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 #ifdef CONFIG_F2FS_FAULT_INJECTION
  51                 if (time_to_inject(sbi, FAULT_CHECKPOINT))
  52                         f2fs_stop_checkpoint(sbi, false);
  53 #endif
  54
  55                 /*
  56                  * [GC triggering condition]
  57                  * 0. GC is not conducted currently.
  58                  * 1. There are enough dirty segments.
  59                  * 2. IO subsystem is idle by checking the # of writeback pages.
  60                  * 3. IO subsystem is idle by checking the # of requests in
  61                  *    bdev's request list.
  62                  *
  63                  * Note) We have to avoid triggering GCs frequently.
  64                  * Because it is possible that some segments can be
  65                  * invalidated soon after by user update or deletion.
  66                  * So, I'd like to wait some time to collect dirty segments.
  67                  */
  68                 if (!mutex_trylock(&sbi->gc_mutex))
  69                         continue;
  70
  71                 if (!is_idle(sbi)) {
  72                         increase_sleep_time(gc_th, &wait_ms);
  73                         mutex_unlock(&sbi->gc_mutex);
  74                         continue;
  75                 }
  76
  77                 if (has_enough_invalid_blocks(sbi))
  78                         decrease_sleep_time(gc_th, &wait_ms);
  79                 else
  80                         increase_sleep_time(gc_th, &wait_ms);
  81
  82                 stat_inc_bggc_count(sbi);
  83
  84                 /* if return value is not zero, no victim was selected */
  85                 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true))
  86                         wait_ms = gc_th->no_gc_sleep_time;
  87
  88                 trace_f2fs_background_gc(sbi->sb, wait_ms,
  89                                 prefree_segments(sbi), free_segments(sbi));
  90
  91                 /* balancing f2fs's metadata periodically */
  92                 f2fs_balance_fs_bg(sbi);
  93
  94         } while (!kthread_should_stop());
  95         return 0;
  96 }
  97
  98 int start_gc_thread(struct f2fs_sb_info *sbi)
  99 {
 100         struct f2fs_gc_kthread *gc_th;
 101         dev_t dev = sbi->sb->s_bdev->bd_dev;
 102         int err = 0;
 103
 104         gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 105         if (!gc_th) {
 106                 err = -ENOMEM;
 107                 goto out;
 108         }
 109
 110         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 111         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 112         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 113
 114         gc_th->gc_idle = 0;
 115
 116         sbi->gc_thread = gc_th;
 117         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 118         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 119                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 120         if (IS_ERR(gc_th->f2fs_gc_task)) {
 121                 err = PTR_ERR(gc_th->f2fs_gc_task);
 122                 kfree(gc_th);
 123                 sbi->gc_thread = NULL;
 124         }
 125 out:
 126         return err;
 127 }
 128
 129 void stop_gc_thread(struct f2fs_sb_info *sbi)
 130 {
 131         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 132         if (!gc_th)
 133                 return;
 134         kthread_stop(gc_th->f2fs_gc_task);
 135         kfree(gc_th);
 136         sbi->gc_thread = NULL;
 137 }
 138
 139 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 140 {
 141         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 142
 143         if (gc_th && gc_th->gc_idle) {
 144                 if (gc_th->gc_idle == 1)
 145                         gc_mode = GC_CB;
 146                 else if (gc_th->gc_idle == 2)
 147                         gc_mode = GC_GREEDY;
 148         }
 149         return gc_mode;
 150 }
 151
 152 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 153                         int type, struct victim_sel_policy *p)
 154 {
 155         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 156
 157         if (p->alloc_mode == SSR) {
 158                 p->gc_mode = GC_GREEDY;
 159                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 160                 p->max_search = dirty_i->nr_dirty[type];
 161                 p->ofs_unit = 1;
 162         } else {
 163                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 164                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 165                 p->max_search = dirty_i->nr_dirty[DIRTY];
 166                 p->ofs_unit = sbi->segs_per_sec;
 167         }
 168
 169         /* we need to check every dirty segments in the FG_GC case */
 170         if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
 171                 p->max_search = sbi->max_victim_search;
 172
 173         p->offset = sbi->last_victim[p->gc_mode];
 174 }
 175
 176 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 177                                 struct victim_sel_policy *p)
 178 {
 179         /* SSR allocates in a segment unit */
 180         if (p->alloc_mode == SSR)
 181                 return sbi->blocks_per_seg;
 182         if (p->gc_mode == GC_GREEDY)
 183                 return sbi->blocks_per_seg * p->ofs_unit;
 184         else if (p->gc_mode == GC_CB)
 185                 return UINT_MAX;
 186         else /* No other gc_mode */
 187                 return 0;
 188 }
 189
 190 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 191 {
 192         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 193         unsigned int secno;
 194
 195         /*
 196          * If the gc_type is FG_GC, we can select victim segments
 197          * selected by background GC before.
