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
  20 #include "f2fs.h"
  21 #include "node.h"
  22 #include "segment.h"
  23 #include "gc.h"
  24 #include <trace/events/f2fs.h>
  25
  26 static int gc_thread_func(void *data)
  27 {
  28         struct f2fs_sb_info *sbi = data;
  29         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
  30         wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
  31         long wait_ms;
  32
  33         wait_ms = gc_th->min_sleep_time;
  34
  35         do {
  36                 if (try_to_freeze())
  37                         continue;
  38                 else
  39                         wait_event_interruptible_timeout(*wq,
  40                                                 kthread_should_stop(),
  41                                                 msecs_to_jiffies(wait_ms));
  42                 if (kthread_should_stop())
  43                         break;
  44
  45                 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
  46                         increase_sleep_time(gc_th, &wait_ms);
  47                         continue;
  48                 }
  49
  50                 /*
  51                  * [GC triggering condition]
  52                  * 0. GC is not conducted currently.
  53                  * 1. There are enough dirty segments.
  54                  * 2. IO subsystem is idle by checking the # of writeback pages.
  55                  * 3. IO subsystem is idle by checking the # of requests in
  56                  *    bdev's request list.
  57                  *
  58                  * Note) We have to avoid triggering GCs frequently.
  59                  * Because it is possible that some segments can be
  60                  * invalidated soon after by user update or deletion.
  61                  * So, I'd like to wait some time to collect dirty segments.
  62                  */
  63                 if (!mutex_trylock(&sbi->gc_mutex))
  64                         continue;
  65
  66                 if (!is_idle(sbi)) {
  67                         increase_sleep_time(gc_th, &wait_ms);
  68                         mutex_unlock(&sbi->gc_mutex);
  69                         continue;
  70                 }
  71
  72                 if (has_enough_invalid_blocks(sbi))
  73                         decrease_sleep_time(gc_th, &wait_ms);
  74                 else
  75                         increase_sleep_time(gc_th, &wait_ms);
  76
  77                 stat_inc_bggc_count(sbi);
  78
  79                 /* if return value is not zero, no victim was selected */
  80                 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC)))
  81                         wait_ms = gc_th->no_gc_sleep_time;
  82
  83                 trace_f2fs_background_gc(sbi->sb, wait_ms,
  84                                 prefree_segments(sbi), free_segments(sbi));
  85
  86                 /* balancing f2fs's metadata periodically */
  87                 f2fs_balance_fs_bg(sbi);
  88
  89         } while (!kthread_should_stop());
  90         return 0;
  91 }
  92
  93 int start_gc_thread(struct f2fs_sb_info *sbi)
  94 {
  95         struct f2fs_gc_kthread *gc_th;
  96         dev_t dev = sbi->sb->s_bdev->bd_dev;
  97         int err = 0;
  98
  99         gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 100         if (!gc_th) {
 101                 err = -ENOMEM;
 102                 goto out;
 103         }
 104
 105         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 106         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 107         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 108
 109         gc_th->gc_idle = 0;
 110
 111         sbi->gc_thread = gc_th;
 112         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 113         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 114                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 115         if (IS_ERR(gc_th->f2fs_gc_task)) {
 116                 err = PTR_ERR(gc_th->f2fs_gc_task);
 117                 kfree(gc_th);
 118                 sbi->gc_thread = NULL;
 119         }
 120 out:
 121         return err;
 122 }
 123
 124 void stop_gc_thread(struct f2fs_sb_info *sbi)
 125 {
 126         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 127         if (!gc_th)
 128                 return;
 129         kthread_stop(gc_th->f2fs_gc_task);
 130         kfree(gc_th);
 131         sbi->gc_thread = NULL;
 132 }
 133
 134 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 135 {
 136         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 137
 138         if (gc_th && gc_th->gc_idle) {
