aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--fs/ext4/mballoc.c296
-rw-r--r--fs/ext4/mballoc.h304
2 files changed, 305 insertions, 295 deletions
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index d4ae948606e8..11e1fd59acbd 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -21,21 +21,7 @@
21 * mballoc.c contains the multiblocks allocation routines 21 * mballoc.c contains the multiblocks allocation routines
22 */ 22 */
23 23
24#include <linux/time.h> 24#include "mballoc.h"
25#include <linux/fs.h>
26#include <linux/namei.h>
27#include <linux/quotaops.h>
28#include <linux/buffer_head.h>
29#include <linux/module.h>
30#include <linux/swap.h>
31#include <linux/proc_fs.h>
32#include <linux/pagemap.h>
33#include <linux/seq_file.h>
34#include <linux/version.h>
35#include "ext4_jbd2.h"
36#include "ext4.h"
37#include "group.h"
38
39/* 25/*
40 * MUSTDO: 26 * MUSTDO:
41 * - test ext4_ext_search_left() and ext4_ext_search_right() 27 * - test ext4_ext_search_left() and ext4_ext_search_right()
@@ -345,286 +331,6 @@
345 * 331 *
346 */ 332 */
347 333
348/*
349 * with AGGRESSIVE_CHECK allocator runs consistency checks over
350 * structures. these checks slow things down a lot
351 */
352#define AGGRESSIVE_CHECK__
353
354/*
355 * with DOUBLE_CHECK defined mballoc creates persistent in-core
356 * bitmaps, maintains and uses them to check for double allocations
357 */
358#define DOUBLE_CHECK__
359
360/*
361 */
362#define MB_DEBUG__
363#ifdef MB_DEBUG
364#define mb_debug(fmt, a...) printk(fmt, ##a)
365#else
366#define mb_debug(fmt, a...)
367#endif
368
369/*
370 * with EXT4_MB_HISTORY mballoc stores last N allocations in memory
371 * and you can monitor it in /proc/fs/ext4/<dev>/mb_history
372 */
373#define EXT4_MB_HISTORY
374#define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
375#define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
376#define EXT4_MB_HISTORY_DISCARD 4 /* preallocation discarded */
377#define EXT4_MB_HISTORY_FREE 8 /* free */
378
379#define EXT4_MB_HISTORY_DEFAULT (EXT4_MB_HISTORY_ALLOC | \
380 EXT4_MB_HISTORY_PREALLOC)
381
382/*
383 * How long mballoc can look for a best extent (in found extents)
384 */
385#define MB_DEFAULT_MAX_TO_SCAN 200
386
387/*
388 * How long mballoc must look for a best extent
389 */
390#define MB_DEFAULT_MIN_TO_SCAN 10
391
392/*
393 * How many groups mballoc will scan looking for the best chunk
394 */
395#define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
396
397/*
398 * with 'ext4_mb_stats' allocator will collect stats that will be
399 * shown at umount. The collecting costs though!
400 */
401#define MB_DEFAULT_STATS 1
402
403/*
404 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
405 * by the stream allocator, which purpose is to pack requests
406 * as close each to other as possible to produce smooth I/O traffic
407 * We use locality group prealloc space for stream request.
