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/init.h>
15 #include <linux/f2fs_fs.h>
16 #include <linux/kthread.h>
17 #include <linux/delay.h>
18 #include <linux/freezer.h>
19 #include <linux/blkdev.h>
25 #include <trace/events/f2fs.h>
27 static struct kmem_cache
*winode_slab
;
29 static int gc_thread_func(void *data
)
31 struct f2fs_sb_info
*sbi
= data
;
32 struct f2fs_gc_kthread
*gc_th
= sbi
->gc_thread
;
33 wait_queue_head_t
*wq
= &sbi
->gc_thread
->gc_wait_queue_head
;
36 wait_ms
= gc_th
->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
= increase_sleep_time(gc_th
, wait_ms
);
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 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(gc_th
, wait_ms
);
71 mutex_unlock(&sbi
->gc_mutex
);
75 if (has_enough_invalid_blocks(sbi
))
76 wait_ms
= decrease_sleep_time(gc_th
, wait_ms
);
78 wait_ms
= increase_sleep_time(gc_th
, wait_ms
);
80 stat_inc_bggc_count(sbi
);
82 /* if return value is not zero, no victim was selected */
84 wait_ms
= gc_th
->no_gc_sleep_time
;
86 /* balancing f2fs's metadata periodically */
87 f2fs_balance_fs_bg(sbi
);
89 } while (!kthread_should_stop());
93 int start_gc_thread(struct f2fs_sb_info
*sbi
)
95 struct f2fs_gc_kthread
*gc_th
;
96 dev_t dev
= sbi
->sb
->s_bdev
->bd_dev
;
99 gc_th
= kmalloc(sizeof(struct f2fs_gc_kthread
), GFP_KERNEL
);
105 gc_th
->min_sleep_time
= DEF_GC_THREAD_MIN_SLEEP_TIME
;
106 gc_th
->max_sleep_time
= DEF_GC_THREAD_MAX_SLEEP_TIME
;
107 gc_th
->no_gc_sleep_time
= DEF_GC_THREAD_NOGC_SLEEP_TIME
;
111 sbi
->gc_thread
= gc_th
;
112 init_waitqueue_head(&sbi
->gc_thread
->gc_wait_queue_head
);
113 sbi
->gc_thread
->f2fs_gc_task
= kthread_run(gc_thread_func
, sbi
,
114 "f2fs_gc-%u:%u", MAJOR(dev
), MINOR(dev
));
115 if (IS_ERR(gc_th
->f2fs_gc_task
)) {
116 err
= PTR_ERR(gc_th
->f2fs_gc_task
);
118 sbi
->gc_thread
= NULL
;
124 void stop_gc_thread(struct f2fs_sb_info
*sbi
)
126 struct f2fs_gc_kthread
*gc_th
= sbi
->gc_thread
;
129 kthread_stop(gc_th
->f2fs_gc_task
);
131 sbi
->gc_thread
= NULL
;
134 static int select_gc_type(struct f2fs_gc_kthread
*gc_th
, int gc_type
)
136 int gc_mode
= (gc_type
== BG_GC
) ? GC_CB
: GC_GREEDY
;
138 if (gc_th
&& gc_th
->gc_idle
) {
139 if (gc_th
->gc_idle
== 1)
141 else if (gc_th
->gc_idle
== 2)
147 static void select_policy(struct f2fs_sb_info
*sbi
, int gc_type
,
148 int type
, struct victim_sel_policy
*p
)
150 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
152 if (p
->alloc_mode
== SSR
) {
153 p
->gc_mode
= GC_GREEDY
;
154 p
->dirty_segmap
= dirty_i
->dirty_segmap
[type
];
155 p
->max_search
= dirty_i
->nr_dirty
[type
];
158 p
->gc_mode
= select_gc_type(sbi
->gc_thread
, gc_type
);
159 p
->dirty_segmap
= dirty_i
->dirty_segmap
[DIRTY
];
160 p
->max_search
= dirty_i
->nr_dirty
[DIRTY
];
161 p
->ofs_unit
= sbi
->segs_per_sec
;
164 if (p
->max_search
> sbi
->max_victim_search
)
165 p
->max_search
= sbi
->max_victim_search
;
167 p
->offset
= sbi
->last_victim
[p
->gc_mode
];
170 static unsigned int get_max_cost(struct f2fs_sb_info
*sbi
,
171 struct victim_sel_policy
*p
)
173 /* SSR allocates in a segment unit */
174 if (p
->alloc_mode
== SSR
)
175 return 1 << sbi
->log_blocks_per_seg
;
176 if (p
->gc_mode
== GC_GREEDY
)
177 return (1 << sbi
->log_blocks_per_seg
) * p
->ofs_unit
;
178 else if (p
->gc_mode
== GC_CB
)
180 else /* No other gc_mode */
184 static unsigned int check_bg_victims(struct f2fs_sb_info
*sbi
)
186 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
190 * If the gc_type is FG_GC, we can select victim segments
191 * selected by background GC before.
