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         unsigned int wait_ms;
  32
  33         wait_ms = gc_th->min_sleep_time;
  34
  35         set_freezable();
  36         do {
  37                 wait_event_interruptible_timeout(*wq,
  38                                 kthread_should_stop() || freezing(current) ||
  39                                 gc_th->gc_wake,
  40                                 msecs_to_jiffies(wait_ms));
  41
  42                 /* give it a try one time */
  43                 if (gc_th->gc_wake)
  44                         gc_th->gc_wake = 0;
  45
  46                 if (try_to_freeze())
  47                         continue;
  48                 if (kthread_should_stop())
  49                         break;
  50
  51                 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
  52                         increase_sleep_time(gc_th, &wait_ms);
  53                         continue;
  54                 }
  55
  56 #ifdef CONFIG_F2FS_FAULT_INJECTION
  57                 if (time_to_inject(sbi, FAULT_CHECKPOINT)) {
  58                         f2fs_show_injection_info(FAULT_CHECKPOINT);
  59                         f2fs_stop_checkpoint(sbi, false);
  60                 }
  61 #endif
  62
  63                 if (!sb_start_write_trylock(sbi->sb))
  64                         continue;
  65
  66                 /*
  67                  * [GC triggering condition]
  68                  * 0. GC is not conducted currently.
  69                  * 1. There are enough dirty segments.
  70                  * 2. IO subsystem is idle by checking the # of writeback pages.
  71                  * 3. IO subsystem is idle by checking the # of requests in
  72                  *    bdev's request list.
  73                  *
  74                  * Note) We have to avoid triggering GCs frequently.
  75                  * Because it is possible that some segments can be
  76                  * invalidated soon after by user update or deletion.
  77                  * So, I'd like to wait some time to collect dirty segments.
  78                  */
  79                 if (!mutex_trylock(&sbi->gc_mutex))
  80                         goto next;
  81
  82                 if (gc_th->gc_urgent) {
  83                         wait_ms = gc_th->urgent_sleep_time;
  84                         goto do_gc;
  85                 }
  86
  87                 if (!is_idle(sbi)) {
  88                         increase_sleep_time(gc_th, &wait_ms);
  89                         mutex_unlock(&sbi->gc_mutex);
  90                         goto next;
  91                 }
  92
  93                 if (has_enough_invalid_blocks(sbi))
  94                         decrease_sleep_time(gc_th, &wait_ms);
  95                 else
  96                         increase_sleep_time(gc_th, &wait_ms);
  97 do_gc:
  98                 stat_inc_bggc_count(sbi);
  99
 100                 /* if return value is not zero, no victim was selected */
 101                 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC), true, NULL_SEGNO))
 102                         wait_ms = gc_th->no_gc_sleep_time;
 103
 104                 trace_f2fs_background_gc(sbi->sb, wait_ms,
 105                                 prefree_segments(sbi), free_segments(sbi));
 106
 107                 /* balancing f2fs's metadata periodically */
 108                 f2fs_balance_fs_bg(sbi);
 109 next:
 110                 sb_end_write(sbi->sb);
 111
 112         } while (!kthread_should_stop());
 113         return 0;
 114 }
 115
 116 int start_gc_thread(struct f2fs_sb_info *sbi)
 117 {
 118         struct f2fs_gc_kthread *gc_th;
 119         dev_t dev = sbi->sb->s_bdev->bd_dev;
 120         int err = 0;
 121
 122         gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 123         if (!gc_th) {
 124                 err = -ENOMEM;
 125                 goto out;
 126         }
 127
 128         gc_th->urgent_sleep_time = DEF_GC_THREAD_URGENT_SLEEP_TIME;
 129         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 130         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 131         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 132
 133         gc_th->gc_idle = 0;
 134         gc_th->gc_urgent = 0;
 135         gc_th->gc_wake= 0;
 136
 137         sbi->gc_thread = gc_th;
 138         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 139         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 140                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 141         if (IS_ERR(gc_th->f2fs_gc_task)) {
 142                 err = PTR_ERR(gc_th->f2fs_gc_task);
 143                 kfree(gc_th);
 144                 sbi->gc_thread = NULL;
 145         }
 146 out:
 147         return err;
 148 }
 149
 150 void stop_gc_thread(struct f2fs_sb_info *sbi)
 151 {
 152         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 153         if (!gc_th)
 154                 return;
 155         kthread_stop(gc_th->f2fs_gc_task);
 156         kfree(gc_th);
 157         sbi->gc_thread = NULL;
 158 }
 159
 160 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 161 {
 162         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 163
