmap_bh);
2989         return err;
2990 }
2991 
2992 /*
2993  * here we normalize request for locality group
2994  * Group request are normalized to s_mb_group_prealloc, which goes to
2995  * s_strip if we set the same via mount option.
2996  * s_mb_group_prealloc can be configured via
2997  * /sys/fs/ext4/<partition>/mb_group_prealloc
2998  *
2999  * XXX: should we try to preallocate more than the group has now?
3000  */
3001 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3002 {
3003         struct super_block *sb = ac->ac_sb;
3004         struct ext4_locality_group *lg = ac->ac_lg;
3005 
3006         BUG_ON(lg == NULL);
3007         ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3008         mb_debug(1, "#%u: goal %u blocks for locality group\n",
3009                 current->pid, ac->ac_g_ex.fe_len);
3010 }
3011 
3012 /*
3013  * Normalization means making request better in terms of
3014  * size and alignment
3015  */
3016 static noinline_for_stack void
3017 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3018                                 struct ext4_allocation_request *ar)
3019 {
3020         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3021         int bsbits, max;
3022         ext4_lblk_t end;
3023         loff_t size, start_off;
3024         loff_t orig_size __maybe_unused;
3025         ext4_lblk_t start;
3026         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3027         struct ext4_prealloc_space *pa;
3028 
3029         /* do normalize only data requests, metadata requests
3030            do not need preallocation */
3031         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3032                 return;
3033 
3034         /* sometime caller may want exact blocks */
3035         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3036                 return;
3037 
3038         /* caller may indicate that preallocation isn't
3039          * required (it's a tail, for example) */
3040         if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3041                 return;
3042 
3043         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3044                 ext4_mb_normalize_group_request(ac);
3045                 return ;
3046         }
3047 
3048         bsbits = ac->ac_sb->s_blocksize_bits;
3049 
3050         /* first, let's learn actual file size
3051          * given current request is allocated */
3052         size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3053         size = size << bsbits;
3054         if (size < i_size_read(ac->ac_inode))
3055                 size = i_size_read(ac->ac_inode);
3056         orig_size = size;
3057 
3058         /* max size of free chunks */
3059         max = 2 << bsbits;
3060 
3061 #define NRL_CHECK_SIZE(req, size, max, chunk_size)      \
3062                 (req <= (size) || max <= (chunk_size))
3063 
3064         /* first, try to predict filesize */
3065         /* XXX: should this table be tunable? */
3066         start_off = 0;
3067         if (size <= 16 * 1024) {
3068                 size = 16 * 1024;
3069         } else if (size <= 32 * 1024) {
3070                 size = 32 * 1024;
3071         } else if (size <= 64 * 1024) {
3072                 size = 64 * 1024;
3073         } else if (size <= 128 * 1024) {
3074                 size = 128 * 1024;
3075         } else if (size <= 256 * 1024) {
3076                 size = 256 * 1024;
3077         } else if (size <= 512 * 1024) {
3078                 size = 512 * 1024;
3079         } else if (size <= 1024 * 1024) {
3080                 size = 1024 * 1024;
3081         } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3082                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3083                                                 (21 - bsbits)) << 21;
3084                 size = 2 * 1024 * 1024;
3085         } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3086                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3087                                                         (22 - bsbits)) << 22;
3088                 size = 4 * 1024 * 1024;
3089         } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3090                                         (8<<20)>>bsbits, max, 8 * 1024)) {
3091                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3092                                                         (23 - bsbits)) << 23;
3093                 size = 8 * 1024 * 1024;
3094         } else {
3095                 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3096                 size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3097                                               ac->ac_o_ex.fe_len) << bsbits;
3098         }
3099         size = size >> bsbits;
3100         start = start_off >> bsbits;
3101 
3102         /* don't cover already allocated blocks in selected range */
3103         if (ar->pleft && start <= ar->lleft) {
3104                 size -= ar->lleft + 1 - start;
3105                 start = ar->lleft + 1;
3106         }
3107         if (ar->pright && start + size - 1 >= ar->lright)
3108                 size -= start + size - ar->lright;
3109 
3110         end = start + size;
3111 
3112         /* check we don't cross already preallocated blocks */
3113         rcu_read_lock();
3114         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3115                 ext4_lblk_t pa_end;
3116 
3117                 if (pa->pa_deleted)
3118                         continue;
3119                 spin_lock(&pa->pa_lock);
3120                 if (pa->pa_deleted) {
3121                         spin_unlock(&pa->pa_lock);
3122                         continue;
3123                 }
3124 
3125                 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3126                                                   pa->pa_len);
3127 
3128                 /* PA must not overlap original request */
3129                 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3130                         ac->ac_o_ex.fe_logical < pa->pa_lstart));
3131 
3132                 /* skip PAs this normalized request doesn't overlap with */
3133                 if (pa->pa_lstart >= end || pa_end <= start) {
3134                         spin_unlock(&pa->pa_lock);
3135                         continue;
3136                 }
3137                 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3138 
3139                 /* adjust start or end to be adjacent to this pa */
3140                 if (pa_end <= ac->ac_o_ex.fe_logical) {
3141                         BUG_ON(pa_end < start);
3142                         start = pa_end;
3143                 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3144                         BUG_ON(pa->pa_lstart > end);
3145                         end = pa->pa_lstart;
3146                 }
3147                 spin_unlock(&pa->pa_lock);
3148         }
3149         rcu_read_unlock();
3150         size = end - start;
3151 
3152         /* XXX: extra loop to check we really don't overlap preallocations */
3153         rcu_read_lock();
3154         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3155                 ext4_lblk_t pa_end;
3156 
3157                 spin_lock(&pa->pa_lock);
3158                 if (pa->pa_deleted == 0) {
3159                         pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3160                                                           pa->pa_len);
3161                         BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3162                 }
3163                 spin_unlock(&pa->pa_lock);
3164         }
3165         rcu_read_unlock();
3166 
3167         if (start + size <= ac->ac_o_ex.fe_logical &&
3168                         start > ac->ac_o_ex.fe_logical) {
3169                 ext4_msg(ac->ac_sb, KERN_ERR,
3170                          "start %lu, size %lu, fe_logical %lu",
3171                          (unsigned long) start, (unsigned long) size,
3172                          (unsigned long) ac->ac_o_ex.fe_logical);
3173                 BUG();
3174         }
3175         BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3176 
3177         /* now prepare goal request */
3178 
3179         /* XXX: is it better to align blocks WRT to logical
3180          * placement or satisfy big request as is */
3181         ac->ac_g_ex.fe_logical = start;
3182         ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3183 
3184         /* define goal start in order to merge */
3185         if (ar->pright && (ar->lright == (start + size))) {
3186                 /* merge to the right */
3187                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3188                                                 &ac->ac_f_ex.fe_group,
3189                                                 &ac->ac_f_ex.fe_start);
3190                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3191         }
3192         if (ar->pleft && (ar->lleft + 1 == start)) {
3193                 /* merge to the left */
3194                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3195                                                 &ac->ac_f_ex.fe_group,
3196                                                 &ac->ac_f_ex.fe_start);
3197                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3198         }
3199 
3200         mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3201                 (unsigned) orig_size, (unsigned) start);
3202 }
3203 
3204 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3205 {
3206         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3207 
3208         if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3209                 atomic_inc(&sbi->s_bal_reqs);
3210                 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3211                 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3212                         atomic_inc(&sbi->s_bal_success);
3213                 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3214                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3215                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3216                         atomic_inc(&sbi->s_bal_goals);
3217                 if (ac->ac_found > sbi->s_mb_max_to_scan)
3218                         atomic_inc(&sbi->s_bal_breaks);
3219         }
3220 
3221         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3222                 trace_ext4_mballoc_alloc(ac);
3223         else
3224                 trace_ext4_mballoc_prealloc(ac);
3225 }
3226 
3227 /*
3228  * Called on failure; free up any blocks from the inode PA for this
3229  * context.  We don't need this for MB_GROUP_PA because we only change
3230  * pa_free in ext4_mb_release_context(), but on failure, we've already
3231  * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3232  */
3233 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3234 {
3235         struct ext4_prealloc_space *pa = ac->ac_pa;
3236         struct ext4_buddy e4b;
3237         int err;
3238 
3239         if (pa == NULL) {
3240                 if (ac->ac_f_ex.fe_len == 0)
3241                         return;
3242                 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3243                 if (err) {
3244                         /*
3245                          * This should never happen since we pin the
3246                          * pages in the ext4_allocation_context so
3247                          * ext4_mb_load_buddy() should never fail.
