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

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

Version: ~ [ linux-5.4-rc7 ] ~ [ linux-5.3.11 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.84 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.154 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.201 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.201 ] ~ [ 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-3.16.77 ] ~ [ 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 License
 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_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_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, gfp_t gfp)
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_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);
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                 bh[i] = ext4_read_block_bitmap_nowait(sb, group);
878                 if (IS_ERR(bh[i])) {
879                         err = PTR_ERR(bh[i]);
880                         bh[i] = NULL;
881                         goto out;
882                 }
883                 mb_debug(1, "read bitmap for group %u\n", group);
884         }
885 
886         /* wait for I/O completion */
887         for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
888                 int err2;
889 
890                 if (!bh[i])
891                         continue;
892                 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
893                 if (!err)
894                         err = err2;
895         }
896 
897         first_block = page->index * blocks_per_page;
898         for (i = 0; i < blocks_per_page; i++) {
899                 group = (first_block + i) >> 1;
900                 if (group >= ngroups)
901                         break;
902 
903                 if (!bh[group - first_group])
904                         /* skip initialized uptodate buddy */
905                         continue;
906 
907                 if (!buffer_verified(bh[group - first_group]))
908                         /* Skip faulty bitmaps */
909                         continue;
910                 err = 0;
911 
912                 /*
913                  * data carry information regarding this
914                  * particular group in the format specified
915                  * above
916                  *
917                  */
918                 data = page_address(page) + (i * blocksize);
919                 bitmap = bh[group - first_group]->b_data;
920 
921                 /*
922                  * We place the buddy block and bitmap block
923                  * close together
924                  */
925                 if ((first_block + i) & 1) {
926                         /* this is block of buddy */
927                         BUG_ON(incore == NULL);
928                         mb_debug(1, "put buddy for group %u in page %lu/%x\n",
929                                 group, page->index, i * blocksize);
930                         trace_ext4_mb_buddy_bitmap_load(sb, group);
931                         grinfo = ext4_get_group_info(sb, group);
932                         grinfo->bb_fragments = 0;
933                         memset(grinfo->bb_counters, 0,
934                                sizeof(*grinfo->bb_counters) *
935                                 (sb->s_blocksize_bits+2));
936                         /*
937                          * incore got set to the group block bitmap below
938                          */
939                         ext4_lock_group(sb, group);
940                         /* init the buddy */
941                         memset(data, 0xff, blocksize);
942                         ext4_mb_generate_buddy(sb, data, incore, group);
943                         ext4_unlock_group(sb, group);
944                         incore = NULL;
945                 } else {
946                         /* this is block of bitmap */
947                         BUG_ON(incore != NULL);
948                         mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
949                                 group, page->index, i * blocksize);
950                         trace_ext4_mb_bitmap_load(sb, group);
951 
952                         /* see comments in ext4_mb_put_pa() */
953                         ext4_lock_group(sb, group);
954                         memcpy(data, bitmap, blocksize);
955 
956                         /* mark all preallocated blks used in in-core bitmap */
957                         ext4_mb_generate_from_pa(sb, data, group);
958                         ext4_mb_generate_from_freelist(sb, data, group);
959                         ext4_unlock_group(sb, group);
960 
961                         /* set incore so that the buddy information can be
962                          * generated using this
963                          */
964                         incore = data;
965                 }
966         }
967         SetPageUptodate(page);
968 
969 out:
970         if (bh) {
971                 for (i = 0; i < groups_per_page; i++)
972                         brelse(bh[i]);
973                 if (bh != &bhs)
974                         kfree(bh);
975         }
976         return err;
977 }
978 
979 /*
980  * Lock the buddy and bitmap pages. This make sure other parallel init_group
981  * on the same buddy page doesn't happen whild holding the buddy page lock.
982  * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
983  * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
984  */
985 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
986                 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
987 {
988         struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
989         int block, pnum, poff;
990         int blocks_per_page;
991         struct page *page;
992 
993         e4b->bd_buddy_page = NULL;
994         e4b->bd_bitmap_page = NULL;
995 
996         blocks_per_page = PAGE_SIZE / sb->s_blocksize;
997         /*
998          * the buddy cache inode stores the block bitmap
999          * and buddy information in consecutive blocks.
1000          * So for each group we need two blocks.
1001          */
1002         block = group * 2;
1003         pnum = block / blocks_per_page;
1004         poff = block % blocks_per_page;
1005         page = find_or_create_page(inode->i_mapping, pnum, gfp);
1006         if (!page)
1007                 return -ENOMEM;
1008         BUG_ON(page->mapping != inode->i_mapping);
1009         e4b->bd_bitmap_page = page;
1010         e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1011 
1012         if (blocks_per_page >= 2) {
1013                 /* buddy and bitmap are on the same page */
1014                 return 0;
1015         }
1016 
1017         block++;
1018         pnum = block / blocks_per_page;
1019         page = find_or_create_page(inode->i_mapping, pnum, gfp);
1020         if (!page)
1021                 return -ENOMEM;
1022         BUG_ON(page->mapping != inode->i_mapping);
1023         e4b->bd_buddy_page = page;
1024         return 0;
1025 }
1026 
1027 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1028 {
1029         if (e4b->bd_bitmap_page) {
1030                 unlock_page(e4b->bd_bitmap_page);
1031                 put_page(e4b->bd_bitmap_page);
1032         }
1033         if (e4b->bd_buddy_page) {
1034                 unlock_page(e4b->bd_buddy_page);
1035                 put_page(e4b->bd_buddy_page);
1036         }
1037 }
1038 
1039 /*
1040  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1041  * block group lock of all groups for this page; do not hold the BG lock when
1042  * calling this routine!
1043  */
1044 static noinline_for_stack
1045 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1046 {
1047 
1048         struct ext4_group_info *this_grp;
1049         struct ext4_buddy e4b;
1050         struct page *page;
1051         int ret = 0;
1052 
1053         might_sleep();
1054         mb_debug(1, "init group %u\n", group);
1055         this_grp = ext4_get_group_info(sb, group);
1056         /*
1057          * This ensures that we don't reinit the buddy cache
1058          * page which map to the group from which we are already
1059          * allocating. If we are looking at the buddy cache we would
1060          * have taken a reference using ext4_mb_load_buddy and that
1061          * would have pinned buddy page to page cache.
1062          * The call to ext4_mb_get_buddy_page_lock will mark the
1063          * page accessed.
1064          */
1065         ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1066         if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1067                 /*
1068                  * somebody initialized the group
1069                  * return without doing anything
1070                  */
1071                 goto err;
1072         }
1073 
1074         page = e4b.bd_bitmap_page;
1075         ret = ext4_mb_init_cache(page, NULL, gfp);
1076         if (ret)
1077                 goto err;
1078         if (!PageUptodate(page)) {
1079                 ret = -EIO;
1080                 goto err;
1081         }
1082 
1083         if (e4b.bd_buddy_page == NULL) {
1084                 /*
1085                  * If both the bitmap and buddy are in
1086                  * the same page we don't need to force
1087                  * init the buddy
1088                  */
1089                 ret = 0;
1090                 goto err;
1091         }
1092         /* init buddy cache */
1093         page = e4b.bd_buddy_page;
1094         ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1095         if (ret)
1096                 goto err;
1097         if (!PageUptodate(page)) {
1098                 ret = -EIO;
1099                 goto err;
1100         }
1101 err:
1102         ext4_mb_put_buddy_page_lock(&e4b);
1103         return ret;
1104 }
1105 
1106 /*
1107  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1108  * block group lock of all groups for this page; do not hold the BG lock when
1109  * calling this routine!
1110  */
1111 static noinline_for_stack int
1112 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1113                        struct ext4_buddy *e4b, gfp_t gfp)
1114 {
1115         int blocks_per_page;
1116         int block;
1117         int pnum;
1118         int poff;
1119         struct page *page;
1120         int ret;
1121         struct ext4_group_info *grp;
1122         struct ext4_sb_info *sbi = EXT4_SB(sb);
1123         struct inode *inode = sbi->s_buddy_cache;
1124 
1125         might_sleep();
1126         mb_debug(1, "load group %u\n", group);
1127 
1128         blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1129         grp = ext4_get_group_info(sb, group);
1130 
1131         e4b->bd_blkbits = sb->s_blocksize_bits;
1132         e4b->bd_info = grp;
1133         e4b->bd_sb = sb;
1134         e4b->bd_group = group;
1135         e4b->bd_buddy_page = NULL;
1136         e4b->bd_bitmap_page = NULL;
1137 
1138         if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1139                 /*
1140                  * we need full data about the group
1141                  * to make a good selection
1142                  */
1143                 ret = ext4_mb_init_group(sb, group, gfp);
1144                 if (ret)
1145                         return ret;
1146         }
1147 
1148         /*
1149          * the buddy cache inode stores the block bitmap
1150          * and buddy information in consecutive blocks.
1151          * So for each group we need two blocks.
1152          */
1153         block = group * 2;
1154         pnum = block / blocks_per_page;
1155         poff = block % blocks_per_page;
1156 
1157         /* we could use find_or_create_page(), but it locks page
1158          * what we'd like to avoid in fast path ... */
1159         page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1160         if (page == NULL || !PageUptodate(page)) {
1161                 if (page)
1162                         /*
1163                          * drop the page reference and try
1164                          * to get the page with lock. If we
1165                          * are not uptodate that implies
1166                          * somebody just created the page but
1167                          * is yet to initialize the same. So
1168                          * wait for it to initialize.
1169                          */
1170                         put_page(page);
1171                 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1172                 if (page) {
1173                         BUG_ON(page->mapping != inode->i_mapping);
1174                         if (!PageUptodate(page)) {
1175                                 ret = ext4_mb_init_cache(page, NULL, gfp);
1176                                 if (ret) {
1177                                         unlock_page(page);
1178                                         goto err;
1179                                 }
1180                                 mb_cmp_bitmaps(e4b, page_address(page) +
1181                                                (poff * sb->s_blocksize));
1182                         }
1183                         unlock_page(page);
1184                 }
1185         }
1186         if (page == NULL) {
1187                 ret = -ENOMEM;
1188                 goto err;
1189         }
1190         if (!PageUptodate(page)) {
1191                 ret = -EIO;
1192                 goto err;
1193         }
1194 
1195         /* Pages marked accessed already */
1196         e4b->bd_bitmap_page = page;
1197         e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1198 
1199         block++;
1200         pnum = block / blocks_per_page;
1201         poff = block % blocks_per_page;
1202 
1203         page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1204         if (page == NULL || !PageUptodate(page)) {
1205                 if (page)
1206                         put_page(page);
1207                 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1208                 if (page) {
1209                         BUG_ON(page->mapping != inode->i_mapping);
1210                         if (!PageUptodate(page)) {
1211                                 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1212                                                          gfp);
1213                                 if (ret) {
1214                                         unlock_page(page);
1215                                         goto err;
1216                                 }
1217                         }
1218                         unlock_page(page);
1219                 }
1220         }
1221         if (page == NULL) {
1222                 ret = -ENOMEM;
1223                 goto err;
1224         }
1225         if (!PageUptodate(page)) {
1226                 ret = -EIO;
1227                 goto err;
1228         }
1229 
1230         /* Pages marked accessed already */
1231         e4b->bd_buddy_page = page;
1232         e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1233 
1234         BUG_ON(e4b->bd_bitmap_page == NULL);
1235         BUG_ON(e4b->bd_buddy_page == NULL);
1236 
1237         return 0;
1238 
1239 err:
1240         if (page)
1241                 put_page(page);
1242         if (e4b->bd_bitmap_page)
1243                 put_page(e4b->bd_bitmap_page);
1244         if (e4b->bd_buddy_page)
1245                 put_page(e4b->bd_buddy_page);
1246         e4b->bd_buddy = NULL;
1247         e4b->bd_bitmap = NULL;
1248         return ret;
1249 }
1250 
1251 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1252                               struct ext4_buddy *e4b)
1253 {
1254         return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1255 }
1256 
1257 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1258 {
1259         if (e4b->bd_bitmap_page)
1260                 put_page(e4b->bd_bitmap_page);
1261         if (e4b->bd_buddy_page)
1262                 put_page(e4b->bd_buddy_page);
1263 }
1264 
1265 
1266 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1267 {
1268         int order = 1;
1269         int bb_incr = 1 << (e4b->bd_blkbits - 1);
1270         void *bb;
1271 
1272         BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1273         BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1274 
1275         bb = e4b->bd_buddy;
1276         while (order <= e4b->bd_blkbits + 1) {
1277                 block = block >> 1;
1278                 if (!mb_test_bit(block, bb)) {
1279                         /* this block is part of buddy of order 'order' */
1280                         return order;
1281                 }
1282                 bb += bb_incr;
1283                 bb_incr >>= 1;
1284                 order++;
1285         }
1286         return 0;
1287 }
1288 
1289 static void mb_clear_bits(void *bm, int cur, int len)
1290 {
1291         __u32 *addr;
1292 
1293         len = cur + len;
1294         while (cur < len) {
1295                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1296                         /* fast path: clear whole word at once */
1297                         addr = bm + (cur >> 3);
1298                         *addr = 0;
1299                         cur += 32;
1300                         continue;
1301                 }
1302                 mb_clear_bit(cur, bm);
1303                 cur++;
1304         }
1305 }
1306 
1307 /* clear bits in given range
1308  * will return first found zero bit if any, -1 otherwise
1309  */
1310 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1311 {
1312         __u32 *addr;
1313         int zero_bit = -1;
1314 
1315         len = cur + len;
1316         while (cur < len) {
1317                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1318                         /* fast path: clear whole word at once */
1319                         addr = bm + (cur >> 3);
1320                         if (*addr != (__u32)(-1) && zero_bit == -1)
1321                                 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1322                         *addr = 0;
1323                         cur += 32;
1324                         continue;
1325                 }
1326                 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1327                         zero_bit = cur;
1328                 cur++;
1329         }
1330 
1331         return zero_bit;
1332 }
1333 
1334 void ext4_set_bits(void *bm, int cur, int len)
1335 {
1336         __u32 *addr;
1337 
1338         len = cur + len;
1339         while (cur < len) {
1340                 if ((cur & 31) == 0 && (len - cur) >= 32) {
1341                         /* fast path: set whole word at once */
1342                         addr = bm + (cur >> 3);
1343                         *addr = 0xffffffff;
1344                         cur += 32;
1345                         continue;
1346                 }
1347                 mb_set_bit(cur, bm);
1348                 cur++;
1349         }
1350 }
1351 
1352 /*
1353  * _________________________________________________________________ */
1354 
1355 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1356 {
1357         if (mb_test_bit(*bit + side, bitmap)) {
1358                 mb_clear_bit(*bit, bitmap);
1359                 (*bit) -= side;
1360                 return 1;
1361         }
1362         else {
1363                 (*bit) += side;
1364                 mb_set_bit(*bit, bitmap);
1365                 return -1;
1366         }
1367 }
1368 
1369 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1370 {
1371         int max;
1372         int order = 1;
1373         void *buddy = mb_find_buddy(e4b, order, &max);
1374 
1375         while (buddy) {
1376                 void *buddy2;
1377 
1378                 /* Bits in range [first; last] are known to be set since
1379                  * corresponding blocks were allocated. Bits in range
1380                  * (first; last) will stay set because they form buddies on
1381                  * upper layer. We just deal with borders if they don't
1382                  * align with upper layer and then go up.
1383                  * Releasing entire group is all about clearing
1384                  * single bit of highest order buddy.
1385                  */
1386 
1387                 /* Example:
1388                  * ---------------------------------
1389                  * |   1   |   1   |   1   |   1   |
1390                  * ---------------------------------
1391                  * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1392                  * ---------------------------------
1393                  *   0   1   2   3   4   5   6   7
1394                  *      \_____________________/
1395                  *
1396                  * Neither [1] nor [6] is aligned to above layer.
1397                  * Left neighbour [0] is free, so mark it busy,
1398                  * decrease bb_counters and extend range to
1399                  * [0; 6]
1400                  * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1401                  * mark [6] free, increase bb_counters and shrink range to
1402                  * [0; 5].
1403                  * Then shift range to [0; 2], go up and do the same.
1404                  */
1405 
1406 
1407                 if (first & 1)
1408                         e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1409                 if (!(last & 1))
1410                         e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1411                 if (first > last)
1412                         break;
1413                 order++;
1414 
1415                 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1416                         mb_clear_bits(buddy, first, last - first + 1);
1417                         e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1418                         break;
1419                 }
1420                 first >>= 1;
1421                 last >>= 1;
1422                 buddy = buddy2;
1423         }
1424 }
1425 
1426 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1427                            int first, int count)
1428 {
1429         int left_is_free = 0;
1430         int right_is_free = 0;
1431         int block;
1432         int last = first + count - 1;
1433         struct super_block *sb = e4b->bd_sb;
1434 
1435         if (WARN_ON(count == 0))
1436                 return;
1437         BUG_ON(last >= (sb->s_blocksize << 3));
1438         assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1439         /* Don't bother if the block group is corrupt. */
1440         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1441                 return;
1442 
1443         mb_check_buddy(e4b);
1444         mb_free_blocks_double(inode, e4b, first, count);
1445 
1446         e4b->bd_info->bb_free += count;
1447         if (first < e4b->bd_info->bb_first_free)
1448                 e4b->bd_info->bb_first_free = first;
1449 
1450         /* access memory sequentially: check left neighbour,
1451          * clear range and then check right neighbour
1452          */
1453         if (first != 0)
1454                 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1455         block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1456         if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1457                 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1458 
1459         if (unlikely(block != -1)) {
1460                 struct ext4_sb_info *sbi = EXT4_SB(sb);
1461                 ext4_fsblk_t blocknr;
1462 
1463                 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1464                 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1465                 ext4_grp_locked_error(sb, e4b->bd_group,
1466                                       inode ? inode->i_ino : 0,
1467                                       blocknr,
1468                                       "freeing already freed block "
1469                                       "(bit %u); block bitmap corrupt.",
1470                                       block);
1471                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1472                         percpu_counter_sub(&sbi->s_freeclusters_counter,
1473                                            e4b->bd_info->bb_free);
1474                 /* Mark the block group as corrupt. */
1475                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1476                         &e4b->bd_info->bb_state);
1477                 mb_regenerate_buddy(e4b);
1478                 goto done;
1479         }
1480 
1481         /* let's maintain fragments counter */
1482         if (left_is_free && right_is_free)
1483                 e4b->bd_info->bb_fragments--;
1484         else if (!left_is_free && !right_is_free)
1485                 e4b->bd_info->bb_fragments++;
1486 
1487         /* buddy[0] == bd_bitmap is a special case, so handle
1488          * it right away and let mb_buddy_mark_free stay free of
1489          * zero order checks.
