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 <linux/freezer.h>
20 #include <trace/events/ext4.h>
21 #include <kunit/static_stub.h>
22
23 /*
24 * MUSTDO:
25 * - test ext4_ext_search_left() and ext4_ext_search_right()
26 * - search for metadata in few groups
27 *
28 * TODO v4:
29 * - normalization should take into account whether file is still open
30 * - discard preallocations if no free space left (policy?)
31 * - don't normalize tails
32 * - quota
33 * - reservation for superuser
34 *
35 * TODO v3:
36 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
37 * - track min/max extents in each group for better group selection
38 * - mb_mark_used() may allocate chunk right after splitting buddy
39 * - tree of groups sorted by number of free blocks
40 * - error handling
41 */
42
43 /*
44 * The allocation request involve request for multiple number of blocks
45 * near to the goal(block) value specified.
46 *
47 * During initialization phase of the allocator we decide to use the
48 * group preallocation or inode preallocation depending on the size of
49 * the file. The size of the file could be the resulting file size we
50 * would have after allocation, or the current file size, which ever
51 * is larger. If the size is less than sbi->s_mb_stream_request we
52 * select to use the group preallocation. The default value of
53 * s_mb_stream_request is 16 blocks. This can also be tuned via
54 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
55 * terms of number of blocks.
56 *
57 * The main motivation for having small file use group preallocation is to
58 * ensure that we have small files closer together on the disk.
59 *
60 * First stage the allocator looks at the inode prealloc list,
61 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
62 * spaces for this particular inode. The inode prealloc space is
63 * represented as:
64 *
65 * pa_lstart -> the logical start block for this prealloc space
66 * pa_pstart -> the physical start block for this prealloc space
67 * pa_len -> length for this prealloc space (in clusters)
68 * pa_free -> free space available in this prealloc space (in clusters)
69 *
70 * The inode preallocation space is used looking at the _logical_ start
71 * block. If only the logical file block falls within the range of prealloc
72 * space we will consume the particular prealloc space. This makes sure that
73 * we have contiguous physical blocks representing the file blocks
74 *
75 * The important thing to be noted in case of inode prealloc space is that
76 * we don't modify the values associated to inode prealloc space except
77 * pa_free.
78 *
79 * If we are not able to find blocks in the inode prealloc space and if we
80 * have the group allocation flag set then we look at the locality group
81 * prealloc space. These are per CPU prealloc list represented as
82 *
83 * ext4_sb_info.s_locality_groups[smp_processor_id()]
84 *
85 * The reason for having a per cpu locality group is to reduce the contention
86 * between CPUs. It is possible to get scheduled at this point.
87 *
88 * The locality group prealloc space is used looking at whether we have
89 * enough free space (pa_free) within the prealloc space.
90 *
91 * If we can't allocate blocks via inode prealloc or/and locality group
92 * prealloc then we look at the buddy cache. The buddy cache is represented
93 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
94 * mapped to the buddy and bitmap information regarding different
95 * groups. The buddy information is attached to buddy cache inode so that
96 * we can access them through the page cache. The information regarding
97 * each group is loaded via ext4_mb_load_buddy. The information involve
98 * block bitmap and buddy information. The information are stored in the
99 * inode as:
100 *
101 * { page }
102 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
103 *
104 *
105 * one block each for bitmap and buddy information. So for each group we
106 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
107 * blocksize) blocks. So it can have information regarding groups_per_page
108 * which is blocks_per_page/2
109 *
110 * The buddy cache inode is not stored on disk. The inode is thrown
111 * away when the filesystem is unmounted.
112 *
113 * We look for count number of blocks in the buddy cache. If we were able
114 * to locate that many free blocks we return with additional information
115 * regarding rest of the contiguous physical block available
116 *
117 * Before allocating blocks via buddy cache we normalize the request
118 * blocks. This ensure we ask for more blocks that we needed. The extra
119 * blocks that we get after allocation is added to the respective prealloc
120 * list. In case of inode preallocation we follow a list of heuristics
121 * based on file size. This can be found in ext4_mb_normalize_request. If
122 * we are doing a group prealloc we try to normalize the request to
123 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
124 * dependent on the cluster size; for non-bigalloc file systems, it is
125 * 512 blocks. This can be tuned via
126 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * smallest multiple of the stripe value (sbi->s_stripe) which is
130 * greater than the default mb_group_prealloc.
131 *
132 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
133 * structures in two data structures:
134 *
135 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
136 *
137 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
138 *
139 * This is an array of lists where the index in the array represents the
140 * largest free order in the buddy bitmap of the participating group infos of
141 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
142 * number of buddy bitmap orders possible) number of lists. Group-infos are
143 * placed in appropriate lists.
144 *
145 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
146 *
147 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
148 *
149 * This is an array of lists where in the i-th list there are groups with
150 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
151 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
152 * Note that we don't bother with a special list for completely empty groups
153 * so we only have MB_NUM_ORDERS(sb) lists.
154 *
155 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
156 * structures to decide the order in which groups are to be traversed for
157 * fulfilling an allocation request.
158 *
159 * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order
160 * >= the order of the request. We directly look at the largest free order list
161 * in the data structure (1) above where largest_free_order = order of the
162 * request. If that list is empty, we look at remaining list in the increasing
163 * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED
164 * lookup in O(1) time.
165 *
166 * At CR_GOAL_LEN_FAST, we only consider groups where
167 * average fragment size > request size. So, we lookup a group which has average
168 * fragment size just above or equal to request size using our average fragment
169 * size group lists (data structure 2) in O(1) time.
170 *
171 * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied
172 * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in
173 * CR_GOAL_LEN_FAST suggests that there is no BG that has avg
174 * fragment size > goal length. So before falling to the slower
175 * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and
176 * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big
177 * enough average fragment size. This increases the chances of finding a
178 * suitable block group in O(1) time and results in faster allocation at the
179 * cost of reduced size of allocation.
180 *
181 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
182 * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and
183 * CR_GOAL_LEN_FAST phase.
184 *
185 * The regular allocator (using the buddy cache) supports a few tunables.
186 *
187 * /sys/fs/ext4/<partition>/mb_min_to_scan
188 * /sys/fs/ext4/<partition>/mb_max_to_scan
189 * /sys/fs/ext4/<partition>/mb_order2_req
190 * /sys/fs/ext4/<partition>/mb_linear_limit
191 *
192 * The regular allocator uses buddy scan only if the request len is power of
193 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
194 * value of s_mb_order2_reqs can be tuned via
195 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
196 * stripe size (sbi->s_stripe), we try to search for contiguous block in
197 * stripe size. This should result in better allocation on RAID setups. If
198 * not, we search in the specific group using bitmap for best extents. The
199 * tunable min_to_scan and max_to_scan control the behaviour here.
200 * min_to_scan indicate how long the mballoc __must__ look for a best
201 * extent and max_to_scan indicates how long the mballoc __can__ look for a
202 * best extent in the found extents. Searching for the blocks starts with
203 * the group specified as the goal value in allocation context via
204 * ac_g_ex. Each group is first checked based on the criteria whether it
205 * can be used for allocation. ext4_mb_good_group explains how the groups are
206 * checked.
207 *
208 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
209 * get traversed linearly. That may result in subsequent allocations being not
210 * close to each other. And so, the underlying device may get filled up in a
211 * non-linear fashion. While that may not matter on non-rotational devices, for
212 * rotational devices that may result in higher seek times. "mb_linear_limit"
213 * tells mballoc how many groups mballoc should search linearly before
214 * performing consulting above data structures for more efficient lookups. For
215 * non rotational devices, this value defaults to 0 and for rotational devices
216 * this is set to MB_DEFAULT_LINEAR_LIMIT.
217 *
218 * Both the prealloc space are getting populated as above. So for the first
219 * request we will hit the buddy cache which will result in this prealloc
220 * space getting filled. The prealloc space is then later used for the
221 * subsequent request.
222 */
223
224 /*
225 * mballoc operates on the following data:
226 * - on-disk bitmap
227 * - in-core buddy (actually includes buddy and bitmap)
228 * - preallocation descriptors (PAs)
229 *
230 * there are two types of preallocations:
231 * - inode
232 * assiged to specific inode and can be used for this inode only.
233 * it describes part of inode's space preallocated to specific
234 * physical blocks. any block from that preallocated can be used
235 * independent. the descriptor just tracks number of blocks left
236 * unused. so, before taking some block from descriptor, one must
237 * make sure corresponded logical block isn't allocated yet. this
238 * also means that freeing any block within descriptor's range
239 * must discard all preallocated blocks.
240 * - locality group
241 * assigned to specific locality group which does not translate to
242 * permanent set of inodes: inode can join and leave group. space
243 * from this type of preallocation can be used for any inode. thus
244 * it's consumed from the beginning to the end.
245 *
246 * relation between them can be expressed as:
247 * in-core buddy = on-disk bitmap + preallocation descriptors
248 *
249 * this mean blocks mballoc considers used are:
250 * - allocated blocks (persistent)
251 * - preallocated blocks (non-persistent)
252 *
253 * consistency in mballoc world means that at any time a block is either
254 * free or used in ALL structures. notice: "any time" should not be read
255 * literally -- time is discrete and delimited by locks.
256 *
257 * to keep it simple, we don't use block numbers, instead we count number of
258 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
259 *
260 * all operations can be expressed as:
261 * - init buddy: buddy = on-disk + PAs
262 * - new PA: buddy += N; PA = N
263 * - use inode PA: on-disk += N; PA -= N
264 * - discard inode PA buddy -= on-disk - PA; PA = 0
265 * - use locality group PA on-disk += N; PA -= N
266 * - discard locality group PA buddy -= PA; PA = 0
267 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
268 * is used in real operation because we can't know actual used
269 * bits from PA, only from on-disk bitmap
270 *
271 * if we follow this strict logic, then all operations above should be atomic.
272 * given some of them can block, we'd have to use something like semaphores
273 * killing performance on high-end SMP hardware. let's try to relax it using
274 * the following knowledge:
275 * 1) if buddy is referenced, it's already initialized
276 * 2) while block is used in buddy and the buddy is referenced,
277 * nobody can re-allocate that block
278 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
279 * bit set and PA claims same block, it's OK. IOW, one can set bit in
280 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
281 * block
282 *
283 * so, now we're building a concurrency table:
284 * - init buddy vs.
285 * - new PA
286 * blocks for PA are allocated in the buddy, buddy must be referenced
287 * until PA is linked to allocation group to avoid concurrent buddy init
288 * - use inode PA
289 * we need to make sure that either on-disk bitmap or PA has uptodate data
290 * given (3) we care that PA-=N operation doesn't interfere with init
291 * - discard inode PA
292 * the simplest way would be to have buddy initialized by the discard
293 * - use locality group PA
294 * again PA-=N must be serialized with init
295 * - discard locality group PA
296 * the simplest way would be to have buddy initialized by the discard
297 * - new PA vs.
298 * - use inode PA
299 * i_data_sem serializes them
300 * - discard inode PA
301 * discard process must wait until PA isn't used by another process
302 * - use locality group PA
303 * some mutex should serialize them
304 * - discard locality group PA
305 * discard process must wait until PA isn't used by another process
306 * - use inode PA
307 * - use inode PA
308 * i_data_sem or another mutex should serializes them
309 * - discard inode PA
310 * discard process must wait until PA isn't used by another process
311 * - use locality group PA
312 * nothing wrong here -- they're different PAs covering different blocks
313 * - discard locality group PA
314 * discard process must wait until PA isn't used by another process
315 *
316 * now we're ready to make few consequences:
317 * - PA is referenced and while it is no discard is possible
318 * - PA is referenced until block isn't marked in on-disk bitmap
319 * - PA changes only after on-disk bitmap
320 * - discard must not compete with init. either init is done before
321 * any discard or they're serialized somehow
322 * - buddy init as sum of on-disk bitmap and PAs is done atomically
323 *
324 * a special case when we've used PA to emptiness. no need to modify buddy
325 * in this case, but we should care about concurrent init
326 *
327 */
328
329 /*
330 * Logic in few words:
331 *
332 * - allocation:
333 * load group
334 * find blocks
335 * mark bits in on-disk bitmap
336 * release group
337 *
338 * - use preallocation:
339 * find proper PA (per-inode or group)
340 * load group
341 * mark bits in on-disk bitmap
342 * release group
343 * release PA
344 *
345 * - free:
346 * load group
347 * mark bits in on-disk bitmap
348 * release group
349 *
350 * - discard preallocations in group:
351 * mark PAs deleted
352 * move them onto local list
353 * load on-disk bitmap
354 * load group
355 * remove PA from object (inode or locality group)
356 * mark free blocks in-core
357 *
358 * - discard inode's preallocations:
359 */
360
361 /*
362 * Locking rules
363 *
364 * Locks:
365 * - bitlock on a group (group)
366 * - object (inode/locality) (object)
367 * - per-pa lock (pa)
368 * - cr_power2_aligned lists lock (cr_power2_aligned)
369 * - cr_goal_len_fast lists lock (cr_goal_len_fast)
370 *
371 * Paths:
372 * - new pa
373 * object
374 * group
375 *
376 * - find and use pa:
377 * pa
378 *
379 * - release consumed pa:
380 * pa
381 * group
382 * object
383 *
384 * - generate in-core bitmap:
385 * group
386 * pa
387 *
388 * - discard all for given object (inode, locality group):
389 * object
390 * pa
391 * group
392 *
393 * - discard all for given group:
394 * group
395 * pa
396 * group
397 * object
398 *
399 * - allocation path (ext4_mb_regular_allocator)
400 * group
401 * cr_power2_aligned/cr_goal_len_fast
402 */
403 static struct kmem_cache *ext4_pspace_cachep;
404 static struct kmem_cache *ext4_ac_cachep;
405 static struct kmem_cache *ext4_free_data_cachep;
406
407 /* We create slab caches for groupinfo data structures based on the
408 * superblock block size. There will be one per mounted filesystem for
409 * each unique s_blocksize_bits */
410 #define NR_GRPINFO_CACHES 8
411 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
412
413 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
414 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
415 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
416 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
417 };
418
419 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
420 ext4_group_t group);
421 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
422
423 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
424 ext4_group_t group, enum criteria cr);
425
426 static int ext4_try_to_trim_range(struct super_block *sb,
427 struct ext4_buddy *e4b, ext4_grpblk_t start,
428 ext4_grpblk_t max, ext4_grpblk_t minblocks);
429
430 /*
431 * The algorithm using this percpu seq counter goes below:
432 * 1. We sample the percpu discard_pa_seq counter before trying for block
433 * allocation in ext4_mb_new_blocks().
434 * 2. We increment this percpu discard_pa_seq counter when we either allocate
435 * or free these blocks i.e. while marking those blocks as used/free in
436 * mb_mark_used()/mb_free_blocks().
437 * 3. We also increment this percpu seq counter when we successfully identify
438 * that the bb_prealloc_list is not empty and hence proceed for discarding
439 * of those PAs inside ext4_mb_discard_group_preallocations().
440 *
441 * Now to make sure that the regular fast path of block allocation is not
442 * affected, as a small optimization we only sample the percpu seq counter
443 * on that cpu. Only when the block allocation fails and when freed blocks
444 * found were 0, that is when we sample percpu seq counter for all cpus using
445 * below function ext4_get_discard_pa_seq_sum(). This happens after making
446 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
447 */
448 static DEFINE_PER_CPU(u64, discard_pa_seq);
449 static inline u64 ext4_get_discard_pa_seq_sum(void)
450 {
451 int __cpu;
452 u64 __seq = 0;
453
454 for_each_possible_cpu(__cpu)
455 __seq += per_cpu(discard_pa_seq, __cpu);
456 return __seq;
457 }
458
459 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
460 {
461 #if BITS_PER_LONG == 64
462 *bit += ((unsigned long) addr & 7UL) << 3;
463 addr = (void *) ((unsigned long) addr & ~7UL);
464 #elif BITS_PER_LONG == 32
465 *bit += ((unsigned long) addr & 3UL) << 3;
466 addr = (void *) ((unsigned long) addr & ~3UL);
467 #else
468 #error "how many bits you are?!"
469 #endif
470 return addr;
471 }
472
473 static inline int mb_test_bit(int bit, void *addr)
474 {
475 /*
476 * ext4_test_bit on architecture like powerpc
477 * needs unsigned long aligned address
478 */
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 return ext4_test_bit(bit, addr);
481 }
482
483 static inline void mb_set_bit(int bit, void *addr)
484 {
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 ext4_set_bit(bit, addr);
487 }
488
489 static inline void mb_clear_bit(int bit, void *addr)
490 {
491 addr = mb_correct_addr_and_bit(&bit, addr);
492 ext4_clear_bit(bit, addr);
493 }
494
495 static inline int mb_test_and_clear_bit(int bit, void *addr)
496 {
497 addr = mb_correct_addr_and_bit(&bit, addr);
498 return ext4_test_and_clear_bit(bit, addr);
499 }
500
501 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
502 {
503 int fix = 0, ret, tmpmax;
504 addr = mb_correct_addr_and_bit(&fix, addr);
505 tmpmax = max + fix;
506 start += fix;
507
508 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
509 if (ret > max)
510 return max;
511 return ret;
512 }
513
514 static inline int mb_find_next_bit(void *addr, int max, int start)
515 {
516 int fix = 0, ret, tmpmax;
517 addr = mb_correct_addr_and_bit(&fix, addr);
518 tmpmax = max + fix;
519 start += fix;
520
521 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
522 if (ret > max)
523 return max;
524 return ret;
525 }
526
527 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
528 {
529 char *bb;
530
531 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
532 BUG_ON(max == NULL);
533
534 if (order > e4b->bd_blkbits + 1) {
535 *max = 0;
536 return NULL;
537 }
538
539 /* at order 0 we see each particular block */
540 if (order == 0) {
541 *max = 1 << (e4b->bd_blkbits + 3);
542 return e4b->bd_bitmap;
543 }
544
545 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
546 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
547
548 return bb;
549 }
550
551 #ifdef DOUBLE_CHECK
552 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
553 int first, int count)
554 {
555 int i;
556 struct super_block *sb = e4b->bd_sb;
557
558 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
559 return;
560 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
561 for (i = 0; i < count; i++) {
562 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
563 ext4_fsblk_t blocknr;
564
565 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
566 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
567 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
568 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
569 ext4_grp_locked_error(sb, e4b->bd_group,
570 inode ? inode->i_ino : 0,
571 blocknr,
572 "freeing block already freed "
573 "(bit %u)",
574 first + i);
575 }
576 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
577 }
578 }
579
580 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
581 {
582 int i;
583
584 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
585 return;
586 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
587 for (i = 0; i < count; i++) {
588 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
589 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
590 }
591 }
592
593 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
594 {
595 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
596 return;
597 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
598 unsigned char *b1, *b2;
599 int i;
600 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
601 b2 = (unsigned char *) bitmap;
602 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
603 if (b1[i] != b2[i]) {
604 ext4_msg(e4b->bd_sb, KERN_ERR,
605 "corruption in group %u "
606 "at byte %u(%u): %x in copy != %x "
607 "on disk/prealloc",
608 e4b->bd_group, i, i * 8, b1[i], b2[i]);
609 BUG();
610 }
611 }
612 }
613 }
614
615 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
616 struct ext4_group_info *grp, ext4_group_t group)
617 {
618 struct buffer_head *bh;
619
620 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
621 if (!grp->bb_bitmap)
622 return;
623
624 bh = ext4_read_block_bitmap(sb, group);
625 if (IS_ERR_OR_NULL(bh)) {
626 kfree(grp->bb_bitmap);
627 grp->bb_bitmap = NULL;
628 return;
629 }
630
631 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
632 put_bh(bh);
633 }
634
635 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
636 {
637 kfree(grp->bb_bitmap);
638 }
639
640 #else
641 static inline void mb_free_blocks_double(struct inode *inode,
642 struct ext4_buddy *e4b, int first, int count)
643 {
644 return;
645 }
646 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
647 int first, int count)
648 {
649 return;
650 }
651 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
652 {
653 return;
654 }
655
656 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
657 struct ext4_group_info *grp, ext4_group_t group)
658 {
659 return;
660 }
661
662 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
663 {
664 return;
665 }
666 #endif
667
668 #ifdef AGGRESSIVE_CHECK
669
670 #define MB_CHECK_ASSERT(assert) \
671 do { \
672 if (!(assert)) { \
673 printk(KERN_EMERG \
674 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
675 function, file, line, # assert); \
676 BUG(); \
677 } \
678 } while (0)
679
680 static void __mb_check_buddy(struct ext4_buddy *e4b, char *file,
681 const char *function, int line)
682 {
683 struct super_block *sb = e4b->bd_sb;
684 int order = e4b->bd_blkbits + 1;
685 int max;
686 int max2;
687 int i;
688 int j;
689 int k;
690 int count;
691 struct ext4_group_info *grp;
692 int fragments = 0;
693 int fstart;
694 struct list_head *cur;
695 void *buddy;
696 void *buddy2;
697
698 if (e4b->bd_info->bb_check_counter++ % 10)
699 return;
700
701 while (order > 1) {
702 buddy = mb_find_buddy(e4b, order, &max);
703 MB_CHECK_ASSERT(buddy);
704 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
705 MB_CHECK_ASSERT(buddy2);
706 MB_CHECK_ASSERT(buddy != buddy2);
707 MB_CHECK_ASSERT(max * 2 == max2);
708
709 count = 0;
710 for (i = 0; i < max; i++) {
711
712 if (mb_test_bit(i, buddy)) {
713 /* only single bit in buddy2 may be 0 */
714 if (!mb_test_bit(i << 1, buddy2)) {
715 MB_CHECK_ASSERT(
716 mb_test_bit((i<<1)+1, buddy2));
717 }
718 continue;
719 }
720
721 /* both bits in buddy2 must be 1 */
722 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
723 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
724
725 for (j = 0; j < (1 << order); j++) {
726 k = (i * (1 << order)) + j;
727 MB_CHECK_ASSERT(
728 !mb_test_bit(k, e4b->bd_bitmap));
729 }
730 count++;
731 }
732 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
733 order--;
734 }
735
736 fstart = -1;
737 buddy = mb_find_buddy(e4b, 0, &max);
738 for (i = 0; i < max; i++) {
739 if (!mb_test_bit(i, buddy)) {
740 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
741 if (fstart == -1) {
742 fragments++;
743 fstart = i;
744 }
745 continue;
746 }
747 fstart = -1;
748 /* check used bits only */
749 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
750 buddy2 = mb_find_buddy(e4b, j, &max2);
751 k = i >> j;
752 MB_CHECK_ASSERT(k < max2);
753 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
754 }
755 }
756 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
757 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
758
759 grp = ext4_get_group_info(sb, e4b->bd_group);
760 if (!grp)
761 return;
762 list_for_each(cur, &grp->bb_prealloc_list) {
763 ext4_group_t groupnr;
764 struct ext4_prealloc_space *pa;
765 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
766 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
767 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
768 for (i = 0; i < pa->pa_len; i++)
769 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
770 }
771 }
772 #undef MB_CHECK_ASSERT
773 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
774 __FILE__, __func__, __LINE__)
775 #else
776 #define mb_check_buddy(e4b)
777 #endif
778
779 /*
780 * Divide blocks started from @first with length @len into
781 * smaller chunks with power of 2 blocks.
782 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
783 * then increase bb_counters[] for corresponded chunk size.
784 */
785 static void ext4_mb_mark_free_simple(struct super_block *sb,
786 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
787 struct ext4_group_info *grp)
788 {
789 struct ext4_sb_info *sbi = EXT4_SB(sb);
790 ext4_grpblk_t min;
791 ext4_grpblk_t max;
792 ext4_grpblk_t chunk;
793 unsigned int border;
794
795 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
796
797 border = 2 << sb->s_blocksize_bits;
798
799 while (len > 0) {
800 /* find how many blocks can be covered since this position */
801 max = ffs(first | border) - 1;
802
803 /* find how many blocks of power 2 we need to mark */
804 min = fls(len) - 1;
805
806 if (max < min)
807 min = max;
808 chunk = 1 << min;
809
810 /* mark multiblock chunks only */
811 grp->bb_counters[min]++;
812 if (min > 0)
813 mb_clear_bit(first >> min,
814 buddy + sbi->s_mb_offsets[min]);
815
816 len -= chunk;
817 first += chunk;
818 }
819 }
820
821 static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
822 {
823 int order;
824
825 /*
826 * We don't bother with a special lists groups with only 1 block free
827 * extents and for completely empty groups.
828 */
829 order = fls(len) - 2;
830 if (order < 0)
831 return 0;
832 if (order == MB_NUM_ORDERS(sb))
833 order--;
834 if (WARN_ON_ONCE(order > MB_NUM_ORDERS(sb)))
835 order = MB_NUM_ORDERS(sb) - 1;
836 return order;
837 }
838
839 /* Move group to appropriate avg_fragment_size list */
840 static void
841 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
842 {
843 struct ext4_sb_info *sbi = EXT4_SB(sb);
844 int new_order;
845
846 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 0)
847 return;
848
849 new_order = mb_avg_fragment_size_order(sb,
850 grp->bb_free / grp->bb_fragments);
851 if (new_order == grp->bb_avg_fragment_size_order)
852 return;
853
854 if (grp->bb_avg_fragment_size_order != -1) {
855 write_lock(&sbi->s_mb_avg_fragment_size_locks[
856 grp->bb_avg_fragment_size_order]);
857 list_del(&grp->bb_avg_fragment_size_node);
858 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
859 grp->bb_avg_fragment_size_order]);
860 }
861 grp->bb_avg_fragment_size_order = new_order;
862 write_lock(&sbi->s_mb_avg_fragment_size_locks[
863 grp->bb_avg_fragment_size_order]);
864 list_add_tail(&grp->bb_avg_fragment_size_node,
865 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
866 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
867 grp->bb_avg_fragment_size_order]);
868 }
869
870 /*
871 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
872 * cr level needs an update.
873 */
874 static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac,
875 enum criteria *new_cr, ext4_group_t *group)
876 {
877 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
878 struct ext4_group_info *iter;
879 int i;
880
881 if (ac->ac_status == AC_STATUS_FOUND)
882 return;
883
884 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED))
885 atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions);
886
887 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
888 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
889 continue;
890 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
891 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
892 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
893 continue;
894 }
895 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
896 bb_largest_free_order_node) {
897 if (sbi->s_mb_stats)
898 atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]);
899 if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) {
900 *group = iter->bb_group;
901 ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED;
902 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
903 return;
904 }
905 }
906 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
907 }
908
909 /* Increment cr and search again if no group is found */
910 *new_cr = CR_GOAL_LEN_FAST;
911 }
912
913 /*
914 * Find a suitable group of given order from the average fragments list.
915 */
916 static struct ext4_group_info *
917 ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order)
918 {
919 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
920 struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order];
921 rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order];
922 struct ext4_group_info *grp = NULL, *iter;
923 enum criteria cr = ac->ac_criteria;
924
925 if (list_empty(frag_list))
926 return NULL;
927 read_lock(frag_list_lock);
928 if (list_empty(frag_list)) {
929 read_unlock(frag_list_lock);
930 return NULL;
931 }
932 list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) {
933 if (sbi->s_mb_stats)
934 atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]);
935 if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) {
936 grp = iter;
937 break;
938 }
939 }
940 read_unlock(frag_list_lock);
941 return grp;
942 }
943
944 /*
945 * Choose next group by traversing average fragment size list of suitable
946 * order. Updates *new_cr if cr level needs an update.
