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_show_injection_info(FAULT_CHECKPOINT);
  53                         f2fs_stop_checkpoint(sbi, false);
  54                 }
  55 #endif
  56
  57                 /*
  58                  * [GC triggering condition]
  59                  * 0. GC is not conducted currently.
  60                  * 1. There are enough dirty segments.
  61                  * 2. IO subsystem is idle by checking the # of writeback pages.
  62                  * 3. IO subsystem is idle by checking the # of requests in
  63                  *    bdev's request list.
  64                  *
  65                  * Note) We have to avoid triggering GCs frequently.
  66                  * Because it is possible that some segments can be
  67                  * invalidated soon after by user update or deletion.
  68                  * So, I'd like to wait some time to collect dirty segments.
  69                  */
  70                 if (!mutex_trylock(&sbi->gc_mutex))
  71                         continue;
  72
  73                 if (!is_idle(sbi)) {
  74                         increase_sleep_time(gc_th, &wait_ms);
  75                         mutex_unlock(&sbi->gc_mutex);
  76                         continue;
  77                 }
  78
  79                 if (has_enough_invalid_blocks(sbi))
  80                         decrease_sleep_time(gc_th, &wait_ms);
  81                 else
  82                         increase_sleep_time(gc_th, &wait_ms);
  83
  84                 stat_inc_bggc_count(sbi);
  85
  86                 /* if return value is not zero, no victim was selected */
  87                 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true))
  88                         wait_ms = gc_th->no_gc_sleep_time;
  89
  90                 trace_f2fs_background_gc(sbi->sb, wait_ms,
  91                                 prefree_segments(sbi), free_segments(sbi));
  92
  93                 /* balancing f2fs's metadata periodically */
  94                 f2fs_balance_fs_bg(sbi);
  95
  96         } while (!kthread_should_stop());
  97         return 0;
  98 }
  99
 100 int start_gc_thread(struct f2fs_sb_info *sbi)
 101 {
 102         struct f2fs_gc_kthread *gc_th;
 103         dev_t dev = sbi->sb->s_bdev->bd_dev;
 104         int err = 0;
 105
 106         gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 107         if (!gc_th) {
 108                 err = -ENOMEM;
 109                 goto out;
 110         }
 111
 112         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 113         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 114         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 115
 116         gc_th->gc_idle = 0;
 117
 118         sbi->gc_thread = gc_th;
 119         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 120         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 121                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 122         if (IS_ERR(gc_th->f2fs_gc_task)) {
 123                 err = PTR_ERR(gc_th->f2fs_gc_task);
 124                 kfree(gc_th);
 125                 sbi->gc_thread = NULL;
 126         }
 127 out:
 128         return err;
 129 }
 130
 131 void stop_gc_thread(struct f2fs_sb_info *sbi)
 132 {
 133         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 134         if (!gc_th)
 135                 return;
 136         kthread_stop(gc_th->f2fs_gc_task);
 137         kfree(gc_th);
 138         sbi->gc_thread = NULL;
 139 }
 140
 141 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 142 {
 143         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 144
 145         if (gc_th && gc_th->gc_idle) {
 146                 if (gc_th->gc_idle == 1)
 147                         gc_mode = GC_CB;
 148                 else if (gc_th->gc_idle == 2)
 149                         gc_mode = GC_GREEDY;
 150         }
 151         return gc_mode;
 152 }
 153
 154 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 155                         int type, struct victim_sel_policy *p)
 156 {
 157         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 158
 159         if (p->alloc_mode == SSR) {
 160                 p->gc_mode = GC_GREEDY;
 161                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 162                 p->max_search = dirty_i->nr_dirty[type];
 163                 p->ofs_unit = 1;
 164         } else {
 165                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 166                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 167                 p->max_search = dirty_i->nr_dirty[DIRTY];
 168                 p->ofs_unit = sbi->segs_per_sec;
 169         }
 170
 171         /* we need to check every dirty segments in the FG_GC case */
 172         if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
 173                 p->max_search = sbi->max_victim_search;
 174
 175         p->offset = sbi->last_victim[p->gc_mode];
 176 }
 177
 178 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 179                                 struct victim_sel_policy *p)
 180 {
 181         /* SSR allocates in a segment unit */
 182         if (p->alloc_mode == SSR)
 183                 return sbi->blocks_per_seg;
 184         if (p->gc_mode == GC_GREEDY)
 185                 return sbi->blocks_per_seg * p->ofs_unit;
 186         else if (p->gc_mode == GC_CB)
 187                 return UINT_MAX;
 188         else /* No other gc_mode */
 189                 return 0;
 190 }
 191
 192 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 193 {
 194         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 195         unsigned int secno;
 196
 197         /*
 198          * If the gc_type is FG_GC, we can select victim segments
 199          * selected by background GC before.