192 * Those segments guarantee they have small valid blocks.
194 for_each_set_bit(secno
, dirty_i
->victim_secmap
, MAIN_SECS(sbi
)) {
195 if (sec_usage_check(sbi
, secno
))
197 clear_bit(secno
, dirty_i
->victim_secmap
);
198 return secno
* sbi
->segs_per_sec
;
203 static unsigned int get_cb_cost(struct f2fs_sb_info
*sbi
, unsigned int segno
)
205 struct sit_info
*sit_i
= SIT_I(sbi
);
206 unsigned int secno
= GET_SECNO(sbi
, segno
);
207 unsigned int start
= secno
* sbi
->segs_per_sec
;
208 unsigned long long mtime
= 0;
209 unsigned int vblocks
;
210 unsigned char age
= 0;
214 for (i
= 0; i
< sbi
->segs_per_sec
; i
++)
215 mtime
+= get_seg_entry(sbi
, start
+ i
)->mtime
;
216 vblocks
= get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
);
218 mtime
= div_u64(mtime
, sbi
->segs_per_sec
);
219 vblocks
= div_u64(vblocks
, sbi
->segs_per_sec
);
221 u
= (vblocks
* 100) >> sbi
->log_blocks_per_seg
;
223 /* Handle if the system time has changed by the user */
224 if (mtime
< sit_i
->min_mtime
)
225 sit_i
->min_mtime
= mtime
;
226 if (mtime
> sit_i
->max_mtime
)
227 sit_i
->max_mtime
= mtime
;
228 if (sit_i
->max_mtime
!= sit_i
->min_mtime
)
229 age
= 100 - div64_u64(100 * (mtime
- sit_i
->min_mtime
),
230 sit_i
->max_mtime
- sit_i
->min_mtime
);
232 return UINT_MAX
- ((100 * (100 - u
) * age
) / (100 + u
));
235 static inline unsigned int get_gc_cost(struct f2fs_sb_info
*sbi
,
236 unsigned int segno
, struct victim_sel_policy
*p
)
238 if (p
->alloc_mode
== SSR
)
239 return get_seg_entry(sbi
, segno
)->ckpt_valid_blocks
;
241 /* alloc_mode == LFS */
242 if (p
->gc_mode
== GC_GREEDY
)
243 return get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
);
245 return get_cb_cost(sbi
, segno
);
249 * This function is called from two paths.
250 * One is garbage collection and the other is SSR segment selection.
251 * When it is called during GC, it just gets a victim segment
252 * and it does not remove it from dirty seglist.
253 * When it is called from SSR segment selection, it finds a segment
254 * which has minimum valid blocks and removes it from dirty seglist.
256 static int get_victim_by_default(struct f2fs_sb_info
*sbi
,
257 unsigned int *result
, int gc_type
, int type
, char alloc_mode
)
259 struct dirty_seglist_info
*dirty_i
= DIRTY_I(sbi
);
260 struct victim_sel_policy p
;
261 unsigned int secno
, max_cost
;
264 mutex_lock(&dirty_i
->seglist_lock
);
266 p
.alloc_mode
= alloc_mode
;
267 select_policy(sbi
, gc_type
, type
, &p
);
269 p
.min_segno
= NULL_SEGNO
;
270 p
.min_cost
= max_cost
= get_max_cost(sbi
, &p
);
272 if (p
.alloc_mode
== LFS
&& gc_type
== FG_GC
) {
273 p
.min_segno
= check_bg_victims(sbi
);
274 if (p
.min_segno
!= NULL_SEGNO
)
282 segno
= find_next_bit(p
.dirty_segmap
, MAIN_SEGS(sbi
), p
.offset
);
283 if (segno
>= MAIN_SEGS(sbi
)) {
284 if (sbi
->last_victim
[p
.gc_mode
]) {
285 sbi
->last_victim
[p
.gc_mode
] = 0;
292 p
.offset
= segno
+ p
.ofs_unit
;
294 p
.offset
-= segno
% p
.ofs_unit
;
296 secno
= GET_SECNO(sbi
, segno
);
298 if (sec_usage_check(sbi
, secno
))
300 if (gc_type
== BG_GC
&& test_bit(secno
, dirty_i
->victim_secmap
))
303 cost
= get_gc_cost(sbi
, segno
, &p
);
305 if (p
.min_cost
> cost
) {
308 } else if (unlikely(cost
== max_cost
)) {
312 if (nsearched
++ >= p
.max_search
) {
313 sbi
->last_victim
