fs/nilfs2/segment.c

   1 /*
   2  * segment.c - NILFS segment constructor.
   3  *
   4  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * Written by Ryusuke Konishi.
  17  *
  18  */
  19
  20 #include <linux/pagemap.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/writeback.h>
  23 #include <linux/bitops.h>
  24 #include <linux/bio.h>
  25 #include <linux/completion.h>
  26 #include <linux/blkdev.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/freezer.h>
  29 #include <linux/kthread.h>
  30 #include <linux/crc32.h>
  31 #include <linux/pagevec.h>
  32 #include <linux/slab.h>
  33 #include <linux/sched/signal.h>
  34
  35 #include "nilfs.h"
  36 #include "btnode.h"
  37 #include "page.h"
  38 #include "segment.h"
  39 #include "sufile.h"
  40 #include "cpfile.h"
  41 #include "ifile.h"
  42 #include "segbuf.h"
  43
  44
  45 /*
  46  * Segment constructor
  47  */
  48 #define SC_N_INODEVEC   16   /* Size of locally allocated inode vector */
  49
  50 #define SC_MAX_SEGDELTA 64   /*
  51                               * Upper limit of the number of segments
  52                               * appended in collection retry loop
  53                               */
  54
  55 /* Construction mode */
  56 enum {
  57         SC_LSEG_SR = 1, /* Make a logical segment having a super root */
  58         SC_LSEG_DSYNC,  /*
  59                          * Flush data blocks of a given file and make
  60                          * a logical segment without a super root.
  61                          */
  62         SC_FLUSH_FILE,  /*
  63                          * Flush data files, leads to segment writes without
  64                          * creating a checkpoint.
  65                          */
  66         SC_FLUSH_DAT,   /*
  67                          * Flush DAT file.  This also creates segments
  68                          * without a checkpoint.
  69                          */
  70 };
  71
  72 /* Stage numbers of dirty block collection */
  73 enum {
  74         NILFS_ST_INIT = 0,
  75         NILFS_ST_GC,            /* Collecting dirty blocks for GC */
  76         NILFS_ST_FILE,
  77         NILFS_ST_IFILE,
  78         NILFS_ST_CPFILE,
  79         NILFS_ST_SUFILE,
  80         NILFS_ST_DAT,
  81         NILFS_ST_SR,            /* Super root */
  82         NILFS_ST_DSYNC,         /* Data sync blocks */
  83         NILFS_ST_DONE,
  84 };
  85
  86 #define CREATE_TRACE_POINTS
  87 #include <trace/events/nilfs2.h>
  88
  89 /*
  90  * nilfs_sc_cstage_inc(), nilfs_sc_cstage_set(), nilfs_sc_cstage_get() are
  91  * wrapper functions of stage count (nilfs_sc_info->sc_stage.scnt). Users of
  92  * the variable must use them because transition of stage count must involve
  93  * trace events (trace_nilfs2_collection_stage_transition).
  94  *
  95  * nilfs_sc_cstage_get() isn't required for the above purpose because it doesn't
  96  * produce tracepoint events. It is provided just for making the intention
  97  * clear.
  98  */
  99 static inline void nilfs_sc_cstage_inc(struct nilfs_sc_info *sci)
 100 {
 101         sci->sc_stage.scnt++;
 102         trace_nilfs2_collection_stage_transition(sci);
 103 }
 104
 105 static inline void nilfs_sc_cstage_set(struct nilfs_sc_info *sci, int next_scnt)
 106 {
 107         sci->sc_stage.scnt = next_scnt;
 108         trace_nilfs2_collection_stage_transition(sci);
 109 }
 110
 111 static inline int nilfs_sc_cstage_get(struct nilfs_sc_info *sci)
 112 {
 113         return sci->sc_stage.scnt;
 114 }
 115
 116 /* State flags of collection */
 117 #define NILFS_CF_NODE           0x0001  /* Collecting node blocks */
 118 #define NILFS_CF_IFILE_STARTED  0x0002  /* IFILE stage has started */
 119 #define NILFS_CF_SUFREED        0x0004  /* segment usages has been freed */
 120 #define NILFS_CF_HISTORY_MASK   (NILFS_CF_IFILE_STARTED | NILFS_CF_SUFREED)
 121
 122 /* Operations depending on the construction mode and file type */
 123 struct nilfs_sc_operations {
 124         int (*collect_data)(struct nilfs_sc_info *, struct buffer_head *,
 125                             struct inode *);
 126         int (*collect_node)(struct nilfs_sc_info *, struct buffer_head *,
 127                             struct inode *);
 128         int (*collect_bmap)(struct nilfs_sc_info *, struct buffer_head *,
 129                             struct inode *);
 130         void (*write_data_binfo)(struct nilfs_sc_info *,
 131                                  struct nilfs_segsum_pointer *,
 132                                  union nilfs_binfo *);
 133         void (*write_node_binfo)(struct nilfs_sc_info *,
 134                                  struct nilfs_segsum_pointer *,
 135                                  union nilfs_binfo *);
 136 };
 137
 138 /*
 139  * Other definitions
 140  */
 141 static void nilfs_segctor_start_timer(struct nilfs_sc_info *);
 142 static void nilfs_segctor_do_flush(struct nilfs_sc_info *, int);
 143 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *);
 144 static void nilfs_dispose_list(struct the_nilfs *, struct list_head *, int);
 145
 146 #define nilfs_cnt32_gt(a, b)   \
 147         (typecheck(__u32, a) && typecheck(__u32, b) && \
 148          ((__s32)(b) - (__s32)(a) < 0))
 149 #define nilfs_cnt32_ge(a, b)   \
 150         (typecheck(__u32, a) && typecheck(__u32, b) && \
 151          ((__s32)(a) - (__s32)(b) >= 0))
 152 #define nilfs_cnt32_lt(a, b)  nilfs_cnt32_gt(b, a)
 153 #define nilfs_cnt32_le(a, b)  nilfs_cnt32_ge(b, a)
 154
 155 static int nilfs_prepare_segment_lock(struct super_block *sb,
 156                                       struct nilfs_transaction_info *ti)
 157 {
 158         struct nilfs_transaction_info *cur_ti = current->journal_info;
 159         void *save = NULL;
 160
 161         if (cur_ti) {
 162                 if (cur_ti->ti_magic == NILFS_TI_MAGIC)
 163                         return ++cur_ti->ti_count;
 164
 165                 /*
 166                  * If journal_info field is occupied by other FS,
 167                  * it is saved and will be restored on
 168                  * nilfs_transaction_commit().
 169                  */
 170                 nilfs_msg(sb, KERN_WARNING, "journal info from a different FS");
 171                 save = current->journal_info;
 172         }
 173         if (!ti) {
 174                 ti = kmem_cache_alloc(nilfs_transaction_cachep, GFP_NOFS);
 175                 if (!ti)
 176                         return -ENOMEM;
 177                 ti->ti_flags = NILFS_TI_DYNAMIC_ALLOC;
 178         } else {
 179                 ti->ti_flags = 0;
 180         }
 181         ti->ti_count = 0;
 182         ti->ti_save = save;
 183         ti->ti_magic = NILFS_TI_MAGIC;
 184         current->journal_info = ti;
 185         return 0;
 186 }
 187
 188 /**
 189  * nilfs_transaction_begin - start indivisible file operations.
 190  * @sb: super block
 191  * @ti: nilfs_transaction_info
 192  * @vacancy_check: flags for vacancy rate checks
 193  *
 194  * nilfs_transaction_begin() acquires a reader/writer semaphore, called
 195  * the segment semaphore, to make a segment construction and write tasks
 196  * exclusive.  The function is used with nilfs_transaction_commit() in pairs.
 197  * The region enclosed by these two functions can be nested.  To avoid a
 198  * deadlock, the semaphore is only acquired or released in the outermost call.
 199  *
 200  * This function allocates a nilfs_transaction_info struct to keep context
 201  * information on it.  It is initialized and hooked onto the current task in
 202  * the outermost call.  If a pre-allocated struct is given to @ti, it is used
 203  * instead; otherwise a new struct is assigned from a slab.
 204  *
 205  * When @vacancy_check flag is set, this function will check the amount of
 206  * free space, and will wait for the GC to reclaim disk space if low capacity.
 207  *
 208  * Return Value: On success, 0 is returned. On error, one of the following
 209  * negative error code is returned.
 210  *
 211  * %-ENOMEM - Insufficient memory available.
 212  *
 213  * %-ENOSPC - No space left on device
 214  */
 215 int nilfs_transaction_begin(struct super_block *sb,
 216                             struct nilfs_transaction_info *ti,
 217                             int vacancy_check)
 218 {
 219         struct the_nilfs *nilfs;
 220         int ret = nilfs_prepare_segment_lock(sb, ti);
 221         struct nilfs_transaction_info *trace_ti;
 222
 223         if (unlikely(ret < 0))
 224                 return ret;
 225         if (ret > 0) {
 226                 trace_ti = current->journal_info;
 227
 228                 trace_nilfs2_transaction_transition(sb, trace_ti,
 229                                     trace_ti->ti_count, trace_ti->ti_flags,
 230                                     TRACE_NILFS2_TRANSACTION_BEGIN);
 231                 return 0;
 232         }
 233
 234         sb_start_intwrite(sb);
 235
 236         nilfs = sb->s_fs_info;
 237         down_read(&nilfs->ns_segctor_sem);
 238         if (vacancy_check && nilfs_near_disk_full(nilfs)) {
 239                 up_read(&nilfs->ns_segctor_sem);
 240                 ret = -ENOSPC;
 241                 goto failed;
 242         }
 243
 244         trace_ti = current->journal_info;
 245         trace_nilfs2_transaction_transition(sb, trace_ti, trace_ti->ti_count,
 246                                             trace_ti->ti_flags,
 247                                             TRACE_NILFS2_TRANSACTION_BEGIN);
 248         return 0;
 249
 250  failed:
 251         ti = current->journal_info;
 252         current->journal_info = ti->ti_save;
 253         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 254                 kmem_cache_free(nilfs_transaction_cachep, ti);
 255         sb_end_intwrite(sb);
 256         return ret;
 257 }
 258
 259 /**
 260  * nilfs_transaction_commit - commit indivisible file operations.
 261  * @sb: super block
 262  *
 263  * nilfs_transaction_commit() releases the read semaphore which is
 264  * acquired by nilfs_transaction_begin(). This is only performed
 265  * in outermost call of this function.  If a commit flag is set,
 266  * nilfs_transaction_commit() sets a timer to start the segment
 267  * constructor.  If a sync flag is set, it starts construction
 268  * directly.