1490          * Check if neighbours are to be coaleasced,
1491          * adjust bitmap bb_counters and borders appropriately.
1492          */
1493         if (first & 1) {
1494                 first += !left_is_free;
1495                 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1496         }
1497         if (!(last & 1)) {
1498                 last -= !right_is_free;
1499                 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1500         }
1501 
1502         if (first <= last)
1503                 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1504 
1505 done:
1506         mb_set_largest_free_order(sb, e4b->bd_info);
1507         mb_check_buddy(e4b);
1508 }
1509 
1510 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1511                                 int needed, struct ext4_free_extent *ex)
1512 {
1513         int next = block;
1514         int max, order;
1515         void *buddy;
1516 
1517         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1518         BUG_ON(ex == NULL);
1519 
1520         buddy = mb_find_buddy(e4b, 0, &max);
1521         BUG_ON(buddy == NULL);
1522         BUG_ON(block >= max);
1523         if (mb_test_bit(block, buddy)) {
1524                 ex->fe_len = 0;
1525                 ex->fe_start = 0;
1526                 ex->fe_group = 0;
1527                 return 0;
1528         }
1529 
1530         /* find actual order */
1531         order = mb_find_order_for_block(e4b, block);
1532         block = block >> order;
1533 
1534         ex->fe_len = 1 << order;
1535         ex->fe_start = block << order;
1536         ex->fe_group = e4b->bd_group;
1537 
1538         /* calc difference from given start */
1539         next = next - ex->fe_start;
1540         ex->fe_len -= next;
1541         ex->fe_start += next;
1542 
1543         while (needed > ex->fe_len &&
1544                mb_find_buddy(e4b, order, &max)) {
1545 
1546                 if (block + 1 >= max)
1547                         break;
1548 
1549                 next = (block + 1) * (1 << order);
1550                 if (mb_test_bit(next, e4b->bd_bitmap))
1551                         break;
1552 
1553                 order = mb_find_order_for_block(e4b, next);
1554 
1555                 block = next >> order;
1556                 ex->fe_len += 1 << order;
1557         }
1558 
1559         BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1560         return ex->fe_len;
1561 }
1562 
1563 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1564 {
1565         int ord;
1566         int mlen = 0;
1567         int max = 0;
1568         int cur;
1569         int start = ex->fe_start;
1570         int len = ex->fe_len;
1571         unsigned ret = 0;
1572         int len0 = len;
1573         void *buddy;
1574 
1575         BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1576         BUG_ON(e4b->bd_group != ex->fe_group);
1577         assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1578         mb_check_buddy(e4b);
1579         mb_mark_used_double(e4b, start, len);
1580 
1581         e4b->bd_info->bb_free -= len;
1582         if (e4b->bd_info->bb_first_free == start)
1583                 e4b->bd_info->bb_first_free += len;
1584 
1585         /* let's maintain fragments counter */
1586         if (start != 0)
1587                 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1588         if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1589                 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1590         if (mlen && max)
1591                 e4b->bd_info->bb_fragments++;
1592         else if (!mlen && !max)
1593                 e4b->bd_info->bb_fragments--;
1594 
1595         /* let's maintain buddy itself */
1596         while (len) {
1597                 ord = mb_find_order_for_block(e4b, start);
1598 
1599                 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1600                         /* the whole chunk may be allocated at once! */
1601                         mlen = 1 << ord;
1602                         buddy = mb_find_buddy(e4b, ord, &max);
1603                         BUG_ON((start >> ord) >= max);
1604                         mb_set_bit(start >> ord, buddy);
1605                         e4b->bd_info->bb_counters[ord]--;
1606                         start += mlen;
1607                         len -= mlen;
1608                         BUG_ON(len < 0);
1609                         continue;
1610                 }
1611 
1612                 /* store for history */
1613                 if (ret == 0)
1614                         ret = len | (ord << 16);
1615 
1616                 /* we have to split large buddy */
1617                 BUG_ON(ord <= 0);
1618                 buddy = mb_find_buddy(e4b, ord, &max);
1619                 mb_set_bit(start >> ord, buddy);
1620                 e4b->bd_info->bb_counters[ord]--;
1621 
1622                 ord--;
1623                 cur = (start >> ord) & ~1U;
1624                 buddy = mb_find_buddy(e4b, ord, &max);
1625                 mb_clear_bit(cur, buddy);
1626                 mb_clear_bit(cur + 1, buddy);
1627                 e4b->bd_info->bb_counters[ord]++;
1628                 e4b->bd_info->bb_counters[ord]++;
1629         }
1630         mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1631 
1632         ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1633         mb_check_buddy(e4b);
1634 
1635         return ret;
1636 }
1637 
1638 /*
1639  * Must be called under group lock!
1640  */
1641 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1642                                         struct ext4_buddy *e4b)
1643 {
1644         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1645         int ret;
1646 
1647         BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1648         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1649 
1650         ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1651         ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1652         ret = mb_mark_used(e4b, &ac->ac_b_ex);
1653 
1654         /* preallocation can change ac_b_ex, thus we store actually
1655          * allocated blocks for history */
1656         ac->ac_f_ex = ac->ac_b_ex;
1657 
1658         ac->ac_status = AC_STATUS_FOUND;
1659         ac->ac_tail = ret & 0xffff;
1660         ac->ac_buddy = ret >> 16;
1661 
1662         /*
1663          * take the page reference. We want the page to be pinned
1664          * so that we don't get a ext4_mb_init_cache_call for this
1665          * group until we update the bitmap. That would mean we
1666          * double allocate blocks. The reference is dropped
1667          * in ext4_mb_release_context
1668          */
1669         ac->ac_bitmap_page = e4b->bd_bitmap_page;
1670         get_page(ac->ac_bitmap_page);
1671         ac->ac_buddy_page = e4b->bd_buddy_page;
1672         get_page(ac->ac_buddy_page);
1673         /* store last allocated for subsequent stream allocation */
1674         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1675                 spin_lock(&sbi->s_md_lock);
1676                 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1677                 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1678                 spin_unlock(&sbi->s_md_lock);
1679         }
1680 }
1681 
1682 /*
1683  * regular allocator, for general purposes allocation
1684  */
1685 
1686 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1687                                         struct ext4_buddy *e4b,
1688                                         int finish_group)
1689 {
1690         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1691         struct ext4_free_extent *bex = &ac->ac_b_ex;
1692         struct ext4_free_extent *gex = &ac->ac_g_ex;
1693         struct ext4_free_extent ex;
1694         int max;
1695 
1696         if (ac->ac_status == AC_STATUS_FOUND)
1697                 return;
1698         /*
1699          * We don't want to scan for a whole year
1700          */
1701         if (ac->ac_found > sbi->s_mb_max_to_scan &&
1702                         !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1703                 ac->ac_status = AC_STATUS_BREAK;
1704                 return;
1705         }
1706 
1707         /*
1708          * Haven't found good chunk so far, let's continue
1709          */
1710         if (bex->fe_len < gex->fe_len)
1711                 return;
1712 
1713         if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1714                         && bex->fe_group == e4b->bd_group) {
1715                 /* recheck chunk's availability - we don't know
1716                  * when it was found (within this lock-unlock
1717                  * period or not) */
1718                 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1719                 if (max >= gex->fe_len) {
1720                         ext4_mb_use_best_found(ac, e4b);
1721                         return;
1722                 }
1723         }
1724 }
1725 
1726 /*
1727  * The routine checks whether found extent is good enough. If it is,
1728  * then the extent gets marked used and flag is set to the context
1729  * to stop scanning. Otherwise, the extent is compared with the
1730  * previous found extent and if new one is better, then it's stored
1731  * in the context. Later, the best found extent will be used, if
1732  * mballoc can't find good enough extent.
1733  *
1734  * FIXME: real allocation policy is to be designed yet!
1735  */
1736 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1737                                         struct ext4_free_extent *ex,
1738                                         struct ext4_buddy *e4b)
1739 {
1740         struct ext4_free_extent *bex = &ac->ac_b_ex;
1741         struct ext4_free_extent *gex = &ac->ac_g_ex;
1742 
1743         BUG_ON(ex->fe_len <= 0);
1744         BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1745         BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1746         BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1747 
1748         ac->ac_found++;
1749 
1750         /*
1751          * The special case - take what you catch first
1752          */
1753         if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1754                 *bex = *ex;
1755                 ext4_mb_use_best_found(ac, e4b);
1756                 return;
1757         }
1758 
1759         /*
1760          * Let's check whether the chuck is good enough
1761          */
1762         if (ex->fe_len == gex->fe_len) {
1763                 *bex = *ex;
1764                 ext4_mb_use_best_found(ac, e4b);
1765                 return;
1766         }
1767 
1768         /*
1769          * If this is first found extent, just store it in the context
1770          */
1771         if (bex->fe_len == 0) {
1772                 *bex = *ex;
1773                 return;
1774         }
1775 
1776         /*
1777          * If new found extent is better, store it in the context
1778          */
1779         if (bex->fe_len < gex->fe_len) {
1780                 /* if the request isn't satisfied, any found extent
1781                  * larger than previous best one is better */
1782                 if (ex->fe_len > bex->fe_len)
1783                         *bex = *ex;
1784         } else if (ex->fe_len > gex->fe_len) {
1785                 /* if the request is satisfied, then we try to find
1786                  * an extent that still satisfy the request, but is
1787                  * smaller than previous one */
1788                 if (ex->fe_len < bex->fe_len)
1789                         *bex = *ex;
1790         }
1791 
1792         ext4_mb_check_limits(ac, e4b, 0);
1793 }
1794 
1795 static noinline_for_stack
1796 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1797                                         struct ext4_buddy *e4b)
1798 {
1799         struct ext4_free_extent ex = ac->ac_b_ex;
1800         ext4_group_t group = ex.fe_group;
1801         int max;
1802         int err;
1803 
1804         BUG_ON(ex.fe_len <= 0);
1805         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1806         if (err)
1807                 return err;
1808 
1809         ext4_lock_group(ac->ac_sb, group);
1810         max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1811 
1812         if (max > 0) {
1813                 ac->ac_b_ex = ex;
1814                 ext4_mb_use_best_found(ac, e4b);
1815         }
1816 
1817         ext4_unlock_group(ac->ac_sb, group);
1818         ext4_mb_unload_buddy(e4b);
1819 
1820         return 0;
1821 }
1822 
1823 static noinline_for_stack
1824 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1825                                 struct ext4_buddy *e4b)
1826 {
1827         ext4_group_t group = ac->ac_g_ex.fe_group;
1828         int max;
1829         int err;
1830         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1831         struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1832         struct ext4_free_extent ex;
1833 
1834         if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1835                 return 0;
1836         if (grp->bb_free == 0)
1837                 return 0;
1838 
1839         err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1840         if (err)
1841                 return err;
1842 
1843         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1844                 ext4_mb_unload_buddy(e4b);
1845                 return 0;
1846         }
1847 
1848         ext4_lock_group(ac->ac_sb, group);
1849         max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1850                              ac->ac_g_ex.fe_len, &ex);
1851         ex.fe_logical = 0xDEADFA11; /* debug value */
1852 
1853         if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1854                 ext4_fsblk_t start;
1855 
1856                 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1857                         ex.fe_start;
1858                 /* use do_div to get remainder (would be 64-bit modulo) */
1859                 if (do_div(start, sbi->s_stripe) == 0) {
1860                         ac->ac_found++;
1861                         ac->ac_b_ex = ex;
1862                         ext4_mb_use_best_found(ac, e4b);
1863                 }
1864         } else if (max >= ac->ac_g_ex.fe_len) {
1865                 BUG_ON(ex.fe_len <= 0);
1866                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1867                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1868                 ac->ac_found++;
1869                 ac->ac_b_ex = ex;
1870                 ext4_mb_use_best_found(ac, e4b);
1871         } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1872                 /* Sometimes, caller may want to merge even small
1873                  * number of blocks to an existing extent */
1874                 BUG_ON(ex.fe_len <= 0);
1875                 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1876                 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1877                 ac->ac_found++;
1878                 ac->ac_b_ex = ex;
1879                 ext4_mb_use_best_found(ac, e4b);
1880         }
1881         ext4_unlock_group(ac->ac_sb, group);
1882         ext4_mb_unload_buddy(e4b);
1883 
1884         return 0;
1885 }
1886 
1887 /*
1888  * The routine scans buddy structures (not bitmap!) from given order
1889  * to max order and tries to find big enough chunk to satisfy the req
1890  */
1891 static noinline_for_stack
1892 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1893                                         struct ext4_buddy *e4b)
1894 {
1895         struct super_block *sb = ac->ac_sb;
1896         struct ext4_group_info *grp = e4b->bd_info;
1897         void *buddy;
1898         int i;
1899         int k;
1900         int max;
1901 
1902         BUG_ON(ac->ac_2order <= 0);
1903         for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1904                 if (grp->bb_counters[i] == 0)
1905                         continue;
1906 
1907                 buddy = mb_find_buddy(e4b, i, &max);
1908                 BUG_ON(buddy == NULL);
1909 
1910                 k = mb_find_next_zero_bit(buddy, max, 0);
1911                 BUG_ON(k >= max);
1912 
1913                 ac->ac_found++;
1914 
1915                 ac->ac_b_ex.fe_len = 1 << i;
1916                 ac->ac_b_ex.fe_start = k << i;
1917                 ac->ac_b_ex.fe_group = e4b->bd_group;
1918 
1919                 ext4_mb_use_best_found(ac, e4b);
1920 
1921                 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1922 
1923                 if (EXT4_SB(sb)->s_mb_stats)
1924                         atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1925 
1926                 break;
1927         }
1928 }
1929 
1930 /*
1931  * The routine scans the group and measures all found extents.
1932  * In order to optimize scanning, caller must pass number of
1933  * free blocks in the group, so the routine can know upper limit.
1934  */
1935 static noinline_for_stack
1936 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1937                                         struct ext4_buddy *e4b)
1938 {
1939         struct super_block *sb = ac->ac_sb;
1940         void *bitmap = e4b->bd_bitmap;
1941         struct ext4_free_extent ex;
1942         int i;
1943         int free;
1944 
1945         free = e4b->bd_info->bb_free;
1946         BUG_ON(free <= 0);
1947 
1948         i = e4b->bd_info->bb_first_free;
1949 
1950         while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1951                 i = mb_find_next_zero_bit(bitmap,
1952                                                 EXT4_CLUSTERS_PER_GROUP(sb), i);
1953                 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1954                         /*
1955                          * IF we have corrupt bitmap, we won't find any
1956                          * free blocks even though group info says we
1957                          * we have free blocks
1958                          */
1959                         ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1960                                         "%d free clusters as per "
1961                                         "group info. But bitmap says 0",
1962                                         free);
1963                         break;
1964                 }
1965 
1966                 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1967                 BUG_ON(ex.fe_len <= 0);
1968                 if (free < ex.fe_len) {
1969                         ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1970                                         "%d free clusters as per "
1971                                         "group info. But got %d blocks",
1972                                         free, ex.fe_len);
1973                         /*
1974                          * The number of free blocks differs. This mostly
1975                          * indicate that the bitmap is corrupt. So exit
1976                          * without claiming the space.
1977                          */
1978                         break;
1979                 }
1980                 ex.fe_logical = 0xDEADC0DE; /* debug value */
1981                 ext4_mb_measure_extent(ac, &ex, e4b);
1982 
1983                 i += ex.fe_len;
1984                 free -= ex.fe_len;
1985         }
1986 
1987         ext4_mb_check_limits(ac, e4b, 1);
1988 }
1989 
1990 /*
1991  * This is a special case for storages like raid5
1992  * we try to find stripe-aligned chunks for stripe-size-multiple requests
1993  */
1994 static noinline_for_stack
1995 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1996                                  struct ext4_buddy *e4b)
1997 {
1998         struct super_block *sb = ac->ac_sb;
1999         struct ext4_sb_info *sbi = EXT4_SB(sb);
2000         void *bitmap = e4b->bd_bitmap;
2001         struct ext4_free_extent ex;
2002         ext4_fsblk_t first_group_block;
2003         ext4_fsblk_t a;
2004         ext4_grpblk_t i;
2005         int max;
2006 
2007         BUG_ON(sbi->s_stripe == 0);
2008 
2009         /* find first stripe-aligned block in group */
2010         first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2011 
2012         a = first_group_block + sbi->s_stripe - 1;
2013         do_div(a, sbi->s_stripe);
2014         i = (a * sbi->s_stripe) - first_group_block;
2015 
2016         while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2017                 if (!mb_test_bit(i, bitmap)) {
2018                         max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2019                         if (max >= sbi->s_stripe) {
2020                                 ac->ac_found++;
2021                                 ex.fe_logical = 0xDEADF00D; /* debug value */
2022                                 ac->ac_b_ex = ex;
2023                                 ext4_mb_use_best_found(ac, e4b);
2024                                 break;
2025                         }
2026                 }
2027                 i += sbi->s_stripe;
2028         }
2029 }
2030 
2031 /*
2032  * This is now called BEFORE we load the buddy bitmap.