[p
.gc_mode
] = segno
;
317 if (p
.min_segno
!= NULL_SEGNO
) {
319 if (p
.alloc_mode
== LFS
) {
320 secno
= GET_SECNO(sbi
, p
.min_segno
);
321 if (gc_type
== FG_GC
)
322 sbi
->cur_victim_sec
= secno
;
324 set_bit(secno
, dirty_i
->victim_secmap
);
326 *result
= (p
.min_segno
/ p
.ofs_unit
) * p
.ofs_unit
;
328 trace_f2fs_get_victim(sbi
->sb
, type
, gc_type
, &p
,
330 prefree_segments(sbi
), free_segments(sbi
));
332 mutex_unlock(&dirty_i
->seglist_lock
);
334 return (p
.min_segno
== NULL_SEGNO
) ? 0 : 1;
337 static const struct victim_selection default_v_ops
= {
338 .get_victim
= get_victim_by_default
,
341 static struct inode
*find_gc_inode(struct gc_inode_list
*gc_list
, nid_t ino
)
343 struct inode_entry
*ie
;
345 ie
= radix_tree_lookup(&gc_list
->iroot
, ino
);
351 static void add_gc_inode(struct gc_inode_list
*gc_list
, struct inode
*inode
)
353 struct inode_entry
*new_ie
;
355 if (inode
== find_gc_inode(gc_list
, inode
->i_ino
)) {
359 new_ie
= f2fs_kmem_cache_alloc(winode_slab
, GFP_NOFS
);
360 new_ie
->inode
= inode
;
362 if (radix_tree_insert(&gc_list
->iroot
, inode
->i_ino
, new_ie
)) {
366 list_add_tail(&new_ie
->list
, &gc_list
->ilist
);
369 static void put_gc_inode(struct gc_inode_list
*gc_list
)
371 struct inode_entry
*ie
, *next_ie
;
372 list_for_each_entry_safe(ie
, next_ie
, &gc_list
->ilist
, list
) {
373 radix_tree_delete(&gc_list
->iroot
, ie
->inode
->i_ino
);
376 kmem_cache_free(winode_slab
, ie
);
380 static int check_valid_map(struct f2fs_sb_info
*sbi
,
381 unsigned int segno
, int offset
)
383 struct sit_info
*sit_i
= SIT_I(sbi
);
384 struct seg_entry
*sentry
;
387 mutex_lock(&sit_i
->sentry_lock
);
388 sentry
= get_seg_entry(sbi
, segno
);
389 ret
= f2fs_test_bit(offset
, sentry
->cur_valid_map
);
390 mutex_unlock(&sit_i
->sentry_lock
);
395 * This function compares node address got in summary with that in NAT.
396 * On validity, copy that node with cold status, otherwise (invalid node)
399 static void gc_node_segment(struct f2fs_sb_info
*sbi
,
400 struct f2fs_summary
*sum
, unsigned int segno
, int gc_type
)
403 struct f2fs_summary
*entry
;
409 for (off
= 0; off
< sbi
->blocks_per_seg
; off
++, entry
++) {
410 nid_t nid
= le32_to_cpu(entry
->nid
);
411 struct page
*node_page
;
413 /* stop BG_GC if there is not enough free sections. */
414 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, 0))
417 if (check_valid_map(sbi
, segno
, off
) == 0)
421 ra_node_page(sbi
, nid
);
424 node_page
= get_node_page(sbi
, nid
);
425 if (IS_ERR(node_page
))
428 /* block may become invalid during get_node_page */
429 if (check_valid_map(sbi
, segno
, off
) == 0) {
430 f2fs_put_page(node_page
, 1);
434 /* set page dirty and write it */
435 if (gc_type
== FG_GC
) {
436 f2fs_wait_on_page_writeback(node_page
, NODE
);
437 set_page_dirty(node_page
);
439 if (!PageWriteback(node_page
))
440 set_page_dirty(node_page
);
442 f2fs_put_page(node_page
, 1);
443 stat_inc_node_blk_count(sbi
, 1);
451 if (gc_type
== FG_GC
) {
452 struct writeback_control wbc
= {
453 .sync_mode
= WB_SYNC_ALL
,
454 .nr_to_write
= LONG_MAX
,
457 sync_node_pages(sbi
, 0, &wbc
);
460 * In the case of FG_GC, it'd be better to reclaim this victim
463 if (get_valid_blocks(sbi
, segno
, 1) != 0)
469 * Calculate start block index indicating the given node offset.