 269  */
 270 int nilfs_transaction_commit(struct super_block *sb)
 271 {
 272         struct nilfs_transaction_info *ti = current->journal_info;
 273         struct the_nilfs *nilfs = sb->s_fs_info;
 274         int err = 0;
 275
 276         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 277         ti->ti_flags |= NILFS_TI_COMMIT;
 278         if (ti->ti_count > 0) {
 279                 ti->ti_count--;
 280                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 281                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_COMMIT);
 282                 return 0;
 283         }
 284         if (nilfs->ns_writer) {
 285                 struct nilfs_sc_info *sci = nilfs->ns_writer;
 286
 287                 if (ti->ti_flags & NILFS_TI_COMMIT)
 288                         nilfs_segctor_start_timer(sci);
 289                 if (atomic_read(&nilfs->ns_ndirtyblks) > sci->sc_watermark)
 290                         nilfs_segctor_do_flush(sci, 0);
 291         }
 292         up_read(&nilfs->ns_segctor_sem);
 293         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 294                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_COMMIT);
 295
 296         current->journal_info = ti->ti_save;
 297
 298         if (ti->ti_flags & NILFS_TI_SYNC)
 299                 err = nilfs_construct_segment(sb);
 300         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 301                 kmem_cache_free(nilfs_transaction_cachep, ti);
 302         sb_end_intwrite(sb);
 303         return err;
 304 }
 305
 306 void nilfs_transaction_abort(struct super_block *sb)
 307 {
 308         struct nilfs_transaction_info *ti = current->journal_info;
 309         struct the_nilfs *nilfs = sb->s_fs_info;
 310
 311         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 312         if (ti->ti_count > 0) {
 313                 ti->ti_count--;
 314                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 315                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_ABORT);
 316                 return;
 317         }
 318         up_read(&nilfs->ns_segctor_sem);
 319
 320         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 321                     ti->ti_flags, TRACE_NILFS2_TRANSACTION_ABORT);
 322
 323         current->journal_info = ti->ti_save;
 324         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 325                 kmem_cache_free(nilfs_transaction_cachep, ti);
 326         sb_end_intwrite(sb);
 327 }
 328
 329 void nilfs_relax_pressure_in_lock(struct super_block *sb)
 330 {
 331         struct the_nilfs *nilfs = sb->s_fs_info;
 332         struct nilfs_sc_info *sci = nilfs->ns_writer;
 333
 334         if (sb_rdonly(sb) || unlikely(!sci) || !sci->sc_flush_request)
 335                 return;
 336
 337         set_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
 338         up_read(&nilfs->ns_segctor_sem);
 339
 340         down_write(&nilfs->ns_segctor_sem);
 341         if (sci->sc_flush_request &&
 342             test_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags)) {
 343                 struct nilfs_transaction_info *ti = current->journal_info;
 344
 345                 ti->ti_flags |= NILFS_TI_WRITER;
 346                 nilfs_segctor_do_immediate_flush(sci);
 347                 ti->ti_flags &= ~NILFS_TI_WRITER;
 348         }
 349         downgrade_write(&nilfs->ns_segctor_sem);
 350 }
 351
 352 static void nilfs_transaction_lock(struct super_block *sb,
 353                                    struct nilfs_transaction_info *ti,
 354                                    int gcflag)
 355 {
 356         struct nilfs_transaction_info *cur_ti = current->journal_info;
 357         struct the_nilfs *nilfs = sb->s_fs_info;
 358         struct nilfs_sc_info *sci = nilfs->ns_writer;
 359
 360         WARN_ON(cur_ti);
 361         ti->ti_flags = NILFS_TI_WRITER;
 362         ti->ti_count = 0;
 363         ti->ti_save = cur_ti;
 364         ti->ti_magic = NILFS_TI_MAGIC;
 365         current->journal_info = ti;
 366
 367         for (;;) {
 368                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 369                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_TRYLOCK);
 370
 371                 down_write(&nilfs->ns_segctor_sem);
 372                 if (!test_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags))
 373                         break;
 374
 375                 nilfs_segctor_do_immediate_flush(sci);
 376
 377                 up_write(&nilfs->ns_segctor_sem);
 378                 cond_resched();
 379         }
 380         if (gcflag)
 381                 ti->ti_flags |= NILFS_TI_GC;
 382
 383         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 384                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_LOCK);
 385 }
 386
 387 static void nilfs_transaction_unlock(struct super_block *sb)
 388 {
 389         struct nilfs_transaction_info *ti = current->journal_info;
 390         struct the_nilfs *nilfs = sb->s_fs_info;
 391
 392         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 393         BUG_ON(ti->ti_count > 0);
 394
 395         up_write(&nilfs->ns_segctor_sem);
 396         current->journal_info = ti->ti_save;
 397
 398         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 399                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_UNLOCK);
 400 }
 401
 402 static void *nilfs_segctor_map_segsum_entry(struct nilfs_sc_info *sci,
 403                                             struct nilfs_segsum_pointer *ssp,
 404                                             unsigned int bytes)
 405 {
 406         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 407         unsigned int blocksize = sci->sc_super->s_blocksize;
 408         void *p;
 409
 410         if (unlikely(ssp->offset + bytes > blocksize)) {
 411                 ssp->offset = 0;
 412                 BUG_ON(NILFS_SEGBUF_BH_IS_LAST(ssp->bh,
 413                                                &segbuf->sb_segsum_buffers));
 414                 ssp->bh = NILFS_SEGBUF_NEXT_BH(ssp->bh);
 415         }
 416         p = ssp->bh->b_data + ssp->offset;
 417         ssp->offset += bytes;
 418         return p;
 419 }
 420
 421 /**
 422  * nilfs_segctor_reset_segment_buffer - reset the current segment buffer
 423  * @sci: nilfs_sc_info
 424  */
 425 static int nilfs_segctor_reset_segment_buffer(struct nilfs_sc_info *sci)
 426 {
 427         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 428         struct buffer_head *sumbh;
 429         unsigned int sumbytes;
 430         unsigned int flags = 0;
 431         int err;
 432
 433         if (nilfs_doing_gc())
 434                 flags = NILFS_SS_GC;
 435         err = nilfs_segbuf_reset(segbuf, flags, sci->sc_seg_ctime, sci->sc_cno);
 436         if (unlikely(err))
 437                 return err;
 438
 439         sumbh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
 440         sumbytes = segbuf->sb_sum.sumbytes;
 441         sci->sc_finfo_ptr.bh = sumbh;  sci->sc_finfo_ptr.offset = sumbytes;
 442         sci->sc_binfo_ptr.bh = sumbh;  sci->sc_binfo_ptr.offset = sumbytes;
 443         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 444         return 0;
 445 }
 446
 447 static int nilfs_segctor_feed_segment(struct nilfs_sc_info *sci)
 448 {
 449         sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
 450         if (NILFS_SEGBUF_IS_LAST(sci->sc_curseg, &sci->sc_segbufs))
 451                 return -E2BIG; /*
 452                                 * The current segment is filled up
 453                                 * (internal code)
 454                                 */
 455         sci->sc_curseg = NILFS_NEXT_SEGBUF(sci->sc_curseg);
 456         return nilfs_segctor_reset_segment_buffer(sci);
 457 }
 458
 459 static int nilfs_segctor_add_super_root(struct nilfs_sc_info *sci)
 460 {
 461         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 462         int err;
 463
 464         if (segbuf->sb_sum.nblocks >= segbuf->sb_rest_blocks) {
 465                 err = nilfs_segctor_feed_segment(sci);
 466                 if (err)
 467                         return err;
 468                 segbuf = sci->sc_curseg;
 469         }
 470         err = nilfs_segbuf_extend_payload(segbuf, &segbuf->sb_super_root);
 471         if (likely(!err))
 472                 segbuf->sb_sum.flags |= NILFS_SS_SR;
 473         return err;
 474 }
 475
 476 /*
 477  * Functions for making segment summary and payloads
 478  */
 479 static int nilfs_segctor_segsum_block_required(
 480         struct nilfs_sc_info *sci, const struct nilfs_segsum_pointer *ssp,
 481         unsigned int binfo_size)
 482 {
 483         unsigned int blocksize = sci->sc_super->s_blocksize;
 484         /* Size of finfo and binfo is enough small against blocksize */
 485
 486         return ssp->offset + binfo_size +
 487                 (!sci->sc_blk_cnt ? sizeof(struct nilfs_finfo) : 0) >
 488                 blocksize;
 489 }
 490
 491 static void nilfs_segctor_begin_finfo(struct nilfs_sc_info *sci,
 492                                       struct inode *inode)
 493 {
 494         sci->sc_curseg->sb_sum.nfinfo++;
 495         sci->sc_binfo_ptr = sci->sc_finfo_ptr;
 496         nilfs_segctor_map_segsum_entry(
 497                 sci, &sci->sc_binfo_ptr, sizeof(struct nilfs_finfo));
 498
 499         if (NILFS_I(inode)->i_root &&
 500             !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 501                 set_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
 502         /* skip finfo */
 503 }
 504
 505 static void nilfs_segctor_end_finfo(struct nilfs_sc_info *sci,
 506                                     struct inode *inode)
 507 {
 508         struct nilfs_finfo *finfo;
 509         struct nilfs_inode_info *ii;
 510         struct nilfs_segment_buffer *segbuf;
 511         __u64 cno;
 512
 513         if (sci->sc_blk_cnt == 0)
 514                 return;
 515
 516         ii = NILFS_I(inode);
 517
 518         if (test_bit(NILFS_I_GCINODE, &ii->i_state))
 519                 cno = ii->i_cno;
 520         else if (NILFS_ROOT_METADATA_FILE(inode->i_ino))
 521                 cno = 0;
 522         else
 523                 cno = sci->sc_cno;
 524
 525         finfo = nilfs_segctor_map_segsum_entry(sci, &sci->sc_finfo_ptr,
 526                                                  sizeof(*finfo));
 527         finfo->fi_ino = cpu_to_le64(inode->i_ino);
 528         finfo->fi_nblocks = cpu_to_le32(sci->sc_blk_cnt);
 529         finfo->fi_ndatablk = cpu_to_le32(sci->sc_datablk_cnt);
 530         finfo->fi_cno = cpu_to_le64(cno);
 531
 532         segbuf = sci->sc_curseg;
 533         segbuf->sb_sum.sumbytes = sci->sc_binfo_ptr.offset +
 534                 sci->sc_super->s_blocksize * (segbuf->sb_sum.nsumblk - 1);
 535         sci->sc_finfo_ptr = sci->sc_binfo_ptr;
 536         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 537 }
 538
 539 static int nilfs_segctor_add_file_block(struct nilfs_sc_info *sci,
 540                                         struct buffer_head *bh,
 541                                         struct inode *inode,
 542                                         unsigned int binfo_size)
 543 {
 544         struct nilfs_segment_buffer *segbuf;
 545         int required, err = 0;
 546
 547  retry:
 548         segbuf = sci->sc_curseg;
 549         required = nilfs_segctor_segsum_block_required(
 550                 sci, &sci->sc_binfo_ptr, binfo_size);
 551         if (segbuf->sb_sum.nblocks + required + 1 > segbuf->sb_rest_blocks) {
 552                 nilfs_segctor_end_finfo(sci, inode);
 553                 err = nilfs_segctor_feed_segment(sci);
 554                 if (err)
 555                         return err;
 556                 goto retry;
 557         }
 558         if (unlikely(required)) {
 559                 err = nilfs_segbuf_extend_segsum(segbuf);
 560                 if (unlikely(err))
 561                         goto failed;
 562         }
 563         if (sci->sc_blk_cnt == 0)
 564                 nilfs_segctor_begin_finfo(sci, inode);
 565
 566         nilfs_segctor_map_segsum_entry(sci, &sci->sc_binfo_ptr, binfo_size);
 567         /* Substitution to vblocknr is delayed until update_blocknr() */
 568         nilfs_segbuf_add_file_buffer(segbuf, bh);
 569         sci->sc_blk_cnt++;
 570  failed:
 571         return err;
 572 }
 573
 574 /*
 575  * Callback functions that enumerate, mark, and collect dirty blocks
 576  */
 577 static int nilfs_collect_file_data(struct nilfs_sc_info *sci,
 578                                    struct buffer_head *bh, struct inode *inode)
 579 {
 580         int err;
 581
 582         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 583         if (err < 0)
 584                 return err;
 585
 586         err = nilfs_segctor_add_file_block(sci, bh, inode,
 587                                            sizeof(struct nilfs_binfo_v));
 588         if (!err)
 589                 sci->sc_datablk_cnt++;
 590         return err;
 591 }
 592
 593 static int nilfs_collect_file_node(struct nilfs_sc_info *sci,
 594                                    struct buffer_head *bh,
 595                                    struct inode *inode)
 596 {
 597         return nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 598 }
 599
 600 static int nilfs_collect_file_bmap(struct nilfs_sc_info *sci,
 601                                    struct buffer_head *bh,
 602                                    struct inode *inode)
 603 {
 604         WARN_ON(!buffer_dirty(bh));
 605         return nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 606 }
 607
 608 static void nilfs_write_file_data_binfo(struct nilfs_sc_info *sci,
 609                                         struct nilfs_segsum_pointer *ssp,
 610                                         union nilfs_binfo *binfo)
 611 {
 612         struct nilfs_binfo_v *binfo_v = nilfs_segctor_map_segsum_entry(
 613                 sci, ssp, sizeof(*binfo_v));
 614         *binfo_v = binfo->bi_v;
 615 }
 616
 617 static void nilfs_write_file_node_binfo(struct nilfs_sc_info *sci,
 618                                         struct nilfs_segsum_pointer *ssp,
 619                                         union nilfs_binfo *binfo)
 620 {
 621         __le64 *vblocknr = nilfs_segctor_map_segsum_entry(
 622                 sci, ssp, sizeof(*vblocknr));
 623         *vblocknr = binfo->bi_v.bi_vblocknr;
 624 }
 625
 626 static const struct nilfs_sc_operations nilfs_sc_file_ops = {
 627         .collect_data = nilfs_collect_file_data,
 628         .collect_node = nilfs_collect_file_node,
 629         .collect_bmap = nilfs_collect_file_bmap,
 630         .write_data_binfo = nilfs_write_file_data_binfo,
 631         .write_node_binfo = nilfs_write_file_node_binfo,
 632 };
 633
 634 static int nilfs_collect_dat_data(struct nilfs_sc_info *sci,
 635                                   struct buffer_head *bh, struct inode *inode)
 636 {
 637         int err;
 638
 639         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 640         if (err < 0)
 641                 return err;
 642
 643         err = nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 644         if (!err)
 645                 sci->sc_datablk_cnt++;
 646         return err;
 647 }
 648
 649 static int nilfs_collect_dat_bmap(struct nilfs_sc_info *sci,
 650                                   struct buffer_head *bh, struct inode *inode)
 651 {
 652         WARN_ON(!buffer_dirty(bh));
 653         return nilfs_segctor_add_file_block(sci, bh, inode,
 654                                             sizeof(struct nilfs_binfo_dat));
 655 }
 656
 657 static void nilfs_write_dat_data_binfo(struct nilfs_sc_info *sci,
 658                                        struct nilfs_segsum_pointer *ssp,
 659                                        union nilfs_binfo *binfo)
 660 {
 661         __le64 *blkoff = nilfs_segctor_map_segsum_entry(sci, ssp,
 662                                                           sizeof(*blkoff));
 663         *blkoff = binfo->bi_dat.bi_blkoff;
 664 }
 665
 666 static void nilfs_write_dat_node_binfo(struct nilfs_sc_info *sci,
 667                                        struct nilfs_segsum_pointer *ssp,
 668                                        union nilfs_binfo *binfo)
 669 {
 670         struct nilfs_binfo_dat *binfo_dat =
 671                 nilfs_segctor_map_segsum_entry(sci, ssp, sizeof(*binfo_dat));
 672         *binfo_dat = binfo->bi_dat;
 673 }
 674
 675 static const struct nilfs_sc_operations nilfs_sc_dat_ops = {
 676         .collect_data = nilfs_collect_dat_data,
 677         .collect_node = nilfs_collect_file_node,
 678         .collect_bmap = nilfs_collect_dat_bmap,
 679         .write_data_binfo = nilfs_write_dat_data_binfo,
 680         .write_node_binfo = nilfs_write_dat_node_binfo,
 681 };
 682
 683 static const struct nilfs_sc_operations nilfs_sc_dsync_ops = {
 684         .collect_data = nilfs_collect_file_data,
 685         .collect_node = NULL,
 686         .collect_bmap = NULL,
 687         .write_data_binfo = nilfs_write_file_data_binfo,
 688         .write_node_binfo = NULL,
 689 };
 690
 691 static size_t nilfs_lookup_dirty_data_buffers(struct inode *inode,
 692                                               struct list_head *listp,
 693                                               size_t nlimit,
 694                                               loff_t start, loff_t end)
 695 {
 696         struct address_space *mapping = inode->i_mapping;
 697         struct pagevec pvec;
 698         pgoff_t index = 0, last = ULONG_MAX;
 699         size_t ndirties = 0;
 700         int i;
 701
 702         if (unlikely(start != 0 || end != LLONG_MAX)) {
 703                 /*
 704                  * A valid range is given for sync-ing data pages. The
 705                  * range is rounded to per-page; extra dirty buffers
 706                  * may be included if blocksize < pagesize.