2033  * Returns either 1 or 0 indicating that the group is either suitable
2034  * for the allocation or not. In addition it can also return negative
2035  * error code when something goes wrong.
2036  */
2037 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2038                                 ext4_group_t group, int cr)
2039 {
2040         unsigned free, fragments;
2041         int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2042         struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2043 
2044         BUG_ON(cr < 0 || cr >= 4);
2045 
2046         free = grp->bb_free;
2047         if (free == 0)
2048                 return 0;
2049         if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2050                 return 0;
2051 
2052         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2053                 return 0;
2054 
2055         /* We only do this if the grp has never been initialized */
2056         if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2057                 int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2058                 if (ret)
2059                         return ret;
2060         }
2061 
2062         fragments = grp->bb_fragments;
2063         if (fragments == 0)
2064                 return 0;
2065 
2066         switch (cr) {
2067         case 0:
2068                 BUG_ON(ac->ac_2order == 0);
2069 
2070                 /* Avoid using the first bg of a flexgroup for data files */
2071                 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2072                     (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2073                     ((group % flex_size) == 0))
2074                         return 0;
2075 
2076                 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2077                     (free / fragments) >= ac->ac_g_ex.fe_len)
2078                         return 1;
2079 
2080                 if (grp->bb_largest_free_order < ac->ac_2order)
2081                         return 0;
2082 
2083                 return 1;
2084         case 1:
2085                 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2086                         return 1;
2087                 break;
2088         case 2:
2089                 if (free >= ac->ac_g_ex.fe_len)
2090                         return 1;
2091                 break;
2092         case 3:
2093                 return 1;
2094         default:
2095                 BUG();
2096         }
2097 
2098         return 0;
2099 }
2100 
2101 static noinline_for_stack int
2102 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2103 {
2104         ext4_group_t ngroups, group, i;
2105         int cr;
2106         int err = 0, first_err = 0;
2107         struct ext4_sb_info *sbi;
2108         struct super_block *sb;
2109         struct ext4_buddy e4b;
2110 
2111         sb = ac->ac_sb;
2112         sbi = EXT4_SB(sb);
2113         ngroups = ext4_get_groups_count(sb);
2114         /* non-extent files are limited to low blocks/groups */
2115         if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2116                 ngroups = sbi->s_blockfile_groups;
2117 
2118         BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2119 
2120         /* first, try the goal */
2121         err = ext4_mb_find_by_goal(ac, &e4b);
2122         if (err || ac->ac_status == AC_STATUS_FOUND)
2123                 goto out;
2124 
2125         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2126                 goto out;
2127 
2128         /*
2129          * ac->ac2_order is set only if the fe_len is a power of 2
2130          * if ac2_order is set we also set criteria to 0 so that we
2131          * try exact allocation using buddy.
2132          */
2133         i = fls(ac->ac_g_ex.fe_len);
2134         ac->ac_2order = 0;
2135         /*
2136          * We search using buddy data only if the order of the request
2137          * is greater than equal to the sbi_s_mb_order2_reqs
2138          * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2139          */
2140         if (i >= sbi->s_mb_order2_reqs) {
2141                 /*
2142                  * This should tell if fe_len is exactly power of 2
2143                  */
2144                 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2145                         ac->ac_2order = i - 1;
2146         }
2147 
2148         /* if stream allocation is enabled, use global goal */
2149         if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2150                 /* TBD: may be hot point */
2151                 spin_lock(&sbi->s_md_lock);
2152                 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2153                 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2154                 spin_unlock(&sbi->s_md_lock);
2155         }
2156 
2157         /* Let's just scan groups to find more-less suitable blocks */
2158         cr = ac->ac_2order ? 0 : 1;
2159         /*
2160          * cr == 0 try to get exact allocation,
2161          * cr == 3  try to get anything
2162          */
2163 repeat:
2164         for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2165                 ac->ac_criteria = cr;
2166                 /*
2167                  * searching for the right group start
2168                  * from the goal value specified
2169                  */
2170                 group = ac->ac_g_ex.fe_group;
2171 
2172                 for (i = 0; i < ngroups; group++, i++) {
2173                         int ret = 0;
2174                         cond_resched();
2175                         /*
2176                          * Artificially restricted ngroups for non-extent
2177                          * files makes group > ngroups possible on first loop.
2178                          */
2179                         if (group >= ngroups)
2180                                 group = 0;
2181 
2182                         /* This now checks without needing the buddy page */
2183                         ret = ext4_mb_good_group(ac, group, cr);
2184                         if (ret <= 0) {
2185                                 if (!first_err)
2186                                         first_err = ret;
2187                                 continue;
2188                         }
2189 
2190                         err = ext4_mb_load_buddy(sb, group, &e4b);
2191                         if (err)
2192                                 goto out;
2193 
2194                         ext4_lock_group(sb, group);
2195 
2196                         /*
2197                          * We need to check again after locking the
2198                          * block group
2199                          */
2200                         ret = ext4_mb_good_group(ac, group, cr);
2201                         if (ret <= 0) {
2202                                 ext4_unlock_group(sb, group);
2203                                 ext4_mb_unload_buddy(&e4b);
2204                                 if (!first_err)
2205                                         first_err = ret;
2206                                 continue;
2207                         }
2208 
2209                         ac->ac_groups_scanned++;
2210                         if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2211                                 ext4_mb_simple_scan_group(ac, &e4b);
2212                         else if (cr == 1 && sbi->s_stripe &&
2213                                         !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2214                                 ext4_mb_scan_aligned(ac, &e4b);
2215                         else
2216                                 ext4_mb_complex_scan_group(ac, &e4b);
2217 
2218                         ext4_unlock_group(sb, group);
2219                         ext4_mb_unload_buddy(&e4b);
2220 
2221                         if (ac->ac_status != AC_STATUS_CONTINUE)
2222                                 break;
2223                 }
2224         }
2225 
2226         if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2227             !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2228                 /*
2229                  * We've been searching too long. Let's try to allocate
2230                  * the best chunk we've found so far
2231                  */
2232 
2233                 ext4_mb_try_best_found(ac, &e4b);
2234                 if (ac->ac_status != AC_STATUS_FOUND) {
2235                         /*
2236                          * Someone more lucky has already allocated it.
2237                          * The only thing we can do is just take first
2238                          * found block(s)
2239                         printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2240                          */
2241                         ac->ac_b_ex.fe_group = 0;
2242                         ac->ac_b_ex.fe_start = 0;
2243                         ac->ac_b_ex.fe_len = 0;
2244                         ac->ac_status = AC_STATUS_CONTINUE;
2245                         ac->ac_flags |= EXT4_MB_HINT_FIRST;
2246                         cr = 3;
2247                         atomic_inc(&sbi->s_mb_lost_chunks);
2248                         goto repeat;
2249                 }
2250         }
2251 out:
2252         if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2253                 err = first_err;
2254         return err;
2255 }
2256 
2257 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2258 {
2259         struct super_block *sb = seq->private;
2260         ext4_group_t group;
2261 
2262         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2263                 return NULL;
2264         group = *pos + 1;
2265         return (void *) ((unsigned long) group);
2266 }
2267 
2268 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2269 {
2270         struct super_block *sb = seq->private;
2271         ext4_group_t group;
2272 
2273         ++*pos;
2274         if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2275                 return NULL;
2276         group = *pos + 1;
2277         return (void *) ((unsigned long) group);
2278 }
2279 
2280 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2281 {
2282         struct super_block *sb = seq->private;
2283         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2284         int i;
2285         int err, buddy_loaded = 0;
2286         struct ext4_buddy e4b;
2287         struct ext4_group_info *grinfo;
2288         struct sg {
2289                 struct ext4_group_info info;
2290                 ext4_grpblk_t counters[16];
2291         } sg;
2292 
2293         group--;
2294         if (group == 0)
2295                 seq_puts(seq, "#group: free  frags first ["
2296                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
2297                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
2298 
2299         i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2300                 sizeof(struct ext4_group_info);
2301         grinfo = ext4_get_group_info(sb, group);
2302         /* Load the group info in memory only if not already loaded. */
2303         if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2304                 err = ext4_mb_load_buddy(sb, group, &e4b);
2305                 if (err) {
2306                         seq_printf(seq, "#%-5u: I/O error\n", group);
2307                         return 0;
2308                 }
2309                 buddy_loaded = 1;
2310         }
2311 
2312         memcpy(&sg, ext4_get_group_info(sb, group), i);
2313 
2314         if (buddy_loaded)
2315                 ext4_mb_unload_buddy(&e4b);
2316 
2317         seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2318                         sg.info.bb_fragments, sg.info.bb_first_free);
2319         for (i = 0; i <= 13; i++)
2320                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2321                                 sg.info.bb_counters[i] : 0);
2322         seq_printf(seq, " ]\n");
2323 
2324         return 0;
2325 }
2326 
2327 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2328 {
2329 }
2330 
2331 static const struct seq_operations ext4_mb_seq_groups_ops = {
2332         .start  = ext4_mb_seq_groups_start,
2333         .next   = ext4_mb_seq_groups_next,
2334         .stop   = ext4_mb_seq_groups_stop,
2335         .show   = ext4_mb_seq_groups_show,
2336 };
2337 
2338 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2339 {
2340         struct super_block *sb = PDE_DATA(inode);
2341         int rc;
2342 
2343         rc = seq_open(file, &ext4_mb_seq_groups_ops);
2344         if (rc == 0) {
2345                 struct seq_file *m = file->private_data;
2346                 m->private = sb;
2347         }
2348         return rc;
2349 
2350 }
2351 
2352 const struct file_operations ext4_seq_mb_groups_fops = {
2353         .owner          = THIS_MODULE,
2354         .open           = ext4_mb_seq_groups_open,
2355         .read           = seq_read,
2356         .llseek         = seq_lseek,
2357         .release        = seq_release,
2358 };
2359 
2360 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2361 {
2362         int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2363         struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2364 
2365         BUG_ON(!cachep);
2366         return cachep;
2367 }
2368 
2369 /*
2370  * Allocate the top-level s_group_info array for the specified number
2371  * of groups
2372  */
2373 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2374 {
2375         struct ext4_sb_info *sbi = EXT4_SB(sb);
2376         unsigned size;
2377         struct ext4_group_info ***new_groupinfo;
2378 
2379         size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2380                 EXT4_DESC_PER_BLOCK_BITS(sb);
2381         if (size <= sbi->s_group_info_size)
2382                 return 0;
2383 
2384         size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2385         new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2386         if (!new_groupinfo) {
2387                 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2388                 return -ENOMEM;
2389         }
2390         if (sbi->s_group_info) {
2391                 memcpy(new_groupinfo, sbi->s_group_info,
2392                        sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2393                 kvfree(sbi->s_group_info);
2394         }
2395         sbi->s_group_info = new_groupinfo;
2396         sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2397         ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 
2398                    sbi->s_group_info_size);
2399         return 0;
2400 }
2401 
2402 /* Create and initialize ext4_group_info data for the given group. */
2403 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2404                           struct ext4_group_desc *desc)
2405 {
2406         int i;
2407         int metalen = 0;
2408         struct ext4_sb_info *sbi = EXT4_SB(sb);
2409         struct ext4_group_info **meta_group_info;
2410         struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2411 
2412         /*
2413          * First check if this group is the first of a reserved block.
2414          * If it's true, we have to allocate a new table of pointers
2415          * to ext4_group_info structures
2416          */
2417         if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2418                 metalen = sizeof(*meta_group_info) <<
2419                         EXT4_DESC_PER_BLOCK_BITS(sb);
2420                 meta_group_info = kmalloc(metalen, GFP_NOFS);
2421                 if (meta_group_info == NULL) {
2422                         ext4_msg(sb, KERN_ERR, "can't allocate mem "
2423                                  "for a buddy group");
2424                         goto exit_meta_group_info;
2425                 }
2426                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2427                         meta_group_info;
2428         }
2429 
2430         meta_group_info =
2431                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2432         i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2433 
2434         meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2435         if (meta_group_info[i] == NULL) {
2436                 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2437                 goto exit_group_info;
2438         }
2439         set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2440                 &(meta_group_info[i]->bb_state));
2441 
2442         /*
2443          * initialize bb_free to be able to skip
2444          * empty groups without initialization
2445          */
2446         if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2447                 meta_group_info[i]->bb_free =
2448                         ext4_free_clusters_after_init(sb, group, desc);
2449         } else {
2450                 meta_group_info[i]->bb_free =
2451                         ext4_free_group_clusters(sb, desc);
2452         }
2453 
2454         INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2455         init_rwsem(&meta_group_info[i]->alloc_sem);
2456         meta_group_info[i]->bb_free_root = RB_ROOT;
2457         meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2458 
2459 #ifdef DOUBLE_CHECK
2460         {
2461                 struct buffer_head *bh;
2462                 meta_group_info[i]->bb_bitmap =
2463                         kmalloc(sb->s_blocksize, GFP_NOFS);
2464                 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2465                 bh = ext4_read_block_bitmap(sb, group);
2466                 BUG_ON(IS_ERR_OR_NULL(bh));
2467                 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2468                         sb->s_blocksize);
2469                 put_bh(bh);
2470         }
2471 #endif
2472 
2473         return 0;
2474 
2475 exit_group_info:
2476         /* If a meta_group_info table has been allocated, release it now */
2477         if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2478                 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2479                 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2480         }
2481 exit_meta_group_info:
2482         return -ENOMEM;
2483 } /* ext4_mb_add_groupinfo */
2484 
2485 static int ext4_mb_init_backend(struct super_block *sb)
2486 {
2487         ext4_group_t ngroups = ext4_get_groups_count(sb);
2488         ext4_group_t i;
2489         struct ext4_sb_info *sbi = EXT4_SB(sb);
2490         int err;
2491         struct ext4_group_desc *desc;
2492         struct kmem_cache *cachep;
2493 
2494         err = ext4_mb_alloc_groupinfo(sb, ngroups);
2495         if (err)
2496                 return err;
2497 
2498         sbi->s_buddy_cache = new_inode(sb);
2499         if (sbi->s_buddy_cache == NULL) {
2500                 ext4_msg(sb, KERN_ERR, "can't get new inode");
2501                 goto err_freesgi;
2502         }
2503         /* To avoid potentially colliding with an valid on-disk inode number,
2504          * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2505          * not in the inode hash, so it should never be found by iget(), but
2506          * this will avoid confusion if it ever shows up during debugging. */
2507         sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2508         EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2509         for (i = 0; i < ngroups; i++) {
2510                 desc = ext4_get_group_desc(sb, i, NULL);
2511                 if (desc == NULL) {
2512                         ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2513                         goto err_freebuddy;
2514                 }
2515                 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2516                         goto err_freebuddy;
2517         }
2518 
2519         return 0;
2520 
2521 err_freebuddy:
2522         cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2523         while (i-- > 0)
2524                 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2525         i = sbi->s_group_info_size;
2526         while (i-- > 0)
2527                 kfree(sbi->s_group_info[i]);
2528         iput(sbi->s_buddy_cache);
2529 err_freesgi:
2530         kvfree(sbi->s_group_info);
2531         return -ENOMEM;
2532 }
2533 
2534 static void ext4_groupinfo_destroy_slabs(void)
2535 {
2536         int i;
2537 
2538         for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2539                 if (ext4_groupinfo_caches[i])
2540                         kmem_cache_destroy(ext4_groupinfo_caches[i]);
2541                 ext4_groupinfo_caches[i] = NULL;
2542         }
2543 }
2544 
2545 static int ext4_groupinfo_create_slab(size_t size)
2546 {
2547         static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2548         int slab_size;
2549         int blocksize_bits = order_base_2(size);
2550         int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2551         struct kmem_cache *cachep;
2552 
2553         if (cache_index >= NR_GRPINFO_CACHES)
2554                 return -EINVAL;
2555 
2556         if (unlikely(cache_index < 0))
2557                 cache_index = 0;
2558 
2559         mutex_lock(&ext4_grpinfo_slab_create_mutex);
2560         if (ext4_groupinfo_caches[cache_index]) {
2561                 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2562                 return 0;       /* Already created */
2563         }
2564 
2565         slab_size = offsetof(struct ext4_group_info,
2566                                 bb_counters[blocksize_bits + 2]);
2567 
2568         cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2569                                         slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2570                                         NULL);
2571 
2572         ext4_groupinfo_caches[cache_index] = cachep;
2573 
2574         mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2575         if (!cachep) {
2576                 printk(KERN_EMERG
2577                        "EXT4-fs: no memory for groupinfo slab cache\n");
2578                 return -ENOMEM;
2579         }
2580 
2581         return 0;
2582 }
2583 
2584 int ext4_mb_init(struct super_block *sb)
2585 {
2586         struct ext4_sb_info *sbi = EXT4_SB(sb);
2587         unsigned i, j;
2588         unsigned offset, offset_incr;
2589         unsigned max;
2590         int ret;
2591 
2592         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2593 
2594         sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2595         if (sbi->s_mb_offsets == NULL) {
2596                 ret = -ENOMEM;
2597                 goto out;
2598         }
2599 
2600         i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2601         sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2602         if (sbi->s_mb_maxs == NULL) {
2603                 ret = -ENOMEM;
2604                 goto out;
2605         }
2606 
2607         ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2608         if (ret < 0)
2609                 goto out;
2610 
2611         /* order 0 is regular bitmap */
2612         sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2613         sbi->s_mb_offsets[0] = 0;
2614 
2615         i = 1;
2616         offset = 0;
2617         offset_incr = 1 << (sb->s_blocksize_bits - 1);
2618         max = sb->s_blocksize << 2;
2619         do {
2620                 sbi->s_mb_offsets[i] = offset;
2621                 sbi->s_mb_maxs[i] = max;
2622                 offset += offset_incr;
2623                 offset_incr = offset_incr >> 1;
2624                 max = max >> 1;
2625                 i++;
2626         } while (i <= sb->s_blocksize_bits + 1);
2627 
2628         spin_lock_init(&sbi->s_md_lock);
2629         spin_lock_init(&sbi->s_bal_lock);
2630 
2631         sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2632         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2633         sbi->s_mb_stats = MB_DEFAULT_STATS;
2634         sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2635         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2636         /*
2637          * The default group preallocation is 512, which for 4k block
2638          * sizes translates to 2 megabytes.  However for bigalloc file
2639          * systems, this is probably too big (i.e, if the cluster size
2640          * is 1 megabyte, then group preallocation size becomes half a
2641          * gigabyte!).  As a default, we will keep a two megabyte
2642          * group pralloc size for cluster sizes up to 64k, and after
2643          * that, we will force a minimum group preallocation size of
2644          * 32 clusters.  This translates to 8 megs when the cluster
2645          * size is 256k, and 32 megs when the cluster size is 1 meg,
2646          * which seems reasonable as a default.
