xref: /openbmc/linux/fs/ext4/mballoc.c (revision d3741027)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4  * Written by Alex Tomas <alex@clusterfs.com>
5  */
6 
7 
8 /*
9  * mballoc.c contains the multiblocks allocation routines
10  */
11 
12 #include "ext4_jbd2.h"
13 #include "mballoc.h"
14 #include <linux/log2.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/nospec.h>
18 #include <linux/backing-dev.h>
19 #include <trace/events/ext4.h>
20 
21 /*
22  * MUSTDO:
23  *   - test ext4_ext_search_left() and ext4_ext_search_right()
24  *   - search for metadata in few groups
25  *
26  * TODO v4:
27  *   - normalization should take into account whether file is still open
28  *   - discard preallocations if no free space left (policy?)
29  *   - don't normalize tails
30  *   - quota
31  *   - reservation for superuser
32  *
33  * TODO v3:
34  *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
35  *   - track min/max extents in each group for better group selection
36  *   - mb_mark_used() may allocate chunk right after splitting buddy
37  *   - tree of groups sorted by number of free blocks
38  *   - error handling
39  */
40 
41 /*
42  * The allocation request involve request for multiple number of blocks
43  * near to the goal(block) value specified.
44  *
45  * During initialization phase of the allocator we decide to use the
46  * group preallocation or inode preallocation depending on the size of
47  * the file. The size of the file could be the resulting file size we
48  * would have after allocation, or the current file size, which ever
49  * is larger. If the size is less than sbi->s_mb_stream_request we
50  * select to use the group preallocation. The default value of
51  * s_mb_stream_request is 16 blocks. This can also be tuned via
52  * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
53  * terms of number of blocks.
54  *
55  * The main motivation for having small file use group preallocation is to
56  * ensure that we have small files closer together on the disk.
57  *
58  * First stage the allocator looks at the inode prealloc list,
59  * ext4_inode_info->i_prealloc_list, which contains list of prealloc
60  * spaces for this particular inode. The inode prealloc space is
61  * represented as:
62  *
63  * pa_lstart -> the logical start block for this prealloc space
64  * pa_pstart -> the physical start block for this prealloc space
65  * pa_len    -> length for this prealloc space (in clusters)
66  * pa_free   ->  free space available in this prealloc space (in clusters)
67  *
68  * The inode preallocation space is used looking at the _logical_ start
69  * block. If only the logical file block falls within the range of prealloc
70  * space we will consume the particular prealloc space. This makes sure that
71  * we have contiguous physical blocks representing the file blocks
72  *
73  * The important thing to be noted in case of inode prealloc space is that
74  * we don't modify the values associated to inode prealloc space except
75  * pa_free.
76  *
77  * If we are not able to find blocks in the inode prealloc space and if we
78  * have the group allocation flag set then we look at the locality group
79  * prealloc space. These are per CPU prealloc list represented as
80  *
81  * ext4_sb_info.s_locality_groups[smp_processor_id()]
82  *
83  * The reason for having a per cpu locality group is to reduce the contention
84  * between CPUs. It is possible to get scheduled at this point.
85  *
86  * The locality group prealloc space is used looking at whether we have
87  * enough free space (pa_free) within the prealloc space.
88  *
89  * If we can't allocate blocks via inode prealloc or/and locality group
90  * prealloc then we look at the buddy cache. The buddy cache is represented
91  * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
92  * mapped to the buddy and bitmap information regarding different
93  * groups. The buddy information is attached to buddy cache inode so that
94  * we can access them through the page cache. The information regarding
95  * each group is loaded via ext4_mb_load_buddy.  The information involve
96  * block bitmap and buddy information. The information are stored in the
97  * inode as:
98  *
99  *  {                        page                        }
100  *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
101  *
102  *
103  * one block each for bitmap and buddy information.  So for each group we
104  * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
105  * blocksize) blocks.  So it can have information regarding groups_per_page
106  * which is blocks_per_page/2
107  *
108  * The buddy cache inode is not stored on disk. The inode is thrown
109  * away when the filesystem is unmounted.
110  *
111  * We look for count number of blocks in the buddy cache. If we were able
112  * to locate that many free blocks we return with additional information
113  * regarding rest of the contiguous physical block available
114  *
115  * Before allocating blocks via buddy cache we normalize the request
116  * blocks. This ensure we ask for more blocks that we needed. The extra
117  * blocks that we get after allocation is added to the respective prealloc
118  * list. In case of inode preallocation we follow a list of heuristics
119  * based on file size. This can be found in ext4_mb_normalize_request. If
120  * we are doing a group prealloc we try to normalize the request to
121  * sbi->s_mb_group_prealloc.  The default value of s_mb_group_prealloc is
122  * dependent on the cluster size; for non-bigalloc file systems, it is
123  * 512 blocks. This can be tuned via
124  * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
125  * terms of number of blocks. If we have mounted the file system with -O
126  * stripe=<value> option the group prealloc request is normalized to the
127  * smallest multiple of the stripe value (sbi->s_stripe) which is
128  * greater than the default mb_group_prealloc.
129  *
130  * If "mb_optimize_scan" mount option is set, we maintain in memory group info
131  * structures in two data structures:
132  *
133  * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
134  *
135  *    Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
136  *
137  *    This is an array of lists where the index in the array represents the
138  *    largest free order in the buddy bitmap of the participating group infos of
139  *    that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
140  *    number of buddy bitmap orders possible) number of lists. Group-infos are
141  *    placed in appropriate lists.
142  *
143  * 2) Average fragment size rb tree (sbi->s_mb_avg_fragment_size_root)
144  *
145  *    Locking: sbi->s_mb_rb_lock (rwlock)
146  *
147  *    This is a red black tree consisting of group infos and the tree is sorted
148  *    by average fragment sizes (which is calculated as ext4_group_info->bb_free
149  *    / ext4_group_info->bb_fragments).
150  *
151  * When "mb_optimize_scan" mount option is set, mballoc consults the above data
152  * structures to decide the order in which groups are to be traversed for
153  * fulfilling an allocation request.
154  *
155  * At CR = 0, we look for groups which have the largest_free_order >= the order
156  * of the request. We directly look at the largest free order list in the data
157  * structure (1) above where largest_free_order = order of the request. If that
158  * list is empty, we look at remaining list in the increasing order of
159  * largest_free_order. This allows us to perform CR = 0 lookup in O(1) time.
160  *
161  * At CR = 1, we only consider groups where average fragment size > request
162  * size. So, we lookup a group which has average fragment size just above or
163  * equal to request size using our rb tree (data structure 2) in O(log N) time.
164  *
165  * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
166  * linear order which requires O(N) search time for each CR 0 and CR 1 phase.
167  *
168  * The regular allocator (using the buddy cache) supports a few tunables.
169  *
170  * /sys/fs/ext4/<partition>/mb_min_to_scan
171  * /sys/fs/ext4/<partition>/mb_max_to_scan
172  * /sys/fs/ext4/<partition>/mb_order2_req
173  * /sys/fs/ext4/<partition>/mb_linear_limit
174  *
175  * The regular allocator uses buddy scan only if the request len is power of
176  * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
177  * value of s_mb_order2_reqs can be tuned via
178  * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
179  * stripe size (sbi->s_stripe), we try to search for contiguous block in
180  * stripe size. This should result in better allocation on RAID setups. If
181  * not, we search in the specific group using bitmap for best extents. The
182  * tunable min_to_scan and max_to_scan control the behaviour here.
183  * min_to_scan indicate how long the mballoc __must__ look for a best
184  * extent and max_to_scan indicates how long the mballoc __can__ look for a
185  * best extent in the found extents. Searching for the blocks starts with
186  * the group specified as the goal value in allocation context via
187  * ac_g_ex. Each group is first checked based on the criteria whether it
188  * can be used for allocation. ext4_mb_good_group explains how the groups are
189  * checked.
190  *
191  * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
192  * get traversed linearly. That may result in subsequent allocations being not
193  * close to each other. And so, the underlying device may get filled up in a
194  * non-linear fashion. While that may not matter on non-rotational devices, for
195  * rotational devices that may result in higher seek times. "mb_linear_limit"
196  * tells mballoc how many groups mballoc should search linearly before
197  * performing consulting above data structures for more efficient lookups. For
198  * non rotational devices, this value defaults to 0 and for rotational devices
199  * this is set to MB_DEFAULT_LINEAR_LIMIT.
200  *
201  * Both the prealloc space are getting populated as above. So for the first
202  * request we will hit the buddy cache which will result in this prealloc
203  * space getting filled. The prealloc space is then later used for the
204  * subsequent request.
205  */
206 
207 /*
208  * mballoc operates on the following data:
209  *  - on-disk bitmap
210  *  - in-core buddy (actually includes buddy and bitmap)
211  *  - preallocation descriptors (PAs)
212  *
213  * there are two types of preallocations:
214  *  - inode
215  *    assiged to specific inode and can be used for this inode only.
216  *    it describes part of inode's space preallocated to specific
217  *    physical blocks. any block from that preallocated can be used
218  *    independent. the descriptor just tracks number of blocks left
219  *    unused. so, before taking some block from descriptor, one must
220  *    make sure corresponded logical block isn't allocated yet. this
221  *    also means that freeing any block within descriptor's range
222  *    must discard all preallocated blocks.
223  *  - locality group
224  *    assigned to specific locality group which does not translate to
225  *    permanent set of inodes: inode can join and leave group. space
226  *    from this type of preallocation can be used for any inode. thus
227  *    it's consumed from the beginning to the end.
228  *
229  * relation between them can be expressed as:
230  *    in-core buddy = on-disk bitmap + preallocation descriptors
231  *
232  * this mean blocks mballoc considers used are:
233  *  - allocated blocks (persistent)
234  *  - preallocated blocks (non-persistent)
235  *
236  * consistency in mballoc world means that at any time a block is either
237  * free or used in ALL structures. notice: "any time" should not be read
238  * literally -- time is discrete and delimited by locks.
239  *
240  *  to keep it simple, we don't use block numbers, instead we count number of
241  *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
242  *
243  * all operations can be expressed as:
244  *  - init buddy:			buddy = on-disk + PAs
245  *  - new PA:				buddy += N; PA = N
246  *  - use inode PA:			on-disk += N; PA -= N
247  *  - discard inode PA			buddy -= on-disk - PA; PA = 0
248  *  - use locality group PA		on-disk += N; PA -= N
249  *  - discard locality group PA		buddy -= PA; PA = 0
250  *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
251  *        is used in real operation because we can't know actual used
252  *        bits from PA, only from on-disk bitmap
253  *
254  * if we follow this strict logic, then all operations above should be atomic.
255  * given some of them can block, we'd have to use something like semaphores
256  * killing performance on high-end SMP hardware. let's try to relax it using
257  * the following knowledge:
258  *  1) if buddy is referenced, it's already initialized
259  *  2) while block is used in buddy and the buddy is referenced,
260  *     nobody can re-allocate that block
261  *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
262  *     bit set and PA claims same block, it's OK. IOW, one can set bit in
263  *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
264  *     block
265  *
266  * so, now we're building a concurrency table:
267  *  - init buddy vs.
268  *    - new PA
269  *      blocks for PA are allocated in the buddy, buddy must be referenced
270  *      until PA is linked to allocation group to avoid concurrent buddy init
271  *    - use inode PA
272  *      we need to make sure that either on-disk bitmap or PA has uptodate data
273  *      given (3) we care that PA-=N operation doesn't interfere with init
274  *    - discard inode PA
275  *      the simplest way would be to have buddy initialized by the discard
276  *    - use locality group PA
277  *      again PA-=N must be serialized with init
278  *    - discard locality group PA
279  *      the simplest way would be to have buddy initialized by the discard
280  *  - new PA vs.
281  *    - use inode PA
282  *      i_data_sem serializes them
283  *    - discard inode PA
284  *      discard process must wait until PA isn't used by another process
285  *    - use locality group PA
286  *      some mutex should serialize them
287  *    - discard locality group PA
288  *      discard process must wait until PA isn't used by another process
289  *  - use inode PA
290  *    - use inode PA
291  *      i_data_sem or another mutex should serializes them
292  *    - discard inode PA
293  *      discard process must wait until PA isn't used by another process
294  *    - use locality group PA
295  *      nothing wrong here -- they're different PAs covering different blocks
296  *    - discard locality group PA
297  *      discard process must wait until PA isn't used by another process
298  *
299  * now we're ready to make few consequences:
300  *  - PA is referenced and while it is no discard is possible
301  *  - PA is referenced until block isn't marked in on-disk bitmap
302  *  - PA changes only after on-disk bitmap
303  *  - discard must not compete with init. either init is done before
304  *    any discard or they're serialized somehow
305  *  - buddy init as sum of on-disk bitmap and PAs is done atomically
306  *
307  * a special case when we've used PA to emptiness. no need to modify buddy
308  * in this case, but we should care about concurrent init
309  *
310  */
311 
312  /*
313  * Logic in few words:
314  *
315  *  - allocation:
316  *    load group
317  *    find blocks
318  *    mark bits in on-disk bitmap
319  *    release group
320  *
321  *  - use preallocation:
322  *    find proper PA (per-inode or group)
323  *    load group
324  *    mark bits in on-disk bitmap
325  *    release group
326  *    release PA
327  *
328  *  - free:
329  *    load group
330  *    mark bits in on-disk bitmap
331  *    release group
332  *
333  *  - discard preallocations in group:
334  *    mark PAs deleted
335  *    move them onto local list
336  *    load on-disk bitmap
337  *    load group
338  *    remove PA from object (inode or locality group)
339  *    mark free blocks in-core
340  *
341  *  - discard inode's preallocations:
342  */
343 
344 /*
345  * Locking rules
346  *
347  * Locks:
348  *  - bitlock on a group	(group)
349  *  - object (inode/locality)	(object)
350  *  - per-pa lock		(pa)
351  *  - cr0 lists lock		(cr0)
352  *  - cr1 tree lock		(cr1)
353  *
354  * Paths:
355  *  - new pa
356  *    object
357  *    group
358  *
359  *  - find and use pa:
360  *    pa
361  *
362  *  - release consumed pa:
363  *    pa
364  *    group
365  *    object
366  *
367  *  - generate in-core bitmap:
368  *    group
369  *        pa
370  *
371  *  - discard all for given object (inode, locality group):
372  *    object
373  *        pa
374  *    group
375  *
376  *  - discard all for given group:
377  *    group
378  *        pa
379  *    group
380  *        object
381  *
382  *  - allocation path (ext4_mb_regular_allocator)
383  *    group
384  *    cr0/cr1
385  */
386 static struct kmem_cache *ext4_pspace_cachep;
387 static struct kmem_cache *ext4_ac_cachep;
388 static struct kmem_cache *ext4_free_data_cachep;
389 
390 /* We create slab caches for groupinfo data structures based on the
391  * superblock block size.  There will be one per mounted filesystem for
392  * each unique s_blocksize_bits */
393 #define NR_GRPINFO_CACHES 8
394 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
395 
396 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
397 	"ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
398 	"ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
399 	"ext4_groupinfo_64k", "ext4_groupinfo_128k"
400 };
401 
402 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
403 					ext4_group_t group);
404 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
405 						ext4_group_t group);
406 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
407 
408 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
409 			       ext4_group_t group, int cr);
410 
411 static int ext4_try_to_trim_range(struct super_block *sb,
412 		struct ext4_buddy *e4b, ext4_grpblk_t start,
413 		ext4_grpblk_t max, ext4_grpblk_t minblocks);
414 
415 /*
416  * The algorithm using this percpu seq counter goes below:
417  * 1. We sample the percpu discard_pa_seq counter before trying for block
418  *    allocation in ext4_mb_new_blocks().
419  * 2. We increment this percpu discard_pa_seq counter when we either allocate
420  *    or free these blocks i.e. while marking those blocks as used/free in
421  *    mb_mark_used()/mb_free_blocks().
422  * 3. We also increment this percpu seq counter when we successfully identify
423  *    that the bb_prealloc_list is not empty and hence proceed for discarding
424  *    of those PAs inside ext4_mb_discard_group_preallocations().
425  *
426  * Now to make sure that the regular fast path of block allocation is not
427  * affected, as a small optimization we only sample the percpu seq counter
428  * on that cpu. Only when the block allocation fails and when freed blocks
429  * found were 0, that is when we sample percpu seq counter for all cpus using
430  * below function ext4_get_discard_pa_seq_sum(). This happens after making
431  * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
432  */
433 static DEFINE_PER_CPU(u64, discard_pa_seq);
434 static inline u64 ext4_get_discard_pa_seq_sum(void)
435 {
436 	int __cpu;
437 	u64 __seq = 0;
438 
439 	for_each_possible_cpu(__cpu)
440 		__seq += per_cpu(discard_pa_seq, __cpu);
441 	return __seq;
442 }
443 
444 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
445 {
446 #if BITS_PER_LONG == 64
447 	*bit += ((unsigned long) addr & 7UL) << 3;
448 	addr = (void *) ((unsigned long) addr & ~7UL);
449 #elif BITS_PER_LONG == 32
450 	*bit += ((unsigned long) addr & 3UL) << 3;
451 	addr = (void *) ((unsigned long) addr & ~3UL);
452 #else
453 #error "how many bits you are?!"
454 #endif
455 	return addr;
456 }
457 
458 static inline int mb_test_bit(int bit, void *addr)
459 {
460 	/*
461 	 * ext4_test_bit on architecture like powerpc
462 	 * needs unsigned long aligned address
463 	 */
464 	addr = mb_correct_addr_and_bit(&bit, addr);
465 	return ext4_test_bit(bit, addr);
466 }
467 
468 static inline void mb_set_bit(int bit, void *addr)
469 {
470 	addr = mb_correct_addr_and_bit(&bit, addr);
471 	ext4_set_bit(bit, addr);
472 }
473 
474 static inline void mb_clear_bit(int bit, void *addr)
475 {
476 	addr = mb_correct_addr_and_bit(&bit, addr);
477 	ext4_clear_bit(bit, addr);
478 }
479 
480 static inline int mb_test_and_clear_bit(int bit, void *addr)
481 {
482 	addr = mb_correct_addr_and_bit(&bit, addr);
483 	return ext4_test_and_clear_bit(bit, addr);
484 }
485 
486 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
487 {
488 	int fix = 0, ret, tmpmax;
489 	addr = mb_correct_addr_and_bit(&fix, addr);
490 	tmpmax = max + fix;
491 	start += fix;
492 
493 	ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
494 	if (ret > max)
495 		return max;
496 	return ret;
497 }
498 
499 static inline int mb_find_next_bit(void *addr, int max, int start)
500 {
501 	int fix = 0, ret, tmpmax;
502 	addr = mb_correct_addr_and_bit(&fix, addr);
503 	tmpmax = max + fix;
504 	start += fix;
505 
506 	ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
507 	if (ret > max)
508 		return max;
509 	return ret;
510 }
511 
512 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
513 {
514 	char *bb;
515 
516 	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
517 	BUG_ON(max == NULL);
518 
519 	if (order > e4b->bd_blkbits + 1) {
520 		*max = 0;
521 		return NULL;
522 	}
523 
524 	/* at order 0 we see each particular block */
525 	if (order == 0) {
526 		*max = 1 << (e4b->bd_blkbits + 3);
527 		return e4b->bd_bitmap;
528 	}
529 
530 	bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
531 	*max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
532 
533 	return bb;
534 }
535 
536 #ifdef DOUBLE_CHECK
537 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
538 			   int first, int count)
539 {
540 	int i;
541 	struct super_block *sb = e4b->bd_sb;
542 
543 	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
544 		return;
545 	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
546 	for (i = 0; i < count; i++) {
547 		if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
548 			ext4_fsblk_t blocknr;
549 
550 			blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
551 			blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
552 			ext4_grp_locked_error(sb, e4b->bd_group,
553 					      inode ? inode->i_ino : 0,
554 					      blocknr,
555 					      "freeing block already freed "
556 					      "(bit %u)",
557 					      first + i);
558 			ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
559 					EXT4_GROUP_INFO_BBITMAP_CORRUPT);
560 		}
561 		mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
562 	}
563 }
564 
565 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
566 {
567 	int i;
568 
569 	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
570 		return;
571 	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
572 	for (i = 0; i < count; i++) {
573 		BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
574 		mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
575 	}
576 }
577 
578 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
579 {
580 	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
581 		return;
582 	if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
583 		unsigned char *b1, *b2;
584 		int i;
585 		b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
586 		b2 = (unsigned char *) bitmap;
587 		for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
588 			if (b1[i] != b2[i]) {
589 				ext4_msg(e4b->bd_sb, KERN_ERR,
590 					 "corruption in group %u "
591 					 "at byte %u(%u): %x in copy != %x "
592 					 "on disk/prealloc",
593 					 e4b->bd_group, i, i * 8, b1[i], b2[i]);
594 				BUG();
595 			}
596 		}
597 	}
598 }
599 
600 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
601 			struct ext4_group_info *grp, ext4_group_t group)
602 {
603 	struct buffer_head *bh;
604 
605 	grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
606 	if (!grp->bb_bitmap)
607 		return;
608 
609 	bh = ext4_read_block_bitmap(sb, group);
610 	if (IS_ERR_OR_NULL(bh)) {
611 		kfree(grp->bb_bitmap);
612 		grp->bb_bitmap = NULL;
613 		return;
614 	}
615 
616 	memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
617 	put_bh(bh);
618 }
619 
620 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
621 {
622 	kfree(grp->bb_bitmap);
623 }
624 
625 #else
626 static inline void mb_free_blocks_double(struct inode *inode,
627 				struct ext4_buddy *e4b, int first, int count)
628 {
629 	return;
630 }
631 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
632 						int first, int count)
633 {
634 	return;
635 }
636 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
637 {
638 	return;
639 }
640 
641 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
642 			struct ext4_group_info *grp, ext4_group_t group)
643 {
644 	return;
645 }
646 
647 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
648 {
649 	return;
650 }
651 #endif
652 
653 #ifdef AGGRESSIVE_CHECK
654 
655 #define MB_CHECK_ASSERT(assert)						\
656 do {									\
657 	if (!(assert)) {						\
658 		printk(KERN_EMERG					\
659 			"Assertion failure in %s() at %s:%d: \"%s\"\n",	\
660 			function, file, line, # assert);		\
661 		BUG();							\
662 	}								\
663 } while (0)
664 
665 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
666 				const char *function, int line)
667 {
668 	struct super_block *sb = e4b->bd_sb;
669 	int order = e4b->bd_blkbits + 1;
670 	int max;
671 	int max2;
672 	int i;
673 	int j;
674 	int k;
675 	int count;
676 	struct ext4_group_info *grp;
677 	int fragments = 0;
678 	int fstart;
679 	struct list_head *cur;
680 	void *buddy;
681 	void *buddy2;
682 
683 	if (e4b->bd_info->bb_check_counter++ % 10)
684 		return 0;
685 
686 	while (order > 1) {
687 		buddy = mb_find_buddy(e4b, order, &max);
688 		MB_CHECK_ASSERT(buddy);
689 		buddy2 = mb_find_buddy(e4b, order - 1, &max2);
690 		MB_CHECK_ASSERT(buddy2);
691 		MB_CHECK_ASSERT(buddy != buddy2);
692 		MB_CHECK_ASSERT(max * 2 == max2);
693 
694 		count = 0;
695 		for (i = 0; i < max; i++) {
696 
697 			if (mb_test_bit(i, buddy)) {
698 				/* only single bit in buddy2 may be 0 */
699 				if (!mb_test_bit(i << 1, buddy2)) {
700 					MB_CHECK_ASSERT(
701 						mb_test_bit((i<<1)+1, buddy2));
702 				}
703 				continue;
704 			}
705 
706 			/* both bits in buddy2 must be 1 */
707 			MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
708 			MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
709 
710 			for (j = 0; j < (1 << order); j++) {
711 				k = (i * (1 << order)) + j;
712 				MB_CHECK_ASSERT(
713 					!mb_test_bit(k, e4b->bd_bitmap));
714 			}
715 			count++;
716 		}
717 		MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
718 		order--;
719 	}
720 
721 	fstart = -1;
722 	buddy = mb_find_buddy(e4b, 0, &max);
723 	for (i = 0; i < max; i++) {
724 		if (!mb_test_bit(i, buddy)) {
725 			MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
726 			if (fstart == -1) {
727 				fragments++;
728 				fstart = i;
729 			}
730 			continue;
731 		}
732 		fstart = -1;
733 		/* check used bits only */
734 		for (j = 0; j < e4b->bd_blkbits + 1; j++) {
735 			buddy2 = mb_find_buddy(e4b, j, &max2);
736 			k = i >> j;
737 			MB_CHECK_ASSERT(k < max2);
738 			MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
739 		}
740 	}
741 	MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
742 	MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
743 
744 	grp = ext4_get_group_info(sb, e4b->bd_group);
745 	list_for_each(cur, &grp->bb_prealloc_list) {
746 		ext4_group_t groupnr;
747 		struct ext4_prealloc_space *pa;
748 		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
749 		ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
750 		MB_CHECK_ASSERT(groupnr == e4b->bd_group);
751 		for (i = 0; i < pa->pa_len; i++)
752 			MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
753 	}
754 	return 0;
755 }
756 #undef MB_CHECK_ASSERT
757 #define mb_check_buddy(e4b) __mb_check_buddy(e4b,	\
758 					__FILE__, __func__, __LINE__)
759 #else
760 #define mb_check_buddy(e4b)
761 #endif
762 
763 /*
764  * Divide blocks started from @first with length @len into
765  * smaller chunks with power of 2 blocks.
766  * Clear the bits in bitmap which the blocks of the chunk(s) covered,
767  * then increase bb_counters[] for corresponded chunk size.
768  */
769 static void ext4_mb_mark_free_simple(struct super_block *sb,
770 				void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
771 					struct ext4_group_info *grp)
772 {
773 	struct ext4_sb_info *sbi = EXT4_SB(sb);
774 	ext4_grpblk_t min;
775 	ext4_grpblk_t max;
776 	ext4_grpblk_t chunk;
777 	unsigned int border;
778 
779 	BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
780 
781 	border = 2 << sb->s_blocksize_bits;
782 
783 	while (len > 0) {
784 		/* find how many blocks can be covered since this position */
785 		max = ffs(first | border) - 1;
786 
787 		/* find how many blocks of power 2 we need to mark */
788 		min = fls(len) - 1;
789 
790 		if (max < min)
791 			min = max;
792 		chunk = 1 << min;
793 
794 		/* mark multiblock chunks only */
795 		grp->bb_counters[min]++;
796 		if (min > 0)
797 			mb_clear_bit(first >> min,
798 				     buddy + sbi->s_mb_offsets[min]);
799 
800 		len -= chunk;
801 		first += chunk;
802 	}
803 }
804 
805 static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *new,
806 			int (*cmp)(struct rb_node *, struct rb_node *))
807 {
808 	struct rb_node **iter = &root->rb_node, *parent = NULL;
809 
810 	while (*iter) {
811 		parent = *iter;
812 		if (cmp(new, *iter) > 0)
813 			iter = &((*iter)->rb_left);
814 		else
815 			iter = &((*iter)->rb_right);
816 	}
817 
818 	rb_link_node(new, parent, iter);
819 	rb_insert_color(new, root);
820 }
821 
822 static int
823 ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2)
824 {
825 	struct ext4_group_info *grp1 = rb_entry(rb1,
826 						struct ext4_group_info,
827 						bb_avg_fragment_size_rb);
828 	struct ext4_group_info *grp2 = rb_entry(rb2,
829 						struct ext4_group_info,
830 						bb_avg_fragment_size_rb);
831 	int num_frags_1, num_frags_2;
832 
833 	num_frags_1 = grp1->bb_fragments ?
834 		grp1->bb_free / grp1->bb_fragments : 0;
835 	num_frags_2 = grp2->bb_fragments ?
836 		grp2->bb_free / grp2->bb_fragments : 0;
837 
838 	return (num_frags_2 - num_frags_1);
839 }
840 
841 /*
842  * Reinsert grpinfo into the avg_fragment_size tree with new average
843  * fragment size.
