xref: /openbmc/linux/fs/f2fs/gc.c (revision 239480ab)
1 /*
2  * fs/f2fs/gc.c
3  *
4  * Copyright (c) 2012 Samsung Electronics Co., Ltd.
5  *             http://www.samsung.com/
6  *
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.
10  */
11 #include <linux/fs.h>
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 
20 #include "f2fs.h"
21 #include "node.h"
22 #include "segment.h"
23 #include "gc.h"
24 #include <trace/events/f2fs.h>
25 
26 static int gc_thread_func(void *data)
27 {
28 	struct f2fs_sb_info *sbi = data;
29 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
30 	wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
31 	long wait_ms;
32 
33 	wait_ms = gc_th->min_sleep_time;
34 
35 	do {
36 		if (try_to_freeze())
37 			continue;
38 		else
39 			wait_event_interruptible_timeout(*wq,
40 						kthread_should_stop(),
41 						msecs_to_jiffies(wait_ms));
42 		if (kthread_should_stop())
43 			break;
44 
45 		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
46 			increase_sleep_time(gc_th, &wait_ms);
47 			continue;
48 		}
49 
50 #ifdef CONFIG_F2FS_FAULT_INJECTION
51 		if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
52 			f2fs_show_injection_info(FAULT_CHECKPOINT);
53 			f2fs_stop_checkpoint(sbi, false);
54 		}
55 #endif
56 
57 		/*
58 		 * [GC triggering condition]
59 		 * 0. GC is not conducted currently.
60 		 * 1. There are enough dirty segments.
61 		 * 2. IO subsystem is idle by checking the # of writeback pages.
62 		 * 3. IO subsystem is idle by checking the # of requests in
63 		 *    bdev's request list.
64 		 *
65 		 * Note) We have to avoid triggering GCs frequently.
66 		 * Because it is possible that some segments can be
67 		 * invalidated soon after by user update or deletion.
68 		 * So, I'd like to wait some time to collect dirty segments.
69 		 */
70 		if (!mutex_trylock(&sbi->gc_mutex))
71 			continue;
72 
73 		if (!is_idle(sbi)) {
74 			increase_sleep_time(gc_th, &wait_ms);
75 			mutex_unlock(&sbi->gc_mutex);
76 			continue;
77 		}
78 
79 		if (has_enough_invalid_blocks(sbi))
80 			decrease_sleep_time(gc_th, &wait_ms);
81 		else
82 			increase_sleep_time(gc_th, &wait_ms);
83 
84 		stat_inc_bggc_count(sbi);
85 
86 		/* if return value is not zero, no victim was selected */
87 		if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO))
88 			wait_ms = gc_th->no_gc_sleep_time;
89 
90 		trace_f2fs_background_gc(sbi->sb, wait_ms,
91 				prefree_segments(sbi), free_segments(sbi));
92 
93 		/* balancing f2fs's metadata periodically */
94 		f2fs_balance_fs_bg(sbi);
95 
96 	} while (!kthread_should_stop());
97 	return 0;
98 }
99 
100 int start_gc_thread(struct f2fs_sb_info *sbi)
101 {
102 	struct f2fs_gc_kthread *gc_th;
103 	dev_t dev = sbi->sb->s_bdev->bd_dev;
104 	int err = 0;
105 
106 	gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
107 	if (!gc_th) {
108 		err = -ENOMEM;
109 		goto out;
110 	}
111 
112 	gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
113 	gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
114 	gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
115 
116 	gc_th->gc_idle = 0;
117 
118 	sbi->gc_thread = gc_th;
119 	init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
120 	sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
121 			"f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
122 	if (IS_ERR(gc_th->f2fs_gc_task)) {
123 		err = PTR_ERR(gc_th->f2fs_gc_task);
124 		kfree(gc_th);
125 		sbi->gc_thread = NULL;
126 	}
127 out:
128 	return err;
129 }
130 
131 void stop_gc_thread(struct f2fs_sb_info *sbi)
132 {
133 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
134 	if (!gc_th)
135 		return;
136 	kthread_stop(gc_th->f2fs_gc_task);
137 	kfree(gc_th);
138 	sbi->gc_thread = NULL;
139 }
140 
141 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
142 {
143 	int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
144 
145 	if (gc_th && gc_th->gc_idle) {
146 		if (gc_th->gc_idle == 1)
147 			gc_mode = GC_CB;
148 		else if (gc_th->gc_idle == 2)
149 			gc_mode = GC_GREEDY;
150 	}
151 	return gc_mode;
152 }
153 
154 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