 139                 if (gc_th->gc_idle == 1)
 140                         gc_mode = GC_CB;
 141                 else if (gc_th->gc_idle == 2)
 142                         gc_mode = GC_GREEDY;
 143         }
 144         return gc_mode;
 145 }
 146
 147 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 148                         int type, struct victim_sel_policy *p)
 149 {
 150         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 151
 152         if (p->alloc_mode == SSR) {
 153                 p->gc_mode = GC_GREEDY;
 154                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 155                 p->max_search = dirty_i->nr_dirty[type];
 156                 p->ofs_unit = 1;
 157         } else {
 158                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 159                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 160                 p->max_search = dirty_i->nr_dirty[DIRTY];
 161                 p->ofs_unit = sbi->segs_per_sec;
 162         }
 163
 164         if (p->max_search > sbi->max_victim_search)
 165                 p->max_search = sbi->max_victim_search;
 166
 167         p->offset = sbi->last_victim[p->gc_mode];
 168 }
 169
 170 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 171                                 struct victim_sel_policy *p)
 172 {
 173         /* SSR allocates in a segment unit */
 174         if (p->alloc_mode == SSR)
 175                 return sbi->blocks_per_seg;
 176         if (p->gc_mode == GC_GREEDY)
 177                 return sbi->blocks_per_seg * p->ofs_unit;
 178         else if (p->gc_mode == GC_CB)
 179                 return UINT_MAX;
 180         else /* No other gc_mode */
 181                 return 0;
 182 }
 183
 184 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 185 {
 186         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 187         unsigned int secno;
 188
 189         /*
 190          * If the gc_type is FG_GC, we can select victim segments
 191          * selected by background GC before.
 192          * Those segments guarantee they have small valid blocks.
 193          */
 194         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 195                 if (sec_usage_check(sbi, secno))
 196                         continue;
 197                 clear_bit(secno, dirty_i->victim_secmap);
 198                 return secno * sbi->segs_per_sec;
 199         }
 200         return NULL_SEGNO;
 201 }
 202
 203 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 204 {
 205         struct sit_info *sit_i = SIT_I(sbi);
 206         unsigned int secno = GET_SECNO(sbi, segno);
 207         unsigned int start = secno * sbi->segs_per_sec;
 208         unsigned long long mtime = 0;
 209         unsigned int vblocks;
 210         unsigned char age = 0;
 211         unsigned char u;
 212         unsigned int i;
 213
 214         for (i = 0; i < sbi->segs_per_sec; i++)
 215                 mtime += get_seg_entry(sbi, start + i)->mtime;
 216         vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 217
 218         mtime = div_u64(mtime, sbi->segs_per_sec);
 219         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 220
 221         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 222
 223         /* Handle if the system time has changed by the user */
 224         if (mtime < sit_i->min_mtime)
 225                 sit_i->min_mtime = mtime;
 226         if (mtime > sit_i->max_mtime)
 227                 sit_i->max_mtime = mtime;
 228         if (sit_i->max_mtime != sit_i->min_mtime)
 229                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 230                                 sit_i->max_mtime - sit_i->min_mtime);
 231
 232         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 233 }
 234
 235 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 236                         unsigned int segno, struct victim_sel_policy *p)
 237 {
 238         if (p->alloc_mode == SSR)
 239                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 240
 241         /* alloc_mode == LFS */
 242         if (p->gc_mode == GC_GREEDY)
 243                 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 244         else
 245                 return get_cb_cost(sbi, segno);
 246 }
 247
 248 /*
 249  * This function is called from two paths.
 250  * One is garbage collection and the other is SSR segment selection.
 251  * When it is called during GC, it just gets a victim segment
 252  * and it does not remove it from dirty seglist.
 253  * When it is called from SSR segment selection, it finds a segment
 254  * which has minimum valid blocks and removes it from dirty seglist.