947 */
948 static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac,
949 enum criteria *new_cr, ext4_group_t *group)
950 {
951 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
952 struct ext4_group_info *grp = NULL;
953 int i;
954
955 if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) {
956 if (sbi->s_mb_stats)
957 atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions);
958 }
959
960 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
961 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
962 grp = ext4_mb_find_good_group_avg_frag_lists(ac, i);
963 if (grp) {
964 *group = grp->bb_group;
965 ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED;
966 return;
967 }
968 }
969
970 /*
971 * CR_BEST_AVAIL_LEN works based on the concept that we have
972 * a larger normalized goal len request which can be trimmed to
973 * a smaller goal len such that it can still satisfy original
974 * request len. However, allocation request for non-regular
975 * files never gets normalized.
976 * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA).
977 */
978 if (ac->ac_flags & EXT4_MB_HINT_DATA)
979 *new_cr = CR_BEST_AVAIL_LEN;
980 else
981 *new_cr = CR_GOAL_LEN_SLOW;
982 }
983
984 /*
985 * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment
986 * order we have and proactively trim the goal request length to that order to
987 * find a suitable group faster.
988 *
989 * This optimizes allocation speed at the cost of slightly reduced
990 * preallocations. However, we make sure that we don't trim the request too
991 * much and fall to CR_GOAL_LEN_SLOW in that case.
992 */
993 static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac,
994 enum criteria *new_cr, ext4_group_t *group)
995 {
996 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
997 struct ext4_group_info *grp = NULL;
998 int i, order, min_order;
999 unsigned long num_stripe_clusters = 0;
1000
1001 if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) {
1002 if (sbi->s_mb_stats)
1003 atomic_inc(&sbi->s_bal_best_avail_bad_suggestions);
1004 }
1005
1006 /*
1007 * mb_avg_fragment_size_order() returns order in a way that makes
1008 * retrieving back the length using (1 << order) inaccurate. Hence, use
1009 * fls() instead since we need to know the actual length while modifying
1010 * goal length.
1011 */
1012 order = fls(ac->ac_g_ex.fe_len) - 1;
1013 if (WARN_ON_ONCE(order - 1 > MB_NUM_ORDERS(ac->ac_sb)))
1014 order = MB_NUM_ORDERS(ac->ac_sb);
1015 min_order = order - sbi->s_mb_best_avail_max_trim_order;
1016 if (min_order < 0)
1017 min_order = 0;
1018
1019 if (sbi->s_stripe > 0) {
1020 /*
1021 * We are assuming that stripe size is always a multiple of
1022 * cluster ratio otherwise __ext4_fill_super exists early.
1023 */
1024 num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe);
1025 if (1 << min_order < num_stripe_clusters)
1026 /*
1027 * We consider 1 order less because later we round
1028 * up the goal len to num_stripe_clusters
1029 */
1030 min_order = fls(num_stripe_clusters) - 1;
1031 }
1032
1033 if (1 << min_order < ac->ac_o_ex.fe_len)
1034 min_order = fls(ac->ac_o_ex.fe_len);
1035
1036 for (i = order; i >= min_order; i--) {
1037 int frag_order;
1038 /*
1039 * Scale down goal len to make sure we find something
1040 * in the free fragments list. Basically, reduce
1041 * preallocations.
1042 */
1043 ac->ac_g_ex.fe_len = 1 << i;
1044
1045 if (num_stripe_clusters > 0) {
1046 /*
1047 * Try to round up the adjusted goal length to
1048 * stripe size (in cluster units) multiple for
1049 * efficiency.
1050 */
1051 ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len,
1052 num_stripe_clusters);
1053 }
1054
1055 frag_order = mb_avg_fragment_size_order(ac->ac_sb,
1056 ac->ac_g_ex.fe_len);
1057
1058 grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order);
1059 if (grp) {
1060 *group = grp->bb_group;
1061 ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED;
1062 return;
1063 }
1064 }
1065
1066 /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */
1067 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
1068 *new_cr = CR_GOAL_LEN_SLOW;
1069 }
1070
1071 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
1072 {
1073 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
1074 return 0;
1075 if (ac->ac_criteria >= CR_GOAL_LEN_SLOW)
1076 return 0;
1077 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1078 return 0;
1079 return 1;
1080 }
1081
1082 /*
1083 * Return next linear group for allocation.
1084 */
1085 static ext4_group_t
1086 next_linear_group(ext4_group_t group, ext4_group_t ngroups)
1087 {
1088 /*
1089 * Artificially restricted ngroups for non-extent
1090 * files makes group > ngroups possible on first loop.
1091 */
1092 return group + 1 >= ngroups ? 0 : group + 1;
1093 }
1094
1095 /*
1096 * ext4_mb_choose_next_group: choose next group for allocation.
1097 *
1098 * @ac Allocation Context
1099 * @new_cr This is an output parameter. If the there is no good group
1100 * available at current CR level, this field is updated to indicate
1101 * the new cr level that should be used.
1102 * @group This is an input / output parameter. As an input it indicates the
1103 * next group that the allocator intends to use for allocation. As
1104 * output, this field indicates the next group that should be used as
1105 * determined by the optimization functions.
1106 * @ngroups Total number of groups
1107 */
1108 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1109 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1110 {
1111 *new_cr = ac->ac_criteria;
1112
1113 if (!should_optimize_scan(ac)) {
1114 *group = next_linear_group(*group, ngroups);
1115 return;
1116 }
1117
1118 /*
1119 * Optimized scanning can return non adjacent groups which can cause
1120 * seek overhead for rotational disks. So try few linear groups before
1121 * trying optimized scan.
1122 */
1123 if (ac->ac_groups_linear_remaining) {
1124 *group = next_linear_group(*group, ngroups);
1125 ac->ac_groups_linear_remaining--;
1126 return;
1127 }
1128
1129 if (*new_cr == CR_POWER2_ALIGNED) {
1130 ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group);
1131 } else if (*new_cr == CR_GOAL_LEN_FAST) {
1132 ext4_mb_choose_next_group_goal_fast(ac, new_cr, group);
1133 } else if (*new_cr == CR_BEST_AVAIL_LEN) {
1134 ext4_mb_choose_next_group_best_avail(ac, new_cr, group);
1135 } else {
1136 /*
1137 * TODO: For CR_GOAL_LEN_SLOW, we can arrange groups in an
1138 * rb tree sorted by bb_free. But until that happens, we should
1139 * never come here.
1140 */
1141 WARN_ON(1);
1142 }
1143 }
1144
1145 /*
1146 * Cache the order of the largest free extent we have available in this block
1147 * group.
1148 */
1149 static void
1150 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1151 {
1152 struct ext4_sb_info *sbi = EXT4_SB(sb);
1153 int i;
1154
1155 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1156 if (grp->bb_counters[i] > 0)
1157 break;
1158 /* No need to move between order lists? */
1159 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1160 i == grp->bb_largest_free_order) {
1161 grp->bb_largest_free_order = i;
1162 return;
1163 }
1164
1165 if (grp->bb_largest_free_order >= 0) {
1166 write_lock(&sbi->s_mb_largest_free_orders_locks[
1167 grp->bb_largest_free_order]);
1168 list_del_init(&grp->bb_largest_free_order_node);
1169 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1170 grp->bb_largest_free_order]);
1171 }
1172 grp->bb_largest_free_order = i;
1173 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1174 write_lock(&sbi->s_mb_largest_free_orders_locks[
1175 grp->bb_largest_free_order]);
1176 list_add_tail(&grp->bb_largest_free_order_node,
1177 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1178 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1179 grp->bb_largest_free_order]);
1180 }
1181 }
1182
1183 static noinline_for_stack
1184 void ext4_mb_generate_buddy(struct super_block *sb,
1185 void *buddy, void *bitmap, ext4_group_t group,
1186 struct ext4_group_info *grp)
1187 {
1188 struct ext4_sb_info *sbi = EXT4_SB(sb);
1189 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1190 ext4_grpblk_t i = 0;
1191 ext4_grpblk_t first;
1192 ext4_grpblk_t len;
1193 unsigned free = 0;
1194 unsigned fragments = 0;
1195 unsigned long long period = get_cycles();
1196
1197 /* initialize buddy from bitmap which is aggregation
1198 * of on-disk bitmap and preallocations */
1199 i = mb_find_next_zero_bit(bitmap, max, 0);
1200 grp->bb_first_free = i;
1201 while (i < max) {
1202 fragments++;
1203 first = i;
1204 i = mb_find_next_bit(bitmap, max, i);
1205 len = i - first;
1206 free += len;
1207 if (len > 1)
1208 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1209 else
1210 grp->bb_counters[0]++;
1211 if (i < max)
1212 i = mb_find_next_zero_bit(bitmap, max, i);
1213 }
1214 grp->bb_fragments = fragments;
1215
1216 if (free != grp->bb_free) {
1217 ext4_grp_locked_error(sb, group, 0, 0,
1218 "block bitmap and bg descriptor "
1219 "inconsistent: %u vs %u free clusters",
1220 free, grp->bb_free);
1221 /*
1222 * If we intend to continue, we consider group descriptor
1223 * corrupt and update bb_free using bitmap value
1224 */
1225 grp->bb_free = free;
1226 ext4_mark_group_bitmap_corrupted(sb, group,
1227 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1228 }
1229 mb_set_largest_free_order(sb, grp);
1230 mb_update_avg_fragment_size(sb, grp);
1231
1232 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1233
1234 period = get_cycles() - period;
1235 atomic_inc(&sbi->s_mb_buddies_generated);
1236 atomic64_add(period, &sbi->s_mb_generation_time);
1237 }
1238
1239 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
1240 {
1241 int count;
1242 int order = 1;
1243 void *buddy;
1244
1245 while ((buddy = mb_find_buddy(e4b, order++, &count)))
1246 mb_set_bits(buddy, 0, count);
1247
1248 e4b->bd_info->bb_fragments = 0;
1249 memset(e4b->bd_info->bb_counters, 0,
1250 sizeof(*e4b->bd_info->bb_counters) *
1251 (e4b->bd_sb->s_blocksize_bits + 2));
1252
1253 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
1254 e4b->bd_bitmap, e4b->bd_group, e4b->bd_info);
1255 }
1256
1257 /* The buddy information is attached the buddy cache inode
1258 * for convenience. The information regarding each group
1259 * is loaded via ext4_mb_load_buddy. The information involve
1260 * block bitmap and buddy information. The information are
1261 * stored in the inode as
1262 *
1263 * { page }
1264 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1265 *
1266 *
1267 * one block each for bitmap and buddy information.
1268 * So for each group we take up 2 blocks. A page can
1269 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1270 * So it can have information regarding groups_per_page which
1271 * is blocks_per_page/2
1272 *
1273 * Locking note: This routine takes the block group lock of all groups
1274 * for this page; do not hold this lock when calling this routine!
1275 */
1276
1277 static int ext4_mb_init_cache(struct folio *folio, char *incore, gfp_t gfp)
1278 {
1279 ext4_group_t ngroups;
1280 unsigned int blocksize;
1281 int blocks_per_page;
1282 int groups_per_page;
1283 int err = 0;
1284 int i;
1285 ext4_group_t first_group, group;
1286 int first_block;
1287 struct super_block *sb;
1288 struct buffer_head *bhs;
1289 struct buffer_head **bh = NULL;
1290 struct inode *inode;
1291 char *data;
1292 char *bitmap;
1293 struct ext4_group_info *grinfo;
1294
1295 inode = folio->mapping->host;
1296 sb = inode->i_sb;
1297 ngroups = ext4_get_groups_count(sb);
1298 blocksize = i_blocksize(inode);
1299 blocks_per_page = PAGE_SIZE / blocksize;
1300
1301 mb_debug(sb, "init folio %lu\n", folio->index);
1302
1303 groups_per_page = blocks_per_page >> 1;
1304 if (groups_per_page == 0)
1305 groups_per_page = 1;
1306
1307 /* allocate buffer_heads to read bitmaps */
1308 if (groups_per_page > 1) {
1309 i = sizeof(struct buffer_head *) * groups_per_page;
1310 bh = kzalloc(i, gfp);
1311 if (bh == NULL)
1312 return -ENOMEM;
1313 } else
1314 bh = &bhs;
1315
1316 first_group = folio->index * blocks_per_page / 2;
1317
1318 /* read all groups the folio covers into the cache */
1319 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1320 if (group >= ngroups)
1321 break;
1322
1323 grinfo = ext4_get_group_info(sb, group);
1324 if (!grinfo)
1325 continue;
1326 /*
1327 * If page is uptodate then we came here after online resize
1328 * which added some new uninitialized group info structs, so
1329 * we must skip all initialized uptodate buddies on the folio,
1330 * which may be currently in use by an allocating task.
1331 */
1332 if (folio_test_uptodate(folio) &&
1333 !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1334 bh[i] = NULL;
1335 continue;
1336 }
1337 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1338 if (IS_ERR(bh[i])) {
1339 err = PTR_ERR(bh[i]);
1340 bh[i] = NULL;
1341 goto out;
1342 }
1343 mb_debug(sb, "read bitmap for group %u\n", group);
1344 }
1345
1346 /* wait for I/O completion */
1347 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1348 int err2;
1349
1350 if (!bh[i])
1351 continue;
1352 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1353 if (!err)
1354 err = err2;
1355 }
1356
1357 first_block = folio->index * blocks_per_page;
1358 for (i = 0; i < blocks_per_page; i++) {
1359 group = (first_block + i) >> 1;
1360 if (group >= ngroups)
1361 break;
1362
1363 if (!bh[group - first_group])
1364 /* skip initialized uptodate buddy */
1365 continue;
1366
1367 if (!buffer_verified(bh[group - first_group]))
1368 /* Skip faulty bitmaps */
1369 continue;
1370 err = 0;
1371
1372 /*
1373 * data carry information regarding this
1374 * particular group in the format specified
1375 * above
1376 *
1377 */
1378 data = folio_address(folio) + (i * blocksize);
1379 bitmap = bh[group - first_group]->b_data;
1380
1381 /*
1382 * We place the buddy block and bitmap block
1383 * close together
1384 */
1385 grinfo = ext4_get_group_info(sb, group);
1386 if (!grinfo) {
1387 err = -EFSCORRUPTED;
1388 goto out;
1389 }
1390 if ((first_block + i) & 1) {
1391 /* this is block of buddy */
1392 BUG_ON(incore == NULL);
1393 mb_debug(sb, "put buddy for group %u in folio %lu/%x\n",
1394 group, folio->index, i * blocksize);
1395 trace_ext4_mb_buddy_bitmap_load(sb, group);
1396 grinfo->bb_fragments = 0;
1397 memset(grinfo->bb_counters, 0,
1398 sizeof(*grinfo->bb_counters) *
1399 (MB_NUM_ORDERS(sb)));
1400 /*
1401 * incore got set to the group block bitmap below
1402 */
1403 ext4_lock_group(sb, group);
1404 /* init the buddy */
1405 memset(data, 0xff, blocksize);
1406 ext4_mb_generate_buddy(sb, data, incore, group, grinfo);
1407 ext4_unlock_group(sb, group);
1408 incore = NULL;
1409 } else {
1410 /* this is block of bitmap */
1411 BUG_ON(incore != NULL);
1412 mb_debug(sb, "put bitmap for group %u in folio %lu/%x\n",
1413 group, folio->index, i * blocksize);
1414 trace_ext4_mb_bitmap_load(sb, group);
1415
1416 /* see comments in ext4_mb_put_pa() */
1417 ext4_lock_group(sb, group);
1418 memcpy(data, bitmap, blocksize);
1419
1420 /* mark all preallocated blks used in in-core bitmap */
1421 ext4_mb_generate_from_pa(sb, data, group);
1422 WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root));
1423 ext4_unlock_group(sb, group);
1424
1425 /* set incore so that the buddy information can be
1426 * generated using this
1427 */
1428 incore = data;
1429 }
1430 }
1431 folio_mark_uptodate(folio);
1432
1433 out:
1434 if (bh) {
1435 for (i = 0; i < groups_per_page; i++)
1436 brelse(bh[i]);
1437 if (bh != &bhs)
1438 kfree(bh);
1439 }
1440 return err;
1441 }
1442
1443 /*
1444 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1445 * on the same buddy page doesn't happen whild holding the buddy page lock.
1446 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1447 * are on the same page e4b->bd_buddy_folio is NULL and return value is 0.
1448 */
1449 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1450 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1451 {
1452 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1453 int block, pnum, poff;
1454 int blocks_per_page;
1455 struct folio *folio;
1456
1457 e4b->bd_buddy_folio = NULL;
1458 e4b->bd_bitmap_folio = NULL;
1459
1460 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1461 /*
1462 * the buddy cache inode stores the block bitmap
1463 * and buddy information in consecutive blocks.
1464 * So for each group we need two blocks.
1465 */
1466 block = group * 2;
1467 pnum = block / blocks_per_page;
1468 poff = block % blocks_per_page;
1469 folio = __filemap_get_folio(inode->i_mapping, pnum,
1470 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1471 if (IS_ERR(folio))
1472 return PTR_ERR(folio);
1473 BUG_ON(folio->mapping != inode->i_mapping);
1474 e4b->bd_bitmap_folio = folio;
1475 e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize);
1476
1477 if (blocks_per_page >= 2) {
1478 /* buddy and bitmap are on the same page */
1479 return 0;
1480 }
1481
1482 /* blocks_per_page == 1, hence we need another page for the buddy */
1483 folio = __filemap_get_folio(inode->i_mapping, block + 1,
1484 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1485 if (IS_ERR(folio))
1486 return PTR_ERR(folio);
1487 BUG_ON(folio->mapping != inode->i_mapping);
1488 e4b->bd_buddy_folio = folio;
1489 return 0;
1490 }
1491
1492 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1493 {
1494 if (e4b->bd_bitmap_folio) {
1495 folio_unlock(e4b->bd_bitmap_folio);
1496 folio_put(e4b->bd_bitmap_folio);
1497 }
1498 if (e4b->bd_buddy_folio) {
1499 folio_unlock(e4b->bd_buddy_folio);
1500 folio_put(e4b->bd_buddy_folio);
1501 }
1502 }
1503
1504 /*
1505 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1506 * block group lock of all groups for this page; do not hold the BG lock when
1507 * calling this routine!
1508 */
1509 static noinline_for_stack
1510 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1511 {
1512
1513 struct ext4_group_info *this_grp;
1514 struct ext4_buddy e4b;
1515 struct folio *folio;
1516 int ret = 0;
1517
1518 might_sleep();
1519 mb_debug(sb, "init group %u\n", group);
1520 this_grp = ext4_get_group_info(sb, group);
1521 if (!this_grp)
1522 return -EFSCORRUPTED;
1523
1524 /*
1525 * This ensures that we don't reinit the buddy cache
1526 * page which map to the group from which we are already
1527 * allocating. If we are looking at the buddy cache we would
1528 * have taken a reference using ext4_mb_load_buddy and that
1529 * would have pinned buddy page to page cache.
1530 * The call to ext4_mb_get_buddy_page_lock will mark the
1531 * page accessed.
1532 */
1533 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1534 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1535 /*
1536 * somebody initialized the group
1537 * return without doing anything
1538 */
1539 goto err;
1540 }
1541
1542 folio = e4b.bd_bitmap_folio;
1543 ret = ext4_mb_init_cache(folio, NULL, gfp);
1544 if (ret)
1545 goto err;
1546 if (!folio_test_uptodate(folio)) {
1547 ret = -EIO;
1548 goto err;
1549 }
1550
1551 if (e4b.bd_buddy_folio == NULL) {
1552 /*
1553 * If both the bitmap and buddy are in
1554 * the same page we don't need to force
1555 * init the buddy
1556 */
1557 ret = 0;
1558 goto err;
1559 }
1560 /* init buddy cache */
1561 folio = e4b.bd_buddy_folio;
1562 ret = ext4_mb_init_cache(folio, e4b.bd_bitmap, gfp);
1563 if (ret)
1564 goto err;
1565 if (!folio_test_uptodate(folio)) {
1566 ret = -EIO;
1567 goto err;
1568 }
1569 err:
1570 ext4_mb_put_buddy_page_lock(&e4b);
1571 return ret;
1572 }
1573
1574 /*
1575 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1576 * block group lock of all groups for this page; do not hold the BG lock when
1577 * calling this routine!
1578 */
1579 static noinline_for_stack int
1580 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1581 struct ext4_buddy *e4b, gfp_t gfp)
1582 {
1583 int blocks_per_page;
1584 int block;
1585 int pnum;
1586 int poff;
1587 struct folio *folio;
1588 int ret;
1589 struct ext4_group_info *grp;
1590 struct ext4_sb_info *sbi = EXT4_SB(sb);
1591 struct inode *inode = sbi->s_buddy_cache;
1592
1593 might_sleep();
1594 mb_debug(sb, "load group %u\n", group);
1595
1596 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1597 grp = ext4_get_group_info(sb, group);
1598 if (!grp)
1599 return -EFSCORRUPTED;
1600
1601 e4b->bd_blkbits = sb->s_blocksize_bits;
1602 e4b->bd_info = grp;
1603 e4b->bd_sb = sb;
1604 e4b->bd_group = group;
1605 e4b->bd_buddy_folio = NULL;
1606 e4b->bd_bitmap_folio = NULL;
1607
1608 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1609 /*
1610 * we need full data about the group
1611 * to make a good selection
1612 */
1613 ret = ext4_mb_init_group(sb, group, gfp);
1614 if (ret)
1615 return ret;
1616 }
1617
1618 /*
1619 * the buddy cache inode stores the block bitmap
1620 * and buddy information in consecutive blocks.
1621 * So for each group we need two blocks.
1622 */
1623 block = group * 2;
1624 pnum = block / blocks_per_page;
1625 poff = block % blocks_per_page;
1626
1627 /* Avoid locking the folio in the fast path ... */
1628 folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0);
1629 if (IS_ERR(folio) || !folio_test_uptodate(folio)) {
1630 if (!IS_ERR(folio))
1631 /*
1632 * drop the folio reference and try
1633 * to get the folio with lock. If we
1634 * are not uptodate that implies
1635 * somebody just created the folio but
1636 * is yet to initialize it. So
1637 * wait for it to initialize.
1638 */
1639 folio_put(folio);
1640 folio = __filemap_get_folio(inode->i_mapping, pnum,
1641 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1642 if (!IS_ERR(folio)) {
1643 if (WARN_RATELIMIT(folio->mapping != inode->i_mapping,
1644 "ext4: bitmap's mapping != inode->i_mapping\n")) {
1645 /* should never happen */
1646 folio_unlock(folio);
1647 ret = -EINVAL;
1648 goto err;
1649 }
1650 if (!folio_test_uptodate(folio)) {
1651 ret = ext4_mb_init_cache(folio, NULL, gfp);
1652 if (ret) {
1653 folio_unlock(folio);
1654 goto err;
1655 }
1656 mb_cmp_bitmaps(e4b, folio_address(folio) +
1657 (poff * sb->s_blocksize));
1658 }
1659 folio_unlock(folio);
1660 }
1661 }
1662 if (IS_ERR(folio)) {
1663 ret = PTR_ERR(folio);
1664 goto err;
1665 }
1666 if (!folio_test_uptodate(folio)) {
1667 ret = -EIO;
1668 goto err;
1669 }
1670
1671 /* Folios marked accessed already */
1672 e4b->bd_bitmap_folio = folio;
1673 e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize);
1674
1675 block++;
1676 pnum = block / blocks_per_page;
1677 poff = block % blocks_per_page;
1678
1679 folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0);
1680 if (IS_ERR(folio) || !folio_test_uptodate(folio)) {
1681 if (!IS_ERR(folio))
1682 folio_put(folio);
1683 folio = __filemap_get_folio(inode->i_mapping, pnum,
1684 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1685 if (!IS_ERR(folio)) {
1686 if (WARN_RATELIMIT(folio->mapping != inode->i_mapping,
1687 "ext4: buddy bitmap's mapping != inode->i_mapping\n")) {
1688 /* should never happen */
1689 folio_unlock(folio);
1690 ret = -EINVAL;
1691 goto err;
1692 }
1693 if (!folio_test_uptodate(folio)) {
1694 ret = ext4_mb_init_cache(folio, e4b->bd_bitmap,
1695 gfp);
1696 if (ret) {
1697 folio_unlock(folio);
1698 goto err;
1699 }
1700 }
1701 folio_unlock(folio);
1702 }
1703 }
1704 if (IS_ERR(folio)) {
1705 ret = PTR_ERR(folio);
1706 goto err;
1707 }
1708 if (!folio_test_uptodate(folio)) {
1709 ret = -EIO;
1710 goto err;
1711 }
1712
1713 /* Folios marked accessed already */
1714 e4b->bd_buddy_folio = folio;
1715 e4b->bd_buddy = folio_address(folio) + (poff * sb->s_blocksize);
1716
1717 return 0;
1718
1719 err:
1720 if (!IS_ERR_OR_NULL(folio))
1721 folio_put(folio);
1722 if (e4b->bd_bitmap_folio)
1723 folio_put(e4b->bd_bitmap_folio);
1724
1725 e4b->bd_buddy = NULL;
1726 e4b->bd_bitmap = NULL;
1727 return ret;
1728 }
1729
1730 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1731 struct ext4_buddy *e4b)
1732 {
1733 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1734 }
1735
1736 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1737 {
1738 if (e4b->bd_bitmap_folio)
1739 folio_put(e4b->bd_bitmap_folio);
1740 if (e4b->bd_buddy_folio)
1741 folio_put(e4b->bd_buddy_folio);
1742 }
1743
1744
1745 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1746 {
1747 int order = 1, max;
1748 void *bb;
1749
1750 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1751 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1752
1753 while (order <= e4b->bd_blkbits + 1) {
1754 bb = mb_find_buddy(e4b, order, &max);
1755 if (!mb_test_bit(block >> order, bb)) {
1756 /* this block is part of buddy of order 'order' */
1757 return order;
1758 }
1759 order++;
1760 }
1761 return 0;
1762 }
1763
1764 static void mb_clear_bits(void *bm, int cur, int len)
1765 {
1766 __u32 *addr;
1767
1768 len = cur + len;
1769 while (cur < len) {
1770 if ((cur & 31) == 0 && (len - cur) >= 32) {
1771 /* fast path: clear whole word at once */
1772 addr = bm + (cur >> 3);
1773 *addr = 0;
1774 cur += 32;
1775 continue;
1776 }
1777 mb_clear_bit(cur, bm);
1778 cur++;
1779 }
1780 }
1781
1782 /* clear bits in given range
1783 * will return first found zero bit if any, -1 otherwise
1784 */
1785 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1786 {
1787 __u32 *addr;
1788 int zero_bit = -1;
1789
1790 len = cur + len;
1791 while (cur < len) {
1792 if ((cur & 31) == 0 && (len - cur) >= 32) {
1793 /* fast path: clear whole word at once */
1794 addr = bm + (cur >> 3);
1795 if (*addr != (__u32)(-1) && zero_bit == -1)
1796 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1797 *addr = 0;
1798 cur += 32;
1799 continue;
1800 }
1801 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1802 zero_bit = cur;
1803 cur++;
1804 }
1805
1806 return zero_bit;
1807 }
1808
1809 void mb_set_bits(void *bm, int cur, int len)
1810 {
1811 __u32 *addr;
1812
1813 len = cur + len;
1814 while (cur < len) {
1815 if ((cur & 31) == 0 && (len - cur) >= 32) {
1816 /* fast path: set whole word at once */
1817 addr = bm + (cur >> 3);
1818 *addr = 0xffffffff;
1819 cur += 32;
1820 continue;
1821 }
1822 mb_set_bit(cur, bm);
1823 cur++;
1824 }
1825 }
1826
1827 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1828 {
1829 if (mb_test_bit(*bit + side, bitmap)) {
1830 mb_clear_bit(*bit, bitmap);
1831 (*bit) -= side;
1832 return 1;
1833 }
1834 else {
1835 (*bit) += side;
1836 mb_set_bit(*bit, bitmap);
1837 return -1;
1838 }
1839 }
1840
1841 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1842 {
1843 int max;
1844 int order = 1;
1845 void *buddy = mb_find_buddy(e4b, order, &max);
1846
1847 while (buddy) {
1848 void *buddy2;
1849
1850 /* Bits in range [first; last] are known to be set since
1851 * corresponding blocks were allocated. Bits in range
1852 * (first; last) will stay set because they form buddies on
1853 * upper layer. We just deal with borders if they don't
1854 * align with upper layer and then go up.