3248                          */
3249                         WARN(1, "mb_load_buddy failed (%d)", err);
3250                         return;
3251                 }
3252                 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3253                 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3254                                ac->ac_f_ex.fe_len);
3255                 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3256                 ext4_mb_unload_buddy(&e4b);
3257                 return;
3258         }
3259         if (pa->pa_type == MB_INODE_PA)
3260                 pa->pa_free += ac->ac_b_ex.fe_len;
3261 }
3262 
3263 /*
3264  * use blocks preallocated to inode
3265  */
3266 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3267                                 struct ext4_prealloc_space *pa)
3268 {
3269         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3270         ext4_fsblk_t start;
3271         ext4_fsblk_t end;
3272         int len;
3273 
3274         /* found preallocated blocks, use them */
3275         start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3276         end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3277                   start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3278         len = EXT4_NUM_B2C(sbi, end - start);
3279         ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3280                                         &ac->ac_b_ex.fe_start);
3281         ac->ac_b_ex.fe_len = len;
3282         ac->ac_status = AC_STATUS_FOUND;
3283         ac->ac_pa = pa;
3284 
3285         BUG_ON(start < pa->pa_pstart);
3286         BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3287         BUG_ON(pa->pa_free < len);
3288         pa->pa_free -= len;
3289 
3290         mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3291 }
3292 
3293 /*
3294  * use blocks preallocated to locality group
3295  */
3296 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3297                                 struct ext4_prealloc_space *pa)
3298 {
3299         unsigned int len = ac->ac_o_ex.fe_len;
3300 
3301         ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3302                                         &ac->ac_b_ex.fe_group,
3303                                         &ac->ac_b_ex.fe_start);
3304         ac->ac_b_ex.fe_len = len;
3305         ac->ac_status = AC_STATUS_FOUND;
3306         ac->ac_pa = pa;
3307 
3308         /* we don't correct pa_pstart or pa_plen here to avoid
3309          * possible race when the group is being loaded concurrently
3310          * instead we correct pa later, after blocks are marked
3311          * in on-disk bitmap -- see ext4_mb_release_context()
3312          * Other CPUs are prevented from allocating from this pa by lg_mutex
3313          */
3314         mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3315 }
3316 
3317 /*
3318  * Return the prealloc space that have minimal distance
3319  * from the goal block. @cpa is the prealloc
3320  * space that is having currently known minimal distance
3321  * from the goal block.
3322  */
3323 static struct ext4_prealloc_space *
3324 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3325                         struct ext4_prealloc_space *pa,
3326                         struct ext4_prealloc_space *cpa)
3327 {
3328         ext4_fsblk_t cur_distance, new_distance;
3329 
3330         if (cpa == NULL) {
3331                 atomic_inc(&pa->pa_count);
3332                 return pa;
3333         }
3334         cur_distance = abs(goal_block - cpa->pa_pstart);
3335         new_distance = abs(goal_block - pa->pa_pstart);
3336 
3337         if (cur_distance <= new_distance)
3338                 return cpa;
3339 
3340         /* drop the previous reference */
3341         atomic_dec(&cpa->pa_count);
3342         atomic_inc(&pa->pa_count);
3343         return pa;
3344 }
3345 
3346 /*
3347  * search goal blocks in preallocated space
3348  */
3349 static noinline_for_stack int
3350 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3351 {
3352         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3353         int order, i;
3354         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3355         struct ext4_locality_group *lg;
3356         struct ext4_prealloc_space *pa, *cpa = NULL;
3357         ext4_fsblk_t goal_block;
3358 
3359         /* only data can be preallocated */
3360         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3361                 return 0;
3362 
3363         /* first, try per-file preallocation */
3364         rcu_read_lock();
3365         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3366 
3367                 /* all fields in this condition don't change,
3368                  * so we can skip locking for them */
3369                 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3370                     ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3371                                                EXT4_C2B(sbi, pa->pa_len)))
3372                         continue;
3373 
3374                 /* non-extent files can't have physical blocks past 2^32 */
3375                 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3376                     (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3377                      EXT4_MAX_BLOCK_FILE_PHYS))
3378                         continue;
3379 
3380                 /* found preallocated blocks, use them */
3381                 spin_lock(&pa->pa_lock);
3382                 if (pa->pa_deleted == 0 && pa->pa_free) {
3383                         atomic_inc(&pa->pa_count);
3384                         ext4_mb_use_inode_pa(ac, pa);
3385                         spin_unlock(&pa->pa_lock);
3386                         ac->ac_criteria = 10;
3387                         rcu_read_unlock();
3388                         return 1;
3389                 }
3390                 spin_unlock(&pa->pa_lock);
3391         }
3392         rcu_read_unlock();
3393 
3394         /* can we use group allocation? */
3395         if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3396                 return 0;
3397 
3398         /* inode may have no locality group for some reason */
3399         lg = ac->ac_lg;
3400         if (lg == NULL)
3401                 return 0;
3402         order  = fls(ac->ac_o_ex.fe_len) - 1;
3403         if (order > PREALLOC_TB_SIZE - 1)
3404                 /* The max size of hash table is PREALLOC_TB_SIZE */
3405                 order = PREALLOC_TB_SIZE - 1;
3406 
3407         goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3408         /*
3409          * search for the prealloc space that is having
3410          * minimal distance from the goal block.
3411          */
3412         for (i = order; i < PREALLOC_TB_SIZE; i++) {
3413                 rcu_read_lock();
3414                 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3415                                         pa_inode_list) {
3416                         spin_lock(&pa->pa_lock);
3417                         if (pa->pa_deleted == 0 &&
3418                                         pa->pa_free >= ac->ac_o_ex.fe_len) {
3419 
3420                                 cpa = ext4_mb_check_group_pa(goal_block,
3421                                                                 pa, cpa);
3422                         }
3423                         spin_unlock(&pa->pa_lock);
3424                 }
3425                 rcu_read_unlock();
3426         }
3427         if (cpa) {
3428                 ext4_mb_use_group_pa(ac, cpa);
3429                 ac->ac_criteria = 20;
3430                 return 1;
3431         }
3432         return 0;
3433 }
3434 
3435 /*
3436  * the function goes through all block freed in the group
3437  * but not yet committed and marks them used in in-core bitmap.
3438  * buddy must be generated from this bitmap
3439  * Need to be called with the ext4 group lock held
3440  */
3441 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3442                                                 ext4_group_t group)
3443 {
3444         struct rb_node *n;
3445         struct ext4_group_info *grp;
3446         struct ext4_free_data *entry;
3447 
3448         grp = ext4_get_group_info(sb, group);
3449         n = rb_first(&(grp->bb_free_root));
3450 
3451         while (n) {
3452                 entry = rb_entry(n, struct ext4_free_data, efd_node);
3453                 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3454                 n = rb_next(n);
3455         }
3456         return;
3457 }
3458 
3459 /*
3460  * the function goes through all preallocation in this group and marks them
3461  * used in in-core bitmap. buddy must be generated from this bitmap
3462  * Need to be called with ext4 group lock held
3463  */
3464 static noinline_for_stack
3465 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3466                                         ext4_group_t group)
3467 {
3468         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3469         struct ext4_prealloc_space *pa;
3470         struct list_head *cur;
3471         ext4_group_t groupnr;
3472         ext4_grpblk_t start;
3473         int preallocated = 0;
3474         int len;
3475 
3476         /* all form of preallocation discards first load group,
3477          * so the only competing code is preallocation use.
3478          * we don't need any locking here
3479          * notice we do NOT ignore preallocations with pa_deleted
3480          * otherwise we could leave used blocks available for
3481          * allocation in buddy when concurrent ext4_mb_put_pa()
3482          * is dropping preallocation
3483          */
3484         list_for_each(cur, &grp->bb_prealloc_list) {
3485                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3486                 spin_lock(&pa->pa_lock);
3487                 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3488                                              &groupnr, &start);
3489                 len = pa->pa_len;
3490                 spin_unlock(&pa->pa_lock);
3491                 if (unlikely(len == 0))
3492                         continue;
3493                 BUG_ON(groupnr != group);
3494                 ext4_set_bits(bitmap, start, len);
3495                 preallocated += len;
3496         }
3497         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3498 }
3499 
3500 static void ext4_mb_pa_callback(struct rcu_head *head)
3501 {
3502         struct ext4_prealloc_space *pa;
3503         pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3504 
3505         BUG_ON(atomic_read(&pa->pa_count));
3506         BUG_ON(pa->pa_deleted == 0);
3507         kmem_cache_free(ext4_pspace_cachep, pa);
3508 }
3509 
3510 /*
3511  * drops a reference to preallocated space descriptor
3512  * if this was the last reference and the space is consumed
3513  */
3514 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3515                         struct super_block *sb, struct ext4_prealloc_space *pa)
3516 {
3517         ext4_group_t grp;
3518         ext4_fsblk_t grp_blk;
3519 
3520         /* in this short window concurrent discard can set pa_deleted */
3521         spin_lock(&pa->pa_lock);
3522         if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3523                 spin_unlock(&pa->pa_lock);
3524                 return;
3525         }
3526 
3527         if (pa->pa_deleted == 1) {
3528                 spin_unlock(&pa->pa_lock);
3529                 return;
3530         }
3531 
3532         pa->pa_deleted = 1;
3533         spin_unlock(&pa->pa_lock);
3534 
3535         grp_blk = pa->pa_pstart;
3536         /*
3537          * If doing group-based preallocation, pa_pstart may be in the
3538          * next group when pa is used up
3539          */
3540         if (pa->pa_type == MB_GROUP_PA)
3541                 grp_blk--;
3542 
3543         grp = ext4_get_group_number(sb, grp_blk);
3544 
3545         /*
3546          * possible race:
3547          *
3548          *  P1 (buddy init)                     P2 (regular allocation)
3549          *                                      find block B in PA
3550          *  copy on-disk bitmap to buddy
3551          *                                      mark B in on-disk bitmap
3552          *                                      drop PA from group
3553          *  mark all PAs in buddy
3554          *
3555          * thus, P1 initializes buddy with B available. to prevent this
3556          * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3557          * against that pair
3558          */
3559         ext4_lock_group(sb, grp);
3560         list_del(&pa->pa_group_list);
3561         ext4_unlock_group(sb, grp);
3562 
3563         spin_lock(pa->pa_obj_lock);
3564         list_del_rcu(&pa->pa_inode_list);
3565         spin_unlock(pa->pa_obj_lock);
3566 
3567         call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3568 }
3569 
3570 /*
3571  * creates new preallocated space for given inode
3572  */
3573 static noinline_for_stack int
3574 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3575 {
3576         struct super_block *sb = ac->ac_sb;
3577         struct ext4_sb_info *sbi = EXT4_SB(sb);
3578         struct ext4_prealloc_space *pa;
3579         struct ext4_group_info *grp;
3580         struct ext4_inode_info *ei;
3581 
3582         /* preallocate only when found space is larger then requested */
3583         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3584         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3585         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3586 
3587         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3588         if (pa == NULL)
3589                 return -ENOMEM;
3590 
3591         if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3592                 int winl;
3593                 int wins;
3594                 int win;
3595                 int offs;
3596 
3597                 /* we can't allocate as much as normalizer wants.