155 			int type, struct victim_sel_policy *p)
156 {
157 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
158 
159 	if (p->alloc_mode == SSR) {
160 		p->gc_mode = GC_GREEDY;
161 		p->dirty_segmap = dirty_i->dirty_segmap[type];
162 		p->max_search = dirty_i->nr_dirty[type];
163 		p->ofs_unit = 1;
164 	} else {
165 		p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
166 		p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
167 		p->max_search = dirty_i->nr_dirty[DIRTY];
168 		p->ofs_unit = sbi->segs_per_sec;
169 	}
170 
171 	/* we need to check every dirty segments in the FG_GC case */
172 	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
173 		p->max_search = sbi->max_victim_search;
174 
175 	/* let's select beginning hot/small space first */
176 	if (type == CURSEG_HOT_DATA || IS_NODESEG(type))
177 		p->offset = 0;
178 	else
179 		p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
180 }
181 
182 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
183 				struct victim_sel_policy *p)
184 {
185 	/* SSR allocates in a segment unit */
186 	if (p->alloc_mode == SSR)
187 		return sbi->blocks_per_seg;
188 	if (p->gc_mode == GC_GREEDY)
189 		return 2 * sbi->blocks_per_seg * p->ofs_unit;
190 	else if (p->gc_mode == GC_CB)
191 		return UINT_MAX;
192 	else /* No other gc_mode */
193 		return 0;
194 }
195 
196 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
197 {
198 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
199 	unsigned int secno;
200 
201 	/*
202 	 * If the gc_type is FG_GC, we can select victim segments
203 	 * selected by background GC before.
204 	 * Those segments guarantee they have small valid blocks.
205 	 */
206 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
207 		if (sec_usage_check(sbi, secno))
208 			continue;
209 
210 		if (no_fggc_candidate(sbi, secno))
211 			continue;
212 
213 		clear_bit(secno, dirty_i->victim_secmap);
214 		return GET_SEG_FROM_SEC(sbi, secno);
215 	}
216 	return NULL_SEGNO;
217 }
218 
219 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
220 {
221 	struct sit_info *sit_i = SIT_I(sbi);
222 	unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
223 	unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
224 	unsigned long long mtime = 0;
225 	unsigned int vblocks;
226 	unsigned char age = 0;
227 	unsigned char u;
228 	unsigned int i;
229 
230 	for (i = 0; i < sbi->segs_per_sec; i++)
231 		mtime += get_seg_entry(sbi, start + i)->mtime;
232 	vblocks = get_valid_blocks(sbi, segno, true);
233 
234 	mtime = div_u64(mtime, sbi->segs_per_sec);
235 	vblocks = div_u64(vblocks, sbi->segs_per_sec);
236 
237 	u = (vblocks * 100) >> sbi->log_blocks_per_seg;
238 
239 	/* Handle if the system time has changed by the user */
240 	if (mtime < sit_i->min_mtime)
241 		sit_i->min_mtime = mtime;
242 	if (mtime > sit_i->max_mtime)
243 		sit_i->max_mtime = mtime;
244 	if (sit_i->max_mtime != sit_i->min_mtime)
245 		age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
246 				sit_i->max_mtime - sit_i->min_mtime);
247 
248 	return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
249 }
250 
251 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
252 						unsigned int segno)
253 {
254 	unsigned int valid_blocks =
255 			get_valid_blocks(sbi, segno, true);
256 
257 	return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
258 				valid_blocks * 2 : valid_blocks;
259 }
260 
261 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
262 			unsigned int segno, struct victim_sel_policy *p)
263 {
264 	if (p->alloc_mode == SSR)
265 		return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
266 
267 	/* alloc_mode == LFS */
268 	if (p->gc_mode == GC_GREEDY)
269 		return get_greedy_cost(sbi, segno);
270 	else
271 		return get_cb_cost(sbi, segno);
272 }
273 
274 static unsigned int count_bits(const unsigned long *addr,
275 				unsigned int offset, unsigned int len)
276 {
277 	unsigned int end = offset + len, sum = 0;
278 
279 	while (offset < end) {
280 		if (test_bit(offset++, addr))
281 			++sum;
282 	}
283 	return sum;
284 }
285 
286 /*
287  * This function is called from two paths.