 200          * Those segments guarantee they have small valid blocks.
 201          */
 202         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 203                 if (sec_usage_check(sbi, secno))
 204                         continue;
 205
 206                 if (no_fggc_candidate(sbi, secno))
 207                         continue;
 208
 209                 clear_bit(secno, dirty_i->victim_secmap);
 210                 return secno * sbi->segs_per_sec;
 211         }
 212         return NULL_SEGNO;
 213 }
 214
 215 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 216 {
 217         struct sit_info *sit_i = SIT_I(sbi);
 218         unsigned int secno = GET_SECNO(sbi, segno);
 219         unsigned int start = secno * sbi->segs_per_sec;
 220         unsigned long long mtime = 0;
 221         unsigned int vblocks;
 222         unsigned char age = 0;
 223         unsigned char u;
 224         unsigned int i;
 225
 226         for (i = 0; i < sbi->segs_per_sec; i++)
 227                 mtime += get_seg_entry(sbi, start + i)->mtime;
 228         vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 229
 230         mtime = div_u64(mtime, sbi->segs_per_sec);
 231         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 232
 233         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 234
 235         /* Handle if the system time has changed by the user */
 236         if (mtime < sit_i->min_mtime)
 237                 sit_i->min_mtime = mtime;
 238         if (mtime > sit_i->max_mtime)
 239                 sit_i->max_mtime = mtime;
 240         if (sit_i->max_mtime != sit_i->min_mtime)
 241                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 242                                 sit_i->max_mtime - sit_i->min_mtime);
 243
 244         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 245 }
 246
 247 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
 248                                                 unsigned int segno)
 249 {
 250         unsigned int valid_blocks =
 251                         get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 252
 253         return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 254                                 valid_blocks * 2 : valid_blocks;
 255 }
 256
 257 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 258                         unsigned int segno, struct victim_sel_policy *p)
 259 {
 260         if (p->alloc_mode == SSR)
 261                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 262
 263         /* alloc_mode == LFS */
 264         if (p->gc_mode == GC_GREEDY)
 265                 return get_greedy_cost(sbi, segno);
 266         else
 267                 return get_cb_cost(sbi, segno);
 268 }
 269
 270 static unsigned int count_bits(const unsigned long *addr,
 271                                 unsigned int offset, unsigned int len)
 272 {
 273         unsigned int end = offset + len, sum = 0;
 274
 275         while (offset < end) {
 276                 if (test_bit(offset++, addr))
 277                         ++sum;
 278         }
 279         return sum;
 280 }
 281
 282 /*
 283  * This function is called from two paths.
 284  * One is garbage collection and the other is SSR segment selection.
 285  * When it is called during GC, it just gets a victim segment
 286  * and it does not remove it from dirty seglist.
 287  * When it is called from SSR segment selection, it finds a segment
 288  * which has minimum valid blocks and removes it from dirty seglist.