3598                  * so, found space must get proper lstart
3599                  * to cover original request */
3600                 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3601                 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3602 
3603                 /* we're limited by original request in that
3604                  * logical block must be covered any way
3605                  * winl is window we can move our chunk within */
3606                 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3607 
3608                 /* also, we should cover whole original request */
3609                 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3610 
3611                 /* the smallest one defines real window */
3612                 win = min(winl, wins);
3613 
3614                 offs = ac->ac_o_ex.fe_logical %
3615                         EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3616                 if (offs && offs < win)
3617                         win = offs;
3618 
3619                 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3620                         EXT4_NUM_B2C(sbi, win);
3621                 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3622                 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3623         }
3624 
3625         /* preallocation can change ac_b_ex, thus we store actually
3626          * allocated blocks for history */
3627         ac->ac_f_ex = ac->ac_b_ex;
3628 
3629         pa->pa_lstart = ac->ac_b_ex.fe_logical;
3630         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3631         pa->pa_len = ac->ac_b_ex.fe_len;
3632         pa->pa_free = pa->pa_len;
3633         atomic_set(&pa->pa_count, 1);
3634         spin_lock_init(&pa->pa_lock);
3635         INIT_LIST_HEAD(&pa->pa_inode_list);
3636         INIT_LIST_HEAD(&pa->pa_group_list);
3637         pa->pa_deleted = 0;
3638         pa->pa_type = MB_INODE_PA;
3639 
3640         mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3641                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3642         trace_ext4_mb_new_inode_pa(ac, pa);
3643 
3644         ext4_mb_use_inode_pa(ac, pa);
3645         atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3646 
3647         ei = EXT4_I(ac->ac_inode);
3648         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3649 
3650         pa->pa_obj_lock = &ei->i_prealloc_lock;
3651         pa->pa_inode = ac->ac_inode;
3652 
3653         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3654         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3655         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3656 
3657         spin_lock(pa->pa_obj_lock);
3658         list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3659         spin_unlock(pa->pa_obj_lock);
3660 
3661         return 0;
3662 }
3663 
3664 /*
3665  * creates new preallocated space for locality group inodes belongs to
3666  */
3667 static noinline_for_stack int
3668 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3669 {
3670         struct super_block *sb = ac->ac_sb;
3671         struct ext4_locality_group *lg;
3672         struct ext4_prealloc_space *pa;
3673         struct ext4_group_info *grp;
3674 
3675         /* preallocate only when found space is larger then requested */
3676         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3677         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3678         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3679 
3680         BUG_ON(ext4_pspace_cachep == NULL);
3681         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3682         if (pa == NULL)
3683                 return -ENOMEM;
3684 
3685         /* preallocation can change ac_b_ex, thus we store actually
3686          * allocated blocks for history */
3687         ac->ac_f_ex = ac->ac_b_ex;
3688 
3689         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3690         pa->pa_lstart = pa->pa_pstart;
3691         pa->pa_len = ac->ac_b_ex.fe_len;
3692         pa->pa_free = pa->pa_len;
3693         atomic_set(&pa->pa_count, 1);
3694         spin_lock_init(&pa->pa_lock);
3695         INIT_LIST_HEAD(&pa->pa_inode_list);
3696         INIT_LIST_HEAD(&pa->pa_group_list);
3697         pa->pa_deleted = 0;
3698         pa->pa_type = MB_GROUP_PA;
3699 
3700         mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3701                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3702         trace_ext4_mb_new_group_pa(ac, pa);
3703 
3704         ext4_mb_use_group_pa(ac, pa);
3705         atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3706 
3707         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3708         lg = ac->ac_lg;
3709         BUG_ON(lg == NULL);
3710 
3711         pa->pa_obj_lock = &lg->lg_prealloc_lock;
3712         pa->pa_inode = NULL;
3713 
3714         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3715         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3716         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3717 
3718         /*
3719          * We will later add the new pa to the right bucket
3720          * after updating the pa_free in ext4_mb_release_context
3721          */
3722         return 0;
3723 }
3724 
3725 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3726 {
3727         int err;
3728 
3729         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3730                 err = ext4_mb_new_group_pa(ac);
3731         else
3732                 err = ext4_mb_new_inode_pa(ac);
3733         return err;
3734 }
3735 
3736 /*
3737  * finds all unused blocks in on-disk bitmap, frees them in
3738  * in-core bitmap and buddy.
3739  * @pa must be unlinked from inode and group lists, so that
3740  * nobody else can find/use it.
3741  * the caller MUST hold group/inode locks.
3742  * TODO: optimize the case when there are no in-core structures yet
3743  */
3744 static noinline_for_stack int
3745 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3746                         struct ext4_prealloc_space *pa)
3747 {
3748         struct super_block *sb = e4b->bd_sb;
3749         struct ext4_sb_info *sbi = EXT4_SB(sb);
3750         unsigned int end;
3751         unsigned int next;
3752         ext4_group_t group;
3753         ext4_grpblk_t bit;
3754         unsigned long long grp_blk_start;
3755         int err = 0;
3756         int free = 0;
3757 
3758         BUG_ON(pa->pa_deleted == 0);
3759         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3760         grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3761         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3762         end = bit + pa->pa_len;
3763 
3764         while (bit < end) {
3765                 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3766                 if (bit >= end)
3767                         break;
3768                 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3769                 mb_debug(1, "    free preallocated %u/%u in group %u\n",
3770                          (unsigned) ext4_group_first_block_no(sb, group) + bit,
3771                          (unsigned) next - bit, (unsigned) group);
3772                 free += next - bit;
3773 
3774                 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3775                 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3776                                                     EXT4_C2B(sbi, bit)),
3777                                                next - bit);
3778                 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3779                 bit = next + 1;
3780         }
3781         if (free != pa->pa_free) {
3782                 ext4_msg(e4b->bd_sb, KERN_CRIT,
3783                          "pa %p: logic %lu, phys. %lu, len %lu",
3784                          pa, (unsigned long) pa->pa_lstart,
3785                          (unsigned long) pa->pa_pstart,
3786                          (unsigned long) pa->pa_len);
3787                 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3788                                         free, pa->pa_free);
3789                 /*
3790                  * pa is already deleted so we use the value obtained
3791                  * from the bitmap and continue.
3792                  */
3793         }
3794         atomic_add(free, &sbi->s_mb_discarded);
3795 
3796         return err;
3797 }
3798 
3799 static noinline_for_stack int
3800 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3801                                 struct ext4_prealloc_space *pa)
3802 {
3803         struct super_block *sb = e4b->bd_sb;
3804         ext4_group_t group;
3805         ext4_grpblk_t bit;
3806 
3807         trace_ext4_mb_release_group_pa(sb, pa);
3808         BUG_ON(pa->pa_deleted == 0);
3809         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3810         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3811         mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3812         atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3813         trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3814 
3815         return 0;
3816 }
3817 
3818 /*
3819  * releases all preallocations in given group
3820  *
3821  * first, we need to decide discard policy:
3822  * - when do we discard
3823  *   1) ENOSPC
3824  * - how many do we discard
3825  *   1) how many requested
3826  */
3827 static noinline_for_stack int
3828 ext4_mb_discard_group_preallocations(struct super_block *sb,
3829                                         ext4_group_t group, int needed)
3830 {
3831         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3832         struct buffer_head *bitmap_bh = NULL;
3833         struct ext4_prealloc_space *pa, *tmp;
3834         struct list_head list;
3835         struct ext4_buddy e4b;
3836         int err;
3837         int busy = 0;
3838         int free = 0;
3839 
3840         mb_debug(1, "discard preallocation for group %u\n", group);
3841 
3842         if (list_empty(&grp->bb_prealloc_list))
3843                 return 0;
3844 
3845         bitmap_bh = ext4_read_block_bitmap(sb, group);
3846         if (bitmap_bh == NULL) {
3847                 ext4_error(sb, "Error reading block bitmap for %u", group);
3848                 return 0;
3849         }
3850 
3851         err = ext4_mb_load_buddy(sb, group, &e4b);
3852         if (err) {
3853                 ext4_error(sb, "Error loading buddy information for %u", group);
3854                 put_bh(bitmap_bh);
3855                 return 0;
3856         }
3857 
3858         if (needed == 0)
3859                 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3860 
3861         INIT_LIST_HEAD(&list);
3862 repeat:
3863         ext4_lock_group(sb, group);
3864         list_for_each_entry_safe(pa, tmp,
3865                                 &grp->bb_prealloc_list, pa_group_list) {
3866                 spin_lock(&pa->pa_lock);
3867                 if (atomic_read(&pa->pa_count)) {
3868                         spin_unlock(&pa->pa_lock);
3869                         busy = 1;
3870                         continue;
3871                 }
3872                 if (pa->pa_deleted) {
3873                         spin_unlock(&pa->pa_lock);
3874                         continue;
3875                 }
3876 
3877                 /* seems this one can be freed ... */
3878                 pa->pa_deleted = 1;
3879 
3880                 /* we can trust pa_free ... */
3881                 free += pa->pa_free;
3882 
3883                 spin_unlock(&pa->pa_lock);
3884 
3885                 list_del(&pa->pa_group_list);
3886                 list_add(&pa->u.pa_tmp_list, &list);
3887         }
3888 
3889         /* if we still need more blocks and some PAs were used, try again */
3890         if (free < needed && busy) {
3891                 busy = 0;
3892                 ext4_unlock_group(sb, group);
3893                 cond_resched();
3894                 goto repeat;
3895         }
3896 
3897         /* found anything to free? */
3898         if (list_empty(&list)) {
3899                 BUG_ON(free != 0);
3900                 goto out;
3901         }
3902 
3903         /* now free all selected PAs */
3904         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3905 
3906                 /* remove from object (inode or locality group) */
3907                 spin_lock(pa->pa_obj_lock);
3908                 list_del_rcu(&pa->pa_inode_list);
3909                 spin_unlock(pa->pa_obj_lock);
3910 
3911                 if (pa->pa_type == MB_GROUP_PA)
3912                         ext4_mb_release_group_pa(&e4b, pa);
3913                 else
3914                         ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3915 
3916                 list_del(&pa->u.pa_tmp_list);
3917                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3918         }
3919 
3920 out:
3921         ext4_unlock_group(sb, group);
3922         ext4_mb_unload_buddy(&e4b);
3923         put_bh(bitmap_bh);
3924         return free;
3925 }
3926 
3927 /*
3928  * releases all non-used preallocated blocks for given inode
3929  *
3930  * It's important to discard preallocations under i_data_sem
3931  * We don't want another block to be served from the prealloc
3932  * space when we are discarding the inode prealloc space.
