xref: /openbmc/linux/fs/f2fs/gc.c (revision 5d0e4d78)
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 	set_freezable();
36 	do {
37 		wait_event_interruptible_timeout(*wq,
38 				kthread_should_stop() || freezing(current),
39 				msecs_to_jiffies(wait_ms));
40 
41 		if (try_to_freeze())
42 			continue;
43 		if (kthread_should_stop())
44 			break;
45 
46 		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
47 			increase_sleep_time(gc_th, &wait_ms);
48 			continue;
49 		}
50 
51 #ifdef CONFIG_F2FS_FAULT_INJECTION
52 		if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
53 			f2fs_show_injection_info(FAULT_CHECKPOINT);
54 			f2fs_stop_checkpoint(sbi, false);
55 		}
56 #endif
57 
58 		/*
59 		 * [GC triggering condition]
60 		 * 0. GC is not conducted currently.
61 		 * 1. There are enough dirty segments.
62 		 * 2. IO subsystem is idle by checking the # of writeback pages.
63 		 * 3. IO subsystem is idle by checking the # of requests in
64 		 *    bdev's request list.
65 		 *
66 		 * Note) We have to avoid triggering GCs frequently.
67 		 * Because it is possible that some segments can be
68 		 * invalidated soon after by user update or deletion.
69 		 * So, I'd like to wait some time to collect dirty segments.
70 		 */
71 		if (!mutex_trylock(&sbi->gc_mutex))
72 			continue;
73 
74 		if (!is_idle(sbi)) {
75 			increase_sleep_time(gc_th, &wait_ms);
76 			mutex_unlock(&sbi->gc_mutex);
77 			continue;
78 		}
79 
80 		if (has_enough_invalid_blocks(sbi))
81 			decrease_sleep_time(gc_th, &wait_ms);
82 		else
83 			increase_sleep_time(gc_th, &wait_ms);
84 
85 		stat_inc_bggc_count(sbi);
86 
87 		/* if return value is not zero, no victim was selected */
88 		if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO))
89 			wait_ms = gc_th->no_gc_sleep_time;
90 
91 		trace_f2fs_background_gc(sbi->sb, wait_ms,
92 				prefree_segments(sbi), free_segments(sbi));
93 
94 		/* balancing f2fs's metadata periodically */
95 		f2fs_balance_fs_bg(sbi);
96 
97 	} while (!kthread_should_stop());
98 	return 0;
99 }
100 
101 int start_gc_thread(struct f2fs_sb_info *sbi)
102 {
103 	struct f2fs_gc_kthread *gc_th;
104 	dev_t dev = sbi->sb->s_bdev->bd_dev;
105 	int err = 0;
106 
107 	gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
108 	if (!gc_th) {
109 		err = -ENOMEM;
110 		goto out;
111 	}
112 
113 	gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
114 	gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
115 	gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
116 
117 	gc_th->gc_idle = 0;
118 
119 	sbi->gc_thread = gc_th;
120 	init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
121 	sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
122 			"f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
123 	if (IS_ERR(gc_th->f2fs_gc_task)) {
124 		err = PTR_ERR(gc_th->f2fs_gc_task);
125 		kfree(gc_th);
126 		sbi->gc_thread = NULL;
127 	}
128 out:
129 	return err;
130 }
131 
132 void stop_gc_thread(struct f2fs_sb_info *sbi)
133 {
134 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
135 	if (!gc_th)
136 		return;
137 	kthread_stop(gc_th->f2fs_gc_task);
138 	kfree(gc_th);
139 	sbi->gc_thread = NULL;
140 }
141 
142 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
143 {
144 	int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
145 
146 	if (gc_th && gc_th->gc_idle) {
147 		if (gc_th->gc_idle == 1)
148 			gc_mode = GC_CB;
149 		else if (gc_th->gc_idle == 2)
150 			gc_mode = GC_GREEDY;
151 	}
152 	return gc_mode;
153 }
154 
155 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
156 			int type, struct victim_sel_policy *p)
157 {
158 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
159 
160 	if (p->alloc_mode == SSR) {
161 		p->gc_mode = GC_GREEDY;
162 		p->dirty_segmap = dirty_i->dirty_segmap[type];
163 		p->max_search = dirty_i->nr_dirty[type];
164 		p->ofs_unit = 1;
165 	} else {
166 		p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
