fs/f2fs/gc.c

   1 /*
   2  * fs/f2fs/gc.c
   3  *
   4  * Copyright (c) 2012 Samsung Electronics Co., Ltd.
   5  *             http://www.samsung.com/
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  */
  11 #include <linux/fs.h>
  12 #include <linux/module.h>
  13 #include <linux/backing-dev.h>
  14 #include <linux/init.h>
  15 #include <linux/f2fs_fs.h>
  16 #include <linux/kthread.h>
  17 #include <linux/delay.h>
  18 #include <linux/freezer.h>
  19
  20 #include "f2fs.h"
  21 #include "node.h"
  22 #include "segment.h"
  23 #include "gc.h"
  24 #include <trace/events/f2fs.h>
  25
  26 static int gc_thread_func(void *data)
  27 {
  28         struct f2fs_sb_info *sbi = data;
  29         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
  30         wait_queue_head_t *wq = &sbi->gc_thread->gc_wait_queue_head;
  31         long wait_ms;
  32
  33         wait_ms = gc_th->min_sleep_time;
  34
  35         do {
  36                 if (try_to_freeze())
  37                         continue;
  38                 else
  39                         wait_event_interruptible_timeout(*wq,
  40                                                 kthread_should_stop(),
  41                                                 msecs_to_jiffies(wait_ms));
  42                 if (kthread_should_stop())
  43                         break;
  44
  45                 if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
  46                         increase_sleep_time(gc_th, &wait_ms);
  47                         continue;
  48                 }
  49
  50                 /*
  51                  * [GC triggering condition]
  52                  * 0. GC is not conducted currently.
  53                  * 1. There are enough dirty segments.
  54                  * 2. IO subsystem is idle by checking the # of writeback pages.
  55                  * 3. IO subsystem is idle by checking the # of requests in
  56                  *    bdev's request list.
  57                  *
  58                  * Note) We have to avoid triggering GCs frequently.
  59                  * Because it is possible that some segments can be
  60                  * invalidated soon after by user update or deletion.
  61                  * So, I'd like to wait some time to collect dirty segments.
  62                  */
  63                 if (!mutex_trylock(&sbi->gc_mutex))
  64                         continue;
  65
  66                 if (!is_idle(sbi)) {
  67                         increase_sleep_time(gc_th, &wait_ms);
  68                         mutex_unlock(&sbi->gc_mutex);
  69                         continue;
  70                 }
  71
  72                 if (has_enough_invalid_blocks(sbi))
  73                         decrease_sleep_time(gc_th, &wait_ms);
  74                 else
  75                         increase_sleep_time(gc_th, &wait_ms);
  76
  77                 stat_inc_bggc_count(sbi);
  78
  79                 /* if return value is not zero, no victim was selected */
  80                 if (f2fs_gc(sbi, test_opt(sbi, FORCE_FG_GC)))
  81                         wait_ms = gc_th->no_gc_sleep_time;
  82
  83                 trace_f2fs_background_gc(sbi->sb, wait_ms,
  84                                 prefree_segments(sbi), free_segments(sbi));
  85
  86                 /* balancing f2fs's metadata periodically */
  87                 f2fs_balance_fs_bg(sbi);
  88
  89         } while (!kthread_should_stop());
  90         return 0;
  91 }
  92
  93 int start_gc_thread(struct f2fs_sb_info *sbi)
  94 {
  95         struct f2fs_gc_kthread *gc_th;
  96         dev_t dev = sbi->sb->s_bdev->bd_dev;
  97         int err = 0;
  98
  99         gc_th = kmalloc(sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
 100         if (!gc_th) {
 101                 err = -ENOMEM;
 102                 goto out;
 103         }
 104
 105         gc_th->min_sleep_time = DEF_GC_THREAD_MIN_SLEEP_TIME;
 106         gc_th->max_sleep_time = DEF_GC_THREAD_MAX_SLEEP_TIME;
 107         gc_th->no_gc_sleep_time = DEF_GC_THREAD_NOGC_SLEEP_TIME;
 108
 109         gc_th->gc_idle = 0;
 110
 111         sbi->gc_thread = gc_th;
 112         init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
 113         sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
 114                         "f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
 115         if (IS_ERR(gc_th->f2fs_gc_task)) {
 116                 err = PTR_ERR(gc_th->f2fs_gc_task);
 117                 kfree(gc_th);
 118                 sbi->gc_thread = NULL;
 119         }
 120 out:
 121         return err;
 122 }
 123
 124 void stop_gc_thread(struct f2fs_sb_info *sbi)
 125 {
 126         struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
 127         if (!gc_th)
 128                 return;
 129         kthread_stop(gc_th->f2fs_gc_task);
 130         kfree(gc_th);
 131         sbi->gc_thread = NULL;
 132 }
 133
