xref: /linux/fs/f2fs/gc.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 #include <linux/blkdev.h>
20 
21 #include "f2fs.h"
22 #include "node.h"
23 #include "segment.h"
24 #include "gc.h"
25 #include <trace/events/f2fs.h>
26 
27 static int gc_thread_func(void *data)
28 {
29 	struct f2fs_sb_info *sbi = data;
30 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
31 	wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
32 	long wait_ms;
33 
34 	wait_ms = gc_th->min_sleep_time;
35 
36 	do {
37 		if (try_to_freeze())
38 			continue;
39 		else
40 			wait_event_interruptible_timeout(*wq,
41 						kthread_should_stop(),
42 						msecs_to_jiffies(wait_ms));
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 		/*
52 		 * [GC triggering condition]
53 		 * 0. GC is not conducted currently.
54 		 * 1. There are enough dirty segments.
55 		 * 2. IO subsystem is idle by checking the # of writeback pages.
56 		 * 3. IO subsystem is idle by checking the # of requests in
57 		 *    bdev's request list.
58 		 *
59 		 * Note) We have to avoid triggering GCs frequently.
60 		 * Because it is possible that some segments can be
61 		 * invalidated soon after by user update or deletion.
62 		 * So, I'd like to wait some time to collect dirty segments.
63 		 */
64 		if (!mutex_trylock(&sbi->gc_mutex))
65 			continue;
66 
67 		if (!is_idle(sbi)) {
68 			increase_sleep_time(gc_th, &wait_ms);
69 			mutex_unlock(&sbi->gc_mutex);
70 			continue;
71 		}
72 
73 		if (has_enough_invalid_blocks(sbi))
74 			decrease_sleep_time(gc_th, &wait_ms);
75 		else
76 			increase_sleep_time(gc_th, &wait_ms);
77 
78 		stat_inc_bggc_count(sbi);
79 
80 		/* if return value is not zero, no victim was selected */
81 		if (f2fs_gc(sbi))
82 			wait_ms = gc_th->no_gc_sleep_time;
83 
84 		/* balancing f2fs's metadata periodically */
85 		f2fs_balance_fs_bg(sbi);
86 
87 	} while (!kthread_should_stop());
88 	return 0;
89 }
90 
91 int start_gc_thread(struct f2fs_sb_info *sbi)
92 {
93 	struct f2fs_gc_kthread *gc_th;
94 	dev_t dev = sbi->sb->s_bdev->bd_dev;
95 	int err = 0;
96 
97 	gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
98 	if (!gc_th) {
99 		err = -ENOMEM;
100 		goto out;
101 	}
102 
103 	gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
104 	gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
105 	gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
106 
107 	gc_th->gc_idle = 0;
108 
109 	sbi->gc_thread = gc_th;
110 	init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
111 	sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
112 			"f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
113 	if (IS_ERR(gc_th->f2fs_gc_task)) {
114 		err = PTR_ERR(gc_th->f2fs_gc_task);
115 		kfree(gc_th);
116 		sbi->gc_thread = NULL;
117 	}
118 out:
119 	return err;
120 }
121 
122 void stop_gc_thread(struct f2fs_sb_info *sbi)
123 {
124 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
125 	if (!gc_th)
126 		return;
127 	kthread_stop(gc_th->f2fs_gc_task);
128 	kfree(gc_th);
129 	sbi->gc_thread = NULL;
130 }
131 
132 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
133 {
134 	int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
135 
136 	if (gc_th && gc_th->gc_idle) {
137 		if (gc_th->gc_idle == 1)
138 			gc_mode = GC_CB;
139 		else if (gc_th->gc_idle == 2)
140 			gc_mode = GC_GREEDY;
141 	}
142 	return gc_mode;
143 }
144 
145 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
146 			int type, struct victim_sel_policy *p)
147 {
148 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
149 
150 	if (p->alloc_mode == SSR) {
151 		p->gc_mode = GC_GREEDY;
152 		p->dirty_segmap = dirty_i->dirty_segmap[type];
153 		p->max_search = dirty_i->nr_dirty[type];
154 		p->ofs_unit = 1;
155 	} else {
156 		p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
157 		p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
158 		p->max_search = dirty_i->nr_dirty[DIRTY];
159 		p->ofs_unit = sbi->segs_per_sec;
160 	}
161 
162 	if (p->max_search > sbi->max_victim_search)
163 		p->max_search = sbi->max_victim_search;
164 
165 	p->offset = sbi->last_victim[p->gc_mode];
166 }
167 
168 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
169 				struct victim_sel_policy *p)
170 {
171 	/* SSR allocates in a segment unit */
172 	if (p->alloc_mode == SSR)
173 		return 1 << sbi->log_blocks_per_seg;
174 	if (p->gc_mode == GC_GREEDY)
175 		return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
176 	else if (p->gc_mode == GC_CB)
177 		return UINT_MAX;
178 	else /* No other gc_mode */
179 		return 0;
180 }
181 
182 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
183 {
184 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
185 	unsigned int secno;
186 
187 	/*
188 	 * If the gc_type is FG_GC, we can select victim segments
189 	 * selected by background GC before.