288  * One is garbage collection and the other is SSR segment selection.
289  * When it is called during GC, it just gets a victim segment
290  * and it does not remove it from dirty seglist.
291  * When it is called from SSR segment selection, it finds a segment
292  * which has minimum valid blocks and removes it from dirty seglist.
293  */
294 static int get_victim_by_default(struct f2fs_sb_info *sbi,
295 		unsigned int *result, int gc_type, int type, char alloc_mode)
296 {
297 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
298 	struct sit_info *sm = SIT_I(sbi);
299 	struct victim_sel_policy p;
300 	unsigned int secno, last_victim;
301 	unsigned int last_segment = MAIN_SEGS(sbi);
302 	unsigned int nsearched = 0;
303 
304 	mutex_lock(&dirty_i->seglist_lock);
305 
306 	p.alloc_mode = alloc_mode;
307 	select_policy(sbi, gc_type, type, &p);
308 
309 	p.min_segno = NULL_SEGNO;
310 	p.min_cost = get_max_cost(sbi, &p);
311 
312 	if (*result != NULL_SEGNO) {
313 		if (IS_DATASEG(get_seg_entry(sbi, *result)->type) &&
314 			get_valid_blocks(sbi, *result, false) &&
315 			!sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
316 			p.min_segno = *result;
317 		goto out;
318 	}
319 
320 	if (p.max_search == 0)
321 		goto out;
322 
323 	last_victim = sm->last_victim[p.gc_mode];
324 	if (p.alloc_mode == LFS && gc_type == FG_GC) {
325 		p.min_segno = check_bg_victims(sbi);
326 		if (p.min_segno != NULL_SEGNO)
327 			goto got_it;
328 	}
329 
330 	while (1) {
331 		unsigned long cost;
332 		unsigned int segno;
333 
334 		segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
335 		if (segno >= last_segment) {
336 			if (sm->last_victim[p.gc_mode]) {
337 				last_segment =
338 					sm->last_victim[p.gc_mode];
339 				sm->last_victim[p.gc_mode] = 0;
340 				p.offset = 0;
341 				continue;
342 			}
343 			break;
344 		}
345 
346 		p.offset = segno + p.ofs_unit;
347 		if (p.ofs_unit > 1) {
348 			p.offset -= segno % p.ofs_unit;
349 			nsearched += count_bits(p.dirty_segmap,
350 						p.offset - p.ofs_unit,
351 						p.ofs_unit);
352 		} else {
353 			nsearched++;
354 		}
355 
356 		secno = GET_SEC_FROM_SEG(sbi, segno);
357 
358 		if (sec_usage_check(sbi, secno))
359 			goto next;
360 		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
361 			goto next;
362 		if (gc_type == FG_GC && p.alloc_mode == LFS &&
363 					no_fggc_candidate(sbi, secno))
364 			goto next;
365 
366 		cost = get_gc_cost(sbi, segno, &p);
367 
368 		if (p.min_cost > cost) {
369 			p.min_segno = segno;
370 			p.min_cost = cost;
371 		}
372 next:
373 		if (nsearched >= p.max_search) {
374 			if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
375 				sm->last_victim[p.gc_mode] = last_victim + 1;
376 			else
377 				sm->last_victim[p.gc_mode] = segno + 1;
378 			sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
379 			break;
380 		}
381 	}
382 	if (p.min_segno != NULL_SEGNO) {
383 got_it:
384 		if (p.alloc_mode == LFS) {
385 			secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
386 			if (gc_type == FG_GC)
387 				sbi->cur_victim_sec = secno;
388 			else
389 				set_bit(secno, dirty_i->victim_secmap);
390 		}
391 		*result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
392 
393 		trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
394 				sbi->cur_victim_sec,
395 				prefree_segments(sbi), free_segments(sbi));
396 	}
397 out:
398 	mutex_unlock(&dirty_i->seglist_lock);
399 
400 	return (p.min_segno == NULL_SEGNO) ? 0 : 1;
401 }
402 