 134 static int select_gc_type(struct f2fs_gc_kthread *gc_th, int gc_type)
 135 {
 136         int gc_mode = (gc_type == BG_GC) ? GC_CB : GC_GREEDY;
 137
 138         if (gc_th && gc_th->gc_idle) {
 139                 if (gc_th->gc_idle == 1)
 140                         gc_mode = GC_CB;
 141                 else if (gc_th->gc_idle == 2)
 142                         gc_mode = GC_GREEDY;
 143         }
 144         return gc_mode;
 145 }
 146
 147 static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 148                         int type, struct victim_sel_policy *p)
 149 {
 150         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 151
 152         if (p->alloc_mode == SSR) {
 153                 p->gc_mode = GC_GREEDY;
 154                 p->dirty_segmap = dirty_i->dirty_segmap[type];
 155                 p->max_search = dirty_i->nr_dirty[type];
 156                 p->ofs_unit = 1;
 157         } else {
 158                 p->gc_mode = select_gc_type(sbi->gc_thread, gc_type);
 159                 p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
 160                 p->max_search = dirty_i->nr_dirty[DIRTY];
 161                 p->ofs_unit = sbi->segs_per_sec;
 162         }
 163
 164         if (p->max_search > sbi->max_victim_search)
 165                 p->max_search = sbi->max_victim_search;
 166
 167         p->offset = sbi->last_victim[p->gc_mode];
 168 }
 169
 170 static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
 171                                 struct victim_sel_policy *p)
 172 {
 173         /* SSR allocates in a segment unit */
 174         if (p->alloc_mode == SSR)
 175                 return sbi->blocks_per_seg;
 176         if (p->gc_mode == GC_GREEDY)
 177                 return sbi->blocks_per_seg * p->ofs_unit;
 178         else if (p->gc_mode == GC_CB)
 179                 return UINT_MAX;
 180         else /* No other gc_mode */
 181                 return 0;
 182 }
 183
 184 static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 185 {
 186         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 187         unsigned int secno;
 188
 189         /*
 190          * If the gc_type is FG_GC, we can select victim segments
 191          * selected by background GC before.
 192          * Those segments guarantee they have small valid blocks.
 193          */
 194         for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 195                 if (sec_usage_check(sbi, secno))
 196                         continue;
 197                 clear_bit(secno, dirty_i->victim_secmap);
 198                 return secno * sbi->segs_per_sec;
 199         }
 200         return NULL_SEGNO;
 201 }
 202
 203 static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
 204 {
 205         struct sit_info *sit_i = SIT_I(sbi);
 206         unsigned int secno = GET_SECNO(sbi, segno);
 207         unsigned int start = secno * sbi->segs_per_sec;
 208         unsigned long long mtime = 0;
 209         unsigned int vblocks;
 210         unsigned char age = 0;
 211         unsigned char u;
 212         unsigned int i;
 213
 214         for (i = 0; i < sbi->segs_per_sec; i++)
 215                 mtime += get_seg_entry(sbi, start + i)->mtime;
 216         vblocks = get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 217
 218         mtime = div_u64(mtime, sbi->segs_per_sec);
 219         vblocks = div_u64(vblocks, sbi->segs_per_sec);
 220
 221         u = (vblocks * 100) >> sbi->log_blocks_per_seg;
 222
 223         /* Handle if the system time has changed by the user */
 224         if (mtime < sit_i->min_mtime)
 225                 sit_i->min_mtime = mtime;
 226         if (mtime > sit_i->max_mtime)
 227                 sit_i->max_mtime = mtime;
 228         if (sit_i->max_mtime != sit_i->min_mtime)
 229                 age = 100 - div64_u64(100 * (mtime - sit_i->min_mtime),
 230                                 sit_i->max_mtime - sit_i->min_mtime);
 231
 232         return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
 233 }
 234
 235 static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
 236                         unsigned int segno, struct victim_sel_policy *p)
 237 {
 238         if (p->alloc_mode == SSR)
 239                 return get_seg_entry(sbi, segno)->ckpt_valid_blocks;
 240
 241         /* alloc_mode == LFS */
 242         if (p->gc_mode == GC_GREEDY)
 243                 return get_valid_blocks(sbi, segno, sbi->segs_per_sec);
 244         else
 245                 return get_cb_cost(sbi, segno);
 246 }
 247
 248 static unsigned int count_bits(const unsigned long *addr,
 249                                 unsigned int offset, unsigned int len)
 250 {
 251         unsigned int end = offset + len, sum = 0;
 252
 253         while (offset < end) {
 254                 if (test_bit(offset++, addr))
 255                         ++sum;
 256         }
 257         return sum;
 258 }
 259
 260 /*
 261  * This function is called from two paths.