 198          * Those segments guarantee they have small valid blocks.
 199          */
 200         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 201                 if (sec_usage_check(sbi, secno))
 202                         continue;
 203
 204                 if (no_fggc_candidate(sbi, secno))
 205                         continue;
 206
 207                 clear_bit(secno, dirty_i->victim_secmap);
 208                 return secno * sbi->segs_per_sec;
 209         }
 210         return NULL_SEGNO;
 211 }
 212
 213 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 214 {
 215         struct sit_info *sit_i = SIT_I(sbi);
 216         unsigned int secno = GET_SECNO(sbi, segno);
 217         unsigned int start = secno * sbi->segs_per_sec;
 218         unsigned long long mtime = 0;
 219         unsigned int vblocks;
 220         unsigned char age = 0;
 221         unsigned char u;
 222         unsigned int i;
 223
 224         for (i = 0; i < sbi->segs_per_sec; i++)
 225                 mtime += get_seg_entry(sbi, start + i)->mtime;
 226         vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 227
 228         mtime = div_u64(mtime, sbi->segs_per_sec);
 229         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 230
 231         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 232
 233         /* Handle if the system time has changed by the user */
 234         if (mtime < sit_i->min_mtime)
 235                 sit_i->min_mtime = mtime;
 236         if (mtime > sit_i->max_mtime)
 237                 sit_i->max_mtime = mtime;
 238         if (sit_i->max_mtime != sit_i->min_mtime)
 239                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 240                                 sit_i->max_mtime - sit_i->min_mtime);
 241
 242         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 243 }
 244
 245 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
 246                                                 unsigned int segno)
 247 {
 248         unsigned int valid_blocks =
 249                         get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 250
 251         return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 252                                 valid_blocks * 2 : valid_blocks;
 253 }
 254
 255 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 256                         unsigned int segno, struct victim_sel_policy *p)
 257 {
 258         if (p->alloc_mode == SSR)
 259                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 260
 261         /* alloc_mode == LFS */
 262         if (p->gc_mode == GC_GREEDY)
 263                 return get_greedy_cost(sbi, segno);
 264         else
 265                 return get_cb_cost(sbi, segno);
 266 }
 267
 268 static unsigned int count_bits(const unsigned long *addr,
 269                                 unsigned int offset, unsigned int len)
 270 {
 271         unsigned int end = offset + len, sum = 0;
 272
 273         while (offset < end) {
 274                 if (test_bit(offset++, addr))
 275                         ++sum;
 276         }
 277         return sum;
 278 }
 279
 280 /*
 281  * This function is called from two paths.
 282  * One is garbage collection and the other is SSR segment selection.
 283  * When it is called during GC, it just gets a victim segment
 284  * and it does not remove it from dirty seglist.
 285  * When it is called from SSR segment selection, it finds a segment
 286  * which has minimum valid blocks and removes it from dirty seglist.