1855 * Releasing entire group is all about clearing
1856 * single bit of highest order buddy.
1857 */
1858
1859 /* Example:
1860 * ---------------------------------
1861 * | 1 | 1 | 1 | 1 |
1862 * ---------------------------------
1863 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1864 * ---------------------------------
1865 * 0 1 2 3 4 5 6 7
1866 * \_____________________/
1867 *
1868 * Neither [1] nor [6] is aligned to above layer.
1869 * Left neighbour [0] is free, so mark it busy,
1870 * decrease bb_counters and extend range to
1871 * [0; 6]
1872 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1873 * mark [6] free, increase bb_counters and shrink range to
1874 * [0; 5].
1875 * Then shift range to [0; 2], go up and do the same.
1876 */
1877
1878
1879 if (first & 1)
1880 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1881 if (!(last & 1))
1882 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1883 if (first > last)
1884 break;
1885 order++;
1886
1887 buddy2 = mb_find_buddy(e4b, order, &max);
1888 if (!buddy2) {
1889 mb_clear_bits(buddy, first, last - first + 1);
1890 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1891 break;
1892 }
1893 first >>= 1;
1894 last >>= 1;
1895 buddy = buddy2;
1896 }
1897 }
1898
1899 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1900 int first, int count)
1901 {
1902 int left_is_free = 0;
1903 int right_is_free = 0;
1904 int block;
1905 int last = first + count - 1;
1906 struct super_block *sb = e4b->bd_sb;
1907
1908 if (WARN_ON(count == 0))
1909 return;
1910 BUG_ON(last >= (sb->s_blocksize << 3));
1911 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1912 /* Don't bother if the block group is corrupt. */
1913 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1914 return;
1915
1916 mb_check_buddy(e4b);
1917 mb_free_blocks_double(inode, e4b, first, count);
1918
1919 /* access memory sequentially: check left neighbour,
1920 * clear range and then check right neighbour
1921 */
1922 if (first != 0)
1923 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1924 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1925 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1926 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1927
1928 if (unlikely(block != -1)) {
1929 struct ext4_sb_info *sbi = EXT4_SB(sb);
1930 ext4_fsblk_t blocknr;
1931
1932 /*
1933 * Fastcommit replay can free already freed blocks which
1934 * corrupts allocation info. Regenerate it.
1935 */
1936 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
1937 mb_regenerate_buddy(e4b);
1938 goto check;
1939 }
1940
1941 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1942 blocknr += EXT4_C2B(sbi, block);
1943 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1944 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1945 ext4_grp_locked_error(sb, e4b->bd_group,
1946 inode ? inode->i_ino : 0, blocknr,
1947 "freeing already freed block (bit %u); block bitmap corrupt.",
1948 block);
1949 return;
1950 }
1951
1952 this_cpu_inc(discard_pa_seq);
1953 e4b->bd_info->bb_free += count;
1954 if (first < e4b->bd_info->bb_first_free)
1955 e4b->bd_info->bb_first_free = first;
1956
1957 /* let's maintain fragments counter */
1958 if (left_is_free && right_is_free)
1959 e4b->bd_info->bb_fragments--;
1960 else if (!left_is_free && !right_is_free)
1961 e4b->bd_info->bb_fragments++;
1962
1963 /* buddy[0] == bd_bitmap is a special case, so handle
1964 * it right away and let mb_buddy_mark_free stay free of
1965 * zero order checks.
1966 * Check if neighbours are to be coaleasced,
1967 * adjust bitmap bb_counters and borders appropriately.
1968 */
1969 if (first & 1) {
1970 first += !left_is_free;
1971 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1972 }
1973 if (!(last & 1)) {
1974 last -= !right_is_free;
1975 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1976 }
1977
1978 if (first <= last)
1979 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1980
1981 mb_set_largest_free_order(sb, e4b->bd_info);
1982 mb_update_avg_fragment_size(sb, e4b->bd_info);
1983 check:
1984 mb_check_buddy(e4b);
1985 }
1986
1987 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1988 int needed, struct ext4_free_extent *ex)
1989 {
1990 int max, order, next;
1991 void *buddy;
1992
1993 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1994 BUG_ON(ex == NULL);
1995
1996 buddy = mb_find_buddy(e4b, 0, &max);
1997 BUG_ON(buddy == NULL);
1998 BUG_ON(block >= max);
1999 if (mb_test_bit(block, buddy)) {
2000 ex->fe_len = 0;
2001 ex->fe_start = 0;
2002 ex->fe_group = 0;
2003 return 0;
2004 }
2005
2006 /* find actual order */
2007 order = mb_find_order_for_block(e4b, block);
2008
2009 ex->fe_len = (1 << order) - (block & ((1 << order) - 1));
2010 ex->fe_start = block;
2011 ex->fe_group = e4b->bd_group;
2012
2013 block = block >> order;
2014
2015 while (needed > ex->fe_len &&
2016 mb_find_buddy(e4b, order, &max)) {
2017
2018 if (block + 1 >= max)
2019 break;
2020
2021 next = (block + 1) * (1 << order);
2022 if (mb_test_bit(next, e4b->bd_bitmap))
2023 break;
2024
2025 order = mb_find_order_for_block(e4b, next);
2026
2027 block = next >> order;
2028 ex->fe_len += 1 << order;
2029 }
2030
2031 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
2032 /* Should never happen! (but apparently sometimes does?!?) */
2033 WARN_ON(1);
2034 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
2035 "corruption or bug in mb_find_extent "
2036 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
2037 block, order, needed, ex->fe_group, ex->fe_start,
2038 ex->fe_len, ex->fe_logical);
2039 ex->fe_len = 0;
2040 ex->fe_start = 0;
2041 ex->fe_group = 0;
2042 }
2043 return ex->fe_len;
2044 }
2045
2046 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
2047 {
2048 int ord;
2049 int mlen = 0;
2050 int max = 0;
2051 int start = ex->fe_start;
2052 int len = ex->fe_len;
2053 unsigned ret = 0;
2054 int len0 = len;
2055 void *buddy;
2056 int ord_start, ord_end;
2057
2058 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
2059 BUG_ON(e4b->bd_group != ex->fe_group);
2060 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
2061 mb_check_buddy(e4b);
2062 mb_mark_used_double(e4b, start, len);
2063
2064 this_cpu_inc(discard_pa_seq);
2065 e4b->bd_info->bb_free -= len;
2066 if (e4b->bd_info->bb_first_free == start)
2067 e4b->bd_info->bb_first_free += len;
2068
2069 /* let's maintain fragments counter */
2070 if (start != 0)
2071 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
2072 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
2073 max = !mb_test_bit(start + len, e4b->bd_bitmap);
2074 if (mlen && max)
2075 e4b->bd_info->bb_fragments++;
2076 else if (!mlen && !max)
2077 e4b->bd_info->bb_fragments--;
2078
2079 /* let's maintain buddy itself */
2080 while (len) {
2081 ord = mb_find_order_for_block(e4b, start);
2082
2083 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
2084 /* the whole chunk may be allocated at once! */
2085 mlen = 1 << ord;
2086 buddy = mb_find_buddy(e4b, ord, &max);
2087 BUG_ON((start >> ord) >= max);
2088 mb_set_bit(start >> ord, buddy);
2089 e4b->bd_info->bb_counters[ord]--;
2090 start += mlen;
2091 len -= mlen;
2092 BUG_ON(len < 0);
2093 continue;
2094 }
2095
2096 /* store for history */
2097 if (ret == 0)
2098 ret = len | (ord << 16);
2099
2100 BUG_ON(ord <= 0);
2101 buddy = mb_find_buddy(e4b, ord, &max);
2102 mb_set_bit(start >> ord, buddy);
2103 e4b->bd_info->bb_counters[ord]--;
2104
2105 ord_start = (start >> ord) << ord;
2106 ord_end = ord_start + (1 << ord);
2107 /* first chunk */
2108 if (start > ord_start)
2109 ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy,
2110 ord_start, start - ord_start,
2111 e4b->bd_info);
2112
2113 /* last chunk */
2114 if (start + len < ord_end) {
2115 ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy,
2116 start + len,
2117 ord_end - (start + len),
2118 e4b->bd_info);
2119 break;
2120 }
2121 len = start + len - ord_end;
2122 start = ord_end;
2123 }
2124 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
2125
2126 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2127 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2128 mb_check_buddy(e4b);
2129
2130 return ret;
2131 }
2132
2133 /*
2134 * Must be called under group lock!
2135 */
2136 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2137 struct ext4_buddy *e4b)
2138 {
2139 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2140 int ret;
2141
2142 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2143 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2144
2145 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2146 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2147 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2148
2149 /* preallocation can change ac_b_ex, thus we store actually
2150 * allocated blocks for history */
2151 ac->ac_f_ex = ac->ac_b_ex;
2152
2153 ac->ac_status = AC_STATUS_FOUND;
2154 ac->ac_tail = ret & 0xffff;
2155 ac->ac_buddy = ret >> 16;
2156
2157 /*
2158 * take the page reference. We want the page to be pinned
2159 * so that we don't get a ext4_mb_init_cache_call for this
2160 * group until we update the bitmap. That would mean we
2161 * double allocate blocks. The reference is dropped
2162 * in ext4_mb_release_context
2163 */
2164 ac->ac_bitmap_folio = e4b->bd_bitmap_folio;
2165 folio_get(ac->ac_bitmap_folio);
2166 ac->ac_buddy_folio = e4b->bd_buddy_folio;
2167 folio_get(ac->ac_buddy_folio);
2168 /* store last allocated for subsequent stream allocation */
2169 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2170 spin_lock(&sbi->s_md_lock);
2171 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2172 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2173 spin_unlock(&sbi->s_md_lock);
2174 }
2175 /*
2176 * As we've just preallocated more space than
2177 * user requested originally, we store allocated
2178 * space in a special descriptor.
2179 */
2180 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2181 ext4_mb_new_preallocation(ac);
2182
2183 }
2184
2185 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2186 struct ext4_buddy *e4b,
2187 int finish_group)
2188 {
2189 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2190 struct ext4_free_extent *bex = &ac->ac_b_ex;
2191 struct ext4_free_extent *gex = &ac->ac_g_ex;
2192
2193 if (ac->ac_status == AC_STATUS_FOUND)
2194 return;
2195 /*
2196 * We don't want to scan for a whole year
2197 */
2198 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2199 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2200 ac->ac_status = AC_STATUS_BREAK;
2201 return;
2202 }
2203
2204 /*
2205 * Haven't found good chunk so far, let's continue
2206 */
2207 if (bex->fe_len < gex->fe_len)
2208 return;
2209
2210 if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2211 ext4_mb_use_best_found(ac, e4b);
2212 }
2213
2214 /*
2215 * The routine checks whether found extent is good enough. If it is,
2216 * then the extent gets marked used and flag is set to the context
2217 * to stop scanning. Otherwise, the extent is compared with the
2218 * previous found extent and if new one is better, then it's stored
2219 * in the context. Later, the best found extent will be used, if
2220 * mballoc can't find good enough extent.
2221 *
2222 * The algorithm used is roughly as follows:
2223 *
2224 * * If free extent found is exactly as big as goal, then
2225 * stop the scan and use it immediately
2226 *
2227 * * If free extent found is smaller than goal, then keep retrying
2228 * upto a max of sbi->s_mb_max_to_scan times (default 200). After
2229 * that stop scanning and use whatever we have.
2230 *
2231 * * If free extent found is bigger than goal, then keep retrying
2232 * upto a max of sbi->s_mb_min_to_scan times (default 10) before
2233 * stopping the scan and using the extent.
2234 *
2235 *
2236 * FIXME: real allocation policy is to be designed yet!
2237 */
2238 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2239 struct ext4_free_extent *ex,
2240 struct ext4_buddy *e4b)
2241 {
2242 struct ext4_free_extent *bex = &ac->ac_b_ex;
2243 struct ext4_free_extent *gex = &ac->ac_g_ex;
2244
2245 BUG_ON(ex->fe_len <= 0);
2246 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2247 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2248 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2249
2250 ac->ac_found++;
2251 ac->ac_cX_found[ac->ac_criteria]++;
2252
2253 /*
2254 * The special case - take what you catch first
2255 */
2256 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2257 *bex = *ex;
2258 ext4_mb_use_best_found(ac, e4b);
2259 return;
2260 }
2261
2262 /*
2263 * Let's check whether the chuck is good enough
2264 */
2265 if (ex->fe_len == gex->fe_len) {
2266 *bex = *ex;
2267 ext4_mb_use_best_found(ac, e4b);
2268 return;
2269 }
2270
2271 /*
2272 * If this is first found extent, just store it in the context
2273 */
2274 if (bex->fe_len == 0) {
2275 *bex = *ex;
2276 return;
2277 }
2278
2279 /*
2280 * If new found extent is better, store it in the context
2281 */
2282 if (bex->fe_len < gex->fe_len) {
2283 /* if the request isn't satisfied, any found extent
2284 * larger than previous best one is better */
2285 if (ex->fe_len > bex->fe_len)
2286 *bex = *ex;
2287 } else if (ex->fe_len > gex->fe_len) {
2288 /* if the request is satisfied, then we try to find
2289 * an extent that still satisfy the request, but is
2290 * smaller than previous one */
2291 if (ex->fe_len < bex->fe_len)
2292 *bex = *ex;
2293 }
2294
2295 ext4_mb_check_limits(ac, e4b, 0);
2296 }
2297
2298 static noinline_for_stack
2299 void ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2300 struct ext4_buddy *e4b)
2301 {
2302 struct ext4_free_extent ex = ac->ac_b_ex;
2303 ext4_group_t group = ex.fe_group;
2304 int max;
2305 int err;
2306
2307 BUG_ON(ex.fe_len <= 0);
2308 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2309 if (err)
2310 return;
2311
2312 ext4_lock_group(ac->ac_sb, group);
2313 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2314 goto out;
2315
2316 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2317
2318 if (max > 0) {
2319 ac->ac_b_ex = ex;
2320 ext4_mb_use_best_found(ac, e4b);
2321 }
2322
2323 out:
2324 ext4_unlock_group(ac->ac_sb, group);
2325 ext4_mb_unload_buddy(e4b);
2326 }
2327
2328 static noinline_for_stack
2329 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2330 struct ext4_buddy *e4b)
2331 {
2332 ext4_group_t group = ac->ac_g_ex.fe_group;
2333 int max;
2334 int err;
2335 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2336 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2337 struct ext4_free_extent ex;
2338
2339 if (!grp)
2340 return -EFSCORRUPTED;
2341 if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY)))
2342 return 0;
2343 if (grp->bb_free == 0)
2344 return 0;
2345
2346 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2347 if (err)
2348 return err;
2349
2350 ext4_lock_group(ac->ac_sb, group);
2351 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2352 goto out;
2353
2354 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2355 ac->ac_g_ex.fe_len, &ex);
2356 ex.fe_logical = 0xDEADFA11; /* debug value */
2357
2358 if (max >= ac->ac_g_ex.fe_len &&
2359 ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) {
2360 ext4_fsblk_t start;
2361
2362 start = ext4_grp_offs_to_block(ac->ac_sb, &ex);
2363 /* use do_div to get remainder (would be 64-bit modulo) */
2364 if (do_div(start, sbi->s_stripe) == 0) {
2365 ac->ac_found++;
2366 ac->ac_b_ex = ex;
2367 ext4_mb_use_best_found(ac, e4b);
2368 }
2369 } else if (max >= ac->ac_g_ex.fe_len) {
2370 BUG_ON(ex.fe_len <= 0);
2371 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2372 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2373 ac->ac_found++;
2374 ac->ac_b_ex = ex;
2375 ext4_mb_use_best_found(ac, e4b);
2376 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2377 /* Sometimes, caller may want to merge even small
2378 * number of blocks to an existing extent */
2379 BUG_ON(ex.fe_len <= 0);
2380 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2381 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2382 ac->ac_found++;
2383 ac->ac_b_ex = ex;
2384 ext4_mb_use_best_found(ac, e4b);
2385 }
2386 out:
2387 ext4_unlock_group(ac->ac_sb, group);
2388 ext4_mb_unload_buddy(e4b);
2389
2390 return 0;
2391 }
2392
2393 /*
2394 * The routine scans buddy structures (not bitmap!) from given order
2395 * to max order and tries to find big enough chunk to satisfy the req
2396 */
2397 static noinline_for_stack
2398 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2399 struct ext4_buddy *e4b)
2400 {
2401 struct super_block *sb = ac->ac_sb;
2402 struct ext4_group_info *grp = e4b->bd_info;
2403 void *buddy;
2404 int i;
2405 int k;
2406 int max;
2407
2408 BUG_ON(ac->ac_2order <= 0);
2409 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2410 if (grp->bb_counters[i] == 0)
2411 continue;
2412
2413 buddy = mb_find_buddy(e4b, i, &max);
2414 if (WARN_RATELIMIT(buddy == NULL,
2415 "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i))
2416 continue;
2417
2418 k = mb_find_next_zero_bit(buddy, max, 0);
2419 if (k >= max) {
2420 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2421 e4b->bd_group,
2422 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2423 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2424 "%d free clusters of order %d. But found 0",
2425 grp->bb_counters[i], i);
2426 break;
2427 }
2428 ac->ac_found++;
2429 ac->ac_cX_found[ac->ac_criteria]++;
2430
2431 ac->ac_b_ex.fe_len = 1 << i;
2432 ac->ac_b_ex.fe_start = k << i;
2433 ac->ac_b_ex.fe_group = e4b->bd_group;
2434
2435 ext4_mb_use_best_found(ac, e4b);
2436
2437 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2438
2439 if (EXT4_SB(sb)->s_mb_stats)
2440 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2441
2442 break;
2443 }
2444 }
2445
2446 /*
2447 * The routine scans the group and measures all found extents.
2448 * In order to optimize scanning, caller must pass number of
2449 * free blocks in the group, so the routine can know upper limit.
2450 */
2451 static noinline_for_stack
2452 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2453 struct ext4_buddy *e4b)
2454 {
2455 struct super_block *sb = ac->ac_sb;
2456 void *bitmap = e4b->bd_bitmap;
2457 struct ext4_free_extent ex;
2458 int i, j, freelen;
2459 int free;
2460
2461 free = e4b->bd_info->bb_free;
2462 if (WARN_ON(free <= 0))
2463 return;
2464
2465 i = e4b->bd_info->bb_first_free;
2466
2467 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2468 i = mb_find_next_zero_bit(bitmap,
2469 EXT4_CLUSTERS_PER_GROUP(sb), i);
2470 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2471 /*
2472 * IF we have corrupt bitmap, we won't find any
2473 * free blocks even though group info says we
2474 * have free blocks
2475 */
2476 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2477 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2478 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2479 "%d free clusters as per "
2480 "group info. But bitmap says 0",
2481 free);
2482 break;
2483 }
2484
2485 if (!ext4_mb_cr_expensive(ac->ac_criteria)) {
2486 /*
2487 * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are
2488 * sure that this group will have a large enough
2489 * continuous free extent, so skip over the smaller free
2490 * extents
2491 */
2492 j = mb_find_next_bit(bitmap,
2493 EXT4_CLUSTERS_PER_GROUP(sb), i);
2494 freelen = j - i;
2495
2496 if (freelen < ac->ac_g_ex.fe_len) {
2497 i = j;
2498 free -= freelen;
2499 continue;
2500 }
2501 }
2502
2503 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2504 if (WARN_ON(ex.fe_len <= 0))
2505 break;
2506 if (free < ex.fe_len) {
2507 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2508 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2509 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2510 "%d free clusters as per "
2511 "group info. But got %d blocks",
2512 free, ex.fe_len);
2513 /*
2514 * The number of free blocks differs. This mostly
2515 * indicate that the bitmap is corrupt. So exit
2516 * without claiming the space.
2517 */
2518 break;
2519 }
2520 ex.fe_logical = 0xDEADC0DE; /* debug value */
2521 ext4_mb_measure_extent(ac, &ex, e4b);
2522
2523 i += ex.fe_len;
2524 free -= ex.fe_len;
2525 }
2526
2527 ext4_mb_check_limits(ac, e4b, 1);
2528 }
2529
2530 /*
2531 * This is a special case for storages like raid5
2532 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2533 */
2534 static noinline_for_stack
2535 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2536 struct ext4_buddy *e4b)
2537 {
2538 struct super_block *sb = ac->ac_sb;
2539 struct ext4_sb_info *sbi = EXT4_SB(sb);
2540 void *bitmap = e4b->bd_bitmap;
2541 struct ext4_free_extent ex;
2542 ext4_fsblk_t first_group_block;
2543 ext4_fsblk_t a;
2544 ext4_grpblk_t i, stripe;
2545 int max;
2546
2547 BUG_ON(sbi->s_stripe == 0);
2548
2549 /* find first stripe-aligned block in group */
2550 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2551
2552 a = first_group_block + sbi->s_stripe - 1;
2553 do_div(a, sbi->s_stripe);
2554 i = (a * sbi->s_stripe) - first_group_block;
2555
2556 stripe = EXT4_B2C(sbi, sbi->s_stripe);
2557 i = EXT4_B2C(sbi, i);
2558 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2559 if (!mb_test_bit(i, bitmap)) {
2560 max = mb_find_extent(e4b, i, stripe, &ex);
2561 if (max >= stripe) {
2562 ac->ac_found++;
2563 ac->ac_cX_found[ac->ac_criteria]++;
2564 ex.fe_logical = 0xDEADF00D; /* debug value */
2565 ac->ac_b_ex = ex;
2566 ext4_mb_use_best_found(ac, e4b);
2567 break;
2568 }
2569 }
2570 i += stripe;
2571 }
2572 }
2573
2574 /*
2575 * This is also called BEFORE we load the buddy bitmap.
2576 * Returns either 1 or 0 indicating that the group is either suitable
2577 * for the allocation or not.
2578 */
2579 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2580 ext4_group_t group, enum criteria cr)
2581 {
2582 ext4_grpblk_t free, fragments;
2583 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2584 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2585
2586 BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS);
2587
2588 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2589 return false;
2590
2591 free = grp->bb_free;
2592 if (free == 0)
2593 return false;
2594
2595 fragments = grp->bb_fragments;
2596 if (fragments == 0)
2597 return false;
2598
2599 switch (cr) {
2600 case CR_POWER2_ALIGNED:
2601 BUG_ON(ac->ac_2order == 0);
2602
2603 /* Avoid using the first bg of a flexgroup for data files */
2604 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2605 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2606 ((group % flex_size) == 0))
2607 return false;
2608
2609 if (free < ac->ac_g_ex.fe_len)
2610 return false;
2611
2612 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2613 return true;
2614
2615 if (grp->bb_largest_free_order < ac->ac_2order)
2616 return false;
2617
2618 return true;
2619 case CR_GOAL_LEN_FAST:
2620 case CR_BEST_AVAIL_LEN:
2621 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2622 return true;
2623 break;
2624 case CR_GOAL_LEN_SLOW:
2625 if (free >= ac->ac_g_ex.fe_len)
2626 return true;
2627 break;
2628 case CR_ANY_FREE:
2629 return true;
2630 default:
2631 BUG();
2632 }
2633
2634 return false;
2635 }
2636
2637 /*
2638 * This could return negative error code if something goes wrong
2639 * during ext4_mb_init_group(). This should not be called with
2640 * ext4_lock_group() held.
2641 *
2642 * Note: because we are conditionally operating with the group lock in
2643 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2644 * function using __acquire and __release. This means we need to be
2645 * super careful before messing with the error path handling via "goto
2646 * out"!
2647 */
2648 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2649 ext4_group_t group, enum criteria cr)
2650 {
2651 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2652 struct super_block *sb = ac->ac_sb;
2653 struct ext4_sb_info *sbi = EXT4_SB(sb);
2654 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2655 ext4_grpblk_t free;
2656 int ret = 0;
2657
2658 if (!grp)
2659 return -EFSCORRUPTED;
2660 if (sbi->s_mb_stats)
2661 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2662 if (should_lock) {
2663 ext4_lock_group(sb, group);
2664 __release(ext4_group_lock_ptr(sb, group));
2665 }
2666 free = grp->bb_free;
2667 if (free == 0)
2668 goto out;
2669 /*
2670 * In all criterias except CR_ANY_FREE we try to avoid groups that
2671 * can't possibly satisfy the full goal request due to insufficient
2672 * free blocks.
2673 */
2674 if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len)
2675 goto out;
2676 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2677 goto out;
2678 if (should_lock) {
2679 __acquire(ext4_group_lock_ptr(sb, group));
2680 ext4_unlock_group(sb, group);
2681 }
2682
2683 /* We only do this if the grp has never been initialized */
2684 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2685 struct ext4_group_desc *gdp =
2686 ext4_get_group_desc(sb, group, NULL);
2687 int ret;
2688
2689 /*
2690 * CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic
2691 * search to find large good chunks almost for free. If buddy
2692 * data is not ready, then this optimization makes no sense. But
2693 * we never skip the first block group in a flex_bg, since this
2694 * gets used for metadata block allocation, and we want to make
2695 * sure we locate metadata blocks in the first block group in
2696 * the flex_bg if possible.
2697 */
2698 if (!ext4_mb_cr_expensive(cr) &&
2699 (!sbi->s_log_groups_per_flex ||
2700 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2701 !(ext4_has_group_desc_csum(sb) &&
2702 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2703 return 0;
2704 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2705 if (ret)
2706 return ret;
2707 }
2708
2709 if (should_lock) {
2710 ext4_lock_group(sb, group);
2711 __release(ext4_group_lock_ptr(sb, group));
2712 }
2713 ret = ext4_mb_good_group(ac, group, cr);
2714 out:
2715 if (should_lock) {
2716 __acquire(ext4_group_lock_ptr(sb, group));
2717 ext4_unlock_group(sb, group);
2718 }
2719 return ret;
2720 }
2721
2722 /*
2723 * Start prefetching @nr block bitmaps starting at @group.
2724 * Return the next group which needs to be prefetched.