 262  * One is garbage collection and the other is SSR segment selection.
 263  * When it is called during GC, it just gets a victim segment
 264  * and it does not remove it from dirty seglist.
 265  * When it is called from SSR segment selection, it finds a segment
 266  * which has minimum valid blocks and removes it from dirty seglist.
 267  */
 268 static int get_victim_by_default(struct f2fs_sb_info *sbi,
 269                 unsigned int *result, int gc_type, int type, char alloc_mode)
 270 {
 271         struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
 272         struct victim_sel_policy p;
 273         unsigned int secno, max_cost, last_victim;
 274         unsigned int last_segment = MAIN_SEGS(sbi);
 275         unsigned int nsearched = 0;
 276
 277         mutex_lock(&dirty_i->seglist_lock);
 278
 279         p.alloc_mode = alloc_mode;
 280         select_policy(sbi, gc_type, type, &p);
 281
 282         p.min_segno = NULL_SEGNO;
 283         p.min_cost = max_cost = get_max_cost(sbi, &p);
 284
 285         if (p.max_search == 0)
 286                 goto out;
 287
 288         last_victim = sbi->last_victim[p.gc_mode];
 289         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 290                 p.min_segno = check_bg_victims(sbi);
 291                 if (p.min_segno != NULL_SEGNO)
 292                         goto got_it;
 293         }
 294
 295         while (1) {
 296                 unsigned long cost;
 297                 unsigned int segno;
 298
 299                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 300                 if (segno >= last_segment) {
 301                         if (sbi->last_victim[p.gc_mode]) {
 302                                 last_segment = sbi->last_victim[p.gc_mode];
 303                                 sbi->last_victim[p.gc_mode] = 0;
 304                                 p.offset = 0;
 305                                 continue;
 306                         }
 307                         break;
 308                 }
 309
 310                 p.offset = segno + p.ofs_unit;
 311                 if (p.ofs_unit > 1) {
 312                         p.offset -= segno % p.ofs_unit;
 313                         nsearched += count_bits(p.dirty_segmap,
 314                                                 p.offset - p.ofs_unit,
 315                                                 p.ofs_unit);
 316                 } else {
 317                         nsearched++;
 318                 }
 319
 320
 321                 secno = GET_SECNO(sbi, segno);
 322
 323                 if (sec_usage_check(sbi, secno))
 324                         goto next;
 325                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 326                         goto next;
 327
 328                 cost = get_gc_cost(sbi, segno, &p);
 329
 330                 if (p.min_cost > cost) {
 331                         p.min_segno = segno;
 332                         p.min_cost = cost;
 333                 }
 334 next:
 335                 if (nsearched >= p.max_search) {
 336                         if (!sbi->last_victim[p.gc_mode] && segno <= last_victim)
 337                                 sbi->last_victim[p.gc_mode] = last_victim + 1;
 338                         else
 339                                 sbi->last_victim[p.gc_mode] = segno + 1;
 340                         break;
 341                 }
 342         }
 343         if (p.min_segno != NULL_SEGNO) {
 344 got_it:
 345                 if (p.alloc_mode == LFS) {
 346                         secno = GET_SECNO(sbi, p.min_segno);
 347                         if (gc_type == FG_GC)
 348                                 sbi->cur_victim_sec = secno;
 349                         else
 350                                 set_bit(secno, dirty_i->victim_secmap);
 351                 }
 352                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 353
 354                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 355                                 sbi->cur_victim_sec,
 356                                 prefree_segments(sbi), free_segments(sbi));
 357         }
 358 out:
 359         mutex_unlock(&dirty_i->seglist_lock);
 360
 361         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 362 }
 363
 364 static const struct victim_selection default_v_ops = {
 365         .get_victim = get_victim_by_default,
 366 };
 367
 368 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 369 {