2647          */
2648         sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2649                                        sbi->s_cluster_bits, 32);
2650         /*
2651          * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2652          * to the lowest multiple of s_stripe which is bigger than
2653          * the s_mb_group_prealloc as determined above. We want
2654          * the preallocation size to be an exact multiple of the
2655          * RAID stripe size so that preallocations don't fragment
2656          * the stripes.
2657          */
2658         if (sbi->s_stripe > 1) {
2659                 sbi->s_mb_group_prealloc = roundup(
2660                         sbi->s_mb_group_prealloc, sbi->s_stripe);
2661         }
2662 
2663         sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2664         if (sbi->s_locality_groups == NULL) {
2665                 ret = -ENOMEM;
2666                 goto out;
2667         }
2668         for_each_possible_cpu(i) {
2669                 struct ext4_locality_group *lg;
2670                 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2671                 mutex_init(&lg->lg_mutex);
2672                 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2673                         INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2674                 spin_lock_init(&lg->lg_prealloc_lock);
2675         }
2676 
2677         /* init file for buddy data */
2678         ret = ext4_mb_init_backend(sb);
2679         if (ret != 0)
2680                 goto out_free_locality_groups;
2681 
2682         return 0;
2683 
2684 out_free_locality_groups:
2685         free_percpu(sbi->s_locality_groups);
2686         sbi->s_locality_groups = NULL;
2687 out:
2688         kfree(sbi->s_mb_offsets);
2689         sbi->s_mb_offsets = NULL;
2690         kfree(sbi->s_mb_maxs);
2691         sbi->s_mb_maxs = NULL;
2692         return ret;
2693 }
2694 
2695 /* need to called with the ext4 group lock held */
2696 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2697 {
2698         struct ext4_prealloc_space *pa;
2699         struct list_head *cur, *tmp;
2700         int count = 0;
2701 
2702         list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2703                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2704                 list_del(&pa->pa_group_list);
2705                 count++;
2706                 kmem_cache_free(ext4_pspace_cachep, pa);
2707         }
2708         if (count)
2709                 mb_debug(1, "mballoc: %u PAs left\n", count);
2710 
2711 }
2712 
2713 int ext4_mb_release(struct super_block *sb)
2714 {
2715         ext4_group_t ngroups = ext4_get_groups_count(sb);
2716         ext4_group_t i;
2717         int num_meta_group_infos;
2718         struct ext4_group_info *grinfo;
2719         struct ext4_sb_info *sbi = EXT4_SB(sb);
2720         struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2721 
2722         if (sbi->s_group_info) {
2723                 for (i = 0; i < ngroups; i++) {
2724                         grinfo = ext4_get_group_info(sb, i);
2725 #ifdef DOUBLE_CHECK
2726                         kfree(grinfo->bb_bitmap);
2727 #endif
2728                         ext4_lock_group(sb, i);
2729                         ext4_mb_cleanup_pa(grinfo);
2730                         ext4_unlock_group(sb, i);
2731                         kmem_cache_free(cachep, grinfo);
2732                 }
2733                 num_meta_group_infos = (ngroups +
2734                                 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2735                         EXT4_DESC_PER_BLOCK_BITS(sb);
2736                 for (i = 0; i < num_meta_group_infos; i++)
2737                         kfree(sbi->s_group_info[i]);
2738                 kvfree(sbi->s_group_info);
2739         }
2740         kfree(sbi->s_mb_offsets);
2741         kfree(sbi->s_mb_maxs);
2742         iput(sbi->s_buddy_cache);
2743         if (sbi->s_mb_stats) {
2744                 ext4_msg(sb, KERN_INFO,
2745                        "mballoc: %u blocks %u reqs (%u success)",
2746                                 atomic_read(&sbi->s_bal_allocated),
2747                                 atomic_read(&sbi->s_bal_reqs),
2748                                 atomic_read(&sbi->s_bal_success));
2749                 ext4_msg(sb, KERN_INFO,
2750                       "mballoc: %u extents scanned, %u goal hits, "
2751                                 "%u 2^N hits, %u breaks, %u lost",
2752                                 atomic_read(&sbi->s_bal_ex_scanned),
2753                                 atomic_read(&sbi->s_bal_goals),
2754                                 atomic_read(&sbi->s_bal_2orders),
2755                                 atomic_read(&sbi->s_bal_breaks),
2756                                 atomic_read(&sbi->s_mb_lost_chunks));
2757                 ext4_msg(sb, KERN_INFO,
2758                        "mballoc: %lu generated and it took %Lu",
2759                                 sbi->s_mb_buddies_generated,
2760                                 sbi->s_mb_generation_time);
2761                 ext4_msg(sb, KERN_INFO,
2762                        "mballoc: %u preallocated, %u discarded",
2763                                 atomic_read(&sbi->s_mb_preallocated),
2764                                 atomic_read(&sbi->s_mb_discarded));
2765         }
2766 
2767         free_percpu(sbi->s_locality_groups);
2768 
2769         return 0;
2770 }
2771 
2772 static inline int ext4_issue_discard(struct super_block *sb,
2773                 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2774 {
2775         ext4_fsblk_t discard_block;
2776 
2777         discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2778                          ext4_group_first_block_no(sb, block_group));
2779         count = EXT4_C2B(EXT4_SB(sb), count);
2780         trace_ext4_discard_blocks(sb,
2781                         (unsigned long long) discard_block, count);
2782         return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2783 }
2784 
2785 /*
2786  * This function is called by the jbd2 layer once the commit has finished,
2787  * so we know we can free the blocks that were released with that commit.
2788  */
2789 static void ext4_free_data_callback(struct super_block *sb,
2790                                     struct ext4_journal_cb_entry *jce,
2791                                     int rc)
2792 {
2793         struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2794         struct ext4_buddy e4b;
2795         struct ext4_group_info *db;
2796         int err, count = 0, count2 = 0;
2797 
2798         mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2799                  entry->efd_count, entry->efd_group, entry);
2800 
2801         if (test_opt(sb, DISCARD)) {
2802                 err = ext4_issue_discard(sb, entry->efd_group,
2803                                          entry->efd_start_cluster,
2804                                          entry->efd_count);
2805                 if (err && err != -EOPNOTSUPP)
2806                         ext4_msg(sb, KERN_WARNING, "discard request in"
2807                                  " group:%d block:%d count:%d failed"
2808                                  " with %d", entry->efd_group,
2809                                  entry->efd_start_cluster,
2810                                  entry->efd_count, err);
2811         }
2812 
2813         err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2814         /* we expect to find existing buddy because it's pinned */
2815         BUG_ON(err != 0);
2816 
2817 
2818         db = e4b.bd_info;
2819         /* there are blocks to put in buddy to make them really free */
2820         count += entry->efd_count;
2821         count2++;
2822         ext4_lock_group(sb, entry->efd_group);
2823         /* Take it out of per group rb tree */
2824         rb_erase(&entry->efd_node, &(db->bb_free_root));
2825         mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2826 
2827         /*
2828          * Clear the trimmed flag for the group so that the next
2829          * ext4_trim_fs can trim it.
2830          * If the volume is mounted with -o discard, online discard
2831          * is supported and the free blocks will be trimmed online.
2832          */
2833         if (!test_opt(sb, DISCARD))
2834                 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2835 
2836         if (!db->bb_free_root.rb_node) {
2837                 /* No more items in the per group rb tree
2838                  * balance refcounts from ext4_mb_free_metadata()
2839                  */
2840                 put_page(e4b.bd_buddy_page);
2841                 put_page(e4b.bd_bitmap_page);
2842         }
2843         ext4_unlock_group(sb, entry->efd_group);
2844         kmem_cache_free(ext4_free_data_cachep, entry);
2845         ext4_mb_unload_buddy(&e4b);
2846 
2847         mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2848 }
2849 
2850 int __init ext4_init_mballoc(void)
2851 {
2852         ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2853                                         SLAB_RECLAIM_ACCOUNT);
2854         if (ext4_pspace_cachep == NULL)
2855                 return -ENOMEM;
2856 
2857         ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2858                                     SLAB_RECLAIM_ACCOUNT);
2859         if (ext4_ac_cachep == NULL) {
2860                 kmem_cache_destroy(ext4_pspace_cachep);
2861                 return -ENOMEM;
2862         }
2863 
2864         ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2865                                            SLAB_RECLAIM_ACCOUNT);
2866         if (ext4_free_data_cachep == NULL) {
2867                 kmem_cache_destroy(ext4_pspace_cachep);
2868                 kmem_cache_destroy(ext4_ac_cachep);
2869                 return -ENOMEM;
2870         }
2871         return 0;
2872 }
2873 
2874 void ext4_exit_mballoc(void)
2875 {
2876         /*
2877          * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2878          * before destroying the slab cache.
2879          */
2880         rcu_barrier();
2881         kmem_cache_destroy(ext4_pspace_cachep);
2882         kmem_cache_destroy(ext4_ac_cachep);
2883         kmem_cache_destroy(ext4_free_data_cachep);
2884         ext4_groupinfo_destroy_slabs();
2885 }
2886 
2887 
2888 /*
2889  * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2890  * Returns 0 if success or error code
2891  */
2892 static noinline_for_stack int
2893 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2894                                 handle_t *handle, unsigned int reserv_clstrs)
2895 {
2896         struct buffer_head *bitmap_bh = NULL;
2897         struct ext4_group_desc *gdp;
2898         struct buffer_head *gdp_bh;
2899         struct ext4_sb_info *sbi;
2900         struct super_block *sb;
2901         ext4_fsblk_t block;
2902         int err, len;
2903 
2904         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2905         BUG_ON(ac->ac_b_ex.fe_len <= 0);
2906 
2907         sb = ac->ac_sb;
2908         sbi = EXT4_SB(sb);
2909 
2910         bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2911         if (IS_ERR(bitmap_bh)) {
2912                 err = PTR_ERR(bitmap_bh);
2913                 bitmap_bh = NULL;
2914                 goto out_err;
2915         }
2916 
2917         BUFFER_TRACE(bitmap_bh, "getting write access");
2918         err = ext4_journal_get_write_access(handle, bitmap_bh);
2919         if (err)
2920                 goto out_err;
2921 
2922         err = -EIO;
2923         gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2924         if (!gdp)
2925                 goto out_err;
2926 
2927         ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2928                         ext4_free_group_clusters(sb, gdp));
2929 
2930         BUFFER_TRACE(gdp_bh, "get_write_access");
2931         err = ext4_journal_get_write_access(handle, gdp_bh);
2932         if (err)
2933                 goto out_err;
2934 
2935         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2936 
2937         len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2938         if (!ext4_data_block_valid(sbi, block, len)) {
2939                 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2940                            "fs metadata", block, block+len);
2941                 /* File system mounted not to panic on error
2942                  * Fix the bitmap and return EFSCORRUPTED
2943                  * We leak some of the blocks here.
2944                  */
2945                 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2946                 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2947                               ac->ac_b_ex.fe_len);
2948                 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2949                 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2950                 if (!err)
2951                         err = -EFSCORRUPTED;
2952                 goto out_err;
2953         }
2954 
2955         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2956 #ifdef AGGRESSIVE_CHECK
2957         {
2958                 int i;
2959                 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2960                         BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2961                                                 bitmap_bh->b_data));
2962                 }
2963         }
2964 #endif
2965         ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2966                       ac->ac_b_ex.fe_len);
2967         if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2968                 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2969                 ext4_free_group_clusters_set(sb, gdp,
2970                                              ext4_free_clusters_after_init(sb,
2971                                                 ac->ac_b_ex.fe_group, gdp));
2972         }
2973         len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2974         ext4_free_group_clusters_set(sb, gdp, len);
2975         ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2976         ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2977 
2978         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2979         percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2980         /*
2981          * Now reduce the dirty block count also. Should not go negative
2982          */
2983         if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2984                 /* release all the reserved blocks if non delalloc */
2985                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2986                                    reserv_clstrs);
2987 
2988         if (sbi->s_log_groups_per_flex) {
2989                 ext4_group_t flex_group = ext4_flex_group(sbi,
2990                                                           ac->ac_b_ex.fe_group);
2991                 atomic64_sub(ac->ac_b_ex.fe_len,
2992                              &sbi->s_flex_groups[flex_group].free_clusters);
2993         }
2994 
2995         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2996         if (err)
2997                 goto out_err;
2998         err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2999 
3000 out_err:
3001         brelse(bitmap_bh);
3002         return err;
3003 }
3004 
3005 /*
3006  * here we normalize request for locality group
3007  * Group request are normalized to s_mb_group_prealloc, which goes to
3008  * s_strip if we set the same via mount option.
3009  * s_mb_group_prealloc can be configured via
3010  * /sys/fs/ext4/<partition>/mb_group_prealloc
3011  *
3012  * XXX: should we try to preallocate more than the group has now?
