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  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  19  *
  20  * Written by Ryusuke Konishi <ryusuke@osrg.net>
  21  *
  22  */
  23
  24 #include <linux/pagemap.h>
  25 #include <linux/buffer_head.h>
  26 #include <linux/writeback.h>
  27 #include <linux/bio.h>
  28 #include <linux/completion.h>
  29 #include <linux/blkdev.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/freezer.h>
  32 #include <linux/kthread.h>
  33 #include <linux/crc32.h>
  34 #include <linux/pagevec.h>
  35 #include <linux/slab.h>
  36 #include "nilfs.h"
  37 #include "btnode.h"
  38 #include "page.h"
  39 #include "segment.h"
  40 #include "sufile.h"
  41 #include "cpfile.h"
  42 #include "ifile.h"
  43 #include "segbuf.h"
  44
  45
  46 /*
  47  * Segment constructor
  48  */
  49 #define SC_N_INODEVEC   16   /* Size of locally allocated inode vector */
  50
  51 #define SC_MAX_SEGDELTA 64   /* Upper limit of the number of segments
  52                                 appended in collection retry loop */
  53
  54 /* Construction mode */
  55 enum {
  56         SC_LSEG_SR = 1, /* Make a logical segment having a super root */
  57         SC_LSEG_DSYNC,  /* Flush data blocks of a given file and make
  58                            a logical segment without a super root */
  59         SC_FLUSH_FILE,  /* Flush data files, leads to segment writes without
  60                            creating a checkpoint */
  61         SC_FLUSH_DAT,   /* Flush DAT file. This also creates segments without
  62                            a checkpoint */
  63 };
  64
  65 /* Stage numbers of dirty block collection */
  66 enum {
  67         NILFS_ST_INIT = 0,
  68         NILFS_ST_GC,            /* Collecting dirty blocks for GC */
  69         NILFS_ST_FILE,
  70         NILFS_ST_IFILE,
  71         NILFS_ST_CPFILE,
  72         NILFS_ST_SUFILE,
  73         NILFS_ST_DAT,
  74         NILFS_ST_SR,            /* Super root */
  75         NILFS_ST_DSYNC,         /* Data sync blocks */
  76         NILFS_ST_DONE,
  77 };
  78
  79 /* State flags of collection */
  80 #define NILFS_CF_NODE           0x0001  /* Collecting node blocks */
  81 #define NILFS_CF_IFILE_STARTED  0x0002  /* IFILE stage has started */
  82 #define NILFS_CF_SUFREED        0x0004  /* segment usages has been freed */
  83 #define NILFS_CF_HISTORY_MASK   (NILFS_CF_IFILE_STARTED | NILFS_CF_SUFREED)
  84
  85 /* Operations depending on the construction mode and file type */
  86 struct nilfs_sc_operations {
  87         int (*collect_data)(struct nilfs_sc_info *, struct buffer_head *,
  88                             struct inode *);
  89         int (*collect_node)(struct nilfs_sc_info *, struct buffer_head *,
  90                             struct inode *);
  91         int (*collect_bmap)(struct nilfs_sc_info *, struct buffer_head *,
  92                             struct inode *);
  93         void (*write_data_binfo)(struct nilfs_sc_info *,
  94                                  struct nilfs_segsum_pointer *,
  95                                  union nilfs_binfo *);
  96         void (*write_node_binfo)(struct nilfs_sc_info *,
  97                                  struct nilfs_segsum_pointer *,
  98                                  union nilfs_binfo *);
  99 };
 100
 101 /*
 102  * Other definitions
 103  */
 104 static void nilfs_segctor_start_timer(struct nilfs_sc_info *);
 105 static void nilfs_segctor_do_flush(struct nilfs_sc_info *, int);
 106 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *);
 107 static void nilfs_dispose_list(struct nilfs_sb_info *, struct list_head *,
 108                                int);
 109
 110 #define nilfs_cnt32_gt(a, b)   \
 111         (typecheck(__u32, a) && typecheck(__u32, b) && \
 112          ((__s32)(b) - (__s32)(a) < 0))
 113 #define nilfs_cnt32_ge(a, b)   \
 114         (typecheck(__u32, a) && typecheck(__u32, b) && \
 115          ((__s32)(a) - (__s32)(b) >= 0))
 116 #define nilfs_cnt32_lt(a, b)  nilfs_cnt32_gt(b, a)
 117 #define nilfs_cnt32_le(a, b)  nilfs_cnt32_ge(b, a)
 118
 119 static int nilfs_prepare_segment_lock(struct nilfs_transaction_info *ti)
 120 {
 121         struct nilfs_transaction_info *cur_ti = current->journal_info;
 122         void *save = NULL;
 123
 124         if (cur_ti) {
 125                 if (cur_ti->ti_magic == NILFS_TI_MAGIC)
 126                         return ++cur_ti->ti_count;
 127                 else {
 128                         /*
 129                          * If journal_info field is occupied by other FS,
 130                          * it is saved and will be restored on
 131                          * nilfs_transaction_commit().
 132                          */
 133                         printk(KERN_WARNING
 134                                "NILFS warning: journal info from a different "
 135                                "FS\n");
 136                         save = current->journal_info;
 137                 }
 138         }
 139         if (!ti) {
 140                 ti = kmem_cache_alloc(nilfs_transaction_cachep, GFP_NOFS);
 141                 if (!ti)
 142                         return -ENOMEM;
 143                 ti->ti_flags = NILFS_TI_DYNAMIC_ALLOC;
 144         } else {
 145                 ti->ti_flags = 0;
 146         }
 147         ti->ti_count = 0;
 148         ti->ti_save = save;
 149         ti->ti_magic = NILFS_TI_MAGIC;
 150         current->journal_info = ti;
 151         return 0;
 152 }
 153
 154 /**
 155  * nilfs_transaction_begin - start indivisible file operations.
 156  * @sb: super block
 157  * @ti: nilfs_transaction_info
 158  * @vacancy_check: flags for vacancy rate checks
 159  *
 160  * nilfs_transaction_begin() acquires a reader/writer semaphore, called
 161  * the segment semaphore, to make a segment construction and write tasks
 162  * exclusive.  The function is used with nilfs_transaction_commit() in pairs.
 163  * The region enclosed by these two functions can be nested.  To avoid a
 164  * deadlock, the semaphore is only acquired or released in the outermost call.
 165  *
 166  * This function allocates a nilfs_transaction_info struct to keep context
 167  * information on it.  It is initialized and hooked onto the current task in
 168  * the outermost call.  If a pre-allocated struct is given to @ti, it is used
 169  * instead; otherwise a new struct is assigned from a slab.
 170  *
 171  * When @vacancy_check flag is set, this function will check the amount of
 172  * free space, and will wait for the GC to reclaim disk space if low capacity.
 173  *
 174  * Return Value: On success, 0 is returned. On error, one of the following
 175  * negative error code is returned.
 176  *
 177  * %-ENOMEM - Insufficient memory available.
 178  *
 179  * %-ENOSPC - No space left on device
 180  */
 181 int nilfs_transaction_begin(struct super_block *sb,
 182                             struct nilfs_transaction_info *ti,
 183                             int vacancy_check)
 184 {
 185         struct nilfs_sb_info *sbi;
 186         struct the_nilfs *nilfs;
 187         int ret = nilfs_prepare_segment_lock(ti);
 188
 189         if (unlikely(ret < 0))
 190                 return ret;
 191         if (ret > 0)
 192                 return 0;
 193
 194         sbi = NILFS_SB(sb);
 195         nilfs = sbi->s_nilfs;
 196         down_read(&nilfs->ns_segctor_sem);
 197         if (vacancy_check && nilfs_near_disk_full(nilfs)) {
 198                 up_read(&nilfs->ns_segctor_sem);
 199                 ret = -ENOSPC;
 200                 goto failed;
 201         }
 202         return 0;
 203
 204  failed:
 205         ti = current->journal_info;
 206         current->journal_info = ti->ti_save;
 207         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 208                 kmem_cache_free(nilfs_transaction_cachep, ti);
 209         return ret;
 210 }
 211
 212 /**
 213  * nilfs_transaction_commit - commit indivisible file operations.
 214  * @sb: super block
 215  *
 216  * nilfs_transaction_commit() releases the read semaphore which is
 217  * acquired by nilfs_transaction_begin(). This is only performed
 218  * in outermost call of this function.  If a commit flag is set,
 219  * nilfs_transaction_commit() sets a timer to start the segment
 220  * constructor.  If a sync flag is set, it starts construction
 221  * directly.
 222  */
 223 int nilfs_transaction_commit(struct super_block *sb)
 224 {
 225         struct nilfs_transaction_info *ti = current->journal_info;
 226         struct nilfs_sb_info *sbi;
 227         struct nilfs_sc_info *sci;
 228         int err = 0;
 229
 230         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 231         ti->ti_flags |= NILFS_TI_COMMIT;
 232         if (ti->ti_count > 0) {
 233                 ti->ti_count--;
 234                 return 0;
 235         }
 236         sbi = NILFS_SB(sb);
 237         sci = NILFS_SC(sbi);
 238         if (sci != NULL) {
 239                 if (ti->ti_flags & NILFS_TI_COMMIT)
 240                         nilfs_segctor_start_timer(sci);
 241                 if (atomic_read(&sbi->s_nilfs->ns_ndirtyblks) >
 242                     sci->sc_watermark)
 243                         nilfs_segctor_do_flush(sci, 0);
 244         }
 245         up_read(&sbi->s_nilfs->ns_segctor_sem);
 246         current->journal_info = ti->ti_save;
 247
 248         if (ti->ti_flags & NILFS_TI_SYNC)
 249                 err = nilfs_construct_segment(sb);
 250         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 251                 kmem_cache_free(nilfs_transaction_cachep, ti);
 252         return err;
 253 }
 254
 255 void nilfs_transaction_abort(struct super_block *sb)
 256 {
 257         struct nilfs_transaction_info *ti = current->journal_info;
 258
 259         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 260         if (ti->ti_count > 0) {
 261                 ti->ti_count--;
 262                 return;
 263         }
 264         up_read(&NILFS_SB(sb)->s_nilfs->ns_segctor_sem);
 265
 266         current->journal_info = ti->ti_save;
 267         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 268                 kmem_cache_free(nilfs_transaction_cachep, ti);
 269 }
 270
 271 void nilfs_relax_pressure_in_lock(struct super_block *sb)
 272 {
 273         struct nilfs_sb_info *sbi = NILFS_SB(sb);
 274         struct nilfs_sc_info *sci = NILFS_SC(sbi);
 275         struct the_nilfs *nilfs = sbi->s_nilfs;
 276
 277         if (!sci || !sci->sc_flush_request)
 278                 return;
 279
 280         set_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
 281         up_read(&nilfs->ns_segctor_sem);
 282
 283         down_write(&nilfs->ns_segctor_sem);
 284         if (sci->sc_flush_request &&
 285             test_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags)) {
 286                 struct nilfs_transaction_info *ti = current->journal_info;
 287
 288                 ti->ti_flags |= NILFS_TI_WRITER;
 289                 nilfs_segctor_do_immediate_flush(sci);
 290                 ti->ti_flags &= ~NILFS_TI_WRITER;
 291         }
 292         downgrade_write(&nilfs->ns_segctor_sem);
 293 }
 294
 295 static void nilfs_transaction_lock(struct nilfs_sb_info *sbi,
 296                                    struct nilfs_transaction_info *ti,
 297                                    int gcflag)
 298 {
 299         struct nilfs_transaction_info *cur_ti = current->journal_info;
 300
 301         WARN_ON(cur_ti);
 302         ti->ti_flags = NILFS_TI_WRITER;
 303         ti->ti_count = 0;
 304         ti->ti_save = cur_ti;
 305         ti->ti_magic = NILFS_TI_MAGIC;
 306         INIT_LIST_HEAD(&ti->ti_garbage);
 307         current->journal_info = ti;
 308
 309         for (;;) {
 310                 down_write(&sbi->s_nilfs->ns_segctor_sem);
 311                 if (!test_bit(NILFS_SC_PRIOR_FLUSH, &NILFS_SC(sbi)->sc_flags))
 312                         break;
 313
 314                 nilfs_segctor_do_immediate_flush(NILFS_SC(sbi));
 315
 316                 up_write(&sbi->s_nilfs->ns_segctor_sem);
 317                 yield();
 318         }
 319         if (gcflag)
 320                 ti->ti_flags |= NILFS_TI_GC;
 321 }
 322
 323 static void nilfs_transaction_unlock(struct nilfs_sb_info *sbi)
 324 {
 325         struct nilfs_transaction_info *ti = current->journal_info;
 326
 327         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 328         BUG_ON(ti->ti_count > 0);
 329
 330         up_write(&sbi->s_nilfs->ns_segctor_sem);
 331         current->journal_info = ti->ti_save;
 332         if (!list_empty(&ti->ti_garbage))
 333                 nilfs_dispose_list(sbi, &ti->ti_garbage, 0);
 334 }
 335
 336 static void *nilfs_segctor_map_segsum_entry(struct nilfs_sc_info *sci,
 337                                             struct nilfs_segsum_pointer *ssp,
 338                                             unsigned bytes)
 339 {
 340         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 341         unsigned blocksize = sci->sc_super->s_blocksize;