403 static const struct victim_selection default_v_ops = {
404 	.get_victim = get_victim_by_default,
405 };
406 
407 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
408 {
409 	struct inode_entry *ie;
410 
411 	ie = radix_tree_lookup(&gc_list->iroot, ino);
412 	if (ie)
413 		return ie->inode;
414 	return NULL;
415 }
416 
417 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
418 {
419 	struct inode_entry *new_ie;
420 
421 	if (inode == find_gc_inode(gc_list, inode->i_ino)) {
422 		iput(inode);
423 		return;
424 	}
425 	new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
426 	new_ie->inode = inode;
427 
428 	f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
429 	list_add_tail(&new_ie->list, &gc_list->ilist);
430 }
431 
432 static void put_gc_inode(struct gc_inode_list *gc_list)
433 {
434 	struct inode_entry *ie, *next_ie;
435 	list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
436 		radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
437 		iput(ie->inode);
438 		list_del(&ie->list);
439 		kmem_cache_free(inode_entry_slab, ie);
440 	}
441 }
442 
443 static int check_valid_map(struct f2fs_sb_info *sbi,
444 				unsigned int segno, int offset)
445 {
446 	struct sit_info *sit_i = SIT_I(sbi);
447 	struct seg_entry *sentry;
448 	int ret;
449 
450 	mutex_lock(&sit_i->sentry_lock);
451 	sentry = get_seg_entry(sbi, segno);
452 	ret = f2fs_test_bit(offset, sentry->cur_valid_map);
453 	mutex_unlock(&sit_i->sentry_lock);
454 	return ret;
455 }
456 
457 /*
458  * This function compares node address got in summary with that in NAT.
459  * On validity, copy that node with cold status, otherwise (invalid node)
460  * ignore that.
461  */
462 static void gc_node_segment(struct f2fs_sb_info *sbi,
463 		struct f2fs_summary *sum, unsigned int segno, int gc_type)
464 {
465 	struct f2fs_summary *entry;
466 	block_t start_addr;
467 	int off;
468 	int phase = 0;
469 
470 	start_addr = START_BLOCK(sbi, segno);
471 
472 next_step:
473 	entry = sum;
474 
475 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
476 		nid_t nid = le32_to_cpu(entry->nid);
477 		struct page *node_page;
478 		struct node_info ni;
479 
480 		/* stop BG_GC if there is not enough free sections. */
481 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
482 			return;
483 
484 		if (check_valid_map(sbi, segno, off) == 0)
485 			continue;
486 
487 		if (phase == 0) {
488 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
489 							META_NAT, true);
490 			continue;
491 		}
492 
493 		if (phase == 1) {
494 			ra_node_page(sbi, nid);
495 			continue;
496 		}
497 
498 		/* phase == 2 */
499 		node_page = get_node_page(sbi, nid);
500 		if (IS_ERR(node_page))
501 			continue;
502 
503 		/* block may become invalid during get_node_page */
504 		if (check_valid_map(sbi, segno, off) == 0) {
505 			f2fs_put_page(node_page, 1);
506 			continue;
507 		}
508 
509 		get_node_info(sbi, nid, &ni);
510 		if (ni.blk_addr != start_addr + off) {
511 			f2fs_put_page(node_page, 1);
512 			continue;
513 		}
514 
515 		move_node_page(node_page, gc_type);
516 		stat_inc_node_blk_count(sbi, 1, gc_type);
517 	}
518 
519 	if (++phase < 3)
520 		goto next_step;
521 }
522 
523 /*
524  * Calculate start block index indicating the given node offset.
525  * Be careful, caller should give this node offset only indicating direct node
526  * blocks. If any node offsets, which point the other types of node blocks such
527  * as indirect or double indirect node blocks, are given, it must be a caller's
528  * bug.
