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 in no_heap mode*/
 195         if (test_opt(sbi, NOHEAP) &&
 196                 (type == CURSEG_HOT_DATA || IS_NODESEG(type)))
 197                 p->offset = 0;
 198         else
 199                 p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
 200 }
 201
 202 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 203                                 struct victim_sel_policy *p)
 204 {
 205         /* SSR allocates in a segment unit */
 206         if (p->alloc_mode == SSR)
 207                 return sbi->blocks_per_seg;
 208         if (p->gc_mode == GC_GREEDY)
 209                 return 2 * sbi->blocks_per_seg * p->ofs_unit;
 210         else if (p->gc_mode == GC_CB)
 211                 return UINT_MAX;
 212         else /* No other gc_mode */
 213                 return 0;
 214 }
 215
 216 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 217 {
 218         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 219         unsigned int secno;
 220
 221         /*
 222          * If the gc_type is FG_GC, we can select victim segments
 223          * selected by background GC before.
 224          * Those segments guarantee they have small valid blocks.
 225          */
 226         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 227                 if (sec_usage_check(sbi, secno))
 228                         continue;
 229
 230                 if (no_fggc_candidate(sbi, secno))
 231                         continue;
 232
 233                 clear_bit(secno, dirty_i->victim_secmap);
 234                 return GET_SEG_FROM_SEC(sbi, secno);
 235         }
 236         return NULL_SEGNO;
 237 }
 238
 239 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 240 {
 241         struct sit_info *sit_i = SIT_I(sbi);
 242         unsigned int secno = GET_SEC_FROM_SEG(sbi, segno);
 243         unsigned int start = GET_SEG_FROM_SEC(sbi, secno);
 244         unsigned long long mtime = 0;
 245         unsigned int vblocks;
 246         unsigned char age = 0;
 247         unsigned char u;
 248         unsigned int i;
 249
 250         for (i = 0; i < sbi->segs_per_sec; i++)
 251                 mtime += get_seg_entry(sbi, start + i)->mtime;
 252         vblocks = get_valid_blocks(sbi, segno, true);
 253
 254         mtime = div_u64(mtime, sbi->segs_per_sec);
 255         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 256
 257         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 258
 259         /* Handle if the system time has changed by the user */
 260         if (mtime < sit_i->min_mtime)
 261                 sit_i->min_mtime = mtime;
 262         if (mtime > sit_i->max_mtime)
 263                 sit_i->max_mtime = mtime;
 264         if (sit_i->max_mtime != sit_i->min_mtime)
 265                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 266                                 sit_i->max_mtime - sit_i->min_mtime);
 267
 268         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 269 }
 270
 271 static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
 272                                                 unsigned int segno)
 273 {
 274         unsigned int valid_blocks =
 275                         get_valid_blocks(sbi, segno, true);
 276
 277         return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 278                                 valid_blocks * 2 : valid_blocks;
 279 }
 280
 281 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 282                         unsigned int segno, struct victim_sel_policy *p)
 283 {
 284         if (p->alloc_mode == SSR)
 285                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 286
 287         /* alloc_mode == LFS */
 288         if (p->gc_mode == GC_GREEDY)
 289                 return get_greedy_cost(sbi, segno);
 290         else
 291                 return get_cb_cost(sbi, segno);
 292 }
 293
 294 static unsigned int count_bits(const unsigned long *addr,
 295                                 unsigned int offset, unsigned int len)
 296 {
 297         unsigned int end = offset + len, sum = 0;
 298
 299         while (offset < end) {
 300                 if (test_bit(offset++, addr))
 301                         ++sum;
 302         }
 303         return sum;
 304 }
 305
 306 /*
 307  * This function is called from two paths.
 308  * One is garbage collection and the other is SSR segment selection.
 309  * When it is called during GC, it just gets a victim segment
 310  * and it does not remove it from dirty seglist.
 311  * When it is called from SSR segment selection, it finds a segment
 312  * which has minimum valid blocks and removes it from dirty seglist.
 313  */
 314 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 315                 unsigned int *result, int gc_type, int type, char alloc_mode)
 316 {
 317         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 318         struct sit_info *sm = SIT_I(sbi);
 319         struct victim_sel_policy p;
 320         unsigned int secno, last_victim;
 321         unsigned int last_segment = MAIN_SEGS(sbi);
 322         unsigned int nsearched = 0;
 323
 324         mutex_lock(&dirty_i->seglist_lock);
 325
 326         p.alloc_mode = alloc_mode;
 327         select_policy(sbi, gc_type, type, &p);
 328
 329         p.min_segno = NULL_SEGNO;
 330         p.min_cost = get_max_cost(sbi, &p);
 331
 332         if (*result != NULL_SEGNO) {
 333                 if (get_valid_blocks(sbi, *result, false) &&
 334                         !sec_usage_check(sbi, GET_SEC_FROM_SEG(sbi, *result)))
 335                         p.min_segno = *result;
 336                 goto out;
 337         }
 338
 339         if (p.max_search == 0)
 340                 goto out;
 341
 342         last_victim = sm->last_victim[p.gc_mode];
 343         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 344                 p.min_segno = check_bg_victims(sbi);
 345                 if (p.min_segno != NULL_SEGNO)
 346                         goto got_it;
 347         }
 348
 349         while (1) {
 350                 unsigned long cost;
 351                 unsigned int segno;
 352
 353                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 354                 if (segno >= last_segment) {
 355                         if (sm->last_victim[p.gc_mode]) {
 356                                 last_segment =
 357                                         sm->last_victim[p.gc_mode];
 358                                 sm->last_victim[p.gc_mode] = 0;
 359                                 p.offset = 0;
 360                                 continue;
 361                         }
 362                         break;
 363                 }
 364
 365                 p.offset = segno + p.ofs_unit;
 366                 if (p.ofs_unit > 1) {
 367                         p.offset -= segno % p.ofs_unit;
 368                         nsearched += count_bits(p.dirty_segmap,
 369                                                 p.offset - p.ofs_unit,
 370                                                 p.ofs_unit);
 371                 } else {
 372                         nsearched++;
 373                 }
 374
 375                 secno = GET_SEC_FROM_SEG(sbi, segno);
 376
 377                 if (sec_usage_check(sbi, secno))
 378                         goto next;
 379                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 380                         goto next;
 381                 if (gc_type == FG_GC && p.alloc_mode == LFS &&
 382                                         no_fggc_candidate(sbi, secno))
 383                         goto next;
 384
 385                 cost = get_gc_cost(sbi, segno, &p);
 386
 387                 if (p.min_cost > cost) {
 388                         p.min_segno = segno;
 389                         p.min_cost = cost;
 390                 }
 391 next:
 392                 if (nsearched >= p.max_search) {
 393                         if (!sm->last_victim[p.gc_mode] && segno <= last_victim)
 394                                 sm->last_victim[p.gc_mode] = last_victim + 1;
 395                         else
 396                                 sm->last_victim[p.gc_mode] = segno + 1;
 397                         sm->last_victim[p.gc_mode] %= MAIN_SEGS(sbi);
 398                         break;
 399                 }
 400         }
 401         if (p.min_segno != NULL_SEGNO) {
 402 got_it:
 403                 if (p.alloc_mode == LFS) {
 404                         secno = GET_SEC_FROM_SEG(sbi, p.min_segno);
 405                         if (gc_type == FG_GC)
 406                                 sbi->cur_victim_sec = secno;
 407                         else
 408                                 set_bit(secno, dirty_i->victim_secmap);
 409                 }
 410                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 411
 412                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 413                                 sbi->cur_victim_sec,
 414                                 prefree_segments(sbi), free_segments(sbi));
 415         }
 416 out:
 417         mutex_unlock(&dirty_i->seglist_lock);
 418
 419         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 420 }
 421
 422 static const struct victim_selection default_v_ops = {
 423         .get_victim = get_victim_by_default,
 424 };
 425
 426 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 427 {
 428         struct inode_entry *ie;
 429
 430         ie = radix_tree_lookup(&gc_list->iroot, ino);
 431         if (ie)
 432                 return ie->inode;
 433         return NULL;
 434 }
 435
 436 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 437 {
 438         struct inode_entry *new_ie;
 439
 440         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 441                 iput(inode);
 442                 return;
 443         }
 444         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 445         new_ie->inode = inode;
 446
 447         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 448         list_add_tail(&new_ie->list, &gc_list->ilist);
 449 }
 450
 451 static void put_gc_inode(struct gc_inode_list *gc_list)
 452 {
 453         struct inode_entry *ie, *next_ie;
 454         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 455                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 456                 iput(ie->inode);
 457                 list_del(&ie->list);
 458                 kmem_cache_free(inode_entry_slab, ie);
 459         }
 460 }
 461
 462 static int check_valid_map(struct f2fs_sb_info *sbi,
 463                                 unsigned int segno, int offset)
 464 {
 465         struct sit_info *sit_i = SIT_I(sbi);
 466         struct seg_entry *sentry;
 467         int ret;
 468
 469         mutex_lock(&sit_i->sentry_lock);
 470         sentry = get_seg_entry(sbi, segno);
 471         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 472         mutex_unlock(&sit_i->sentry_lock);
 473         return ret;
 474 }
 475
 476 /*
 477  * This function compares node address got in summary with that in NAT.
 478  * On validity, copy that node with cold status, otherwise (invalid node)
 479  * ignore that.
 480  */
 481 static void gc_node_segment(struct f2fs_sb_info *sbi,
 482                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 483 {
 484         struct f2fs_summary *entry;
 485         block_t start_addr;
 486         int off;
 487         int phase = 0;
 488
 489         start_addr = START_BLOCK(sbi, segno);
 490
 491 next_step:
 492         entry = sum;
 493
 494         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 495                 nid_t nid = le32_to_cpu(entry->nid);
 496                 struct page *node_page;
 497                 struct node_info ni;
 498
 499                 /* stop BG_GC if there is not enough free sections. */
 500                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 501                         return;
 502
 503                 if (check_valid_map(sbi, segno, off) == 0)
 504                         continue;
 505
 506                 if (phase == 0) {
 507                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 508                                                         META_NAT, true);
 509                         continue;
 510                 }
 511
 512                 if (phase == 1) {
 513                         ra_node_page(sbi, nid);
 514                         continue;
 515                 }
 516
 517                 /* phase == 2 */
 518                 node_page = get_node_page(sbi, nid);
 519                 if (IS_ERR(node_page))
 520                         continue;
 521
 522                 /* block may become invalid during get_node_page */
 523                 if (check_valid_map(sbi, segno, off) == 0) {
 524                         f2fs_put_page(node_page, 1);
 525                         continue;
 526                 }
 527
 528                 get_node_info(sbi, nid, &ni);
 529                 if (ni.blk_addr != start_addr + off) {
 530                         f2fs_put_page(node_page, 1);
 531                         continue;
 532                 }
 533
 534                 move_node_page(node_page, gc_type);
 535                 stat_inc_node_blk_count(sbi, 1, gc_type);
 536         }
 537
 538         if (++phase < 3)
 539                 goto next_step;
 540 }
 541
 542 /*
 543  * Calculate start block index indicating the given node offset.
 544  * Be careful, caller should give this node offset only indicating direct node
 545  * blocks. If any node offsets, which point the other types of node blocks such
 546  * as indirect or double indirect node blocks, are given, it must be a caller's
 547  * bug.
 548  */
 549 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 550 {
 551         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 552         unsigned int bidx;
 553
 554         if (node_ofs == 0)
 555                 return 0;
 556
 557         if (node_ofs <= 2) {
 558                 bidx = node_ofs - 1;
 559         } else if (node_ofs <= indirect_blks) {
 560                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 561                 bidx = node_ofs - 2 - dec;
 562         } else {
 563                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 564                 bidx = node_ofs - 5 - dec;
 565         }
 566         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 567 }
 568
 569 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 570                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 571 {
 572         struct page *node_page;
 573         nid_t nid;
 574         unsigned int ofs_in_node;
 575         block_t source_blkaddr;
 576
 577         nid = le32_to_cpu(sum->nid);
 578         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 579
 580         node_page = get_node_page(sbi, nid);
 581         if (IS_ERR(node_page))
 582                 return false;
 583
 584         get_node_info(sbi, nid, dni);
 585
 586         if (sum->version != dni->version) {
 587                 f2fs_msg(sbi->sb, KERN_WARNING,
 588                                 "%s: valid data with mismatched node version.",
 589                                 __func__);
 590                 set_sbi_flag(sbi, SBI_NEED_FSCK);
 591         }
 592
 593         *nofs = ofs_of_node(node_page);
 594         source_blkaddr = datablock_addr(NULL, node_page, ofs_in_node);
 595         f2fs_put_page(node_page, 1);
 596
 597         if (source_blkaddr != blkaddr)
 598                 return false;
 599         return true;
 600 }
 601
 602 /*
 603  * Move data block via META_MAPPING while keeping locked data page.
 604  * This can be used to move blocks, aka LBAs, directly on disk.
 605  */
 606 static void move_data_block(struct inode *inode, block_t bidx,
 607                                         unsigned int segno, int off)
 608 {
 609         struct f2fs_io_info fio = {
 610                 .sbi = F2FS_I_SB(inode),
 611                 .type = DATA,
 612                 .temp = COLD,
 613                 .op = REQ_OP_READ,
 614                 .op_flags = 0,
 615                 .encrypted_page = NULL,
 616                 .in_list = false,
 617         };
 618         struct dnode_of_data dn;
 619         struct f2fs_summary sum;
 620         struct node_info ni;
 621         struct page *page;
 622         block_t newaddr;
 623         int err;
 624
 625         /* do not read out */
 626         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 627         if (!page)
 628                 return;
 629
 630         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 631                 goto out;
 632
 633         if (f2fs_is_atomic_file(inode))
 634                 goto out;
 635
 636         set_new_dnode(&dn, inode, NULL, NULL, 0);
 637         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 638         if (err)
 639                 goto out;
 640
 641         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 642                 ClearPageUptodate(page);
 643                 goto put_out;
 644         }
 645
 646         /*
 647          * don't cache encrypted data into meta inode until previous dirty
 648          * data were writebacked to avoid racing between GC and flush.
 649          */
 650         f2fs_wait_on_page_writeback(page, DATA, true);
 651
 652         get_node_info(fio.sbi, dn.nid, &ni);
 653         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 654
 655         /* read page */
 656         fio.page = page;
 657         fio.new_blkaddr = fio.old_blkaddr = dn.data_blkaddr;
 658
 659         allocate_data_block(fio.sbi, NULL, fio.old_blkaddr, &newaddr,
 660                                         &sum, CURSEG_COLD_DATA, NULL, false);
 661
 662         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi), newaddr,
 663                                         FGP_LOCK | FGP_CREAT, GFP_NOFS);
 664         if (!fio.encrypted_page) {
 665                 err = -ENOMEM;
 666                 goto recover_block;
 667         }
 668
 669         err = f2fs_submit_page_bio(&fio);
 670         if (err)
 671                 goto put_page_out;
 672
 673         /* write page */
 674         lock_page(fio.encrypted_page);
 675
 676         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi))) {
 677                 err = -EIO;
 678                 goto put_page_out;
 679         }
 680         if (unlikely(!PageUptodate(fio.encrypted_page))) {
 681                 err = -EIO;
 682                 goto put_page_out;
 683         }
 684
 685         set_page_dirty(fio.encrypted_page);
 686         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 687         if (clear_page_dirty_for_io(fio.encrypted_page))
 688                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 689
 690         set_page_writeback(fio.encrypted_page);
 691
 692         /* allocate block address */
 693         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 694
 695         fio.op = REQ_OP_WRITE;
 696         fio.op_flags = REQ_SYNC;
 697         fio.new_blkaddr = newaddr;
 698         err = f2fs_submit_page_write(&fio);
 699         if (err) {
 700                 if (PageWriteback(fio.encrypted_page))
 701                         end_page_writeback(fio.encrypted_page);
 702                 goto put_page_out;
 703         }
 704
 705         f2fs_update_iostat(fio.sbi, FS_GC_DATA_IO, F2FS_BLKSIZE);
 706
 707         f2fs_update_data_blkaddr(&dn, newaddr);
 708         set_inode_flag(inode, FI_APPEND_WRITE);
 709         if (page->index == 0)
 710                 set_inode_flag(inode, FI_FIRST_BLOCK_WRITTEN);
 711 put_page_out:
 712         f2fs_put_page(fio.encrypted_page, 1);
 713 recover_block:
 714         if (err)
 715                 __f2fs_replace_block(fio.sbi, &sum, newaddr, fio.old_blkaddr,
 716                                                                 true, true);
 717 put_out:
 718         f2fs_put_dnode(&dn);
 719 out:
 720         f2fs_put_page(page, 1);
 721 }
 722
 723 static void move_data_page(struct inode *inode, block_t bidx, int gc_type,
 724                                                         unsigned int segno, int off)
 725 {
 726         struct page *page;
 727
 728         page = get_lock_data_page(inode, bidx, true);
 729         if (IS_ERR(page))
 730                 return;
 731
 732         if (!check_valid_map(F2FS_I_SB(inode), segno, off))
 733                 goto out;
 734
 735         if (f2fs_is_atomic_file(inode))
 736                 goto out;
 737
 738         if (gc_type == BG_GC) {
 739                 if (PageWriteback(page))
 740                         goto out;
 741                 set_page_dirty(page);
 742                 set_cold_data(page);
 743         } else {
 744                 struct f2fs_io_info fio = {
 745                         .sbi = F2FS_I_SB(inode),
 746                         .type = DATA,
 747                         .temp = COLD,
 748                         .op = REQ_OP_WRITE,
 749                         .op_flags = REQ_SYNC,
 750                         .old_blkaddr = NULL_ADDR,
 751                         .page = page,
 752                         .encrypted_page = NULL,
 753                         .need_lock = LOCK_REQ,
 754                         .io_type = FS_GC_DATA_IO,
 755                 };
 756                 bool is_dirty = PageDirty(page);
 757                 int err;
 758
 759 retry:
 760                 set_page_dirty(page);
 761                 f2fs_wait_on_page_writeback(page, DATA, true);
 762                 if (clear_page_dirty_for_io(page)) {
 763                         inode_dec_dirty_pages(inode);
 764                         remove_dirty_inode(inode);
 765                 }
 766
 767                 set_cold_data(page);
 768
 769                 err = do_write_data_page(&fio);
 770                 if (err) {
 771                         clear_cold_data(page);
 772                         if (err == -ENOMEM) {
 773                                 congestion_wait(BLK_RW_ASYNC, HZ/50);
 774                                 goto retry;
 775                         }
 776                         if (is_dirty)
 777                                 set_page_dirty(page);
 778                 }
 779         }
 780 out:
 781         f2fs_put_page(page, 1);
 782 }
 783
 784 /*
 785  * This function tries to get parent node of victim data block, and identifies
 786  * data block validity. If the block is valid, copy that with cold status and
 787  * modify parent node.
 788  * If the parent node is not valid or the data block address is different,
 789  * the victim data block is ignored.
 790  */
 791 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 792                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 793 {
 794         struct super_block *sb = sbi->sb;
 795         struct f2fs_summary *entry;
 796         block_t start_addr;
 797         int off;
 798         int phase = 0;
 799
 800         start_addr = START_BLOCK(sbi, segno);
 801
 802 next_step:
 803         entry = sum;
 804
 805         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 806                 struct page *data_page;
 807                 struct inode *inode;
 808                 struct node_info dni; /* dnode info for the data */
 809                 unsigned int ofs_in_node, nofs;
 810                 block_t start_bidx;
 811                 nid_t nid = le32_to_cpu(entry->nid);
 812
 813                 /* stop BG_GC if there is not enough free sections. */
 814                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
 815                         return;
 816
 817                 if (check_valid_map(sbi, segno, off) == 0)
 818                         continue;
 819
 820                 if (phase == 0) {
 821                         ra_meta_pages(sbi, NAT_BLOCK_OFFSET(nid), 1,
 822                                                         META_NAT, true);
 823                         continue;
 824                 }
 825
 826                 if (phase == 1) {
 827                         ra_node_page(sbi, nid);
 828                         continue;
 829                 }
 830
 831                 /* Get an inode by ino with checking validity */
 832                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 833                         continue;
 834
 835                 if (phase == 2) {
 836                         ra_node_page(sbi, dni.ino);
 837                         continue;
 838                 }
 839
 840                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 841
 842                 if (phase == 3) {
 843                         inode = f2fs_iget(sb, dni.ino);
 844                         if (IS_ERR(inode) || is_bad_inode(inode) ||
 845                                         special_file(inode->i_mode))
 846                                 continue;
 847
 848                         /* if encrypted inode, let's go phase 3 */
 849                         if (f2fs_encrypted_file(inode)) {
 850                                 add_gc_inode(gc_list, inode);
 851                                 continue;
 852                         }
 853
 854                         start_bidx = start_bidx_of_node(nofs, inode);
 855                         data_page = get_read_data_page(inode,
 856                                         start_bidx + ofs_in_node, REQ_RAHEAD,
 857                                         true);
 858                         if (IS_ERR(data_page)) {
 859                                 iput(inode);
 860                                 continue;
 861                         }
 862
 863                         f2fs_put_page(data_page, 0);
 864                         add_gc_inode(gc_list, inode);
 865                         continue;
 866                 }
 867
 868                 /* phase 4 */
 869                 inode = find_gc_inode(gc_list, dni.ino);
 870                 if (inode) {
 871                         struct f2fs_inode_info *fi = F2FS_I(inode);
 872                         bool locked = false;
 873
 874                         if (S_ISREG(inode->i_mode)) {
 875                                 if (!down_write_trylock(&fi->dio_rwsem[READ]))
 876                                         continue;
 877                                 if (!down_write_trylock(
 878                                                 &fi->dio_rwsem[WRITE])) {
 879                                         up_write(&fi->dio_rwsem[READ]);
 880                                         continue;
 881                                 }
 882                                 locked = true;
 883
 884                                 /* wait for all inflight aio data */
 885                                 inode_dio_wait(inode);
 886                         }
 887
 888                         start_bidx = start_bidx_of_node(nofs, inode)
 889                                                                 + ofs_in_node;
 890                         if (f2fs_encrypted_file(inode))
 891                                 move_data_block(inode, start_bidx, segno, off);
 892                         else
 893                                 move_data_page(inode, start_bidx, gc_type,
 894                                                                 segno, off);
 895
 896                         if (locked) {
 897                                 up_write(&fi->dio_rwsem[WRITE]);
 898                                 up_write(&fi->dio_rwsem[READ]);
 899                         }
 900
 901                         stat_inc_data_blk_count(sbi, 1, gc_type);
 902                 }
 903         }
 904
 905         if (++phase < 5)
 906                 goto next_step;
 907 }
 908
 909 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 910                         int gc_type)
 911 {
 912         struct sit_info *sit_i = SIT_I(sbi);
 913         int ret;
 914
 915         mutex_lock(&sit_i->sentry_lock);
 916         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 917                                               NO_CHECK_TYPE, LFS);
 918         mutex_unlock(&sit_i->sentry_lock);
 919         return ret;
 920 }
 921
 922 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 923                                 unsigned int start_segno,
 924                                 struct gc_inode_list *gc_list, int gc_type)
 925 {
 926         struct page *sum_page;
 927         struct f2fs_summary_block *sum;
 928         struct blk_plug plug;
 929         unsigned int segno = start_segno;
 930         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 931         int seg_freed = 0;
 932         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 933                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 934
 935         /* readahead multi ssa blocks those have contiguous address */
 936         if (sbi->segs_per_sec > 1)
 937                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 938                                         sbi->segs_per_sec, META_SSA, true);
 939
 940         /* reference all summary page */
 941         while (segno < end_segno) {
 942                 sum_page = get_sum_page(sbi, segno++);
 943                 unlock_page(sum_page);
 944         }
 945
 946         blk_start_plug(&plug);
 947
 948         for (segno = start_segno; segno < end_segno; segno++) {
 949
 950                 /* find segment summary of victim */
 951                 sum_page = find_get_page(META_MAPPING(sbi),
 952                                         GET_SUM_BLOCK(sbi, segno));
 953                 f2fs_put_page(sum_page, 0);
 954
 955                 if (get_valid_blocks(sbi, segno, false) == 0)
 956                         goto freed;
 957                 if (!PageUptodate(sum_page) || unlikely(f2fs_cp_error(sbi)))
 958                         goto next;
 959
 960                 sum = page_address(sum_page);
 961                 if (type != GET_SUM_TYPE((&sum->footer))) {
 962                         f2fs_msg(sbi->sb, KERN_ERR, "Inconsistent segment (%u) "
 963                                 "type [%d, %d] in SSA and SIT",
 964                                 segno, type, GET_SUM_TYPE((&sum->footer)));
 965                         set_sbi_flag(sbi, SBI_NEED_FSCK);
 966                         goto next;
 967                 }
 968
 969                 /*
 970                  * this is to avoid deadlock:
 971                  * - lock_page(sum_page)         - f2fs_replace_block
 972                  *  - check_valid_map()            - mutex_lock(sentry_lock)
 973                  *   - mutex_lock(sentry_lock)     - change_curseg()
 974                  *                                  - lock_page(sum_page)
 975                  */
 976                 if (type == SUM_TYPE_NODE)
 977                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 978                 else
 979                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 980                                                                 gc_type);
 981
 982                 stat_inc_seg_count(sbi, type, gc_type);
 983
 984 freed:
 985                 if (gc_type == FG_GC &&
 986                                 get_valid_blocks(sbi, segno, false) == 0)
 987                         seg_freed++;
 988 next:
 989                 f2fs_put_page(sum_page, 0);
 990         }
 991
 992         if (gc_type == FG_GC)
 993                 f2fs_submit_merged_write(sbi,
 994                                 (type == SUM_TYPE_NODE) ? NODE : DATA);
 995
 996         blk_finish_plug(&plug);
 997
 998         stat_inc_call_count(sbi->stat_info);
 999
1000         return seg_freed;
1001 }
1002
1003 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
1004                         bool background, unsigned int segno)
1005 {
1006         int gc_type = sync ? FG_GC : BG_GC;
1007         int sec_freed = 0, seg_freed = 0, total_freed = 0;
1008         int ret = 0;
1009         struct cp_control cpc;
1010         unsigned int init_segno = segno;
1011         struct gc_inode_list gc_list = {
1012                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
1013                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
1014         };
1015
1016         trace_f2fs_gc_begin(sbi->sb, sync, background,
1017                                 get_pages(sbi, F2FS_DIRTY_NODES),
1018                                 get_pages(sbi, F2FS_DIRTY_DENTS),
1019                                 get_pages(sbi, F2FS_DIRTY_IMETA),
1020                                 free_sections(sbi),
1021                                 free_segments(sbi),
1022                                 reserved_segments(sbi),
1023                                 prefree_segments(sbi));
1024
1025         cpc.reason = __get_cp_reason(sbi);
1026 gc_more:
1027         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE))) {
1028                 ret = -EINVAL;
1029                 goto stop;
1030         }
1031         if (unlikely(f2fs_cp_error(sbi))) {
1032                 ret = -EIO;
1033                 goto stop;
1034         }
1035
1036         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
1037                 /*
1038                  * For example, if there are many prefree_segments below given
1039                  * threshold, we can make them free by checkpoint. Then, we
1040                  * secure free segments which doesn't need fggc any more.
1041                  */
1042                 if (prefree_segments(sbi)) {
1043                         ret = write_checkpoint(sbi, &cpc);
1044                         if (ret)
1045                                 goto stop;
1046                 }
1047                 if (has_not_enough_free_secs(sbi, 0, 0))
1048                         gc_type = FG_GC;
1049         }
1050
1051         /* f2fs_balance_fs doesn't need to do BG_GC in critical path. */
1052         if (gc_type == BG_GC && !background) {
1053                 ret = -EINVAL;
1054                 goto stop;
1055         }
1056         if (!__get_victim(sbi, &segno, gc_type)) {
1057                 ret = -ENODATA;
1058                 goto stop;
1059         }
1060
1061         seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
1062         if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
1063                 sec_freed++;
1064         total_freed += seg_freed;
1065
1066         if (gc_type == FG_GC)
1067                 sbi->cur_victim_sec = NULL_SEGNO;
1068
1069         if (!sync) {
1070                 if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
1071                         segno = NULL_SEGNO;
1072                         goto gc_more;
1073                 }
1074
1075                 if (gc_type == FG_GC)
1076                         ret = write_checkpoint(sbi, &cpc);
1077         }
1078 stop:
1079         SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
1080         SIT_I(sbi)->last_victim[FLUSH_DEVICE] = init_segno;
1081
1082         trace_f2fs_gc_end(sbi->sb, ret, total_freed, sec_freed,
1083                                 get_pages(sbi, F2FS_DIRTY_NODES),
1084                                 get_pages(sbi, F2FS_DIRTY_DENTS),
1085                                 get_pages(sbi, F2FS_DIRTY_IMETA),
1086                                 free_sections(sbi),
1087                                 free_segments(sbi),
1088                                 reserved_segments(sbi),
1089                                 prefree_segments(sbi));
1090
1091         mutex_unlock(&sbi->gc_mutex);
1092
1093         put_gc_inode(&gc_list);
1094
1095         if (sync && !ret)
1096                 ret = sec_freed ? 0 : -EAGAIN;
1097         return ret;
1098 }
1099
1100 void build_gc_manager(struct f2fs_sb_info *sbi)
1101 {
1102         u64 main_count, resv_count, ovp_count;
1103
1104         DIRTY_I(sbi)->v_ops = &default_v_ops;
1105
1106         /* threshold of # of valid blocks in a section for victims of FG_GC */
1107         main_count = SM_I(sbi)->main_segments << sbi->log_blocks_per_seg;
1108         resv_count = SM_I(sbi)->reserved_segments << sbi->log_blocks_per_seg;
1109         ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
1110
1111         sbi->fggc_threshold = div64_u64((main_count - ovp_count) *
1112                                 BLKS_PER_SEC(sbi), (main_count - resv_count));
1113
1114         /* give warm/cold data area from slower device */
1115         if (f2fs_is_multi_device(sbi) && sbi->segs_per_sec == 1)
1116                 SIT_I(sbi)->last_victim[ALLOC_NEXT] =
1117                                 GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
1118 }