 342         void *p;
 343
 344         if (unlikely(ssp->offset + bytes > blocksize)) {
 345                 ssp->offset = 0;
 346                 BUG_ON(NILFS_SEGBUF_BH_IS_LAST(ssp->bh,
 347                                                &segbuf->sb_segsum_buffers));
 348                 ssp->bh = NILFS_SEGBUF_NEXT_BH(ssp->bh);
 349         }
 350         p = ssp->bh->b_data + ssp->offset;
 351         ssp->offset += bytes;
 352         return p;
 353 }
 354
 355 /**
 356  * nilfs_segctor_reset_segment_buffer - reset the current segment buffer
 357  * @sci: nilfs_sc_info
 358  */
 359 static int nilfs_segctor_reset_segment_buffer(struct nilfs_sc_info *sci)
 360 {
 361         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 362         struct buffer_head *sumbh;
 363         unsigned sumbytes;
 364         unsigned flags = 0;
 365         int err;
 366
 367         if (nilfs_doing_gc())
 368                 flags = NILFS_SS_GC;
 369         err = nilfs_segbuf_reset(segbuf, flags, sci->sc_seg_ctime, sci->sc_cno);
 370         if (unlikely(err))
 371                 return err;
 372
 373         sumbh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
 374         sumbytes = segbuf->sb_sum.sumbytes;
 375         sci->sc_finfo_ptr.bh = sumbh;  sci->sc_finfo_ptr.offset = sumbytes;
 376         sci->sc_binfo_ptr.bh = sumbh;  sci->sc_binfo_ptr.offset = sumbytes;
 377         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 378         return 0;
 379 }
 380
 381 static int nilfs_segctor_feed_segment(struct nilfs_sc_info *sci)
 382 {
 383         sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
 384         if (NILFS_SEGBUF_IS_LAST(sci->sc_curseg, &sci->sc_segbufs))
 385                 return -E2BIG; /* The current segment is filled up
 386                                   (internal code) */
 387         sci->sc_curseg = NILFS_NEXT_SEGBUF(sci->sc_curseg);
 388         return nilfs_segctor_reset_segment_buffer(sci);
 389 }
 390
 391 static int nilfs_segctor_add_super_root(struct nilfs_sc_info *sci)
 392 {
 393         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 394         int err;
 395
 396         if (segbuf->sb_sum.nblocks >= segbuf->sb_rest_blocks) {
 397                 err = nilfs_segctor_feed_segment(sci);
 398                 if (err)
 399                         return err;
 400                 segbuf = sci->sc_curseg;
 401         }
 402         err = nilfs_segbuf_extend_payload(segbuf, &segbuf->sb_super_root);
 403         if (likely(!err))
 404                 segbuf->sb_sum.flags |= NILFS_SS_SR;
 405         return err;
 406 }
 407
 408 /*
 409  * Functions for making segment summary and payloads
 410  */
 411 static int nilfs_segctor_segsum_block_required(
 412         struct nilfs_sc_info *sci, const struct nilfs_segsum_pointer *ssp,
 413         unsigned binfo_size)
 414 {
 415         unsigned blocksize = sci->sc_super->s_blocksize;
 416         /* Size of finfo and binfo is enough small against blocksize */
 417
 418         return ssp->offset + binfo_size +
 419                 (!sci->sc_blk_cnt ? sizeof(struct nilfs_finfo) : 0) >
 420                 blocksize;
 421 }
 422
 423 static void nilfs_segctor_begin_finfo(struct nilfs_sc_info *sci,
 424                                       struct inode *inode)
 425 {
 426         sci->sc_curseg->sb_sum.nfinfo++;
 427         sci->sc_binfo_ptr = sci->sc_finfo_ptr;
 428         nilfs_segctor_map_segsum_entry(
 429                 sci, &sci->sc_binfo_ptr, sizeof(struct nilfs_finfo));
 430
 431         if (inode->i_sb && !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 432                 set_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
 433         /* skip finfo */
 434 }
 435
 436 static void nilfs_segctor_end_finfo(struct nilfs_sc_info *sci,
 437                                     struct inode *inode)
 438 {
 439         struct nilfs_finfo *finfo;
 440         struct nilfs_inode_info *ii;
 441         struct nilfs_segment_buffer *segbuf;
 442         __u64 cno;
 443
 444         if (sci->sc_blk_cnt == 0)
 445                 return;
 446
 447         ii = NILFS_I(inode);
 448
 449         if (test_bit(NILFS_I_GCINODE, &ii->i_state))
 450                 cno = ii->i_cno;
 451         else if (NILFS_ROOT_METADATA_FILE(inode->i_ino))
 452                 cno = 0;
 453         else
 454                 cno = sci->sc_cno;
 455
 456         finfo = nilfs_segctor_map_segsum_entry(sci, &sci->sc_finfo_ptr,
 457                                                  sizeof(*finfo));
 458         finfo->fi_ino = cpu_to_le64(inode->i_ino);
 459         finfo->fi_nblocks = cpu_to_le32(sci->sc_blk_cnt);
 460         finfo->fi_ndatablk = cpu_to_le32(sci->sc_datablk_cnt);
 461         finfo->fi_cno = cpu_to_le64(cno);
 462
 463         segbuf = sci->sc_curseg;
 464         segbuf->sb_sum.sumbytes = sci->sc_binfo_ptr.offset +
 465                 sci->sc_super->s_blocksize * (segbuf->sb_sum.nsumblk - 1);
 466         sci->sc_finfo_ptr = sci->sc_binfo_ptr;
 467         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 468 }
 469
 470 static int nilfs_segctor_add_file_block(struct nilfs_sc_info *sci,
 471                                         struct buffer_head *bh,
 472                                         struct inode *inode,
 473                                         unsigned binfo_size)
 474 {
 475         struct nilfs_segment_buffer *segbuf;
 476         int required, err = 0;
 477
 478  retry:
 479         segbuf = sci->sc_curseg;
 480         required = nilfs_segctor_segsum_block_required(
 481                 sci, &sci->sc_binfo_ptr, binfo_size);
 482         if (segbuf->sb_sum.nblocks + required + 1 > segbuf->sb_rest_blocks) {
 483                 nilfs_segctor_end_finfo(sci, inode);
 484                 err = nilfs_segctor_feed_segment(sci);
 485                 if (err)
 486                         return err;
 487                 goto retry;
 488         }
 489         if (unlikely(required)) {
 490                 err = nilfs_segbuf_extend_segsum(segbuf);
 491                 if (unlikely(err))
 492                         goto failed;
 493         }
 494         if (sci->sc_blk_cnt == 0)
 495                 nilfs_segctor_begin_finfo(sci, inode);
 496
 497         nilfs_segctor_map_segsum_entry(sci, &sci->sc_binfo_ptr, binfo_size);
 498         /* Substitution to vblocknr is delayed until update_blocknr() */
 499         nilfs_segbuf_add_file_buffer(segbuf, bh);
 500         sci->sc_blk_cnt++;
 501  failed:
 502         return err;
 503 }
 504
 505 static int nilfs_handle_bmap_error(int err, const char *fname,
 506                                    struct inode *inode, struct super_block *sb)
 507 {
 508         if (err == -EINVAL) {
 509                 nilfs_error(sb, fname, "broken bmap (inode=%lu)\n",
 510                             inode->i_ino);
 511                 err = -EIO;
 512         }
 513         return err;
 514 }
 515
 516 /*
 517  * Callback functions that enumerate, mark, and collect dirty blocks
 518  */
 519 static int nilfs_collect_file_data(struct nilfs_sc_info *sci,
 520                                    struct buffer_head *bh, struct inode *inode)
 521 {
 522         int err;
 523
 524         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 525         if (unlikely(err < 0))
 526                 return nilfs_handle_bmap_error(err, __func__, inode,
 527                                                sci->sc_super);
 528
 529         err = nilfs_segctor_add_file_block(sci, bh, inode,
 530                                            sizeof(struct nilfs_binfo_v));
 531         if (!err)
 532                 sci->sc_datablk_cnt++;
 533         return err;
 534 }
 535
 536 static int nilfs_collect_file_node(struct nilfs_sc_info *sci,
 537                                    struct buffer_head *bh,
 538                                    struct inode *inode)
 539 {
 540         int err;
 541
 542         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 543         if (unlikely(err < 0))
 544                 return nilfs_handle_bmap_error(err, __func__, inode,
 545                                                sci->sc_super);
 546         return 0;
 547 }
 548
 549 static int nilfs_collect_file_bmap(struct nilfs_sc_info *sci,
 550                                    struct buffer_head *bh,
 551                                    struct inode *inode)
 552 {
 553         WARN_ON(!buffer_dirty(bh));
 554         return nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 555 }
 556
 557 static void nilfs_write_file_data_binfo(struct nilfs_sc_info *sci,
 558                                         struct nilfs_segsum_pointer *ssp,
 559                                         union nilfs_binfo *binfo)
 560 {
 561         struct nilfs_binfo_v *binfo_v = nilfs_segctor_map_segsum_entry(
 562                 sci, ssp, sizeof(*binfo_v));
 563         *binfo_v = binfo->bi_v;
 564 }
 565
 566 static void nilfs_write_file_node_binfo(struct nilfs_sc_info *sci,
 567                                         struct nilfs_segsum_pointer *ssp,
 568                                         union nilfs_binfo *binfo)
 569 {
 570         __le64 *vblocknr = nilfs_segctor_map_segsum_entry(
 571                 sci, ssp, sizeof(*vblocknr));
 572         *vblocknr = binfo->bi_v.bi_vblocknr;
 573 }
 574
 575 static struct nilfs_sc_operations nilfs_sc_file_ops = {
 576         .collect_data = nilfs_collect_file_data,
 577         .collect_node = nilfs_collect_file_node,
 578         .collect_bmap = nilfs_collect_file_bmap,
 579         .write_data_binfo = nilfs_write_file_data_binfo,
 580         .write_node_binfo = nilfs_write_file_node_binfo,
 581 };
 582
 583 static int nilfs_collect_dat_data(struct nilfs_sc_info *sci,
 584                                   struct buffer_head *bh, struct inode *inode)
 585 {
 586         int err;
 587
 588         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 589         if (unlikely(err < 0))
 590                 return nilfs_handle_bmap_error(err, __func__, inode,
 591                                                sci->sc_super);
 592
 593         err = nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 594         if (!err)
 595                 sci->sc_datablk_cnt++;
 596         return err;
 597 }
 598
 599 static int nilfs_collect_dat_bmap(struct nilfs_sc_info *sci,
 600                                   struct buffer_head *bh, struct inode *inode)
 601 {
 602         WARN_ON(!buffer_dirty(bh));
 603         return nilfs_segctor_add_file_block(sci, bh, inode,
 604                                             sizeof(struct nilfs_binfo_dat));
 605 }
 606
 607 static void nilfs_write_dat_data_binfo(struct nilfs_sc_info *sci,
 608                                        struct nilfs_segsum_pointer *ssp,
 609                                        union nilfs_binfo *binfo)
 610 {
 611         __le64 *blkoff = nilfs_segctor_map_segsum_entry(sci, ssp,
 612                                                           sizeof(*blkoff));
 613         *blkoff = binfo->bi_dat.bi_blkoff;
 614 }
 615
 616 static void nilfs_write_dat_node_binfo(struct nilfs_sc_info *sci,
 617                                        struct nilfs_segsum_pointer *ssp,
 618                                        union nilfs_binfo *binfo)
 619 {
 620         struct nilfs_binfo_dat *binfo_dat =
 621                 nilfs_segctor_map_segsum_entry(sci, ssp, sizeof(*binfo_dat));
 622         *binfo_dat = binfo->bi_dat;
 623 }
 624
 625 static struct nilfs_sc_operations nilfs_sc_dat_ops = {
 626         .collect_data = nilfs_collect_dat_data,
 627         .collect_node = nilfs_collect_file_node,
 628         .collect_bmap = nilfs_collect_dat_bmap,
 629         .write_data_binfo = nilfs_write_dat_data_binfo,
 630         .write_node_binfo = nilfs_write_dat_node_binfo,
 631 };
 632
 633 static struct nilfs_sc_operations nilfs_sc_dsync_ops = {
 634         .collect_data = nilfs_collect_file_data,
 635         .collect_node = NULL,
 636         .collect_bmap = NULL,
 637         .write_data_binfo = nilfs_write_file_data_binfo,
 638         .write_node_binfo = NULL,
 639 };
 640
 641 static size_t nilfs_lookup_dirty_data_buffers(struct inode *inode,
 642                                               struct list_head *listp,
 643                                               size_t nlimit,
 644                                               loff_t start, loff_t end)
 645 {
 646         struct address_space *mapping = inode->i_mapping;
 647         struct pagevec pvec;
 648         pgoff_t index = 0, last = ULONG_MAX;
 649         size_t ndirties = 0;
 650         int i;
 651
 652         if (unlikely(start != 0 || end != LLONG_MAX)) {
 653                 /*
 654                  * A valid range is given for sync-ing data pages. The
 655                  * range is rounded to per-page; extra dirty buffers
 656                  * may be included if blocksize < pagesize.
