xref: /openbmc/linux/fs/f2fs/gc.c (revision a34a3ed7)
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 	unsigned int wait_ms;
32 
33 	wait_ms = gc_th->min_sleep_time;
34 
35 	set_freezable();
36 	do {
37 		wait_event_interruptible_timeout(*wq,
38 				kthread_should_stop() || freezing(current) ||
39 				gc_th->gc_wake,
40 				msecs_to_jiffies(wait_ms));
41 
42 		/* give it a try one time */
43 		if (gc_th->gc_wake)
44 			gc_th->gc_wake = 0;
45 
46 		if (try_to_freeze())
47 			continue;
48 		if (kthread_should_stop())
49 			break;
50 
51 		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
52 			increase_sleep_time(gc_th, &wait_ms);
53 			continue;
54 		}
55 
56 #ifdef CONFIG_F2FS_FAULT_INJECTION
57 		if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
58 			f2fs_show_injection_info(FAULT_CHECKPOINT);
59 			f2fs_stop_checkpoint(sbi, false);
60 		}
61 #endif
62 
63 		if (!sb_start_write_trylock(sbi->sb))
64 			continue;
65 
66 		/*
67 		 * [GC triggering condition]
68 		 * 0. GC is not conducted currently.
69 		 * 1. There are enough dirty segments.
70 		 * 2. IO subsystem is idle by checking the # of writeback pages.
71 		 * 3. IO subsystem is idle by checking the # of requests in
72 		 *    bdev's request list.
73 		 *
74 		 * Note) We have to avoid triggering GCs frequently.
75 		 * Because it is possible that some segments can be
76 		 * invalidated soon after by user update or deletion.
77 		 * So, I'd like to wait some time to collect dirty segments.
78 		 */
79 		if (!mutex_trylock(&sbi->gc_mutex))
80 			goto next;
81 
82 		if (gc_th->gc_urgent) {
83 			wait_ms = gc_th->urgent_sleep_time;
84 			goto do_gc;
85 		}
86 
87 		if (!is_idle(sbi)) {
88 			increase_sleep_time(gc_th, &wait_ms);
89 			mutex_unlock(&sbi->gc_mutex);
90 			goto next;
91 		}
92 
93 		if (has_enough_invalid_blocks(sbi))
94 			decrease_sleep_time(gc_th, &wait_ms);
95 		else
96 			increase_sleep_time(gc_th, &wait_ms);
97 do_gc:
98 		stat_inc_bggc_count(sbi);
99 
100 		/* if return value is not zero, no victim was selected */
101 		if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO))
102 			wait_ms = gc_th->no_gc_sleep_time;
103 
104 		trace_f2fs_background_gc(sbi->sb, wait_ms,
105 				prefree_segments(sbi), free_segments(sbi));
106 
107 		/* balancing f2fs's metadata periodically */
108 		f2fs_balance_fs_bg(sbi);
109 next:
110 		sb_end_write(sbi->sb);
111 
112 	} while (!kthread_should_stop());
113 	return 0;
114 }
115 
116 int start_gc_thread(struct f2fs_sb_info *sbi)
117 {
118 	struct f2fs_gc_kthread *gc_th;
119 	dev_t dev = sbi->sb->s_bdev->bd_dev;
120 	int err = 0;
121 
122 	gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
123 	if (!gc_th) {
124 		err = -ENOMEM;
125 		goto out;
126 	}
127 
128 	gc_th->urgent_sleep_time = DEF_GC_THREAD_URGENT_SLEEP_TIME;
129 	gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
130 	gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
131 	gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
132 
133 	gc_th->gc_idle = 0;
134 	gc_th->gc_urgent = 0;
135 	gc_th->gc_wake= 0;
136 
137 	sbi->gc_thread = gc_th;
138 	init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
139 	sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
140 			"f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
141 	if (IS_ERR(gc_th->f2fs_gc_task)) {
142 		err = PTR_ERR(gc_th->f2fs_gc_task);
143 		kfree(gc_th);
144 		sbi->gc_thread = NULL;
145 	}
146 out:
147 	return err;
148 }
149 
150 void stop_gc_thread(struct f2fs_sb_info *sbi)
151 {
152 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
153 	if (!gc_th)
154 		return;
155 	kthread_stop(gc_th->f2fs_gc_task);
156 	kfree(gc_th);
157 	sbi->gc_thread = NULL;
158 }
159 
160 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
161 {
162 	int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
163 
164 	if (gc_th && gc_th->gc_idle) {
165 		if (gc_th->gc_idle == 1)
166 			gc_mode = GC_CB;
167 		else if (gc_th->gc_idle == 2)
168 			gc_mode = GC_GREEDY;
169 	}
170 	return gc_mode;
171 }
172 