3933  *
3934  * FIXME!! Make sure it is valid at all the call sites
3935  */
3936 void ext4_discard_preallocations(struct inode *inode)
3937 {
3938         struct ext4_inode_info *ei = EXT4_I(inode);
3939         struct super_block *sb = inode->i_sb;
3940         struct buffer_head *bitmap_bh = NULL;
3941         struct ext4_prealloc_space *pa, *tmp;
3942         ext4_group_t group = 0;
3943         struct list_head list;
3944         struct ext4_buddy e4b;
3945         int err;
3946 
3947         if (!S_ISREG(inode->i_mode)) {
3948                 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3949                 return;
3950         }
3951 
3952         mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3953         trace_ext4_discard_preallocations(inode);
3954 
3955         INIT_LIST_HEAD(&list);
3956 
3957 repeat:
3958         /* first, collect all pa's in the inode */
3959         spin_lock(&ei->i_prealloc_lock);
3960         while (!list_empty(&ei->i_prealloc_list)) {
3961                 pa = list_entry(ei->i_prealloc_list.next,
3962                                 struct ext4_prealloc_space, pa_inode_list);
3963                 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3964                 spin_lock(&pa->pa_lock);
3965                 if (atomic_read(&pa->pa_count)) {
3966                         /* this shouldn't happen often - nobody should
3967                          * use preallocation while we're discarding it */
3968                         spin_unlock(&pa->pa_lock);
3969                         spin_unlock(&ei->i_prealloc_lock);
3970                         ext4_msg(sb, KERN_ERR,
3971                                  "uh-oh! used pa while discarding");
3972                         WARN_ON(1);
3973                         schedule_timeout_uninterruptible(HZ);
3974                         goto repeat;
3975 
3976                 }
3977                 if (pa->pa_deleted == 0) {
3978                         pa->pa_deleted = 1;
3979                         spin_unlock(&pa->pa_lock);
3980                         list_del_rcu(&pa->pa_inode_list);
3981                         list_add(&pa->u.pa_tmp_list, &list);
3982                         continue;
3983                 }
3984 
3985                 /* someone is deleting pa right now */
3986                 spin_unlock(&pa->pa_lock);
3987                 spin_unlock(&ei->i_prealloc_lock);
3988 
3989                 /* we have to wait here because pa_deleted
3990                  * doesn't mean pa is already unlinked from
3991                  * the list. as we might be called from
3992                  * ->clear_inode() the inode will get freed
3993                  * and concurrent thread which is unlinking
3994                  * pa from inode's list may access already
3995                  * freed memory, bad-bad-bad */
3996 
3997                 /* XXX: if this happens too often, we can
3998                  * add a flag to force wait only in case
3999                  * of ->clear_inode(), but not in case of
4000                  * regular truncate */
4001                 schedule_timeout_uninterruptible(HZ);
4002                 goto repeat;
4003         }
4004         spin_unlock(&ei->i_prealloc_lock);
4005 
4006         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4007                 BUG_ON(pa->pa_type != MB_INODE_PA);
4008                 group = ext4_get_group_number(sb, pa->pa_pstart);
4009 
4010                 err = ext4_mb_load_buddy(sb, group, &e4b);
4011                 if (err) {
4012                         ext4_error(sb, "Error loading buddy information for %u",
4013                                         group);
4014                         continue;
4015                 }
4016 
4017                 bitmap_bh = ext4_read_block_bitmap(sb, group);
4018                 if (bitmap_bh == NULL) {
4019                         ext4_error(sb, "Error reading block bitmap for %u",
4020                                         group);
4021                         ext4_mb_unload_buddy(&e4b);
4022                         continue;
4023                 }
4024 
4025                 ext4_lock_group(sb, group);
4026                 list_del(&pa->pa_group_list);
4027                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4028                 ext4_unlock_group(sb, group);
4029 
4030                 ext4_mb_unload_buddy(&e4b);
4031                 put_bh(bitmap_bh);
4032 
4033                 list_del(&pa->u.pa_tmp_list);
4034                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4035         }
4036 }
4037 
4038 #ifdef CONFIG_EXT4_DEBUG
4039 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4040 {
4041         struct super_block *sb = ac->ac_sb;
4042         ext4_group_t ngroups, i;
4043 
4044         if (!ext4_mballoc_debug ||
4045             (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4046                 return;
4047 
4048         ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4049                         " Allocation context details:");
4050         ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4051                         ac->ac_status, ac->ac_flags);
4052         ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4053                         "goal %lu/%lu/%lu@%lu, "
4054                         "best %lu/%lu/%lu@%lu cr %d",
4055                         (unsigned long)ac->ac_o_ex.fe_group,
4056                         (unsigned long)ac->ac_o_ex.fe_start,
4057                         (unsigned long)ac->ac_o_ex.fe_len,
4058                         (unsigned long)ac->ac_o_ex.fe_logical,
4059                         (unsigned long)ac->ac_g_ex.fe_group,
4060                         (unsigned long)ac->ac_g_ex.fe_start,
4061                         (unsigned long)ac->ac_g_ex.fe_len,
4062                         (unsigned long)ac->ac_g_ex.fe_logical,
4063                         (unsigned long)ac->ac_b_ex.fe_group,
4064                         (unsigned long)ac->ac_b_ex.fe_start,
4065                         (unsigned long)ac->ac_b_ex.fe_len,
4066                         (unsigned long)ac->ac_b_ex.fe_logical,
4067                         (int)ac->ac_criteria);
4068         ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4069         ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4070         ngroups = ext4_get_groups_count(sb);
4071         for (i = 0; i < ngroups; i++) {
4072                 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4073                 struct ext4_prealloc_space *pa;
4074                 ext4_grpblk_t start;
4075                 struct list_head *cur;
4076                 ext4_lock_group(sb, i);
4077                 list_for_each(cur, &grp->bb_prealloc_list) {
4078                         pa = list_entry(cur, struct ext4_prealloc_space,
4079                                         pa_group_list);
4080                         spin_lock(&pa->pa_lock);
4081                         ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4082                                                      NULL, &start);
4083                         spin_unlock(&pa->pa_lock);
4084                         printk(KERN_ERR "PA:%u:%d:%u \n", i,
4085                                start, pa->pa_len);
4086                 }
4087                 ext4_unlock_group(sb, i);
4088 
4089                 if (grp->bb_free == 0)
4090                         continue;
4091                 printk(KERN_ERR "%u: %d/%d \n",
4092                        i, grp->bb_free, grp->bb_fragments);
4093         }
4094         printk(KERN_ERR "\n");
4095 }
4096 #else
4097 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4098 {
4099         return;
4100 }
4101 #endif
4102 
4103 /*
4104  * We use locality group preallocation for small size file. The size of the
4105  * file is determined by the current size or the resulting size after
4106  * allocation which ever is larger
4107  *
4108  * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4109  */
4110 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4111 {
4112         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4113         int bsbits = ac->ac_sb->s_blocksize_bits;
4114         loff_t size, isize;
4115 
4116         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4117                 return;
4118 
4119         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4120                 return;
4121 
4122         size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4123         isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4124                 >> bsbits;
4125 
4126         if ((size == isize) &&
4127             !ext4_fs_is_busy(sbi) &&
4128             (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4129                 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4130                 return;
4131         }
4132 
4133         if (sbi->s_mb_group_prealloc <= 0) {
4134                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4135                 return;
4136         }
4137 
4138         /* don't use group allocation for large files */
4139         size = max(size, isize);
4140         if (size > sbi->s_mb_stream_request) {
4141                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4142                 return;
4143         }
4144 
4145         BUG_ON(ac->ac_lg != NULL);
4146         /*
4147          * locality group prealloc space are per cpu. The reason for having
4148          * per cpu locality group is to reduce the contention between block
4149          * request from multiple CPUs.