2725 */
2726 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2727 unsigned int nr, int *cnt)
2728 {
2729 ext4_group_t ngroups = ext4_get_groups_count(sb);
2730 struct buffer_head *bh;
2731 struct blk_plug plug;
2732
2733 blk_start_plug(&plug);
2734 while (nr-- > 0) {
2735 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2736 NULL);
2737 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2738
2739 /*
2740 * Prefetch block groups with free blocks; but don't
2741 * bother if it is marked uninitialized on disk, since
2742 * it won't require I/O to read. Also only try to
2743 * prefetch once, so we avoid getblk() call, which can
2744 * be expensive.
2745 */
2746 if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2747 EXT4_MB_GRP_NEED_INIT(grp) &&
2748 ext4_free_group_clusters(sb, gdp) > 0 ) {
2749 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2750 if (bh && !IS_ERR(bh)) {
2751 if (!buffer_uptodate(bh) && cnt)
2752 (*cnt)++;
2753 brelse(bh);
2754 }
2755 }
2756 if (++group >= ngroups)
2757 group = 0;
2758 }
2759 blk_finish_plug(&plug);
2760 return group;
2761 }
2762
2763 /*
2764 * Prefetching reads the block bitmap into the buffer cache; but we
2765 * need to make sure that the buddy bitmap in the page cache has been
2766 * initialized. Note that ext4_mb_init_group() will block if the I/O
2767 * is not yet completed, or indeed if it was not initiated by
2768 * ext4_mb_prefetch did not start the I/O.
2769 *
2770 * TODO: We should actually kick off the buddy bitmap setup in a work
2771 * queue when the buffer I/O is completed, so that we don't block
2772 * waiting for the block allocation bitmap read to finish when
2773 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2774 */
2775 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2776 unsigned int nr)
2777 {
2778 struct ext4_group_desc *gdp;
2779 struct ext4_group_info *grp;
2780
2781 while (nr-- > 0) {
2782 if (!group)
2783 group = ext4_get_groups_count(sb);
2784 group--;
2785 gdp = ext4_get_group_desc(sb, group, NULL);
2786 grp = ext4_get_group_info(sb, group);
2787
2788 if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) &&
2789 ext4_free_group_clusters(sb, gdp) > 0) {
2790 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2791 break;
2792 }
2793 }
2794 }
2795
2796 static noinline_for_stack int
2797 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2798 {
2799 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2800 enum criteria new_cr, cr = CR_GOAL_LEN_FAST;
2801 int err = 0, first_err = 0;
2802 unsigned int nr = 0, prefetch_ios = 0;
2803 struct ext4_sb_info *sbi;
2804 struct super_block *sb;
2805 struct ext4_buddy e4b;
2806 int lost;
2807
2808 sb = ac->ac_sb;
2809 sbi = EXT4_SB(sb);
2810 ngroups = ext4_get_groups_count(sb);
2811 /* non-extent files are limited to low blocks/groups */
2812 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2813 ngroups = sbi->s_blockfile_groups;
2814
2815 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2816
2817 /* first, try the goal */
2818 err = ext4_mb_find_by_goal(ac, &e4b);
2819 if (err || ac->ac_status == AC_STATUS_FOUND)
2820 goto out;
2821
2822 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2823 goto out;
2824
2825 /*
2826 * ac->ac_2order is set only if the fe_len is a power of 2
2827 * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED
2828 * so that we try exact allocation using buddy.
2829 */
2830 i = fls(ac->ac_g_ex.fe_len);
2831 ac->ac_2order = 0;
2832 /*
2833 * We search using buddy data only if the order of the request
2834 * is greater than equal to the sbi_s_mb_order2_reqs
2835 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2836 * We also support searching for power-of-two requests only for
2837 * requests upto maximum buddy size we have constructed.
2838 */
2839 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2840 if (is_power_of_2(ac->ac_g_ex.fe_len))
2841 ac->ac_2order = array_index_nospec(i - 1,
2842 MB_NUM_ORDERS(sb));
2843 }
2844
2845 /* if stream allocation is enabled, use global goal */
2846 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2847 /* TBD: may be hot point */
2848 spin_lock(&sbi->s_md_lock);
2849 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2850 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2851 spin_unlock(&sbi->s_md_lock);
2852 }
2853
2854 /*
2855 * Let's just scan groups to find more-less suitable blocks We
2856 * start with CR_GOAL_LEN_FAST, unless it is power of 2
2857 * aligned, in which case let's do that faster approach first.
2858 */
2859 if (ac->ac_2order)
2860 cr = CR_POWER2_ALIGNED;
2861 repeat:
2862 for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2863 ac->ac_criteria = cr;
2864 /*
2865 * searching for the right group start
2866 * from the goal value specified
2867 */
2868 group = ac->ac_g_ex.fe_group;
2869 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2870 prefetch_grp = group;
2871 nr = 0;
2872
2873 for (i = 0, new_cr = cr; i < ngroups; i++,
2874 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2875 int ret = 0;
2876
2877 cond_resched();
2878 if (new_cr != cr) {
2879 cr = new_cr;
2880 goto repeat;
2881 }
2882
2883 /*
2884 * Batch reads of the block allocation bitmaps
2885 * to get multiple READs in flight; limit
2886 * prefetching at inexpensive CR, otherwise mballoc
2887 * can spend a lot of time loading imperfect groups
2888 */
2889 if ((prefetch_grp == group) &&
2890 (ext4_mb_cr_expensive(cr) ||
2891 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2892 nr = sbi->s_mb_prefetch;
2893 if (ext4_has_feature_flex_bg(sb)) {
2894 nr = 1 << sbi->s_log_groups_per_flex;
2895 nr -= group & (nr - 1);
2896 nr = min(nr, sbi->s_mb_prefetch);
2897 }
2898 prefetch_grp = ext4_mb_prefetch(sb, group,
2899 nr, &prefetch_ios);
2900 }
2901
2902 /* This now checks without needing the buddy page */
2903 ret = ext4_mb_good_group_nolock(ac, group, cr);
2904 if (ret <= 0) {
2905 if (!first_err)
2906 first_err = ret;
2907 continue;
2908 }
2909
2910 err = ext4_mb_load_buddy(sb, group, &e4b);
2911 if (err)
2912 goto out;
2913
2914 ext4_lock_group(sb, group);
2915
2916 /*
2917 * We need to check again after locking the
2918 * block group
2919 */
2920 ret = ext4_mb_good_group(ac, group, cr);
2921 if (ret == 0) {
2922 ext4_unlock_group(sb, group);
2923 ext4_mb_unload_buddy(&e4b);
2924 continue;
2925 }
2926
2927 ac->ac_groups_scanned++;
2928 if (cr == CR_POWER2_ALIGNED)
2929 ext4_mb_simple_scan_group(ac, &e4b);
2930 else {
2931 bool is_stripe_aligned = sbi->s_stripe &&
2932 !(ac->ac_g_ex.fe_len %
2933 EXT4_B2C(sbi, sbi->s_stripe));
2934
2935 if ((cr == CR_GOAL_LEN_FAST ||
2936 cr == CR_BEST_AVAIL_LEN) &&
2937 is_stripe_aligned)
2938 ext4_mb_scan_aligned(ac, &e4b);
2939
2940 if (ac->ac_status == AC_STATUS_CONTINUE)
2941 ext4_mb_complex_scan_group(ac, &e4b);
2942 }
2943
2944 ext4_unlock_group(sb, group);
2945 ext4_mb_unload_buddy(&e4b);
2946
2947 if (ac->ac_status != AC_STATUS_CONTINUE)
2948 break;
2949 }
2950 /* Processed all groups and haven't found blocks */
2951 if (sbi->s_mb_stats && i == ngroups)
2952 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2953
2954 if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN)
2955 /* Reset goal length to original goal length before
2956 * falling into CR_GOAL_LEN_SLOW */
2957 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
2958 }
2959
2960 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2961 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2962 /*
2963 * We've been searching too long. Let's try to allocate
2964 * the best chunk we've found so far
2965 */
2966 ext4_mb_try_best_found(ac, &e4b);
2967 if (ac->ac_status != AC_STATUS_FOUND) {
2968 /*
2969 * Someone more lucky has already allocated it.
2970 * The only thing we can do is just take first
2971 * found block(s)
2972 */
2973 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2974 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2975 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2976 ac->ac_b_ex.fe_len, lost);
2977
2978 ac->ac_b_ex.fe_group = 0;
2979 ac->ac_b_ex.fe_start = 0;
2980 ac->ac_b_ex.fe_len = 0;
2981 ac->ac_status = AC_STATUS_CONTINUE;
2982 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2983 cr = CR_ANY_FREE;
2984 goto repeat;
2985 }
2986 }
2987
2988 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2989 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2990 out:
2991 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2992 err = first_err;
2993
2994 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2995 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2996 ac->ac_flags, cr, err);
2997
2998 if (nr)
2999 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
3000
3001 return err;
3002 }
3003
3004 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
3005 {
3006 struct super_block *sb = pde_data(file_inode(seq->file));
3007 ext4_group_t group;
3008
3009 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3010 return NULL;
3011 group = *pos + 1;
3012 return (void *) ((unsigned long) group);
3013 }
3014
3015 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
3016 {
3017 struct super_block *sb = pde_data(file_inode(seq->file));
3018 ext4_group_t group;
3019
3020 ++*pos;
3021 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3022 return NULL;
3023 group = *pos + 1;
3024 return (void *) ((unsigned long) group);
3025 }
3026
3027 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
3028 {
3029 struct super_block *sb = pde_data(file_inode(seq->file));
3030 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
3031 int i, err;
3032 char nbuf[16];
3033 struct ext4_buddy e4b;
3034 struct ext4_group_info *grinfo;
3035 unsigned char blocksize_bits = min_t(unsigned char,
3036 sb->s_blocksize_bits,
3037 EXT4_MAX_BLOCK_LOG_SIZE);
3038 struct sg {
3039 struct ext4_group_info info;
3040 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
3041 } sg;
3042
3043 group--;
3044 if (group == 0)
3045 seq_puts(seq, "#group: free frags first ["
3046 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
3047 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
3048
3049 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
3050 sizeof(struct ext4_group_info);
3051
3052 grinfo = ext4_get_group_info(sb, group);
3053 if (!grinfo)
3054 return 0;
3055 /* Load the group info in memory only if not already loaded. */
3056 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
3057 err = ext4_mb_load_buddy(sb, group, &e4b);
3058 if (err) {
3059 seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf));
3060 return 0;
3061 }
3062 ext4_mb_unload_buddy(&e4b);
3063 }
3064
3065 /*
3066 * We care only about free space counters in the group info and
3067 * these are safe to access even after the buddy has been unloaded
3068 */
3069 memcpy(&sg, grinfo, i);
3070 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
3071 sg.info.bb_fragments, sg.info.bb_first_free);
3072 for (i = 0; i <= 13; i++)
3073 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
3074 sg.info.bb_counters[i] : 0);
3075 seq_puts(seq, " ]");
3076 if (EXT4_MB_GRP_BBITMAP_CORRUPT(&sg.info))
3077 seq_puts(seq, " Block bitmap corrupted!");
3078 seq_puts(seq, "\n");
3079
3080 return 0;
3081 }
3082
3083 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
3084 {
3085 }
3086
3087 const struct seq_operations ext4_mb_seq_groups_ops = {
3088 .start = ext4_mb_seq_groups_start,
3089 .next = ext4_mb_seq_groups_next,
3090 .stop = ext4_mb_seq_groups_stop,
3091 .show = ext4_mb_seq_groups_show,
3092 };
3093
3094 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
3095 {
3096 struct super_block *sb = seq->private;
3097 struct ext4_sb_info *sbi = EXT4_SB(sb);
3098
3099 seq_puts(seq, "mballoc:\n");
3100 if (!sbi->s_mb_stats) {
3101 seq_puts(seq, "\tmb stats collection turned off.\n");
3102 seq_puts(
3103 seq,
3104 "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
3105 return 0;
3106 }
3107 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
3108 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
3109
3110 seq_printf(seq, "\tgroups_scanned: %u\n",
3111 atomic_read(&sbi->s_bal_groups_scanned));
3112
3113 /* CR_POWER2_ALIGNED stats */
3114 seq_puts(seq, "\tcr_p2_aligned_stats:\n");
3115 seq_printf(seq, "\t\thits: %llu\n",
3116 atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED]));
3117 seq_printf(
3118 seq, "\t\tgroups_considered: %llu\n",
3119 atomic64_read(
3120 &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]));
3121 seq_printf(seq, "\t\textents_scanned: %u\n",
3122 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED]));
3123 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3124 atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED]));
3125 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3126 atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions));
3127
3128 /* CR_GOAL_LEN_FAST stats */
3129 seq_puts(seq, "\tcr_goal_fast_stats:\n");
3130 seq_printf(seq, "\t\thits: %llu\n",
3131 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST]));
3132 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3133 atomic64_read(
3134 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST]));
3135 seq_printf(seq, "\t\textents_scanned: %u\n",
3136 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST]));
3137 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3138 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST]));
3139 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3140 atomic_read(&sbi->s_bal_goal_fast_bad_suggestions));
3141
3142 /* CR_BEST_AVAIL_LEN stats */
3143 seq_puts(seq, "\tcr_best_avail_stats:\n");
3144 seq_printf(seq, "\t\thits: %llu\n",
3145 atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN]));
3146 seq_printf(
3147 seq, "\t\tgroups_considered: %llu\n",
3148 atomic64_read(
3149 &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN]));
3150 seq_printf(seq, "\t\textents_scanned: %u\n",
3151 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN]));
3152 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3153 atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN]));
3154 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3155 atomic_read(&sbi->s_bal_best_avail_bad_suggestions));
3156
3157 /* CR_GOAL_LEN_SLOW stats */
3158 seq_puts(seq, "\tcr_goal_slow_stats:\n");
3159 seq_printf(seq, "\t\thits: %llu\n",
3160 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW]));
3161 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3162 atomic64_read(
3163 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW]));
3164 seq_printf(seq, "\t\textents_scanned: %u\n",
3165 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW]));
3166 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3167 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW]));
3168
3169 /* CR_ANY_FREE stats */
3170 seq_puts(seq, "\tcr_any_free_stats:\n");
3171 seq_printf(seq, "\t\thits: %llu\n",
3172 atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE]));
3173 seq_printf(
3174 seq, "\t\tgroups_considered: %llu\n",
3175 atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE]));
3176 seq_printf(seq, "\t\textents_scanned: %u\n",
3177 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE]));
3178 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3179 atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE]));
3180
3181 /* Aggregates */
3182 seq_printf(seq, "\textents_scanned: %u\n",
3183 atomic_read(&sbi->s_bal_ex_scanned));
3184 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
3185 seq_printf(seq, "\t\tlen_goal_hits: %u\n",
3186 atomic_read(&sbi->s_bal_len_goals));
3187 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
3188 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
3189 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
3190 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
3191 atomic_read(&sbi->s_mb_buddies_generated),
3192 ext4_get_groups_count(sb));
3193 seq_printf(seq, "\tbuddies_time_used: %llu\n",
3194 atomic64_read(&sbi->s_mb_generation_time));
3195 seq_printf(seq, "\tpreallocated: %u\n",
3196 atomic_read(&sbi->s_mb_preallocated));
3197 seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded));
3198 return 0;
3199 }
3200
3201 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
3202 {
3203 struct super_block *sb = pde_data(file_inode(seq->file));
3204 unsigned long position;
3205
3206 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3207 return NULL;
3208 position = *pos + 1;
3209 return (void *) ((unsigned long) position);
3210 }
3211
3212 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3213 {
3214 struct super_block *sb = pde_data(file_inode(seq->file));
3215 unsigned long position;
3216
3217 ++*pos;
3218 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3219 return NULL;
3220 position = *pos + 1;
3221 return (void *) ((unsigned long) position);
3222 }
3223
3224 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3225 {
3226 struct super_block *sb = pde_data(file_inode(seq->file));
3227 struct ext4_sb_info *sbi = EXT4_SB(sb);
3228 unsigned long position = ((unsigned long) v);
3229 struct ext4_group_info *grp;
3230 unsigned int count;
3231
3232 position--;
3233 if (position >= MB_NUM_ORDERS(sb)) {
3234 position -= MB_NUM_ORDERS(sb);
3235 if (position == 0)
3236 seq_puts(seq, "avg_fragment_size_lists:\n");
3237
3238 count = 0;
3239 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
3240 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
3241 bb_avg_fragment_size_node)
3242 count++;
3243 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
3244 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3245 (unsigned int)position, count);
3246 return 0;
3247 }
3248
3249 if (position == 0) {
3250 seq_printf(seq, "optimize_scan: %d\n",
3251 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3252 seq_puts(seq, "max_free_order_lists:\n");
3253 }
3254 count = 0;
3255 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
3256 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3257 bb_largest_free_order_node)
3258 count++;
3259 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
3260 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3261 (unsigned int)position, count);
3262
3263 return 0;
3264 }
3265
3266 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3267 {
3268 }
3269
3270 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3271 .start = ext4_mb_seq_structs_summary_start,
3272 .next = ext4_mb_seq_structs_summary_next,
3273 .stop = ext4_mb_seq_structs_summary_stop,
3274 .show = ext4_mb_seq_structs_summary_show,
3275 };
3276
3277 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3278 {
3279 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3280 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3281
3282 BUG_ON(!cachep);
3283 return cachep;
3284 }
3285
3286 /*
3287 * Allocate the top-level s_group_info array for the specified number
3288 * of groups
3289 */
3290 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3291 {
3292 struct ext4_sb_info *sbi = EXT4_SB(sb);
3293 unsigned size;
3294 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3295
3296 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3297 EXT4_DESC_PER_BLOCK_BITS(sb);
3298 if (size <= sbi->s_group_info_size)
3299 return 0;
3300
3301 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3302 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3303 if (!new_groupinfo) {
3304 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3305 return -ENOMEM;
3306 }
3307 rcu_read_lock();
3308 old_groupinfo = rcu_dereference(sbi->s_group_info);
3309 if (old_groupinfo)
3310 memcpy(new_groupinfo, old_groupinfo,
3311 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3312 rcu_read_unlock();
3313 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3314 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3315 if (old_groupinfo)
3316 ext4_kvfree_array_rcu(old_groupinfo);
3317 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3318 sbi->s_group_info_size);
3319 return 0;
3320 }
3321
3322 /* Create and initialize ext4_group_info data for the given group. */
3323 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3324 struct ext4_group_desc *desc)
3325 {
3326 int i;
3327 int metalen = 0;
3328 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3329 struct ext4_sb_info *sbi = EXT4_SB(sb);
3330 struct ext4_group_info **meta_group_info;
3331 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3332
3333 /*
3334 * First check if this group is the first of a reserved block.
3335 * If it's true, we have to allocate a new table of pointers
3336 * to ext4_group_info structures
3337 */
3338 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3339 metalen = sizeof(*meta_group_info) <<
3340 EXT4_DESC_PER_BLOCK_BITS(sb);
3341 meta_group_info = kmalloc(metalen, GFP_NOFS);
3342 if (meta_group_info == NULL) {
3343 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3344 "for a buddy group");
3345 return -ENOMEM;
3346 }
3347 rcu_read_lock();
3348 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3349 rcu_read_unlock();
3350 }
3351
3352 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3353 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3354
3355 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3356 if (meta_group_info[i] == NULL) {
3357 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3358 goto exit_group_info;
3359 }
3360 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3361 &(meta_group_info[i]->bb_state));
3362
3363 /*
3364 * initialize bb_free to be able to skip
3365 * empty groups without initialization
3366 */
3367 if (ext4_has_group_desc_csum(sb) &&
3368 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3369 meta_group_info[i]->bb_free =
3370 ext4_free_clusters_after_init(sb, group, desc);
3371 } else {
3372 meta_group_info[i]->bb_free =
3373 ext4_free_group_clusters(sb, desc);
3374 }
3375
3376 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3377 init_rwsem(&meta_group_info[i]->alloc_sem);
3378 meta_group_info[i]->bb_free_root = RB_ROOT;
3379 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3380 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3381 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3382 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3383 meta_group_info[i]->bb_group = group;
3384
3385 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3386 return 0;
3387
3388 exit_group_info:
3389 /* If a meta_group_info table has been allocated, release it now */
3390 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3391 struct ext4_group_info ***group_info;
3392
3393 rcu_read_lock();
3394 group_info = rcu_dereference(sbi->s_group_info);
3395 kfree(group_info[idx]);
3396 group_info[idx] = NULL;
3397 rcu_read_unlock();
3398 }
3399 return -ENOMEM;
3400 } /* ext4_mb_add_groupinfo */
3401
3402 static int ext4_mb_init_backend(struct super_block *sb)
3403 {
3404 ext4_group_t ngroups = ext4_get_groups_count(sb);
3405 ext4_group_t i;
3406 struct ext4_sb_info *sbi = EXT4_SB(sb);
3407 int err;
3408 struct ext4_group_desc *desc;
3409 struct ext4_group_info ***group_info;
3410 struct kmem_cache *cachep;
3411
3412 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3413 if (err)
3414 return err;
3415
3416 sbi->s_buddy_cache = new_inode(sb);
3417 if (sbi->s_buddy_cache == NULL) {
3418 ext4_msg(sb, KERN_ERR, "can't get new inode");
3419 goto err_freesgi;
3420 }
3421 /* To avoid potentially colliding with an valid on-disk inode number,
3422 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3423 * not in the inode hash, so it should never be found by iget(), but
3424 * this will avoid confusion if it ever shows up during debugging. */
3425 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3426 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3427 for (i = 0; i < ngroups; i++) {
3428 cond_resched();
3429 desc = ext4_get_group_desc(sb, i, NULL);
3430 if (desc == NULL) {
3431 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3432 goto err_freebuddy;
3433 }
3434 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3435 goto err_freebuddy;
3436 }
3437
3438 if (ext4_has_feature_flex_bg(sb)) {
3439 /* a single flex group is supposed to be read by a single IO.
3440 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3441 * unsigned integer, so the maximum shift is 32.
3442 */
3443 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3444 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3445 goto err_freebuddy;
3446 }
3447 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3448 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3449 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3450 } else {
3451 sbi->s_mb_prefetch = 32;
3452 }
3453 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3454 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3455 /*
3456 * now many real IOs to prefetch within a single allocation at
3457 * CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related
3458 * optimization we shouldn't try to load too many groups, at some point
3459 * we should start to use what we've got in memory.
3460 * with an average random access time 5ms, it'd take a second to get
3461 * 200 groups (* N with flex_bg), so let's make this limit 4
3462 */
3463 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3464 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3465 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3466
3467 return 0;
3468
3469 err_freebuddy:
3470 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3471 while (i-- > 0) {
3472 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3473
3474 if (grp)
3475 kmem_cache_free(cachep, grp);
3476 }
3477 i = sbi->s_group_info_size;
3478 rcu_read_lock();
3479 group_info = rcu_dereference(sbi->s_group_info);
3480 while (i-- > 0)
3481 kfree(group_info[i]);
3482 rcu_read_unlock();
3483 iput(sbi->s_buddy_cache);
3484 err_freesgi:
3485 rcu_read_lock();
3486 kvfree(rcu_dereference(sbi->s_group_info));
3487 rcu_read_unlock();
3488 return -ENOMEM;
3489 }
3490
3491 static void ext4_groupinfo_destroy_slabs(void)
3492 {
3493 int i;
3494
3495 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3496 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3497 ext4_groupinfo_caches[i] = NULL;
3498 }
3499 }
3500
3501 static int ext4_groupinfo_create_slab(size_t size)
3502 {
3503 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3504 int slab_size;
3505 int blocksize_bits = order_base_2(size);
3506 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3507 struct kmem_cache *cachep;
3508
3509 if (cache_index >= NR_GRPINFO_CACHES)
3510 return -EINVAL;
3511
3512 if (unlikely(cache_index < 0))
3513 cache_index = 0;
3514
3515 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3516 if (ext4_groupinfo_caches[cache_index]) {
3517 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3518 return 0; /* Already created */
3519 }
3520
3521 slab_size = offsetof(struct ext4_group_info,
3522 bb_counters[blocksize_bits + 2]);
3523
3524 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3525 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3526 NULL);
3527
3528 ext4_groupinfo_caches[cache_index] = cachep;
3529
3530 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3531 if (!cachep) {
3532 printk(KERN_EMERG
3533 "EXT4-fs: no memory for groupinfo slab cache\n");
3534 return -ENOMEM;
3535 }
3536
3537 return 0;
3538 }
3539
3540 static void ext4_discard_work(struct work_struct *work)
3541 {
3542 struct ext4_sb_info *sbi = container_of(work,
3543 struct ext4_sb_info, s_discard_work);
3544 struct super_block *sb = sbi->s_sb;
3545 struct ext4_free_data *fd, *nfd;
3546 struct ext4_buddy e4b;
3547 LIST_HEAD(discard_list);
3548 ext4_group_t grp, load_grp;
3549 int err = 0;
3550
3551 spin_lock(&sbi->s_md_lock);
3552 list_splice_init(&sbi->s_discard_list, &discard_list);
3553 spin_unlock(&sbi->s_md_lock);
3554
3555 load_grp = UINT_MAX;
3556 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3557 /*
3558 * If filesystem is umounting or no memory or suffering
3559 * from no space, give up the discard
3560 */
3561 if ((sb->s_flags & SB_ACTIVE) && !err &&
3562 !atomic_read(&sbi->s_retry_alloc_pending)) {
3563 grp = fd->efd_group;
3564 if (grp != load_grp) {
3565 if (load_grp != UINT_MAX)
3566 ext4_mb_unload_buddy(&e4b);
3567
3568 err = ext4_mb_load_buddy(sb, grp, &e4b);
3569 if (err) {
3570 kmem_cache_free(ext4_free_data_cachep, fd);
3571 load_grp = UINT_MAX;
3572 continue;
3573 } else {
3574 load_grp = grp;
3575 }
3576 }
3577
3578 ext4_lock_group(sb, grp);
3579 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3580 fd->efd_start_cluster + fd->efd_count - 1, 1);
3581 ext4_unlock_group(sb, grp);
3582 }
3583 kmem_cache_free(ext4_free_data_cachep, fd);
3584 }
3585
3586 if (load_grp != UINT_MAX)
3587 ext4_mb_unload_buddy(&e4b);
3588 }
3589
3590 int ext4_mb_init(struct super_block *sb)
3591 {
3592 struct ext4_sb_info *sbi = EXT4_SB(sb);
3593 unsigned i, j;
3594 unsigned offset, offset_incr;
3595 unsigned max;
3596 int ret;
3597
3598 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3599
3600 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3601 if (sbi->s_mb_offsets == NULL) {
3602 ret = -ENOMEM;
3603 goto out;
3604 }
3605
3606 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3607 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3608 if (sbi->s_mb_maxs == NULL) {
3609 ret = -ENOMEM;
3610 goto out;
3611 }
3612
3613 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3614 if (ret < 0)
3615 goto out;
3616
3617 /* order 0 is regular bitmap */
3618 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3619 sbi->s_mb_offsets[0] = 0;
3620
3621 i = 1;
3622 offset = 0;
3623 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3624 max = sb->s_blocksize << 2;
3625 do {
3626 sbi->s_mb_offsets[i] = offset;
3627 sbi->s_mb_maxs[i] = max;
3628 offset += offset_incr;
3629 offset_incr = offset_incr >> 1;
3630 max = max >> 1;
3631 i++;
3632 } while (i < MB_NUM_ORDERS(sb));
3633
3634 sbi->s_mb_avg_fragment_size =
3635 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3636 GFP_KERNEL);
3637 if (!sbi->s_mb_avg_fragment_size) {
3638 ret = -ENOMEM;
3639 goto out;
3640 }
3641 sbi->s_mb_avg_fragment_size_locks =
3642 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3643 GFP_KERNEL);
3644 if (!sbi->s_mb_avg_fragment_size_locks) {
3645 ret = -ENOMEM;
3646 goto out;
3647 }
3648 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3649 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3650 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3651 }
3652 sbi->s_mb_largest_free_orders =
3653 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3654 GFP_KERNEL);
3655 if (!sbi->s_mb_largest_free_orders) {
3656 ret = -ENOMEM;
3657 goto out;
3658 }
3659 sbi->s_mb_largest_free_orders_locks =
3660 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3661 GFP_KERNEL);
3662 if (!sbi->s_mb_largest_free_orders_locks) {
3663 ret = -ENOMEM;
3664 goto out;
3665 }
3666 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3667 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3668 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3669 }
3670
3671 spin_lock_init(&sbi->s_md_lock);
3672 sbi->s_mb_free_pending = 0;
3673 INIT_LIST_HEAD(&sbi->s_freed_data_list[0]);
3674 INIT_LIST_HEAD(&sbi->s_freed_data_list[1]);
3675 INIT_LIST_HEAD(&sbi->s_discard_list);
3676 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3677 atomic_set(&sbi->s_retry_alloc_pending, 0);
3678
3679 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3680 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3681 sbi->s_mb_stats = MB_DEFAULT_STATS;
3682 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3683 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3684 sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER;
3685
3686 /*
3687 * The default group preallocation is 512, which for 4k block
3688 * sizes translates to 2 megabytes. However for bigalloc file
3689 * systems, this is probably too big (i.e, if the cluster size
3690 * is 1 megabyte, then group preallocation size becomes half a
3691 * gigabyte!). As a default, we will keep a two megabyte
3692 * group pralloc size for cluster sizes up to 64k, and after
3693 * that, we will force a minimum group preallocation size of
3694 * 32 clusters. This translates to 8 megs when the cluster
3695 * size is 256k, and 32 megs when the cluster size is 1 meg,
3696 * which seems reasonable as a default.