 287  */
 288 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 289                 unsigned int *result, int gc_type, int type, char alloc_mode)
 290 {
 291         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 292         struct victim_sel_policy p;
 293         unsigned int secno, last_victim;
 294         unsigned int last_segment = MAIN_SEGS(sbi);
 295         unsigned int nsearched = 0;
 296
 297         mutex_lock(&dirty_i->seglist_lock);
 298
 299         p.alloc_mode = alloc_mode;
 300         select_policy(sbi, gc_type, type, &p);
 301
 302         p.min_segno = NULL_SEGNO;
 303         p.min_cost = get_max_cost(sbi, &p);
 304
 305         if (p.max_search == 0)
 306                 goto out;
 307
 308         last_victim = sbi->last_victim[p.gc_mode];
 309         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 310                 p.min_segno = check_bg_victims(sbi);
 311                 if (p.min_segno != NULL_SEGNO)
 312                         goto got_it;
 313         }
 314
 315         while (1) {
 316                 unsigned long cost;
 317                 unsigned int segno;
 318
 319                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 320                 if (segno >= last_segment) {
 321                         if (sbi->last_victim[p.gc_mode]) {
 322                                 last_segment = sbi->last_victim[p.gc_mode];
 323                                 sbi->last_victim[p.gc_mode] = 0;
 324                                 p.offset = 0;
 325                                 continue;
 326                         }
 327                         break;
 328                 }
 329
 330                 p.offset = segno + p.ofs_unit;
 331                 if (p.ofs_unit > 1) {
 332                         p.offset -= segno % p.ofs_unit;
 333                         nsearched += count_bits(p.dirty_segmap,
 334                                                 p.offset - p.ofs_unit,
 335                                                 p.ofs_unit);
 336                 } else {
 337                         nsearched++;
 338                 }
 339
 340                 secno = GET_SECNO(sbi, segno);
 341
 342                 if (sec_usage_check(sbi, secno))
 343                         goto next;
 344                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 345                         goto next;
 346                 if (gc_type == FG_GC && p.alloc_mode == LFS &&
 347                                         no_fggc_candidate(sbi, secno))
 348                         goto next;
 349
 350                 cost = get_gc_cost(sbi, segno, &p);
 351
 352                 if (p.min_cost > cost) {
 353                         p.min_segno = segno;
 354                         p.min_cost = cost;
 355                 }
 356 next:
 357                 if (nsearched >= p.max_search) {
 358                         if (!sbi->last_victim[p.gc_mode] && segno <= last_victim)
 359                                 sbi->last_victim[p.gc_mode] = last_victim + 1;
 360                         else
 361                                 sbi->last_victim[p.gc_mode] = segno + 1;
 362                         break;
 363                 }
 364         }
 365         if (p.min_segno != NULL_SEGNO) {
 366 got_it:
 367                 if (p.alloc_mode == LFS) {
 368                         secno = GET_SECNO(sbi, p.min_segno);
 369                         if (gc_type == FG_GC)
 370                                 sbi->cur_victim_sec = secno;
 371                         else
 372                                 set_bit(secno, dirty_i->victim_secmap);
 373                 }
 374                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 375
 376                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 377                                 sbi->cur_victim_sec,
 378                                 prefree_segments(sbi), free_segments(sbi));
 379         }
 380 out:
 381         mutex_unlock(&dirty_i->seglist_lock);
 382
 383         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 384 }
 385
 386 static const struct victim_selection default_v_ops = {
 387         .get_victim = get_victim_by_default,
 388 };
 389
 390 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 391 {
 392         struct inode_entry *ie;
 393
 394         ie = radix_tree_lookup(&gc_list->iroot, ino);
 395         if (ie)
 396                 return ie->inode;
 397         return NULL;
 398 }
 399
 400 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 401 {
 402         struct inode_entry *new_ie;
 403
 404         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 405                 iput(inode);
 406                 return;
 407         }
 408         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 409         new_ie->inode = inode;
 410
 411         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 412         list_add_tail(&new_ie->list, &gc_list->ilist);
 413 }
 414
 415 static void put_gc_inode(struct gc_inode_list *gc_list)
 416 {
 417         struct inode_entry *ie, *next_ie;
 418         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 419                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 420                 iput(ie->inode);
 421                 list_del(&ie->list);
 422                 kmem_cache_free(inode_entry_slab, ie);
 423         }
 424 }
 425
 426 static int check_valid_map(struct f2fs_sb_info *sbi,
 427                                 unsigned int segno, int offset)
 428 {
 429         struct sit_info *sit_i = SIT_I(sbi);
 430         struct seg_entry *sentry;
 431         int ret;
 432
 433         mutex_lock(&sit_i->sentry_lock);
 434         sentry = get_seg_entry(sbi, segno);
 435         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 436         mutex_unlock(&sit_i->sentry_lock);
 437         return ret;
 438 }
 439
 440 /*
 441  * This function compares node address got in summary with that in NAT.
 442  * On validity, copy that node with cold status, otherwise (invalid node)
 443  * ignore that.