4150          */
4151         ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4152 
4153         /* we're going to use group allocation */
4154         ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4155 
4156         /* serialize all allocations in the group */
4157         mutex_lock(&ac->ac_lg->lg_mutex);
4158 }
4159 
4160 static noinline_for_stack int
4161 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4162                                 struct ext4_allocation_request *ar)
4163 {
4164         struct super_block *sb = ar->inode->i_sb;
4165         struct ext4_sb_info *sbi = EXT4_SB(sb);
4166         struct ext4_super_block *es = sbi->s_es;
4167         ext4_group_t group;
4168         unsigned int len;
4169         ext4_fsblk_t goal;
4170         ext4_grpblk_t block;
4171 
4172         /* we can't allocate > group size */
4173         len = ar->len;
4174 
4175         /* just a dirty hack to filter too big requests  */
4176         if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4177                 len = EXT4_CLUSTERS_PER_GROUP(sb);
4178 
4179         /* start searching from the goal */
4180         goal = ar->goal;
4181         if (goal < le32_to_cpu(es->s_first_data_block) ||
4182                         goal >= ext4_blocks_count(es))
4183                 goal = le32_to_cpu(es->s_first_data_block);
4184         ext4_get_group_no_and_offset(sb, goal, &group, &block);
4185 
4186         /* set up allocation goals */
4187         ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4188         ac->ac_status = AC_STATUS_CONTINUE;
4189         ac->ac_sb = sb;
4190         ac->ac_inode = ar->inode;
4191         ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4192         ac->ac_o_ex.fe_group = group;
4193         ac->ac_o_ex.fe_start = block;
4194         ac->ac_o_ex.fe_len = len;
4195         ac->ac_g_ex = ac->ac_o_ex;
4196         ac->ac_flags = ar->flags;
4197 
4198         /* we have to define context: we'll we work with a file or
4199          * locality group. this is a policy, actually */
4200         ext4_mb_group_or_file(ac);
4201 
4202         mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4203                         "left: %u/%u, right %u/%u to %swritable\n",
4204                         (unsigned) ar->len, (unsigned) ar->logical,
4205                         (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4206                         (unsigned) ar->lleft, (unsigned) ar->pleft,
4207                         (unsigned) ar->lright, (unsigned) ar->pright,
4208                         atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4209         return 0;
4210 
4211 }
4212 
4213 static noinline_for_stack void
4214 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4215                                         struct ext4_locality_group *lg,
4216                                         int order, int total_entries)
4217 {
4218         ext4_group_t group = 0;
4219         struct ext4_buddy e4b;
4220         struct list_head discard_list;
4221         struct ext4_prealloc_space *pa, *tmp;
4222 
4223         mb_debug(1, "discard locality group preallocation\n");
4224 
4225         INIT_LIST_HEAD(&discard_list);
4226 
4227         spin_lock(&lg->lg_prealloc_lock);
4228         list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4229                                                 pa_inode_list) {
4230                 spin_lock(&pa->pa_lock);
4231                 if (atomic_read(&pa->pa_count)) {
4232                         /*
4233                          * This is the pa that we just used
4234                          * for block allocation. So don't
4235                          * free that
4236                          */
4237                         spin_unlock(&pa->pa_lock);
4238                         continue;
4239                 }
4240                 if (pa->pa_deleted) {
4241                         spin_unlock(&pa->pa_lock);
4242                         continue;
4243                 }
4244                 /* only lg prealloc space */
4245                 BUG_ON(pa->pa_type != MB_GROUP_PA);
4246 
4247                 /* seems this one can be freed ... */
4248                 pa->pa_deleted = 1;
4249                 spin_unlock(&pa->pa_lock);
4250 
4251                 list_del_rcu(&pa->pa_inode_list);
4252                 list_add(&pa->u.pa_tmp_list, &discard_list);
4253 
4254                 total_entries--;
4255                 if (total_entries <= 5) {
4256                         /*
4257                          * we want to keep only 5 entries
4258                          * allowing it to grow to 8. This
4259                          * mak sure we don't call discard
4260                          * soon for this list.
4261                          */
4262                         break;
4263                 }
4264         }
4265         spin_unlock(&lg->lg_prealloc_lock);
4266 
4267         list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4268 
4269                 group = ext4_get_group_number(sb, pa->pa_pstart);
4270                 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4271                         ext4_error(sb, "Error loading buddy information for %u",
4272                                         group);
4273                         continue;
4274                 }
4275                 ext4_lock_group(sb, group);
4276                 list_del(&pa->pa_group_list);
4277                 ext4_mb_release_group_pa(&e4b, pa);
4278                 ext4_unlock_group(sb, group);
4279 
4280                 ext4_mb_unload_buddy(&e4b);
4281                 list_del(&pa->u.pa_tmp_list);
4282                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4283         }
4284 }
4285 
4286 /*
4287  * We have incremented pa_count. So it cannot be freed at this
4288  * point. Also we hold lg_mutex. So no parallel allocation is
4289  * possible from this lg. That means pa_free cannot be updated.
4290  *
4291  * A parallel ext4_mb_discard_group_preallocations is possible.
4292  * which can cause the lg_prealloc_list to be updated.
4293  */
4294 
4295 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4296 {
4297         int order, added = 0, lg_prealloc_count = 1;
4298         struct super_block *sb = ac->ac_sb;
4299         struct ext4_locality_group *lg = ac->ac_lg;
4300         struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4301 
4302         order = fls(pa->pa_free) - 1;
4303         if (order > PREALLOC_TB_SIZE - 1)
4304                 /* The max size of hash table is PREALLOC_TB_SIZE */
4305                 order = PREALLOC_TB_SIZE - 1;
4306         /* Add the prealloc space to lg */
4307         spin_lock(&lg->lg_prealloc_lock);
4308         list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4309                                                 pa_inode_list) {
4310                 spin_lock(&tmp_pa->pa_lock);
4311                 if (tmp_pa->pa_deleted) {
4312                         spin_unlock(&tmp_pa->pa_lock);
4313                         continue;
4314                 }
4315                 if (!added && pa->pa_free < tmp_pa->pa_free) {
4316                         /* Add to the tail of the previous entry */
4317                         list_add_tail_rcu(&pa->pa_inode_list,
4318                                                 &tmp_pa->pa_inode_list);
4319                         added = 1;
4320                         /*
4321                          * we want to count the total
4322                          * number of entries in the list
4323                          */
4324                 }
4325                 spin_unlock(&tmp_pa->pa_lock);
4326                 lg_prealloc_count++;
4327         }
4328         if (!added)
4329                 list_add_tail_rcu(&pa->pa_inode_list,
4330                                         &lg->lg_prealloc_list[order]);
4331         spin_unlock(&lg->lg_prealloc_lock);
4332 
4333         /* Now trim the list to be not more than 8 elements */
4334         if (lg_prealloc_count > 8) {
4335                 ext4_mb_discard_lg_preallocations(sb, lg,
4336                                                   order, lg_prealloc_count);
4337                 return;
4338         }
4339         return ;
4340 }
4341 
4342 /*
4343  * release all resource we used in allocation
4344  */
4345 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4346 {
4347         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4348         struct ext4_prealloc_space *pa = ac->ac_pa;
4349         if (pa) {
4350                 if (pa->pa_type == MB_GROUP_PA) {
4351                         /* see comment in ext4_mb_use_group_pa() */
4352                         spin_lock(&pa->pa_lock);
4353                         pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4354                         pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4355                         pa->pa_free -= ac->ac_b_ex.fe_len;
4356                         pa->pa_len -= ac->ac_b_ex.fe_len;
4357                         spin_unlock(&pa->pa_lock);
4358                 }
4359         }
4360         if (pa) {
4361                 /*
4362                  * We want to add the pa to the right bucket.
4363                  * Remove it from the list and while adding
4364                  * make sure the list to which we are adding
4365                  * doesn't grow big.