3697 */
3698 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3699 sbi->s_cluster_bits, 32);
3700 /*
3701 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3702 * to the lowest multiple of s_stripe which is bigger than
3703 * the s_mb_group_prealloc as determined above. We want
3704 * the preallocation size to be an exact multiple of the
3705 * RAID stripe size so that preallocations don't fragment
3706 * the stripes.
3707 */
3708 if (sbi->s_stripe > 1) {
3709 sbi->s_mb_group_prealloc = roundup(
3710 sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe));
3711 }
3712
3713 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3714 if (sbi->s_locality_groups == NULL) {
3715 ret = -ENOMEM;
3716 goto out;
3717 }
3718 for_each_possible_cpu(i) {
3719 struct ext4_locality_group *lg;
3720 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3721 mutex_init(&lg->lg_mutex);
3722 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3723 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3724 spin_lock_init(&lg->lg_prealloc_lock);
3725 }
3726
3727 if (bdev_nonrot(sb->s_bdev))
3728 sbi->s_mb_max_linear_groups = 0;
3729 else
3730 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3731 /* init file for buddy data */
3732 ret = ext4_mb_init_backend(sb);
3733 if (ret != 0)
3734 goto out_free_locality_groups;
3735
3736 return 0;
3737
3738 out_free_locality_groups:
3739 free_percpu(sbi->s_locality_groups);
3740 sbi->s_locality_groups = NULL;
3741 out:
3742 kfree(sbi->s_mb_avg_fragment_size);
3743 kfree(sbi->s_mb_avg_fragment_size_locks);
3744 kfree(sbi->s_mb_largest_free_orders);
3745 kfree(sbi->s_mb_largest_free_orders_locks);
3746 kfree(sbi->s_mb_offsets);
3747 sbi->s_mb_offsets = NULL;
3748 kfree(sbi->s_mb_maxs);
3749 sbi->s_mb_maxs = NULL;
3750 return ret;
3751 }
3752
3753 /* need to called with the ext4 group lock held */
3754 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3755 {
3756 struct ext4_prealloc_space *pa;
3757 struct list_head *cur, *tmp;
3758 int count = 0;
3759
3760 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3761 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3762 list_del(&pa->pa_group_list);
3763 count++;
3764 kmem_cache_free(ext4_pspace_cachep, pa);
3765 }
3766 return count;
3767 }
3768
3769 void ext4_mb_release(struct super_block *sb)
3770 {
3771 ext4_group_t ngroups = ext4_get_groups_count(sb);
3772 ext4_group_t i;
3773 int num_meta_group_infos;
3774 struct ext4_group_info *grinfo, ***group_info;
3775 struct ext4_sb_info *sbi = EXT4_SB(sb);
3776 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3777 int count;
3778
3779 if (test_opt(sb, DISCARD)) {
3780 /*
3781 * wait the discard work to drain all of ext4_free_data
3782 */
3783 flush_work(&sbi->s_discard_work);
3784 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3785 }
3786
3787 if (sbi->s_group_info) {
3788 for (i = 0; i < ngroups; i++) {
3789 cond_resched();
3790 grinfo = ext4_get_group_info(sb, i);
3791 if (!grinfo)
3792 continue;
3793 mb_group_bb_bitmap_free(grinfo);
3794 ext4_lock_group(sb, i);
3795 count = ext4_mb_cleanup_pa(grinfo);
3796 if (count)
3797 mb_debug(sb, "mballoc: %d PAs left\n",
3798 count);
3799 ext4_unlock_group(sb, i);
3800 kmem_cache_free(cachep, grinfo);
3801 }
3802 num_meta_group_infos = (ngroups +
3803 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3804 EXT4_DESC_PER_BLOCK_BITS(sb);
3805 rcu_read_lock();
3806 group_info = rcu_dereference(sbi->s_group_info);
3807 for (i = 0; i < num_meta_group_infos; i++)
3808 kfree(group_info[i]);
3809 kvfree(group_info);
3810 rcu_read_unlock();
3811 }
3812 kfree(sbi->s_mb_avg_fragment_size);
3813 kfree(sbi->s_mb_avg_fragment_size_locks);
3814 kfree(sbi->s_mb_largest_free_orders);
3815 kfree(sbi->s_mb_largest_free_orders_locks);
3816 kfree(sbi->s_mb_offsets);
3817 kfree(sbi->s_mb_maxs);
3818 iput(sbi->s_buddy_cache);
3819 if (sbi->s_mb_stats) {
3820 ext4_msg(sb, KERN_INFO,
3821 "mballoc: %u blocks %u reqs (%u success)",
3822 atomic_read(&sbi->s_bal_allocated),
3823 atomic_read(&sbi->s_bal_reqs),
3824 atomic_read(&sbi->s_bal_success));
3825 ext4_msg(sb, KERN_INFO,
3826 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3827 "%u 2^N hits, %u breaks, %u lost",
3828 atomic_read(&sbi->s_bal_ex_scanned),
3829 atomic_read(&sbi->s_bal_groups_scanned),
3830 atomic_read(&sbi->s_bal_goals),
3831 atomic_read(&sbi->s_bal_2orders),
3832 atomic_read(&sbi->s_bal_breaks),
3833 atomic_read(&sbi->s_mb_lost_chunks));
3834 ext4_msg(sb, KERN_INFO,
3835 "mballoc: %u generated and it took %llu",
3836 atomic_read(&sbi->s_mb_buddies_generated),
3837 atomic64_read(&sbi->s_mb_generation_time));
3838 ext4_msg(sb, KERN_INFO,
3839 "mballoc: %u preallocated, %u discarded",
3840 atomic_read(&sbi->s_mb_preallocated),
3841 atomic_read(&sbi->s_mb_discarded));
3842 }
3843
3844 free_percpu(sbi->s_locality_groups);
3845 }
3846
3847 static inline int ext4_issue_discard(struct super_block *sb,
3848 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
3849 {
3850 ext4_fsblk_t discard_block;
3851
3852 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3853 ext4_group_first_block_no(sb, block_group));
3854 count = EXT4_C2B(EXT4_SB(sb), count);
3855 trace_ext4_discard_blocks(sb,
3856 (unsigned long long) discard_block, count);
3857
3858 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3859 }
3860
3861 static void ext4_free_data_in_buddy(struct super_block *sb,
3862 struct ext4_free_data *entry)
3863 {
3864 struct ext4_buddy e4b;
3865 struct ext4_group_info *db;
3866 int err, count = 0;
3867
3868 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3869 entry->efd_count, entry->efd_group, entry);
3870
3871 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3872 /* we expect to find existing buddy because it's pinned */
3873 BUG_ON(err != 0);
3874
3875 spin_lock(&EXT4_SB(sb)->s_md_lock);
3876 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3877 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3878
3879 db = e4b.bd_info;
3880 /* there are blocks to put in buddy to make them really free */
3881 count += entry->efd_count;
3882 ext4_lock_group(sb, entry->efd_group);
3883 /* Take it out of per group rb tree */
3884 rb_erase(&entry->efd_node, &(db->bb_free_root));
3885 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3886
3887 /*
3888 * Clear the trimmed flag for the group so that the next
3889 * ext4_trim_fs can trim it.
3890 */
3891 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3892
3893 if (!db->bb_free_root.rb_node) {
3894 /* No more items in the per group rb tree
3895 * balance refcounts from ext4_mb_free_metadata()
3896 */
3897 folio_put(e4b.bd_buddy_folio);
3898 folio_put(e4b.bd_bitmap_folio);
3899 }
3900 ext4_unlock_group(sb, entry->efd_group);
3901 ext4_mb_unload_buddy(&e4b);
3902
3903 mb_debug(sb, "freed %d blocks in 1 structures\n", count);
3904 }
3905
3906 /*
3907 * This function is called by the jbd2 layer once the commit has finished,
3908 * so we know we can free the blocks that were released with that commit.
3909 */
3910 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3911 {
3912 struct ext4_sb_info *sbi = EXT4_SB(sb);
3913 struct ext4_free_data *entry, *tmp;
3914 LIST_HEAD(freed_data_list);
3915 struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1];
3916 bool wake;
3917
3918 list_replace_init(s_freed_head, &freed_data_list);
3919
3920 list_for_each_entry(entry, &freed_data_list, efd_list)
3921 ext4_free_data_in_buddy(sb, entry);
3922
3923 if (test_opt(sb, DISCARD)) {
3924 spin_lock(&sbi->s_md_lock);
3925 wake = list_empty(&sbi->s_discard_list);
3926 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3927 spin_unlock(&sbi->s_md_lock);
3928 if (wake)
3929 queue_work(system_unbound_wq, &sbi->s_discard_work);
3930 } else {
3931 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3932 kmem_cache_free(ext4_free_data_cachep, entry);
3933 }
3934 }
3935
3936 int __init ext4_init_mballoc(void)
3937 {
3938 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3939 SLAB_RECLAIM_ACCOUNT);
3940 if (ext4_pspace_cachep == NULL)
3941 goto out;
3942
3943 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3944 SLAB_RECLAIM_ACCOUNT);
3945 if (ext4_ac_cachep == NULL)
3946 goto out_pa_free;
3947
3948 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3949 SLAB_RECLAIM_ACCOUNT);
3950 if (ext4_free_data_cachep == NULL)
3951 goto out_ac_free;
3952
3953 return 0;
3954
3955 out_ac_free:
3956 kmem_cache_destroy(ext4_ac_cachep);
3957 out_pa_free:
3958 kmem_cache_destroy(ext4_pspace_cachep);
3959 out:
3960 return -ENOMEM;
3961 }
3962
3963 void ext4_exit_mballoc(void)
3964 {
3965 /*
3966 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3967 * before destroying the slab cache.
3968 */
3969 rcu_barrier();
3970 kmem_cache_destroy(ext4_pspace_cachep);
3971 kmem_cache_destroy(ext4_ac_cachep);
3972 kmem_cache_destroy(ext4_free_data_cachep);
3973 ext4_groupinfo_destroy_slabs();
3974 }
3975
3976 #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001
3977 #define EXT4_MB_SYNC_UPDATE 0x0002
3978 static int
3979 ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state,
3980 ext4_group_t group, ext4_grpblk_t blkoff,
3981 ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed)
3982 {
3983 struct ext4_sb_info *sbi = EXT4_SB(sb);
3984 struct buffer_head *bitmap_bh = NULL;
3985 struct ext4_group_desc *gdp;
3986 struct buffer_head *gdp_bh;
3987 int err;
3988 unsigned int i, already, changed = len;
3989
3990 KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context,
3991 handle, sb, state, group, blkoff, len,
3992 flags, ret_changed);
3993
3994 if (ret_changed)
3995 *ret_changed = 0;
3996 bitmap_bh = ext4_read_block_bitmap(sb, group);
3997 if (IS_ERR(bitmap_bh))
3998 return PTR_ERR(bitmap_bh);
3999
4000 if (handle) {
4001 BUFFER_TRACE(bitmap_bh, "getting write access");
4002 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
4003 EXT4_JTR_NONE);
4004 if (err)
4005 goto out_err;
4006 }
4007
4008 err = -EIO;
4009 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
4010 if (!gdp)
4011 goto out_err;
4012
4013 if (handle) {
4014 BUFFER_TRACE(gdp_bh, "get_write_access");
4015 err = ext4_journal_get_write_access(handle, sb, gdp_bh,
4016 EXT4_JTR_NONE);
4017 if (err)
4018 goto out_err;
4019 }
4020
4021 ext4_lock_group(sb, group);
4022 if (ext4_has_group_desc_csum(sb) &&
4023 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
4024 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
4025 ext4_free_group_clusters_set(sb, gdp,
4026 ext4_free_clusters_after_init(sb, group, gdp));
4027 }
4028
4029 if (flags & EXT4_MB_BITMAP_MARKED_CHECK) {
4030 already = 0;
4031 for (i = 0; i < len; i++)
4032 if (mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
4033 state)
4034 already++;
4035 changed = len - already;
4036 }
4037
4038 if (state) {
4039 mb_set_bits(bitmap_bh->b_data, blkoff, len);
4040 ext4_free_group_clusters_set(sb, gdp,
4041 ext4_free_group_clusters(sb, gdp) - changed);
4042 } else {
4043 mb_clear_bits(bitmap_bh->b_data, blkoff, len);
4044 ext4_free_group_clusters_set(sb, gdp,
4045 ext4_free_group_clusters(sb, gdp) + changed);
4046 }
4047
4048 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh);
4049 ext4_group_desc_csum_set(sb, group, gdp);
4050 ext4_unlock_group(sb, group);
4051 if (ret_changed)
4052 *ret_changed = changed;
4053
4054 if (sbi->s_log_groups_per_flex) {
4055 ext4_group_t flex_group = ext4_flex_group(sbi, group);
4056 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
4057 s_flex_groups, flex_group);
4058
4059 if (state)
4060 atomic64_sub(changed, &fg->free_clusters);
4061 else
4062 atomic64_add(changed, &fg->free_clusters);
4063 }
4064
4065 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4066 if (err)
4067 goto out_err;
4068 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
4069 if (err)
4070 goto out_err;
4071
4072 if (flags & EXT4_MB_SYNC_UPDATE) {
4073 sync_dirty_buffer(bitmap_bh);
4074 sync_dirty_buffer(gdp_bh);
4075 }
4076
4077 out_err:
4078 brelse(bitmap_bh);
4079 return err;
4080 }
4081
4082 /*
4083 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
4084 * Returns 0 if success or error code
4085 */
4086 static noinline_for_stack int
4087 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
4088 handle_t *handle, unsigned int reserv_clstrs)
4089 {
4090 struct ext4_group_desc *gdp;
4091 struct ext4_sb_info *sbi;
4092 struct super_block *sb;
4093 ext4_fsblk_t block;
4094 int err, len;
4095 int flags = 0;
4096 ext4_grpblk_t changed;
4097
4098 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4099 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4100
4101 sb = ac->ac_sb;
4102 sbi = EXT4_SB(sb);
4103
4104 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL);
4105 if (!gdp)
4106 return -EIO;
4107 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
4108 ext4_free_group_clusters(sb, gdp));
4109
4110 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4111 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4112 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
4113 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
4114 "fs metadata", block, block+len);
4115 /* File system mounted not to panic on error
4116 * Fix the bitmap and return EFSCORRUPTED
4117 * We leak some of the blocks here.
4118 */
4119 err = ext4_mb_mark_context(handle, sb, true,
4120 ac->ac_b_ex.fe_group,
4121 ac->ac_b_ex.fe_start,
4122 ac->ac_b_ex.fe_len,
4123 0, NULL);
4124 if (!err)
4125 err = -EFSCORRUPTED;
4126 return err;
4127 }
4128
4129 #ifdef AGGRESSIVE_CHECK
4130 flags |= EXT4_MB_BITMAP_MARKED_CHECK;
4131 #endif
4132 err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group,
4133 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len,
4134 flags, &changed);
4135
4136 if (err && changed == 0)
4137 return err;
4138
4139 #ifdef AGGRESSIVE_CHECK
4140 BUG_ON(changed != ac->ac_b_ex.fe_len);
4141 #endif
4142 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
4143 /*
4144 * Now reduce the dirty block count also. Should not go negative
4145 */
4146 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
4147 /* release all the reserved blocks if non delalloc */
4148 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4149 reserv_clstrs);
4150
4151 return err;
4152 }
4153
4154 /*
4155 * Idempotent helper for Ext4 fast commit replay path to set the state of
4156 * blocks in bitmaps and update counters.
4157 */
4158 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
4159 int len, bool state)
4160 {
4161 struct ext4_sb_info *sbi = EXT4_SB(sb);
4162 ext4_group_t group;
4163 ext4_grpblk_t blkoff;
4164 int err = 0;
4165 unsigned int clen, thisgrp_len;
4166
4167 while (len > 0) {
4168 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
4169
4170 /*
4171 * Check to see if we are freeing blocks across a group
4172 * boundary.
4173 * In case of flex_bg, this can happen that (block, len) may
4174 * span across more than one group. In that case we need to
4175 * get the corresponding group metadata to work with.
4176 * For this we have goto again loop.
4177 */
4178 thisgrp_len = min_t(unsigned int, (unsigned int)len,
4179 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
4180 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
4181
4182 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
4183 ext4_error(sb, "Marking blocks in system zone - "
4184 "Block = %llu, len = %u",
4185 block, thisgrp_len);
4186 break;
4187 }
4188
4189 err = ext4_mb_mark_context(NULL, sb, state,
4190 group, blkoff, clen,
4191 EXT4_MB_BITMAP_MARKED_CHECK |
4192 EXT4_MB_SYNC_UPDATE,
4193 NULL);
4194 if (err)
4195 break;
4196
4197 block += thisgrp_len;
4198 len -= thisgrp_len;
4199 BUG_ON(len < 0);
4200 }
4201 }
4202
4203 /*
4204 * here we normalize request for locality group
4205 * Group request are normalized to s_mb_group_prealloc, which goes to
4206 * s_strip if we set the same via mount option.
4207 * s_mb_group_prealloc can be configured via
4208 * /sys/fs/ext4/<partition>/mb_group_prealloc
4209 *
4210 * XXX: should we try to preallocate more than the group has now?
4211 */
4212 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4213 {
4214 struct super_block *sb = ac->ac_sb;
4215 struct ext4_locality_group *lg = ac->ac_lg;
4216
4217 BUG_ON(lg == NULL);
4218 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4219 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4220 }
4221
4222 /*
4223 * This function returns the next element to look at during inode
4224 * PA rbtree walk. We assume that we have held the inode PA rbtree lock
4225 * (ei->i_prealloc_lock)
4226 *
4227 * new_start The start of the range we want to compare
4228 * cur_start The existing start that we are comparing against
4229 * node The node of the rb_tree
4230 */
4231 static inline struct rb_node*
4232 ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node)
4233 {
4234 if (new_start < cur_start)
4235 return node->rb_left;
4236 else
4237 return node->rb_right;
4238 }
4239
4240 static inline void
4241 ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
4242 ext4_lblk_t start, loff_t end)
4243 {
4244 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4245 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4246 struct ext4_prealloc_space *tmp_pa;
4247 ext4_lblk_t tmp_pa_start;
4248 loff_t tmp_pa_end;
4249 struct rb_node *iter;
4250
4251 read_lock(&ei->i_prealloc_lock);
4252 for (iter = ei->i_prealloc_node.rb_node; iter;
4253 iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) {
4254 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4255 pa_node.inode_node);
4256 tmp_pa_start = tmp_pa->pa_lstart;
4257 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4258
4259 spin_lock(&tmp_pa->pa_lock);
4260 if (tmp_pa->pa_deleted == 0)
4261 BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
4262 spin_unlock(&tmp_pa->pa_lock);
4263 }
4264 read_unlock(&ei->i_prealloc_lock);
4265 }
4266
4267 /*
4268 * Given an allocation context "ac" and a range "start", "end", check
4269 * and adjust boundaries if the range overlaps with any of the existing
4270 * preallocatoins stored in the corresponding inode of the allocation context.
4271 *
4272 * Parameters:
4273 * ac allocation context
4274 * start start of the new range
4275 * end end of the new range
4276 */
4277 static inline void
4278 ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
4279 ext4_lblk_t *start, loff_t *end)
4280 {
4281 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4282 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4283 struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL;
4284 struct rb_node *iter;
4285 ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1;
4286 loff_t new_end, tmp_pa_end, left_pa_end = -1;
4287
4288 new_start = *start;
4289 new_end = *end;
4290
4291 /*
4292 * Adjust the normalized range so that it doesn't overlap with any
4293 * existing preallocated blocks(PAs). Make sure to hold the rbtree lock
4294 * so it doesn't change underneath us.
4295 */
4296 read_lock(&ei->i_prealloc_lock);
4297
4298 /* Step 1: find any one immediate neighboring PA of the normalized range */
4299 for (iter = ei->i_prealloc_node.rb_node; iter;
4300 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4301 tmp_pa_start, iter)) {
4302 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4303 pa_node.inode_node);
4304 tmp_pa_start = tmp_pa->pa_lstart;
4305 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4306
4307 /* PA must not overlap original request */
4308 spin_lock(&tmp_pa->pa_lock);
4309 if (tmp_pa->pa_deleted == 0)
4310 BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
4311 ac->ac_o_ex.fe_logical < tmp_pa_start));
4312 spin_unlock(&tmp_pa->pa_lock);
4313 }
4314
4315 /*
4316 * Step 2: check if the found PA is left or right neighbor and
4317 * get the other neighbor
4318 */
4319 if (tmp_pa) {
4320 if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) {
4321 struct rb_node *tmp;
4322
4323 left_pa = tmp_pa;
4324 tmp = rb_next(&left_pa->pa_node.inode_node);
4325 if (tmp) {
4326 right_pa = rb_entry(tmp,
4327 struct ext4_prealloc_space,
4328 pa_node.inode_node);
4329 }
4330 } else {
4331 struct rb_node *tmp;
4332
4333 right_pa = tmp_pa;
4334 tmp = rb_prev(&right_pa->pa_node.inode_node);
4335 if (tmp) {
4336 left_pa = rb_entry(tmp,
4337 struct ext4_prealloc_space,
4338 pa_node.inode_node);
4339 }
4340 }
4341 }
4342
4343 /* Step 3: get the non deleted neighbors */
4344 if (left_pa) {
4345 for (iter = &left_pa->pa_node.inode_node;;
4346 iter = rb_prev(iter)) {
4347 if (!iter) {
4348 left_pa = NULL;
4349 break;
4350 }
4351
4352 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4353 pa_node.inode_node);
4354 left_pa = tmp_pa;
4355 spin_lock(&tmp_pa->pa_lock);
4356 if (tmp_pa->pa_deleted == 0) {
4357 spin_unlock(&tmp_pa->pa_lock);
4358 break;
4359 }
4360 spin_unlock(&tmp_pa->pa_lock);
4361 }
4362 }
4363
4364 if (right_pa) {
4365 for (iter = &right_pa->pa_node.inode_node;;
4366 iter = rb_next(iter)) {
4367 if (!iter) {
4368 right_pa = NULL;
4369 break;
4370 }
4371
4372 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4373 pa_node.inode_node);
4374 right_pa = tmp_pa;
4375 spin_lock(&tmp_pa->pa_lock);
4376 if (tmp_pa->pa_deleted == 0) {
4377 spin_unlock(&tmp_pa->pa_lock);
4378 break;
4379 }
4380 spin_unlock(&tmp_pa->pa_lock);
4381 }
4382 }
4383
4384 if (left_pa) {
4385 left_pa_end = pa_logical_end(sbi, left_pa);
4386 BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical);
4387 }
4388
4389 if (right_pa) {
4390 right_pa_start = right_pa->pa_lstart;
4391 BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical);
4392 }
4393
4394 /* Step 4: trim our normalized range to not overlap with the neighbors */
4395 if (left_pa) {
4396 if (left_pa_end > new_start)
4397 new_start = left_pa_end;
4398 }
4399
4400 if (right_pa) {
4401 if (right_pa_start < new_end)
4402 new_end = right_pa_start;
4403 }
4404 read_unlock(&ei->i_prealloc_lock);
4405
4406 /* XXX: extra loop to check we really don't overlap preallocations */
4407 ext4_mb_pa_assert_overlap(ac, new_start, new_end);
4408
4409 *start = new_start;
4410 *end = new_end;
4411 }
4412
4413 /*
4414 * Normalization means making request better in terms of
4415 * size and alignment
4416 */
4417 static noinline_for_stack void
4418 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4419 struct ext4_allocation_request *ar)
4420 {
4421 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4422 struct ext4_super_block *es = sbi->s_es;
4423 int bsbits, max;
4424 loff_t size, start_off, end;
4425 loff_t orig_size __maybe_unused;
4426 ext4_lblk_t start;
4427
4428 /* do normalize only data requests, metadata requests
4429 do not need preallocation */
4430 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4431 return;
4432
4433 /* sometime caller may want exact blocks */
4434 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4435 return;
4436
4437 /* caller may indicate that preallocation isn't
4438 * required (it's a tail, for example) */
4439 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4440 return;
4441
4442 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4443 ext4_mb_normalize_group_request(ac);
4444 return ;
4445 }
4446
4447 bsbits = ac->ac_sb->s_blocksize_bits;
4448
4449 /* first, let's learn actual file size
4450 * given current request is allocated */
4451 size = extent_logical_end(sbi, &ac->ac_o_ex);
4452 size = size << bsbits;
4453 if (size < i_size_read(ac->ac_inode))
4454 size = i_size_read(ac->ac_inode);
4455 orig_size = size;
4456
4457 /* max size of free chunks */
4458 max = 2 << bsbits;
4459
4460 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4461 (req <= (size) || max <= (chunk_size))
4462
4463 /* first, try to predict filesize */
4464 /* XXX: should this table be tunable? */
4465 start_off = 0;
4466 if (size <= 16 * 1024) {
4467 size = 16 * 1024;
4468 } else if (size <= 32 * 1024) {
4469 size = 32 * 1024;
4470 } else if (size <= 64 * 1024) {
4471 size = 64 * 1024;
4472 } else if (size <= 128 * 1024) {
4473 size = 128 * 1024;
4474 } else if (size <= 256 * 1024) {
4475 size = 256 * 1024;
4476 } else if (size <= 512 * 1024) {
4477 size = 512 * 1024;
4478 } else if (size <= 1024 * 1024) {
4479 size = 1024 * 1024;
4480 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4481 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4482 (21 - bsbits)) << 21;
4483 size = 2 * 1024 * 1024;
4484 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4485 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4486 (22 - bsbits)) << 22;
4487 size = 4 * 1024 * 1024;
4488 } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len),
4489 (8<<20)>>bsbits, max, 8 * 1024)) {
4490 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4491 (23 - bsbits)) << 23;
4492 size = 8 * 1024 * 1024;
4493 } else {
4494 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4495 size = (loff_t) EXT4_C2B(sbi,
4496 ac->ac_o_ex.fe_len) << bsbits;
4497 }
4498 size = size >> bsbits;
4499 start = start_off >> bsbits;
4500
4501 /*
4502 * For tiny groups (smaller than 8MB) the chosen allocation
4503 * alignment may be larger than group size. Make sure the
4504 * alignment does not move allocation to a different group which
4505 * makes mballoc fail assertions later.
4506 */
4507 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4508 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4509
4510 /* avoid unnecessary preallocation that may trigger assertions */
4511 if (start + size > EXT_MAX_BLOCKS)
4512 size = EXT_MAX_BLOCKS - start;
4513
4514 /* don't cover already allocated blocks in selected range */
4515 if (ar->pleft && start <= ar->lleft) {
4516 size -= ar->lleft + 1 - start;
4517 start = ar->lleft + 1;
4518 }
4519 if (ar->pright && start + size - 1 >= ar->lright)
4520 size -= start + size - ar->lright;
4521
4522 /*
4523 * Trim allocation request for filesystems with artificially small
4524 * groups.
4525 */
4526 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4527 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4528
4529 end = start + size;
4530
4531 ext4_mb_pa_adjust_overlap(ac, &start, &end);
4532
4533 size = end - start;
4534
4535 /*
4536 * In this function "start" and "size" are normalized for better
4537 * alignment and length such that we could preallocate more blocks.
4538 * This normalization is done such that original request of
4539 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4540 * "size" boundaries.
4541 * (Note fe_len can be relaxed since FS block allocation API does not
4542 * provide gurantee on number of contiguous blocks allocation since that
4543 * depends upon free space left, etc).
4544 * In case of inode pa, later we use the allocated blocks
4545 * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated
4546 * range of goal/best blocks [start, size] to put it at the
4547 * ac_o_ex.fe_logical extent of this inode.
4548 * (See ext4_mb_use_inode_pa() for more details)
4549 */
4550 if (start + size <= ac->ac_o_ex.fe_logical ||
4551 start > ac->ac_o_ex.fe_logical) {
4552 ext4_msg(ac->ac_sb, KERN_ERR,
4553 "start %lu, size %lu, fe_logical %lu",
4554 (unsigned long) start, (unsigned long) size,
4555 (unsigned long) ac->ac_o_ex.fe_logical);
4556 BUG();
4557 }
4558 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4559
4560 /* now prepare goal request */
4561
4562 /* XXX: is it better to align blocks WRT to logical
4563 * placement or satisfy big request as is */
4564 ac->ac_g_ex.fe_logical = start;
4565 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4566 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
4567
4568 /* define goal start in order to merge */
4569 if (ar->pright && (ar->lright == (start + size)) &&
4570 ar->pright >= size &&
4571 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
4572 /* merge to the right */
4573 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4574 &ac->ac_g_ex.fe_group,
4575 &ac->ac_g_ex.fe_start);
4576 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4577 }
4578 if (ar->pleft && (ar->lleft + 1 == start) &&
4579 ar->pleft + 1 < ext4_blocks_count(es)) {
4580 /* merge to the left */
4581 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4582 &ac->ac_g_ex.fe_group,
4583 &ac->ac_g_ex.fe_start);
4584 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4585 }
4586
4587 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4588 orig_size, start);
4589 }
4590
4591 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4592 {
4593 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4594
4595 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4596 atomic_inc(&sbi->s_bal_reqs);
4597 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4598 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4599 atomic_inc(&sbi->s_bal_success);
4600
4601 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4602 for (int i=0; i<EXT4_MB_NUM_CRS; i++) {
4603 atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]);
4604 }
4605
4606 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4607 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4608 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4609 atomic_inc(&sbi->s_bal_goals);
4610 /* did we allocate as much as normalizer originally wanted? */
4611 if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len)
4612 atomic_inc(&sbi->s_bal_len_goals);
4613
4614 if (ac->ac_found > sbi->s_mb_max_to_scan)
4615 atomic_inc(&sbi->s_bal_breaks);
4616 }
4617
4618 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4619 trace_ext4_mballoc_alloc(ac);
4620 else
4621 trace_ext4_mballoc_prealloc(ac);
4622 }
4623
4624 /*
4625 * Called on failure; free up any blocks from the inode PA for this
4626 * context. We don't need this for MB_GROUP_PA because we only change
4627 * pa_free in ext4_mb_release_context(), but on failure, we've already
4628 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4629 */
4630 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4631 {
4632 struct ext4_prealloc_space *pa = ac->ac_pa;
4633 struct ext4_buddy e4b;
4634 int err;
4635
4636 if (pa == NULL) {
4637 if (ac->ac_f_ex.fe_len == 0)
4638 return;
4639 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4640 if (WARN_RATELIMIT(err,
4641 "ext4: mb_load_buddy failed (%d)", err))
4642 /*
4643 * This should never happen since we pin the
4644 * pages in the ext4_allocation_context so
4645 * ext4_mb_load_buddy() should never fail.
4646 */
4647 return;
4648 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4649 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4650 ac->ac_f_ex.fe_len);
4651 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4652 ext4_mb_unload_buddy(&e4b);
4653 return;
4654 }
4655 if (pa->pa_type == MB_INODE_PA) {
4656 spin_lock(&pa->pa_lock);
4657 pa->pa_free += ac->ac_b_ex.fe_len;
4658 spin_unlock(&pa->pa_lock);
4659 }
4660 }
4661
4662 /*
4663 * use blocks preallocated to inode
4664 */
4665 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4666 struct ext4_prealloc_space *pa)
4667 {
4668 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4669 ext4_fsblk_t start;
4670 ext4_fsblk_t end;
4671 int len;
4672
4673 /* found preallocated blocks, use them */
4674 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4675 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4676 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4677 len = EXT4_NUM_B2C(sbi, end - start);
4678 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4679 &ac->ac_b_ex.fe_start);
4680 ac->ac_b_ex.fe_len = len;
4681 ac->ac_status = AC_STATUS_FOUND;
4682 ac->ac_pa = pa;
4683
4684 BUG_ON(start < pa->pa_pstart);
4685 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4686 BUG_ON(pa->pa_free < len);
4687 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4688 pa->pa_free -= len;
4689
4690 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4691 }
4692
4693 /*
4694 * use blocks preallocated to locality group
4695 */
4696 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4697 struct ext4_prealloc_space *pa)
4698 {
4699 unsigned int len = ac->ac_o_ex.fe_len;
4700
4701 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4702 &ac->ac_b_ex.fe_group,
4703 &ac->ac_b_ex.fe_start);
4704 ac->ac_b_ex.fe_len = len;
4705 ac->ac_status = AC_STATUS_FOUND;
4706 ac->ac_pa = pa;
4707
4708 /* we don't correct pa_pstart or pa_len here to avoid
4709 * possible race when the group is being loaded concurrently
4710 * instead we correct pa later, after blocks are marked
4711 * in on-disk bitmap -- see ext4_mb_release_context()
4712 * Other CPUs are prevented from allocating from this pa by lg_mutex
4713 */
4714 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4715 pa->pa_lstart, len, pa);
4716 }
4717
4718 /*
4719 * Return the prealloc space that have minimal distance
4720 * from the goal block. @cpa is the prealloc
4721 * space that is having currently known minimal distance
4722 * from the goal block.
4723 */
4724 static struct ext4_prealloc_space *
4725 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4726 struct ext4_prealloc_space *pa,
4727 struct ext4_prealloc_space *cpa)
4728 {
4729 ext4_fsblk_t cur_distance, new_distance;
4730
4731 if (cpa == NULL) {
4732 atomic_inc(&pa->pa_count);
4733 return pa;
4734 }
4735 cur_distance = abs(goal_block - cpa->pa_pstart);
4736 new_distance = abs(goal_block - pa->pa_pstart);
4737
4738 if (cur_distance <= new_distance)
4739 return cpa;
4740
4741 /* drop the previous reference */
4742 atomic_dec(&cpa->pa_count);
4743 atomic_inc(&pa->pa_count);
4744 return pa;
4745 }
4746
4747 /*
4748 * check if found pa meets EXT4_MB_HINT_GOAL_ONLY
4749 */
4750 static bool
4751 ext4_mb_pa_goal_check(struct ext4_allocation_context *ac,
4752 struct ext4_prealloc_space *pa)
4753 {
4754 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4755 ext4_fsblk_t start;
4756
4757 if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)))
4758 return true;
4759
4760 /*
4761 * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted
4762 * in ext4_mb_normalize_request and will keep same with ac_o_ex
4763 * from ext4_mb_initialize_context. Choose ac_g_ex here to keep
4764 * consistent with ext4_mb_find_by_goal.
4765 */
4766 start = pa->pa_pstart +
4767 (ac->ac_g_ex.fe_logical - pa->pa_lstart);
4768 if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start)
4769 return false;
4770
4771 if (ac->ac_g_ex.fe_len > pa->pa_len -
4772 EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart))
4773 return false;
4774
4775 return true;
4776 }
4777
4778 /*
4779 * search goal blocks in preallocated space
4780 */
4781 static noinline_for_stack bool
4782 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4783 {
4784 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4785 int order, i;
4786 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4787 struct ext4_locality_group *lg;
4788 struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL;
4789 struct rb_node *iter;
4790 ext4_fsblk_t goal_block;
4791
4792 /* only data can be preallocated */
4793 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4794 return false;
4795
4796 /*
4797 * first, try per-file preallocation by searching the inode pa rbtree.
4798 *
4799 * Here, we can't do a direct traversal of the tree because
4800 * ext4_mb_discard_group_preallocation() can paralelly mark the pa
4801 * deleted and that can cause direct traversal to skip some entries.
4802 */
4803 read_lock(&ei->i_prealloc_lock);
4804
4805 if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) {
4806 goto try_group_pa;
4807 }
4808
4809 /*
4810 * Step 1: Find a pa with logical start immediately adjacent to the
4811 * original logical start. This could be on the left or right.
4812 *
4813 * (tmp_pa->pa_lstart never changes so we can skip locking for it).
4814 */
4815 for (iter = ei->i_prealloc_node.rb_node; iter;
4816 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4817 tmp_pa->pa_lstart, iter)) {
4818 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4819 pa_node.inode_node);
4820 }
4821
4822 /*
4823 * Step 2: The adjacent pa might be to the right of logical start, find
4824 * the left adjacent pa. After this step we'd have a valid tmp_pa whose
4825 * logical start is towards the left of original request's logical start
4826 */
4827 if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4828 struct rb_node *tmp;
4829 tmp = rb_prev(&tmp_pa->pa_node.inode_node);
4830
4831 if (tmp) {
4832 tmp_pa = rb_entry(tmp, struct ext4_prealloc_space,
4833 pa_node.inode_node);
4834 } else {
4835 /*
4836 * If there is no adjacent pa to the left then finding
4837 * an overlapping pa is not possible hence stop searching
4838 * inode pa tree
4839 */
4840 goto try_group_pa;
4841 }
4842 }
4843
4844 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4845
4846 /*
4847 * Step 3: If the left adjacent pa is deleted, keep moving left to find
4848 * the first non deleted adjacent pa. After this step we should have a
4849 * valid tmp_pa which is guaranteed to be non deleted.
4850 */
4851 for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) {
4852 if (!iter) {
4853 /*
4854 * no non deleted left adjacent pa, so stop searching
4855 * inode pa tree
4856 */
4857 goto try_group_pa;
4858 }
4859 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4860 pa_node.inode_node);
4861 spin_lock(&tmp_pa->pa_lock);
4862 if (tmp_pa->pa_deleted == 0) {
4863 /*
4864 * We will keep holding the pa_lock from
4865 * this point on because we don't want group discard
4866 * to delete this pa underneath us. Since group
4867 * discard is anyways an ENOSPC operation it
4868 * should be okay for it to wait a few more cycles.
4869 */
4870 break;
4871 } else {
4872 spin_unlock(&tmp_pa->pa_lock);
4873 }
4874 }
4875
4876 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4877 BUG_ON(tmp_pa->pa_deleted == 1);
4878
4879 /*
4880 * Step 4: We now have the non deleted left adjacent pa. Only this
4881 * pa can possibly satisfy the request hence check if it overlaps
4882 * original logical start and stop searching if it doesn't.
4883 */
4884 if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) {
4885 spin_unlock(&tmp_pa->pa_lock);
4886 goto try_group_pa;
4887 }
4888
4889 /* non-extent files can't have physical blocks past 2^32 */
4890 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4891 (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
4892 EXT4_MAX_BLOCK_FILE_PHYS)) {
4893 /*
4894 * Since PAs don't overlap, we won't find any other PA to
4895 * satisfy this.
4896 */
4897 spin_unlock(&tmp_pa->pa_lock);
4898 goto try_group_pa;
4899 }
4900
4901 if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) {
4902 atomic_inc(&tmp_pa->pa_count);
4903 ext4_mb_use_inode_pa(ac, tmp_pa);
4904 spin_unlock(&tmp_pa->pa_lock);
4905 read_unlock(&ei->i_prealloc_lock);
4906 return true;
4907 } else {
4908 /*
4909 * We found a valid overlapping pa but couldn't use it because
4910 * it had no free blocks. This should ideally never happen
4911 * because:
4912 *
4913 * 1. When a new inode pa is added to rbtree it must have
4914 * pa_free > 0 since otherwise we won't actually need
4915 * preallocation.
4916 *
4917 * 2. An inode pa that is in the rbtree can only have it's
4918 * pa_free become zero when another thread calls:
4919 * ext4_mb_new_blocks
4920 * ext4_mb_use_preallocated
4921 * ext4_mb_use_inode_pa
4922 *
4923 * 3. Further, after the above calls make pa_free == 0, we will
4924 * immediately remove it from the rbtree in:
4925 * ext4_mb_new_blocks
4926 * ext4_mb_release_context
4927 * ext4_mb_put_pa
4928 *
4929 * 4. Since the pa_free becoming 0 and pa_free getting removed
4930 * from tree both happen in ext4_mb_new_blocks, which is always
4931 * called with i_data_sem held for data allocations, we can be
4932 * sure that another process will never see a pa in rbtree with
4933 * pa_free == 0.
4934 */
4935 WARN_ON_ONCE(tmp_pa->pa_free == 0);
4936 }
4937 spin_unlock(&tmp_pa->pa_lock);
4938 try_group_pa:
4939 read_unlock(&ei->i_prealloc_lock);
4940
4941 /* can we use group allocation? */
4942 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4943 return false;
4944
4945 /* inode may have no locality group for some reason */
4946 lg = ac->ac_lg;
4947 if (lg == NULL)
4948 return false;
4949 order = fls(ac->ac_o_ex.fe_len) - 1;
4950 if (order > PREALLOC_TB_SIZE - 1)
4951 /* The max size of hash table is PREALLOC_TB_SIZE */
4952 order = PREALLOC_TB_SIZE - 1;
4953
4954 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4955 /*
4956 * search for the prealloc space that is having
4957 * minimal distance from the goal block.
4958 */
4959 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4960 rcu_read_lock();
4961 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
4962 pa_node.lg_list) {
4963 spin_lock(&tmp_pa->pa_lock);
4964 if (tmp_pa->pa_deleted == 0 &&
4965 tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
4966
4967 cpa = ext4_mb_check_group_pa(goal_block,
4968 tmp_pa, cpa);
4969 }
4970 spin_unlock(&tmp_pa->pa_lock);
4971 }
4972 rcu_read_unlock();
4973 }
4974 if (cpa) {
4975 ext4_mb_use_group_pa(ac, cpa);
4976 return true;
4977 }
4978 return false;
4979 }
4980
4981 /*
4982 * the function goes through all preallocation in this group and marks them
4983 * used in in-core bitmap. buddy must be generated from this bitmap
4984 * Need to be called with ext4 group lock held
4985 */
4986 static noinline_for_stack
4987 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4988 ext4_group_t group)
4989 {
4990 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4991 struct ext4_prealloc_space *pa;
4992 struct list_head *cur;
4993 ext4_group_t groupnr;
4994 ext4_grpblk_t start;
4995 int preallocated = 0;
4996 int len;
4997
4998 if (!grp)
4999 return;
5000
5001 /* all form of preallocation discards first load group,
5002 * so the only competing code is preallocation use.
5003 * we don't need any locking here
5004 * notice we do NOT ignore preallocations with pa_deleted
5005 * otherwise we could leave used blocks available for
5006 * allocation in buddy when concurrent ext4_mb_put_pa()
5007 * is dropping preallocation
5008 */
5009 list_for_each(cur, &grp->bb_prealloc_list) {
5010 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
5011 spin_lock(&pa->pa_lock);
5012 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5013 &groupnr, &start);
5014 len = pa->pa_len;
5015 spin_unlock(&pa->pa_lock);
5016 if (unlikely(len == 0))
5017 continue;
5018 BUG_ON(groupnr != group);
5019 mb_set_bits(bitmap, start, len);
5020 preallocated += len;
5021 }
5022 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
5023 }
5024
5025 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
5026 struct ext4_prealloc_space *pa)
5027 {
5028 struct ext4_inode_info *ei;
5029
5030 if (pa->pa_deleted) {
5031 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
5032 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
5033 pa->pa_len);
5034 return;
5035 }
5036
5037 pa->pa_deleted = 1;
5038
5039 if (pa->pa_type == MB_INODE_PA) {
5040 ei = EXT4_I(pa->pa_inode);
5041 atomic_dec(&ei->i_prealloc_active);
5042 }
5043 }
5044
5045 static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
5046 {
5047 BUG_ON(!pa);
5048 BUG_ON(atomic_read(&pa->pa_count));
5049 BUG_ON(pa->pa_deleted == 0);
5050 kmem_cache_free(ext4_pspace_cachep, pa);
5051 }
5052
5053 static void ext4_mb_pa_callback(struct rcu_head *head)
5054 {
5055 struct ext4_prealloc_space *pa;
5056
5057 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
5058 ext4_mb_pa_free(pa);
5059 }
5060
5061 /*
5062 * drops a reference to preallocated space descriptor
5063 * if this was the last reference and the space is consumed
5064 */
5065 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
5066 struct super_block *sb, struct ext4_prealloc_space *pa)
5067 {
5068 ext4_group_t grp;
5069 ext4_fsblk_t grp_blk;
5070 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
5071
5072 /* in this short window concurrent discard can set pa_deleted */
5073 spin_lock(&pa->pa_lock);
5074 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
5075 spin_unlock(&pa->pa_lock);
5076 return;
5077 }
5078
5079 if (pa->pa_deleted == 1) {
5080 spin_unlock(&pa->pa_lock);
5081 return;
5082 }
5083
5084 ext4_mb_mark_pa_deleted(sb, pa);
5085 spin_unlock(&pa->pa_lock);
5086
5087 grp_blk = pa->pa_pstart;
5088 /*
5089 * If doing group-based preallocation, pa_pstart may be in the
5090 * next group when pa is used up
5091 */
5092 if (pa->pa_type == MB_GROUP_PA)
5093 grp_blk--;
5094
5095 grp = ext4_get_group_number(sb, grp_blk);
5096
5097 /*
5098 * possible race:
5099 *
5100 * P1 (buddy init) P2 (regular allocation)
5101 * find block B in PA
5102 * copy on-disk bitmap to buddy
5103 * mark B in on-disk bitmap
5104 * drop PA from group
5105 * mark all PAs in buddy
5106 *
5107 * thus, P1 initializes buddy with B available. to prevent this
5108 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
5109 * against that pair
5110 */
5111 ext4_lock_group(sb, grp);
5112 list_del(&pa->pa_group_list);
5113 ext4_unlock_group(sb, grp);
5114
5115 if (pa->pa_type == MB_INODE_PA) {
5116 write_lock(pa->pa_node_lock.inode_lock);
5117 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5118 write_unlock(pa->pa_node_lock.inode_lock);
5119 ext4_mb_pa_free(pa);
5120 } else {
5121 spin_lock(pa->pa_node_lock.lg_lock);
5122 list_del_rcu(&pa->pa_node.lg_list);
5123 spin_unlock(pa->pa_node_lock.lg_lock);
5124 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5125 }
5126 }
5127
5128 static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new)
5129 {
5130 struct rb_node **iter = &root->rb_node, *parent = NULL;
5131 struct ext4_prealloc_space *iter_pa, *new_pa;
5132 ext4_lblk_t iter_start, new_start;
5133
5134 while (*iter) {
5135 iter_pa = rb_entry(*iter, struct ext4_prealloc_space,
5136 pa_node.inode_node);
5137 new_pa = rb_entry(new, struct ext4_prealloc_space,
5138 pa_node.inode_node);
5139 iter_start = iter_pa->pa_lstart;
5140 new_start = new_pa->pa_lstart;
5141
5142 parent = *iter;
5143 if (new_start < iter_start)
5144 iter = &((*iter)->rb_left);
5145 else
5146 iter = &((*iter)->rb_right);
5147 }
5148
5149 rb_link_node(new, parent, iter);
5150 rb_insert_color(new, root);
5151 }
5152
5153 /*
5154 * creates new preallocated space for given inode
5155 */
5156 static noinline_for_stack void
5157 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
5158 {
5159 struct super_block *sb = ac->ac_sb;
5160 struct ext4_sb_info *sbi = EXT4_SB(sb);
5161 struct ext4_prealloc_space *pa;
5162 struct ext4_group_info *grp;
5163 struct ext4_inode_info *ei;
5164
5165 /* preallocate only when found space is larger then requested */
5166 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5167 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5168 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5169 BUG_ON(ac->ac_pa == NULL);
5170
5171 pa = ac->ac_pa;
5172
5173 if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) {
5174 struct ext4_free_extent ex = {
5175 .fe_logical = ac->ac_g_ex.fe_logical,
5176 .fe_len = ac->ac_orig_goal_len,
5177 };
5178 loff_t orig_goal_end = extent_logical_end(sbi, &ex);
5179 loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex);
5180
5181 /*
5182 * We can't allocate as much as normalizer wants, so we try
5183 * to get proper lstart to cover the original request, except
5184 * when the goal doesn't cover the original request as below:
5185 *
5186 * orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048
5187 * best_ex:0/200(200) -> adjusted: 1848/2048(200)
5188 */
5189 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
5190 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
5191
5192 /*
5193 * Use the below logic for adjusting best extent as it keeps
5194 * fragmentation in check while ensuring logical range of best
5195 * extent doesn't overflow out of goal extent:
5196 *
5197 * 1. Check if best ex can be kept at end of goal (before
5198 * cr_best_avail trimmed it) and still cover original start
5199 * 2. Else, check if best ex can be kept at start of goal and
5200 * still cover original end
5201 * 3. Else, keep the best ex at start of original request.
5202 */
5203 ex.fe_len = ac->ac_b_ex.fe_len;
5204
5205 ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len);
5206 if (ac->ac_o_ex.fe_logical >= ex.fe_logical)
5207 goto adjust_bex;
5208
5209 ex.fe_logical = ac->ac_g_ex.fe_logical;
5210 if (o_ex_end <= extent_logical_end(sbi, &ex))
5211 goto adjust_bex;
5212
5213 ex.fe_logical = ac->ac_o_ex.fe_logical;
5214 adjust_bex:
5215 ac->ac_b_ex.fe_logical = ex.fe_logical;
5216
5217 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
5218 BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end);
5219 }
5220
5221 pa->pa_lstart = ac->ac_b_ex.fe_logical;
5222 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5223 pa->pa_len = ac->ac_b_ex.fe_len;
5224 pa->pa_free = pa->pa_len;
5225 spin_lock_init(&pa->pa_lock);
5226 INIT_LIST_HEAD(&pa->pa_group_list);
5227 pa->pa_deleted = 0;
5228 pa->pa_type = MB_INODE_PA;
5229
5230 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5231 pa->pa_len, pa->pa_lstart);
5232 trace_ext4_mb_new_inode_pa(ac, pa);
5233
5234 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
5235 ext4_mb_use_inode_pa(ac, pa);
5236
5237 ei = EXT4_I(ac->ac_inode);
5238 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5239 if (!grp)
5240 return;
5241
5242 pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
5243 pa->pa_inode = ac->ac_inode;
5244
5245 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5246
5247 write_lock(pa->pa_node_lock.inode_lock);
5248 ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node);
5249 write_unlock(pa->pa_node_lock.inode_lock);
5250 atomic_inc(&ei->i_prealloc_active);
5251 }
5252
5253 /*
5254 * creates new preallocated space for locality group inodes belongs to
5255 */
5256 static noinline_for_stack void
5257 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
5258 {
5259 struct super_block *sb = ac->ac_sb;
5260 struct ext4_locality_group *lg;
5261 struct ext4_prealloc_space *pa;
5262 struct ext4_group_info *grp;
5263
5264 /* preallocate only when found space is larger then requested */
5265 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5266 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5267 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5268 BUG_ON(ac->ac_pa == NULL);
5269
5270 pa = ac->ac_pa;
5271
5272 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5273 pa->pa_lstart = pa->pa_pstart;
5274 pa->pa_len = ac->ac_b_ex.fe_len;
5275 pa->pa_free = pa->pa_len;
5276 spin_lock_init(&pa->pa_lock);
5277 INIT_LIST_HEAD(&pa->pa_node.lg_list);
5278 INIT_LIST_HEAD(&pa->pa_group_list);
5279 pa->pa_deleted = 0;
5280 pa->pa_type = MB_GROUP_PA;
5281
5282 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5283 pa->pa_len, pa->pa_lstart);
5284 trace_ext4_mb_new_group_pa(ac, pa);
5285
5286 ext4_mb_use_group_pa(ac, pa);
5287 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
5288
5289 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5290 if (!grp)
5291 return;
5292 lg = ac->ac_lg;
5293 BUG_ON(lg == NULL);
5294
5295 pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
5296 pa->pa_inode = NULL;
5297
5298 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5299
5300 /*
5301 * We will later add the new pa to the right bucket
5302 * after updating the pa_free in ext4_mb_release_context
5303 */
5304 }
5305
5306 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
5307 {
5308 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5309 ext4_mb_new_group_pa(ac);
5310 else
5311 ext4_mb_new_inode_pa(ac);
5312 }
5313
5314 /*
5315 * finds all unused blocks in on-disk bitmap, frees them in
5316 * in-core bitmap and buddy.
5317 * @pa must be unlinked from inode and group lists, so that
5318 * nobody else can find/use it.
5319 * the caller MUST hold group/inode locks.
5320 * TODO: optimize the case when there are no in-core structures yet
5321 */
5322 static noinline_for_stack void
5323 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
5324 struct ext4_prealloc_space *pa)
5325 {
5326 struct super_block *sb = e4b->bd_sb;
5327 struct ext4_sb_info *sbi = EXT4_SB(sb);
5328 unsigned int end;
5329 unsigned int next;
5330 ext4_group_t group;
5331 ext4_grpblk_t bit;
5332 unsigned long long grp_blk_start;
5333 int free = 0;
5334
5335 BUG_ON(pa->pa_deleted == 0);
5336 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5337 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
5338 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
5339 end = bit + pa->pa_len;
5340
5341 while (bit < end) {
5342 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
5343 if (bit >= end)
5344 break;
5345 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
5346 mb_debug(sb, "free preallocated %u/%u in group %u\n",
5347 (unsigned) ext4_group_first_block_no(sb, group) + bit,
5348 (unsigned) next - bit, (unsigned) group);
5349 free += next - bit;
5350
5351 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
5352 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
5353 EXT4_C2B(sbi, bit)),
5354 next - bit);
5355 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
5356 bit = next + 1;
5357 }
5358 if (free != pa->pa_free) {
5359 ext4_msg(e4b->bd_sb, KERN_CRIT,
5360 "pa %p: logic %lu, phys. %lu, len %d",
5361 pa, (unsigned long) pa->pa_lstart,
5362 (unsigned long) pa->pa_pstart,
5363 pa->pa_len);
5364 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
5365 free, pa->pa_free);
5366 /*
5367 * pa is already deleted so we use the value obtained
5368 * from the bitmap and continue.
5369 */
5370 }
5371 atomic_add(free, &sbi->s_mb_discarded);
5372 }
5373
5374 static noinline_for_stack void
5375 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
5376 struct ext4_prealloc_space *pa)
5377 {
5378 struct super_block *sb = e4b->bd_sb;
5379 ext4_group_t group;
5380 ext4_grpblk_t bit;
5381
5382 trace_ext4_mb_release_group_pa(sb, pa);
5383 BUG_ON(pa->pa_deleted == 0);
5384 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5385 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
5386 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
5387 e4b->bd_group, group, pa->pa_pstart);
5388 return;
5389 }
5390 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
5391 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
5392 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
5393 }
5394
5395 /*
5396 * releases all preallocations in given group
5397 *
5398 * first, we need to decide discard policy:
5399 * - when do we discard
5400 * 1) ENOSPC
5401 * - how many do we discard
5402 * 1) how many requested
5403 */
5404 static noinline_for_stack int
5405 ext4_mb_discard_group_preallocations(struct super_block *sb,
5406 ext4_group_t group, int *busy)
5407 {
5408 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
5409 struct buffer_head *bitmap_bh = NULL;
5410 struct ext4_prealloc_space *pa, *tmp;
5411 LIST_HEAD(list);
5412 struct ext4_buddy e4b;
5413 struct ext4_inode_info *ei;
5414 int err;
5415 int free = 0;
5416
5417 if (!grp)
5418 return 0;
5419 mb_debug(sb, "discard preallocation for group %u\n", group);
5420 if (list_empty(&grp->bb_prealloc_list))
5421 goto out_dbg;
5422
5423 bitmap_bh = ext4_read_block_bitmap(sb, group);
5424 if (IS_ERR(bitmap_bh)) {
5425 err = PTR_ERR(bitmap_bh);
5426 ext4_error_err(sb, -err,
5427 "Error %d reading block bitmap for %u",
5428 err, group);
5429 goto out_dbg;
5430 }
5431
5432 err = ext4_mb_load_buddy(sb, group, &e4b);
5433 if (err) {
5434 ext4_warning(sb, "Error %d loading buddy information for %u",
5435 err, group);
5436 put_bh(bitmap_bh);
5437 goto out_dbg;
5438 }
5439
5440 ext4_lock_group(sb, group);
5441 list_for_each_entry_safe(pa, tmp,
5442 &grp->bb_prealloc_list, pa_group_list) {
5443 spin_lock(&pa->pa_lock);
5444 if (atomic_read(&pa->pa_count)) {
5445 spin_unlock(&pa->pa_lock);
5446 *busy = 1;
5447 continue;
5448 }
5449 if (pa->pa_deleted) {
5450 spin_unlock(&pa->pa_lock);
5451 continue;
5452 }
5453
5454 /* seems this one can be freed ... */
5455 ext4_mb_mark_pa_deleted(sb, pa);
5456
5457 if (!free)
5458 this_cpu_inc(discard_pa_seq);
5459
5460 /* we can trust pa_free ... */
5461 free += pa->pa_free;
5462
5463 spin_unlock(&pa->pa_lock);
5464
5465 list_del(&pa->pa_group_list);
5466 list_add(&pa->u.pa_tmp_list, &list);
5467 }
5468
5469 /* now free all selected PAs */
5470 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5471
5472 /* remove from object (inode or locality group) */
5473 if (pa->pa_type == MB_GROUP_PA) {
5474 spin_lock(pa->pa_node_lock.lg_lock);
5475 list_del_rcu(&pa->pa_node.lg_list);
5476 spin_unlock(pa->pa_node_lock.lg_lock);
5477 } else {
5478 write_lock(pa->pa_node_lock.inode_lock);
5479 ei = EXT4_I(pa->pa_inode);
5480 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5481 write_unlock(pa->pa_node_lock.inode_lock);
5482 }
5483
5484 list_del(&pa->u.pa_tmp_list);
5485
5486 if (pa->pa_type == MB_GROUP_PA) {
5487 ext4_mb_release_group_pa(&e4b, pa);
5488 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5489 } else {
5490 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5491 ext4_mb_pa_free(pa);
5492 }
5493 }
5494
5495 ext4_unlock_group(sb, group);
5496 ext4_mb_unload_buddy(&e4b);
5497 put_bh(bitmap_bh);
5498 out_dbg:
5499 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
5500 free, group, grp->bb_free);
5501 return free;
5502 }
5503
5504 /*
5505 * releases all non-used preallocated blocks for given inode
5506 *
5507 * It's important to discard preallocations under i_data_sem
5508 * We don't want another block to be served from the prealloc
5509 * space when we are discarding the inode prealloc space.
5510 *
5511 * FIXME!! Make sure it is valid at all the call sites
5512 */
5513 void ext4_discard_preallocations(struct inode *inode)
5514 {
5515 struct ext4_inode_info *ei = EXT4_I(inode);
5516 struct super_block *sb = inode->i_sb;
5517 struct buffer_head *bitmap_bh = NULL;
5518 struct ext4_prealloc_space *pa, *tmp;
5519 ext4_group_t group = 0;
5520 LIST_HEAD(list);
5521 struct ext4_buddy e4b;
5522 struct rb_node *iter;
5523 int err;
5524
5525 if (!S_ISREG(inode->i_mode))
5526 return;
5527
5528 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
5529 return;
5530
5531 mb_debug(sb, "discard preallocation for inode %lu\n",
5532 inode->i_ino);
5533 trace_ext4_discard_preallocations(inode,
5534 atomic_read(&ei->i_prealloc_active));
5535
5536 repeat:
5537 /* first, collect all pa's in the inode */
5538 write_lock(&ei->i_prealloc_lock);
5539 for (iter = rb_first(&ei->i_prealloc_node); iter;
5540 iter = rb_next(iter)) {
5541 pa = rb_entry(iter, struct ext4_prealloc_space,
5542 pa_node.inode_node);
5543 BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
5544
5545 spin_lock(&pa->pa_lock);
5546 if (atomic_read(&pa->pa_count)) {
5547 /* this shouldn't happen often - nobody should
5548 * use preallocation while we're discarding it */
5549 spin_unlock(&pa->pa_lock);
5550 write_unlock(&ei->i_prealloc_lock);
5551 ext4_msg(sb, KERN_ERR,
5552 "uh-oh! used pa while discarding");
5553 WARN_ON(1);
5554 schedule_timeout_uninterruptible(HZ);
5555 goto repeat;
5556
5557 }
5558 if (pa->pa_deleted == 0) {
5559 ext4_mb_mark_pa_deleted(sb, pa);
5560 spin_unlock(&pa->pa_lock);
5561 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5562 list_add(&pa->u.pa_tmp_list, &list);
5563 continue;
5564 }
5565
5566 /* someone is deleting pa right now */
5567 spin_unlock(&pa->pa_lock);
5568 write_unlock(&ei->i_prealloc_lock);
5569
5570 /* we have to wait here because pa_deleted
5571 * doesn't mean pa is already unlinked from
5572 * the list. as we might be called from
5573 * ->clear_inode() the inode will get freed
5574 * and concurrent thread which is unlinking
5575 * pa from inode's list may access already
5576 * freed memory, bad-bad-bad */
5577
5578 /* XXX: if this happens too often, we can
5579 * add a flag to force wait only in case
5580 * of ->clear_inode(), but not in case of
5581 * regular truncate */
5582 schedule_timeout_uninterruptible(HZ);
5583 goto repeat;
5584 }
5585 write_unlock(&ei->i_prealloc_lock);
5586
5587 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5588 BUG_ON(pa->pa_type != MB_INODE_PA);
5589 group = ext4_get_group_number(sb, pa->pa_pstart);
5590
5591 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5592 GFP_NOFS|__GFP_NOFAIL);
5593 if (err) {
5594 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5595 err, group);
5596 continue;
5597 }
5598
5599 bitmap_bh = ext4_read_block_bitmap(sb, group);
5600 if (IS_ERR(bitmap_bh)) {
5601 err = PTR_ERR(bitmap_bh);
5602 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5603 err, group);
5604 ext4_mb_unload_buddy(&e4b);
5605 continue;
5606 }
5607
5608 ext4_lock_group(sb, group);
5609 list_del(&pa->pa_group_list);
5610 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5611 ext4_unlock_group(sb, group);
5612
5613 ext4_mb_unload_buddy(&e4b);
5614 put_bh(bitmap_bh);
5615
5616 list_del(&pa->u.pa_tmp_list);
5617 ext4_mb_pa_free(pa);
5618 }
5619 }
5620
5621 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5622 {
5623 struct ext4_prealloc_space *pa;
5624
5625 BUG_ON(ext4_pspace_cachep == NULL);
5626 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5627 if (!pa)
5628 return -ENOMEM;
5629 atomic_set(&pa->pa_count, 1);
5630 ac->ac_pa = pa;
5631 return 0;
5632 }
5633
5634 static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
5635 {
5636 struct ext4_prealloc_space *pa = ac->ac_pa;
5637
5638 BUG_ON(!pa);
5639 ac->ac_pa = NULL;
5640 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5641 /*
5642 * current function is only called due to an error or due to
5643 * len of found blocks < len of requested blocks hence the PA has not
5644 * been added to grp->bb_prealloc_list. So we don't need to lock it
5645 */
5646 pa->pa_deleted = 1;
5647 ext4_mb_pa_free(pa);
5648 }
5649
5650 #ifdef CONFIG_EXT4_DEBUG
5651 static inline void ext4_mb_show_pa(struct super_block *sb)
5652 {
5653 ext4_group_t i, ngroups;
5654
5655 if (ext4_forced_shutdown(sb))
5656 return;
5657
5658 ngroups = ext4_get_groups_count(sb);
5659 mb_debug(sb, "groups: ");
5660 for (i = 0; i < ngroups; i++) {
5661 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5662 struct ext4_prealloc_space *pa;
5663 ext4_grpblk_t start;
5664 struct list_head *cur;
5665
5666 if (!grp)
5667 continue;
5668 ext4_lock_group(sb, i);
5669 list_for_each(cur, &grp->bb_prealloc_list) {
5670 pa = list_entry(cur, struct ext4_prealloc_space,
5671 pa_group_list);
5672 spin_lock(&pa->pa_lock);
5673 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5674 NULL, &start);
5675 spin_unlock(&pa->pa_lock);
5676 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5677 pa->pa_len);
5678 }
5679 ext4_unlock_group(sb, i);
5680 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5681 grp->bb_fragments);
5682 }
5683 }
5684
5685 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5686 {
5687 struct super_block *sb = ac->ac_sb;
5688
5689 if (ext4_forced_shutdown(sb))
5690 return;
5691
5692 mb_debug(sb, "Can't allocate:"
5693 " Allocation context details:");
5694 mb_debug(sb, "status %u flags 0x%x",
5695 ac->ac_status, ac->ac_flags);
5696 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5697 "goal %lu/%lu/%lu@%lu, "
5698 "best %lu/%lu/%lu@%lu cr %d",
5699 (unsigned long)ac->ac_o_ex.fe_group,
5700 (unsigned long)ac->ac_o_ex.fe_start,
5701 (unsigned long)ac->ac_o_ex.fe_len,
5702 (unsigned long)ac->ac_o_ex.fe_logical,
5703 (unsigned long)ac->ac_g_ex.fe_group,
5704 (unsigned long)ac->ac_g_ex.fe_start,
5705 (unsigned long)ac->ac_g_ex.fe_len,
5706 (unsigned long)ac->ac_g_ex.fe_logical,
5707 (unsigned long)ac->ac_b_ex.fe_group,
5708 (unsigned long)ac->ac_b_ex.fe_start,
5709 (unsigned long)ac->ac_b_ex.fe_len,
5710 (unsigned long)ac->ac_b_ex.fe_logical,
5711 (int)ac->ac_criteria);
5712 mb_debug(sb, "%u found", ac->ac_found);
5713 mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no");
5714 if (ac->ac_pa)
5715 mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ?
5716 "group pa" : "inode pa");
5717 ext4_mb_show_pa(sb);
5718 }
5719 #else
5720 static inline void ext4_mb_show_pa(struct super_block *sb)
5721 {
5722 }
5723 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5724 {
5725 ext4_mb_show_pa(ac->ac_sb);
5726 }
5727 #endif
5728
5729 /*
5730 * We use locality group preallocation for small size file. The size of the
5731 * file is determined by the current size or the resulting size after
5732 * allocation which ever is larger
5733 *
5734 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5735 */
5736 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5737 {
5738 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5739 int bsbits = ac->ac_sb->s_blocksize_bits;
5740 loff_t size, isize;
5741 bool inode_pa_eligible, group_pa_eligible;
5742
5743 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5744 return;
5745
5746 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5747 return;
5748
5749 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5750 inode_pa_eligible = true;
5751 size = extent_logical_end(sbi, &ac->ac_o_ex);
5752 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5753 >> bsbits;
5754
5755 /* No point in using inode preallocation for closed files */
5756 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5757 !inode_is_open_for_write(ac->ac_inode))
5758 inode_pa_eligible = false;
5759
5760 size = max(size, isize);
5761 /* Don't use group allocation for large files */
5762 if (size > sbi->s_mb_stream_request)
5763 group_pa_eligible = false;
5764
5765 if (!group_pa_eligible) {
5766 if (inode_pa_eligible)
5767 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5768 else
5769 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5770 return;
5771 }
5772
5773 BUG_ON(ac->ac_lg != NULL);
5774 /*
5775 * locality group prealloc space are per cpu. The reason for having
5776 * per cpu locality group is to reduce the contention between block
5777 * request from multiple CPUs.
5778 */
5779 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5780
5781 /* we're going to use group allocation */
5782 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5783
5784 /* serialize all allocations in the group */
5785 mutex_lock(&ac->ac_lg->lg_mutex);
5786 }
5787
5788 static noinline_for_stack void
5789 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5790 struct ext4_allocation_request *ar)
5791 {
5792 struct super_block *sb = ar->inode->i_sb;
5793 struct ext4_sb_info *sbi = EXT4_SB(sb);
5794 struct ext4_super_block *es = sbi->s_es;
5795 ext4_group_t group;
5796 unsigned int len;
5797 ext4_fsblk_t goal;
5798 ext4_grpblk_t block;
5799
5800 /* we can't allocate > group size */
5801 len = ar->len;
5802
5803 /* just a dirty hack to filter too big requests */
5804 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5805 len = EXT4_CLUSTERS_PER_GROUP(sb);
5806
5807 /* start searching from the goal */
5808 goal = ar->goal;
5809 if (goal < le32_to_cpu(es->s_first_data_block) ||
5810 goal >= ext4_blocks_count(es))
5811 goal = le32_to_cpu(es->s_first_data_block);
5812 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5813
5814 /* set up allocation goals */
5815 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5816 ac->ac_status = AC_STATUS_CONTINUE;
5817 ac->ac_sb = sb;
5818 ac->ac_inode = ar->inode;
5819 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5820 ac->ac_o_ex.fe_group = group;
5821 ac->ac_o_ex.fe_start = block;
5822 ac->ac_o_ex.fe_len = len;
5823 ac->ac_g_ex = ac->ac_o_ex;
5824 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
5825 ac->ac_flags = ar->flags;
5826
5827 /* we have to define context: we'll work with a file or
5828 * locality group. this is a policy, actually */
5829 ext4_mb_group_or_file(ac);
5830
5831 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5832 "left: %u/%u, right %u/%u to %swritable\n",
5833 (unsigned) ar->len, (unsigned) ar->logical,
5834 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5835 (unsigned) ar->lleft, (unsigned) ar->pleft,
5836 (unsigned) ar->lright, (unsigned) ar->pright,
5837 inode_is_open_for_write(ar->inode) ? "" : "non-");
5838 }
5839
5840 static noinline_for_stack void
5841 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5842 struct ext4_locality_group *lg,
5843 int order, int total_entries)
5844 {
5845 ext4_group_t group = 0;
5846 struct ext4_buddy e4b;
5847 LIST_HEAD(discard_list);
5848 struct ext4_prealloc_space *pa, *tmp;
5849
5850 mb_debug(sb, "discard locality group preallocation\n");
5851
5852 spin_lock(&lg->lg_prealloc_lock);
5853 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5854 pa_node.lg_list,
5855 lockdep_is_held(&lg->lg_prealloc_lock)) {
5856 spin_lock(&pa->pa_lock);
5857 if (atomic_read(&pa->pa_count)) {
5858 /*
5859 * This is the pa that we just used
5860 * for block allocation. So don't
5861 * free that
5862 */
5863 spin_unlock(&pa->pa_lock);
5864 continue;
5865 }
5866 if (pa->pa_deleted) {
5867 spin_unlock(&pa->pa_lock);
5868 continue;
5869 }
5870 /* only lg prealloc space */
5871 BUG_ON(pa->pa_type != MB_GROUP_PA);
5872
5873 /* seems this one can be freed ... */
5874 ext4_mb_mark_pa_deleted(sb, pa);
5875 spin_unlock(&pa->pa_lock);
5876
5877 list_del_rcu(&pa->pa_node.lg_list);
5878 list_add(&pa->u.pa_tmp_list, &discard_list);
5879
5880 total_entries--;
5881 if (total_entries <= 5) {
5882 /*
5883 * we want to keep only 5 entries
5884 * allowing it to grow to 8. This
5885 * mak sure we don't call discard
5886 * soon for this list.
5887 */
5888 break;
5889 }
5890 }
5891 spin_unlock(&lg->lg_prealloc_lock);
5892
5893 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5894 int err;
5895
5896 group = ext4_get_group_number(sb, pa->pa_pstart);
5897 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5898 GFP_NOFS|__GFP_NOFAIL);
5899 if (err) {
5900 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5901 err, group);
5902 continue;
5903 }
5904 ext4_lock_group(sb, group);
5905 list_del(&pa->pa_group_list);
5906 ext4_mb_release_group_pa(&e4b, pa);
5907 ext4_unlock_group(sb, group);
5908
5909 ext4_mb_unload_buddy(&e4b);
5910 list_del(&pa->u.pa_tmp_list);
5911 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5912 }
5913 }
5914
5915 /*
5916 * We have incremented pa_count. So it cannot be freed at this
5917 * point. Also we hold lg_mutex. So no parallel allocation is
5918 * possible from this lg. That means pa_free cannot be updated.
5919 *
5920 * A parallel ext4_mb_discard_group_preallocations is possible.
5921 * which can cause the lg_prealloc_list to be updated.
5922 */
5923
5924 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5925 {
5926 int order, added = 0, lg_prealloc_count = 1;
5927 struct super_block *sb = ac->ac_sb;
5928 struct ext4_locality_group *lg = ac->ac_lg;
5929 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5930
5931 order = fls(pa->pa_free) - 1;
5932 if (order > PREALLOC_TB_SIZE - 1)
5933 /* The max size of hash table is PREALLOC_TB_SIZE */
5934 order = PREALLOC_TB_SIZE - 1;
5935 /* Add the prealloc space to lg */
5936 spin_lock(&lg->lg_prealloc_lock);
5937 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5938 pa_node.lg_list,
5939 lockdep_is_held(&lg->lg_prealloc_lock)) {
5940 spin_lock(&tmp_pa->pa_lock);
5941 if (tmp_pa->pa_deleted) {
5942 spin_unlock(&tmp_pa->pa_lock);
5943 continue;
5944 }
5945 if (!added && pa->pa_free < tmp_pa->pa_free) {
5946 /* Add to the tail of the previous entry */
5947 list_add_tail_rcu(&pa->pa_node.lg_list,
5948 &tmp_pa->pa_node.lg_list);
5949 added = 1;
5950 /*
5951 * we want to count the total
5952 * number of entries in the list
5953 */
5954 }
5955 spin_unlock(&tmp_pa->pa_lock);
5956 lg_prealloc_count++;
5957 }
5958 if (!added)
5959 list_add_tail_rcu(&pa->pa_node.lg_list,
5960 &lg->lg_prealloc_list[order]);
5961 spin_unlock(&lg->lg_prealloc_lock);
5962
5963 /* Now trim the list to be not more than 8 elements */
5964 if (lg_prealloc_count > 8)
5965 ext4_mb_discard_lg_preallocations(sb, lg,
5966 order, lg_prealloc_count);
5967 }
5968
5969 /*
5970 * release all resource we used in allocation
5971 */
5972 static void ext4_mb_release_context(struct ext4_allocation_context *ac)
5973 {
5974 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5975 struct ext4_prealloc_space *pa = ac->ac_pa;
5976 if (pa) {
5977 if (pa->pa_type == MB_GROUP_PA) {
5978 /* see comment in ext4_mb_use_group_pa() */
5979 spin_lock(&pa->pa_lock);
5980 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5981 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5982 pa->pa_free -= ac->ac_b_ex.fe_len;
5983 pa->pa_len -= ac->ac_b_ex.fe_len;
5984 spin_unlock(&pa->pa_lock);
5985
5986 /*
5987 * We want to add the pa to the right bucket.
5988 * Remove it from the list and while adding
5989 * make sure the list to which we are adding
5990 * doesn't grow big.
5991 */
5992 if (likely(pa->pa_free)) {
5993 spin_lock(pa->pa_node_lock.lg_lock);
5994 list_del_rcu(&pa->pa_node.lg_list);
5995 spin_unlock(pa->pa_node_lock.lg_lock);
5996 ext4_mb_add_n_trim(ac);
5997 }
5998 }
5999
6000 ext4_mb_put_pa(ac, ac->ac_sb, pa);
6001 }
6002 if (ac->ac_bitmap_folio)
6003 folio_put(ac->ac_bitmap_folio);
6004 if (ac->ac_buddy_folio)
6005 folio_put(ac->ac_buddy_folio);
6006 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
6007 mutex_unlock(&ac->ac_lg->lg_mutex);
6008 ext4_mb_collect_stats(ac);
6009 }
6010
6011 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
6012 {
6013 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
6014 int ret;
6015 int freed = 0, busy = 0;
6016 int retry = 0;
6017
6018 trace_ext4_mb_discard_preallocations(sb, needed);
6019
6020 if (needed == 0)
6021 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
6022 repeat:
6023 for (i = 0; i < ngroups && needed > 0; i++) {
6024 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
6025 freed += ret;
6026 needed -= ret;
6027 cond_resched();
6028 }
6029
6030 if (needed > 0 && busy && ++retry < 3) {
6031 busy = 0;
6032 goto repeat;
6033 }
6034
6035 return freed;
6036 }
6037
6038 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
6039 struct ext4_allocation_context *ac, u64 *seq)
6040 {
6041 int freed;
6042 u64 seq_retry = 0;
6043 bool ret = false;
6044
6045 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
6046 if (freed) {
6047 ret = true;
6048 goto out_dbg;
6049 }
6050 seq_retry = ext4_get_discard_pa_seq_sum();
6051 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
6052 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
6053 *seq = seq_retry;
6054 ret = true;
6055 }
6056
6057 out_dbg:
6058 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
6059 return ret;
6060 }
6061
6062 /*
6063 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
6064 * linearly starting at the goal block and also excludes the blocks which
6065 * are going to be in use after fast commit replay.
6066 */
6067 static ext4_fsblk_t
6068 ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp)
6069 {
6070 struct buffer_head *bitmap_bh;
6071 struct super_block *sb = ar->inode->i_sb;
6072 struct ext4_sb_info *sbi = EXT4_SB(sb);
6073 ext4_group_t group, nr;
6074 ext4_grpblk_t blkoff;
6075 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
6076 ext4_grpblk_t i = 0;
6077 ext4_fsblk_t goal, block;
6078 struct ext4_super_block *es = sbi->s_es;
6079
6080 goal = ar->goal;
6081 if (goal < le32_to_cpu(es->s_first_data_block) ||
6082 goal >= ext4_blocks_count(es))
6083 goal = le32_to_cpu(es->s_first_data_block);
6084
6085 ar->len = 0;
6086 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
6087 for (nr = ext4_get_groups_count(sb); nr > 0; nr--) {
6088 bitmap_bh = ext4_read_block_bitmap(sb, group);
6089 if (IS_ERR(bitmap_bh)) {
6090 *errp = PTR_ERR(bitmap_bh);
6091 pr_warn("Failed to read block bitmap\n");
6092 return 0;
6093 }
6094
6095 while (1) {
6096 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
6097 blkoff);
6098 if (i >= max)
6099 break;
6100 if (ext4_fc_replay_check_excluded(sb,
6101 ext4_group_first_block_no(sb, group) +
6102 EXT4_C2B(sbi, i))) {
6103 blkoff = i + 1;
6104 } else
6105 break;
6106 }
6107 brelse(bitmap_bh);
6108 if (i < max)
6109 break;
6110
6111 if (++group >= ext4_get_groups_count(sb))
6112 group = 0;
6113
6114 blkoff = 0;
6115 }
6116
6117 if (i >= max) {
6118 *errp = -ENOSPC;
6119 return 0;
6120 }
6121
6122 block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i);
6123 ext4_mb_mark_bb(sb, block, 1, true);
6124 ar->len = 1;
6125
6126 *errp = 0;
6127 return block;
6128 }
6129
6130 /*
6131 * Main entry point into mballoc to allocate blocks
6132 * it tries to use preallocation first, then falls back
6133 * to usual allocation
6134 */
6135 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
6136 struct ext4_allocation_request *ar, int *errp)
6137 {
6138 struct ext4_allocation_context *ac = NULL;
6139 struct ext4_sb_info *sbi;
6140 struct super_block *sb;
6141 ext4_fsblk_t block = 0;
6142 unsigned int inquota = 0;
6143 unsigned int reserv_clstrs = 0;
6144 int retries = 0;
6145 u64 seq;
6146
6147 might_sleep();
6148 sb = ar->inode->i_sb;
6149 sbi = EXT4_SB(sb);
6150
6151 trace_ext4_request_blocks(ar);
6152 if (sbi->s_mount_state & EXT4_FC_REPLAY)
6153 return ext4_mb_new_blocks_simple(ar, errp);
6154
6155 /* Allow to use superuser reservation for quota file */
6156 if (ext4_is_quota_file(ar->inode))
6157 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
6158
6159 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
6160 /* Without delayed allocation we need to verify
6161 * there is enough free blocks to do block allocation
6162 * and verify allocation doesn't exceed the quota limits.