 657                  */
 658                 index = start >> PAGE_SHIFT;
 659                 last = end >> PAGE_SHIFT;
 660         }
 661         pagevec_init(&pvec, 0);
 662  repeat:
 663         if (unlikely(index > last) ||
 664             !pagevec_lookup_tag(&pvec, mapping, &index, PAGECACHE_TAG_DIRTY,
 665                                 min_t(pgoff_t, last - index,
 666                                       PAGEVEC_SIZE - 1) + 1))
 667                 return ndirties;
 668
 669         for (i = 0; i < pagevec_count(&pvec); i++) {
 670                 struct buffer_head *bh, *head;
 671                 struct page *page = pvec.pages[i];
 672
 673                 if (unlikely(page->index > last))
 674                         break;
 675
 676                 if (mapping->host) {
 677                         lock_page(page);
 678                         if (!page_has_buffers(page))
 679                                 create_empty_buffers(page,
 680                                                      1 << inode->i_blkbits, 0);
 681                         unlock_page(page);
 682                 }
 683
 684                 bh = head = page_buffers(page);
 685                 do {
 686                         if (!buffer_dirty(bh))
 687                                 continue;
 688                         get_bh(bh);
 689                         list_add_tail(&bh->b_assoc_buffers, listp);
 690                         ndirties++;
 691                         if (unlikely(ndirties >= nlimit)) {
 692                                 pagevec_release(&pvec);
 693                                 cond_resched();
 694                                 return ndirties;
 695                         }
 696                 } while (bh = bh->b_this_page, bh != head);
 697         }
 698         pagevec_release(&pvec);
 699         cond_resched();
 700         goto repeat;
 701 }
 702
 703 static void nilfs_lookup_dirty_node_buffers(struct inode *inode,
 704                                             struct list_head *listp)
 705 {
 706         struct nilfs_inode_info *ii = NILFS_I(inode);
 707         struct address_space *mapping = &ii->i_btnode_cache;
 708         struct pagevec pvec;
 709         struct buffer_head *bh, *head;
 710         unsigned int i;
 711         pgoff_t index = 0;
 712
 713         pagevec_init(&pvec, 0);
 714
 715         while (pagevec_lookup_tag(&pvec, mapping, &index, PAGECACHE_TAG_DIRTY,
 716                                   PAGEVEC_SIZE)) {
 717                 for (i = 0; i < pagevec_count(&pvec); i++) {
 718                         bh = head = page_buffers(pvec.pages[i]);
 719                         do {
 720                                 if (buffer_dirty(bh)) {
 721                                         get_bh(bh);
 722                                         list_add_tail(&bh->b_assoc_buffers,
 723                                                       listp);
 724                                 }
 725                                 bh = bh->b_this_page;
 726                         } while (bh != head);
 727                 }
 728                 pagevec_release(&pvec);
 729                 cond_resched();
 730         }
 731 }
 732
 733 static void nilfs_dispose_list(struct nilfs_sb_info *sbi,
 734                                struct list_head *head, int force)
 735 {
 736         struct nilfs_inode_info *ii, *n;
 737         struct nilfs_inode_info *ivec[SC_N_INODEVEC], **pii;
 738         unsigned nv = 0;
 739
 740         while (!list_empty(head)) {
 741                 spin_lock(&sbi->s_inode_lock);
 742                 list_for_each_entry_safe(ii, n, head, i_dirty) {
 743                         list_del_init(&ii->i_dirty);
 744                         if (force) {
 745                                 if (unlikely(ii->i_bh)) {
 746                                         brelse(ii->i_bh);
 747                                         ii->i_bh = NULL;
 748                                 }
 749                         } else if (test_bit(NILFS_I_DIRTY, &ii->i_state)) {
 750                                 set_bit(NILFS_I_QUEUED, &ii->i_state);
 751                                 list_add_tail(&ii->i_dirty,
 752                                               &sbi->s_dirty_files);
 753                                 continue;
 754                         }
 755                         ivec[nv++] = ii;
 756                         if (nv == SC_N_INODEVEC)
 757                                 break;
 758                 }
 759                 spin_unlock(&sbi->s_inode_lock);
 760
 761                 for (pii = ivec; nv > 0; pii++, nv--)
 762                         iput(&(*pii)->vfs_inode);
 763         }
 764 }
 765
 766 static int nilfs_test_metadata_dirty(struct the_nilfs *nilfs,
 767                                      struct nilfs_root *root)
 768 {
 769         int ret = 0;
 770
 771         if (nilfs_mdt_fetch_dirty(root->ifile))
 772                 ret++;
 773         if (nilfs_mdt_fetch_dirty(nilfs->ns_cpfile))
 774                 ret++;
 775         if (nilfs_mdt_fetch_dirty(nilfs->ns_sufile))
 776                 ret++;
 777         if (ret || nilfs_doing_gc())
 778                 if (nilfs_mdt_fetch_dirty(nilfs_dat_inode(nilfs)))
 779                         ret++;
 780         return ret;
 781 }
 782
 783 static int nilfs_segctor_clean(struct nilfs_sc_info *sci)
 784 {
 785         return list_empty(&sci->sc_dirty_files) &&
 786                 !test_bit(NILFS_SC_DIRTY, &sci->sc_flags) &&
 787                 sci->sc_nfreesegs == 0 &&
 788                 (!nilfs_doing_gc() || list_empty(&sci->sc_gc_inodes));
 789 }
 790
 791 static int nilfs_segctor_confirm(struct nilfs_sc_info *sci)
 792 {
 793         struct nilfs_sb_info *sbi = sci->sc_sbi;
 794         int ret = 0;
 795
 796         if (nilfs_test_metadata_dirty(sbi->s_nilfs, sci->sc_root))
 797                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
 798
 799         spin_lock(&sbi->s_inode_lock);
 800         if (list_empty(&sbi->s_dirty_files) && nilfs_segctor_clean(sci))
 801                 ret++;
 802
 803         spin_unlock(&sbi->s_inode_lock);
 804         return ret;
 805 }
 806
 807 static void nilfs_segctor_clear_metadata_dirty(struct nilfs_sc_info *sci)
 808 {
 809         struct nilfs_sb_info *sbi = sci->sc_sbi;
 810         struct the_nilfs *nilfs = sbi->s_nilfs;
 811
 812         nilfs_mdt_clear_dirty(sci->sc_root->ifile);
 813         nilfs_mdt_clear_dirty(nilfs->ns_cpfile);
 814         nilfs_mdt_clear_dirty(nilfs->ns_sufile);
 815         nilfs_mdt_clear_dirty(nilfs_dat_inode(nilfs));
 816 }
 817
 818 static int nilfs_segctor_create_checkpoint(struct nilfs_sc_info *sci)
 819 {
 820         struct the_nilfs *nilfs = sci->sc_sbi->s_nilfs;
 821         struct buffer_head *bh_cp;
 822         struct nilfs_checkpoint *raw_cp;
 823         int err;
 824
 825         /* XXX: this interface will be changed */
 826         err = nilfs_cpfile_get_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, 1,
 827                                           &raw_cp, &bh_cp);
 828         if (likely(!err)) {
 829                 /* The following code is duplicated with cpfile.  But, it is
 830                    needed to collect the checkpoint even if it was not newly
 831                    created */
 832                 nilfs_mdt_mark_buffer_dirty(bh_cp);
 833                 nilfs_mdt_mark_dirty(nilfs->ns_cpfile);
 834                 nilfs_cpfile_put_checkpoint(
 835                         nilfs->ns_cpfile, nilfs->ns_cno, bh_cp);
 836         } else
 837                 WARN_ON(err == -EINVAL || err == -ENOENT);
 838
 839         return err;
 840 }
 841
 842 static int nilfs_segctor_fill_in_checkpoint(struct nilfs_sc_info *sci)
 843 {
 844         struct nilfs_sb_info *sbi = sci->sc_sbi;
 845         struct the_nilfs *nilfs = sbi->s_nilfs;
 846         struct buffer_head *bh_cp;
 847         struct nilfs_checkpoint *raw_cp;
 848         int err;
 849
 850         err = nilfs_cpfile_get_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, 0,
 851                                           &raw_cp, &bh_cp);
 852         if (unlikely(err)) {
 853                 WARN_ON(err == -EINVAL || err == -ENOENT);
 854                 goto failed_ibh;
 855         }
 856         raw_cp->cp_snapshot_list.ssl_next = 0;
 857         raw_cp->cp_snapshot_list.ssl_prev = 0;
 858         raw_cp->cp_inodes_count =
 859                 cpu_to_le64(atomic_read(&sci->sc_root->inodes_count));
 860         raw_cp->cp_blocks_count =
 861                 cpu_to_le64(atomic_read(&sci->sc_root->blocks_count));
 862         raw_cp->cp_nblk_inc =
 863                 cpu_to_le64(sci->sc_nblk_inc + sci->sc_nblk_this_inc);
 864         raw_cp->cp_create = cpu_to_le64(sci->sc_seg_ctime);
 865         raw_cp->cp_cno = cpu_to_le64(nilfs->ns_cno);
 866
 867         if (test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 868                 nilfs_checkpoint_clear_minor(raw_cp);
 869         else
 870                 nilfs_checkpoint_set_minor(raw_cp);
 871
 872         nilfs_write_inode_common(sci->sc_root->ifile,
 873                                  &raw_cp->cp_ifile_inode, 1);
 874         nilfs_cpfile_put_checkpoint(nilfs->ns_cpfile, nilfs->ns_cno, bh_cp);
 875         return 0;
 876
 877  failed_ibh:
 878         return err;
 879 }
 880
 881 static void nilfs_fill_in_file_bmap(struct inode *ifile,
 882                                     struct nilfs_inode_info *ii)
 883
 884 {
 885         struct buffer_head *ibh;
 886         struct nilfs_inode *raw_inode;
 887
 888         if (test_bit(NILFS_I_BMAP, &ii->i_state)) {
 889                 ibh = ii->i_bh;
 890                 BUG_ON(!ibh);
 891                 raw_inode = nilfs_ifile_map_inode(ifile, ii->vfs_inode.i_ino,
 892                                                   ibh);
 893                 nilfs_bmap_write(ii->i_bmap, raw_inode);
 894                 nilfs_ifile_unmap_inode(ifile, ii->vfs_inode.i_ino, ibh);
 895         }
 896 }
 897
 898 static void nilfs_segctor_fill_in_file_bmap(struct nilfs_sc_info *sci)
 899 {
 900         struct nilfs_inode_info *ii;
 901
 902         list_for_each_entry(ii, &sci->sc_dirty_files, i_dirty) {
 903                 nilfs_fill_in_file_bmap(sci->sc_root->ifile, ii);
 904                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
 905         }
 906 }
 907
 908 static void nilfs_segctor_fill_in_super_root(struct nilfs_sc_info *sci,
 909                                              struct the_nilfs *nilfs)
 910 {
 911         struct buffer_head *bh_sr;
 912         struct nilfs_super_root *raw_sr;
 913         unsigned isz = nilfs->ns_inode_size;
 914
 915         bh_sr = NILFS_LAST_SEGBUF(&sci->sc_segbufs)->sb_super_root;
 916         raw_sr = (struct nilfs_super_root *)bh_sr->b_data;
 917
 918         raw_sr->sr_bytes = cpu_to_le16(NILFS_SR_BYTES);
 919         raw_sr->sr_nongc_ctime
 920                 = cpu_to_le64(nilfs_doing_gc() ?