173 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
174 			int type, struct victim_sel_policy *p)
175 {
176 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
177 
178 	if (p->alloc_mode == SSR) {
179 		p->gc_mode = GC_GREEDY;
180 		p->dirty_segmap = dirty_i->dirty_segmap[type];
181 		p->max_search = dirty_i->nr_dirty[type];
182 		p->ofs_unit = 1;
183 	} else {
184 		p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
185 		p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
186 		p->max_search = dirty_i->nr_dirty[DIRTY];
187 		p->ofs_unit = sbi->segs_per_sec;
188 	}
189 
190 	/* we need to check every dirty segments in the FG_GC case */
191 	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
192 		p->max_search = sbi->max_victim_search;
193 
194 	/* let's select beginning hot/small space first */
195 	if (type == CURSEG_HOT_DATA || IS_NODESEG(type))
196 		p->offset = 0;
197 	else
198 		p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
199 }
200 
201 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
202 				struct victim_sel_policy *p)
203 {
204 	/* SSR allocates in a segment unit */
205 	if (p->alloc_mode == SSR)
206 		return sbi->blocks_per_seg;
207 	if (p->gc_mode == GC_GREEDY)
208 		return 2 * sbi->blocks_per_seg * p->ofs_unit;
209 	else if (p->gc_mode == GC_CB)
210 		return UINT_MAX;
211 	else /* No other gc_mode */
212 		return 0;
213 }
214 
215 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
216 {
217 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
218 	unsigned int secno;
219 
220 	/*
221 	 * If the gc_type is FG_GC, we can select victim segments
222 	 * selected by background GC before.
223 	 * Those segments guarantee they have small valid blocks.
224 	 */
225 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
226 		if (sec_usage_check(sbi, secno))
227 			continue;
228 
229 		if (no_fggc_candidate(sbi, secno))
230 			continue;
231 
232 		clear_bit(secno, dirty_i->victim_secmap);
233 		return GET_SEG_FROM_SEC(sbi, secno);
234 	}
235 	return NULL_SEGNO;
236 }
237 
238 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
239 {
240 	struct sit_info *sit_i = SIT_I(sbi);
241 	unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
242 	unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
243 	unsigned long long mtime = 0;
244 	unsigned int vblocks;
245 	unsigned char age = 0;
246 	unsigned char u;
247 	unsigned int i;
248 
249 	for (i = 0; i < sbi->segs_per_sec; i++)
250 		mtime += get_seg_entry(sbi, start + i)->mtime;
251 	vblocks = get_valid_blocks(sbi, segno, true);
252 
253 	mtime = div_u64(mtime, sbi->segs_per_sec);
254 	vblocks = div_u64(vblocks, sbi->segs_per_sec);
255 
256 	u = (vblocks * 100) >> sbi->log_blocks_per_seg;
257 
258 	/* Handle if the system time has changed by the user */
259 	if (mtime < sit_i->min_mtime)
260 		sit_i->min_mtime = mtime;
261 	if (mtime > sit_i->max_mtime)
262 		sit_i->max_mtime = mtime;
263 	if (sit_i->max_mtime != sit_i->min_mtime)
264 		age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
265 				sit_i->max_mtime - sit_i->min_mtime);
266 
267 	return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
268 }
269 
270 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
271 			unsigned int segno, struct victim_sel_policy *p)
272 {
273 	if (p->alloc_mode == SSR)
274 		return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
275 
276 	/* alloc_mode == LFS */
277 	if (p->gc_mode == GC_GREEDY)
278 		return get_valid_blocks(sbi, segno, true);
279 	else
280 		return get_cb_cost(sbi, segno);
281 }
282 
283 static unsigned int count_bits(const unsigned long *addr,
284 				unsigned int offset, unsigned int len)
285 {
286 	unsigned int end = offset + len, sum = 0;
287 
288 	while (offset < end) {
289 		if (test_bit(offset++, addr))
290 			++sum;
291 	}
292 	return sum;
293 }
294 
295 /*
296  * This function is called from two paths.
297  * One is garbage collection and the other is SSR segment selection.
298  * When it is called during GC, it just gets a victim segment
299  * and it does not remove it from dirty seglist.
300  * When it is called from SSR segment selection, it finds a segment
301  * which has minimum valid blocks and removes it from dirty seglist.
302  */
303 static int get_victim_by_default(struct f2fs_sb_info *sbi,
304 		unsigned int *result, int gc_type, int type, char alloc_mode)
305 {
306 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
307 	struct sit_info *sm = SIT_I(sbi);
308 	struct victim_sel_policy p;
309 	unsigned int secno, last_victim;
310 	unsigned int last_segment = MAIN_SEGS(sbi);
311 	unsigned int nsearched = 0;
312 
313 	mutex_lock(&dirty_i->seglist_lock);
314 
315 	p.alloc_mode = alloc_mode;
316 	select_policy(sbi, gc_type, type, &p);
317 
318 	p.min_segno = NULL_SEGNO;
319 	p.min_cost = get_max_cost(sbi, &p);
320 
321 	if (*result != NULL_SEGNO) {
322 		if (IS_DATASEG(get_seg_entry(sbi, *result)->type) &&
323 			get_valid_blocks(sbi, *result, false) &&
324 			!sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
325 			p.min_segno = *result;
326 		goto out;
327 	}
328 
329 	if (p.max_search == 0)
330 		goto out;
331 
332 	last_victim = sm->last_victim[p.gc_mode];
333 	if (p.alloc_mode == LFS && gc_type == FG_GC) {
334 		p.min_segno = check_bg_victims(sbi);
335 		if (p.min_segno != NULL_SEGNO)
336 			goto got_it;
337 	}
338 
339 	while (1) {
340 		unsigned long cost;
341 		unsigned int segno;
342 
343 		segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
344 		if (segno >= last_segment) {
345 			if (sm->last_victim[p.gc_mode]) {
346 				last_segment =
347 					sm->last_victim[p.gc_mode];
348 				sm->last_victim[p.gc_mode] = 0;
349 				p.offset = 0;
350 				continue;
351 			}
352 			break;
353 		}
354 
355 		p.offset = segno + p.ofs_unit;
356 		if (p.ofs_unit > 1) {
357 			p.offset -= segno % p.ofs_unit;
358 			nsearched += count_bits(p.dirty_segmap,
359 						p.offset - p.ofs_unit,
360 						p.ofs_unit);
361 		} else {
362 			nsearched++;
363 		}
364 
365 		secno = GET_SEC_FROM_SEG(sbi, segno);
366 
367 		if (sec_usage_check(sbi, secno))
368 			goto next;
369 		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
370 			goto next;
371 		if (gc_type == FG_GC && p.alloc_mode == LFS &&
372 					no_fggc_candidate(sbi, secno))
373 			goto next;
374 
375 		cost = get_gc_cost(sbi, segno, &p);
376 
377 		if (p.min_cost > cost) {
378 			p.min_segno = segno;
379 			p.min_cost = cost;
380 		}
381 next:
382 		if (nsearched >= p.max_search) {
383 			if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
384 				sm->last_victim[p.gc_mode] = last_victim + 1;
385 			else
386 				sm->last_victim[p.gc_mode] = segno + 1;
387 			sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
388 			break;
389 		}
390 	}
391 	if (p.min_segno != NULL_SEGNO) {
392 got_it:
393 		if (p.alloc_mode == LFS) {
394 			secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
395 			if (gc_type == FG_GC)
396 				sbi->cur_victim_sec = secno;
397 			else
398 				set_bit(secno, dirty_i->victim_secmap);
399 		}
400 		*result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
401 
402 		trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
403 				sbi->cur_victim_sec,
404 				prefree_segments(sbi), free_segments(sbi));
405 	}
406 out:
407 	mutex_unlock(&dirty_i->seglist_lock);
408 
409 	return (p.min_segno == NULL_SEGNO) ? 0 : 1;
410 }
411 
412 static const struct victim_selection default_v_ops = {
413 	.get_victim = get_victim_by_default,
414 };
415 
416 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
417 {
418 	struct inode_entry *ie;
419 
420 	ie = radix_tree_lookup(&gc_list->iroot, ino);
421 	if (ie)
422 		return ie->inode;
423 	return NULL;
424 }
425 
426 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
427 {
428 	struct inode_entry *new_ie;
429 
430 	if (inode == find_gc_inode(gc_list, inode->i_ino)) {
431 		iput(inode);
432 		return;
433 	}
434 	new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
435 	new_ie->inode = inode;
436 
437 	f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
438 	list_add_tail(&new_ie->list, &gc_list->ilist);
439 }
440 
441 static void put_gc_inode(struct gc_inode_list *gc_list)
442 {
443 	struct inode_entry *ie, *next_ie;
444 	list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
445 		radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
446 		iput(ie->inode);
447 		list_del(&ie->list);
448 		kmem_cache_free(inode_entry_slab, ie);
449 	}
450 }
451 
452 static int check_valid_map(struct f2fs_sb_info *sbi,
453 				unsigned int segno, int offset)
454 {
455 	struct sit_info *sit_i = SIT_I(sbi);
456 	struct seg_entry *sentry;
457 	int ret;
458 
459 	down_read(&sit_i->sentry_lock);
460 	sentry = get_seg_entry(sbi, segno);
461 	ret = f2fs_test_bit(offset, sentry->cur_valid_map);
462 	up_read(&sit_i->sentry_lock);
463 	return ret;
464 }
465 
466 /*
467  * This function compares node address got in summary with that in NAT.
468  * On validity, copy that node with cold status, otherwise (invalid node)
469  * ignore that.
470  */
471 static void gc_node_segment(struct f2fs_sb_info *sbi,
472 		struct f2fs_summary *sum, unsigned int segno, int gc_type)
473 {
474 	struct f2fs_summary *entry;
475 	block_t start_addr;
476 	int off;
477 	int phase = 0;
478 
479 	start_addr = START_BLOCK(sbi, segno);
480 
481 next_step:
482 	entry = sum;
483 
484 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
485 		nid_t nid = le32_to_cpu(entry->nid);
486 		struct page *node_page;
487 		struct node_info ni;
488 
489 		/* stop BG_GC if there is not enough free sections. */
490 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
491 			return;
492 
493 		if (check_valid_map(sbi, segno, off) == 0)
494 			continue;
495 
496 		if (phase == 0) {
497 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
498 							META_NAT, true);
499 			continue;
500 		}
501 
502 		if (phase == 1) {
503 			ra_node_page(sbi, nid);
504 			continue;
505 		}
506 
507 		/* phase == 2 */
508 		node_page = get_node_page(sbi, nid);
509 		if (IS_ERR(node_page))
510 			continue;
511 
512 		/* block may become invalid during get_node_page */
513 		if (check_valid_map(sbi, segno, off) == 0) {
514 			f2fs_put_page(node_page, 1);
515 			continue;
516 		}
517 
518 		get_node_info(sbi, nid, &ni);
519 		if (ni.blk_addr != start_addr + off) {
520 			f2fs_put_page(node_page, 1);
521 			continue;
522 		}
523 
524 		move_node_page(node_page, gc_type);
525 		stat_inc_node_blk_count(sbi, 1, gc_type);
526 	}
527 
528 	if (++phase < 3)
529 		goto next_step;
530 }
531 
532 /*
533  * Calculate start block index indicating the given node offset.
534  * Be careful, caller should give this node offset only indicating direct node
535  * blocks. If any node offsets, which point the other types of node blocks such
536  * as indirect or double indirect node blocks, are given, it must be a caller's
537  * bug.
538  */
539 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
540 {
541 	unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
542 	unsigned int bidx;
543 
544 	if (node_ofs == 0)
545 		return 0;
546 
547 	if (node_ofs <= 2) {
548 		bidx = node_ofs - 1;
549 	} else if (node_ofs <= indirect_blks) {
550 		int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
551 		bidx = node_ofs - 2 - dec;
552 	} else {
553 		int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
554 		bidx = node_ofs - 5 - dec;
555 	}
556 	return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
557 }
558 
559 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
560 		struct node_info *dni, block_t blkaddr, unsigned int *nofs)
561 {
562 	struct page *node_page;
563 	nid_t nid;
564 	unsigned int ofs_in_node;
565 	block_t source_blkaddr;
566 
567 	nid = le32_to_cpu(sum->nid);
568 	ofs_in_node = le16_to_cpu(sum->ofs_in_node);
569 
570 	node_page = get_node_page(sbi, nid);
571 	if (IS_ERR(node_page))
572 		return false;
573 
574 	get_node_info(sbi, nid, dni);
575 
576 	if (sum->version != dni->version) {
577 		f2fs_msg(sbi->sb, KERN_WARNING,
578 				"%s: valid data with mismatched node version.",
579 				__func__);
580 		set_sbi_flag(sbi, SBI_NEED_FSCK);
581 	}
582 
583 	*nofs = ofs_of_node(node_page);
584 	source_blkaddr = datablock_addr(NULL, node_page, ofs_in_node);
585 	f2fs_put_page(node_page, 1);
586 
587 	if (source_blkaddr != blkaddr)
588 		return false;
589 	return true;
590 }
591 
592 /*
593  * Move data block via META_MAPPING while keeping locked data page.
594  * This can be used to move blocks, aka LBAs, directly on disk.
595  */
596 static void move_data_block(struct inode *inode, block_t bidx,
597 					unsigned int segno, int off)
598 {
599 	struct f2fs_io_info fio = {
600 		.sbi = F2FS_I_SB(inode),
601 		.ino = inode->i_ino,
602 		.type = DATA,
603 		.temp = COLD,
604 		.op = REQ_OP_READ,
605 		.op_flags = 0,
606 		.encrypted_page = NULL,
607 		.in_list = false,
608 	};
609 	struct dnode_of_data dn;
610 	struct f2fs_summary sum;
611 	struct node_info ni;
612 	struct page *page;
613 	block_t newaddr;
614 	int err;
615 
616 	/* do not read out */
617 	page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
618 	if (!page)
619 		return;
620 
621 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
622 		goto out;
623 
624 	if (f2fs_is_atomic_file(inode))
625 		goto out;
626 
627 	if (f2fs_is_pinned_file(inode)) {
628 		f2fs_pin_file_control(inode, true);
629 		goto out;
630 	}
631 
632 	set_new_dnode(&dn, inode, NULL, NULL, 0);
633 	err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
634 	if (err)
635 		goto out;
636 
637 	if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
638 		ClearPageUptodate(page);
639 		goto put_out;
640 	}
641 
642 	/*
643 	 * don't cache encrypted data into meta inode until previous dirty
644 	 * data were writebacked to avoid racing between GC and flush.
645 	 */
646 	f2fs_wait_on_page_writeback(page, DATA, true);
647 
648 	get_node_info(fio.sbi, dn.nid, &ni);
649 	set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
650 
651 	/* read page */
652 	fio.page = page;
653 	fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
654 
655 	allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
656 					&sum, CURSEG_COLD_DATA, NULL, false);
657 
658 	fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(fio.sbi),
659 				newaddr, FGP_LOCK | FGP_CREAT, GFP_NOFS);
660 	if (!fio.encrypted_page) {
661 		err = -ENOMEM;
662 		goto recover_block;
663 	}
664 
665 	err = f2fs_submit_page_bio(&fio);
666 	if (err)
667 		goto put_page_out;
668 
669 	/* write page */
670 	lock_page(fio.encrypted_page);
671 
672 	if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
673 		err = -EIO;
674 		goto put_page_out;
675 	}
676 	if (unlikely(!PageUptodate(fio.encrypted_page))) {
677 		err = -EIO;
678 		goto put_page_out;
679 	}
680 
681 	set_page_dirty(fio.encrypted_page);
682 	f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
683 	if (clear_page_dirty_for_io(fio.encrypted_page))
684 		dec_page_count(fio.sbi, F2FS_DIRTY_META);
685 
686 	set_page_writeback(fio.encrypted_page);
687 
688 	/* allocate block address */
689 	f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
690 
691 	fio.op = REQ_OP_WRITE;
692 	fio.op_flags = REQ_SYNC;
693 	fio.new_blkaddr = newaddr;
694 	err = f2fs_submit_page_write(&fio);
695 	if (err) {
696 		if (PageWriteback(fio.encrypted_page))
697 			end_page_writeback(fio.encrypted_page);
698 		goto put_page_out;
699 	}
700 
701 	f2fs_update_iostat(fio.sbi, FS_GC_DATA_IO, F2FS_BLKSIZE);
702 
703 	f2fs_update_data_blkaddr(&dn, newaddr);
704 	set_inode_flag(inode, FI_APPEND_WRITE);
705 	if (page->index == 0)
706 		set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
707 put_page_out:
708 	f2fs_put_page(fio.encrypted_page, 1);
709 recover_block:
710 	if (err)
711 		__f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
712 								true, true);
713 put_out:
714 	f2fs_put_dnode(&dn);
715 out:
716 	f2fs_put_page(page, 1);
717 }
718 
719 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
720 							unsigned int segno, int off)
721 {
722 	struct page *page;
723 
724 	page = get_lock_data_page(inode, bidx, true);
725 	if (IS_ERR(page))
726 		return;
727 
728 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
729 		goto out;
730 
731 	if (f2fs_is_atomic_file(inode))
732 		goto out;
733 	if (f2fs_is_pinned_file(inode)) {
734 		if (gc_type == FG_GC)
735 			f2fs_pin_file_control(inode, true);
736 		goto out;
737 	}
738 
739 	if (gc_type == BG_GC) {
740 		if (PageWriteback(page))
741 			goto out;
742 		set_page_dirty(page);
743 		set_cold_data(page);
744 	} else {
745 		struct f2fs_io_info fio = {
746 			.sbi = F2FS_I_SB(inode),
747 			.ino = inode->i_ino,
748 			.type = DATA,
749 			.temp = COLD,
750 			.op = REQ_OP_WRITE,
751 			.op_flags = REQ_SYNC,
752 			.old_blkaddr = NULL_ADDR,
753 			.page = page,
754 			.encrypted_page = NULL,
755 			.need_lock = LOCK_REQ,
756 			.io_type = FS_GC_DATA_IO,
757 		};
758 		bool is_dirty = PageDirty(page);
759 		int err;
760 
761 retry:
762 		set_page_dirty(page);
763 		f2fs_wait_on_page_writeback(page, DATA, true);
764 		if (clear_page_dirty_for_io(page)) {
765 			inode_dec_dirty_pages(inode);
766 			remove_dirty_inode(inode);
767 		}
768 
769 		set_cold_data(page);
770 
771 		err = do_write_data_page(&fio);
772 		if (err == -ENOMEM && is_dirty) {
773 			congestion_wait(BLK_RW_ASYNC, HZ/50);
774 			goto retry;
775 		}
776 	}
777 out:
778 	f2fs_put_page(page, 1);
779 }
780 
781 /*
782  * This function tries to get parent node of victim data block, and identifies
783  * data block validity. If the block is valid, copy that with cold status and
784  * modify parent node.
785  * If the parent node is not valid or the data block address is different,
786  * the victim data block is ignored.
787  */
788 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
789 		struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
790 {
791 	struct super_block *sb = sbi->sb;
792 	struct f2fs_summary *entry;
793 	block_t start_addr;
794 	int off;
795 	int phase = 0;
796 
797 	start_addr = START_BLOCK(sbi, segno);
798 
799 next_step:
800 	entry = sum;
801 
802 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
803 		struct page *data_page;
804 		struct inode *inode;
805 		struct node_info dni; /* dnode info for the data */
806 		unsigned int ofs_in_node, nofs;
807 		block_t start_bidx;
808 		nid_t nid = le32_to_cpu(entry->nid);
809 
810 		/* stop BG_GC if there is not enough free sections. */
811 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
812 			return;
813 
814 		if (check_valid_map(sbi, segno, off) == 0)
815 			continue;
816 
817 		if (phase == 0) {
818 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
819 							META_NAT, true);
820 			continue;
821 		}
822 
823 		if (phase == 1) {
824 			ra_node_page(sbi, nid);
825 			continue;
826 		}
827 
828 		/* Get an inode by ino with checking validity */
829 		if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
830 			continue;
831 
832 		if (phase == 2) {
833 			ra_node_page(sbi, dni.ino);
834 			continue;
835 		}
836 
837 		ofs_in_node = le16_to_cpu(entry->ofs_in_node);
838 
839 		if (phase == 3) {
840 			inode = f2fs_iget(sb, dni.ino);
841 			if (IS_ERR(inode) || is_bad_inode(inode))
842 				continue;
843 
844 			/* if encrypted inode, let's go phase 3 */
845 			if (f2fs_encrypted_file(inode)) {
846 				add_gc_inode(gc_list, inode);
847 				continue;
848 			}
849 
850 			if (!down_write_trylock(
851 				&F2FS_I(inode)->dio_rwsem[WRITE])) {
852 				iput(inode);
853 				continue;
854 			}
855 
856 			start_bidx = start_bidx_of_node(nofs, inode);
857 			data_page = get_read_data_page(inode,
858 					start_bidx + ofs_in_node, REQ_RAHEAD,
859 					true);
860 			up_write(&F2FS_I(inode)->dio_rwsem[WRITE]);
861 			if (IS_ERR(data_page)) {
862 				iput(inode);
863 				continue;
864 			}
865 
866 			f2fs_put_page(data_page, 0);
867 			add_gc_inode(gc_list, inode);
868 			continue;
869 		}
870 
871 		/* phase 4 */
872 		inode = find_gc_inode(gc_list, dni.ino);
873 		if (inode) {
874 			struct f2fs_inode_info *fi = F2FS_I(inode);
875 			bool locked = false;
876 
877 			if (S_ISREG(inode->i_mode)) {
878 				if (!down_write_trylock(&fi->dio_rwsem[READ]))
879 					continue;
880 				if (!down_write_trylock(
881 						&fi->dio_rwsem[WRITE])) {
882 					up_write(&fi->dio_rwsem[READ]);
883 					continue;
884 				}
885 				locked = true;
886 
887 				/* wait for all inflight aio data */
888 				inode_dio_wait(inode);
889 			}
890 
891 			start_bidx = start_bidx_of_node(nofs, inode)
892 								+ ofs_in_node;
893 			if (f2fs_encrypted_file(inode))
894 				move_data_block(inode, start_bidx, segno, off);
895 			else
896 				move_data_page(inode, start_bidx, gc_type,
897 								segno, off);
898 
899 			if (locked) {
900 				up_write(&fi->dio_rwsem[WRITE]);
901 				up_write(&fi->dio_rwsem[READ]);
902 			}
903 
904 			stat_inc_data_blk_count(sbi, 1, gc_type);
905 		}
906 	}
907 
908 	if (++phase < 5)
909 		goto next_step;
910 }
911 
912 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
913 			int gc_type)
914 {
915 	struct sit_info *sit_i = SIT_I(sbi);
916 	int ret;
917 
918 	down_write(&sit_i->sentry_lock);
919 	ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
920 					      NO_CHECK_TYPE, LFS);
921 	up_write(&sit_i->sentry_lock);
922 	return ret;
923 }
924 
925 static int do_garbage_collect(struct f2fs_sb_info *sbi,
926 				unsigned int start_segno,
927 				struct gc_inode_list *gc_list, int gc_type)
928 {
929 	struct page *sum_page;
930 	struct f2fs_summary_block *sum;
931 	struct blk_plug plug;
932 	unsigned int segno = start_segno;
933 	unsigned int end_segno = start_segno + sbi->segs_per_sec;
934 	int seg_freed = 0;
935 	unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
936 						SUM_TYPE_DATA : SUM_TYPE_NODE;
937 
938 	/* readahead multi ssa blocks those have contiguous address */
939 	if (sbi->segs_per_sec > 1)
940 		ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
941 					sbi->segs_per_sec, META_SSA, true);
942 
943 	/* reference all summary page */
944 	while (segno < end_segno) {
945 		sum_page = get_sum_page(sbi, segno++);
946 		unlock_page(sum_page);
947 	}
948 
949 	blk_start_plug(&plug);
950 
951 	for (segno = start_segno; segno < end_segno; segno++) {
952 
953 		/* find segment summary of victim */
954 		sum_page = find_get_page(META_MAPPING(sbi),
955 					GET_SUM_BLOCK(sbi, segno));
956 		f2fs_put_page(sum_page, 0);
957 
958 		if (get_valid_blocks(sbi, segno, false) == 0 ||
959 				!PageUptodate(sum_page) ||
960 				unlikely(f2fs_cp_error(sbi)))
961 			goto next;
962 
963 		sum = page_address(sum_page);
964 		f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
965 
966 		/*
967 		 * this is to avoid deadlock:
968 		 * - lock_page(sum_page)         - f2fs_replace_block
969 		 *  - check_valid_map()            - down_write(sentry_lock)
970 		 *   - down_read(sentry_lock)     - change_curseg()
971 		 *                                  - lock_page(sum_page)
972 		 */
973 		if (type == SUM_TYPE_NODE)
974 			gc_node_segment(sbi, sum->entries, segno, gc_type);
975 		else
976 			gc_data_segment(sbi, sum->entries, gc_list, segno,
977 								gc_type);
978 
979 		stat_inc_seg_count(sbi, type, gc_type);
980 
981 		if (gc_type == FG_GC &&
982 				get_valid_blocks(sbi, segno, false) == 0)
983 			seg_freed++;
984 next:
985 		f2fs_put_page(sum_page, 0);
986 	}
987 
988 	if (gc_type == FG_GC)
989 		f2fs_submit_merged_write(sbi,
990 				(type == SUM_TYPE_NODE) ? NODE : DATA);
991 
992 	blk_finish_plug(&plug);
993 
994 	stat_inc_call_count(sbi->stat_info);
995 
996 	return seg_freed;
997 }
998 
999 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
1000 			bool background, unsigned int segno)
1001 {
1002 	int gc_type = sync ? FG_GC : BG_GC;
1003 	int sec_freed = 0, seg_freed = 0, total_freed = 0;
1004 	int ret = 0;
1005 	struct cp_control cpc;
1006 	unsigned int init_segno = segno;
1007 	struct gc_inode_list gc_list = {
1008 		.ilist = LIST_HEAD_INIT(gc_list.ilist),
1009 		.iroot = RADIX_TREE_INIT(GFP_NOFS),
1010 	};
1011 
1012 	trace_f2fs_gc_begin(sbi->sb, sync, background,
1013 				get_pages(sbi, F2FS_DIRTY_NODES),
1014 				get_pages(sbi, F2FS_DIRTY_DENTS),
1015 				get_pages(sbi, F2FS_DIRTY_IMETA),
1016 				free_sections(sbi),
1017 				free_segments(sbi),
1018 				reserved_segments(sbi),
1019 				prefree_segments(sbi));
1020 
1021 	cpc.reason = __get_cp_reason(sbi);
1022 gc_more:
1023 	if (unlikely(!(sbi->sb->s_flags & SB_ACTIVE))) {
1024 		ret = -EINVAL;
1025 		goto stop;
1026 	}
1027 	if (unlikely(f2fs_cp_error(sbi))) {
1028 		ret = -EIO;
1029 		goto stop;
1030 	}
1031 
1032 	if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
1033 		/*
1034 		 * For example, if there are many prefree_segments below given
1035 		 * threshold, we can make them free by checkpoint. Then, we
1036 		 * secure free segments which doesn't need fggc any more.
1037 		 */
1038 		if (prefree_segments(sbi)) {
1039 			ret = write_checkpoint(sbi, &cpc);
1040 			if (ret)
1041 				goto stop;
1042 		}
1043 		if (has_not_enough_free_secs(sbi, 0, 0))
1044 			gc_type = FG_GC;
1045 	}
1046 
1047 	/* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
1048 	if (gc_type == BG_GC && !background) {
1049 		ret = -EINVAL;
1050 		goto stop;
1051 	}
1052 	if (!__get_victim(sbi, &segno, gc_type)) {
1053 		ret = -ENODATA;
1054 		goto stop;
1055 	}
1056 
1057 	seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
1058 	if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
1059 		sec_freed++;
1060 	total_freed += seg_freed;
1061 
1062 	if (gc_type == FG_GC)
1063 		sbi->cur_victim_sec = NULL_SEGNO;
1064 
1065 	if (!sync) {
1066 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1067 			segno = NULL_SEGNO;
1068 			goto gc_more;
1069 		}
1070 
1071 		if (gc_type == FG_GC)
1072 			ret = write_checkpoint(sbi, &cpc);
1073 	}
1074 stop:
1075 	SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1076 	SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1077 
1078 	trace_f2fs_gc_end(sbi->sb, ret, total_freed, sec_freed,
1079 				get_pages(sbi, F2FS_DIRTY_NODES),
1080 				get_pages(sbi, F2FS_DIRTY_DENTS),
1081 				get_pages(sbi, F2FS_DIRTY_IMETA),
1082 				free_sections(sbi),
1083 				free_segments(sbi),
1084 				reserved_segments(sbi),
1085 				prefree_segments(sbi));
1086 
1087 	mutex_unlock(&sbi->gc_mutex);
1088 
1089 	put_gc_inode(&gc_list);
1090 
1091 	if (sync)
1092 		ret = sec_freed ? 0 : -EAGAIN;
1093 	return ret;
1094 }
1095 
1096 void build_gc_manager(struct f2fs_sb_info *sbi)
1097 {
1098 	u64 main_count, resv_count, ovp_count;
1099 
1100 	DIRTY_I(sbi)->v_ops = &default_v_ops;
1101 
1102 	/* threshold of # of valid blocks in a section for victims of FG_GC */
1103 	main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1104 	resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1105 	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1106 
1107 	sbi->fggc_threshold = div64_u64((main_count - ovp_count) *
1108 				BLKS_PER_SEC(sbi), (main_count - resv_count));
1109 	sbi->gc_pin_file_threshold = DEF_GC_FAILED_PINNED_FILES;
1110 
1111 	/* give warm/cold data area from slower device */
1112 	if (sbi->s_ndevs && sbi->segs_per_sec == 1)
1113 		SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1114 				GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
1115 }
1116