529  */
530 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
531 {
532 	unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
533 	unsigned int bidx;
534 
535 	if (node_ofs == 0)
536 		return 0;
537 
538 	if (node_ofs <= 2) {
539 		bidx = node_ofs - 1;
540 	} else if (node_ofs <= indirect_blks) {
541 		int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
542 		bidx = node_ofs - 2 - dec;
543 	} else {
544 		int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
545 		bidx = node_ofs - 5 - dec;
546 	}
547 	return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
548 }
549 
550 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
551 		struct node_info *dni, block_t blkaddr, unsigned int *nofs)
552 {
553 	struct page *node_page;
554 	nid_t nid;
555 	unsigned int ofs_in_node;
556 	block_t source_blkaddr;
557 
558 	nid = le32_to_cpu(sum->nid);
559 	ofs_in_node = le16_to_cpu(sum->ofs_in_node);
560 
561 	node_page = get_node_page(sbi, nid);
562 	if (IS_ERR(node_page))
563 		return false;
564 
565 	get_node_info(sbi, nid, dni);
566 
567 	if (sum->version != dni->version) {
568 		f2fs_msg(sbi->sb, KERN_WARNING,
569 				"%s: valid data with mismatched node version.",
570 				__func__);
571 		set_sbi_flag(sbi, SBI_NEED_FSCK);
572 	}
573 
574 	*nofs = ofs_of_node(node_page);
575 	source_blkaddr = datablock_addr(node_page, ofs_in_node);
576 	f2fs_put_page(node_page, 1);
577 
578 	if (source_blkaddr != blkaddr)
579 		return false;
580 	return true;
581 }
582 
583 static void move_encrypted_block(struct inode *inode, block_t bidx,
584 							unsigned int segno, int off)
585 {
586 	struct f2fs_io_info fio = {
587 		.sbi = F2FS_I_SB(inode),
588 		.type = DATA,
589 		.op = REQ_OP_READ,
590 		.op_flags = 0,
591 		.encrypted_page = NULL,
592 	};
593 	struct dnode_of_data dn;
594 	struct f2fs_summary sum;
595 	struct node_info ni;
596 	struct page *page;
597 	block_t newaddr;
598 	int err;
599 
600 	/* do not read out */
601 	page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
602 	if (!page)
603 		return;
604 
605 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
606 		goto out;
607 
608 	if (f2fs_is_atomic_file(inode))
609 		goto out;
610 
611 	set_new_dnode(&dn, inode, NULL, NULL, 0);
612 	err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
613 	if (err)
614 		goto out;
615 
616 	if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
617 		ClearPageUptodate(page);
618 		goto put_out;
619 	}
620 
621 	/*
622 	 * don't cache encrypted data into meta inode until previous dirty
623 	 * data were writebacked to avoid racing between GC and flush.
624 	 */
625 	f2fs_wait_on_page_writeback(page, DATA, true);
626 
627 	get_node_info(fio.sbi, dn.nid, &ni);
628 	set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
629 
630 	/* read page */
631 	fio.page = page;
632 	fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
633 
634 	allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
635 							&sum, CURSEG_COLD_DATA);
636 
637 	fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr,
638 					FGP_LOCK | FGP_CREAT, GFP_NOFS);
639 	if (!fio.encrypted_page) {
640 		err = -ENOMEM;
641 		goto recover_block;
642 	}
643 
644 	err = f2fs_submit_page_bio(&fio);
645 	if (err)
646 		goto put_page_out;
647 
648 	/* write page */
649 	lock_page(fio.encrypted_page);
650 
651 	if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
652 		err = -EIO;
653 		goto put_page_out;
654 	}
655 	if (unlikely(!PageUptodate(fio.encrypted_page))) {
656 		err = -EIO;
657 		goto put_page_out;
658 	}
659 
660 	set_page_dirty(fio.encrypted_page);
661 	f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