 921                               nilfs->ns_nongc_ctime : sci->sc_seg_ctime);
 922         raw_sr->sr_flags = 0;
 923
 924         nilfs_write_inode_common(nilfs_dat_inode(nilfs), (void *)raw_sr +
 925                                  NILFS_SR_DAT_OFFSET(isz), 1);
 926         nilfs_write_inode_common(nilfs->ns_cpfile, (void *)raw_sr +
 927                                  NILFS_SR_CPFILE_OFFSET(isz), 1);
 928         nilfs_write_inode_common(nilfs->ns_sufile, (void *)raw_sr +
 929                                  NILFS_SR_SUFILE_OFFSET(isz), 1);
 930 }
 931
 932 static void nilfs_redirty_inodes(struct list_head *head)
 933 {
 934         struct nilfs_inode_info *ii;
 935
 936         list_for_each_entry(ii, head, i_dirty) {
 937                 if (test_bit(NILFS_I_COLLECTED, &ii->i_state))
 938                         clear_bit(NILFS_I_COLLECTED, &ii->i_state);
 939         }
 940 }
 941
 942 static void nilfs_drop_collected_inodes(struct list_head *head)
 943 {
 944         struct nilfs_inode_info *ii;
 945
 946         list_for_each_entry(ii, head, i_dirty) {
 947                 if (!test_and_clear_bit(NILFS_I_COLLECTED, &ii->i_state))
 948                         continue;
 949
 950                 clear_bit(NILFS_I_INODE_DIRTY, &ii->i_state);
 951                 set_bit(NILFS_I_UPDATED, &ii->i_state);
 952         }
 953 }
 954
 955 static int nilfs_segctor_apply_buffers(struct nilfs_sc_info *sci,
 956                                        struct inode *inode,
 957                                        struct list_head *listp,
 958                                        int (*collect)(struct nilfs_sc_info *,
 959                                                       struct buffer_head *,
 960                                                       struct inode *))
 961 {
 962         struct buffer_head *bh, *n;
 963         int err = 0;
 964
 965         if (collect) {
 966                 list_for_each_entry_safe(bh, n, listp, b_assoc_buffers) {
 967                         list_del_init(&bh->b_assoc_buffers);
 968                         err = collect(sci, bh, inode);
 969                         brelse(bh);
 970                         if (unlikely(err))
 971                                 goto dispose_buffers;
 972                 }
 973                 return 0;
 974         }
 975
 976  dispose_buffers:
 977         while (!list_empty(listp)) {
 978                 bh = list_entry(listp->next, struct buffer_head,
 979                                 b_assoc_buffers);
 980                 list_del_init(&bh->b_assoc_buffers);
 981                 brelse(bh);
 982         }
 983         return err;
 984 }
 985
 986 static size_t nilfs_segctor_buffer_rest(struct nilfs_sc_info *sci)
 987 {
 988         /* Remaining number of blocks within segment buffer */
 989         return sci->sc_segbuf_nblocks -
 990                 (sci->sc_nblk_this_inc + sci->sc_curseg->sb_sum.nblocks);
 991 }
 992
 993 static int nilfs_segctor_scan_file(struct nilfs_sc_info *sci,
 994                                    struct inode *inode,
 995                                    struct nilfs_sc_operations *sc_ops)
 996 {
 997         LIST_HEAD(data_buffers);
 998         LIST_HEAD(node_buffers);
 999         int err;
1000
1001         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1002                 size_t n, rest = nilfs_segctor_buffer_rest(sci);
1003
1004                 n = nilfs_lookup_dirty_data_buffers(
1005                         inode, &data_buffers, rest + 1, 0, LLONG_MAX);
1006                 if (n > rest) {
1007                         err = nilfs_segctor_apply_buffers(
1008                                 sci, inode, &data_buffers,
1009                                 sc_ops->collect_data);
1010                         BUG_ON(!err); /* always receive -E2BIG or true error */
1011                         goto break_or_fail;
1012                 }
1013         }
1014         nilfs_lookup_dirty_node_buffers(inode, &node_buffers);
1015
1016         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1017                 err = nilfs_segctor_apply_buffers(
1018                         sci, inode, &data_buffers, sc_ops->collect_data);
1019                 if (unlikely(err)) {
1020                         /* dispose node list */
1021                         nilfs_segctor_apply_buffers(
1022                                 sci, inode, &node_buffers, NULL);
1023                         goto break_or_fail;
1024                 }
1025                 sci->sc_stage.flags |= NILFS_CF_NODE;
1026         }
1027         /* Collect node */
1028         err = nilfs_segctor_apply_buffers(
1029                 sci, inode, &node_buffers, sc_ops->collect_node);
1030         if (unlikely(err))
1031                 goto break_or_fail;
1032
1033         nilfs_bmap_lookup_dirty_buffers(NILFS_I(inode)->i_bmap, &node_buffers);
1034         err = nilfs_segctor_apply_buffers(
1035                 sci, inode, &node_buffers, sc_ops->collect_bmap);
1036         if (unlikely(err))
1037                 goto break_or_fail;
1038
1039         nilfs_segctor_end_finfo(sci, inode);
1040         sci->sc_stage.flags &= ~NILFS_CF_NODE;
1041
1042  break_or_fail:
1043         return err;
1044 }
1045
1046 static int nilfs_segctor_scan_file_dsync(struct nilfs_sc_info *sci,
1047                                          struct inode *inode)
1048 {
1049         LIST_HEAD(data_buffers);
1050         size_t n, rest = nilfs_segctor_buffer_rest(sci);
1051         int err;
1052
1053         n = nilfs_lookup_dirty_data_buffers(inode, &data_buffers, rest + 1,
1054                                             sci->sc_dsync_start,
1055                                             sci->sc_dsync_end);
1056
1057         err = nilfs_segctor_apply_buffers(sci, inode, &data_buffers,
1058                                           nilfs_collect_file_data);
1059         if (!err) {
1060                 nilfs_segctor_end_finfo(sci, inode);
1061                 BUG_ON(n > rest);
1062                 /* always receive -E2BIG or true error if n > rest */
1063         }
1064         return err;
1065 }
1066
1067 static int nilfs_segctor_collect_blocks(struct nilfs_sc_info *sci, int mode)
1068 {
1069         struct nilfs_sb_info *sbi = sci->sc_sbi;
1070         struct the_nilfs *nilfs = sbi->s_nilfs;
1071         struct list_head *head;
1072         struct nilfs_inode_info *ii;
1073         size_t ndone;
1074         int err = 0;
1075
1076         switch (sci->sc_stage.scnt) {
1077         case NILFS_ST_INIT:
1078                 /* Pre-processes */
1079                 sci->sc_stage.flags = 0;
1080
1081                 if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags)) {
1082                         sci->sc_nblk_inc = 0;
1083                         sci->sc_curseg->sb_sum.flags = NILFS_SS_LOGBGN;
1084                         if (mode == SC_LSEG_DSYNC) {
1085                                 sci->sc_stage.scnt = NILFS_ST_DSYNC;
1086                                 goto dsync_mode;
1087                         }
1088                 }
1089
1090                 sci->sc_stage.dirty_file_ptr = NULL;
1091                 sci->sc_stage.gc_inode_ptr = NULL;
1092                 if (mode == SC_FLUSH_DAT) {
1093                         sci->sc_stage.scnt = NILFS_ST_DAT;
1094                         goto dat_stage;
1095                 }
1096                 sci->sc_stage.scnt++;  /* Fall through */
1097         case NILFS_ST_GC:
1098                 if (nilfs_doing_gc()) {
1099                         head = &sci->sc_gc_inodes;
1100                         ii = list_prepare_entry(sci->sc_stage.gc_inode_ptr,
1101                                                 head, i_dirty);
1102                         list_for_each_entry_continue(ii, head, i_dirty) {
1103                                 err = nilfs_segctor_scan_file(
1104                                         sci, &ii->vfs_inode,
1105                                         &nilfs_sc_file_ops);
1106                                 if (unlikely(err)) {
1107                                         sci->sc_stage.gc_inode_ptr = list_entry(
1108                                                 ii->i_dirty.prev,
1109                                                 struct nilfs_inode_info,
1110                                                 i_dirty);
1111                                         goto break_or_fail;
1112                                 }
1113                                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
1114                         }
1115                         sci->sc_stage.gc_inode_ptr = NULL;
1116                 }
1117                 sci->sc_stage.scnt++;  /* Fall through */
1118         case NILFS_ST_FILE:
1119                 head = &sci->sc_dirty_files;
1120                 ii = list_prepare_entry(sci->sc_stage.dirty_file_ptr, head,
1121                                         i_dirty);
1122                 list_for_each_entry_continue(ii, head, i_dirty) {
1123                         clear_bit(NILFS_I_DIRTY, &ii->i_state);
1124
1125                         err = nilfs_segctor_scan_file(sci, &ii->vfs_inode,
1126                                                       &nilfs_sc_file_ops);
1127                         if (unlikely(err)) {
1128                                 sci->sc_stage.dirty_file_ptr =
1129                                         list_entry(ii->i_dirty.prev,
1130                                                    struct nilfs_inode_info,
1131                                                    i_dirty);
1132                                 goto break_or_fail;
1133                         }
1134                         /* sci->sc_stage.dirty_file_ptr = NILFS_I(inode); */
1135                         /* XXX: required ? */
1136                 }
1137                 sci->sc_stage.dirty_file_ptr = NULL;
1138                 if (mode == SC_FLUSH_FILE) {
1139                         sci->sc_stage.scnt = NILFS_ST_DONE;
1140                         return 0;
1141                 }
1142                 sci->sc_stage.scnt++;
1143                 sci->sc_stage.flags |= NILFS_CF_IFILE_STARTED;
1144                 /* Fall through */
1145         case NILFS_ST_IFILE:
1146                 err = nilfs_segctor_scan_file(sci, sci->sc_root->ifile,
1147                                               &nilfs_sc_file_ops);
1148                 if (unlikely(err))
1149                         break;
1150                 sci->sc_stage.scnt++;
1151                 /* Creating a checkpoint */
1152                 err = nilfs_segctor_create_checkpoint(sci);
1153                 if (unlikely(err))
1154                         break;
1155                 /* Fall through */
1156         case NILFS_ST_CPFILE:
1157                 err = nilfs_segctor_scan_file(sci, nilfs->ns_cpfile,
1158                                               &nilfs_sc_file_ops);
1159                 if (unlikely(err))
1160                         break;
1161                 sci->sc_stage.scnt++;  /* Fall through */
1162         case NILFS_ST_SUFILE:
1163                 err = nilfs_sufile_freev(nilfs->ns_sufile, sci->sc_freesegs,
1164                                          sci->sc_nfreesegs, &ndone);
1165                 if (unlikely(err)) {
1166                         nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1167                                                   sci->sc_freesegs, ndone,
1168                                                   NULL);
1169                         break;
1170                 }
1171                 sci->sc_stage.flags |= NILFS_CF_SUFREED;
1172
1173                 err = nilfs_segctor_scan_file(sci, nilfs->ns_sufile,
1174                                               &nilfs_sc_file_ops);
1175                 if (unlikely(err))
1176                         break;
1177                 sci->sc_stage.scnt++;  /* Fall through */
1178         case NILFS_ST_DAT:
1179  dat_stage:
1180                 err = nilfs_segctor_scan_file(sci, nilfs_dat_inode(nilfs),
1181                                               &nilfs_sc_dat_ops);
1182                 if (unlikely(err))
1183                         break;
1184                 if (mode == SC_FLUSH_DAT) {
1185                         sci->sc_stage.scnt = NILFS_ST_DONE;
1186                         return 0;
1187                 }
1188                 sci->sc_stage.scnt++;  /* Fall through */
1189         case NILFS_ST_SR:
1190                 if (mode == SC_LSEG_SR) {
1191                         /* Appending a super root */
1192                         err = nilfs_segctor_add_super_root(sci);
1193                         if (unlikely(err))
1194                                 break;
1195                 }
1196                 /* End of a logical segment */
1197                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1198                 sci->sc_stage.scnt = NILFS_ST_DONE;
1199                 return 0;
1200         case NILFS_ST_DSYNC:
1201  dsync_mode:
1202                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_SYNDT;
1203                 ii = sci->sc_dsync_inode;
1204                 if (!test_bit(NILFS_I_BUSY, &ii->i_state))
1205                         break;
1206
1207                 err = nilfs_segctor_scan_file_dsync(sci, &ii->vfs_inode);
1208                 if (unlikely(err))
1209                         break;
1210                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1211                 sci->sc_stage.scnt = NILFS_ST_DONE;
1212                 return 0;
1213         case NILFS_ST_DONE:
1214                 return 0;
1215         default:
1216                 BUG();
1217         }
1218
1219  break_or_fail:
1220         return err;
1221 }
1222
1223 /**
1224  * nilfs_segctor_begin_construction - setup segment buffer to make a new log
1225  * @sci: nilfs_sc_info
1226  * @nilfs: nilfs object
1227  */
1228 static int nilfs_segctor_begin_construction(struct nilfs_sc_info *sci,
1229                                             struct the_nilfs *nilfs)
1230 {
1231         struct nilfs_segment_buffer *segbuf, *prev;
1232         __u64 nextnum;
1233         int err, alloc = 0;
1234
1235         segbuf = nilfs_segbuf_new(sci->sc_super);
1236         if (unlikely(!segbuf))
1237                 return -ENOMEM;
1238
1239         if (list_empty(&sci->sc_write_logs)) {
1240                 nilfs_segbuf_map(segbuf, nilfs->ns_segnum,
1241                                  nilfs->ns_pseg_offset, nilfs);
1242                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1243                         nilfs_shift_to_next_segment(nilfs);
1244                         nilfs_segbuf_map(segbuf, nilfs->ns_segnum, 0, nilfs);
1245                 }
1246
1247                 segbuf->sb_sum.seg_seq = nilfs->ns_seg_seq;
1248                 nextnum = nilfs->ns_nextnum;
1249
1250                 if (nilfs->ns_segnum == nilfs->ns_nextnum)
1251                         /* Start from the head of a new full segment */
1252                         alloc++;
1253         } else {
1254                 /* Continue logs */
1255                 prev = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1256                 nilfs_segbuf_map_cont(segbuf, prev);
1257                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq;
1258                 nextnum = prev->sb_nextnum;
1259
1260                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1261                         nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1262                         segbuf->sb_sum.seg_seq++;
1263                         alloc++;
1264                 }
1265         }
1266
1267         err = nilfs_sufile_mark_dirty(nilfs->ns_sufile, segbuf->sb_segnum);
1268         if (err)
1269                 goto failed;
1270
1271         if (alloc) {
1272                 err = nilfs_sufile_alloc(nilfs->ns_sufile, &nextnum);
1273                 if (err)
1274                         goto failed;
1275         }
1276         nilfs_segbuf_set_next_segnum(segbuf, nextnum, nilfs);
1277
1278         BUG_ON(!list_empty(&sci->sc_segbufs));
1279         list_add_tail(&segbuf->sb_list, &sci->sc_segbufs);
1280         sci->sc_segbuf_nblocks = segbuf->sb_rest_blocks;
1281         return 0;
1282
1283  failed:
1284         nilfs_segbuf_free(segbuf);
1285         return err;
1286 }
1287
1288 static int nilfs_segctor_extend_segments(struct nilfs_sc_info *sci,
1289                                          struct the_nilfs *nilfs, int nadd)
1290 {
1291         struct nilfs_segment_buffer *segbuf, *prev;
1292         struct inode *sufile = nilfs->ns_sufile;
1293         __u64 nextnextnum;
1294         LIST_HEAD(list);
1295         int err, ret, i;
1296
1297         prev = NILFS_LAST_SEGBUF(&sci->sc_segbufs);
1298         /*
1299          * Since the segment specified with nextnum might be allocated during
1300          * the previous construction, the buffer including its segusage may
1301          * not be dirty.  The following call ensures that the buffer is dirty
1302          * and will pin the buffer on memory until the sufile is written.