844  */
845 static void
846 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
847 {
848 	struct ext4_sb_info *sbi = EXT4_SB(sb);
849 
850 	if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0)
851 		return;
852 
853 	write_lock(&sbi->s_mb_rb_lock);
854 	if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
855 		rb_erase(&grp->bb_avg_fragment_size_rb,
856 				&sbi->s_mb_avg_fragment_size_root);
857 		RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
858 	}
859 
860 	ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root,
861 		&grp->bb_avg_fragment_size_rb,
862 		ext4_mb_avg_fragment_size_cmp);
863 	write_unlock(&sbi->s_mb_rb_lock);
864 }
865 
866 /*
867  * Choose next group by traversing largest_free_order lists. Updates *new_cr if
868  * cr level needs an update.
869  */
870 static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac,
871 			int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
872 {
873 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
874 	struct ext4_group_info *iter, *grp;
875 	int i;
876 
877 	if (ac->ac_status == AC_STATUS_FOUND)
878 		return;
879 
880 	if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED))
881 		atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
882 
883 	grp = NULL;
884 	for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
885 		if (list_empty(&sbi->s_mb_largest_free_orders[i]))
886 			continue;
887 		read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
888 		if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
889 			read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
890 			continue;
891 		}
892 		grp = NULL;
893 		list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
894 				    bb_largest_free_order_node) {
895 			if (sbi->s_mb_stats)
896 				atomic64_inc(&sbi->s_bal_cX_groups_considered[0]);
897 			if (likely(ext4_mb_good_group(ac, iter->bb_group, 0))) {
898 				grp = iter;
899 				break;
900 			}
901 		}
902 		read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
903 		if (grp)
904 			break;
905 	}
906 
907 	if (!grp) {
908 		/* Increment cr and search again */
909 		*new_cr = 1;
910 	} else {
911 		*group = grp->bb_group;
912 		ac->ac_last_optimal_group = *group;
913 		ac->ac_flags |= EXT4_MB_CR0_OPTIMIZED;
914 	}
915 }
916 
917 /*
918  * Choose next group by traversing average fragment size tree. Updates *new_cr
919  * if cr lvel needs an update. Sets EXT4_MB_SEARCH_NEXT_LINEAR to indicate that
920  * the linear search should continue for one iteration since there's lock
921  * contention on the rb tree lock.
922  */
923 static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac,
924 		int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
925 {
926 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
927 	int avg_fragment_size, best_so_far;
928 	struct rb_node *node, *found;
929 	struct ext4_group_info *grp;
930 
931 	/*
932 	 * If there is contention on the lock, instead of waiting for the lock
933 	 * to become available, just continue searching lineraly. We'll resume
934 	 * our rb tree search later starting at ac->ac_last_optimal_group.
935 	 */
936 	if (!read_trylock(&sbi->s_mb_rb_lock)) {
937 		ac->ac_flags |= EXT4_MB_SEARCH_NEXT_LINEAR;
938 		return;
939 	}
940 
941 	if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
942 		if (sbi->s_mb_stats)
943 			atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
944 		/* We have found something at CR 1 in the past */
945 		grp = ext4_get_group_info(ac->ac_sb, ac->ac_last_optimal_group);
946 		for (found = rb_next(&grp->bb_avg_fragment_size_rb); found != NULL;
947 		     found = rb_next(found)) {
948 			grp = rb_entry(found, struct ext4_group_info,
949 				       bb_avg_fragment_size_rb);
950 			if (sbi->s_mb_stats)
951 				atomic64_inc(&sbi->s_bal_cX_groups_considered[1]);
952 			if (likely(ext4_mb_good_group(ac, grp->bb_group, 1)))
953 				break;
954 		}
955 		goto done;
956 	}
957 
958 	node = sbi->s_mb_avg_fragment_size_root.rb_node;
959 	best_so_far = 0;
960 	found = NULL;
961 
962 	while (node) {
963 		grp = rb_entry(node, struct ext4_group_info,
964 			       bb_avg_fragment_size_rb);
965 		avg_fragment_size = 0;
966 		if (ext4_mb_good_group(ac, grp->bb_group, 1)) {
967 			avg_fragment_size = grp->bb_fragments ?
968 				grp->bb_free / grp->bb_fragments : 0;
969 			if (!best_so_far || avg_fragment_size < best_so_far) {
970 				best_so_far = avg_fragment_size;
971 				found = node;
972 			}
973 		}
974 		if (avg_fragment_size > ac->ac_g_ex.fe_len)
975 			node = node->rb_right;
976 		else
977 			node = node->rb_left;
978 	}
979 
980 done:
981 	if (found) {
982 		grp = rb_entry(found, struct ext4_group_info,
983 			       bb_avg_fragment_size_rb);
984 		*group = grp->bb_group;
985 		ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED;
986 	} else {
987 		*new_cr = 2;
988 	}
989 
990 	read_unlock(&sbi->s_mb_rb_lock);
991 	ac->ac_last_optimal_group = *group;
992 }
993 
994 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
995 {
996 	if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
997 		return 0;
998 	if (ac->ac_criteria >= 2)
999 		return 0;
1000 	if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1001 		return 0;
1002 	return 1;
1003 }
1004 
1005 /*
1006  * Return next linear group for allocation. If linear traversal should not be
1007  * performed, this function just returns the same group
1008  */
1009 static int
1010 next_linear_group(struct ext4_allocation_context *ac, int group, int ngroups)
1011 {
1012 	if (!should_optimize_scan(ac))
1013 		goto inc_and_return;
1014 
1015 	if (ac->ac_groups_linear_remaining) {
1016 		ac->ac_groups_linear_remaining--;
1017 		goto inc_and_return;
1018 	}
1019 
1020 	if (ac->ac_flags & EXT4_MB_SEARCH_NEXT_LINEAR) {
1021 		ac->ac_flags &= ~EXT4_MB_SEARCH_NEXT_LINEAR;
1022 		goto inc_and_return;
1023 	}
1024 
1025 	return group;
1026 inc_and_return:
1027 	/*
1028 	 * Artificially restricted ngroups for non-extent
1029 	 * files makes group > ngroups possible on first loop.
1030 	 */
1031 	return group + 1 >= ngroups ? 0 : group + 1;
1032 }
1033 
1034 /*
1035  * ext4_mb_choose_next_group: choose next group for allocation.
1036  *
1037  * @ac        Allocation Context
1038  * @new_cr    This is an output parameter. If the there is no good group
1039  *            available at current CR level, this field is updated to indicate
1040  *            the new cr level that should be used.
1041  * @group     This is an input / output parameter. As an input it indicates the
1042  *            next group that the allocator intends to use for allocation. As
1043  *            output, this field indicates the next group that should be used as
1044  *            determined by the optimization functions.
1045  * @ngroups   Total number of groups
1046  */
1047 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1048 		int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1049 {
1050 	*new_cr = ac->ac_criteria;
1051 
1052 	if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining)
1053 		return;
1054 
1055 	if (*new_cr == 0) {
1056 		ext4_mb_choose_next_group_cr0(ac, new_cr, group, ngroups);
1057 	} else if (*new_cr == 1) {
1058 		ext4_mb_choose_next_group_cr1(ac, new_cr, group, ngroups);
1059 	} else {
1060 		/*
1061 		 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1062 		 * bb_free. But until that happens, we should never come here.
1063 		 */
1064 		WARN_ON(1);
1065 	}
1066 }
1067 
1068 /*
1069  * Cache the order of the largest free extent we have available in this block
1070  * group.
1071  */
1072 static void
1073 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1074 {
1075 	struct ext4_sb_info *sbi = EXT4_SB(sb);
1076 	int i;
1077 
1078 	if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order >= 0) {
1079 		write_lock(&sbi->s_mb_largest_free_orders_locks[
1080 					      grp->bb_largest_free_order]);
1081 		list_del_init(&grp->bb_largest_free_order_node);
1082 		write_unlock(&sbi->s_mb_largest_free_orders_locks[
1083 					      grp->bb_largest_free_order]);
1084 	}
1085 	grp->bb_largest_free_order = -1; /* uninit */
1086 
1087 	for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) {
1088 		if (grp->bb_counters[i] > 0) {
1089 			grp->bb_largest_free_order = i;
1090 			break;
1091 		}
1092 	}
1093 	if (test_opt2(sb, MB_OPTIMIZE_SCAN) &&
1094 	    grp->bb_largest_free_order >= 0 && grp->bb_free) {
1095 		write_lock(&sbi->s_mb_largest_free_orders_locks[
1096 					      grp->bb_largest_free_order]);
1097 		list_add_tail(&grp->bb_largest_free_order_node,
1098 		      &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1099 		write_unlock(&sbi->s_mb_largest_free_orders_locks[
1100 					      grp->bb_largest_free_order]);
1101 	}
1102 }
1103 
1104 static noinline_for_stack
1105 void ext4_mb_generate_buddy(struct super_block *sb,
1106 				void *buddy, void *bitmap, ext4_group_t group)
1107 {
1108 	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
1109 	struct ext4_sb_info *sbi = EXT4_SB(sb);
1110 	ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1111 	ext4_grpblk_t i = 0;
1112 	ext4_grpblk_t first;
1113 	ext4_grpblk_t len;
1114 	unsigned free = 0;
1115 	unsigned fragments = 0;
1116 	unsigned long long period = get_cycles();
1117 
1118 	/* initialize buddy from bitmap which is aggregation
1119 	 * of on-disk bitmap and preallocations */
1120 	i = mb_find_next_zero_bit(bitmap, max, 0);
1121 	grp->bb_first_free = i;
1122 	while (i < max) {
1123 		fragments++;
1124 		first = i;
1125 		i = mb_find_next_bit(bitmap, max, i);
1126 		len = i - first;
1127 		free += len;
1128 		if (len > 1)
1129 			ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1130 		else
1131 			grp->bb_counters[0]++;
1132 		if (i < max)
1133 			i = mb_find_next_zero_bit(bitmap, max, i);
1134 	}
1135 	grp->bb_fragments = fragments;
1136 
1137 	if (free != grp->bb_free) {
1138 		ext4_grp_locked_error(sb, group, 0, 0,
1139 				      "block bitmap and bg descriptor "
1140 				      "inconsistent: %u vs %u free clusters",
1141 				      free, grp->bb_free);
1142 		/*
1143 		 * If we intend to continue, we consider group descriptor
1144 		 * corrupt and update bb_free using bitmap value
1145 		 */
1146 		grp->bb_free = free;
1147 		ext4_mark_group_bitmap_corrupted(sb, group,
1148 					EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1149 	}
1150 	mb_set_largest_free_order(sb, grp);
1151 
1152 	clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1153 
1154 	period = get_cycles() - period;
1155 	atomic_inc(&sbi->s_mb_buddies_generated);
1156 	atomic64_add(period, &sbi->s_mb_generation_time);
1157 	mb_update_avg_fragment_size(sb, grp);
1158 }
1159 
1160 /* The buddy information is attached the buddy cache inode
1161  * for convenience. The information regarding each group
1162  * is loaded via ext4_mb_load_buddy. The information involve
1163  * block bitmap and buddy information. The information are
1164  * stored in the inode as
1165  *
1166  * {                        page                        }
1167  * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1168  *
1169  *
1170  * one block each for bitmap and buddy information.
1171  * So for each group we take up 2 blocks. A page can
1172  * contain blocks_per_page (PAGE_SIZE / blocksize)  blocks.
1173  * So it can have information regarding groups_per_page which
1174  * is blocks_per_page/2
1175  *
1176  * Locking note:  This routine takes the block group lock of all groups
1177  * for this page; do not hold this lock when calling this routine!
1178  */
1179 
1180 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1181 {
1182 	ext4_group_t ngroups;
1183 	int blocksize;
1184 	int blocks_per_page;
1185 	int groups_per_page;
1186 	int err = 0;
1187 	int i;
1188 	ext4_group_t first_group, group;
1189 	int first_block;
1190 	struct super_block *sb;
1191 	struct buffer_head *bhs;
1192 	struct buffer_head **bh = NULL;
1193 	struct inode *inode;
1194 	char *data;
1195 	char *bitmap;
1196 	struct ext4_group_info *grinfo;
1197 
1198 	inode = page->mapping->host;
1199 	sb = inode->i_sb;
1200 	ngroups = ext4_get_groups_count(sb);
1201 	blocksize = i_blocksize(inode);
1202 	blocks_per_page = PAGE_SIZE / blocksize;
1203 
1204 	mb_debug(sb, "init page %lu\n", page->index);
1205 
1206 	groups_per_page = blocks_per_page >> 1;
1207 	if (groups_per_page == 0)
1208 		groups_per_page = 1;
1209 
1210 	/* allocate buffer_heads to read bitmaps */
1211 	if (groups_per_page > 1) {
1212 		i = sizeof(struct buffer_head *) * groups_per_page;
1213 		bh = kzalloc(i, gfp);
1214 		if (bh == NULL) {
1215 			err = -ENOMEM;
1216 			goto out;
1217 		}
1218 	} else
1219 		bh = &bhs;
1220 
1221 	first_group = page->index * blocks_per_page / 2;
1222 
1223 	/* read all groups the page covers into the cache */
1224 	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1225 		if (group >= ngroups)
1226 			break;
1227 
1228 		grinfo = ext4_get_group_info(sb, group);
1229 		/*
1230 		 * If page is uptodate then we came here after online resize
1231 		 * which added some new uninitialized group info structs, so
1232 		 * we must skip all initialized uptodate buddies on the page,
1233 		 * which may be currently in use by an allocating task.
1234 		 */
1235 		if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1236 			bh[i] = NULL;
1237 			continue;
1238 		}
1239 		bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1240 		if (IS_ERR(bh[i])) {
1241 			err = PTR_ERR(bh[i]);
1242 			bh[i] = NULL;
1243 			goto out;
1244 		}
1245 		mb_debug(sb, "read bitmap for group %u\n", group);
1246 	}
1247 
1248 	/* wait for I/O completion */
1249 	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1250 		int err2;
1251 
1252 		if (!bh[i])
1253 			continue;
1254 		err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1255 		if (!err)
1256 			err = err2;
1257 	}
1258 
1259 	first_block = page->index * blocks_per_page;
1260 	for (i = 0; i < blocks_per_page; i++) {
1261 		group = (first_block + i) >> 1;
1262 		if (group >= ngroups)
1263 			break;
1264 
1265 		if (!bh[group - first_group])
1266 			/* skip initialized uptodate buddy */
1267 			continue;
1268 
1269 		if (!buffer_verified(bh[group - first_group]))
1270 			/* Skip faulty bitmaps */
1271 			continue;
1272 		err = 0;
1273 
1274 		/*
1275 		 * data carry information regarding this
1276 		 * particular group in the format specified
1277 		 * above
1278 		 *
1279 		 */
1280 		data = page_address(page) + (i * blocksize);
1281 		bitmap = bh[group - first_group]->b_data;
1282 
1283 		/*
1284 		 * We place the buddy block and bitmap block
1285 		 * close together
1286 		 */
1287 		if ((first_block + i) & 1) {
1288 			/* this is block of buddy */
1289 			BUG_ON(incore == NULL);
1290 			mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1291 				group, page->index, i * blocksize);
1292 			trace_ext4_mb_buddy_bitmap_load(sb, group);
1293 			grinfo = ext4_get_group_info(sb, group);
1294 			grinfo->bb_fragments = 0;
1295 			memset(grinfo->bb_counters, 0,
1296 			       sizeof(*grinfo->bb_counters) *
1297 			       (MB_NUM_ORDERS(sb)));
1298 			/*
1299 			 * incore got set to the group block bitmap below
1300 			 */
1301 			ext4_lock_group(sb, group);
1302 			/* init the buddy */
1303 			memset(data, 0xff, blocksize);
1304 			ext4_mb_generate_buddy(sb, data, incore, group);
1305 			ext4_unlock_group(sb, group);
1306 			incore = NULL;
1307 		} else {
1308 			/* this is block of bitmap */
1309 			BUG_ON(incore != NULL);
1310 			mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1311 				group, page->index, i * blocksize);
1312 			trace_ext4_mb_bitmap_load(sb, group);
1313 
1314 			/* see comments in ext4_mb_put_pa() */
1315 			ext4_lock_group(sb, group);
1316 			memcpy(data, bitmap, blocksize);
1317 
1318 			/* mark all preallocated blks used in in-core bitmap */
1319 			ext4_mb_generate_from_pa(sb, data, group);
1320 			ext4_mb_generate_from_freelist(sb, data, group);
1321 			ext4_unlock_group(sb, group);
1322 
1323 			/* set incore so that the buddy information can be
1324 			 * generated using this
1325 			 */
1326 			incore = data;
1327 		}
1328 	}
1329 	SetPageUptodate(page);
1330 
1331 out:
1332 	if (bh) {
1333 		for (i = 0; i < groups_per_page; i++)
1334 			brelse(bh[i]);
1335 		if (bh != &bhs)
1336 			kfree(bh);
1337 	}
1338 	return err;
1339 }
1340 
1341 /*
1342  * Lock the buddy and bitmap pages. This make sure other parallel init_group
1343  * on the same buddy page doesn't happen whild holding the buddy page lock.
1344  * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1345  * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1346  */
1347 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1348 		ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1349 {
1350 	struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1351 	int block, pnum, poff;
1352 	int blocks_per_page;
1353 	struct page *page;
1354 
1355 	e4b->bd_buddy_page = NULL;
1356 	e4b->bd_bitmap_page = NULL;
1357 
1358 	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1359 	/*
1360 	 * the buddy cache inode stores the block bitmap
1361 	 * and buddy information in consecutive blocks.
1362 	 * So for each group we need two blocks.
1363 	 */
1364 	block = group * 2;
1365 	pnum = block / blocks_per_page;
1366 	poff = block % blocks_per_page;
1367 	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1368 	if (!page)
1369 		return -ENOMEM;
1370 	BUG_ON(page->mapping != inode->i_mapping);
1371 	e4b->bd_bitmap_page = page;
1372 	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1373 
1374 	if (blocks_per_page >= 2) {
1375 		/* buddy and bitmap are on the same page */
1376 		return 0;
1377 	}
1378 
1379 	block++;
1380 	pnum = block / blocks_per_page;
1381 	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1382 	if (!page)
1383 		return -ENOMEM;
1384 	BUG_ON(page->mapping != inode->i_mapping);
1385 	e4b->bd_buddy_page = page;
1386 	return 0;
1387 }
1388 
1389 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1390 {
1391 	if (e4b->bd_bitmap_page) {
1392 		unlock_page(e4b->bd_bitmap_page);
1393 		put_page(e4b->bd_bitmap_page);
1394 	}
1395 	if (e4b->bd_buddy_page) {
1396 		unlock_page(e4b->bd_buddy_page);
1397 		put_page(e4b->bd_buddy_page);
1398 	}
1399 }
1400 
1401 /*
1402  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1403  * block group lock of all groups for this page; do not hold the BG lock when
1404  * calling this routine!
1405  */
1406 static noinline_for_stack
1407 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1408 {
1409 
1410 	struct ext4_group_info *this_grp;
1411 	struct ext4_buddy e4b;
1412 	struct page *page;
1413 	int ret = 0;
1414 
1415 	might_sleep();
1416 	mb_debug(sb, "init group %u\n", group);
1417 	this_grp = ext4_get_group_info(sb, group);
1418 	/*
1419 	 * This ensures that we don't reinit the buddy cache
1420 	 * page which map to the group from which we are already
1421 	 * allocating. If we are looking at the buddy cache we would
1422 	 * have taken a reference using ext4_mb_load_buddy and that
1423 	 * would have pinned buddy page to page cache.
1424 	 * The call to ext4_mb_get_buddy_page_lock will mark the
1425 	 * page accessed.
1426 	 */
1427 	ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1428 	if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1429 		/*
1430 		 * somebody initialized the group
1431 		 * return without doing anything
1432 		 */
1433 		goto err;
1434 	}
1435 
1436 	page = e4b.bd_bitmap_page;
1437 	ret = ext4_mb_init_cache(page, NULL, gfp);
1438 	if (ret)
1439 		goto err;
1440 	if (!PageUptodate(page)) {
1441 		ret = -EIO;
1442 		goto err;
1443 	}
1444 
1445 	if (e4b.bd_buddy_page == NULL) {
1446 		/*
1447 		 * If both the bitmap and buddy are in
1448 		 * the same page we don't need to force
1449 		 * init the buddy
1450 		 */
1451 		ret = 0;
1452 		goto err;
1453 	}
1454 	/* init buddy cache */
1455 	page = e4b.bd_buddy_page;
1456 	ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1457 	if (ret)
1458 		goto err;
1459 	if (!PageUptodate(page)) {
1460 		ret = -EIO;
1461 		goto err;
1462 	}
1463 err:
1464 	ext4_mb_put_buddy_page_lock(&e4b);
1465 	return ret;
1466 }
1467 
1468 /*
1469  * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1470  * block group lock of all groups for this page; do not hold the BG lock when
1471  * calling this routine!
1472  */
1473 static noinline_for_stack int
1474 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1475 		       struct ext4_buddy *e4b, gfp_t gfp)
1476 {
1477 	int blocks_per_page;
1478 	int block;
1479 	int pnum;
1480 	int poff;
1481 	struct page *page;
1482 	int ret;
1483 	struct ext4_group_info *grp;
1484 	struct ext4_sb_info *sbi = EXT4_SB(sb);
1485 	struct inode *inode = sbi->s_buddy_cache;
1486 
1487 	might_sleep();
1488 	mb_debug(sb, "load group %u\n", group);
1489 
1490 	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1491 	grp = ext4_get_group_info(sb, group);
1492 
1493 	e4b->bd_blkbits = sb->s_blocksize_bits;
1494 	e4b->bd_info = grp;
1495 	e4b->bd_sb = sb;
1496 	e4b->bd_group = group;
1497 	e4b->bd_buddy_page = NULL;
1498 	e4b->bd_bitmap_page = NULL;
1499 
1500 	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1501 		/*
1502 		 * we need full data about the group
1503 		 * to make a good selection
1504 		 */
1505 		ret = ext4_mb_init_group(sb, group, gfp);
1506 		if (ret)
1507 			return ret;
1508 	}
1509 
1510 	/*
1511 	 * the buddy cache inode stores the block bitmap
1512 	 * and buddy information in consecutive blocks.
1513 	 * So for each group we need two blocks.
1514 	 */
1515 	block = group * 2;
1516 	pnum = block / blocks_per_page;
1517 	poff = block % blocks_per_page;
1518 
1519 	/* we could use find_or_create_page(), but it locks page
1520 	 * what we'd like to avoid in fast path ... */
1521 	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1522 	if (page == NULL || !PageUptodate(page)) {
1523 		if (page)
1524 			/*
1525 			 * drop the page reference and try
1526 			 * to get the page with lock. If we
1527 			 * are not uptodate that implies
1528 			 * somebody just created the page but
1529 			 * is yet to initialize the same. So
1530 			 * wait for it to initialize.
1531 			 */
1532 			put_page(page);
1533 		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1534 		if (page) {
1535 			BUG_ON(page->mapping != inode->i_mapping);
1536 			if (!PageUptodate(page)) {
1537 				ret = ext4_mb_init_cache(page, NULL, gfp);
1538 				if (ret) {
1539 					unlock_page(page);
1540 					goto err;
1541 				}
1542 				mb_cmp_bitmaps(e4b, page_address(page) +
1543 					       (poff * sb->s_blocksize));
1544 			}
1545 			unlock_page(page);
1546 		}
1547 	}
1548 	if (page == NULL) {
1549 		ret = -ENOMEM;
1550 		goto err;
1551 	}
1552 	if (!PageUptodate(page)) {
1553 		ret = -EIO;
1554 		goto err;
1555 	}
1556 
1557 	/* Pages marked accessed already */
1558 	e4b->bd_bitmap_page = page;
1559 	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1560 
1561 	block++;
1562 	pnum = block / blocks_per_page;
1563 	poff = block % blocks_per_page;
1564 
1565 	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1566 	if (page == NULL || !PageUptodate(page)) {
1567 		if (page)
1568 			put_page(page);
1569 		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1570 		if (page) {
1571 			BUG_ON(page->mapping != inode->i_mapping);
1572 			if (!PageUptodate(page)) {
1573 				ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1574 							 gfp);
1575 				if (ret) {
1576 					unlock_page(page);
1577 					goto err;
1578 				}
1579 			}
1580 			unlock_page(page);
1581 		}
1582 	}
1583 	if (page == NULL) {
1584 		ret = -ENOMEM;
1585 		goto err;
1586 	}
1587 	if (!PageUptodate(page)) {
1588 		ret = -EIO;
1589 		goto err;
1590 	}
1591 
1592 	/* Pages marked accessed already */
1593 	e4b->bd_buddy_page = page;
1594 	e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1595 
1596 	return 0;
1597 
1598 err:
1599 	if (page)
1600 		put_page(page);
1601 	if (e4b->bd_bitmap_page)
1602 		put_page(e4b->bd_bitmap_page);
1603 	if (e4b->bd_buddy_page)
1604 		put_page(e4b->bd_buddy_page);
1605 	e4b->bd_buddy = NULL;
1606 	e4b->bd_bitmap = NULL;
1607 	return ret;
1608 }
1609 
1610 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1611 			      struct ext4_buddy *e4b)
1612 {
1613 	return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1614 }
1615 
1616 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1617 {
1618 	if (e4b->bd_bitmap_page)
1619 		put_page(e4b->bd_bitmap_page);
1620 	if (e4b->bd_buddy_page)
1621 		put_page(e4b->bd_buddy_page);
1622 }
1623 
1624 
1625 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1626 {
1627 	int order = 1, max;
1628 	void *bb;
1629 
1630 	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1631 	BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1632 
1633 	while (order <= e4b->bd_blkbits + 1) {
1634 		bb = mb_find_buddy(e4b, order, &max);
1635 		if (!mb_test_bit(block >> order, bb)) {
1636 			/* this block is part of buddy of order 'order' */
1637 			return order;
1638 		}
1639 		order++;
1640 	}
1641 	return 0;
1642 }
1643 
1644 static void mb_clear_bits(void *bm, int cur, int len)
1645 {
1646 	__u32 *addr;
1647 
1648 	len = cur + len;
1649 	while (cur < len) {
1650 		if ((cur & 31) == 0 && (len - cur) >= 32) {
1651 			/* fast path: clear whole word at once */
1652 			addr = bm + (cur >> 3);
1653 			*addr = 0;
1654 			cur += 32;
1655 			continue;
1656 		}
1657 		mb_clear_bit(cur, bm);
1658 		cur++;
1659 	}
1660 }
1661 
1662 /* clear bits in given range
1663  * will return first found zero bit if any, -1 otherwise
1664  */
1665 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1666 {
1667 	__u32 *addr;
1668 	int zero_bit = -1;
1669 
1670 	len = cur + len;
1671 	while (cur < len) {
1672 		if ((cur & 31) == 0 && (len - cur) >= 32) {
1673 			/* fast path: clear whole word at once */
1674 			addr = bm + (cur >> 3);
1675 			if (*addr != (__u32)(-1) && zero_bit == -1)
1676 				zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1677 			*addr = 0;
1678 			cur += 32;
1679 			continue;
1680 		}
1681 		if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1682 			zero_bit = cur;
1683 		cur++;
1684 	}
1685 
1686 	return zero_bit;
1687 }
1688 
1689 void mb_set_bits(void *bm, int cur, int len)
1690 {
1691 	__u32 *addr;
1692 
1693 	len = cur + len;
1694 	while (cur < len) {
1695 		if ((cur & 31) == 0 && (len - cur) >= 32) {
1696 			/* fast path: set whole word at once */
1697 			addr = bm + (cur >> 3);
1698 			*addr = 0xffffffff;
1699 			cur += 32;
1700 			continue;
1701 		}
1702 		mb_set_bit(cur, bm);
1703 		cur++;
1704 	}
1705 }
1706 
1707 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1708 {
1709 	if (mb_test_bit(*bit + side, bitmap)) {
1710 		mb_clear_bit(*bit, bitmap);
1711 		(*bit) -= side;
1712 		return 1;
1713 	}
1714 	else {
1715 		(*bit) += side;
1716 		mb_set_bit(*bit, bitmap);
1717 		return -1;
1718 	}
1719 }
1720 
1721 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1722 {
1723 	int max;
1724 	int order = 1;
1725 	void *buddy = mb_find_buddy(e4b, order, &max);
1726 
1727 	while (buddy) {
1728 		void *buddy2;
1729 
1730 		/* Bits in range [first; last] are known to be set since
1731 		 * corresponding blocks were allocated. Bits in range
1732 		 * (first; last) will stay set because they form buddies on
1733 		 * upper layer. We just deal with borders if they don't
1734 		 * align with upper layer and then go up.
1735 		 * Releasing entire group is all about clearing
1736 		 * single bit of highest order buddy.
1737 		 */
1738 
1739 		/* Example:
1740 		 * ---------------------------------
1741 		 * |   1   |   1   |   1   |   1   |
1742 		 * ---------------------------------
1743 		 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1744 		 * ---------------------------------
1745 		 *   0   1   2   3   4   5   6   7
1746 		 *      \_____________________/
1747 		 *
1748 		 * Neither [1] nor [6] is aligned to above layer.
1749 		 * Left neighbour [0] is free, so mark it busy,
1750 		 * decrease bb_counters and extend range to
1751 		 * [0; 6]
1752 		 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1753 		 * mark [6] free, increase bb_counters and shrink range to
1754 		 * [0; 5].
1755 		 * Then shift range to [0; 2], go up and do the same.
1756 		 */
1757 
1758 
1759 		if (first & 1)
1760 			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1761 		if (!(last & 1))
1762 			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1763 		if (first > last)
1764 			break;
1765 		order++;
1766 
1767 		if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1768 			mb_clear_bits(buddy, first, last - first + 1);
1769 			e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1770 			break;
1771 		}
1772 		first >>= 1;
1773 		last >>= 1;
1774 		buddy = buddy2;
1775 	}
1776 }
1777 
1778 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1779 			   int first, int count)
1780 {
1781 	int left_is_free = 0;
1782 	int right_is_free = 0;
1783 	int block;
1784 	int last = first + count - 1;
1785 	struct super_block *sb = e4b->bd_sb;
1786 
1787 	if (WARN_ON(count == 0))
1788 		return;
1789 	BUG_ON(last >= (sb->s_blocksize << 3));
1790 	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1791 	/* Don't bother if the block group is corrupt. */
1792 	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1793 		return;
1794 
1795 	mb_check_buddy(e4b);
1796 	mb_free_blocks_double(inode, e4b, first, count);
1797 
1798 	this_cpu_inc(discard_pa_seq);
1799 	e4b->bd_info->bb_free += count;
1800 	if (first < e4b->bd_info->bb_first_free)
1801 		e4b->bd_info->bb_first_free = first;
1802 
1803 	/* access memory sequentially: check left neighbour,
1804 	 * clear range and then check right neighbour
1805 	 */
1806 	if (first != 0)
1807 		left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1808 	block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1809 	if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1810 		right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1811 
1812 	if (unlikely(block != -1)) {
1813 		struct ext4_sb_info *sbi = EXT4_SB(sb);
1814 		ext4_fsblk_t blocknr;
1815 
1816 		blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1817 		blocknr += EXT4_C2B(sbi, block);
1818 		if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1819 			ext4_grp_locked_error(sb, e4b->bd_group,
1820 					      inode ? inode->i_ino : 0,
1821 					      blocknr,
1822 					      "freeing already freed block (bit %u); block bitmap corrupt.",
1823 					      block);
1824 			ext4_mark_group_bitmap_corrupted(
1825 				sb, e4b->bd_group,
1826 				EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1827 		}
1828 		goto done;
1829 	}
1830 
1831 	/* let's maintain fragments counter */
1832 	if (left_is_free && right_is_free)
1833 		e4b->bd_info->bb_fragments--;
1834 	else if (!left_is_free && !right_is_free)
1835 		e4b->bd_info->bb_fragments++;
1836 
1837 	/* buddy[0] == bd_bitmap is a special case, so handle
1838 	 * it right away and let mb_buddy_mark_free stay free of
1839 	 * zero order checks.
1840 	 * Check if neighbours are to be coaleasced,
1841 	 * adjust bitmap bb_counters and borders appropriately.
1842 	 */
1843 	if (first & 1) {
1844 		first += !left_is_free;
1845 		e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1846 	}
1847 	if (!(last & 1)) {
1848 		last -= !right_is_free;
1849 		e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1850 	}
1851 
1852 	if (first <= last)
1853 		mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1854 
1855 done:
1856 	mb_set_largest_free_order(sb, e4b->bd_info);
1857 	mb_update_avg_fragment_size(sb, e4b->bd_info);
1858 	mb_check_buddy(e4b);
1859 }
1860 
1861 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1862 				int needed, struct ext4_free_extent *ex)
1863 {
1864 	int next = block;
1865 	int max, order;
1866 	void *buddy;
1867 
1868 	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1869 	BUG_ON(ex == NULL);
1870 
1871 	buddy = mb_find_buddy(e4b, 0, &max);
1872 	BUG_ON(buddy == NULL);
1873 	BUG_ON(block >= max);
1874 	if (mb_test_bit(block, buddy)) {
1875 		ex->fe_len = 0;
1876 		ex->fe_start = 0;
1877 		ex->fe_group = 0;
1878 		return 0;
1879 	}
1880 
1881 	/* find actual order */
1882 	order = mb_find_order_for_block(e4b, block);
1883 	block = block >> order;
1884 
1885 	ex->fe_len = 1 << order;
1886 	ex->fe_start = block << order;
1887 	ex->fe_group = e4b->bd_group;
1888 
1889 	/* calc difference from given start */
1890 	next = next - ex->fe_start;
1891 	ex->fe_len -= next;
1892 	ex->fe_start += next;
1893 
1894 	while (needed > ex->fe_len &&
1895 	       mb_find_buddy(e4b, order, &max)) {
1896 
1897 		if (block + 1 >= max)
1898 			break;
1899 
1900 		next = (block + 1) * (1 << order);
1901 		if (mb_test_bit(next, e4b->bd_bitmap))
1902 			break;
1903 
1904 		order = mb_find_order_for_block(e4b, next);
1905 
1906 		block = next >> order;
1907 		ex->fe_len += 1 << order;
1908 	}
1909 
1910 	if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1911 		/* Should never happen! (but apparently sometimes does?!?) */
1912 		WARN_ON(1);
1913 		ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
1914 			"corruption or bug in mb_find_extent "
1915 			"block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1916 			block, order, needed, ex->fe_group, ex->fe_start,
1917 			ex->fe_len, ex->fe_logical);
1918 		ex->fe_len = 0;
1919 		ex->fe_start = 0;
1920 		ex->fe_group = 0;
1921 	}
1922 	return ex->fe_len;
1923 }
1924 
1925 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1926 {
1927 	int ord;
1928 	int mlen = 0;
1929 	int max = 0;
1930 	int cur;
1931 	int start = ex->fe_start;
1932 	int len = ex->fe_len;
1933 	unsigned ret = 0;
1934 	int len0 = len;
1935 	void *buddy;
1936 
1937 	BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1938 	BUG_ON(e4b->bd_group != ex->fe_group);
1939 	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1940 	mb_check_buddy(e4b);
1941 	mb_mark_used_double(e4b, start, len);
1942 
1943 	this_cpu_inc(discard_pa_seq);
1944 	e4b->bd_info->bb_free -= len;
1945 	if (e4b->bd_info->bb_first_free == start)
1946 		e4b->bd_info->bb_first_free += len;
1947 
1948 	/* let's maintain fragments counter */
1949 	if (start != 0)
1950 		mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1951 	if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1952 		max = !mb_test_bit(start + len, e4b->bd_bitmap);
1953 	if (mlen && max)
1954 		e4b->bd_info->bb_fragments++;
1955 	else if (!mlen && !max)
1956 		e4b->bd_info->bb_fragments--;
1957 
1958 	/* let's maintain buddy itself */
1959 	while (len) {
1960 		ord = mb_find_order_for_block(e4b, start);
1961 
1962 		if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1963 			/* the whole chunk may be allocated at once! */
1964 			mlen = 1 << ord;
1965 			buddy = mb_find_buddy(e4b, ord, &max);
1966 			BUG_ON((start >> ord) >= max);
1967 			mb_set_bit(start >> ord, buddy);
1968 			e4b->bd_info->bb_counters[ord]--;
1969 			start += mlen;
1970 			len -= mlen;
1971 			BUG_ON(len < 0);
1972 			continue;
1973 		}
1974 
1975 		/* store for history */
1976 		if (ret == 0)
1977 			ret = len | (ord << 16);
1978 
1979 		/* we have to split large buddy */
1980 		BUG_ON(ord <= 0);
1981 		buddy = mb_find_buddy(e4b, ord, &max);
1982 		mb_set_bit(start >> ord, buddy);
1983 		e4b->bd_info->bb_counters[ord]--;
1984 
1985 		ord--;
1986 		cur = (start >> ord) & ~1U;
1987 		buddy = mb_find_buddy(e4b, ord, &max);
1988 		mb_clear_bit(cur, buddy);
1989 		mb_clear_bit(cur + 1, buddy);
1990 		e4b->bd_info->bb_counters[ord]++;
1991 		e4b->bd_info->bb_counters[ord]++;
1992 	}
1993 	mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1994 
1995 	mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
1996 	mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1997 	mb_check_buddy(e4b);
1998 
1999 	return ret;
2000 }
2001 
2002 /*
2003  * Must be called under group lock!
2004  */
2005 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2006 					struct ext4_buddy *e4b)
2007 {
2008 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2009 	int ret;
2010 
2011 	BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2012 	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2013 
2014 	ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2015 	ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2016 	ret = mb_mark_used(e4b, &ac->ac_b_ex);
2017 
2018 	/* preallocation can change ac_b_ex, thus we store actually
2019 	 * allocated blocks for history */
2020 	ac->ac_f_ex = ac->ac_b_ex;
2021 
2022 	ac->ac_status = AC_STATUS_FOUND;
2023 	ac->ac_tail = ret & 0xffff;
2024 	ac->ac_buddy = ret >> 16;
2025 
2026 	/*
2027 	 * take the page reference. We want the page to be pinned
2028 	 * so that we don't get a ext4_mb_init_cache_call for this
2029 	 * group until we update the bitmap. That would mean we
2030 	 * double allocate blocks. The reference is dropped
2031 	 * in ext4_mb_release_context
2032 	 */
2033 	ac->ac_bitmap_page = e4b->bd_bitmap_page;
2034 	get_page(ac->ac_bitmap_page);
2035 	ac->ac_buddy_page = e4b->bd_buddy_page;
2036 	get_page(ac->ac_buddy_page);
2037 	/* store last allocated for subsequent stream allocation */
2038 	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2039 		spin_lock(&sbi->s_md_lock);
2040 		sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2041 		sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2042 		spin_unlock(&sbi->s_md_lock);
2043 	}
2044 	/*
2045 	 * As we've just preallocated more space than
2046 	 * user requested originally, we store allocated
2047 	 * space in a special descriptor.
2048 	 */
2049 	if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2050 		ext4_mb_new_preallocation(ac);
2051 
2052 }
2053 
2054 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2055 					struct ext4_buddy *e4b,
2056 					int finish_group)
2057 {
2058 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2059 	struct ext4_free_extent *bex = &ac->ac_b_ex;
2060 	struct ext4_free_extent *gex = &ac->ac_g_ex;
2061 	struct ext4_free_extent ex;
2062 	int max;
2063 
2064 	if (ac->ac_status == AC_STATUS_FOUND)
2065 		return;
2066 	/*
2067 	 * We don't want to scan for a whole year
2068 	 */
2069 	if (ac->ac_found > sbi->s_mb_max_to_scan &&
2070 			!(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2071 		ac->ac_status = AC_STATUS_BREAK;
2072 		return;
2073 	}
2074 
2075 	/*
2076 	 * Haven't found good chunk so far, let's continue
2077 	 */
2078 	if (bex->fe_len < gex->fe_len)
2079 		return;
2080 
2081 	if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2082 			&& bex->fe_group == e4b->bd_group) {
2083 		/* recheck chunk's availability - we don't know
2084 		 * when it was found (within this lock-unlock
2085 		 * period or not) */
2086 		max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
2087 		if (max >= gex->fe_len) {
2088 			ext4_mb_use_best_found(ac, e4b);
2089 			return;
2090 		}
2091 	}
2092 }
2093 
2094 /*
2095  * The routine checks whether found extent is good enough. If it is,
2096  * then the extent gets marked used and flag is set to the context
2097  * to stop scanning. Otherwise, the extent is compared with the
2098  * previous found extent and if new one is better, then it's stored
2099  * in the context. Later, the best found extent will be used, if
2100  * mballoc can't find good enough extent.
2101  *
2102  * FIXME: real allocation policy is to be designed yet!
2103  */
2104 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2105 					struct ext4_free_extent *ex,
2106 					struct ext4_buddy *e4b)
2107 {
2108 	struct ext4_free_extent *bex = &ac->ac_b_ex;
2109 	struct ext4_free_extent *gex = &ac->ac_g_ex;
2110 
2111 	BUG_ON(ex->fe_len <= 0);
2112 	BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2113 	BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2114 	BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2115 
2116 	ac->ac_found++;
2117 
2118 	/*
2119 	 * The special case - take what you catch first
2120 	 */
2121 	if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2122 		*bex = *ex;
2123 		ext4_mb_use_best_found(ac, e4b);
2124 		return;
2125 	}
2126 
2127 	/*
2128 	 * Let's check whether the chuck is good enough
2129 	 */
2130 	if (ex->fe_len == gex->fe_len) {
2131 		*bex = *ex;
2132 		ext4_mb_use_best_found(ac, e4b);
2133 		return;
2134 	}
2135 
2136 	/*
2137 	 * If this is first found extent, just store it in the context
2138 	 */
2139 	if (bex->fe_len == 0) {
2140 		*bex = *ex;
2141 		return;
2142 	}
2143 
2144 	/*
2145 	 * If new found extent is better, store it in the context
2146 	 */
2147 	if (bex->fe_len < gex->fe_len) {
2148 		/* if the request isn't satisfied, any found extent
2149 		 * larger than previous best one is better */
2150 		if (ex->fe_len > bex->fe_len)
2151 			*bex = *ex;
2152 	} else if (ex->fe_len > gex->fe_len) {
2153 		/* if the request is satisfied, then we try to find
2154 		 * an extent that still satisfy the request, but is
2155 		 * smaller than previous one */
2156 		if (ex->fe_len < bex->fe_len)
2157 			*bex = *ex;
2158 	}
2159 
2160 	ext4_mb_check_limits(ac, e4b, 0);
2161 }
2162 
2163 static noinline_for_stack
2164 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2165 					struct ext4_buddy *e4b)
2166 {
2167 	struct ext4_free_extent ex = ac->ac_b_ex;
2168 	ext4_group_t group = ex.fe_group;
2169 	int max;
2170 	int err;
2171 
2172 	BUG_ON(ex.fe_len <= 0);
2173 	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2174 	if (err)
2175 		return err;
2176 
2177 	ext4_lock_group(ac->ac_sb, group);
2178 	max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2179 
2180 	if (max > 0) {
2181 		ac->ac_b_ex = ex;
2182 		ext4_mb_use_best_found(ac, e4b);
2183 	}
2184 
2185 	ext4_unlock_group(ac->ac_sb, group);
2186 	ext4_mb_unload_buddy(e4b);
2187 
2188 	return 0;
2189 }
2190 
2191 static noinline_for_stack
2192 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2193 				struct ext4_buddy *e4b)
2194 {
2195 	ext4_group_t group = ac->ac_g_ex.fe_group;
2196 	int max;
2197 	int err;
2198 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2199 	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2200 	struct ext4_free_extent ex;
2201 
2202 	if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
2203 		return 0;
2204 	if (grp->bb_free == 0)
2205 		return 0;
2206 
2207 	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2208 	if (err)
2209 		return err;
2210 
2211 	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
2212 		ext4_mb_unload_buddy(e4b);
2213 		return 0;
2214 	}
2215 
2216 	ext4_lock_group(ac->ac_sb, group);
2217 	max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2218 			     ac->ac_g_ex.fe_len, &ex);
2219 	ex.fe_logical = 0xDEADFA11; /* debug value */
2220 
2221 	if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
2222 		ext4_fsblk_t start;
2223 
2224 		start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
2225 			ex.fe_start;
2226 		/* use do_div to get remainder (would be 64-bit modulo) */
2227 		if (do_div(start, sbi->s_stripe) == 0) {
2228 			ac->ac_found++;
2229 			ac->ac_b_ex = ex;
2230 			ext4_mb_use_best_found(ac, e4b);
2231 		}
2232 	} else if (max >= ac->ac_g_ex.fe_len) {
2233 		BUG_ON(ex.fe_len <= 0);
2234 		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2235 		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2236 		ac->ac_found++;
2237 		ac->ac_b_ex = ex;
2238 		ext4_mb_use_best_found(ac, e4b);
2239 	} else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2240 		/* Sometimes, caller may want to merge even small
2241 		 * number of blocks to an existing extent */
2242 		BUG_ON(ex.fe_len <= 0);
2243 		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2244 		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2245 		ac->ac_found++;
2246 		ac->ac_b_ex = ex;
2247 		ext4_mb_use_best_found(ac, e4b);
2248 	}
2249 	ext4_unlock_group(ac->ac_sb, group);
2250 	ext4_mb_unload_buddy(e4b);
2251 
2252 	return 0;
2253 }
2254 
2255 /*
2256  * The routine scans buddy structures (not bitmap!) from given order
2257  * to max order and tries to find big enough chunk to satisfy the req
2258  */
2259 static noinline_for_stack
2260 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2261 					struct ext4_buddy *e4b)
2262 {
2263 	struct super_block *sb = ac->ac_sb;
2264 	struct ext4_group_info *grp = e4b->bd_info;
2265 	void *buddy;
2266 	int i;
2267 	int k;
2268 	int max;
2269 
2270 	BUG_ON(ac->ac_2order <= 0);
2271 	for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2272 		if (grp->bb_counters[i] == 0)
2273 			continue;
2274 
2275 		buddy = mb_find_buddy(e4b, i, &max);
2276 		BUG_ON(buddy == NULL);
2277 
2278 		k = mb_find_next_zero_bit(buddy, max, 0);
2279 		if (k >= max) {
2280 			ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2281 				"%d free clusters of order %d. But found 0",
2282 				grp->bb_counters[i], i);
2283 			ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2284 					 e4b->bd_group,
2285 					EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2286 			break;
2287 		}
2288 		ac->ac_found++;
2289 
2290 		ac->ac_b_ex.fe_len = 1 << i;
2291 		ac->ac_b_ex.fe_start = k << i;
2292 		ac->ac_b_ex.fe_group = e4b->bd_group;
2293 
2294 		ext4_mb_use_best_found(ac, e4b);
2295 
2296 		BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2297 
2298 		if (EXT4_SB(sb)->s_mb_stats)
2299 			atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2300 
2301 		break;
2302 	}
2303 }
2304 
2305 /*
2306  * The routine scans the group and measures all found extents.
2307  * In order to optimize scanning, caller must pass number of
2308  * free blocks in the group, so the routine can know upper limit.
2309  */
2310 static noinline_for_stack
2311 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2312 					struct ext4_buddy *e4b)
2313 {
2314 	struct super_block *sb = ac->ac_sb;
2315 	void *bitmap = e4b->bd_bitmap;
2316 	struct ext4_free_extent ex;
2317 	int i;
2318 	int free;
2319 
2320 	free = e4b->bd_info->bb_free;
2321 	if (WARN_ON(free <= 0))
2322 		return;
2323 
2324 	i = e4b->bd_info->bb_first_free;
2325 
2326 	while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2327 		i = mb_find_next_zero_bit(bitmap,
2328 						EXT4_CLUSTERS_PER_GROUP(sb), i);
2329 		if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2330 			/*
2331 			 * IF we have corrupt bitmap, we won't find any
2332 			 * free blocks even though group info says we
2333 			 * have free blocks
2334 			 */
2335 			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2336 					"%d free clusters as per "
2337 					"group info. But bitmap says 0",
2338 					free);
2339 			ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2340 					EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2341 			break;
2342 		}
2343 
2344 		mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2345 		if (WARN_ON(ex.fe_len <= 0))
2346 			break;
2347 		if (free < ex.fe_len) {
2348 			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2349 					"%d free clusters as per "
2350 					"group info. But got %d blocks",
2351 					free, ex.fe_len);
2352 			ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2353 					EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2354 			/*
2355 			 * The number of free blocks differs. This mostly
2356 			 * indicate that the bitmap is corrupt. So exit
2357 			 * without claiming the space.
2358 			 */
2359 			break;
2360 		}
2361 		ex.fe_logical = 0xDEADC0DE; /* debug value */
2362 		ext4_mb_measure_extent(ac, &ex, e4b);
2363 
2364 		i += ex.fe_len;
2365 		free -= ex.fe_len;
2366 	}
2367 
2368 	ext4_mb_check_limits(ac, e4b, 1);
2369 }
2370 
2371 /*
2372  * This is a special case for storages like raid5
2373  * we try to find stripe-aligned chunks for stripe-size-multiple requests
2374  */
2375 static noinline_for_stack
2376 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2377 				 struct ext4_buddy *e4b)
2378 {
2379 	struct super_block *sb = ac->ac_sb;
2380 	struct ext4_sb_info *sbi = EXT4_SB(sb);
2381 	void *bitmap = e4b->bd_bitmap;
2382 	struct ext4_free_extent ex;
2383 	ext4_fsblk_t first_group_block;
2384 	ext4_fsblk_t a;
2385 	ext4_grpblk_t i;
2386 	int max;
2387 
2388 	BUG_ON(sbi->s_stripe == 0);
2389 
2390 	/* find first stripe-aligned block in group */
2391 	first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2392 
2393 	a = first_group_block + sbi->s_stripe - 1;
2394 	do_div(a, sbi->s_stripe);
2395 	i = (a * sbi->s_stripe) - first_group_block;
2396 
2397 	while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2398 		if (!mb_test_bit(i, bitmap)) {
2399 			max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2400 			if (max >= sbi->s_stripe) {
2401 				ac->ac_found++;
2402 				ex.fe_logical = 0xDEADF00D; /* debug value */
2403 				ac->ac_b_ex = ex;
2404 				ext4_mb_use_best_found(ac, e4b);
2405 				break;
2406 			}
2407 		}
2408 		i += sbi->s_stripe;
2409 	}
2410 }
2411 
2412 /*
2413  * This is also called BEFORE we load the buddy bitmap.
2414  * Returns either 1 or 0 indicating that the group is either suitable
2415  * for the allocation or not.
2416  */
2417 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2418 				ext4_group_t group, int cr)
2419 {
2420 	ext4_grpblk_t free, fragments;
2421 	int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2422 	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2423 
2424 	BUG_ON(cr < 0 || cr >= 4);
2425 
2426 	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2427 		return false;
2428 
2429 	free = grp->bb_free;
2430 	if (free == 0)
2431 		return false;
2432 
2433 	fragments = grp->bb_fragments;
2434 	if (fragments == 0)
2435 		return false;
2436 
2437 	switch (cr) {
2438 	case 0:
2439 		BUG_ON(ac->ac_2order == 0);
2440 
2441 		/* Avoid using the first bg of a flexgroup for data files */
2442 		if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2443 		    (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2444 		    ((group % flex_size) == 0))
2445 			return false;
2446 
2447 		if (free < ac->ac_g_ex.fe_len)
2448 			return false;
2449 
2450 		if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2451 			return true;
2452 
2453 		if (grp->bb_largest_free_order < ac->ac_2order)
2454 			return false;
2455 
2456 		return true;
2457 	case 1:
2458 		if ((free / fragments) >= ac->ac_g_ex.fe_len)
2459 			return true;
2460 		break;
2461 	case 2:
2462 		if (free >= ac->ac_g_ex.fe_len)
2463 			return true;
2464 		break;
2465 	case 3:
2466 		return true;
2467 	default:
2468 		BUG();
2469 	}
2470 
2471 	return false;
2472 }
2473 
2474 /*
2475  * This could return negative error code if something goes wrong
2476  * during ext4_mb_init_group(). This should not be called with
2477  * ext4_lock_group() held.
2478  *
2479  * Note: because we are conditionally operating with the group lock in
2480  * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2481  * function using __acquire and __release.  This means we need to be
2482  * super careful before messing with the error path handling via "goto
2483  * out"!
2484  */
2485 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2486 				     ext4_group_t group, int cr)
2487 {
2488 	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2489 	struct super_block *sb = ac->ac_sb;
2490 	struct ext4_sb_info *sbi = EXT4_SB(sb);
2491 	bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2492 	ext4_grpblk_t free;
2493 	int ret = 0;
2494 
2495 	if (sbi->s_mb_stats)
2496 		atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2497 	if (should_lock) {
2498 		ext4_lock_group(sb, group);
2499 		__release(ext4_group_lock_ptr(sb, group));
2500 	}
2501 	free = grp->bb_free;
2502 	if (free == 0)
2503 		goto out;
2504 	if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2505 		goto out;
2506 	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2507 		goto out;
2508 	if (should_lock) {
2509 		__acquire(ext4_group_lock_ptr(sb, group));
2510 		ext4_unlock_group(sb, group);
2511 	}
2512 
2513 	/* We only do this if the grp has never been initialized */
2514 	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2515 		struct ext4_group_desc *gdp =
2516 			ext4_get_group_desc(sb, group, NULL);
2517 		int ret;
2518 
2519 		/* cr=0/1 is a very optimistic search to find large
2520 		 * good chunks almost for free.  If buddy data is not
2521 		 * ready, then this optimization makes no sense.  But
2522 		 * we never skip the first block group in a flex_bg,
2523 		 * since this gets used for metadata block allocation,
2524 		 * and we want to make sure we locate metadata blocks
2525 		 * in the first block group in the flex_bg if possible.
2526 		 */
2527 		if (cr < 2 &&
2528 		    (!sbi->s_log_groups_per_flex ||
2529 		     ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2530 		    !(ext4_has_group_desc_csum(sb) &&
2531 		      (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2532 			return 0;
2533 		ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2534 		if (ret)
2535 			return ret;
2536 	}
2537 
2538 	if (should_lock) {
2539 		ext4_lock_group(sb, group);
2540 		__release(ext4_group_lock_ptr(sb, group));
2541 	}
2542 	ret = ext4_mb_good_group(ac, group, cr);
2543 out:
2544 	if (should_lock) {
2545 		__acquire(ext4_group_lock_ptr(sb, group));
2546 		ext4_unlock_group(sb, group);
2547 	}
2548 	return ret;
2549 }
2550 
2551 /*
2552  * Start prefetching @nr block bitmaps starting at @group.
2553  * Return the next group which needs to be prefetched.
2554  */
2555 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2556 			      unsigned int nr, int *cnt)
2557 {
2558 	ext4_group_t ngroups = ext4_get_groups_count(sb);
2559 	struct buffer_head *bh;
2560 	struct blk_plug plug;
2561 
2562 	blk_start_plug(&plug);
2563 	while (nr-- > 0) {
2564 		struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2565 								  NULL);
2566 		struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2567 
2568 		/*
2569 		 * Prefetch block groups with free blocks; but don't
2570 		 * bother if it is marked uninitialized on disk, since
2571 		 * it won't require I/O to read.  Also only try to
2572 		 * prefetch once, so we avoid getblk() call, which can
2573 		 * be expensive.
2574 		 */
2575 		if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2576 		    EXT4_MB_GRP_NEED_INIT(grp) &&
2577 		    ext4_free_group_clusters(sb, gdp) > 0 &&
2578 		    !(ext4_has_group_desc_csum(sb) &&
2579 		      (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2580 			bh = ext4_read_block_bitmap_nowait(sb, group, true);
2581 			if (bh && !IS_ERR(bh)) {
2582 				if (!buffer_uptodate(bh) && cnt)
2583 					(*cnt)++;
2584 				brelse(bh);
2585 			}
2586 		}
2587 		if (++group >= ngroups)
2588 			group = 0;
2589 	}
2590 	blk_finish_plug(&plug);
2591 	return group;
2592 }
2593 
2594 /*
2595  * Prefetching reads the block bitmap into the buffer cache; but we
2596  * need to make sure that the buddy bitmap in the page cache has been
2597  * initialized.  Note that ext4_mb_init_group() will block if the I/O
2598  * is not yet completed, or indeed if it was not initiated by
2599  * ext4_mb_prefetch did not start the I/O.
2600  *
2601  * TODO: We should actually kick off the buddy bitmap setup in a work
2602  * queue when the buffer I/O is completed, so that we don't block
2603  * waiting for the block allocation bitmap read to finish when
2604  * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2605  */
2606 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2607 			   unsigned int nr)
2608 {
2609 	while (nr-- > 0) {
2610 		struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2611 								  NULL);
2612 		struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2613 
2614 		if (!group)
2615 			group = ext4_get_groups_count(sb);
2616 		group--;
2617 		grp = ext4_get_group_info(sb, group);
2618 
2619 		if (EXT4_MB_GRP_NEED_INIT(grp) &&
2620 		    ext4_free_group_clusters(sb, gdp) > 0 &&
2621 		    !(ext4_has_group_desc_csum(sb) &&
2622 		      (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2623 			if (ext4_mb_init_group(sb, group, GFP_NOFS))
2624 				break;
2625 		}
2626 	}
2627 }
2628 
2629 static noinline_for_stack int
2630 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2631 {
2632 	ext4_group_t prefetch_grp = 0, ngroups, group, i;
2633 	int cr = -1;
2634 	int err = 0, first_err = 0;
2635 	unsigned int nr = 0, prefetch_ios = 0;
2636 	struct ext4_sb_info *sbi;
2637 	struct super_block *sb;
2638 	struct ext4_buddy e4b;
2639 	int lost;
2640 
2641 	sb = ac->ac_sb;
2642 	sbi = EXT4_SB(sb);
2643 	ngroups = ext4_get_groups_count(sb);
2644 	/* non-extent files are limited to low blocks/groups */
2645 	if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2646 		ngroups = sbi->s_blockfile_groups;
2647 
2648 	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2649 
2650 	/* first, try the goal */
2651 	err = ext4_mb_find_by_goal(ac, &e4b);
2652 	if (err || ac->ac_status == AC_STATUS_FOUND)
2653 		goto out;
2654 
2655 	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2656 		goto out;
2657 
2658 	/*
2659 	 * ac->ac_2order is set only if the fe_len is a power of 2
2660 	 * if ac->ac_2order is set we also set criteria to 0 so that we
2661 	 * try exact allocation using buddy.
2662 	 */
2663 	i = fls(ac->ac_g_ex.fe_len);
2664 	ac->ac_2order = 0;
2665 	/*
2666 	 * We search using buddy data only if the order of the request
2667 	 * is greater than equal to the sbi_s_mb_order2_reqs
2668 	 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2669 	 * We also support searching for power-of-two requests only for
2670 	 * requests upto maximum buddy size we have constructed.
2671 	 */
2672 	if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2673 		/*
2674 		 * This should tell if fe_len is exactly power of 2
2675 		 */
2676 		if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2677 			ac->ac_2order = array_index_nospec(i - 1,
2678 							   MB_NUM_ORDERS(sb));
2679 	}
2680 
2681 	/* if stream allocation is enabled, use global goal */
2682 	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2683 		/* TBD: may be hot point */
2684 		spin_lock(&sbi->s_md_lock);
2685 		ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2686 		ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2687 		spin_unlock(&sbi->s_md_lock);
2688 	}
2689 
2690 	/* Let's just scan groups to find more-less suitable blocks */
2691 	cr = ac->ac_2order ? 0 : 1;
2692 	/*
2693 	 * cr == 0 try to get exact allocation,
2694 	 * cr == 3  try to get anything
2695 	 */
2696 repeat:
2697 	for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2698 		ac->ac_criteria = cr;
2699 		/*
2700 		 * searching for the right group start
2701 		 * from the goal value specified
2702 		 */
2703 		group = ac->ac_g_ex.fe_group;
2704 		ac->ac_last_optimal_group = group;
2705 		ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2706 		prefetch_grp = group;
2707 
2708 		for (i = 0; i < ngroups; group = next_linear_group(ac, group, ngroups),
2709 			     i++) {
2710 			int ret = 0, new_cr;
2711 
2712 			cond_resched();
2713 
2714 			ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups);
2715 			if (new_cr != cr) {
2716 				cr = new_cr;
2717 				goto repeat;
2718 			}
2719 
2720 			/*
2721 			 * Batch reads of the block allocation bitmaps
2722 			 * to get multiple READs in flight; limit
2723 			 * prefetching at cr=0/1, otherwise mballoc can
2724 			 * spend a lot of time loading imperfect groups
2725 			 */
2726 			if ((prefetch_grp == group) &&
2727 			    (cr > 1 ||
2728 			     prefetch_ios < sbi->s_mb_prefetch_limit)) {
2729 				unsigned int curr_ios = prefetch_ios;
2730 
2731 				nr = sbi->s_mb_prefetch;
2732 				if (ext4_has_feature_flex_bg(sb)) {
2733 					nr = 1 << sbi->s_log_groups_per_flex;
2734 					nr -= group & (nr - 1);
2735 					nr = min(nr, sbi->s_mb_prefetch);
2736 				}
2737 				prefetch_grp = ext4_mb_prefetch(sb, group,
2738 							nr, &prefetch_ios);
2739 				if (prefetch_ios == curr_ios)
2740 					nr = 0;
2741 			}
2742 
2743 			/* This now checks without needing the buddy page */
2744 			ret = ext4_mb_good_group_nolock(ac, group, cr);
2745 			if (ret <= 0) {
2746 				if (!first_err)
2747 					first_err = ret;
2748 				continue;
2749 			}
2750 
2751 			err = ext4_mb_load_buddy(sb, group, &e4b);
2752 			if (err)
2753 				goto out;
2754 
2755 			ext4_lock_group(sb, group);
2756 
2757 			/*
2758 			 * We need to check again after locking the
2759 			 * block group
2760 			 */
2761 			ret = ext4_mb_good_group(ac, group, cr);
2762 			if (ret == 0) {
2763 				ext4_unlock_group(sb, group);
2764 				ext4_mb_unload_buddy(&e4b);
2765 				continue;
2766 			}
2767 
2768 			ac->ac_groups_scanned++;
2769 			if (cr == 0)
2770 				ext4_mb_simple_scan_group(ac, &e4b);
2771 			else if (cr == 1 && sbi->s_stripe &&
2772 					!(ac->ac_g_ex.fe_len % sbi->s_stripe))
2773 				ext4_mb_scan_aligned(ac, &e4b);
2774 			else
2775 				ext4_mb_complex_scan_group(ac, &e4b);
2776 
2777 			ext4_unlock_group(sb, group);
2778 			ext4_mb_unload_buddy(&e4b);
2779 
2780 			if (ac->ac_status != AC_STATUS_CONTINUE)
2781 				break;
2782 		}
2783 		/* Processed all groups and haven't found blocks */
2784 		if (sbi->s_mb_stats && i == ngroups)
2785 			atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2786 	}
2787 
2788 	if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2789 	    !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2790 		/*
2791 		 * We've been searching too long. Let's try to allocate
2792 		 * the best chunk we've found so far
2793 		 */
2794 		ext4_mb_try_best_found(ac, &e4b);
2795 		if (ac->ac_status != AC_STATUS_FOUND) {
2796 			/*
2797 			 * Someone more lucky has already allocated it.
2798 			 * The only thing we can do is just take first
2799 			 * found block(s)
2800 			 */
2801 			lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2802 			mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2803 				 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2804 				 ac->ac_b_ex.fe_len, lost);
2805 
2806 			ac->ac_b_ex.fe_group = 0;
2807 			ac->ac_b_ex.fe_start = 0;
2808 			ac->ac_b_ex.fe_len = 0;
2809 			ac->ac_status = AC_STATUS_CONTINUE;
2810 			ac->ac_flags |= EXT4_MB_HINT_FIRST;
2811 			cr = 3;
2812 			goto repeat;
2813 		}
2814 	}
2815 
2816 	if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2817 		atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2818 out:
2819 	if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2820 		err = first_err;
2821 
2822 	mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2823 		 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2824 		 ac->ac_flags, cr, err);
2825 
2826 	if (nr)
2827 		ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2828 
2829 	return err;
2830 }
2831 
2832 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2833 {
2834 	struct super_block *sb = pde_data(file_inode(seq->file));
2835 	ext4_group_t group;
2836 
2837 	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2838 		return NULL;
2839 	group = *pos + 1;
2840 	return (void *) ((unsigned long) group);
2841 }
2842 
2843 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2844 {
2845 	struct super_block *sb = pde_data(file_inode(seq->file));
2846 	ext4_group_t group;
2847 
2848 	++*pos;
2849 	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2850 		return NULL;
2851 	group = *pos + 1;
2852 	return (void *) ((unsigned long) group);
2853 }
2854 
2855 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2856 {
2857 	struct super_block *sb = pde_data(file_inode(seq->file));
2858 	ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2859 	int i;
2860 	int err, buddy_loaded = 0;
2861 	struct ext4_buddy e4b;
2862 	struct ext4_group_info *grinfo;
2863 	unsigned char blocksize_bits = min_t(unsigned char,
2864 					     sb->s_blocksize_bits,
2865 					     EXT4_MAX_BLOCK_LOG_SIZE);
2866 	struct sg {
2867 		struct ext4_group_info info;
2868 		ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2869 	} sg;
2870 
2871 	group--;
2872 	if (group == 0)
2873 		seq_puts(seq, "#group: free  frags first ["
2874 			      " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
2875 			      " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
2876 
2877 	i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2878 		sizeof(struct ext4_group_info);
2879 
2880 	grinfo = ext4_get_group_info(sb, group);
2881 	/* Load the group info in memory only if not already loaded. */
2882 	if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2883 		err = ext4_mb_load_buddy(sb, group, &e4b);
2884 		if (err) {
2885 			seq_printf(seq, "#%-5u: I/O error\n", group);
2886 			return 0;
2887 		}
2888 		buddy_loaded = 1;
2889 	}
2890 
2891 	memcpy(&sg, ext4_get_group_info(sb, group), i);
2892 
2893 	if (buddy_loaded)
2894 		ext4_mb_unload_buddy(&e4b);
2895 
2896 	seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2897 			sg.info.bb_fragments, sg.info.bb_first_free);
2898 	for (i = 0; i <= 13; i++)
2899 		seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2900 				sg.info.bb_counters[i] : 0);
2901 	seq_puts(seq, " ]\n");
2902 
2903 	return 0;
2904 }
2905 
2906 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2907 {
2908 }
2909 
2910 const struct seq_operations ext4_mb_seq_groups_ops = {
2911 	.start  = ext4_mb_seq_groups_start,
2912 	.next   = ext4_mb_seq_groups_next,
2913 	.stop   = ext4_mb_seq_groups_stop,
2914 	.show   = ext4_mb_seq_groups_show,
2915 };
2916 
2917 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
2918 {
2919 	struct super_block *sb = seq->private;
2920 	struct ext4_sb_info *sbi = EXT4_SB(sb);
2921 
2922 	seq_puts(seq, "mballoc:\n");
2923 	if (!sbi->s_mb_stats) {
2924 		seq_puts(seq, "\tmb stats collection turned off.\n");
2925 		seq_puts(seq, "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
2926 		return 0;
2927 	}
2928 	seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
2929 	seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
2930 
2931 	seq_printf(seq, "\tgroups_scanned: %u\n",  atomic_read(&sbi->s_bal_groups_scanned));
2932 
2933 	seq_puts(seq, "\tcr0_stats:\n");
2934 	seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[0]));
2935 	seq_printf(seq, "\t\tgroups_considered: %llu\n",
2936 		   atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
2937 	seq_printf(seq, "\t\tuseless_loops: %llu\n",
2938 		   atomic64_read(&sbi->s_bal_cX_failed[0]));
2939 	seq_printf(seq, "\t\tbad_suggestions: %u\n",
2940 		   atomic_read(&sbi->s_bal_cr0_bad_suggestions));
2941 
2942 	seq_puts(seq, "\tcr1_stats:\n");
2943 	seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[1]));
2944 	seq_printf(seq, "\t\tgroups_considered: %llu\n",
2945 		   atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
2946 	seq_printf(seq, "\t\tuseless_loops: %llu\n",
2947 		   atomic64_read(&sbi->s_bal_cX_failed[1]));
2948 	seq_printf(seq, "\t\tbad_suggestions: %u\n",
2949 		   atomic_read(&sbi->s_bal_cr1_bad_suggestions));
2950 
2951 	seq_puts(seq, "\tcr2_stats:\n");
2952 	seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[2]));
2953 	seq_printf(seq, "\t\tgroups_considered: %llu\n",
2954 		   atomic64_read(&sbi->s_bal_cX_groups_considered[2]));
2955 	seq_printf(seq, "\t\tuseless_loops: %llu\n",
2956 		   atomic64_read(&sbi->s_bal_cX_failed[2]));
2957 
2958 	seq_puts(seq, "\tcr3_stats:\n");
2959 	seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_hits[3]));
2960 	seq_printf(seq, "\t\tgroups_considered: %llu\n",
2961 		   atomic64_read(&sbi->s_bal_cX_groups_considered[3]));
2962 	seq_printf(seq, "\t\tuseless_loops: %llu\n",
2963 		   atomic64_read(&sbi->s_bal_cX_failed[3]));
2964 	seq_printf(seq, "\textents_scanned: %u\n", atomic_read(&sbi->s_bal_ex_scanned));
2965 	seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
2966 	seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
2967 	seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
2968 	seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
2969 
2970 	seq_printf(seq, "\tbuddies_generated: %u/%u\n",
2971 		   atomic_read(&sbi->s_mb_buddies_generated),
2972 		   ext4_get_groups_count(sb));
2973 	seq_printf(seq, "\tbuddies_time_used: %llu\n",
2974 		   atomic64_read(&sbi->s_mb_generation_time));
2975 	seq_printf(seq, "\tpreallocated: %u\n",
2976 		   atomic_read(&sbi->s_mb_preallocated));
2977 	seq_printf(seq, "\tdiscarded: %u\n",
2978 		   atomic_read(&sbi->s_mb_discarded));
2979 	return 0;
2980 }
2981 
2982 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
2983 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
2984 {
2985 	struct super_block *sb = pde_data(file_inode(seq->file));
2986 	unsigned long position;
2987 
2988 	read_lock(&EXT4_SB(sb)->s_mb_rb_lock);
2989 
2990 	if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
2991 		return NULL;
2992 	position = *pos + 1;
2993 	return (void *) ((unsigned long) position);
2994 }
2995 
2996 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
2997 {
2998 	struct super_block *sb = pde_data(file_inode(seq->file));
2999 	unsigned long position;
3000 
3001 	++*pos;
3002 	if (*pos < 0 || *pos >= MB_NUM_ORDERS(sb) + 1)
3003 		return NULL;
3004 	position = *pos + 1;
3005 	return (void *) ((unsigned long) position);
3006 }
3007 
3008 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3009 {
3010 	struct super_block *sb = pde_data(file_inode(seq->file));
3011 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3012 	unsigned long position = ((unsigned long) v);
3013 	struct ext4_group_info *grp;
3014 	struct rb_node *n;
3015 	unsigned int count, min, max;
3016 
3017 	position--;
3018 	if (position >= MB_NUM_ORDERS(sb)) {
3019 		seq_puts(seq, "fragment_size_tree:\n");
3020 		n = rb_first(&sbi->s_mb_avg_fragment_size_root);
3021 		if (!n) {
3022 			seq_puts(seq, "\ttree_min: 0\n\ttree_max: 0\n\ttree_nodes: 0\n");
3023 			return 0;
3024 		}
3025 		grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3026 		min = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3027 		count = 1;
3028 		while (rb_next(n)) {
3029 			count++;
3030 			n = rb_next(n);
3031 		}
3032 		grp = rb_entry(n, struct ext4_group_info, bb_avg_fragment_size_rb);
3033 		max = grp->bb_fragments ? grp->bb_free / grp->bb_fragments : 0;
3034 
3035 		seq_printf(seq, "\ttree_min: %u\n\ttree_max: %u\n\ttree_nodes: %u\n",
3036 			   min, max, count);
3037 		return 0;
3038 	}
3039 
3040 	if (position == 0) {
3041 		seq_printf(seq, "optimize_scan: %d\n",
3042 			   test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3043 		seq_puts(seq, "max_free_order_lists:\n");
3044 	}
3045 	count = 0;
3046 	list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3047 			    bb_largest_free_order_node)
3048 		count++;
3049 	seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3050 		   (unsigned int)position, count);
3051 
3052 	return 0;
3053 }
3054 
3055 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3056 __releases(&EXT4_SB(sb)->s_mb_rb_lock)
3057 {
3058 	struct super_block *sb = pde_data(file_inode(seq->file));
3059 
3060 	read_unlock(&EXT4_SB(sb)->s_mb_rb_lock);
3061 }
3062 
3063 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3064 	.start  = ext4_mb_seq_structs_summary_start,
3065 	.next   = ext4_mb_seq_structs_summary_next,
3066 	.stop   = ext4_mb_seq_structs_summary_stop,
3067 	.show   = ext4_mb_seq_structs_summary_show,
3068 };
3069 
3070 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3071 {
3072 	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3073 	struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3074 
3075 	BUG_ON(!cachep);
3076 	return cachep;
3077 }
3078 
3079 /*
3080  * Allocate the top-level s_group_info array for the specified number
3081  * of groups
3082  */
3083 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3084 {
3085 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3086 	unsigned size;
3087 	struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3088 
3089 	size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3090 		EXT4_DESC_PER_BLOCK_BITS(sb);
3091 	if (size <= sbi->s_group_info_size)
3092 		return 0;
3093 
3094 	size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3095 	new_groupinfo = kvzalloc(size, GFP_KERNEL);
3096 	if (!new_groupinfo) {
3097 		ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3098 		return -ENOMEM;
3099 	}
3100 	rcu_read_lock();
3101 	old_groupinfo = rcu_dereference(sbi->s_group_info);
3102 	if (old_groupinfo)
3103 		memcpy(new_groupinfo, old_groupinfo,
3104 		       sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3105 	rcu_read_unlock();
3106 	rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3107 	sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3108 	if (old_groupinfo)
3109 		ext4_kvfree_array_rcu(old_groupinfo);
3110 	ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3111 		   sbi->s_group_info_size);
3112 	return 0;
3113 }
3114 
3115 /* Create and initialize ext4_group_info data for the given group. */
3116 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3117 			  struct ext4_group_desc *desc)
3118 {
3119 	int i;
3120 	int metalen = 0;
3121 	int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3122 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3123 	struct ext4_group_info **meta_group_info;
3124 	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3125 
3126 	/*
3127 	 * First check if this group is the first of a reserved block.
3128 	 * If it's true, we have to allocate a new table of pointers
3129 	 * to ext4_group_info structures
3130 	 */
3131 	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3132 		metalen = sizeof(*meta_group_info) <<
3133 			EXT4_DESC_PER_BLOCK_BITS(sb);
3134 		meta_group_info = kmalloc(metalen, GFP_NOFS);
3135 		if (meta_group_info == NULL) {
3136 			ext4_msg(sb, KERN_ERR, "can't allocate mem "
3137 				 "for a buddy group");
3138 			goto exit_meta_group_info;
3139 		}
3140 		rcu_read_lock();
3141 		rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3142 		rcu_read_unlock();
3143 	}
3144 
3145 	meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3146 	i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3147 
3148 	meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3149 	if (meta_group_info[i] == NULL) {
3150 		ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3151 		goto exit_group_info;
3152 	}
3153 	set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3154 		&(meta_group_info[i]->bb_state));
3155 
3156 	/*
3157 	 * initialize bb_free to be able to skip
3158 	 * empty groups without initialization
3159 	 */
3160 	if (ext4_has_group_desc_csum(sb) &&
3161 	    (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3162 		meta_group_info[i]->bb_free =
3163 			ext4_free_clusters_after_init(sb, group, desc);
3164 	} else {
3165 		meta_group_info[i]->bb_free =
3166 			ext4_free_group_clusters(sb, desc);
3167 	}
3168 
3169 	INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3170 	init_rwsem(&meta_group_info[i]->alloc_sem);
3171 	meta_group_info[i]->bb_free_root = RB_ROOT;
3172 	INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3173 	RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
3174 	meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
3175 	meta_group_info[i]->bb_group = group;
3176 
3177 	mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3178 	return 0;
3179 
3180 exit_group_info:
3181 	/* If a meta_group_info table has been allocated, release it now */
3182 	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3183 		struct ext4_group_info ***group_info;
3184 
3185 		rcu_read_lock();
3186 		group_info = rcu_dereference(sbi->s_group_info);
3187 		kfree(group_info[idx]);
3188 		group_info[idx] = NULL;
3189 		rcu_read_unlock();
3190 	}
3191 exit_meta_group_info:
3192 	return -ENOMEM;
3193 } /* ext4_mb_add_groupinfo */
3194 
3195 static int ext4_mb_init_backend(struct super_block *sb)
3196 {
3197 	ext4_group_t ngroups = ext4_get_groups_count(sb);
3198 	ext4_group_t i;
3199 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3200 	int err;
3201 	struct ext4_group_desc *desc;
3202 	struct ext4_group_info ***group_info;
3203 	struct kmem_cache *cachep;
3204 
3205 	err = ext4_mb_alloc_groupinfo(sb, ngroups);
3206 	if (err)
3207 		return err;
3208 
3209 	sbi->s_buddy_cache = new_inode(sb);
3210 	if (sbi->s_buddy_cache == NULL) {
3211 		ext4_msg(sb, KERN_ERR, "can't get new inode");
3212 		goto err_freesgi;
3213 	}
3214 	/* To avoid potentially colliding with an valid on-disk inode number,
3215 	 * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
3216 	 * not in the inode hash, so it should never be found by iget(), but
3217 	 * this will avoid confusion if it ever shows up during debugging. */
3218 	sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3219 	EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3220 	for (i = 0; i < ngroups; i++) {
3221 		cond_resched();
3222 		desc = ext4_get_group_desc(sb, i, NULL);
3223 		if (desc == NULL) {
3224 			ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3225 			goto err_freebuddy;
3226 		}
3227 		if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3228 			goto err_freebuddy;
3229 	}
3230 
3231 	if (ext4_has_feature_flex_bg(sb)) {
3232 		/* a single flex group is supposed to be read by a single IO.
3233 		 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3234 		 * unsigned integer, so the maximum shift is 32.
3235 		 */
3236 		if (sbi->s_es->s_log_groups_per_flex >= 32) {
3237 			ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3238 			goto err_freebuddy;
3239 		}
3240 		sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3241 			BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3242 		sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3243 	} else {
3244 		sbi->s_mb_prefetch = 32;
3245 	}
3246 	if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3247 		sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3248 	/* now many real IOs to prefetch within a single allocation at cr=0
3249 	 * given cr=0 is an CPU-related optimization we shouldn't try to
3250 	 * load too many groups, at some point we should start to use what
3251 	 * we've got in memory.
3252 	 * with an average random access time 5ms, it'd take a second to get
3253 	 * 200 groups (* N with flex_bg), so let's make this limit 4
3254 	 */
3255 	sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3256 	if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3257 		sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3258 
3259 	return 0;
3260 
3261 err_freebuddy:
3262 	cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3263 	while (i-- > 0)
3264 		kmem_cache_free(cachep, ext4_get_group_info(sb, i));
3265 	i = sbi->s_group_info_size;
3266 	rcu_read_lock();
3267 	group_info = rcu_dereference(sbi->s_group_info);
3268 	while (i-- > 0)
3269 		kfree(group_info[i]);
3270 	rcu_read_unlock();
3271 	iput(sbi->s_buddy_cache);
3272 err_freesgi:
3273 	rcu_read_lock();
3274 	kvfree(rcu_dereference(sbi->s_group_info));
3275 	rcu_read_unlock();
3276 	return -ENOMEM;
3277 }
3278 
3279 static void ext4_groupinfo_destroy_slabs(void)
3280 {
3281 	int i;
3282 
3283 	for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3284 		kmem_cache_destroy(ext4_groupinfo_caches[i]);
3285 		ext4_groupinfo_caches[i] = NULL;
3286 	}
3287 }
3288 
3289 static int ext4_groupinfo_create_slab(size_t size)
3290 {
3291 	static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3292 	int slab_size;
3293 	int blocksize_bits = order_base_2(size);
3294 	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3295 	struct kmem_cache *cachep;
3296 
3297 	if (cache_index >= NR_GRPINFO_CACHES)
3298 		return -EINVAL;
3299 
3300 	if (unlikely(cache_index < 0))
3301 		cache_index = 0;
3302 
3303 	mutex_lock(&ext4_grpinfo_slab_create_mutex);
3304 	if (ext4_groupinfo_caches[cache_index]) {
3305 		mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3306 		return 0;	/* Already created */
3307 	}
3308 
3309 	slab_size = offsetof(struct ext4_group_info,
3310 				bb_counters[blocksize_bits + 2]);
3311 
3312 	cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3313 					slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3314 					NULL);
3315 
3316 	ext4_groupinfo_caches[cache_index] = cachep;
3317 
3318 	mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3319 	if (!cachep) {
3320 		printk(KERN_EMERG
3321 		       "EXT4-fs: no memory for groupinfo slab cache\n");
3322 		return -ENOMEM;
3323 	}
3324 
3325 	return 0;
3326 }
3327 
3328 static void ext4_discard_work(struct work_struct *work)
3329 {
3330 	struct ext4_sb_info *sbi = container_of(work,
3331 			struct ext4_sb_info, s_discard_work);
3332 	struct super_block *sb = sbi->s_sb;
3333 	struct ext4_free_data *fd, *nfd;
3334 	struct ext4_buddy e4b;
3335 	struct list_head discard_list;
3336 	ext4_group_t grp, load_grp;
3337 	int err = 0;
3338 
3339 	INIT_LIST_HEAD(&discard_list);
3340 	spin_lock(&sbi->s_md_lock);
3341 	list_splice_init(&sbi->s_discard_list, &discard_list);
3342 	spin_unlock(&sbi->s_md_lock);
3343 
3344 	load_grp = UINT_MAX;
3345 	list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3346 		/*
3347 		 * If filesystem is umounting or no memory or suffering
3348 		 * from no space, give up the discard
3349 		 */
3350 		if ((sb->s_flags & SB_ACTIVE) && !err &&
3351 		    !atomic_read(&sbi->s_retry_alloc_pending)) {
3352 			grp = fd->efd_group;
3353 			if (grp != load_grp) {
3354 				if (load_grp != UINT_MAX)
3355 					ext4_mb_unload_buddy(&e4b);
3356 
3357 				err = ext4_mb_load_buddy(sb, grp, &e4b);
3358 				if (err) {
3359 					kmem_cache_free(ext4_free_data_cachep, fd);
3360 					load_grp = UINT_MAX;
3361 					continue;
3362 				} else {
3363 					load_grp = grp;
3364 				}
3365 			}
3366 
3367 			ext4_lock_group(sb, grp);
3368 			ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3369 						fd->efd_start_cluster + fd->efd_count - 1, 1);
3370 			ext4_unlock_group(sb, grp);
3371 		}
3372 		kmem_cache_free(ext4_free_data_cachep, fd);
3373 	}
3374 
3375 	if (load_grp != UINT_MAX)
3376 		ext4_mb_unload_buddy(&e4b);
3377 }
3378 
3379 int ext4_mb_init(struct super_block *sb)
3380 {
3381 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3382 	unsigned i, j;
3383 	unsigned offset, offset_incr;
3384 	unsigned max;
3385 	int ret;
3386 
3387 	i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3388 
3389 	sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3390 	if (sbi->s_mb_offsets == NULL) {
3391 		ret = -ENOMEM;
3392 		goto out;
3393 	}
3394 
3395 	i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3396 	sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3397 	if (sbi->s_mb_maxs == NULL) {
3398 		ret = -ENOMEM;
3399 		goto out;
3400 	}
3401 
3402 	ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3403 	if (ret < 0)
3404 		goto out;
3405 
3406 	/* order 0 is regular bitmap */
3407 	sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3408 	sbi->s_mb_offsets[0] = 0;
3409 
3410 	i = 1;
3411 	offset = 0;
3412 	offset_incr = 1 << (sb->s_blocksize_bits - 1);
3413 	max = sb->s_blocksize << 2;
3414 	do {
3415 		sbi->s_mb_offsets[i] = offset;
3416 		sbi->s_mb_maxs[i] = max;
3417 		offset += offset_incr;
3418 		offset_incr = offset_incr >> 1;
3419 		max = max >> 1;
3420 		i++;
3421 	} while (i < MB_NUM_ORDERS(sb));
3422 
3423 	sbi->s_mb_avg_fragment_size_root = RB_ROOT;
3424 	sbi->s_mb_largest_free_orders =
3425 		kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3426 			GFP_KERNEL);
3427 	if (!sbi->s_mb_largest_free_orders) {
3428 		ret = -ENOMEM;
3429 		goto out;
3430 	}
3431 	sbi->s_mb_largest_free_orders_locks =
3432 		kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3433 			GFP_KERNEL);
3434 	if (!sbi->s_mb_largest_free_orders_locks) {
3435 		ret = -ENOMEM;
3436 		goto out;
3437 	}
3438 	for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3439 		INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3440 		rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3441 	}
3442 	rwlock_init(&sbi->s_mb_rb_lock);
3443 
3444 	spin_lock_init(&sbi->s_md_lock);
3445 	sbi->s_mb_free_pending = 0;
3446 	INIT_LIST_HEAD(&sbi->s_freed_data_list);
3447 	INIT_LIST_HEAD(&sbi->s_discard_list);
3448 	INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3449 	atomic_set(&sbi->s_retry_alloc_pending, 0);
3450 
3451 	sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3452 	sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3453 	sbi->s_mb_stats = MB_DEFAULT_STATS;
3454 	sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3455 	sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3456 	sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
3457 	/*
3458 	 * The default group preallocation is 512, which for 4k block
3459 	 * sizes translates to 2 megabytes.  However for bigalloc file
3460 	 * systems, this is probably too big (i.e, if the cluster size
3461 	 * is 1 megabyte, then group preallocation size becomes half a
3462 	 * gigabyte!).  As a default, we will keep a two megabyte
3463 	 * group pralloc size for cluster sizes up to 64k, and after
3464 	 * that, we will force a minimum group preallocation size of
3465 	 * 32 clusters.  This translates to 8 megs when the cluster
3466 	 * size is 256k, and 32 megs when the cluster size is 1 meg,
3467 	 * which seems reasonable as a default.
3468 	 */
3469 	sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3470 				       sbi->s_cluster_bits, 32);
3471 	/*
3472 	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3473 	 * to the lowest multiple of s_stripe which is bigger than
3474 	 * the s_mb_group_prealloc as determined above. We want
3475 	 * the preallocation size to be an exact multiple of the
3476 	 * RAID stripe size so that preallocations don't fragment
3477 	 * the stripes.
3478 	 */
3479 	if (sbi->s_stripe > 1) {
3480 		sbi->s_mb_group_prealloc = roundup(
3481 			sbi->s_mb_group_prealloc, sbi->s_stripe);
3482 	}
3483 
3484 	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3485 	if (sbi->s_locality_groups == NULL) {
3486 		ret = -ENOMEM;
3487 		goto out;
3488 	}
3489 	for_each_possible_cpu(i) {
3490 		struct ext4_locality_group *lg;
3491 		lg = per_cpu_ptr(sbi->s_locality_groups, i);
3492 		mutex_init(&lg->lg_mutex);
3493 		for (j = 0; j < PREALLOC_TB_SIZE; j++)
3494 			INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3495 		spin_lock_init(&lg->lg_prealloc_lock);
3496 	}
3497 
3498 	if (bdev_nonrot(sb->s_bdev))
3499 		sbi->s_mb_max_linear_groups = 0;
3500 	else
3501 		sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3502 	/* init file for buddy data */
3503 	ret = ext4_mb_init_backend(sb);
3504 	if (ret != 0)
3505 		goto out_free_locality_groups;
3506 
3507 	return 0;
3508 
3509 out_free_locality_groups:
3510 	free_percpu(sbi->s_locality_groups);
3511 	sbi->s_locality_groups = NULL;
3512 out:
3513 	kfree(sbi->s_mb_largest_free_orders);
3514 	kfree(sbi->s_mb_largest_free_orders_locks);
3515 	kfree(sbi->s_mb_offsets);
3516 	sbi->s_mb_offsets = NULL;
3517 	kfree(sbi->s_mb_maxs);
3518 	sbi->s_mb_maxs = NULL;
3519 	return ret;
3520 }
3521 
3522 /* need to called with the ext4 group lock held */
3523 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3524 {
3525 	struct ext4_prealloc_space *pa;
3526 	struct list_head *cur, *tmp;
3527 	int count = 0;
3528 
3529 	list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3530 		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3531 		list_del(&pa->pa_group_list);
3532 		count++;
3533 		kmem_cache_free(ext4_pspace_cachep, pa);
3534 	}
3535 	return count;
3536 }
3537 
3538 int ext4_mb_release(struct super_block *sb)
3539 {
3540 	ext4_group_t ngroups = ext4_get_groups_count(sb);
3541 	ext4_group_t i;
3542 	int num_meta_group_infos;
3543 	struct ext4_group_info *grinfo, ***group_info;
3544 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3545 	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3546 	int count;
3547 
3548 	if (test_opt(sb, DISCARD)) {
3549 		/*
3550 		 * wait the discard work to drain all of ext4_free_data
3551 		 */
3552 		flush_work(&sbi->s_discard_work);
3553 		WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3554 	}
3555 
3556 	if (sbi->s_group_info) {
3557 		for (i = 0; i < ngroups; i++) {
3558 			cond_resched();
3559 			grinfo = ext4_get_group_info(sb, i);
3560 			mb_group_bb_bitmap_free(grinfo);
3561 			ext4_lock_group(sb, i);
3562 			count = ext4_mb_cleanup_pa(grinfo);
3563 			if (count)
3564 				mb_debug(sb, "mballoc: %d PAs left\n",
3565 					 count);
3566 			ext4_unlock_group(sb, i);
3567 			kmem_cache_free(cachep, grinfo);
3568 		}
3569 		num_meta_group_infos = (ngroups +
3570 				EXT4_DESC_PER_BLOCK(sb) - 1) >>
3571 			EXT4_DESC_PER_BLOCK_BITS(sb);
3572 		rcu_read_lock();
3573 		group_info = rcu_dereference(sbi->s_group_info);
3574 		for (i = 0; i < num_meta_group_infos; i++)
3575 			kfree(group_info[i]);
3576 		kvfree(group_info);
3577 		rcu_read_unlock();
3578 	}
3579 	kfree(sbi->s_mb_largest_free_orders);
3580 	kfree(sbi->s_mb_largest_free_orders_locks);
3581 	kfree(sbi->s_mb_offsets);
3582 	kfree(sbi->s_mb_maxs);
3583 	iput(sbi->s_buddy_cache);
3584 	if (sbi->s_mb_stats) {
3585 		ext4_msg(sb, KERN_INFO,
3586 		       "mballoc: %u blocks %u reqs (%u success)",
3587 				atomic_read(&sbi->s_bal_allocated),
3588 				atomic_read(&sbi->s_bal_reqs),
3589 				atomic_read(&sbi->s_bal_success));
3590 		ext4_msg(sb, KERN_INFO,
3591 		      "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3592 				"%u 2^N hits, %u breaks, %u lost",
3593 				atomic_read(&sbi->s_bal_ex_scanned),
3594 				atomic_read(&sbi->s_bal_groups_scanned),
3595 				atomic_read(&sbi->s_bal_goals),
3596 				atomic_read(&sbi->s_bal_2orders),
3597 				atomic_read(&sbi->s_bal_breaks),
3598 				atomic_read(&sbi->s_mb_lost_chunks));
3599 		ext4_msg(sb, KERN_INFO,
3600 		       "mballoc: %u generated and it took %llu",
3601 				atomic_read(&sbi->s_mb_buddies_generated),
3602 				atomic64_read(&sbi->s_mb_generation_time));
3603 		ext4_msg(sb, KERN_INFO,
3604 		       "mballoc: %u preallocated, %u discarded",
3605 				atomic_read(&sbi->s_mb_preallocated),
3606 				atomic_read(&sbi->s_mb_discarded));
3607 	}
3608 
3609 	free_percpu(sbi->s_locality_groups);
3610 
3611 	return 0;
3612 }
3613 
3614 static inline int ext4_issue_discard(struct super_block *sb,
3615 		ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3616 		struct bio **biop)
3617 {
3618 	ext4_fsblk_t discard_block;
3619 
3620 	discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3621 			 ext4_group_first_block_no(sb, block_group));
3622 	count = EXT4_C2B(EXT4_SB(sb), count);
3623 	trace_ext4_discard_blocks(sb,
3624 			(unsigned long long) discard_block, count);
3625 	if (biop) {
3626 		return __blkdev_issue_discard(sb->s_bdev,
3627 			(sector_t)discard_block << (sb->s_blocksize_bits - 9),
3628 			(sector_t)count << (sb->s_blocksize_bits - 9),
3629 			GFP_NOFS, biop);
3630 	} else
3631 		return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3632 }
3633 
3634 static void ext4_free_data_in_buddy(struct super_block *sb,
3635 				    struct ext4_free_data *entry)
3636 {
3637 	struct ext4_buddy e4b;
3638 	struct ext4_group_info *db;
3639 	int err, count = 0, count2 = 0;
3640 
3641 	mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3642 		 entry->efd_count, entry->efd_group, entry);
3643 
3644 	err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3645 	/* we expect to find existing buddy because it's pinned */
3646 	BUG_ON(err != 0);
3647 
3648 	spin_lock(&EXT4_SB(sb)->s_md_lock);
3649 	EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3650 	spin_unlock(&EXT4_SB(sb)->s_md_lock);
3651 
3652 	db = e4b.bd_info;
3653 	/* there are blocks to put in buddy to make them really free */
3654 	count += entry->efd_count;
3655 	count2++;
3656 	ext4_lock_group(sb, entry->efd_group);
3657 	/* Take it out of per group rb tree */
3658 	rb_erase(&entry->efd_node, &(db->bb_free_root));
3659 	mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3660 
3661 	/*
3662 	 * Clear the trimmed flag for the group so that the next
3663 	 * ext4_trim_fs can trim it.
3664 	 * If the volume is mounted with -o discard, online discard
3665 	 * is supported and the free blocks will be trimmed online.
3666 	 */
3667 	if (!test_opt(sb, DISCARD))
3668 		EXT4_MB_GRP_CLEAR_TRIMMED(db);
3669 
3670 	if (!db->bb_free_root.rb_node) {
3671 		/* No more items in the per group rb tree
3672 		 * balance refcounts from ext4_mb_free_metadata()
3673 		 */
3674 		put_page(e4b.bd_buddy_page);
3675 		put_page(e4b.bd_bitmap_page);
3676 	}
3677 	ext4_unlock_group(sb, entry->efd_group);
3678 	ext4_mb_unload_buddy(&e4b);
3679 
3680 	mb_debug(sb, "freed %d blocks in %d structures\n", count,
3681 		 count2);
3682 }
3683 
3684 /*
3685  * This function is called by the jbd2 layer once the commit has finished,
3686  * so we know we can free the blocks that were released with that commit.
3687  */
3688 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3689 {
3690 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3691 	struct ext4_free_data *entry, *tmp;
3692 	struct list_head freed_data_list;
3693 	struct list_head *cut_pos = NULL;
3694 	bool wake;
3695 
3696 	INIT_LIST_HEAD(&freed_data_list);
3697 
3698 	spin_lock(&sbi->s_md_lock);
3699 	list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3700 		if (entry->efd_tid != commit_tid)
3701 			break;
3702 		cut_pos = &entry->efd_list;
3703 	}
3704 	if (cut_pos)
3705 		list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3706 				  cut_pos);
3707 	spin_unlock(&sbi->s_md_lock);
3708 
3709 	list_for_each_entry(entry, &freed_data_list, efd_list)
3710 		ext4_free_data_in_buddy(sb, entry);
3711 
3712 	if (test_opt(sb, DISCARD)) {
3713 		spin_lock(&sbi->s_md_lock);
3714 		wake = list_empty(&sbi->s_discard_list);
3715 		list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3716 		spin_unlock(&sbi->s_md_lock);
3717 		if (wake)
3718 			queue_work(system_unbound_wq, &sbi->s_discard_work);
3719 	} else {
3720 		list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3721 			kmem_cache_free(ext4_free_data_cachep, entry);
3722 	}
3723 }
3724 
3725 int __init ext4_init_mballoc(void)
3726 {
3727 	ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3728 					SLAB_RECLAIM_ACCOUNT);
3729 	if (ext4_pspace_cachep == NULL)
3730 		goto out;
3731 
3732 	ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3733 				    SLAB_RECLAIM_ACCOUNT);
3734 	if (ext4_ac_cachep == NULL)
3735 		goto out_pa_free;
3736 
3737 	ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3738 					   SLAB_RECLAIM_ACCOUNT);
3739 	if (ext4_free_data_cachep == NULL)
3740 		goto out_ac_free;
3741 
3742 	return 0;
3743 
3744 out_ac_free:
3745 	kmem_cache_destroy(ext4_ac_cachep);
3746 out_pa_free:
3747 	kmem_cache_destroy(ext4_pspace_cachep);
3748 out:
3749 	return -ENOMEM;
3750 }
3751 
3752 void ext4_exit_mballoc(void)
3753 {
3754 	/*
3755 	 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3756 	 * before destroying the slab cache.
3757 	 */
3758 	rcu_barrier();
3759 	kmem_cache_destroy(ext4_pspace_cachep);
3760 	kmem_cache_destroy(ext4_ac_cachep);
3761 	kmem_cache_destroy(ext4_free_data_cachep);
3762 	ext4_groupinfo_destroy_slabs();
3763 }
3764 
3765 
3766 /*
3767  * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3768  * Returns 0 if success or error code
3769  */
3770 static noinline_for_stack int
3771 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3772 				handle_t *handle, unsigned int reserv_clstrs)
3773 {
3774 	struct buffer_head *bitmap_bh = NULL;
3775 	struct ext4_group_desc *gdp;
3776 	struct buffer_head *gdp_bh;
3777 	struct ext4_sb_info *sbi;
3778 	struct super_block *sb;
3779 	ext4_fsblk_t block;
3780 	int err, len;
3781 
3782 	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3783 	BUG_ON(ac->ac_b_ex.fe_len <= 0);
3784 
3785 	sb = ac->ac_sb;
3786 	sbi = EXT4_SB(sb);
3787 
3788 	bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3789 	if (IS_ERR(bitmap_bh)) {
3790 		err = PTR_ERR(bitmap_bh);
3791 		bitmap_bh = NULL;
3792 		goto out_err;
3793 	}
3794 
3795 	BUFFER_TRACE(bitmap_bh, "getting write access");
3796 	err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
3797 					    EXT4_JTR_NONE);
3798 	if (err)
3799 		goto out_err;
3800 
3801 	err = -EIO;
3802 	gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3803 	if (!gdp)
3804 		goto out_err;
3805 
3806 	ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3807 			ext4_free_group_clusters(sb, gdp));
3808 
3809 	BUFFER_TRACE(gdp_bh, "get_write_access");
3810 	err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE);
3811 	if (err)
3812 		goto out_err;
3813 
3814 	block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3815 
3816 	len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3817 	if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3818 		ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3819 			   "fs metadata", block, block+len);
3820 		/* File system mounted not to panic on error
3821 		 * Fix the bitmap and return EFSCORRUPTED
3822 		 * We leak some of the blocks here.
3823 		 */
3824 		ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3825 		mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3826 			      ac->ac_b_ex.fe_len);
3827 		ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3828 		err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3829 		if (!err)
3830 			err = -EFSCORRUPTED;
3831 		goto out_err;
3832 	}
3833 
3834 	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3835 #ifdef AGGRESSIVE_CHECK
3836 	{
3837 		int i;
3838 		for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3839 			BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3840 						bitmap_bh->b_data));
3841 		}
3842 	}
3843 #endif
3844 	mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3845 		      ac->ac_b_ex.fe_len);
3846 	if (ext4_has_group_desc_csum(sb) &&
3847 	    (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3848 		gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3849 		ext4_free_group_clusters_set(sb, gdp,
3850 					     ext4_free_clusters_after_init(sb,
3851 						ac->ac_b_ex.fe_group, gdp));
3852 	}
3853 	len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3854 	ext4_free_group_clusters_set(sb, gdp, len);
3855 	ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3856 	ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3857 
3858 	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3859 	percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3860 	/*
3861 	 * Now reduce the dirty block count also. Should not go negative
3862 	 */
3863 	if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3864 		/* release all the reserved blocks if non delalloc */
3865 		percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3866 				   reserv_clstrs);
3867 
3868 	if (sbi->s_log_groups_per_flex) {
3869 		ext4_group_t flex_group = ext4_flex_group(sbi,
3870 							  ac->ac_b_ex.fe_group);
3871 		atomic64_sub(ac->ac_b_ex.fe_len,
3872 			     &sbi_array_rcu_deref(sbi, s_flex_groups,
3873 						  flex_group)->free_clusters);
3874 	}
3875 
3876 	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3877 	if (err)
3878 		goto out_err;
3879 	err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3880 
3881 out_err:
3882 	brelse(bitmap_bh);
3883 	return err;
3884 }
3885 
3886 /*
3887  * Idempotent helper for Ext4 fast commit replay path to set the state of
3888  * blocks in bitmaps and update counters.
3889  */
3890 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3891 			int len, int state)
3892 {
3893 	struct buffer_head *bitmap_bh = NULL;
3894 	struct ext4_group_desc *gdp;
3895 	struct buffer_head *gdp_bh;
3896 	struct ext4_sb_info *sbi = EXT4_SB(sb);
3897 	ext4_group_t group;
3898 	ext4_grpblk_t blkoff;
3899 	int i, err;
3900 	int already;
3901 	unsigned int clen, clen_changed, thisgrp_len;
3902 
3903 	while (len > 0) {
3904 		ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3905 
3906 		/*
3907 		 * Check to see if we are freeing blocks across a group
3908 		 * boundary.
3909 		 * In case of flex_bg, this can happen that (block, len) may
3910 		 * span across more than one group. In that case we need to
3911 		 * get the corresponding group metadata to work with.
3912 		 * For this we have goto again loop.
3913 		 */
3914 		thisgrp_len = min_t(unsigned int, (unsigned int)len,
3915 			EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
3916 		clen = EXT4_NUM_B2C(sbi, thisgrp_len);
3917 
3918 		if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
3919 			ext4_error(sb, "Marking blocks in system zone - "
3920 				   "Block = %llu, len = %u",
3921 				   block, thisgrp_len);
3922 			bitmap_bh = NULL;
3923 			break;
3924 		}
3925 
3926 		bitmap_bh = ext4_read_block_bitmap(sb, group);
3927 		if (IS_ERR(bitmap_bh)) {
3928 			err = PTR_ERR(bitmap_bh);
3929 			bitmap_bh = NULL;
3930 			break;
3931 		}
3932 
3933 		err = -EIO;
3934 		gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3935 		if (!gdp)
3936 			break;
3937 
3938 		ext4_lock_group(sb, group);
3939 		already = 0;
3940 		for (i = 0; i < clen; i++)
3941 			if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
3942 					 !state)
3943 				already++;
3944 
3945 		clen_changed = clen - already;
3946 		if (state)
3947 			mb_set_bits(bitmap_bh->b_data, blkoff, clen);
3948 		else
3949 			mb_clear_bits(bitmap_bh->b_data, blkoff, clen);
3950 		if (ext4_has_group_desc_csum(sb) &&
3951 		    (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3952 			gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3953 			ext4_free_group_clusters_set(sb, gdp,
3954 			     ext4_free_clusters_after_init(sb, group, gdp));
3955 		}
3956 		if (state)
3957 			clen = ext4_free_group_clusters(sb, gdp) - clen_changed;
3958 		else
3959 			clen = ext4_free_group_clusters(sb, gdp) + clen_changed;
3960 
3961 		ext4_free_group_clusters_set(sb, gdp, clen);
3962 		ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3963 		ext4_group_desc_csum_set(sb, group, gdp);
3964 
3965 		ext4_unlock_group(sb, group);
3966 
3967 		if (sbi->s_log_groups_per_flex) {
3968 			ext4_group_t flex_group = ext4_flex_group(sbi, group);
3969 			struct flex_groups *fg = sbi_array_rcu_deref(sbi,
3970 						   s_flex_groups, flex_group);
3971 
3972 			if (state)
3973 				atomic64_sub(clen_changed, &fg->free_clusters);
3974 			else
3975 				atomic64_add(clen_changed, &fg->free_clusters);
3976 
3977 		}
3978 
3979 		err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3980 		if (err)
3981 			break;
3982 		sync_dirty_buffer(bitmap_bh);
3983 		err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3984 		sync_dirty_buffer(gdp_bh);
3985 		if (err)
3986 			break;
3987 
3988 		block += thisgrp_len;
3989 		len -= thisgrp_len;
3990 		brelse(bitmap_bh);
3991 		BUG_ON(len < 0);
3992 	}
3993 
3994 	if (err)
3995 		brelse(bitmap_bh);
3996 }
3997 
3998 /*
3999  * here we normalize request for locality group
4000  * Group request are normalized to s_mb_group_prealloc, which goes to
4001  * s_strip if we set the same via mount option.
4002  * s_mb_group_prealloc can be configured via
4003  * /sys/fs/ext4/<partition>/mb_group_prealloc
4004  *
4005  * XXX: should we try to preallocate more than the group has now?
4006  */
4007 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4008 {
4009 	struct super_block *sb = ac->ac_sb;
4010 	struct ext4_locality_group *lg = ac->ac_lg;
4011 
4012 	BUG_ON(lg == NULL);
4013 	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4014 	mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4015 }
4016 
4017 /*
4018  * Normalization means making request better in terms of
4019  * size and alignment
4020  */
4021 static noinline_for_stack void
4022 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4023 				struct ext4_allocation_request *ar)
4024 {
4025 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4026 	int bsbits, max;
4027 	ext4_lblk_t end;
4028 	loff_t size, start_off;
4029 	loff_t orig_size __maybe_unused;
4030 	ext4_lblk_t start;
4031 	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4032 	struct ext4_prealloc_space *pa;
4033 
4034 	/* do normalize only data requests, metadata requests
4035 	   do not need preallocation */
4036 	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4037 		return;
4038 
4039 	/* sometime caller may want exact blocks */
4040 	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4041 		return;
4042 
4043 	/* caller may indicate that preallocation isn't
4044 	 * required (it's a tail, for example) */
4045 	if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4046 		return;
4047 
4048 	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4049 		ext4_mb_normalize_group_request(ac);
4050 		return ;
4051 	}
4052 
4053 	bsbits = ac->ac_sb->s_blocksize_bits;
4054 
4055 	/* first, let's learn actual file size
4056 	 * given current request is allocated */
4057 	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4058 	size = size << bsbits;
4059 	if (size < i_size_read(ac->ac_inode))
4060 		size = i_size_read(ac->ac_inode);
4061 	orig_size = size;
4062 
4063 	/* max size of free chunks */
4064 	max = 2 << bsbits;
4065 
4066 #define NRL_CHECK_SIZE(req, size, max, chunk_size)	\
4067 		(req <= (size) || max <= (chunk_size))
4068 
4069 	/* first, try to predict filesize */
4070 	/* XXX: should this table be tunable? */
4071 	start_off = 0;
4072 	if (size <= 16 * 1024) {
4073 		size = 16 * 1024;
4074 	} else if (size <= 32 * 1024) {
4075 		size = 32 * 1024;
4076 	} else if (size <= 64 * 1024) {
4077 		size = 64 * 1024;
4078 	} else if (size <= 128 * 1024) {
4079 		size = 128 * 1024;
4080 	} else if (size <= 256 * 1024) {
4081 		size = 256 * 1024;
4082 	} else if (size <= 512 * 1024) {
4083 		size = 512 * 1024;
4084 	} else if (size <= 1024 * 1024) {
4085 		size = 1024 * 1024;
4086 	} else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4087 		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4088 						(21 - bsbits)) << 21;
4089 		size = 2 * 1024 * 1024;
4090 	} else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4091 		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4092 							(22 - bsbits)) << 22;
4093 		size = 4 * 1024 * 1024;
4094 	} else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
4095 					(8<<20)>>bsbits, max, 8 * 1024)) {
4096 		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4097 							(23 - bsbits)) << 23;
4098 		size = 8 * 1024 * 1024;
4099 	} else {
4100 		start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4101 		size	  = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
4102 					      ac->ac_o_ex.fe_len) << bsbits;
4103 	}
4104 	size = size >> bsbits;
4105 	start = start_off >> bsbits;
4106 
4107 	/*
4108 	 * For tiny groups (smaller than 8MB) the chosen allocation
4109 	 * alignment may be larger than group size. Make sure the
4110 	 * alignment does not move allocation to a different group which
4111 	 * makes mballoc fail assertions later.
4112 	 */
4113 	start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4114 			(ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4115 
4116 	/* don't cover already allocated blocks in selected range */
4117 	if (ar->pleft && start <= ar->lleft) {
4118 		size -= ar->lleft + 1 - start;
4119 		start = ar->lleft + 1;
4120 	}
4121 	if (ar->pright && start + size - 1 >= ar->lright)
4122 		size -= start + size - ar->lright;
4123 
4124 	/*
4125 	 * Trim allocation request for filesystems with artificially small
4126 	 * groups.
4127 	 */
4128 	if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4129 		size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4130 
4131 	end = start + size;
4132 
4133 	/* check we don't cross already preallocated blocks */
4134 	rcu_read_lock();
4135 	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4136 		ext4_lblk_t pa_end;
4137 
4138 		if (pa->pa_deleted)
4139 			continue;
4140 		spin_lock(&pa->pa_lock);
4141 		if (pa->pa_deleted) {
4142 			spin_unlock(&pa->pa_lock);
4143 			continue;
4144 		}
4145 
4146 		pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4147 						  pa->pa_len);
4148 
4149 		/* PA must not overlap original request */
4150 		BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
4151 			ac->ac_o_ex.fe_logical < pa->pa_lstart));
4152 
4153 		/* skip PAs this normalized request doesn't overlap with */
4154 		if (pa->pa_lstart >= end || pa_end <= start) {
4155 			spin_unlock(&pa->pa_lock);
4156 			continue;
4157 		}
4158 		BUG_ON(pa->pa_lstart <= start && pa_end >= end);
4159 
4160 		/* adjust start or end to be adjacent to this pa */
4161 		if (pa_end <= ac->ac_o_ex.fe_logical) {
4162 			BUG_ON(pa_end < start);
4163 			start = pa_end;
4164 		} else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4165 			BUG_ON(pa->pa_lstart > end);
4166 			end = pa->pa_lstart;
4167 		}
4168 		spin_unlock(&pa->pa_lock);
4169 	}
4170 	rcu_read_unlock();
4171 	size = end - start;
4172 
4173 	/* XXX: extra loop to check we really don't overlap preallocations */
4174 	rcu_read_lock();
4175 	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4176 		ext4_lblk_t pa_end;
4177 
4178 		spin_lock(&pa->pa_lock);
4179 		if (pa->pa_deleted == 0) {
4180 			pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
4181 							  pa->pa_len);
4182 			BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
4183 		}
4184 		spin_unlock(&pa->pa_lock);
4185 	}
4186 	rcu_read_unlock();
4187 
4188 	/*
4189 	 * In this function "start" and "size" are normalized for better
4190 	 * alignment and length such that we could preallocate more blocks.
4191 	 * This normalization is done such that original request of
4192 	 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4193 	 * "size" boundaries.
4194 	 * (Note fe_len can be relaxed since FS block allocation API does not
4195 	 * provide gurantee on number of contiguous blocks allocation since that
4196 	 * depends upon free space left, etc).
4197 	 * In case of inode pa, later we use the allocated blocks
4198 	 * [pa_start + fe_logical - pa_lstart, fe_len/size] from the preallocated
4199 	 * range of goal/best blocks [start, size] to put it at the
4200 	 * ac_o_ex.fe_logical extent of this inode.
4201 	 * (See ext4_mb_use_inode_pa() for more details)
4202 	 */
4203 	if (start + size <= ac->ac_o_ex.fe_logical ||
4204 			start > ac->ac_o_ex.fe_logical) {
4205 		ext4_msg(ac->ac_sb, KERN_ERR,
4206 			 "start %lu, size %lu, fe_logical %lu",
4207 			 (unsigned long) start, (unsigned long) size,
4208 			 (unsigned long) ac->ac_o_ex.fe_logical);
4209 		BUG();
4210 	}
4211 	BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4212 
4213 	/* now prepare goal request */
4214 
4215 	/* XXX: is it better to align blocks WRT to logical
4216 	 * placement or satisfy big request as is */
4217 	ac->ac_g_ex.fe_logical = start;
4218 	ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4219 
4220 	/* define goal start in order to merge */
4221 	if (ar->pright && (ar->lright == (start + size))) {
4222 		/* merge to the right */
4223 		ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4224 						&ac->ac_f_ex.fe_group,
4225 						&ac->ac_f_ex.fe_start);
4226 		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4227 	}
4228 	if (ar->pleft && (ar->lleft + 1 == start)) {
4229 		/* merge to the left */
4230 		ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4231 						&ac->ac_f_ex.fe_group,
4232 						&ac->ac_f_ex.fe_start);
4233 		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4234 	}
4235 
4236 	mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4237 		 orig_size, start);
4238 }
4239 
4240 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4241 {
4242 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4243 
4244 	if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4245 		atomic_inc(&sbi->s_bal_reqs);
4246 		atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4247 		if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4248 			atomic_inc(&sbi->s_bal_success);
4249 		atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4250 		atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4251 		if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4252 				ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4253 			atomic_inc(&sbi->s_bal_goals);
4254 		if (ac->ac_found > sbi->s_mb_max_to_scan)
4255 			atomic_inc(&sbi->s_bal_breaks);
4256 	}
4257 
4258 	if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4259 		trace_ext4_mballoc_alloc(ac);
4260 	else
4261 		trace_ext4_mballoc_prealloc(ac);
4262 }
4263 
4264 /*
4265  * Called on failure; free up any blocks from the inode PA for this
4266  * context.  We don't need this for MB_GROUP_PA because we only change
4267  * pa_free in ext4_mb_release_context(), but on failure, we've already
4268  * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4269  */
4270 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4271 {
4272 	struct ext4_prealloc_space *pa = ac->ac_pa;
4273 	struct ext4_buddy e4b;
4274 	int err;
4275 
4276 	if (pa == NULL) {
4277 		if (ac->ac_f_ex.fe_len == 0)
4278 			return;
4279 		err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4280 		if (err) {
4281 			/*
4282 			 * This should never happen since we pin the
4283 			 * pages in the ext4_allocation_context so
4284 			 * ext4_mb_load_buddy() should never fail.
4285 			 */
4286 			WARN(1, "mb_load_buddy failed (%d)", err);
4287 			return;
4288 		}
4289 		ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4290 		mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4291 			       ac->ac_f_ex.fe_len);
4292 		ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4293 		ext4_mb_unload_buddy(&e4b);
4294 		return;
4295 	}
4296 	if (pa->pa_type == MB_INODE_PA)
4297 		pa->pa_free += ac->ac_b_ex.fe_len;
4298 }
4299 
4300 /*
4301  * use blocks preallocated to inode
4302  */
4303 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4304 				struct ext4_prealloc_space *pa)
4305 {
4306 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4307 	ext4_fsblk_t start;
4308 	ext4_fsblk_t end;
4309 	int len;
4310 
4311 	/* found preallocated blocks, use them */
4312 	start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4313 	end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4314 		  start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4315 	len = EXT4_NUM_B2C(sbi, end - start);
4316 	ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4317 					&ac->ac_b_ex.fe_start);
4318 	ac->ac_b_ex.fe_len = len;
4319 	ac->ac_status = AC_STATUS_FOUND;
4320 	ac->ac_pa = pa;
4321 
4322 	BUG_ON(start < pa->pa_pstart);
4323 	BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4324 	BUG_ON(pa->pa_free < len);
4325 	pa->pa_free -= len;
4326 
4327 	mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4328 }
4329 
4330 /*
4331  * use blocks preallocated to locality group
4332  */
4333 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4334 				struct ext4_prealloc_space *pa)
4335 {
4336 	unsigned int len = ac->ac_o_ex.fe_len;
4337 
4338 	ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4339 					&ac->ac_b_ex.fe_group,
4340 					&ac->ac_b_ex.fe_start);
4341 	ac->ac_b_ex.fe_len = len;
4342 	ac->ac_status = AC_STATUS_FOUND;
4343 	ac->ac_pa = pa;
4344 
4345 	/* we don't correct pa_pstart or pa_plen here to avoid
4346 	 * possible race when the group is being loaded concurrently
4347 	 * instead we correct pa later, after blocks are marked
4348 	 * in on-disk bitmap -- see ext4_mb_release_context()
4349 	 * Other CPUs are prevented from allocating from this pa by lg_mutex
4350 	 */
4351 	mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4352 		 pa->pa_lstart-len, len, pa);
4353 }
4354 
4355 /*
4356  * Return the prealloc space that have minimal distance
4357  * from the goal block. @cpa is the prealloc
4358  * space that is having currently known minimal distance
4359  * from the goal block.
4360  */
4361 static struct ext4_prealloc_space *
4362 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4363 			struct ext4_prealloc_space *pa,
4364 			struct ext4_prealloc_space *cpa)
4365 {
4366 	ext4_fsblk_t cur_distance, new_distance;
4367 
4368 	if (cpa == NULL) {
4369 		atomic_inc(&pa->pa_count);
4370 		return pa;
4371 	}
4372 	cur_distance = abs(goal_block - cpa->pa_pstart);
4373 	new_distance = abs(goal_block - pa->pa_pstart);
4374 
4375 	if (cur_distance <= new_distance)
4376 		return cpa;
4377 
4378 	/* drop the previous reference */
4379 	atomic_dec(&cpa->pa_count);
4380 	atomic_inc(&pa->pa_count);
4381 	return pa;
4382 }
4383 
4384 /*
4385  * search goal blocks in preallocated space
4386  */
4387 static noinline_for_stack bool
4388 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4389 {
4390 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4391 	int order, i;
4392 	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4393 	struct ext4_locality_group *lg;
4394 	struct ext4_prealloc_space *pa, *cpa = NULL;
4395 	ext4_fsblk_t goal_block;
4396 
4397 	/* only data can be preallocated */
4398 	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4399 		return false;
4400 
4401 	/* first, try per-file preallocation */
4402 	rcu_read_lock();
4403 	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
4404 
4405 		/* all fields in this condition don't change,
4406 		 * so we can skip locking for them */
4407 		if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
4408 		    ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
4409 					       EXT4_C2B(sbi, pa->pa_len)))
4410 			continue;
4411 
4412 		/* non-extent files can't have physical blocks past 2^32 */
4413 		if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4414 		    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
4415 		     EXT4_MAX_BLOCK_FILE_PHYS))
4416 			continue;
4417 
4418 		/* found preallocated blocks, use them */
4419 		spin_lock(&pa->pa_lock);
4420 		if (pa->pa_deleted == 0 && pa->pa_free) {
4421 			atomic_inc(&pa->pa_count);
4422 			ext4_mb_use_inode_pa(ac, pa);
4423 			spin_unlock(&pa->pa_lock);
4424 			ac->ac_criteria = 10;
4425 			rcu_read_unlock();
4426 			return true;
4427 		}
4428 		spin_unlock(&pa->pa_lock);
4429 	}
4430 	rcu_read_unlock();
4431 
4432 	/* can we use group allocation? */
4433 	if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4434 		return false;
4435 
4436 	/* inode may have no locality group for some reason */
4437 	lg = ac->ac_lg;
4438 	if (lg == NULL)
4439 		return false;
4440 	order  = fls(ac->ac_o_ex.fe_len) - 1;
4441 	if (order > PREALLOC_TB_SIZE - 1)
4442 		/* The max size of hash table is PREALLOC_TB_SIZE */
4443 		order = PREALLOC_TB_SIZE - 1;
4444 
4445 	goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4446 	/*
4447 	 * search for the prealloc space that is having
4448 	 * minimal distance from the goal block.
4449 	 */
4450 	for (i = order; i < PREALLOC_TB_SIZE; i++) {
4451 		rcu_read_lock();
4452 		list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
4453 					pa_inode_list) {
4454 			spin_lock(&pa->pa_lock);
4455 			if (pa->pa_deleted == 0 &&
4456 					pa->pa_free >= ac->ac_o_ex.fe_len) {
4457 
4458 				cpa = ext4_mb_check_group_pa(goal_block,
4459 								pa, cpa);
4460 			}
4461 			spin_unlock(&pa->pa_lock);
4462 		}
4463 		rcu_read_unlock();
4464 	}
4465 	if (cpa) {
4466 		ext4_mb_use_group_pa(ac, cpa);
4467 		ac->ac_criteria = 20;
4468 		return true;
4469 	}
4470 	return false;
4471 }
4472 
4473 /*
4474  * the function goes through all block freed in the group
4475  * but not yet committed and marks them used in in-core bitmap.
4476  * buddy must be generated from this bitmap
4477  * Need to be called with the ext4 group lock held
4478  */
4479 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
4480 						ext4_group_t group)
4481 {
4482 	struct rb_node *n;
4483 	struct ext4_group_info *grp;
4484 	struct ext4_free_data *entry;
4485 
4486 	grp = ext4_get_group_info(sb, group);
4487 	n = rb_first(&(grp->bb_free_root));
4488 
4489 	while (n) {
4490 		entry = rb_entry(n, struct ext4_free_data, efd_node);
4491 		mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
4492 		n = rb_next(n);
4493 	}
4494 	return;
4495 }
4496 
4497 /*
4498  * the function goes through all preallocation in this group and marks them
4499  * used in in-core bitmap. buddy must be generated from this bitmap
4500  * Need to be called with ext4 group lock held
4501  */
4502 static noinline_for_stack
4503 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4504 					ext4_group_t group)
4505 {
4506 	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4507 	struct ext4_prealloc_space *pa;
4508 	struct list_head *cur;
4509 	ext4_group_t groupnr;
4510 	ext4_grpblk_t start;
4511 	int preallocated = 0;
4512 	int len;
4513 
4514 	/* all form of preallocation discards first load group,
4515 	 * so the only competing code is preallocation use.
4516 	 * we don't need any locking here
4517 	 * notice we do NOT ignore preallocations with pa_deleted
4518 	 * otherwise we could leave used blocks available for
4519 	 * allocation in buddy when concurrent ext4_mb_put_pa()
4520 	 * is dropping preallocation
4521 	 */
4522 	list_for_each(cur, &grp->bb_prealloc_list) {
4523 		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
4524 		spin_lock(&pa->pa_lock);
4525 		ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4526 					     &groupnr, &start);
4527 		len = pa->pa_len;
4528 		spin_unlock(&pa->pa_lock);
4529 		if (unlikely(len == 0))
4530 			continue;
4531 		BUG_ON(groupnr != group);
4532 		mb_set_bits(bitmap, start, len);
4533 		preallocated += len;
4534 	}
4535 	mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
4536 }
4537 
4538 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
4539 				    struct ext4_prealloc_space *pa)
4540 {
4541 	struct ext4_inode_info *ei;
4542 
4543 	if (pa->pa_deleted) {
4544 		ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
4545 			     pa->pa_type, pa->pa_pstart, pa->pa_lstart,
4546 			     pa->pa_len);
4547 		return;
4548 	}
4549 
4550 	pa->pa_deleted = 1;
4551 
4552 	if (pa->pa_type == MB_INODE_PA) {
4553 		ei = EXT4_I(pa->pa_inode);
4554 		atomic_dec(&ei->i_prealloc_active);
4555 	}
4556 }
4557 
4558 static void ext4_mb_pa_callback(struct rcu_head *head)
4559 {
4560 	struct ext4_prealloc_space *pa;
4561 	pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
4562 
4563 	BUG_ON(atomic_read(&pa->pa_count));
4564 	BUG_ON(pa->pa_deleted == 0);
4565 	kmem_cache_free(ext4_pspace_cachep, pa);
4566 }
4567 
4568 /*
4569  * drops a reference to preallocated space descriptor
4570  * if this was the last reference and the space is consumed
4571  */
4572 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
4573 			struct super_block *sb, struct ext4_prealloc_space *pa)
4574 {
4575 	ext4_group_t grp;
4576 	ext4_fsblk_t grp_blk;
4577 
4578 	/* in this short window concurrent discard can set pa_deleted */
4579 	spin_lock(&pa->pa_lock);
4580 	if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
4581 		spin_unlock(&pa->pa_lock);
4582 		return;
4583 	}
4584 
4585 	if (pa->pa_deleted == 1) {
4586 		spin_unlock(&pa->pa_lock);
4587 		return;
4588 	}
4589 
4590 	ext4_mb_mark_pa_deleted(sb, pa);
4591 	spin_unlock(&pa->pa_lock);
4592 
4593 	grp_blk = pa->pa_pstart;
4594 	/*
4595 	 * If doing group-based preallocation, pa_pstart may be in the
4596 	 * next group when pa is used up
4597 	 */
4598 	if (pa->pa_type == MB_GROUP_PA)
4599 		grp_blk--;
4600 
4601 	grp = ext4_get_group_number(sb, grp_blk);
4602 
4603 	/*
4604 	 * possible race:
4605 	 *
4606 	 *  P1 (buddy init)			P2 (regular allocation)
4607 	 *					find block B in PA
4608 	 *  copy on-disk bitmap to buddy
4609 	 *  					mark B in on-disk bitmap
4610 	 *					drop PA from group
4611 	 *  mark all PAs in buddy
4612 	 *
4613 	 * thus, P1 initializes buddy with B available. to prevent this
4614 	 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
4615 	 * against that pair
4616 	 */
4617 	ext4_lock_group(sb, grp);
4618 	list_del(&pa->pa_group_list);
4619 	ext4_unlock_group(sb, grp);
4620 
4621 	spin_lock(pa->pa_obj_lock);
4622 	list_del_rcu(&pa->pa_inode_list);
4623 	spin_unlock(pa->pa_obj_lock);
4624 
4625 	call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4626 }
4627 
4628 /*
4629  * creates new preallocated space for given inode
4630  */
4631 static noinline_for_stack void
4632 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
4633 {
4634 	struct super_block *sb = ac->ac_sb;
4635 	struct ext4_sb_info *sbi = EXT4_SB(sb);
4636 	struct ext4_prealloc_space *pa;
4637 	struct ext4_group_info *grp;
4638 	struct ext4_inode_info *ei;
4639 
4640 	/* preallocate only when found space is larger then requested */
4641 	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4642 	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4643 	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4644 	BUG_ON(ac->ac_pa == NULL);
4645 
4646 	pa = ac->ac_pa;
4647 
4648 	if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
4649 		int winl;
4650 		int wins;
4651 		int win;
4652 		int offs;
4653 
4654 		/* we can't allocate as much as normalizer wants.
4655 		 * so, found space must get proper lstart
4656 		 * to cover original request */
4657 		BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
4658 		BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
4659 
4660 		/* we're limited by original request in that
4661 		 * logical block must be covered any way
4662 		 * winl is window we can move our chunk within */
4663 		winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
4664 
4665 		/* also, we should cover whole original request */
4666 		wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4667 
4668 		/* the smallest one defines real window */
4669 		win = min(winl, wins);
4670 
4671 		offs = ac->ac_o_ex.fe_logical %
4672 			EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4673 		if (offs && offs < win)
4674 			win = offs;
4675 
4676 		ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4677 			EXT4_NUM_B2C(sbi, win);
4678 		BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4679 		BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4680 	}
4681 
4682 	/* preallocation can change ac_b_ex, thus we store actually
4683 	 * allocated blocks for history */
4684 	ac->ac_f_ex = ac->ac_b_ex;
4685 
4686 	pa->pa_lstart = ac->ac_b_ex.fe_logical;
4687 	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4688 	pa->pa_len = ac->ac_b_ex.fe_len;
4689 	pa->pa_free = pa->pa_len;
4690 	spin_lock_init(&pa->pa_lock);
4691 	INIT_LIST_HEAD(&pa->pa_inode_list);
4692 	INIT_LIST_HEAD(&pa->pa_group_list);
4693 	pa->pa_deleted = 0;
4694 	pa->pa_type = MB_INODE_PA;
4695 
4696 	mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4697 		 pa->pa_len, pa->pa_lstart);
4698 	trace_ext4_mb_new_inode_pa(ac, pa);
4699 
4700 	ext4_mb_use_inode_pa(ac, pa);
4701 	atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4702 
4703 	ei = EXT4_I(ac->ac_inode);
4704 	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4705 
4706 	pa->pa_obj_lock = &ei->i_prealloc_lock;
4707 	pa->pa_inode = ac->ac_inode;
4708 
4709 	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4710 
4711 	spin_lock(pa->pa_obj_lock);
4712 	list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4713 	spin_unlock(pa->pa_obj_lock);
4714 	atomic_inc(&ei->i_prealloc_active);
4715 }
4716 
4717 /*
4718  * creates new preallocated space for locality group inodes belongs to
4719  */
4720 static noinline_for_stack void
4721 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4722 {
4723 	struct super_block *sb = ac->ac_sb;
4724 	struct ext4_locality_group *lg;
4725 	struct ext4_prealloc_space *pa;
4726 	struct ext4_group_info *grp;
4727 
4728 	/* preallocate only when found space is larger then requested */
4729 	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4730 	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4731 	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4732 	BUG_ON(ac->ac_pa == NULL);
4733 
4734 	pa = ac->ac_pa;
4735 
4736 	/* preallocation can change ac_b_ex, thus we store actually
4737 	 * allocated blocks for history */
4738 	ac->ac_f_ex = ac->ac_b_ex;
4739 
4740 	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4741 	pa->pa_lstart = pa->pa_pstart;
4742 	pa->pa_len = ac->ac_b_ex.fe_len;
4743 	pa->pa_free = pa->pa_len;
4744 	spin_lock_init(&pa->pa_lock);
4745 	INIT_LIST_HEAD(&pa->pa_inode_list);
4746 	INIT_LIST_HEAD(&pa->pa_group_list);
4747 	pa->pa_deleted = 0;
4748 	pa->pa_type = MB_GROUP_PA;
4749 
4750 	mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4751 		 pa->pa_len, pa->pa_lstart);
4752 	trace_ext4_mb_new_group_pa(ac, pa);
4753 
4754 	ext4_mb_use_group_pa(ac, pa);
4755 	atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4756 
4757 	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4758 	lg = ac->ac_lg;
4759 	BUG_ON(lg == NULL);
4760 
4761 	pa->pa_obj_lock = &lg->lg_prealloc_lock;
4762 	pa->pa_inode = NULL;
4763 
4764 	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4765 
4766 	/*
4767 	 * We will later add the new pa to the right bucket
4768 	 * after updating the pa_free in ext4_mb_release_context
4769 	 */
4770 }
4771 
4772 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4773 {
4774 	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4775 		ext4_mb_new_group_pa(ac);
4776 	else
4777 		ext4_mb_new_inode_pa(ac);
4778 }
4779 
4780 /*
4781  * finds all unused blocks in on-disk bitmap, frees them in
4782  * in-core bitmap and buddy.
4783  * @pa must be unlinked from inode and group lists, so that
4784  * nobody else can find/use it.
4785  * the caller MUST hold group/inode locks.
4786  * TODO: optimize the case when there are no in-core structures yet
4787  */
4788 static noinline_for_stack int
4789 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4790 			struct ext4_prealloc_space *pa)
4791 {
4792 	struct super_block *sb = e4b->bd_sb;
4793 	struct ext4_sb_info *sbi = EXT4_SB(sb);
4794 	unsigned int end;
4795 	unsigned int next;
4796 	ext4_group_t group;
4797 	ext4_grpblk_t bit;
4798 	unsigned long long grp_blk_start;
4799 	int free = 0;
4800 
4801 	BUG_ON(pa->pa_deleted == 0);
4802 	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4803 	grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4804 	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4805 	end = bit + pa->pa_len;
4806 
4807 	while (bit < end) {
4808 		bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4809 		if (bit >= end)
4810 			break;
4811 		next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4812 		mb_debug(sb, "free preallocated %u/%u in group %u\n",
4813 			 (unsigned) ext4_group_first_block_no(sb, group) + bit,
4814 			 (unsigned) next - bit, (unsigned) group);
4815 		free += next - bit;
4816 
4817 		trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4818 		trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4819 						    EXT4_C2B(sbi, bit)),
4820 					       next - bit);
4821 		mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4822 		bit = next + 1;
4823 	}
4824 	if (free != pa->pa_free) {
4825 		ext4_msg(e4b->bd_sb, KERN_CRIT,
4826 			 "pa %p: logic %lu, phys. %lu, len %d",
4827 			 pa, (unsigned long) pa->pa_lstart,
4828 			 (unsigned long) pa->pa_pstart,
4829 			 pa->pa_len);
4830 		ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4831 					free, pa->pa_free);
4832 		/*
4833 		 * pa is already deleted so we use the value obtained
4834 		 * from the bitmap and continue.
4835 		 */
4836 	}
4837 	atomic_add(free, &sbi->s_mb_discarded);
4838 
4839 	return 0;
4840 }
4841 
4842 static noinline_for_stack int
4843 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4844 				struct ext4_prealloc_space *pa)
4845 {
4846 	struct super_block *sb = e4b->bd_sb;
4847 	ext4_group_t group;
4848 	ext4_grpblk_t bit;
4849 
4850 	trace_ext4_mb_release_group_pa(sb, pa);
4851 	BUG_ON(pa->pa_deleted == 0);
4852 	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4853 	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4854 	mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4855 	atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4856 	trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4857 
4858 	return 0;
4859 }
4860 
4861 /*
4862  * releases all preallocations in given group
4863  *
4864  * first, we need to decide discard policy:
4865  * - when do we discard
4866  *   1) ENOSPC
4867  * - how many do we discard
4868  *   1) how many requested
4869  */
4870 static noinline_for_stack int
4871 ext4_mb_discard_group_preallocations(struct super_block *sb,
4872 				     ext4_group_t group, int *busy)
4873 {
4874 	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4875 	struct buffer_head *bitmap_bh = NULL;
4876 	struct ext4_prealloc_space *pa, *tmp;
4877 	struct list_head list;
4878 	struct ext4_buddy e4b;
4879 	int err;
4880 	int free = 0;
4881 
4882 	mb_debug(sb, "discard preallocation for group %u\n", group);
4883 	if (list_empty(&grp->bb_prealloc_list))
4884 		goto out_dbg;
4885 
4886 	bitmap_bh = ext4_read_block_bitmap(sb, group);
4887 	if (IS_ERR(bitmap_bh)) {
4888 		err = PTR_ERR(bitmap_bh);
4889 		ext4_error_err(sb, -err,
4890 			       "Error %d reading block bitmap for %u",
4891 			       err, group);
4892 		goto out_dbg;
4893 	}
4894 
4895 	err = ext4_mb_load_buddy(sb, group, &e4b);
4896 	if (err) {
4897 		ext4_warning(sb, "Error %d loading buddy information for %u",
4898 			     err, group);
4899 		put_bh(bitmap_bh);
4900 		goto out_dbg;
4901 	}
4902 
4903 	INIT_LIST_HEAD(&list);
4904 	ext4_lock_group(sb, group);
4905 	list_for_each_entry_safe(pa, tmp,
4906 				&grp->bb_prealloc_list, pa_group_list) {
4907 		spin_lock(&pa->pa_lock);
4908 		if (atomic_read(&pa->pa_count)) {
4909 			spin_unlock(&pa->pa_lock);
4910 			*busy = 1;
4911 			continue;
4912 		}
4913 		if (pa->pa_deleted) {
4914 			spin_unlock(&pa->pa_lock);
4915 			continue;
4916 		}
4917 
4918 		/* seems this one can be freed ... */
4919 		ext4_mb_mark_pa_deleted(sb, pa);
4920 
4921 		if (!free)
4922 			this_cpu_inc(discard_pa_seq);
4923 
4924 		/* we can trust pa_free ... */
4925 		free += pa->pa_free;
4926 
4927 		spin_unlock(&pa->pa_lock);
4928 
4929 		list_del(&pa->pa_group_list);
4930 		list_add(&pa->u.pa_tmp_list, &list);
4931 	}
4932 
4933 	/* now free all selected PAs */
4934 	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4935 
4936 		/* remove from object (inode or locality group) */
4937 		spin_lock(pa->pa_obj_lock);
4938 		list_del_rcu(&pa->pa_inode_list);
4939 		spin_unlock(pa->pa_obj_lock);
4940 
4941 		if (pa->pa_type == MB_GROUP_PA)
4942 			ext4_mb_release_group_pa(&e4b, pa);
4943 		else
4944 			ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4945 
4946 		list_del(&pa->u.pa_tmp_list);
4947 		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4948 	}
4949 
4950 	ext4_unlock_group(sb, group);
4951 	ext4_mb_unload_buddy(&e4b);
4952 	put_bh(bitmap_bh);
4953 out_dbg:
4954 	mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4955 		 free, group, grp->bb_free);
4956 	return free;
4957 }
4958 
4959 /*
4960  * releases all non-used preallocated blocks for given inode
4961  *
4962  * It's important to discard preallocations under i_data_sem
4963  * We don't want another block to be served from the prealloc
4964  * space when we are discarding the inode prealloc space.
4965  *
4966  * FIXME!! Make sure it is valid at all the call sites
4967  */
4968 void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4969 {
4970 	struct ext4_inode_info *ei = EXT4_I(inode);
4971 	struct super_block *sb = inode->i_sb;
4972 	struct buffer_head *bitmap_bh = NULL;
4973 	struct ext4_prealloc_space *pa, *tmp;
4974 	ext4_group_t group = 0;
4975 	struct list_head list;
4976 	struct ext4_buddy e4b;
4977 	int err;
4978 
4979 	if (!S_ISREG(inode->i_mode)) {
4980 		/*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4981 		return;
4982 	}
4983 
4984 	if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4985 		return;
4986 
4987 	mb_debug(sb, "discard preallocation for inode %lu\n",
4988 		 inode->i_ino);
4989 	trace_ext4_discard_preallocations(inode,
4990 			atomic_read(&ei->i_prealloc_active), needed);
4991 
4992 	INIT_LIST_HEAD(&list);
4993 
4994 	if (needed == 0)
4995 		needed = UINT_MAX;
4996 
4997 repeat:
4998 	/* first, collect all pa's in the inode */
4999 	spin_lock(&ei->i_prealloc_lock);
5000 	while (!list_empty(&ei->i_prealloc_list) && needed) {
5001 		pa = list_entry(ei->i_prealloc_list.prev,
5002 				struct ext4_prealloc_space, pa_inode_list);
5003 		BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
5004 		spin_lock(&pa->pa_lock);
5005 		if (atomic_read(&pa->pa_count)) {
5006 			/* this shouldn't happen often - nobody should
5007 			 * use preallocation while we're discarding it */
5008 			spin_unlock(&pa->pa_lock);
5009 			spin_unlock(&ei->i_prealloc_lock);
5010 			ext4_msg(sb, KERN_ERR,
5011 				 "uh-oh! used pa while discarding");
5012 			WARN_ON(1);
5013 			schedule_timeout_uninterruptible(HZ);
5014 			goto repeat;
5015 
5016 		}
5017 		if (pa->pa_deleted == 0) {
5018 			ext4_mb_mark_pa_deleted(sb, pa);
5019 			spin_unlock(&pa->pa_lock);
5020 			list_del_rcu(&pa->pa_inode_list);
5021 			list_add(&pa->u.pa_tmp_list, &list);
5022 			needed--;
5023 			continue;
5024 		}
5025 
5026 		/* someone is deleting pa right now */
5027 		spin_unlock(&pa->pa_lock);
5028 		spin_unlock(&ei->i_prealloc_lock);
5029 
5030 		/* we have to wait here because pa_deleted
5031 		 * doesn't mean pa is already unlinked from
5032 		 * the list. as we might be called from
5033 		 * ->clear_inode() the inode will get freed
5034 		 * and concurrent thread which is unlinking
5035 		 * pa from inode's list may access already
5036 		 * freed memory, bad-bad-bad */
5037 
5038 		/* XXX: if this happens too often, we can
5039 		 * add a flag to force wait only in case
5040 		 * of ->clear_inode(), but not in case of
5041 		 * regular truncate */
5042 		schedule_timeout_uninterruptible(HZ);
5043 		goto repeat;
5044 	}
5045 	spin_unlock(&ei->i_prealloc_lock);
5046 
5047 	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5048 		BUG_ON(pa->pa_type != MB_INODE_PA);
5049 		group = ext4_get_group_number(sb, pa->pa_pstart);
5050 
5051 		err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5052 					     GFP_NOFS|__GFP_NOFAIL);
5053 		if (err) {
5054 			ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5055 				       err, group);
5056 			continue;
5057 		}
5058 
5059 		bitmap_bh = ext4_read_block_bitmap(sb, group);
5060 		if (IS_ERR(bitmap_bh)) {
5061 			err = PTR_ERR(bitmap_bh);
5062 			ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5063 				       err, group);
5064 			ext4_mb_unload_buddy(&e4b);
5065 			continue;
5066 		}
5067 
5068 		ext4_lock_group(sb, group);
5069 		list_del(&pa->pa_group_list);
5070 		ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5071 		ext4_unlock_group(sb, group);
5072 
5073 		ext4_mb_unload_buddy(&e4b);
5074 		put_bh(bitmap_bh);
5075 
5076 		list_del(&pa->u.pa_tmp_list);
5077 		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5078 	}
5079 }
5080 
5081 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5082 {
5083 	struct ext4_prealloc_space *pa;
5084 
5085 	BUG_ON(ext4_pspace_cachep == NULL);
5086 	pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5087 	if (!pa)
5088 		return -ENOMEM;
5089 	atomic_set(&pa->pa_count, 1);
5090 	ac->ac_pa = pa;
5091 	return 0;
5092 }
5093 
5094 static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
5095 {
5096 	struct ext4_prealloc_space *pa = ac->ac_pa;
5097 
5098 	BUG_ON(!pa);
5099 	ac->ac_pa = NULL;
5100 	WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5101 	kmem_cache_free(ext4_pspace_cachep, pa);
5102 }
5103 
5104 #ifdef CONFIG_EXT4_DEBUG
5105 static inline void ext4_mb_show_pa(struct super_block *sb)
5106 {
5107 	ext4_group_t i, ngroups;
5108 
5109 	if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5110 		return;
5111 
5112 	ngroups = ext4_get_groups_count(sb);
5113 	mb_debug(sb, "groups: ");
5114 	for (i = 0; i < ngroups; i++) {
5115 		struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5116 		struct ext4_prealloc_space *pa;
5117 		ext4_grpblk_t start;
5118 		struct list_head *cur;
5119 		ext4_lock_group(sb, i);
5120 		list_for_each(cur, &grp->bb_prealloc_list) {
5121 			pa = list_entry(cur, struct ext4_prealloc_space,
5122 					pa_group_list);
5123 			spin_lock(&pa->pa_lock);
5124 			ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5125 						     NULL, &start);
5126 			spin_unlock(&pa->pa_lock);
5127 			mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5128 				 pa->pa_len);
5129 		}
5130 		ext4_unlock_group(sb, i);
5131 		mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5132 			 grp->bb_fragments);
5133 	}
5134 }
5135 
5136 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5137 {
5138 	struct super_block *sb = ac->ac_sb;
5139 
5140 	if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
5141 		return;
5142 
5143 	mb_debug(sb, "Can't allocate:"
5144 			" Allocation context details:");
5145 	mb_debug(sb, "status %u flags 0x%x",
5146 			ac->ac_status, ac->ac_flags);
5147 	mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5148 			"goal %lu/%lu/%lu@%lu, "
5149 			"best %lu/%lu/%lu@%lu cr %d",
5150 			(unsigned long)ac->ac_o_ex.fe_group,
5151 			(unsigned long)ac->ac_o_ex.fe_start,
5152 			(unsigned long)ac->ac_o_ex.fe_len,
5153 			(unsigned long)ac->ac_o_ex.fe_logical,
5154 			(unsigned long)ac->ac_g_ex.fe_group,
5155 			(unsigned long)ac->ac_g_ex.fe_start,
5156 			(unsigned long)ac->ac_g_ex.fe_len,
5157 			(unsigned long)ac->ac_g_ex.fe_logical,
5158 			(unsigned long)ac->ac_b_ex.fe_group,
5159 			(unsigned long)ac->ac_b_ex.fe_start,
5160 			(unsigned long)ac->ac_b_ex.fe_len,
5161 			(unsigned long)ac->ac_b_ex.fe_logical,
5162 			(int)ac->ac_criteria);
5163 	mb_debug(sb, "%u found", ac->ac_found);
5164 	ext4_mb_show_pa(sb);
5165 }
5166 #else
5167 static inline void ext4_mb_show_pa(struct super_block *sb)
5168 {
5169 	return;
5170 }
5171 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5172 {
5173 	ext4_mb_show_pa(ac->ac_sb);
5174 	return;
5175 }
5176 #endif
5177 
5178 /*
5179  * We use locality group preallocation for small size file. The size of the
5180  * file is determined by the current size or the resulting size after
5181  * allocation which ever is larger
5182  *
5183  * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5184  */
5185 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5186 {
5187 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5188 	int bsbits = ac->ac_sb->s_blocksize_bits;
5189 	loff_t size, isize;
5190 
5191 	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5192 		return;
5193 
5194 	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5195 		return;
5196 
5197 	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
5198 	isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5199 		>> bsbits;
5200 
5201 	if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5202 	    !inode_is_open_for_write(ac->ac_inode)) {
5203 		ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5204 		return;
5205 	}
5206 
5207 	if (sbi->s_mb_group_prealloc <= 0) {
5208 		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5209 		return;
5210 	}
5211 
5212 	/* don't use group allocation for large files */
5213 	size = max(size, isize);
5214 	if (size > sbi->s_mb_stream_request) {
5215 		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5216 		return;
5217 	}
5218 
5219 	BUG_ON(ac->ac_lg != NULL);
5220 	/*
5221 	 * locality group prealloc space are per cpu. The reason for having
5222 	 * per cpu locality group is to reduce the contention between block
5223 	 * request from multiple CPUs.
5224 	 */
5225 	ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5226 
5227 	/* we're going to use group allocation */
5228 	ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5229 
5230 	/* serialize all allocations in the group */
5231 	mutex_lock(&ac->ac_lg->lg_mutex);
5232 }
5233 
5234 static noinline_for_stack int
5235 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5236 				struct ext4_allocation_request *ar)
5237 {
5238 	struct super_block *sb = ar->inode->i_sb;
5239 	struct ext4_sb_info *sbi = EXT4_SB(sb);
5240 	struct ext4_super_block *es = sbi->s_es;
5241 	ext4_group_t group;
5242 	unsigned int len;
5243 	ext4_fsblk_t goal;
5244 	ext4_grpblk_t block;
5245 
5246 	/* we can't allocate > group size */
5247 	len = ar->len;
5248 
5249 	/* just a dirty hack to filter too big requests  */
5250 	if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5251 		len = EXT4_CLUSTERS_PER_GROUP(sb);
5252 
5253 	/* start searching from the goal */
5254 	goal = ar->goal;
5255 	if (goal < le32_to_cpu(es->s_first_data_block) ||
5256 			goal >= ext4_blocks_count(es))
5257 		goal = le32_to_cpu(es->s_first_data_block);
5258 	ext4_get_group_no_and_offset(sb, goal, &group, &block);
5259 
5260 	/* set up allocation goals */
5261 	ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5262 	ac->ac_status = AC_STATUS_CONTINUE;
5263 	ac->ac_sb = sb;
5264 	ac->ac_inode = ar->inode;
5265 	ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5266 	ac->ac_o_ex.fe_group = group;
5267 	ac->ac_o_ex.fe_start = block;
5268 	ac->ac_o_ex.fe_len = len;
5269 	ac->ac_g_ex = ac->ac_o_ex;
5270 	ac->ac_flags = ar->flags;
5271 
5272 	/* we have to define context: we'll work with a file or
5273 	 * locality group. this is a policy, actually */
5274 	ext4_mb_group_or_file(ac);
5275 
5276 	mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5277 			"left: %u/%u, right %u/%u to %swritable\n",
5278 			(unsigned) ar->len, (unsigned) ar->logical,
5279 			(unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5280 			(unsigned) ar->lleft, (unsigned) ar->pleft,
5281 			(unsigned) ar->lright, (unsigned) ar->pright,
5282 			inode_is_open_for_write(ar->inode) ? "" : "non-");
5283 	return 0;
5284 
5285 }
5286 
5287 static noinline_for_stack void
5288 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5289 					struct ext4_locality_group *lg,
5290 					int order, int total_entries)
5291 {
5292 	ext4_group_t group = 0;
5293 	struct ext4_buddy e4b;
5294 	struct list_head discard_list;
5295 	struct ext4_prealloc_space *pa, *tmp;
5296 
5297 	mb_debug(sb, "discard locality group preallocation\n");
5298 
5299 	INIT_LIST_HEAD(&discard_list);
5300 
5301 	spin_lock(&lg->lg_prealloc_lock);
5302 	list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5303 				pa_inode_list,
5304 				lockdep_is_held(&lg->lg_prealloc_lock)) {
5305 		spin_lock(&pa->pa_lock);
5306 		if (atomic_read(&pa->pa_count)) {
5307 			/*
5308 			 * This is the pa that we just used
5309 			 * for block allocation. So don't
5310 			 * free that
5311 			 */
5312 			spin_unlock(&pa->pa_lock);
5313 			continue;
5314 		}
5315 		if (pa->pa_deleted) {
5316 			spin_unlock(&pa->pa_lock);
5317 			continue;
5318 		}
5319 		/* only lg prealloc space */
5320 		BUG_ON(pa->pa_type != MB_GROUP_PA);
5321 
5322 		/* seems this one can be freed ... */
5323 		ext4_mb_mark_pa_deleted(sb, pa);
5324 		spin_unlock(&pa->pa_lock);
5325 
5326 		list_del_rcu(&pa->pa_inode_list);
5327 		list_add(&pa->u.pa_tmp_list, &discard_list);
5328 
5329 		total_entries--;
5330 		if (total_entries <= 5) {
5331 			/*
5332 			 * we want to keep only 5 entries
5333 			 * allowing it to grow to 8. This
5334 			 * mak sure we don't call discard
5335 			 * soon for this list.
5336 			 */
5337 			break;
5338 		}
5339 	}
5340 	spin_unlock(&lg->lg_prealloc_lock);
5341 
5342 	list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5343 		int err;
5344 
5345 		group = ext4_get_group_number(sb, pa->pa_pstart);
5346 		err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5347 					     GFP_NOFS|__GFP_NOFAIL);
5348 		if (err) {
5349 			ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5350 				       err, group);
5351 			continue;
5352 		}
5353 		ext4_lock_group(sb, group);
5354 		list_del(&pa->pa_group_list);
5355 		ext4_mb_release_group_pa(&e4b, pa);
5356 		ext4_unlock_group(sb, group);
5357 
5358 		ext4_mb_unload_buddy(&e4b);
5359 		list_del(&pa->u.pa_tmp_list);
5360 		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5361 	}
5362 }
5363 
5364 /*
5365  * We have incremented pa_count. So it cannot be freed at this
5366  * point. Also we hold lg_mutex. So no parallel allocation is
5367  * possible from this lg. That means pa_free cannot be updated.
5368  *
5369  * A parallel ext4_mb_discard_group_preallocations is possible.
5370  * which can cause the lg_prealloc_list to be updated.
5371  */
5372 
5373 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5374 {
5375 	int order, added = 0, lg_prealloc_count = 1;
5376 	struct super_block *sb = ac->ac_sb;
5377 	struct ext4_locality_group *lg = ac->ac_lg;
5378 	struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5379 
5380 	order = fls(pa->pa_free) - 1;
5381 	if (order > PREALLOC_TB_SIZE - 1)
5382 		/* The max size of hash table is PREALLOC_TB_SIZE */
5383 		order = PREALLOC_TB_SIZE - 1;
5384 	/* Add the prealloc space to lg */
5385 	spin_lock(&lg->lg_prealloc_lock);
5386 	list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5387 				pa_inode_list,
5388 				lockdep_is_held(&lg->lg_prealloc_lock)) {
5389 		spin_lock(&tmp_pa->pa_lock);
5390 		if (tmp_pa->pa_deleted) {
5391 			spin_unlock(&tmp_pa->pa_lock);
5392 			continue;
5393 		}
5394 		if (!added && pa->pa_free < tmp_pa->pa_free) {
5395 			/* Add to the tail of the previous entry */
5396 			list_add_tail_rcu(&pa->pa_inode_list,
5397 						&tmp_pa->pa_inode_list);
5398 			added = 1;
5399 			/*
5400 			 * we want to count the total
5401 			 * number of entries in the list
5402 			 */
5403 		}
5404 		spin_unlock(&tmp_pa->pa_lock);
5405 		lg_prealloc_count++;
5406 	}
5407 	if (!added)
5408 		list_add_tail_rcu(&pa->pa_inode_list,
5409 					&lg->lg_prealloc_list[order]);
5410 	spin_unlock(&lg->lg_prealloc_lock);
5411 
5412 	/* Now trim the list to be not more than 8 elements */
5413 	if (lg_prealloc_count > 8) {
5414 		ext4_mb_discard_lg_preallocations(sb, lg,
5415 						  order, lg_prealloc_count);
5416 		return;
5417 	}
5418 	return ;
5419 }
5420 
5421 /*
5422  * if per-inode prealloc list is too long, trim some PA
5423  */
5424 static void ext4_mb_trim_inode_pa(struct inode *inode)
5425 {
5426 	struct ext4_inode_info *ei = EXT4_I(inode);
5427 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5428 	int count, delta;
5429 
5430 	count = atomic_read(&ei->i_prealloc_active);
5431 	delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
5432 	if (count > sbi->s_mb_max_inode_prealloc + delta) {
5433 		count -= sbi->s_mb_max_inode_prealloc;
5434 		ext4_discard_preallocations(inode, count);
5435 	}
5436 }
5437 
5438 /*
5439  * release all resource we used in allocation
5440  */
5441 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
5442 {
5443 	struct inode *inode = ac->ac_inode;
5444 	struct ext4_inode_info *ei = EXT4_I(inode);
5445 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5446 	struct ext4_prealloc_space *pa = ac->ac_pa;
5447 	if (pa) {
5448 		if (pa->pa_type == MB_GROUP_PA) {
5449 			/* see comment in ext4_mb_use_group_pa() */
5450 			spin_lock(&pa->pa_lock);
5451 			pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5452 			pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5453 			pa->pa_free -= ac->ac_b_ex.fe_len;
5454 			pa->pa_len -= ac->ac_b_ex.fe_len;
5455 			spin_unlock(&pa->pa_lock);
5456 
5457 			/*
5458 			 * We want to add the pa to the right bucket.
5459 			 * Remove it from the list and while adding
5460 			 * make sure the list to which we are adding
5461 			 * doesn't grow big.
5462 			 */
5463 			if (likely(pa->pa_free)) {
5464 				spin_lock(pa->pa_obj_lock);
5465 				list_del_rcu(&pa->pa_inode_list);
5466 				spin_unlock(pa->pa_obj_lock);
5467 				ext4_mb_add_n_trim(ac);
5468 			}
5469 		}
5470 
5471 		if (pa->pa_type == MB_INODE_PA) {
5472 			/*
5473 			 * treat per-inode prealloc list as a lru list, then try
5474 			 * to trim the least recently used PA.
5475 			 */
5476 			spin_lock(pa->pa_obj_lock);
5477 			list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
5478 			spin_unlock(pa->pa_obj_lock);
5479 		}
5480 
5481 		ext4_mb_put_pa(ac, ac->ac_sb, pa);
5482 	}
5483 	if (ac->ac_bitmap_page)
5484 		put_page(ac->ac_bitmap_page);
5485 	if (ac->ac_buddy_page)
5486 		put_page(ac->ac_buddy_page);
5487 	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5488 		mutex_unlock(&ac->ac_lg->lg_mutex);
5489 	ext4_mb_collect_stats(ac);
5490 	ext4_mb_trim_inode_pa(inode);
5491 	return 0;
5492 }
5493 
5494 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
5495 {
5496 	ext4_group_t i, ngroups = ext4_get_groups_count(sb);
5497 	int ret;
5498 	int freed = 0, busy = 0;
5499 	int retry = 0;
5500 
5501 	trace_ext4_mb_discard_preallocations(sb, needed);
5502 
5503 	if (needed == 0)
5504 		needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
5505  repeat:
5506 	for (i = 0; i < ngroups && needed > 0; i++) {
5507 		ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
5508 		freed += ret;
5509 		needed -= ret;
5510 		cond_resched();
5511 	}
5512 
5513 	if (needed > 0 && busy && ++retry < 3) {
5514 		busy = 0;
5515 		goto repeat;
5516 	}
5517 
5518 	return freed;
5519 }
5520 
5521 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
5522 			struct ext4_allocation_context *ac, u64 *seq)
5523 {
5524 	int freed;
5525 	u64 seq_retry = 0;
5526 	bool ret = false;
5527 
5528 	freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
5529 	if (freed) {
5530 		ret = true;
5531 		goto out_dbg;
5532 	}
5533 	seq_retry = ext4_get_discard_pa_seq_sum();
5534 	if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
5535 		ac->ac_flags |= EXT4_MB_STRICT_CHECK;
5536 		*seq = seq_retry;
5537 		ret = true;
5538 	}
5539 
5540 out_dbg:
5541 	mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
5542 	return ret;
5543 }
5544 
5545 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5546 				struct ext4_allocation_request *ar, int *errp);
5547 
5548 /*
5549  * Main entry point into mballoc to allocate blocks
5550  * it tries to use preallocation first, then falls back
5551  * to usual allocation
5552  */
5553 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
5554 				struct ext4_allocation_request *ar, int *errp)
5555 {
5556 	struct ext4_allocation_context *ac = NULL;
5557 	struct ext4_sb_info *sbi;
5558 	struct super_block *sb;
5559 	ext4_fsblk_t block = 0;
5560 	unsigned int inquota = 0;
5561 	unsigned int reserv_clstrs = 0;
5562 	u64 seq;
5563 
5564 	might_sleep();
5565 	sb = ar->inode->i_sb;
5566 	sbi = EXT4_SB(sb);
5567 
5568 	trace_ext4_request_blocks(ar);
5569 	if (sbi->s_mount_state & EXT4_FC_REPLAY)
5570 		return ext4_mb_new_blocks_simple(handle, ar, errp);
5571 
5572 	/* Allow to use superuser reservation for quota file */
5573 	if (ext4_is_quota_file(ar->inode))
5574 		ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
5575 
5576 	if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
5577 		/* Without delayed allocation we need to verify
5578 		 * there is enough free blocks to do block allocation
5579 		 * and verify allocation doesn't exceed the quota limits.
5580 		 */
5581 		while (ar->len &&
5582 			ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
5583 
5584 			/* let others to free the space */
5585 			cond_resched();
5586 			ar->len = ar->len >> 1;
5587 		}
5588 		if (!ar->len) {
5589 			ext4_mb_show_pa(sb);
5590 			*errp = -ENOSPC;
5591 			return 0;
5592 		}
5593 		reserv_clstrs = ar->len;
5594 		if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
5595 			dquot_alloc_block_nofail(ar->inode,
5596 						 EXT4_C2B(sbi, ar->len));
5597 		} else {
5598 			while (ar->len &&
5599 				dquot_alloc_block(ar->inode,
5600 						  EXT4_C2B(sbi, ar->len))) {
5601 
5602 				ar->flags |= EXT4_MB_HINT_NOPREALLOC;
5603 				ar->len--;
5604 			}
5605 		}
5606 		inquota = ar->len;
5607 		if (ar->len == 0) {
5608 			*errp = -EDQUOT;
5609 			goto out;
5610 		}
5611 	}
5612 
5613 	ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
5614 	if (!ac) {
5615 		ar->len = 0;
5616 		*errp = -ENOMEM;
5617 		goto out;
5618 	}
5619 
5620 	*errp = ext4_mb_initialize_context(ac, ar);
5621 	if (*errp) {
5622 		ar->len = 0;
5623 		goto out;
5624 	}
5625 
5626 	ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
5627 	seq = this_cpu_read(discard_pa_seq);
5628 	if (!ext4_mb_use_preallocated(ac)) {
5629 		ac->ac_op = EXT4_MB_HISTORY_ALLOC;
5630 		ext4_mb_normalize_request(ac, ar);
5631 
5632 		*errp = ext4_mb_pa_alloc(ac);
5633 		if (*errp)
5634 			goto errout;
5635 repeat:
5636 		/* allocate space in core */
5637 		*errp = ext4_mb_regular_allocator(ac);
5638 		/*
5639 		 * pa allocated above is added to grp->bb_prealloc_list only
5640 		 * when we were able to allocate some block i.e. when
5641 		 * ac->ac_status == AC_STATUS_FOUND.
5642 		 * And error from above mean ac->ac_status != AC_STATUS_FOUND
5643 		 * So we have to free this pa here itself.
5644 		 */
5645 		if (*errp) {
5646 			ext4_mb_pa_free(ac);
5647 			ext4_discard_allocated_blocks(ac);
5648 			goto errout;
5649 		}
5650 		if (ac->ac_status == AC_STATUS_FOUND &&
5651 			ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
5652 			ext4_mb_pa_free(ac);
5653 	}
5654 	if (likely(ac->ac_status == AC_STATUS_FOUND)) {
5655 		*errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
5656 		if (*errp) {
5657 			ext4_discard_allocated_blocks(ac);
5658 			goto errout;
5659 		} else {
5660 			block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5661 			ar->len = ac->ac_b_ex.fe_len;
5662 		}
5663 	} else {
5664 		if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5665 			goto repeat;
5666 		/*
5667 		 * If block allocation fails then the pa allocated above
5668 		 * needs to be freed here itself.
5669 		 */
5670 		ext4_mb_pa_free(ac);
5671 		*errp = -ENOSPC;
5672 	}
5673 
5674 errout:
5675 	if (*errp) {
5676 		ac->ac_b_ex.fe_len = 0;
5677 		ar->len = 0;
5678 		ext4_mb_show_ac(ac);
5679 	}
5680 	ext4_mb_release_context(ac);
5681 out:
5682 	if (ac)
5683 		kmem_cache_free(ext4_ac_cachep, ac);
5684 	if (inquota && ar->len < inquota)
5685 		dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5686 	if (!ar->len) {
5687 		if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5688 			/* release all the reserved blocks if non delalloc */
5689 			percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5690 						reserv_clstrs);
5691 	}
5692 
5693 	trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5694 
5695 	return block;
5696 }
5697 
5698 /*
5699  * We can merge two free data extents only if the physical blocks
5700  * are contiguous, AND the extents were freed by the same transaction,
5701  * AND the blocks are associated with the same group.
5702  */
5703 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5704 					struct ext4_free_data *entry,
5705 					struct ext4_free_data *new_entry,
5706 					struct rb_root *entry_rb_root)
5707 {
5708 	if ((entry->efd_tid != new_entry->efd_tid) ||
5709 	    (entry->efd_group != new_entry->efd_group))
5710 		return;
5711 	if (entry->efd_start_cluster + entry->efd_count ==
5712 	    new_entry->efd_start_cluster) {
5713 		new_entry->efd_start_cluster = entry->efd_start_cluster;
5714 		new_entry->efd_count += entry->efd_count;
5715 	} else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5716 		   entry->efd_start_cluster) {
5717 		new_entry->efd_count += entry->efd_count;
5718 	} else
5719 		return;
5720 	spin_lock(&sbi->s_md_lock);
5721 	list_del(&entry->efd_list);
5722 	spin_unlock(&sbi->s_md_lock);
5723 	rb_erase(&entry->efd_node, entry_rb_root);
5724 	kmem_cache_free(ext4_free_data_cachep, entry);
5725 }
5726 
5727 static noinline_for_stack int
5728 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5729 		      struct ext4_free_data *new_entry)
5730 {
5731 	ext4_group_t group = e4b->bd_group;
5732 	ext4_grpblk_t cluster;
5733 	ext4_grpblk_t clusters = new_entry->efd_count;
5734 	struct ext4_free_data *entry;
5735 	struct ext4_group_info *db = e4b->bd_info;
5736 	struct super_block *sb = e4b->bd_sb;
5737 	struct ext4_sb_info *sbi = EXT4_SB(sb);
5738 	struct rb_node **n = &db->bb_free_root.rb_node, *node;
5739 	struct rb_node *parent = NULL, *new_node;
5740 
5741 	BUG_ON(!ext4_handle_valid(handle));
5742 	BUG_ON(e4b->bd_bitmap_page == NULL);
5743 	BUG_ON(e4b->bd_buddy_page == NULL);
5744 
5745 	new_node = &new_entry->efd_node;
5746 	cluster = new_entry->efd_start_cluster;
5747 
5748 	if (!*n) {
5749 		/* first free block exent. We need to
5750 		   protect buddy cache from being freed,
5751 		 * otherwise we'll refresh it from
5752 		 * on-disk bitmap and lose not-yet-available
5753 		 * blocks */
5754 		get_page(e4b->bd_buddy_page);
5755 		get_page(e4b->bd_bitmap_page);
5756 	}
5757 	while (*n) {
5758 		parent = *n;
5759 		entry = rb_entry(parent, struct ext4_free_data, efd_node);
5760 		if (cluster < entry->efd_start_cluster)
5761 			n = &(*n)->rb_left;
5762 		else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5763 			n = &(*n)->rb_right;
5764 		else {
5765 			ext4_grp_locked_error(sb, group, 0,
5766 				ext4_group_first_block_no(sb, group) +
5767 				EXT4_C2B(sbi, cluster),
5768 				"Block already on to-be-freed list");
5769 			kmem_cache_free(ext4_free_data_cachep, new_entry);
5770 			return 0;
5771 		}
5772 	}
5773 
5774 	rb_link_node(new_node, parent, n);
5775 	rb_insert_color(new_node, &db->bb_free_root);
5776 
5777 	/* Now try to see the extent can be merged to left and right */
5778 	node = rb_prev(new_node);
5779 	if (node) {
5780 		entry = rb_entry(node, struct ext4_free_data, efd_node);
5781 		ext4_try_merge_freed_extent(sbi, entry, new_entry,
5782 					    &(db->bb_free_root));
5783 	}
5784 
5785 	node = rb_next(new_node);
5786 	if (node) {
5787 		entry = rb_entry(node, struct ext4_free_data, efd_node);
5788 		ext4_try_merge_freed_extent(sbi, entry, new_entry,
5789 					    &(db->bb_free_root));
5790 	}
5791 
5792 	spin_lock(&sbi->s_md_lock);
5793 	list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5794 	sbi->s_mb_free_pending += clusters;
5795 	spin_unlock(&sbi->s_md_lock);
5796 	return 0;
5797 }
5798 
5799 /*
5800  * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5801  * linearly starting at the goal block and also excludes the blocks which
5802  * are going to be in use after fast commit replay.
5803  */
5804 static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5805 				struct ext4_allocation_request *ar, int *errp)
5806 {
5807 	struct buffer_head *bitmap_bh;
5808 	struct super_block *sb = ar->inode->i_sb;
5809 	ext4_group_t group;
5810 	ext4_grpblk_t blkoff;
5811 	ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
5812 	ext4_grpblk_t i = 0;
5813 	ext4_fsblk_t goal, block;
5814 	struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5815 
5816 	goal = ar->goal;
5817 	if (goal < le32_to_cpu(es->s_first_data_block) ||
5818 			goal >= ext4_blocks_count(es))
5819 		goal = le32_to_cpu(es->s_first_data_block);
5820 
5821 	ar->len = 0;
5822 	ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5823 	for (; group < ext4_get_groups_count(sb); group++) {
5824 		bitmap_bh = ext4_read_block_bitmap(sb, group);
5825 		if (IS_ERR(bitmap_bh)) {
5826 			*errp = PTR_ERR(bitmap_bh);
5827 			pr_warn("Failed to read block bitmap\n");
5828 			return 0;
5829 		}
5830 
5831 		ext4_get_group_no_and_offset(sb,
5832 			max(ext4_group_first_block_no(sb, group), goal),
5833 			NULL, &blkoff);
5834 		while (1) {
5835 			i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
5836 						blkoff);
5837 			if (i >= max)
5838 				break;
5839 			if (ext4_fc_replay_check_excluded(sb,
5840 				ext4_group_first_block_no(sb, group) + i)) {
5841 				blkoff = i + 1;
5842 			} else
5843 				break;
5844 		}
5845 		brelse(bitmap_bh);
5846 		if (i < max)
5847 			break;
5848 	}
5849 
5850 	if (group >= ext4_get_groups_count(sb) || i >= max) {
5851 		*errp = -ENOSPC;
5852 		return 0;
5853 	}
5854 
5855 	block = ext4_group_first_block_no(sb, group) + i;
5856 	ext4_mb_mark_bb(sb, block, 1, 1);
5857 	ar->len = 1;
5858 
5859 	return block;
5860 }
5861 
5862 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5863 					unsigned long count)
5864 {
5865 	struct buffer_head *bitmap_bh;
5866 	struct super_block *sb = inode->i_sb;
5867 	struct ext4_group_desc *gdp;
5868 	struct buffer_head *gdp_bh;
5869 	ext4_group_t group;
5870 	ext4_grpblk_t blkoff;
5871 	int already_freed = 0, err, i;
5872 
5873 	ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5874 	bitmap_bh = ext4_read_block_bitmap(sb, group);
5875 	if (IS_ERR(bitmap_bh)) {
5876 		err = PTR_ERR(bitmap_bh);
5877 		pr_warn("Failed to read block bitmap\n");
5878 		return;
5879 	}
5880 	gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5881 	if (!gdp)
5882 		return;
5883 
5884 	for (i = 0; i < count; i++) {
5885 		if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5886 			already_freed++;
5887 	}
5888 	mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5889 	err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5890 	if (err)
5891 		return;
5892 	ext4_free_group_clusters_set(
5893 		sb, gdp, ext4_free_group_clusters(sb, gdp) +
5894 		count - already_freed);
5895 	ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5896 	ext4_group_desc_csum_set(sb, group, gdp);
5897 	ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5898 	sync_dirty_buffer(bitmap_bh);
5899 	sync_dirty_buffer(gdp_bh);
5900 	brelse(bitmap_bh);
5901 }
5902 
5903 /**
5904  * ext4_mb_clear_bb() -- helper function for freeing blocks.
5905  *			Used by ext4_free_blocks()
5906  * @handle:		handle for this transaction
5907  * @inode:		inode
5908  * @block:		starting physical block to be freed
5909  * @count:		number of blocks to be freed
5910  * @flags:		flags used by ext4_free_blocks
5911  */
5912 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
5913 			       ext4_fsblk_t block, unsigned long count,
5914 			       int flags)
5915 {
5916 	struct buffer_head *bitmap_bh = NULL;
5917 	struct super_block *sb = inode->i_sb;
5918 	struct ext4_group_desc *gdp;
5919 	unsigned int overflow;
5920 	ext4_grpblk_t bit;
5921 	struct buffer_head *gd_bh;
5922 	ext4_group_t block_group;
5923 	struct ext4_sb_info *sbi;
5924 	struct ext4_buddy e4b;
5925 	unsigned int count_clusters;
5926 	int err = 0;
5927 	int ret;
5928 
5929 	sbi = EXT4_SB(sb);
5930 
5931 do_more:
5932 	overflow = 0;
5933 	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5934 
5935 	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5936 			ext4_get_group_info(sb, block_group))))
5937 		return;
5938 
5939 	/*
5940 	 * Check to see if we are freeing blocks across a group
5941 	 * boundary.
5942 	 */
5943 	if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5944 		overflow = EXT4_C2B(sbi, bit) + count -
5945 			EXT4_BLOCKS_PER_GROUP(sb);
5946 		count -= overflow;
5947 	}
5948 	count_clusters = EXT4_NUM_B2C(sbi, count);
5949 	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5950 	if (IS_ERR(bitmap_bh)) {
5951 		err = PTR_ERR(bitmap_bh);
5952 		bitmap_bh = NULL;
5953 		goto error_return;
5954 	}
5955 	gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5956 	if (!gdp) {
5957 		err = -EIO;
5958 		goto error_return;
5959 	}
5960 
5961 	if (!ext4_inode_block_valid(inode, block, count)) {
5962 		ext4_error(sb, "Freeing blocks in system zone - "
5963 			   "Block = %llu, count = %lu", block, count);
5964 		/* err = 0. ext4_std_error should be a no op */
5965 		goto error_return;
5966 	}
5967 
5968 	BUFFER_TRACE(bitmap_bh, "getting write access");
5969 	err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
5970 					    EXT4_JTR_NONE);
5971 	if (err)
5972 		goto error_return;
5973 
5974 	/*
5975 	 * We are about to modify some metadata.  Call the journal APIs
5976 	 * to unshare ->b_data if a currently-committing transaction is
5977 	 * using it
5978 	 */
5979 	BUFFER_TRACE(gd_bh, "get_write_access");
5980 	err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
5981 	if (err)
5982 		goto error_return;
5983 #ifdef AGGRESSIVE_CHECK
5984 	{
5985 		int i;
5986 		for (i = 0; i < count_clusters; i++)
5987 			BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5988 	}
5989 #endif
5990 	trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5991 
5992 	/* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5993 	err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5994 				     GFP_NOFS|__GFP_NOFAIL);
5995 	if (err)
5996 		goto error_return;
5997 
5998 	/*
5999 	 * We need to make sure we don't reuse the freed block until after the
6000 	 * transaction is committed. We make an exception if the inode is to be
6001 	 * written in writeback mode since writeback mode has weak data
6002 	 * consistency guarantees.
6003 	 */
6004 	if (ext4_handle_valid(handle) &&
6005 	    ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6006 	     !ext4_should_writeback_data(inode))) {
6007 		struct ext4_free_data *new_entry;
6008 		/*
6009 		 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6010 		 * to fail.
6011 		 */
6012 		new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6013 				GFP_NOFS|__GFP_NOFAIL);
6014 		new_entry->efd_start_cluster = bit;
6015 		new_entry->efd_group = block_group;
6016 		new_entry->efd_count = count_clusters;
6017 		new_entry->efd_tid = handle->h_transaction->t_tid;
6018 
6019 		ext4_lock_group(sb, block_group);
6020 		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6021 		ext4_mb_free_metadata(handle, &e4b, new_entry);
6022 	} else {
6023 		/* need to update group_info->bb_free and bitmap
6024 		 * with group lock held. generate_buddy look at
6025 		 * them with group lock_held
6026 		 */
6027 		if (test_opt(sb, DISCARD)) {
6028 			err = ext4_issue_discard(sb, block_group, bit, count,
6029 						 NULL);
6030 			if (err && err != -EOPNOTSUPP)
6031 				ext4_msg(sb, KERN_WARNING, "discard request in"
6032 					 " group:%u block:%d count:%lu failed"
6033 					 " with %d", block_group, bit, count,
6034 					 err);
6035 		} else
6036 			EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6037 
6038 		ext4_lock_group(sb, block_group);
6039 		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
6040 		mb_free_blocks(inode, &e4b, bit, count_clusters);
6041 	}
6042 
6043 	ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
6044 	ext4_free_group_clusters_set(sb, gdp, ret);
6045 	ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
6046 	ext4_group_desc_csum_set(sb, block_group, gdp);
6047 	ext4_unlock_group(sb, block_group);
6048 
6049 	if (sbi->s_log_groups_per_flex) {
6050 		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6051 		atomic64_add(count_clusters,
6052 			     &sbi_array_rcu_deref(sbi, s_flex_groups,
6053 						  flex_group)->free_clusters);
6054 	}
6055 
6056 	/*
6057 	 * on a bigalloc file system, defer the s_freeclusters_counter
6058 	 * update to the caller (ext4_remove_space and friends) so they
6059 	 * can determine if a cluster freed here should be rereserved
6060 	 */
6061 	if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6062 		if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6063 			dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6064 		percpu_counter_add(&sbi->s_freeclusters_counter,
6065 				   count_clusters);
6066 	}
6067 
6068 	ext4_mb_unload_buddy(&e4b);
6069 
6070 	/* We dirtied the bitmap block */
6071 	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6072 	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6073 
6074 	/* And the group descriptor block */
6075 	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6076 	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6077 	if (!err)
6078 		err = ret;
6079 
6080 	if (overflow && !err) {
6081 		block += count;
6082 		count = overflow;
6083 		put_bh(bitmap_bh);
6084 		goto do_more;
6085 	}
6086 error_return:
6087 	brelse(bitmap_bh);
6088 	ext4_std_error(sb, err);
6089 	return;
6090 }
6091 
6092 /**
6093  * ext4_free_blocks() -- Free given blocks and update quota
6094  * @handle:		handle for this transaction
6095  * @inode:		inode
6096  * @bh:			optional buffer of the block to be freed
6097  * @block:		starting physical block to be freed
6098  * @count:		number of blocks to be freed
6099  * @flags:		flags used by ext4_free_blocks
6100  */
6101 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6102 		      struct buffer_head *bh, ext4_fsblk_t block,
6103 		      unsigned long count, int flags)
6104 {
6105 	struct super_block *sb = inode->i_sb;
6106 	unsigned int overflow;
6107 	struct ext4_sb_info *sbi;
6108 
6109 	sbi = EXT4_SB(sb);
6110 
6111 	if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6112 		ext4_free_blocks_simple(inode, block, count);
6113 		return;
6114 	}
6115 
6116 	might_sleep();
6117 	if (bh) {
6118 		if (block)
6119 			BUG_ON(block != bh->b_blocknr);
6120 		else
6121 			block = bh->b_blocknr;
6122 	}
6123 
6124 	if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6125 	    !ext4_inode_block_valid(inode, block, count)) {
6126 		ext4_error(sb, "Freeing blocks not in datazone - "
6127 			   "block = %llu, count = %lu", block, count);
6128 		return;
6129 	}
6130 
6131 	ext4_debug("freeing block %llu\n", block);
6132 	trace_ext4_free_blocks(inode, block, count, flags);
6133 
6134 	if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6135 		BUG_ON(count > 1);
6136 
6137 		ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6138 			    inode, bh, block);
6139 	}
6140 
6141 	/*
6142 	 * If the extent to be freed does not begin on a cluster
6143 	 * boundary, we need to deal with partial clusters at the
6144 	 * beginning and end of the extent.  Normally we will free
6145 	 * blocks at the beginning or the end unless we are explicitly
6146 	 * requested to avoid doing so.
6147 	 */
6148 	overflow = EXT4_PBLK_COFF(sbi, block);
6149 	if (overflow) {
6150 		if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6151 			overflow = sbi->s_cluster_ratio - overflow;
6152 			block += overflow;
6153 			if (count > overflow)
6154 				count -= overflow;
6155 			else
6156 				return;
6157 		} else {
6158 			block -= overflow;
6159 			count += overflow;
6160 		}
6161 	}
6162 	overflow = EXT4_LBLK_COFF(sbi, count);
6163 	if (overflow) {
6164 		if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6165 			if (count > overflow)
6166 				count -= overflow;
6167 			else
6168 				return;
6169 		} else
6170 			count += sbi->s_cluster_ratio - overflow;
6171 	}
6172 
6173 	if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6174 		int i;
6175 		int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6176 
6177 		for (i = 0; i < count; i++) {
6178 			cond_resched();
6179 			if (is_metadata)
6180 				bh = sb_find_get_block(inode->i_sb, block + i);
6181 			ext4_forget(handle, is_metadata, inode, bh, block + i);
6182 		}
6183 	}
6184 
6185 	ext4_mb_clear_bb(handle, inode, block, count, flags);
6186 	return;
6187 }
6188 
6189 /**
6190  * ext4_group_add_blocks() -- Add given blocks to an existing group
6191  * @handle:			handle to this transaction
6192  * @sb:				super block
6193  * @block:			start physical block to add to the block group
6194  * @count:			number of blocks to free
6195  *
6196  * This marks the blocks as free in the bitmap and buddy.
6197  */
6198 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6199 			 ext4_fsblk_t block, unsigned long count)
6200 {
6201 	struct buffer_head *bitmap_bh = NULL;
6202 	struct buffer_head *gd_bh;
6203 	ext4_group_t block_group;
6204 	ext4_grpblk_t bit;
6205 	unsigned int i;
6206 	struct ext4_group_desc *desc;
6207 	struct ext4_sb_info *sbi = EXT4_SB(sb);
6208 	struct ext4_buddy e4b;
6209 	int err = 0, ret, free_clusters_count;
6210 	ext4_grpblk_t clusters_freed;
6211 	ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6212 	ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6213 	unsigned long cluster_count = last_cluster - first_cluster + 1;
6214 
6215 	ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6216 
6217 	if (count == 0)
6218 		return 0;
6219 
6220 	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6221 	/*
6222 	 * Check to see if we are freeing blocks across a group
6223 	 * boundary.
6224 	 */
6225 	if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6226 		ext4_warning(sb, "too many blocks added to group %u",
6227 			     block_group);
6228 		err = -EINVAL;
6229 		goto error_return;
6230 	}
6231 
6232 	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
6233 	if (IS_ERR(bitmap_bh)) {
6234 		err = PTR_ERR(bitmap_bh);
6235 		bitmap_bh = NULL;
6236 		goto error_return;
6237 	}
6238 
6239 	desc = ext4_get_group_desc(sb, block_group, &gd_bh);
6240 	if (!desc) {
6241 		err = -EIO;
6242 		goto error_return;
6243 	}
6244 
6245 	if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6246 		ext4_error(sb, "Adding blocks in system zones - "
6247 			   "Block = %llu, count = %lu",
6248 			   block, count);
6249 		err = -EINVAL;
6250 		goto error_return;
6251 	}
6252 
6253 	BUFFER_TRACE(bitmap_bh, "getting write access");
6254 	err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
6255 					    EXT4_JTR_NONE);
6256 	if (err)
6257 		goto error_return;
6258 
6259 	/*
6260 	 * We are about to modify some metadata.  Call the journal APIs
6261 	 * to unshare ->b_data if a currently-committing transaction is
6262 	 * using it
6263 	 */
6264 	BUFFER_TRACE(gd_bh, "get_write_access");
6265 	err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE);
6266 	if (err)
6267 		goto error_return;
6268 
6269 	for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
6270 		BUFFER_TRACE(bitmap_bh, "clear bit");
6271 		if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
6272 			ext4_error(sb, "bit already cleared for block %llu",
6273 				   (ext4_fsblk_t)(block + i));
6274 			BUFFER_TRACE(bitmap_bh, "bit already cleared");
6275 		} else {
6276 			clusters_freed++;
6277 		}
6278 	}
6279 
6280 	err = ext4_mb_load_buddy(sb, block_group, &e4b);
6281 	if (err)
6282 		goto error_return;
6283 
6284 	/*
6285 	 * need to update group_info->bb_free and bitmap
6286 	 * with group lock held. generate_buddy look at
6287 	 * them with group lock_held
6288 	 */
6289 	ext4_lock_group(sb, block_group);
6290 	mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
6291 	mb_free_blocks(NULL, &e4b, bit, cluster_count);
6292 	free_clusters_count = clusters_freed +
6293 		ext4_free_group_clusters(sb, desc);
6294 	ext4_free_group_clusters_set(sb, desc, free_clusters_count);
6295 	ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
6296 	ext4_group_desc_csum_set(sb, block_group, desc);
6297 	ext4_unlock_group(sb, block_group);
6298 	percpu_counter_add(&sbi->s_freeclusters_counter,
6299 			   clusters_freed);
6300 
6301 	if (sbi->s_log_groups_per_flex) {
6302 		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
6303 		atomic64_add(clusters_freed,
6304 			     &sbi_array_rcu_deref(sbi, s_flex_groups,
6305 						  flex_group)->free_clusters);
6306 	}
6307 
6308 	ext4_mb_unload_buddy(&e4b);
6309 
6310 	/* We dirtied the bitmap block */
6311 	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
6312 	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
6313 
6314 	/* And the group descriptor block */
6315 	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
6316 	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
6317 	if (!err)
6318 		err = ret;
6319 
6320 error_return:
6321 	brelse(bitmap_bh);
6322 	ext4_std_error(sb, err);
6323 	return err;
6324 }
6325 
6326 /**
6327  * ext4_trim_extent -- function to TRIM one single free extent in the group
6328  * @sb:		super block for the file system
6329  * @start:	starting block of the free extent in the alloc. group
6330  * @count:	number of blocks to TRIM
6331  * @e4b:	ext4 buddy for the group
6332  *
6333  * Trim "count" blocks starting at "start" in the "group". To assure that no
6334  * one will allocate those blocks, mark it as used in buddy bitmap. This must
6335  * be called with under the group lock.
6336  */
6337 static int ext4_trim_extent(struct super_block *sb,
6338 		int start, int count, struct ext4_buddy *e4b)
6339 __releases(bitlock)
6340 __acquires(bitlock)
6341 {
6342 	struct ext4_free_extent ex;
6343 	ext4_group_t group = e4b->bd_group;
6344 	int ret = 0;
6345 
6346 	trace_ext4_trim_extent(sb, group, start, count);
6347 
6348 	assert_spin_locked(ext4_group_lock_ptr(sb, group));
6349 
6350 	ex.fe_start = start;
6351 	ex.fe_group = group;
6352 	ex.fe_len = count;
6353 
6354 	/*
6355 	 * Mark blocks used, so no one can reuse them while
6356 	 * being trimmed.
6357 	 */
6358 	mb_mark_used(e4b, &ex);
6359 	ext4_unlock_group(sb, group);
6360 	ret = ext4_issue_discard(sb, group, start, count, NULL);
6361 	ext4_lock_group(sb, group);
6362 	mb_free_blocks(NULL, e4b, start, ex.fe_len);
6363 	return ret;
6364 }
6365 
6366 static int ext4_try_to_trim_range(struct super_block *sb,
6367 		struct ext4_buddy *e4b, ext4_grpblk_t start,
6368 		ext4_grpblk_t max, ext4_grpblk_t minblocks)
6369 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6370 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6371 {
6372 	ext4_grpblk_t next, count, free_count;
6373 	void *bitmap;
6374 
6375 	bitmap = e4b->bd_bitmap;
6376 	start = (e4b->bd_info->bb_first_free > start) ?
6377 		e4b->bd_info->bb_first_free : start;
6378 	count = 0;
6379 	free_count = 0;
6380 
6381 	while (start <= max) {
6382 		start = mb_find_next_zero_bit(bitmap, max + 1, start);
6383 		if (start > max)
6384 			break;
6385 		next = mb_find_next_bit(bitmap, max + 1, start);
6386 
6387 		if ((next - start) >= minblocks) {
6388 			int ret = ext4_trim_extent(sb, start, next - start, e4b);
6389 
6390 			if (ret && ret != -EOPNOTSUPP)
6391 				break;
6392 			count += next - start;
6393 		}
6394 		free_count += next - start;
6395 		start = next + 1;
6396 
6397 		if (fatal_signal_pending(current)) {
6398 			count = -ERESTARTSYS;
6399 			break;
6400 		}
6401 
6402 		if (need_resched()) {
6403 			ext4_unlock_group(sb, e4b->bd_group);
6404 			cond_resched();
6405 			ext4_lock_group(sb, e4b->bd_group);
6406 		}
6407 
6408 		if ((e4b->bd_info->bb_free - free_count) < minblocks)
6409 			break;
6410 	}
6411 
6412 	return count;
6413 }
6414 
6415 /**
6416  * ext4_trim_all_free -- function to trim all free space in alloc. group
6417  * @sb:			super block for file system
6418  * @group:		group to be trimmed
6419  * @start:		first group block to examine
6420  * @max:		last group block to examine
6421  * @minblocks:		minimum extent block count
6422  * @set_trimmed:	set the trimmed flag if at least one block is trimmed
6423  *
6424  * ext4_trim_all_free walks through group's block bitmap searching for free
6425  * extents. When the free extent is found, mark it as used in group buddy
6426  * bitmap. Then issue a TRIM command on this extent and free the extent in
6427  * the group buddy bitmap.
6428  */
6429 static ext4_grpblk_t
6430 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6431 		   ext4_grpblk_t start, ext4_grpblk_t max,
6432 		   ext4_grpblk_t minblocks, bool set_trimmed)
6433 {
6434 	struct ext4_buddy e4b;
6435 	int ret;
6436 
6437 	trace_ext4_trim_all_free(sb, group, start, max);
6438 
6439 	ret = ext4_mb_load_buddy(sb, group, &e4b);
6440 	if (ret) {
6441 		ext4_warning(sb, "Error %d loading buddy information for %u",
6442 			     ret, group);
6443 		return ret;
6444 	}
6445 
6446 	ext4_lock_group(sb, group);
6447 
6448 	if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6449 	    minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
6450 		ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6451 		if (ret >= 0 && set_trimmed)
6452 			EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
6453 	} else {
6454 		ret = 0;
6455 	}
6456 
6457 	ext4_unlock_group(sb, group);
6458 	ext4_mb_unload_buddy(&e4b);
6459 
6460 	ext4_debug("trimmed %d blocks in the group %d\n",
6461 		ret, group);
6462 
6463 	return ret;
6464 }
6465 
6466 /**
6467  * ext4_trim_fs() -- trim ioctl handle function
6468  * @sb:			superblock for filesystem
6469  * @range:		fstrim_range structure
6470  *
6471  * start:	First Byte to trim
6472  * len:		number of Bytes to trim from start
6473  * minlen:	minimum extent length in Bytes
6474  * ext4_trim_fs goes through all allocation groups containing Bytes from
6475  * start to start+len. For each such a group ext4_trim_all_free function
6476  * is invoked to trim all free space.
6477  */
6478 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6479 {
6480 	unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6481 	struct ext4_group_info *grp;
6482 	ext4_group_t group, first_group, last_group;
6483 	ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6484 	uint64_t start, end, minlen, trimmed = 0;
6485 	ext4_fsblk_t first_data_blk =
6486 			le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6487 	ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6488 	bool whole_group, eof = false;
6489 	int ret = 0;
6490 
6491 	start = range->start >> sb->s_blocksize_bits;
6492 	end = start + (range->len >> sb->s_blocksize_bits) - 1;
6493 	minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6494 			      range->minlen >> sb->s_blocksize_bits);
6495 
6496 	if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6497 	    start >= max_blks ||
6498 	    range->len < sb->s_blocksize)
6499 		return -EINVAL;
6500 	/* No point to try to trim less than discard granularity */
6501 	if (range->minlen < discard_granularity) {
6502 		minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6503 				discard_granularity >> sb->s_blocksize_bits);
6504 		if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6505 			goto out;
6506 	}
6507 	if (end >= max_blks - 1) {
6508 		end = max_blks - 1;
6509 		eof = true;
6510 	}
6511 	if (end <= first_data_blk)
6512 		goto out;
6513 	if (start < first_data_blk)
6514 		start = first_data_blk;
6515 
6516 	/* Determine first and last group to examine based on start and end */
6517 	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6518 				     &first_group, &first_cluster);
6519 	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6520 				     &last_group, &last_cluster);
6521 
6522 	/* end now represents the last cluster to discard in this group */
6523 	end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6524 	whole_group = true;
6525 
6526 	for (group = first_group; group <= last_group; group++) {
6527 		grp = ext4_get_group_info(sb, group);
6528 		/* We only do this if the grp has never been initialized */
6529 		if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6530 			ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6531 			if (ret)
6532 				break;
6533 		}
6534 
6535 		/*
6536 		 * For all the groups except the last one, last cluster will
6537 		 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6538 		 * change it for the last group, note that last_cluster is
6539 		 * already computed earlier by ext4_get_group_no_and_offset()
6540 		 */
6541 		if (group == last_group) {
6542 			end = last_cluster;
6543 			whole_group = eof ? true : end == EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6544 		}
6545 		if (grp->bb_free >= minlen) {
6546 			cnt = ext4_trim_all_free(sb, group, first_cluster,
6547 						 end, minlen, whole_group);
6548 			if (cnt < 0) {
6549 				ret = cnt;
6550 				break;
6551 			}
6552 			trimmed += cnt;
6553 		}
6554 
6555 		/*
6556 		 * For every group except the first one, we are sure
6557 		 * that the first cluster to discard will be cluster #0.
6558 		 */
6559 		first_cluster = 0;
6560 	}
6561 
6562 	if (!ret)
6563 		EXT4_SB(sb)->s_last_trim_minblks = minlen;
6564 
6565 out:
6566 	range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6567 	return ret;
6568 }
6569 
6570 /* Iterate all the free extents in the group. */
6571 int
6572 ext4_mballoc_query_range(
6573 	struct super_block		*sb,
6574 	ext4_group_t			group,
6575 	ext4_grpblk_t			start,
6576 	ext4_grpblk_t			end,
6577 	ext4_mballoc_query_range_fn	formatter,
6578 	void				*priv)
6579 {
6580 	void				*bitmap;
6581 	ext4_grpblk_t			next;
6582 	struct ext4_buddy		e4b;
6583 	int				error;
6584 
6585 	error = ext4_mb_load_buddy(sb, group, &e4b);
6586 	if (error)
6587 		return error;
6588 	bitmap = e4b.bd_bitmap;
6589 
6590 	ext4_lock_group(sb, group);
6591 
6592 	start = (e4b.bd_info->bb_first_free > start) ?
6593 		e4b.bd_info->bb_first_free : start;
6594 	if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
6595 		end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6596 
6597 	while (start <= end) {
6598 		start = mb_find_next_zero_bit(bitmap, end + 1, start);
6599 		if (start > end)
6600 			break;
6601 		next = mb_find_next_bit(bitmap, end + 1, start);
6602 
6603 		ext4_unlock_group(sb, group);
6604 		error = formatter(sb, group, start, next - start, priv);
6605 		if (error)
6606 			goto out_unload;
6607 		ext4_lock_group(sb, group);
6608 
6609 		start = next + 1;
6610 	}
6611 
6612 	ext4_unlock_group(sb, group);
6613 out_unload:
6614 	ext4_mb_unload_buddy(&e4b);
6615 
6616 	return error;
6617 }
6618