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;
 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         if (p.alloc_mode == LFS && gc_type == FG_GC) {
 289                 p.min_segno = check_bg_victims(sbi);
 290                 if (p.min_segno != NULL_SEGNO)
 291                         goto got_it;
 292         }
 293
 294         while (1) {
 295                 unsigned long cost;
 296                 unsigned int segno;
 297
 298                 segno = find_next_bit(p.dirty_segmap, last_segment, p.offset);
 299                 if (segno >= last_segment) {
 300                         if (sbi->last_victim[p.gc_mode]) {
 301                                 last_segment = sbi->last_victim[p.gc_mode];
 302                                 sbi->last_victim[p.gc_mode] = 0;
 303                                 p.offset = 0;
 304                                 continue;
 305                         }
 306                         break;
 307                 }
 308
 309                 p.offset = segno + p.ofs_unit;
 310                 if (p.ofs_unit > 1) {
 311                         p.offset -= segno % p.ofs_unit;
 312                         nsearched += count_bits(p.dirty_segmap,
 313                                                 p.offset - p.ofs_unit,
 314                                                 p.ofs_unit);
 315                 } else {
 316                         nsearched++;
 317                 }
 318
 319
 320                 secno = GET_SECNO(sbi, segno);
 321
 322                 if (sec_usage_check(sbi, secno))
 323                         goto next;
 324                 if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 325                         goto next;
 326
 327                 cost = get_gc_cost(sbi, segno, &p);
 328
 329                 if (p.min_cost > cost) {
 330                         p.min_segno = segno;
 331                         p.min_cost = cost;
 332                 }
 333 next:
 334                 if (nsearched >= p.max_search) {
 335                         sbi->last_victim[p.gc_mode] = segno;
 336                         break;
 337                 }
 338         }
 339         if (p.min_segno != NULL_SEGNO) {
 340 got_it:
 341                 if (p.alloc_mode == LFS) {
 342                         secno = GET_SECNO(sbi, p.min_segno);
 343                         if (gc_type == FG_GC)
 344                                 sbi->cur_victim_sec = secno;
 345                         else
 346                                 set_bit(secno, dirty_i->victim_secmap);
 347                 }
 348                 *result = (p.min_segno / p.ofs_unit) * p.ofs_unit;
 349
 350                 trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
 351                                 sbi->cur_victim_sec,
 352                                 prefree_segments(sbi), free_segments(sbi));
 353         }
 354 out:
 355         mutex_unlock(&dirty_i->seglist_lock);
 356
 357         return (p.min_segno == NULL_SEGNO) ? 0 : 1;
 358 }
 359
 360 static const struct victim_selection default_v_ops = {
 361         .get_victim = get_victim_by_default,
 362 };
 363
 364 static struct inode *find_gc_inode(struct gc_inode_list *gc_list, nid_t ino)
 365 {
 366         struct inode_entry *ie;
 367
 368         ie = radix_tree_lookup(&gc_list->iroot, ino);
 369         if (ie)
 370                 return ie->inode;
 371         return NULL;
 372 }
 373
 374 static void add_gc_inode(struct gc_inode_list *gc_list, struct inode *inode)
 375 {
 376         struct inode_entry *new_ie;
 377
 378         if (inode == find_gc_inode(gc_list, inode->i_ino)) {
 379                 iput(inode);
 380                 return;
 381         }
 382         new_ie = f2fs_kmem_cache_alloc(inode_entry_slab, GFP_NOFS);
 383         new_ie->inode = inode;
 384
 385         f2fs_radix_tree_insert(&gc_list->iroot, inode->i_ino, new_ie);
 386         list_add_tail(&new_ie->list, &gc_list->ilist);
 387 }
 388
 389 static void put_gc_inode(struct gc_inode_list *gc_list)
 390 {
 391         struct inode_entry *ie, *next_ie;
 392         list_for_each_entry_safe(ie, next_ie, &gc_list->ilist, list) {
 393                 radix_tree_delete(&gc_list->iroot, ie->inode->i_ino);
 394                 iput(ie->inode);
 395                 list_del(&ie->list);
 396                 kmem_cache_free(inode_entry_slab, ie);
 397         }
 398 }
 399
 400 static int check_valid_map(struct f2fs_sb_info *sbi,
 401                                 unsigned int segno, int offset)
 402 {
 403         struct sit_info *sit_i = SIT_I(sbi);
 404         struct seg_entry *sentry;
 405         int ret;
 406
 407         mutex_lock(&sit_i->sentry_lock);
 408         sentry = get_seg_entry(sbi, segno);
 409         ret = f2fs_test_bit(offset, sentry->cur_valid_map);
 410         mutex_unlock(&sit_i->sentry_lock);
 411         return ret;
 412 }
 413
 414 /*
 415  * This function compares node address got in summary with that in NAT.