408 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
409 */
410#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
411
412/*
413 * for which requests use 2^N search using buddies
414 */
415#define MB_DEFAULT_ORDER2_REQS 2
416
417/*
418 * default group prealloc size 512 blocks
419 */
420#define MB_DEFAULT_GROUP_PREALLOC 512
421
422static struct kmem_cache *ext4_pspace_cachep;
423static struct kmem_cache *ext4_ac_cachep;
424
425#ifdef EXT4_BB_MAX_BLOCKS
426#undef EXT4_BB_MAX_BLOCKS
427#endif
428#define EXT4_BB_MAX_BLOCKS 30
429
430struct ext4_free_metadata {
431 ext4_group_t group;
432 unsigned short num;
433 ext4_grpblk_t blocks[EXT4_BB_MAX_BLOCKS];
434 struct list_head list;
435};
436
437struct ext4_group_info {
438 unsigned long bb_state;
439 unsigned long bb_tid;
440 struct ext4_free_metadata *bb_md_cur;
441 unsigned short bb_first_free;
442 unsigned short bb_free;
443 unsigned short bb_fragments;
444 struct list_head bb_prealloc_list;
445#ifdef DOUBLE_CHECK
446 void *bb_bitmap;
447#endif
448 unsigned short bb_counters[];
449};
450
451#define EXT4_GROUP_INFO_NEED_INIT_BIT 0
452#define EXT4_GROUP_INFO_LOCKED_BIT 1
453
454#define EXT4_MB_GRP_NEED_INIT(grp) \
455 (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
456
457
458struct ext4_prealloc_space {
459 struct list_head pa_inode_list;
460 struct list_head pa_group_list;
461 union {
462 struct list_head pa_tmp_list;
463 struct rcu_head pa_rcu;
464 } u;
465 spinlock_t pa_lock;
466 atomic_t pa_count;
467 unsigned pa_deleted;
468 ext4_fsblk_t pa_pstart; /* phys. block */
469 ext4_lblk_t pa_lstart; /* log. block */
470 unsigned short pa_len; /* len of preallocated chunk */
471 unsigned short pa_free; /* how many blocks are free */
472 unsigned short pa_linear; /* consumed in one direction
473 * strictly, for grp prealloc */
474 spinlock_t *pa_obj_lock;
475 struct inode *pa_inode; /* hack, for history only */
476};
477
478
479struct ext4_free_extent {
480 ext4_lblk_t fe_logical;
481 ext4_grpblk_t fe_start;
482 ext4_group_t fe_group;
483 int fe_len;
484};
485
486/*
487 * Locality group:
488 * we try to group all related changes together
489 * so that writeback can flush/allocate them together as well
490 */
491struct ext4_locality_group {
492 /* for allocator */
493 struct mutex lg_mutex; /* to serialize allocates */
494 struct list_head lg_prealloc_list;/* list of preallocations */
495 spinlock_t lg_prealloc_lock;
496};
497
498struct ext4_allocation_context {
499 struct inode *ac_inode;
500 struct super_block *ac_sb;
501
502 /* original request */
503 struct ext4_free_extent ac_o_ex;
504
505 /* goal request (after normalization) */
506 struct ext4_free_extent ac_g_ex;
507
508 /* the best found extent */
509 struct ext4_free_extent ac_b_ex;
510
511 /* copy of the bext found extent taken before preallocation efforts */
512 struct ext4_free_extent ac_f_ex;
513
514 /* number of iterations done. we have to track to limit searching */
515 unsigned long ac_ex_scanned;
516 __u16 ac_groups_scanned;
517 __u16 ac_found;
518 __u16 ac_tail;
519 __u16 ac_buddy;
520 __u16 ac_flags; /* allocation hints */
521 __u8 ac_status;
522 __u8 ac_criteria;
523 __u8 ac_repeats;
524 __u8 ac_2order; /* if request is to allocate 2^N blocks and
525 * N > 0, the field stores N, otherwise 0 */
526 __u8 ac_op; /* operation, for history only */
527 struct page *ac_bitmap_page;
528 struct page *ac_buddy_page;
529 struct ext4_prealloc_space *ac_pa;
530 struct ext4_locality_group *ac_lg;
531};
532
533#define AC_STATUS_CONTINUE 1
534#define AC_STATUS_FOUND 2
535#define AC_STATUS_BREAK 3
536
537struct ext4_mb_history {
538 struct ext4_free_extent orig; /* orig allocation */
539 struct ext4_free_extent goal; /* goal allocation */
540 struct ext4_free_extent result; /* result allocation */
541 unsigned pid;
542 unsigned ino;
543 __u16 found; /* how many extents have been found */
544 __u16 groups; /* how many groups have been scanned */
545 __u16 tail; /* what tail broke some buddy */