470 * Be careful, caller should give this node offset only indicating direct node
471 * blocks. If any node offsets, which point the other types of node blocks such
472 * as indirect or double indirect node blocks, are given, it must be a caller's
475 block_t
start_bidx_of_node(unsigned int node_ofs
, struct f2fs_inode_info
*fi
)
477 unsigned int indirect_blks
= 2 * NIDS_PER_BLOCK
+ 4;
485 } else if (node_ofs
<= indirect_blks
) {
486 int dec
= (node_ofs
- 4) / (NIDS_PER_BLOCK
+ 1);
487 bidx
= node_ofs
- 2 - dec
;
489 int dec
= (node_ofs
- indirect_blks
- 3) / (NIDS_PER_BLOCK
+ 1);
490 bidx
= node_ofs
- 5 - dec
;
492 return bidx
* ADDRS_PER_BLOCK
+ ADDRS_PER_INODE(fi
);
495 static int check_dnode(struct f2fs_sb_info
*sbi
, struct f2fs_summary
*sum
,
496 struct node_info
*dni
, block_t blkaddr
, unsigned int *nofs
)
498 struct page
*node_page
;
500 unsigned int ofs_in_node
;
501 block_t source_blkaddr
;
503 nid
= le32_to_cpu(sum
->nid
);
504 ofs_in_node
= le16_to_cpu(sum
->ofs_in_node
);
506 node_page
= get_node_page(sbi
, nid
);
507 if (IS_ERR(node_page
))
510 get_node_info(sbi
, nid
, dni
);
512 if (sum
->version
!= dni
->version
) {
513 f2fs_put_page(node_page
, 1);
517 *nofs
= ofs_of_node(node_page
);
518 source_blkaddr
= datablock_addr(node_page
, ofs_in_node
);
519 f2fs_put_page(node_page
, 1);
521 if (source_blkaddr
!= blkaddr
)
526 static void move_data_page(struct inode
*inode
, struct page
*page
, int gc_type
)
528 struct f2fs_io_info fio
= {
533 if (gc_type
== BG_GC
) {
534 if (PageWriteback(page
))
536 set_page_dirty(page
);
539 f2fs_wait_on_page_writeback(page
, DATA
);
541 if (clear_page_dirty_for_io(page
))
542 inode_dec_dirty_pages(inode
);
544 do_write_data_page(page
, &fio
);
545 clear_cold_data(page
);
548 f2fs_put_page(page
, 1);
552 * This function tries to get parent node of victim data block, and identifies
553 * data block validity. If the block is valid, copy that with cold status and
554 * modify parent node.
555 * If the parent node is not valid or the data block address is different,
556 * the victim data block is ignored.
558 static void gc_data_segment(struct f2fs_sb_info
*sbi
, struct f2fs_summary
*sum
,
559 struct gc_inode_list
*gc_list
, unsigned int segno
, int gc_type
)
561 struct super_block
*sb
= sbi
->sb
;
562 struct f2fs_summary
*entry
;
567 start_addr
= START_BLOCK(sbi
, segno
);
572 for (off
= 0; off
< sbi
->blocks_per_seg
; off
++, entry
++) {
573 struct page
*data_page
;
575 struct node_info dni
; /* dnode info for the data */
576 unsigned int ofs_in_node
, nofs
;
579 /* stop BG_GC if there is not enough free sections. */
580 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, 0))
583 if (check_valid_map(sbi
, segno
, off
) == 0)
587 ra_node_page(sbi
, le32_to_cpu(entry
->nid
));
591 /* Get an inode by ino with checking validity */
592 if (check_dnode(sbi
, entry
, &dni
, start_addr
+ off
, &nofs
) == 0)
596 ra_node_page(sbi
, dni
.ino
);
600 ofs_in_node
= le16_to_cpu(entry
->ofs_in_node
);
603 inode
= f2fs_iget(sb
, dni
.ino
);
604 if (IS_ERR(inode
) || is_bad_inode(inode
))
607 start_bidx
= start_bidx_of_node(nofs