 707                  */
 708                 index = start >> PAGE_SHIFT;
 709                 last = end >> PAGE_SHIFT;
 710         }
 711         pagevec_init(&pvec, 0);
 712  repeat:
 713         if (unlikely(index > last) ||
 714             !pagevec_lookup_tag(&pvec, mapping, &index, PAGECACHE_TAG_DIRTY,
 715                                 min_t(pgoff_t, last - index,
 716                                       PAGEVEC_SIZE - 1) + 1))
 717                 return ndirties;
 718
 719         for (i = 0; i < pagevec_count(&pvec); i++) {
 720                 struct buffer_head *bh, *head;
 721                 struct page *page = pvec.pages[i];
 722
 723                 if (unlikely(page->index > last))
 724                         break;
 725
 726                 lock_page(page);
 727                 if (!page_has_buffers(page))
 728                         create_empty_buffers(page, i_blocksize(inode), 0);
 729                 unlock_page(page);
 730
 731                 bh = head = page_buffers(page);
 732                 do {
 733                         if (!buffer_dirty(bh) || buffer_async_write(bh))
 734                                 continue;
 735                         get_bh(bh);
 736                         list_add_tail(&bh->b_assoc_buffers, listp);
 737                         ndirties++;
 738                         if (unlikely(ndirties >= nlimit)) {
 739                                 pagevec_release(&pvec);
 740                                 cond_resched();
 741                                 return ndirties;
 742                         }
 743                 } while (bh = bh->b_this_page, bh != head);
 744         }
 745         pagevec_release(&pvec);
 746         cond_resched();
 747         goto repeat;
 748 }
 749
 750 static void nilfs_lookup_dirty_node_buffers(struct inode *inode,
 751                                             struct list_head *listp)
 752 {
 753         struct nilfs_inode_info *ii = NILFS_I(inode);
 754         struct inode *btnc_inode = ii->i_assoc_inode;
 755         struct pagevec pvec;
 756         struct buffer_head *bh, *head;
 757         unsigned int i;
 758         pgoff_t index = 0;
 759
 760         if (!btnc_inode)
 761                 return;
 762
 763         pagevec_init(&pvec, 0);
 764
 765         while (pagevec_lookup_tag(&pvec, btnc_inode->i_mapping, &index,
 766                                   PAGECACHE_TAG_DIRTY, PAGEVEC_SIZE)) {
 767                 for (i = 0; i < pagevec_count(&pvec); i++) {
 768                         bh = head = page_buffers(pvec.pages[i]);
 769                         do {
 770                                 if (buffer_dirty(bh) &&
 771                                                 !buffer_async_write(bh)) {
 772                                         get_bh(bh);
 773                                         list_add_tail(&bh->b_assoc_buffers,
 774                                                       listp);
 775                                 }
 776                                 bh = bh->b_this_page;
 777                         } while (bh != head);
 778                 }
 779                 pagevec_release(&pvec);
 780                 cond_resched();
 781         }
 782 }
 783
 784 static void nilfs_dispose_list(struct the_nilfs *nilfs,
 785                                struct list_head *head, int force)
 786 {
 787         struct nilfs_inode_info *ii, *n;
 788         struct nilfs_inode_info *ivec[SC_N_INODEVEC], **pii;
 789         unsigned int nv = 0;
 790
 791         while (!list_empty(head)) {
 792                 spin_lock(&nilfs->ns_inode_lock);
 793                 list_for_each_entry_safe(ii, n, head, i_dirty) {
 794                         list_del_init(&ii->i_dirty);
 795                         if (force) {
 796                                 if (unlikely(ii->i_bh)) {
 797                                         brelse(ii->i_bh);
 798                                         ii->i_bh = NULL;
 799                                 }
 800                         } else if (test_bit(NILFS_I_DIRTY, &ii->i_state)) {
 801                                 set_bit(NILFS_I_QUEUED, &ii->i_state);
 802                                 list_add_tail(&ii->i_dirty,
 803                                               &nilfs->ns_dirty_files);
 804                                 continue;
 805                         }
 806                         ivec[nv++] = ii;
 807                         if (nv == SC_N_INODEVEC)
 808                                 break;
 809                 }
 810                 spin_unlock(&nilfs->ns_inode_lock);
 811
 812                 for (pii = ivec; nv > 0; pii++, nv--)
 813                         iput(&(*pii)->vfs_inode);
 814         }
 815 }
 816
 817 static void nilfs_iput_work_func(struct work_struct *work)
 818 {
 819         struct nilfs_sc_info *sci = container_of(work, struct nilfs_sc_info,
 820                                                  sc_iput_work);
 821         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 822
 823         nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 0);
 824 }
 825
 826 static int nilfs_test_metadata_dirty(struct the_nilfs *nilfs,
 827                                      struct nilfs_root *root)
 828 {
 829         int ret = 0;
 830
 831         if (nilfs_mdt_fetch_dirty(root->ifile))
 832                 ret++;
 833         if (nilfs_mdt_fetch_dirty(nilfs->ns_cpfile))
 834                 ret++;
 835         if (nilfs_mdt_fetch_dirty(nilfs->ns_sufile))
 836                 ret++;
 837         if ((ret || nilfs_doing_gc()) && nilfs_mdt_fetch_dirty(nilfs->ns_dat))
 838                 ret++;
 839         return ret;
 840 }
 841
 842 static int nilfs_segctor_clean(struct nilfs_sc_info *sci)
 843 {
 844         return list_empty(&sci->sc_dirty_files) &&
 845                 !test_bit(NILFS_SC_DIRTY, &sci->sc_flags) &&
 846                 sci->sc_nfreesegs == 0 &&
 847                 (!nilfs_doing_gc() || list_empty(&sci->sc_gc_inodes));
 848 }
 849
 850 static int nilfs_segctor_confirm(struct nilfs_sc_info *sci)
 851 {
 852         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 853         int ret = 0;
 854
 855         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
 856                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
 857
 858         spin_lock(&nilfs->ns_inode_lock);
 859         if (list_empty(&nilfs->ns_dirty_files) && nilfs_segctor_clean(sci))
 860                 ret++;
 861
 862         spin_unlock(&nilfs->ns_inode_lock);
 863         return ret;
 864 }
 865
 866 static void nilfs_segctor_clear_metadata_dirty(struct nilfs_sc_info *sci)
 867 {
 868         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 869
 870         nilfs_mdt_clear_dirty(sci->sc_root->ifile);
 871         nilfs_mdt_clear_dirty(nilfs->ns_cpfile);
 872         nilfs_mdt_clear_dirty(nilfs->ns_sufile);
 873         nilfs_mdt_clear_dirty(nilfs->ns_dat);
 874 }
 875
 876 static int nilfs_segctor_create_checkpoint(struct nilfs_sc_info *sci)
 877 {
 878         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 879         struct buffer_head *bh_cp;
 880         struct nilfs_checkpoint *raw_cp;
 881         int err;
 882
 883         /* XXX: this interface will be changed */
 884         err = nilfs_cpfile_get_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, 1,
 885                                           &raw_cp, &bh_cp);
 886         if (likely(!err)) {
 887                 /*
 888                  * The following code is duplicated with cpfile.  But, it is
 889                  * needed to collect the checkpoint even if it was not newly
 890                  * created.
 891                  */
 892                 mark_buffer_dirty(bh_cp);
 893                 nilfs_mdt_mark_dirty(nilfs->ns_cpfile);
 894                 nilfs_cpfile_put_checkpoint(
 895                         nilfs->ns_cpfile, nilfs->ns_cno, bh_cp);
 896         } else if (err == -EINVAL || err == -ENOENT) {
 897                 nilfs_error(sci->sc_super,
 898                             "checkpoint creation failed due to metadata corruption.");
 899                 err = -EIO;
 900         }
 901         return err;
 902 }
 903
 904 static int nilfs_segctor_fill_in_checkpoint(struct nilfs_sc_info *sci)
 905 {
 906         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 907         struct buffer_head *bh_cp;
 908         struct nilfs_checkpoint *raw_cp;
 909         int err;
 910
 911         err = nilfs_cpfile_get_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, 0,
 912                                           &raw_cp, &bh_cp);
 913         if (unlikely(err)) {
 914                 if (err == -EINVAL || err == -ENOENT) {
 915                         nilfs_error(sci->sc_super,
 916                                     "checkpoint finalization failed due to metadata corruption.");
 917                         err = -EIO;
 918                 }
 919                 goto failed_ibh;
 920         }
 921         raw_cp->cp_snapshot_list.ssl_next = 0;
 922         raw_cp->cp_snapshot_list.ssl_prev = 0;
 923         raw_cp->cp_inodes_count =
 924                 cpu_to_le64(atomic64_read(&sci->sc_root->inodes_count));
 925         raw_cp->cp_blocks_count =
 926                 cpu_to_le64(atomic64_read(&sci->sc_root->blocks_count));
 927         raw_cp->cp_nblk_inc =
 928                 cpu_to_le64(sci->sc_nblk_inc + sci->sc_nblk_this_inc);
 929         raw_cp->cp_create = cpu_to_le64(sci->sc_seg_ctime);
 930         raw_cp->cp_cno = cpu_to_le64(nilfs->ns_cno);
 931
 932         if (test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 933                 nilfs_checkpoint_clear_minor(raw_cp);
 934         else
 935                 nilfs_checkpoint_set_minor(raw_cp);
 936
 937         nilfs_write_inode_common(sci->sc_root->ifile,
 938                                  &raw_cp->cp_ifile_inode, 1);
 939         nilfs_cpfile_put_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, bh_cp);
 940         return 0;
 941
 942  failed_ibh:
 943         return err;
 944 }
 945
 946 static void nilfs_fill_in_file_bmap(struct inode *ifile,
 947                                     struct nilfs_inode_info *ii)
 948
 949 {
 950         struct buffer_head *ibh;
 951         struct nilfs_inode *raw_inode;
 952
 953         if (test_bit(NILFS_I_BMAP, &ii->i_state)) {
 954                 ibh = ii->i_bh;
 955                 BUG_ON(!ibh);
 956                 raw_inode = nilfs_ifile_map_inode(ifile, ii->vfs_inode.i_ino,
 957                                                   ibh);
 958                 nilfs_bmap_write(ii->i_bmap, raw_inode);
 959                 nilfs_ifile_unmap_inode(ifile, ii->vfs_inode.i_ino, ibh);
 960         }
 961 }
 962
 963 static void nilfs_segctor_fill_in_file_bmap(struct nilfs_sc_info *sci)
 964 {
 965         struct nilfs_inode_info *ii;
 966
 967         list_for_each_entry(ii, &sci->sc_dirty_files, i_dirty) {
 968                 nilfs_fill_in_file_bmap(sci->sc_root->ifile, ii);
 969                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
 970         }
 971 }
 972
 973 static void nilfs_segctor_fill_in_super_root(struct nilfs_sc_info *sci,
 974                                              struct the_nilfs *nilfs)
 975 {
 976         struct buffer_head *bh_sr;
 977         struct nilfs_super_root *raw_sr;
 978         unsigned int isz, srsz;
 979
 980         bh_sr = NILFS_LAST_SEGBUF(&sci->sc_segbufs)->sb_super_root;
 981         raw_sr = (struct nilfs_super_root *)bh_sr->b_data;
 982         isz = nilfs->ns_inode_size;
 983         srsz = NILFS_SR_BYTES(isz);
 984
 985         raw_sr->sr_bytes = cpu_to_le16(srsz);
 986         raw_sr->sr_nongc_ctime
 987                 = cpu_to_le64(nilfs_doing_gc() ?