 370         struct inode_entry *ie;
 371
 372         ie = radix_tree_lookup(&gc_list->iroot, ino);
 373         if (ie)
 374                 return ie->inode;
 375         return NULL;
 376 }
 377
 378 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 379 {
 380         struct inode_entry *new_ie;
 381
 382         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 383                 iput(inode);
 384                 return;
 385         }
 386         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 387         new_ie->inode = inode;
 388
 389         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 390         list_add_tail(&new_ie->list, &gc_list->ilist);
 391 }
 392
 393 static void put_gc_inode(struct gc_inode_list *gc_list)
 394 {
 395         struct inode_entry *ie, *next_ie;
 396         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 397                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 398                 iput(ie->inode);
 399                 list_del(&ie->list);
 400                 kmem_cache_free(inode_entry_slab, ie);
 401         }
 402 }
 403
 404 static int check_valid_map(struct f2fs_sb_info *sbi,
 405                                 unsigned int segno, int offset)
 406 {
 407         struct sit_info *sit_i = SIT_I(sbi);
 408         struct seg_entry *sentry;
 409         int ret;
 410
 411         mutex_lock(&sit_i->sentry_lock);
 412         sentry = get_seg_entry(sbi, segno);
 413         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 414         mutex_unlock(&sit_i->sentry_lock);
 415         return ret;
 416 }
 417
 418 /*
 419  * This function compares node address got in summary with that in NAT.
 420  * On validity, copy that node with cold status, otherwise (invalid node)
 421  * ignore that.
 422  */
 423 static void gc_node_segment(struct f2fs_sb_info *sbi,
 424                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 425 {
 426         bool initial = true;
 427         struct f2fs_summary *entry;
 428         block_t start_addr;
 429         int off;
 430
 431         start_addr = START_BLOCK(sbi, segno);
 432
 433 next_step:
 434         entry = sum;
 435
 436         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 437                 nid_t nid = le32_to_cpu(entry->nid);
 438                 struct page *node_page;
 439                 struct node_info ni;
 440
 441                 /* stop BG_GC if there is not enough free sections. */
 442                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 443                         return;
 444
 445                 if (check_valid_map(sbi, segno, off) == 0)
 446                         continue;
 447
 448                 if (initial) {
 449                         ra_node_page(sbi, nid);
 450                         continue;
 451                 }
 452                 node_page = get_node_page(sbi, nid);
 453                 if (IS_ERR(node_page))
 454                         continue;
 455
 456                 /* block may become invalid during get_node_page */
 457                 if (check_valid_map(sbi, segno, off) == 0) {
 458                         f2fs_put_page(node_page, 1);
 459                         continue;
 460                 }
 461
 462                 get_node_info(sbi, nid, &ni);
 463                 if (ni.blk_addr != start_addr + off) {
 464                         f2fs_put_page(node_page, 1);
 465                         continue;
 466                 }
 467
 468                 /* set page dirty and write it */
 469                 if (gc_type == FG_GC) {
 470                         f2fs_wait_on_page_writeback(node_page, NODE, true);
 471                         set_page_dirty(node_page);
 472                 } else {
 473                         if (!PageWriteback(node_page))
 474                                 set_page_dirty(node_page);
 475                 }
 476                 f2fs_put_page(node_page, 1);
 477                 stat_inc_node_blk_count(sbi, 1, gc_type);
 478         }
 479
 480         if (initial) {
 481                 initial = false;
 482                 goto next_step;
 483         }
 484 }
 485
 486 /*
 487  * Calculate start block index indicating the given node offset.
 488  * Be careful, caller should give this node offset only indicating direct node
 489  * blocks. If any node offsets, which point the other types of node blocks such
 490  * as indirect or double indirect node blocks, are given, it must be a caller's
 491  * bug.