 164         if (gc_th && gc_th->gc_idle) {
 165                 if (gc_th->gc_idle == 1)
 166                         gc_mode = GC_CB;
 167                 else if (gc_th->gc_idle == 2)
 168                         gc_mode = GC_GREEDY;
 169         }
 170         return gc_mode;
 171 }
 172
 173 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 174                         int type, struct victim_sel_policy *p)
 175 {
 176         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 177
 178         if (p->alloc_mode == SSR) {
 179                 p->gc_mode = GC_GREEDY;
 180                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 181                 p->max_search = dirty_i->nr_dirty[type];
 182                 p->ofs_unit = 1;
 183         } else {
 184                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 185                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 186                 p->max_search = dirty_i->nr_dirty[DIRTY];
 187                 p->ofs_unit = sbi->segs_per_sec;
 188         }
 189
 190         /* we need to check every dirty segments in the FG_GC case */
 191         if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
 192                 p->max_search = sbi->max_victim_search;
 193
 194         /* let's select beginning hot/small space first */
 195         if (type == CURSEG_HOT_DATA || IS_NODESEG(type))
 196                 p->offset = 0;
 197         else
 198                 p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
 199 }
 200
 201 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 202                                 struct victim_sel_policy *p)
 203 {
 204         /* SSR allocates in a segment unit */
 205         if (p->alloc_mode == SSR)
 206                 return sbi->blocks_per_seg;
 207         if (p->gc_mode == GC_GREEDY)
 208                 return 2 * sbi->blocks_per_seg * p->ofs_unit;
 209         else if (p->gc_mode == GC_CB)
 210                 return UINT_MAX;
 211         else /* No other gc_mode */
 212                 return 0;
 213 }
 214
 215 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 216 {
 217         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 218         unsigned int secno;
 219
 220         /*
 221          * If the gc_type is FG_GC, we can select victim segments
 222          * selected by background GC before.
 223          * Those segments guarantee they have small valid blocks.
 224          */
 225         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 226                 if (sec_usage_check(sbi, secno))
 227                         continue;
 228
 229                 if (no_fggc_candidate(sbi, secno))
 230                         continue;
 231
 232                 clear_bit(secno, dirty_i->victim_secmap);
 233                 return GET_SEG_FROM_SEC(sbi, secno);
 234         }
 235         return NULL_SEGNO;
 236 }
 237
 238 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 239 {
 240         struct sit_info *sit_i = SIT_I(sbi);
 241         unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
 242         unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
 243         unsigned long long mtime = 0;
 244         unsigned int vblocks;
 245         unsigned char age = 0;
 246         unsigned char u;
 247         unsigned int i;
 248
 249         for (i = 0; i < sbi->segs_per_sec; i++)
 250                 mtime += get_seg_entry(sbi, start + i)->mtime;
 251         vblocks = get_valid_blocks(sbi, segno, true);
 252
 253         mtime = div_u64(mtime, sbi->segs_per_sec);
 254         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 255
 256         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 257
 258         /* Handle if the system time has changed by the user */
 259         if (mtime < sit_i->min_mtime)
 260                 sit_i->min_mtime = mtime;
 261         if (mtime > sit_i->max_mtime)
 262                 sit_i->max_mtime = mtime;
 263         if (sit_i->max_mtime != sit_i->min_mtime)
 264                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 265                                 sit_i->max_mtime - sit_i->min_mtime);
 266
 267         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 268 }
 269
 270 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 271                         unsigned int segno, struct victim_sel_policy *p)
 272 {
 273         if (p->alloc_mode == SSR)
 274                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 275
 276         /* alloc_mode == LFS */
 277         if (p->gc_mode == GC_GREEDY)
 278                 return get_valid_blocks(sbi, segno, true);
 279         else
 280                 return get_cb_cost(sbi, segno);
 281 }
 282
 283 static unsigned int count_bits(const unsigned long *addr,
 284                                 unsigned int offset, unsigned int len)
 285 {
 286         unsigned int end = offset + len, sum = 0;
 287
 288         while (offset < end) {
 289                 if (test_bit(offset++, addr))
 290                         ++sum;
 291         }
 292         return sum;
 293 }
 294
 295 /*
 296  * This function is called from two paths.
 297  * One is garbage collection and the other is SSR segment selection.
 298  * When it is called during GC, it just gets a victim segment
 299  * and it does not remove it from dirty seglist.
 300  * When it is called from SSR segment selection, it finds a segment
 301  * which has minimum valid blocks and removes it from dirty seglist.
 302  */
 303 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 304                 unsigned int *result, int gc_type, int type, char alloc_mode)
 305 {
 306         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 307         struct sit_info *sm = SIT_I(sbi);
 308         struct victim_sel_policy p;
 309         unsigned int secno, last_victim;
 310         unsigned int last_segment = MAIN_SEGS(sbi);
 311         unsigned int nsearched = 0;
 312
 313         mutex_lock(&dirty_i->seglist_lock);
 314
 315         p.alloc_mode = alloc_mode;
 316         select_policy(sbi, gc_type, type, &p);
 317
 318         p.min_segno = NULL_SEGNO;
 319         p.min_cost = get_max_cost(sbi, &p);
 320
 321         if (*result != NULL_SEGNO) {
 322                 if (IS_DATASEG(get_seg_entry(sbi, *result)->type) &&
 323                         get_valid_blocks(sbi, *result, false) &&
 324                         !sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
 325                         p.min_segno = *result;
 326                 goto out;
 327         }
 328
 329         if (p.max_search == 0)
 330                 goto out;
 331
 332         last_victim = sm->last_victim[p.gc_mode];
 333         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 334                 p.min_segno = check_bg_victims(sbi);
 335                 if (p.min_segno != NULL_SEGNO)
 336                         goto got_it;
 337         }
 338
 339         while (1) {
 340                 unsigned long cost;
 341                 unsigned int segno;
 342
 343                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 344                 if (segno >= last_segment) {
 345                         if (sm->last_victim[p.gc_mode]) {
 346                                 last_segment =
 347                                         sm->last_victim[p.gc_mode];
 348                                 sm->last_victim[p.gc_mode] = 0;
 349                                 p.offset = 0;
 350                                 continue;
 351                         }
 352                         break;
 353                 }
 354
 355                 p.offset = segno + p.ofs_unit;
 356                 if (p.ofs_unit > 1) {
 357                         p.offset -= segno % p.ofs_unit;
 358                         nsearched += count_bits(p.dirty_segmap,
 359                                                 p.offset - p.ofs_unit,
 360                                                 p.ofs_unit);
 361                 } else {
 362                         nsearched++;
 363                 }
 364
 365                 secno = GET_SEC_FROM_SEG(sbi, segno);
 366
 367                 if (sec_usage_check(sbi, secno))
 368                         goto next;
 369                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 370                         goto next;
 371                 if (gc_type == FG_GC && p.alloc_mode == LFS &&
 372                                         no_fggc_candidate(sbi, secno))
 373                         goto next;
 374
 375                 cost = get_gc_cost(sbi, segno, &p);
 376
 377                 if (p.min_cost > cost) {
 378                         p.min_segno = segno;
 379                         p.min_cost = cost;
 380                 }
 381 next:
 382                 if (nsearched >= p.max_search) {
 383                         if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
 384                                 sm->last_victim[p.gc_mode] = last_victim + 1;
 385                         else
 386                                 sm->last_victim[p.gc_mode] = segno + 1;
 387                         sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
 388                         break;
 389                 }
 390         }
 391         if (p.min_segno != NULL_SEGNO) {
 392 got_it:
 393                 if (p.alloc_mode == LFS) {
 394                         secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
 395                         if (gc_type == FG_GC)
 396                                 sbi->cur_victim_sec = secno;
 397                         else
 398                                 set_bit(secno, dirty_i->victim_secmap);
 399                 }
 400                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 401
 402                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 403                                 sbi->cur_victim_sec,
 404                                 prefree_segments(sbi), free_segments(sbi));
 405         }
 406 out:
 407         mutex_unlock(&dirty_i->seglist_lock);
 408
 409         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 410 }
 411
 412 static const struct victim_selection default_v_ops = {
 413         .get_victim = get_victim_by_default,
 414 };
 415
 416 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 417 {
 418         struct inode_entry *ie;
 419
 420         ie = radix_tree_lookup(&gc_list->iroot, ino);
 421         if (ie)
 422                 return ie->inode;
 423         return NULL;
 424 }
 425
 426 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 427 {
 428         struct inode_entry *new_ie;
 429
 430         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 431                 iput(inode);
 432                 return;
 433         }
 434         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 435         new_ie->inode = inode;
 436
 437         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 438         list_add_tail(&new_ie->list, &gc_list->ilist);
 439 }
 440
 441 static void put_gc_inode(struct gc_inode_list *gc_list)
 442 {
 443         struct inode_entry *ie, *next_ie;
 444         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 445                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 446                 iput(ie->inode);
 447                 list_del(&ie->list);
 448                 kmem_cache_free(inode_entry_slab, ie);
 449         }
 450 }
 451
 452 static int check_valid_map(struct f2fs_sb_info *sbi,
 453                                 unsigned int segno, int offset)
 454 {
 455         struct sit_info *sit_i = SIT_I(sbi);
 456         struct seg_entry *sentry;
 457         int ret;
 458
 459         down_read(&sit_i->sentry_lock);
 460         sentry = get_seg_entry(sbi, segno);
 461         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 462         up_read(&sit_i->sentry_lock);
 463         return ret;
 464 }
 465
 466 /*
 467  * This function compares node address got in summary with that in NAT.
 468  * On validity, copy that node with cold status, otherwise (invalid node)
 469  * ignore that.
 470  */
 471 static void gc_node_segment(struct f2fs_sb_info *sbi,
 472                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 473 {
 474         struct f2fs_summary *entry;
 475         block_t start_addr;
 476         int off;
 477         int phase = 0;
 478
 479         start_addr = START_BLOCK(sbi, segno);
 480
 481 next_step:
 482         entry = sum;
 483
 484         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 485                 nid_t nid = le32_to_cpu(entry->nid);
 486                 struct page *node_page;
 487                 struct node_info ni;
 488
 489                 /* stop BG_GC if there is not enough free sections. */
 490                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 491                         return;
 492
 493                 if (check_valid_map(sbi, segno, off) == 0)
 494                         continue;
 495
 496                 if (phase == 0) {
 497                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 498                                                         META_NAT, true);
 499                         continue;
 500                 }
 501
 502                 if (phase == 1) {
 503                         ra_node_page(sbi, nid);
 504                         continue;
 505                 }
 506
 507                 /* phase == 2 */
 508                 node_page = get_node_page(sbi, nid);
 509                 if (IS_ERR(node_page))
 510                         continue;
 511
 512                 /* block may become invalid during get_node_page */
 513                 if (check_valid_map(sbi, segno, off) == 0) {
 514                         f2fs_put_page(node_page, 1);
 515                         continue;
 516                 }
 517
 518                 get_node_info(sbi, nid, &ni);
 519                 if (ni.blk_addr != start_addr + off) {
 520                         f2fs_put_page(node_page, 1);
 521                         continue;
 522                 }
 523
 524                 move_node_page(node_page, gc_type);
 525                 stat_inc_node_blk_count(sbi, 1, gc_type);
 526         }
 527
 528         if (++phase < 3)
 529                 goto next_step;
 530 }
 531
 532 /*
 533  * Calculate start block index indicating the given node offset.
 534  * Be careful, caller should give this node offset only indicating direct node
 535  * blocks. If any node offsets, which point the other types of node blocks such
 536  * as indirect or double indirect node blocks, are given, it must be a caller's
 537  * bug.
 538  */
 539 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 540 {
 541         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 542         unsigned int bidx;
 543
 544         if (node_ofs == 0)
 545                 return 0;
 546
 547         if (node_ofs <= 2) {
 548                 bidx = node_ofs - 1;
 549         } else if (node_ofs <= indirect_blks) {
 550                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 551                 bidx = node_ofs - 2 - dec;
 552         } else {
 553                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 554                 bidx = node_ofs - 5 - dec;
 555         }
 556         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 557 }
 558
 559 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 560                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 561 {
 562         struct page *node_page;
 563         nid_t nid;
 564         unsigned int ofs_in_node;
 565         block_t source_blkaddr;
 566
 567         nid = le32_to_cpu(sum->nid);
 568         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 569
 570         node_page = get_node_page(sbi, nid);
 571         if (IS_ERR(node_page))
 572                 return false;
 573
 574         get_node_info(sbi, nid, dni);
 575
 576         if (sum->version != dni->version) {
 577                 f2fs_msg(sbi->sb, KERN_WARNING,
 578                                 "%s: valid data with mismatched node version.",
 579                                 __func__);
 580                 set_sbi_flag(sbi, SBI_NEED_FSCK);
 581         }
 582
 583         *nofs = ofs_of_node(node_page);
 584         source_blkaddr = datablock_addr(NULL, node_page, ofs_in_node);
 585         f2fs_put_page(node_page, 1);
 586
 587         if (source_blkaddr != blkaddr)
 588                 return false;
 589         return true;
 590 }
 591
 592 /*
 593  * Move data block via META_MAPPING while keeping locked data page.
 594  * This can be used to move blocks, aka LBAs, directly on disk.
 595  */
 596 static void move_data_block(struct inode *inode, block_t bidx,
 597                                         unsigned int segno, int off)
 598 {
 599         struct f2fs_io_info fio = {
 600                 .sbi = F2FS_I_SB(inode),
 601                 .ino = inode->i_ino,
 602                 .type = DATA,
 603                 .temp = COLD,
 604                 .op = REQ_OP_READ,
 605                 .op_flags = 0,
 606                 .encrypted_page = NULL,
 607                 .in_list = false,
 608         };
 609         struct dnode_of_data dn;
 610         struct f2fs_summary sum;
 611         struct node_info ni;
 612         struct page *page;
 613         block_t newaddr;
 614         int err;
 615
 616         /* do not read out */
 617         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 618         if (!page)
 619                 return;
 620
 621         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 622                 goto out;
 623
 624         if (f2fs_is_atomic_file(inode))
 625                 goto out;
 626
 627         set_new_dnode(&dn, inode, NULL, NULL, 0);
 628         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 629         if (err)
 630                 goto out;
 631
 632         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 633                 ClearPageUptodate(page);
 634                 goto put_out;
 635         }
 636
 637         /*
 638          * don't cache encrypted data into meta inode until previous dirty
 639          * data were writebacked to avoid racing between GC and flush.
 640          */
 641         f2fs_wait_on_page_writeback(page, DATA, true);
 642
 643         get_node_info(fio.sbi, dn.nid, &ni);
 644         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 645
 646         /* read page */
 647         fio.page = page;
 648         fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
 649
 650         allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
 651                                         &sum, CURSEG_COLD_DATA, NULL, false);
 652
 653         fio.encrypted_page = f2fs_pagecache_get_page(META_MAPPING(fio.sbi),
 654                                 newaddr, FGP_LOCK | FGP_CREAT, GFP_NOFS);
 655         if (!fio.encrypted_page) {
 656                 err = -ENOMEM;
 657                 goto recover_block;
 658         }
 659
 660         err = f2fs_submit_page_bio(&fio);
 661         if (err)
 662                 goto put_page_out;
 663
 664         /* write page */
 665         lock_page(fio.encrypted_page);
 666
 667         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
 668                 err = -EIO;
 669                 goto put_page_out;
 670         }
 671         if (unlikely(!PageUptodate(fio.encrypted_page))) {
 672                 err = -EIO;
 673                 goto put_page_out;
 674         }
 675
 676         set_page_dirty(fio.encrypted_page);
 677         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 678         if (clear_page_dirty_for_io(fio.encrypted_page))
 679                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 680
 681         set_page_writeback(fio.encrypted_page);
 682
 683         /* allocate block address */
 684         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 685
 686         fio.op = REQ_OP_WRITE;
 687         fio.op_flags = REQ_SYNC;
 688         fio.new_blkaddr = newaddr;
 689         f2fs_submit_page_write(&fio);
 690
 691         f2fs_update_iostat(fio.sbi, FS_GC_DATA_IO, F2FS_BLKSIZE);
 692
 693         f2fs_update_data_blkaddr(&dn, newaddr);
 694         set_inode_flag(inode, FI_APPEND_WRITE);
 695         if (page->index == 0)
 696                 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
 697 put_page_out:
 698         f2fs_put_page(fio.encrypted_page, 1);
 699 recover_block:
 700         if (err)
 701                 __f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
 702                                                                 true, true);
 703 put_out:
 704         f2fs_put_dnode(&dn);
 705 out:
 706         f2fs_put_page(page, 1);
 707 }
 708
 709 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 710                                                         unsigned int segno, int off)
 711 {
 712         struct page *page;
 713
 714         page = get_lock_data_page(inode, bidx, true);
 715         if (IS_ERR(page))
 716                 return;
 717
 718         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 719                 goto out;
 720
 721         if (f2fs_is_atomic_file(inode))
 722                 goto out;
 723
 724         if (gc_type == BG_GC) {
 725                 if (PageWriteback(page))
 726                         goto out;
 727                 set_page_dirty(page);
 728                 set_cold_data(page);
 729         } else {
 730                 struct f2fs_io_info fio = {
 731                         .sbi = F2FS_I_SB(inode),
 732                         .ino = inode->i_ino,
 733                         .type = DATA,
 734                         .temp = COLD,
 735                         .op = REQ_OP_WRITE,
 736                         .op_flags = REQ_SYNC,
 737                         .old_blkaddr = NULL_ADDR,
 738                         .page = page,
 739                         .encrypted_page = NULL,
 740                         .need_lock = LOCK_REQ,
 741                         .io_type = FS_GC_DATA_IO,
 742                 };
 743                 bool is_dirty = PageDirty(page);
 744                 int err;
 745
 746 retry:
 747                 set_page_dirty(page);
 748                 f2fs_wait_on_page_writeback(page, DATA, true);
 749                 if (clear_page_dirty_for_io(page)) {
 750                         inode_dec_dirty_pages(inode);
 751                         remove_dirty_inode(inode);
 752                 }
 753
 754                 set_cold_data(page);
 755
 756                 err = do_write_data_page(&fio);
 757                 if (err == -ENOMEM && is_dirty) {
 758                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 759                         goto retry;
 760                 }
 761         }
 762 out:
 763         f2fs_put_page(page, 1);
 764 }
 765
 766 /*
 767  * This function tries to get parent node of victim data block, and identifies
 768  * data block validity. If the block is valid, copy that with cold status and
 769  * modify parent node.
 770  * If the parent node is not valid or the data block address is different,
 771  * the victim data block is ignored.
 772  */
 773 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 774                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 775 {
 776         struct super_block *sb = sbi->sb;
 777         struct f2fs_summary *entry;
 778         block_t start_addr;
 779         int off;
 780         int phase = 0;
 781
 782         start_addr = START_BLOCK(sbi, segno);
 783
 784 next_step:
 785         entry = sum;
 786
 787         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 788                 struct page *data_page;
 789                 struct inode *inode;
 790                 struct node_info dni; /* dnode info for the data */
 791                 unsigned int ofs_in_node, nofs;
 792                 block_t start_bidx;
 793                 nid_t nid = le32_to_cpu(entry->nid);
 794
 795                 /* stop BG_GC if there is not enough free sections. */
 796                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 797                         return;
 798
 799                 if (check_valid_map(sbi, segno, off) == 0)
 800                         continue;
 801
 802                 if (phase == 0) {
 803                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 804                                                         META_NAT, true);
 805                         continue;
 806                 }
 807
 808                 if (phase == 1) {
 809                         ra_node_page(sbi, nid);
 810                         continue;
 811                 }
 812
 813                 /* Get an inode by ino with checking validity */
 814                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 815                         continue;
 816
 817                 if (phase == 2) {
 818                         ra_node_page(sbi, dni.ino);
 819                         continue;
 820                 }
 821
 822                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 823
 824                 if (phase == 3) {
 825                         inode = f2fs_iget(sb, dni.ino);
 826                         if (IS_ERR(inode) || is_bad_inode(inode))
 827                                 continue;
 828
 829                         /* if encrypted inode, let's go phase 3 */
 830                         if (f2fs_encrypted_file(inode)) {
 831                                 add_gc_inode(gc_list, inode);
 832                                 continue;
 833                         }
 834
 835                         if (!down_write_trylock(
 836                                 &F2FS_I(inode)->dio_rwsem[WRITE])) {
 837                                 iput(inode);
 838                                 continue;
 839                         }
 840
 841                         start_bidx = start_bidx_of_node(nofs, inode);
 842                         data_page = get_read_data_page(inode,
 843                                         start_bidx + ofs_in_node, REQ_RAHEAD,
 844                                         true);
 845                         up_write(&F2FS_I(inode)->dio_rwsem[WRITE]);
 846                         if (IS_ERR(data_page)) {
 847                                 iput(inode);
 848                                 continue;
 849                         }
 850
 851                         f2fs_put_page(data_page, 0);
 852                         add_gc_inode(gc_list, inode);
 853                         continue;
 854                 }
 855
 856                 /* phase 4 */
 857                 inode = find_gc_inode(gc_list, dni.ino);
 858                 if (inode) {
 859                         struct f2fs_inode_info *fi = F2FS_I(inode);
 860                         bool locked = false;
 861
 862                         if (S_ISREG(inode->i_mode)) {
 863                                 if (!down_write_trylock(&fi->dio_rwsem[READ]))
 864                                         continue;
 865                                 if (!down_write_trylock(
 866                                                 &fi->dio_rwsem[WRITE])) {
 867                                         up_write(&fi->dio_rwsem[READ]);
 868                                         continue;
 869                                 }
 870                                 locked = true;
 871
 872                                 /* wait for all inflight aio data */
 873                                 inode_dio_wait(inode);
 874                         }
 875
 876                         start_bidx = start_bidx_of_node(nofs, inode)
 877                                                                 + ofs_in_node;
 878                         if (f2fs_encrypted_file(inode))
 879                                 move_data_block(inode, start_bidx, segno, off);
 880                         else
 881                                 move_data_page(inode, start_bidx, gc_type,
 882                                                                 segno, off);
 883
 884                         if (locked) {
 885                                 up_write(&fi->dio_rwsem[WRITE]);
 886                                 up_write(&fi->dio_rwsem[READ]);
 887                         }
 888
 889                         stat_inc_data_blk_count(sbi, 1, gc_type);
 890                 }
 891         }
 892
 893         if (++phase < 5)
 894                 goto next_step;
 895 }
 896
 897 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 898                         int gc_type)
 899 {
 900         struct sit_info *sit_i = SIT_I(sbi);
 901         int ret;
 902
 903         down_write(&sit_i->sentry_lock);
 904         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 905                                               NO_CHECK_TYPE, LFS);
 906         up_write(&sit_i->sentry_lock);
 907         return ret;
 908 }
 909
 910 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 911                                 unsigned int start_segno,
 912                                 struct gc_inode_list *gc_list, int gc_type)
 913 {
 914         struct page *sum_page;
 915         struct f2fs_summary_block *sum;
 916         struct blk_plug plug;
 917         unsigned int segno = start_segno;
 918         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 919         int seg_freed = 0;
 920         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 921                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 922
 923         /* readahead multi ssa blocks those have contiguous address */
 924         if (sbi->segs_per_sec > 1)
 925                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 926                                         sbi->segs_per_sec, META_SSA, true);
 927
 928         /* reference all summary page */
 929         while (segno < end_segno) {
 930                 sum_page = get_sum_page(sbi, segno++);
 931                 unlock_page(sum_page);
 932         }
 933
 934         blk_start_plug(&plug);
 935
 936         for (segno = start_segno; segno < end_segno; segno++) {
 937
 938                 /* find segment summary of victim */
 939                 sum_page = find_get_page(META_MAPPING(sbi),
 940                                         GET_SUM_BLOCK(sbi, segno));
 941                 f2fs_put_page(sum_page, 0);
 942
 943                 if (get_valid_blocks(sbi, segno, false) == 0 ||
 944                                 !PageUptodate(sum_page) ||
 945                                 unlikely(f2fs_cp_error(sbi)))
 946                         goto next;
 947
 948                 sum = page_address(sum_page);
 949                 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
 950
 951                 /*
 952                  * this is to avoid deadlock:
 953                  * - lock_page(sum_page)         - f2fs_replace_block
 954                  *  - check_valid_map()            - down_write(sentry_lock)
 955                  *   - down_read(sentry_lock)     - change_curseg()
 956                  *                                  - lock_page(sum_page)
 957                  */
 958                 if (type == SUM_TYPE_NODE)
 959                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 960                 else
 961                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 962                                                                 gc_type);
 963
 964                 stat_inc_seg_count(sbi, type, gc_type);
 965
 966                 if (gc_type == FG_GC &&
 967                                 get_valid_blocks(sbi, segno, false) == 0)
 968                         seg_freed++;
 969 next:
 970                 f2fs_put_page(sum_page, 0);
 971         }
 972
 973         if (gc_type == FG_GC)
 974                 f2fs_submit_merged_write(sbi,
 975                                 (type == SUM_TYPE_NODE) ? NODE : DATA);
 976
 977         blk_finish_plug(&plug);
 978
 979         stat_inc_call_count(sbi->stat_info);
 980
 981         return seg_freed;
 982 }
 983
 984 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 985                         bool background, unsigned int segno)
 986 {
 987         int gc_type = sync ? FG_GC : BG_GC;
 988         int sec_freed = 0, seg_freed = 0, total_freed = 0;
 989         int ret = 0;
 990         struct cp_control cpc;
 991         unsigned int init_segno = segno;
 992         struct gc_inode_list gc_list = {
 993                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 994                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 995         };
 996
 997         trace_f2fs_gc_begin(sbi->sb, sync, background,
 998                                 get_pages(sbi, F2FS_DIRTY_NODES),
 999                                 get_pages(sbi, F2FS_DIRTY_DENTS),
1000                                 get_pages(sbi, F2FS_DIRTY_IMETA),
1001                                 free_sections(sbi),
1002                                 free_segments(sbi),
1003                                 reserved_segments(sbi),
1004                                 prefree_segments(sbi));
1005
1006         cpc.reason = __get_cp_reason(sbi);
1007 gc_more:
1008         if (unlikely(!(sbi->sb->s_flags & SB_ACTIVE))) {
1009                 ret = -EINVAL;
1010                 goto stop;
1011         }
1012         if (unlikely(f2fs_cp_error(sbi))) {
1013                 ret = -EIO;
1014                 goto stop;
1015         }
1016
1017         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
1018                 /*
1019                  * For example, if there are many prefree_segments below given
1020                  * threshold, we can make them free by checkpoint. Then, we
1021                  * secure free segments which doesn't need fggc any more.
1022                  */
1023                 if (prefree_segments(sbi)) {
1024                         ret = write_checkpoint(sbi, &cpc);
1025                         if (ret)
1026                                 goto stop;
1027                 }
1028                 if (has_not_enough_free_secs(sbi, 0, 0))
1029                         gc_type = FG_GC;
1030         }
1031
1032         /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
1033         if (gc_type == BG_GC && !background) {
1034                 ret = -EINVAL;
1035                 goto stop;
1036         }
1037         if (!__get_victim(sbi, &segno, gc_type)) {
1038                 ret = -ENODATA;
1039                 goto stop;
1040         }
1041
1042         seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
1043         if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
1044                 sec_freed++;
1045         total_freed += seg_freed;
1046
1047         if (gc_type == FG_GC)
1048                 sbi->cur_victim_sec = NULL_SEGNO;
1049
1050         if (!sync) {
1051                 if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1052                         segno = NULL_SEGNO;
1053                         goto gc_more;
1054                 }
1055
1056                 if (gc_type == FG_GC)
1057                         ret = write_checkpoint(sbi, &cpc);
1058         }
1059 stop:
1060         SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1061         SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1062
1063         trace_f2fs_gc_end(sbi->sb, ret, total_freed, sec_freed,
1064                                 get_pages(sbi, F2FS_DIRTY_NODES),
1065                                 get_pages(sbi, F2FS_DIRTY_DENTS),
1066                                 get_pages(sbi, F2FS_DIRTY_IMETA),
1067                                 free_sections(sbi),
1068                                 free_segments(sbi),
1069                                 reserved_segments(sbi),
1070                                 prefree_segments(sbi));
1071
1072         mutex_unlock(&sbi->gc_mutex);
1073
1074         put_gc_inode(&gc_list);
1075
1076         if (sync)
1077                 ret = sec_freed ? 0 : -EAGAIN;
1078         return ret;
1079 }
1080
1081 void build_gc_manager(struct f2fs_sb_info *sbi)
1082 {
1083         u64 main_count, resv_count, ovp_count;
1084
1085         DIRTY_I(sbi)->v_ops = &default_v_ops;
1086
1087         /* threshold of # of valid blocks in a section for victims of FG_GC */
1088         main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1089         resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1090         ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1091
1092         sbi->fggc_threshold = div64_u64((main_count - ovp_count) *
1093                                 BLKS_PER_SEC(sbi), (main_count - resv_count));
1094
1095         /* give warm/cold data area from slower device */
1096         if (sbi->s_ndevs && sbi->segs_per_sec == 1)
1097                 SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1098                                 GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
1099 }