546 __u16 buddy; /* buddy the tail ^^^ broke */
547 __u16 flags;
548 __u8 cr:3; /* which phase the result extent was found at */
549 __u8 op:4;
550 __u8 merged:1;
551};
552
553struct ext4_buddy {
554 struct page *bd_buddy_page;
555 void *bd_buddy;
556 struct page *bd_bitmap_page;
557 void *bd_bitmap;
558 struct ext4_group_info *bd_info;
559 struct super_block *bd_sb;
560 __u16 bd_blkbits;
561 ext4_group_t bd_group;
562};
563#define EXT4_MB_BITMAP(e4b) ((e4b)->bd_bitmap)
564#define EXT4_MB_BUDDY(e4b) ((e4b)->bd_buddy)
565
566#ifndef EXT4_MB_HISTORY
567static inline void ext4_mb_store_history(struct ext4_allocation_context *ac)
568{
569 return;
570}
571#else
572static void ext4_mb_store_history(struct ext4_allocation_context *ac);
573#endif
574
575#define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
576
577static struct proc_dir_entry *proc_root_ext4;
578struct buffer_head *read_block_bitmap(struct super_block *, ext4_group_t);
579
580static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
581 ext4_group_t group);
582static void ext4_mb_poll_new_transaction(struct super_block *, handle_t *);
583static void ext4_mb_free_committed_blocks(struct super_block *);
584static void ext4_mb_return_to_preallocation(struct inode *inode,
585 struct ext4_buddy *e4b, sector_t block,
586 int count);
587static void ext4_mb_put_pa(struct ext4_allocation_context *,
588 struct super_block *, struct ext4_prealloc_space *pa);
589static int ext4_mb_init_per_dev_proc(struct super_block *sb);
590static int ext4_mb_destroy_per_dev_proc(struct super_block *sb);
591
592
593static inline void ext4_lock_group(struct super_block *sb, ext4_group_t group)
594{
595 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
596
597 bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
598}
599
600static inline void ext4_unlock_group(struct super_block *sb,
601 ext4_group_t group)
602{
603 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
604
605 bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
606}
607
608static inline int ext4_is_group_locked(struct super_block *sb,
609 ext4_group_t group)
610{
611 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
612
613 return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT,
614 &(grinfo->bb_state));
615}
616
617static ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
618 struct ext4_free_extent *fex)
619{
620 ext4_fsblk_t block;
621
622 block = (ext4_fsblk_t) fex->fe_group * EXT4_BLOCKS_PER_GROUP(sb)
623 + fex->fe_start
624 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
625 return block;
626}
627
628static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 334static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
629{ 335{
630#if BITS_PER_LONG == 64 336#if BITS_PER_LONG == 64
diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
new file mode 100644
index 000000000000..bfe6add46bcf
--- /dev/null
+++ b/fs/ext4/mballoc.h
@@ -0,0 +1,304 @@
1/*
2 * fs/ext4/mballoc.h
3 *
4 * Written by: Alex Tomas <alex@clusterfs.com>
5 *
6 */
7#ifndef _EXT4_MBALLOC_H
8#define _EXT4_MBALLOC_H
9
10#include <linux/time.h>
11#include <linux/fs.h>
12#include <linux/namei.h>
13#include <linux/quotaops.h>
14#include <linux/buffer_head.h>
15#include <linux/module.h>
16#include <linux/swap.h>
17#include <linux/proc_fs.h>
18#include <linux/pagemap.h>
19#include <linux/seq_file.h>
20#include <linux/version.h>
21#include "ext4_jbd2.h"
22#include "ext4.h"
23#include "group.h"
24
25/*
26 * with AGGRESSIVE_CHECK allocator runs consistency checks over
27 * structures. these checks slow things down a lot
28 */
29#define AGGRESSIVE_CHECK__
30
31/*
32 * with DOUBLE_CHECK defined mballoc creates persistent in-core
33 * bitmaps, maintains and uses them to check for double allocations
34 */
35#define DOUBLE_CHECK__
36
37/*
38 */
39#define MB_DEBUG__
40#ifdef MB_DEBUG
41#define mb_debug(fmt, a...) printk(fmt, ##a)
42#else
43#define mb_debug(fmt, a...)