 444  */
 445 static void gc_node_segment(struct f2fs_sb_info *sbi,
 446                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 447 {
 448         struct f2fs_summary *entry;
 449         block_t start_addr;
 450         int off;
 451         int phase = 0;
 452
 453         start_addr = START_BLOCK(sbi, segno);
 454
 455 next_step:
 456         entry = sum;
 457
 458         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 459                 nid_t nid = le32_to_cpu(entry->nid);
 460                 struct page *node_page;
 461                 struct node_info ni;
 462
 463                 /* stop BG_GC if there is not enough free sections. */
 464                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 465                         return;
 466
 467                 if (check_valid_map(sbi, segno, off) == 0)
 468                         continue;
 469
 470                 if (phase == 0) {
 471                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 472                                                         META_NAT, true);
 473                         continue;
 474                 }
 475
 476                 if (phase == 1) {
 477                         ra_node_page(sbi, nid);
 478                         continue;
 479                 }
 480
 481                 /* phase == 2 */
 482                 node_page = get_node_page(sbi, nid);
 483                 if (IS_ERR(node_page))
 484                         continue;
 485
 486                 /* block may become invalid during get_node_page */
 487                 if (check_valid_map(sbi, segno, off) == 0) {
 488                         f2fs_put_page(node_page, 1);
 489                         continue;
 490                 }
 491
 492                 get_node_info(sbi, nid, &ni);
 493                 if (ni.blk_addr != start_addr + off) {
 494                         f2fs_put_page(node_page, 1);
 495                         continue;
 496                 }
 497
 498                 move_node_page(node_page, gc_type);
 499                 stat_inc_node_blk_count(sbi, 1, gc_type);
 500         }
 501
 502         if (++phase < 3)
 503                 goto next_step;
 504 }
 505
 506 /*
 507  * Calculate start block index indicating the given node offset.
 508  * Be careful, caller should give this node offset only indicating direct node
 509  * blocks. If any node offsets, which point the other types of node blocks such
 510  * as indirect or double indirect node blocks, are given, it must be a caller's
 511  * bug.
 512  */
 513 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 514 {
 515         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 516         unsigned int bidx;
 517
 518         if (node_ofs == 0)
 519                 return 0;
 520
 521         if (node_ofs <= 2) {
 522                 bidx = node_ofs - 1;
 523         } else if (node_ofs <= indirect_blks) {
 524                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 525                 bidx = node_ofs - 2 - dec;
 526         } else {
 527                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 528                 bidx = node_ofs - 5 - dec;
 529         }
 530         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 531 }
 532
 533 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 534                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 535 {
 536         struct page *node_page;
 537         nid_t nid;
 538         unsigned int ofs_in_node;
 539         block_t source_blkaddr;
 540
 541         nid = le32_to_cpu(sum->nid);
 542         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 543
 544         node_page = get_node_page(sbi, nid);
 545         if (IS_ERR(node_page))
 546                 return false;
 547
 548         get_node_info(sbi, nid, dni);
 549
 550         if (sum->version != dni->version) {
 551                 f2fs_put_page(node_page, 1);
 552                 return false;
 553         }
 554
 555         *nofs = ofs_of_node(node_page);
 556         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 557         f2fs_put_page(node_page, 1);
 558
 559         if (source_blkaddr != blkaddr)
 560                 return false;
 561         return true;
 562 }
 563
 564 static void move_encrypted_block(struct inode *inode, block_t bidx,
 565                                                         unsigned int segno, int off)
 566 {
 567         struct f2fs_io_info fio = {
 568                 .sbi = F2FS_I_SB(inode),
 569                 .type = DATA,
 570                 .op = REQ_OP_READ,
 571                 .op_flags = 0,
 572                 .encrypted_page = NULL,
 573         };
 574         struct dnode_of_data dn;
 575         struct f2fs_summary sum;
 576         struct node_info ni;
 577         struct page *page;
 578         block_t newaddr;
 579         int err;
 580
 581         /* do not read out */
 582         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 583         if (!page)
 584                 return;
 585
 586         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 587                 goto out;
 588
 589         if (f2fs_is_atomic_file(inode))
 590                 goto out;
 591
 592         set_new_dnode(&dn, inode, NULL, NULL, 0);
 593         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 594         if (err)
 595                 goto out;
 596
 597         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 598                 ClearPageUptodate(page);
 599                 goto put_out;
 600         }
 601
 602         /*
 603          * don't cache encrypted data into meta inode until previous dirty
 604          * data were writebacked to avoid racing between GC and flush.