190 	 * Those segments guarantee they have small valid blocks.
191 	 */
192 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
193 		if (sec_usage_check(sbi, secno))
194 			continue;
195 		clear_bit(secno, dirty_i->victim_secmap);
196 		return secno * sbi->segs_per_sec;
197 	}
198 	return NULL_SEGNO;
199 }
200 
201 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
202 {
203 	struct sit_info *sit_i = SIT_I(sbi);
204 	unsigned int secno = GET_SECNO(sbi, segno);
205 	unsigned int start = secno * sbi->segs_per_sec;
206 	unsigned long long mtime = 0;
207 	unsigned int vblocks;
208 	unsigned char age = 0;
209 	unsigned char u;
210 	unsigned int i;
211 
212 	for (i = 0; i < sbi->segs_per_sec; i++)
213 		mtime += get_seg_entry(sbi, start + i)->mtime;
214 	vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
215 
216 	mtime = div_u64(mtime, sbi->segs_per_sec);
217 	vblocks = div_u64(vblocks, sbi->segs_per_sec);
218 
219 	u = (vblocks * 100) >> sbi->log_blocks_per_seg;
220 
221 	/* Handle if the system time has changed by the user */
222 	if (mtime < sit_i->min_mtime)
223 		sit_i->min_mtime = mtime;
224 	if (mtime > sit_i->max_mtime)
225 		sit_i->max_mtime = mtime;
226 	if (sit_i->max_mtime != sit_i->min_mtime)
227 		age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
228 				sit_i->max_mtime - sit_i->min_mtime);
229 
230 	return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
231 }
232 
233 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
234 			unsigned int segno, struct victim_sel_policy *p)
235 {
236 	if (p->alloc_mode == SSR)
237 		return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
238 
239 	/* alloc_mode == LFS */
240 	if (p->gc_mode == GC_GREEDY)
241 		return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
242 	else
243 		return get_cb_cost(sbi, segno);
244 }
245 
246 /*
247  * This function is called from two paths.
248  * One is garbage collection and the other is SSR segment selection.
249  * When it is called during GC, it just gets a victim segment
250  * and it does not remove it from dirty seglist.
251  * When it is called from SSR segment selection, it finds a segment
252  * which has minimum valid blocks and removes it from dirty seglist.
253  */
254 static int get_victim_by_default(struct f2fs_sb_info *sbi,
255 		unsigned int *result, int gc_type, int type, char alloc_mode)
256 {
257 	struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
258 	struct victim_sel_policy p;
259 	unsigned int secno, max_cost;
260 	int nsearched = 0;
261 
262 	mutex_lock(&dirty_i->seglist_lock);
263 
264 	p.alloc_mode = alloc_mode;
265 	select_policy(sbi, gc_type, type, &p);
266 
267 	p.min_segno = NULL_SEGNO;
268 	p.min_cost = max_cost = get_max_cost(sbi, &p);
269 
270 	if (p.alloc_mode == LFS && gc_type == FG_GC) {
271 		p.min_segno = check_bg_victims(sbi);
272 		if (p.min_segno != NULL_SEGNO)
273 			goto got_it;
274 	}
275 
276 	while (1) {
277 		unsigned long cost;
278 		unsigned int segno;
279 
280 		segno = find_next_bit(p.dirty_segmap, MAIN_SEGS(sbi), p.offset);
281 		if (segno >= MAIN_SEGS(sbi)) {
282 			if (sbi->last_victim[p.gc_mode]) {
283 				sbi->last_victim[p.gc_mode] = 0;
284 				p.offset = 0;
285 				continue;
286 			}
287 			break;
288 		}
289 
290 		p.offset = segno + p.ofs_unit;
291 		if (p.ofs_unit > 1)
292 			p.offset -= segno % p.ofs_unit;
293 
294 		secno = GET_SECNO(sbi, segno);
295 
296 		if (sec_usage_check(sbi, secno))
297 			continue;
298 		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
299 			continue;
300 
301 		cost = get_gc_cost(sbi, segno, &p);
302 
303 		if (p.min_cost > cost) {
304 			p.min_segno = segno;
305 			p.min_cost = cost;
306 		} else if (unlikely(cost == max_cost)) {
307 			continue;
308 		}
309 
310 		if (nsearched++ >= p.max_search) {
311 			sbi->last_victim[p.gc_mode] = segno;
312 			break;
313 		}
314 	}
315 	if (p.min_segno != NULL_SEGNO) {
316 got_it:
317 		if (p.alloc_mode == LFS) {
318 			secno = GET_SECNO(sbi, p.min_segno);
319 			if (gc_type == FG_GC)
320 				sbi->cur_victim_sec = secno;
321 			else
322 				set_bit(secno, dirty_i->victim_secmap);
