4 * Copyright (c) 2012 Samsung Electronics Co., Ltd.
5 * http://www.samsung.com/
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
12 #include <linux/module.h>
13 #include <linux/backing-dev.h>
14 #include <linux/proc_fs.h>
15 #include <linux/init.h>
16 #include <linux/f2fs_fs.h>
17 #include <linux/kthread.h>
18 #include <linux/delay.h>
19 #include <linux/freezer.h>
20 #include <linux/blkdev.h>
26 #include <trace/events/f2fs.h>
28 static struct kmem_cache
*winode_slab
;
30 static int gc_thread_func(void *data
)
32 struct f2fs_sb_info
*sbi
= data
;
33 wait_queue_head_t
*wq
= &sbi
->gc_thread
->gc_wait_queue_head
;
36 wait_ms
= GC_THREAD_MIN_SLEEP_TIME
;
42 wait_event_interruptible_timeout(*wq
,
43 kthread_should_stop(),
44 msecs_to_jiffies(wait_ms
));
45 if (kthread_should_stop())
48 if (sbi
->sb
->s_writers
.frozen
>= SB_FREEZE_WRITE
) {
49 wait_ms
= GC_THREAD_MAX_SLEEP_TIME
;
54 * [GC triggering condition]
55 * 0. GC is not conducted currently.
56 * 1. There are enough dirty segments.
57 * 2. IO subsystem is idle by checking the # of writeback pages.
58 * 3. IO subsystem is idle by checking the # of requests in
59 * bdev's request list.
61 * Note) We have to avoid triggering GCs too much frequently.
62 * Because it is possible that some segments can be
63 * invalidated soon after by user update or deletion.
64 * So, I'd like to wait some time to collect dirty segments.
66 if (!mutex_trylock(&sbi
->gc_mutex
))
70 wait_ms
= increase_sleep_time(wait_ms
);
71 mutex_unlock(&sbi
->gc_mutex
);
75 if (has_enough_invalid_blocks(sbi
))
76 wait_ms
= decrease_sleep_time(wait_ms
);
78 wait_ms
= increase_sleep_time(wait_ms
);
82 /* if return value is not zero, no victim was selected */
84 wait_ms
= GC_THREAD_NOGC_SLEEP_TIME
;
85 } while (!kthread_should_stop());
89 int start_gc_thread(struct f2fs_sb_info
*sbi
)
91 struct f2fs_gc_kthread
*gc_th
;
92 dev_t dev
= sbi
->sb
->s_bdev
->bd_dev
;
94 if (!test_opt(sbi
, BG_GC
))
96 gc_th
= kmalloc(sizeof(struct f2fs_gc_kthread
), GFP_KERNEL
);
100 sbi
->gc_thread
= gc_th
;
101 init_waitqueue_head(&sbi
->gc_thread
->gc_wait_queue_head
);
102 sbi
->gc_thread
->f2fs_gc_task
= kthread_run(gc_thread_func
, sbi
,
103 "f2fs_gc-%u:%u", MAJOR(dev
), MINOR(dev
));
104 if (IS_ERR(gc_th
->f2fs_gc_task
)) {
106 sbi
->gc_thread
= NULL
;
112 void stop_gc_thread(struct f2fs_sb_info
*sbi
)
114 struct f2fs_gc_kthread
*gc_th
= sbi
->gc_thread
;
117 kthread_stop(gc_th
->f2fs_gc_task
);
119 sbi
->gc_thread
= NULL
;
122 static int select_gc_type(int gc_type
)
124 return (gc_type
== BG_GC
) ? GC_CB
: GC_GREEDY
;
127 static void select_policy(struct f2fs_sb_info
*sbi
, int gc_type
,
128 int type
, struct victim_sel_policy
*p
)
130 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
132 if (p
->alloc_mode
== SSR
) {
133 p
->gc_mode
= GC_GREEDY
;
134 p
->dirty_segmap
= dirty_i
->dirty_segmap
[type
];
137 p
->gc_mode
= select_gc_type(gc_type
);
138 p
->dirty_segmap
= dirty_i
->dirty_segmap
[DIRTY
];
139 p
->ofs_unit
= sbi
->segs_per_sec
;
141 p
->offset
= sbi
->last_victim
[p
->gc_mode
];
144 static unsigned int get_max_cost(struct f2fs_sb_info
*sbi
,
145 struct victim_sel_policy
*p
)
147 /* SSR allocates in a segment unit */
148 if (p
->alloc_mode
== SSR
)
149 return 1 << sbi
->log_blocks_per_seg
;
150 if (p
->gc_mode
== GC_GREEDY
)
151 return (1 << sbi
->log_blocks_per_seg
) * p
->ofs_unit
;
152 else if (p
->gc_mode
== GC_CB
)
154 else /* No other gc_mode */
158 static unsigned int check_bg_victims(struct f2fs_sb_info
*sbi
)
160 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
161 unsigned int hint
= 0;
165 * If the gc_type is FG_GC, we can select victim segments
166 * selected by background GC before.