 605          */
 606         f2fs_wait_on_page_writeback(page, DATA, true);
 607
 608         get_node_info(fio.sbi, dn.nid, &ni);
 609         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 610
 611         /* read page */
 612         fio.page = page;
 613         fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
 614
 615         allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
 616                                                         &sum, CURSEG_COLD_DATA);
 617
 618         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr,
 619                                         FGP_LOCK | FGP_CREAT, GFP_NOFS);
 620         if (!fio.encrypted_page) {
 621                 err = -ENOMEM;
 622                 goto recover_block;
 623         }
 624
 625         err = f2fs_submit_page_bio(&fio);
 626         if (err)
 627                 goto put_page_out;
 628
 629         /* write page */
 630         lock_page(fio.encrypted_page);
 631
 632         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
 633                 err = -EIO;
 634                 goto put_page_out;
 635         }
 636         if (unlikely(!PageUptodate(fio.encrypted_page))) {
 637                 err = -EIO;
 638                 goto put_page_out;
 639         }
 640
 641         set_page_dirty(fio.encrypted_page);
 642         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 643         if (clear_page_dirty_for_io(fio.encrypted_page))
 644                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 645
 646         set_page_writeback(fio.encrypted_page);
 647
 648         /* allocate block address */
 649         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 650
 651         fio.op = REQ_OP_WRITE;
 652         fio.op_flags = REQ_SYNC;
 653         fio.new_blkaddr = newaddr;
 654         f2fs_submit_page_mbio(&fio);
 655
 656         f2fs_update_data_blkaddr(&dn, newaddr);
 657         set_inode_flag(inode, FI_APPEND_WRITE);
 658         if (page->index == 0)
 659                 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
 660 put_page_out:
 661         f2fs_put_page(fio.encrypted_page, 1);
 662 recover_block:
 663         if (err)
 664                 __f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
 665                                                                 true, true);
 666 put_out:
 667         f2fs_put_dnode(&dn);
 668 out:
 669         f2fs_put_page(page, 1);
 670 }
 671
 672 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 673                                                         unsigned int segno, int off)
 674 {
 675         struct page *page;
 676
 677         page = get_lock_data_page(inode, bidx, true);
 678         if (IS_ERR(page))
 679                 return;
 680
 681         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 682                 goto out;
 683
 684         if (f2fs_is_atomic_file(inode))
 685                 goto out;
 686
 687         if (gc_type == BG_GC) {
 688                 if (PageWriteback(page))
 689                         goto out;
 690                 set_page_dirty(page);
 691                 set_cold_data(page);
 692         } else {
 693                 struct f2fs_io_info fio = {
 694                         .sbi = F2FS_I_SB(inode),
 695                         .type = DATA,
 696                         .op = REQ_OP_WRITE,
 697                         .op_flags = REQ_SYNC,
 698                         .page = page,
 699                         .encrypted_page = NULL,
 700                 };
 701                 bool is_dirty = PageDirty(page);
 702                 int err;
 703
 704 retry:
 705                 set_page_dirty(page);
 706                 f2fs_wait_on_page_writeback(page, DATA, true);
 707                 if (clear_page_dirty_for_io(page)) {
 708                         inode_dec_dirty_pages(inode);
 709                         remove_dirty_inode(inode);
 710                 }
 711
 712                 set_cold_data(page);
 713
 714                 err = do_write_data_page(&fio);
 715                 if (err == -ENOMEM && is_dirty) {
 716                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 717                         goto retry;
 718                 }
 719         }
 720 out:
 721         f2fs_put_page(page, 1);
 722 }
 723
 724 /*
 725  * This function tries to get parent node of victim data block, and identifies
 726  * data block validity. If the block is valid, copy that with cold status and
 727  * modify parent node.
 728  * If the parent node is not valid or the data block address is different,
 729  * the victim data block is ignored.