3013  */
3014 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3015 {
3016         struct super_block *sb = ac->ac_sb;
3017         struct ext4_locality_group *lg = ac->ac_lg;
3018 
3019         BUG_ON(lg == NULL);
3020         ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3021         mb_debug(1, "#%u: goal %u blocks for locality group\n",
3022                 current->pid, ac->ac_g_ex.fe_len);
3023 }
3024 
3025 /*
3026  * Normalization means making request better in terms of
3027  * size and alignment
3028  */
3029 static noinline_for_stack void
3030 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3031                                 struct ext4_allocation_request *ar)
3032 {
3033         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3034         int bsbits, max;
3035         ext4_lblk_t end;
3036         loff_t size, start_off;
3037         loff_t orig_size __maybe_unused;
3038         ext4_lblk_t start;
3039         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3040         struct ext4_prealloc_space *pa;
3041 
3042         /* do normalize only data requests, metadata requests
3043            do not need preallocation */
3044         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3045                 return;
3046 
3047         /* sometime caller may want exact blocks */
3048         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3049                 return;
3050 
3051         /* caller may indicate that preallocation isn't
3052          * required (it's a tail, for example) */
3053         if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3054                 return;
3055 
3056         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3057                 ext4_mb_normalize_group_request(ac);
3058                 return ;
3059         }
3060 
3061         bsbits = ac->ac_sb->s_blocksize_bits;
3062 
3063         /* first, let's learn actual file size
3064          * given current request is allocated */
3065         size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3066         size = size << bsbits;
3067         if (size < i_size_read(ac->ac_inode))
3068                 size = i_size_read(ac->ac_inode);
3069         orig_size = size;
3070 
3071         /* max size of free chunks */
3072         max = 2 << bsbits;
3073 
3074 #define NRL_CHECK_SIZE(req, size, max, chunk_size)      \
3075                 (req <= (size) || max <= (chunk_size))
3076 
3077         /* first, try to predict filesize */
3078         /* XXX: should this table be tunable? */
3079         start_off = 0;
3080         if (size <= 16 * 1024) {
3081                 size = 16 * 1024;
3082         } else if (size <= 32 * 1024) {
3083                 size = 32 * 1024;
3084         } else if (size <= 64 * 1024) {
3085                 size = 64 * 1024;
3086         } else if (size <= 128 * 1024) {
3087                 size = 128 * 1024;
3088         } else if (size <= 256 * 1024) {
3089                 size = 256 * 1024;
3090         } else if (size <= 512 * 1024) {
3091                 size = 512 * 1024;
3092         } else if (size <= 1024 * 1024) {
3093                 size = 1024 * 1024;
3094         } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3095                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3096                                                 (21 - bsbits)) << 21;
3097                 size = 2 * 1024 * 1024;
3098         } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3099                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3100                                                         (22 - bsbits)) << 22;
3101                 size = 4 * 1024 * 1024;
3102         } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3103                                         (8<<20)>>bsbits, max, 8 * 1024)) {
3104                 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3105                                                         (23 - bsbits)) << 23;
3106                 size = 8 * 1024 * 1024;
3107         } else {
3108                 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3109                 size      = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3110                                               ac->ac_o_ex.fe_len) << bsbits;
3111         }
3112         size = size >> bsbits;
3113         start = start_off >> bsbits;
3114 
3115         /* don't cover already allocated blocks in selected range */
3116         if (ar->pleft && start <= ar->lleft) {
3117                 size -= ar->lleft + 1 - start;
3118                 start = ar->lleft + 1;
3119         }
3120         if (ar->pright && start + size - 1 >= ar->lright)
3121                 size -= start + size - ar->lright;
3122 
3123         end = start + size;
3124 
3125         /* check we don't cross already preallocated blocks */
3126         rcu_read_lock();
3127         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3128                 ext4_lblk_t pa_end;
3129 
3130                 if (pa->pa_deleted)
3131                         continue;
3132                 spin_lock(&pa->pa_lock);
3133                 if (pa->pa_deleted) {
3134                         spin_unlock(&pa->pa_lock);
3135                         continue;
3136                 }
3137 
3138                 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3139                                                   pa->pa_len);
3140 
3141                 /* PA must not overlap original request */
3142                 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3143                         ac->ac_o_ex.fe_logical < pa->pa_lstart));
3144 
3145                 /* skip PAs this normalized request doesn't overlap with */
3146                 if (pa->pa_lstart >= end || pa_end <= start) {
3147                         spin_unlock(&pa->pa_lock);
3148                         continue;
3149                 }
3150                 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3151 
3152                 /* adjust start or end to be adjacent to this pa */
3153                 if (pa_end <= ac->ac_o_ex.fe_logical) {
3154                         BUG_ON(pa_end < start);
3155                         start = pa_end;
3156                 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3157                         BUG_ON(pa->pa_lstart > end);
3158                         end = pa->pa_lstart;
3159                 }
3160                 spin_unlock(&pa->pa_lock);
3161         }
3162         rcu_read_unlock();
3163         size = end - start;
3164 
3165         /* XXX: extra loop to check we really don't overlap preallocations */
3166         rcu_read_lock();
3167         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3168                 ext4_lblk_t pa_end;
3169 
3170                 spin_lock(&pa->pa_lock);
3171                 if (pa->pa_deleted == 0) {
3172                         pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3173                                                           pa->pa_len);
3174                         BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3175                 }
3176                 spin_unlock(&pa->pa_lock);
3177         }
3178         rcu_read_unlock();
3179 
3180         if (start + size <= ac->ac_o_ex.fe_logical &&
3181                         start > ac->ac_o_ex.fe_logical) {
3182                 ext4_msg(ac->ac_sb, KERN_ERR,
3183                          "start %lu, size %lu, fe_logical %lu",
3184                          (unsigned long) start, (unsigned long) size,
3185                          (unsigned long) ac->ac_o_ex.fe_logical);
3186                 BUG();
3187         }
3188         BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3189 
3190         /* now prepare goal request */
3191 
3192         /* XXX: is it better to align blocks WRT to logical
3193          * placement or satisfy big request as is */
3194         ac->ac_g_ex.fe_logical = start;
3195         ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3196 
3197         /* define goal start in order to merge */
3198         if (ar->pright && (ar->lright == (start + size))) {
3199                 /* merge to the right */
3200                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3201                                                 &ac->ac_f_ex.fe_group,
3202                                                 &ac->ac_f_ex.fe_start);
3203                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3204         }
3205         if (ar->pleft && (ar->lleft + 1 == start)) {
3206                 /* merge to the left */
3207                 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3208                                                 &ac->ac_f_ex.fe_group,
3209                                                 &ac->ac_f_ex.fe_start);
3210                 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3211         }
3212 
3213         mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3214                 (unsigned) orig_size, (unsigned) start);
3215 }
3216 
3217 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3218 {
3219         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3220 
3221         if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3222                 atomic_inc(&sbi->s_bal_reqs);
3223                 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3224                 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3225                         atomic_inc(&sbi->s_bal_success);
3226                 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3227                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3228                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3229                         atomic_inc(&sbi->s_bal_goals);
3230                 if (ac->ac_found > sbi->s_mb_max_to_scan)
3231                         atomic_inc(&sbi->s_bal_breaks);
3232         }
3233 
3234         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3235                 trace_ext4_mballoc_alloc(ac);
3236         else
3237                 trace_ext4_mballoc_prealloc(ac);
3238 }
3239 
3240 /*
3241  * Called on failure; free up any blocks from the inode PA for this
3242  * context.  We don't need this for MB_GROUP_PA because we only change
3243  * pa_free in ext4_mb_release_context(), but on failure, we've already
3244  * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3245  */
3246 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3247 {
3248         struct ext4_prealloc_space *pa = ac->ac_pa;
3249         struct ext4_buddy e4b;
3250         int err;
3251 
3252         if (pa == NULL) {
3253                 if (ac->ac_f_ex.fe_len == 0)
3254                         return;
3255                 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3256                 if (err) {
3257                         /*
3258                          * This should never happen since we pin the
3259                          * pages in the ext4_allocation_context so
3260                          * ext4_mb_load_buddy() should never fail.
3261                          */
3262                         WARN(1, "mb_load_buddy failed (%d)", err);
3263                         return;
3264                 }
3265                 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3266                 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3267                                ac->ac_f_ex.fe_len);
3268                 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3269                 ext4_mb_unload_buddy(&e4b);
3270                 return;
3271         }
3272         if (pa->pa_type == MB_INODE_PA)
3273                 pa->pa_free += ac->ac_b_ex.fe_len;
3274 }
3275 
3276 /*
3277  * use blocks preallocated to inode
3278  */
3279 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3280                                 struct ext4_prealloc_space *pa)
3281 {
3282         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3283         ext4_fsblk_t start;
3284         ext4_fsblk_t end;
3285         int len;
3286 
3287         /* found preallocated blocks, use them */
3288         start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3289         end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3290                   start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3291         len = EXT4_NUM_B2C(sbi, end - start);
3292         ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3293                                         &ac->ac_b_ex.fe_start);
3294         ac->ac_b_ex.fe_len = len;
3295         ac->ac_status = AC_STATUS_FOUND;
3296         ac->ac_pa = pa;
3297 
3298         BUG_ON(start < pa->pa_pstart);
3299         BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3300         BUG_ON(pa->pa_free < len);
3301         pa->pa_free -= len;
3302 
3303         mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3304 }
3305 
3306 /*
3307  * use blocks preallocated to locality group
3308  */
3309 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3310                                 struct ext4_prealloc_space *pa)
3311 {
3312         unsigned int len = ac->ac_o_ex.fe_len;
3313 
3314         ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3315                                         &ac->ac_b_ex.fe_group,
3316                                         &ac->ac_b_ex.fe_start);
3317         ac->ac_b_ex.fe_len = len;
3318         ac->ac_status = AC_STATUS_FOUND;
3319         ac->ac_pa = pa;
3320 
3321         /* we don't correct pa_pstart or pa_plen here to avoid
3322          * possible race when the group is being loaded concurrently
3323          * instead we correct pa later, after blocks are marked
3324          * in on-disk bitmap -- see ext4_mb_release_context()
3325          * Other CPUs are prevented from allocating from this pa by lg_mutex
3326          */
3327         mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3328 }
3329 
3330 /*
3331  * Return the prealloc space that have minimal distance
3332  * from the goal block. @cpa is the prealloc
3333  * space that is having currently known minimal distance
3334  * from the goal block.
3335  */
3336 static struct ext4_prealloc_space *
3337 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3338                         struct ext4_prealloc_space *pa,
3339                         struct ext4_prealloc_space *cpa)
3340 {
3341         ext4_fsblk_t cur_distance, new_distance;
3342 
3343         if (cpa == NULL) {
3344                 atomic_inc(&pa->pa_count);
3345                 return pa;
3346         }
3347         cur_distance = abs(goal_block - cpa->pa_pstart);
3348         new_distance = abs(goal_block - pa->pa_pstart);
3349 
3350         if (cur_distance <= new_distance)
3351                 return cpa;
3352 
3353         /* drop the previous reference */
3354         atomic_dec(&cpa->pa_count);
3355         atomic_inc(&pa->pa_count);
3356         return pa;
3357 }
3358 
3359 /*
3360  * search goal blocks in preallocated space
3361  */
3362 static noinline_for_stack int
3363 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3364 {
3365         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3366         int order, i;
3367         struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3368         struct ext4_locality_group *lg;
3369         struct ext4_prealloc_space *pa, *cpa = NULL;
3370         ext4_fsblk_t goal_block;
3371 
3372         /* only data can be preallocated */
3373         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3374                 return 0;
3375 
3376         /* first, try per-file preallocation */
3377         rcu_read_lock();
3378         list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3379 
3380                 /* all fields in this condition don't change,
3381                  * so we can skip locking for them */
3382                 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3383                     ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3384                                                EXT4_C2B(sbi, pa->pa_len)))
3385                         continue;
3386 
3387                 /* non-extent files can't have physical blocks past 2^32 */
3388                 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3389                     (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3390                      EXT4_MAX_BLOCK_FILE_PHYS))
3391                         continue;
3392 
3393                 /* found preallocated blocks, use them */
3394                 spin_lock(&pa->pa_lock);
3395                 if (pa->pa_deleted == 0 && pa->pa_free) {
3396                         atomic_inc(&pa->pa_count);
3397                         ext4_mb_use_inode_pa(ac, pa);
3398                         spin_unlock(&pa->pa_lock);
3399                         ac->ac_criteria = 10;
3400                         rcu_read_unlock();
3401                         return 1;
3402                 }
3403                 spin_unlock(&pa->pa_lock);
3404         }
3405         rcu_read_unlock();
3406 
3407         /* can we use group allocation? */
3408         if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3409                 return 0;
3410 
3411         /* inode may have no locality group for some reason */
3412         lg = ac->ac_lg;
3413         if (lg == NULL)
3414                 return 0;
3415         order  = fls(ac->ac_o_ex.fe_len) - 1;
3416         if (order > PREALLOC_TB_SIZE - 1)
3417                 /* The max size of hash table is PREALLOC_TB_SIZE */
3418                 order = PREALLOC_TB_SIZE - 1;
3419 
3420         goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3421         /*
3422          * search for the prealloc space that is having
3423          * minimal distance from the goal block.
3424          */
3425         for (i = order; i < PREALLOC_TB_SIZE; i++) {
3426                 rcu_read_lock();
3427                 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3428                                         pa_inode_list) {
3429                         spin_lock(&pa->pa_lock);
3430                         if (pa->pa_deleted == 0 &&
3431                                         pa->pa_free >= ac->ac_o_ex.fe_len) {
3432 
3433                                 cpa = ext4_mb_check_group_pa(goal_block,
3434                                                                 pa, cpa);
3435                         }
3436                         spin_unlock(&pa->pa_lock);
3437                 }
3438                 rcu_read_unlock();
3439         }
3440         if (cpa) {
3441                 ext4_mb_use_group_pa(ac, cpa);
3442                 ac->ac_criteria = 20;
3443                 return 1;
3444         }
3445         return 0;
3446 }
3447 
3448 /*
3449  * the function goes through all block freed in the group
3450  * but not yet committed and marks them used in in-core bitmap.
3451  * buddy must be generated from this bitmap
3452  * Need to be called with the ext4 group lock held
3453  */
3454 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3455                                                 ext4_group_t group)
3456 {
3457         struct rb_node *n;
3458         struct ext4_group_info *grp;
3459         struct ext4_free_data *entry;
3460 
3461         grp = ext4_get_group_info(sb, group);
3462         n = rb_first(&(grp->bb_free_root));
3463 
3464         while (n) {
3465                 entry = rb_entry(n, struct ext4_free_data, efd_node);
3466                 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3467                 n = rb_next(n);
3468         }
3469         return;
3470 }
3471 
3472 /*
3473  * the function goes through all preallocation in this group and marks them
3474  * used in in-core bitmap. buddy must be generated from this bitmap
3475  * Need to be called with ext4 group lock held
3476  */
3477 static noinline_for_stack
3478 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3479                                         ext4_group_t group)
3480 {
3481         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3482         struct ext4_prealloc_space *pa;
3483         struct list_head *cur;
3484         ext4_group_t groupnr;
3485         ext4_grpblk_t start;
3486         int preallocated = 0;
3487         int len;
3488 
3489         /* all form of preallocation discards first load group,
3490          * so the only competing code is preallocation use.
3491          * we don't need any locking here
3492          * notice we do NOT ignore preallocations with pa_deleted
3493          * otherwise we could leave used blocks available for
3494          * allocation in buddy when concurrent ext4_mb_put_pa()
3495          * is dropping preallocation
3496          */
3497         list_for_each(cur, &grp->bb_prealloc_list) {
3498                 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3499                 spin_lock(&pa->pa_lock);
3500                 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3501                                              &groupnr, &start);
3502                 len = pa->pa_len;
3503                 spin_unlock(&pa->pa_lock);
3504                 if (unlikely(len == 0))
3505                         continue;
3506                 BUG_ON(groupnr != group);
3507                 ext4_set_bits(bitmap, start, len);
3508                 preallocated += len;
3509         }
3510         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3511 }
3512 
3513 static void ext4_mb_pa_callback(struct rcu_head *head)
3514 {
3515         struct ext4_prealloc_space *pa;
3516         pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3517 
3518         BUG_ON(atomic_read(&pa->pa_count));
3519         BUG_ON(pa->pa_deleted == 0);
3520         kmem_cache_free(ext4_pspace_cachep, pa);
3521 }
3522 
3523 /*
3524  * drops a reference to preallocated space descriptor
3525  * if this was the last reference and the space is consumed
3526  */
3527 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3528                         struct super_block *sb, struct ext4_prealloc_space *pa)
3529 {
3530         ext4_group_t grp;
3531         ext4_fsblk_t grp_blk;
3532 
3533         /* in this short window concurrent discard can set pa_deleted */
3534         spin_lock(&pa->pa_lock);
3535         if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3536                 spin_unlock(&pa->pa_lock);
3537                 return;
3538         }
3539 
3540         if (pa->pa_deleted == 1) {
3541                 spin_unlock(&pa->pa_lock);
3542                 return;
3543         }
3544 
3545         pa->pa_deleted = 1;
3546         spin_unlock(&pa->pa_lock);
3547 
3548         grp_blk = pa->pa_pstart;
3549         /*
3550          * If doing group-based preallocation, pa_pstart may be in the
3551          * next group when pa is used up
3552          */
3553         if (pa->pa_type == MB_GROUP_PA)
3554                 grp_blk--;
3555 
3556         grp = ext4_get_group_number(sb, grp_blk);
3557 
3558         /*
3559          * possible race:
3560          *
3561          *  P1 (buddy init)                     P2 (regular allocation)
3562          *                                      find block B in PA
3563          *  copy on-disk bitmap to buddy
3564          *                                      mark B in on-disk bitmap
3565          *                                      drop PA from group
3566          *  mark all PAs in buddy
3567          *
3568          * thus, P1 initializes buddy with B available. to prevent this
3569          * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3570          * against that pair
3571          */
3572         ext4_lock_group(sb, grp);
3573         list_del(&pa->pa_group_list);
3574         ext4_unlock_group(sb, grp);
3575 
3576         spin_lock(pa->pa_obj_lock);
3577         list_del_rcu(&pa->pa_inode_list);
3578         spin_unlock(pa->pa_obj_lock);
3579 
3580         call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3581 }
3582 
3583 /*
3584  * creates new preallocated space for given inode
3585  */
3586 static noinline_for_stack int
3587 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3588 {
3589         struct super_block *sb = ac->ac_sb;
3590         struct ext4_sb_info *sbi = EXT4_SB(sb);
3591         struct ext4_prealloc_space *pa;
3592         struct ext4_group_info *grp;
3593         struct ext4_inode_info *ei;
3594 
3595         /* preallocate only when found space is larger then requested */
3596         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3597         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3598         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3599 
3600         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3601         if (pa == NULL)
3602                 return -ENOMEM;
3603 
3604         if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3605                 int winl;
3606                 int wins;
3607                 int win;
3608                 int offs;
3609 
3610                 /* we can't allocate as much as normalizer wants.