167 * Those segments guarantee they have small valid blocks.
170 secno
= find_next_bit(dirty_i
->victim_secmap
, TOTAL_SECS(sbi
), hint
++);
171 if (secno
< TOTAL_SECS(sbi
)) {
172 if (sec_usage_check(sbi
, secno
))
174 clear_bit(secno
, dirty_i
->victim_secmap
);
175 return secno
* sbi
->segs_per_sec
;
180 static unsigned int get_cb_cost(struct f2fs_sb_info
*sbi
, unsigned int segno
)
182 struct sit_info
*sit_i
= SIT_I(sbi
);
183 unsigned int secno
= GET_SECNO(sbi
, segno
);
184 unsigned int start
= secno
* sbi
->segs_per_sec
;
185 unsigned long long mtime
= 0;
186 unsigned int vblocks
;
187 unsigned char age
= 0;
191 for (i
= 0; i
< sbi
->segs_per_sec
; i
++)
192 mtime
+= get_seg_entry(sbi
, start
+ i
)->mtime
;
193 vblocks
= get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
);
195 mtime
= div_u64(mtime
, sbi
->segs_per_sec
);
196 vblocks
= div_u64(vblocks
, sbi
->segs_per_sec
);
198 u
= (vblocks
* 100) >> sbi
->log_blocks_per_seg
;
200 /* Handle if the system time is changed by user */
201 if (mtime
< sit_i
->min_mtime
)
202 sit_i
->min_mtime
= mtime
;
203 if (mtime
> sit_i
->max_mtime
)
204 sit_i
->max_mtime
= mtime
;
205 if (sit_i
->max_mtime
!= sit_i
->min_mtime
)
206 age
= 100 - div64_u64(100 * (mtime
- sit_i
->min_mtime
),
207 sit_i
->max_mtime
- sit_i
->min_mtime
);
209 return UINT_MAX
- ((100 * (100 - u
) * age
) / (100 + u
));
212 static unsigned int get_gc_cost(struct f2fs_sb_info
*sbi
, unsigned int segno
,
213 struct victim_sel_policy
*p
)
215 if (p
->alloc_mode
== SSR
)
216 return get_seg_entry(sbi
, segno
)->ckpt_valid_blocks
;
218 /* alloc_mode == LFS */
219 if (p
->gc_mode
== GC_GREEDY
)
220 return get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
);
222 return get_cb_cost(sbi
, segno
);
226 * This function is called from two paths.
227 * One is garbage collection and the other is SSR segment selection.
228 * When it is called during GC, it just gets a victim segment
229 * and it does not remove it from dirty seglist.
230 * When it is called from SSR segment selection, it finds a segment
231 * which has minimum valid blocks and removes it from dirty seglist.