 988                               nilfs->ns_nongc_ctime : sci->sc_seg_ctime);
 989         raw_sr->sr_flags = 0;
 990
 991         nilfs_write_inode_common(nilfs->ns_dat, (void *)raw_sr +
 992                                  NILFS_SR_DAT_OFFSET(isz), 1);
 993         nilfs_write_inode_common(nilfs->ns_cpfile, (void *)raw_sr +
 994                                  NILFS_SR_CPFILE_OFFSET(isz), 1);
 995         nilfs_write_inode_common(nilfs->ns_sufile, (void *)raw_sr +
 996                                  NILFS_SR_SUFILE_OFFSET(isz), 1);
 997         memset((void *)raw_sr + srsz, 0, nilfs->ns_blocksize - srsz);
 998 }
 999
1000 static void nilfs_redirty_inodes(struct list_head *head)
1001 {
1002         struct nilfs_inode_info *ii;
1003
1004         list_for_each_entry(ii, head, i_dirty) {
1005                 if (test_bit(NILFS_I_COLLECTED, &ii->i_state))
1006                         clear_bit(NILFS_I_COLLECTED, &ii->i_state);
1007         }
1008 }
1009
1010 static void nilfs_drop_collected_inodes(struct list_head *head)
1011 {
1012         struct nilfs_inode_info *ii;
1013
1014         list_for_each_entry(ii, head, i_dirty) {
1015                 if (!test_and_clear_bit(NILFS_I_COLLECTED, &ii->i_state))
1016                         continue;
1017
1018                 clear_bit(NILFS_I_INODE_SYNC, &ii->i_state);
1019                 set_bit(NILFS_I_UPDATED, &ii->i_state);
1020         }
1021 }
1022
1023 static int nilfs_segctor_apply_buffers(struct nilfs_sc_info *sci,
1024                                        struct inode *inode,
1025                                        struct list_head *listp,
1026                                        int (*collect)(struct nilfs_sc_info *,
1027                                                       struct buffer_head *,
1028                                                       struct inode *))
1029 {
1030         struct buffer_head *bh, *n;
1031         int err = 0;
1032
1033         if (collect) {
1034                 list_for_each_entry_safe(bh, n, listp, b_assoc_buffers) {
1035                         list_del_init(&bh->b_assoc_buffers);
1036                         err = collect(sci, bh, inode);
1037                         brelse(bh);
1038                         if (unlikely(err))
1039                                 goto dispose_buffers;
1040                 }
1041                 return 0;
1042         }
1043
1044  dispose_buffers:
1045         while (!list_empty(listp)) {
1046                 bh = list_first_entry(listp, struct buffer_head,
1047                                       b_assoc_buffers);
1048                 list_del_init(&bh->b_assoc_buffers);
1049                 brelse(bh);
1050         }
1051         return err;
1052 }
1053
1054 static size_t nilfs_segctor_buffer_rest(struct nilfs_sc_info *sci)
1055 {
1056         /* Remaining number of blocks within segment buffer */
1057         return sci->sc_segbuf_nblocks -
1058                 (sci->sc_nblk_this_inc + sci->sc_curseg->sb_sum.nblocks);
1059 }
1060
1061 static int nilfs_segctor_scan_file(struct nilfs_sc_info *sci,
1062                                    struct inode *inode,
1063                                    const struct nilfs_sc_operations *sc_ops)
1064 {
1065         LIST_HEAD(data_buffers);
1066         LIST_HEAD(node_buffers);
1067         int err;
1068
1069         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1070                 size_t n, rest = nilfs_segctor_buffer_rest(sci);
1071
1072                 n = nilfs_lookup_dirty_data_buffers(
1073                         inode, &data_buffers, rest + 1, 0, LLONG_MAX);
1074                 if (n > rest) {
1075                         err = nilfs_segctor_apply_buffers(
1076                                 sci, inode, &data_buffers,
1077                                 sc_ops->collect_data);
1078                         BUG_ON(!err); /* always receive -E2BIG or true error */
1079                         goto break_or_fail;
1080                 }
1081         }
1082         nilfs_lookup_dirty_node_buffers(inode, &node_buffers);
1083
1084         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1085                 err = nilfs_segctor_apply_buffers(
1086                         sci, inode, &data_buffers, sc_ops->collect_data);
1087                 if (unlikely(err)) {
1088                         /* dispose node list */
1089                         nilfs_segctor_apply_buffers(
1090                                 sci, inode, &node_buffers, NULL);
1091                         goto break_or_fail;
1092                 }
1093                 sci->sc_stage.flags |= NILFS_CF_NODE;
1094         }
1095         /* Collect node */
1096         err = nilfs_segctor_apply_buffers(
1097                 sci, inode, &node_buffers, sc_ops->collect_node);
1098         if (unlikely(err))
1099                 goto break_or_fail;
1100
1101         nilfs_bmap_lookup_dirty_buffers(NILFS_I(inode)->i_bmap, &node_buffers);
1102         err = nilfs_segctor_apply_buffers(
1103                 sci, inode, &node_buffers, sc_ops->collect_bmap);
1104         if (unlikely(err))
1105                 goto break_or_fail;
1106
1107         nilfs_segctor_end_finfo(sci, inode);
1108         sci->sc_stage.flags &= ~NILFS_CF_NODE;
1109
1110  break_or_fail:
1111         return err;
1112 }
1113
1114 static int nilfs_segctor_scan_file_dsync(struct nilfs_sc_info *sci,
1115                                          struct inode *inode)
1116 {
1117         LIST_HEAD(data_buffers);
1118         size_t n, rest = nilfs_segctor_buffer_rest(sci);
1119         int err;
1120
1121         n = nilfs_lookup_dirty_data_buffers(inode, &data_buffers, rest + 1,
1122                                             sci->sc_dsync_start,
1123                                             sci->sc_dsync_end);
1124
1125         err = nilfs_segctor_apply_buffers(sci, inode, &data_buffers,
1126                                           nilfs_collect_file_data);
1127         if (!err) {
1128                 nilfs_segctor_end_finfo(sci, inode);
1129                 BUG_ON(n > rest);
1130                 /* always receive -E2BIG or true error if n > rest */
1131         }
1132         return err;
1133 }
1134
1135 static int nilfs_segctor_collect_blocks(struct nilfs_sc_info *sci, int mode)
1136 {
1137         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1138         struct list_head *head;
1139         struct nilfs_inode_info *ii;
1140         size_t ndone;
1141         int err = 0;
1142
1143         switch (nilfs_sc_cstage_get(sci)) {
1144         case NILFS_ST_INIT:
1145                 /* Pre-processes */
1146                 sci->sc_stage.flags = 0;
1147
1148                 if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags)) {
1149                         sci->sc_nblk_inc = 0;
1150                         sci->sc_curseg->sb_sum.flags = NILFS_SS_LOGBGN;
1151                         if (mode == SC_LSEG_DSYNC) {
1152                                 nilfs_sc_cstage_set(sci, NILFS_ST_DSYNC);
1153                                 goto dsync_mode;
1154                         }
1155                 }
1156
1157                 sci->sc_stage.dirty_file_ptr = NULL;
1158                 sci->sc_stage.gc_inode_ptr = NULL;
1159                 if (mode == SC_FLUSH_DAT) {
1160                         nilfs_sc_cstage_set(sci, NILFS_ST_DAT);
1161                         goto dat_stage;
1162                 }
1163                 nilfs_sc_cstage_inc(sci);  /* Fall through */
1164         case NILFS_ST_GC:
1165                 if (nilfs_doing_gc()) {
1166                         head = &sci->sc_gc_inodes;
1167                         ii = list_prepare_entry(sci->sc_stage.gc_inode_ptr,
1168                                                 head, i_dirty);
1169                         list_for_each_entry_continue(ii, head, i_dirty) {
1170                                 err = nilfs_segctor_scan_file(
1171                                         sci, &ii->vfs_inode,
1172                                         &nilfs_sc_file_ops);
1173                                 if (unlikely(err)) {
1174                                         sci->sc_stage.gc_inode_ptr = list_entry(
1175                                                 ii->i_dirty.prev,
1176                                                 struct nilfs_inode_info,
1177                                                 i_dirty);
1178                                         goto break_or_fail;
1179                                 }
1180                                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
1181                         }
1182                         sci->sc_stage.gc_inode_ptr = NULL;
1183                 }
1184                 nilfs_sc_cstage_inc(sci);  /* Fall through */
1185         case NILFS_ST_FILE:
1186                 head = &sci->sc_dirty_files;
1187                 ii = list_prepare_entry(sci->sc_stage.dirty_file_ptr, head,
1188                                         i_dirty);
1189                 list_for_each_entry_continue(ii, head, i_dirty) {
1190                         clear_bit(NILFS_I_DIRTY, &ii->i_state);
1191
1192                         err = nilfs_segctor_scan_file(sci, &ii->vfs_inode,
1193                                                       &nilfs_sc_file_ops);
1194                         if (unlikely(err)) {
1195                                 sci->sc_stage.dirty_file_ptr =
1196                                         list_entry(ii->i_dirty.prev,
1197                                                    struct nilfs_inode_info,
1198                                                    i_dirty);
1199                                 goto break_or_fail;
1200                         }
1201                         /* sci->sc_stage.dirty_file_ptr = NILFS_I(inode); */
1202                         /* XXX: required ? */
1203                 }
1204                 sci->sc_stage.dirty_file_ptr = NULL;
1205                 if (mode == SC_FLUSH_FILE) {
1206                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1207                         return 0;
1208                 }
1209                 nilfs_sc_cstage_inc(sci);
1210                 sci->sc_stage.flags |= NILFS_CF_IFILE_STARTED;
1211                 /* Fall through */
1212         case NILFS_ST_IFILE:
1213                 err = nilfs_segctor_scan_file(sci, sci->sc_root->ifile,
1214                                               &nilfs_sc_file_ops);
1215                 if (unlikely(err))
1216                         break;
1217                 nilfs_sc_cstage_inc(sci);
1218                 /* Creating a checkpoint */
1219                 err = nilfs_segctor_create_checkpoint(sci);
1220                 if (unlikely(err))
1221                         break;
1222                 /* Fall through */
1223         case NILFS_ST_CPFILE:
1224                 err = nilfs_segctor_scan_file(sci, nilfs->ns_cpfile,
1225                                               &nilfs_sc_file_ops);
1226                 if (unlikely(err))
1227                         break;
1228                 nilfs_sc_cstage_inc(sci);  /* Fall through */
1229         case NILFS_ST_SUFILE:
1230                 err = nilfs_sufile_freev(nilfs->ns_sufile, sci->sc_freesegs,
1231                                          sci->sc_nfreesegs, &ndone);
1232                 if (unlikely(err)) {
1233                         nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1234                                                   sci->sc_freesegs, ndone,
1235                                                   NULL);
1236                         break;
1237                 }
1238                 sci->sc_stage.flags |= NILFS_CF_SUFREED;
1239
1240                 err = nilfs_segctor_scan_file(sci, nilfs->ns_sufile,
1241                                               &nilfs_sc_file_ops);
1242                 if (unlikely(err))
1243                         break;
1244                 nilfs_sc_cstage_inc(sci);  /* Fall through */
1245         case NILFS_ST_DAT:
1246  dat_stage:
1247                 err = nilfs_segctor_scan_file(sci, nilfs->ns_dat,
1248                                               &nilfs_sc_dat_ops);
1249                 if (unlikely(err))
1250                         break;
1251                 if (mode == SC_FLUSH_DAT) {
1252                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1253                         return 0;
1254                 }
1255                 nilfs_sc_cstage_inc(sci);  /* Fall through */
1256         case NILFS_ST_SR:
1257                 if (mode == SC_LSEG_SR) {
1258                         /* Appending a super root */
1259                         err = nilfs_segctor_add_super_root(sci);
1260                         if (unlikely(err))
1261                                 break;
1262                 }
1263                 /* End of a logical segment */
1264                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1265                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1266                 return 0;
1267         case NILFS_ST_DSYNC:
1268  dsync_mode:
1269                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_SYNDT;
1270                 ii = sci->sc_dsync_inode;
1271                 if (!test_bit(NILFS_I_BUSY, &ii->i_state))
1272                         break;
1273
1274                 err = nilfs_segctor_scan_file_dsync(sci, &ii->vfs_inode);
1275                 if (unlikely(err))
1276                         break;
1277                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1278                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1279                 return 0;
1280         case NILFS_ST_DONE:
1281                 return 0;
1282         default:
1283                 BUG();
1284         }
1285
1286  break_or_fail:
1287         return err;
1288 }
1289
1290 /**
1291  * nilfs_segctor_begin_construction - setup segment buffer to make a new log
1292  * @sci: nilfs_sc_info
1293  * @nilfs: nilfs object
1294  */
1295 static int nilfs_segctor_begin_construction(struct nilfs_sc_info *sci,
1296                                             struct the_nilfs *nilfs)
1297 {
1298         struct nilfs_segment_buffer *segbuf, *prev;
1299         __u64 nextnum;
1300         int err, alloc = 0;
1301
1302         segbuf = nilfs_segbuf_new(sci->sc_super);
1303         if (unlikely(!segbuf))
1304                 return -ENOMEM;
1305
1306         if (list_empty(&sci->sc_write_logs)) {
1307                 nilfs_segbuf_map(segbuf, nilfs->ns_segnum,
1308                                  nilfs->ns_pseg_offset, nilfs);
1309                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1310                         nilfs_shift_to_next_segment(nilfs);
1311                         nilfs_segbuf_map(segbuf, nilfs->ns_segnum, 0, nilfs);
1312                 }
1313
1314                 segbuf->sb_sum.seg_seq = nilfs->ns_seg_seq;
1315                 nextnum = nilfs->ns_nextnum;
1316
1317                 if (nilfs->ns_segnum == nilfs->ns_nextnum)
1318                         /* Start from the head of a new full segment */
1319                         alloc++;
1320         } else {
1321                 /* Continue logs */
1322                 prev = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1323                 nilfs_segbuf_map_cont(segbuf, prev);
1324                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq;
1325                 nextnum = prev->sb_nextnum;
1326
1327                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1328                         nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1329                         segbuf->sb_sum.seg_seq++;
1330                         alloc++;
1331                 }
1332         }
1333
1334         err = nilfs_sufile_mark_dirty(nilfs->ns_sufile, segbuf->sb_segnum);
1335         if (err)
1336                 goto failed;
1337
1338         if (alloc) {
1339                 err = nilfs_sufile_alloc(nilfs->ns_sufile, &nextnum);
1340                 if (err)
1341                         goto failed;
1342         }
1343         nilfs_segbuf_set_next_segnum(segbuf, nextnum, nilfs);
1344
1345         BUG_ON(!list_empty(&sci->sc_segbufs));
1346         list_add_tail(&segbuf->sb_list, &sci->sc_segbufs);
1347         sci->sc_segbuf_nblocks = segbuf->sb_rest_blocks;
1348         return 0;
1349
1350  failed:
1351         nilfs_segbuf_free(segbuf);
1352         return err;
1353 }
1354
1355 static int nilfs_segctor_extend_segments(struct nilfs_sc_info *sci,
1356                                          struct the_nilfs *nilfs, int nadd)
1357 {
1358         struct nilfs_segment_buffer *segbuf, *prev;
1359         struct inode *sufile = nilfs->ns_sufile;
1360         __u64 nextnextnum;
1361         LIST_HEAD(list);
1362         int err, ret, i;
1363
1364         prev = NILFS_LAST_SEGBUF(&sci->sc_segbufs);
1365         /*
1366          * Since the segment specified with nextnum might be allocated during
1367          * the previous construction, the buffer including its segusage may
1368          * not be dirty.  The following call ensures that the buffer is dirty
1369          * and will pin the buffer on memory until the sufile is written.