1303          */
1304         err = nilfs_sufile_mark_dirty(sufile, prev->sb_nextnum);
1305         if (unlikely(err))
1306                 return err;
1307
1308         for (i = 0; i < nadd; i++) {
1309                 /* extend segment info */
1310                 err = -ENOMEM;
1311                 segbuf = nilfs_segbuf_new(sci->sc_super);
1312                 if (unlikely(!segbuf))
1313                         goto failed;
1314
1315                 /* map this buffer to region of segment on-disk */
1316                 nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1317                 sci->sc_segbuf_nblocks += segbuf->sb_rest_blocks;
1318
1319                 /* allocate the next next full segment */
1320                 err = nilfs_sufile_alloc(sufile, &nextnextnum);
1321                 if (unlikely(err))
1322                         goto failed_segbuf;
1323
1324                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq + 1;
1325                 nilfs_segbuf_set_next_segnum(segbuf, nextnextnum, nilfs);
1326
1327                 list_add_tail(&segbuf->sb_list, &list);
1328                 prev = segbuf;
1329         }
1330         list_splice_tail(&list, &sci->sc_segbufs);
1331         return 0;
1332
1333  failed_segbuf:
1334         nilfs_segbuf_free(segbuf);
1335  failed:
1336         list_for_each_entry(segbuf, &list, sb_list) {
1337                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1338                 WARN_ON(ret); /* never fails */
1339         }
1340         nilfs_destroy_logs(&list);
1341         return err;
1342 }
1343
1344 static void nilfs_free_incomplete_logs(struct list_head *logs,
1345                                        struct the_nilfs *nilfs)
1346 {
1347         struct nilfs_segment_buffer *segbuf, *prev;
1348         struct inode *sufile = nilfs->ns_sufile;
1349         int ret;
1350
1351         segbuf = NILFS_FIRST_SEGBUF(logs);
1352         if (nilfs->ns_nextnum != segbuf->sb_nextnum) {
1353                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1354                 WARN_ON(ret); /* never fails */
1355         }
1356         if (atomic_read(&segbuf->sb_err)) {
1357                 /* Case 1: The first segment failed */
1358                 if (segbuf->sb_pseg_start != segbuf->sb_fseg_start)
1359                         /* Case 1a:  Partial segment appended into an existing
1360                            segment */
1361                         nilfs_terminate_segment(nilfs, segbuf->sb_fseg_start,
1362                                                 segbuf->sb_fseg_end);
1363                 else /* Case 1b:  New full segment */
1364                         set_nilfs_discontinued(nilfs);
1365         }
1366
1367         prev = segbuf;
1368         list_for_each_entry_continue(segbuf, logs, sb_list) {
1369                 if (prev->sb_nextnum != segbuf->sb_nextnum) {
1370                         ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1371                         WARN_ON(ret); /* never fails */
1372                 }
1373                 if (atomic_read(&segbuf->sb_err) &&
1374                     segbuf->sb_segnum != nilfs->ns_nextnum)
1375                         /* Case 2: extended segment (!= next) failed */
1376                         nilfs_sufile_set_error(sufile, segbuf->sb_segnum);
1377                 prev = segbuf;
1378         }
1379 }
1380
1381 static void nilfs_segctor_update_segusage(struct nilfs_sc_info *sci,
1382                                           struct inode *sufile)
1383 {
1384         struct nilfs_segment_buffer *segbuf;
1385         unsigned long live_blocks;
1386         int ret;
1387
1388         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1389                 live_blocks = segbuf->sb_sum.nblocks +
1390                         (segbuf->sb_pseg_start - segbuf->sb_fseg_start);
1391                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1392                                                      live_blocks,
1393                                                      sci->sc_seg_ctime);
1394                 WARN_ON(ret); /* always succeed because the segusage is dirty */
1395         }
1396 }
1397
1398 static void nilfs_cancel_segusage(struct list_head *logs, struct inode *sufile)
1399 {
1400         struct nilfs_segment_buffer *segbuf;
1401         int ret;
1402
1403         segbuf = NILFS_FIRST_SEGBUF(logs);
1404         ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1405                                              segbuf->sb_pseg_start -
1406                                              segbuf->sb_fseg_start, 0);
1407         WARN_ON(ret); /* always succeed because the segusage is dirty */
1408
1409         list_for_each_entry_continue(segbuf, logs, sb_list) {
1410                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1411                                                      0, 0);
1412                 WARN_ON(ret); /* always succeed */
1413         }
1414 }
1415
1416 static void nilfs_segctor_truncate_segments(struct nilfs_sc_info *sci,
1417                                             struct nilfs_segment_buffer *last,
1418                                             struct inode *sufile)
1419 {
1420         struct nilfs_segment_buffer *segbuf = last;
1421         int ret;
1422
1423         list_for_each_entry_continue(segbuf, &sci->sc_segbufs, sb_list) {
1424                 sci->sc_segbuf_nblocks -= segbuf->sb_rest_blocks;
1425                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1426                 WARN_ON(ret);
1427         }
1428         nilfs_truncate_logs(&sci->sc_segbufs, last);
1429 }
1430
1431
1432 static int nilfs_segctor_collect(struct nilfs_sc_info *sci,
1433                                  struct the_nilfs *nilfs, int mode)
1434 {
1435         struct nilfs_cstage prev_stage = sci->sc_stage;
1436         int err, nadd = 1;
1437
1438         /* Collection retry loop */
1439         for (;;) {
1440                 sci->sc_nblk_this_inc = 0;
1441                 sci->sc_curseg = NILFS_FIRST_SEGBUF(&sci->sc_segbufs);
1442
1443                 err = nilfs_segctor_reset_segment_buffer(sci);
1444                 if (unlikely(err))
1445                         goto failed;
1446
1447                 err = nilfs_segctor_collect_blocks(sci, mode);
1448                 sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
1449                 if (!err)
1450                         break;
1451
1452                 if (unlikely(err != -E2BIG))
1453                         goto failed;
1454
1455                 /* The current segment is filled up */
1456                 if (mode != SC_LSEG_SR || sci->sc_stage.scnt < NILFS_ST_CPFILE)
1457                         break;
1458
1459                 nilfs_clear_logs(&sci->sc_segbufs);
1460
1461                 err = nilfs_segctor_extend_segments(sci, nilfs, nadd);
1462                 if (unlikely(err))
1463                         return err;
1464
1465                 if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1466                         err = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1467                                                         sci->sc_freesegs,
1468                                                         sci->sc_nfreesegs,
1469                                                         NULL);
1470                         WARN_ON(err); /* do not happen */
1471                 }
1472                 nadd = min_t(int, nadd << 1, SC_MAX_SEGDELTA);
1473                 sci->sc_stage = prev_stage;
1474         }
1475         nilfs_segctor_truncate_segments(sci, sci->sc_curseg, nilfs->ns_sufile);
1476         return 0;
1477
1478  failed:
1479         return err;
1480 }
1481
1482 static void nilfs_list_replace_buffer(struct buffer_head *old_bh,
1483                                       struct buffer_head *new_bh)
1484 {
1485         BUG_ON(!list_empty(&new_bh->b_assoc_buffers));
1486
1487         list_replace_init(&old_bh->b_assoc_buffers, &new_bh->b_assoc_buffers);
1488         /* The caller must release old_bh */
1489 }
1490
1491 static int
1492 nilfs_segctor_update_payload_blocknr(struct nilfs_sc_info *sci,
1493                                      struct nilfs_segment_buffer *segbuf,
1494                                      int mode)
1495 {
1496         struct inode *inode = NULL;
1497         sector_t blocknr;
1498         unsigned long nfinfo = segbuf->sb_sum.nfinfo;
1499         unsigned long nblocks = 0, ndatablk = 0;
1500         struct nilfs_sc_operations *sc_op = NULL;
1501         struct nilfs_segsum_pointer ssp;
1502         struct nilfs_finfo *finfo = NULL;
1503         union nilfs_binfo binfo;
1504         struct buffer_head *bh, *bh_org;
1505         ino_t ino = 0;
1506         int err = 0;
1507
1508         if (!nfinfo)
1509                 goto out;
1510
1511         blocknr = segbuf->sb_pseg_start + segbuf->sb_sum.nsumblk;
1512         ssp.bh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
1513         ssp.offset = sizeof(struct nilfs_segment_summary);
1514
1515         list_for_each_entry(bh, &segbuf->sb_payload_buffers, b_assoc_buffers) {
1516                 if (bh == segbuf->sb_super_root)
1517                         break;
1518                 if (!finfo) {
1519                         finfo = nilfs_segctor_map_segsum_entry(
1520                                 sci, &ssp, sizeof(*finfo));
1521                         ino = le64_to_cpu(finfo->fi_ino);
1522                         nblocks = le32_to_cpu(finfo->fi_nblocks);
1523                         ndatablk = le32_to_cpu(finfo->fi_ndatablk);
1524
1525                         if (buffer_nilfs_node(bh))
1526                                 inode = NILFS_BTNC_I(bh->b_page->mapping);
1527                         else
1528                                 inode = NILFS_AS_I(bh->b_page->mapping);
1529
1530                         if (mode == SC_LSEG_DSYNC)
1531                                 sc_op = &nilfs_sc_dsync_ops;
1532                         else if (ino == NILFS_DAT_INO)
1533                                 sc_op = &nilfs_sc_dat_ops;
1534                         else /* file blocks */
1535                                 sc_op = &nilfs_sc_file_ops;
1536                 }
1537                 bh_org = bh;
1538                 get_bh(bh_org);
1539                 err = nilfs_bmap_assign(NILFS_I(inode)->i_bmap, &bh, blocknr,
1540                                         &binfo);
1541                 if (bh != bh_org)
1542                         nilfs_list_replace_buffer(bh_org, bh);
1543                 brelse(bh_org);
1544                 if (unlikely(err))
1545                         goto failed_bmap;
1546
1547                 if (ndatablk > 0)
1548                         sc_op->write_data_binfo(sci, &ssp, &binfo);
1549                 else
1550                         sc_op->write_node_binfo(sci, &ssp, &binfo);
1551
1552                 blocknr++;
1553                 if (--nblocks == 0) {
1554                         finfo = NULL;
1555                         if (--nfinfo == 0)
1556                                 break;
1557                 } else if (ndatablk > 0)
1558                         ndatablk--;
1559         }
1560  out:
1561         return 0;
1562
1563  failed_bmap:
1564         err = nilfs_handle_bmap_error(err, __func__, inode, sci->sc_super);
1565         return err;
1566 }
1567
1568 static int nilfs_segctor_assign(struct nilfs_sc_info *sci, int mode)
1569 {
1570         struct nilfs_segment_buffer *segbuf;
1571         int err;
1572
1573         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1574                 err = nilfs_segctor_update_payload_blocknr(sci, segbuf, mode);
1575                 if (unlikely(err))
1576                         return err;
1577                 nilfs_segbuf_fill_in_segsum(segbuf);
1578         }
1579         return 0;
1580 }
1581
1582 static int
1583 nilfs_copy_replace_page_buffers(struct page *page, struct list_head *out)
1584 {
1585         struct page *clone_page;
1586         struct buffer_head *bh, *head, *bh2;
1587         void *kaddr;
1588
1589         bh = head = page_buffers(page);
1590
1591         clone_page = nilfs_alloc_private_page(bh->b_bdev, bh->b_size, 0);
1592         if (unlikely(!clone_page))
1593                 return -ENOMEM;
1594
1595         bh2 = page_buffers(clone_page);
1596         kaddr = kmap_atomic(page, KM_USER0);
1597         do {
1598                 if (list_empty(&bh->b_assoc_buffers))
1599                         continue;
1600                 get_bh(bh2);
1601                 page_cache_get(clone_page); /* for each bh */
1602                 memcpy(bh2->b_data, kaddr + bh_offset(bh), bh2->b_size);
1603                 bh2->b_blocknr = bh->b_blocknr;
1604                 list_replace(&bh->b_assoc_buffers, &bh2->b_assoc_buffers);
1605                 list_add_tail(&bh->b_assoc_buffers, out);
1606         } while (bh = bh->b_this_page, bh2 = bh2->b_this_page, bh != head);
1607         kunmap_atomic(kaddr, KM_USER0);
1608
1609         if (!TestSetPageWriteback(clone_page))
1610                 inc_zone_page_state(clone_page, NR_WRITEBACK);
1611         unlock_page(clone_page);
1612
1613         return 0;
1614 }
1615
1616 static int nilfs_test_page_to_be_frozen(struct page *page)
1617 {
1618         struct address_space *mapping = page->mapping;
1619
1620         if (!mapping || !mapping->host || S_ISDIR(mapping->host->i_mode))
1621                 return 0;
1622
1623         if (page_mapped(page)) {
1624                 ClearPageChecked(page);
1625                 return 1;
1626         }
1627         return PageChecked(page);
1628 }
1629
1630 static int nilfs_begin_page_io(struct page *page, struct list_head *out)