233 static int get_victim_by_default(struct f2fs_sb_info
*sbi
,
234 unsigned int *result
, int gc_type
, int type
, char alloc_mode
)
236 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
237 struct victim_sel_policy p
;
241 p
.alloc_mode
= alloc_mode
;
242 select_policy(sbi
, gc_type
, type
, &p
);
244 p
.min_segno
= NULL_SEGNO
;
245 p
.min_cost
= get_max_cost(sbi
, &p
);
247 mutex_lock(&dirty_i
->seglist_lock
);
249 if (p
.alloc_mode
== LFS
&& gc_type
== FG_GC
) {
250 p
.min_segno
= check_bg_victims(sbi
);
251 if (p
.min_segno
!= NULL_SEGNO
)
259 segno
= find_next_bit(p
.dirty_segmap
,
260 TOTAL_SEGS(sbi
), p
.offset
);
261 if (segno
>= TOTAL_SEGS(sbi
)) {
262 if (sbi
->last_victim
[p
.gc_mode
]) {
263 sbi
->last_victim
[p
.gc_mode
] = 0;
269 p
.offset
= ((segno
/ p
.ofs_unit
) * p
.ofs_unit
) + p
.ofs_unit
;
270 secno
= GET_SECNO(sbi
, segno
);
272 if (sec_usage_check(sbi
, secno
))
274 if (gc_type
== BG_GC
&& test_bit(secno
, dirty_i
->victim_secmap
))
277 cost
= get_gc_cost(sbi
, segno
, &p
);
279 if (p
.min_cost
> cost
) {
284 if (cost
== get_max_cost(sbi
, &p
))
287 if (nsearched
++ >= MAX_VICTIM_SEARCH
) {
288 sbi
->last_victim
[p
.gc_mode
] = segno
;
293 if (p
.min_segno
!= NULL_SEGNO
) {
294 if (p
.alloc_mode
== LFS
) {
295 secno
= GET_SECNO(sbi
, p
.min_segno
);
296 if (gc_type
== FG_GC
)
297 sbi
->cur_victim_sec
= secno
;
299 set_bit(secno
, dirty_i
->victim_secmap
);
301 *result
= (p
.min_segno
/ p
.ofs_unit
) * p
.ofs_unit
;
303 trace_f2fs_get_victim(sbi
->sb
, type
, gc_type
, &p
,
305 prefree_segments(sbi
), free_segments(sbi
));
307 mutex_unlock(&dirty_i
->seglist_lock
);
309 return (p
.min_segno
== NULL_SEGNO
) ? 0 : 1;
312 static const struct victim_selection default_v_ops
= {
313 .get_victim
= get_victim_by_default
,
316 static struct inode
*find_gc_inode(nid_t ino
, struct list_head
*ilist
)
318 struct list_head
*this;
319 struct inode_entry
*ie
;
321 list_for_each(this, ilist
) {
322 ie
= list_entry(this, struct inode_entry
, list
);
323 if (ie
->inode
->i_ino
== ino
)
329 static void add_gc_inode(struct inode
*inode
, struct list_head
*ilist
)
331 struct list_head
*this;
332 struct inode_entry
*new_ie
, *ie
;
334 list_for_each(this, ilist
) {
335 ie
= list_entry(this, struct inode_entry
, list
);
336 if (ie
->inode
== inode
) {
342 new_ie
= kmem_cache_alloc(winode_slab
, GFP_NOFS
);
347 new_ie
->inode
= inode
;
348 list_add_tail(&new_ie
->list
, ilist
);
351 static void put_gc_inode(struct list_head
*ilist
)
353 struct inode_entry
*ie
, *next_ie
;
354 list_for_each_entry_safe(ie
, next_ie
, ilist
, list
) {
357 kmem_cache_free(winode_slab
, ie
);
361 static int check_valid_map(struct f2fs_sb_info
*sbi
,
362 unsigned int segno
, int offset
)
364 struct sit_info
*sit_i
= SIT_I(sbi
);
365 struct seg_entry
*sentry
;
368 mutex_lock(&sit_i
->sentry_lock
);
369 sentry
= get_seg_entry(sbi
, segno
);
370 ret
= f2fs_test_bit(offset
, sentry
->cur_valid_map
);
371 mutex_unlock(&sit_i
->sentry_lock
);
376 * This function compares node address got in summary with that in NAT.
377 * On validity, copy that node with cold status, otherwise (invalid node)
380 static void gc_node_segment(struct f2fs_sb_info
*sbi
,
381 struct f2fs_summary
*sum
, unsigned int segno
, int gc_type
)
384 struct f2fs_summary
*entry
;
390 for (off
= 0; off
< sbi
->blocks_per_seg
; off
++, entry
++) {
391 nid_t nid
= le32_to_cpu(entry
->nid
);
392 struct page
*node_page
;
394 /* stop BG_GC if there is not enough free sections. */
395 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, 0))
398 if (check_valid_map(sbi
, segno
, off
) == 0)
402 ra_node_page(sbi
, nid
);
405 node_page
= get_node_page(sbi
, nid
);
406 if (IS_ERR(node_page
))
409 /* set page dirty and write it */
410 if (gc_type
== FG_GC
) {
411 f2fs_submit_bio(sbi
, NODE
, true);
412 wait_on_page_writeback(node_page
);
413 set_page_dirty(node_page
);
415 if (!PageWriteback(node_page
))
416 set_page_dirty(node_page
);
418 f2fs_put_page(node_page
, 1);
419 stat_inc_node_blk_count(sbi
, 1);
427 if (gc_type
== FG_GC
) {
428 struct writeback_control wbc
= {
429 .sync_mode
= WB_SYNC_ALL
,
430 .nr_to_write
= LONG_MAX
,
433 sync_node_pages(sbi
, 0, &wbc
);
436 * In the case of FG_GC, it'd be better to reclaim this victim
439 if (get_valid_blocks(sbi
, segno
, 1) != 0)
445 * Calculate start block index indicating the given node offset.