 492  */
 493 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 494 {
 495         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 496         unsigned int bidx;
 497
 498         if (node_ofs == 0)
 499                 return 0;
 500
 501         if (node_ofs <= 2) {
 502                 bidx = node_ofs - 1;
 503         } else if (node_ofs <= indirect_blks) {
 504                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 505                 bidx = node_ofs - 2 - dec;
 506         } else {
 507                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 508                 bidx = node_ofs - 5 - dec;
 509         }
 510         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 511 }
 512
 513 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 514                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 515 {
 516         struct page *node_page;
 517         nid_t nid;
 518         unsigned int ofs_in_node;
 519         block_t source_blkaddr;
 520
 521         nid = le32_to_cpu(sum->nid);
 522         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 523
 524         node_page = get_node_page(sbi, nid);
 525         if (IS_ERR(node_page))
 526                 return false;
 527
 528         get_node_info(sbi, nid, dni);
 529
 530         if (sum->version != dni->version) {
 531                 f2fs_put_page(node_page, 1);
 532                 return false;
 533         }
 534
 535         *nofs = ofs_of_node(node_page);
 536         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 537         f2fs_put_page(node_page, 1);
 538
 539         if (source_blkaddr != blkaddr)
 540                 return false;
 541         return true;
 542 }
 543
 544 static void move_encrypted_block(struct inode *inode, block_t bidx)
 545 {
 546         struct f2fs_io_info fio = {
 547                 .sbi = F2FS_I_SB(inode),
 548                 .type = DATA,
 549                 .rw = READ_SYNC,
 550                 .encrypted_page = NULL,
 551         };
 552         struct dnode_of_data dn;
 553         struct f2fs_summary sum;
 554         struct node_info ni;
 555         struct page *page;
 556         int err;
 557
 558         /* do not read out */
 559         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 560         if (!page)
 561                 return;
 562
 563         set_new_dnode(&dn, inode, NULL, NULL, 0);
 564         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 565         if (err)
 566                 goto out;
 567
 568         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 569                 ClearPageUptodate(page);
 570                 goto put_out;
 571         }
 572
 573         /*
 574          * don't cache encrypted data into meta inode until previous dirty
 575          * data were writebacked to avoid racing between GC and flush.
 576          */
 577         f2fs_wait_on_page_writeback(page, DATA, true);
 578
 579         get_node_info(fio.sbi, dn.nid, &ni);
 580         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 581
 582         /* read page */
 583         fio.page = page;
 584         fio.blk_addr = dn.data_blkaddr;
 585
 586         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
 587                                         fio.blk_addr,
 588                                         FGP_LOCK|FGP_CREAT,
 589                                         GFP_NOFS);
 590         if (!fio.encrypted_page)
 591                 goto put_out;
 592
 593         err = f2fs_submit_page_bio(&fio);
 594         if (err)
 595                 goto put_page_out;
 596
 597         /* write page */
 598         lock_page(fio.encrypted_page);
 599
 600         if (unlikely(!PageUptodate(fio.encrypted_page)))
 601                 goto put_page_out;
 602         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
 603                 goto put_page_out;
 604
 605         set_page_dirty(fio.encrypted_page);
 606         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 607         if (clear_page_dirty_for_io(fio.encrypted_page))
 608                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 609
 610         set_page_writeback(fio.encrypted_page);
 611
 612         /* allocate block address */
 613         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 614         allocate_data_block(fio.sbi, NULL, fio.blk_addr,
 615                                         &fio.blk_addr, &sum, CURSEG_COLD_DATA);
 616         fio.rw = WRITE_SYNC;
 617         f2fs_submit_page_mbio(&fio);
 618
 619         dn.data_blkaddr = fio.blk_addr;
 620         set_data_blkaddr(&dn);
 621         f2fs_update_extent_cache(&dn);
 622         set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
 623         if (page->index == 0)
 624                 set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
 625 put_page_out:
 626         f2fs_put_page(fio.encrypted_page, 1);
 627 put_out:
 628         f2fs_put_dnode(&dn);
 629 out:
 630         f2fs_put_page(page, 1);
 631 }
 632
 633 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
 634 {
 635         struct page *page;
 636
 637         page = get_lock_data_page(inode, bidx, true);
 638         if (IS_ERR(page))
 639                 return;
 640
 641         if (gc_type == BG_GC) {
 642                 if (PageWriteback(page))
 643                         goto out;
 644                 set_page_dirty(page);
 645                 set_cold_data(page);
 646         } else {
 647                 struct f2fs_io_info fio = {
 648                         .sbi = F2FS_I_SB(inode),
 649                         .type = DATA,
 650                         .rw = WRITE_SYNC,
 651                         .page = page,
 652                         .encrypted_page = NULL,
 653                 };
 654                 set_page_dirty(page);
 655                 f2fs_wait_on_page_writeback(page, DATA, true);
 656                 if (clear_page_dirty_for_io(page))
 657                         inode_dec_dirty_pages(inode);
 658                 set_cold_data(page);
 659                 do_write_data_page(&fio);
 660                 clear_cold_data(page);
 661         }
 662 out:
 663         f2fs_put_page(page, 1);
 664 }
 665
 666 /*
 667  * This function tries to get parent node of victim data block, and identifies
 668  * data block validity. If the block is valid, copy that with cold status and
 669  * modify parent node.