 289  */
 290 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 291                 unsigned int *result, int gc_type, int type, char alloc_mode)
 292 {
 293         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 294         struct victim_sel_policy p;
 295         unsigned int secno, last_victim;
 296         unsigned int last_segment = MAIN_SEGS(sbi);
 297         unsigned int nsearched = 0;
 298
 299         mutex_lock(&dirty_i->seglist_lock);
 300
 301         p.alloc_mode = alloc_mode;
 302         select_policy(sbi, gc_type, type, &p);
 303
 304         p.min_segno = NULL_SEGNO;
 305         p.min_cost = get_max_cost(sbi, &p);
 306
 307         if (p.max_search == 0)
 308                 goto out;
 309
 310         last_victim = sbi->last_victim[p.gc_mode];
 311         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 312                 p.min_segno = check_bg_victims(sbi);
 313                 if (p.min_segno != NULL_SEGNO)
 314                         goto got_it;
 315         }
 316
 317         while (1) {
 318                 unsigned long cost;
 319                 unsigned int segno;
 320
 321                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 322                 if (segno >= last_segment) {
 323                         if (sbi->last_victim[p.gc_mode]) {
 324                                 last_segment = sbi->last_victim[p.gc_mode];
 325                                 sbi->last_victim[p.gc_mode] = 0;
 326                                 p.offset = 0;
 327                                 continue;
 328                         }
 329                         break;
 330                 }
 331
 332                 p.offset = segno + p.ofs_unit;
 333                 if (p.ofs_unit > 1) {
 334                         p.offset -= segno % p.ofs_unit;
 335                         nsearched += count_bits(p.dirty_segmap,
 336                                                 p.offset - p.ofs_unit,
 337                                                 p.ofs_unit);
 338                 } else {
 339                         nsearched++;
 340                 }
 341
 342                 secno = GET_SECNO(sbi, segno);
 343
 344                 if (sec_usage_check(sbi, secno))
 345                         goto next;
 346                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 347                         goto next;
 348                 if (gc_type == FG_GC && p.alloc_mode == LFS &&
 349                                         no_fggc_candidate(sbi, secno))
 350                         goto next;
 351
 352                 cost = get_gc_cost(sbi, segno, &p);
 353
 354                 if (p.min_cost > cost) {
 355                         p.min_segno = segno;
 356                         p.min_cost = cost;
 357                 }
 358 next:
 359                 if (nsearched >= p.max_search) {
 360                         if (!sbi->last_victim[p.gc_mode] && segno <= last_victim)
 361                                 sbi->last_victim[p.gc_mode] = last_victim + 1;
 362                         else
 363                                 sbi->last_victim[p.gc_mode] = segno + 1;
 364                         sbi->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
 365                         break;
 366                 }
 367         }
 368         if (p.min_segno != NULL_SEGNO) {
 369 got_it:
 370                 if (p.alloc_mode == LFS) {
 371                         secno = GET_SECNO(sbi, p.min_segno);
 372                         if (gc_type == FG_GC)
 373                                 sbi->cur_victim_sec = secno;
 374                         else
 375                                 set_bit(secno, dirty_i->victim_secmap);
 376                 }
 377                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 378
 379                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 380                                 sbi->cur_victim_sec,
 381                                 prefree_segments(sbi), free_segments(sbi));
 382         }
 383 out:
 384         mutex_unlock(&dirty_i->seglist_lock);
 385
 386         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 387 }
 388
 389 static const struct victim_selection default_v_ops = {
 390         .get_victim = get_victim_by_default,
 391 };
 392
 393 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 394 {
 395         struct inode_entry *ie;
 396
 397         ie = radix_tree_lookup(&gc_list->iroot, ino);
 398         if (ie)
 399                 return ie->inode;
 400         return NULL;
 401 }
 402
 403 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 404 {
 405         struct inode_entry *new_ie;
 406
 407         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 408                 iput(inode);
 409                 return;
 410         }
 411         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 412         new_ie->inode = inode;
 413
 414         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 415         list_add_tail(&new_ie->list, &gc_list->ilist);
 416 }
 417
 418 static void put_gc_inode(struct gc_inode_list *gc_list)
 419 {
 420         struct inode_entry *ie, *next_ie;
 421         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 422                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 423                 iput(ie->inode);
 424                 list_del(&ie->list);
 425                 kmem_cache_free(inode_entry_slab, ie);
 426         }
 427 }
 428
 429 static int check_valid_map(struct f2fs_sb_info *sbi,
 430                                 unsigned int segno, int offset)
 431 {
 432         struct sit_info *sit_i = SIT_I(sbi);
 433         struct seg_entry *sentry;
 434         int ret;
 435
 436         mutex_lock(&sit_i->sentry_lock);
 437         sentry = get_seg_entry(sbi, segno);
 438         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 439         mutex_unlock(&sit_i->sentry_lock);
 440         return ret;
 441 }
 442
 443 /*
 444  * This function compares node address got in summary with that in NAT.
 445  * On validity, copy that node with cold status, otherwise (invalid node)
 446  * ignore that.