1370          */
1371         err = nilfs_sufile_mark_dirty(sufile, prev->sb_nextnum);
1372         if (unlikely(err))
1373                 return err;
1374
1375         for (i = 0; i < nadd; i++) {
1376                 /* extend segment info */
1377                 err = -ENOMEM;
1378                 segbuf = nilfs_segbuf_new(sci->sc_super);
1379                 if (unlikely(!segbuf))
1380                         goto failed;
1381
1382                 /* map this buffer to region of segment on-disk */
1383                 nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1384                 sci->sc_segbuf_nblocks += segbuf->sb_rest_blocks;
1385
1386                 /* allocate the next next full segment */
1387                 err = nilfs_sufile_alloc(sufile, &nextnextnum);
1388                 if (unlikely(err))
1389                         goto failed_segbuf;
1390
1391                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq + 1;
1392                 nilfs_segbuf_set_next_segnum(segbuf, nextnextnum, nilfs);
1393
1394                 list_add_tail(&segbuf->sb_list, &list);
1395                 prev = segbuf;
1396         }
1397         list_splice_tail(&list, &sci->sc_segbufs);
1398         return 0;
1399
1400  failed_segbuf:
1401         nilfs_segbuf_free(segbuf);
1402  failed:
1403         list_for_each_entry(segbuf, &list, sb_list) {
1404                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1405                 WARN_ON(ret); /* never fails */
1406         }
1407         nilfs_destroy_logs(&list);
1408         return err;
1409 }
1410
1411 static void nilfs_free_incomplete_logs(struct list_head *logs,
1412                                        struct the_nilfs *nilfs)
1413 {
1414         struct nilfs_segment_buffer *segbuf, *prev;
1415         struct inode *sufile = nilfs->ns_sufile;
1416         int ret;
1417
1418         segbuf = NILFS_FIRST_SEGBUF(logs);
1419         if (nilfs->ns_nextnum != segbuf->sb_nextnum) {
1420                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1421                 WARN_ON(ret); /* never fails */
1422         }
1423         if (atomic_read(&segbuf->sb_err)) {
1424                 /* Case 1: The first segment failed */
1425                 if (segbuf->sb_pseg_start != segbuf->sb_fseg_start)
1426                         /*
1427                          * Case 1a:  Partial segment appended into an existing
1428                          * segment
1429                          */
1430                         nilfs_terminate_segment(nilfs, segbuf->sb_fseg_start,
1431                                                 segbuf->sb_fseg_end);
1432                 else /* Case 1b:  New full segment */
1433                         set_nilfs_discontinued(nilfs);
1434         }
1435
1436         prev = segbuf;
1437         list_for_each_entry_continue(segbuf, logs, sb_list) {
1438                 if (prev->sb_nextnum != segbuf->sb_nextnum) {
1439                         ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1440                         WARN_ON(ret); /* never fails */
1441                 }
1442                 if (atomic_read(&segbuf->sb_err) &&
1443                     segbuf->sb_segnum != nilfs->ns_nextnum)
1444                         /* Case 2: extended segment (!= next) failed */
1445                         nilfs_sufile_set_error(sufile, segbuf->sb_segnum);
1446                 prev = segbuf;
1447         }
1448 }
1449
1450 static void nilfs_segctor_update_segusage(struct nilfs_sc_info *sci,
1451                                           struct inode *sufile)
1452 {
1453         struct nilfs_segment_buffer *segbuf;
1454         unsigned long live_blocks;
1455         int ret;
1456
1457         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1458                 live_blocks = segbuf->sb_sum.nblocks +
1459                         (segbuf->sb_pseg_start - segbuf->sb_fseg_start);
1460                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1461                                                      live_blocks,
1462                                                      sci->sc_seg_ctime);
1463                 WARN_ON(ret); /* always succeed because the segusage is dirty */
1464         }
1465 }
1466
1467 static void nilfs_cancel_segusage(struct list_head *logs, struct inode *sufile)
1468 {
1469         struct nilfs_segment_buffer *segbuf;
1470         int ret;
1471
1472         segbuf = NILFS_FIRST_SEGBUF(logs);
1473         ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1474                                              segbuf->sb_pseg_start -
1475                                              segbuf->sb_fseg_start, 0);
1476         WARN_ON(ret); /* always succeed because the segusage is dirty */
1477
1478         list_for_each_entry_continue(segbuf, logs, sb_list) {
1479                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1480                                                      0, 0);
1481                 WARN_ON(ret); /* always succeed */
1482         }
1483 }
1484
1485 static void nilfs_segctor_truncate_segments(struct nilfs_sc_info *sci,
1486                                             struct nilfs_segment_buffer *last,
1487                                             struct inode *sufile)
1488 {
1489         struct nilfs_segment_buffer *segbuf = last;
1490         int ret;
1491
1492         list_for_each_entry_continue(segbuf, &sci->sc_segbufs, sb_list) {
1493                 sci->sc_segbuf_nblocks -= segbuf->sb_rest_blocks;
1494                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1495                 WARN_ON(ret);
1496         }
1497         nilfs_truncate_logs(&sci->sc_segbufs, last);
1498 }
1499
1500
1501 static int nilfs_segctor_collect(struct nilfs_sc_info *sci,
1502                                  struct the_nilfs *nilfs, int mode)
1503 {
1504         struct nilfs_cstage prev_stage = sci->sc_stage;
1505         int err, nadd = 1;
1506
1507         /* Collection retry loop */
1508         for (;;) {
1509                 sci->sc_nblk_this_inc = 0;
1510                 sci->sc_curseg = NILFS_FIRST_SEGBUF(&sci->sc_segbufs);
1511
1512                 err = nilfs_segctor_reset_segment_buffer(sci);
1513                 if (unlikely(err))
1514                         goto failed;
1515
1516                 err = nilfs_segctor_collect_blocks(sci, mode);
1517                 sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
1518                 if (!err)
1519                         break;
1520
1521                 if (unlikely(err != -E2BIG))
1522                         goto failed;
1523
1524                 /* The current segment is filled up */
1525                 if (mode != SC_LSEG_SR ||
1526                     nilfs_sc_cstage_get(sci) < NILFS_ST_CPFILE)
1527                         break;
1528
1529                 nilfs_clear_logs(&sci->sc_segbufs);
1530
1531                 if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1532                         err = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1533                                                         sci->sc_freesegs,
1534                                                         sci->sc_nfreesegs,
1535                                                         NULL);
1536                         WARN_ON(err); /* do not happen */
1537                         sci->sc_stage.flags &= ~NILFS_CF_SUFREED;
1538                 }
1539
1540                 err = nilfs_segctor_extend_segments(sci, nilfs, nadd);
1541                 if (unlikely(err))
1542                         return err;
1543
1544                 nadd = min_t(int, nadd << 1, SC_MAX_SEGDELTA);
1545                 sci->sc_stage = prev_stage;
1546         }
1547         nilfs_segctor_truncate_segments(sci, sci->sc_curseg, nilfs->ns_sufile);
1548         return 0;
1549
1550  failed:
1551         return err;
1552 }
1553
1554 static void nilfs_list_replace_buffer(struct buffer_head *old_bh,
1555                                       struct buffer_head *new_bh)
1556 {
1557         BUG_ON(!list_empty(&new_bh->b_assoc_buffers));
1558
1559         list_replace_init(&old_bh->b_assoc_buffers, &new_bh->b_assoc_buffers);
1560         /* The caller must release old_bh */
1561 }
1562
1563 static int
1564 nilfs_segctor_update_payload_blocknr(struct nilfs_sc_info *sci,
1565                                      struct nilfs_segment_buffer *segbuf,
1566                                      int mode)
1567 {
1568         struct inode *inode = NULL;
1569         sector_t blocknr;
1570         unsigned long nfinfo = segbuf->sb_sum.nfinfo;
1571         unsigned long nblocks = 0, ndatablk = 0;
1572         const struct nilfs_sc_operations *sc_op = NULL;
1573         struct nilfs_segsum_pointer ssp;
1574         struct nilfs_finfo *finfo = NULL;
1575         union nilfs_binfo binfo;
1576         struct buffer_head *bh, *bh_org;
1577         ino_t ino = 0;
1578         int err = 0;
1579
1580         if (!nfinfo)
1581                 goto out;
1582
1583         blocknr = segbuf->sb_pseg_start + segbuf->sb_sum.nsumblk;
1584         ssp.bh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
1585         ssp.offset = sizeof(struct nilfs_segment_summary);
1586
1587         list_for_each_entry(bh, &segbuf->sb_payload_buffers, b_assoc_buffers) {
1588                 if (bh == segbuf->sb_super_root)
1589                         break;
1590                 if (!finfo) {
1591                         finfo = nilfs_segctor_map_segsum_entry(
1592                                 sci, &ssp, sizeof(*finfo));
1593                         ino = le64_to_cpu(finfo->fi_ino);
1594                         nblocks = le32_to_cpu(finfo->fi_nblocks);
1595                         ndatablk = le32_to_cpu(finfo->fi_ndatablk);
1596
1597                         inode = bh->b_page->mapping->host;
1598
1599                         if (mode == SC_LSEG_DSYNC)
1600                                 sc_op = &nilfs_sc_dsync_ops;
1601                         else if (ino == NILFS_DAT_INO)
1602                                 sc_op = &nilfs_sc_dat_ops;
1603                         else /* file blocks */
1604                                 sc_op = &nilfs_sc_file_ops;
1605                 }
1606                 bh_org = bh;
1607                 get_bh(bh_org);
1608                 err = nilfs_bmap_assign(NILFS_I(inode)->i_bmap, &bh, blocknr,
1609                                         &binfo);
1610                 if (bh != bh_org)
1611                         nilfs_list_replace_buffer(bh_org, bh);
1612                 brelse(bh_org);
1613                 if (unlikely(err))
1614                         goto failed_bmap;
1615
1616                 if (ndatablk > 0)
1617                         sc_op->write_data_binfo(sci, &ssp, &binfo);
1618                 else
1619                         sc_op->write_node_binfo(sci, &ssp, &binfo);
1620
1621                 blocknr++;
1622                 if (--nblocks == 0) {
1623                         finfo = NULL;
1624                         if (--nfinfo == 0)
1625                                 break;
1626                 } else if (ndatablk > 0)
1627                         ndatablk--;
1628         }
1629  out:
1630         return 0;
1631
1632  failed_bmap:
1633         return err;
1634 }
1635
1636 static int nilfs_segctor_assign(struct nilfs_sc_info *sci, int mode)
1637 {
1638         struct nilfs_segment_buffer *segbuf;
1639         int err;
1640
1641         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1642                 err = nilfs_segctor_update_payload_blocknr(sci, segbuf, mode);
1643                 if (unlikely(err))
1644                         return err;
1645                 nilfs_segbuf_fill_in_segsum(segbuf);
1646         }
1647         return 0;
1648 }
1649
1650 static void nilfs_begin_page_io(struct page *page)
1651 {
1652         if (!page || PageWriteback(page))
1653                 /*
1654                  * For split b-tree node pages, this function may be called
1655                  * twice.  We ignore the 2nd or later calls by this check.
1656                  */
1657                 return;
1658
1659         lock_page(page);
1660         clear_page_dirty_for_io(page);
1661         set_page_writeback(page);
1662         unlock_page(page);
1663 }
1664
1665 static void nilfs_segctor_prepare_write(struct nilfs_sc_info *sci)
1666 {
1667         struct nilfs_segment_buffer *segbuf;
1668         struct page *bd_page = NULL, *fs_page = NULL;
1669
1670         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1671                 struct buffer_head *bh;
1672
1673                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1674                                     b_assoc_buffers) {
1675                         if (bh->b_page != bd_page) {
1676                                 if (bd_page) {
1677                                         lock_page(bd_page);
1678                                         clear_page_dirty_for_io(bd_page);
1679                                         set_page_writeback(bd_page);
1680                                         unlock_page(bd_page);
1681                                 }
1682                                 bd_page = bh->b_page;
1683                         }
1684                 }
1685
1686                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1687                                     b_assoc_buffers) {
1688                         set_buffer_async_write(bh);
1689                         if (bh == segbuf->sb_super_root) {
1690                                 if (bh->b_page != bd_page) {
1691                                         lock_page(bd_page);
1692                                         clear_page_dirty_for_io(bd_page);
1693                                         set_page_writeback(bd_page);
1694                                         unlock_page(bd_page);
1695                                         bd_page = bh->b_page;
1696                                 }
1697                                 break;
1698                         }
1699                         if (bh->b_page != fs_page) {
1700                                 nilfs_begin_page_io(fs_page);
1701                                 fs_page = bh->b_page;
1702                         }
1703                 }
1704         }
1705         if (bd_page) {
1706                 lock_page(bd_page);
1707                 clear_page_dirty_for_io(bd_page);
1708                 set_page_writeback(bd_page);
1709                 unlock_page(bd_page);
1710         }
1711         nilfs_begin_page_io(fs_page);
1712 }
1713
1714 static int nilfs_segctor_write(struct nilfs_sc_info *sci,
1715                                struct the_nilfs *nilfs)
1716 {
1717         int ret;
1718
1719         ret = nilfs_write_logs(&sci->sc_segbufs, nilfs);
1720         list_splice_tail_init(&sci->sc_segbufs, &sci->sc_write_logs);
1721         return ret;
1722 }
1723
1724 static void nilfs_end_page_io(struct page *page, int err)
1725 {
1726         if (!page)
1727                 return;
1728
1729         if (buffer_nilfs_node(page_buffers(page)) && !PageWriteback(page)) {
1730                 /*
1731                  * For b-tree node pages, this function may be called twice
1732                  * or more because they might be split in a segment.
1733                  */
1734                 if (PageDirty(page)) {
1735                         /*
1736                          * For pages holding split b-tree node buffers, dirty
1737                          * flag on the buffers may be cleared discretely.
1738                          * In that case, the page is once redirtied for
1739                          * remaining buffers, and it must be cancelled if
1740                          * all the buffers get cleaned later.