 670  * If the parent node is not valid or the data block address is different,
 671  * the victim data block is ignored.
 672  */
 673 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 674                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 675 {
 676         struct super_block *sb = sbi->sb;
 677         struct f2fs_summary *entry;
 678         block_t start_addr;
 679         int off;
 680         int phase = 0;
 681
 682         start_addr = START_BLOCK(sbi, segno);
 683
 684 next_step:
 685         entry = sum;
 686
 687         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 688                 struct page *data_page;
 689                 struct inode *inode;
 690                 struct node_info dni; /* dnode info for the data */
 691                 unsigned int ofs_in_node, nofs;
 692                 block_t start_bidx;
 693
 694                 /* stop BG_GC if there is not enough free sections. */
 695                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 696                         return;
 697
 698                 if (check_valid_map(sbi, segno, off) == 0)
 699                         continue;
 700
 701                 if (phase == 0) {
 702                         ra_node_page(sbi, le32_to_cpu(entry->nid));
 703                         continue;
 704                 }
 705
 706                 /* Get an inode by ino with checking validity */
 707                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 708                         continue;
 709
 710                 if (phase == 1) {
 711                         ra_node_page(sbi, dni.ino);
 712                         continue;
 713                 }
 714
 715                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 716
 717                 if (phase == 2) {
 718                         inode = f2fs_iget(sb, dni.ino);
 719                         if (IS_ERR(inode) || is_bad_inode(inode))
 720                                 continue;
 721
 722                         /* if encrypted inode, let's go phase 3 */
 723                         if (f2fs_encrypted_inode(inode) &&
 724                                                 S_ISREG(inode->i_mode)) {
 725                                 add_gc_inode(gc_list, inode);
 726                                 continue;
 727                         }
 728
 729                         start_bidx = start_bidx_of_node(nofs, inode);
 730                         data_page = get_read_data_page(inode,
 731                                         start_bidx + ofs_in_node, READA, true);
 732                         if (IS_ERR(data_page)) {
 733                                 iput(inode);
 734                                 continue;
 735                         }
 736
 737                         f2fs_put_page(data_page, 0);
 738                         add_gc_inode(gc_list, inode);
 739                         continue;
 740                 }
 741
 742                 /* phase 3 */
 743                 inode = find_gc_inode(gc_list, dni.ino);
 744                 if (inode) {
 745                         start_bidx = start_bidx_of_node(nofs, inode)
 746                                                                 + ofs_in_node;
 747                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 748                                 move_encrypted_block(inode, start_bidx);
 749                         else
 750                                 move_data_page(inode, start_bidx, gc_type);
 751                         stat_inc_data_blk_count(sbi, 1, gc_type);
 752                 }
 753         }
 754
 755         if (++phase < 4)
 756                 goto next_step;
 757 }
 758
 759 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 760                         int gc_type)
 761 {
 762         struct sit_info *sit_i = SIT_I(sbi);
 763         int ret;
 764
 765         mutex_lock(&sit_i->sentry_lock);
 766         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 767                                               NO_CHECK_TYPE, LFS);
 768         mutex_unlock(&sit_i->sentry_lock);
 769         return ret;
 770 }
 771
 772 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 773                                 unsigned int start_segno,
 774                                 struct gc_inode_list *gc_list, int gc_type)
 775 {
 776         struct page *sum_page;
 777         struct f2fs_summary_block *sum;
 778         struct blk_plug plug;
 779         unsigned int segno = start_segno;
 780         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 781         int seg_freed = 0;
 782         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 783                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 784
 785         /* readahead multi ssa blocks those have contiguous address */
 786         if (sbi->segs_per_sec > 1)
 787                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 788                                         sbi->segs_per_sec, META_SSA, true);