 447  */
 448 static void gc_node_segment(struct f2fs_sb_info *sbi,
 449                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 450 {
 451         struct f2fs_summary *entry;
 452         block_t start_addr;
 453         int off;
 454         int phase = 0;
 455
 456         start_addr = START_BLOCK(sbi, segno);
 457
 458 next_step:
 459         entry = sum;
 460
 461         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 462                 nid_t nid = le32_to_cpu(entry->nid);
 463                 struct page *node_page;
 464                 struct node_info ni;
 465
 466                 /* stop BG_GC if there is not enough free sections. */
 467                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 468                         return;
 469
 470                 if (check_valid_map(sbi, segno, off) == 0)
 471                         continue;
 472
 473                 if (phase == 0) {
 474                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 475                                                         META_NAT, true);
 476                         continue;
 477                 }
 478
 479                 if (phase == 1) {
 480                         ra_node_page(sbi, nid);
 481                         continue;
 482                 }
 483
 484                 /* phase == 2 */
 485                 node_page = get_node_page(sbi, nid);
 486                 if (IS_ERR(node_page))
 487                         continue;
 488
 489                 /* block may become invalid during get_node_page */
 490                 if (check_valid_map(sbi, segno, off) == 0) {
 491                         f2fs_put_page(node_page, 1);
 492                         continue;
 493                 }
 494
 495                 get_node_info(sbi, nid, &ni);
 496                 if (ni.blk_addr != start_addr + off) {
 497                         f2fs_put_page(node_page, 1);
 498                         continue;
 499                 }
 500
 501                 move_node_page(node_page, gc_type);
 502                 stat_inc_node_blk_count(sbi, 1, gc_type);
 503         }
 504
 505         if (++phase < 3)
 506                 goto next_step;
 507 }
 508
 509 /*
 510  * Calculate start block index indicating the given node offset.
 511  * Be careful, caller should give this node offset only indicating direct node
 512  * blocks. If any node offsets, which point the other types of node blocks such
 513  * as indirect or double indirect node blocks, are given, it must be a caller's
 514  * bug.
 515  */
 516 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 517 {
 518         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 519         unsigned int bidx;
 520
 521         if (node_ofs == 0)
 522                 return 0;
 523
 524         if (node_ofs <= 2) {
 525                 bidx = node_ofs - 1;
 526         } else if (node_ofs <= indirect_blks) {
 527                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 528                 bidx = node_ofs - 2 - dec;
 529         } else {
 530                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 531                 bidx = node_ofs - 5 - dec;
 532         }
 533         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 534 }
 535
 536 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 537                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 538 {
 539         struct page *node_page;
 540         nid_t nid;
 541         unsigned int ofs_in_node;
 542         block_t source_blkaddr;
 543
 544         nid = le32_to_cpu(sum->nid);
 545         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 546
 547         node_page = get_node_page(sbi, nid);
 548         if (IS_ERR(node_page))
 549                 return false;
 550
 551         get_node_info(sbi, nid, dni);
 552
 553         if (sum->version != dni->version) {
 554                 f2fs_put_page(node_page, 1);
 555                 return false;
 556         }
 557
 558         *nofs = ofs_of_node(node_page);
 559         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 560         f2fs_put_page(node_page, 1);
 561
 562         if (source_blkaddr != blkaddr)
 563                 return false;
 564         return true;
 565 }
 566
 567 static void move_encrypted_block(struct inode *inode, block_t bidx,
 568                                                         unsigned int segno, int off)
 569 {
 570         struct f2fs_io_info fio = {
 571                 .sbi = F2FS_I_SB(inode),
 572                 .type = DATA,
 573                 .op = REQ_OP_READ,
 574                 .op_flags = 0,
 575                 .encrypted_page = NULL,
 576         };
 577         struct dnode_of_data dn;
 578         struct f2fs_summary sum;
 579         struct node_info ni;
 580         struct page *page;
 581         block_t newaddr;
 582         int err;
 583
 584         /* do not read out */
 585         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 586         if (!page)
 587                 return;
 588
 589         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 590                 goto out;
 591
 592         if (f2fs_is_atomic_file(inode))
 593                 goto out;
 594
 595         set_new_dnode(&dn, inode, NULL, NULL, 0);
 596         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 597         if (err)
 598                 goto out;
 599
 600         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 601                 ClearPageUptodate(page);
 602                 goto put_out;
 603         }
 604
 605         /*
 606          * don't cache encrypted data into meta inode until previous dirty
 607          * data were writebacked to avoid racing between GC and flush.