1741                          */
1742                         lock_page(page);
1743                         if (nilfs_page_buffers_clean(page))
1744                                 __nilfs_clear_page_dirty(page);
1745                         unlock_page(page);
1746                 }
1747                 return;
1748         }
1749
1750         if (!err) {
1751                 if (!nilfs_page_buffers_clean(page))
1752                         __set_page_dirty_nobuffers(page);
1753                 ClearPageError(page);
1754         } else {
1755                 __set_page_dirty_nobuffers(page);
1756                 SetPageError(page);
1757         }
1758
1759         end_page_writeback(page);
1760 }
1761
1762 static void nilfs_abort_logs(struct list_head *logs, int err)
1763 {
1764         struct nilfs_segment_buffer *segbuf;
1765         struct page *bd_page = NULL, *fs_page = NULL;
1766         struct buffer_head *bh;
1767
1768         if (list_empty(logs))
1769                 return;
1770
1771         list_for_each_entry(segbuf, logs, sb_list) {
1772                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1773                                     b_assoc_buffers) {
1774                         if (bh->b_page != bd_page) {
1775                                 if (bd_page)
1776                                         end_page_writeback(bd_page);
1777                                 bd_page = bh->b_page;
1778                         }
1779                 }
1780
1781                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1782                                     b_assoc_buffers) {
1783                         clear_buffer_async_write(bh);
1784                         if (bh == segbuf->sb_super_root) {
1785                                 if (bh->b_page != bd_page) {
1786                                         end_page_writeback(bd_page);
1787                                         bd_page = bh->b_page;
1788                                 }
1789                                 break;
1790                         }
1791                         if (bh->b_page != fs_page) {
1792                                 nilfs_end_page_io(fs_page, err);
1793                                 fs_page = bh->b_page;
1794                         }
1795                 }
1796         }
1797         if (bd_page)
1798                 end_page_writeback(bd_page);
1799
1800         nilfs_end_page_io(fs_page, err);
1801 }
1802
1803 static void nilfs_segctor_abort_construction(struct nilfs_sc_info *sci,
1804                                              struct the_nilfs *nilfs, int err)
1805 {
1806         LIST_HEAD(logs);
1807         int ret;
1808
1809         list_splice_tail_init(&sci->sc_write_logs, &logs);
1810         ret = nilfs_wait_on_logs(&logs);
1811         nilfs_abort_logs(&logs, ret ? : err);
1812
1813         list_splice_tail_init(&sci->sc_segbufs, &logs);
1814         nilfs_cancel_segusage(&logs, nilfs->ns_sufile);
1815         nilfs_free_incomplete_logs(&logs, nilfs);
1816
1817         if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1818                 ret = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1819                                                 sci->sc_freesegs,
1820                                                 sci->sc_nfreesegs,
1821                                                 NULL);
1822                 WARN_ON(ret); /* do not happen */
1823         }
1824
1825         nilfs_destroy_logs(&logs);
1826 }
1827
1828 static void nilfs_set_next_segment(struct the_nilfs *nilfs,
1829                                    struct nilfs_segment_buffer *segbuf)
1830 {
1831         nilfs->ns_segnum = segbuf->sb_segnum;
1832         nilfs->ns_nextnum = segbuf->sb_nextnum;
1833         nilfs->ns_pseg_offset = segbuf->sb_pseg_start - segbuf->sb_fseg_start
1834                 + segbuf->sb_sum.nblocks;
1835         nilfs->ns_seg_seq = segbuf->sb_sum.seg_seq;
1836         nilfs->ns_ctime = segbuf->sb_sum.ctime;
1837 }
1838
1839 static void nilfs_segctor_complete_write(struct nilfs_sc_info *sci)
1840 {
1841         struct nilfs_segment_buffer *segbuf;
1842         struct page *bd_page = NULL, *fs_page = NULL;
1843         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1844         int update_sr = false;
1845
1846         list_for_each_entry(segbuf, &sci->sc_write_logs, sb_list) {
1847                 struct buffer_head *bh;
1848
1849                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1850                                     b_assoc_buffers) {
1851                         set_buffer_uptodate(bh);
1852                         clear_buffer_dirty(bh);
1853                         if (bh->b_page != bd_page) {
1854                                 if (bd_page)
1855                                         end_page_writeback(bd_page);
1856                                 bd_page = bh->b_page;
1857                         }
1858                 }
1859                 /*
1860                  * We assume that the buffers which belong to the same page
1861                  * continue over the buffer list.
1862                  * Under this assumption, the last BHs of pages is
1863                  * identifiable by the discontinuity of bh->b_page
1864                  * (page != fs_page).
1865                  *
1866                  * For B-tree node blocks, however, this assumption is not
1867                  * guaranteed.  The cleanup code of B-tree node pages needs
1868                  * special care.
1869                  */
1870                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1871                                     b_assoc_buffers) {
1872                         const unsigned long set_bits = BIT(BH_Uptodate);
1873                         const unsigned long clear_bits =
1874                                 (BIT(BH_Dirty) | BIT(BH_Async_Write) |
1875                                  BIT(BH_Delay) | BIT(BH_NILFS_Volatile) |
1876                                  BIT(BH_NILFS_Redirected));
1877
1878                         set_mask_bits(&bh->b_state, clear_bits, set_bits);
1879                         if (bh == segbuf->sb_super_root) {
1880                                 if (bh->b_page != bd_page) {
1881                                         end_page_writeback(bd_page);
1882                                         bd_page = bh->b_page;
1883                                 }
1884                                 update_sr = true;
1885                                 break;
1886                         }
1887                         if (bh->b_page != fs_page) {
1888                                 nilfs_end_page_io(fs_page, 0);
1889                                 fs_page = bh->b_page;
1890                         }
1891                 }
1892
1893                 if (!nilfs_segbuf_simplex(segbuf)) {
1894                         if (segbuf->sb_sum.flags & NILFS_SS_LOGBGN) {
1895                                 set_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1896                                 sci->sc_lseg_stime = jiffies;
1897                         }
1898                         if (segbuf->sb_sum.flags & NILFS_SS_LOGEND)
1899                                 clear_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1900                 }
1901         }
1902         /*
1903          * Since pages may continue over multiple segment buffers,
1904          * end of the last page must be checked outside of the loop.
1905          */
1906         if (bd_page)
1907                 end_page_writeback(bd_page);
1908
1909         nilfs_end_page_io(fs_page, 0);
1910
1911         nilfs_drop_collected_inodes(&sci->sc_dirty_files);
1912
1913         if (nilfs_doing_gc())
1914                 nilfs_drop_collected_inodes(&sci->sc_gc_inodes);
1915         else
1916                 nilfs->ns_nongc_ctime = sci->sc_seg_ctime;
1917
1918         sci->sc_nblk_inc += sci->sc_nblk_this_inc;
1919
1920         segbuf = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1921         nilfs_set_next_segment(nilfs, segbuf);
1922
1923         if (update_sr) {
1924                 nilfs->ns_flushed_device = 0;
1925                 nilfs_set_last_segment(nilfs, segbuf->sb_pseg_start,
1926                                        segbuf->sb_sum.seg_seq, nilfs->ns_cno++);
1927
1928                 clear_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
1929                 clear_bit(NILFS_SC_DIRTY, &sci->sc_flags);
1930                 set_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1931                 nilfs_segctor_clear_metadata_dirty(sci);
1932         } else
1933                 clear_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1934 }
1935
1936 static int nilfs_segctor_wait(struct nilfs_sc_info *sci)
1937 {
1938         int ret;
1939
1940         ret = nilfs_wait_on_logs(&sci->sc_write_logs);
1941         if (!ret) {
1942                 nilfs_segctor_complete_write(sci);
1943                 nilfs_destroy_logs(&sci->sc_write_logs);
1944         }
1945         return ret;
1946 }
1947
1948 static int nilfs_segctor_collect_dirty_files(struct nilfs_sc_info *sci,
1949                                              struct the_nilfs *nilfs)
1950 {
1951         struct nilfs_inode_info *ii, *n;
1952         struct inode *ifile = sci->sc_root->ifile;
1953
1954         spin_lock(&nilfs->ns_inode_lock);
1955  retry:
1956         list_for_each_entry_safe(ii, n, &nilfs->ns_dirty_files, i_dirty) {
1957                 if (!ii->i_bh) {
1958                         struct buffer_head *ibh;
1959                         int err;
1960
1961                         spin_unlock(&nilfs->ns_inode_lock);
1962                         err = nilfs_ifile_get_inode_block(
1963                                 ifile, ii->vfs_inode.i_ino, &ibh);
1964                         if (unlikely(err)) {
1965                                 nilfs_msg(sci->sc_super, KERN_WARNING,
1966                                           "log writer: error %d getting inode block (ino=%lu)",
1967                                           err, ii->vfs_inode.i_ino);
1968                                 return err;
1969                         }
1970                         spin_lock(&nilfs->ns_inode_lock);
1971                         if (likely(!ii->i_bh))
1972                                 ii->i_bh = ibh;
1973                         else
1974                                 brelse(ibh);
1975                         goto retry;
1976                 }
1977
1978                 // Always redirty the buffer to avoid race condition
1979                 mark_buffer_dirty(ii->i_bh);
1980                 nilfs_mdt_mark_dirty(ifile);
1981
1982                 clear_bit(NILFS_I_QUEUED, &ii->i_state);
1983                 set_bit(NILFS_I_BUSY, &ii->i_state);
1984                 list_move_tail(&ii->i_dirty, &sci->sc_dirty_files);
1985         }
1986         spin_unlock(&nilfs->ns_inode_lock);
1987
1988         return 0;
1989 }
1990
1991 static void nilfs_segctor_drop_written_files(struct nilfs_sc_info *sci,
1992                                              struct the_nilfs *nilfs)
1993 {
1994         struct nilfs_inode_info *ii, *n;
1995         int during_mount = !(sci->sc_super->s_flags & MS_ACTIVE);
1996         int defer_iput = false;
1997
1998         spin_lock(&nilfs->ns_inode_lock);
1999         list_for_each_entry_safe(ii, n, &sci->sc_dirty_files, i_dirty) {
2000                 if (!test_and_clear_bit(NILFS_I_UPDATED, &ii->i_state) ||
2001                     test_bit(NILFS_I_DIRTY, &ii->i_state))
2002                         continue;
2003
2004                 clear_bit(NILFS_I_BUSY, &ii->i_state);
2005                 brelse(ii->i_bh);
2006                 ii->i_bh = NULL;
2007                 list_del_init(&ii->i_dirty);
2008                 if (!ii->vfs_inode.i_nlink || during_mount) {
2009                         /*
2010                          * Defer calling iput() to avoid deadlocks if
2011                          * i_nlink == 0 or mount is not yet finished.
2012                          */
2013                         list_add_tail(&ii->i_dirty, &sci->sc_iput_queue);
2014                         defer_iput = true;
2015                 } else {
2016                         spin_unlock(&nilfs->ns_inode_lock);
2017                         iput(&ii->vfs_inode);
2018                         spin_lock(&nilfs->ns_inode_lock);
2019                 }
2020         }
2021         spin_unlock(&nilfs->ns_inode_lock);
2022
2023         if (defer_iput)
2024                 schedule_work(&sci->sc_iput_work);
2025 }
2026
2027 /*
2028  * Main procedure of segment constructor
2029  */
2030 static int nilfs_segctor_do_construct(struct nilfs_sc_info *sci, int mode)
2031 {
2032         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2033         int err;
2034
2035         nilfs_sc_cstage_set(sci, NILFS_ST_INIT);
2036         sci->sc_cno = nilfs->ns_cno;
2037
2038         err = nilfs_segctor_collect_dirty_files(sci, nilfs);
2039         if (unlikely(err))
2040                 goto out;
2041
2042         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
2043                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
2044
2045         if (nilfs_segctor_clean(sci))
2046                 goto out;
2047
2048         do {
2049                 sci->sc_stage.flags &= ~NILFS_CF_HISTORY_MASK;
2050
2051                 err = nilfs_segctor_begin_construction(sci, nilfs);
2052                 if (unlikely(err))
2053                         goto out;
2054
2055                 /* Update time stamp */
2056                 sci->sc_seg_ctime = get_seconds();
2057
2058                 err = nilfs_segctor_collect(sci, nilfs, mode);
2059                 if (unlikely(err))
2060                         goto failed;
2061
2062                 /* Avoid empty segment */
2063                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE &&
2064                     nilfs_segbuf_empty(sci->sc_curseg)) {
2065                         nilfs_segctor_abort_construction(sci, nilfs, 1);
2066                         goto out;
2067                 }
2068
2069                 err = nilfs_segctor_assign(sci, mode);
2070                 if (unlikely(err))
2071                         goto failed;
2072
2073                 if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2074                         nilfs_segctor_fill_in_file_bmap(sci);
2075
2076                 if (mode == SC_LSEG_SR &&
2077                     nilfs_sc_cstage_get(sci) >= NILFS_ST_CPFILE) {
2078                         err = nilfs_segctor_fill_in_checkpoint(sci);
2079                         if (unlikely(err))
2080                                 goto failed_to_write;
2081
2082                         nilfs_segctor_fill_in_super_root(sci, nilfs);
2083                 }
2084                 nilfs_segctor_update_segusage(sci, nilfs->ns_sufile);
2085
2086                 /* Write partial segments */
2087                 nilfs_segctor_prepare_write(sci);
2088
2089                 nilfs_add_checksums_on_logs(&sci->sc_segbufs,
2090                                             nilfs->ns_crc_seed);
2091
2092                 err = nilfs_segctor_write(sci, nilfs);
2093                 if (unlikely(err))
2094                         goto failed_to_write;
2095
2096                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE ||
2097                     nilfs->ns_blocksize_bits != PAGE_SHIFT) {
2098                         /*
2099                          * At this point, we avoid double buffering
2100                          * for blocksize < pagesize because page dirty
2101                          * flag is turned off during write and dirty
2102                          * buffers are not properly collected for
2103                          * pages crossing over segments.
2104                          */
2105                         err = nilfs_segctor_wait(sci);
2106                         if (err)
2107                                 goto failed_to_write;
2108                 }
2109         } while (nilfs_sc_cstage_get(sci) != NILFS_ST_DONE);
2110
2111  out:
2112         nilfs_segctor_drop_written_files(sci, nilfs);
2113         return err;
2114
2115  failed_to_write:
2116         if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2117                 nilfs_redirty_inodes(&sci->sc_dirty_files);
2118
2119  failed:
2120         if (nilfs_doing_gc())
2121                 nilfs_redirty_inodes(&sci->sc_gc_inodes);
2122         nilfs_segctor_abort_construction(sci, nilfs, err);
2123         goto out;
2124 }
2125
2126 /**
2127  * nilfs_segctor_start_timer - set timer of background write
2128  * @sci: nilfs_sc_info
2129  *
2130  * If the timer has already been set, it ignores the new request.
2131  * This function MUST be called within a section locking the segment
2132  * semaphore.