3611                  * so, found space must get proper lstart
3612                  * to cover original request */
3613                 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3614                 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3615 
3616                 /* we're limited by original request in that
3617                  * logical block must be covered any way
3618                  * winl is window we can move our chunk within */
3619                 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3620 
3621                 /* also, we should cover whole original request */
3622                 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3623 
3624                 /* the smallest one defines real window */
3625                 win = min(winl, wins);
3626 
3627                 offs = ac->ac_o_ex.fe_logical %
3628                         EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3629                 if (offs && offs < win)
3630                         win = offs;
3631 
3632                 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3633                         EXT4_NUM_B2C(sbi, win);
3634                 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3635                 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3636         }
3637 
3638         /* preallocation can change ac_b_ex, thus we store actually
3639          * allocated blocks for history */
3640         ac->ac_f_ex = ac->ac_b_ex;
3641 
3642         pa->pa_lstart = ac->ac_b_ex.fe_logical;
3643         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3644         pa->pa_len = ac->ac_b_ex.fe_len;
3645         pa->pa_free = pa->pa_len;
3646         atomic_set(&pa->pa_count, 1);
3647         spin_lock_init(&pa->pa_lock);
3648         INIT_LIST_HEAD(&pa->pa_inode_list);
3649         INIT_LIST_HEAD(&pa->pa_group_list);
3650         pa->pa_deleted = 0;
3651         pa->pa_type = MB_INODE_PA;
3652 
3653         mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3654                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3655         trace_ext4_mb_new_inode_pa(ac, pa);
3656 
3657         ext4_mb_use_inode_pa(ac, pa);
3658         atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3659 
3660         ei = EXT4_I(ac->ac_inode);
3661         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3662 
3663         pa->pa_obj_lock = &ei->i_prealloc_lock;
3664         pa->pa_inode = ac->ac_inode;
3665 
3666         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3667         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3668         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3669 
3670         spin_lock(pa->pa_obj_lock);
3671         list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3672         spin_unlock(pa->pa_obj_lock);
3673 
3674         return 0;
3675 }
3676 
3677 /*
3678  * creates new preallocated space for locality group inodes belongs to
3679  */
3680 static noinline_for_stack int
3681 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3682 {
3683         struct super_block *sb = ac->ac_sb;
3684         struct ext4_locality_group *lg;
3685         struct ext4_prealloc_space *pa;
3686         struct ext4_group_info *grp;
3687 
3688         /* preallocate only when found space is larger then requested */
3689         BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3690         BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3691         BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3692 
3693         BUG_ON(ext4_pspace_cachep == NULL);
3694         pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3695         if (pa == NULL)
3696                 return -ENOMEM;
3697 
3698         /* preallocation can change ac_b_ex, thus we store actually
3699          * allocated blocks for history */
3700         ac->ac_f_ex = ac->ac_b_ex;
3701 
3702         pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3703         pa->pa_lstart = pa->pa_pstart;
3704         pa->pa_len = ac->ac_b_ex.fe_len;
3705         pa->pa_free = pa->pa_len;
3706         atomic_set(&pa->pa_count, 1);
3707         spin_lock_init(&pa->pa_lock);
3708         INIT_LIST_HEAD(&pa->pa_inode_list);
3709         INIT_LIST_HEAD(&pa->pa_group_list);
3710         pa->pa_deleted = 0;
3711         pa->pa_type = MB_GROUP_PA;
3712 
3713         mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3714                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3715         trace_ext4_mb_new_group_pa(ac, pa);
3716 
3717         ext4_mb_use_group_pa(ac, pa);
3718         atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3719 
3720         grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3721         lg = ac->ac_lg;
3722         BUG_ON(lg == NULL);
3723 
3724         pa->pa_obj_lock = &lg->lg_prealloc_lock;
3725         pa->pa_inode = NULL;
3726 
3727         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3728         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3729         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3730 
3731         /*
3732          * We will later add the new pa to the right bucket
3733          * after updating the pa_free in ext4_mb_release_context
3734          */
3735         return 0;
3736 }
3737 
3738 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3739 {
3740         int err;
3741 
3742         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3743                 err = ext4_mb_new_group_pa(ac);
3744         else
3745                 err = ext4_mb_new_inode_pa(ac);
3746         return err;
3747 }
3748 
3749 /*
3750  * finds all unused blocks in on-disk bitmap, frees them in
3751  * in-core bitmap and buddy.
3752  * @pa must be unlinked from inode and group lists, so that
3753  * nobody else can find/use it.
3754  * the caller MUST hold group/inode locks.
3755  * TODO: optimize the case when there are no in-core structures yet
3756  */
3757 static noinline_for_stack int
3758 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3759                         struct ext4_prealloc_space *pa)
3760 {
3761         struct super_block *sb = e4b->bd_sb;
3762         struct ext4_sb_info *sbi = EXT4_SB(sb);
3763         unsigned int end;
3764         unsigned int next;
3765         ext4_group_t group;
3766         ext4_grpblk_t bit;
3767         unsigned long long grp_blk_start;
3768         int err = 0;
3769         int free = 0;
3770 
3771         BUG_ON(pa->pa_deleted == 0);
3772         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3773         grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3774         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3775         end = bit + pa->pa_len;
3776 
3777         while (bit < end) {
3778                 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3779                 if (bit >= end)
3780                         break;
3781                 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3782                 mb_debug(1, "    free preallocated %u/%u in group %u\n",
3783                          (unsigned) ext4_group_first_block_no(sb, group) + bit,
3784                          (unsigned) next - bit, (unsigned) group);
3785                 free += next - bit;
3786 
3787                 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3788                 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3789                                                     EXT4_C2B(sbi, bit)),
3790                                                next - bit);
3791                 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3792                 bit = next + 1;
3793         }
3794         if (free != pa->pa_free) {
3795                 ext4_msg(e4b->bd_sb, KERN_CRIT,
3796                          "pa %p: logic %lu, phys. %lu, len %lu",
3797                          pa, (unsigned long) pa->pa_lstart,
3798                          (unsigned long) pa->pa_pstart,
3799                          (unsigned long) pa->pa_len);
3800                 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3801                                         free, pa->pa_free);
3802                 /*
3803                  * pa is already deleted so we use the value obtained
3804                  * from the bitmap and continue.
3805                  */
3806         }
3807         atomic_add(free, &sbi->s_mb_discarded);
3808 
3809         return err;
3810 }
3811 
3812 static noinline_for_stack int
3813 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3814                                 struct ext4_prealloc_space *pa)
3815 {
3816         struct super_block *sb = e4b->bd_sb;
3817         ext4_group_t group;
3818         ext4_grpblk_t bit;
3819 
3820         trace_ext4_mb_release_group_pa(sb, pa);
3821         BUG_ON(pa->pa_deleted == 0);
3822         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3823         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3824         mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3825         atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3826         trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3827 
3828         return 0;
3829 }
3830 
3831 /*
3832  * releases all preallocations in given group
3833  *
3834  * first, we need to decide discard policy:
3835  * - when do we discard
3836  *   1) ENOSPC
3837  * - how many do we discard
3838  *   1) how many requested
3839  */
3840 static noinline_for_stack int
3841 ext4_mb_discard_group_preallocations(struct super_block *sb,
3842                                         ext4_group_t group, int needed)
3843 {
3844         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3845         struct buffer_head *bitmap_bh = NULL;
3846         struct ext4_prealloc_space *pa, *tmp;
3847         struct list_head list;
3848         struct ext4_buddy e4b;
3849         int err;
3850         int busy = 0;
3851         int free = 0;
3852 
3853         mb_debug(1, "discard preallocation for group %u\n", group);
3854 
3855         if (list_empty(&grp->bb_prealloc_list))
3856                 return 0;
3857 
3858         bitmap_bh = ext4_read_block_bitmap(sb, group);
3859         if (IS_ERR(bitmap_bh)) {
3860                 err = PTR_ERR(bitmap_bh);
3861                 ext4_error(sb, "Error %d reading block bitmap for %u",
3862                            err, group);
3863                 return 0;
3864         }
3865 
3866         err = ext4_mb_load_buddy(sb, group, &e4b);
3867         if (err) {
3868                 ext4_error(sb, "Error loading buddy information for %u", group);
3869                 put_bh(bitmap_bh);
3870                 return 0;
3871         }
3872 
3873         if (needed == 0)
3874                 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3875 
3876         INIT_LIST_HEAD(&list);
3877 repeat:
3878         ext4_lock_group(sb, group);
3879         list_for_each_entry_safe(pa, tmp,
3880                                 &grp->bb_prealloc_list, pa_group_list) {
3881                 spin_lock(&pa->pa_lock);
3882                 if (atomic_read(&pa->pa_count)) {
3883                         spin_unlock(&pa->pa_lock);
3884                         busy = 1;
3885                         continue;
3886                 }
3887                 if (pa->pa_deleted) {
3888                         spin_unlock(&pa->pa_lock);
3889                         continue;
3890                 }
3891 
3892                 /* seems this one can be freed ... */
3893                 pa->pa_deleted = 1;
3894 
3895                 /* we can trust pa_free ... */
3896                 free += pa->pa_free;
3897 
3898                 spin_unlock(&pa->pa_lock);
3899 
3900                 list_del(&pa->pa_group_list);
3901                 list_add(&pa->u.pa_tmp_list, &list);
3902         }
3903 
3904         /* if we still need more blocks and some PAs were used, try again */
3905         if (free < needed && busy) {
3906                 busy = 0;
3907                 ext4_unlock_group(sb, group);
3908                 cond_resched();
3909                 goto repeat;
3910         }
3911 
3912         /* found anything to free? */
3913         if (list_empty(&list)) {
3914                 BUG_ON(free != 0);
3915                 goto out;
3916         }
3917 
3918         /* now free all selected PAs */
3919         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3920 
3921                 /* remove from object (inode or locality group) */
3922                 spin_lock(pa->pa_obj_lock);
3923                 list_del_rcu(&pa->pa_inode_list);
3924                 spin_unlock(pa->pa_obj_lock);
3925 
3926                 if (pa->pa_type == MB_GROUP_PA)
3927                         ext4_mb_release_group_pa(&e4b, pa);
3928                 else
3929                         ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3930 
3931                 list_del(&pa->u.pa_tmp_list);
3932                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3933         }
3934 
3935 out:
3936         ext4_unlock_group(sb, group);
3937         ext4_mb_unload_buddy(&e4b);
3938         put_bh(bitmap_bh);
3939         return free;
3940 }
3941 
3942 /*
3943  * releases all non-used preallocated blocks for given inode
3944  *
3945  * It's important to discard preallocations under i_data_sem
3946  * We don't want another block to be served from the prealloc
3947  * space when we are discarding the inode prealloc space.
3948  *
3949  * FIXME!! Make sure it is valid at all the call sites
3950  */
3951 void ext4_discard_preallocations(struct inode *inode)
3952 {
3953         struct ext4_inode_info *ei = EXT4_I(inode);
3954         struct super_block *sb = inode->i_sb;
3955         struct buffer_head *bitmap_bh = NULL;
3956         struct ext4_prealloc_space *pa, *tmp;
3957         ext4_group_t group = 0;
3958         struct list_head list;
3959         struct ext4_buddy e4b;
3960         int err;
3961 
3962         if (!S_ISREG(inode->i_mode)) {
3963                 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3964                 return;
3965         }
3966 
3967         mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3968         trace_ext4_discard_preallocations(inode);
3969 
3970         INIT_LIST_HEAD(&list);
3971 
3972 repeat:
3973         /* first, collect all pa's in the inode */
3974         spin_lock(&ei->i_prealloc_lock);
3975         while (!list_empty(&ei->i_prealloc_list)) {
3976                 pa = list_entry(ei->i_prealloc_list.next,
3977                                 struct ext4_prealloc_space, pa_inode_list);
3978                 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3979                 spin_lock(&pa->pa_lock);
3980                 if (atomic_read(&pa->pa_count)) {
3981                         /* this shouldn't happen often - nobody should
3982                          * use preallocation while we're discarding it */
3983                         spin_unlock(&pa->pa_lock);
3984                         spin_unlock(&ei->i_prealloc_lock);
3985                         ext4_msg(sb, KERN_ERR,
3986                                  "uh-oh! used pa while discarding");
3987                         WARN_ON(1);
3988                         schedule_timeout_uninterruptible(HZ);
3989                         goto repeat;
3990 
3991                 }
3992                 if (pa->pa_deleted == 0) {
3993                         pa->pa_deleted = 1;
3994                         spin_unlock(&pa->pa_lock);
3995                         list_del_rcu(&pa->pa_inode_list);
3996                         list_add(&pa->u.pa_tmp_list, &list);
3997                         continue;
3998                 }
3999 
4000                 /* someone is deleting pa right now */
4001                 spin_unlock(&pa->pa_lock);
4002                 spin_unlock(&ei->i_prealloc_lock);
4003 
4004                 /* we have to wait here because pa_deleted
4005                  * doesn't mean pa is already unlinked from
4006                  * the list. as we might be called from
4007                  * ->clear_inode() the inode will get freed
4008                  * and concurrent thread which is unlinking
4009                  * pa from inode's list may access already
4010                  * freed memory, bad-bad-bad */
4011 
4012                 /* XXX: if this happens too often, we can
4013                  * add a flag to force wait only in case
4014                  * of ->clear_inode(), but not in case of
4015                  * regular truncate */
4016                 schedule_timeout_uninterruptible(HZ);
4017                 goto repeat;
4018         }
4019         spin_unlock(&ei->i_prealloc_lock);
4020 
4021         list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4022                 BUG_ON(pa->pa_type != MB_INODE_PA);
4023                 group = ext4_get_group_number(sb, pa->pa_pstart);
4024 
4025                 err = ext4_mb_load_buddy(sb, group, &e4b);
4026                 if (err) {
4027                         ext4_error(sb, "Error loading buddy information for %u",
4028                                         group);
4029                         continue;
4030                 }
4031 
4032                 bitmap_bh = ext4_read_block_bitmap(sb, group);
4033                 if (IS_ERR(bitmap_bh)) {
4034                         err = PTR_ERR(bitmap_bh);
4035                         ext4_error(sb, "Error %d reading block bitmap for %u",
4036                                         err, group);
4037                         ext4_mb_unload_buddy(&e4b);
4038                         continue;
4039                 }
4040 
4041                 ext4_lock_group(sb, group);
4042                 list_del(&pa->pa_group_list);
4043                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4044                 ext4_unlock_group(sb, group);
4045 
4046                 ext4_mb_unload_buddy(&e4b);
4047                 put_bh(bitmap_bh);
4048 
4049                 list_del(&pa->u.pa_tmp_list);
4050                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4051         }
4052 }
4053 
4054 #ifdef CONFIG_EXT4_DEBUG
4055 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4056 {
4057         struct super_block *sb = ac->ac_sb;
4058         ext4_group_t ngroups, i;
4059 
4060         if (!ext4_mballoc_debug ||
4061             (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4062                 return;
4063 
4064         ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4065                         " Allocation context details:");
4066         ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4067                         ac->ac_status, ac->ac_flags);
4068         ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4069                         "goal %lu/%lu/%lu@%lu, "
4070                         "best %lu/%lu/%lu@%lu cr %d",
4071                         (unsigned long)ac->ac_o_ex.fe_group,
4072                         (unsigned long)ac->ac_o_ex.fe_start,
4073                         (unsigned long)ac->ac_o_ex.fe_len,
4074                         (unsigned long)ac->ac_o_ex.fe_logical,
4075                         (unsigned long)ac->ac_g_ex.fe_group,
4076                         (unsigned long)ac->ac_g_ex.fe_start,
4077                         (unsigned long)ac->ac_g_ex.fe_len,
4078                         (unsigned long)ac->ac_g_ex.fe_logical,
4079                         (unsigned long)ac->ac_b_ex.fe_group,
4080                         (unsigned long)ac->ac_b_ex.fe_start,
4081                         (unsigned long)ac->ac_b_ex.fe_len,
4082                         (unsigned long)ac->ac_b_ex.fe_logical,
4083                         (int)ac->ac_criteria);
4084         ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4085         ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4086         ngroups = ext4_get_groups_count(sb);
4087         for (i = 0; i < ngroups; i++) {
4088                 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4089                 struct ext4_prealloc_space *pa;
4090                 ext4_grpblk_t start;
4091                 struct list_head *cur;
4092                 ext4_lock_group(sb, i);
4093                 list_for_each(cur, &grp->bb_prealloc_list) {
4094                         pa = list_entry(cur, struct ext4_prealloc_space,
4095                                         pa_group_list);
4096                         spin_lock(&pa->pa_lock);
4097                         ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4098                                                      NULL, &start);
4099                         spin_unlock(&pa->pa_lock);
4100                         printk(KERN_ERR "PA:%u:%d:%u \n", i,
4101                                start, pa->pa_len);
4102                 }
4103                 ext4_unlock_group(sb, i);
4104 
4105                 if (grp->bb_free == 0)
4106                         continue;
4107                 printk(KERN_ERR "%u: %d/%d \n",
4108                        i, grp->bb_free, grp->bb_fragments);
4109         }
4110         printk(KERN_ERR "\n");
4111 }
4112 #else
4113 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4114 {
4115         return;
4116 }
4117 #endif
4118 
4119 /*
4120  * We use locality group preallocation for small size file. The size of the
4121  * file is determined by the current size or the resulting size after
4122  * allocation which ever is larger
4123  *
4124  * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4125  */
4126 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4127 {
4128         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4129         int bsbits = ac->ac_sb->s_blocksize_bits;
4130         loff_t size, isize;
4131 
4132         if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4133                 return;
4134 
4135         if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4136                 return;
4137 
4138         size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4139         isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4140                 >> bsbits;
4141 
4142         if ((size == isize) &&
4143             !ext4_fs_is_busy(sbi) &&
4144             (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4145                 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4146                 return;
4147         }
4148 
4149         if (sbi->s_mb_group_prealloc <= 0) {
4150                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4151                 return;
4152         }
4153 
4154         /* don't use group allocation for large files */
4155         size = max(size, isize);
4156         if (size > sbi->s_mb_stream_request) {
4157                 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4158                 return;
4159         }
4160 
4161         BUG_ON(ac->ac_lg != NULL);
4162         /*
4163          * locality group prealloc space are per cpu. The reason for having
4164          * per cpu locality group is to reduce the contention between block
4165          * request from multiple CPUs.
