fs/f2fs/gc.c

   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, test_opt(sbi, FORCE_FG_GC)))
  82                         wait_ms = gc_th->no_gc_sleep_time;
  83
  84                 trace_f2fs_background_gc(sbi->sb, wait_ms,
  85                                 prefree_segments(sbi), free_segments(sbi));
  86
  87                 /* balancing f2fs's metadata periodically */
  88                 f2fs_balance_fs_bg(sbi);
  89
  90         } while (!kthread_should_stop());
  91         return 0;
  92 }
  93
  94 int start_gc_thread(struct f2fs_sb_info *sbi)
  95 {
  96         struct f2fs_gc_kthread *gc_th;
  97         dev_t dev = sbi->sb->s_bdev->bd_dev;
  98         int err = 0;
  99
 100         gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 101         if (!gc_th) {
 102                 err = -ENOMEM;
 103                 goto out;
 104         }
 105
 106         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 107         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 108         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 109
 110         gc_th->gc_idle = 0;
 111
 112         sbi->gc_thread = gc_th;
 113         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 114         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 115                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 116         if (IS_ERR(gc_th->f2fs_gc_task)) {
 117                 err = PTR_ERR(gc_th->f2fs_gc_task);
 118                 kfree(gc_th);
 119                 sbi->gc_thread = NULL;
 120         }
 121 out:
 122         return err;
 123 }
 124
 125 void stop_gc_thread(struct f2fs_sb_info *sbi)
 126 {
 127         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 128         if (!gc_th)
 129                 return;
 130         kthread_stop(gc_th->f2fs_gc_task);
 131         kfree(gc_th);
 132         sbi->gc_thread = NULL;
 133 }
 134
 135 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 136 {
 137         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 138
 139         if (gc_th && gc_th->gc_idle) {
 140                 if (gc_th->gc_idle == 1)
 141                         gc_mode = GC_CB;
 142                 else if (gc_th->gc_idle == 2)
 143                         gc_mode = GC_GREEDY;
 144         }
 145         return gc_mode;
 146 }
 147
 148 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 149                         int type, struct victim_sel_policy *p)
 150 {
 151         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 152
 153         if (p->alloc_mode == SSR) {
 154                 p->gc_mode = GC_GREEDY;
 155                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 156                 p->max_search = dirty_i->nr_dirty[type];
 157                 p->ofs_unit = 1;
 158         } else {
 159                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 160                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 161                 p->max_search = dirty_i->nr_dirty[DIRTY];
 162                 p->ofs_unit = sbi->segs_per_sec;
 163         }
 164
 165         if (p->max_search > sbi->max_victim_search)
 166                 p->max_search = sbi->max_victim_search;
 167
 168         p->offset = sbi->last_victim[p->gc_mode];
 169 }
 170
 171 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 172                                 struct victim_sel_policy *p)
 173 {
 174         /* SSR allocates in a segment unit */
 175         if (p->alloc_mode == SSR)
 176                 return 1 << sbi->log_blocks_per_seg;
 177         if (p->gc_mode == GC_GREEDY)
 178                 return (1 << sbi->log_blocks_per_seg) * p->ofs_unit;
 179         else if (p->gc_mode == GC_CB)
 180                 return UINT_MAX;
 181         else /* No other gc_mode */
 182                 return 0;
 183 }
 184
 185 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 186 {
 187         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 188         unsigned int secno;
 189
 190         /*
 191          * If the gc_type is FG_GC, we can select victim segments
 192          * selected by background GC before.
 193          * Those segments guarantee they have small valid blocks.