6163 */
6164 while (ar->len &&
6165 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
6166
6167 /* let others to free the space */
6168 cond_resched();
6169 ar->len = ar->len >> 1;
6170 }
6171 if (!ar->len) {
6172 ext4_mb_show_pa(sb);
6173 *errp = -ENOSPC;
6174 return 0;
6175 }
6176 reserv_clstrs = ar->len;
6177 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
6178 dquot_alloc_block_nofail(ar->inode,
6179 EXT4_C2B(sbi, ar->len));
6180 } else {
6181 while (ar->len &&
6182 dquot_alloc_block(ar->inode,
6183 EXT4_C2B(sbi, ar->len))) {
6184
6185 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
6186 ar->len--;
6187 }
6188 }
6189 inquota = ar->len;
6190 if (ar->len == 0) {
6191 *errp = -EDQUOT;
6192 goto out;
6193 }
6194 }
6195
6196 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
6197 if (!ac) {
6198 ar->len = 0;
6199 *errp = -ENOMEM;
6200 goto out;
6201 }
6202
6203 ext4_mb_initialize_context(ac, ar);
6204
6205 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
6206 seq = this_cpu_read(discard_pa_seq);
6207 if (!ext4_mb_use_preallocated(ac)) {
6208 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
6209 ext4_mb_normalize_request(ac, ar);
6210
6211 *errp = ext4_mb_pa_alloc(ac);
6212 if (*errp)
6213 goto errout;
6214 repeat:
6215 /* allocate space in core */
6216 *errp = ext4_mb_regular_allocator(ac);
6217 /*
6218 * pa allocated above is added to grp->bb_prealloc_list only
6219 * when we were able to allocate some block i.e. when
6220 * ac->ac_status == AC_STATUS_FOUND.
6221 * And error from above mean ac->ac_status != AC_STATUS_FOUND
6222 * So we have to free this pa here itself.
6223 */
6224 if (*errp) {
6225 ext4_mb_pa_put_free(ac);
6226 ext4_discard_allocated_blocks(ac);
6227 goto errout;
6228 }
6229 if (ac->ac_status == AC_STATUS_FOUND &&
6230 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
6231 ext4_mb_pa_put_free(ac);
6232 }
6233 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
6234 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
6235 if (*errp) {
6236 ext4_discard_allocated_blocks(ac);
6237 goto errout;
6238 } else {
6239 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
6240 ar->len = ac->ac_b_ex.fe_len;
6241 }
6242 } else {
6243 if (++retries < 3 &&
6244 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
6245 goto repeat;
6246 /*
6247 * If block allocation fails then the pa allocated above
6248 * needs to be freed here itself.
6249 */
6250 ext4_mb_pa_put_free(ac);
6251 *errp = -ENOSPC;
6252 }
6253
6254 if (*errp) {
6255 errout:
6256 ac->ac_b_ex.fe_len = 0;
6257 ar->len = 0;
6258 ext4_mb_show_ac(ac);
6259 }
6260 ext4_mb_release_context(ac);
6261 kmem_cache_free(ext4_ac_cachep, ac);
6262 out:
6263 if (inquota && ar->len < inquota)
6264 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
6265 if (!ar->len) {
6266 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
6267 /* release all the reserved blocks if non delalloc */
6268 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
6269 reserv_clstrs);
6270 }
6271
6272 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
6273
6274 return block;
6275 }
6276
6277 /*
6278 * We can merge two free data extents only if the physical blocks
6279 * are contiguous, AND the extents were freed by the same transaction,
6280 * AND the blocks are associated with the same group.
6281 */
6282 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
6283 struct ext4_free_data *entry,
6284 struct ext4_free_data *new_entry,
6285 struct rb_root *entry_rb_root)
6286 {
6287 if ((entry->efd_tid != new_entry->efd_tid) ||
6288 (entry->efd_group != new_entry->efd_group))
6289 return;
6290 if (entry->efd_start_cluster + entry->efd_count ==
6291 new_entry->efd_start_cluster) {
6292 new_entry->efd_start_cluster = entry->efd_start_cluster;
6293 new_entry->efd_count += entry->efd_count;
6294 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
6295 entry->efd_start_cluster) {
6296 new_entry->efd_count += entry->efd_count;
6297 } else
6298 return;
6299 spin_lock(&sbi->s_md_lock);
6300 list_del(&entry->efd_list);
6301 spin_unlock(&sbi->s_md_lock);
6302 rb_erase(&entry->efd_node, entry_rb_root);
6303 kmem_cache_free(ext4_free_data_cachep, entry);
6304 }
6305
6306 static noinline_for_stack void
6307 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
6308 struct ext4_free_data *new_entry)
6309 {
6310 ext4_group_t group = e4b->bd_group;
6311 ext4_grpblk_t cluster;
6312 ext4_grpblk_t clusters = new_entry->efd_count;
6313 struct ext4_free_data *entry;
6314 struct ext4_group_info *db = e4b->bd_info;
6315 struct super_block *sb = e4b->bd_sb;
6316 struct ext4_sb_info *sbi = EXT4_SB(sb);
6317 struct rb_node **n = &db->bb_free_root.rb_node, *node;
6318 struct rb_node *parent = NULL, *new_node;
6319
6320 BUG_ON(!ext4_handle_valid(handle));
6321 BUG_ON(e4b->bd_bitmap_folio == NULL);
6322 BUG_ON(e4b->bd_buddy_folio == NULL);
6323
6324 new_node = &new_entry->efd_node;
6325 cluster = new_entry->efd_start_cluster;
6326
6327 if (!*n) {
6328 /* first free block exent. We need to
6329 protect buddy cache from being freed,
6330 * otherwise we'll refresh it from
6331 * on-disk bitmap and lose not-yet-available
6332 * blocks */
6333 folio_get(e4b->bd_buddy_folio);
6334 folio_get(e4b->bd_bitmap_folio);
6335 }
6336 while (*n) {
6337 parent = *n;
6338 entry = rb_entry(parent, struct ext4_free_data, efd_node);
6339 if (cluster < entry->efd_start_cluster)
6340 n = &(*n)->rb_left;
6341 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
6342 n = &(*n)->rb_right;
6343 else {
6344 ext4_grp_locked_error(sb, group, 0,
6345 ext4_group_first_block_no(sb, group) +
6346 EXT4_C2B(sbi, cluster),
6347 "Block already on to-be-freed list");
6348 kmem_cache_free(ext4_free_data_cachep, new_entry);
6349 return;
6350 }
6351 }
6352
6353 rb_link_node(new_node, parent, n);
6354 rb_insert_color(new_node, &db->bb_free_root);
6355
6356 /* Now try to see the extent can be merged to left and right */
6357 node = rb_prev(new_node);
6358 if (node) {
6359 entry = rb_entry(node, struct ext4_free_data, efd_node);
6360 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6361 &(db->bb_free_root));
6362 }
6363
6364 node = rb_next(new_node);
6365 if (node) {
6366 entry = rb_entry(node, struct ext4_free_data, efd_node);
6367 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6368 &(db->bb_free_root));
6369 }
6370
6371 spin_lock(&sbi->s_md_lock);
6372 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]);
6373 sbi->s_mb_free_pending += clusters;
6374 spin_unlock(&sbi->s_md_lock);
6375 }
6376
6377 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
6378 unsigned long count)
6379 {
6380 struct super_block *sb = inode->i_sb;
6381 ext4_group_t group;
6382 ext4_grpblk_t blkoff;
6383
6384 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
6385 ext4_mb_mark_context(NULL, sb, false, group, blkoff, count,
6386 EXT4_MB_BITMAP_MARKED_CHECK |
6387 EXT4_MB_SYNC_UPDATE,
6388 NULL);
6389 }
6390
6391 /**
6392 * ext4_mb_clear_bb() -- helper function for freeing blocks.
6393 * Used by ext4_free_blocks()
6394 * @handle: handle for this transaction
6395 * @inode: inode
6396 * @block: starting physical block to be freed
6397 * @count: number of blocks to be freed
6398 * @flags: flags used by ext4_free_blocks
6399 */
6400 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
6401 ext4_fsblk_t block, unsigned long count,
6402 int flags)
6403 {
6404 struct super_block *sb = inode->i_sb;
6405 struct ext4_group_info *grp;
6406 unsigned int overflow;
6407 ext4_grpblk_t bit;
6408 ext4_group_t block_group;
6409 struct ext4_sb_info *sbi;
6410 struct ext4_buddy e4b;
6411 unsigned int count_clusters;
6412 int err = 0;
6413 int mark_flags = 0;
6414 ext4_grpblk_t changed;
6415
6416 sbi = EXT4_SB(sb);
6417
6418 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6419 !ext4_inode_block_valid(inode, block, count)) {
6420 ext4_error(sb, "Freeing blocks in system zone - "
6421 "Block = %llu, count = %lu", block, count);
6422 /* err = 0. ext4_std_error should be a no op */
6423 goto error_out;
6424 }
6425 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6426
6427 do_more:
6428 overflow = 0;
6429 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6430
6431 grp = ext4_get_group_info(sb, block_group);
6432 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
6433 return;
6434
6435 /*
6436 * Check to see if we are freeing blocks across a group
6437 * boundary.
6438 */
6439 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
6440 overflow = EXT4_C2B(sbi, bit) + count -
6441 EXT4_BLOCKS_PER_GROUP(sb);
6442 count -= overflow;
6443 /* The range changed so it's no longer validated */
6444 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6445 }
6446 count_clusters = EXT4_NUM_B2C(sbi, count);
6447 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6448
6449 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6450 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6451 GFP_NOFS|__GFP_NOFAIL);
6452 if (err)
6453 goto error_out;
6454
6455 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6456 !ext4_inode_block_valid(inode, block, count)) {
6457 ext4_error(sb, "Freeing blocks in system zone - "
6458 "Block = %llu, count = %lu", block, count);
6459 /* err = 0. ext4_std_error should be a no op */
6460 goto error_clean;
6461 }
6462
6463 #ifdef AGGRESSIVE_CHECK
6464 mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK;
6465 #endif
6466 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6467 count_clusters, mark_flags, &changed);
6468
6469
6470 if (err && changed == 0)
6471 goto error_clean;
6472
6473 #ifdef AGGRESSIVE_CHECK
6474 BUG_ON(changed != count_clusters);
6475 #endif
6476
6477 /*
6478 * We need to make sure we don't reuse the freed block until after the
6479 * transaction is committed. We make an exception if the inode is to be
6480 * written in writeback mode since writeback mode has weak data
6481 * consistency guarantees.
6482 */
6483 if (ext4_handle_valid(handle) &&
6484 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6485 !ext4_should_writeback_data(inode))) {
6486 struct ext4_free_data *new_entry;
6487 /*
6488 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6489 * to fail.
6490 */
6491 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6492 GFP_NOFS|__GFP_NOFAIL);
6493 new_entry->efd_start_cluster = bit;
6494 new_entry->efd_group = block_group;
6495 new_entry->efd_count = count_clusters;
6496 new_entry->efd_tid = handle->h_transaction->t_tid;
6497
6498 ext4_lock_group(sb, block_group);
6499 ext4_mb_free_metadata(handle, &e4b, new_entry);
6500 } else {
6501 if (test_opt(sb, DISCARD)) {
6502 err = ext4_issue_discard(sb, block_group, bit,
6503 count_clusters);
6504 /*
6505 * Ignore EOPNOTSUPP error. This is consistent with
6506 * what happens when using journal.
6507 */
6508 if (err == -EOPNOTSUPP)
6509 err = 0;
6510 if (err)
6511 ext4_msg(sb, KERN_WARNING, "discard request in"
6512 " group:%u block:%d count:%lu failed"
6513 " with %d", block_group, bit, count,
6514 err);
6515 }
6516
6517 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6518
6519 ext4_lock_group(sb, block_group);
6520 mb_free_blocks(inode, &e4b, bit, count_clusters);
6521 }
6522
6523 ext4_unlock_group(sb, block_group);
6524
6525 /*
6526 * on a bigalloc file system, defer the s_freeclusters_counter
6527 * update to the caller (ext4_remove_space and friends) so they
6528 * can determine if a cluster freed here should be rereserved
6529 */
6530 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6531 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6532 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6533 percpu_counter_add(&sbi->s_freeclusters_counter,
6534 count_clusters);
6535 }
6536
6537 if (overflow && !err) {
6538 block += count;
6539 count = overflow;
6540 ext4_mb_unload_buddy(&e4b);
6541 /* The range changed so it's no longer validated */
6542 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6543 goto do_more;
6544 }
6545
6546 error_clean:
6547 ext4_mb_unload_buddy(&e4b);
6548 error_out:
6549 ext4_std_error(sb, err);
6550 }
6551
6552 /**
6553 * ext4_free_blocks() -- Free given blocks and update quota
6554 * @handle: handle for this transaction
6555 * @inode: inode
6556 * @bh: optional buffer of the block to be freed
6557 * @block: starting physical block to be freed
6558 * @count: number of blocks to be freed
6559 * @flags: flags used by ext4_free_blocks
6560 */
6561 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6562 struct buffer_head *bh, ext4_fsblk_t block,
6563 unsigned long count, int flags)
6564 {
6565 struct super_block *sb = inode->i_sb;
6566 unsigned int overflow;
6567 struct ext4_sb_info *sbi;
6568
6569 sbi = EXT4_SB(sb);
6570
6571 if (bh) {
6572 if (block)
6573 BUG_ON(block != bh->b_blocknr);
6574 else
6575 block = bh->b_blocknr;
6576 }
6577
6578 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6579 ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count));
6580 return;
6581 }
6582
6583 might_sleep();
6584
6585 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6586 !ext4_inode_block_valid(inode, block, count)) {
6587 ext4_error(sb, "Freeing blocks not in datazone - "
6588 "block = %llu, count = %lu", block, count);
6589 return;
6590 }
6591 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6592
6593 ext4_debug("freeing block %llu\n", block);
6594 trace_ext4_free_blocks(inode, block, count, flags);
6595
6596 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6597 BUG_ON(count > 1);
6598
6599 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6600 inode, bh, block);
6601 }
6602
6603 /*
6604 * If the extent to be freed does not begin on a cluster
6605 * boundary, we need to deal with partial clusters at the
6606 * beginning and end of the extent. Normally we will free
6607 * blocks at the beginning or the end unless we are explicitly
6608 * requested to avoid doing so.
6609 */
6610 overflow = EXT4_PBLK_COFF(sbi, block);
6611 if (overflow) {
6612 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6613 overflow = sbi->s_cluster_ratio - overflow;
6614 block += overflow;
6615 if (count > overflow)
6616 count -= overflow;
6617 else
6618 return;
6619 } else {
6620 block -= overflow;
6621 count += overflow;
6622 }
6623 /* The range changed so it's no longer validated */
6624 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6625 }
6626 overflow = EXT4_LBLK_COFF(sbi, count);
6627 if (overflow) {
6628 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6629 if (count > overflow)
6630 count -= overflow;
6631 else
6632 return;
6633 } else
6634 count += sbi->s_cluster_ratio - overflow;
6635 /* The range changed so it's no longer validated */
6636 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6637 }
6638
6639 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6640 int i;
6641 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6642
6643 for (i = 0; i < count; i++) {
6644 cond_resched();
6645 if (is_metadata)
6646 bh = sb_find_get_block(inode->i_sb, block + i);
6647 ext4_forget(handle, is_metadata, inode, bh, block + i);
6648 }
6649 }
6650
6651 ext4_mb_clear_bb(handle, inode, block, count, flags);
6652 }
6653
6654 /**
6655 * ext4_group_add_blocks() -- Add given blocks to an existing group
6656 * @handle: handle to this transaction
6657 * @sb: super block
6658 * @block: start physical block to add to the block group
6659 * @count: number of blocks to free
6660 *
6661 * This marks the blocks as free in the bitmap and buddy.
6662 */
6663 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6664 ext4_fsblk_t block, unsigned long count)
6665 {
6666 ext4_group_t block_group;
6667 ext4_grpblk_t bit;
6668 struct ext4_sb_info *sbi = EXT4_SB(sb);
6669 struct ext4_buddy e4b;
6670 int err = 0;
6671 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6672 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6673 unsigned long cluster_count = last_cluster - first_cluster + 1;
6674 ext4_grpblk_t changed;
6675
6676 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6677
6678 if (cluster_count == 0)
6679 return 0;
6680
6681 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6682 /*
6683 * Check to see if we are freeing blocks across a group
6684 * boundary.
6685 */
6686 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6687 ext4_warning(sb, "too many blocks added to group %u",
6688 block_group);
6689 err = -EINVAL;
6690 goto error_out;
6691 }
6692
6693 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6694 if (err)
6695 goto error_out;
6696
6697 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6698 ext4_error(sb, "Adding blocks in system zones - "
6699 "Block = %llu, count = %lu",
6700 block, count);
6701 err = -EINVAL;
6702 goto error_clean;
6703 }
6704
6705 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6706 cluster_count, EXT4_MB_BITMAP_MARKED_CHECK,
6707 &changed);
6708 if (err && changed == 0)
6709 goto error_clean;
6710
6711 if (changed != cluster_count)
6712 ext4_error(sb, "bit already cleared in group %u", block_group);
6713
6714 ext4_lock_group(sb, block_group);
6715 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6716 ext4_unlock_group(sb, block_group);
6717 percpu_counter_add(&sbi->s_freeclusters_counter,
6718 changed);
6719
6720 error_clean:
6721 ext4_mb_unload_buddy(&e4b);
6722 error_out:
6723 ext4_std_error(sb, err);
6724 return err;
6725 }
6726
6727 /**
6728 * ext4_trim_extent -- function to TRIM one single free extent in the group
6729 * @sb: super block for the file system
6730 * @start: starting block of the free extent in the alloc. group
6731 * @count: number of blocks to TRIM
6732 * @e4b: ext4 buddy for the group
6733 *
6734 * Trim "count" blocks starting at "start" in the "group". To assure that no
6735 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6736 * be called with under the group lock.
6737 */
6738 static int ext4_trim_extent(struct super_block *sb,
6739 int start, int count, struct ext4_buddy *e4b)
6740 __releases(bitlock)
6741 __acquires(bitlock)
6742 {
6743 struct ext4_free_extent ex;
6744 ext4_group_t group = e4b->bd_group;
6745 int ret = 0;
6746
6747 trace_ext4_trim_extent(sb, group, start, count);
6748
6749 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6750
6751 ex.fe_start = start;
6752 ex.fe_group = group;
6753 ex.fe_len = count;
6754
6755 /*
6756 * Mark blocks used, so no one can reuse them while
6757 * being trimmed.
6758 */
6759 mb_mark_used(e4b, &ex);
6760 ext4_unlock_group(sb, group);
6761 ret = ext4_issue_discard(sb, group, start, count);
6762 ext4_lock_group(sb, group);
6763 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6764 return ret;
6765 }
6766
6767 static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb,
6768 ext4_group_t grp)
6769 {
6770 unsigned long nr_clusters_in_group;
6771
6772 if (grp < (ext4_get_groups_count(sb) - 1))
6773 nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb);
6774 else
6775 nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) -
6776 ext4_group_first_block_no(sb, grp))
6777 >> EXT4_CLUSTER_BITS(sb);
6778
6779 return nr_clusters_in_group - 1;
6780 }
6781
6782 static bool ext4_trim_interrupted(void)
6783 {
6784 return fatal_signal_pending(current) || freezing(current);
6785 }
6786
6787 static int ext4_try_to_trim_range(struct super_block *sb,
6788 struct ext4_buddy *e4b, ext4_grpblk_t start,
6789 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6790 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6791 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6792 {
6793 ext4_grpblk_t next, count, free_count, last, origin_start;
6794 bool set_trimmed = false;
6795 void *bitmap;
6796
6797 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
6798 return 0;
6799
6800 last = ext4_last_grp_cluster(sb, e4b->bd_group);
6801 bitmap = e4b->bd_bitmap;
6802 if (start == 0 && max >= last)
6803 set_trimmed = true;
6804 origin_start = start;
6805 start = max(e4b->bd_info->bb_first_free, start);
6806 count = 0;
6807 free_count = 0;
6808
6809 while (start <= max) {
6810 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6811 if (start > max)
6812 break;
6813
6814 next = mb_find_next_bit(bitmap, last + 1, start);
6815 if (origin_start == 0 && next >= last)
6816 set_trimmed = true;
6817
6818 if ((next - start) >= minblocks) {
6819 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6820
6821 if (ret && ret != -EOPNOTSUPP)
6822 return count;
6823 count += next - start;
6824 }
6825 free_count += next - start;
6826 start = next + 1;
6827
6828 if (ext4_trim_interrupted())
6829 return count;
6830
6831 if (need_resched()) {
6832 ext4_unlock_group(sb, e4b->bd_group);
6833 cond_resched();
6834 ext4_lock_group(sb, e4b->bd_group);
6835 }
6836
6837 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6838 break;
6839 }
6840
6841 if (set_trimmed)
6842 EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info);
6843
6844 return count;
6845 }
6846
6847 /**
6848 * ext4_trim_all_free -- function to trim all free space in alloc. group
6849 * @sb: super block for file system
6850 * @group: group to be trimmed
6851 * @start: first group block to examine
6852 * @max: last group block to examine
6853 * @minblocks: minimum extent block count
6854 *
6855 * ext4_trim_all_free walks through group's block bitmap searching for free
6856 * extents. When the free extent is found, mark it as used in group buddy
6857 * bitmap. Then issue a TRIM command on this extent and free the extent in
6858 * the group buddy bitmap.
6859 */
6860 static ext4_grpblk_t
6861 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6862 ext4_grpblk_t start, ext4_grpblk_t max,
6863 ext4_grpblk_t minblocks)
6864 {
6865 struct ext4_buddy e4b;
6866 int ret;
6867
6868 trace_ext4_trim_all_free(sb, group, start, max);
6869
6870 ret = ext4_mb_load_buddy(sb, group, &e4b);
6871 if (ret) {
6872 ext4_warning(sb, "Error %d loading buddy information for %u",
6873 ret, group);
6874 return ret;
6875 }
6876
6877 ext4_lock_group(sb, group);
6878
6879 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6880 minblocks < EXT4_SB(sb)->s_last_trim_minblks)
6881 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6882 else
6883 ret = 0;
6884
6885 ext4_unlock_group(sb, group);
6886 ext4_mb_unload_buddy(&e4b);
6887
6888 ext4_debug("trimmed %d blocks in the group %d\n",
6889 ret, group);
6890
6891 return ret;
6892 }
6893
6894 /**
6895 * ext4_trim_fs() -- trim ioctl handle function
6896 * @sb: superblock for filesystem
6897 * @range: fstrim_range structure
6898 *
6899 * start: First Byte to trim
6900 * len: number of Bytes to trim from start
6901 * minlen: minimum extent length in Bytes
6902 * ext4_trim_fs goes through all allocation groups containing Bytes from
6903 * start to start+len. For each such a group ext4_trim_all_free function
6904 * is invoked to trim all free space.
6905 */
6906 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6907 {
6908 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6909 struct ext4_group_info *grp;
6910 ext4_group_t group, first_group, last_group;
6911 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6912 uint64_t start, end, minlen, trimmed = 0;
6913 ext4_fsblk_t first_data_blk =
6914 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6915 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6916 int ret = 0;
6917
6918 start = range->start >> sb->s_blocksize_bits;
6919 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6920 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6921 range->minlen >> sb->s_blocksize_bits);
6922
6923 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6924 start >= max_blks ||
6925 range->len < sb->s_blocksize)
6926 return -EINVAL;
6927 /* No point to try to trim less than discard granularity */
6928 if (range->minlen < discard_granularity) {
6929 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6930 discard_granularity >> sb->s_blocksize_bits);
6931 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6932 goto out;
6933 }
6934 if (end >= max_blks - 1)
6935 end = max_blks - 1;
6936 if (end <= first_data_blk)
6937 goto out;
6938 if (start < first_data_blk)
6939 start = first_data_blk;
6940
6941 /* Determine first and last group to examine based on start and end */
6942 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6943 &first_group, &first_cluster);
6944 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6945 &last_group, &last_cluster);
6946
6947 /* end now represents the last cluster to discard in this group */
6948 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6949
6950 for (group = first_group; group <= last_group; group++) {
6951 if (ext4_trim_interrupted())
6952 break;
6953 grp = ext4_get_group_info(sb, group);
6954 if (!grp)
6955 continue;
6956 /* We only do this if the grp has never been initialized */
6957 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6958 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6959 if (ret)
6960 break;
6961 }
6962
6963 /*
6964 * For all the groups except the last one, last cluster will
6965 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6966 * change it for the last group, note that last_cluster is
6967 * already computed earlier by ext4_get_group_no_and_offset()
6968 */
6969 if (group == last_group)
6970 end = last_cluster;
6971 if (grp->bb_free >= minlen) {
6972 cnt = ext4_trim_all_free(sb, group, first_cluster,
6973 end, minlen);
6974 if (cnt < 0) {
6975 ret = cnt;
6976 break;
6977 }
6978 trimmed += cnt;
6979 }
6980
6981 /*
6982 * For every group except the first one, we are sure
6983 * that the first cluster to discard will be cluster #0.
6984 */
6985 first_cluster = 0;
6986 }
6987
6988 if (!ret)
6989 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6990
6991 out:
6992 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6993 return ret;
6994 }
6995
6996 /* Iterate all the free extents in the group. */
6997 int
6998 ext4_mballoc_query_range(
6999 struct super_block *sb,
7000 ext4_group_t group,
7001 ext4_grpblk_t start,
7002 ext4_grpblk_t end,
7003 ext4_mballoc_query_range_fn formatter,
7004 void *priv)
7005 {
7006 void *bitmap;
7007 ext4_grpblk_t next;
7008 struct ext4_buddy e4b;
7009 int error;
7010
7011 error = ext4_mb_load_buddy(sb, group, &e4b);
7012 if (error)
7013 return error;
7014 bitmap = e4b.bd_bitmap;
7015
7016 ext4_lock_group(sb, group);
7017
7018 start = max(e4b.bd_info->bb_first_free, start);
7019 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
7020 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
7021
7022 while (start <= end) {
7023 start = mb_find_next_zero_bit(bitmap, end + 1, start);
7024 if (start > end)
7025 break;
7026 next = mb_find_next_bit(bitmap, end + 1, start);
7027
7028 ext4_unlock_group(sb, group);
7029 error = formatter(sb, group, start, next - start, priv);
7030 if (error)
7031 goto out_unload;
7032 ext4_lock_group(sb, group);
7033
7034 start = next + 1;
7035 }
7036
7037 ext4_unlock_group(sb, group);
7038 out_unload:
7039 ext4_mb_unload_buddy(&e4b);
7040
7041 return error;
7042 }
7043
7044 #ifdef CONFIG_EXT4_KUNIT_TESTS
7045 #include "mballoc-test.c"
7046 #endif
7047
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.