 608          */
 609         f2fs_wait_on_page_writeback(page, DATA, true);
 610
 611         get_node_info(fio.sbi, dn.nid, &ni);
 612         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 613
 614         /* read page */
 615         fio.page = page;
 616         fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
 617
 618         allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
 619                                                         &sum, CURSEG_COLD_DATA);
 620
 621         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr,
 622                                         FGP_LOCK | FGP_CREAT, GFP_NOFS);
 623         if (!fio.encrypted_page) {
 624                 err = -ENOMEM;
 625                 goto recover_block;
 626         }
 627
 628         err = f2fs_submit_page_bio(&fio);
 629         if (err)
 630                 goto put_page_out;
 631
 632         /* write page */
 633         lock_page(fio.encrypted_page);
 634
 635         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
 636                 err = -EIO;
 637                 goto put_page_out;
 638         }
 639         if (unlikely(!PageUptodate(fio.encrypted_page))) {
 640                 err = -EIO;
 641                 goto put_page_out;
 642         }
 643
 644         set_page_dirty(fio.encrypted_page);
 645         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 646         if (clear_page_dirty_for_io(fio.encrypted_page))
 647                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 648
 649         set_page_writeback(fio.encrypted_page);
 650
 651         /* allocate block address */
 652         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 653
 654         fio.op = REQ_OP_WRITE;
 655         fio.op_flags = REQ_SYNC;
 656         fio.new_blkaddr = newaddr;
 657         f2fs_submit_page_mbio(&fio);
 658
 659         f2fs_update_data_blkaddr(&dn, newaddr);
 660         set_inode_flag(inode, FI_APPEND_WRITE);
 661         if (page->index == 0)
 662                 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
 663 put_page_out:
 664         f2fs_put_page(fio.encrypted_page, 1);
 665 recover_block:
 666         if (err)
 667                 __f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
 668                                                                 true, true);
 669 put_out:
 670         f2fs_put_dnode(&dn);
 671 out:
 672         f2fs_put_page(page, 1);
 673 }
 674
 675 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 676                                                         unsigned int segno, int off)
 677 {
 678         struct page *page;
 679
 680         page = get_lock_data_page(inode, bidx, true);
 681         if (IS_ERR(page))
 682                 return;
 683
 684         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 685                 goto out;
 686
 687         if (f2fs_is_atomic_file(inode))
 688                 goto out;
 689
 690         if (gc_type == BG_GC) {
 691                 if (PageWriteback(page))
 692                         goto out;
 693                 set_page_dirty(page);
 694                 set_cold_data(page);
 695         } else {
 696                 struct f2fs_io_info fio = {
 697                         .sbi = F2FS_I_SB(inode),
 698                         .type = DATA,
 699                         .op = REQ_OP_WRITE,
 700                         .op_flags = REQ_SYNC,
 701                         .page = page,
 702                         .encrypted_page = NULL,
 703                 };
 704                 bool is_dirty = PageDirty(page);
 705                 int err;
 706
 707 retry:
 708                 set_page_dirty(page);
 709                 f2fs_wait_on_page_writeback(page, DATA, true);
 710                 if (clear_page_dirty_for_io(page)) {
 711                         inode_dec_dirty_pages(inode);
 712                         remove_dirty_inode(inode);
 713                 }
 714
 715                 set_cold_data(page);
 716
 717                 err = do_write_data_page(&fio);
 718                 if (err == -ENOMEM && is_dirty) {
 719                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 720                         goto retry;
 721                 }
 722         }
 723 out:
 724         f2fs_put_page(page, 1);
 725 }
 726
 727 /*
 728  * This function tries to get parent node of victim data block, and identifies
 729  * data block validity. If the block is valid, copy that with cold status and
 730  * modify parent node.
 731  * If the parent node is not valid or the data block address is different,
 732  * the victim data block is ignored.