323 		}
324 		*result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
325 
326 		trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
327 				sbi->cur_victim_sec,
328 				prefree_segments(sbi), free_segments(sbi));
329 	}
330 	mutex_unlock(&dirty_i->seglist_lock);
331 
332 	return (p.min_segno == NULL_SEGNO) ? 0 : 1;
333 }
334 
335 static const struct victim_selection default_v_ops = {
336 	.get_victim = get_victim_by_default,
337 };
338 
339 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
340 {
341 	struct inode_entry *ie;
342 
343 	ie = radix_tree_lookup(&gc_list->iroot, ino);
344 	if (ie)
345 		return ie->inode;
346 	return NULL;
347 }
348 
349 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
350 {
351 	struct inode_entry *new_ie;
352 
353 	if (inode == find_gc_inode(gc_list, inode->i_ino)) {
354 		iput(inode);
355 		return;
356 	}
357 	new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
358 	new_ie->inode = inode;
359 
360 	f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
361 	list_add_tail(&new_ie->list, &gc_list->ilist);
362 }
363 
364 static void put_gc_inode(struct gc_inode_list *gc_list)
365 {
366 	struct inode_entry *ie, *next_ie;
367 	list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
368 		radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
369 		iput(ie->inode);
370 		list_del(&ie->list);
371 		kmem_cache_free(inode_entry_slab, ie);
372 	}
373 }
374 
375 static int check_valid_map(struct f2fs_sb_info *sbi,
376 				unsigned int segno, int offset)
377 {
378 	struct sit_info *sit_i = SIT_I(sbi);
379 	struct seg_entry *sentry;
380 	int ret;
381 
382 	mutex_lock(&sit_i->sentry_lock);
383 	sentry = get_seg_entry(sbi, segno);
384 	ret = f2fs_test_bit(offset, sentry->cur_valid_map);
385 	mutex_unlock(&sit_i->sentry_lock);
386 	return ret;
387 }
388 
389 /*
390  * This function compares node address got in summary with that in NAT.
391  * On validity, copy that node with cold status, otherwise (invalid node)
392  * ignore that.
393  */
394 static int gc_node_segment(struct f2fs_sb_info *sbi,
395 		struct f2fs_summary *sum, unsigned int segno, int gc_type)
396 {
397 	bool initial = true;
398 	struct f2fs_summary *entry;
399 	block_t start_addr;
400 	int off;
401 
402 	start_addr = START_BLOCK(sbi, segno);
403 
404 next_step:
405 	entry = sum;
406 
407 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
408 		nid_t nid = le32_to_cpu(entry->nid);
409 		struct page *node_page;
410 		struct node_info ni;
411 
412 		/* stop BG_GC if there is not enough free sections. */
413 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
414 			return 0;
415 
416 		if (check_valid_map(sbi, segno, off) == 0)
417 			continue;
418 
419 		if (initial) {
420 			ra_node_page(sbi, nid);
421 			continue;
422 		}
423 		node_page = get_node_page(sbi, nid);
424 		if (IS_ERR(node_page))
425 			continue;
426 
427 		/* block may become invalid during get_node_page */
428 		if (check_valid_map(sbi, segno, off) == 0) {
429 			f2fs_put_page(node_page, 1);
430 			continue;
431 		}
432 
433 		get_node_info(sbi, nid, &ni);
434 		if (ni.blk_addr != start_addr + off) {
435 			f2fs_put_page(node_page, 1);
436 			continue;
437 		}
438 
439 		/* set page dirty and write it */
440 		if (gc_type == FG_GC) {
441 			f2fs_wait_on_page_writeback(node_page, NODE);
442 			set_page_dirty(node_page);
443 		} else {
444 			if (!PageWriteback(node_page))
445 				set_page_dirty(node_page);
446 		}
447 		f2fs_put_page(node_page, 1);
448 		stat_inc_node_blk_count(sbi, 1, gc_type);
449 	}
450 
451 	if (initial) {
452 		initial = false;
453 		goto next_step;
454 	}
455 
456 	if (gc_type == FG_GC) {
457 		struct writeback_control wbc = {
458 			.sync_mode = WB_SYNC_ALL,
459 			.nr_to_write = LONG_MAX,
460 			.for_reclaim = 0,
461 		};
462 		sync_node_pages(sbi, 0, &wbc);
463 
464 		/* return 1 only if FG_GC succefully reclaimed one */
465 		if (get_valid_blocks(sbi, segno, 1) == 0)
466 			return 1;
467 	}
468 	return 0;
469 }
470 
471 /*
472  * Calculate start block index indicating the given node offset.