 194          */
 195         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 196                 if (sec_usage_check(sbi, secno))
 197                         continue;
 198                 clear_bit(secno, dirty_i->victim_secmap);
 199                 return secno * sbi->segs_per_sec;
 200         }
 201         return NULL_SEGNO;
 202 }
 203
 204 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 205 {
 206         struct sit_info *sit_i = SIT_I(sbi);
 207         unsigned int secno = GET_SECNO(sbi, segno);
 208         unsigned int start = secno * sbi->segs_per_sec;
 209         unsigned long long mtime = 0;
 210         unsigned int vblocks;
 211         unsigned char age = 0;
 212         unsigned char u;
 213         unsigned int i;
 214
 215         for (i = 0; i < sbi->segs_per_sec; i++)
 216                 mtime += get_seg_entry(sbi, start + i)->mtime;
 217         vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 218
 219         mtime = div_u64(mtime, sbi->segs_per_sec);
 220         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 221
 222         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 223
 224         /* Handle if the system time has changed by the user */
 225         if (mtime < sit_i->min_mtime)
 226                 sit_i->min_mtime = mtime;
 227         if (mtime > sit_i->max_mtime)
 228                 sit_i->max_mtime = mtime;
 229         if (sit_i->max_mtime != sit_i->min_mtime)
 230                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 231                                 sit_i->max_mtime - sit_i->min_mtime);
 232
 233         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 234 }
 235
 236 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 237                         unsigned int segno, struct victim_sel_policy *p)
 238 {
 239         if (p->alloc_mode == SSR)
 240                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 241
 242         /* alloc_mode == LFS */
 243         if (p->gc_mode == GC_GREEDY)
 244                 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 245         else
 246                 return get_cb_cost(sbi, segno);
 247 }
 248
 249 /*
 250  * This function is called from two paths.
 251  * One is garbage collection and the other is SSR segment selection.
 252  * When it is called during GC, it just gets a victim segment
 253  * and it does not remove it from dirty seglist.
 254  * When it is called from SSR segment selection, it finds a segment
 255  * which has minimum valid blocks and removes it from dirty seglist.
 256  */
 257 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 258                 unsigned int *result, int gc_type, int type, char alloc_mode)
 259 {
 260         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 261         struct victim_sel_policy p;
 262         unsigned int secno, max_cost;
 263         unsigned int last_segment = MAIN_SEGS(sbi);
 264         int nsearched = 0;
 265
 266         mutex_lock(&dirty_i->seglist_lock);
 267
 268         p.alloc_mode = alloc_mode;
 269         select_policy(sbi, gc_type, type, &p);
 270
 271         p.min_segno = NULL_SEGNO;
 272         p.min_cost = max_cost = get_max_cost(sbi, &p);
 273
 274         if (p.max_search == 0)
 275                 goto out;
 276
 277         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 278                 p.min_segno = check_bg_victims(sbi);
 279                 if (p.min_segno != NULL_SEGNO)
 280                         goto got_it;
 281         }
 282
 283         while (1) {
 284                 unsigned long cost;
 285                 unsigned int segno;
 286
 287                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 288                 if (segno >= last_segment) {
 289                         if (sbi->last_victim[p.gc_mode]) {
 290                                 last_segment = sbi->last_victim[p.gc_mode];
 291                                 sbi->last_victim[p.gc_mode] = 0;
 292                                 p.offset = 0;
 293                                 continue;
 294                         }
 295                         break;
 296                 }
 297
 298                 p.offset = segno + p.ofs_unit;
 299                 if (p.ofs_unit > 1)
 300                         p.offset -= segno % p.ofs_unit;
 301
 302                 secno = GET_SECNO(sbi, segno);
 303
 304                 if (sec_usage_check(sbi, secno))
 305                         continue;
 306                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 307                         continue;
 308
 309                 cost = get_gc_cost(sbi, segno, &p);
 310
 311                 if (p.min_cost > cost) {
 312                         p.min_segno = segno;
 313                         p.min_cost = cost;
 314                 } else if (unlikely(cost == max_cost)) {
 315                         continue;
 316                 }
 317
 318                 if (nsearched++ >= p.max_search) {
 319                         sbi->last_victim[p.gc_mode] = segno;
 320                         break;
 321                 }
 322         }
 323         if (p.min_segno != NULL_SEGNO) {
 324 got_it:
 325                 if (p.alloc_mode == LFS) {
 326                         secno = GET_SECNO(sbi, p.min_segno);
 327                         if (gc_type == FG_GC)
 328                                 sbi->cur_victim_sec = secno;
 329                         else
 330                                 set_bit(secno, dirty_i->victim_secmap);
 331                 }
 332                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 333