 730  */
 731 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 732                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 733 {
 734         struct super_block *sb = sbi->sb;
 735         struct f2fs_summary *entry;
 736         block_t start_addr;
 737         int off;
 738         int phase = 0;
 739
 740         start_addr = START_BLOCK(sbi, segno);
 741
 742 next_step:
 743         entry = sum;
 744
 745         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 746                 struct page *data_page;
 747                 struct inode *inode;
 748                 struct node_info dni; /* dnode info for the data */
 749                 unsigned int ofs_in_node, nofs;
 750                 block_t start_bidx;
 751                 nid_t nid = le32_to_cpu(entry->nid);
 752
 753                 /* stop BG_GC if there is not enough free sections. */
 754                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 755                         return;
 756
 757                 if (check_valid_map(sbi, segno, off) == 0)
 758                         continue;
 759
 760                 if (phase == 0) {
 761                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 762                                                         META_NAT, true);
 763                         continue;
 764                 }
 765
 766                 if (phase == 1) {
 767                         ra_node_page(sbi, nid);
 768                         continue;
 769                 }
 770
 771                 /* Get an inode by ino with checking validity */
 772                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 773                         continue;
 774
 775                 if (phase == 2) {
 776                         ra_node_page(sbi, dni.ino);
 777                         continue;
 778                 }
 779
 780                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 781
 782                 if (phase == 3) {
 783                         inode = f2fs_iget(sb, dni.ino);
 784                         if (IS_ERR(inode) || is_bad_inode(inode))
 785                                 continue;
 786
 787                         /* if encrypted inode, let's go phase 3 */
 788                         if (f2fs_encrypted_inode(inode) &&
 789                                                 S_ISREG(inode->i_mode)) {
 790                                 add_gc_inode(gc_list, inode);
 791                                 continue;
 792                         }
 793
 794                         start_bidx = start_bidx_of_node(nofs, inode);
 795                         data_page = get_read_data_page(inode,
 796                                         start_bidx + ofs_in_node, REQ_RAHEAD,
 797                                         true);
 798                         if (IS_ERR(data_page)) {
 799                                 iput(inode);
 800                                 continue;
 801                         }
 802
 803                         f2fs_put_page(data_page, 0);
 804                         add_gc_inode(gc_list, inode);
 805                         continue;
 806                 }
 807
 808                 /* phase 4 */
 809                 inode = find_gc_inode(gc_list, dni.ino);
 810                 if (inode) {
 811                         struct f2fs_inode_info *fi = F2FS_I(inode);
 812                         bool locked = false;
 813
 814                         if (S_ISREG(inode->i_mode)) {
 815                                 if (!down_write_trylock(&fi->dio_rwsem[READ]))
 816                                         continue;
 817                                 if (!down_write_trylock(
 818                                                 &fi->dio_rwsem[WRITE])) {
 819                                         up_write(&fi->dio_rwsem[READ]);
 820                                         continue;
 821                                 }
 822                                 locked = true;
 823                         }
 824
 825                         start_bidx = start_bidx_of_node(nofs, inode)
 826                                                                 + ofs_in_node;
 827                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 828                                 move_encrypted_block(inode, start_bidx, segno, off);
 829                         else
 830                                 move_data_page(inode, start_bidx, gc_type, segno, off);
 831
 832                         if (locked) {
 833                                 up_write(&fi->dio_rwsem[WRITE]);
 834                                 up_write(&fi->dio_rwsem[READ]);
 835                         }
 836
 837                         stat_inc_data_blk_count(sbi, 1, gc_type);
 838                 }
 839         }
 840
 841         if (++phase < 5)
 842                 goto next_step;
 843 }
 844
 845 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 846                         int gc_type)
 847 {
 848         struct sit_info *sit_i = SIT_I(sbi);
 849         int ret;
 850
 851         mutex_lock(&sit_i->sentry_lock);
 852         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 853                                               NO_CHECK_TYPE, LFS);
 854         mutex_unlock(&sit_i->sentry_lock);
 855         return ret;
 856 }
 857
 858 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 859                                 unsigned int start_segno,
 860                                 struct gc_inode_list *gc_list, int gc_type)
 861 {
 862         struct page *sum_page;
 863         struct f2fs_summary_block *sum;
 864         struct blk_plug plug;
 865         unsigned int segno = start_segno;