4166          */
4167         ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4168 
4169         /* we're going to use group allocation */
4170         ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4171 
4172         /* serialize all allocations in the group */
4173         mutex_lock(&ac->ac_lg->lg_mutex);
4174 }
4175 
4176 static noinline_for_stack int
4177 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4178                                 struct ext4_allocation_request *ar)
4179 {
4180         struct super_block *sb = ar->inode->i_sb;
4181         struct ext4_sb_info *sbi = EXT4_SB(sb);
4182         struct ext4_super_block *es = sbi->s_es;
4183         ext4_group_t group;
4184         unsigned int len;
4185         ext4_fsblk_t goal;
4186         ext4_grpblk_t block;
4187 
4188         /* we can't allocate > group size */
4189         len = ar->len;
4190 
4191         /* just a dirty hack to filter too big requests  */
4192         if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4193                 len = EXT4_CLUSTERS_PER_GROUP(sb);
4194 
4195         /* start searching from the goal */
4196         goal = ar->goal;
4197         if (goal < le32_to_cpu(es->s_first_data_block) ||
4198                         goal >= ext4_blocks_count(es))
4199                 goal = le32_to_cpu(es->s_first_data_block);
4200         ext4_get_group_no_and_offset(sb, goal, &group, &block);
4201 
4202         /* set up allocation goals */
4203         ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4204         ac->ac_status = AC_STATUS_CONTINUE;
4205         ac->ac_sb = sb;
4206         ac->ac_inode = ar->inode;
4207         ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4208         ac->ac_o_ex.fe_group = group;
4209         ac->ac_o_ex.fe_start = block;
4210         ac->ac_o_ex.fe_len = len;
4211         ac->ac_g_ex = ac->ac_o_ex;
4212         ac->ac_flags = ar->flags;
4213 
4214         /* we have to define context: we'll we work with a file or
4215          * locality group. this is a policy, actually */
4216         ext4_mb_group_or_file(ac);
4217 
4218         mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4219                         "left: %u/%u, right %u/%u to %swritable\n",
4220                         (unsigned) ar->len, (unsigned) ar->logical,
4221                         (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4222                         (unsigned) ar->lleft, (unsigned) ar->pleft,
4223                         (unsigned) ar->lright, (unsigned) ar->pright,
4224                         atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4225         return 0;
4226 
4227 }
4228 
4229 static noinline_for_stack void
4230 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4231                                         struct ext4_locality_group *lg,
4232                                         int order, int total_entries)
4233 {
4234         ext4_group_t group = 0;
4235         struct ext4_buddy e4b;
4236         struct list_head discard_list;
4237         struct ext4_prealloc_space *pa, *tmp;
4238 
4239         mb_debug(1, "discard locality group preallocation\n");
4240 
4241         INIT_LIST_HEAD(&discard_list);
4242 
4243         spin_lock(&lg->lg_prealloc_lock);
4244         list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4245                                                 pa_inode_list) {
4246                 spin_lock(&pa->pa_lock);
4247                 if (atomic_read(&pa->pa_count)) {
4248                         /*
4249                          * This is the pa that we just used
4250                          * for block allocation. So don't
4251                          * free that
4252                          */
4253                         spin_unlock(&pa->pa_lock);
4254                         continue;
4255                 }
4256                 if (pa->pa_deleted) {
4257                         spin_unlock(&pa->pa_lock);
4258                         continue;
4259                 }
4260                 /* only lg prealloc space */
4261                 BUG_ON(pa->pa_type != MB_GROUP_PA);
4262 
4263                 /* seems this one can be freed ... */
4264                 pa->pa_deleted = 1;
4265                 spin_unlock(&pa->pa_lock);
4266 
4267                 list_del_rcu(&pa->pa_inode_list);
4268                 list_add(&pa->u.pa_tmp_list, &discard_list);
4269 
4270                 total_entries--;
4271                 if (total_entries <= 5) {
4272                         /*
4273                          * we want to keep only 5 entries
4274                          * allowing it to grow to 8. This
4275                          * mak sure we don't call discard
4276                          * soon for this list.
4277                          */
4278                         break;
4279                 }
4280         }
4281         spin_unlock(&lg->lg_prealloc_lock);
4282 
4283         list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4284 
4285                 group = ext4_get_group_number(sb, pa->pa_pstart);
4286                 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4287                         ext4_error(sb, "Error loading buddy information for %u",
4288                                         group);
4289                         continue;
4290                 }
4291                 ext4_lock_group(sb, group);
4292                 list_del(&pa->pa_group_list);
4293                 ext4_mb_release_group_pa(&e4b, pa);
4294                 ext4_unlock_group(sb, group);
4295 
4296                 ext4_mb_unload_buddy(&e4b);
4297                 list_del(&pa->u.pa_tmp_list);
4298                 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4299         }
4300 }
4301 
4302 /*
4303  * We have incremented pa_count. So it cannot be freed at this
4304  * point. Also we hold lg_mutex. So no parallel allocation is
4305  * possible from this lg. That means pa_free cannot be updated.
4306  *
4307  * A parallel ext4_mb_discard_group_preallocations is possible.
4308  * which can cause the lg_prealloc_list to be updated.
4309  */
4310 
4311 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4312 {
4313         int order, added = 0, lg_prealloc_count = 1;
4314         struct super_block *sb = ac->ac_sb;
4315         struct ext4_locality_group *lg = ac->ac_lg;
4316         struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4317 
4318         order = fls(pa->pa_free) - 1;
4319         if (order > PREALLOC_TB_SIZE - 1)
4320                 /* The max size of hash table is PREALLOC_TB_SIZE */
4321                 order = PREALLOC_TB_SIZE - 1;
4322         /* Add the prealloc space to lg */
4323         spin_lock(&lg->lg_prealloc_lock);
4324         list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4325                                                 pa_inode_list) {
4326                 spin_lock(&tmp_pa->pa_lock);
4327                 if (tmp_pa->pa_deleted) {
4328                         spin_unlock(&tmp_pa->pa_lock);
4329                         continue;
4330                 }
4331                 if (!added && pa->pa_free < tmp_pa->pa_free) {
4332                         /* Add to the tail of the previous entry */
4333                         list_add_tail_rcu(&pa->pa_inode_list,
4334                                                 &tmp_pa->pa_inode_list);
4335                         added = 1;
4336                         /*
4337                          * we want to count the total
4338                          * number of entries in the list
4339                          */
4340                 }
4341                 spin_unlock(&tmp_pa->pa_lock);
4342                 lg_prealloc_count++;
4343         }
4344         if (!added)
4345                 list_add_tail_rcu(&pa->pa_inode_list,
4346                                         &lg->lg_prealloc_list[order]);
4347         spin_unlock(&lg->lg_prealloc_lock);
4348 
4349         /* Now trim the list to be not more than 8 elements */
4350         if (lg_prealloc_count > 8) {
4351                 ext4_mb_discard_lg_preallocations(sb, lg,
4352                                                   order, lg_prealloc_count);
4353                 return;
4354         }
4355         return ;
4356 }
4357 
4358 /*
4359  * release all resource we used in allocation
4360  */
4361 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4362 {
4363         struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4364         struct ext4_prealloc_space *pa = ac->ac_pa;
4365         if (pa) {
4366                 if (pa->pa_type == MB_GROUP_PA) {
4367                         /* see comment in ext4_mb_use_group_pa() */
4368                         spin_lock(&pa->pa_lock);
4369                         pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4370                         pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4371                         pa->pa_free -= ac->ac_b_ex.fe_len;
4372                         pa->pa_len -= ac->ac_b_ex.fe_len;
4373                         spin_unlock(&pa->pa_lock);
4374                 }
4375         }
4376         if (pa) {
4377                 /*
4378                  * We want to add the pa to the right bucket.
4379                  * Remove it from the list and while adding
4380                  * make sure the list to which we are adding
4381                  * doesn't grow big.
4382                  */
4383                 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4384                         spin_lock(pa->pa_obj_lock);
4385                         list_del_rcu(&pa->pa_inode_list);
4386                         spin_unlock(pa->pa_obj_lock);
4387                         ext4_mb_add_n_trim(ac);
4388                 }
4389                 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4390         }
4391         if (ac->ac_bitmap_page)
4392                 put_page(ac->ac_bitmap_page);
4393         if (ac->ac_buddy_page)
4394                 put_page(ac->ac_buddy_page);
4395         if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4396                 mutex_unlock(&ac->ac_lg->lg_mutex);
4397         ext4_mb_collect_stats(ac);
4398         return 0;
4399 }
4400 
4401 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4402 {
4403         ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4404         int ret;
4405         int freed = 0;
4406 
4407         trace_ext4_mb_discard_preallocations(sb, needed);
4408         for (i = 0; i < ngroups && needed > 0; i++) {
4409                 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4410                 freed += ret;
4411                 needed -= ret;
4412         }
4413 
4414         return freed;
4415 }
4416 
4417 /*
4418  * Main entry point into mballoc to allocate blocks
4419  * it tries to use preallocation first, then falls back
4420  * to usual allocation
4421  */
4422 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4423                                 struct ext4_allocation_request *ar, int *errp)
4424 {
4425         int freed;
4426         struct ext4_allocation_context *ac = NULL;
4427         struct ext4_sb_info *sbi;
4428         struct super_block *sb;
4429         ext4_fsblk_t block = 0;
4430         unsigned int inquota = 0;
4431         unsigned int reserv_clstrs = 0;
4432 
4433         might_sleep();
4434         sb = ar->inode->i_sb;
4435         sbi = EXT4_SB(sb);
4436 
4437         trace_ext4_request_blocks(ar);
4438 
4439         /* Allow to use superuser reservation for quota file */
4440         if (IS_NOQUOTA(ar->inode))
4441                 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4442 
4443         if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4444                 /* Without delayed allocation we need to verify
4445                  * there is enough free blocks to do block allocation
4446                  * and verify allocation doesn't exceed the quota limits.
4447                  */
4448                 while (ar->len &&
4449                         ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4450 
4451                         /* let others to free the space */
4452                         cond_resched();
4453                         ar->len = ar->len >> 1;
4454                 }
4455                 if (!ar->len) {
4456                         *errp = -ENOSPC;
4457                         return 0;
4458                 }
4459                 reserv_clstrs = ar->len;
4460                 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4461                         dquot_alloc_block_nofail(ar->inode,
4462                                                  EXT4_C2B(sbi, ar->len));
4463                 } else {
4464                         while (ar->len &&
4465                                 dquot_alloc_block(ar->inode,
4466                                                   EXT4_C2B(sbi, ar->len))) {
4467 
4468                                 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4469                                 ar->len--;
4470                         }
4471                 }
4472                 inquota = ar->len;
4473                 if (ar->len == 0) {
4474                         *errp = -EDQUOT;
4475                         goto out;
4476                 }
4477         }
4478 
4479         ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4480         if (!ac) {
4481                 ar->len = 0;
4482                 *errp = -ENOMEM;
4483                 goto out;
4484         }
4485 
4486         *errp = ext4_mb_initialize_context(ac, ar);
4487         if (*errp) {
4488                 ar->len = 0;
4489                 goto out;
4490         }
4491 
4492         ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4493         if (!ext4_mb_use_preallocated(ac)) {
4494                 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4495                 ext4_mb_normalize_request(ac, ar);
4496 repeat:
4497                 /* allocate space in core */
4498                 *errp = ext4_mb_regular_allocator(ac);
4499                 if (*errp)
4500                         goto discard_and_exit;
4501 
4502                 /* as we've just preallocated more space than
4503                  * user requested originally, we store allocated
4504                  * space in a special descriptor */
4505                 if (ac->ac_status == AC_STATUS_FOUND &&
4506                     ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4507                         *errp = ext4_mb_new_preallocation(ac);
4508                 if (*errp) {
4509                 discard_and_exit:
4510                         ext4_discard_allocated_blocks(ac);
4511                         goto errout;
4512                 }
4513         }
4514         if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4515                 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4516                 if (*errp) {
4517                         ext4_discard_allocated_blocks(ac);
4518                         goto errout;
4519                 } else {
4520                         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4521                         ar->len = ac->ac_b_ex.fe_len;
4522                 }
4523         } else {
4524                 freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4525                 if (freed)
4526                         goto repeat;
4527                 *errp = -ENOSPC;
4528         }
4529 
4530 errout:
4531         if (*errp) {
4532                 ac->ac_b_ex.fe_len = 0;
4533                 ar->len = 0;
4534                 ext4_mb_show_ac(ac);
4535         }
4536         ext4_mb_release_context(ac);
4537 out:
4538         if (ac)
4539                 kmem_cache_free(ext4_ac_cachep, ac);
4540         if (inquota && ar->len < inquota)
4541                 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4542         if (!ar->len) {
4543                 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4544                         /* release all the reserved blocks if non delalloc */
4545                         percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4546                                                 reserv_clstrs);
4547         }
4548 
4549         trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4550 
4551         return block;
4552 }
4553 
4554 /*
4555  * We can merge two free data extents only if the physical blocks
4556  * are contiguous, AND the extents were freed by the same transaction,
4557  * AND the blocks are associated with the same group.
4558  */
4559 static int can_merge(struct ext4_free_data *entry1,
4560                         struct ext4_free_data *entry2)
4561 {
4562         if ((entry1->efd_tid == entry2->efd_tid) &&
4563             (entry1->efd_group == entry2->efd_group) &&
4564             ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4565                 return 1;
4566         return 0;
4567 }
4568 
4569 static noinline_for_stack int
4570 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4571                       struct ext4_free_data *new_entry)
4572 {
4573         ext4_group_t group = e4b->bd_group;
4574         ext4_grpblk_t cluster;
4575         struct ext4_free_data *entry;
4576         struct ext4_group_info *db = e4b->bd_info;
4577         struct super_block *sb = e4b->bd_sb;
4578         struct ext4_sb_info *sbi = EXT4_SB(sb);
4579         struct rb_node **n = &db->bb_free_root.rb_node, *node;
4580         struct rb_node *parent = NULL, *new_node;
4581 
4582         BUG_ON(!ext4_handle_valid(handle));
4583         BUG_ON(e4b->bd_bitmap_page == NULL);
4584         BUG_ON(e4b->bd_buddy_page == NULL);
4585 
4586         new_node = &new_entry->efd_node;
4587         cluster = new_entry->efd_start_cluster;
4588 
4589         if (!*n) {
4590                 /* first free block exent. We need to
4591                    protect buddy cache from being freed,
4592                  * otherwise we'll refresh it from
4593                  * on-disk bitmap and lose not-yet-available
4594                  * blocks */
4595                 get_page(e4b->bd_buddy_page);
4596                 get_page(e4b->bd_bitmap_page);
4597         }
4598         while (*n) {
4599                 parent = *n;
4600                 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4601                 if (cluster < entry->efd_start_cluster)
4602                         n = &(*n)->rb_left;
4603                 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4604                         n = &(*n)->rb_right;
4605                 else {
4606                         ext4_grp_locked_error(sb, group, 0,
4607                                 ext4_group_first_block_no(sb, group) +
4608                                 EXT4_C2B(sbi, cluster),
4609                                 "Block already on to-be-freed list");
4610                         return 0;
4611                 }
4612         }
4613 
4614         rb_link_node(new_node, parent, n);
4615         rb_insert_color(new_node, &db->bb_free_root);
4616 
4617         /* Now try to see the extent can be merged to left and right */
4618         node = rb_prev(new_node);
4619         if (node) {
4620                 entry = rb_entry(node, struct ext4_free_data, efd_node);
4621                 if (can_merge(entry, new_entry) &&
4622                     ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4623                         new_entry->efd_start_cluster = entry->efd_start_cluster;
4624                         new_entry->efd_count += entry->efd_count;
4625                         rb_erase(node, &(db->bb_free_root));
4626                         kmem_cache_free(ext4_free_data_cachep, entry);
4627                 }
4628         }
4629 
4630         node = rb_next(new_node);
4631         if (node) {
4632                 entry = rb_entry(node, struct ext4_free_data, efd_node);
4633                 if (can_merge(new_entry, entry) &&
4634                     ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4635                         new_entry->efd_count += entry->efd_count;
4636                         rb_erase(node, &(db->bb_free_root));
4637                         kmem_cache_free(ext4_free_data_cachep, entry);
4638                 }
4639         }
4640         /* Add the extent to transaction's private list */
4641         ext4_journal_callback_add(handle, ext4_free_data_callback,
4642                                   &new_entry->efd_jce);
4643         return 0;
4644 }
4645 
4646 /**
4647  * ext4_free_blocks() -- Free given blocks and update quota
4648  * @handle:             handle for this transaction
4649  * @inode:              inode
4650  * @block:              start physical block to free
4651  * @count:              number of blocks to count
4652  * @flags:              flags used by ext4_free_blocks
4653  */
4654 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4655                       struct buffer_head *bh, ext4_fsblk_t block,
4656                       unsigned long count, int flags)
4657 {
4658         struct buffer_head *bitmap_bh = NULL;
4659         struct super_block *sb = inode->i_sb;
4660         struct ext4_group_desc *gdp;
4661         unsigned int overflow;
4662         ext4_grpblk_t bit;
4663         struct buffer_head *gd_bh;
4664         ext4_group_t block_group;
4665         struct ext4_sb_info *sbi;
4666         struct ext4_buddy e4b;
4667         unsigned int count_clusters;
4668         int err = 0;
4669         int ret;
4670 
4671         might_sleep();
4672         if (bh) {
4673                 if (block)
4674                         BUG_ON(block != bh->b_blocknr);
4675                 else
4676                         block = bh->b_blocknr;
4677         }
4678 
4679         sbi = EXT4_SB(sb);
4680         if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4681             !ext4_data_block_valid(sbi, block, count)) {
4682                 ext4_error(sb, "Freeing blocks not in datazone - "
4683                            "block = %llu, count = %lu", block, count);
4684                 goto error_return;
4685         }
4686 
4687         ext4_debug("freeing block %llu\n", block);
4688         trace_ext4_free_blocks(inode, block, count, flags);
4689 
4690         if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4691                 BUG_ON(count > 1);
4692 
4693                 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4694                             inode, bh, block);
4695         }
4696 
4697         /*
4698          * If the extent to be freed does not begin on a cluster
4699          * boundary, we need to deal with partial clusters at the
4700          * beginning and end of the extent.  Normally we will free
4701          * blocks at the beginning or the end unless we are explicitly
4702          * requested to avoid doing so.