1631 {
1632         if (!page || PageWriteback(page))
1633                 /* For split b-tree node pages, this function may be called
1634                    twice.  We ignore the 2nd or later calls by this check. */
1635                 return 0;
1636
1637         lock_page(page);
1638         clear_page_dirty_for_io(page);
1639         set_page_writeback(page);
1640         unlock_page(page);
1641
1642         if (nilfs_test_page_to_be_frozen(page)) {
1643                 int err = nilfs_copy_replace_page_buffers(page, out);
1644                 if (unlikely(err))
1645                         return err;
1646         }
1647         return 0;
1648 }
1649
1650 static int nilfs_segctor_prepare_write(struct nilfs_sc_info *sci,
1651                                        struct page **failed_page)
1652 {
1653         struct nilfs_segment_buffer *segbuf;
1654         struct page *bd_page = NULL, *fs_page = NULL;
1655         struct list_head *list = &sci->sc_copied_buffers;
1656         int err;
1657
1658         *failed_page = NULL;
1659         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1660                 struct buffer_head *bh;
1661
1662                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1663                                     b_assoc_buffers) {
1664                         if (bh->b_page != bd_page) {
1665                                 if (bd_page) {
1666                                         lock_page(bd_page);
1667                                         clear_page_dirty_for_io(bd_page);
1668                                         set_page_writeback(bd_page);
1669                                         unlock_page(bd_page);
1670                                 }
1671                                 bd_page = bh->b_page;
1672                         }
1673                 }
1674
1675                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1676                                     b_assoc_buffers) {
1677                         if (bh == segbuf->sb_super_root) {
1678                                 if (bh->b_page != bd_page) {
1679                                         lock_page(bd_page);
1680                                         clear_page_dirty_for_io(bd_page);
1681                                         set_page_writeback(bd_page);
1682                                         unlock_page(bd_page);
1683                                         bd_page = bh->b_page;
1684                                 }
1685                                 break;
1686                         }
1687                         if (bh->b_page != fs_page) {
1688                                 err = nilfs_begin_page_io(fs_page, list);
1689                                 if (unlikely(err)) {
1690                                         *failed_page = fs_page;
1691                                         goto out;
1692                                 }
1693                                 fs_page = bh->b_page;
1694                         }
1695                 }
1696         }
1697         if (bd_page) {
1698                 lock_page(bd_page);
1699                 clear_page_dirty_for_io(bd_page);
1700                 set_page_writeback(bd_page);
1701                 unlock_page(bd_page);
1702         }
1703         err = nilfs_begin_page_io(fs_page, list);
1704         if (unlikely(err))
1705                 *failed_page = fs_page;
1706  out:
1707         return err;
1708 }
1709
1710 static int nilfs_segctor_write(struct nilfs_sc_info *sci,
1711                                struct the_nilfs *nilfs)
1712 {
1713         int ret;
1714
1715         ret = nilfs_write_logs(&sci->sc_segbufs, nilfs);
1716         list_splice_tail_init(&sci->sc_segbufs, &sci->sc_write_logs);
1717         return ret;
1718 }
1719
1720 static void __nilfs_end_page_io(struct page *page, int err)
1721 {
1722         if (!err) {
1723                 if (!nilfs_page_buffers_clean(page))
1724                         __set_page_dirty_nobuffers(page);
1725                 ClearPageError(page);
1726         } else {
1727                 __set_page_dirty_nobuffers(page);
1728                 SetPageError(page);
1729         }
1730
1731         if (buffer_nilfs_allocated(page_buffers(page))) {
1732                 if (TestClearPageWriteback(page))
1733                         dec_zone_page_state(page, NR_WRITEBACK);
1734         } else
1735                 end_page_writeback(page);
1736 }
1737
1738 static void nilfs_end_page_io(struct page *page, int err)
1739 {
1740         if (!page)
1741                 return;
1742
1743         if (buffer_nilfs_node(page_buffers(page)) && !PageWriteback(page)) {
1744                 /*
1745                  * For b-tree node pages, this function may be called twice
1746                  * or more because they might be split in a segment.
1747                  */
1748                 if (PageDirty(page)) {
1749                         /*
1750                          * For pages holding split b-tree node buffers, dirty
1751                          * flag on the buffers may be cleared discretely.
1752                          * In that case, the page is once redirtied for
1753                          * remaining buffers, and it must be cancelled if
1754                          * all the buffers get cleaned later.
1755                          */
1756                         lock_page(page);
1757                         if (nilfs_page_buffers_clean(page))
1758                                 __nilfs_clear_page_dirty(page);
1759                         unlock_page(page);
1760                 }
1761                 return;
1762         }
1763
1764         __nilfs_end_page_io(page, err);
1765 }
1766
1767 static void nilfs_clear_copied_buffers(struct list_head *list, int err)
1768 {
1769         struct buffer_head *bh, *head;
1770         struct page *page;
1771
1772         while (!list_empty(list)) {
1773                 bh = list_entry(list->next, struct buffer_head,
1774                                 b_assoc_buffers);
1775                 page = bh->b_page;
1776                 page_cache_get(page);
1777                 head = bh = page_buffers(page);
1778                 do {
1779                         if (!list_empty(&bh->b_assoc_buffers)) {
1780                                 list_del_init(&bh->b_assoc_buffers);
1781                                 if (!err) {
1782                                         set_buffer_uptodate(bh);
1783                                         clear_buffer_dirty(bh);
1784                                         clear_buffer_nilfs_volatile(bh);
1785                                 }
1786                                 brelse(bh); /* for b_assoc_buffers */
1787                         }
1788                 } while ((bh = bh->b_this_page) != head);
1789
1790                 __nilfs_end_page_io(page, err);
1791                 page_cache_release(page);
1792         }
1793 }
1794
1795 static void nilfs_abort_logs(struct list_head *logs, struct page *failed_page,
1796                              int err)
1797 {
1798         struct nilfs_segment_buffer *segbuf;
1799         struct page *bd_page = NULL, *fs_page = NULL;
1800         struct buffer_head *bh;
1801
1802         if (list_empty(logs))
1803                 return;
1804
1805         list_for_each_entry(segbuf, logs, sb_list) {
1806                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1807                                     b_assoc_buffers) {
1808                         if (bh->b_page != bd_page) {
1809                                 if (bd_page)
1810                                         end_page_writeback(bd_page);
1811                                 bd_page = bh->b_page;
1812                         }
1813                 }
1814
1815                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1816                                     b_assoc_buffers) {
1817                         if (bh == segbuf->sb_super_root) {
1818                                 if (bh->b_page != bd_page) {
1819                                         end_page_writeback(bd_page);
1820                                         bd_page = bh->b_page;
1821                                 }
1822                                 break;
1823                         }
1824                         if (bh->b_page != fs_page) {
1825                                 nilfs_end_page_io(fs_page, err);
1826                                 if (fs_page && fs_page == failed_page)
1827                                         return;
1828                                 fs_page = bh->b_page;
1829                         }
1830                 }
1831         }
1832         if (bd_page)
1833                 end_page_writeback(bd_page);
1834
1835         nilfs_end_page_io(fs_page, err);
1836 }
1837
1838 static void nilfs_segctor_abort_construction(struct nilfs_sc_info *sci,
1839                                              struct the_nilfs *nilfs, int err)
1840 {
1841         LIST_HEAD(logs);
1842         int ret;
1843
1844         list_splice_tail_init(&sci->sc_write_logs, &logs);
1845         ret = nilfs_wait_on_logs(&logs);
1846         nilfs_abort_logs(&logs, NULL, ret ? : err);
1847
1848         list_splice_tail_init(&sci->sc_segbufs, &logs);
1849         nilfs_cancel_segusage(&logs, nilfs->ns_sufile);
1850         nilfs_free_incomplete_logs(&logs, nilfs);
1851         nilfs_clear_copied_buffers(&sci->sc_copied_buffers, err);
1852
1853         if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1854                 ret = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1855                                                 sci->sc_freesegs,
1856                                                 sci->sc_nfreesegs,
1857                                                 NULL);
1858                 WARN_ON(ret); /* do not happen */
1859         }
1860
1861         nilfs_destroy_logs(&logs);
1862 }
1863
1864 static void nilfs_set_next_segment(struct the_nilfs *nilfs,
1865                                    struct nilfs_segment_buffer *segbuf)
1866 {
1867         nilfs->ns_segnum = segbuf->sb_segnum;
1868         nilfs->ns_nextnum = segbuf->sb_nextnum;
1869         nilfs->ns_pseg_offset = segbuf->sb_pseg_start - segbuf->sb_fseg_start
1870                 + segbuf->sb_sum.nblocks;
1871         nilfs->ns_seg_seq = segbuf->sb_sum.seg_seq;
1872         nilfs->ns_ctime = segbuf->sb_sum.ctime;
1873 }
1874
1875 static void nilfs_segctor_complete_write(struct nilfs_sc_info *sci)
1876 {
1877         struct nilfs_segment_buffer *segbuf;
1878         struct page *bd_page = NULL, *fs_page = NULL;
1879         struct the_nilfs *nilfs = sci->sc_sbi->s_nilfs;
1880         int update_sr = false;
1881
1882         list_for_each_entry(segbuf, &sci->sc_write_logs, sb_list) {
1883                 struct buffer_head *bh;
1884
1885                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1886                                     b_assoc_buffers) {
1887                         set_buffer_uptodate(bh);
1888                         clear_buffer_dirty(bh);
1889                         if (bh->b_page != bd_page) {
1890                                 if (bd_page)
1891                                         end_page_writeback(bd_page);
1892                                 bd_page = bh->b_page;
1893                         }
1894                 }
1895                 /*
1896                  * We assume that the buffers which belong to the same page
1897                  * continue over the buffer list.
1898                  * Under this assumption, the last BHs of pages is
1899                  * identifiable by the discontinuity of bh->b_page
1900                  * (page != fs_page).
1901                  *
1902                  * For B-tree node blocks, however, this assumption is not
1903                  * guaranteed.  The cleanup code of B-tree node pages needs
1904                  * special care.
1905                  */
1906                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1907                                     b_assoc_buffers) {
1908                         set_buffer_uptodate(bh);
1909                         clear_buffer_dirty(bh);
1910                         clear_buffer_nilfs_volatile(bh);
1911                         clear_buffer_nilfs_redirected(bh);
1912                         if (bh == segbuf->sb_super_root) {
1913                                 if (bh->b_page != bd_page) {
1914                                         end_page_writeback(bd_page);
1915                                         bd_page = bh->b_page;
1916                                 }
1917                                 update_sr = true;
1918                                 break;
1919                         }
1920                         if (bh->b_page != fs_page) {
1921                                 nilfs_end_page_io(fs_page, 0);
1922                                 fs_page = bh->b_page;
1923                         }
1924                 }
1925
1926                 if (!nilfs_segbuf_simplex(segbuf)) {
1927                         if (segbuf->sb_sum.flags & NILFS_SS_LOGBGN) {
1928                                 set_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1929                                 sci->sc_lseg_stime = jiffies;
1930                         }
1931                         if (segbuf->sb_sum.flags & NILFS_SS_LOGEND)
1932                                 clear_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1933                 }
1934         }
1935         /*
1936          * Since pages may continue over multiple segment buffers,
1937          * end of the last page must be checked outside of the loop.