 255  */
 256 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 257                 unsigned int *result, int gc_type, int type, char alloc_mode)
 258 {
 259         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 260         struct victim_sel_policy p;
 261         unsigned int secno, max_cost;
 262         unsigned int last_segment = MAIN_SEGS(sbi);
 263         int nsearched = 0;
 264
 265         mutex_lock(&dirty_i->seglist_lock);
 266
 267         p.alloc_mode = alloc_mode;
 268         select_policy(sbi, gc_type, type, &p);
 269
 270         p.min_segno = NULL_SEGNO;
 271         p.min_cost = max_cost = get_max_cost(sbi, &p);
 272
 273         if (p.max_search == 0)
 274                 goto out;
 275
 276         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 277                 p.min_segno = check_bg_victims(sbi);
 278                 if (p.min_segno != NULL_SEGNO)
 279                         goto got_it;
 280         }
 281
 282         while (1) {
 283                 unsigned long cost;
 284                 unsigned int segno;
 285
 286                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 287                 if (segno >= last_segment) {
 288                         if (sbi->last_victim[p.gc_mode]) {
 289                                 last_segment = sbi->last_victim[p.gc_mode];
 290                                 sbi->last_victim[p.gc_mode] = 0;
 291                                 p.offset = 0;
 292                                 continue;
 293                         }
 294                         break;
 295                 }
 296
 297                 p.offset = segno + p.ofs_unit;
 298                 if (p.ofs_unit > 1)
 299                         p.offset -= segno % p.ofs_unit;
 300
 301                 secno = GET_SECNO(sbi, segno);
 302
 303                 if (sec_usage_check(sbi, secno))
 304                         continue;
 305                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 306                         continue;
 307
 308                 cost = get_gc_cost(sbi, segno, &p);
 309
 310                 if (p.min_cost > cost) {
 311                         p.min_segno = segno;
 312                         p.min_cost = cost;
 313                 } else if (unlikely(cost == max_cost)) {
 314                         continue;
 315                 }
 316
 317                 if (nsearched++ >= p.max_search) {
 318                         sbi->last_victim[p.gc_mode] = segno;
 319                         break;
 320                 }
 321         }
 322         if (p.min_segno != NULL_SEGNO) {
 323 got_it:
 324                 if (p.alloc_mode == LFS) {
 325                         secno = GET_SECNO(sbi, p.min_segno);
 326                         if (gc_type == FG_GC)
 327                                 sbi->cur_victim_sec = secno;
 328                         else
 329                                 set_bit(secno, dirty_i->victim_secmap);
 330                 }
 331                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 332
 333                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 334                                 sbi->cur_victim_sec,
 335                                 prefree_segments(sbi), free_segments(sbi));
 336         }
 337 out:
 338         mutex_unlock(&dirty_i->seglist_lock);
 339
 340         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 341 }
 342
 343 static const struct victim_selection default_v_ops = {
 344         .get_victim = get_victim_by_default,
 345 };
 346
 347 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 348 {
 349         struct inode_entry *ie;
 350
 351         ie = radix_tree_lookup(&gc_list->iroot, ino);
 352         if (ie)
 353                 return ie->inode;
 354         return NULL;
 355 }
 356
 357 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 358 {
 359         struct inode_entry *new_ie;
 360
 361         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 362                 iput(inode);
 363                 return;
 364         }
 365         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 366         new_ie->inode = inode;
 367
 368         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 369         list_add_tail(&new_ie->list, &gc_list->ilist);
 370 }
 371
 372 static void put_gc_inode(struct gc_inode_list *gc_list)
 373 {
 374         struct inode_entry *ie, *next_ie;
 375         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 376                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 377                 iput(ie->inode);
 378                 list_del(&ie->list);
 379                 kmem_cache_free(inode_entry_slab, ie);
 380         }
 381 }
 382
 383 static int check_valid_map(struct f2fs_sb_info *sbi,
 384                                 unsigned int segno, int offset)
 385 {
 386         struct sit_info *sit_i = SIT_I(sbi);
 387         struct seg_entry *sentry;
 388         int ret;
 389
 390         mutex_lock(&sit_i->sentry_lock);
 391         sentry = get_seg_entry(sbi, segno);
 392         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 393         mutex_unlock(&sit_i->sentry_lock);
 394         return ret;
 395 }
 396
 397 /*
 398  * This function compares node address got in summary with that in NAT.