167 		p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
168 		p->max_search = dirty_i->nr_dirty[DIRTY];
169 		p->ofs_unit = sbi->segs_per_sec;
170 	}
171 
172 	/* we need to check every dirty segments in the FG_GC case */
173 	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
174 		p->max_search = sbi->max_victim_search;
175 
176 	/* let's select beginning hot/small space first */
177 	if (type == CURSEG_HOT_DATA || IS_NODESEG(type))
178 		p->offset = 0;
179 	else
180 		p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
181 }
182 
183 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
184 				struct victim_sel_policy *p)
185 {
186 	/* SSR allocates in a segment unit */
187 	if (p->alloc_mode == SSR)
188 		return sbi->blocks_per_seg;
189 	if (p->gc_mode == GC_GREEDY)
190 		return 2 * sbi->blocks_per_seg * p->ofs_unit;
191 	else if (p->gc_mode == GC_CB)
192 		return UINT_MAX;
193 	else /* No other gc_mode */
194 		return 0;
195 }
196 
197 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
198 {
199 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
200 	unsigned int secno;
201 
202 	/*
203 	 * If the gc_type is FG_GC, we can select victim segments
204 	 * selected by background GC before.
205 	 * Those segments guarantee they have small valid blocks.
206 	 */
207 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
208 		if (sec_usage_check(sbi, secno))
209 			continue;
210 
211 		if (no_fggc_candidate(sbi, secno))
212 			continue;
213 
214 		clear_bit(secno, dirty_i->victim_secmap);
215 		return GET_SEG_FROM_SEC(sbi, secno);
216 	}
217 	return NULL_SEGNO;
218 }
219 
220 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
221 {
222 	struct sit_info *sit_i = SIT_I(sbi);
223 	unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
224 	unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
225 	unsigned long long mtime = 0;
226 	unsigned int vblocks;
227 	unsigned char age = 0;
228 	unsigned char u;
229 	unsigned int i;
230 
231 	for (i = 0; i < sbi->segs_per_sec; i++)
232 		mtime += get_seg_entry(sbi, start + i)->mtime;
233 	vblocks = get_valid_blocks(sbi, segno, true);
234 
235 	mtime = div_u64(mtime, sbi->segs_per_sec);
236 	vblocks = div_u64(vblocks, sbi->segs_per_sec);
237 
238 	u = (vblocks * 100) >> sbi->log_blocks_per_seg;
239 
240 	/* Handle if the system time has changed by the user */
241 	if (mtime < sit_i->min_mtime)
242 		sit_i->min_mtime = mtime;
243 	if (mtime > sit_i->max_mtime)
244 		sit_i->max_mtime = mtime;
245 	if (sit_i->max_mtime != sit_i->min_mtime)
246 		age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
247 				sit_i->max_mtime - sit_i->min_mtime);
248 
249 	return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
250 }
251 
252 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
253 						unsigned int segno)
254 {
255 	unsigned int valid_blocks =
256 			get_valid_blocks(sbi, segno, true);
257 
258 	return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
259 				valid_blocks * 2 : valid_blocks;
260 }
261 
262 static unsigned int get_ssr_cost(struct f2fs_sb_info *sbi,
263 						unsigned int segno)
264 {
265 	struct seg_entry *se = get_seg_entry(sbi, segno);
266 
267 	return se->ckpt_valid_blocks > se->valid_blocks ?
268 				se->ckpt_valid_blocks : se->valid_blocks;
269 }
270 
271 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
272 			unsigned int segno, struct victim_sel_policy *p)
273 {
274 	if (p->alloc_mode == SSR)
275 		return get_ssr_cost(sbi, segno);
276 
277 	/* alloc_mode == LFS */
278 	if (p->gc_mode == GC_GREEDY)
279 		return get_greedy_cost(sbi, segno);
280 	else
281 		return get_cb_cost(sbi, segno);
282 }
283 
284 static unsigned int count_bits(const unsigned long *addr,
285 				unsigned int offset, unsigned int len)
286 {
287 	unsigned int end = offset + len, sum = 0;
288 
289 	while (offset < end) {
290 		if (test_bit(offset++, addr))
291 			++sum;
292 	}
293 	return sum;
294 }
295 
296 /*
297  * This function is called from two paths.
298  * One is garbage collection and the other is SSR segment selection.
299  * When it is called during GC, it just gets a victim segment
300  * and it does not remove it from dirty seglist.
301  * When it is called from SSR segment selection, it finds a segment
302  * which has minimum valid blocks and removes it from dirty seglist.
303  */
304 static int get_victim_by_default(struct f2fs_sb_info *sbi,
305 		unsigned int *result, int gc_type, int type, char alloc_mode)
306 {
307 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
308 	struct sit_info *sm = SIT_I(sbi);
309 	struct victim_sel_policy p;
310 	unsigned int secno, last_victim;
311 	unsigned int last_segment = MAIN_SEGS(sbi);
312 	unsigned int nsearched = 0;
313 
314 	mutex_lock(&dirty_i->seglist_lock);
315 
316 	p.alloc_mode = alloc_mode;
317 	select_policy(sbi, gc_type, type, &p);
318 
319 	p.min_segno = NULL_SEGNO;
320 	p.min_cost = get_max_cost(sbi, &p);
321 
322 	if (*result != NULL_SEGNO) {
323 		if (IS_DATASEG(get_seg_entry(sbi, *result)->type) &&
324 			get_valid_blocks(sbi, *result, false) &&
325 			!sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
326 			p.min_segno = *result;
327 		goto out;
328 	}
329 
330 	if (p.max_search == 0)
331 		goto out;
332 
333 	last_victim = sm->last_victim[p.gc_mode];
334 	if (p.alloc_mode == LFS && gc_type == FG_GC) {
335 		p.min_segno = check_bg_victims(sbi);
336 		if (p.min_segno != NULL_SEGNO)
337 			goto got_it;
338 	}
339 
340 	while (1) {
341 		unsigned long cost;
342 		unsigned int segno;
343 
344 		segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
345 		if (segno >= last_segment) {
346 			if (sm->last_victim[p.gc_mode]) {
347 				last_segment =
348 					sm->last_victim[p.gc_mode];
349 				sm->last_victim[p.gc_mode] = 0;
350 				p.offset = 0;
351 				continue;
352 			}
353 			break;
354 		}
355 
356 		p.offset = segno + p.ofs_unit;
357 		if (p.ofs_unit > 1) {
358 			p.offset -= segno % p.ofs_unit;
359 			nsearched += count_bits(p.dirty_segmap,
360 						p.offset - p.ofs_unit,
361 						p.ofs_unit);
362 		} else {
363 			nsearched++;
364 		}
365 
366 		secno = GET_SEC_FROM_SEG(sbi, segno);
367 
368 		if (sec_usage_check(sbi, secno))
369 			goto next;
370 		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
371 			goto next;
372 		if (gc_type == FG_GC && p.alloc_mode == LFS &&
373 					no_fggc_candidate(sbi, secno))
374 			goto next;
375 
376 		cost = get_gc_cost(sbi, segno, &p);
377 
378 		if (p.min_cost > cost) {
379 			p.min_segno = segno;
380 			p.min_cost = cost;
381 		}
382 next:
383 		if (nsearched >= p.max_search) {
384 			if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
385 				sm->last_victim[p.gc_mode] = last_victim + 1;
386 			else
387 				sm->last_victim[p.gc_mode] = segno + 1;
388 			sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
389 			break;
390 		}
391 	}
392 	if (p.min_segno != NULL_SEGNO) {
393 got_it:
394 		if (p.alloc_mode == LFS) {
395 			secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
396 			if (gc_type == FG_GC)
397 				sbi->cur_victim_sec = secno;
398 			else
399 				set_bit(secno, dirty_i->victim_secmap);
400 		}
401 		*result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
402 
403 		trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
404 				sbi->cur_victim_sec,
405 				prefree_segments(sbi), free_segments(sbi));
406 	}
407 out:
408 	mutex_unlock(&dirty_i->seglist_lock);
409 
410 	return (p.min_segno == NULL_SEGNO) ? 0 : 1;
411 }
412 
413 static const struct victim_selection default_v_ops = {
414 	.get_victim = get_victim_by_default,
415 };
416 
417 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
418 {
419 	struct inode_entry *ie;
420 
421 	ie = radix_tree_lookup(&gc_list->iroot, ino);
422 	if (ie)
423 		return ie->inode;
424 	return NULL;
425 }
426 
427 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
428 {
429 	struct inode_entry *new_ie;
430 
431 	if (inode == find_gc_inode(gc_list, inode->i_ino)) {
432 		iput(inode);
433 		return;
434 	}
435 	new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
436 	new_ie->inode = inode;
437 
438 	f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
439 	list_add_tail(&new_ie->list, &gc_list->ilist);
440 }
441 
442 static void put_gc_inode(struct gc_inode_list *gc_list)
443 {
444 	struct inode_entry *ie, *next_ie;
445 	list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
446 		radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
447 		iput(ie->inode);
448 		list_del(&ie->list);
449 		kmem_cache_free(inode_entry_slab, ie);
450 	}
451 }
452 
453 static int check_valid_map(struct f2fs_sb_info *sbi,
454 				unsigned int segno, int offset)
455 {
456 	struct sit_info *sit_i = SIT_I(sbi);
457 	struct seg_entry *sentry;
458 	int ret;
459 
460 	mutex_lock(&sit_i->sentry_lock);
461 	sentry = get_seg_entry(sbi, segno);
462 	ret = f2fs_test_bit(offset, sentry->cur_valid_map);
463 	mutex_unlock(&sit_i->sentry_lock);
464 	return ret;
465 }
466 
467 /*
468  * This function compares node address got in summary with that in NAT.
469  * On validity, copy that node with cold status, otherwise (invalid node)
470  * ignore that.
471  */
472 static void gc_node_segment(struct f2fs_sb_info *sbi,
473 		struct f2fs_summary *sum, unsigned int segno, int gc_type)
474 {
475 	struct f2fs_summary *entry;
476 	block_t start_addr;
477 	int off;
478 	int phase = 0;
479 
480 	start_addr = START_BLOCK(sbi, segno);
481 
482 next_step:
483 	entry = sum;
484 
485 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
486 		nid_t nid = le32_to_cpu(entry->nid);
487 		struct page *node_page;
488 		struct node_info ni;
489 
490 		/* stop BG_GC if there is not enough free sections. */
491 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
492 			return;
493 
494 		if (check_valid_map(sbi, segno, off) == 0)
495 			continue;
496 
497 		if (phase == 0) {
498 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
499 							META_NAT, true);
500 			continue;
501 		}
502 
503 		if (phase == 1) {
504 			ra_node_page(sbi, nid);
505 			continue;
506 		}
507 
508 		/* phase == 2 */
509 		node_page = get_node_page(sbi, nid);
510 		if (IS_ERR(node_page))
511 			continue;
512 
513 		/* block may become invalid during get_node_page */
514 		if (check_valid_map(sbi, segno, off) == 0) {
515 			f2fs_put_page(node_page, 1);
516 			continue;
517 		}
518 
519 		get_node_info(sbi, nid, &ni);
520 		if (ni.blk_addr != start_addr + off) {
521 			f2fs_put_page(node_page, 1);
522 			continue;
523 		}
524 
525 		move_node_page(node_page, gc_type);
526 		stat_inc_node_blk_count(sbi, 1, gc_type);
527 	}
528 
529 	if (++phase < 3)
530 		goto next_step;
531 }
532 
533 /*
534  * Calculate start block index indicating the given node offset.
535  * Be careful, caller should give this node offset only indicating direct node
536  * blocks. If any node offsets, which point the other types of node blocks such
537  * as indirect or double indirect node blocks, are given, it must be a caller's
538  * bug.
539  */
540 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
541 {
542 	unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
543 	unsigned int bidx;
544 
545 	if (node_ofs == 0)
546 		return 0;
547 
548 	if (node_ofs <= 2) {
549 		bidx = node_ofs - 1;
550 	} else if (node_ofs <= indirect_blks) {
551 		int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
552 		bidx = node_ofs - 2 - dec;
553 	} else {
554 		int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
555 		bidx = node_ofs - 5 - dec;
556 	}
557 	return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
558 }
559 
560 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
561 		struct node_info *dni, block_t blkaddr, unsigned int *nofs)
562 {
563 	struct page *node_page;
564 	nid_t nid;
565 	unsigned int ofs_in_node;
566 	block_t source_blkaddr;
567 
568 	nid = le32_to_cpu(sum->nid);
569 	ofs_in_node = le16_to_cpu(sum->ofs_in_node);
570 
571 	node_page = get_node_page(sbi, nid);
572 	if (IS_ERR(node_page))
573 		return false;
574 
575 	get_node_info(sbi, nid, dni);
576 
577 	if (sum->version != dni->version) {
578 		f2fs_msg(sbi->sb, KERN_WARNING,
579 				"%s: valid data with mismatched node version.",
580 				__func__);
581 		set_sbi_flag(sbi, SBI_NEED_FSCK);
582 	}
583 
584 	*nofs = ofs_of_node(node_page);
585 	source_blkaddr = datablock_addr(node_page, ofs_in_node);
586 	f2fs_put_page(node_page, 1);
587 
588 	if (source_blkaddr != blkaddr)
589 		return false;
590 	return true;
591 }
592 
593 static void move_encrypted_block(struct inode *inode, block_t bidx,
594 							unsigned int segno, int off)
595 {
596 	struct f2fs_io_info fio = {
597 		.sbi = F2FS_I_SB(inode),
598 		.type = DATA,
599 		.temp = COLD,
600 		.op = REQ_OP_READ,
601 		.op_flags = 0,
602 		.encrypted_page = NULL,
603 		.in_list = false,
604 	};
605 	struct dnode_of_data dn;
606 	struct f2fs_summary sum;
607 	struct node_info ni;
608 	struct page *page;
609 	block_t newaddr;
610 	int err;
611 
612 	/* do not read out */
613 	page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
614 	if (!page)
615 		return;
616 
617 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
618 		goto out;
619 
620 	if (f2fs_is_atomic_file(inode))
621 		goto out;
622 
623 	set_new_dnode(&dn, inode, NULL, NULL, 0);
624 	err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
625 	if (err)
626 		goto out;
627 
628 	if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
629 		ClearPageUptodate(page);
630 		goto put_out;
631 	}
632 
633 	/*
634 	 * don't cache encrypted data into meta inode until previous dirty
635 	 * data were writebacked to avoid racing between GC and flush.
636 	 */
637 	f2fs_wait_on_page_writeback(page, DATA, true);
638 
639 	get_node_info(fio.sbi, dn.nid, &ni);
640 	set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
641 
642 	/* read page */
643 	fio.page = page;
644 	fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
645 
646 	allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
647 					&sum, CURSEG_COLD_DATA, NULL, false);
648 
649 	fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr,
650 					FGP_LOCK | FGP_CREAT, GFP_NOFS);
651 	if (!fio.encrypted_page) {
652 		err = -ENOMEM;
653 		goto recover_block;
654 	}
655 
656 	err = f2fs_submit_page_bio(&fio);
657 	if (err)
658 		goto put_page_out;
659 
660 	/* write page */
661 	lock_page(fio.encrypted_page);
662 
663 	if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
664 		err = -EIO;
665 		goto put_page_out;
666 	}
667 	if (unlikely(!PageUptodate(fio.encrypted_page))) {
668 		err = -EIO;
669 		goto put_page_out;
670 	}
671 
672 	set_page_dirty(fio.encrypted_page);
673 	f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
674 	if (clear_page_dirty_for_io(fio.encrypted_page))
675 		dec_page_count(fio.sbi, F2FS_DIRTY_META);
676 
677 	set_page_writeback(fio.encrypted_page);
678 
679 	/* allocate block address */
680 	f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
681 
682 	fio.op = REQ_OP_WRITE;
683 	fio.op_flags = REQ_SYNC;
684 	fio.new_blkaddr = newaddr;
685 	f2fs_submit_page_write(&fio);
686 
687 	f2fs_update_data_blkaddr(&dn, newaddr);
688 	set_inode_flag(inode, FI_APPEND_WRITE);
689 	if (page->index == 0)
690 		set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
691 put_page_out:
692 	f2fs_put_page(fio.encrypted_page, 1);
693 recover_block:
694 	if (err)
695 		__f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
696 								true, true);
697 put_out:
698 	f2fs_put_dnode(&dn);
699 out:
700 	f2fs_put_page(page, 1);
701 }
702 
703 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
704 							unsigned int segno, int off)
705 {
706 	struct page *page;
707 
708 	page = get_lock_data_page(inode, bidx, true);
709 	if (IS_ERR(page))
710 		return;
711 
712 	if (!check_valid_map(F2FS_I_SB(inode), segno, off))
713 		goto out;
714 
715 	if (f2fs_is_atomic_file(inode))
716 		goto out;
717 
718 	if (gc_type == BG_GC) {
719 		if (PageWriteback(page))
720 			goto out;
721 		set_page_dirty(page);
722 		set_cold_data(page);
723 	} else {
724 		struct f2fs_io_info fio = {
725 			.sbi = F2FS_I_SB(inode),
726 			.type = DATA,
727 			.temp = COLD,
728 			.op = REQ_OP_WRITE,
729 			.op_flags = REQ_SYNC,
730 			.old_blkaddr = NULL_ADDR,
731 			.page = page,
732 			.encrypted_page = NULL,
733 			.need_lock = LOCK_REQ,
734 		};
735 		bool is_dirty = PageDirty(page);
736 		int err;
737 
738 retry:
739 		set_page_dirty(page);
740 		f2fs_wait_on_page_writeback(page, DATA, true);
741 		if (clear_page_dirty_for_io(page)) {
742 			inode_dec_dirty_pages(inode);
743 			remove_dirty_inode(inode);
744 		}
745 
746 		set_cold_data(page);
747 
748 		err = do_write_data_page(&fio);
749 		if (err == -ENOMEM && is_dirty) {
750 			congestion_wait(BLK_RW_ASYNC, HZ/50);
751 			goto retry;
752 		}
753 	}
754 out:
755 	f2fs_put_page(page, 1);
756 }
757 
758 /*
759  * This function tries to get parent node of victim data block, and identifies
760  * data block validity. If the block is valid, copy that with cold status and
761  * modify parent node.
762  * If the parent node is not valid or the data block address is different,
763  * the victim data block is ignored.
764  */
765 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
766 		struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
767 {
768 	struct super_block *sb = sbi->sb;
769 	struct f2fs_summary *entry;
770 	block_t start_addr;
771 	int off;
772 	int phase = 0;
773 
774 	start_addr = START_BLOCK(sbi, segno);
775 
776 next_step:
777 	entry = sum;
778 
779 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
780 		struct page *data_page;
781 		struct inode *inode;
782 		struct node_info dni; /* dnode info for the data */
783 		unsigned int ofs_in_node, nofs;
784 		block_t start_bidx;
785 		nid_t nid = le32_to_cpu(entry->nid);
786 
787 		/* stop BG_GC if there is not enough free sections. */
788 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
789 			return;
790 
791 		if (check_valid_map(sbi, segno, off) == 0)
792 			continue;
793 
794 		if (phase == 0) {
795 			ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
796 							META_NAT, true);
797 			continue;
798 		}
799 
800 		if (phase == 1) {
801 			ra_node_page(sbi, nid);
802 			continue;
803 		}
804 
805 		/* Get an inode by ino with checking validity */
806 		if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
807 			continue;
808 
809 		if (phase == 2) {
810 			ra_node_page(sbi, dni.ino);
811 			continue;
812 		}
813 
814 		ofs_in_node = le16_to_cpu(entry->ofs_in_node);
815 
816 		if (phase == 3) {
817 			inode = f2fs_iget(sb, dni.ino);
818 			if (IS_ERR(inode) || is_bad_inode(inode))
819 				continue;
820 
821 			/* if encrypted inode, let's go phase 3 */
822 			if (f2fs_encrypted_inode(inode) &&
823 						S_ISREG(inode->i_mode)) {
824 				add_gc_inode(gc_list, inode);
825 				continue;
826 			}
827 
828 			start_bidx = start_bidx_of_node(nofs, inode);
829 			data_page = get_read_data_page(inode,
830 					start_bidx + ofs_in_node, REQ_RAHEAD,
831 					true);
832 			if (IS_ERR(data_page)) {
833 				iput(inode);
834 				continue;
835 			}
836 
837 			f2fs_put_page(data_page, 0);
838 			add_gc_inode(gc_list, inode);
839 			continue;
840 		}
841 
842 		/* phase 4 */
843 		inode = find_gc_inode(gc_list, dni.ino);
844 		if (inode) {
845 			struct f2fs_inode_info *fi = F2FS_I(inode);
846 			bool locked = false;
847 
848 			if (S_ISREG(inode->i_mode)) {
849 				if (!down_write_trylock(&fi->dio_rwsem[READ]))
850 					continue;
851 				if (!down_write_trylock(
852 						&fi->dio_rwsem[WRITE])) {
853 					up_write(&fi->dio_rwsem[READ]);
854 					continue;
855 				}
856 				locked = true;
857 			}
858 
859 			start_bidx = start_bidx_of_node(nofs, inode)
860 								+ ofs_in_node;
861 			if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
862 				move_encrypted_block(inode, start_bidx, segno, off);
863 			else
864 				move_data_page(inode, start_bidx, gc_type, segno, off);
865 
866 			if (locked) {
867 				up_write(&fi->dio_rwsem[WRITE]);
868 				up_write(&fi->dio_rwsem[READ]);
869 			}
870 
871 			stat_inc_data_blk_count(sbi, 1, gc_type);
872 		}
873 	}
874 
875 	if (++phase < 5)
876 		goto next_step;
877 }
878 
879 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
880 			int gc_type)
881 {
882 	struct sit_info *sit_i = SIT_I(sbi);
883 	int ret;
884 
885 	mutex_lock(&sit_i->sentry_lock);
886 	ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
887 					      NO_CHECK_TYPE, LFS);
888 	mutex_unlock(&sit_i->sentry_lock);
889 	return ret;
890 }
891 
892 static int do_garbage_collect(struct f2fs_sb_info *sbi,
893 				unsigned int start_segno,
894 				struct gc_inode_list *gc_list, int gc_type)
895 {
896 	struct page *sum_page;
897 	struct f2fs_summary_block *sum;
898 	struct blk_plug plug;
899 	unsigned int segno = start_segno;
900 	unsigned int end_segno = start_segno + sbi->segs_per_sec;
901 	int sec_freed = 0;
902 	unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
903 						SUM_TYPE_DATA : SUM_TYPE_NODE;
904 
905 	/* readahead multi ssa blocks those have contiguous address */
906 	if (sbi->segs_per_sec > 1)
907 		ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
908 					sbi->segs_per_sec, META_SSA, true);
909 
910 	/* reference all summary page */
911 	while (segno < end_segno) {
912 		sum_page = get_sum_page(sbi, segno++);
913 		unlock_page(sum_page);
914 	}
915 
916 	blk_start_plug(&plug);
917 
918 	for (segno = start_segno; segno < end_segno; segno++) {
919 
920 		/* find segment summary of victim */
921 		sum_page = find_get_page(META_MAPPING(sbi),
922 					GET_SUM_BLOCK(sbi, segno));
923 		f2fs_put_page(sum_page, 0);
924 
925 		if (get_valid_blocks(sbi, segno, false) == 0 ||
926 				!PageUptodate(sum_page) ||
927 				unlikely(f2fs_cp_error(sbi)))
928 			goto next;
929 
930 		sum = page_address(sum_page);
931 		f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
932 
933 		/*
934 		 * this is to avoid deadlock:
935 		 * - lock_page(sum_page)         - f2fs_replace_block
936 		 *  - check_valid_map()            - mutex_lock(sentry_lock)
937 		 *   - mutex_lock(sentry_lock)     - change_curseg()
938 		 *                                  - lock_page(sum_page)
939 		 */
940 		if (type == SUM_TYPE_NODE)
941 			gc_node_segment(sbi, sum->entries, segno, gc_type);
942 		else
943 			gc_data_segment(sbi, sum->entries, gc_list, segno,
944 								gc_type);
945 
946 		stat_inc_seg_count(sbi, type, gc_type);
947 next:
948 		f2fs_put_page(sum_page, 0);
949 	}
950 
951 	if (gc_type == FG_GC)
952 		f2fs_submit_merged_write(sbi,
953 				(type == SUM_TYPE_NODE) ? NODE : DATA);
954 
955 	blk_finish_plug(&plug);
956 
957 	if (gc_type == FG_GC &&
958 		get_valid_blocks(sbi, start_segno, true) == 0)
959 		sec_freed = 1;
960 
961 	stat_inc_call_count(sbi->stat_info);
962 
963 	return sec_freed;
964 }
965 
966 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
967 			bool background, unsigned int segno)
968 {
969 	int gc_type = sync ? FG_GC : BG_GC;
970 	int sec_freed = 0;
971 	int ret;
972 	struct cp_control cpc;
973 	unsigned int init_segno = segno;
974 	struct gc_inode_list gc_list = {
975 		.ilist = LIST_HEAD_INIT(gc_list.ilist),
976 		.iroot = RADIX_TREE_INIT(GFP_NOFS),
977 	};
978 
979 	cpc.reason = __get_cp_reason(sbi);
980 gc_more:
981 	if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE))) {
982 		ret = -EINVAL;
983 		goto stop;
984 	}
985 	if (unlikely(f2fs_cp_error(sbi))) {
986 		ret = -EIO;
987 		goto stop;
988 	}
989 
990 	if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
991 		/*
992 		 * For example, if there are many prefree_segments below given
993 		 * threshold, we can make them free by checkpoint. Then, we
994 		 * secure free segments which doesn't need fggc any more.
995 		 */
996 		if (prefree_segments(sbi)) {
997 			ret = write_checkpoint(sbi, &cpc);
998 			if (ret)
999 				goto stop;
1000 		}
1001 		if (has_not_enough_free_secs(sbi, 0, 0))
1002 			gc_type = FG_GC;
1003 	}
1004 
1005 	ret = -EINVAL;
1006 	/* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
1007 	if (gc_type == BG_GC && !background)
1008 		goto stop;
1009 	if (!__get_victim(sbi, &segno, gc_type))
1010 		goto stop;
1011 	ret = 0;
1012 
1013 	if (do_garbage_collect(sbi, segno, &gc_list, gc_type) &&
1014 			gc_type == FG_GC)
1015 		sec_freed++;
1016 
1017 	if (gc_type == FG_GC)
1018 		sbi->cur_victim_sec = NULL_SEGNO;
1019 
1020 	if (!sync) {
1021 		if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1022 			segno = NULL_SEGNO;
1023 			goto gc_more;
1024 		}
1025 
1026 		if (gc_type == FG_GC)
1027 			ret = write_checkpoint(sbi, &cpc);
1028 	}
1029 stop:
1030 	SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1031 	SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1032 	mutex_unlock(&sbi->gc_mutex);
1033 
1034 	put_gc_inode(&gc_list);
1035 
1036 	if (sync)
1037 		ret = sec_freed ? 0 : -EAGAIN;
1038 	return ret;
1039 }
1040 
1041 void build_gc_manager(struct f2fs_sb_info *sbi)
1042 {
1043 	u64 main_count, resv_count, ovp_count;
1044 
1045 	DIRTY_I(sbi)->v_ops = &default_v_ops;
1046 
1047 	/* threshold of # of valid blocks in a section for victims of FG_GC */
1048 	main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1049 	resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1050 	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1051 
1052 	sbi->fggc_threshold = div64_u64((main_count - ovp_count) *
1053 				BLKS_PER_SEC(sbi), (main_count - resv_count));
1054 
1055 	/* give warm/cold data area from slower device */
1056 	if (sbi->s_ndevs && sbi->segs_per_sec == 1)
1057 		SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1058 				GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
1059 }
1060