4366                  */
4367                 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4368                         spin_lock(pa->pa_obj_lock);
4369                         list_del_rcu(&pa->pa_inode_list);
4370                         spin_unlock(pa->pa_obj_lock);
4371                         ext4_mb_add_n_trim(ac);
4372                 }
4373                 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4374         }
4375         if (ac->ac_bitmap_page)
4376                 page_cache_release(ac->ac_bitmap_page);
4377         if (ac->ac_buddy_page)
4378                 page_cache_release(ac->ac_buddy_page);
4379         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4380                 mutex_unlock(&ac->ac_lg->lg_mutex);
4381         ext4_mb_collect_stats(ac);
4382         return 0;
4383 }
4384 
4385 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4386 {
4387         ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4388         int ret;
4389         int freed = 0;
4390 
4391         trace_ext4_mb_discard_preallocations(sb, needed);
4392         for (i = 0; i < ngroups && needed > 0; i++) {
4393                 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4394                 freed += ret;
4395                 needed -= ret;
4396         }
4397 
4398         return freed;
4399 }
4400 
4401 /*
4402  * Main entry point into mballoc to allocate blocks
4403  * it tries to use preallocation first, then falls back
4404  * to usual allocation
4405  */
4406 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4407                                 struct ext4_allocation_request *ar, int *errp)
4408 {
4409         int freed;
4410         struct ext4_allocation_context *ac = NULL;
4411         struct ext4_sb_info *sbi;
4412         struct super_block *sb;
4413         ext4_fsblk_t block = 0;
4414         unsigned int inquota = 0;
4415         unsigned int reserv_clstrs = 0;
4416 
4417         might_sleep();
4418         sb = ar->inode->i_sb;
4419         sbi = EXT4_SB(sb);
4420 
4421         trace_ext4_request_blocks(ar);
4422 
4423         /* Allow to use superuser reservation for quota file */
4424         if (IS_NOQUOTA(ar->inode))
4425                 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4426 
4427         if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4428                 /* Without delayed allocation we need to verify
4429                  * there is enough free blocks to do block allocation
4430                  * and verify allocation doesn't exceed the quota limits.
4431                  */
4432                 while (ar->len &&
4433                         ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4434 
4435                         /* let others to free the space */
4436                         cond_resched();
4437                         ar->len = ar->len >> 1;
4438                 }
4439                 if (!ar->len) {
4440                         *errp = -ENOSPC;
4441                         return 0;
4442                 }
4443                 reserv_clstrs = ar->len;
4444                 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4445                         dquot_alloc_block_nofail(ar->inode,
4446                                                  EXT4_C2B(sbi, ar->len));
4447                 } else {
4448                         while (ar->len &&
4449                                 dquot_alloc_block(ar->inode,
4450                                                   EXT4_C2B(sbi, ar->len))) {
4451 
4452                                 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4453                                 ar->len--;
4454                         }
4455                 }
4456                 inquota = ar->len;
4457                 if (ar->len == 0) {
4458                         *errp = -EDQUOT;
4459                         goto out;
4460                 }
4461         }
4462 
4463         ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4464         if (!ac) {
4465                 ar->len = 0;
4466                 *errp = -ENOMEM;
4467                 goto out;
4468         }
4469 
4470         *errp = ext4_mb_initialize_context(ac, ar);
4471         if (*errp) {
4472                 ar->len = 0;
4473                 goto out;
4474         }
4475 
4476         ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4477         if (!ext4_mb_use_preallocated(ac)) {
4478                 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4479                 ext4_mb_normalize_request(ac, ar);
4480 repeat:
4481                 /* allocate space in core */
4482                 *errp = ext4_mb_regular_allocator(ac);
4483                 if (*errp)
4484                         goto discard_and_exit;
4485 
4486                 /* as we've just preallocated more space than
4487                  * user requested originally, we store allocated
4488                  * space in a special descriptor */
4489                 if (ac->ac_status == AC_STATUS_FOUND &&
4490                     ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4491                         *errp = ext4_mb_new_preallocation(ac);
4492                 if (*errp) {
4493                 discard_and_exit:
4494                         ext4_discard_allocated_blocks(ac);
4495                         goto errout;
4496                 }
4497         }
4498         if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4499                 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4500                 if (*errp == -EAGAIN) {
4501                         /*
4502                          * drop the reference that we took
4503                          * in ext4_mb_use_best_found
4504                          */
4505                         ext4_mb_release_context(ac);
4506                         ac->ac_b_ex.fe_group = 0;
4507                         ac->ac_b_ex.fe_start = 0;
4508                         ac->ac_b_ex.fe_len = 0;
4509                         ac->ac_status = AC_STATUS_CONTINUE;
4510                         goto repeat;
4511                 } else if (*errp) {
4512                         ext4_discard_allocated_blocks(ac);
4513                         goto errout;
4514                 } else {
4515                         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4516                         ar->len = ac->ac_b_ex.fe_len;
4517                 }
4518         } else {
4519                 freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4520                 if (freed)
4521                         goto repeat;
4522                 *errp = -ENOSPC;
4523         }
4524 
4525 errout:
4526         if (*errp) {
4527                 ac->ac_b_ex.fe_len = 0;
4528                 ar->len = 0;
4529                 ext4_mb_show_ac(ac);
4530         }
4531         ext4_mb_release_context(ac);
4532 out:
4533         if (ac)
4534                 kmem_cache_free(ext4_ac_cachep, ac);
4535         if (inquota && ar->len < inquota)
4536                 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4537         if (!ar->len) {
4538                 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4539                         /* release all the reserved blocks if non delalloc */
4540                         percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4541                                                 reserv_clstrs);
4542         }
4543 
4544         trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4545 
4546         return block;
4547 }
4548 
4549 /*
4550  * We can merge two free data extents only if the physical blocks
4551  * are contiguous, AND the extents were freed by the same transaction,
4552  * AND the blocks are associated with the same group.
4553  */
4554 static int can_merge(struct ext4_free_data *entry1,
4555                         struct ext4_free_data *entry2)
4556 {
4557         if ((entry1->efd_tid == entry2->efd_tid) &&
4558             (entry1->efd_group == entry2->efd_group) &&
4559             ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4560                 return 1;
4561         return 0;
4562 }
4563 
4564 static noinline_for_stack int
4565 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4566                       struct ext4_free_data *new_entry)
4567 {
4568         ext4_group_t group = e4b->bd_group;
4569         ext4_grpblk_t cluster;
4570         struct ext4_free_data *entry;
4571         struct ext4_group_info *db = e4b->bd_info;
4572         struct super_block *sb = e4b->bd_sb;
4573         struct ext4_sb_info *sbi = EXT4_SB(sb);
4574         struct rb_node **n = &db->bb_free_root.rb_node, *node;
4575         struct rb_node *parent = NULL, *new_node;
4576 
4577         BUG_ON(!ext4_handle_valid(handle));
4578         BUG_ON(e4b->bd_bitmap_page == NULL);
4579         BUG_ON(e4b->bd_buddy_page == NULL);
4580 
4581         new_node = &new_entry->efd_node;
4582         cluster = new_entry->efd_start_cluster;
4583 
4584         if (!*n) {
4585                 /* first free block exent. We need to
4586                    protect buddy cache from being freed,
4587                  * otherwise we'll refresh it from
4588                  * on-disk bitmap and lose not-yet-available
4589                  * blocks */
4590                 page_cache_get(e4b->bd_buddy_page);
4591                 page_cache_get(e4b->bd_bitmap_page);
4592         }
4593         while (*n) {
4594                 parent = *n;
4595                 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4596                 if (cluster < entry->efd_start_cluster)
4597                         n = &(*n)->rb_left;
4598                 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4599                         n = &(*n)->rb_right;
4600                 else {
4601                         ext4_grp_locked_error(sb, group, 0,
4602                                 ext4_group_first_block_no(sb, group) +
4603                                 EXT4_C2B(sbi, cluster),
4604                                 "Block already on to-be-freed list");
4605                         return 0;
4606                 }
4607         }
4608 
4609         rb_link_node(new_node, parent, n);
4610         rb_insert_color(new_node, &db->bb_free_root);
4611 
4612         /* Now try to see the extent can be merged to left and right */
4613         node = rb_prev(new_node);
4614         if (node) {
4615                 entry = rb_entry(node, struct ext4_free_data, efd_node);
4616                 if (can_merge(entry, new_entry) &&
4617                     ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4618                         new_entry->efd_start_cluster = entry->efd_start_cluster;
4619                         new_entry->efd_count += entry->efd_count;
4620                         rb_erase(node, &(db->bb_free_root));
4621                         kmem_cache_free(ext4_free_data_cachep, entry);
4622                 }
4623         }
4624 
4625         node = rb_next(new_node);
4626         if (node) {
4627                 entry = rb_entry(node, struct ext4_free_data, efd_node);
4628                 if (can_merge(new_entry, entry) &&
4629                     ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4630                         new_entry->efd_count += entry->efd_count;
4631                         rb_erase(node, &(db->bb_free_root));
4632                         kmem_cache_free(ext4_free_data_cachep, entry);
4633                 }
4634         }
4635         /* Add the extent to transaction's private list */
4636         ext4_journal_callback_add(handle, ext4_free_data_callback,
4637                                   &new_entry->efd_jce);
4638         return 0;
4639 }
4640 
4641 /**
4642  * ext4_free_blocks() -- Free given blocks and update quota
4643  * @handle:             handle for this transaction
4644  * @inode:              inode
4645  * @block:              start physical block to free
4646  * @count:              number of blocks to count
4647  * @flags:              flags used by ext4_free_blocks
4648  */
4649 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4650                       struct buffer_head *bh, ext4_fsblk_t block,
4651                       unsigned long count, int flags)
4652 {
4653         struct buffer_head *bitmap_bh = NULL;
4654         struct super_block *sb = inode->i_sb;
4655         struct ext4_group_desc *gdp;
4656         unsigned int overflow;
4657         ext4_grpblk_t bit;
4658         struct buffer_head *gd_bh;
4659         ext4_group_t block_group;
4660         struct ext4_sb_info *sbi;
4661         struct ext4_buddy e4b;
4662         unsigned int count_clusters;
4663         int err = 0;
4664         int ret;
4665 
4666         might_sleep();
4667         if (bh) {
4668                 if (block)
4669                         BUG_ON(block != bh->b_blocknr);
4670                 else
4671                         block = bh->b_blocknr;
4672         }
4673 
4674         sbi = EXT4_SB(sb);
4675         if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4676             !ext4_data_block_valid(sbi, block, count)) {
4677                 ext4_error(sb, "Freeing blocks not in datazone - "
4678                            "block = %llu, count = %lu", block, count);
4679                 goto error_return;
4680         }
4681 
4682         ext4_debug("freeing block %llu\n", block);
4683         trace_ext4_free_blocks(inode, block, count, flags);
4684 
4685         if (flags & EXT4_FREE_BLOCKS_FORGET) {
4686                 struct buffer_head *tbh = bh;
4687                 int i;
4688 
4689                 BUG_ON(bh && (count > 1));
4690 
4691                 for (i = 0; i < count; i++) {
4692                         cond_resched();
4693                         if (!bh)
4694                                 tbh = sb_find_get_block(inode->i_sb,
4695                                                         block + i);
4696                         if (!tbh)
4697                                 continue;
4698                         ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4699                                     inode, tbh, block + i);
4700                 }
4701         }
4702 
4703         /*
4704          * We need to make sure we don't reuse the freed block until
4705          * after the transaction is committed, which we can do by
4706          * treating the block as metadata, below.  We make an
4707          * exception if the inode is to be written in writeback mode
4708          * since writeback mode has weak data consistency guarantees.