 866         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 867         int sec_freed = 0;
 868         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 869                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 870
 871         /* readahead multi ssa blocks those have contiguous address */
 872         if (sbi->segs_per_sec > 1)
 873                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 874                                         sbi->segs_per_sec, META_SSA, true);
 875
 876         /* reference all summary page */
 877         while (segno < end_segno) {
 878                 sum_page = get_sum_page(sbi, segno++);
 879                 unlock_page(sum_page);
 880         }
 881
 882         blk_start_plug(&plug);
 883
 884         for (segno = start_segno; segno < end_segno; segno++) {
 885
 886                 /* find segment summary of victim */
 887                 sum_page = find_get_page(META_MAPPING(sbi),
 888                                         GET_SUM_BLOCK(sbi, segno));
 889                 f2fs_put_page(sum_page, 0);
 890
 891                 if (get_valid_blocks(sbi, segno, 1) == 0 ||
 892                                 !PageUptodate(sum_page) ||
 893                                 unlikely(f2fs_cp_error(sbi)))
 894                         goto next;
 895
 896                 sum = page_address(sum_page);
 897                 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
 898
 899                 /*
 900                  * this is to avoid deadlock:
 901                  * - lock_page(sum_page)         - f2fs_replace_block
 902                  *  - check_valid_map()            - mutex_lock(sentry_lock)
 903                  *   - mutex_lock(sentry_lock)     - change_curseg()
 904                  *                                  - lock_page(sum_page)
 905                  */
 906
 907                 if (type == SUM_TYPE_NODE)
 908                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 909                 else
 910                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 911                                                                 gc_type);
 912
 913                 stat_inc_seg_count(sbi, type, gc_type);
 914 next:
 915                 f2fs_put_page(sum_page, 0);
 916         }
 917
 918         if (gc_type == FG_GC)
 919                 f2fs_submit_merged_bio(sbi,
 920                                 (type == SUM_TYPE_NODE) ? NODE : DATA, WRITE);
 921
 922         blk_finish_plug(&plug);
 923
 924         if (gc_type == FG_GC &&
 925                 get_valid_blocks(sbi, start_segno, sbi->segs_per_sec) == 0)
 926                 sec_freed = 1;
 927
 928         stat_inc_call_count(sbi->stat_info);
 929
 930         return sec_freed;
 931 }
 932
 933 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
 934 {
 935         unsigned int segno;
 936         int gc_type = sync ? FG_GC : BG_GC;
 937         int sec_freed = 0;
 938         int ret = -EINVAL;
 939         struct cp_control cpc;
 940         struct gc_inode_list gc_list = {
 941                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 942                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 943         };
 944
 945         cpc.reason = __get_cp_reason(sbi);
 946 gc_more:
 947         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 948                 goto stop;
 949         if (unlikely(f2fs_cp_error(sbi))) {
 950                 ret = -EIO;
 951                 goto stop;
 952         }
 953
 954         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed, 0)) {
 955                 gc_type = FG_GC;
 956                 /*
 957                  * If there is no victim and no prefree segment but still not
 958                  * enough free sections, we should flush dent/node blocks and do
 959                  * garbage collections.
 960                  */
 961                 ret = write_checkpoint(sbi, &cpc);
 962                 if (ret)
 963                         goto stop;
 964         } else if (gc_type == BG_GC && !background) {
 965                 /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
 966                 goto stop;
 967         }
 968
 969         if (!__get_victim(sbi, &segno, gc_type))
 970                 goto stop;
 971         ret = 0;
 972
 973         if (do_garbage_collect(sbi, segno, &gc_list, gc_type) &&
 974                         gc_type == FG_GC)
 975                 sec_freed++;
 976
 977         if (gc_type == FG_GC)
 978                 sbi->cur_victim_sec = NULL_SEGNO;
 979
 980         if (!sync) {
 981                 if (has_not_enough_free_secs(sbi, sec_freed, 0))
 982                         goto gc_more;
 983
 984                 if (gc_type == FG_GC)
 985                         ret = write_checkpoint(sbi, &cpc);
 986         }
 987 stop:
 988         mutex_unlock(&sbi->gc_mutex);
 989
 990         put_gc_inode(&gc_list);
 991
 992         if (sync)
 993                 ret = sec_freed ? 0 : -EAGAIN;
 994         return ret;
 995 }
 996
 997 void build_gc_manager(struct f2fs_sb_info *sbi)
 998 {
 999         u64 main_count, resv_count, ovp_count, blocks_per_sec;
1000
1001         DIRTY_I(sbi)->v_ops = &default_v_ops;
1002
1003         /* threshold of # of valid blocks in a section for victims of FG_GC */
1004         main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1005         resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1006         ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1007         blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
1008
1009         sbi->fggc_threshold = div_u64((main_count - ovp_count) * blocks_per_sec,
1010                                         (main_count - resv_count));
1011 }