 399  * On validity, copy that node with cold status, otherwise (invalid node)
 400  * ignore that.
 401  */
 402 static int gc_node_segment(struct f2fs_sb_info *sbi,
 403                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 404 {
 405         bool initial = true;
 406         struct f2fs_summary *entry;
 407         block_t start_addr;
 408         int off;
 409
 410         start_addr = START_BLOCK(sbi, segno);
 411
 412 next_step:
 413         entry = sum;
 414
 415         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 416                 nid_t nid = le32_to_cpu(entry->nid);
 417                 struct page *node_page;
 418                 struct node_info ni;
 419
 420                 /* stop BG_GC if there is not enough free sections. */
 421                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 422                         return 0;
 423
 424                 if (check_valid_map(sbi, segno, off) == 0)
 425                         continue;
 426
 427                 if (initial) {
 428                         ra_node_page(sbi, nid);
 429                         continue;
 430                 }
 431                 node_page = get_node_page(sbi, nid);
 432                 if (IS_ERR(node_page))
 433                         continue;
 434
 435                 /* block may become invalid during get_node_page */
 436                 if (check_valid_map(sbi, segno, off) == 0) {
 437                         f2fs_put_page(node_page, 1);
 438                         continue;
 439                 }
 440
 441                 get_node_info(sbi, nid, &ni);
 442                 if (ni.blk_addr != start_addr + off) {
 443                         f2fs_put_page(node_page, 1);
 444                         continue;
 445                 }
 446
 447                 /* set page dirty and write it */
 448                 if (gc_type == FG_GC) {
 449                         f2fs_wait_on_page_writeback(node_page, NODE);
 450                         set_page_dirty(node_page);
 451                 } else {
 452                         if (!PageWriteback(node_page))
 453                                 set_page_dirty(node_page);
 454                 }
 455                 f2fs_put_page(node_page, 1);
 456                 stat_inc_node_blk_count(sbi, 1, gc_type);
 457         }
 458
 459         if (initial) {
 460                 initial = false;
 461                 goto next_step;
 462         }
 463
 464         if (gc_type == FG_GC) {
 465                 struct writeback_control wbc = {
 466                         .sync_mode = WB_SYNC_ALL,
 467                         .nr_to_write = LONG_MAX,
 468                         .for_reclaim = 0,
 469                 };
 470                 sync_node_pages(sbi, 0, &wbc);
 471
 472                 /* return 1 only if FG_GC succefully reclaimed one */
 473                 if (get_valid_blocks(sbi, segno, 1) == 0)
 474                         return 1;
 475         }
 476         return 0;
 477 }
 478
 479 /*
 480  * Calculate start block index indicating the given node offset.
 481  * Be careful, caller should give this node offset only indicating direct node
 482  * blocks. If any node offsets, which point the other types of node blocks such
 483  * as indirect or double indirect node blocks, are given, it must be a caller's
 484  * bug.