473  * Be careful, caller should give this node offset only indicating direct node
474  * blocks. If any node offsets, which point the other types of node blocks such
475  * as indirect or double indirect node blocks, are given, it must be a caller's
476  * bug.
477  */
478 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
479 {
480 	unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
481 	unsigned int bidx;
482 
483 	if (node_ofs == 0)
484 		return 0;
485 
486 	if (node_ofs <= 2) {
487 		bidx = node_ofs - 1;
488 	} else if (node_ofs <= indirect_blks) {
489 		int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
490 		bidx = node_ofs - 2 - dec;
491 	} else {
492 		int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
493 		bidx = node_ofs - 5 - dec;
494 	}
495 	return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
496 }
497 
498 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
499 		struct node_info *dni, block_t blkaddr, unsigned int *nofs)
500 {
501 	struct page *node_page;
502 	nid_t nid;
503 	unsigned int ofs_in_node;
504 	block_t source_blkaddr;
505 
506 	nid = le32_to_cpu(sum->nid);
507 	ofs_in_node = le16_to_cpu(sum->ofs_in_node);
508 
509 	node_page = get_node_page(sbi, nid);
510 	if (IS_ERR(node_page))
511 		return false;
512 
513 	get_node_info(sbi, nid, dni);
514 
515 	if (sum->version != dni->version) {
516 		f2fs_put_page(node_page, 1);
517 		return false;
518 	}
519 
520 	*nofs = ofs_of_node(node_page);
521 	source_blkaddr = datablock_addr(node_page, ofs_in_node);
522 	f2fs_put_page(node_page, 1);
523 
524 	if (source_blkaddr != blkaddr)
525 		return false;
526 	return true;
527 }
528 
529 static void move_encrypted_block(struct inode *inode, block_t bidx)
530 {
531 	struct f2fs_io_info fio = {
532 		.sbi = F2FS_I_SB(inode),
533 		.type = DATA,
534 		.rw = READ_SYNC,
535 		.encrypted_page = NULL,
536 	};
537 	struct dnode_of_data dn;
538 	struct f2fs_summary sum;
539 	struct node_info ni;
540 	struct page *page;
541 	int err;
542 
543 	/* do not read out */
544 	page = grab_cache_page(inode->i_mapping, bidx);
545 	if (!page)
546 		return;
547 
548 	set_new_dnode(&dn, inode, NULL, NULL, 0);
549 	err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
550 	if (err)
551 		goto out;
552 
553 	if (unlikely(dn.data_blkaddr == NULL_ADDR))
554 		goto put_out;
555 
556 	get_node_info(fio.sbi, dn.nid, &ni);
557 	set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
558 
559 	/* read page */
560 	fio.page = page;
561 	fio.blk_addr = dn.data_blkaddr;
562 
563 	fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
564 					fio.blk_addr,
565 					FGP_LOCK|FGP_CREAT,
566 					GFP_NOFS);
567 	if (!fio.encrypted_page)
568 		goto put_out;
569 
570 	err = f2fs_submit_page_bio(&fio);
571 	if (err)
572 		goto put_page_out;
573 
574 	/* write page */
575 	lock_page(fio.encrypted_page);
576 
577 	if (unlikely(!PageUptodate(fio.encrypted_page)))
578 		goto put_page_out;
579 	if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
580 		goto put_page_out;
581 
582 	set_page_dirty(fio.encrypted_page);
583 	f2fs_wait_on_page_writeback(fio.encrypted_page, META);
584 	if (clear_page_dirty_for_io(fio.encrypted_page))
585 		dec_page_count(fio.sbi, F2FS_DIRTY_META);
586 
587 	set_page_writeback(fio.encrypted_page);
588 
589 	/* allocate block address */
590 	f2fs_wait_on_page_writeback(dn.node_page, NODE);
591 	allocate_data_block(fio.sbi, NULL, fio.blk_addr,
592 					&fio.blk_addr, &sum, CURSEG_COLD_DATA);
593 	fio.rw = WRITE_SYNC;
594 	f2fs_submit_page_mbio(&fio);
595 
596 	dn.data_blkaddr = fio.blk_addr;