662 	if (clear_page_dirty_for_io(fio.encrypted_page))
663 		dec_page_count(fio.sbi, F2FS_DIRTY_META);
664 
665 	set_page_writeback(fio.encrypted_page);
666 
667 	/* allocate block address */
668 	f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
669 
670 	fio.op = REQ_OP_WRITE;
671 	fio.op_flags = REQ_SYNC;
672 	fio.new_blkaddr = newaddr;
673 	f2fs_submit_page_mbio(&fio);
674 
675 	f2fs_update_data_blkaddr(&dn, newaddr);
676 	set_inode_flag(inode, FI_APPEND_WRITE);
677 	if (page->index == 0)
678 		set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
679 put_page_out:
680 	f2fs_put_page(fio.encrypted_page, 1);
681 recover_block:
682 	if (err)
683 		__f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
684 								true, true);
685 put_out:
686 	f2fs_put_dnode(&dn);
687 out:
688 	f2fs_put_page(page, 1);
689 }
690 
691 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
692 							unsigned int segno, int off)
693 {
694 	struct page *page;
695 
696 	page = get_lock_data_page(inode, bidx, true);
697 	if (IS_ERR(page))
698 		return;
699 
700 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
701 		goto out;
702 
703 	if (f2fs_is_atomic_file(inode))
704 		goto out;
705 
706 	if (gc_type == BG_GC) {
707 		if (PageWriteback(page))
708 			goto out;
709 		set_page_dirty(page);
710 		set_cold_data(page);
711 	} else {
712 		struct f2fs_io_info fio = {
713 			.sbi = F2FS_I_SB(inode),
714 			.type = DATA,
715 			.op = REQ_OP_WRITE,
716 			.op_flags = REQ_SYNC,
717 			.old_blkaddr = NULL_ADDR,
718 			.page = page,
719 			.encrypted_page = NULL,
720 			.need_lock = true,
721 		};
722 		bool is_dirty = PageDirty(page);
723 		int err;
724 
725 retry:
726 		set_page_dirty(page);
727 		f2fs_wait_on_page_writeback(page, DATA, true);
728 		if (clear_page_dirty_for_io(page)) {
729 			inode_dec_dirty_pages(inode);
730 			remove_dirty_inode(inode);
731 		}
732 
733 		set_cold_data(page);
734 
735 		err = do_write_data_page(&fio);
736 		if (err == -ENOMEM && is_dirty) {
737 			congestion_wait(BLK_RW_ASYNC, HZ/50);
738 			goto retry;
739 		}
740 	}
741 out:
742 	f2fs_put_page(page, 1);
743 }
744 
745 /*
746  * This function tries to get parent node of victim data block, and identifies
747  * data block validity. If the block is valid, copy that with cold status and
748  * modify parent node.
749  * If the parent node is not valid or the data block address is different,
750  * the victim data block is ignored.
751  */
752 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
753 		struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
754 {
755 	struct super_block *sb = sbi->sb;
756 	struct f2fs_summary *entry;
757 	block_t start_addr;
758 	int off;
759 	int phase = 0;
760 
761 	start_addr = START_BLOCK(sbi, segno);
762 
763 next_step:
764 	entry = sum;
765 
766 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
767 		struct page *data_page;
768 		struct inode *inode;
769 		struct node_info dni; /* dnode info for the data */
770 		unsigned int ofs_in_node, nofs;
771 		block_t start_bidx;
772 		nid_t nid = le32_to_cpu(entry->nid);
773 
774 		/* stop BG_GC if there is not enough free sections. */
775 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
776 			return;
777 
778 		if (check_valid_map(sbi, segno, off) == 0)
779 			continue;
780 
781 		if (phase == 0) {
782 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
783 							META_NAT, true);
784 			continue;
785 		}
786 
787 		if (phase == 1) {
788 			ra_node_page(sbi, nid);
789 			continue;
790 		}
791 
792 		/* Get an inode by ino with checking validity */
793 		if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