 485  */
 486 block_t start_bidx_of_node(unsigned int node_ofs, struct f2fs_inode_info *fi)
 487 {
 488         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 489         unsigned int bidx;
 490
 491         if (node_ofs == 0)
 492                 return 0;
 493
 494         if (node_ofs <= 2) {
 495                 bidx = node_ofs - 1;
 496         } else if (node_ofs <= indirect_blks) {
 497                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 498                 bidx = node_ofs - 2 - dec;
 499         } else {
 500                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 501                 bidx = node_ofs - 5 - dec;
 502         }
 503         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(fi);
 504 }
 505
 506 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 507                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 508 {
 509         struct page *node_page;
 510         nid_t nid;
 511         unsigned int ofs_in_node;
 512         block_t source_blkaddr;
 513
 514         nid = le32_to_cpu(sum->nid);
 515         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 516
 517         node_page = get_node_page(sbi, nid);
 518         if (IS_ERR(node_page))
 519                 return false;
 520
 521         get_node_info(sbi, nid, dni);
 522
 523         if (sum->version != dni->version) {
 524                 f2fs_put_page(node_page, 1);
 525                 return false;
 526         }
 527
 528         *nofs = ofs_of_node(node_page);
 529         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 530         f2fs_put_page(node_page, 1);
 531
 532         if (source_blkaddr != blkaddr)
 533                 return false;
 534         return true;
 535 }
 536
 537 static void move_encrypted_block(struct inode *inode, block_t bidx)
 538 {
 539         struct f2fs_io_info fio = {
 540                 .sbi = F2FS_I_SB(inode),
 541                 .type = DATA,
 542                 .rw = READ_SYNC,
 543                 .encrypted_page = NULL,
 544         };
 545         struct dnode_of_data dn;
 546         struct f2fs_summary sum;
 547         struct node_info ni;
 548         struct page *page;
 549         int err;
 550
 551         /* do not read out */
 552         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 553         if (!page)
 554                 return;
 555
 556         set_new_dnode(&dn, inode, NULL, NULL, 0);
 557         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 558         if (err)
 559                 goto out;
 560
 561         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 562                 ClearPageUptodate(page);
 563                 goto put_out;
 564         }
 565
 566         /*
 567          * don't cache encrypted data into meta inode until previous dirty
 568          * data were writebacked to avoid racing between GC and flush.
 569          */
 570         f2fs_wait_on_page_writeback(page, DATA);
 571
 572         get_node_info(fio.sbi, dn.nid, &ni);
 573         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 574
 575         /* read page */
 576         fio.page = page;
 577         fio.blk_addr = dn.data_blkaddr;
 578
 579         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
 580                                         fio.blk_addr,
 581                                         FGP_LOCK|FGP_CREAT,
 582                                         GFP_NOFS);
 583         if (!fio.encrypted_page)
 584                 goto put_out;
 585
 586         err = f2fs_submit_page_bio(&fio);
 587         if (err)
 588                 goto put_page_out;
 589
 590         /* write page */
 591         lock_page(fio.encrypted_page);
 592
 593         if (unlikely(!PageUptodate(fio.encrypted_page)))
 594                 goto put_page_out;
 595         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
 596                 goto put_page_out;
 597
 598         set_page_dirty(fio.encrypted_page);
 599         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA);
 600         if (clear_page_dirty_for_io(fio.encrypted_page))
 601                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 602
 603         set_page_writeback(fio.encrypted_page);
 604
 605         /* allocate block address */
 606         f2fs_wait_on_page_writeback(dn.node_page, NODE);
 607         allocate_data_block(fio.sbi, NULL, fio.blk_addr,
 608                                         &fio.blk_addr, &sum, CURSEG_COLD_DATA);
 609         fio.rw = WRITE_SYNC;
 610         f2fs_submit_page_mbio(&fio);
 611
 612         dn.data_blkaddr = fio.blk_addr;
 613         set_data_blkaddr(&dn);
 614         f2fs_update_extent_cache(&dn);
 615         set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
 616         if (page->index == 0)
 617                 set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
 618 put_page_out:
 619         f2fs_put_page(fio.encrypted_page, 1);
 620 put_out:
 621         f2fs_put_dnode(&dn);
 622 out:
 623         f2fs_put_page(page, 1);
 624 }
 625
 626 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
 627 {
 628         struct page *page;
 629
 630         page = get_lock_data_page(inode, bidx, true);
 631         if (IS_ERR(page))
 632                 return;
 633
 634         if (gc_type == BG_GC) {
 635                 if (PageWriteback(page))
 636                         goto out;
 637                 set_page_dirty(page);
 638                 set_cold_data(page);
 639         } else {
 640                 struct f2fs_io_info fio = {
 641                         .sbi = F2FS_I_SB(inode),
 642                         .type = DATA,
 643                         .rw = WRITE_SYNC,
 644                         .page = page,
 645                         .encrypted_page = NULL,
 646                 };
 647                 set_page_dirty(page);
 648                 f2fs_wait_on_page_writeback(page, DATA);
 649                 if (clear_page_dirty_for_io(page))
 650                         inode_dec_dirty_pages(inode);
 651                 set_cold_data(page);
 652                 do_write_data_page(&fio);
 653                 clear_cold_data(page);
 654         }
 655 out:
 656         f2fs_put_page(page, 1);
 657 }
 658
 659 /*
 660  * This function tries to get parent node of victim data block, and identifies
 661  * data block validity. If the block is valid, copy that with cold status and
 662  * modify parent node.