2133  */
2134 static void nilfs_segctor_start_timer(struct nilfs_sc_info *sci)
2135 {
2136         spin_lock(&sci->sc_state_lock);
2137         if (!(sci->sc_state & NILFS_SEGCTOR_COMMIT)) {
2138                 sci->sc_timer.expires = jiffies + sci->sc_interval;
2139                 add_timer(&sci->sc_timer);
2140                 sci->sc_state |= NILFS_SEGCTOR_COMMIT;
2141         }
2142         spin_unlock(&sci->sc_state_lock);
2143 }
2144
2145 static void nilfs_segctor_do_flush(struct nilfs_sc_info *sci, int bn)
2146 {
2147         spin_lock(&sci->sc_state_lock);
2148         if (!(sci->sc_flush_request & BIT(bn))) {
2149                 unsigned long prev_req = sci->sc_flush_request;
2150
2151                 sci->sc_flush_request |= BIT(bn);
2152                 if (!prev_req)
2153                         wake_up(&sci->sc_wait_daemon);
2154         }
2155         spin_unlock(&sci->sc_state_lock);
2156 }
2157
2158 /**
2159  * nilfs_flush_segment - trigger a segment construction for resource control
2160  * @sb: super block
2161  * @ino: inode number of the file to be flushed out.
2162  */
2163 void nilfs_flush_segment(struct super_block *sb, ino_t ino)
2164 {
2165         struct the_nilfs *nilfs = sb->s_fs_info;
2166         struct nilfs_sc_info *sci = nilfs->ns_writer;
2167
2168         if (!sci || nilfs_doing_construction())
2169                 return;
2170         nilfs_segctor_do_flush(sci, NILFS_MDT_INODE(sb, ino) ? ino : 0);
2171                                         /* assign bit 0 to data files */
2172 }
2173
2174 struct nilfs_segctor_wait_request {
2175         wait_queue_entry_t      wq;
2176         __u32           seq;
2177         int             err;
2178         atomic_t        done;
2179 };
2180
2181 static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
2182 {
2183         struct nilfs_segctor_wait_request wait_req;
2184         int err = 0;
2185
2186         spin_lock(&sci->sc_state_lock);
2187         init_wait(&wait_req.wq);
2188         wait_req.err = 0;
2189         atomic_set(&wait_req.done, 0);
2190         wait_req.seq = ++sci->sc_seq_request;
2191         spin_unlock(&sci->sc_state_lock);
2192
2193         init_waitqueue_entry(&wait_req.wq, current);
2194         add_wait_queue(&sci->sc_wait_request, &wait_req.wq);
2195         set_current_state(TASK_INTERRUPTIBLE);
2196         wake_up(&sci->sc_wait_daemon);
2197
2198         for (;;) {
2199                 if (atomic_read(&wait_req.done)) {
2200                         err = wait_req.err;
2201                         break;
2202                 }
2203                 if (!signal_pending(current)) {
2204                         schedule();
2205                         continue;
2206                 }
2207                 err = -ERESTARTSYS;
2208                 break;
2209         }
2210         finish_wait(&sci->sc_wait_request, &wait_req.wq);
2211         return err;
2212 }
2213
2214 static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err)
2215 {
2216         struct nilfs_segctor_wait_request *wrq, *n;
2217         unsigned long flags;
2218
2219         spin_lock_irqsave(&sci->sc_wait_request.lock, flags);
2220         list_for_each_entry_safe(wrq, n, &sci->sc_wait_request.head, wq.entry) {
2221                 if (!atomic_read(&wrq->done) &&
2222                     nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq)) {
2223                         wrq->err = err;
2224                         atomic_set(&wrq->done, 1);
2225                 }
2226                 if (atomic_read(&wrq->done)) {
2227                         wrq->wq.func(&wrq->wq,
2228                                      TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2229                                      0, NULL);
2230                 }
2231         }
2232         spin_unlock_irqrestore(&sci->sc_wait_request.lock, flags);
2233 }
2234
2235 /**
2236  * nilfs_construct_segment - construct a logical segment
2237  * @sb: super block
2238  *
2239  * Return Value: On success, 0 is retured. On errors, one of the following
2240  * negative error code is returned.
2241  *
2242  * %-EROFS - Read only filesystem.
2243  *
2244  * %-EIO - I/O error
2245  *
2246  * %-ENOSPC - No space left on device (only in a panic state).
2247  *
2248  * %-ERESTARTSYS - Interrupted.
2249  *
2250  * %-ENOMEM - Insufficient memory available.
2251  */
2252 int nilfs_construct_segment(struct super_block *sb)
2253 {
2254         struct the_nilfs *nilfs = sb->s_fs_info;
2255         struct nilfs_sc_info *sci = nilfs->ns_writer;
2256         struct nilfs_transaction_info *ti;
2257         int err;
2258
2259         if (sb_rdonly(sb) || unlikely(!sci))
2260                 return -EROFS;
2261
2262         /* A call inside transactions causes a deadlock. */
2263         BUG_ON((ti = current->journal_info) && ti->ti_magic == NILFS_TI_MAGIC);
2264
2265         err = nilfs_segctor_sync(sci);
2266         return err;
2267 }
2268
2269 /**
2270  * nilfs_construct_dsync_segment - construct a data-only logical segment
2271  * @sb: super block
2272  * @inode: inode whose data blocks should be written out
2273  * @start: start byte offset
2274  * @end: end byte offset (inclusive)
2275  *
2276  * Return Value: On success, 0 is retured. On errors, one of the following
2277  * negative error code is returned.
2278  *
2279  * %-EROFS - Read only filesystem.
2280  *
2281  * %-EIO - I/O error
2282  *
2283  * %-ENOSPC - No space left on device (only in a panic state).
2284  *
2285  * %-ERESTARTSYS - Interrupted.
2286  *
2287  * %-ENOMEM - Insufficient memory available.
2288  */
2289 int nilfs_construct_dsync_segment(struct super_block *sb, struct inode *inode,
2290                                   loff_t start, loff_t end)
2291 {
2292         struct the_nilfs *nilfs = sb->s_fs_info;
2293         struct nilfs_sc_info *sci = nilfs->ns_writer;
2294         struct nilfs_inode_info *ii;
2295         struct nilfs_transaction_info ti;
2296         int err = 0;
2297
2298         if (sb_rdonly(sb) || unlikely(!sci))
2299                 return -EROFS;
2300
2301         nilfs_transaction_lock(sb, &ti, 0);
2302
2303         ii = NILFS_I(inode);
2304         if (test_bit(NILFS_I_INODE_SYNC, &ii->i_state) ||
2305             nilfs_test_opt(nilfs, STRICT_ORDER) ||
2306             test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2307             nilfs_discontinued(nilfs)) {
2308                 nilfs_transaction_unlock(sb);
2309                 err = nilfs_segctor_sync(sci);
2310                 return err;
2311         }
2312
2313         spin_lock(&nilfs->ns_inode_lock);
2314         if (!test_bit(NILFS_I_QUEUED, &ii->i_state) &&
2315             !test_bit(NILFS_I_BUSY, &ii->i_state)) {
2316                 spin_unlock(&nilfs->ns_inode_lock);
2317                 nilfs_transaction_unlock(sb);
2318                 return 0;
2319         }
2320         spin_unlock(&nilfs->ns_inode_lock);
2321         sci->sc_dsync_inode = ii;
2322         sci->sc_dsync_start = start;
2323         sci->sc_dsync_end = end;
2324
2325         err = nilfs_segctor_do_construct(sci, SC_LSEG_DSYNC);
2326         if (!err)
2327                 nilfs->ns_flushed_device = 0;
2328
2329         nilfs_transaction_unlock(sb);
2330         return err;
2331 }
2332
2333 #define FLUSH_FILE_BIT  (0x1) /* data file only */
2334 #define FLUSH_DAT_BIT   BIT(NILFS_DAT_INO) /* DAT only */
2335
2336 /**
2337  * nilfs_segctor_accept - record accepted sequence count of log-write requests
2338  * @sci: segment constructor object
2339  */
2340 static void nilfs_segctor_accept(struct nilfs_sc_info *sci)
2341 {
2342         spin_lock(&sci->sc_state_lock);
2343         sci->sc_seq_accepted = sci->sc_seq_request;
2344         spin_unlock(&sci->sc_state_lock);
2345         del_timer_sync(&sci->sc_timer);
2346 }
2347
2348 /**
2349  * nilfs_segctor_notify - notify the result of request to caller threads
2350  * @sci: segment constructor object
2351  * @mode: mode of log forming
2352  * @err: error code to be notified
2353  */
2354 static void nilfs_segctor_notify(struct nilfs_sc_info *sci, int mode, int err)
2355 {
2356         /* Clear requests (even when the construction failed) */
2357         spin_lock(&sci->sc_state_lock);
2358
2359         if (mode == SC_LSEG_SR) {
2360                 sci->sc_state &= ~NILFS_SEGCTOR_COMMIT;
2361                 sci->sc_seq_done = sci->sc_seq_accepted;
2362                 nilfs_segctor_wakeup(sci, err);
2363                 sci->sc_flush_request = 0;
2364         } else {
2365                 if (mode == SC_FLUSH_FILE)
2366                         sci->sc_flush_request &= ~FLUSH_FILE_BIT;
2367                 else if (mode == SC_FLUSH_DAT)
2368                         sci->sc_flush_request &= ~FLUSH_DAT_BIT;
2369
2370                 /* re-enable timer if checkpoint creation was not done */
2371                 if ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2372                     time_before(jiffies, sci->sc_timer.expires))
2373                         add_timer(&sci->sc_timer);
2374         }
2375         spin_unlock(&sci->sc_state_lock);
2376 }
2377
2378 /**
2379  * nilfs_segctor_construct - form logs and write them to disk
2380  * @sci: segment constructor object
2381  * @mode: mode of log forming
2382  */
2383 static int nilfs_segctor_construct(struct nilfs_sc_info *sci, int mode)
2384 {
2385         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2386         struct nilfs_super_block **sbp;
2387         int err = 0;
2388
2389         nilfs_segctor_accept(sci);
2390
2391         if (nilfs_discontinued(nilfs))
2392                 mode = SC_LSEG_SR;
2393         if (!nilfs_segctor_confirm(sci))
2394                 err = nilfs_segctor_do_construct(sci, mode);
2395
2396         if (likely(!err)) {
2397                 if (mode != SC_FLUSH_DAT)
2398                         atomic_set(&nilfs->ns_ndirtyblks, 0);
2399                 if (test_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags) &&
2400                     nilfs_discontinued(nilfs)) {
2401                         down_write(&nilfs->ns_sem);
2402                         err = -EIO;
2403                         sbp = nilfs_prepare_super(sci->sc_super,
2404                                                   nilfs_sb_will_flip(nilfs));
2405                         if (likely(sbp)) {
2406                                 nilfs_set_log_cursor(sbp[0], nilfs);
2407                                 err = nilfs_commit_super(sci->sc_super,
2408                                                          NILFS_SB_COMMIT);
2409                         }
2410                         up_write(&nilfs->ns_sem);
2411                 }
2412         }
2413
2414         nilfs_segctor_notify(sci, mode, err);
2415         return err;
2416 }
2417
2418 static void nilfs_construction_timeout(unsigned long data)
2419 {
2420         struct task_struct *p = (struct task_struct *)data;
2421
2422         wake_up_process(p);
2423 }
2424
2425 static void
2426 nilfs_remove_written_gcinodes(struct the_nilfs *nilfs, struct list_head *head)
2427 {
2428         struct nilfs_inode_info *ii, *n;
2429
2430         list_for_each_entry_safe(ii, n, head, i_dirty) {
2431                 if (!test_bit(NILFS_I_UPDATED, &ii->i_state))
2432                         continue;
2433                 list_del_init(&ii->i_dirty);
2434                 truncate_inode_pages(&ii->vfs_inode.i_data, 0);
2435                 nilfs_btnode_cache_clear(ii->i_assoc_inode->i_mapping);
2436                 iput(&ii->vfs_inode);
2437         }
2438 }
2439
2440 int nilfs_clean_segments(struct super_block *sb, struct nilfs_argv *argv,
2441                          void **kbufs)
2442 {
2443         struct the_nilfs *nilfs = sb->s_fs_info;
2444         struct nilfs_sc_info *sci = nilfs->ns_writer;
2445         struct nilfs_transaction_info ti;
2446         int err;
2447
2448         if (unlikely(!sci))
2449                 return -EROFS;
2450
2451         nilfs_transaction_lock(sb, &ti, 1);
2452
2453         err = nilfs_mdt_save_to_shadow_map(nilfs->ns_dat);
2454         if (unlikely(err))
2455                 goto out_unlock;
2456
2457         err = nilfs_ioctl_prepare_clean_segments(nilfs, argv, kbufs);
2458         if (unlikely(err)) {
2459                 nilfs_mdt_restore_from_shadow_map(nilfs->ns_dat);
2460                 goto out_unlock;
2461         }
2462
2463         sci->sc_freesegs = kbufs[4];
2464         sci->sc_nfreesegs = argv[4].v_nmembs;
2465         list_splice_tail_init(&nilfs->ns_gc_inodes, &sci->sc_gc_inodes);
2466
2467         for (;;) {
2468                 err = nilfs_segctor_construct(sci, SC_LSEG_SR);
2469                 nilfs_remove_written_gcinodes(nilfs, &sci->sc_gc_inodes);
2470
2471                 if (likely(!err))
2472                         break;
2473
2474                 nilfs_msg(sb, KERN_WARNING, "error %d cleaning segments", err);
2475                 set_current_state(TASK_INTERRUPTIBLE);
2476                 schedule_timeout(sci->sc_interval);
2477         }
2478         if (nilfs_test_opt(nilfs, DISCARD)) {
2479                 int ret = nilfs_discard_segments(nilfs, sci->sc_freesegs,
2480                                                  sci->sc_nfreesegs);
2481                 if (ret) {
2482                         nilfs_msg(sb, KERN_WARNING,
2483                                   "error %d on discard request, turning discards off for the device",
2484                                   ret);
2485                         nilfs_clear_opt(nilfs, DISCARD);
2486                 }
2487         }
2488
2489  out_unlock:
2490         sci->sc_freesegs = NULL;
2491         sci->sc_nfreesegs = 0;
2492         nilfs_mdt_clear_shadow_map(nilfs->ns_dat);
2493         nilfs_transaction_unlock(sb);
2494         return err;
2495 }
2496
2497 static void nilfs_segctor_thread_construct(struct nilfs_sc_info *sci, int mode)
2498 {
2499         struct nilfs_transaction_info ti;
2500
2501         nilfs_transaction_lock(sci->sc_super, &ti, 0);
2502         nilfs_segctor_construct(sci, mode);
2503
2504         /*
2505          * Unclosed segment should be retried.  We do this using sc_timer.