4709          */
4710         if (!ext4_should_writeback_data(inode))
4711                 flags |= EXT4_FREE_BLOCKS_METADATA;
4712 
4713         /*
4714          * If the extent to be freed does not begin on a cluster
4715          * boundary, we need to deal with partial clusters at the
4716          * beginning and end of the extent.  Normally we will free
4717          * blocks at the beginning or the end unless we are explicitly
4718          * requested to avoid doing so.
4719          */
4720         overflow = EXT4_PBLK_COFF(sbi, block);
4721         if (overflow) {
4722                 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4723                         overflow = sbi->s_cluster_ratio - overflow;
4724                         block += overflow;
4725                         if (count > overflow)
4726                                 count -= overflow;
4727                         else
4728                                 return;
4729                 } else {
4730                         block -= overflow;
4731                         count += overflow;
4732                 }
4733         }
4734         overflow = EXT4_LBLK_COFF(sbi, count);
4735         if (overflow) {
4736                 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4737                         if (count > overflow)
4738                                 count -= overflow;
4739                         else
4740                                 return;
4741                 } else
4742                         count += sbi->s_cluster_ratio - overflow;
4743         }
4744 
4745 do_more:
4746         overflow = 0;
4747         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4748 
4749         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4750                         ext4_get_group_info(sb, block_group))))
4751                 return;
4752 
4753         /*
4754          * Check to see if we are freeing blocks across a group
4755          * boundary.
4756          */
4757         if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4758                 overflow = EXT4_C2B(sbi, bit) + count -
4759                         EXT4_BLOCKS_PER_GROUP(sb);
4760                 count -= overflow;
4761         }
4762         count_clusters = EXT4_NUM_B2C(sbi, count);
4763         bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4764         if (!bitmap_bh) {
4765                 err = -EIO;
4766                 goto error_return;
4767         }
4768         gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4769         if (!gdp) {
4770                 err = -EIO;
4771                 goto error_return;
4772         }
4773 
4774         if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4775             in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4776             in_range(block, ext4_inode_table(sb, gdp),
4777                      EXT4_SB(sb)->s_itb_per_group) ||
4778             in_range(block + count - 1, ext4_inode_table(sb, gdp),
4779                      EXT4_SB(sb)->s_itb_per_group)) {
4780 
4781                 ext4_error(sb, "Freeing blocks in system zone - "
4782                            "Block = %llu, count = %lu", block, count);
4783                 /* err = 0. ext4_std_error should be a no op */
4784                 goto error_return;
4785         }
4786 
4787         BUFFER_TRACE(bitmap_bh, "getting write access");
4788         err = ext4_journal_get_write_access(handle, bitmap_bh);
4789         if (err)
4790                 goto error_return;
4791 
4792         /*
4793          * We are about to modify some metadata.  Call the journal APIs
4794          * to unshare ->b_data if a currently-committing transaction is
4795          * using it
4796          */
4797         BUFFER_TRACE(gd_bh, "get_write_access");
4798         err = ext4_journal_get_write_access(handle, gd_bh);
4799         if (err)
4800                 goto error_return;
4801 #ifdef AGGRESSIVE_CHECK
4802         {
4803                 int i;
4804                 for (i = 0; i < count_clusters; i++)
4805                         BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4806         }
4807 #endif
4808         trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4809 
4810         err = ext4_mb_load_buddy(sb, block_group, &e4b);
4811         if (err)
4812                 goto error_return;
4813 
4814         if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4815                 struct ext4_free_data *new_entry;
4816                 /*
4817                  * blocks being freed are metadata. these blocks shouldn't
4818                  * be used until this transaction is committed
4819                  *
4820                  * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4821                  * to fail.
4822                  */
4823                 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4824                                 GFP_NOFS|__GFP_NOFAIL);
4825                 new_entry->efd_start_cluster = bit;
4826                 new_entry->efd_group = block_group;
4827                 new_entry->efd_count = count_clusters;
4828                 new_entry->efd_tid = handle->h_transaction->t_tid;
4829 
4830                 ext4_lock_group(sb, block_group);
4831                 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4832                 ext4_mb_free_metadata(handle, &e4b, new_entry);
4833         } else {
4834                 /* need to update group_info->bb_free and bitmap
4835                  * with group lock held. generate_buddy look at
4836                  * them with group lock_held
4837                  */
4838                 if (test_opt(sb, DISCARD)) {
4839                         err = ext4_issue_discard(sb, block_group, bit, count);
4840                         if (err && err != -EOPNOTSUPP)
4841                                 ext4_msg(sb, KERN_WARNING, "discard request in"
4842                                          " group:%d block:%d count:%lu failed"
4843                                          " with %d", block_group, bit, count,
4844                                          err);
4845                 } else
4846                         EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4847 
4848                 ext4_lock_group(sb, block_group);
4849                 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4850                 mb_free_blocks(inode, &e4b, bit, count_clusters);
4851         }
4852 
4853         ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4854         ext4_free_group_clusters_set(sb, gdp, ret);
4855         ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4856         ext4_group_desc_csum_set(sb, block_group, gdp);
4857         ext4_unlock_group(sb, block_group);
4858 
4859         if (sbi->s_log_groups_per_flex) {
4860                 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4861                 atomic64_add(count_clusters,
4862                              &sbi->s_flex_groups[flex_group].free_clusters);
4863         }
4864 
4865         if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4866                 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4867         percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4868 
4869         ext4_mb_unload_buddy(&e4b);
4870 
4871         /* We dirtied the bitmap block */
4872         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4873         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4874 
4875         /* And the group descriptor block */
4876         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4877         ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4878         if (!err)
4879                 err = ret;
4880 
4881         if (overflow && !err) {
4882                 block += count;
4883                 count = overflow;
4884                 put_bh(bitmap_bh);
4885                 goto do_more;
4886         }
4887 error_return:
4888         brelse(bitmap_bh);
4889         ext4_std_error(sb, err);
4890         return;
4891 }
4892 
4893 /**
4894  * ext4_group_add_blocks() -- Add given blocks to an existing group
4895  * @handle:                     handle to this transaction
4896  * @sb:                         super block
4897  * @block:                      start physical block to add to the block group
4898  * @count:                      number of blocks to free
4899  *
4900  * This marks the blocks as free in the bitmap and buddy.
4901  */
4902 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4903                          ext4_fsblk_t block, unsigned long count)
4904 {
4905         struct buffer_head *bitmap_bh = NULL;
4906         struct buffer_head *gd_bh;
4907         ext4_group_t block_group;
4908         ext4_grpblk_t bit;
4909         unsigned int i;
4910         struct ext4_group_desc *desc;
4911         struct ext4_sb_info *sbi = EXT4_SB(sb);
4912         struct ext4_buddy e4b;
4913         int err = 0, ret, blk_free_count;
4914         ext4_grpblk_t blocks_freed;
4915 
4916         ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4917 
4918         if (count == 0)
4919                 return 0;
4920 
4921         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4922         /*
4923          * Check to see if we are freeing blocks across a group
4924          * boundary.