446 * Be careful, caller should give this node offset only indicating direct node
447 * blocks. If any node offsets, which point the other types of node blocks such
448 * as indirect or double indirect node blocks, are given, it must be a caller's
451 block_t
start_bidx_of_node(unsigned int node_ofs
)
453 unsigned int indirect_blks
= 2 * NIDS_PER_BLOCK
+ 4;
461 } else if (node_ofs
<= indirect_blks
) {
462 int dec
= (node_ofs
- 4) / (NIDS_PER_BLOCK
+ 1);
463 bidx
= node_ofs
- 2 - dec
;
465 int dec
= (node_ofs
- indirect_blks
- 3) / (NIDS_PER_BLOCK
+ 1);
466 bidx
= node_ofs
- 5 - dec
;
468 return bidx
* ADDRS_PER_BLOCK
+ ADDRS_PER_INODE
;
471 static int check_dnode(struct f2fs_sb_info
*sbi
, struct f2fs_summary
*sum
,
472 struct node_info
*dni
, block_t blkaddr
, unsigned int *nofs
)
474 struct page
*node_page
;
476 unsigned int ofs_in_node
;
477 block_t source_blkaddr
;
479 nid
= le32_to_cpu(sum
->nid
);
480 ofs_in_node
= le16_to_cpu(sum
->ofs_in_node
);
482 node_page
= get_node_page(sbi
, nid
);
483 if (IS_ERR(node_page
))
486 get_node_info(sbi
, nid
, dni
);
488 if (sum
->version
!= dni
->version
) {
489 f2fs_put_page(node_page
, 1);
493 *nofs
= ofs_of_node(node_page
);
494 source_blkaddr
= datablock_addr(node_page
, ofs_in_node
);
495 f2fs_put_page(node_page
, 1);
497 if (source_blkaddr
!= blkaddr
)
502 static void move_data_page(struct inode
*inode
, struct page
*page
, int gc_type
)
504 if (gc_type
== BG_GC
) {
505 if (PageWriteback(page
))
507 set_page_dirty(page
);
510 struct f2fs_sb_info
*sbi
= F2FS_SB(inode
->i_sb
);
512 if (PageWriteback(page
)) {
513 f2fs_submit_bio(sbi
, DATA
, true);
514 wait_on_page_writeback(page
);
517 if (clear_page_dirty_for_io(page
) &&
518 S_ISDIR(inode
->i_mode
)) {
519 dec_page_count(sbi
, F2FS_DIRTY_DENTS
);
520 inode_dec_dirty_dents(inode
);
523 do_write_data_page(page
);
524 clear_cold_data(page
);
527 f2fs_put_page(page
, 1);
531 * This function tries to get parent node of victim data block, and identifies
532 * data block validity. If the block is valid, copy that with cold status and
533 * modify parent node.
534 * If the parent node is not valid or the data block address is different,
535 * the victim data block is ignored.