 789
 790         /* reference all summary page */
 791         while (segno < end_segno) {
 792                 sum_page = get_sum_page(sbi, segno++);
 793                 unlock_page(sum_page);
 794         }
 795
 796         blk_start_plug(&plug);
 797
 798         for (segno = start_segno; segno < end_segno; segno++) {
 799                 /* find segment summary of victim */
 800                 sum_page = find_get_page(META_MAPPING(sbi),
 801                                         GET_SUM_BLOCK(sbi, segno));
 802                 f2fs_bug_on(sbi, !PageUptodate(sum_page));
 803                 f2fs_put_page(sum_page, 0);
 804
 805                 sum = page_address(sum_page);
 806                 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
 807
 808                 /*
 809                  * this is to avoid deadlock:
 810                  * - lock_page(sum_page)         - f2fs_replace_block
 811                  *  - check_valid_map()            - mutex_lock(sentry_lock)
 812                  *   - mutex_lock(sentry_lock)     - change_curseg()
 813                  *                                  - lock_page(sum_page)
 814                  */
 815
 816                 if (type == SUM_TYPE_NODE)
 817                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 818                 else
 819                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 820                                                                 gc_type);
 821
 822                 stat_inc_seg_count(sbi, type, gc_type);
 823
 824                 f2fs_put_page(sum_page, 0);
 825         }
 826
 827         if (gc_type == FG_GC) {
 828                 if (type == SUM_TYPE_NODE) {
 829                         struct writeback_control wbc = {
 830                                 .sync_mode = WB_SYNC_ALL,
 831                                 .nr_to_write = LONG_MAX,
 832                                 .for_reclaim = 0,
 833                         };
 834                         sync_node_pages(sbi, 0, &wbc);
 835                 } else {
 836                         f2fs_submit_merged_bio(sbi, DATA, WRITE);
 837                 }
 838         }
 839
 840         blk_finish_plug(&plug);
 841
 842         if (gc_type == FG_GC) {
 843                 while (start_segno < end_segno)
 844                         if (get_valid_blocks(sbi, start_segno++, 1) == 0)
 845                                 seg_freed++;
 846         }
 847
 848         stat_inc_call_count(sbi->stat_info);
 849
 850         return seg_freed;
 851 }
 852
 853 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync)
 854 {
 855         unsigned int segno;
 856         int gc_type = sync ? FG_GC : BG_GC;
 857         int sec_freed = 0, seg_freed;
 858         int ret = -EINVAL;
 859         struct cp_control cpc;
 860         struct gc_inode_list gc_list = {
 861                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 862                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 863         };
 864
 865         cpc.reason = __get_cp_reason(sbi);
 866 gc_more:
 867         segno = NULL_SEGNO;
 868
 869         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 870                 goto stop;
 871         if (unlikely(f2fs_cp_error(sbi))) {
 872                 ret = -EIO;
 873                 goto stop;
 874         }
 875
 876         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed)) {
 877                 gc_type = FG_GC;
 878                 /*
 879                  * If there is no victim and no prefree segment but still not
 880                  * enough free sections, we should flush dent/node blocks and do
 881                  * garbage collections.
 882                  */
 883                 if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
 884                         write_checkpoint(sbi, &cpc);
 885                 else if (has_not_enough_free_secs(sbi, 0))
 886                         write_checkpoint(sbi, &cpc);
 887         }
 888
 889         if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
 890                 goto stop;
 891         ret = 0;
 892
 893         seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
 894
 895         if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
 896                 sec_freed++;
 897
 898         if (gc_type == FG_GC)
 899                 sbi->cur_victim_sec = NULL_SEGNO;
 900
 901         if (!sync) {
 902                 if (has_not_enough_free_secs(sbi, sec_freed))
 903                         goto gc_more;
 904
 905                 if (gc_type == FG_GC)
 906                         write_checkpoint(sbi, &cpc);
 907         }
 908 stop:
 909         mutex_unlock(&sbi->gc_mutex);
 910
 911         put_gc_inode(&gc_list);
 912
 913         if (sync)
 914                 ret = sec_freed ? 0 : -EAGAIN;
 915         return ret;
 916 }
 917
 918 void build_gc_manager(struct f2fs_sb_info *sbi)
 919 {
 920         DIRTY_I(sbi)->v_ops = &default_v_ops;
 921 }