4703          */
4704         overflow = EXT4_PBLK_COFF(sbi, block);
4705         if (overflow) {
4706                 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4707                         overflow = sbi->s_cluster_ratio - overflow;
4708                         block += overflow;
4709                         if (count > overflow)
4710                                 count -= overflow;
4711                         else
4712                                 return;
4713                 } else {
4714                         block -= overflow;
4715                         count += overflow;
4716                 }
4717         }
4718         overflow = EXT4_LBLK_COFF(sbi, count);
4719         if (overflow) {
4720                 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4721                         if (count > overflow)
4722                                 count -= overflow;
4723                         else
4724                                 return;
4725                 } else
4726                         count += sbi->s_cluster_ratio - overflow;
4727         }
4728 
4729         if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4730                 int i;
4731                 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4732 
4733                 for (i = 0; i < count; i++) {
4734                         cond_resched();
4735                         if (is_metadata)
4736                                 bh = sb_find_get_block(inode->i_sb, block + i);
4737                         ext4_forget(handle, is_metadata, inode, bh, block + i);
4738                 }
4739         }
4740 
4741 do_more:
4742         overflow = 0;
4743         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4744 
4745         if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4746                         ext4_get_group_info(sb, block_group))))
4747                 return;
4748 
4749         /*
4750          * Check to see if we are freeing blocks across a group
4751          * boundary.
4752          */
4753         if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4754                 overflow = EXT4_C2B(sbi, bit) + count -
4755                         EXT4_BLOCKS_PER_GROUP(sb);
4756                 count -= overflow;
4757         }
4758         count_clusters = EXT4_NUM_B2C(sbi, count);
4759         bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4760         if (IS_ERR(bitmap_bh)) {
4761                 err = PTR_ERR(bitmap_bh);
4762                 bitmap_bh = NULL;
4763                 goto error_return;
4764         }
4765         gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4766         if (!gdp) {
4767                 err = -EIO;
4768                 goto error_return;
4769         }
4770 
4771         if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4772             in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4773             in_range(block, ext4_inode_table(sb, gdp),
4774                      EXT4_SB(sb)->s_itb_per_group) ||
4775             in_range(block + count - 1, ext4_inode_table(sb, gdp),
4776                      EXT4_SB(sb)->s_itb_per_group)) {
4777 
4778                 ext4_error(sb, "Freeing blocks in system zone - "
4779                            "Block = %llu, count = %lu", block, count);
4780                 /* err = 0. ext4_std_error should be a no op */
4781                 goto error_return;
4782         }
4783 
4784         BUFFER_TRACE(bitmap_bh, "getting write access");
4785         err = ext4_journal_get_write_access(handle, bitmap_bh);
4786         if (err)
4787                 goto error_return;
4788 
4789         /*
4790          * We are about to modify some metadata.  Call the journal APIs
4791          * to unshare ->b_data if a currently-committing transaction is
4792          * using it
4793          */
4794         BUFFER_TRACE(gd_bh, "get_write_access");
4795         err = ext4_journal_get_write_access(handle, gd_bh);
4796         if (err)
4797                 goto error_return;
4798 #ifdef AGGRESSIVE_CHECK
4799         {
4800                 int i;
4801                 for (i = 0; i < count_clusters; i++)
4802                         BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4803         }
4804 #endif
4805         trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4806 
4807         /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4808         err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4809                                      GFP_NOFS|__GFP_NOFAIL);
4810         if (err)
4811                 goto error_return;
4812 
4813         /*
4814          * We need to make sure we don't reuse the freed block until after the
4815          * transaction is committed. We make an exception if the inode is to be
4816          * written in writeback mode since writeback mode has weak data
4817          * consistency guarantees.
4818          */
4819         if (ext4_handle_valid(handle) &&
4820             ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4821              !ext4_should_writeback_data(inode))) {
4822                 struct ext4_free_data *new_entry;
4823                 /*
4824                  * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4825                  * to fail.
4826                  */
4827                 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4828                                 GFP_NOFS|__GFP_NOFAIL);
4829                 new_entry->efd_start_cluster = bit;
4830                 new_entry->efd_group = block_group;
4831                 new_entry->efd_count = count_clusters;
4832                 new_entry->efd_tid = handle->h_transaction->t_tid;
4833 
4834                 ext4_lock_group(sb, block_group);
4835                 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4836                 ext4_mb_free_metadata(handle, &e4b, new_entry);
4837         } else {
4838                 /* need to update group_info->bb_free and bitmap
4839                  * with group lock held. generate_buddy look at
4840                  * them with group lock_held
4841                  */
4842                 if (test_opt(sb, DISCARD)) {
4843                         err = ext4_issue_discard(sb, block_group, bit, count);
4844                         if (err && err != -EOPNOTSUPP)
4845                                 ext4_msg(sb, KERN_WARNING, "discard request in"
4846                                          " group:%d block:%d count:%lu failed"
4847                                          " with %d", block_group, bit, count,
4848                                          err);
4849                 } else
4850                         EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4851 
4852                 ext4_lock_group(sb, block_group);
4853                 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4854                 mb_free_blocks(inode, &e4b, bit, count_clusters);
4855         }
4856 
4857         ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4858         ext4_free_group_clusters_set(sb, gdp, ret);
4859         ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4860         ext4_group_desc_csum_set(sb, block_group, gdp);
4861         ext4_unlock_group(sb, block_group);
4862 
4863         if (sbi->s_log_groups_per_flex) {
4864                 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4865                 atomic64_add(count_clusters,
4866                              &sbi->s_flex_groups[flex_group].free_clusters);
4867         }
4868 
4869         if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4870                 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4871         percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4872 
4873         ext4_mb_unload_buddy(&e4b);
4874 
4875         /* We dirtied the bitmap block */
4876         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4877         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4878 
4879         /* And the group descriptor block */
4880         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4881         ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4882         if (!err)
4883                 err = ret;
4884 
4885         if (overflow && !err) {
4886                 block += count;
4887                 count = overflow;
4888                 put_bh(bitmap_bh);
4889                 goto do_more;
4890         }
4891 error_return:
4892         brelse(bitmap_bh);
4893         ext4_std_error(sb, err);
4894         return;
4895 }
4896 
4897 /**
4898  * ext4_group_add_blocks() -- Add given blocks to an existing group
4899  * @handle:                     handle to this transaction
4900  * @sb:                         super block
4901  * @block:                      start physical block to add to the block group
4902  * @count:                      number of blocks to free
4903  *
4904  * This marks the blocks as free in the bitmap and buddy.
4905  */
4906 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4907                          ext4_fsblk_t block, unsigned long count)
4908 {
4909         struct buffer_head *bitmap_bh = NULL;
4910         struct buffer_head *gd_bh;
4911         ext4_group_t block_group;
4912         ext4_grpblk_t bit;
4913         unsigned int i;
4914         struct ext4_group_desc *desc;
4915         struct ext4_sb_info *sbi = EXT4_SB(sb);
4916         struct ext4_buddy e4b;
4917         int err = 0, ret, blk_free_count;
4918         ext4_grpblk_t blocks_freed;
4919 
4920         ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4921 
4922         if (count == 0)
4923                 return 0;
4924 
4925         ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4926         /*
4927          * Check to see if we are freeing blocks across a group
4928          * boundary.
4929          */
4930         if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4931                 ext4_warning(sb, "too much blocks added to group %u\n",
4932                              block_group);
4933                 err = -EINVAL;
4934                 goto error_return;
4935         }
4936 
4937         bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4938         if (IS_ERR(bitmap_bh)) {
4939                 err = PTR_ERR(bitmap_bh);
4940                 bitmap_bh = NULL;
4941                 goto error_return;
4942         }
4943 
4944         desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4945         if (!desc) {
4946                 err = -EIO;
4947                 goto error_return;
4948         }
4949 
4950         if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4951             in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4952             in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4953             in_range(block + count - 1, ext4_inode_table(sb, desc),
4954                      sbi->s_itb_per_group)) {
4955                 ext4_error(sb, "Adding blocks in system zones - "
4956                            "Block = %llu, count = %lu",
4957                            block, count);
4958                 err = -EINVAL;
4959                 goto error_return;
4960         }
4961 
4962         BUFFER_TRACE(bitmap_bh, "getting write access");
4963         err = ext4_journal_get_write_access(handle, bitmap_bh);
4964         if (err)
4965                 goto error_return;
4966 
4967         /*
4968          * We are about to modify some metadata.  Call the journal APIs
4969          * to unshare ->b_data if a currently-committing transaction is
4970          * using it
4971          */
4972         BUFFER_TRACE(gd_bh, "get_write_access");
4973         err = ext4_journal_get_write_access(handle, gd_bh);
4974         if (err)
4975                 goto error_return;
4976 
4977         for (i = 0, blocks_freed = 0; i < count; i++) {
4978                 BUFFER_TRACE(bitmap_bh, "clear bit");
4979                 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4980                         ext4_error(sb, "bit already cleared for block %llu",
4981                                    (ext4_fsblk_t)(block + i));
4982                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
4983                 } else {
4984                         blocks_freed++;
4985                 }
4986         }
4987 
4988         err = ext4_mb_load_buddy(sb, block_group, &e4b);
4989         if (err)
4990                 goto error_return;
4991 
4992         /*
4993          * need to update group_info->bb_free and bitmap
4994          * with group lock held. generate_buddy look at
4995          * them with group lock_held
4996          */
4997         ext4_lock_group(sb, block_group);
4998         mb_clear_bits(bitmap_bh->b_data, bit, count);
4999         mb_free_blocks(NULL, &e4b, bit, count);
5000         blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5001         ext4_free_group_clusters_set(sb, desc, blk_free_count);
5002         ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5003         ext4_group_desc_csum_set(sb, block_group, desc);
5004         ext4_unlock_group(sb, block_group);
5005         percpu_counter_add(&sbi->s_freeclusters_counter,
5006                            EXT4_NUM_B2C(sbi, blocks_freed));
5007 
5008         if (sbi->s_log_groups_per_flex) {
5009                 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5010                 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5011                              &sbi->s_flex_groups[flex_group].free_clusters);
5012         }
5013 
5014         ext4_mb_unload_buddy(&e4b);
5015 
5016         /* We dirtied the bitmap block */
5017         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5018         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5019 
5020         /* And the group descriptor block */
5021         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5022         ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5023         if (!err)
5024                 err = ret;
5025 
5026 error_return:
5027         brelse(bitmap_bh);
5028         ext4_std_error(sb, err);
5029         return err;
5030 }
5031 
5032 /**
5033  * ext4_trim_extent -- function to TRIM one single free extent in the group
5034  * @sb:         super block for the file system
5035  * @start:      starting block of the free extent in the alloc. group
5036  * @count:      number of blocks to TRIM
5037  * @group:      alloc. group we are working with
5038  * @e4b:        ext4 buddy for the group
5039  *
5040  * Trim "count" blocks starting at "start" in the "group". To assure that no
5041  * one will allocate those blocks, mark it as used in buddy bitmap. This must
5042  * be called with under the group lock.
5043  */
5044 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5045                              ext4_group_t group, struct ext4_buddy *e4b)
5046 __releases(bitlock)
5047 __acquires(bitlock)
5048 {
5049         struct ext4_free_extent ex;
5050         int ret = 0;
5051 
5052         trace_ext4_trim_extent(sb, group, start, count);
5053 
5054         assert_spin_locked(ext4_group_lock_ptr(sb, group));
5055 
5056         ex.fe_start = start;
5057         ex.fe_group = group;
5058         ex.fe_len = count;
5059 
5060         /*
5061          * Mark blocks used, so no one can reuse them while
5062          * being trimmed.
5063          */
5064         mb_mark_used(e4b, &ex);
5065         ext4_unlock_group(sb, group);
5066         ret = ext4_issue_discard(sb, group, start, count);
5067         ext4_lock_group(sb, group);
5068         mb_free_blocks(NULL, e4b, start, ex.fe_len);
5069         return ret;
5070 }
5071 
5072 /**
5073  * ext4_trim_all_free -- function to trim all free space in alloc. group
5074  * @sb:                 super block for file system
5075  * @group:              group to be trimmed
5076  * @start:              first group block to examine
5077  * @max:                last group block to examine
5078  * @minblocks:          minimum extent block count
5079  *
5080  * ext4_trim_all_free walks through group's buddy bitmap searching for free
5081  * extents. When the free block is found, ext4_trim_extent is called to TRIM
5082  * the extent.
5083  *
5084  *
5085  * ext4_trim_all_free walks through group's block bitmap searching for free
5086  * extents. When the free extent is found, mark it as used in group buddy
5087  * bitmap. Then issue a TRIM command on this extent and free the extent in
5088  * the group buddy bitmap. This is done until whole group is scanned.
5089  */
5090 static ext4_grpblk_t
5091 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5092                    ext4_grpblk_t start, ext4_grpblk_t max,
5093                    ext4_grpblk_t minblocks)
5094 {
5095         void *bitmap;
5096         ext4_grpblk_t next, count = 0, free_count = 0;
5097         struct ext4_buddy e4b;
5098         int ret = 0;
5099 
5100         trace_ext4_trim_all_free(sb, group, start, max);
5101 
5102         ret = ext4_mb_load_buddy(sb, group, &e4b);
5103         if (ret) {
5104                 ext4_error(sb, "Error in loading buddy "
5105                                 "information for %u", group);
5106                 return ret;
5107         }
5108         bitmap = e4b.bd_bitmap;
5109 
5110         ext4_lock_group(sb, group);
5111         if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5112             minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5113                 goto out;
5114 
5115         start = (e4b.bd_info->bb_first_free > start) ?
5116                 e4b.bd_info->bb_first_free : start;
5117 
5118         while (start <= max) {
5119                 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5120                 if (start > max)
5121                         break;
5122                 next = mb_find_next_bit(bitmap, max + 1, start);
5123 
5124                 if ((next - start) >= minblocks) {
5125                         ret = ext4_trim_extent(sb, start,
5126                                                next - start, group, &e4b);
5127                         if (ret && ret != -EOPNOTSUPP)
5128                                 break;
5129                         ret = 0;
5130                         count += next - start;
5131                 }
5132                 free_count += next - start;
5133                 start = next + 1;
5134 
5135                 if (fatal_signal_pending(current)) {
5136                         count = -ERESTARTSYS;
5137                         break;
5138                 }
5139 
5140                 if (need_resched()) {
5141                         ext4_unlock_group(sb, group);
5142                         cond_resched();
5143                         ext4_lock_group(sb, group);
5144                 }
5145 
5146                 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5147                         break;
5148         }
5149 
5150         if (!ret) {
5151                 ret = count;
5152                 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5153         }
5154 out:
5155         ext4_unlock_group(sb, group);
5156         ext4_mb_unload_buddy(&e4b);
5157 
5158         ext4_debug("trimmed %d blocks in the group %d\n",
5159                 count, group);
5160 
5161         return ret;
5162 }
5163 
5164 /**
5165  * ext4_trim_fs() -- trim ioctl handle function
5166  * @sb:                 superblock for filesystem
5167  * @range:              fstrim_range structure
5168  *
5169  * start:       First Byte to trim
5170  * len:         number of Bytes to trim from start
5171  * minlen:      minimum extent length in Bytes
5172  * ext4_trim_fs goes through all allocation groups containing Bytes from
5173  * start to start+len. For each such a group ext4_trim_all_free function
5174  * is invoked to trim all free space.
5175  */
5176 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5177 {
5178         struct ext4_group_info *grp;
5179         ext4_group_t group, first_group, last_group;
5180         ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5181         uint64_t start, end, minlen, trimmed = 0;
5182         ext4_fsblk_t first_data_blk =
5183                         le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5184         ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5185         int ret = 0;
5186 
5187         start = range->start >> sb->s_blocksize_bits;
5188         end = start + (range->len >> sb->s_blocksize_bits) - 1;
5189         minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5190                               range->minlen >> sb->s_blocksize_bits);
5191 
5192         if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5193             start >= max_blks ||
5194             range->len < sb->s_blocksize)
5195                 return -EINVAL;
5196         if (end >= max_blks)
5197                 end = max_blks - 1;
5198         if (end <= first_data_blk)
5199                 goto out;
5200         if (start < first_data_blk)
5201                 start = first_data_blk;
5202 
5203         /* Determine first and last group to examine based on start and end */
5204         ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5205                                      &first_group, &first_cluster);
5206         ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5207                                      &last_group, &last_cluster);
5208 
5209         /* end now represents the last cluster to discard in this group */
5210         end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5211 
5212         for (group = first_group; group <= last_group; group++) {
5213                 grp = ext4_get_group_info(sb, group);
5214                 /* We only do this if the grp has never been initialized */
5215                 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5216                         ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5217                         if (ret)
5218                                 break;
5219                 }
5220 
5221                 /*
5222                  * For all the groups except the last one, last cluster will
5223                  * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5224                  * change it for the last group, note that last_cluster is
5225                  * already computed earlier by ext4_get_group_no_and_offset()
5226                  */
5227                 if (group == last_group)
5228                         end = last_cluster;
5229 
5230                 if (grp->bb_free >= minlen) {
5231                         cnt = ext4_trim_all_free(sb, group, first_cluster,
5232                                                 end, minlen);
5233                         if (cnt < 0) {
5234                                 ret = cnt;
5235                                 break;
5236                         }
5237                         trimmed += cnt;
5238                 }
5239 
5240                 /*
5241                  * For every group except the first one, we are sure
5242                  * that the first cluster to discard will be cluster #0.
5243                  */
5244                 first_cluster = 0;
5245         }
5246 
5247         if (!ret)
5248                 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5249 
5250 out:
5251         range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5252         return ret;
5253 }
5254 

~ [ 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