4925          */
4926         if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4927                 ext4_warning(sb, "too much blocks added to group %u\n",
4928                              block_group);
4929                 err = -EINVAL;
4930                 goto error_return;
4931         }
4932 
4933         bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4934         if (!bitmap_bh) {
4935                 err = -EIO;
4936                 goto error_return;
4937         }
4938 
4939         desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4940         if (!desc) {
4941                 err = -EIO;
4942                 goto error_return;
4943         }
4944 
4945         if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4946             in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4947             in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4948             in_range(block + count - 1, ext4_inode_table(sb, desc),
4949                      sbi->s_itb_per_group)) {
4950                 ext4_error(sb, "Adding blocks in system zones - "
4951                            "Block = %llu, count = %lu",
4952                            block, count);
4953                 err = -EINVAL;
4954                 goto error_return;
4955         }
4956 
4957         BUFFER_TRACE(bitmap_bh, "getting write access");
4958         err = ext4_journal_get_write_access(handle, bitmap_bh);
4959         if (err)
4960                 goto error_return;
4961 
4962         /*
4963          * We are about to modify some metadata.  Call the journal APIs
4964          * to unshare ->b_data if a currently-committing transaction is
4965          * using it
4966          */
4967         BUFFER_TRACE(gd_bh, "get_write_access");
4968         err = ext4_journal_get_write_access(handle, gd_bh);
4969         if (err)
4970                 goto error_return;
4971 
4972         for (i = 0, blocks_freed = 0; i < count; i++) {
4973                 BUFFER_TRACE(bitmap_bh, "clear bit");
4974                 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4975                         ext4_error(sb, "bit already cleared for block %llu",
4976                                    (ext4_fsblk_t)(block + i));
4977                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
4978                 } else {
4979                         blocks_freed++;
4980                 }
4981         }
4982 
4983         err = ext4_mb_load_buddy(sb, block_group, &e4b);
4984         if (err)
4985                 goto error_return;
4986 
4987         /*
4988          * need to update group_info->bb_free and bitmap
4989          * with group lock held. generate_buddy look at
4990          * them with group lock_held
4991          */
4992         ext4_lock_group(sb, block_group);
4993         mb_clear_bits(bitmap_bh->b_data, bit, count);
4994         mb_free_blocks(NULL, &e4b, bit, count);
4995         blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4996         ext4_free_group_clusters_set(sb, desc, blk_free_count);
4997         ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4998         ext4_group_desc_csum_set(sb, block_group, desc);
4999         ext4_unlock_group(sb, block_group);
5000         percpu_counter_add(&sbi->s_freeclusters_counter,
5001                            EXT4_NUM_B2C(sbi, blocks_freed));
5002 
5003         if (sbi->s_log_groups_per_flex) {
5004                 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5005                 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5006                              &sbi->s_flex_groups[flex_group].free_clusters);
5007         }
5008 
5009         ext4_mb_unload_buddy(&e4b);
5010 
5011         /* We dirtied the bitmap block */
5012         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5013         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5014 
5015         /* And the group descriptor block */
5016         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5017         ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5018         if (!err)
5019                 err = ret;
5020 
5021 error_return:
5022         brelse(bitmap_bh);
5023         ext4_std_error(sb, err);
5024         return err;
5025 }
5026 
5027 /**
5028  * ext4_trim_extent -- function to TRIM one single free extent in the group
5029  * @sb:         super block for the file system
5030  * @start:      starting block of the free extent in the alloc. group
5031  * @count:      number of blocks to TRIM
5032  * @group:      alloc. group we are working with
5033  * @e4b:        ext4 buddy for the group
5034  *
5035  * Trim "count" blocks starting at "start" in the "group". To assure that no
5036  * one will allocate those blocks, mark it as used in buddy bitmap. This must
5037  * be called with under the group lock.
5038  */
5039 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5040                              ext4_group_t group, struct ext4_buddy *e4b)
5041 __releases(bitlock)
5042 __acquires(bitlock)
5043 {
5044         struct ext4_free_extent ex;
5045         int ret = 0;
5046 
5047         trace_ext4_trim_extent(sb, group, start, count);
5048 
5049         assert_spin_locked(ext4_group_lock_ptr(sb, group));
5050 
5051         ex.fe_start = start;
5052         ex.fe_group = group;
5053         ex.fe_len = count;
5054 
5055         /*
5056          * Mark blocks used, so no one can reuse them while
5057          * being trimmed.
5058          */
5059         mb_mark_used(e4b, &ex);
5060         ext4_unlock_group(sb, group);
5061         ret = ext4_issue_discard(sb, group, start, count);
5062         ext4_lock_group(sb, group);
5063         mb_free_blocks(NULL, e4b, start, ex.fe_len);
5064         return ret;
5065 }
5066 
5067 /**
5068  * ext4_trim_all_free -- function to trim all free space in alloc. group
5069  * @sb:                 super block for file system
5070  * @group:              group to be trimmed
5071  * @start:              first group block to examine
5072  * @max:                last group block to examine
5073  * @minblocks:          minimum extent block count
5074  *
5075  * ext4_trim_all_free walks through group's buddy bitmap searching for free
5076  * extents. When the free block is found, ext4_trim_extent is called to TRIM
5077  * the extent.
5078  *
5079  *
5080  * ext4_trim_all_free walks through group's block bitmap searching for free
5081  * extents. When the free extent is found, mark it as used in group buddy
5082  * bitmap. Then issue a TRIM command on this extent and free the extent in
5083  * the group buddy bitmap. This is done until whole group is scanned.
5084  */
5085 static ext4_grpblk_t
5086 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5087                    ext4_grpblk_t start, ext4_grpblk_t max,
5088                    ext4_grpblk_t minblocks)
5089 {
5090         void *bitmap;
5091         ext4_grpblk_t next, count = 0, free_count = 0;
5092         struct ext4_buddy e4b;
5093         int ret = 0;
5094 
5095         trace_ext4_trim_all_free(sb, group, start, max);
5096 
5097         ret = ext4_mb_load_buddy(sb, group, &e4b);
5098         if (ret) {
5099                 ext4_error(sb, "Error in loading buddy "
5100                                 "information for %u", group);
5101                 return ret;
5102         }
5103         bitmap = e4b.bd_bitmap;
5104 
5105         ext4_lock_group(sb, group);
5106         if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5107             minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5108                 goto out;
5109 
5110         start = (e4b.bd_info->bb_first_free > start) ?
5111                 e4b.bd_info->bb_first_free : start;
5112 
5113         while (start <= max) {
5114                 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5115                 if (start > max)
5116                         break;
5117                 next = mb_find_next_bit(bitmap, max + 1, start);
5118 
5119                 if ((next - start) >= minblocks) {
5120                         ret = ext4_trim_extent(sb, start,
5121                                                next - start, group, &e4b);
5122                         if (ret && ret != -EOPNOTSUPP)
5123                                 break;
5124                         ret = 0;
5125                         count += next - start;
5126                 }
5127                 free_count += next - start;
5128                 start = next + 1;
5129 
5130                 if (fatal_signal_pending(current)) {
5131                         count = -ERESTARTSYS;
5132                         break;
5133                 }
5134 
5135                 if (need_resched()) {
5136                         ext4_unlock_group(sb, group);
5137                         cond_resched();
5138                         ext4_lock_group(sb, group);
5139                 }
5140 
5141                 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5142                         break;
5143         }
5144 
5145         if (!ret) {
5146                 ret = count;
5147                 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5148         }
5149 out:
5150         ext4_unlock_group(sb, group);
5151         ext4_mb_unload_buddy(&e4b);
5152 
5153         ext4_debug("trimmed %d blocks in the group %d\n",
5154                 count, group);
5155 
5156         return ret;
5157 }
5158 
5159 /**
5160  * ext4_trim_fs() -- trim ioctl handle function
5161  * @sb:                 superblock for filesystem
5162  * @range:              fstrim_range structure
5163  *
5164  * start:       First Byte to trim
5165  * len:         number of Bytes to trim from start
5166  * minlen:      minimum extent length in Bytes
5167  * ext4_trim_fs goes through all allocation groups containing Bytes from
5168  * start to start+len. For each such a group ext4_trim_all_free function
5169  * is invoked to trim all free space.
5170  */
5171 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5172 {
5173         struct ext4_group_info *grp;
5174         ext4_group_t group, first_group, last_group;
5175         ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5176         uint64_t start, end, minlen, trimmed = 0;
5177         ext4_fsblk_t first_data_blk =
5178                         le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5179         ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5180         int ret = 0;
5181 
5182         start = range->start >> sb->s_blocksize_bits;
5183         end = start + (range->len >> sb->s_blocksize_bits) - 1;
5184         minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5185                               range->minlen >> sb->s_blocksize_bits);
5186 
5187         if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5188             start >= max_blks ||
5189             range->len < sb->s_blocksize)
5190                 return -EINVAL;
5191         if (end >= max_blks)
5192                 end = max_blks - 1;
5193         if (end <= first_data_blk)
5194                 goto out;
5195         if (start < first_data_blk)
5196                 start = first_data_blk;
5197 
5198         /* Determine first and last group to examine based on start and end */
5199         ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5200                                      &first_group, &first_cluster);
5201         ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5202                                      &last_group, &last_cluster);
5203 
5204         /* end now represents the last cluster to discard in this group */
5205         end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5206 
5207         for (group = first_group; group <= last_group; group++) {
5208                 grp = ext4_get_group_info(sb, group);
5209                 /* We only do this if the grp has never been initialized */
5210                 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5211                         ret = ext4_mb_init_group(sb, group);
5212                         if (ret)
5213                                 break;
5214                 }
5215 
5216                 /*
5217                  * For all the groups except the last one, last cluster will
5218                  * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5219                  * change it for the last group, note that last_cluster is
5220                  * already computed earlier by ext4_get_group_no_and_offset()
5221                  */
5222                 if (group == last_group)
5223                         end = last_cluster;
5224 
5225                 if (grp->bb_free >= minlen) {
5226                         cnt = ext4_trim_all_free(sb, group, first_cluster,
5227                                                 end, minlen);
5228                         if (cnt < 0) {
5229                                 ret = cnt;
5230                                 break;
5231                         }
5232                         trimmed += cnt;
5233                 }
5234 
5235                 /*
5236                  * For every group except the first one, we are sure
5237                  * that the first cluster to discard will be cluster #0.
5238                  */
5239                 first_cluster = 0;
5240         }
5241 
5242         if (!ret)
5243                 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5244 
5245 out:
5246         range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5247         return ret;
5248 }
5249 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp