Line data Source code
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 : * The regular allocator (using the buddy cache) supports a few tunables.
131 : *
132 : * /sys/fs/ext4/<partition>/mb_min_to_scan
133 : * /sys/fs/ext4/<partition>/mb_max_to_scan
134 : * /sys/fs/ext4/<partition>/mb_order2_req
135 : *
136 : * The regular allocator uses buddy scan only if the request len is power of
137 : * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
138 : * value of s_mb_order2_reqs can be tuned via
139 : * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
140 : * stripe size (sbi->s_stripe), we try to search for contiguous block in
141 : * stripe size. This should result in better allocation on RAID setups. If
142 : * not, we search in the specific group using bitmap for best extents. The
143 : * tunable min_to_scan and max_to_scan control the behaviour here.
144 : * min_to_scan indicate how long the mballoc __must__ look for a best
145 : * extent and max_to_scan indicates how long the mballoc __can__ look for a
146 : * best extent in the found extents. Searching for the blocks starts with
147 : * the group specified as the goal value in allocation context via
148 : * ac_g_ex. Each group is first checked based on the criteria whether it
149 : * can be used for allocation. ext4_mb_good_group explains how the groups are
150 : * checked.
151 : *
152 : * Both the prealloc space are getting populated as above. So for the first
153 : * request we will hit the buddy cache which will result in this prealloc
154 : * space getting filled. The prealloc space is then later used for the
155 : * subsequent request.
156 : */
157 :
158 : /*
159 : * mballoc operates on the following data:
160 : * - on-disk bitmap
161 : * - in-core buddy (actually includes buddy and bitmap)
162 : * - preallocation descriptors (PAs)
163 : *
164 : * there are two types of preallocations:
165 : * - inode
166 : * assiged to specific inode and can be used for this inode only.
167 : * it describes part of inode's space preallocated to specific
168 : * physical blocks. any block from that preallocated can be used
169 : * independent. the descriptor just tracks number of blocks left
170 : * unused. so, before taking some block from descriptor, one must
171 : * make sure corresponded logical block isn't allocated yet. this
172 : * also means that freeing any block within descriptor's range
173 : * must discard all preallocated blocks.
174 : * - locality group
175 : * assigned to specific locality group which does not translate to
176 : * permanent set of inodes: inode can join and leave group. space
177 : * from this type of preallocation can be used for any inode. thus
178 : * it's consumed from the beginning to the end.
179 : *
180 : * relation between them can be expressed as:
181 : * in-core buddy = on-disk bitmap + preallocation descriptors
182 : *
183 : * this mean blocks mballoc considers used are:
184 : * - allocated blocks (persistent)
185 : * - preallocated blocks (non-persistent)
186 : *
187 : * consistency in mballoc world means that at any time a block is either
188 : * free or used in ALL structures. notice: "any time" should not be read
189 : * literally -- time is discrete and delimited by locks.
190 : *
191 : * to keep it simple, we don't use block numbers, instead we count number of
192 : * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
193 : *
194 : * all operations can be expressed as:
195 : * - init buddy: buddy = on-disk + PAs
196 : * - new PA: buddy += N; PA = N
197 : * - use inode PA: on-disk += N; PA -= N
198 : * - discard inode PA buddy -= on-disk - PA; PA = 0
199 : * - use locality group PA on-disk += N; PA -= N
200 : * - discard locality group PA buddy -= PA; PA = 0
201 : * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
202 : * is used in real operation because we can't know actual used
203 : * bits from PA, only from on-disk bitmap
204 : *
205 : * if we follow this strict logic, then all operations above should be atomic.
206 : * given some of them can block, we'd have to use something like semaphores
207 : * killing performance on high-end SMP hardware. let's try to relax it using
208 : * the following knowledge:
209 : * 1) if buddy is referenced, it's already initialized
210 : * 2) while block is used in buddy and the buddy is referenced,
211 : * nobody can re-allocate that block
212 : * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
213 : * bit set and PA claims same block, it's OK. IOW, one can set bit in
214 : * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
215 : * block
216 : *
217 : * so, now we're building a concurrency table:
218 : * - init buddy vs.
219 : * - new PA
220 : * blocks for PA are allocated in the buddy, buddy must be referenced
221 : * until PA is linked to allocation group to avoid concurrent buddy init
222 : * - use inode PA
223 : * we need to make sure that either on-disk bitmap or PA has uptodate data
224 : * given (3) we care that PA-=N operation doesn't interfere with init
225 : * - discard inode PA
226 : * the simplest way would be to have buddy initialized by the discard
227 : * - use locality group PA
228 : * again PA-=N must be serialized with init
229 : * - discard locality group PA
230 : * the simplest way would be to have buddy initialized by the discard
231 : * - new PA vs.
232 : * - use inode PA
233 : * i_data_sem serializes them
234 : * - discard inode PA
235 : * discard process must wait until PA isn't used by another process
236 : * - use locality group PA
237 : * some mutex should serialize them
238 : * - discard locality group PA
239 : * discard process must wait until PA isn't used by another process
240 : * - use inode PA
241 : * - use inode PA
242 : * i_data_sem or another mutex should serializes them
243 : * - discard inode PA
244 : * discard process must wait until PA isn't used by another process
245 : * - use locality group PA
246 : * nothing wrong here -- they're different PAs covering different blocks
247 : * - discard locality group PA
248 : * discard process must wait until PA isn't used by another process
249 : *
250 : * now we're ready to make few consequences:
251 : * - PA is referenced and while it is no discard is possible
252 : * - PA is referenced until block isn't marked in on-disk bitmap
253 : * - PA changes only after on-disk bitmap
254 : * - discard must not compete with init. either init is done before
255 : * any discard or they're serialized somehow
256 : * - buddy init as sum of on-disk bitmap and PAs is done atomically
257 : *
258 : * a special case when we've used PA to emptiness. no need to modify buddy
259 : * in this case, but we should care about concurrent init
260 : *
261 : */
262 :
263 : /*
264 : * Logic in few words:
265 : *
266 : * - allocation:
267 : * load group
268 : * find blocks
269 : * mark bits in on-disk bitmap
270 : * release group
271 : *
272 : * - use preallocation:
273 : * find proper PA (per-inode or group)
274 : * load group
275 : * mark bits in on-disk bitmap
276 : * release group
277 : * release PA
278 : *
279 : * - free:
280 : * load group
281 : * mark bits in on-disk bitmap
282 : * release group
283 : *
284 : * - discard preallocations in group:
285 : * mark PAs deleted
286 : * move them onto local list
287 : * load on-disk bitmap
288 : * load group
289 : * remove PA from object (inode or locality group)
290 : * mark free blocks in-core
291 : *
292 : * - discard inode's preallocations:
293 : */
294 :
295 : /*
296 : * Locking rules
297 : *
298 : * Locks:
299 : * - bitlock on a group (group)
300 : * - object (inode/locality) (object)
301 : * - per-pa lock (pa)
302 : *
303 : * Paths:
304 : * - new pa
305 : * object
306 : * group
307 : *
308 : * - find and use pa:
309 : * pa
310 : *
311 : * - release consumed pa:
312 : * pa
313 : * group
314 : * object
315 : *
316 : * - generate in-core bitmap:
317 : * group
318 : * pa
319 : *
320 : * - discard all for given object (inode, locality group):
321 : * object
322 : * pa
323 : * group
324 : *
325 : * - discard all for given group:
326 : * group
327 : * pa
328 : * group
329 : * object
330 : *
331 : */
332 : static struct kmem_cache *ext4_pspace_cachep;
333 : static struct kmem_cache *ext4_ac_cachep;
334 : static struct kmem_cache *ext4_free_data_cachep;
335 :
336 : /* We create slab caches for groupinfo data structures based on the
337 : * superblock block size. There will be one per mounted filesystem for
338 : * each unique s_blocksize_bits */
339 : #define NR_GRPINFO_CACHES 8
340 : static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
341 :
342 : static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
343 : "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
344 : "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
345 : "ext4_groupinfo_64k", "ext4_groupinfo_128k"
346 : };
347 :
348 : static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
349 : ext4_group_t group);
350 : static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
351 : ext4_group_t group);
352 : static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
353 :
354 : /*
355 : * The algorithm using this percpu seq counter goes below:
356 : * 1. We sample the percpu discard_pa_seq counter before trying for block
357 : * allocation in ext4_mb_new_blocks().
358 : * 2. We increment this percpu discard_pa_seq counter when we either allocate
359 : * or free these blocks i.e. while marking those blocks as used/free in
360 : * mb_mark_used()/mb_free_blocks().
361 : * 3. We also increment this percpu seq counter when we successfully identify
362 : * that the bb_prealloc_list is not empty and hence proceed for discarding
363 : * of those PAs inside ext4_mb_discard_group_preallocations().
364 : *
365 : * Now to make sure that the regular fast path of block allocation is not
366 : * affected, as a small optimization we only sample the percpu seq counter
367 : * on that cpu. Only when the block allocation fails and when freed blocks
368 : * found were 0, that is when we sample percpu seq counter for all cpus using
369 : * below function ext4_get_discard_pa_seq_sum(). This happens after making
370 : * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
371 : */
372 : static DEFINE_PER_CPU(u64, discard_pa_seq);
373 0 : static inline u64 ext4_get_discard_pa_seq_sum(void)
374 : {
375 0 : int __cpu;
376 0 : u64 __seq = 0;
377 :
378 0 : for_each_possible_cpu(__cpu)
379 0 : __seq += per_cpu(discard_pa_seq, __cpu);
380 0 : return __seq;
381 : }
382 :
383 57552 : static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
384 : {
385 : #if BITS_PER_LONG == 64
386 57552 : *bit += ((unsigned long) addr & 7UL) << 3;
387 57552 : addr = (void *) ((unsigned long) addr & ~7UL);
388 : #elif BITS_PER_LONG == 32
389 : *bit += ((unsigned long) addr & 3UL) << 3;
390 : addr = (void *) ((unsigned long) addr & ~3UL);
391 : #else
392 : #error "how many bits you are?!"
393 : #endif
394 57552 : return addr;
395 : }
396 :
397 34612 : static inline int mb_test_bit(int bit, void *addr)
398 : {
399 : /*
400 : * ext4_test_bit on architecture like powerpc
401 : * needs unsigned long aligned address
402 : */
403 34612 : addr = mb_correct_addr_and_bit(&bit, addr);
404 34612 : return ext4_test_bit(bit, addr);
405 : }
406 :
407 1149 : static inline void mb_set_bit(int bit, void *addr)
408 : {
409 1149 : addr = mb_correct_addr_and_bit(&bit, addr);
410 1149 : ext4_set_bit(bit, addr);
411 1149 : }
412 :
413 8965 : static inline void mb_clear_bit(int bit, void *addr)
414 : {
415 8965 : addr = mb_correct_addr_and_bit(&bit, addr);
416 8965 : ext4_clear_bit(bit, addr);
417 8965 : }
418 :
419 830 : static inline int mb_test_and_clear_bit(int bit, void *addr)
420 : {
421 830 : addr = mb_correct_addr_and_bit(&bit, addr);
422 830 : return ext4_test_and_clear_bit(bit, addr);
423 : }
424 :
425 7500 : static inline int mb_find_next_zero_bit(void *addr, int max, int start)
426 : {
427 7500 : int fix = 0, ret, tmpmax;
428 7500 : addr = mb_correct_addr_and_bit(&fix, addr);
429 7500 : tmpmax = max + fix;
430 7500 : start += fix;
431 :
432 7500 : ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
433 7500 : if (ret > max)
434 : return max;
435 : return ret;
436 : }
437 :
438 4496 : static inline int mb_find_next_bit(void *addr, int max, int start)
439 : {
440 4496 : int fix = 0, ret, tmpmax;
441 4496 : addr = mb_correct_addr_and_bit(&fix, addr);
442 4496 : tmpmax = max + fix;
443 4496 : start += fix;
444 :
445 4496 : ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
446 4496 : if (ret > max)
447 : return max;
448 : return ret;
449 : }
450 :
451 34354 : static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
452 : {
453 34354 : char *bb;
454 :
455 34354 : BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
456 34354 : BUG_ON(max == NULL);
457 :
458 34354 : if (order > e4b->bd_blkbits + 1) {
459 0 : *max = 0;
460 0 : return NULL;
461 : }
462 :
463 : /* at order 0 we see each particular block */
464 34354 : if (order == 0) {
465 4866 : *max = 1 << (e4b->bd_blkbits + 3);
466 4866 : return e4b->bd_bitmap;
467 : }
468 :
469 29488 : bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
470 29488 : *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
471 :
472 29488 : return bb;
473 : }
474 :
475 : #ifdef DOUBLE_CHECK
476 : static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
477 : int first, int count)
478 : {
479 : int i;
480 : struct super_block *sb = e4b->bd_sb;
481 :
482 : if (unlikely(e4b->bd_info->bb_bitmap == NULL))
483 : return;
484 : assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
485 : for (i = 0; i < count; i++) {
486 : if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
487 : ext4_fsblk_t blocknr;
488 :
489 : blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
490 : blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
491 : ext4_grp_locked_error(sb, e4b->bd_group,
492 : inode ? inode->i_ino : 0,
493 : blocknr,
494 : "freeing block already freed "
495 : "(bit %u)",
496 : first + i);
497 : ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
498 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
499 : }
500 : mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
501 : }
502 : }
503 :
504 : static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
505 : {
506 : int i;
507 :
508 : if (unlikely(e4b->bd_info->bb_bitmap == NULL))
509 : return;
510 : assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
511 : for (i = 0; i < count; i++) {
512 : BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
513 : mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
514 : }
515 : }
516 :
517 : static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
518 : {
519 : if (unlikely(e4b->bd_info->bb_bitmap == NULL))
520 : return;
521 : if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
522 : unsigned char *b1, *b2;
523 : int i;
524 : b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
525 : b2 = (unsigned char *) bitmap;
526 : for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
527 : if (b1[i] != b2[i]) {
528 : ext4_msg(e4b->bd_sb, KERN_ERR,
529 : "corruption in group %u "
530 : "at byte %u(%u): %x in copy != %x "
531 : "on disk/prealloc",
532 : e4b->bd_group, i, i * 8, b1[i], b2[i]);
533 : BUG();
534 : }
535 : }
536 : }
537 : }
538 :
539 : static void mb_group_bb_bitmap_alloc(struct super_block *sb,
540 : struct ext4_group_info *grp, ext4_group_t group)
541 : {
542 : struct buffer_head *bh;
543 :
544 : grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
545 : if (!grp->bb_bitmap)
546 : return;
547 :
548 : bh = ext4_read_block_bitmap(sb, group);
549 : if (IS_ERR_OR_NULL(bh)) {
550 : kfree(grp->bb_bitmap);
551 : grp->bb_bitmap = NULL;
552 : return;
553 : }
554 :
555 : memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
556 : put_bh(bh);
557 : }
558 :
559 : static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
560 : {
561 : kfree(grp->bb_bitmap);
562 : }
563 :
564 : #else
565 166 : static inline void mb_free_blocks_double(struct inode *inode,
566 : struct ext4_buddy *e4b, int first, int count)
567 : {
568 166 : return;
569 : }
570 197 : static inline void mb_mark_used_double(struct ext4_buddy *e4b,
571 : int first, int count)
572 : {
573 197 : return;
574 : }
575 : static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
576 : {
577 : return;
578 : }
579 :
580 16 : static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
581 : struct ext4_group_info *grp, ext4_group_t group)
582 : {
583 16 : return;
584 : }
585 :
586 0 : static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
587 : {
588 0 : return;
589 : }
590 : #endif
591 :
592 : #ifdef AGGRESSIVE_CHECK
593 :
594 : #define MB_CHECK_ASSERT(assert) \
595 : do { \
596 : if (!(assert)) { \
597 : printk(KERN_EMERG \
598 : "Assertion failure in %s() at %s:%d: \"%s\"\n", \
599 : function, file, line, # assert); \
600 : BUG(); \
601 : } \
602 : } while (0)
603 :
604 : static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
605 : const char *function, int line)
606 : {
607 : struct super_block *sb = e4b->bd_sb;
608 : int order = e4b->bd_blkbits + 1;
609 : int max;
610 : int max2;
611 : int i;
612 : int j;
613 : int k;
614 : int count;
615 : struct ext4_group_info *grp;
616 : int fragments = 0;
617 : int fstart;
618 : struct list_head *cur;
619 : void *buddy;
620 : void *buddy2;
621 :
622 : if (e4b->bd_info->bb_check_counter++ % 10)
623 : return 0;
624 :
625 : while (order > 1) {
626 : buddy = mb_find_buddy(e4b, order, &max);
627 : MB_CHECK_ASSERT(buddy);
628 : buddy2 = mb_find_buddy(e4b, order - 1, &max2);
629 : MB_CHECK_ASSERT(buddy2);
630 : MB_CHECK_ASSERT(buddy != buddy2);
631 : MB_CHECK_ASSERT(max * 2 == max2);
632 :
633 : count = 0;
634 : for (i = 0; i < max; i++) {
635 :
636 : if (mb_test_bit(i, buddy)) {
637 : /* only single bit in buddy2 may be 1 */
638 : if (!mb_test_bit(i << 1, buddy2)) {
639 : MB_CHECK_ASSERT(
640 : mb_test_bit((i<<1)+1, buddy2));
641 : } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
642 : MB_CHECK_ASSERT(
643 : mb_test_bit(i << 1, buddy2));
644 : }
645 : continue;
646 : }
647 :
648 : /* both bits in buddy2 must be 1 */
649 : MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
650 : MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
651 :
652 : for (j = 0; j < (1 << order); j++) {
653 : k = (i * (1 << order)) + j;
654 : MB_CHECK_ASSERT(
655 : !mb_test_bit(k, e4b->bd_bitmap));
656 : }
657 : count++;
658 : }
659 : MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
660 : order--;
661 : }
662 :
663 : fstart = -1;
664 : buddy = mb_find_buddy(e4b, 0, &max);
665 : for (i = 0; i < max; i++) {
666 : if (!mb_test_bit(i, buddy)) {
667 : MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
668 : if (fstart == -1) {
669 : fragments++;
670 : fstart = i;
671 : }
672 : continue;
673 : }
674 : fstart = -1;
675 : /* check used bits only */
676 : for (j = 0; j < e4b->bd_blkbits + 1; j++) {
677 : buddy2 = mb_find_buddy(e4b, j, &max2);
678 : k = i >> j;
679 : MB_CHECK_ASSERT(k < max2);
680 : MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
681 : }
682 : }
683 : MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
684 : MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
685 :
686 : grp = ext4_get_group_info(sb, e4b->bd_group);
687 : list_for_each(cur, &grp->bb_prealloc_list) {
688 : ext4_group_t groupnr;
689 : struct ext4_prealloc_space *pa;
690 : pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
691 : ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
692 : MB_CHECK_ASSERT(groupnr == e4b->bd_group);
693 : for (i = 0; i < pa->pa_len; i++)
694 : MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
695 : }
696 : return 0;
697 : }
698 : #undef MB_CHECK_ASSERT
699 : #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
700 : __FILE__, __func__, __LINE__)
701 : #else
702 : #define mb_check_buddy(e4b)
703 : #endif
704 :
705 : /*
706 : * Divide blocks started from @first with length @len into
707 : * smaller chunks with power of 2 blocks.
708 : * Clear the bits in bitmap which the blocks of the chunk(s) covered,
709 : * then increase bb_counters[] for corresponded chunk size.
710 : */
711 3362 : static void ext4_mb_mark_free_simple(struct super_block *sb,
712 : void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
713 : struct ext4_group_info *grp)
714 : {
715 3362 : struct ext4_sb_info *sbi = EXT4_SB(sb);
716 3362 : ext4_grpblk_t min;
717 3362 : ext4_grpblk_t max;
718 3362 : ext4_grpblk_t chunk;
719 3362 : unsigned int border;
720 :
721 3362 : BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
722 :
723 3362 : border = 2 << sb->s_blocksize_bits;
724 :
725 13508 : while (len > 0) {
726 : /* find how many blocks can be covered since this position */
727 10146 : max = ffs(first | border) - 1;
728 :
729 : /* find how many blocks of power 2 we need to mark */
730 10146 : min = fls(len) - 1;
731 :
732 10146 : if (max < min)
733 : min = max;
734 10146 : chunk = 1 << min;
735 :
736 : /* mark multiblock chunks only */
737 10146 : grp->bb_counters[min]++;
738 10146 : if (min > 0)
739 7829 : mb_clear_bit(first >> min,
740 7829 : buddy + sbi->s_mb_offsets[min]);
741 :
742 10146 : len -= chunk;
743 10146 : first += chunk;
744 : }
745 3362 : }
746 :
747 : /*
748 : * Cache the order of the largest free extent we have available in this block
749 : * group.
750 : */
751 : static void
752 379 : mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
753 : {
754 379 : int i;
755 379 : int bits;
756 :
757 379 : grp->bb_largest_free_order = -1; /* uninit */
758 :
759 379 : bits = sb->s_blocksize_bits + 1;
760 2954 : for (i = bits; i >= 0; i--) {
761 2954 : if (grp->bb_counters[i] > 0) {
762 379 : grp->bb_largest_free_order = i;
763 379 : break;
764 : }
765 : }
766 : }
767 :
768 : static noinline_for_stack
769 16 : void ext4_mb_generate_buddy(struct super_block *sb,
770 : void *buddy, void *bitmap, ext4_group_t group)
771 : {
772 16 : struct ext4_group_info *grp = ext4_get_group_info(sb, group);
773 16 : struct ext4_sb_info *sbi = EXT4_SB(sb);
774 16 : ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
775 16 : ext4_grpblk_t i = 0;
776 16 : ext4_grpblk_t first;
777 16 : ext4_grpblk_t len;
778 16 : unsigned free = 0;
779 16 : unsigned fragments = 0;
780 16 : unsigned long long period = get_cycles();
781 :
782 : /* initialize buddy from bitmap which is aggregation
783 : * of on-disk bitmap and preallocations */
784 16 : i = mb_find_next_zero_bit(bitmap, max, 0);
785 16 : grp->bb_first_free = i;
786 4510 : while (i < max) {
787 4494 : fragments++;
788 4494 : first = i;
789 4494 : i = mb_find_next_bit(bitmap, max, i);
790 4494 : len = i - first;
791 4494 : free += len;
792 4494 : if (len > 1)
793 3362 : ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
794 : else
795 1132 : grp->bb_counters[0]++;
796 4494 : if (i < max)
797 4486 : i = mb_find_next_zero_bit(bitmap, max, i);
798 : }
799 16 : grp->bb_fragments = fragments;
800 :
801 16 : if (free != grp->bb_free) {
802 0 : ext4_grp_locked_error(sb, group, 0, 0,
803 : "block bitmap and bg descriptor "
804 : "inconsistent: %u vs %u free clusters",
805 : free, grp->bb_free);
806 : /*
807 : * If we intend to continue, we consider group descriptor
808 : * corrupt and update bb_free using bitmap value
809 : */
810 0 : grp->bb_free = free;
811 0 : ext4_mark_group_bitmap_corrupted(sb, group,
812 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
813 : }
814 16 : mb_set_largest_free_order(sb, grp);
815 :
816 16 : clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
817 :
818 16 : period = get_cycles() - period;
819 16 : spin_lock(&sbi->s_bal_lock);
820 16 : sbi->s_mb_buddies_generated++;
821 16 : sbi->s_mb_generation_time += period;
822 16 : spin_unlock(&sbi->s_bal_lock);
823 16 : }
824 :
825 : /* The buddy information is attached the buddy cache inode
826 : * for convenience. The information regarding each group
827 : * is loaded via ext4_mb_load_buddy. The information involve
828 : * block bitmap and buddy information. The information are
829 : * stored in the inode as
830 : *
831 : * { page }
832 : * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
833 : *
834 : *
835 : * one block each for bitmap and buddy information.
836 : * So for each group we take up 2 blocks. A page can
837 : * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
838 : * So it can have information regarding groups_per_page which
839 : * is blocks_per_page/2
840 : *
841 : * Locking note: This routine takes the block group lock of all groups
842 : * for this page; do not hold this lock when calling this routine!
843 : */
844 :
845 32 : static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
846 : {
847 32 : ext4_group_t ngroups;
848 32 : int blocksize;
849 32 : int blocks_per_page;
850 32 : int groups_per_page;
851 32 : int err = 0;
852 32 : int i;
853 32 : ext4_group_t first_group, group;
854 32 : int first_block;
855 32 : struct super_block *sb;
856 32 : struct buffer_head *bhs;
857 32 : struct buffer_head **bh = NULL;
858 32 : struct inode *inode;
859 32 : char *data;
860 32 : char *bitmap;
861 32 : struct ext4_group_info *grinfo;
862 :
863 32 : inode = page->mapping->host;
864 32 : sb = inode->i_sb;
865 32 : ngroups = ext4_get_groups_count(sb);
866 32 : blocksize = i_blocksize(inode);
867 32 : blocks_per_page = PAGE_SIZE / blocksize;
868 :
869 32 : mb_debug(sb, "init page %lu\n", page->index);
870 :
871 32 : groups_per_page = blocks_per_page >> 1;
872 32 : if (groups_per_page == 0)
873 : groups_per_page = 1;
874 :
875 : /* allocate buffer_heads to read bitmaps */
876 0 : if (groups_per_page > 1) {
877 0 : i = sizeof(struct buffer_head *) * groups_per_page;
878 0 : bh = kzalloc(i, gfp);
879 0 : if (bh == NULL) {
880 0 : err = -ENOMEM;
881 0 : goto out;
882 : }
883 : } else
884 : bh = &bhs;
885 :
886 32 : first_group = page->index * blocks_per_page / 2;
887 :
888 : /* read all groups the page covers into the cache */
889 64 : for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
890 32 : if (group >= ngroups)
891 : break;
892 :
893 32 : grinfo = ext4_get_group_info(sb, group);
894 : /*
895 : * If page is uptodate then we came here after online resize
896 : * which added some new uninitialized group info structs, so
897 : * we must skip all initialized uptodate buddies on the page,
898 : * which may be currently in use by an allocating task.
899 : */
900 32 : if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
901 0 : bh[i] = NULL;
902 0 : continue;
903 : }
904 32 : bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
905 32 : if (IS_ERR(bh[i])) {
906 0 : err = PTR_ERR(bh[i]);
907 0 : bh[i] = NULL;
908 0 : goto out;
909 : }
910 : mb_debug(sb, "read bitmap for group %u\n", group);
911 : }
912 :
913 : /* wait for I/O completion */
914 64 : for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
915 32 : int err2;
916 :
917 32 : if (!bh[i])
918 0 : continue;
919 32 : err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
920 32 : if (!err)
921 32 : err = err2;
922 : }
923 :
924 32 : first_block = page->index * blocks_per_page;
925 64 : for (i = 0; i < blocks_per_page; i++) {
926 32 : group = (first_block + i) >> 1;
927 32 : if (group >= ngroups)
928 : break;
929 :
930 32 : if (!bh[group - first_group])
931 : /* skip initialized uptodate buddy */
932 0 : continue;
933 :
934 32 : if (!buffer_verified(bh[group - first_group]))
935 : /* Skip faulty bitmaps */
936 0 : continue;
937 32 : err = 0;
938 :
939 : /*
940 : * data carry information regarding this
941 : * particular group in the format specified
942 : * above
943 : *
944 : */
945 32 : data = page_address(page) + (i * blocksize);
946 32 : bitmap = bh[group - first_group]->b_data;
947 :
948 : /*
949 : * We place the buddy block and bitmap block
950 : * close together
951 : */
952 32 : if ((first_block + i) & 1) {
953 : /* this is block of buddy */
954 16 : BUG_ON(incore == NULL);
955 16 : mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
956 : group, page->index, i * blocksize);
957 16 : trace_ext4_mb_buddy_bitmap_load(sb, group);
958 16 : grinfo = ext4_get_group_info(sb, group);
959 16 : grinfo->bb_fragments = 0;
960 16 : memset(grinfo->bb_counters, 0,
961 : sizeof(*grinfo->bb_counters) *
962 16 : (sb->s_blocksize_bits+2));
963 : /*
964 : * incore got set to the group block bitmap below
965 : */
966 16 : ext4_lock_group(sb, group);
967 : /* init the buddy */
968 16 : memset(data, 0xff, blocksize);
969 16 : ext4_mb_generate_buddy(sb, data, incore, group);
970 16 : ext4_unlock_group(sb, group);
971 16 : incore = NULL;
972 : } else {
973 : /* this is block of bitmap */
974 16 : BUG_ON(incore != NULL);
975 16 : mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
976 : group, page->index, i * blocksize);
977 16 : trace_ext4_mb_bitmap_load(sb, group);
978 :
979 : /* see comments in ext4_mb_put_pa() */
980 16 : ext4_lock_group(sb, group);
981 16 : memcpy(data, bitmap, blocksize);
982 :
983 : /* mark all preallocated blks used in in-core bitmap */
984 16 : ext4_mb_generate_from_pa(sb, data, group);
985 16 : ext4_mb_generate_from_freelist(sb, data, group);
986 16 : ext4_unlock_group(sb, group);
987 :
988 : /* set incore so that the buddy information can be
989 : * generated using this
990 : */
991 16 : incore = data;
992 : }
993 : }
994 32 : SetPageUptodate(page);
995 :
996 32 : out:
997 32 : if (bh) {
998 64 : for (i = 0; i < groups_per_page; i++)
999 64 : brelse(bh[i]);
1000 32 : if (bh != &bhs)
1001 0 : kfree(bh);
1002 : }
1003 32 : return err;
1004 : }
1005 :
1006 : /*
1007 : * Lock the buddy and bitmap pages. This make sure other parallel init_group
1008 : * on the same buddy page doesn't happen whild holding the buddy page lock.
1009 : * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1010 : * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1011 : */
1012 16 : static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1013 : ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1014 : {
1015 16 : struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1016 16 : int block, pnum, poff;
1017 16 : int blocks_per_page;
1018 16 : struct page *page;
1019 :
1020 16 : e4b->bd_buddy_page = NULL;
1021 16 : e4b->bd_bitmap_page = NULL;
1022 :
1023 16 : blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1024 : /*
1025 : * the buddy cache inode stores the block bitmap
1026 : * and buddy information in consecutive blocks.
1027 : * So for each group we need two blocks.
1028 : */
1029 16 : block = group * 2;
1030 16 : pnum = block / blocks_per_page;
1031 16 : poff = block % blocks_per_page;
1032 16 : page = find_or_create_page(inode->i_mapping, pnum, gfp);
1033 16 : if (!page)
1034 : return -ENOMEM;
1035 16 : BUG_ON(page->mapping != inode->i_mapping);
1036 16 : e4b->bd_bitmap_page = page;
1037 16 : e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1038 :
1039 16 : if (blocks_per_page >= 2) {
1040 : /* buddy and bitmap are on the same page */
1041 : return 0;
1042 : }
1043 :
1044 16 : block++;
1045 16 : pnum = block / blocks_per_page;
1046 16 : page = find_or_create_page(inode->i_mapping, pnum, gfp);
1047 16 : if (!page)
1048 : return -ENOMEM;
1049 16 : BUG_ON(page->mapping != inode->i_mapping);
1050 16 : e4b->bd_buddy_page = page;
1051 16 : return 0;
1052 : }
1053 :
1054 16 : static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1055 : {
1056 16 : if (e4b->bd_bitmap_page) {
1057 16 : unlock_page(e4b->bd_bitmap_page);
1058 16 : put_page(e4b->bd_bitmap_page);
1059 : }
1060 16 : if (e4b->bd_buddy_page) {
1061 16 : unlock_page(e4b->bd_buddy_page);
1062 16 : put_page(e4b->bd_buddy_page);
1063 : }
1064 16 : }
1065 :
1066 : /*
1067 : * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1068 : * block group lock of all groups for this page; do not hold the BG lock when
1069 : * calling this routine!
1070 : */
1071 : static noinline_for_stack
1072 16 : int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1073 : {
1074 :
1075 16 : struct ext4_group_info *this_grp;
1076 16 : struct ext4_buddy e4b;
1077 16 : struct page *page;
1078 16 : int ret = 0;
1079 :
1080 16 : might_sleep();
1081 16 : mb_debug(sb, "init group %u\n", group);
1082 16 : this_grp = ext4_get_group_info(sb, group);
1083 : /*
1084 : * This ensures that we don't reinit the buddy cache
1085 : * page which map to the group from which we are already
1086 : * allocating. If we are looking at the buddy cache we would
1087 : * have taken a reference using ext4_mb_load_buddy and that
1088 : * would have pinned buddy page to page cache.
1089 : * The call to ext4_mb_get_buddy_page_lock will mark the
1090 : * page accessed.
1091 : */
1092 16 : ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1093 16 : if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1094 : /*
1095 : * somebody initialized the group
1096 : * return without doing anything
1097 : */
1098 0 : goto err;
1099 : }
1100 :
1101 16 : page = e4b.bd_bitmap_page;
1102 16 : ret = ext4_mb_init_cache(page, NULL, gfp);
1103 16 : if (ret)
1104 0 : goto err;
1105 16 : if (!PageUptodate(page)) {
1106 0 : ret = -EIO;
1107 0 : goto err;
1108 : }
1109 :
1110 16 : if (e4b.bd_buddy_page == NULL) {
1111 : /*
1112 : * If both the bitmap and buddy are in
1113 : * the same page we don't need to force
1114 : * init the buddy
1115 : */
1116 0 : ret = 0;
1117 0 : goto err;
1118 : }
1119 : /* init buddy cache */
1120 16 : page = e4b.bd_buddy_page;
1121 16 : ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1122 16 : if (ret)
1123 0 : goto err;
1124 16 : if (!PageUptodate(page)) {
1125 0 : ret = -EIO;
1126 0 : goto err;
1127 : }
1128 16 : err:
1129 16 : ext4_mb_put_buddy_page_lock(&e4b);
1130 16 : return ret;
1131 : }
1132 :
1133 : /*
1134 : * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1135 : * block group lock of all groups for this page; do not hold the BG lock when
1136 : * calling this routine!
1137 : */
1138 : static noinline_for_stack int
1139 585 : ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1140 : struct ext4_buddy *e4b, gfp_t gfp)
1141 : {
1142 585 : int blocks_per_page;
1143 585 : int block;
1144 585 : int pnum;
1145 585 : int poff;
1146 585 : struct page *page;
1147 585 : int ret;
1148 585 : struct ext4_group_info *grp;
1149 585 : struct ext4_sb_info *sbi = EXT4_SB(sb);
1150 585 : struct inode *inode = sbi->s_buddy_cache;
1151 :
1152 585 : might_sleep();
1153 585 : mb_debug(sb, "load group %u\n", group);
1154 :
1155 585 : blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1156 585 : grp = ext4_get_group_info(sb, group);
1157 :
1158 585 : e4b->bd_blkbits = sb->s_blocksize_bits;
1159 585 : e4b->bd_info = grp;
1160 585 : e4b->bd_sb = sb;
1161 585 : e4b->bd_group = group;
1162 585 : e4b->bd_buddy_page = NULL;
1163 585 : e4b->bd_bitmap_page = NULL;
1164 :
1165 585 : if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1166 : /*
1167 : * we need full data about the group
1168 : * to make a good selection
1169 : */
1170 1 : ret = ext4_mb_init_group(sb, group, gfp);
1171 1 : if (ret)
1172 : return ret;
1173 : }
1174 :
1175 : /*
1176 : * the buddy cache inode stores the block bitmap
1177 : * and buddy information in consecutive blocks.
1178 : * So for each group we need two blocks.
1179 : */
1180 585 : block = group * 2;
1181 585 : pnum = block / blocks_per_page;
1182 585 : poff = block % blocks_per_page;
1183 :
1184 : /* we could use find_or_create_page(), but it locks page
1185 : * what we'd like to avoid in fast path ... */
1186 585 : page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1187 585 : if (page == NULL || !PageUptodate(page)) {
1188 0 : if (page)
1189 : /*
1190 : * drop the page reference and try
1191 : * to get the page with lock. If we
1192 : * are not uptodate that implies
1193 : * somebody just created the page but
1194 : * is yet to initialize the same. So
1195 : * wait for it to initialize.
1196 : */
1197 0 : put_page(page);
1198 0 : page = find_or_create_page(inode->i_mapping, pnum, gfp);
1199 0 : if (page) {
1200 0 : BUG_ON(page->mapping != inode->i_mapping);
1201 0 : if (!PageUptodate(page)) {
1202 0 : ret = ext4_mb_init_cache(page, NULL, gfp);
1203 0 : if (ret) {
1204 0 : unlock_page(page);
1205 0 : goto err;
1206 : }
1207 0 : mb_cmp_bitmaps(e4b, page_address(page) +
1208 0 : (poff * sb->s_blocksize));
1209 : }
1210 0 : unlock_page(page);
1211 : }
1212 : }
1213 585 : if (page == NULL) {
1214 0 : ret = -ENOMEM;
1215 0 : goto err;
1216 : }
1217 585 : if (!PageUptodate(page)) {
1218 0 : ret = -EIO;
1219 0 : goto err;
1220 : }
1221 :
1222 : /* Pages marked accessed already */
1223 585 : e4b->bd_bitmap_page = page;
1224 585 : e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1225 :
1226 585 : block++;
1227 585 : pnum = block / blocks_per_page;
1228 585 : poff = block % blocks_per_page;
1229 :
1230 585 : page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1231 585 : if (page == NULL || !PageUptodate(page)) {
1232 0 : if (page)
1233 0 : put_page(page);
1234 0 : page = find_or_create_page(inode->i_mapping, pnum, gfp);
1235 0 : if (page) {
1236 0 : BUG_ON(page->mapping != inode->i_mapping);
1237 0 : if (!PageUptodate(page)) {
1238 0 : ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1239 : gfp);
1240 0 : if (ret) {
1241 0 : unlock_page(page);
1242 0 : goto err;
1243 : }
1244 : }
1245 0 : unlock_page(page);
1246 : }
1247 : }
1248 585 : if (page == NULL) {
1249 0 : ret = -ENOMEM;
1250 0 : goto err;
1251 : }
1252 585 : if (!PageUptodate(page)) {
1253 0 : ret = -EIO;
1254 0 : goto err;
1255 : }
1256 :
1257 : /* Pages marked accessed already */
1258 585 : e4b->bd_buddy_page = page;
1259 585 : e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1260 :
1261 585 : return 0;
1262 :
1263 0 : err:
1264 0 : if (page)
1265 0 : put_page(page);
1266 0 : if (e4b->bd_bitmap_page)
1267 0 : put_page(e4b->bd_bitmap_page);
1268 0 : if (e4b->bd_buddy_page)
1269 0 : put_page(e4b->bd_buddy_page);
1270 0 : e4b->bd_buddy = NULL;
1271 0 : e4b->bd_bitmap = NULL;
1272 0 : return ret;
1273 : }
1274 :
1275 377 : static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1276 : struct ext4_buddy *e4b)
1277 : {
1278 377 : return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1279 : }
1280 :
1281 585 : static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1282 : {
1283 585 : if (e4b->bd_bitmap_page)
1284 585 : put_page(e4b->bd_bitmap_page);
1285 585 : if (e4b->bd_buddy_page)
1286 585 : put_page(e4b->bd_buddy_page);
1287 585 : }
1288 :
1289 :
1290 3640 : static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1291 : {
1292 3640 : int order = 1, max;
1293 3640 : void *bb;
1294 :
1295 3640 : BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1296 3640 : BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1297 :
1298 30546 : while (order <= e4b->bd_blkbits + 1) {
1299 28700 : bb = mb_find_buddy(e4b, order, &max);
1300 28700 : if (!mb_test_bit(block >> order, bb)) {
1301 : /* this block is part of buddy of order 'order' */
1302 1794 : return order;
1303 : }
1304 26906 : order++;
1305 : }
1306 : return 0;
1307 : }
1308 :
1309 206 : static void mb_clear_bits(void *bm, int cur, int len)
1310 : {
1311 206 : __u32 *addr;
1312 :
1313 206 : len = cur + len;
1314 1013 : while (cur < len) {
1315 807 : if ((cur & 31) == 0 && (len - cur) >= 32) {
1316 : /* fast path: clear whole word at once */
1317 16 : addr = bm + (cur >> 3);
1318 16 : *addr = 0;
1319 16 : cur += 32;
1320 16 : continue;
1321 : }
1322 791 : mb_clear_bit(cur, bm);
1323 791 : cur++;
1324 : }
1325 206 : }
1326 :
1327 : /* clear bits in given range
1328 : * will return first found zero bit if any, -1 otherwise
1329 : */
1330 166 : static int mb_test_and_clear_bits(void *bm, int cur, int len)
1331 : {
1332 166 : __u32 *addr;
1333 166 : int zero_bit = -1;
1334 :
1335 166 : len = cur + len;
1336 1012 : while (cur < len) {
1337 846 : if ((cur & 31) == 0 && (len - cur) >= 32) {
1338 : /* fast path: clear whole word at once */
1339 16 : addr = bm + (cur >> 3);
1340 16 : if (*addr != (__u32)(-1) && zero_bit == -1)
1341 0 : zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1342 16 : *addr = 0;
1343 16 : cur += 32;
1344 16 : continue;
1345 : }
1346 830 : if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1347 0 : zero_bit = cur;
1348 830 : cur++;
1349 : }
1350 :
1351 166 : return zero_bit;
1352 : }
1353 :
1354 442 : void ext4_set_bits(void *bm, int cur, int len)
1355 : {
1356 442 : __u32 *addr;
1357 :
1358 442 : len = cur + len;
1359 1270 : while (cur < len) {
1360 828 : if ((cur & 31) == 0 && (len - cur) >= 32) {
1361 : /* fast path: set whole word at once */
1362 69 : addr = bm + (cur >> 3);
1363 69 : *addr = 0xffffffff;
1364 69 : cur += 32;
1365 69 : continue;
1366 : }
1367 759 : mb_set_bit(cur, bm);
1368 759 : cur++;
1369 : }
1370 442 : }
1371 :
1372 237 : static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1373 : {
1374 237 : if (mb_test_bit(*bit + side, bitmap)) {
1375 159 : mb_clear_bit(*bit, bitmap);
1376 159 : (*bit) -= side;
1377 159 : return 1;
1378 : }
1379 : else {
1380 78 : (*bit) += side;
1381 78 : mb_set_bit(*bit, bitmap);
1382 78 : return -1;
1383 : }
1384 : }
1385 :
1386 104 : static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1387 : {
1388 104 : int max;
1389 104 : int order = 1;
1390 104 : void *buddy = mb_find_buddy(e4b, order, &max);
1391 :
1392 293 : while (buddy) {
1393 293 : void *buddy2;
1394 :
1395 : /* Bits in range [first; last] are known to be set since
1396 : * corresponding blocks were allocated. Bits in range
1397 : * (first; last) will stay set because they form buddies on
1398 : * upper layer. We just deal with borders if they don't
1399 : * align with upper layer and then go up.
1400 : * Releasing entire group is all about clearing
1401 : * single bit of highest order buddy.
1402 : */
1403 :
1404 : /* Example:
1405 : * ---------------------------------
1406 : * | 1 | 1 | 1 | 1 |
1407 : * ---------------------------------
1408 : * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1409 : * ---------------------------------
1410 : * 0 1 2 3 4 5 6 7
1411 : * \_____________________/
1412 : *
1413 : * Neither [1] nor [6] is aligned to above layer.
1414 : * Left neighbour [0] is free, so mark it busy,
1415 : * decrease bb_counters and extend range to
1416 : * [0; 6]
1417 : * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1418 : * mark [6] free, increase bb_counters and shrink range to
1419 : * [0; 5].
1420 : * Then shift range to [0; 2], go up and do the same.
1421 : */
1422 :
1423 :
1424 293 : if (first & 1)
1425 110 : e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1426 293 : if (!(last & 1))
1427 127 : e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1428 293 : if (first > last)
1429 : break;
1430 189 : order++;
1431 :
1432 189 : if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1433 0 : mb_clear_bits(buddy, first, last - first + 1);
1434 0 : e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1435 0 : break;
1436 : }
1437 189 : first >>= 1;
1438 189 : last >>= 1;
1439 189 : buddy = buddy2;
1440 : }
1441 104 : }
1442 :
1443 166 : static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1444 : int first, int count)
1445 : {
1446 166 : int left_is_free = 0;
1447 166 : int right_is_free = 0;
1448 166 : int block;
1449 166 : int last = first + count - 1;
1450 166 : struct super_block *sb = e4b->bd_sb;
1451 :
1452 166 : if (WARN_ON(count == 0))
1453 : return;
1454 166 : BUG_ON(last >= (sb->s_blocksize << 3));
1455 166 : assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1456 : /* Don't bother if the block group is corrupt. */
1457 166 : if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1458 : return;
1459 :
1460 166 : mb_check_buddy(e4b);
1461 166 : mb_free_blocks_double(inode, e4b, first, count);
1462 :
1463 166 : this_cpu_inc(discard_pa_seq);
1464 166 : e4b->bd_info->bb_free += count;
1465 166 : if (first < e4b->bd_info->bb_first_free)
1466 10 : e4b->bd_info->bb_first_free = first;
1467 :
1468 : /* access memory sequentially: check left neighbour,
1469 : * clear range and then check right neighbour
1470 : */
1471 166 : if (first != 0)
1472 166 : left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1473 166 : block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1474 166 : if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1475 166 : right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1476 :
1477 166 : if (unlikely(block != -1)) {
1478 0 : struct ext4_sb_info *sbi = EXT4_SB(sb);
1479 0 : ext4_fsblk_t blocknr;
1480 :
1481 0 : blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1482 0 : blocknr += EXT4_C2B(sbi, block);
1483 0 : if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) {
1484 0 : ext4_grp_locked_error(sb, e4b->bd_group,
1485 : inode ? inode->i_ino : 0,
1486 : blocknr,
1487 : "freeing already freed block (bit %u); block bitmap corrupt.",
1488 : block);
1489 0 : ext4_mark_group_bitmap_corrupted(
1490 : sb, e4b->bd_group,
1491 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1492 : }
1493 0 : goto done;
1494 : }
1495 :
1496 : /* let's maintain fragments counter */
1497 166 : if (left_is_free && right_is_free)
1498 20 : e4b->bd_info->bb_fragments--;
1499 146 : else if (!left_is_free && !right_is_free)
1500 79 : e4b->bd_info->bb_fragments++;
1501 :
1502 : /* buddy[0] == bd_bitmap is a special case, so handle
1503 : * it right away and let mb_buddy_mark_free stay free of
1504 : * zero order checks.
1505 : * Check if neighbours are to be coaleasced,
1506 : * adjust bitmap bb_counters and borders appropriately.
1507 : */
1508 166 : if (first & 1) {
1509 65 : first += !left_is_free;
1510 104 : e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1511 : }
1512 166 : if (!(last & 1)) {
1513 87 : last -= !right_is_free;
1514 151 : e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1515 : }
1516 :
1517 166 : if (first <= last)
1518 104 : mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1519 :
1520 62 : done:
1521 166 : mb_set_largest_free_order(sb, e4b->bd_info);
1522 166 : mb_check_buddy(e4b);
1523 : }
1524 :
1525 3071 : static int mb_find_extent(struct ext4_buddy *e4b, int block,
1526 : int needed, struct ext4_free_extent *ex)
1527 : {
1528 3071 : int next = block;
1529 3071 : int max, order;
1530 3071 : void *buddy;
1531 :
1532 3071 : assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1533 3071 : BUG_ON(ex == NULL);
1534 :
1535 3071 : buddy = mb_find_buddy(e4b, 0, &max);
1536 3071 : BUG_ON(buddy == NULL);
1537 3071 : BUG_ON(block >= max);
1538 3071 : if (mb_test_bit(block, buddy)) {
1539 8 : ex->fe_len = 0;
1540 8 : ex->fe_start = 0;
1541 8 : ex->fe_group = 0;
1542 8 : return 0;
1543 : }
1544 :
1545 : /* find actual order */
1546 3063 : order = mb_find_order_for_block(e4b, block);
1547 3063 : block = block >> order;
1548 :
1549 3063 : ex->fe_len = 1 << order;
1550 3063 : ex->fe_start = block << order;
1551 3063 : ex->fe_group = e4b->bd_group;
1552 :
1553 : /* calc difference from given start */
1554 3063 : next = next - ex->fe_start;
1555 3063 : ex->fe_len -= next;
1556 3063 : ex->fe_start += next;
1557 :
1558 5206 : while (needed > ex->fe_len &&
1559 1878 : mb_find_buddy(e4b, order, &max)) {
1560 :
1561 1878 : if (block + 1 >= max)
1562 : break;
1563 :
1564 1878 : next = (block + 1) * (1 << order);
1565 1878 : if (mb_test_bit(next, e4b->bd_bitmap))
1566 : break;
1567 :
1568 265 : order = mb_find_order_for_block(e4b, next);
1569 :
1570 265 : block = next >> order;
1571 265 : ex->fe_len += 1 << order;
1572 : }
1573 :
1574 3063 : if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
1575 : /* Should never happen! (but apparently sometimes does?!?) */
1576 0 : WARN_ON(1);
1577 0 : ext4_error(e4b->bd_sb, "corruption or bug in mb_find_extent "
1578 : "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
1579 : block, order, needed, ex->fe_group, ex->fe_start,
1580 : ex->fe_len, ex->fe_logical);
1581 0 : ex->fe_len = 0;
1582 0 : ex->fe_start = 0;
1583 0 : ex->fe_group = 0;
1584 : }
1585 3063 : return ex->fe_len;
1586 : }
1587 :
1588 197 : static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1589 : {
1590 197 : int ord;
1591 197 : int mlen = 0;
1592 197 : int max = 0;
1593 197 : int cur;
1594 197 : int start = ex->fe_start;
1595 197 : int len = ex->fe_len;
1596 197 : unsigned ret = 0;
1597 197 : int len0 = len;
1598 197 : void *buddy;
1599 :
1600 197 : BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1601 197 : BUG_ON(e4b->bd_group != ex->fe_group);
1602 197 : assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1603 197 : mb_check_buddy(e4b);
1604 197 : mb_mark_used_double(e4b, start, len);
1605 :
1606 197 : this_cpu_inc(discard_pa_seq);
1607 197 : e4b->bd_info->bb_free -= len;
1608 197 : if (e4b->bd_info->bb_first_free == start)
1609 6 : e4b->bd_info->bb_first_free += len;
1610 :
1611 : /* let's maintain fragments counter */
1612 197 : if (start != 0)
1613 197 : mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1614 197 : if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1615 197 : max = !mb_test_bit(start + len, e4b->bd_bitmap);
1616 197 : if (mlen && max)
1617 16 : e4b->bd_info->bb_fragments++;
1618 181 : else if (!mlen && !max)
1619 71 : e4b->bd_info->bb_fragments--;
1620 :
1621 : /* let's maintain buddy itself */
1622 509 : while (len) {
1623 312 : ord = mb_find_order_for_block(e4b, start);
1624 :
1625 312 : if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1626 : /* the whole chunk may be allocated at once! */
1627 219 : mlen = 1 << ord;
1628 219 : buddy = mb_find_buddy(e4b, ord, &max);
1629 219 : BUG_ON((start >> ord) >= max);
1630 219 : mb_set_bit(start >> ord, buddy);
1631 219 : e4b->bd_info->bb_counters[ord]--;
1632 219 : start += mlen;
1633 219 : len -= mlen;
1634 219 : BUG_ON(len < 0);
1635 219 : continue;
1636 : }
1637 :
1638 : /* store for history */
1639 93 : if (ret == 0)
1640 66 : ret = len | (ord << 16);
1641 :
1642 : /* we have to split large buddy */
1643 93 : BUG_ON(ord <= 0);
1644 93 : buddy = mb_find_buddy(e4b, ord, &max);
1645 93 : mb_set_bit(start >> ord, buddy);
1646 93 : e4b->bd_info->bb_counters[ord]--;
1647 :
1648 93 : ord--;
1649 93 : cur = (start >> ord) & ~1U;
1650 93 : buddy = mb_find_buddy(e4b, ord, &max);
1651 93 : mb_clear_bit(cur, buddy);
1652 93 : mb_clear_bit(cur + 1, buddy);
1653 93 : e4b->bd_info->bb_counters[ord]++;
1654 93 : e4b->bd_info->bb_counters[ord]++;
1655 : }
1656 197 : mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1657 :
1658 197 : ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1659 197 : mb_check_buddy(e4b);
1660 :
1661 197 : return ret;
1662 : }
1663 :
1664 : /*
1665 : * Must be called under group lock!
1666 : */
1667 197 : static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1668 : struct ext4_buddy *e4b)
1669 : {
1670 197 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1671 197 : int ret;
1672 :
1673 197 : BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1674 197 : BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1675 :
1676 197 : ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1677 197 : ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1678 197 : ret = mb_mark_used(e4b, &ac->ac_b_ex);
1679 :
1680 : /* preallocation can change ac_b_ex, thus we store actually
1681 : * allocated blocks for history */
1682 197 : ac->ac_f_ex = ac->ac_b_ex;
1683 :
1684 197 : ac->ac_status = AC_STATUS_FOUND;
1685 197 : ac->ac_tail = ret & 0xffff;
1686 197 : ac->ac_buddy = ret >> 16;
1687 :
1688 : /*
1689 : * take the page reference. We want the page to be pinned
1690 : * so that we don't get a ext4_mb_init_cache_call for this
1691 : * group until we update the bitmap. That would mean we
1692 : * double allocate blocks. The reference is dropped
1693 : * in ext4_mb_release_context
1694 : */
1695 197 : ac->ac_bitmap_page = e4b->bd_bitmap_page;
1696 197 : get_page(ac->ac_bitmap_page);
1697 197 : ac->ac_buddy_page = e4b->bd_buddy_page;
1698 197 : get_page(ac->ac_buddy_page);
1699 : /* store last allocated for subsequent stream allocation */
1700 197 : if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1701 9 : spin_lock(&sbi->s_md_lock);
1702 9 : sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1703 9 : sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1704 9 : spin_unlock(&sbi->s_md_lock);
1705 : }
1706 : /*
1707 : * As we've just preallocated more space than
1708 : * user requested originally, we store allocated
1709 : * space in a special descriptor.
1710 : */
1711 197 : if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
1712 6 : ext4_mb_new_preallocation(ac);
1713 :
1714 197 : }
1715 :
1716 2900 : static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1717 : struct ext4_buddy *e4b,
1718 : int finish_group)
1719 : {
1720 2900 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1721 2900 : struct ext4_free_extent *bex = &ac->ac_b_ex;
1722 2900 : struct ext4_free_extent *gex = &ac->ac_g_ex;
1723 2900 : struct ext4_free_extent ex;
1724 2900 : int max;
1725 :
1726 2900 : if (ac->ac_status == AC_STATUS_FOUND)
1727 1856 : return;
1728 : /*
1729 : * We don't want to scan for a whole year
1730 : */
1731 2718 : if (ac->ac_found > sbi->s_mb_max_to_scan &&
1732 16 : !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1733 16 : ac->ac_status = AC_STATUS_BREAK;
1734 16 : return;
1735 : }
1736 :
1737 : /*
1738 : * Haven't found good chunk so far, let's continue
1739 : */
1740 2702 : if (bex->fe_len < gex->fe_len)
1741 : return;
1742 :
1743 1110 : if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1744 66 : && bex->fe_group == e4b->bd_group) {
1745 : /* recheck chunk's availability - we don't know
1746 : * when it was found (within this lock-unlock
1747 : * period or not) */
1748 66 : max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1749 66 : if (max >= gex->fe_len) {
1750 66 : ext4_mb_use_best_found(ac, e4b);
1751 66 : return;
1752 : }
1753 : }
1754 : }
1755 :
1756 : /*
1757 : * The routine checks whether found extent is good enough. If it is,
1758 : * then the extent gets marked used and flag is set to the context
1759 : * to stop scanning. Otherwise, the extent is compared with the
1760 : * previous found extent and if new one is better, then it's stored
1761 : * in the context. Later, the best found extent will be used, if
1762 : * mballoc can't find good enough extent.
1763 : *
1764 : * FIXME: real allocation policy is to be designed yet!
1765 : */
1766 2989 : static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1767 : struct ext4_free_extent *ex,
1768 : struct ext4_buddy *e4b)
1769 : {
1770 2989 : struct ext4_free_extent *bex = &ac->ac_b_ex;
1771 2989 : struct ext4_free_extent *gex = &ac->ac_g_ex;
1772 :
1773 2989 : BUG_ON(ex->fe_len <= 0);
1774 2989 : BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1775 2989 : BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1776 2989 : BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1777 :
1778 2989 : ac->ac_found++;
1779 :
1780 : /*
1781 : * The special case - take what you catch first
1782 : */
1783 2989 : if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1784 0 : *bex = *ex;
1785 0 : ext4_mb_use_best_found(ac, e4b);
1786 0 : return;
1787 : }
1788 :
1789 : /*
1790 : * Let's check whether the chuck is good enough
1791 : */
1792 2989 : if (ex->fe_len == gex->fe_len) {
1793 116 : *bex = *ex;
1794 116 : ext4_mb_use_best_found(ac, e4b);
1795 116 : return;
1796 : }
1797 :
1798 : /*
1799 : * If this is first found extent, just store it in the context
1800 : */
1801 2873 : if (bex->fe_len == 0) {
1802 163 : *bex = *ex;
1803 163 : return;
1804 : }
1805 :
1806 : /*
1807 : * If new found extent is better, store it in the context
1808 : */
1809 2710 : if (bex->fe_len < gex->fe_len) {
1810 : /* if the request isn't satisfied, any found extent
1811 : * larger than previous best one is better */
1812 1603 : if (ex->fe_len > bex->fe_len)
1813 16 : *bex = *ex;
1814 1107 : } else if (ex->fe_len > gex->fe_len) {
1815 : /* if the request is satisfied, then we try to find
1816 : * an extent that still satisfy the request, but is
1817 : * smaller than previous one */
1818 1105 : if (ex->fe_len < bex->fe_len)
1819 106 : *bex = *ex;
1820 : }
1821 :
1822 2710 : ext4_mb_check_limits(ac, e4b, 0);
1823 : }
1824 :
1825 : static noinline_for_stack
1826 8 : int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1827 : struct ext4_buddy *e4b)
1828 : {
1829 8 : struct ext4_free_extent ex = ac->ac_b_ex;
1830 8 : ext4_group_t group = ex.fe_group;
1831 8 : int max;
1832 8 : int err;
1833 :
1834 8 : BUG_ON(ex.fe_len <= 0);
1835 8 : err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1836 8 : if (err)
1837 : return err;
1838 :
1839 8 : ext4_lock_group(ac->ac_sb, group);
1840 8 : max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1841 :
1842 8 : if (max > 0) {
1843 8 : ac->ac_b_ex = ex;
1844 8 : ext4_mb_use_best_found(ac, e4b);
1845 : }
1846 :
1847 8 : ext4_unlock_group(ac->ac_sb, group);
1848 8 : ext4_mb_unload_buddy(e4b);
1849 :
1850 8 : return 0;
1851 : }
1852 :
1853 : static noinline_for_stack
1854 197 : int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1855 : struct ext4_buddy *e4b)
1856 : {
1857 197 : ext4_group_t group = ac->ac_g_ex.fe_group;
1858 197 : int max;
1859 197 : int err;
1860 197 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1861 197 : struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1862 197 : struct ext4_free_extent ex;
1863 :
1864 197 : if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1865 : return 0;
1866 8 : if (grp->bb_free == 0)
1867 : return 0;
1868 :
1869 8 : err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1870 8 : if (err)
1871 : return err;
1872 :
1873 8 : if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1874 0 : ext4_mb_unload_buddy(e4b);
1875 0 : return 0;
1876 : }
1877 :
1878 8 : ext4_lock_group(ac->ac_sb, group);
1879 8 : max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1880 : ac->ac_g_ex.fe_len, &ex);
1881 8 : ex.fe_logical = 0xDEADFA11; /* debug value */
1882 :
1883 8 : if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1884 0 : ext4_fsblk_t start;
1885 :
1886 0 : start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1887 0 : ex.fe_start;
1888 : /* use do_div to get remainder (would be 64-bit modulo) */
1889 0 : if (do_div(start, sbi->s_stripe) == 0) {
1890 0 : ac->ac_found++;
1891 0 : ac->ac_b_ex = ex;
1892 0 : ext4_mb_use_best_found(ac, e4b);
1893 : }
1894 8 : } else if (max >= ac->ac_g_ex.fe_len) {
1895 0 : BUG_ON(ex.fe_len <= 0);
1896 0 : BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1897 0 : BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1898 0 : ac->ac_found++;
1899 0 : ac->ac_b_ex = ex;
1900 0 : ext4_mb_use_best_found(ac, e4b);
1901 8 : } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1902 : /* Sometimes, caller may want to merge even small
1903 : * number of blocks to an existing extent */
1904 0 : BUG_ON(ex.fe_len <= 0);
1905 0 : BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1906 0 : BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1907 0 : ac->ac_found++;
1908 0 : ac->ac_b_ex = ex;
1909 0 : ext4_mb_use_best_found(ac, e4b);
1910 : }
1911 8 : ext4_unlock_group(ac->ac_sb, group);
1912 8 : ext4_mb_unload_buddy(e4b);
1913 :
1914 8 : return 0;
1915 : }
1916 :
1917 : /*
1918 : * The routine scans buddy structures (not bitmap!) from given order
1919 : * to max order and tries to find big enough chunk to satisfy the req
1920 : */
1921 : static noinline_for_stack
1922 7 : void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1923 : struct ext4_buddy *e4b)
1924 : {
1925 7 : struct super_block *sb = ac->ac_sb;
1926 7 : struct ext4_group_info *grp = e4b->bd_info;
1927 7 : void *buddy;
1928 7 : int i;
1929 7 : int k;
1930 7 : int max;
1931 :
1932 7 : BUG_ON(ac->ac_2order <= 0);
1933 7 : for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1934 7 : if (grp->bb_counters[i] == 0)
1935 0 : continue;
1936 :
1937 7 : buddy = mb_find_buddy(e4b, i, &max);
1938 7 : BUG_ON(buddy == NULL);
1939 :
1940 7 : k = mb_find_next_zero_bit(buddy, max, 0);
1941 7 : if (k >= max) {
1942 0 : ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
1943 : "%d free clusters of order %d. But found 0",
1944 : grp->bb_counters[i], i);
1945 0 : ext4_mark_group_bitmap_corrupted(ac->ac_sb,
1946 : e4b->bd_group,
1947 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1948 0 : break;
1949 : }
1950 7 : ac->ac_found++;
1951 :
1952 7 : ac->ac_b_ex.fe_len = 1 << i;
1953 7 : ac->ac_b_ex.fe_start = k << i;
1954 7 : ac->ac_b_ex.fe_group = e4b->bd_group;
1955 :
1956 7 : ext4_mb_use_best_found(ac, e4b);
1957 :
1958 7 : BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
1959 :
1960 7 : if (EXT4_SB(sb)->s_mb_stats)
1961 0 : atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1962 :
1963 : break;
1964 : }
1965 7 : }
1966 :
1967 : /*
1968 : * The routine scans the group and measures all found extents.
1969 : * In order to optimize scanning, caller must pass number of
1970 : * free blocks in the group, so the routine can know upper limit.
1971 : */
1972 : static noinline_for_stack
1973 190 : void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1974 : struct ext4_buddy *e4b)
1975 : {
1976 190 : struct super_block *sb = ac->ac_sb;
1977 190 : void *bitmap = e4b->bd_bitmap;
1978 190 : struct ext4_free_extent ex;
1979 190 : int i;
1980 190 : int free;
1981 :
1982 190 : free = e4b->bd_info->bb_free;
1983 190 : if (WARN_ON(free <= 0))
1984 0 : return;
1985 :
1986 190 : i = e4b->bd_info->bb_first_free;
1987 :
1988 3179 : while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1989 2989 : i = mb_find_next_zero_bit(bitmap,
1990 2989 : EXT4_CLUSTERS_PER_GROUP(sb), i);
1991 2989 : if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1992 : /*
1993 : * IF we have corrupt bitmap, we won't find any
1994 : * free blocks even though group info says we
1995 : * have free blocks
1996 : */
1997 0 : ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1998 : "%d free clusters as per "
1999 : "group info. But bitmap says 0",
2000 : free);
2001 0 : ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2002 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2003 0 : break;
2004 : }
2005 :
2006 2989 : mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2007 2989 : if (WARN_ON(ex.fe_len <= 0))
2008 : break;
2009 2989 : if (free < ex.fe_len) {
2010 0 : ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2011 : "%d free clusters as per "
2012 : "group info. But got %d blocks",
2013 : free, ex.fe_len);
2014 0 : ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2015 : EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2016 : /*
2017 : * The number of free blocks differs. This mostly
2018 : * indicate that the bitmap is corrupt. So exit
2019 : * without claiming the space.
2020 : */
2021 0 : break;
2022 : }
2023 2989 : ex.fe_logical = 0xDEADC0DE; /* debug value */
2024 2989 : ext4_mb_measure_extent(ac, &ex, e4b);
2025 :
2026 2989 : i += ex.fe_len;
2027 2989 : free -= ex.fe_len;
2028 : }
2029 :
2030 190 : ext4_mb_check_limits(ac, e4b, 1);
2031 : }
2032 :
2033 : /*
2034 : * This is a special case for storages like raid5
2035 : * we try to find stripe-aligned chunks for stripe-size-multiple requests
2036 : */
2037 : static noinline_for_stack
2038 0 : void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2039 : struct ext4_buddy *e4b)
2040 : {
2041 0 : struct super_block *sb = ac->ac_sb;
2042 0 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2043 0 : void *bitmap = e4b->bd_bitmap;
2044 0 : struct ext4_free_extent ex;
2045 0 : ext4_fsblk_t first_group_block;
2046 0 : ext4_fsblk_t a;
2047 0 : ext4_grpblk_t i;
2048 0 : int max;
2049 :
2050 0 : BUG_ON(sbi->s_stripe == 0);
2051 :
2052 : /* find first stripe-aligned block in group */
2053 0 : first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2054 :
2055 0 : a = first_group_block + sbi->s_stripe - 1;
2056 0 : do_div(a, sbi->s_stripe);
2057 0 : i = (a * sbi->s_stripe) - first_group_block;
2058 :
2059 0 : while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2060 0 : if (!mb_test_bit(i, bitmap)) {
2061 0 : max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2062 0 : if (max >= sbi->s_stripe) {
2063 0 : ac->ac_found++;
2064 0 : ex.fe_logical = 0xDEADF00D; /* debug value */
2065 0 : ac->ac_b_ex = ex;
2066 0 : ext4_mb_use_best_found(ac, e4b);
2067 0 : break;
2068 : }
2069 : }
2070 0 : i += sbi->s_stripe;
2071 : }
2072 0 : }
2073 :
2074 : /*
2075 : * This is also called BEFORE we load the buddy bitmap.
2076 : * Returns either 1 or 0 indicating that the group is either suitable
2077 : * for the allocation or not.
2078 : */
2079 585 : static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2080 : ext4_group_t group, int cr)
2081 : {
2082 585 : ext4_grpblk_t free, fragments;
2083 585 : int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2084 585 : struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2085 :
2086 585 : BUG_ON(cr < 0 || cr >= 4);
2087 :
2088 585 : if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2089 : return false;
2090 :
2091 585 : free = grp->bb_free;
2092 585 : if (free == 0)
2093 : return false;
2094 :
2095 585 : fragments = grp->bb_fragments;
2096 585 : if (fragments == 0)
2097 : return false;
2098 :
2099 585 : switch (cr) {
2100 71 : case 0:
2101 71 : BUG_ON(ac->ac_2order == 0);
2102 :
2103 : /* Avoid using the first bg of a flexgroup for data files */
2104 71 : if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2105 71 : (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2106 71 : ((group % flex_size) == 0))
2107 : return false;
2108 :
2109 70 : if (free < ac->ac_g_ex.fe_len)
2110 : return false;
2111 :
2112 70 : if (ac->ac_2order > ac->ac_sb->s_blocksize_bits+1)
2113 : return true;
2114 :
2115 70 : if (grp->bb_largest_free_order < ac->ac_2order)
2116 56 : return false;
2117 :
2118 : return true;
2119 498 : case 1:
2120 498 : if ((free / fragments) >= ac->ac_g_ex.fe_len)
2121 364 : return true;
2122 : break;
2123 16 : case 2:
2124 16 : if (free >= ac->ac_g_ex.fe_len)
2125 16 : return true;
2126 : break;
2127 : case 3:
2128 : return true;
2129 0 : default:
2130 0 : BUG();
2131 : }
2132 :
2133 : return false;
2134 : }
2135 :
2136 : /*
2137 : * This could return negative error code if something goes wrong
2138 : * during ext4_mb_init_group(). This should not be called with
2139 : * ext4_lock_group() held.
2140 : */
2141 390 : static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2142 : ext4_group_t group, int cr)
2143 : {
2144 390 : struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2145 390 : struct super_block *sb = ac->ac_sb;
2146 390 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2147 390 : bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2148 390 : ext4_grpblk_t free;
2149 390 : int ret = 0;
2150 :
2151 390 : if (should_lock)
2152 0 : ext4_lock_group(sb, group);
2153 390 : free = grp->bb_free;
2154 390 : if (free == 0)
2155 0 : goto out;
2156 390 : if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2157 2 : goto out;
2158 388 : if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2159 0 : goto out;
2160 388 : if (should_lock)
2161 0 : ext4_unlock_group(sb, group);
2162 :
2163 : /* We only do this if the grp has never been initialized */
2164 388 : if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2165 0 : struct ext4_group_desc *gdp =
2166 0 : ext4_get_group_desc(sb, group, NULL);
2167 0 : int ret;
2168 :
2169 : /* cr=0/1 is a very optimistic search to find large
2170 : * good chunks almost for free. If buddy data is not
2171 : * ready, then this optimization makes no sense. But
2172 : * we never skip the first block group in a flex_bg,
2173 : * since this gets used for metadata block allocation,
2174 : * and we want to make sure we locate metadata blocks
2175 : * in the first block group in the flex_bg if possible.
2176 : */
2177 0 : if (cr < 2 &&
2178 0 : (!sbi->s_log_groups_per_flex ||
2179 0 : ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2180 0 : !(ext4_has_group_desc_csum(sb) &&
2181 0 : (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2182 : return 0;
2183 0 : ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2184 0 : if (ret)
2185 : return ret;
2186 : }
2187 :
2188 388 : if (should_lock)
2189 0 : ext4_lock_group(sb, group);
2190 388 : ret = ext4_mb_good_group(ac, group, cr);
2191 390 : out:
2192 390 : if (should_lock)
2193 0 : ext4_unlock_group(sb, group);
2194 : return ret;
2195 : }
2196 :
2197 : /*
2198 : * Start prefetching @nr block bitmaps starting at @group.
2199 : * Return the next group which needs to be prefetched.
2200 : */
2201 213 : ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2202 : unsigned int nr, int *cnt)
2203 : {
2204 213 : ext4_group_t ngroups = ext4_get_groups_count(sb);
2205 213 : struct buffer_head *bh;
2206 213 : struct blk_plug plug;
2207 :
2208 213 : blk_start_plug(&plug);
2209 3497 : while (nr-- > 0) {
2210 3284 : struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2211 : NULL);
2212 3284 : struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2213 :
2214 : /*
2215 : * Prefetch block groups with free blocks; but don't
2216 : * bother if it is marked uninitialized on disk, since
2217 : * it won't require I/O to read. Also only try to
2218 : * prefetch once, so we avoid getblk() call, which can
2219 : * be expensive.
2220 : */
2221 3284 : if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2222 31 : EXT4_MB_GRP_NEED_INIT(grp) &&
2223 30 : ext4_free_group_clusters(sb, gdp) > 0 &&
2224 15 : !(ext4_has_group_desc_csum(sb) &&
2225 15 : (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2226 15 : bh = ext4_read_block_bitmap_nowait(sb, group, true);
2227 15 : if (bh && !IS_ERR(bh)) {
2228 15 : if (!buffer_uptodate(bh) && cnt)
2229 15 : (*cnt)++;
2230 15 : brelse(bh);
2231 : }
2232 : }
2233 3284 : if (++group >= ngroups)
2234 213 : group = 0;
2235 : }
2236 213 : blk_finish_plug(&plug);
2237 213 : return group;
2238 : }
2239 :
2240 : /*
2241 : * Prefetching reads the block bitmap into the buffer cache; but we
2242 : * need to make sure that the buddy bitmap in the page cache has been
2243 : * initialized. Note that ext4_mb_init_group() will block if the I/O
2244 : * is not yet completed, or indeed if it was not initiated by
2245 : * ext4_mb_prefetch did not start the I/O.
2246 : *
2247 : * TODO: We should actually kick off the buddy bitmap setup in a work
2248 : * queue when the buffer I/O is completed, so that we don't block
2249 : * waiting for the block allocation bitmap read to finish when
2250 : * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2251 : */
2252 1 : void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2253 : unsigned int nr)
2254 : {
2255 17 : while (nr-- > 0) {
2256 16 : struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2257 : NULL);
2258 16 : struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2259 :
2260 16 : if (!group)
2261 1 : group = ext4_get_groups_count(sb);
2262 16 : group--;
2263 16 : grp = ext4_get_group_info(sb, group);
2264 :
2265 31 : if (EXT4_MB_GRP_NEED_INIT(grp) &&
2266 30 : ext4_free_group_clusters(sb, gdp) > 0 &&
2267 15 : !(ext4_has_group_desc_csum(sb) &&
2268 15 : (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
2269 15 : if (ext4_mb_init_group(sb, group, GFP_NOFS))
2270 : break;
2271 : }
2272 : }
2273 1 : }
2274 :
2275 : static noinline_for_stack int
2276 197 : ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2277 : {
2278 197 : ext4_group_t prefetch_grp = 0, ngroups, group, i;
2279 197 : int cr = -1;
2280 197 : int err = 0, first_err = 0;
2281 197 : unsigned int nr = 0, prefetch_ios = 0;
2282 197 : struct ext4_sb_info *sbi;
2283 197 : struct super_block *sb;
2284 197 : struct ext4_buddy e4b;
2285 197 : int lost;
2286 :
2287 197 : sb = ac->ac_sb;
2288 197 : sbi = EXT4_SB(sb);
2289 197 : ngroups = ext4_get_groups_count(sb);
2290 : /* non-extent files are limited to low blocks/groups */
2291 197 : if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2292 0 : ngroups = sbi->s_blockfile_groups;
2293 :
2294 197 : BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2295 :
2296 : /* first, try the goal */
2297 197 : err = ext4_mb_find_by_goal(ac, &e4b);
2298 197 : if (err || ac->ac_status == AC_STATUS_FOUND)
2299 0 : goto out;
2300 :
2301 197 : if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2302 0 : goto out;
2303 :
2304 : /*
2305 : * ac->ac_2order is set only if the fe_len is a power of 2
2306 : * if ac->ac_2order is set we also set criteria to 0 so that we
2307 : * try exact allocation using buddy.
2308 : */
2309 197 : i = fls(ac->ac_g_ex.fe_len);
2310 197 : ac->ac_2order = 0;
2311 : /*
2312 : * We search using buddy data only if the order of the request
2313 : * is greater than equal to the sbi_s_mb_order2_reqs
2314 : * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2315 : * We also support searching for power-of-two requests only for
2316 : * requests upto maximum buddy size we have constructed.
2317 : */
2318 197 : if (i >= sbi->s_mb_order2_reqs && i <= sb->s_blocksize_bits + 2) {
2319 : /*
2320 : * This should tell if fe_len is exactly power of 2
2321 : */
2322 20 : if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2323 7 : ac->ac_2order = array_index_nospec(i - 1,
2324 : sb->s_blocksize_bits + 2);
2325 : }
2326 :
2327 : /* if stream allocation is enabled, use global goal */
2328 197 : if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2329 : /* TBD: may be hot point */
2330 9 : spin_lock(&sbi->s_md_lock);
2331 9 : ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2332 9 : ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2333 9 : spin_unlock(&sbi->s_md_lock);
2334 : }
2335 :
2336 : /* Let's just scan groups to find more-less suitable blocks */
2337 197 : cr = ac->ac_2order ? 0 : 1;
2338 : /*
2339 : * cr == 0 try to get exact allocation,
2340 : * cr == 3 try to get anything
2341 : */
2342 : repeat:
2343 607 : for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2344 205 : ac->ac_criteria = cr;
2345 : /*
2346 : * searching for the right group start
2347 : * from the goal value specified
2348 : */
2349 205 : group = ac->ac_g_ex.fe_group;
2350 205 : prefetch_grp = group;
2351 :
2352 398 : for (i = 0; i < ngroups; group++, i++) {
2353 390 : int ret = 0;
2354 390 : cond_resched();
2355 : /*
2356 : * Artificially restricted ngroups for non-extent
2357 : * files makes group > ngroups possible on first loop.
2358 : */
2359 390 : if (group >= ngroups)
2360 8 : group = 0;
2361 :
2362 : /*
2363 : * Batch reads of the block allocation bitmaps
2364 : * to get multiple READs in flight; limit
2365 : * prefetching at cr=0/1, otherwise mballoc can
2366 : * spend a lot of time loading imperfect groups
2367 : */
2368 390 : if ((prefetch_grp == group) &&
2369 205 : (cr > 1 ||
2370 205 : prefetch_ios < sbi->s_mb_prefetch_limit)) {
2371 213 : unsigned int curr_ios = prefetch_ios;
2372 :
2373 213 : nr = sbi->s_mb_prefetch;
2374 213 : if (ext4_has_feature_flex_bg(sb)) {
2375 213 : nr = 1 << sbi->s_log_groups_per_flex;
2376 213 : nr -= group & (nr - 1);
2377 213 : nr = min(nr, sbi->s_mb_prefetch);
2378 : }
2379 213 : prefetch_grp = ext4_mb_prefetch(sb, group,
2380 : nr, &prefetch_ios);
2381 213 : if (prefetch_ios == curr_ios)
2382 212 : nr = 0;
2383 : }
2384 :
2385 : /* This now checks without needing the buddy page */
2386 390 : ret = ext4_mb_good_group_nolock(ac, group, cr);
2387 390 : if (ret <= 0) {
2388 193 : if (!first_err)
2389 193 : first_err = ret;
2390 193 : continue;
2391 : }
2392 :
2393 197 : err = ext4_mb_load_buddy(sb, group, &e4b);
2394 197 : if (err)
2395 0 : goto out;
2396 :
2397 197 : ext4_lock_group(sb, group);
2398 :
2399 : /*
2400 : * We need to check again after locking the
2401 : * block group
2402 : */
2403 197 : ret = ext4_mb_good_group(ac, group, cr);
2404 197 : if (ret == 0) {
2405 0 : ext4_unlock_group(sb, group);
2406 0 : ext4_mb_unload_buddy(&e4b);
2407 0 : continue;
2408 : }
2409 :
2410 197 : ac->ac_groups_scanned++;
2411 197 : if (cr == 0)
2412 7 : ext4_mb_simple_scan_group(ac, &e4b);
2413 190 : else if (cr == 1 && sbi->s_stripe &&
2414 0 : !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2415 0 : ext4_mb_scan_aligned(ac, &e4b);
2416 : else
2417 190 : ext4_mb_complex_scan_group(ac, &e4b);
2418 :
2419 197 : ext4_unlock_group(sb, group);
2420 197 : ext4_mb_unload_buddy(&e4b);
2421 :
2422 197 : if (ac->ac_status != AC_STATUS_CONTINUE)
2423 : break;
2424 : }
2425 : }
2426 :
2427 197 : if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2428 8 : !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2429 : /*
2430 : * We've been searching too long. Let's try to allocate
2431 : * the best chunk we've found so far
2432 : */
2433 8 : ext4_mb_try_best_found(ac, &e4b);
2434 8 : if (ac->ac_status != AC_STATUS_FOUND) {
2435 : /*
2436 : * Someone more lucky has already allocated it.
2437 : * The only thing we can do is just take first
2438 : * found block(s)
2439 : */
2440 0 : lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2441 0 : mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2442 : ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2443 : ac->ac_b_ex.fe_len, lost);
2444 :
2445 0 : ac->ac_b_ex.fe_group = 0;
2446 0 : ac->ac_b_ex.fe_start = 0;
2447 0 : ac->ac_b_ex.fe_len = 0;
2448 0 : ac->ac_status = AC_STATUS_CONTINUE;
2449 0 : ac->ac_flags |= EXT4_MB_HINT_FIRST;
2450 0 : cr = 3;
2451 0 : goto repeat;
2452 : }
2453 : }
2454 197 : out:
2455 197 : if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2456 0 : err = first_err;
2457 :
2458 197 : mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2459 : ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2460 : ac->ac_flags, cr, err);
2461 :
2462 197 : if (nr)
2463 1 : ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2464 :
2465 197 : return err;
2466 : }
2467 :
2468 0 : static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2469 : {
2470 0 : struct super_block *sb = PDE_DATA(file_inode(seq->file));
2471 0 : ext4_group_t group;
2472 :
2473 0 : if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2474 0 : return NULL;
2475 0 : group = *pos + 1;
2476 0 : return (void *) ((unsigned long) group);
2477 : }
2478 :
2479 0 : static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2480 : {
2481 0 : struct super_block *sb = PDE_DATA(file_inode(seq->file));
2482 0 : ext4_group_t group;
2483 :
2484 0 : ++*pos;
2485 0 : if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2486 0 : return NULL;
2487 0 : group = *pos + 1;
2488 0 : return (void *) ((unsigned long) group);
2489 : }
2490 :
2491 0 : static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2492 : {
2493 0 : struct super_block *sb = PDE_DATA(file_inode(seq->file));
2494 0 : ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2495 0 : int i;
2496 0 : int err, buddy_loaded = 0;
2497 0 : struct ext4_buddy e4b;
2498 0 : struct ext4_group_info *grinfo;
2499 0 : unsigned char blocksize_bits = min_t(unsigned char,
2500 : sb->s_blocksize_bits,
2501 : EXT4_MAX_BLOCK_LOG_SIZE);
2502 0 : struct sg {
2503 : struct ext4_group_info info;
2504 : ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2505 : } sg;
2506 :
2507 0 : group--;
2508 0 : if (group == 0)
2509 0 : seq_puts(seq, "#group: free frags first ["
2510 : " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
2511 : " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
2512 :
2513 0 : i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2514 : sizeof(struct ext4_group_info);
2515 :
2516 0 : grinfo = ext4_get_group_info(sb, group);
2517 : /* Load the group info in memory only if not already loaded. */
2518 0 : if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2519 0 : err = ext4_mb_load_buddy(sb, group, &e4b);
2520 0 : if (err) {
2521 0 : seq_printf(seq, "#%-5u: I/O error\n", group);
2522 0 : return 0;
2523 : }
2524 : buddy_loaded = 1;
2525 : }
2526 :
2527 0 : memcpy(&sg, ext4_get_group_info(sb, group), i);
2528 :
2529 0 : if (buddy_loaded)
2530 0 : ext4_mb_unload_buddy(&e4b);
2531 :
2532 0 : seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2533 : sg.info.bb_fragments, sg.info.bb_first_free);
2534 0 : for (i = 0; i <= 13; i++)
2535 0 : seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
2536 : sg.info.bb_counters[i] : 0);
2537 0 : seq_puts(seq, " ]\n");
2538 :
2539 0 : return 0;
2540 : }
2541 :
2542 0 : static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2543 : {
2544 0 : }
2545 :
2546 : const struct seq_operations ext4_mb_seq_groups_ops = {
2547 : .start = ext4_mb_seq_groups_start,
2548 : .next = ext4_mb_seq_groups_next,
2549 : .stop = ext4_mb_seq_groups_stop,
2550 : .show = ext4_mb_seq_groups_show,
2551 : };
2552 :
2553 16 : static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2554 : {
2555 16 : int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2556 16 : struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2557 :
2558 16 : BUG_ON(!cachep);
2559 16 : return cachep;
2560 : }
2561 :
2562 : /*
2563 : * Allocate the top-level s_group_info array for the specified number
2564 : * of groups
2565 : */
2566 1 : int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2567 : {
2568 1 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2569 1 : unsigned size;
2570 1 : struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
2571 :
2572 1 : size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2573 1 : EXT4_DESC_PER_BLOCK_BITS(sb);
2574 1 : if (size <= sbi->s_group_info_size)
2575 : return 0;
2576 :
2577 1 : size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2578 1 : new_groupinfo = kvzalloc(size, GFP_KERNEL);
2579 1 : if (!new_groupinfo) {
2580 0 : ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2581 0 : return -ENOMEM;
2582 : }
2583 1 : rcu_read_lock();
2584 1 : old_groupinfo = rcu_dereference(sbi->s_group_info);
2585 1 : if (old_groupinfo)
2586 0 : memcpy(new_groupinfo, old_groupinfo,
2587 0 : sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2588 1 : rcu_read_unlock();
2589 1 : rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
2590 1 : sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2591 1 : if (old_groupinfo)
2592 0 : ext4_kvfree_array_rcu(old_groupinfo);
2593 : ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2594 : sbi->s_group_info_size);
2595 : return 0;
2596 : }
2597 :
2598 : /* Create and initialize ext4_group_info data for the given group. */
2599 16 : int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2600 : struct ext4_group_desc *desc)
2601 : {
2602 16 : int i;
2603 16 : int metalen = 0;
2604 16 : int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
2605 16 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2606 16 : struct ext4_group_info **meta_group_info;
2607 16 : struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2608 :
2609 : /*
2610 : * First check if this group is the first of a reserved block.
2611 : * If it's true, we have to allocate a new table of pointers
2612 : * to ext4_group_info structures
2613 : */
2614 16 : if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2615 1 : metalen = sizeof(*meta_group_info) <<
2616 1 : EXT4_DESC_PER_BLOCK_BITS(sb);
2617 1 : meta_group_info = kmalloc(metalen, GFP_NOFS);
2618 1 : if (meta_group_info == NULL) {
2619 0 : ext4_msg(sb, KERN_ERR, "can't allocate mem "
2620 : "for a buddy group");
2621 0 : goto exit_meta_group_info;
2622 : }
2623 1 : rcu_read_lock();
2624 1 : rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
2625 1 : rcu_read_unlock();
2626 : }
2627 :
2628 32 : meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
2629 16 : i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2630 :
2631 16 : meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2632 16 : if (meta_group_info[i] == NULL) {
2633 0 : ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2634 0 : goto exit_group_info;
2635 : }
2636 16 : set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2637 16 : &(meta_group_info[i]->bb_state));
2638 :
2639 : /*
2640 : * initialize bb_free to be able to skip
2641 : * empty groups without initialization
2642 : */
2643 16 : if (ext4_has_group_desc_csum(sb) &&
2644 16 : (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
2645 0 : meta_group_info[i]->bb_free =
2646 0 : ext4_free_clusters_after_init(sb, group, desc);
2647 : } else {
2648 32 : meta_group_info[i]->bb_free =
2649 16 : ext4_free_group_clusters(sb, desc);
2650 : }
2651 :
2652 16 : INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2653 16 : init_rwsem(&meta_group_info[i]->alloc_sem);
2654 16 : meta_group_info[i]->bb_free_root = RB_ROOT;
2655 16 : meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2656 :
2657 16 : mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
2658 16 : return 0;
2659 :
2660 0 : exit_group_info:
2661 : /* If a meta_group_info table has been allocated, release it now */
2662 0 : if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2663 0 : struct ext4_group_info ***group_info;
2664 :
2665 0 : rcu_read_lock();
2666 0 : group_info = rcu_dereference(sbi->s_group_info);
2667 0 : kfree(group_info[idx]);
2668 0 : group_info[idx] = NULL;
2669 0 : rcu_read_unlock();
2670 : }
2671 0 : exit_meta_group_info:
2672 : return -ENOMEM;
2673 : } /* ext4_mb_add_groupinfo */
2674 :
2675 1 : static int ext4_mb_init_backend(struct super_block *sb)
2676 : {
2677 1 : ext4_group_t ngroups = ext4_get_groups_count(sb);
2678 1 : ext4_group_t i;
2679 1 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2680 1 : int err;
2681 1 : struct ext4_group_desc *desc;
2682 1 : struct ext4_group_info ***group_info;
2683 1 : struct kmem_cache *cachep;
2684 :
2685 1 : err = ext4_mb_alloc_groupinfo(sb, ngroups);
2686 1 : if (err)
2687 : return err;
2688 :
2689 1 : sbi->s_buddy_cache = new_inode(sb);
2690 1 : if (sbi->s_buddy_cache == NULL) {
2691 0 : ext4_msg(sb, KERN_ERR, "can't get new inode");
2692 0 : goto err_freesgi;
2693 : }
2694 : /* To avoid potentially colliding with an valid on-disk inode number,
2695 : * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2696 : * not in the inode hash, so it should never be found by iget(), but
2697 : * this will avoid confusion if it ever shows up during debugging. */
2698 1 : sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2699 1 : EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2700 17 : for (i = 0; i < ngroups; i++) {
2701 16 : cond_resched();
2702 16 : desc = ext4_get_group_desc(sb, i, NULL);
2703 16 : if (desc == NULL) {
2704 0 : ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2705 0 : goto err_freebuddy;
2706 : }
2707 16 : if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2708 0 : goto err_freebuddy;
2709 : }
2710 :
2711 1 : if (ext4_has_feature_flex_bg(sb)) {
2712 : /* a single flex group is supposed to be read by a single IO */
2713 1 : sbi->s_mb_prefetch = min(1 << sbi->s_es->s_log_groups_per_flex,
2714 : BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
2715 1 : sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
2716 : } else {
2717 0 : sbi->s_mb_prefetch = 32;
2718 : }
2719 1 : if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
2720 1 : sbi->s_mb_prefetch = ext4_get_groups_count(sb);
2721 : /* now many real IOs to prefetch within a single allocation at cr=0
2722 : * given cr=0 is an CPU-related optimization we shouldn't try to
2723 : * load too many groups, at some point we should start to use what
2724 : * we've got in memory.
2725 : * with an average random access time 5ms, it'd take a second to get
2726 : * 200 groups (* N with flex_bg), so let's make this limit 4
2727 : */
2728 1 : sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
2729 1 : if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
2730 1 : sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
2731 :
2732 : return 0;
2733 :
2734 0 : err_freebuddy:
2735 0 : cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2736 0 : while (i-- > 0)
2737 0 : kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2738 0 : i = sbi->s_group_info_size;
2739 0 : rcu_read_lock();
2740 0 : group_info = rcu_dereference(sbi->s_group_info);
2741 0 : while (i-- > 0)
2742 0 : kfree(group_info[i]);
2743 0 : rcu_read_unlock();
2744 0 : iput(sbi->s_buddy_cache);
2745 0 : err_freesgi:
2746 0 : rcu_read_lock();
2747 0 : kvfree(rcu_dereference(sbi->s_group_info));
2748 0 : rcu_read_unlock();
2749 0 : return -ENOMEM;
2750 : }
2751 :
2752 0 : static void ext4_groupinfo_destroy_slabs(void)
2753 : {
2754 0 : int i;
2755 :
2756 0 : for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2757 0 : kmem_cache_destroy(ext4_groupinfo_caches[i]);
2758 0 : ext4_groupinfo_caches[i] = NULL;
2759 : }
2760 0 : }
2761 :
2762 1 : static int ext4_groupinfo_create_slab(size_t size)
2763 : {
2764 1 : static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2765 1 : int slab_size;
2766 1 : int blocksize_bits = order_base_2(size);
2767 1 : int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2768 1 : struct kmem_cache *cachep;
2769 :
2770 1 : if (cache_index >= NR_GRPINFO_CACHES)
2771 : return -EINVAL;
2772 :
2773 1 : if (unlikely(cache_index < 0))
2774 0 : cache_index = 0;
2775 :
2776 1 : mutex_lock(&ext4_grpinfo_slab_create_mutex);
2777 1 : if (ext4_groupinfo_caches[cache_index]) {
2778 0 : mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2779 0 : return 0; /* Already created */
2780 : }
2781 :
2782 1 : slab_size = offsetof(struct ext4_group_info,
2783 : bb_counters[blocksize_bits + 2]);
2784 :
2785 1 : cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2786 : slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2787 : NULL);
2788 :
2789 1 : ext4_groupinfo_caches[cache_index] = cachep;
2790 :
2791 1 : mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2792 1 : if (!cachep) {
2793 0 : printk(KERN_EMERG
2794 : "EXT4-fs: no memory for groupinfo slab cache\n");
2795 0 : return -ENOMEM;
2796 : }
2797 :
2798 : return 0;
2799 : }
2800 :
2801 1 : int ext4_mb_init(struct super_block *sb)
2802 : {
2803 1 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2804 1 : unsigned i, j;
2805 1 : unsigned offset, offset_incr;
2806 1 : unsigned max;
2807 1 : int ret;
2808 :
2809 1 : i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2810 :
2811 1 : sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2812 1 : if (sbi->s_mb_offsets == NULL) {
2813 0 : ret = -ENOMEM;
2814 0 : goto out;
2815 : }
2816 :
2817 1 : i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2818 1 : sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2819 1 : if (sbi->s_mb_maxs == NULL) {
2820 0 : ret = -ENOMEM;
2821 0 : goto out;
2822 : }
2823 :
2824 1 : ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2825 1 : if (ret < 0)
2826 0 : goto out;
2827 :
2828 : /* order 0 is regular bitmap */
2829 1 : sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2830 1 : sbi->s_mb_offsets[0] = 0;
2831 :
2832 1 : i = 1;
2833 1 : offset = 0;
2834 1 : offset_incr = 1 << (sb->s_blocksize_bits - 1);
2835 1 : max = sb->s_blocksize << 2;
2836 13 : do {
2837 13 : sbi->s_mb_offsets[i] = offset;
2838 13 : sbi->s_mb_maxs[i] = max;
2839 13 : offset += offset_incr;
2840 13 : offset_incr = offset_incr >> 1;
2841 13 : max = max >> 1;
2842 13 : i++;
2843 13 : } while (i <= sb->s_blocksize_bits + 1);
2844 :
2845 1 : spin_lock_init(&sbi->s_md_lock);
2846 1 : spin_lock_init(&sbi->s_bal_lock);
2847 1 : sbi->s_mb_free_pending = 0;
2848 1 : INIT_LIST_HEAD(&sbi->s_freed_data_list);
2849 :
2850 1 : sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2851 1 : sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2852 1 : sbi->s_mb_stats = MB_DEFAULT_STATS;
2853 1 : sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2854 1 : sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2855 1 : sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
2856 : /*
2857 : * The default group preallocation is 512, which for 4k block
2858 : * sizes translates to 2 megabytes. However for bigalloc file
2859 : * systems, this is probably too big (i.e, if the cluster size
2860 : * is 1 megabyte, then group preallocation size becomes half a
2861 : * gigabyte!). As a default, we will keep a two megabyte
2862 : * group pralloc size for cluster sizes up to 64k, and after
2863 : * that, we will force a minimum group preallocation size of
2864 : * 32 clusters. This translates to 8 megs when the cluster
2865 : * size is 256k, and 32 megs when the cluster size is 1 meg,
2866 : * which seems reasonable as a default.
2867 : */
2868 1 : sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2869 : sbi->s_cluster_bits, 32);
2870 : /*
2871 : * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2872 : * to the lowest multiple of s_stripe which is bigger than
2873 : * the s_mb_group_prealloc as determined above. We want
2874 : * the preallocation size to be an exact multiple of the
2875 : * RAID stripe size so that preallocations don't fragment
2876 : * the stripes.
2877 : */
2878 1 : if (sbi->s_stripe > 1) {
2879 0 : sbi->s_mb_group_prealloc = roundup(
2880 : sbi->s_mb_group_prealloc, sbi->s_stripe);
2881 : }
2882 :
2883 1 : sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2884 1 : if (sbi->s_locality_groups == NULL) {
2885 0 : ret = -ENOMEM;
2886 0 : goto out;
2887 : }
2888 5 : for_each_possible_cpu(i) {
2889 4 : struct ext4_locality_group *lg;
2890 4 : lg = per_cpu_ptr(sbi->s_locality_groups, i);
2891 4 : mutex_init(&lg->lg_mutex);
2892 48 : for (j = 0; j < PREALLOC_TB_SIZE; j++)
2893 40 : INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2894 5 : spin_lock_init(&lg->lg_prealloc_lock);
2895 : }
2896 :
2897 : /* init file for buddy data */
2898 1 : ret = ext4_mb_init_backend(sb);
2899 1 : if (ret != 0)
2900 0 : goto out_free_locality_groups;
2901 :
2902 : return 0;
2903 :
2904 0 : out_free_locality_groups:
2905 0 : free_percpu(sbi->s_locality_groups);
2906 0 : sbi->s_locality_groups = NULL;
2907 0 : out:
2908 0 : kfree(sbi->s_mb_offsets);
2909 0 : sbi->s_mb_offsets = NULL;
2910 0 : kfree(sbi->s_mb_maxs);
2911 0 : sbi->s_mb_maxs = NULL;
2912 0 : return ret;
2913 : }
2914 :
2915 : /* need to called with the ext4 group lock held */
2916 0 : static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2917 : {
2918 0 : struct ext4_prealloc_space *pa;
2919 0 : struct list_head *cur, *tmp;
2920 0 : int count = 0;
2921 :
2922 0 : list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2923 0 : pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2924 0 : list_del(&pa->pa_group_list);
2925 0 : count++;
2926 0 : kmem_cache_free(ext4_pspace_cachep, pa);
2927 : }
2928 0 : return count;
2929 : }
2930 :
2931 0 : int ext4_mb_release(struct super_block *sb)
2932 : {
2933 0 : ext4_group_t ngroups = ext4_get_groups_count(sb);
2934 0 : ext4_group_t i;
2935 0 : int num_meta_group_infos;
2936 0 : struct ext4_group_info *grinfo, ***group_info;
2937 0 : struct ext4_sb_info *sbi = EXT4_SB(sb);
2938 0 : struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2939 0 : int count;
2940 :
2941 0 : if (sbi->s_group_info) {
2942 0 : for (i = 0; i < ngroups; i++) {
2943 0 : cond_resched();
2944 0 : grinfo = ext4_get_group_info(sb, i);
2945 0 : mb_group_bb_bitmap_free(grinfo);
2946 0 : ext4_lock_group(sb, i);
2947 0 : count = ext4_mb_cleanup_pa(grinfo);
2948 0 : if (count)
2949 : mb_debug(sb, "mballoc: %d PAs left\n",
2950 : count);
2951 0 : ext4_unlock_group(sb, i);
2952 0 : kmem_cache_free(cachep, grinfo);
2953 : }
2954 0 : num_meta_group_infos = (ngroups +
2955 0 : EXT4_DESC_PER_BLOCK(sb) - 1) >>
2956 0 : EXT4_DESC_PER_BLOCK_BITS(sb);
2957 0 : rcu_read_lock();
2958 0 : group_info = rcu_dereference(sbi->s_group_info);
2959 0 : for (i = 0; i < num_meta_group_infos; i++)
2960 0 : kfree(group_info[i]);
2961 0 : kvfree(group_info);
2962 0 : rcu_read_unlock();
2963 : }
2964 0 : kfree(sbi->s_mb_offsets);
2965 0 : kfree(sbi->s_mb_maxs);
2966 0 : iput(sbi->s_buddy_cache);
2967 0 : if (sbi->s_mb_stats) {
2968 0 : ext4_msg(sb, KERN_INFO,
2969 : "mballoc: %u blocks %u reqs (%u success)",
2970 : atomic_read(&sbi->s_bal_allocated),
2971 : atomic_read(&sbi->s_bal_reqs),
2972 : atomic_read(&sbi->s_bal_success));
2973 0 : ext4_msg(sb, KERN_INFO,
2974 : "mballoc: %u extents scanned, %u goal hits, "
2975 : "%u 2^N hits, %u breaks, %u lost",
2976 : atomic_read(&sbi->s_bal_ex_scanned),
2977 : atomic_read(&sbi->s_bal_goals),
2978 : atomic_read(&sbi->s_bal_2orders),
2979 : atomic_read(&sbi->s_bal_breaks),
2980 : atomic_read(&sbi->s_mb_lost_chunks));
2981 0 : ext4_msg(sb, KERN_INFO,
2982 : "mballoc: %lu generated and it took %Lu",
2983 : sbi->s_mb_buddies_generated,
2984 : sbi->s_mb_generation_time);
2985 0 : ext4_msg(sb, KERN_INFO,
2986 : "mballoc: %u preallocated, %u discarded",
2987 : atomic_read(&sbi->s_mb_preallocated),
2988 : atomic_read(&sbi->s_mb_discarded));
2989 : }
2990 :
2991 0 : free_percpu(sbi->s_locality_groups);
2992 :
2993 0 : return 0;
2994 : }
2995 :
2996 0 : static inline int ext4_issue_discard(struct super_block *sb,
2997 : ext4_group_t block_group, ext4_grpblk_t cluster, int count,
2998 : struct bio **biop)
2999 : {
3000 0 : ext4_fsblk_t discard_block;
3001 :
3002 0 : discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3003 0 : ext4_group_first_block_no(sb, block_group));
3004 0 : count = EXT4_C2B(EXT4_SB(sb), count);
3005 0 : trace_ext4_discard_blocks(sb,
3006 : (unsigned long long) discard_block, count);
3007 0 : if (biop) {
3008 0 : return __blkdev_issue_discard(sb->s_bdev,
3009 : (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3010 0 : (sector_t)count << (sb->s_blocksize_bits - 9),
3011 : GFP_NOFS, 0, biop);
3012 : } else
3013 0 : return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3014 : }
3015 :
3016 164 : static void ext4_free_data_in_buddy(struct super_block *sb,
3017 : struct ext4_free_data *entry)
3018 : {
3019 164 : struct ext4_buddy e4b;
3020 164 : struct ext4_group_info *db;
3021 164 : int err, count = 0, count2 = 0;
3022 :
3023 164 : mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3024 : entry->efd_count, entry->efd_group, entry);
3025 :
3026 164 : err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3027 : /* we expect to find existing buddy because it's pinned */
3028 164 : BUG_ON(err != 0);
3029 :
3030 164 : spin_lock(&EXT4_SB(sb)->s_md_lock);
3031 164 : EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3032 164 : spin_unlock(&EXT4_SB(sb)->s_md_lock);
3033 :
3034 164 : db = e4b.bd_info;
3035 : /* there are blocks to put in buddy to make them really free */
3036 164 : count += entry->efd_count;
3037 164 : count2++;
3038 164 : ext4_lock_group(sb, entry->efd_group);
3039 : /* Take it out of per group rb tree */
3040 164 : rb_erase(&entry->efd_node, &(db->bb_free_root));
3041 164 : mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3042 :
3043 : /*
3044 : * Clear the trimmed flag for the group so that the next
3045 : * ext4_trim_fs can trim it.
3046 : * If the volume is mounted with -o discard, online discard
3047 : * is supported and the free blocks will be trimmed online.
3048 : */
3049 164 : if (!test_opt(sb, DISCARD))
3050 164 : EXT4_MB_GRP_CLEAR_TRIMMED(db);
3051 :
3052 164 : if (!db->bb_free_root.rb_node) {
3053 : /* No more items in the per group rb tree
3054 : * balance refcounts from ext4_mb_free_metadata()
3055 : */
3056 79 : put_page(e4b.bd_buddy_page);
3057 79 : put_page(e4b.bd_bitmap_page);
3058 : }
3059 164 : ext4_unlock_group(sb, entry->efd_group);
3060 164 : kmem_cache_free(ext4_free_data_cachep, entry);
3061 164 : ext4_mb_unload_buddy(&e4b);
3062 :
3063 164 : mb_debug(sb, "freed %d blocks in %d structures\n", count,
3064 : count2);
3065 164 : }
3066 :
3067 : /*
3068 : * This function is called by the jbd2 layer once the commit has finished,
3069 : * so we know we can free the blocks that were released with that commit.
3070 : */
3071 58 : void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3072 : {
3073 58 : struct ext4_sb_info *sbi = EXT4_SB(sb);
3074 58 : struct ext4_free_data *entry, *tmp;
3075 58 : struct bio *discard_bio = NULL;
3076 58 : struct list_head freed_data_list;
3077 58 : struct list_head *cut_pos = NULL;
3078 58 : int err;
3079 :
3080 58 : INIT_LIST_HEAD(&freed_data_list);
3081 :
3082 58 : spin_lock(&sbi->s_md_lock);
3083 222 : list_for_each_entry(entry, &sbi->s_freed_data_list, efd_list) {
3084 164 : if (entry->efd_tid != commit_tid)
3085 : break;
3086 164 : cut_pos = &entry->efd_list;
3087 : }
3088 58 : if (cut_pos)
3089 42 : list_cut_position(&freed_data_list, &sbi->s_freed_data_list,
3090 : cut_pos);
3091 58 : spin_unlock(&sbi->s_md_lock);
3092 :
3093 58 : if (test_opt(sb, DISCARD)) {
3094 0 : list_for_each_entry(entry, &freed_data_list, efd_list) {
3095 0 : err = ext4_issue_discard(sb, entry->efd_group,
3096 : entry->efd_start_cluster,
3097 : entry->efd_count,
3098 : &discard_bio);
3099 0 : if (err && err != -EOPNOTSUPP) {
3100 0 : ext4_msg(sb, KERN_WARNING, "discard request in"
3101 : " group:%d block:%d count:%d failed"
3102 : " with %d", entry->efd_group,
3103 : entry->efd_start_cluster,
3104 : entry->efd_count, err);
3105 0 : } else if (err == -EOPNOTSUPP)
3106 : break;
3107 : }
3108 :
3109 0 : if (discard_bio) {
3110 0 : submit_bio_wait(discard_bio);
3111 0 : bio_put(discard_bio);
3112 : }
3113 : }
3114 :
3115 222 : list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3116 164 : ext4_free_data_in_buddy(sb, entry);
3117 58 : }
3118 :
3119 1 : int __init ext4_init_mballoc(void)
3120 : {
3121 1 : ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3122 : SLAB_RECLAIM_ACCOUNT);
3123 1 : if (ext4_pspace_cachep == NULL)
3124 0 : goto out;
3125 :
3126 1 : ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3127 : SLAB_RECLAIM_ACCOUNT);
3128 1 : if (ext4_ac_cachep == NULL)
3129 0 : goto out_pa_free;
3130 :
3131 1 : ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3132 : SLAB_RECLAIM_ACCOUNT);
3133 1 : if (ext4_free_data_cachep == NULL)
3134 0 : goto out_ac_free;
3135 :
3136 : return 0;
3137 :
3138 0 : out_ac_free:
3139 0 : kmem_cache_destroy(ext4_ac_cachep);
3140 0 : out_pa_free:
3141 0 : kmem_cache_destroy(ext4_pspace_cachep);
3142 : out:
3143 : return -ENOMEM;
3144 : }
3145 :
3146 0 : void ext4_exit_mballoc(void)
3147 : {
3148 : /*
3149 : * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3150 : * before destroying the slab cache.
3151 : */
3152 0 : rcu_barrier();
3153 0 : kmem_cache_destroy(ext4_pspace_cachep);
3154 0 : kmem_cache_destroy(ext4_ac_cachep);
3155 0 : kmem_cache_destroy(ext4_free_data_cachep);
3156 0 : ext4_groupinfo_destroy_slabs();
3157 0 : }
3158 :
3159 :
3160 : /*
3161 : * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
3162 : * Returns 0 if success or error code
3163 : */
3164 : static noinline_for_stack int
3165 245 : ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
3166 : handle_t *handle, unsigned int reserv_clstrs)
3167 : {
3168 245 : struct buffer_head *bitmap_bh = NULL;
3169 245 : struct ext4_group_desc *gdp;
3170 245 : struct buffer_head *gdp_bh;
3171 245 : struct ext4_sb_info *sbi;
3172 245 : struct super_block *sb;
3173 245 : ext4_fsblk_t block;
3174 245 : int err, len;
3175 :
3176 245 : BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3177 245 : BUG_ON(ac->ac_b_ex.fe_len <= 0);
3178 :
3179 245 : sb = ac->ac_sb;
3180 245 : sbi = EXT4_SB(sb);
3181 :
3182 245 : bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
3183 245 : if (IS_ERR(bitmap_bh)) {
3184 0 : err = PTR_ERR(bitmap_bh);
3185 0 : bitmap_bh = NULL;
3186 0 : goto out_err;
3187 : }
3188 :
3189 245 : BUFFER_TRACE(bitmap_bh, "getting write access");
3190 245 : err = ext4_journal_get_write_access(handle, bitmap_bh);
3191 245 : if (err)
3192 0 : goto out_err;
3193 :
3194 245 : err = -EIO;
3195 245 : gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
3196 245 : if (!gdp)
3197 0 : goto out_err;
3198 :
3199 245 : ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
3200 : ext4_free_group_clusters(sb, gdp));
3201 :
3202 245 : BUFFER_TRACE(gdp_bh, "get_write_access");
3203 245 : err = ext4_journal_get_write_access(handle, gdp_bh);
3204 245 : if (err)
3205 0 : goto out_err;
3206 :
3207 245 : block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3208 :
3209 245 : len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3210 245 : if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
3211 0 : ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
3212 : "fs metadata", block, block+len);
3213 : /* File system mounted not to panic on error
3214 : * Fix the bitmap and return EFSCORRUPTED
3215 : * We leak some of the blocks here.
3216 : */
3217 0 : ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3218 0 : ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3219 : ac->ac_b_ex.fe_len);
3220 0 : ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3221 0 : err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3222 0 : if (!err)
3223 0 : err = -EFSCORRUPTED;
3224 0 : goto out_err;
3225 : }
3226 :
3227 245 : ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3228 : #ifdef AGGRESSIVE_CHECK
3229 : {
3230 : int i;
3231 : for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
3232 : BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
3233 : bitmap_bh->b_data));
3234 : }
3235 : }
3236 : #endif
3237 245 : ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
3238 : ac->ac_b_ex.fe_len);
3239 245 : if (ext4_has_group_desc_csum(sb) &&
3240 245 : (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3241 0 : gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3242 0 : ext4_free_group_clusters_set(sb, gdp,
3243 : ext4_free_clusters_after_init(sb,
3244 : ac->ac_b_ex.fe_group, gdp));
3245 : }
3246 245 : len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
3247 245 : ext4_free_group_clusters_set(sb, gdp, len);
3248 245 : ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
3249 245 : ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
3250 :
3251 245 : ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3252 245 : percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
3253 : /*
3254 : * Now reduce the dirty block count also. Should not go negative
3255 : */
3256 245 : if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
3257 : /* release all the reserved blocks if non delalloc */
3258 173 : percpu_counter_sub(&sbi->s_dirtyclusters_counter,
3259 : reserv_clstrs);
3260 :
3261 245 : if (sbi->s_log_groups_per_flex) {
3262 245 : ext4_group_t flex_group = ext4_flex_group(sbi,
3263 : ac->ac_b_ex.fe_group);
3264 735 : atomic64_sub(ac->ac_b_ex.fe_len,
3265 490 : &sbi_array_rcu_deref(sbi, s_flex_groups,
3266 : flex_group)->free_clusters);
3267 : }
3268 :
3269 245 : err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
3270 245 : if (err)
3271 0 : goto out_err;
3272 245 : err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3273 :
3274 245 : out_err:
3275 245 : brelse(bitmap_bh);
3276 245 : return err;
3277 : }
3278 :
3279 : /*
3280 : * Idempotent helper for Ext4 fast commit replay path to set the state of
3281 : * blocks in bitmaps and update counters.
3282 : */
3283 0 : void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
3284 : int len, int state)
3285 : {
3286 0 : struct buffer_head *bitmap_bh = NULL;
3287 0 : struct ext4_group_desc *gdp;
3288 0 : struct buffer_head *gdp_bh;
3289 0 : struct ext4_sb_info *sbi = EXT4_SB(sb);
3290 0 : ext4_group_t group;
3291 0 : ext4_grpblk_t blkoff;
3292 0 : int i, clen, err;
3293 0 : int already;
3294 :
3295 0 : clen = EXT4_B2C(sbi, len);
3296 :
3297 0 : ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
3298 0 : bitmap_bh = ext4_read_block_bitmap(sb, group);
3299 0 : if (IS_ERR(bitmap_bh)) {
3300 0 : err = PTR_ERR(bitmap_bh);
3301 0 : bitmap_bh = NULL;
3302 0 : goto out_err;
3303 : }
3304 :
3305 0 : err = -EIO;
3306 0 : gdp = ext4_get_group_desc(sb, group, &gdp_bh);
3307 0 : if (!gdp)
3308 0 : goto out_err;
3309 :
3310 0 : ext4_lock_group(sb, group);
3311 0 : already = 0;
3312 0 : for (i = 0; i < clen; i++)
3313 0 : if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == !state)
3314 0 : already++;
3315 :
3316 0 : if (state)
3317 0 : ext4_set_bits(bitmap_bh->b_data, blkoff, clen);
3318 : else
3319 0 : mb_test_and_clear_bits(bitmap_bh->b_data, blkoff, clen);
3320 0 : if (ext4_has_group_desc_csum(sb) &&
3321 0 : (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3322 0 : gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
3323 0 : ext4_free_group_clusters_set(sb, gdp,
3324 : ext4_free_clusters_after_init(sb,
3325 : group, gdp));
3326 : }
3327 0 : if (state)
3328 0 : clen = ext4_free_group_clusters(sb, gdp) - clen + already;
3329 : else
3330 0 : clen = ext4_free_group_clusters(sb, gdp) + clen - already;
3331 :
3332 0 : ext4_free_group_clusters_set(sb, gdp, clen);
3333 0 : ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
3334 0 : ext4_group_desc_csum_set(sb, group, gdp);
3335 :
3336 0 : ext4_unlock_group(sb, group);
3337 :
3338 0 : if (sbi->s_log_groups_per_flex) {
3339 0 : ext4_group_t flex_group = ext4_flex_group(sbi, group);
3340 :
3341 0 : atomic64_sub(len,
3342 0 : &sbi_array_rcu_deref(sbi, s_flex_groups,
3343 : flex_group)->free_clusters);
3344 : }
3345 :
3346 0 : err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
3347 0 : if (err)
3348 0 : goto out_err;
3349 0 : sync_dirty_buffer(bitmap_bh);
3350 0 : err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
3351 0 : sync_dirty_buffer(gdp_bh);
3352 :
3353 0 : out_err:
3354 0 : brelse(bitmap_bh);
3355 0 : }
3356 :
3357 : /*
3358 : * here we normalize request for locality group
3359 : * Group request are normalized to s_mb_group_prealloc, which goes to
3360 : * s_strip if we set the same via mount option.
3361 : * s_mb_group_prealloc can be configured via
3362 : * /sys/fs/ext4/<partition>/mb_group_prealloc
3363 : *
3364 : * XXX: should we try to preallocate more than the group has now?
3365 : */
3366 4 : static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3367 : {
3368 4 : struct super_block *sb = ac->ac_sb;
3369 4 : struct ext4_locality_group *lg = ac->ac_lg;
3370 :
3371 4 : BUG_ON(lg == NULL);
3372 4 : ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3373 4 : mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
3374 4 : }
3375 :
3376 : /*
3377 : * Normalization means making request better in terms of
3378 : * size and alignment
3379 : */
3380 : static noinline_for_stack void
3381 197 : ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3382 : struct ext4_allocation_request *ar)
3383 : {
3384 197 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3385 197 : int bsbits, max;
3386 197 : ext4_lblk_t end;
3387 197 : loff_t size, start_off;
3388 197 : loff_t orig_size __maybe_unused;
3389 197 : ext4_lblk_t start;
3390 197 : struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3391 197 : struct ext4_prealloc_space *pa;
3392 :
3393 : /* do normalize only data requests, metadata requests
3394 : do not need preallocation */
3395 197 : if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3396 : return;
3397 :
3398 : /* sometime caller may want exact blocks */
3399 22 : if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3400 : return;
3401 :
3402 : /* caller may indicate that preallocation isn't
3403 : * required (it's a tail, for example) */
3404 22 : if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3405 : return;
3406 :
3407 13 : if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3408 4 : ext4_mb_normalize_group_request(ac);
3409 4 : return ;
3410 : }
3411 :
3412 9 : bsbits = ac->ac_sb->s_blocksize_bits;
3413 :
3414 : /* first, let's learn actual file size
3415 : * given current request is allocated */
3416 9 : size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3417 9 : size = size << bsbits;
3418 9 : if (size < i_size_read(ac->ac_inode))
3419 : size = i_size_read(ac->ac_inode);
3420 9 : orig_size = size;
3421 :
3422 : /* max size of free chunks */
3423 9 : max = 2 << bsbits;
3424 :
3425 : #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3426 : (req <= (size) || max <= (chunk_size))
3427 :
3428 : /* first, try to predict filesize */
3429 : /* XXX: should this table be tunable? */
3430 9 : start_off = 0;
3431 9 : if (size <= 16 * 1024) {
3432 : size = 16 * 1024;
3433 9 : } else if (size <= 32 * 1024) {
3434 : size = 32 * 1024;
3435 9 : } else if (size <= 64 * 1024) {
3436 : size = 64 * 1024;
3437 9 : } else if (size <= 128 * 1024) {
3438 : size = 128 * 1024;
3439 9 : } else if (size <= 256 * 1024) {
3440 : size = 256 * 1024;
3441 8 : } else if (size <= 512 * 1024) {
3442 : size = 512 * 1024;
3443 0 : } else if (size <= 1024 * 1024) {
3444 : size = 1024 * 1024;
3445 0 : } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3446 0 : start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3447 0 : (21 - bsbits)) << 21;
3448 0 : size = 2 * 1024 * 1024;
3449 0 : } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3450 0 : start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3451 0 : (22 - bsbits)) << 22;
3452 0 : size = 4 * 1024 * 1024;
3453 0 : } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3454 : (8<<20)>>bsbits, max, 8 * 1024)) {
3455 0 : start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3456 0 : (23 - bsbits)) << 23;
3457 0 : size = 8 * 1024 * 1024;
3458 : } else {
3459 0 : start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3460 0 : size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3461 : ac->ac_o_ex.fe_len) << bsbits;
3462 : }
3463 9 : size = size >> bsbits;
3464 9 : start = start_off >> bsbits;
3465 :
3466 : /* don't cover already allocated blocks in selected range */
3467 9 : if (ar->pleft && start <= ar->lleft) {
3468 8 : size -= ar->lleft + 1 - start;
3469 8 : start = ar->lleft + 1;
3470 : }
3471 9 : if (ar->pright && start + size - 1 >= ar->lright)
3472 0 : size -= start + size - ar->lright;
3473 :
3474 : /*
3475 : * Trim allocation request for filesystems with artificially small
3476 : * groups.
3477 : */
3478 9 : if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3479 0 : size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3480 :
3481 9 : end = start + size;
3482 :
3483 : /* check we don't cross already preallocated blocks */
3484 9 : rcu_read_lock();
3485 9 : list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3486 0 : ext4_lblk_t pa_end;
3487 :
3488 0 : if (pa->pa_deleted)
3489 0 : continue;
3490 0 : spin_lock(&pa->pa_lock);
3491 0 : if (pa->pa_deleted) {
3492 0 : spin_unlock(&pa->pa_lock);
3493 0 : continue;
3494 : }
3495 :
3496 0 : pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3497 : pa->pa_len);
3498 :
3499 : /* PA must not overlap original request */
3500 0 : BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3501 : ac->ac_o_ex.fe_logical < pa->pa_lstart));
3502 :
3503 : /* skip PAs this normalized request doesn't overlap with */
3504 0 : if (pa->pa_lstart >= end || pa_end <= start) {
3505 0 : spin_unlock(&pa->pa_lock);
3506 0 : continue;
3507 : }
3508 0 : BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3509 :
3510 : /* adjust start or end to be adjacent to this pa */
3511 0 : if (pa_end <= ac->ac_o_ex.fe_logical) {
3512 : BUG_ON(pa_end < start);
3513 : start = pa_end;
3514 0 : } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3515 0 : BUG_ON(pa->pa_lstart > end);
3516 0 : end = pa->pa_lstart;
3517 : }
3518 0 : spin_unlock(&pa->pa_lock);
3519 : }
3520 9 : rcu_read_unlock();
3521 9 : size = end - start;
3522 :
3523 : /* XXX: extra loop to check we really don't overlap preallocations */
3524 9 : rcu_read_lock();
3525 9 : list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3526 0 : ext4_lblk_t pa_end;
3527 :
3528 0 : spin_lock(&pa->pa_lock);
3529 0 : if (pa->pa_deleted == 0) {
3530 0 : pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3531 : pa->pa_len);
3532 0 : BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3533 : }
3534 0 : spin_unlock(&pa->pa_lock);
3535 : }
3536 9 : rcu_read_unlock();
3537 :
3538 9 : if (start + size <= ac->ac_o_ex.fe_logical &&
3539 : start > ac->ac_o_ex.fe_logical) {
3540 0 : ext4_msg(ac->ac_sb, KERN_ERR,
3541 : "start %lu, size %lu, fe_logical %lu",
3542 : (unsigned long) start, (unsigned long) size,
3543 : (unsigned long) ac->ac_o_ex.fe_logical);
3544 0 : BUG();
3545 : }
3546 9 : BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3547 :
3548 : /* now prepare goal request */
3549 :
3550 : /* XXX: is it better to align blocks WRT to logical
3551 : * placement or satisfy big request as is */
3552 9 : ac->ac_g_ex.fe_logical = start;
3553 9 : ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3554 :
3555 : /* define goal start in order to merge */
3556 9 : if (ar->pright && (ar->lright == (start + size))) {
3557 : /* merge to the right */
3558 0 : ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3559 : &ac->ac_f_ex.fe_group,
3560 : &ac->ac_f_ex.fe_start);
3561 0 : ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3562 : }
3563 9 : if (ar->pleft && (ar->lleft + 1 == start)) {
3564 : /* merge to the left */
3565 8 : ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3566 : &ac->ac_f_ex.fe_group,
3567 : &ac->ac_f_ex.fe_start);
3568 8 : ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3569 : }
3570 :
3571 : mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
3572 : orig_size, start);
3573 : }
3574 :
3575 245 : static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3576 : {
3577 245 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3578 :
3579 245 : if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3580 0 : atomic_inc(&sbi->s_bal_reqs);
3581 0 : atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3582 0 : if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3583 0 : atomic_inc(&sbi->s_bal_success);
3584 0 : atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3585 0 : if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3586 0 : ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3587 0 : atomic_inc(&sbi->s_bal_goals);
3588 0 : if (ac->ac_found > sbi->s_mb_max_to_scan)
3589 0 : atomic_inc(&sbi->s_bal_breaks);
3590 : }
3591 :
3592 245 : if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3593 197 : trace_ext4_mballoc_alloc(ac);
3594 : else
3595 48 : trace_ext4_mballoc_prealloc(ac);
3596 245 : }
3597 :
3598 : /*
3599 : * Called on failure; free up any blocks from the inode PA for this
3600 : * context. We don't need this for MB_GROUP_PA because we only change
3601 : * pa_free in ext4_mb_release_context(), but on failure, we've already
3602 : * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3603 : */
3604 0 : static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3605 : {
3606 0 : struct ext4_prealloc_space *pa = ac->ac_pa;
3607 0 : struct ext4_buddy e4b;
3608 0 : int err;
3609 :
3610 0 : if (pa == NULL) {
3611 0 : if (ac->ac_f_ex.fe_len == 0)
3612 0 : return;
3613 0 : err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3614 0 : if (err) {
3615 : /*
3616 : * This should never happen since we pin the
3617 : * pages in the ext4_allocation_context so
3618 : * ext4_mb_load_buddy() should never fail.
3619 : */
3620 0 : WARN(1, "mb_load_buddy failed (%d)", err);
3621 0 : return;
3622 : }
3623 0 : ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3624 0 : mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3625 : ac->ac_f_ex.fe_len);
3626 0 : ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3627 0 : ext4_mb_unload_buddy(&e4b);
3628 0 : return;
3629 : }
3630 0 : if (pa->pa_type == MB_INODE_PA)
3631 0 : pa->pa_free += ac->ac_b_ex.fe_len;
3632 : }
3633 :
3634 : /*
3635 : * use blocks preallocated to inode
3636 : */
3637 2 : static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3638 : struct ext4_prealloc_space *pa)
3639 : {
3640 2 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3641 2 : ext4_fsblk_t start;
3642 2 : ext4_fsblk_t end;
3643 2 : int len;
3644 :
3645 : /* found preallocated blocks, use them */
3646 2 : start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3647 2 : end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3648 : start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3649 2 : len = EXT4_NUM_B2C(sbi, end - start);
3650 2 : ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3651 : &ac->ac_b_ex.fe_start);
3652 2 : ac->ac_b_ex.fe_len = len;
3653 2 : ac->ac_status = AC_STATUS_FOUND;
3654 2 : ac->ac_pa = pa;
3655 :
3656 2 : BUG_ON(start < pa->pa_pstart);
3657 2 : BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3658 2 : BUG_ON(pa->pa_free < len);
3659 2 : pa->pa_free -= len;
3660 :
3661 2 : mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
3662 2 : }
3663 :
3664 : /*
3665 : * use blocks preallocated to locality group
3666 : */
3667 52 : static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3668 : struct ext4_prealloc_space *pa)
3669 : {
3670 52 : unsigned int len = ac->ac_o_ex.fe_len;
3671 :
3672 52 : ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3673 : &ac->ac_b_ex.fe_group,
3674 : &ac->ac_b_ex.fe_start);
3675 52 : ac->ac_b_ex.fe_len = len;
3676 52 : ac->ac_status = AC_STATUS_FOUND;
3677 52 : ac->ac_pa = pa;
3678 :
3679 : /* we don't correct pa_pstart or pa_plen here to avoid
3680 : * possible race when the group is being loaded concurrently
3681 : * instead we correct pa later, after blocks are marked
3682 : * in on-disk bitmap -- see ext4_mb_release_context()
3683 : * Other CPUs are prevented from allocating from this pa by lg_mutex
3684 : */
3685 52 : mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
3686 : pa->pa_lstart-len, len, pa);
3687 52 : }
3688 :
3689 : /*
3690 : * Return the prealloc space that have minimal distance
3691 : * from the goal block. @cpa is the prealloc
3692 : * space that is having currently known minimal distance
3693 : * from the goal block.
3694 : */
3695 : static struct ext4_prealloc_space *
3696 48 : ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3697 : struct ext4_prealloc_space *pa,
3698 : struct ext4_prealloc_space *cpa)
3699 : {
3700 48 : ext4_fsblk_t cur_distance, new_distance;
3701 :
3702 48 : if (cpa == NULL) {
3703 48 : atomic_inc(&pa->pa_count);
3704 48 : return pa;
3705 : }
3706 0 : cur_distance = abs(goal_block - cpa->pa_pstart);
3707 0 : new_distance = abs(goal_block - pa->pa_pstart);
3708 :
3709 0 : if (cur_distance <= new_distance)
3710 : return cpa;
3711 :
3712 : /* drop the previous reference */
3713 0 : atomic_dec(&cpa->pa_count);
3714 0 : atomic_inc(&pa->pa_count);
3715 0 : return pa;
3716 : }
3717 :
3718 : /*
3719 : * search goal blocks in preallocated space
3720 : */
3721 : static noinline_for_stack bool
3722 245 : ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3723 : {
3724 245 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3725 245 : int order, i;
3726 245 : struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3727 245 : struct ext4_locality_group *lg;
3728 245 : struct ext4_prealloc_space *pa, *cpa = NULL;
3729 245 : ext4_fsblk_t goal_block;
3730 :
3731 : /* only data can be preallocated */
3732 245 : if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3733 : return false;
3734 :
3735 : /* first, try per-file preallocation */
3736 70 : rcu_read_lock();
3737 70 : list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3738 :
3739 : /* all fields in this condition don't change,
3740 : * so we can skip locking for them */
3741 0 : if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3742 0 : ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3743 0 : EXT4_C2B(sbi, pa->pa_len)))
3744 0 : continue;
3745 :
3746 : /* non-extent files can't have physical blocks past 2^32 */
3747 0 : if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3748 0 : (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3749 : EXT4_MAX_BLOCK_FILE_PHYS))
3750 0 : continue;
3751 :
3752 : /* found preallocated blocks, use them */
3753 0 : spin_lock(&pa->pa_lock);
3754 0 : if (pa->pa_deleted == 0 && pa->pa_free) {
3755 0 : atomic_inc(&pa->pa_count);
3756 0 : ext4_mb_use_inode_pa(ac, pa);
3757 0 : spin_unlock(&pa->pa_lock);
3758 0 : ac->ac_criteria = 10;
3759 0 : rcu_read_unlock();
3760 0 : return true;
3761 : }
3762 0 : spin_unlock(&pa->pa_lock);
3763 : }
3764 70 : rcu_read_unlock();
3765 :
3766 : /* can we use group allocation? */
3767 70 : if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3768 : return false;
3769 :
3770 : /* inode may have no locality group for some reason */
3771 52 : lg = ac->ac_lg;
3772 52 : if (lg == NULL)
3773 : return false;
3774 52 : order = fls(ac->ac_o_ex.fe_len) - 1;
3775 52 : if (order > PREALLOC_TB_SIZE - 1)
3776 : /* The max size of hash table is PREALLOC_TB_SIZE */
3777 : order = PREALLOC_TB_SIZE - 1;
3778 :
3779 52 : goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3780 : /*
3781 : * search for the prealloc space that is having
3782 : * minimal distance from the goal block.
3783 : */
3784 503 : for (i = order; i < PREALLOC_TB_SIZE; i++) {
3785 451 : rcu_read_lock();
3786 499 : list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3787 : pa_inode_list) {
3788 48 : spin_lock(&pa->pa_lock);
3789 48 : if (pa->pa_deleted == 0 &&
3790 48 : pa->pa_free >= ac->ac_o_ex.fe_len) {
3791 :
3792 48 : cpa = ext4_mb_check_group_pa(goal_block,
3793 : pa, cpa);
3794 : }
3795 48 : spin_unlock(&pa->pa_lock);
3796 : }
3797 451 : rcu_read_unlock();
3798 : }
3799 52 : if (cpa) {
3800 48 : ext4_mb_use_group_pa(ac, cpa);
3801 48 : ac->ac_criteria = 20;
3802 48 : return true;
3803 : }
3804 : return false;
3805 : }
3806 :
3807 : /*
3808 : * the function goes through all block freed in the group
3809 : * but not yet committed and marks them used in in-core bitmap.
3810 : * buddy must be generated from this bitmap
3811 : * Need to be called with the ext4 group lock held
3812 : */
3813 16 : static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3814 : ext4_group_t group)
3815 : {
3816 16 : struct rb_node *n;
3817 16 : struct ext4_group_info *grp;
3818 16 : struct ext4_free_data *entry;
3819 :
3820 16 : grp = ext4_get_group_info(sb, group);
3821 16 : n = rb_first(&(grp->bb_free_root));
3822 :
3823 16 : while (n) {
3824 0 : entry = rb_entry(n, struct ext4_free_data, efd_node);
3825 0 : ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3826 0 : n = rb_next(n);
3827 : }
3828 16 : return;
3829 : }
3830 :
3831 : /*
3832 : * the function goes through all preallocation in this group and marks them
3833 : * used in in-core bitmap. buddy must be generated from this bitmap
3834 : * Need to be called with ext4 group lock held
3835 : */
3836 : static noinline_for_stack
3837 16 : void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3838 : ext4_group_t group)
3839 : {
3840 16 : struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3841 16 : struct ext4_prealloc_space *pa;
3842 16 : struct list_head *cur;
3843 16 : ext4_group_t groupnr;
3844 16 : ext4_grpblk_t start;
3845 16 : int preallocated = 0;
3846 16 : int len;
3847 :
3848 : /* all form of preallocation discards first load group,
3849 : * so the only competing code is preallocation use.
3850 : * we don't need any locking here
3851 : * notice we do NOT ignore preallocations with pa_deleted
3852 : * otherwise we could leave used blocks available for
3853 : * allocation in buddy when concurrent ext4_mb_put_pa()
3854 : * is dropping preallocation
3855 : */
3856 16 : list_for_each(cur, &grp->bb_prealloc_list) {
3857 0 : pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3858 0 : spin_lock(&pa->pa_lock);
3859 0 : ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3860 : &groupnr, &start);
3861 0 : len = pa->pa_len;
3862 0 : spin_unlock(&pa->pa_lock);
3863 0 : if (unlikely(len == 0))
3864 0 : continue;
3865 0 : BUG_ON(groupnr != group);
3866 0 : ext4_set_bits(bitmap, start, len);
3867 0 : preallocated += len;
3868 : }
3869 16 : mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
3870 16 : }
3871 :
3872 2 : static void ext4_mb_mark_pa_deleted(struct super_block *sb,
3873 : struct ext4_prealloc_space *pa)
3874 : {
3875 2 : struct ext4_inode_info *ei;
3876 :
3877 2 : if (pa->pa_deleted) {
3878 0 : ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
3879 : pa->pa_type, pa->pa_pstart, pa->pa_lstart,
3880 : pa->pa_len);
3881 0 : return;
3882 : }
3883 :
3884 2 : pa->pa_deleted = 1;
3885 :
3886 2 : if (pa->pa_type == MB_INODE_PA) {
3887 2 : ei = EXT4_I(pa->pa_inode);
3888 2 : atomic_dec(&ei->i_prealloc_active);
3889 : }
3890 : }
3891 :
3892 2 : static void ext4_mb_pa_callback(struct rcu_head *head)
3893 : {
3894 2 : struct ext4_prealloc_space *pa;
3895 2 : pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3896 :
3897 2 : BUG_ON(atomic_read(&pa->pa_count));
3898 2 : BUG_ON(pa->pa_deleted == 0);
3899 2 : kmem_cache_free(ext4_pspace_cachep, pa);
3900 2 : }
3901 :
3902 : /*
3903 : * drops a reference to preallocated space descriptor
3904 : * if this was the last reference and the space is consumed
3905 : */
3906 54 : static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3907 : struct super_block *sb, struct ext4_prealloc_space *pa)
3908 : {
3909 54 : ext4_group_t grp;
3910 54 : ext4_fsblk_t grp_blk;
3911 :
3912 : /* in this short window concurrent discard can set pa_deleted */
3913 54 : spin_lock(&pa->pa_lock);
3914 108 : if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3915 54 : spin_unlock(&pa->pa_lock);
3916 54 : return;
3917 : }
3918 :
3919 0 : if (pa->pa_deleted == 1) {
3920 0 : spin_unlock(&pa->pa_lock);
3921 0 : return;
3922 : }
3923 :
3924 0 : ext4_mb_mark_pa_deleted(sb, pa);
3925 0 : spin_unlock(&pa->pa_lock);
3926 :
3927 0 : grp_blk = pa->pa_pstart;
3928 : /*
3929 : * If doing group-based preallocation, pa_pstart may be in the
3930 : * next group when pa is used up
3931 : */
3932 0 : if (pa->pa_type == MB_GROUP_PA)
3933 0 : grp_blk--;
3934 :
3935 0 : grp = ext4_get_group_number(sb, grp_blk);
3936 :
3937 : /*
3938 : * possible race:
3939 : *
3940 : * P1 (buddy init) P2 (regular allocation)
3941 : * find block B in PA
3942 : * copy on-disk bitmap to buddy
3943 : * mark B in on-disk bitmap
3944 : * drop PA from group
3945 : * mark all PAs in buddy
3946 : *
3947 : * thus, P1 initializes buddy with B available. to prevent this
3948 : * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3949 : * against that pair
3950 : */
3951 0 : ext4_lock_group(sb, grp);
3952 0 : list_del(&pa->pa_group_list);
3953 0 : ext4_unlock_group(sb, grp);
3954 :
3955 0 : spin_lock(pa->pa_obj_lock);
3956 0 : list_del_rcu(&pa->pa_inode_list);
3957 0 : spin_unlock(pa->pa_obj_lock);
3958 :
3959 0 : call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3960 : }
3961 :
3962 : /*
3963 : * creates new preallocated space for given inode
3964 : */
3965 : static noinline_for_stack void
3966 2 : ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3967 : {
3968 2 : struct super_block *sb = ac->ac_sb;
3969 2 : struct ext4_sb_info *sbi = EXT4_SB(sb);
3970 2 : struct ext4_prealloc_space *pa;
3971 2 : struct ext4_group_info *grp;
3972 2 : struct ext4_inode_info *ei;
3973 :
3974 : /* preallocate only when found space is larger then requested */
3975 2 : BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3976 2 : BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3977 2 : BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3978 2 : BUG_ON(ac->ac_pa == NULL);
3979 :
3980 2 : pa = ac->ac_pa;
3981 :
3982 2 : if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3983 1 : int winl;
3984 1 : int wins;
3985 1 : int win;
3986 1 : int offs;
3987 :
3988 : /* we can't allocate as much as normalizer wants.
3989 : * so, found space must get proper lstart
3990 : * to cover original request */
3991 1 : BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3992 1 : BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3993 :
3994 : /* we're limited by original request in that
3995 : * logical block must be covered any way
3996 : * winl is window we can move our chunk within */
3997 1 : winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3998 :
3999 : /* also, we should cover whole original request */
4000 1 : wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
4001 :
4002 : /* the smallest one defines real window */
4003 1 : win = min(winl, wins);
4004 :
4005 1 : offs = ac->ac_o_ex.fe_logical %
4006 1 : EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4007 1 : if (offs && offs < win)
4008 0 : win = offs;
4009 :
4010 1 : ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
4011 1 : EXT4_NUM_B2C(sbi, win);
4012 1 : BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
4013 1 : BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
4014 : }
4015 :
4016 : /* preallocation can change ac_b_ex, thus we store actually
4017 : * allocated blocks for history */
4018 2 : ac->ac_f_ex = ac->ac_b_ex;
4019 :
4020 2 : pa->pa_lstart = ac->ac_b_ex.fe_logical;
4021 2 : pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4022 2 : pa->pa_len = ac->ac_b_ex.fe_len;
4023 2 : pa->pa_free = pa->pa_len;
4024 2 : spin_lock_init(&pa->pa_lock);
4025 2 : INIT_LIST_HEAD(&pa->pa_inode_list);
4026 2 : INIT_LIST_HEAD(&pa->pa_group_list);
4027 2 : pa->pa_deleted = 0;
4028 2 : pa->pa_type = MB_INODE_PA;
4029 :
4030 2 : mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4031 : pa->pa_len, pa->pa_lstart);
4032 2 : trace_ext4_mb_new_inode_pa(ac, pa);
4033 :
4034 2 : ext4_mb_use_inode_pa(ac, pa);
4035 2 : atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
4036 :
4037 2 : ei = EXT4_I(ac->ac_inode);
4038 2 : grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4039 :
4040 2 : pa->pa_obj_lock = &ei->i_prealloc_lock;
4041 2 : pa->pa_inode = ac->ac_inode;
4042 :
4043 2 : list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4044 :
4045 2 : spin_lock(pa->pa_obj_lock);
4046 2 : list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
4047 2 : spin_unlock(pa->pa_obj_lock);
4048 2 : atomic_inc(&ei->i_prealloc_active);
4049 2 : }
4050 :
4051 : /*
4052 : * creates new preallocated space for locality group inodes belongs to
4053 : */
4054 : static noinline_for_stack void
4055 4 : ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
4056 : {
4057 4 : struct super_block *sb = ac->ac_sb;
4058 4 : struct ext4_locality_group *lg;
4059 4 : struct ext4_prealloc_space *pa;
4060 4 : struct ext4_group_info *grp;
4061 :
4062 : /* preallocate only when found space is larger then requested */
4063 4 : BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
4064 4 : BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4065 4 : BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
4066 4 : BUG_ON(ac->ac_pa == NULL);
4067 :
4068 4 : pa = ac->ac_pa;
4069 :
4070 : /* preallocation can change ac_b_ex, thus we store actually
4071 : * allocated blocks for history */
4072 4 : ac->ac_f_ex = ac->ac_b_ex;
4073 :
4074 4 : pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4075 4 : pa->pa_lstart = pa->pa_pstart;
4076 4 : pa->pa_len = ac->ac_b_ex.fe_len;
4077 4 : pa->pa_free = pa->pa_len;
4078 4 : spin_lock_init(&pa->pa_lock);
4079 4 : INIT_LIST_HEAD(&pa->pa_inode_list);
4080 4 : INIT_LIST_HEAD(&pa->pa_group_list);
4081 4 : pa->pa_deleted = 0;
4082 4 : pa->pa_type = MB_GROUP_PA;
4083 :
4084 4 : mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
4085 : pa->pa_len, pa->pa_lstart);
4086 4 : trace_ext4_mb_new_group_pa(ac, pa);
4087 :
4088 4 : ext4_mb_use_group_pa(ac, pa);
4089 4 : atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
4090 :
4091 4 : grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
4092 4 : lg = ac->ac_lg;
4093 4 : BUG_ON(lg == NULL);
4094 :
4095 4 : pa->pa_obj_lock = &lg->lg_prealloc_lock;
4096 4 : pa->pa_inode = NULL;
4097 :
4098 4 : list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
4099 :
4100 : /*
4101 : * We will later add the new pa to the right bucket
4102 : * after updating the pa_free in ext4_mb_release_context
4103 : */
4104 4 : }
4105 :
4106 6 : static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
4107 : {
4108 6 : if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4109 4 : ext4_mb_new_group_pa(ac);
4110 : else
4111 2 : ext4_mb_new_inode_pa(ac);
4112 6 : }
4113 :
4114 : /*
4115 : * finds all unused blocks in on-disk bitmap, frees them in
4116 : * in-core bitmap and buddy.
4117 : * @pa must be unlinked from inode and group lists, so that
4118 : * nobody else can find/use it.
4119 : * the caller MUST hold group/inode locks.
4120 : * TODO: optimize the case when there are no in-core structures yet
4121 : */
4122 : static noinline_for_stack int
4123 2 : ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
4124 : struct ext4_prealloc_space *pa)
4125 : {
4126 2 : struct super_block *sb = e4b->bd_sb;
4127 2 : struct ext4_sb_info *sbi = EXT4_SB(sb);
4128 2 : unsigned int end;
4129 2 : unsigned int next;
4130 2 : ext4_group_t group;
4131 2 : ext4_grpblk_t bit;
4132 2 : unsigned long long grp_blk_start;
4133 2 : int free = 0;
4134 :
4135 2 : BUG_ON(pa->pa_deleted == 0);
4136 2 : ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4137 2 : grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
4138 2 : BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4139 2 : end = bit + pa->pa_len;
4140 :
4141 4 : while (bit < end) {
4142 2 : bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
4143 2 : if (bit >= end)
4144 : break;
4145 2 : next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
4146 2 : mb_debug(sb, "free preallocated %u/%u in group %u\n",
4147 : (unsigned) ext4_group_first_block_no(sb, group) + bit,
4148 : (unsigned) next - bit, (unsigned) group);
4149 2 : free += next - bit;
4150 :
4151 2 : trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
4152 2 : trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
4153 2 : EXT4_C2B(sbi, bit)),
4154 : next - bit);
4155 2 : mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
4156 2 : bit = next + 1;
4157 : }
4158 2 : if (free != pa->pa_free) {
4159 0 : ext4_msg(e4b->bd_sb, KERN_CRIT,
4160 : "pa %p: logic %lu, phys. %lu, len %d",
4161 : pa, (unsigned long) pa->pa_lstart,
4162 : (unsigned long) pa->pa_pstart,
4163 : pa->pa_len);
4164 0 : ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
4165 : free, pa->pa_free);
4166 : /*
4167 : * pa is already deleted so we use the value obtained
4168 : * from the bitmap and continue.
4169 : */
4170 : }
4171 2 : atomic_add(free, &sbi->s_mb_discarded);
4172 :
4173 2 : return 0;
4174 : }
4175 :
4176 : static noinline_for_stack int
4177 0 : ext4_mb_release_group_pa(struct ext4_buddy *e4b,
4178 : struct ext4_prealloc_space *pa)
4179 : {
4180 0 : struct super_block *sb = e4b->bd_sb;
4181 0 : ext4_group_t group;
4182 0 : ext4_grpblk_t bit;
4183 :
4184 0 : trace_ext4_mb_release_group_pa(sb, pa);
4185 0 : BUG_ON(pa->pa_deleted == 0);
4186 0 : ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
4187 0 : BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
4188 0 : mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
4189 0 : atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
4190 0 : trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
4191 :
4192 0 : return 0;
4193 : }
4194 :
4195 : /*
4196 : * releases all preallocations in given group
4197 : *
4198 : * first, we need to decide discard policy:
4199 : * - when do we discard
4200 : * 1) ENOSPC
4201 : * - how many do we discard
4202 : * 1) how many requested
4203 : */
4204 : static noinline_for_stack int
4205 0 : ext4_mb_discard_group_preallocations(struct super_block *sb,
4206 : ext4_group_t group, int needed)
4207 : {
4208 0 : struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4209 0 : struct buffer_head *bitmap_bh = NULL;
4210 0 : struct ext4_prealloc_space *pa, *tmp;
4211 0 : struct list_head list;
4212 0 : struct ext4_buddy e4b;
4213 0 : int err;
4214 0 : int busy = 0;
4215 0 : int free, free_total = 0;
4216 :
4217 0 : mb_debug(sb, "discard preallocation for group %u\n", group);
4218 0 : if (list_empty(&grp->bb_prealloc_list))
4219 0 : goto out_dbg;
4220 :
4221 0 : bitmap_bh = ext4_read_block_bitmap(sb, group);
4222 0 : if (IS_ERR(bitmap_bh)) {
4223 0 : err = PTR_ERR(bitmap_bh);
4224 0 : ext4_error_err(sb, -err,
4225 : "Error %d reading block bitmap for %u",
4226 : err, group);
4227 0 : goto out_dbg;
4228 : }
4229 :
4230 0 : err = ext4_mb_load_buddy(sb, group, &e4b);
4231 0 : if (err) {
4232 0 : ext4_warning(sb, "Error %d loading buddy information for %u",
4233 : err, group);
4234 0 : put_bh(bitmap_bh);
4235 0 : goto out_dbg;
4236 : }
4237 :
4238 0 : if (needed == 0)
4239 0 : needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
4240 :
4241 0 : INIT_LIST_HEAD(&list);
4242 0 : repeat:
4243 0 : free = 0;
4244 0 : ext4_lock_group(sb, group);
4245 0 : list_for_each_entry_safe(pa, tmp,
4246 : &grp->bb_prealloc_list, pa_group_list) {
4247 0 : spin_lock(&pa->pa_lock);
4248 0 : if (atomic_read(&pa->pa_count)) {
4249 0 : spin_unlock(&pa->pa_lock);
4250 0 : busy = 1;
4251 0 : continue;
4252 : }
4253 0 : if (pa->pa_deleted) {
4254 0 : spin_unlock(&pa->pa_lock);
4255 0 : continue;
4256 : }
4257 :
4258 : /* seems this one can be freed ... */
4259 0 : ext4_mb_mark_pa_deleted(sb, pa);
4260 :
4261 0 : if (!free)
4262 0 : this_cpu_inc(discard_pa_seq);
4263 :
4264 : /* we can trust pa_free ... */
4265 0 : free += pa->pa_free;
4266 :
4267 0 : spin_unlock(&pa->pa_lock);
4268 :
4269 0 : list_del(&pa->pa_group_list);
4270 0 : list_add(&pa->u.pa_tmp_list, &list);
4271 : }
4272 :
4273 : /* now free all selected PAs */
4274 0 : list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4275 :
4276 : /* remove from object (inode or locality group) */
4277 0 : spin_lock(pa->pa_obj_lock);
4278 0 : list_del_rcu(&pa->pa_inode_list);
4279 0 : spin_unlock(pa->pa_obj_lock);
4280 :
4281 0 : if (pa->pa_type == MB_GROUP_PA)
4282 0 : ext4_mb_release_group_pa(&e4b, pa);
4283 : else
4284 0 : ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4285 :
4286 0 : list_del(&pa->u.pa_tmp_list);
4287 0 : call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4288 : }
4289 :
4290 0 : free_total += free;
4291 :
4292 : /* if we still need more blocks and some PAs were used, try again */
4293 0 : if (free_total < needed && busy) {
4294 0 : ext4_unlock_group(sb, group);
4295 0 : cond_resched();
4296 0 : busy = 0;
4297 0 : goto repeat;
4298 : }
4299 0 : ext4_unlock_group(sb, group);
4300 0 : ext4_mb_unload_buddy(&e4b);
4301 0 : put_bh(bitmap_bh);
4302 0 : out_dbg:
4303 0 : mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
4304 : free_total, group, grp->bb_free);
4305 0 : return free_total;
4306 : }
4307 :
4308 : /*
4309 : * releases all non-used preallocated blocks for given inode
4310 : *
4311 : * It's important to discard preallocations under i_data_sem
4312 : * We don't want another block to be served from the prealloc
4313 : * space when we are discarding the inode prealloc space.
4314 : *
4315 : * FIXME!! Make sure it is valid at all the call sites
4316 : */
4317 449 : void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
4318 : {
4319 449 : struct ext4_inode_info *ei = EXT4_I(inode);
4320 449 : struct super_block *sb = inode->i_sb;
4321 449 : struct buffer_head *bitmap_bh = NULL;
4322 449 : struct ext4_prealloc_space *pa, *tmp;
4323 449 : ext4_group_t group = 0;
4324 449 : struct list_head list;
4325 449 : struct ext4_buddy e4b;
4326 449 : int err;
4327 :
4328 449 : if (!S_ISREG(inode->i_mode)) {
4329 : /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4330 142 : return;
4331 : }
4332 :
4333 307 : if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
4334 : return;
4335 :
4336 307 : mb_debug(sb, "discard preallocation for inode %lu\n",
4337 : inode->i_ino);
4338 614 : trace_ext4_discard_preallocations(inode,
4339 307 : atomic_read(&ei->i_prealloc_active), needed);
4340 :
4341 307 : INIT_LIST_HEAD(&list);
4342 :
4343 307 : if (needed == 0)
4344 307 : needed = UINT_MAX;
4345 :
4346 0 : repeat:
4347 : /* first, collect all pa's in the inode */
4348 307 : spin_lock(&ei->i_prealloc_lock);
4349 309 : while (!list_empty(&ei->i_prealloc_list) && needed) {
4350 2 : pa = list_entry(ei->i_prealloc_list.prev,
4351 : struct ext4_prealloc_space, pa_inode_list);
4352 2 : BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
4353 2 : spin_lock(&pa->pa_lock);
4354 2 : if (atomic_read(&pa->pa_count)) {
4355 : /* this shouldn't happen often - nobody should
4356 : * use preallocation while we're discarding it */
4357 0 : spin_unlock(&pa->pa_lock);
4358 0 : spin_unlock(&ei->i_prealloc_lock);
4359 0 : ext4_msg(sb, KERN_ERR,
4360 : "uh-oh! used pa while discarding");
4361 0 : WARN_ON(1);
4362 0 : schedule_timeout_uninterruptible(HZ);
4363 0 : goto repeat;
4364 :
4365 : }
4366 2 : if (pa->pa_deleted == 0) {
4367 2 : ext4_mb_mark_pa_deleted(sb, pa);
4368 2 : spin_unlock(&pa->pa_lock);
4369 2 : list_del_rcu(&pa->pa_inode_list);
4370 2 : list_add(&pa->u.pa_tmp_list, &list);
4371 2 : needed--;
4372 2 : continue;
4373 : }
4374 :
4375 : /* someone is deleting pa right now */
4376 0 : spin_unlock(&pa->pa_lock);
4377 0 : spin_unlock(&ei->i_prealloc_lock);
4378 :
4379 : /* we have to wait here because pa_deleted
4380 : * doesn't mean pa is already unlinked from
4381 : * the list. as we might be called from
4382 : * ->clear_inode() the inode will get freed
4383 : * and concurrent thread which is unlinking
4384 : * pa from inode's list may access already
4385 : * freed memory, bad-bad-bad */
4386 :
4387 : /* XXX: if this happens too often, we can
4388 : * add a flag to force wait only in case
4389 : * of ->clear_inode(), but not in case of
4390 : * regular truncate */
4391 0 : schedule_timeout_uninterruptible(HZ);
4392 0 : goto repeat;
4393 : }
4394 307 : spin_unlock(&ei->i_prealloc_lock);
4395 :
4396 309 : list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4397 2 : BUG_ON(pa->pa_type != MB_INODE_PA);
4398 2 : group = ext4_get_group_number(sb, pa->pa_pstart);
4399 :
4400 2 : err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4401 : GFP_NOFS|__GFP_NOFAIL);
4402 2 : if (err) {
4403 0 : ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4404 : err, group);
4405 0 : continue;
4406 : }
4407 :
4408 2 : bitmap_bh = ext4_read_block_bitmap(sb, group);
4409 2 : if (IS_ERR(bitmap_bh)) {
4410 0 : err = PTR_ERR(bitmap_bh);
4411 0 : ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
4412 : err, group);
4413 0 : ext4_mb_unload_buddy(&e4b);
4414 0 : continue;
4415 : }
4416 :
4417 2 : ext4_lock_group(sb, group);
4418 2 : list_del(&pa->pa_group_list);
4419 2 : ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4420 2 : ext4_unlock_group(sb, group);
4421 :
4422 2 : ext4_mb_unload_buddy(&e4b);
4423 2 : put_bh(bitmap_bh);
4424 :
4425 2 : list_del(&pa->u.pa_tmp_list);
4426 2 : call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4427 : }
4428 : }
4429 :
4430 197 : static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
4431 : {
4432 197 : struct ext4_prealloc_space *pa;
4433 :
4434 197 : BUG_ON(ext4_pspace_cachep == NULL);
4435 197 : pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
4436 197 : if (!pa)
4437 : return -ENOMEM;
4438 197 : atomic_set(&pa->pa_count, 1);
4439 197 : ac->ac_pa = pa;
4440 197 : return 0;
4441 : }
4442 :
4443 191 : static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
4444 : {
4445 191 : struct ext4_prealloc_space *pa = ac->ac_pa;
4446 :
4447 191 : BUG_ON(!pa);
4448 191 : ac->ac_pa = NULL;
4449 382 : WARN_ON(!atomic_dec_and_test(&pa->pa_count));
4450 191 : kmem_cache_free(ext4_pspace_cachep, pa);
4451 191 : }
4452 :
4453 : #ifdef CONFIG_EXT4_DEBUG
4454 : static inline void ext4_mb_show_pa(struct super_block *sb)
4455 : {
4456 : ext4_group_t i, ngroups;
4457 :
4458 : if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
4459 : return;
4460 :
4461 : ngroups = ext4_get_groups_count(sb);
4462 : mb_debug(sb, "groups: ");
4463 : for (i = 0; i < ngroups; i++) {
4464 : struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4465 : struct ext4_prealloc_space *pa;
4466 : ext4_grpblk_t start;
4467 : struct list_head *cur;
4468 : ext4_lock_group(sb, i);
4469 : list_for_each(cur, &grp->bb_prealloc_list) {
4470 : pa = list_entry(cur, struct ext4_prealloc_space,
4471 : pa_group_list);
4472 : spin_lock(&pa->pa_lock);
4473 : ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4474 : NULL, &start);
4475 : spin_unlock(&pa->pa_lock);
4476 : mb_debug(sb, "PA:%u:%d:%d\n", i, start,
4477 : pa->pa_len);
4478 : }
4479 : ext4_unlock_group(sb, i);
4480 : mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
4481 : grp->bb_fragments);
4482 : }
4483 : }
4484 :
4485 : static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4486 : {
4487 : struct super_block *sb = ac->ac_sb;
4488 :
4489 : if (ext4_test_mount_flag(sb, EXT4_MF_FS_ABORTED))
4490 : return;
4491 :
4492 : mb_debug(sb, "Can't allocate:"
4493 : " Allocation context details:");
4494 : mb_debug(sb, "status %u flags 0x%x",
4495 : ac->ac_status, ac->ac_flags);
4496 : mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
4497 : "goal %lu/%lu/%lu@%lu, "
4498 : "best %lu/%lu/%lu@%lu cr %d",
4499 : (unsigned long)ac->ac_o_ex.fe_group,
4500 : (unsigned long)ac->ac_o_ex.fe_start,
4501 : (unsigned long)ac->ac_o_ex.fe_len,
4502 : (unsigned long)ac->ac_o_ex.fe_logical,
4503 : (unsigned long)ac->ac_g_ex.fe_group,
4504 : (unsigned long)ac->ac_g_ex.fe_start,
4505 : (unsigned long)ac->ac_g_ex.fe_len,
4506 : (unsigned long)ac->ac_g_ex.fe_logical,
4507 : (unsigned long)ac->ac_b_ex.fe_group,
4508 : (unsigned long)ac->ac_b_ex.fe_start,
4509 : (unsigned long)ac->ac_b_ex.fe_len,
4510 : (unsigned long)ac->ac_b_ex.fe_logical,
4511 : (int)ac->ac_criteria);
4512 : mb_debug(sb, "%u found", ac->ac_found);
4513 : ext4_mb_show_pa(sb);
4514 : }
4515 : #else
4516 0 : static inline void ext4_mb_show_pa(struct super_block *sb)
4517 : {
4518 0 : return;
4519 : }
4520 0 : static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4521 : {
4522 0 : ext4_mb_show_pa(ac->ac_sb);
4523 0 : return;
4524 : }
4525 : #endif
4526 :
4527 : /*
4528 : * We use locality group preallocation for small size file. The size of the
4529 : * file is determined by the current size or the resulting size after
4530 : * allocation which ever is larger
4531 : *
4532 : * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4533 : */
4534 245 : static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4535 : {
4536 245 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4537 245 : int bsbits = ac->ac_sb->s_blocksize_bits;
4538 245 : loff_t size, isize;
4539 :
4540 245 : if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4541 : return;
4542 :
4543 70 : if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4544 : return;
4545 :
4546 70 : size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4547 70 : isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4548 70 : >> bsbits;
4549 :
4550 140 : if ((size == isize) && !ext4_fs_is_busy(sbi) &&
4551 70 : !inode_is_open_for_write(ac->ac_inode)) {
4552 9 : ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4553 9 : return;
4554 : }
4555 :
4556 61 : if (sbi->s_mb_group_prealloc <= 0) {
4557 0 : ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4558 0 : return;
4559 : }
4560 :
4561 : /* don't use group allocation for large files */
4562 61 : size = max(size, isize);
4563 61 : if (size > sbi->s_mb_stream_request) {
4564 9 : ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4565 9 : return;
4566 : }
4567 :
4568 52 : BUG_ON(ac->ac_lg != NULL);
4569 : /*
4570 : * locality group prealloc space are per cpu. The reason for having
4571 : * per cpu locality group is to reduce the contention between block
4572 : * request from multiple CPUs.
4573 : */
4574 52 : ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4575 :
4576 : /* we're going to use group allocation */
4577 52 : ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4578 :
4579 : /* serialize all allocations in the group */
4580 52 : mutex_lock(&ac->ac_lg->lg_mutex);
4581 : }
4582 :
4583 : static noinline_for_stack int
4584 245 : ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4585 : struct ext4_allocation_request *ar)
4586 : {
4587 245 : struct super_block *sb = ar->inode->i_sb;
4588 245 : struct ext4_sb_info *sbi = EXT4_SB(sb);
4589 245 : struct ext4_super_block *es = sbi->s_es;
4590 245 : ext4_group_t group;
4591 245 : unsigned int len;
4592 245 : ext4_fsblk_t goal;
4593 245 : ext4_grpblk_t block;
4594 :
4595 : /* we can't allocate > group size */
4596 245 : len = ar->len;
4597 :
4598 : /* just a dirty hack to filter too big requests */
4599 245 : if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4600 0 : len = EXT4_CLUSTERS_PER_GROUP(sb);
4601 :
4602 : /* start searching from the goal */
4603 245 : goal = ar->goal;
4604 245 : if (goal < le32_to_cpu(es->s_first_data_block) ||
4605 245 : goal >= ext4_blocks_count(es))
4606 : goal = le32_to_cpu(es->s_first_data_block);
4607 245 : ext4_get_group_no_and_offset(sb, goal, &group, &block);
4608 :
4609 : /* set up allocation goals */
4610 245 : ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4611 245 : ac->ac_status = AC_STATUS_CONTINUE;
4612 245 : ac->ac_sb = sb;
4613 245 : ac->ac_inode = ar->inode;
4614 245 : ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4615 245 : ac->ac_o_ex.fe_group = group;
4616 245 : ac->ac_o_ex.fe_start = block;
4617 245 : ac->ac_o_ex.fe_len = len;
4618 245 : ac->ac_g_ex = ac->ac_o_ex;
4619 245 : ac->ac_flags = ar->flags;
4620 :
4621 : /* we have to define context: we'll work with a file or
4622 : * locality group. this is a policy, actually */
4623 245 : ext4_mb_group_or_file(ac);
4624 :
4625 245 : mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
4626 : "left: %u/%u, right %u/%u to %swritable\n",
4627 : (unsigned) ar->len, (unsigned) ar->logical,
4628 : (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4629 : (unsigned) ar->lleft, (unsigned) ar->pleft,
4630 : (unsigned) ar->lright, (unsigned) ar->pright,
4631 : inode_is_open_for_write(ar->inode) ? "" : "non-");
4632 245 : return 0;
4633 :
4634 : }
4635 :
4636 : static noinline_for_stack void
4637 0 : ext4_mb_discard_lg_preallocations(struct super_block *sb,
4638 : struct ext4_locality_group *lg,
4639 : int order, int total_entries)
4640 : {
4641 0 : ext4_group_t group = 0;
4642 0 : struct ext4_buddy e4b;
4643 0 : struct list_head discard_list;
4644 0 : struct ext4_prealloc_space *pa, *tmp;
4645 :
4646 0 : mb_debug(sb, "discard locality group preallocation\n");
4647 :
4648 0 : INIT_LIST_HEAD(&discard_list);
4649 :
4650 0 : spin_lock(&lg->lg_prealloc_lock);
4651 0 : list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4652 : pa_inode_list,
4653 : lockdep_is_held(&lg->lg_prealloc_lock)) {
4654 0 : spin_lock(&pa->pa_lock);
4655 0 : if (atomic_read(&pa->pa_count)) {
4656 : /*
4657 : * This is the pa that we just used
4658 : * for block allocation. So don't
4659 : * free that
4660 : */
4661 0 : spin_unlock(&pa->pa_lock);
4662 0 : continue;
4663 : }
4664 0 : if (pa->pa_deleted) {
4665 0 : spin_unlock(&pa->pa_lock);
4666 0 : continue;
4667 : }
4668 : /* only lg prealloc space */
4669 0 : BUG_ON(pa->pa_type != MB_GROUP_PA);
4670 :
4671 : /* seems this one can be freed ... */
4672 0 : ext4_mb_mark_pa_deleted(sb, pa);
4673 0 : spin_unlock(&pa->pa_lock);
4674 :
4675 0 : list_del_rcu(&pa->pa_inode_list);
4676 0 : list_add(&pa->u.pa_tmp_list, &discard_list);
4677 :
4678 0 : total_entries--;
4679 0 : if (total_entries <= 5) {
4680 : /*
4681 : * we want to keep only 5 entries
4682 : * allowing it to grow to 8. This
4683 : * mak sure we don't call discard
4684 : * soon for this list.
4685 : */
4686 : break;
4687 : }
4688 : }
4689 0 : spin_unlock(&lg->lg_prealloc_lock);
4690 :
4691 0 : list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4692 0 : int err;
4693 :
4694 0 : group = ext4_get_group_number(sb, pa->pa_pstart);
4695 0 : err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
4696 : GFP_NOFS|__GFP_NOFAIL);
4697 0 : if (err) {
4698 0 : ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
4699 : err, group);
4700 0 : continue;
4701 : }
4702 0 : ext4_lock_group(sb, group);
4703 0 : list_del(&pa->pa_group_list);
4704 0 : ext4_mb_release_group_pa(&e4b, pa);
4705 0 : ext4_unlock_group(sb, group);
4706 :
4707 0 : ext4_mb_unload_buddy(&e4b);
4708 0 : list_del(&pa->u.pa_tmp_list);
4709 0 : call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4710 : }
4711 0 : }
4712 :
4713 : /*
4714 : * We have incremented pa_count. So it cannot be freed at this
4715 : * point. Also we hold lg_mutex. So no parallel allocation is
4716 : * possible from this lg. That means pa_free cannot be updated.
4717 : *
4718 : * A parallel ext4_mb_discard_group_preallocations is possible.
4719 : * which can cause the lg_prealloc_list to be updated.
4720 : */
4721 :
4722 52 : static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4723 : {
4724 52 : int order, added = 0, lg_prealloc_count = 1;
4725 52 : struct super_block *sb = ac->ac_sb;
4726 52 : struct ext4_locality_group *lg = ac->ac_lg;
4727 52 : struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4728 :
4729 52 : order = fls(pa->pa_free) - 1;
4730 52 : if (order > PREALLOC_TB_SIZE - 1)
4731 : /* The max size of hash table is PREALLOC_TB_SIZE */
4732 : order = PREALLOC_TB_SIZE - 1;
4733 : /* Add the prealloc space to lg */
4734 52 : spin_lock(&lg->lg_prealloc_lock);
4735 52 : list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4736 : pa_inode_list,
4737 : lockdep_is_held(&lg->lg_prealloc_lock)) {
4738 0 : spin_lock(&tmp_pa->pa_lock);
4739 0 : if (tmp_pa->pa_deleted) {
4740 0 : spin_unlock(&tmp_pa->pa_lock);
4741 0 : continue;
4742 : }
4743 0 : if (!added && pa->pa_free < tmp_pa->pa_free) {
4744 : /* Add to the tail of the previous entry */
4745 0 : list_add_tail_rcu(&pa->pa_inode_list,
4746 : &tmp_pa->pa_inode_list);
4747 0 : added = 1;
4748 : /*
4749 : * we want to count the total
4750 : * number of entries in the list
4751 : */
4752 : }
4753 0 : spin_unlock(&tmp_pa->pa_lock);
4754 0 : lg_prealloc_count++;
4755 : }
4756 52 : if (!added)
4757 52 : list_add_tail_rcu(&pa->pa_inode_list,
4758 : &lg->lg_prealloc_list[order]);
4759 52 : spin_unlock(&lg->lg_prealloc_lock);
4760 :
4761 : /* Now trim the list to be not more than 8 elements */
4762 52 : if (lg_prealloc_count > 8) {
4763 0 : ext4_mb_discard_lg_preallocations(sb, lg,
4764 : order, lg_prealloc_count);
4765 0 : return;
4766 : }
4767 : return ;
4768 : }
4769 :
4770 : /*
4771 : * if per-inode prealloc list is too long, trim some PA
4772 : */
4773 245 : static void ext4_mb_trim_inode_pa(struct inode *inode)
4774 : {
4775 245 : struct ext4_inode_info *ei = EXT4_I(inode);
4776 245 : struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4777 245 : int count, delta;
4778 :
4779 245 : count = atomic_read(&ei->i_prealloc_active);
4780 245 : delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
4781 245 : if (count > sbi->s_mb_max_inode_prealloc + delta) {
4782 0 : count -= sbi->s_mb_max_inode_prealloc;
4783 0 : ext4_discard_preallocations(inode, count);
4784 : }
4785 245 : }
4786 :
4787 : /*
4788 : * release all resource we used in allocation
4789 : */
4790 245 : static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4791 : {
4792 245 : struct inode *inode = ac->ac_inode;
4793 245 : struct ext4_inode_info *ei = EXT4_I(inode);
4794 245 : struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4795 245 : struct ext4_prealloc_space *pa = ac->ac_pa;
4796 245 : if (pa) {
4797 54 : if (pa->pa_type == MB_GROUP_PA) {
4798 : /* see comment in ext4_mb_use_group_pa() */
4799 52 : spin_lock(&pa->pa_lock);
4800 52 : pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4801 52 : pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4802 52 : pa->pa_free -= ac->ac_b_ex.fe_len;
4803 52 : pa->pa_len -= ac->ac_b_ex.fe_len;
4804 52 : spin_unlock(&pa->pa_lock);
4805 :
4806 : /*
4807 : * We want to add the pa to the right bucket.
4808 : * Remove it from the list and while adding
4809 : * make sure the list to which we are adding
4810 : * doesn't grow big.
4811 : */
4812 52 : if (likely(pa->pa_free)) {
4813 52 : spin_lock(pa->pa_obj_lock);
4814 52 : list_del_rcu(&pa->pa_inode_list);
4815 52 : spin_unlock(pa->pa_obj_lock);
4816 52 : ext4_mb_add_n_trim(ac);
4817 : }
4818 : }
4819 :
4820 54 : if (pa->pa_type == MB_INODE_PA) {
4821 : /*
4822 : * treat per-inode prealloc list as a lru list, then try
4823 : * to trim the least recently used PA.
4824 : */
4825 2 : spin_lock(pa->pa_obj_lock);
4826 2 : list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
4827 2 : spin_unlock(pa->pa_obj_lock);
4828 : }
4829 :
4830 54 : ext4_mb_put_pa(ac, ac->ac_sb, pa);
4831 : }
4832 245 : if (ac->ac_bitmap_page)
4833 197 : put_page(ac->ac_bitmap_page);
4834 245 : if (ac->ac_buddy_page)
4835 197 : put_page(ac->ac_buddy_page);
4836 245 : if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4837 52 : mutex_unlock(&ac->ac_lg->lg_mutex);
4838 245 : ext4_mb_collect_stats(ac);
4839 245 : ext4_mb_trim_inode_pa(inode);
4840 245 : return 0;
4841 : }
4842 :
4843 0 : static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4844 : {
4845 0 : ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4846 0 : int ret;
4847 0 : int freed = 0;
4848 :
4849 0 : trace_ext4_mb_discard_preallocations(sb, needed);
4850 0 : for (i = 0; i < ngroups && needed > 0; i++) {
4851 0 : ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4852 0 : freed += ret;
4853 0 : needed -= ret;
4854 : }
4855 :
4856 0 : return freed;
4857 : }
4858 :
4859 0 : static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
4860 : struct ext4_allocation_context *ac, u64 *seq)
4861 : {
4862 0 : int freed;
4863 0 : u64 seq_retry = 0;
4864 0 : bool ret = false;
4865 :
4866 0 : freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4867 0 : if (freed) {
4868 0 : ret = true;
4869 0 : goto out_dbg;
4870 : }
4871 0 : seq_retry = ext4_get_discard_pa_seq_sum();
4872 0 : if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
4873 0 : ac->ac_flags |= EXT4_MB_STRICT_CHECK;
4874 0 : *seq = seq_retry;
4875 0 : ret = true;
4876 : }
4877 :
4878 0 : out_dbg:
4879 0 : mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
4880 0 : return ret;
4881 : }
4882 :
4883 : static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
4884 : struct ext4_allocation_request *ar, int *errp);
4885 :
4886 : /*
4887 : * Main entry point into mballoc to allocate blocks
4888 : * it tries to use preallocation first, then falls back
4889 : * to usual allocation
4890 : */
4891 245 : ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4892 : struct ext4_allocation_request *ar, int *errp)
4893 : {
4894 245 : struct ext4_allocation_context *ac = NULL;
4895 245 : struct ext4_sb_info *sbi;
4896 245 : struct super_block *sb;
4897 245 : ext4_fsblk_t block = 0;
4898 245 : unsigned int inquota = 0;
4899 245 : unsigned int reserv_clstrs = 0;
4900 245 : u64 seq;
4901 :
4902 245 : might_sleep();
4903 245 : sb = ar->inode->i_sb;
4904 245 : sbi = EXT4_SB(sb);
4905 :
4906 245 : trace_ext4_request_blocks(ar);
4907 245 : if (sbi->s_mount_state & EXT4_FC_REPLAY)
4908 0 : return ext4_mb_new_blocks_simple(handle, ar, errp);
4909 :
4910 : /* Allow to use superuser reservation for quota file */
4911 245 : if (ext4_is_quota_file(ar->inode))
4912 0 : ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4913 :
4914 245 : if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4915 : /* Without delayed allocation we need to verify
4916 : * there is enough free blocks to do block allocation
4917 : * and verify allocation doesn't exceed the quota limits.
4918 : */
4919 346 : while (ar->len &&
4920 173 : ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4921 :
4922 : /* let others to free the space */
4923 0 : cond_resched();
4924 0 : ar->len = ar->len >> 1;
4925 : }
4926 173 : if (!ar->len) {
4927 0 : ext4_mb_show_pa(sb);
4928 0 : *errp = -ENOSPC;
4929 0 : return 0;
4930 : }
4931 173 : reserv_clstrs = ar->len;
4932 173 : if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4933 173 : dquot_alloc_block_nofail(ar->inode,
4934 0 : EXT4_C2B(sbi, ar->len));
4935 : } else {
4936 346 : while (ar->len &&
4937 173 : dquot_alloc_block(ar->inode,
4938 173 : EXT4_C2B(sbi, ar->len))) {
4939 :
4940 0 : ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4941 0 : ar->len--;
4942 : }
4943 : }
4944 173 : inquota = ar->len;
4945 173 : if (ar->len == 0) {
4946 0 : *errp = -EDQUOT;
4947 0 : goto out;
4948 : }
4949 : }
4950 :
4951 245 : ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4952 245 : if (!ac) {
4953 0 : ar->len = 0;
4954 0 : *errp = -ENOMEM;
4955 0 : goto out;
4956 : }
4957 :
4958 245 : *errp = ext4_mb_initialize_context(ac, ar);
4959 245 : if (*errp) {
4960 0 : ar->len = 0;
4961 0 : goto out;
4962 : }
4963 :
4964 245 : ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4965 245 : seq = this_cpu_read(discard_pa_seq);
4966 245 : if (!ext4_mb_use_preallocated(ac)) {
4967 197 : ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4968 197 : ext4_mb_normalize_request(ac, ar);
4969 :
4970 197 : *errp = ext4_mb_pa_alloc(ac);
4971 197 : if (*errp)
4972 0 : goto errout;
4973 197 : repeat:
4974 : /* allocate space in core */
4975 197 : *errp = ext4_mb_regular_allocator(ac);
4976 : /*
4977 : * pa allocated above is added to grp->bb_prealloc_list only
4978 : * when we were able to allocate some block i.e. when
4979 : * ac->ac_status == AC_STATUS_FOUND.
4980 : * And error from above mean ac->ac_status != AC_STATUS_FOUND
4981 : * So we have to free this pa here itself.
4982 : */
4983 197 : if (*errp) {
4984 0 : ext4_mb_pa_free(ac);
4985 0 : ext4_discard_allocated_blocks(ac);
4986 0 : goto errout;
4987 : }
4988 197 : if (ac->ac_status == AC_STATUS_FOUND &&
4989 197 : ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
4990 191 : ext4_mb_pa_free(ac);
4991 : }
4992 245 : if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4993 245 : *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4994 245 : if (*errp) {
4995 0 : ext4_discard_allocated_blocks(ac);
4996 0 : goto errout;
4997 : } else {
4998 245 : block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4999 245 : ar->len = ac->ac_b_ex.fe_len;
5000 : }
5001 : } else {
5002 0 : if (ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
5003 0 : goto repeat;
5004 : /*
5005 : * If block allocation fails then the pa allocated above
5006 : * needs to be freed here itself.
5007 : */
5008 0 : ext4_mb_pa_free(ac);
5009 0 : *errp = -ENOSPC;
5010 : }
5011 :
5012 245 : errout:
5013 245 : if (*errp) {
5014 0 : ac->ac_b_ex.fe_len = 0;
5015 0 : ar->len = 0;
5016 0 : ext4_mb_show_ac(ac);
5017 : }
5018 245 : ext4_mb_release_context(ac);
5019 245 : out:
5020 245 : if (ac)
5021 245 : kmem_cache_free(ext4_ac_cachep, ac);
5022 245 : if (inquota && ar->len < inquota)
5023 0 : dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
5024 245 : if (!ar->len) {
5025 0 : if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
5026 : /* release all the reserved blocks if non delalloc */
5027 0 : percpu_counter_sub(&sbi->s_dirtyclusters_counter,
5028 : reserv_clstrs);
5029 : }
5030 :
5031 245 : trace_ext4_allocate_blocks(ar, (unsigned long long)block);
5032 :
5033 245 : return block;
5034 : }
5035 :
5036 : /*
5037 : * We can merge two free data extents only if the physical blocks
5038 : * are contiguous, AND the extents were freed by the same transaction,
5039 : * AND the blocks are associated with the same group.
5040 : */
5041 185 : static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
5042 : struct ext4_free_data *entry,
5043 : struct ext4_free_data *new_entry,
5044 : struct rb_root *entry_rb_root)
5045 : {
5046 185 : if ((entry->efd_tid != new_entry->efd_tid) ||
5047 185 : (entry->efd_group != new_entry->efd_group))
5048 : return;
5049 185 : if (entry->efd_start_cluster + entry->efd_count ==
5050 185 : new_entry->efd_start_cluster) {
5051 30 : new_entry->efd_start_cluster = entry->efd_start_cluster;
5052 30 : new_entry->efd_count += entry->efd_count;
5053 155 : } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
5054 : entry->efd_start_cluster) {
5055 12 : new_entry->efd_count += entry->efd_count;
5056 : } else
5057 : return;
5058 42 : spin_lock(&sbi->s_md_lock);
5059 42 : list_del(&entry->efd_list);
5060 42 : spin_unlock(&sbi->s_md_lock);
5061 42 : rb_erase(&entry->efd_node, entry_rb_root);
5062 42 : kmem_cache_free(ext4_free_data_cachep, entry);
5063 : }
5064 :
5065 : static noinline_for_stack int
5066 206 : ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
5067 : struct ext4_free_data *new_entry)
5068 : {
5069 206 : ext4_group_t group = e4b->bd_group;
5070 206 : ext4_grpblk_t cluster;
5071 206 : ext4_grpblk_t clusters = new_entry->efd_count;
5072 206 : struct ext4_free_data *entry;
5073 206 : struct ext4_group_info *db = e4b->bd_info;
5074 206 : struct super_block *sb = e4b->bd_sb;
5075 206 : struct ext4_sb_info *sbi = EXT4_SB(sb);
5076 206 : struct rb_node **n = &db->bb_free_root.rb_node, *node;
5077 206 : struct rb_node *parent = NULL, *new_node;
5078 :
5079 206 : BUG_ON(!ext4_handle_valid(handle));
5080 206 : BUG_ON(e4b->bd_bitmap_page == NULL);
5081 206 : BUG_ON(e4b->bd_buddy_page == NULL);
5082 :
5083 206 : new_node = &new_entry->efd_node;
5084 206 : cluster = new_entry->efd_start_cluster;
5085 :
5086 206 : if (!*n) {
5087 : /* first free block exent. We need to
5088 : protect buddy cache from being freed,
5089 : * otherwise we'll refresh it from
5090 : * on-disk bitmap and lose not-yet-available
5091 : * blocks */
5092 79 : get_page(e4b->bd_buddy_page);
5093 79 : get_page(e4b->bd_bitmap_page);
5094 : }
5095 503 : while (*n) {
5096 297 : parent = *n;
5097 297 : entry = rb_entry(parent, struct ext4_free_data, efd_node);
5098 297 : if (cluster < entry->efd_start_cluster)
5099 84 : n = &(*n)->rb_left;
5100 213 : else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
5101 213 : n = &(*n)->rb_right;
5102 : else {
5103 0 : ext4_grp_locked_error(sb, group, 0,
5104 : ext4_group_first_block_no(sb, group) +
5105 : EXT4_C2B(sbi, cluster),
5106 : "Block already on to-be-freed list");
5107 0 : kmem_cache_free(ext4_free_data_cachep, new_entry);
5108 0 : return 0;
5109 : }
5110 : }
5111 :
5112 206 : rb_link_node(new_node, parent, n);
5113 206 : rb_insert_color(new_node, &db->bb_free_root);
5114 :
5115 : /* Now try to see the extent can be merged to left and right */
5116 206 : node = rb_prev(new_node);
5117 206 : if (node) {
5118 108 : entry = rb_entry(node, struct ext4_free_data, efd_node);
5119 108 : ext4_try_merge_freed_extent(sbi, entry, new_entry,
5120 : &(db->bb_free_root));
5121 : }
5122 :
5123 206 : node = rb_next(new_node);
5124 206 : if (node) {
5125 77 : entry = rb_entry(node, struct ext4_free_data, efd_node);
5126 77 : ext4_try_merge_freed_extent(sbi, entry, new_entry,
5127 : &(db->bb_free_root));
5128 : }
5129 :
5130 206 : spin_lock(&sbi->s_md_lock);
5131 206 : list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list);
5132 206 : sbi->s_mb_free_pending += clusters;
5133 206 : spin_unlock(&sbi->s_md_lock);
5134 206 : return 0;
5135 : }
5136 :
5137 : /*
5138 : * Simple allocator for Ext4 fast commit replay path. It searches for blocks
5139 : * linearly starting at the goal block and also excludes the blocks which
5140 : * are going to be in use after fast commit replay.
5141 : */
5142 0 : static ext4_fsblk_t ext4_mb_new_blocks_simple(handle_t *handle,
5143 : struct ext4_allocation_request *ar, int *errp)
5144 : {
5145 0 : struct buffer_head *bitmap_bh;
5146 0 : struct super_block *sb = ar->inode->i_sb;
5147 0 : ext4_group_t group;
5148 0 : ext4_grpblk_t blkoff;
5149 0 : int i = sb->s_blocksize;
5150 0 : ext4_fsblk_t goal, block;
5151 0 : struct ext4_super_block *es = EXT4_SB(sb)->s_es;
5152 :
5153 0 : goal = ar->goal;
5154 0 : if (goal < le32_to_cpu(es->s_first_data_block) ||
5155 0 : goal >= ext4_blocks_count(es))
5156 : goal = le32_to_cpu(es->s_first_data_block);
5157 :
5158 0 : ar->len = 0;
5159 0 : ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
5160 0 : for (; group < ext4_get_groups_count(sb); group++) {
5161 0 : bitmap_bh = ext4_read_block_bitmap(sb, group);
5162 0 : if (IS_ERR(bitmap_bh)) {
5163 0 : *errp = PTR_ERR(bitmap_bh);
5164 0 : pr_warn("Failed to read block bitmap\n");
5165 0 : return 0;
5166 : }
5167 :
5168 0 : ext4_get_group_no_and_offset(sb,
5169 0 : max(ext4_group_first_block_no(sb, group), goal),
5170 : NULL, &blkoff);
5171 0 : i = mb_find_next_zero_bit(bitmap_bh->b_data, sb->s_blocksize,
5172 : blkoff);
5173 0 : brelse(bitmap_bh);
5174 0 : if (i >= sb->s_blocksize)
5175 0 : continue;
5176 0 : if (ext4_fc_replay_check_excluded(sb,
5177 0 : ext4_group_first_block_no(sb, group) + i))
5178 0 : continue;
5179 : break;
5180 : }
5181 :
5182 0 : if (group >= ext4_get_groups_count(sb) && i >= sb->s_blocksize)
5183 : return 0;
5184 :
5185 0 : block = ext4_group_first_block_no(sb, group) + i;
5186 0 : ext4_mb_mark_bb(sb, block, 1, 1);
5187 0 : ar->len = 1;
5188 :
5189 0 : return block;
5190 : }
5191 :
5192 0 : static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
5193 : unsigned long count)
5194 : {
5195 0 : struct buffer_head *bitmap_bh;
5196 0 : struct super_block *sb = inode->i_sb;
5197 0 : struct ext4_group_desc *gdp;
5198 0 : struct buffer_head *gdp_bh;
5199 0 : ext4_group_t group;
5200 0 : ext4_grpblk_t blkoff;
5201 0 : int already_freed = 0, err, i;
5202 :
5203 0 : ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
5204 0 : bitmap_bh = ext4_read_block_bitmap(sb, group);
5205 0 : if (IS_ERR(bitmap_bh)) {
5206 0 : err = PTR_ERR(bitmap_bh);
5207 0 : pr_warn("Failed to read block bitmap\n");
5208 0 : return;
5209 : }
5210 0 : gdp = ext4_get_group_desc(sb, group, &gdp_bh);
5211 0 : if (!gdp)
5212 : return;
5213 :
5214 0 : for (i = 0; i < count; i++) {
5215 0 : if (!mb_test_bit(blkoff + i, bitmap_bh->b_data))
5216 0 : already_freed++;
5217 : }
5218 0 : mb_clear_bits(bitmap_bh->b_data, blkoff, count);
5219 0 : err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh);
5220 0 : if (err)
5221 : return;
5222 0 : ext4_free_group_clusters_set(
5223 0 : sb, gdp, ext4_free_group_clusters(sb, gdp) +
5224 : count - already_freed);
5225 0 : ext4_block_bitmap_csum_set(sb, group, gdp, bitmap_bh);
5226 0 : ext4_group_desc_csum_set(sb, group, gdp);
5227 0 : ext4_handle_dirty_metadata(NULL, NULL, gdp_bh);
5228 0 : sync_dirty_buffer(bitmap_bh);
5229 0 : sync_dirty_buffer(gdp_bh);
5230 0 : brelse(bitmap_bh);
5231 : }
5232 :
5233 : /**
5234 : * ext4_free_blocks() -- Free given blocks and update quota
5235 : * @handle: handle for this transaction
5236 : * @inode: inode
5237 : * @bh: optional buffer of the block to be freed
5238 : * @block: starting physical block to be freed
5239 : * @count: number of blocks to be freed
5240 : * @flags: flags used by ext4_free_blocks
5241 : */
5242 206 : void ext4_free_blocks(handle_t *handle, struct inode *inode,
5243 : struct buffer_head *bh, ext4_fsblk_t block,
5244 : unsigned long count, int flags)
5245 : {
5246 206 : struct buffer_head *bitmap_bh = NULL;
5247 206 : struct super_block *sb = inode->i_sb;
5248 206 : struct ext4_group_desc *gdp;
5249 206 : unsigned int overflow;
5250 206 : ext4_grpblk_t bit;
5251 206 : struct buffer_head *gd_bh;
5252 206 : ext4_group_t block_group;
5253 206 : struct ext4_sb_info *sbi;
5254 206 : struct ext4_buddy e4b;
5255 206 : unsigned int count_clusters;
5256 206 : int err = 0;
5257 206 : int ret;
5258 :
5259 206 : sbi = EXT4_SB(sb);
5260 :
5261 206 : if (sbi->s_mount_state & EXT4_FC_REPLAY) {
5262 0 : ext4_free_blocks_simple(inode, block, count);
5263 0 : return;
5264 : }
5265 :
5266 206 : might_sleep();
5267 206 : if (bh) {
5268 0 : if (block)
5269 0 : BUG_ON(block != bh->b_blocknr);
5270 : else
5271 0 : block = bh->b_blocknr;
5272 : }
5273 :
5274 412 : if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
5275 206 : !ext4_inode_block_valid(inode, block, count)) {
5276 0 : ext4_error(sb, "Freeing blocks not in datazone - "
5277 : "block = %llu, count = %lu", block, count);
5278 0 : goto error_return;
5279 : }
5280 :
5281 206 : ext4_debug("freeing block %llu\n", block);
5282 206 : trace_ext4_free_blocks(inode, block, count, flags);
5283 :
5284 206 : if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5285 0 : BUG_ON(count > 1);
5286 :
5287 0 : ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
5288 : inode, bh, block);
5289 : }
5290 :
5291 : /*
5292 : * If the extent to be freed does not begin on a cluster
5293 : * boundary, we need to deal with partial clusters at the
5294 : * beginning and end of the extent. Normally we will free
5295 : * blocks at the beginning or the end unless we are explicitly
5296 : * requested to avoid doing so.
5297 : */
5298 206 : overflow = EXT4_PBLK_COFF(sbi, block);
5299 206 : if (overflow) {
5300 0 : if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
5301 0 : overflow = sbi->s_cluster_ratio - overflow;
5302 0 : block += overflow;
5303 0 : if (count > overflow)
5304 0 : count -= overflow;
5305 : else
5306 : return;
5307 : } else {
5308 0 : block -= overflow;
5309 0 : count += overflow;
5310 : }
5311 : }
5312 206 : overflow = EXT4_LBLK_COFF(sbi, count);
5313 206 : if (overflow) {
5314 0 : if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
5315 0 : if (count > overflow)
5316 0 : count -= overflow;
5317 : else
5318 : return;
5319 : } else
5320 0 : count += sbi->s_cluster_ratio - overflow;
5321 : }
5322 :
5323 206 : if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
5324 77 : int i;
5325 77 : int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
5326 :
5327 154 : for (i = 0; i < count; i++) {
5328 77 : cond_resched();
5329 77 : if (is_metadata)
5330 77 : bh = sb_find_get_block(inode->i_sb, block + i);
5331 77 : ext4_forget(handle, is_metadata, inode, bh, block + i);
5332 : }
5333 : }
5334 :
5335 206 : do_more:
5336 206 : overflow = 0;
5337 206 : ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5338 :
5339 206 : if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
5340 : ext4_get_group_info(sb, block_group))))
5341 : return;
5342 :
5343 : /*
5344 : * Check to see if we are freeing blocks across a group
5345 : * boundary.
5346 : */
5347 206 : if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
5348 0 : overflow = EXT4_C2B(sbi, bit) + count -
5349 0 : EXT4_BLOCKS_PER_GROUP(sb);
5350 0 : count -= overflow;
5351 : }
5352 206 : count_clusters = EXT4_NUM_B2C(sbi, count);
5353 206 : bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5354 206 : if (IS_ERR(bitmap_bh)) {
5355 0 : err = PTR_ERR(bitmap_bh);
5356 0 : bitmap_bh = NULL;
5357 0 : goto error_return;
5358 : }
5359 206 : gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
5360 206 : if (!gdp) {
5361 0 : err = -EIO;
5362 0 : goto error_return;
5363 : }
5364 :
5365 412 : if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
5366 412 : in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
5367 412 : in_range(block, ext4_inode_table(sb, gdp),
5368 206 : sbi->s_itb_per_group) ||
5369 412 : in_range(block + count - 1, ext4_inode_table(sb, gdp),
5370 : sbi->s_itb_per_group)) {
5371 :
5372 0 : ext4_error(sb, "Freeing blocks in system zone - "
5373 : "Block = %llu, count = %lu", block, count);
5374 : /* err = 0. ext4_std_error should be a no op */
5375 0 : goto error_return;
5376 : }
5377 :
5378 206 : BUFFER_TRACE(bitmap_bh, "getting write access");
5379 206 : err = ext4_journal_get_write_access(handle, bitmap_bh);
5380 206 : if (err)
5381 0 : goto error_return;
5382 :
5383 : /*
5384 : * We are about to modify some metadata. Call the journal APIs
5385 : * to unshare ->b_data if a currently-committing transaction is
5386 : * using it
5387 : */
5388 206 : BUFFER_TRACE(gd_bh, "get_write_access");
5389 206 : err = ext4_journal_get_write_access(handle, gd_bh);
5390 206 : if (err)
5391 0 : goto error_return;
5392 : #ifdef AGGRESSIVE_CHECK
5393 : {
5394 : int i;
5395 : for (i = 0; i < count_clusters; i++)
5396 : BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
5397 : }
5398 : #endif
5399 206 : trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
5400 :
5401 : /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
5402 206 : err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
5403 : GFP_NOFS|__GFP_NOFAIL);
5404 206 : if (err)
5405 0 : goto error_return;
5406 :
5407 : /*
5408 : * We need to make sure we don't reuse the freed block until after the
5409 : * transaction is committed. We make an exception if the inode is to be
5410 : * written in writeback mode since writeback mode has weak data
5411 : * consistency guarantees.
5412 : */
5413 206 : if (ext4_handle_valid(handle) &&
5414 206 : ((flags & EXT4_FREE_BLOCKS_METADATA) ||
5415 335 : !ext4_should_writeback_data(inode))) {
5416 206 : struct ext4_free_data *new_entry;
5417 : /*
5418 : * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
5419 : * to fail.
5420 : */
5421 206 : new_entry = kmem_cache_alloc(ext4_free_data_cachep,
5422 : GFP_NOFS|__GFP_NOFAIL);
5423 206 : new_entry->efd_start_cluster = bit;
5424 206 : new_entry->efd_group = block_group;
5425 206 : new_entry->efd_count = count_clusters;
5426 206 : new_entry->efd_tid = handle->h_transaction->t_tid;
5427 :
5428 206 : ext4_lock_group(sb, block_group);
5429 206 : mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5430 206 : ext4_mb_free_metadata(handle, &e4b, new_entry);
5431 : } else {
5432 : /* need to update group_info->bb_free and bitmap
5433 : * with group lock held. generate_buddy look at
5434 : * them with group lock_held
5435 : */
5436 0 : if (test_opt(sb, DISCARD)) {
5437 0 : err = ext4_issue_discard(sb, block_group, bit, count,
5438 : NULL);
5439 0 : if (err && err != -EOPNOTSUPP)
5440 0 : ext4_msg(sb, KERN_WARNING, "discard request in"
5441 : " group:%d block:%d count:%lu failed"
5442 : " with %d", block_group, bit, count,
5443 : err);
5444 : } else
5445 0 : EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
5446 :
5447 0 : ext4_lock_group(sb, block_group);
5448 0 : mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
5449 0 : mb_free_blocks(inode, &e4b, bit, count_clusters);
5450 : }
5451 :
5452 206 : ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
5453 206 : ext4_free_group_clusters_set(sb, gdp, ret);
5454 206 : ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
5455 206 : ext4_group_desc_csum_set(sb, block_group, gdp);
5456 206 : ext4_unlock_group(sb, block_group);
5457 :
5458 206 : if (sbi->s_log_groups_per_flex) {
5459 206 : ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5460 618 : atomic64_add(count_clusters,
5461 412 : &sbi_array_rcu_deref(sbi, s_flex_groups,
5462 : flex_group)->free_clusters);
5463 : }
5464 :
5465 : /*
5466 : * on a bigalloc file system, defer the s_freeclusters_counter
5467 : * update to the caller (ext4_remove_space and friends) so they
5468 : * can determine if a cluster freed here should be rereserved
5469 : */
5470 206 : if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
5471 206 : if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
5472 206 : dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
5473 206 : percpu_counter_add(&sbi->s_freeclusters_counter,
5474 : count_clusters);
5475 : }
5476 :
5477 206 : ext4_mb_unload_buddy(&e4b);
5478 :
5479 : /* We dirtied the bitmap block */
5480 206 : BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5481 206 : err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5482 :
5483 : /* And the group descriptor block */
5484 206 : BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5485 206 : ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5486 206 : if (!err)
5487 206 : err = ret;
5488 :
5489 206 : if (overflow && !err) {
5490 0 : block += count;
5491 0 : count = overflow;
5492 0 : put_bh(bitmap_bh);
5493 0 : goto do_more;
5494 : }
5495 206 : error_return:
5496 206 : brelse(bitmap_bh);
5497 206 : ext4_std_error(sb, err);
5498 : return;
5499 : }
5500 :
5501 : /**
5502 : * ext4_group_add_blocks() -- Add given blocks to an existing group
5503 : * @handle: handle to this transaction
5504 : * @sb: super block
5505 : * @block: start physical block to add to the block group
5506 : * @count: number of blocks to free
5507 : *
5508 : * This marks the blocks as free in the bitmap and buddy.
5509 : */
5510 0 : int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
5511 : ext4_fsblk_t block, unsigned long count)
5512 : {
5513 0 : struct buffer_head *bitmap_bh = NULL;
5514 0 : struct buffer_head *gd_bh;
5515 0 : ext4_group_t block_group;
5516 0 : ext4_grpblk_t bit;
5517 0 : unsigned int i;
5518 0 : struct ext4_group_desc *desc;
5519 0 : struct ext4_sb_info *sbi = EXT4_SB(sb);
5520 0 : struct ext4_buddy e4b;
5521 0 : int err = 0, ret, free_clusters_count;
5522 0 : ext4_grpblk_t clusters_freed;
5523 0 : ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
5524 0 : ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
5525 0 : unsigned long cluster_count = last_cluster - first_cluster + 1;
5526 :
5527 0 : ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
5528 :
5529 0 : if (count == 0)
5530 : return 0;
5531 :
5532 0 : ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
5533 : /*
5534 : * Check to see if we are freeing blocks across a group
5535 : * boundary.
5536 : */
5537 0 : if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
5538 0 : ext4_warning(sb, "too many blocks added to group %u",
5539 : block_group);
5540 0 : err = -EINVAL;
5541 0 : goto error_return;
5542 : }
5543 :
5544 0 : bitmap_bh = ext4_read_block_bitmap(sb, block_group);
5545 0 : if (IS_ERR(bitmap_bh)) {
5546 0 : err = PTR_ERR(bitmap_bh);
5547 0 : bitmap_bh = NULL;
5548 0 : goto error_return;
5549 : }
5550 :
5551 0 : desc = ext4_get_group_desc(sb, block_group, &gd_bh);
5552 0 : if (!desc) {
5553 0 : err = -EIO;
5554 0 : goto error_return;
5555 : }
5556 :
5557 0 : if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
5558 0 : in_range(ext4_inode_bitmap(sb, desc), block, count) ||
5559 0 : in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
5560 0 : in_range(block + count - 1, ext4_inode_table(sb, desc),
5561 : sbi->s_itb_per_group)) {
5562 0 : ext4_error(sb, "Adding blocks in system zones - "
5563 : "Block = %llu, count = %lu",
5564 : block, count);
5565 0 : err = -EINVAL;
5566 0 : goto error_return;
5567 : }
5568 :
5569 0 : BUFFER_TRACE(bitmap_bh, "getting write access");
5570 0 : err = ext4_journal_get_write_access(handle, bitmap_bh);
5571 0 : if (err)
5572 0 : goto error_return;
5573 :
5574 : /*
5575 : * We are about to modify some metadata. Call the journal APIs
5576 : * to unshare ->b_data if a currently-committing transaction is
5577 : * using it
5578 : */
5579 0 : BUFFER_TRACE(gd_bh, "get_write_access");
5580 0 : err = ext4_journal_get_write_access(handle, gd_bh);
5581 0 : if (err)
5582 0 : goto error_return;
5583 :
5584 0 : for (i = 0, clusters_freed = 0; i < cluster_count; i++) {
5585 0 : BUFFER_TRACE(bitmap_bh, "clear bit");
5586 0 : if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
5587 0 : ext4_error(sb, "bit already cleared for block %llu",
5588 : (ext4_fsblk_t)(block + i));
5589 0 : BUFFER_TRACE(bitmap_bh, "bit already cleared");
5590 : } else {
5591 0 : clusters_freed++;
5592 : }
5593 : }
5594 :
5595 0 : err = ext4_mb_load_buddy(sb, block_group, &e4b);
5596 0 : if (err)
5597 0 : goto error_return;
5598 :
5599 : /*
5600 : * need to update group_info->bb_free and bitmap
5601 : * with group lock held. generate_buddy look at
5602 : * them with group lock_held
5603 : */
5604 0 : ext4_lock_group(sb, block_group);
5605 0 : mb_clear_bits(bitmap_bh->b_data, bit, cluster_count);
5606 0 : mb_free_blocks(NULL, &e4b, bit, cluster_count);
5607 0 : free_clusters_count = clusters_freed +
5608 0 : ext4_free_group_clusters(sb, desc);
5609 0 : ext4_free_group_clusters_set(sb, desc, free_clusters_count);
5610 0 : ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5611 0 : ext4_group_desc_csum_set(sb, block_group, desc);
5612 0 : ext4_unlock_group(sb, block_group);
5613 0 : percpu_counter_add(&sbi->s_freeclusters_counter,
5614 : clusters_freed);
5615 :
5616 0 : if (sbi->s_log_groups_per_flex) {
5617 0 : ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5618 0 : atomic64_add(clusters_freed,
5619 0 : &sbi_array_rcu_deref(sbi, s_flex_groups,
5620 : flex_group)->free_clusters);
5621 : }
5622 :
5623 0 : ext4_mb_unload_buddy(&e4b);
5624 :
5625 : /* We dirtied the bitmap block */
5626 0 : BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5627 0 : err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5628 :
5629 : /* And the group descriptor block */
5630 0 : BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5631 0 : ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5632 0 : if (!err)
5633 0 : err = ret;
5634 :
5635 0 : error_return:
5636 0 : brelse(bitmap_bh);
5637 0 : ext4_std_error(sb, err);
5638 : return err;
5639 : }
5640 :
5641 : /**
5642 : * ext4_trim_extent -- function to TRIM one single free extent in the group
5643 : * @sb: super block for the file system
5644 : * @start: starting block of the free extent in the alloc. group
5645 : * @count: number of blocks to TRIM
5646 : * @group: alloc. group we are working with
5647 : * @e4b: ext4 buddy for the group
5648 : *
5649 : * Trim "count" blocks starting at "start" in the "group". To assure that no
5650 : * one will allocate those blocks, mark it as used in buddy bitmap. This must
5651 : * be called with under the group lock.
5652 : */
5653 0 : static int ext4_trim_extent(struct super_block *sb, int start, int count,
5654 : ext4_group_t group, struct ext4_buddy *e4b)
5655 : __releases(bitlock)
5656 : __acquires(bitlock)
5657 : {
5658 0 : struct ext4_free_extent ex;
5659 0 : int ret = 0;
5660 :
5661 0 : trace_ext4_trim_extent(sb, group, start, count);
5662 :
5663 0 : assert_spin_locked(ext4_group_lock_ptr(sb, group));
5664 :
5665 0 : ex.fe_start = start;
5666 0 : ex.fe_group = group;
5667 0 : ex.fe_len = count;
5668 :
5669 : /*
5670 : * Mark blocks used, so no one can reuse them while
5671 : * being trimmed.
5672 : */
5673 0 : mb_mark_used(e4b, &ex);
5674 0 : ext4_unlock_group(sb, group);
5675 0 : ret = ext4_issue_discard(sb, group, start, count, NULL);
5676 0 : ext4_lock_group(sb, group);
5677 0 : mb_free_blocks(NULL, e4b, start, ex.fe_len);
5678 0 : return ret;
5679 : }
5680 :
5681 : /**
5682 : * ext4_trim_all_free -- function to trim all free space in alloc. group
5683 : * @sb: super block for file system
5684 : * @group: group to be trimmed
5685 : * @start: first group block to examine
5686 : * @max: last group block to examine
5687 : * @minblocks: minimum extent block count
5688 : *
5689 : * ext4_trim_all_free walks through group's buddy bitmap searching for free
5690 : * extents. When the free block is found, ext4_trim_extent is called to TRIM
5691 : * the extent.
5692 : *
5693 : *
5694 : * ext4_trim_all_free walks through group's block bitmap searching for free
5695 : * extents. When the free extent is found, mark it as used in group buddy
5696 : * bitmap. Then issue a TRIM command on this extent and free the extent in
5697 : * the group buddy bitmap. This is done until whole group is scanned.
5698 : */
5699 : static ext4_grpblk_t
5700 0 : ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5701 : ext4_grpblk_t start, ext4_grpblk_t max,
5702 : ext4_grpblk_t minblocks)
5703 : {
5704 0 : void *bitmap;
5705 0 : ext4_grpblk_t next, count = 0, free_count = 0;
5706 0 : struct ext4_buddy e4b;
5707 0 : int ret = 0;
5708 :
5709 0 : trace_ext4_trim_all_free(sb, group, start, max);
5710 :
5711 0 : ret = ext4_mb_load_buddy(sb, group, &e4b);
5712 0 : if (ret) {
5713 0 : ext4_warning(sb, "Error %d loading buddy information for %u",
5714 : ret, group);
5715 0 : return ret;
5716 : }
5717 0 : bitmap = e4b.bd_bitmap;
5718 :
5719 0 : ext4_lock_group(sb, group);
5720 0 : if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5721 0 : minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5722 0 : goto out;
5723 :
5724 0 : start = (e4b.bd_info->bb_first_free > start) ?
5725 : e4b.bd_info->bb_first_free : start;
5726 :
5727 0 : while (start <= max) {
5728 0 : start = mb_find_next_zero_bit(bitmap, max + 1, start);
5729 0 : if (start > max)
5730 : break;
5731 0 : next = mb_find_next_bit(bitmap, max + 1, start);
5732 :
5733 0 : if ((next - start) >= minblocks) {
5734 0 : ret = ext4_trim_extent(sb, start,
5735 : next - start, group, &e4b);
5736 0 : if (ret && ret != -EOPNOTSUPP)
5737 : break;
5738 0 : ret = 0;
5739 0 : count += next - start;
5740 : }
5741 0 : free_count += next - start;
5742 0 : start = next + 1;
5743 :
5744 0 : if (fatal_signal_pending(current)) {
5745 : count = -ERESTARTSYS;
5746 : break;
5747 : }
5748 :
5749 0 : if (need_resched()) {
5750 0 : ext4_unlock_group(sb, group);
5751 0 : cond_resched();
5752 0 : ext4_lock_group(sb, group);
5753 : }
5754 :
5755 0 : if ((e4b.bd_info->bb_free - free_count) < minblocks)
5756 : break;
5757 : }
5758 :
5759 0 : if (!ret) {
5760 0 : ret = count;
5761 0 : EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5762 : }
5763 0 : out:
5764 0 : ext4_unlock_group(sb, group);
5765 0 : ext4_mb_unload_buddy(&e4b);
5766 :
5767 0 : ext4_debug("trimmed %d blocks in the group %d\n",
5768 : count, group);
5769 :
5770 0 : return ret;
5771 : }
5772 :
5773 : /**
5774 : * ext4_trim_fs() -- trim ioctl handle function
5775 : * @sb: superblock for filesystem
5776 : * @range: fstrim_range structure
5777 : *
5778 : * start: First Byte to trim
5779 : * len: number of Bytes to trim from start
5780 : * minlen: minimum extent length in Bytes
5781 : * ext4_trim_fs goes through all allocation groups containing Bytes from
5782 : * start to start+len. For each such a group ext4_trim_all_free function
5783 : * is invoked to trim all free space.
5784 : */
5785 0 : int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5786 : {
5787 0 : struct ext4_group_info *grp;
5788 0 : ext4_group_t group, first_group, last_group;
5789 0 : ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5790 0 : uint64_t start, end, minlen, trimmed = 0;
5791 0 : ext4_fsblk_t first_data_blk =
5792 0 : le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5793 0 : ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5794 0 : int ret = 0;
5795 :
5796 0 : start = range->start >> sb->s_blocksize_bits;
5797 0 : end = start + (range->len >> sb->s_blocksize_bits) - 1;
5798 0 : minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5799 : range->minlen >> sb->s_blocksize_bits);
5800 :
5801 0 : if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5802 0 : start >= max_blks ||
5803 0 : range->len < sb->s_blocksize)
5804 : return -EINVAL;
5805 0 : if (end >= max_blks)
5806 0 : end = max_blks - 1;
5807 0 : if (end <= first_data_blk)
5808 0 : goto out;
5809 0 : if (start < first_data_blk)
5810 : start = first_data_blk;
5811 :
5812 : /* Determine first and last group to examine based on start and end */
5813 0 : ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5814 : &first_group, &first_cluster);
5815 0 : ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5816 : &last_group, &last_cluster);
5817 :
5818 : /* end now represents the last cluster to discard in this group */
5819 0 : end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5820 :
5821 0 : for (group = first_group; group <= last_group; group++) {
5822 0 : grp = ext4_get_group_info(sb, group);
5823 : /* We only do this if the grp has never been initialized */
5824 0 : if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5825 0 : ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5826 0 : if (ret)
5827 : break;
5828 : }
5829 :
5830 : /*
5831 : * For all the groups except the last one, last cluster will
5832 : * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5833 : * change it for the last group, note that last_cluster is
5834 : * already computed earlier by ext4_get_group_no_and_offset()
5835 : */
5836 0 : if (group == last_group)
5837 0 : end = last_cluster;
5838 :
5839 0 : if (grp->bb_free >= minlen) {
5840 0 : cnt = ext4_trim_all_free(sb, group, first_cluster,
5841 : end, minlen);
5842 0 : if (cnt < 0) {
5843 : ret = cnt;
5844 : break;
5845 : }
5846 0 : trimmed += cnt;
5847 : }
5848 :
5849 : /*
5850 : * For every group except the first one, we are sure
5851 : * that the first cluster to discard will be cluster #0.
5852 : */
5853 0 : first_cluster = 0;
5854 : }
5855 :
5856 0 : if (!ret)
5857 0 : atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5858 :
5859 0 : out:
5860 0 : range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5861 0 : return ret;
5862 : }
5863 :
5864 : /* Iterate all the free extents in the group. */
5865 : int
5866 0 : ext4_mballoc_query_range(
5867 : struct super_block *sb,
5868 : ext4_group_t group,
5869 : ext4_grpblk_t start,
5870 : ext4_grpblk_t end,
5871 : ext4_mballoc_query_range_fn formatter,
5872 : void *priv)
5873 : {
5874 0 : void *bitmap;
5875 0 : ext4_grpblk_t next;
5876 0 : struct ext4_buddy e4b;
5877 0 : int error;
5878 :
5879 0 : error = ext4_mb_load_buddy(sb, group, &e4b);
5880 0 : if (error)
5881 : return error;
5882 0 : bitmap = e4b.bd_bitmap;
5883 :
5884 0 : ext4_lock_group(sb, group);
5885 :
5886 0 : start = (e4b.bd_info->bb_first_free > start) ?
5887 : e4b.bd_info->bb_first_free : start;
5888 0 : if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
5889 0 : end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5890 :
5891 0 : while (start <= end) {
5892 0 : start = mb_find_next_zero_bit(bitmap, end + 1, start);
5893 0 : if (start > end)
5894 : break;
5895 0 : next = mb_find_next_bit(bitmap, end + 1, start);
5896 :
5897 0 : ext4_unlock_group(sb, group);
5898 0 : error = formatter(sb, group, start, next - start, priv);
5899 0 : if (error)
5900 0 : goto out_unload;
5901 0 : ext4_lock_group(sb, group);
5902 :
5903 0 : start = next + 1;
5904 : }
5905 :
5906 0 : ext4_unlock_group(sb, group);
5907 0 : out_unload:
5908 0 : ext4_mb_unload_buddy(&e4b);
5909 :
5910 0 : return error;
5911 : }
|