 334                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 335                                 sbi->cur_victim_sec,
 336                                 prefree_segments(sbi), free_segments(sbi));
 337         }
 338 out:
 339         mutex_unlock(&dirty_i->seglist_lock);
 340
 341         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 342 }
 343
 344 static const struct victim_selection default_v_ops = {
 345         .get_victim = get_victim_by_default,
 346 };
 347
 348 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 349 {
 350         struct inode_entry *ie;
 351
 352         ie = radix_tree_lookup(&gc_list->iroot, ino);
 353         if (ie)
 354                 return ie->inode;
 355         return NULL;
 356 }
 357
 358 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 359 {
 360         struct inode_entry *new_ie;
 361
 362         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 363                 iput(inode);
 364                 return;
 365         }
 366         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 367         new_ie->inode = inode;
 368
 369         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 370         list_add_tail(&new_ie->list, &gc_list->ilist);
 371 }
 372
 373 static void put_gc_inode(struct gc_inode_list *gc_list)
 374 {
 375         struct inode_entry *ie, *next_ie;
 376         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 377                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 378                 iput(ie->inode);
 379                 list_del(&ie->list);
 380                 kmem_cache_free(inode_entry_slab, ie);
 381         }
 382 }
 383
 384 static int check_valid_map(struct f2fs_sb_info *sbi,
 385                                 unsigned int segno, int offset)
 386 {
 387         struct sit_info *sit_i = SIT_I(sbi);
 388         struct seg_entry *sentry;
 389         int ret;
 390
 391         mutex_lock(&sit_i->sentry_lock);
 392         sentry = get_seg_entry(sbi, segno);
 393         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 394         mutex_unlock(&sit_i->sentry_lock);
 395         return ret;
 396 }
 397
 398 /*
 399  * This function compares node address got in summary with that in NAT.
 400  * On validity, copy that node with cold status, otherwise (invalid node)
 401  * ignore that.
 402  */
 403 static int gc_node_segment(struct f2fs_sb_info *sbi,
 404                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 405 {
 406         bool initial = true;
 407         struct f2fs_summary *entry;
 408         block_t start_addr;
 409         int off;
 410
 411         start_addr = START_BLOCK(sbi, segno);
 412
 413 next_step:
 414         entry = sum;
 415
 416         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 417                 nid_t nid = le32_to_cpu(entry->nid);
 418                 struct page *node_page;
 419                 struct node_info ni;
 420
 421                 /* stop BG_GC if there is not enough free sections. */
 422                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 423                         return 0;
 424
 425                 if (check_valid_map(sbi, segno, off) == 0)
 426                         continue;
 427
 428                 if (initial) {
 429                         ra_node_page(sbi, nid);
 430                         continue;
 431                 }
 432                 node_page = get_node_page(sbi, nid);
 433                 if (IS_ERR(node_page))
 434                         continue;
 435
 436                 /* block may become invalid during get_node_page */
 437                 if (check_valid_map(sbi, segno, off) == 0) {
 438                         f2fs_put_page(node_page, 1);
 439                         continue;
 440                 }
 441
 442                 get_node_info(sbi, nid, &ni);
 443                 if (ni.blk_addr != start_addr + off) {
 444                         f2fs_put_page(node_page, 1);
 445                         continue;
 446                 }
 447
 448                 /* set page dirty and write it */
 449                 if (gc_type == FG_GC) {
 450                         f2fs_wait_on_page_writeback(node_page, NODE);
 451                         set_page_dirty(node_page);
 452                 } else {
 453                         if (!PageWriteback(node_page))
 454                                 set_page_dirty(node_page);
 455                 }
 456                 f2fs_put_page(node_page, 1);
 457                 stat_inc_node_blk_count(sbi, 1, gc_type);
 458         }
 459
 460         if (initial) {
 461                 initial = false;
 462                 goto next_step;
 463         }
 464
 465         if (gc_type == FG_GC) {
 466                 struct writeback_control wbc = {
 467                         .sync_mode = WB_SYNC_ALL,
 468                         .nr_to_write = LONG_MAX,
 469                         .for_reclaim = 0,
 470                 };
 471                 sync_node_pages(sbi, 0, &wbc);
 472
 473                 /* return 1 only if FG_GC succefully reclaimed one */
 474                 if (get_valid_blocks(sbi, segno, 1) == 0)
 475                         return 1;
 476         }
 477         return 0;
 478 }
 479
 480 /*
 481  * Calculate start block index indicating the given node offset.
 482  * Be careful, caller should give this node offset only indicating direct node
 483  * blocks. If any node offsets, which point the other types of node blocks such
 484  * as indirect or double indirect node blocks, are given, it must be a caller's
 485  * bug.