 663  * If the parent node is not valid or the data block address is different,
 664  * the victim data block is ignored.
 665  */
 666 static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 667                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 668 {
 669         struct super_block *sb = sbi->sb;
 670         struct f2fs_summary *entry;
 671         block_t start_addr;
 672         int off;
 673         int phase = 0;
 674
 675         start_addr = START_BLOCK(sbi, segno);
 676
 677 next_step:
 678         entry = sum;
 679
 680         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 681                 struct page *data_page;
 682                 struct inode *inode;
 683                 struct node_info dni; /* dnode info for the data */
 684                 unsigned int ofs_in_node, nofs;
 685                 block_t start_bidx;
 686
 687                 /* stop BG_GC if there is not enough free sections. */
 688                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 689                         return 0;
 690
 691                 if (check_valid_map(sbi, segno, off) == 0)
 692                         continue;
 693
 694                 if (phase == 0) {
 695                         ra_node_page(sbi, le32_to_cpu(entry->nid));
 696                         continue;
 697                 }
 698
 699                 /* Get an inode by ino with checking validity */
 700                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 701                         continue;
 702
 703                 if (phase == 1) {
 704                         ra_node_page(sbi, dni.ino);
 705                         continue;
 706                 }
 707
 708                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 709
 710                 if (phase == 2) {
 711                         inode = f2fs_iget(sb, dni.ino);
 712                         if (IS_ERR(inode) || is_bad_inode(inode))
 713                                 continue;
 714
 715                         /* if encrypted inode, let's go phase 3 */
 716                         if (f2fs_encrypted_inode(inode) &&
 717                                                 S_ISREG(inode->i_mode)) {
 718                                 add_gc_inode(gc_list, inode);
 719                                 continue;
 720                         }
 721
 722                         start_bidx = start_bidx_of_node(nofs, F2FS_I(inode));
 723                         data_page = get_read_data_page(inode,
 724                                         start_bidx + ofs_in_node, READA, true);
 725                         if (IS_ERR(data_page)) {
 726                                 iput(inode);
 727                                 continue;
 728                         }
 729
 730                         f2fs_put_page(data_page, 0);
 731                         add_gc_inode(gc_list, inode);
 732                         continue;
 733                 }
 734
 735                 /* phase 3 */
 736                 inode = find_gc_inode(gc_list, dni.ino);
 737                 if (inode) {
 738                         start_bidx = start_bidx_of_node(nofs, F2FS_I(inode))
 739                                                                 + ofs_in_node;
 740                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 741                                 move_encrypted_block(inode, start_bidx);
 742                         else
 743                                 move_data_page(inode, start_bidx, gc_type);
 744                         stat_inc_data_blk_count(sbi, 1, gc_type);
 745                 }
 746         }
 747
 748         if (++phase < 4)
 749                 goto next_step;
 750
 751         if (gc_type == FG_GC) {
 752                 f2fs_submit_merged_bio(sbi, DATA, WRITE);
 753
 754                 /* return 1 only if FG_GC succefully reclaimed one */
 755                 if (get_valid_blocks(sbi, segno, 1) == 0)
 756                         return 1;
 757         }
 758         return 0;
 759 }
 760
 761 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 762                         int gc_type)
 763 {
 764         struct sit_info *sit_i = SIT_I(sbi);
 765         int ret;
 766
 767         mutex_lock(&sit_i->sentry_lock);
 768         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 769                                               NO_CHECK_TYPE, LFS);