 733  */
 734 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 735                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 736 {
 737         struct super_block *sb = sbi->sb;
 738         struct f2fs_summary *entry;
 739         block_t start_addr;
 740         int off;
 741         int phase = 0;
 742
 743         start_addr = START_BLOCK(sbi, segno);
 744
 745 next_step:
 746         entry = sum;
 747
 748         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 749                 struct page *data_page;
 750                 struct inode *inode;
 751                 struct node_info dni; /* dnode info for the data */
 752                 unsigned int ofs_in_node, nofs;
 753                 block_t start_bidx;
 754                 nid_t nid = le32_to_cpu(entry->nid);
 755
 756                 /* stop BG_GC if there is not enough free sections. */
 757                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 758                         return;
 759
 760                 if (check_valid_map(sbi, segno, off) == 0)
 761                         continue;
 762
 763                 if (phase == 0) {
 764                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 765                                                         META_NAT, true);
 766                         continue;
 767                 }
 768
 769                 if (phase == 1) {
 770                         ra_node_page(sbi, nid);
 771                         continue;
 772                 }
 773
 774                 /* Get an inode by ino with checking validity */
 775                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 776                         continue;
 777
 778                 if (phase == 2) {
 779                         ra_node_page(sbi, dni.ino);
 780                         continue;
 781                 }
 782
 783                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 784
 785                 if (phase == 3) {
 786                         inode = f2fs_iget(sb, dni.ino);
 787                         if (IS_ERR(inode) || is_bad_inode(inode))
 788                                 continue;
 789
 790                         /* if encrypted inode, let's go phase 3 */
 791                         if (f2fs_encrypted_inode(inode) &&
 792                                                 S_ISREG(inode->i_mode)) {
 793                                 add_gc_inode(gc_list, inode);
 794                                 continue;
 795                         }
 796
 797                         start_bidx = start_bidx_of_node(nofs, inode);
 798                         data_page = get_read_data_page(inode,
 799                                         start_bidx + ofs_in_node, REQ_RAHEAD,
 800                                         true);
 801                         if (IS_ERR(data_page)) {
 802                                 iput(inode);
 803                                 continue;
 804                         }
 805
 806                         f2fs_put_page(data_page, 0);
 807                         add_gc_inode(gc_list, inode);
 808                         continue;
 809                 }
 810
 811                 /* phase 4 */
 812                 inode = find_gc_inode(gc_list, dni.ino);
 813                 if (inode) {
 814                         struct f2fs_inode_info *fi = F2FS_I(inode);
 815                         bool locked = false;
 816
 817                         if (S_ISREG(inode->i_mode)) {
 818                                 if (!down_write_trylock(&fi->dio_rwsem[READ]))
 819                                         continue;
 820                                 if (!down_write_trylock(
 821                                                 &fi->dio_rwsem[WRITE])) {
 822                                         up_write(&fi->dio_rwsem[READ]);
 823                                         continue;
 824                                 }
 825                                 locked = true;
 826                         }
 827
 828                         start_bidx = start_bidx_of_node(nofs, inode)
 829                                                                 + ofs_in_node;
 830                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 831                                 move_encrypted_block(inode, start_bidx, segno, off);
 832                         else
 833                                 move_data_page(inode, start_bidx, gc_type, segno, off);
 834
 835                         if (locked) {
 836                                 up_write(&fi->dio_rwsem[WRITE]);
 837                                 up_write(&fi->dio_rwsem[READ]);
 838                         }
 839
 840                         stat_inc_data_blk_count(sbi, 1, gc_type);
 841                 }
 842         }
 843
 844         if (++phase < 5)
 845                 goto next_step;
 846 }
 847
 848 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 849                         int gc_type)
 850 {
 851         struct sit_info *sit_i = SIT_I(sbi);
 852         int ret;
 853
 854         mutex_lock(&sit_i->sentry_lock);
 855         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 856                                               NO_CHECK_TYPE, LFS);
 857         mutex_unlock(&sit_i->sentry_lock);
 858         return ret;
 859 }
 860
 861 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 862                                 unsigned int start_segno,
 863                                 struct gc_inode_list *gc_list, int gc_type)
 864 {
 865         struct page *sum_page;
 866         struct f2fs_summary_block *sum;
 867         struct blk_plug plug;
 868         unsigned int segno = start_segno;
 869         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 870         int sec_freed = 0;