794 			continue;
795 
796 		if (phase == 2) {
797 			ra_node_page(sbi, dni.ino);
798 			continue;
799 		}
800 
801 		ofs_in_node = le16_to_cpu(entry->ofs_in_node);
802 
803 		if (phase == 3) {
804 			inode = f2fs_iget(sb, dni.ino);
805 			if (IS_ERR(inode) || is_bad_inode(inode))
806 				continue;
807 
808 			/* if encrypted inode, let's go phase 3 */
809 			if (f2fs_encrypted_inode(inode) &&
810 						S_ISREG(inode->i_mode)) {
811 				add_gc_inode(gc_list, inode);
812 				continue;
813 			}
814 
815 			start_bidx = start_bidx_of_node(nofs, inode);
816 			data_page = get_read_data_page(inode,
817 					start_bidx + ofs_in_node, REQ_RAHEAD,
818 					true);
819 			if (IS_ERR(data_page)) {
820 				iput(inode);
821 				continue;
822 			}
823 
824 			f2fs_put_page(data_page, 0);
825 			add_gc_inode(gc_list, inode);
826 			continue;
827 		}
828 
829 		/* phase 4 */
830 		inode = find_gc_inode(gc_list, dni.ino);
831 		if (inode) {
832 			struct f2fs_inode_info *fi = F2FS_I(inode);
833 			bool locked = false;
834 
835 			if (S_ISREG(inode->i_mode)) {
836 				if (!down_write_trylock(&fi->dio_rwsem[READ]))
837 					continue;
838 				if (!down_write_trylock(
839 						&fi->dio_rwsem[WRITE])) {
840 					up_write(&fi->dio_rwsem[READ]);
841 					continue;
842 				}
843 				locked = true;
844 			}
845 
846 			start_bidx = start_bidx_of_node(nofs, inode)
847 								+ ofs_in_node;
848 			if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
849 				move_encrypted_block(inode, start_bidx, segno, off);
850 			else
851 				move_data_page(inode, start_bidx, gc_type, segno, off);
852 
853 			if (locked) {
854 				up_write(&fi->dio_rwsem[WRITE]);
855 				up_write(&fi->dio_rwsem[READ]);
856 			}
857 
858 			stat_inc_data_blk_count(sbi, 1, gc_type);
859 		}
860 	}
861 
862 	if (++phase < 5)
863 		goto next_step;
864 }
865 
866 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
867 			int gc_type)
868 {
869 	struct sit_info *sit_i = SIT_I(sbi);
870 	int ret;
871 
872 	mutex_lock(&sit_i->sentry_lock);
873 	ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
874 					      NO_CHECK_TYPE, LFS);
875 	mutex_unlock(&sit_i->sentry_lock);
876 	return ret;
877 }
878 
879 static int do_garbage_collect(struct f2fs_sb_info *sbi,
880 				unsigned int start_segno,
881 				struct gc_inode_list *gc_list, int gc_type)
882 {
883 	struct page *sum_page;
884 	struct f2fs_summary_block *sum;
885 	struct blk_plug plug;
886 	unsigned int segno = start_segno;
887 	unsigned int end_segno = start_segno + sbi->segs_per_sec;
888 	int sec_freed = 0;
889 	unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
890 						SUM_TYPE_DATA : SUM_TYPE_NODE;
891 
892 	/* readahead multi ssa blocks those have contiguous address */
893 	if (sbi->segs_per_sec > 1)
894 		ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
895 					sbi->segs_per_sec, META_SSA, true);
896 
897 	/* reference all summary page */
898 	while (segno < end_segno) {
899 		sum_page = get_sum_page(sbi, segno++);
900 		unlock_page(sum_page);
901 	}
902 
903 	blk_start_plug(&plug);
904 
905 	for (segno = start_segno; segno < end_segno; segno++) {
906 
907 		/* find segment summary of victim */
908 		sum_page = find_get_page(META_MAPPING(sbi),
909 					GET_SUM_BLOCK(sbi, segno));
910 		f2fs_put_page(sum_page, 0);
911 
912 		if (get_valid_blocks(sbi, segno, false) == 0 ||
913 				!PageUptodate(sum_page) ||
914 				unlikely(f2fs_cp_error(sbi)))
915 			goto next;
916 
917 		sum = page_address(sum_page);
918 		f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
919 
920 		/*
921 		 * this is to avoid deadlock:
922 		 * - lock_page(sum_page)         - f2fs_replace_block