1938          */
1939         if (bd_page)
1940                 end_page_writeback(bd_page);
1941
1942         nilfs_end_page_io(fs_page, 0);
1943
1944         nilfs_clear_copied_buffers(&sci->sc_copied_buffers, 0);
1945
1946         nilfs_drop_collected_inodes(&sci->sc_dirty_files);
1947
1948         if (nilfs_doing_gc()) {
1949                 nilfs_drop_collected_inodes(&sci->sc_gc_inodes);
1950                 if (update_sr)
1951                         nilfs_commit_gcdat_inode(nilfs);
1952         } else
1953                 nilfs->ns_nongc_ctime = sci->sc_seg_ctime;
1954
1955         sci->sc_nblk_inc += sci->sc_nblk_this_inc;
1956
1957         segbuf = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1958         nilfs_set_next_segment(nilfs, segbuf);
1959
1960         if (update_sr) {
1961                 nilfs_set_last_segment(nilfs, segbuf->sb_pseg_start,
1962                                        segbuf->sb_sum.seg_seq, nilfs->ns_cno++);
1963
1964                 clear_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
1965                 clear_bit(NILFS_SC_DIRTY, &sci->sc_flags);
1966                 set_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1967                 nilfs_segctor_clear_metadata_dirty(sci);
1968         } else
1969                 clear_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1970 }
1971
1972 static int nilfs_segctor_wait(struct nilfs_sc_info *sci)
1973 {
1974         int ret;
1975
1976         ret = nilfs_wait_on_logs(&sci->sc_write_logs);
1977         if (!ret) {
1978                 nilfs_segctor_complete_write(sci);
1979                 nilfs_destroy_logs(&sci->sc_write_logs);
1980         }
1981         return ret;
1982 }
1983
1984 static int nilfs_segctor_check_in_files(struct nilfs_sc_info *sci,
1985                                         struct nilfs_sb_info *sbi)
1986 {
1987         struct nilfs_inode_info *ii, *n;
1988         struct inode *ifile = sci->sc_root->ifile;
1989
1990         spin_lock(&sbi->s_inode_lock);
1991  retry:
1992         list_for_each_entry_safe(ii, n, &sbi->s_dirty_files, i_dirty) {
1993                 if (!ii->i_bh) {
1994                         struct buffer_head *ibh;
1995                         int err;
1996
1997                         spin_unlock(&sbi->s_inode_lock);
1998                         err = nilfs_ifile_get_inode_block(
1999                                 ifile, ii->vfs_inode.i_ino, &ibh);
2000                         if (unlikely(err)) {
2001                                 nilfs_warning(sbi->s_super, __func__,
2002                                               "failed to get inode block.\n");
2003                                 return err;
2004                         }
2005                         nilfs_mdt_mark_buffer_dirty(ibh);
2006                         nilfs_mdt_mark_dirty(ifile);
2007                         spin_lock(&sbi->s_inode_lock);
2008                         if (likely(!ii->i_bh))
2009                                 ii->i_bh = ibh;
2010                         else
2011                                 brelse(ibh);
2012                         goto retry;
2013                 }
2014
2015                 clear_bit(NILFS_I_QUEUED, &ii->i_state);
2016                 set_bit(NILFS_I_BUSY, &ii->i_state);
2017                 list_del(&ii->i_dirty);
2018                 list_add_tail(&ii->i_dirty, &sci->sc_dirty_files);
2019         }
2020         spin_unlock(&sbi->s_inode_lock);
2021
2022         return 0;
2023 }
2024
2025 static void nilfs_segctor_check_out_files(struct nilfs_sc_info *sci,
2026                                           struct nilfs_sb_info *sbi)
2027 {
2028         struct nilfs_transaction_info *ti = current->journal_info;
2029         struct nilfs_inode_info *ii, *n;
2030
2031         spin_lock(&sbi->s_inode_lock);
2032         list_for_each_entry_safe(ii, n, &sci->sc_dirty_files, i_dirty) {
2033                 if (!test_and_clear_bit(NILFS_I_UPDATED, &ii->i_state) ||
2034                     test_bit(NILFS_I_DIRTY, &ii->i_state))
2035                         continue;
2036
2037                 clear_bit(NILFS_I_BUSY, &ii->i_state);
2038                 brelse(ii->i_bh);
2039                 ii->i_bh = NULL;
2040                 list_del(&ii->i_dirty);
2041                 list_add_tail(&ii->i_dirty, &ti->ti_garbage);
2042         }
2043         spin_unlock(&sbi->s_inode_lock);
2044 }
2045
2046 /*
2047  * Main procedure of segment constructor
2048  */
2049 static int nilfs_segctor_do_construct(struct nilfs_sc_info *sci, int mode)
2050 {
2051         struct nilfs_sb_info *sbi = sci->sc_sbi;
2052         struct the_nilfs *nilfs = sbi->s_nilfs;
2053         struct page *failed_page;
2054         int err;
2055
2056         sci->sc_stage.scnt = NILFS_ST_INIT;
2057         sci->sc_cno = nilfs->ns_cno;
2058
2059         err = nilfs_segctor_check_in_files(sci, sbi);
2060         if (unlikely(err))
2061                 goto out;
2062
2063         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
2064                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
2065
2066         if (nilfs_segctor_clean(sci))
2067                 goto out;
2068
2069         do {
2070                 sci->sc_stage.flags &= ~NILFS_CF_HISTORY_MASK;
2071
2072                 err = nilfs_segctor_begin_construction(sci, nilfs);
2073                 if (unlikely(err))
2074                         goto out;
2075
2076                 /* Update time stamp */
2077                 sci->sc_seg_ctime = get_seconds();
2078
2079                 err = nilfs_segctor_collect(sci, nilfs, mode);
2080                 if (unlikely(err))
2081                         goto failed;
2082
2083                 /* Avoid empty segment */
2084                 if (sci->sc_stage.scnt == NILFS_ST_DONE &&
2085                     nilfs_segbuf_empty(sci->sc_curseg)) {
2086                         nilfs_segctor_abort_construction(sci, nilfs, 1);
2087                         goto out;
2088                 }
2089
2090                 err = nilfs_segctor_assign(sci, mode);
2091                 if (unlikely(err))
2092                         goto failed;
2093
2094                 if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2095                         nilfs_segctor_fill_in_file_bmap(sci);
2096
2097                 if (mode == SC_LSEG_SR &&
2098                     sci->sc_stage.scnt >= NILFS_ST_CPFILE) {
2099                         err = nilfs_segctor_fill_in_checkpoint(sci);
2100                         if (unlikely(err))
2101                                 goto failed_to_write;
2102
2103                         nilfs_segctor_fill_in_super_root(sci, nilfs);
2104                 }
2105                 nilfs_segctor_update_segusage(sci, nilfs->ns_sufile);
2106
2107                 /* Write partial segments */
2108                 err = nilfs_segctor_prepare_write(sci, &failed_page);
2109                 if (err) {
2110                         nilfs_abort_logs(&sci->sc_segbufs, failed_page, err);
2111                         goto failed_to_write;
2112                 }
2113
2114                 nilfs_add_checksums_on_logs(&sci->sc_segbufs,
2115                                             nilfs->ns_crc_seed);
2116
2117                 err = nilfs_segctor_write(sci, nilfs);
2118                 if (unlikely(err))
2119                         goto failed_to_write;
2120
2121                 if (sci->sc_stage.scnt == NILFS_ST_DONE ||
2122                     nilfs->ns_blocksize_bits != PAGE_CACHE_SHIFT) {
2123                         /*
2124                          * At this point, we avoid double buffering
2125                          * for blocksize < pagesize because page dirty
2126                          * flag is turned off during write and dirty
2127                          * buffers are not properly collected for
2128                          * pages crossing over segments.
2129                          */
2130                         err = nilfs_segctor_wait(sci);
2131                         if (err)
2132                                 goto failed_to_write;
2133                 }
2134         } while (sci->sc_stage.scnt != NILFS_ST_DONE);
2135
2136  out:
2137         nilfs_segctor_check_out_files(sci, sbi);
2138         return err;
2139
2140  failed_to_write:
2141         if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2142                 nilfs_redirty_inodes(&sci->sc_dirty_files);
2143
2144  failed:
2145         if (nilfs_doing_gc())
2146                 nilfs_redirty_inodes(&sci->sc_gc_inodes);
2147         nilfs_segctor_abort_construction(sci, nilfs, err);
2148         goto out;
2149 }
2150
2151 /**
2152  * nilfs_segctor_start_timer - set timer of background write
2153  * @sci: nilfs_sc_info
2154  *
2155  * If the timer has already been set, it ignores the new request.
2156  * This function MUST be called within a section locking the segment
2157  * semaphore.
2158  */
2159 static void nilfs_segctor_start_timer(struct nilfs_sc_info *sci)
2160 {
2161         spin_lock(&sci->sc_state_lock);
2162         if (!(sci->sc_state & NILFS_SEGCTOR_COMMIT)) {
2163                 sci->sc_timer.expires = jiffies + sci->sc_interval;
2164                 add_timer(&sci->sc_timer);
2165                 sci->sc_state |= NILFS_SEGCTOR_COMMIT;
2166         }
2167         spin_unlock(&sci->sc_state_lock);
2168 }
2169
2170 static void nilfs_segctor_do_flush(struct nilfs_sc_info *sci, int bn)
2171 {
2172         spin_lock(&sci->sc_state_lock);
2173         if (!(sci->sc_flush_request & (1 << bn))) {
2174                 unsigned long prev_req = sci->sc_flush_request;
2175
2176                 sci->sc_flush_request |= (1 << bn);
2177                 if (!prev_req)
2178                         wake_up(&sci->sc_wait_daemon);
2179         }
2180         spin_unlock(&sci->sc_state_lock);
2181 }
2182
2183 /**
2184  * nilfs_flush_segment - trigger a segment construction for resource control
2185  * @sb: super block
2186  * @ino: inode number of the file to be flushed out.
2187  */
2188 void nilfs_flush_segment(struct super_block *sb, ino_t ino)
2189 {
2190         struct nilfs_sb_info *sbi = NILFS_SB(sb);
2191         struct nilfs_sc_info *sci = NILFS_SC(sbi);
2192
2193         if (!sci || nilfs_doing_construction())
2194                 return;
2195         nilfs_segctor_do_flush(sci, NILFS_MDT_INODE(sb, ino) ? ino : 0);
2196                                         /* assign bit 0 to data files */
2197 }
2198
2199 struct nilfs_segctor_wait_request {
2200         wait_queue_t    wq;
2201         __u32           seq;
2202         int             err;
2203         atomic_t        done;
2204 };
2205
2206 static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
2207 {
2208         struct nilfs_segctor_wait_request wait_req;
2209         int err = 0;
2210
2211         spin_lock(&sci->sc_state_lock);
2212         init_wait(&wait_req.wq);
2213         wait_req.err = 0;
2214         atomic_set(&wait_req.done, 0);
2215         wait_req.seq = ++sci->sc_seq_request;
2216         spin_unlock(&sci->sc_state_lock);
2217
2218         init_waitqueue_entry(&wait_req.wq, current);
2219         add_wait_queue(&sci->sc_wait_request, &wait_req.wq);
2220         set_current_state(TASK_INTERRUPTIBLE);
2221         wake_up(&sci->sc_wait_daemon);
2222
2223         for (;;) {
2224                 if (atomic_read(&wait_req.done)) {
2225                         err = wait_req.err;
2226                         break;
2227                 }
2228                 if (!signal_pending(current)) {
2229                         schedule();
2230                         continue;
2231                 }
2232                 err = -ERESTARTSYS;
2233                 break;
2234         }
2235         finish_wait(&sci->sc_wait_request, &wait_req.wq);
2236         return err;
2237 }
2238
2239 static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err)
2240 {
2241         struct nilfs_segctor_wait_request *wrq, *n;
2242         unsigned long flags;
2243
2244         spin_lock_irqsave(&sci->sc_wait_request.lock, flags);
2245         list_for_each_entry_safe(wrq, n, &sci->sc_wait_request.task_list,
2246                                  wq.task_list) {
2247                 if (!atomic_read(&wrq->done) &&
2248                     nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq)) {
2249                         wrq->err = err;
2250                         atomic_set(&wrq->done, 1);
2251                 }
2252                 if (atomic_read(&wrq->done)) {
2253                         wrq->wq.func(&wrq->wq,
2254                                      TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2255                                      0, NULL);
2256                 }
2257         }
2258         spin_unlock_irqrestore(&sci->sc_wait_request.lock, flags);
2259 }
2260
2261 /**
2262  * nilfs_construct_segment - construct a logical segment
2263  * @sb: super block
2264  *
2265  * Return Value: On success, 0 is retured. On errors, one of the following
2266  * negative error code is returned.
2267  *
2268  * %-EROFS - Read only filesystem.
2269  *
2270  * %-EIO - I/O error
2271  *
2272  * %-ENOSPC - No space left on device (only in a panic state).
2273  *
2274  * %-ERESTARTSYS - Interrupted.
2275  *
2276  * %-ENOMEM - Insufficient memory available.
2277  */
2278 int nilfs_construct_segment(struct super_block *sb)
2279 {
2280         struct nilfs_sb_info *sbi = NILFS_SB(sb);
2281         struct nilfs_sc_info *sci = NILFS_SC(sbi);
2282         struct nilfs_transaction_info *ti;
2283         int err;
2284
2285         if (!sci)
2286                 return -EROFS;
2287
2288         /* A call inside transactions causes a deadlock. */
2289         BUG_ON((ti = current->journal_info) && ti->ti_magic == NILFS_TI_MAGIC);
2290
2291         err = nilfs_segctor_sync(sci);
2292         return err;
2293 }
2294
2295 /**
2296  * nilfs_construct_dsync_segment - construct a data-only logical segment
2297  * @sb: super block
2298  * @inode: inode whose data blocks should be written out
2299  * @start: start byte offset
2300  * @end: end byte offset (inclusive)
2301  *
2302  * Return Value: On success, 0 is retured. On errors, one of the following
2303  * negative error code is returned.
2304  *
2305  * %-EROFS - Read only filesystem.
2306  *
2307  * %-EIO - I/O error
2308  *
2309  * %-ENOSPC - No space left on device (only in a panic state).
2310  *
2311  * %-ERESTARTSYS - Interrupted.
2312  *
2313  * %-ENOMEM - Insufficient memory available.