 770         mutex_unlock(&sit_i->sentry_lock);
 771         return ret;
 772 }
 773
 774 static int do_garbage_collect(struct f2fs_sb_info *sbi, unsigned int segno,
 775                                 struct gc_inode_list *gc_list, int gc_type)
 776 {
 777         struct page *sum_page;
 778         struct f2fs_summary_block *sum;
 779         struct blk_plug plug;
 780         int nfree = 0;
 781
 782         /* read segment summary of victim */
 783         sum_page = get_sum_page(sbi, segno);
 784
 785         blk_start_plug(&plug);
 786
 787         sum = page_address(sum_page);
 788
 789         /*
 790          * this is to avoid deadlock:
 791          * - lock_page(sum_page)         - f2fs_replace_block
 792          *  - check_valid_map()            - mutex_lock(sentry_lock)
 793          *   - mutex_lock(sentry_lock)     - change_curseg()
 794          *                                  - lock_page(sum_page)
 795          */
 796         unlock_page(sum_page);
 797
 798         switch (GET_SUM_TYPE((&sum->footer))) {
 799         case SUM_TYPE_NODE:
 800                 nfree = gc_node_segment(sbi, sum->entries, segno, gc_type);
 801                 break;
 802         case SUM_TYPE_DATA:
 803                 nfree = gc_data_segment(sbi, sum->entries, gc_list,
 804                                                         segno, gc_type);
 805                 break;
 806         }
 807         blk_finish_plug(&plug);
 808
 809         stat_inc_seg_count(sbi, GET_SUM_TYPE((&sum->footer)), gc_type);
 810         stat_inc_call_count(sbi->stat_info);
 811
 812         f2fs_put_page(sum_page, 0);
 813         return nfree;
 814 }
 815
 816 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync)
 817 {
 818         unsigned int segno, i;
 819         int gc_type = sync ? FG_GC : BG_GC;
 820         int sec_freed = 0;
 821         int ret = -EINVAL;
 822         struct cp_control cpc;
 823         struct gc_inode_list gc_list = {
 824                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 825                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 826         };
 827
 828         cpc.reason = __get_cp_reason(sbi);
 829 gc_more:
 830         segno = NULL_SEGNO;
 831
 832         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 833                 goto stop;
 834         if (unlikely(f2fs_cp_error(sbi))) {
 835                 ret = -EIO;
 836                 goto stop;
 837         }
 838
 839         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed)) {
 840                 gc_type = FG_GC;
 841                 if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
 842                         write_checkpoint(sbi, &cpc);
 843         }
 844
 845         if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
 846                 goto stop;
 847         ret = 0;
 848
 849         /* readahead multi ssa blocks those have contiguous address */
 850         if (sbi->segs_per_sec > 1)
 851                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
 852                                                         META_SSA, true);
 853
 854         for (i = 0; i < sbi->segs_per_sec; i++) {
 855                 /*
 856                  * for FG_GC case, halt gcing left segments once failed one
 857                  * of segments in selected section to avoid long latency.
 858                  */
 859                 if (!do_garbage_collect(sbi, segno + i, &gc_list, gc_type) &&
 860                                 gc_type == FG_GC)
 861                         break;
 862         }
 863
 864         if (i == sbi->segs_per_sec && gc_type == FG_GC)
 865                 sec_freed++;
 866
 867         if (gc_type == FG_GC)
 868                 sbi->cur_victim_sec = NULL_SEGNO;
 869
 870         if (!sync) {
 871                 if (has_not_enough_free_secs(sbi, sec_freed))
 872                         goto gc_more;
 873
 874                 if (gc_type == FG_GC)
 875                         write_checkpoint(sbi, &cpc);
 876         }
 877 stop:
 878         mutex_unlock(&sbi->gc_mutex);
 879
 880         put_gc_inode(&gc_list);
 881
 882         if (sync)
 883                 ret = sec_freed ? 0 : -EAGAIN;
 884         return ret;
 885 }
 886
 887 void build_gc_manager(struct f2fs_sb_info *sbi)
 888 {
 889         DIRTY_I(sbi)->v_ops = &default_v_ops;
 890 }