44#endif
45
46/*
47 * with EXT4_MB_HISTORY mballoc stores last N allocations in memory
48 * and you can monitor it in /proc/fs/ext4/<dev>/mb_history
49 */
50#define EXT4_MB_HISTORY
51#define EXT4_MB_HISTORY_ALLOC 1 /* allocation */
52#define EXT4_MB_HISTORY_PREALLOC 2 /* preallocated blocks used */
53#define EXT4_MB_HISTORY_DISCARD 4 /* preallocation discarded */
54#define EXT4_MB_HISTORY_FREE 8 /* free */
55
56#define EXT4_MB_HISTORY_DEFAULT (EXT4_MB_HISTORY_ALLOC | \
57 EXT4_MB_HISTORY_PREALLOC)
58
59/*
60 * How long mballoc can look for a best extent (in found extents)
61 */
62#define MB_DEFAULT_MAX_TO_SCAN 200
63
64/*
65 * How long mballoc must look for a best extent
66 */
67#define MB_DEFAULT_MIN_TO_SCAN 10
68
69/*
70 * How many groups mballoc will scan looking for the best chunk
71 */
72#define MB_DEFAULT_MAX_GROUPS_TO_SCAN 5
73
74/*
75 * with 'ext4_mb_stats' allocator will collect stats that will be
76 * shown at umount. The collecting costs though!
77 */
78#define MB_DEFAULT_STATS 1
79
80/*
81 * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
82 * by the stream allocator, which purpose is to pack requests
83 * as close each to other as possible to produce smooth I/O traffic
84 * We use locality group prealloc space for stream request.
85 * We can tune the same via /proc/fs/ext4/<parition>/stream_req
86 */
87#define MB_DEFAULT_STREAM_THRESHOLD 16 /* 64K */
88
89/*
90 * for which requests use 2^N search using buddies
91 */
92#define MB_DEFAULT_ORDER2_REQS 2
93
94/*
95 * default group prealloc size 512 blocks
96 */
97#define MB_DEFAULT_GROUP_PREALLOC 512
98
99static struct kmem_cache *ext4_pspace_cachep;
100static struct kmem_cache *ext4_ac_cachep;
101
102#ifdef EXT4_BB_MAX_BLOCKS
103#undef EXT4_BB_MAX_BLOCKS
104#endif
105#define EXT4_BB_MAX_BLOCKS 30
106
107struct ext4_free_metadata {
108 ext4_group_t group;
109 unsigned short num;
110 ext4_grpblk_t blocks[EXT4_BB_MAX_BLOCKS];
111 struct list_head list;
112};
113
114struct ext4_group_info {
115 unsigned long bb_state;
116 unsigned long bb_tid;
117 struct ext4_free_metadata *bb_md_cur;
118 unsigned short bb_first_free;
119 unsigned short bb_free;
120 unsigned short bb_fragments;
121 struct list_head bb_prealloc_list;
122#ifdef DOUBLE_CHECK
123 void *bb_bitmap;
124#endif
125 unsigned short bb_counters[];
126};
127
128#define EXT4_GROUP_INFO_NEED_INIT_BIT 0
129#define EXT4_GROUP_INFO_LOCKED_BIT 1
130
131#define EXT4_MB_GRP_NEED_INIT(grp) \
132 (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
133
134
135struct ext4_prealloc_space {
136 struct list_head pa_inode_list;
137 struct list_head pa_group_list;
138 union {
139 struct list_head pa_tmp_list;
140 struct rcu_head pa_rcu;
141 } u;
142 spinlock_t pa_lock;
143 atomic_t pa_count;
144 unsigned pa_deleted;
145 ext4_fsblk_t pa_pstart; /* phys. block */
146 ext4_lblk_t pa_lstart; /* log. block */
147 unsigned short pa_len; /* len of preallocated chunk */
148 unsigned short pa_free; /* how many blocks are free */
149 unsigned short pa_linear; /* consumed in one direction
150 * strictly, for grp prealloc */
151 spinlock_t *pa_obj_lock;
152 struct inode *pa_inode; /* hack, for history only */
153};
154
155
156struct ext4_free_extent {
157 ext4_lblk_t fe_logical;
158 ext4_grpblk_t fe_start;
159 ext4_group_t fe_group;
160 int fe_len;
161};
162
163/*
164 * Locality group:
165 * we try to group all related changes together
166 * so that writeback can flush/allocate them together as well
167 */
168struct ext4_locality_group {
169 /* for allocator */
170 struct mutex lg_mutex; /* to serialize allocates */
171 struct list_head lg_prealloc_list;/* list of preallocations */
172 spinlock_t lg_prealloc_lock;
173};
174
175struct ext4_allocation_context {
176 struct inode *ac_inode;
177 struct super_block *ac_sb;
178
179 /* original request */
180 struct ext4_free_extent ac_o_ex;
181
182 /* goal request (after normalization) */
183 struct ext4_free_extent ac_g_ex;
184