2506          * Timeout of sc_timer will invoke complete construction which leads
2507          * to close the current logical segment.
2508          */
2509         if (test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags))
2510                 nilfs_segctor_start_timer(sci);
2511
2512         nilfs_transaction_unlock(sci->sc_super);
2513 }
2514
2515 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *sci)
2516 {
2517         int mode = 0;
2518
2519         spin_lock(&sci->sc_state_lock);
2520         mode = (sci->sc_flush_request & FLUSH_DAT_BIT) ?
2521                 SC_FLUSH_DAT : SC_FLUSH_FILE;
2522         spin_unlock(&sci->sc_state_lock);
2523
2524         if (mode) {
2525                 nilfs_segctor_do_construct(sci, mode);
2526
2527                 spin_lock(&sci->sc_state_lock);
2528                 sci->sc_flush_request &= (mode == SC_FLUSH_FILE) ?
2529                         ~FLUSH_FILE_BIT : ~FLUSH_DAT_BIT;
2530                 spin_unlock(&sci->sc_state_lock);
2531         }
2532         clear_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
2533 }
2534
2535 static int nilfs_segctor_flush_mode(struct nilfs_sc_info *sci)
2536 {
2537         if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2538             time_before(jiffies, sci->sc_lseg_stime + sci->sc_mjcp_freq)) {
2539                 if (!(sci->sc_flush_request & ~FLUSH_FILE_BIT))
2540                         return SC_FLUSH_FILE;
2541                 else if (!(sci->sc_flush_request & ~FLUSH_DAT_BIT))
2542                         return SC_FLUSH_DAT;
2543         }
2544         return SC_LSEG_SR;
2545 }
2546
2547 /**
2548  * nilfs_segctor_thread - main loop of the segment constructor thread.
2549  * @arg: pointer to a struct nilfs_sc_info.
2550  *
2551  * nilfs_segctor_thread() initializes a timer and serves as a daemon
2552  * to execute segment constructions.
2553  */
2554 static int nilfs_segctor_thread(void *arg)
2555 {
2556         struct nilfs_sc_info *sci = (struct nilfs_sc_info *)arg;
2557         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2558         int timeout = 0;
2559
2560         sci->sc_timer.data = (unsigned long)current;
2561         sci->sc_timer.function = nilfs_construction_timeout;
2562
2563         /* start sync. */
2564         sci->sc_task = current;
2565         wake_up(&sci->sc_wait_task); /* for nilfs_segctor_start_thread() */
2566         nilfs_msg(sci->sc_super, KERN_INFO,
2567                   "segctord starting. Construction interval = %lu seconds, CP frequency < %lu seconds",
2568                   sci->sc_interval / HZ, sci->sc_mjcp_freq / HZ);
2569
2570         spin_lock(&sci->sc_state_lock);
2571  loop:
2572         for (;;) {
2573                 int mode;
2574
2575                 if (sci->sc_state & NILFS_SEGCTOR_QUIT)
2576                         goto end_thread;
2577
2578                 if (timeout || sci->sc_seq_request != sci->sc_seq_done)
2579                         mode = SC_LSEG_SR;
2580                 else if (sci->sc_flush_request)
2581                         mode = nilfs_segctor_flush_mode(sci);
2582                 else
2583                         break;
2584
2585                 spin_unlock(&sci->sc_state_lock);
2586                 nilfs_segctor_thread_construct(sci, mode);
2587                 spin_lock(&sci->sc_state_lock);
2588                 timeout = 0;
2589         }
2590
2591
2592         if (freezing(current)) {
2593                 spin_unlock(&sci->sc_state_lock);
2594                 try_to_freeze();
2595                 spin_lock(&sci->sc_state_lock);
2596         } else {
2597                 DEFINE_WAIT(wait);
2598                 int should_sleep = 1;
2599
2600                 prepare_to_wait(&sci->sc_wait_daemon, &wait,
2601                                 TASK_INTERRUPTIBLE);
2602
2603                 if (sci->sc_seq_request != sci->sc_seq_done)
2604                         should_sleep = 0;
2605                 else if (sci->sc_flush_request)
2606                         should_sleep = 0;
2607                 else if (sci->sc_state & NILFS_SEGCTOR_COMMIT)
2608                         should_sleep = time_before(jiffies,
2609                                         sci->sc_timer.expires);
2610
2611                 if (should_sleep) {
2612                         spin_unlock(&sci->sc_state_lock);
2613                         schedule();
2614                         spin_lock(&sci->sc_state_lock);
2615                 }
2616                 finish_wait(&sci->sc_wait_daemon, &wait);
2617                 timeout = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2618                            time_after_eq(jiffies, sci->sc_timer.expires));
2619
2620                 if (nilfs_sb_dirty(nilfs) && nilfs_sb_need_update(nilfs))
2621                         set_nilfs_discontinued(nilfs);
2622         }
2623         goto loop;
2624
2625  end_thread:
2626         /* end sync. */
2627         sci->sc_task = NULL;
2628         wake_up(&sci->sc_wait_task); /* for nilfs_segctor_kill_thread() */
2629         spin_unlock(&sci->sc_state_lock);
2630         return 0;
2631 }
2632
2633 static int nilfs_segctor_start_thread(struct nilfs_sc_info *sci)
2634 {
2635         struct task_struct *t;
2636
2637         t = kthread_run(nilfs_segctor_thread, sci, "segctord");
2638         if (IS_ERR(t)) {
2639                 int err = PTR_ERR(t);
2640
2641                 nilfs_msg(sci->sc_super, KERN_ERR,
2642                           "error %d creating segctord thread", err);
2643                 return err;
2644         }
2645         wait_event(sci->sc_wait_task, sci->sc_task != NULL);
2646         return 0;
2647 }
2648
2649 static void nilfs_segctor_kill_thread(struct nilfs_sc_info *sci)
2650         __acquires(&sci->sc_state_lock)
2651         __releases(&sci->sc_state_lock)
2652 {
2653         sci->sc_state |= NILFS_SEGCTOR_QUIT;
2654
2655         while (sci->sc_task) {
2656                 wake_up(&sci->sc_wait_daemon);
2657                 spin_unlock(&sci->sc_state_lock);
2658                 wait_event(sci->sc_wait_task, sci->sc_task == NULL);
2659                 spin_lock(&sci->sc_state_lock);
2660         }
2661 }
2662
2663 /*
2664  * Setup & clean-up functions
2665  */
2666 static struct nilfs_sc_info *nilfs_segctor_new(struct super_block *sb,
2667                                                struct nilfs_root *root)
2668 {
2669         struct the_nilfs *nilfs = sb->s_fs_info;
2670         struct nilfs_sc_info *sci;
2671
2672         sci = kzalloc(sizeof(*sci), GFP_KERNEL);
2673         if (!sci)
2674                 return NULL;
2675
2676         sci->sc_super = sb;
2677
2678         nilfs_get_root(root);
2679         sci->sc_root = root;
2680
2681         init_waitqueue_head(&sci->sc_wait_request);
2682         init_waitqueue_head(&sci->sc_wait_daemon);
2683         init_waitqueue_head(&sci->sc_wait_task);
2684         spin_lock_init(&sci->sc_state_lock);
2685         INIT_LIST_HEAD(&sci->sc_dirty_files);
2686         INIT_LIST_HEAD(&sci->sc_segbufs);
2687         INIT_LIST_HEAD(&sci->sc_write_logs);
2688         INIT_LIST_HEAD(&sci->sc_gc_inodes);
2689         INIT_LIST_HEAD(&sci->sc_iput_queue);
2690         INIT_WORK(&sci->sc_iput_work, nilfs_iput_work_func);
2691         init_timer(&sci->sc_timer);
2692
2693         sci->sc_interval = HZ * NILFS_SC_DEFAULT_TIMEOUT;
2694         sci->sc_mjcp_freq = HZ * NILFS_SC_DEFAULT_SR_FREQ;
2695         sci->sc_watermark = NILFS_SC_DEFAULT_WATERMARK;
2696
2697         if (nilfs->ns_interval)
2698                 sci->sc_interval = HZ * nilfs->ns_interval;
2699         if (nilfs->ns_watermark)
2700                 sci->sc_watermark = nilfs->ns_watermark;
2701         return sci;
2702 }
2703
2704 static void nilfs_segctor_write_out(struct nilfs_sc_info *sci)
2705 {
2706         int ret, retrycount = NILFS_SC_CLEANUP_RETRY;
2707
2708         /*
2709          * The segctord thread was stopped and its timer was removed.
2710          * But some tasks remain.
2711          */
2712         do {
2713                 struct nilfs_transaction_info ti;
2714
2715                 nilfs_transaction_lock(sci->sc_super, &ti, 0);
2716                 ret = nilfs_segctor_construct(sci, SC_LSEG_SR);
2717                 nilfs_transaction_unlock(sci->sc_super);
2718
2719                 flush_work(&sci->sc_iput_work);
2720
2721         } while (ret && retrycount-- > 0);
2722 }
2723
2724 /**
2725  * nilfs_segctor_destroy - destroy the segment constructor.
2726  * @sci: nilfs_sc_info
2727  *
2728  * nilfs_segctor_destroy() kills the segctord thread and frees
2729  * the nilfs_sc_info struct.
2730  * Caller must hold the segment semaphore.
2731  */
2732 static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
2733 {
2734         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2735         int flag;
2736
2737         up_write(&nilfs->ns_segctor_sem);
2738
2739         spin_lock(&sci->sc_state_lock);
2740         nilfs_segctor_kill_thread(sci);
2741         flag = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) || sci->sc_flush_request
2742                 || sci->sc_seq_request != sci->sc_seq_done);
2743         spin_unlock(&sci->sc_state_lock);
2744
2745         if (flush_work(&sci->sc_iput_work))
2746                 flag = true;
2747
2748         if (flag || !nilfs_segctor_confirm(sci))
2749                 nilfs_segctor_write_out(sci);
2750
2751         if (!list_empty(&sci->sc_dirty_files)) {
2752                 nilfs_msg(sci->sc_super, KERN_WARNING,
2753                           "disposed unprocessed dirty file(s) when stopping log writer");
2754                 nilfs_dispose_list(nilfs, &sci->sc_dirty_files, 1);
2755         }
2756
2757         if (!list_empty(&sci->sc_iput_queue)) {
2758                 nilfs_msg(sci->sc_super, KERN_WARNING,
2759                           "disposed unprocessed inode(s) in iput queue when stopping log writer");
2760                 nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 1);
2761         }
2762
2763         WARN_ON(!list_empty(&sci->sc_segbufs));
2764         WARN_ON(!list_empty(&sci->sc_write_logs));
2765
2766         nilfs_put_root(sci->sc_root);
2767
2768         down_write(&nilfs->ns_segctor_sem);
2769
2770         del_timer_sync(&sci->sc_timer);
2771         kfree(sci);
2772 }
2773
2774 /**
2775  * nilfs_attach_log_writer - attach log writer
2776  * @sb: super block instance
2777  * @root: root object of the current filesystem tree
2778  *
2779  * This allocates a log writer object, initializes it, and starts the
2780  * log writer.
2781  *
2782  * Return Value: On success, 0 is returned. On error, one of the following
2783  * negative error code is returned.
2784  *
2785  * %-ENOMEM - Insufficient memory available.
2786  */
2787 int nilfs_attach_log_writer(struct super_block *sb, struct nilfs_root *root)
2788 {
2789         struct the_nilfs *nilfs = sb->s_fs_info;
2790         int err;
2791
2792         if (nilfs->ns_writer) {
2793                 /*
2794                  * This happens if the filesystem is made read-only by
2795                  * __nilfs_error or nilfs_remount and then remounted
2796                  * read/write.  In these cases, reuse the existing
2797                  * writer.
2798                  */
2799                 return 0;
2800         }
2801
2802         nilfs->ns_writer = nilfs_segctor_new(sb, root);
2803         if (!nilfs->ns_writer)
2804                 return -ENOMEM;
2805
2806         inode_attach_wb(nilfs->ns_bdev->bd_inode, NULL);
2807
2808         err = nilfs_segctor_start_thread(nilfs->ns_writer);
2809         if (unlikely(err))
2810                 nilfs_detach_log_writer(sb);
2811
2812         return err;
2813 }
2814
2815 /**
2816  * nilfs_detach_log_writer - destroy log writer
2817  * @sb: super block instance
2818  *
2819  * This kills log writer daemon, frees the log writer object, and
2820  * destroys list of dirty files.
2821  */
2822 void nilfs_detach_log_writer(struct super_block *sb)
2823 {
2824         struct the_nilfs *nilfs = sb->s_fs_info;
2825         LIST_HEAD(garbage_list);
2826
2827         down_write(&nilfs->ns_segctor_sem);
2828         if (nilfs->ns_writer) {
2829                 nilfs_segctor_destroy(nilfs->ns_writer);
2830                 nilfs->ns_writer = NULL;
2831         }
2832
2833         /* Force to free the list of dirty files */
2834         spin_lock(&nilfs->ns_inode_lock);
2835         if (!list_empty(&nilfs->ns_dirty_files)) {
2836                 list_splice_init(&nilfs->ns_dirty_files, &garbage_list);
2837                 nilfs_msg(sb, KERN_WARNING,
2838                           "disposed unprocessed dirty file(s) when detaching log writer");
2839         }
2840         spin_unlock(&nilfs->ns_inode_lock);
2841         up_write(&nilfs->ns_segctor_sem);
2842
2843         nilfs_dispose_list(nilfs, &garbage_list, 1);
2844 }