 416  * On validity, copy that node with cold status, otherwise (invalid node)
 417  * ignore that.
 418  */
 419 static void gc_node_segment(struct f2fs_sb_info *sbi,
 420                 struct f2fs_summary *sum, unsigned int segno, int gc_type)
 421 {
 422         bool initial = true;
 423         struct f2fs_summary *entry;
 424         block_t start_addr;
 425         int off;
 426
 427         start_addr = START_BLOCK(sbi, segno);
 428
 429 next_step:
 430         entry = sum;
 431
 432         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 433                 nid_t nid = le32_to_cpu(entry->nid);
 434                 struct page *node_page;
 435                 struct node_info ni;
 436
 437                 /* stop BG_GC if there is not enough free sections. */
 438                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 439                         return;
 440
 441                 if (check_valid_map(sbi, segno, off) == 0)
 442                         continue;
 443
 444                 if (initial) {
 445                         ra_node_page(sbi, nid);
 446                         continue;
 447                 }
 448                 node_page = get_node_page(sbi, nid);
 449                 if (IS_ERR(node_page))
 450                         continue;
 451
 452                 /* block may become invalid during get_node_page */
 453                 if (check_valid_map(sbi, segno, off) == 0) {
 454                         f2fs_put_page(node_page, 1);
 455                         continue;
 456                 }
 457
 458                 get_node_info(sbi, nid, &ni);
 459                 if (ni.blk_addr != start_addr + off) {
 460                         f2fs_put_page(node_page, 1);
 461                         continue;
 462                 }
 463
 464                 /* set page dirty and write it */
 465                 if (gc_type == FG_GC) {
 466                         f2fs_wait_on_page_writeback(node_page, NODE, true);
 467                         set_page_dirty(node_page);
 468                 } else {
 469                         if (!PageWriteback(node_page))
 470                                 set_page_dirty(node_page);
 471                 }
 472                 f2fs_put_page(node_page, 1);
 473                 stat_inc_node_blk_count(sbi, 1, gc_type);
 474         }
 475
 476         if (initial) {
 477                 initial = false;
 478                 goto next_step;
 479         }
 480 }
 481
 482 /*
 483  * Calculate start block index indicating the given node offset.
 484  * Be careful, caller should give this node offset only indicating direct node
 485  * blocks. If any node offsets, which point the other types of node blocks such
 486  * as indirect or double indirect node blocks, are given, it must be a caller's
 487  * bug.