537 static void gc_data_segment(struct f2fs_sb_info
*sbi
, struct f2fs_summary
*sum
,
538 struct list_head
*ilist
, unsigned int segno
, int gc_type
)
540 struct super_block
*sb
= sbi
->sb
;
541 struct f2fs_summary
*entry
;
546 start_addr
= START_BLOCK(sbi
, segno
);
551 for (off
= 0; off
< sbi
->blocks_per_seg
; off
++, entry
++) {
552 struct page
*data_page
;
554 struct node_info dni
; /* dnode info for the data */
555 unsigned int ofs_in_node
, nofs
;
558 /* stop BG_GC if there is not enough free sections. */
559 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, 0))
562 if (check_valid_map(sbi
, segno
, off
) == 0)
566 ra_node_page(sbi
, le32_to_cpu(entry
->nid
));
570 /* Get an inode by ino with checking validity */
571 if (check_dnode(sbi
, entry
, &dni
, start_addr
+ off
, &nofs
) == 0)
575 ra_node_page(sbi
, dni
.ino
);
579 start_bidx
= start_bidx_of_node(nofs
);
580 ofs_in_node
= le16_to_cpu(entry
->ofs_in_node
);
583 inode
= f2fs_iget(sb
, dni
.ino
);
587 data_page
= find_data_page(inode
,
588 start_bidx
+ ofs_in_node
, false);
589 if (IS_ERR(data_page
))
592 f2fs_put_page(data_page
, 0);
593 add_gc_inode(inode
, ilist
);
595 inode
= find_gc_inode(dni
.ino
, ilist
);
597 data_page
= get_lock_data_page(inode
,
598 start_bidx
+ ofs_in_node
);
599 if (IS_ERR(data_page
))
601 move_data_page(inode
, data_page
, gc_type
);
602 stat_inc_data_blk_count(sbi
, 1);
613 if (gc_type
== FG_GC
) {
614 f2fs_submit_bio(sbi
, DATA
, true);
617 * In the case of FG_GC, it'd be better to reclaim this victim
620 if (get_valid_blocks(sbi
, segno
, 1) != 0) {
627 static int __get_victim(struct f2fs_sb_info
*sbi
, unsigned int *victim
,
628 int gc_type
, int type
)
630 struct sit_info
*sit_i
= SIT_I(sbi
);
632 mutex_lock(&sit_i
->sentry_lock
);
633 ret
= DIRTY_I(sbi
)->v_ops
->get_victim(sbi
, victim
, gc_type
, type
, LFS
);
634 mutex_unlock(&sit_i
->sentry_lock
);
638 static void do_garbage_collect(struct f2fs_sb_info
*sbi
, unsigned int segno
,
639 struct list_head
*ilist
, int gc_type
)
641 struct page
*sum_page
;
642 struct f2fs_summary_block
*sum
;
643 struct blk_plug plug
;
645 /* read segment summary of victim */
646 sum_page
= get_sum_page(sbi
, segno
);
647 if (IS_ERR(sum_page
))
650 blk_start_plug(&plug
);
652 sum
= page_address(sum_page
);
654 switch (GET_SUM_TYPE((&sum
->footer
))) {
656 gc_node_segment(sbi
, sum
->entries
, segno
, gc_type
);
659 gc_data_segment(sbi
, sum
->entries
, ilist
, segno
, gc_type
);
662 blk_finish_plug(&plug
);
664 stat_inc_seg_count(sbi
, GET_SUM_TYPE((&sum
->footer
)));
665 stat_inc_call_count(sbi
->stat_info
);
667 f2fs_put_page(sum_page
, 1);
670 int f2fs_gc(struct f2fs_sb_info
*sbi
)
672 struct list_head ilist
;
673 unsigned int segno
, i
;
678 INIT_LIST_HEAD(&ilist
);
680 if (!(sbi
->sb
->s_flags
& MS_ACTIVE
))
683 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, nfree
)) {
685 write_checkpoint(sbi
, false);
688 if (!__get_victim(sbi
, &segno
, gc_type
, NO_CHECK_TYPE
))
692 for (i
= 0; i
< sbi
->segs_per_sec
; i
++)
693 do_garbage_collect(sbi
, segno
+ i
, &ilist
, gc_type
);
695 if (gc_type
== FG_GC
) {
696 sbi
->cur_victim_sec
= NULL_SEGNO
;
698 WARN_ON(get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
));
701 if (has_not_enough_free_secs(sbi
, nfree
))
704 if (gc_type
== FG_GC
)
705 write_checkpoint(sbi
, false);
707 mutex_unlock(&sbi
->gc_mutex
);
709 put_gc_inode(&ilist
);
713 void build_gc_manager(struct f2fs_sb_info
*sbi
)
715 DIRTY_I(sbi
)->v_ops
= &default_v_ops
;
718 int __init
create_gc_caches(void)
720 winode_slab
= f2fs_kmem_cache_create("f2fs_gc_inodes",
721 sizeof(struct inode_entry
), NULL
);
727 void destroy_gc_caches(void)
729 kmem_cache_destroy(winode_slab
);