 486  */
 487 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
 488 {
 489         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 490         unsigned int bidx;
 491
 492         if (node_ofs == 0)
 493                 return 0;
 494
 495         if (node_ofs <= 2) {
 496                 bidx = node_ofs - 1;
 497         } else if (node_ofs <= indirect_blks) {
 498                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 499                 bidx = node_ofs - 2 - dec;
 500         } else {
 501                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 502                 bidx = node_ofs - 5 - dec;
 503         }
 504         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
 505 }
 506
 507 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 508                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 509 {
 510         struct page *node_page;
 511         nid_t nid;
 512         unsigned int ofs_in_node;
 513         block_t source_blkaddr;
 514
 515         nid = le32_to_cpu(sum->nid);
 516         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 517
 518         node_page = get_node_page(sbi, nid);
 519         if (IS_ERR(node_page))
 520                 return false;
 521
 522         get_node_info(sbi, nid, dni);
 523
 524         if (sum->version != dni->version) {
 525                 f2fs_msg(sbi->sb, KERN_WARNING,
 526                                 "%s: valid data with mismatched node version.",
 527                                 __func__);
 528                 set_sbi_flag(sbi, SBI_NEED_FSCK);
 529         }
 530
 531         *nofs = ofs_of_node(node_page);
 532         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 533         f2fs_put_page(node_page, 1);
 534
 535         if (source_blkaddr != blkaddr)
 536                 return false;
 537         return true;
 538 }
 539
 540 static void move_encrypted_block(struct inode *inode, block_t bidx)
 541 {
 542         struct f2fs_io_info fio = {
 543                 .sbi = F2FS_I_SB(inode),
 544                 .type = DATA,
 545                 .rw = READ_SYNC,
 546                 .encrypted_page = NULL,
 547         };
 548         struct dnode_of_data dn;
 549         struct f2fs_summary sum;
 550         struct node_info ni;
 551         struct page *page;
 552         int err;
 553
 554         /* do not read out */
 555         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 556         if (!page)
 557                 return;
 558
 559         set_new_dnode(&dn, inode, NULL, NULL, 0);
 560         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 561         if (err)
 562                 goto out;
 563
 564         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 565                 ClearPageUptodate(page);
 566                 goto put_out;
 567         }
 568
 569         /*
 570          * don't cache encrypted data into meta inode until previous dirty
 571          * data were writebacked to avoid racing between GC and flush.
 572          */
 573         f2fs_wait_on_page_writeback(page, DATA);
 574
 575         get_node_info(fio.sbi, dn.nid, &ni);
 576         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 577
 578         /* read page */
 579         fio.page = page;
 580         fio.blk_addr = dn.data_blkaddr;
 581
 582         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
 583                                         fio.blk_addr,
 584                                         FGP_LOCK|FGP_CREAT,
 585                                         GFP_NOFS);
 586         if (!fio.encrypted_page)
 587                 goto put_out;
 588
 589         err = f2fs_submit_page_bio(&fio);
 590         if (err)
 591                 goto put_page_out;
 592
 593         /* write page */
 594         lock_page(fio.encrypted_page);
 595
 596         if (unlikely(!PageUptodate(fio.encrypted_page)))
 597                 goto put_page_out;
 598         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
 599                 goto put_page_out;
 600
 601         set_page_dirty(fio.encrypted_page);
 602         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA);
 603         if (clear_page_dirty_for_io(fio.encrypted_page))
 604                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 605
 606         set_page_writeback(fio.encrypted_page);
 607
 608         /* allocate block address */
 609         f2fs_wait_on_page_writeback(dn.node_page, NODE);
 610         allocate_data_block(fio.sbi, NULL, fio.blk_addr,
 611                                         &fio.blk_addr, &sum, CURSEG_COLD_DATA);
 612         fio.rw = WRITE_SYNC;
 613         f2fs_submit_page_mbio(&fio);
 614
 615         dn.data_blkaddr = fio.blk_addr;
 616         set_data_blkaddr(&dn);
 617         f2fs_update_extent_cache(&dn);
 618         set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
 619         if (page->index == 0)
 620                 set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
 621 put_page_out:
 622         f2fs_put_page(fio.encrypted_page, 1);
 623 put_out:
 624         f2fs_put_dnode(&dn);
 625 out:
 626         f2fs_put_page(page, 1);
 627 }
 628
 629 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
 630 {
 631         struct page *page;
 632
 633         page = get_lock_data_page(inode, bidx, true);
 634         if (IS_ERR(page))
 635                 return;
 636
 637         if (gc_type == BG_GC) {
 638                 if (PageWriteback(page))
 639                         goto out;
 640                 set_page_dirty(page);
 641                 set_cold_data(page);
 642         } else {
 643                 struct f2fs_io_info fio = {
 644                         .sbi = F2FS_I_SB(inode),
 645                         .type = DATA,
 646                         .rw = WRITE_SYNC,
 647                         .page = page,
 648                         .encrypted_page = NULL,
 649                 };
 650                 set_page_dirty(page);
 651                 f2fs_wait_on_page_writeback(page, DATA);
 652                 if (clear_page_dirty_for_io(page))
 653                         inode_dec_dirty_pages(inode);
 654                 set_cold_data(page);
 655                 do_write_data_page(&fio);
 656                 clear_cold_data(page);
 657         }
 658 out:
 659         f2fs_put_page(page, 1);
 660 }
 661
 662 /*
 663  * This function tries to get parent node of victim data block, and identifies
 664  * data block validity. If the block is valid, copy that with cold status and
 665  * modify parent node.
 666  * If the parent node is not valid or the data block address is different,
 667  * the victim data block is ignored.
 668  */
 669 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 670                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 671 {
 672         struct super_block *sb = sbi->sb;
 673         struct f2fs_summary *entry;
 674         block_t start_addr;
 675         int off;
 676         int phase = 0;
 677
 678         start_addr = START_BLOCK(sbi, segno);
 679
 680 next_step:
 681         entry = sum;
 682
 683         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 684                 struct page *data_page;
 685                 struct inode *inode;
 686                 struct node_info dni; /* dnode info for the data */
 687                 unsigned int ofs_in_node, nofs;
 688                 block_t start_bidx;
 689
 690                 /* stop BG_GC if there is not enough free sections. */
 691                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 692                         return 0;
 693
 694                 if (check_valid_map(sbi, segno, off) == 0)
 695                         continue;
 696
 697                 if (phase == 0) {
 698                         ra_node_page(sbi, le32_to_cpu(entry->nid));
 699                         continue;
 700                 }
 701
 702                 /* Get an inode by ino with checking validity */
 703                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 704                         continue;
 705
 706                 if (phase == 1) {
 707                         ra_node_page(sbi, dni.ino);
 708                         continue;
 709                 }
 710
 711                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 712
 713                 if (phase == 2) {
 714                         inode = f2fs_iget(sb, dni.ino);
 715                         if (IS_ERR(inode) || is_bad_inode(inode))
 716                                 continue;
 717
 718                         /* if encrypted inode, let's go phase 3 */
 719                         if (f2fs_encrypted_inode(inode) &&
 720                                                 S_ISREG(inode->i_mode)) {
 721                                 add_gc_inode(gc_list, inode);
 722                                 continue;
 723                         }
 724
 725                         start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
 726                         data_page = get_read_data_page(inode,
 727                                         start_bidx + ofs_in_node, READA, true);
 728                         if (IS_ERR(data_page)) {
 729                                 iput(inode);
 730                                 continue;
 731                         }
 732
 733                         f2fs_put_page(data_page, 0);
 734                         add_gc_inode(gc_list, inode);
 735                         continue;
 736                 }
 737
 738                 /* phase 3 */
 739                 inode = find_gc_inode(gc_list, dni.ino);
 740                 if (inode) {
 741                         start_bidx = start_bidx_of_node(nofs, F2FS_I(inode))
 742                                                                 + ofs_in_node;
 743                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 744                                 move_encrypted_block(inode, start_bidx);
 745                         else
 746                                 move_data_page(inode, start_bidx, gc_type);
 747                         stat_inc_data_blk_count(sbi, 1, gc_type);
 748                 }