 488  */
 489 block_t start_bidx_of_node(unsigned int node_ofs, struct inode *inode)
 490 {
 491         unsigned int indirect_blks = 2 * NIDS_PER_BLOCK + 4;
 492         unsigned int bidx;
 493
 494         if (node_ofs == 0)
 495                 return 0;
 496
 497         if (node_ofs <= 2) {
 498                 bidx = node_ofs - 1;
 499         } else if (node_ofs <= indirect_blks) {
 500                 int dec = (node_ofs - 4) / (NIDS_PER_BLOCK + 1);
 501                 bidx = node_ofs - 2 - dec;
 502         } else {
 503                 int dec = (node_ofs - indirect_blks - 3) / (NIDS_PER_BLOCK + 1);
 504                 bidx = node_ofs - 5 - dec;
 505         }
 506         return bidx * ADDRS_PER_BLOCK + ADDRS_PER_INODE(inode);
 507 }
 508
 509 static bool is_alive(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 510                 struct node_info *dni, block_t blkaddr, unsigned int *nofs)
 511 {
 512         struct page *node_page;
 513         nid_t nid;
 514         unsigned int ofs_in_node;
 515         block_t source_blkaddr;
 516
 517         nid = le32_to_cpu(sum->nid);
 518         ofs_in_node = le16_to_cpu(sum->ofs_in_node);
 519
 520         node_page = get_node_page(sbi, nid);
 521         if (IS_ERR(node_page))
 522                 return false;
 523
 524         get_node_info(sbi, nid, dni);
 525
 526         if (sum->version != dni->version) {
 527                 f2fs_put_page(node_page, 1);
 528                 return false;
 529         }
 530
 531         *nofs = ofs_of_node(node_page);
 532         source_blkaddr = datablock_addr(node_page, ofs_in_node);
 533         f2fs_put_page(node_page, 1);
 534
 535         if (source_blkaddr != blkaddr)
 536                 return false;
 537         return true;
 538 }
 539
 540 static void move_encrypted_block(struct inode *inode, block_t bidx)
 541 {
 542         struct f2fs_io_info fio = {
 543                 .sbi = F2FS_I_SB(inode),
 544                 .type = DATA,
 545                 .rw = READ_SYNC,
 546                 .encrypted_page = NULL,
 547         };
 548         struct dnode_of_data dn;
 549         struct f2fs_summary sum;
 550         struct node_info ni;
 551         struct page *page;
 552         int err;
 553
 554         /* do not read out */
 555         page = f2fs_grab_cache_page(inode->i_mapping, bidx, false);
 556         if (!page)
 557                 return;
 558
 559         set_new_dnode(&dn, inode, NULL, NULL, 0);
 560         err = get_dnode_of_data(&dn, bidx, LOOKUP_NODE);
 561         if (err)
 562                 goto out;
 563
 564         if (unlikely(dn.data_blkaddr == NULL_ADDR)) {
 565                 ClearPageUptodate(page);
 566                 goto put_out;
 567         }
 568
 569         /*
 570          * don't cache encrypted data into meta inode until previous dirty
 571          * data were writebacked to avoid racing between GC and flush.
 572          */
 573         f2fs_wait_on_page_writeback(page, DATA, true);
 574
 575         get_node_info(fio.sbi, dn.nid, &ni);
 576         set_summary(&sum, dn.nid, dn.ofs_in_node, ni.version);
 577
 578         /* read page */
 579         fio.page = page;
 580         fio.blk_addr = dn.data_blkaddr;
 581
 582         fio.encrypted_page = pagecache_get_page(META_MAPPING(fio.sbi),
 583                                         fio.blk_addr,
 584                                         FGP_LOCK|FGP_CREAT,
 585                                         GFP_NOFS);
 586         if (!fio.encrypted_page)
 587                 goto put_out;
 588
 589         err = f2fs_submit_page_bio(&fio);
 590         if (err)
 591                 goto put_page_out;
 592
 593         /* write page */
 594         lock_page(fio.encrypted_page);
 595
 596         if (unlikely(!PageUptodate(fio.encrypted_page)))
 597                 goto put_page_out;
 598         if (unlikely(fio.encrypted_page->mapping != META_MAPPING(fio.sbi)))
 599                 goto put_page_out;
 600
 601         set_page_dirty(fio.encrypted_page);
 602         f2fs_wait_on_page_writeback(fio.encrypted_page, DATA, true);
 603         if (clear_page_dirty_for_io(fio.encrypted_page))