 871         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 872                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 873
 874         /* readahead multi ssa blocks those have contiguous address */
 875         if (sbi->segs_per_sec > 1)
 876                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 877                                         sbi->segs_per_sec, META_SSA, true);
 878
 879         /* reference all summary page */
 880         while (segno < end_segno) {
 881                 sum_page = get_sum_page(sbi, segno++);
 882                 unlock_page(sum_page);
 883         }
 884
 885         blk_start_plug(&plug);
 886
 887         for (segno = start_segno; segno < end_segno; segno++) {
 888
 889                 /* find segment summary of victim */
 890                 sum_page = find_get_page(META_MAPPING(sbi),
 891                                         GET_SUM_BLOCK(sbi, segno));
 892                 f2fs_put_page(sum_page, 0);
 893
 894                 if (get_valid_blocks(sbi, segno, 1) == 0 ||
 895                                 !PageUptodate(sum_page) ||
 896                                 unlikely(f2fs_cp_error(sbi)))
 897                         goto next;
 898
 899                 sum = page_address(sum_page);
 900                 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
 901
 902                 /*
 903                  * this is to avoid deadlock:
 904                  * - lock_page(sum_page)         - f2fs_replace_block
 905                  *  - check_valid_map()            - mutex_lock(sentry_lock)
 906                  *   - mutex_lock(sentry_lock)     - change_curseg()
 907                  *                                  - lock_page(sum_page)
 908                  */
 909
 910                 if (type == SUM_TYPE_NODE)
 911                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 912                 else
 913                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 914                                                                 gc_type);
 915
 916                 stat_inc_seg_count(sbi, type, gc_type);
 917 next:
 918                 f2fs_put_page(sum_page, 0);
 919         }
 920
 921         if (gc_type == FG_GC)
 922                 f2fs_submit_merged_bio(sbi,
 923                                 (type == SUM_TYPE_NODE) ? NODE : DATA, WRITE);
 924
 925         blk_finish_plug(&plug);
 926
 927         if (gc_type == FG_GC &&
 928                 get_valid_blocks(sbi, start_segno, sbi->segs_per_sec) == 0)
 929                 sec_freed = 1;
 930
 931         stat_inc_call_count(sbi->stat_info);
 932
 933         return sec_freed;
 934 }
 935
 936 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
 937 {
 938         unsigned int segno;
 939         int gc_type = sync ? FG_GC : BG_GC;
 940         int sec_freed = 0;
 941         int ret = -EINVAL;
 942         struct cp_control cpc;
 943         struct gc_inode_list gc_list = {
 944                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 945                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 946         };
 947
 948         cpc.reason = __get_cp_reason(sbi);
 949 gc_more:
 950         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 951                 goto stop;
 952         if (unlikely(f2fs_cp_error(sbi))) {
 953                 ret = -EIO;
 954                 goto stop;
 955         }
 956
 957         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
 958                 /*
 959                  * For example, if there are many prefree_segments below given
 960                  * threshold, we can make them free by checkpoint. Then, we
 961                  * secure free segments which doesn't need fggc any more.
 962                  */
 963                 ret = write_checkpoint(sbi, &cpc);
 964                 if (ret)
 965                         goto stop;
 966                 if (has_not_enough_free_secs(sbi, 0, 0))
 967                         gc_type = FG_GC;
 968         }
 969
 970         /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
 971         if (gc_type == BG_GC && !background)
 972                 goto stop;
 973         if (!__get_victim(sbi, &segno, gc_type))
 974                 goto stop;
 975         ret = 0;
 976
 977         if (do_garbage_collect(sbi, segno, &gc_list, gc_type) &&
 978                         gc_type == FG_GC)
 979                 sec_freed++;
 980
 981         if (gc_type == FG_GC)
 982                 sbi->cur_victim_sec = NULL_SEGNO;
 983
 984         if (!sync) {
 985                 if (has_not_enough_free_secs(sbi, sec_freed, 0))
 986                         goto gc_more;
 987
 988                 if (gc_type == FG_GC)
 989                         ret = write_checkpoint(sbi, &cpc);
 990         }
 991 stop:
 992         mutex_unlock(&sbi->gc_mutex);
 993
 994         put_gc_inode(&gc_list);
 995
 996         if (sync)
 997                 ret = sec_freed ? 0 : -EAGAIN;
 998         return ret;
 999 }
1000
1001 void build_gc_manager(struct f2fs_sb_info *sbi)
1002 {
1003         u64 main_count, resv_count, ovp_count, blocks_per_sec;
1004
1005         DIRTY_I(sbi)->v_ops = &default_v_ops;
1006
1007         /* threshold of # of valid blocks in a section for victims of FG_GC */
1008         main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1009         resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1010         ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1011         blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
1012
1013         sbi->fggc_threshold = div64_u64((main_count - ovp_count) * blocks_per_sec,
1014                                         (main_count - resv_count));
1015 }