185 /* the best found extent */
186 struct ext4_free_extent ac_b_ex;
187
188 /* copy of the bext found extent taken before preallocation efforts */
189 struct ext4_free_extent ac_f_ex;
190
191 /* number of iterations done. we have to track to limit searching */
192 unsigned long ac_ex_scanned;
193 __u16 ac_groups_scanned;
194 __u16 ac_found;
195 __u16 ac_tail;
196 __u16 ac_buddy;
197 __u16 ac_flags; /* allocation hints */
198 __u8 ac_status;
199 __u8 ac_criteria;
200 __u8 ac_repeats;
201 __u8 ac_2order; /* if request is to allocate 2^N blocks and
202 * N > 0, the field stores N, otherwise 0 */
203 __u8 ac_op; /* operation, for history only */
204 struct page *ac_bitmap_page;
205 struct page *ac_buddy_page;
206 struct ext4_prealloc_space *ac_pa;
207 struct ext4_locality_group *ac_lg;
208};
209
210#define AC_STATUS_CONTINUE 1
211#define AC_STATUS_FOUND 2
212#define AC_STATUS_BREAK 3
213
214struct ext4_mb_history {
215 struct ext4_free_extent orig; /* orig allocation */
216 struct ext4_free_extent goal; /* goal allocation */
217 struct ext4_free_extent result; /* result allocation */
218 unsigned pid;
219 unsigned ino;
220 __u16 found; /* how many extents have been found */
221 __u16 groups; /* how many groups have been scanned */
222 __u16 tail; /* what tail broke some buddy */
223 __u16 buddy; /* buddy the tail ^^^ broke */
224 __u16 flags;
225 __u8 cr:3; /* which phase the result extent was found at */
226 __u8 op:4;
227 __u8 merged:1;
228};
229
230struct ext4_buddy {
231 struct page *bd_buddy_page;
232 void *bd_buddy;
233 struct page *bd_bitmap_page;
234 void *bd_bitmap;
235 struct ext4_group_info *bd_info;
236 struct super_block *bd_sb;
237 __u16 bd_blkbits;
238 ext4_group_t bd_group;
239};
240#define EXT4_MB_BITMAP(e4b) ((e4b)->bd_bitmap)
241#define EXT4_MB_BUDDY(e4b) ((e4b)->bd_buddy)
242
243#ifndef EXT4_MB_HISTORY
244static inline void ext4_mb_store_history(struct ext4_allocation_context *ac)
245{
246 return;
247}
248#else
249static void ext4_mb_store_history(struct ext4_allocation_context *ac);
250#endif
251
252#define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
253
254static struct proc_dir_entry *proc_root_ext4;
255struct buffer_head *read_block_bitmap(struct super_block *, ext4_group_t);
256
257static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
258 ext4_group_t group);
259static void ext4_mb_poll_new_transaction(struct super_block *, handle_t *);
260static void ext4_mb_free_committed_blocks(struct super_block *);
261static void ext4_mb_return_to_preallocation(struct inode *inode,
262 struct ext4_buddy *e4b, sector_t block,
263 int count);
264static void ext4_mb_put_pa(struct ext4_allocation_context *,
265 struct super_block *, struct ext4_prealloc_space *pa);
266static int ext4_mb_init_per_dev_proc(struct super_block *sb);
267static int ext4_mb_destroy_per_dev_proc(struct super_block *sb);
268
269
270static inline void ext4_lock_group(struct super_block *sb, ext4_group_t group)
271{
272 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
273
274 bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
275}
276
277static inline void ext4_unlock_group(struct super_block *sb,
278 ext4_group_t group)
279{
280 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
281
282 bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
283}
284
285static inline int ext4_is_group_locked(struct super_block *sb,
286 ext4_group_t group)
287{
288 struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
289
290 return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT,
291 &(grinfo->bb_state));
292}
293
294static ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
295 struct ext4_free_extent *fex)
296{
297 ext4_fsblk_t block;
298
299 block = (ext4_fsblk_t) fex->fe_group * EXT4_BLOCKS_PER_GROUP(sb)
300 + fex->fe_start
301 + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
302 return block;
303}
304#endif
generated by cgit v1.2.2 (git 2.25.0) at 2025-11-23 04:10:46 -0500