 604                 dec_page_count(fio.sbi, F2FS_DIRTY_META);
 605
 606         set_page_writeback(fio.encrypted_page);
 607
 608         /* allocate block address */
 609         f2fs_wait_on_page_writeback(dn.node_page, NODE, true);
 610         allocate_data_block(fio.sbi, NULL, fio.blk_addr,
 611                                         &fio.blk_addr, &sum, CURSEG_COLD_DATA);
 612         fio.rw = WRITE_SYNC;
 613         f2fs_submit_page_mbio(&fio);
 614
 615         dn.data_blkaddr = fio.blk_addr;
 616         set_data_blkaddr(&dn);
 617         f2fs_update_extent_cache(&dn);
 618         set_inode_flag(F2FS_I(inode), FI_APPEND_WRITE);
 619         if (page->index == 0)
 620                 set_inode_flag(F2FS_I(inode), FI_FIRST_BLOCK_WRITTEN);
 621 put_page_out:
 622         f2fs_put_page(fio.encrypted_page, 1);
 623 put_out:
 624         f2fs_put_dnode(&dn);
 625 out:
 626         f2fs_put_page(page, 1);
 627 }
 628
 629 static void move_data_page(struct inode *inode, block_t bidx, int gc_type)
 630 {
 631         struct page *page;
 632
 633         page = get_lock_data_page(inode, bidx, true);
 634         if (IS_ERR(page))
 635                 return;
 636
 637         if (gc_type == BG_GC) {
 638                 if (PageWriteback(page))
 639                         goto out;
 640                 set_page_dirty(page);
 641                 set_cold_data(page);
 642         } else {
 643                 struct f2fs_io_info fio = {
 644                         .sbi = F2FS_I_SB(inode),
 645                         .type = DATA,
 646                         .rw = WRITE_SYNC,
 647                         .page = page,
 648                         .encrypted_page = NULL,
 649                 };
 650                 set_page_dirty(page);
 651                 f2fs_wait_on_page_writeback(page, DATA, true);
 652                 if (clear_page_dirty_for_io(page))
 653                         inode_dec_dirty_pages(inode);
 654                 set_cold_data(page);
 655                 do_write_data_page(&fio);
 656                 clear_cold_data(page);
 657         }
 658 out:
 659         f2fs_put_page(page, 1);
 660 }
 661
 662 /*
 663  * This function tries to get parent node of victim data block, and identifies
 664  * data block validity. If the block is valid, copy that with cold status and
 665  * modify parent node.
 666  * If the parent node is not valid or the data block address is different,
 667  * the victim data block is ignored.
 668  */
 669 static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 670                 struct gc_inode_list *gc_list, unsigned int segno, int gc_type)
 671 {
 672         struct super_block *sb = sbi->sb;
 673         struct f2fs_summary *entry;
 674         block_t start_addr;
 675         int off;
 676         int phase = 0;
 677
 678         start_addr = START_BLOCK(sbi, segno);
 679
 680 next_step:
 681         entry = sum;
 682
 683         for (off = 0; off < sbi->blocks_per_seg; off++, entry++) {
 684                 struct page *data_page;
 685                 struct inode *inode;
 686                 struct node_info dni; /* dnode info for the data */
 687                 unsigned int ofs_in_node, nofs;
 688                 block_t start_bidx;
 689
 690                 /* stop BG_GC if there is not enough free sections. */
 691                 if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0))
 692                         return;
 693
 694                 if (check_valid_map(sbi, segno, off) == 0)
 695                         continue;
 696
 697                 if (phase == 0) {
 698                         ra_node_page(sbi, le32_to_cpu(entry->nid));
 699                         continue;
 700                 }
 701
 702                 /* Get an inode by ino with checking validity */
 703                 if (!is_alive(sbi, entry, &dni, start_addr + off, &nofs))
 704                         continue;
 705
 706                 if (phase == 1) {
 707                         ra_node_page(sbi, dni.ino);
 708                         continue;
 709                 }
 710
 711                 ofs_in_node = le16_to_cpu(entry->ofs_in_node);
 712
 713                 if (phase == 2) {
 714                         inode = f2fs_iget(sb, dni.ino);
 715                         if (IS_ERR(inode) || is_bad_inode(inode))
 716                                 continue;
 717
 718                         /* if encrypted inode, let's go phase 3 */
 719                         if (f2fs_encrypted_inode(inode) &&
 720                                                 S_ISREG(inode->i_mode)) {
 721                                 add_gc_inode(gc_list, inode);
 722                                 continue;
 723                         }
 724
 725                         start_bidx = start_bidx_of_node(nofs, inode);
 726                         data_page = get_read_data_page(inode,
 727                                         start_bidx + ofs_in_node, READA, true);
 728                         if (IS_ERR(data_page)) {
 729                                 iput(inode);
 730                                 continue;
 731                         }
 732
 733                         f2fs_put_page(data_page, 0);
 734                         add_gc_inode(gc_list, inode);
 735                         continue;
 736                 }
 737
 738                 /* phase 3 */
 739                 inode = find_gc_inode(gc_list, dni.ino);
 740                 if (inode) {
 741                         start_bidx = start_bidx_of_node(nofs, inode)
 742                                                                 + ofs_in_node;
 743                         if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
 744                                 move_encrypted_block(inode, start_bidx);
 745                         else
 746                                 move_data_page(inode, start_bidx, gc_type);
 747                         stat_inc_data_blk_count(sbi, 1, gc_type);
 748                 }
 749         }
 750
 751         if (++phase < 4)
 752                 goto next_step;
 753 }
 754
 755 static int __get_victim(struct f2fs_sb_info *sbi, unsigned int *victim,
 756                         int gc_type)
 757 {
 758         struct sit_info *sit_i = SIT_I(sbi);
 759         int ret;
 760
 761         mutex_lock(&sit_i->sentry_lock);
 762         ret = DIRTY_I(sbi)->v_ops->get_victim(sbi, victim, gc_type,
 763                                               NO_CHECK_TYPE, LFS);
 764         mutex_unlock(&sit_i->sentry_lock);
 765         return ret;
 766 }
 767
 768 static int do_garbage_collect(struct f2fs_sb_info *sbi,
 769                                 unsigned int start_segno,
 770                                 struct gc_inode_list *gc_list, int gc_type)
 771 {
 772         struct page *sum_page;
 773         struct f2fs_summary_block *sum;
 774         struct blk_plug plug;
 775         unsigned int segno = start_segno;
 776         unsigned int end_segno = start_segno + sbi->segs_per_sec;
 777         int seg_freed = 0;
 778         unsigned char type = IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
 779                                                 SUM_TYPE_DATA : SUM_TYPE_NODE;
 780
 781         /* readahead multi ssa blocks those have contiguous address */
 782         if (sbi->segs_per_sec > 1)
 783                 ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno),
 784                                         sbi->segs_per_sec, META_SSA, true);
 785
 786         /* reference all summary page */
 787         while (segno < end_segno) {
 788                 sum_page = get_sum_page(sbi, segno++);
 789                 unlock_page(sum_page);
 790         }
 791
 792         blk_start_plug(&plug);
 793
 794         for (segno = start_segno; segno < end_segno; segno++) {
 795                 /* find segment summary of victim */
 796                 sum_page = find_get_page(META_MAPPING(sbi),
 797                                         GET_SUM_BLOCK(sbi, segno));
 798                 f2fs_bug_on(sbi, !PageUptodate(sum_page));
 799                 f2fs_put_page(sum_page, 0);
 800
 801                 sum = page_address(sum_page);
 802                 f2fs_bug_on(sbi, type != GET_SUM_TYPE((&sum->footer)));
 803
 804                 /*
 805                  * this is to avoid deadlock:
 806                  * - lock_page(sum_page)         - f2fs_replace_block
 807                  *  - check_valid_map()            - mutex_lock(sentry_lock)
 808                  *   - mutex_lock(sentry_lock)     - change_curseg()
 809                  *                                  - lock_page(sum_page)
 810                  */
 811
 812                 if (type == SUM_TYPE_NODE)
 813                         gc_node_segment(sbi, sum->entries, segno, gc_type);
 814                 else
 815                         gc_data_segment(sbi, sum->entries, gc_list, segno,
 816                                                                 gc_type);
 817
 818                 stat_inc_seg_count(sbi, type, gc_type);
 819                 stat_inc_call_count(sbi->stat_info);
 820
 821                 f2fs_put_page(sum_page, 0);
 822         }
 823
 824         if (gc_type == FG_GC) {
 825                 if (type == SUM_TYPE_NODE) {
 826                         struct writeback_control wbc = {
 827                                 .sync_mode = WB_SYNC_ALL,
 828                                 .nr_to_write = LONG_MAX,
 829                                 .for_reclaim = 0,
 830                         };
 831                         sync_node_pages(sbi, 0, &wbc);
 832                 } else {
 833                         f2fs_submit_merged_bio(sbi, DATA, WRITE);
 834                 }
 835         }
 836
 837         blk_finish_plug(&plug);
 838
 839         if (gc_type == FG_GC) {
 840                 while (start_segno < end_segno)
 841                         if (get_valid_blocks(sbi, start_segno++, 1) == 0)
 842                                 seg_freed++;
 843         }
 844         return seg_freed;
 845 }
 846
 847 int f2fs_gc(struct f2fs_sb_info *sbi, bool sync)
 848 {
 849         unsigned int segno;
 850         int gc_type = sync ? FG_GC : BG_GC;
 851         int sec_freed = 0, seg_freed;
 852         int ret = -EINVAL;
 853         struct cp_control cpc;
 854         struct gc_inode_list gc_list = {
 855                 .ilist = LIST_HEAD_INIT(gc_list.ilist),
 856                 .iroot = RADIX_TREE_INIT(GFP_NOFS),
 857         };
 858
 859         cpc.reason = __get_cp_reason(sbi);
 860 gc_more:
 861         segno = NULL_SEGNO;
 862
 863         if (unlikely(!(sbi->sb->s_flags & MS_ACTIVE)))
 864                 goto stop;
 865         if (unlikely(f2fs_cp_error(sbi))) {
 866                 ret = -EIO;
 867                 goto stop;
 868         }
 869
 870         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed)) {
 871                 gc_type = FG_GC;
 872                 /*
 873                  * If there is no victim and no prefree segment but still not
 874                  * enough free sections, we should flush dent/node blocks and do
 875                  * garbage collections.
 876                  */
 877                 if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
 878                         write_checkpoint(sbi, &cpc);
 879                 else if (has_not_enough_free_secs(sbi, 0))
 880                         write_checkpoint(sbi, &cpc);
 881         }
 882
 883         if (segno == NULL_SEGNO && !__get_victim(sbi, &segno, gc_type))
 884                 goto stop;
 885         ret = 0;
 886
 887         seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type);
 888
 889         if (gc_type == FG_GC && seg_freed == sbi->segs_per_sec)
 890                 sec_freed++;
 891
 892         if (gc_type == FG_GC)
 893                 sbi->cur_victim_sec = NULL_SEGNO;
 894
 895         if (!sync) {
 896                 if (has_not_enough_free_secs(sbi, sec_freed))
 897                         goto gc_more;
 898
 899                 if (gc_type == FG_GC)
 900                         write_checkpoint(sbi, &cpc);
 901         }
 902 stop:
 903         mutex_unlock(&sbi->gc_mutex);
 904
 905         put_gc_inode(&gc_list);
 906
 907         if (sync)
 908                 ret = sec_freed ? 0 : -EAGAIN;
 909         return ret;
 910 }
 911
 912 void build_gc_manager(struct f2fs_sb_info *sbi)
 913 {
 914         DIRTY_I(sbi)->v_ops = &default_v_ops;
 915 }