 749         }
 750
 751         if (++phase < 4)
 752                 goto next_step;
 753
 754         if (gc_type == FG_GC) {
 755                 f2fs_submit_merged_bio(sbi, DATA, WRITE);
 756
 757                 /* return 1 only if FG_GC succefully reclaimed one */
 758                 if (get_valid_blocks(sbi, segno, 1) == 0)
 759                         return 1;
 760         }
 761         return 0;
 762 }
 763
 764 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 765                         int gc_type)
 766 {
 767         struct sit_info *sit_i = SIT_I(sbi);
 768         int ret;
 769
 770         mutex_lock(&sit_i->sentry_lock);
 771         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 772                                               NO_CHECK_TYPE, LFS);
 773         mutex_unlock(&sit_i->sentry_lock);
 774         return ret;
 775 }
 776
 777 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
 778                                 struct gc_inode_list *gc_list, int gc_type)
 779 {
 780         struct page *sum_page;
 781         struct f2fs_summary_block *sum;
 782         struct blk_plug plug;
 783         int nfree = 0;
 784
 785         /* read segment summary of victim */
 786         sum_page = get_sum_page(sbi, segno);
 787
 788         blk_start_plug(&plug);
 789
 790         sum = page_address(sum_page);
 791
 792         /*
 793          * this is to avoid deadlock:
 794          * - lock_page(sum_page)         - f2fs_replace_block
 795          *  - check_valid_map()            - mutex_lock(sentry_lock)
 796          *   - mutex_lock(sentry_lock)     - change_curseg()
 797          *                                  - lock_page(sum_page)
 798          */
 799         unlock_page(sum_page);
 800
 801         switch (GET_SUM_TYPE((&sum->footer))) {
 802         case SUM_TYPE_NODE:
 803                 nfree = gc_node_segment(sbi, sum->entries, segno, gc_type);
 804                 break;
 805         case SUM_TYPE_DATA:
 806                 nfree = gc_data_segment(sbi, sum->entries, gc_list,
 807                                                         segno, gc_type);
 808                 break;
 809         }
 810         blk_finish_plug(&plug);
 811
 812         stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)), gc_type);
 813         stat_inc_call_count(sbi->stat_info);
 814
 815         f2fs_put_page(sum_page, 0);
 816         return nfree;
 817 }
 818
 819 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync)
 820 {
 821         unsigned int segno, i;
 822         int gc_type = sync ? FG_GC : BG_GC;
 823         int sec_freed = 0;
 824         int ret = -EINVAL;
 825         struct cp_control cpc;
 826         struct gc_inode_list gc_list = {
 827                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 828                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 829         };
 830
 831         cpc.reason = __get_cp_reason(sbi);
 832 gc_more:
 833         segno = NULL_SEGNO;
 834
 835         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 836                 goto stop;
 837         if (unlikely(f2fs_cp_error(sbi)))
 838                 goto stop;
 839
 840         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed)) {
 841                 gc_type = FG_GC;
 842                 if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
 843                         write_checkpoint(sbi, &cpc);
 844         }
 845
 846         if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
 847                 goto stop;
 848         ret = 0;
 849
 850         /* readahead multi ssa blocks those have contiguous address */
 851         if (sbi->segs_per_sec > 1)
 852                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
 853                                                         META_SSA, true);
 854
 855         for (i = 0; i < sbi->segs_per_sec; i++) {
 856                 /*
 857                  * for FG_GC case, halt gcing left segments once failed one
 858                  * of segments in selected section to avoid long latency.
 859                  */
 860                 if (!do_garbage_collect(sbi, segno + i, &gc_list, gc_type) &&
 861                                 gc_type == FG_GC)
 862                         break;
 863         }
 864
 865         if (i == sbi->segs_per_sec && gc_type == FG_GC)
 866                 sec_freed++;
 867
 868         if (gc_type == FG_GC)
 869                 sbi->cur_victim_sec = NULL_SEGNO;
 870
 871         if (!sync) {
 872                 if (has_not_enough_free_secs(sbi, sec_freed))
 873                         goto gc_more;
 874
 875                 if (gc_type == FG_GC)
 876                         write_checkpoint(sbi, &cpc);
 877         }
 878 stop:
 879         mutex_unlock(&sbi->gc_mutex);
 880
 881         put_gc_inode(&gc_list);
 882
 883         if (sync && !ret)
 884                 ret = sec_freed ? 0 : -EAGAIN;
 885         return ret;
 886 }
 887
 888 void build_gc_manager(struct f2fs_sb_info *sbi)
 889 {
 890         DIRTY_I(sbi)->v_ops = &default_v_ops;
 891 }