597 	set_data_blkaddr(&dn);
598 	f2fs_update_extent_cache(&dn);
599 	set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
600 	if (page->index == 0)
601 		set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
602 put_page_out:
603 	f2fs_put_page(fio.encrypted_page, 1);
604 put_out:
605 	f2fs_put_dnode(&dn);
606 out:
607 	f2fs_put_page(page, 1);
608 }
609 
610 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
611 {
612 	struct page *page;
613 
614 	page = get_lock_data_page(inode, bidx);
615 	if (IS_ERR(page))
616 		return;
617 
618 	if (gc_type == BG_GC) {
619 		if (PageWriteback(page))
620 			goto out;
621 		set_page_dirty(page);
622 		set_cold_data(page);
623 	} else {
624 		struct f2fs_io_info fio = {
625 			.sbi = F2FS_I_SB(inode),
626 			.type = DATA,
627 			.rw = WRITE_SYNC,
628 			.page = page,
629 			.encrypted_page = NULL,
630 		};
631 		set_page_dirty(page);
632 		f2fs_wait_on_page_writeback(page, DATA);
633 		if (clear_page_dirty_for_io(page))
634 			inode_dec_dirty_pages(inode);
635 		set_cold_data(page);
636 		do_write_data_page(&fio);
637 		clear_cold_data(page);
638 	}
639 out:
640 	f2fs_put_page(page, 1);
641 }
642 
643 /*
644  * This function tries to get parent node of victim data block, and identifies
645  * data block validity. If the block is valid, copy that with cold status and
646  * modify parent node.
647  * If the parent node is not valid or the data block address is different,
648  * the victim data block is ignored.
649  */
650 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
651 		struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
652 {
653 	struct super_block *sb = sbi->sb;
654 	struct f2fs_summary *entry;
655 	block_t start_addr;
656 	int off;
657 	int phase = 0;
658 
659 	start_addr = START_BLOCK(sbi, segno);
660 
661 next_step:
662 	entry = sum;
663 
664 	for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
665 		struct page *data_page;
666 		struct inode *inode;
667 		struct node_info dni; /* dnode info for the data */
668 		unsigned int ofs_in_node, nofs;
669 		block_t start_bidx;
670 
671 		/* stop BG_GC if there is not enough free sections. */
672 		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
673 			return 0;
674 
675 		if (check_valid_map(sbi, segno, off) == 0)
676 			continue;
677 
678 		if (phase == 0) {
679 			ra_node_page(sbi, le32_to_cpu(entry->nid));
680 			continue;
681 		}
682 
683 		/* Get an inode by ino with checking validity */
684 		if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
685 			continue;
686 
687 		if (phase == 1) {
688 			ra_node_page(sbi, dni.ino);
689 			continue;
690 		}
691 
692 		ofs_in_node = le16_to_cpu(entry->ofs_in_node);
693 
694 		if (phase == 2) {
695 			inode = f2fs_iget(sb, dni.ino);
696 			if (IS_ERR(inode) || is_bad_inode(inode))
697 				continue;
698 
699 			/* if encrypted inode, let's go phase 3 */
700 			if (f2fs_encrypted_inode(inode) &&
701 						S_ISREG(inode->i_mode)) {
702 				add_gc_inode(gc_list, inode);
703 				continue;
704 			}
705 
706 			start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
707 			data_page = get_read_data_page(inode,
708 					start_bidx + ofs_in_node, READA);
709 			if (IS_ERR(data_page)) {
710 				iput(inode);
711 				continue;
712 			}
713 
714 			f2fs_put_page(data_page, 0);
715 			add_gc_inode(gc_list, inode);
716 			continue;
717 		}
718 
719 		/* phase 3 */
720 		inode = find_gc_inode(gc_list, dni.ino);
721 		if (inode) {
722 			start_bidx = start_bidx_of_node(nofs, F2FS_I(inode))
723 								+ ofs_in_node;
724 			if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