, F2FS_I(inode
));
609 data_page
= find_data_page(inode
,
610 start_bidx
+ ofs_in_node
, false);
611 if (IS_ERR(data_page
)) {
616 f2fs_put_page(data_page
, 0);
617 add_gc_inode(gc_list
, inode
);
622 inode
= find_gc_inode(gc_list
, dni
.ino
);
624 start_bidx
= start_bidx_of_node(nofs
, F2FS_I(inode
));
625 data_page
= get_lock_data_page(inode
,
626 start_bidx
+ ofs_in_node
);
627 if (IS_ERR(data_page
))
629 move_data_page(inode
, data_page
, gc_type
);
630 stat_inc_data_blk_count(sbi
, 1);
637 if (gc_type
== FG_GC
) {
638 f2fs_submit_merged_bio(sbi
, DATA
, WRITE
);
641 * In the case of FG_GC, it'd be better to reclaim this victim
644 if (get_valid_blocks(sbi
, segno
, 1) != 0) {
651 static int __get_victim(struct f2fs_sb_info
*sbi
, unsigned int *victim
,
654 struct sit_info
*sit_i
= SIT_I(sbi
);
657 mutex_lock(&sit_i
->sentry_lock
);
658 ret
= DIRTY_I(sbi
)->v_ops
->get_victim(sbi
, victim
, gc_type
,
660 mutex_unlock(&sit_i
->sentry_lock
);
664 static void do_garbage_collect(struct f2fs_sb_info
*sbi
, unsigned int segno
,
665 struct gc_inode_list
*gc_list
, int gc_type
)
667 struct page
*sum_page
;
668 struct f2fs_summary_block
*sum
;
669 struct blk_plug plug
;
671 /* read segment summary of victim */
672 sum_page
= get_sum_page(sbi
, segno
);
674 blk_start_plug(&plug
);
676 sum
= page_address(sum_page
);
678 switch (GET_SUM_TYPE((&sum
->footer
))) {
680 gc_node_segment(sbi
, sum
->entries
, segno
, gc_type
);
683 gc_data_segment(sbi
, sum
->entries
, gc_list
, segno
, gc_type
);
686 blk_finish_plug(&plug
);
688 stat_inc_seg_count(sbi
, GET_SUM_TYPE((&sum
->footer
)));
689 stat_inc_call_count(sbi
->stat_info
);
691 f2fs_put_page(sum_page
, 1);
694 int f2fs_gc(struct f2fs_sb_info
*sbi
)
696 unsigned int segno
, i
;
700 struct cp_control cpc
;
701 struct gc_inode_list gc_list
= {
702 .ilist
= LIST_HEAD_INIT(gc_list
.ilist
),
703 .iroot
= RADIX_TREE_INIT(GFP_NOFS
),
706 cpc
.reason
= test_opt(sbi
, FASTBOOT
) ? CP_UMOUNT
: CP_SYNC
;
709 if (unlikely(!(sbi
->sb
->s_flags
& MS_ACTIVE
)))
711 if (unlikely(f2fs_cp_error(sbi
)))
714 if (gc_type
== BG_GC
&& has_not_enough_free_secs(sbi
, nfree
)) {
716 write_checkpoint(sbi
, &cpc
);
719 if (!__get_victim(sbi
, &segno
, gc_type
))
723 /* readahead multi ssa blocks those have contiguous address */
724 if (sbi
->segs_per_sec
> 1)
725 ra_meta_pages(sbi
, GET_SUM_BLOCK(sbi
, segno
), sbi
->segs_per_sec
,
728 for (i
= 0; i
< sbi
->segs_per_sec
; i
++)
729 do_garbage_collect(sbi
, segno
+ i
, &gc_list
, gc_type
);
731 if (gc_type
== FG_GC
) {
732 sbi
->cur_victim_sec
= NULL_SEGNO
;
734 WARN_ON(get_valid_blocks(sbi
, segno
, sbi
->segs_per_sec
));
737 if (has_not_enough_free_secs(sbi
, nfree
))
740 if (gc_type
== FG_GC
)
741 write_checkpoint(sbi
, &cpc
);
743 mutex_unlock(&sbi
->gc_mutex
);
745 put_gc_inode(&gc_list
);
749 void build_gc_manager(struct f2fs_sb_info
*sbi
)
751 DIRTY_I(sbi
)->v_ops
= &default_v_ops
;
754 int __init
create_gc_caches(void)
756 winode_slab
= f2fs_kmem_cache_create("f2fs_gc_inodes",
757 sizeof(struct inode_entry
));
763 void destroy_gc_caches(void)
765 kmem_cache_destroy(winode_slab
);