923 		 *  - check_valid_map()            - mutex_lock(sentry_lock)
924 		 *   - mutex_lock(sentry_lock)     - change_curseg()
925 		 *                                  - lock_page(sum_page)
926 		 */
927 		if (type == SUM_TYPE_NODE)
928 			gc_node_segment(sbi, sum->entries, segno, gc_type);
929 		else
930 			gc_data_segment(sbi, sum->entries, gc_list, segno,
931 								gc_type);
932 
933 		stat_inc_seg_count(sbi, type, gc_type);
934 next:
935 		f2fs_put_page(sum_page, 0);
936 	}
937 
938 	if (gc_type == FG_GC)
939 		f2fs_submit_merged_bio(sbi,
940 				(type == SUM_TYPE_NODE) ? NODE : DATA, WRITE);
941 
942 	blk_finish_plug(&plug);
943 
944 	if (gc_type == FG_GC &&
945 		get_valid_blocks(sbi, start_segno, true) == 0)
946 		sec_freed = 1;
947 
948 	stat_inc_call_count(sbi->stat_info);
949 
950 	return sec_freed;
951 }
952 
953 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
954 			bool background, unsigned int segno)
955 {
956 	int gc_type = sync ? FG_GC : BG_GC;
957 	int sec_freed = 0;
958 	int ret = -EINVAL;
959 	struct cp_control cpc;
960 	unsigned int init_segno = segno;
961 	struct gc_inode_list gc_list = {
962 		.ilist = LIST_HEAD_INIT(gc_list.ilist),
963 		.iroot = RADIX_TREE_INIT(GFP_NOFS),
964 	};
965 
966 	cpc.reason = __get_cp_reason(sbi);
967 gc_more:
968 	if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
969 		goto stop;
970 	if (unlikely(f2fs_cp_error(sbi))) {
971 		ret = -EIO;
972 		goto stop;
973 	}
974 
975 	if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
976 		/*
977 		 * For example, if there are many prefree_segments below given
978 		 * threshold, we can make them free by checkpoint. Then, we
979 		 * secure free segments which doesn't need fggc any more.
980 		 */
981 		if (prefree_segments(sbi)) {
982 			ret = write_checkpoint(sbi, &cpc);
983 			if (ret)
984 				goto stop;
985 		}
986 		if (has_not_enough_free_secs(sbi, 0, 0))
987 			gc_type = FG_GC;
988 	}
989 
990 	/* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
991 	if (gc_type == BG_GC && !background)
992 		goto stop;
993 	if (!__get_victim(sbi, &segno, gc_type))
994 		goto stop;
995 	ret = 0;
996 
997 	if (do_garbage_collect(sbi, segno, &gc_list, gc_type) &&
998 			gc_type == FG_GC)
999 		sec_freed++;
1000 
1001 	if (gc_type == FG_GC)
1002 		sbi->cur_victim_sec = NULL_SEGNO;
1003 
1004 	if (!sync) {
1005 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1006 			segno = NULL_SEGNO;
1007 			goto gc_more;
1008 		}
1009 
1010 		if (gc_type == FG_GC)
1011 			ret = write_checkpoint(sbi, &cpc);
1012 	}
1013 stop:
1014 	SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1015 	SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1016 	mutex_unlock(&sbi->gc_mutex);
1017 
1018 	put_gc_inode(&gc_list);
1019 
1020 	if (sync)
1021 		ret = sec_freed ? 0 : -EAGAIN;
1022 	return ret;
1023 }
1024 
1025 void build_gc_manager(struct f2fs_sb_info *sbi)
1026 {
1027 	u64 main_count, resv_count, ovp_count;
1028 
1029 	DIRTY_I(sbi)->v_ops = &default_v_ops;
1030 
1031 	/* threshold of # of valid blocks in a section for victims of FG_GC */
1032 	main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1033 	resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1034 	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1035 
1036 	sbi->fggc_threshold = div64_u64((main_count - ovp_count) *
1037 				BLKS_PER_SEC(sbi), (main_count - resv_count));
1038 
1039 	/* give warm/cold data area from slower device */
1040 	if (sbi->s_ndevs && sbi->segs_per_sec == 1)
1041 		SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1042 				GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
1043 }
1044