725 				move_encrypted_block(inode, start_bidx);
726 			else
727 				move_data_page(inode, start_bidx, gc_type);
728 			stat_inc_data_blk_count(sbi, 1, gc_type);
729 		}
730 	}
731 
732 	if (++phase < 4)
733 		goto next_step;
734 
735 	if (gc_type == FG_GC) {
736 		f2fs_submit_merged_bio(sbi, DATA, WRITE);
737 
738 		/* return 1 only if FG_GC succefully reclaimed one */
739 		if (get_valid_blocks(sbi, segno, 1) == 0)
740 			return 1;
741 	}
742 	return 0;
743 }
744 
745 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
746 			int gc_type)
747 {
748 	struct sit_info *sit_i = SIT_I(sbi);
749 	int ret;
750 
751 	mutex_lock(&sit_i->sentry_lock);
752 	ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
753 					      NO_CHECK_TYPE, LFS);
754 	mutex_unlock(&sit_i->sentry_lock);
755 	return ret;
756 }
757 
758 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
759 				struct gc_inode_list *gc_list, int gc_type)
760 {
761 	struct page *sum_page;
762 	struct f2fs_summary_block *sum;
763 	struct blk_plug plug;
764 	int nfree = 0;
765 
766 	/* read segment summary of victim */
767 	sum_page = get_sum_page(sbi, segno);
768 
769 	blk_start_plug(&plug);
770 
771 	sum = page_address(sum_page);
772 
773 	/*
774 	 * this is to avoid deadlock:
775 	 * - lock_page(sum_page)         - f2fs_replace_block
776 	 *  - check_valid_map()            - mutex_lock(sentry_lock)
777 	 *   - mutex_lock(sentry_lock)     - change_curseg()
778 	 *                                  - lock_page(sum_page)
779 	 */
780 	unlock_page(sum_page);
781 
782 	switch (GET_SUM_TYPE((&sum->footer))) {
783 	case SUM_TYPE_NODE:
784 		nfree = gc_node_segment(sbi, sum->entries, segno, gc_type);
785 		break;
786 	case SUM_TYPE_DATA:
787 		nfree = gc_data_segment(sbi, sum->entries, gc_list,
788 							segno, gc_type);
789 		break;
790 	}
791 	blk_finish_plug(&plug);
792 
793 	stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)), gc_type);
794 	stat_inc_call_count(sbi->stat_info);
795 
796 	f2fs_put_page(sum_page, 0);
797 	return nfree;
798 }
799 
800 int f2fs_gc(struct f2fs_sb_info *sbi)
801 {
802 	unsigned int segno = NULL_SEGNO;
803 	unsigned int i;
804 	int gc_type = BG_GC;
805 	int nfree = 0;
806 	int ret = -1;
807 	struct cp_control cpc;
808 	struct gc_inode_list gc_list = {
809 		.ilist = LIST_HEAD_INIT(gc_list.ilist),
810 		.iroot = RADIX_TREE_INIT(GFP_NOFS),
811 	};
812 
813 	cpc.reason = __get_cp_reason(sbi);
814 gc_more:
815 	if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
816 		goto stop;
817 	if (unlikely(f2fs_cp_error(sbi)))
818 		goto stop;
819 
820 	if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
821 		gc_type = FG_GC;
822 		if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
823 			write_checkpoint(sbi, &cpc);
824 	}
825 
826 	if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
827 		goto stop;
828 	ret = 0;
829 
830 	/* readahead multi ssa blocks those have contiguous address */
831 	if (sbi->segs_per_sec > 1)
832 		ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
833 								META_SSA);
834 
835 	for (i = 0; i < sbi->segs_per_sec; i++)
836 		nfree += do_garbage_collect(sbi, segno + i, &gc_list, gc_type);
837 
838 	if (gc_type == FG_GC)
839 		sbi->cur_victim_sec = NULL_SEGNO;
840 
841 	if (has_not_enough_free_secs(sbi, nfree))
842 		goto gc_more;
843 
844 	if (gc_type == FG_GC)
845 		write_checkpoint(sbi, &cpc);
846 stop:
847 	mutex_unlock(&sbi->gc_mutex);
848 
849 	put_gc_inode(&gc_list);
850 	return ret;
851 }
852 
853 void build_gc_manager(struct f2fs_sb_info *sbi)
854 {
855 	DIRTY_I(sbi)->v_ops = &default_v_ops;
856 }
857