2314  */
2315 int nilfs_construct_dsync_segment(struct super_block *sb, struct inode *inode,
2316                                   loff_t start, loff_t end)
2317 {
2318         struct nilfs_sb_info *sbi = NILFS_SB(sb);
2319         struct nilfs_sc_info *sci = NILFS_SC(sbi);
2320         struct nilfs_inode_info *ii;
2321         struct nilfs_transaction_info ti;
2322         int err = 0;
2323
2324         if (!sci)
2325                 return -EROFS;
2326
2327         nilfs_transaction_lock(sbi, &ti, 0);
2328
2329         ii = NILFS_I(inode);
2330         if (test_bit(NILFS_I_INODE_DIRTY, &ii->i_state) ||
2331             nilfs_test_opt(sbi, STRICT_ORDER) ||
2332             test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2333             nilfs_discontinued(sbi->s_nilfs)) {
2334                 nilfs_transaction_unlock(sbi);
2335                 err = nilfs_segctor_sync(sci);
2336                 return err;
2337         }
2338
2339         spin_lock(&sbi->s_inode_lock);
2340         if (!test_bit(NILFS_I_QUEUED, &ii->i_state) &&
2341             !test_bit(NILFS_I_BUSY, &ii->i_state)) {
2342                 spin_unlock(&sbi->s_inode_lock);
2343                 nilfs_transaction_unlock(sbi);
2344                 return 0;
2345         }
2346         spin_unlock(&sbi->s_inode_lock);
2347         sci->sc_dsync_inode = ii;
2348         sci->sc_dsync_start = start;
2349         sci->sc_dsync_end = end;
2350
2351         err = nilfs_segctor_do_construct(sci, SC_LSEG_DSYNC);
2352
2353         nilfs_transaction_unlock(sbi);
2354         return err;
2355 }
2356
2357 #define FLUSH_FILE_BIT  (0x1) /* data file only */
2358 #define FLUSH_DAT_BIT   (1 << NILFS_DAT_INO) /* DAT only */
2359
2360 /**
2361  * nilfs_segctor_accept - record accepted sequence count of log-write requests
2362  * @sci: segment constructor object
2363  */
2364 static void nilfs_segctor_accept(struct nilfs_sc_info *sci)
2365 {
2366         spin_lock(&sci->sc_state_lock);
2367         sci->sc_seq_accepted = sci->sc_seq_request;
2368         spin_unlock(&sci->sc_state_lock);
2369         del_timer_sync(&sci->sc_timer);
2370 }
2371
2372 /**
2373  * nilfs_segctor_notify - notify the result of request to caller threads
2374  * @sci: segment constructor object
2375  * @mode: mode of log forming
2376  * @err: error code to be notified
2377  */
2378 static void nilfs_segctor_notify(struct nilfs_sc_info *sci, int mode, int err)
2379 {
2380         /* Clear requests (even when the construction failed) */
2381         spin_lock(&sci->sc_state_lock);
2382
2383         if (mode == SC_LSEG_SR) {
2384                 sci->sc_state &= ~NILFS_SEGCTOR_COMMIT;
2385                 sci->sc_seq_done = sci->sc_seq_accepted;
2386                 nilfs_segctor_wakeup(sci, err);
2387                 sci->sc_flush_request = 0;
2388         } else {
2389                 if (mode == SC_FLUSH_FILE)
2390                         sci->sc_flush_request &= ~FLUSH_FILE_BIT;
2391                 else if (mode == SC_FLUSH_DAT)
2392                         sci->sc_flush_request &= ~FLUSH_DAT_BIT;
2393
2394                 /* re-enable timer if checkpoint creation was not done */
2395                 if ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2396                     time_before(jiffies, sci->sc_timer.expires))
2397                         add_timer(&sci->sc_timer);
2398         }
2399         spin_unlock(&sci->sc_state_lock);
2400 }
2401
2402 /**
2403  * nilfs_segctor_construct - form logs and write them to disk
2404  * @sci: segment constructor object
2405  * @mode: mode of log forming
2406  */
2407 static int nilfs_segctor_construct(struct nilfs_sc_info *sci, int mode)
2408 {
2409         struct nilfs_sb_info *sbi = sci->sc_sbi;
2410         struct the_nilfs *nilfs = sbi->s_nilfs;
2411         struct nilfs_super_block **sbp;
2412         int err = 0;
2413
2414         nilfs_segctor_accept(sci);
2415
2416         if (nilfs_discontinued(nilfs))
2417                 mode = SC_LSEG_SR;
2418         if (!nilfs_segctor_confirm(sci))
2419                 err = nilfs_segctor_do_construct(sci, mode);
2420
2421         if (likely(!err)) {
2422                 if (mode != SC_FLUSH_DAT)
2423                         atomic_set(&nilfs->ns_ndirtyblks, 0);
2424                 if (test_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags) &&
2425                     nilfs_discontinued(nilfs)) {
2426                         down_write(&nilfs->ns_sem);
2427                         err = -EIO;
2428                         sbp = nilfs_prepare_super(sbi,
2429                                                   nilfs_sb_will_flip(nilfs));
2430                         if (likely(sbp)) {
2431                                 nilfs_set_log_cursor(sbp[0], nilfs);
2432                                 err = nilfs_commit_super(sbi, NILFS_SB_COMMIT);
2433                         }
2434                         up_write(&nilfs->ns_sem);
2435                 }
2436         }
2437
2438         nilfs_segctor_notify(sci, mode, err);
2439         return err;
2440 }
2441
2442 static void nilfs_construction_timeout(unsigned long data)
2443 {
2444         struct task_struct *p = (struct task_struct *)data;
2445         wake_up_process(p);
2446 }
2447
2448 static void
2449 nilfs_remove_written_gcinodes(struct the_nilfs *nilfs, struct list_head *head)
2450 {
2451         struct nilfs_inode_info *ii, *n;
2452
2453         list_for_each_entry_safe(ii, n, head, i_dirty) {
2454                 if (!test_bit(NILFS_I_UPDATED, &ii->i_state))
2455                         continue;
2456                 list_del_init(&ii->i_dirty);
2457                 iput(&ii->vfs_inode);
2458         }
2459 }
2460
2461 int nilfs_clean_segments(struct super_block *sb, struct nilfs_argv *argv,
2462                          void **kbufs)
2463 {
2464         struct nilfs_sb_info *sbi = NILFS_SB(sb);
2465         struct nilfs_sc_info *sci = NILFS_SC(sbi);
2466         struct the_nilfs *nilfs = sbi->s_nilfs;
2467         struct nilfs_transaction_info ti;
2468         int err;
2469
2470         if (unlikely(!sci))
2471                 return -EROFS;
2472
2473         nilfs_transaction_lock(sbi, &ti, 1);
2474
2475         err = nilfs_init_gcdat_inode(nilfs);
2476         if (unlikely(err))
2477                 goto out_unlock;
2478
2479         err = nilfs_ioctl_prepare_clean_segments(nilfs, argv, kbufs);
2480         if (unlikely(err))
2481                 goto out_unlock;
2482
2483         sci->sc_freesegs = kbufs[4];
2484         sci->sc_nfreesegs = argv[4].v_nmembs;
2485         list_splice_tail_init(&nilfs->ns_gc_inodes, &sci->sc_gc_inodes);
2486
2487         for (;;) {
2488                 err = nilfs_segctor_construct(sci, SC_LSEG_SR);
2489                 nilfs_remove_written_gcinodes(nilfs, &sci->sc_gc_inodes);
2490
2491                 if (likely(!err))
2492                         break;
2493
2494                 nilfs_warning(sb, __func__,
2495                               "segment construction failed. (err=%d)", err);
2496                 set_current_state(TASK_INTERRUPTIBLE);
2497                 schedule_timeout(sci->sc_interval);
2498         }
2499         if (nilfs_test_opt(sbi, DISCARD)) {
2500                 int ret = nilfs_discard_segments(nilfs, sci->sc_freesegs,
2501                                                  sci->sc_nfreesegs);
2502                 if (ret) {
2503                         printk(KERN_WARNING
2504                                "NILFS warning: error %d on discard request, "
2505                                "turning discards off for the device\n", ret);
2506                         nilfs_clear_opt(sbi, DISCARD);
2507                 }
2508         }
2509
2510  out_unlock:
2511         sci->sc_freesegs = NULL;
2512         sci->sc_nfreesegs = 0;
2513         nilfs_clear_gcdat_inode(nilfs);
2514         nilfs_transaction_unlock(sbi);
2515         return err;
2516 }
2517
2518 static void nilfs_segctor_thread_construct(struct nilfs_sc_info *sci, int mode)
2519 {
2520         struct nilfs_sb_info *sbi = sci->sc_sbi;
2521         struct nilfs_transaction_info ti;
2522
2523         nilfs_transaction_lock(sbi, &ti, 0);
2524         nilfs_segctor_construct(sci, mode);
2525
2526         /*
2527          * Unclosed segment should be retried.  We do this using sc_timer.
2528          * Timeout of sc_timer will invoke complete construction which leads
2529          * to close the current logical segment.
2530          */
2531         if (test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags))
2532                 nilfs_segctor_start_timer(sci);
2533
2534         nilfs_transaction_unlock(sbi);
2535 }
2536
2537 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *sci)
2538 {
2539         int mode = 0;
2540         int err;
2541
2542         spin_lock(&sci->sc_state_lock);
2543         mode = (sci->sc_flush_request & FLUSH_DAT_BIT) ?
2544                 SC_FLUSH_DAT : SC_FLUSH_FILE;
2545         spin_unlock(&sci->sc_state_lock);
2546
2547         if (mode) {
2548                 err = nilfs_segctor_do_construct(sci, mode);
2549
2550                 spin_lock(&sci->sc_state_lock);
2551                 sci->sc_flush_request &= (mode == SC_FLUSH_FILE) ?
2552                         ~FLUSH_FILE_BIT : ~FLUSH_DAT_BIT;
2553                 spin_unlock(&sci->sc_state_lock);
2554         }
2555         clear_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
2556 }
2557
2558 static int nilfs_segctor_flush_mode(struct nilfs_sc_info *sci)
2559 {
2560         if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2561             time_before(jiffies, sci->sc_lseg_stime + sci->sc_mjcp_freq)) {
2562                 if (!(sci->sc_flush_request & ~FLUSH_FILE_BIT))
2563                         return SC_FLUSH_FILE;
2564                 else if (!(sci->sc_flush_request & ~FLUSH_DAT_BIT))
2565                         return SC_FLUSH_DAT;
2566         }
2567         return SC_LSEG_SR;
2568 }
2569
2570 /**
2571  * nilfs_segctor_thread - main loop of the segment constructor thread.
2572  * @arg: pointer to a struct nilfs_sc_info.
2573  *
2574  * nilfs_segctor_thread() initializes a timer and serves as a daemon
2575  * to execute segment constructions.
2576  */
2577 static int nilfs_segctor_thread(void *arg)
2578 {
2579         struct nilfs_sc_info *sci = (struct nilfs_sc_info *)arg;
2580         struct the_nilfs *nilfs = sci->sc_sbi->s_nilfs;
2581         int timeout = 0;
2582
2583         sci->sc_timer.data = (unsigned long)current;
2584         sci->sc_timer.function = nilfs_construction_timeout;
2585
2586         /* start sync. */
2587         sci->sc_task = current;
2588         wake_up(&sci->sc_wait_task); /* for nilfs_segctor_start_thread() */
2589         printk(KERN_INFO
2590                "segctord starting. Construction interval = %lu seconds, "
2591                "CP frequency < %lu seconds\n",
2592                sci->sc_interval / HZ, sci->sc_mjcp_freq / HZ);
2593
2594         spin_lock(&sci->sc_state_lock);
2595  loop:
2596         for (;;) {
2597                 int mode;
2598
2599                 if (sci->sc_state & NILFS_SEGCTOR_QUIT)
2600                         goto end_thread;
2601
2602                 if (timeout || sci->sc_seq_request != sci->sc_seq_done)
2603                         mode = SC_LSEG_SR;
2604                 else if (!sci->sc_flush_request)
2605                         break;
2606                 else
2607                         mode = nilfs_segctor_flush_mode(sci);
2608
2609                 spin_unlock(&sci->sc_state_lock);
2610                 nilfs_segctor_thread_construct(sci, mode);
2611                 spin_lock(&sci->sc_state_lock);
2612                 timeout = 0;
2613         }
2614
2615
2616         if (freezing(current)) {
2617                 spin_unlock(&sci->sc_state_lock);
2618                 refrigerator();
2619                 spin_lock(&sci->sc_state_lock);
2620         } else {
2621                 DEFINE_WAIT(wait);
2622                 int should_sleep = 1;
2623
2624                 prepare_to_wait(&sci->sc_wait_daemon, &wait,
2625                                 TASK_INTERRUPTIBLE);
2626
2627                 if (sci->sc_seq_request != sci->sc_seq_done)
2628                         should_sleep = 0;
2629                 else if (sci->sc_flush_request)
2630                         should_sleep = 0;
2631                 else if (sci->sc_state & NILFS_SEGCTOR_COMMIT)
2632                         should_sleep = time_before(jiffies,
2633                                         sci->sc_timer.expires);
2634
2635                 if (should_sleep) {
2636                         spin_unlock(&sci->sc_state_lock);
2637                         schedule();
2638                         spin_lock(&sci->sc_state_lock);
2639                 }
2640                 finish_wait(&sci->sc_wait_daemon, &wait);
2641                 timeout = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2642                            time_after_eq(jiffies, sci->sc_timer.expires));
2643
2644                 if (nilfs_sb_dirty(nilfs) && nilfs_sb_need_update(nilfs))
2645                         set_nilfs_discontinued(nilfs);
2646         }
2647         goto loop;
2648
2649  end_thread:
2650         spin_unlock(&sci->sc_state_lock);
2651
2652         /* end sync. */
2653         sci->sc_task = NULL;
2654         wake_up(&sci->sc_wait_task); /* for nilfs_segctor_kill_thread() */
2655         return 0;
2656 }
2657
2658 static int nilfs_segctor_start_thread(struct nilfs_sc_info *sci)
2659 {
2660         struct task_struct *t;
2661
2662         t = kthread_run(nilfs_segctor_thread, sci, "segctord");
2663         if (IS_ERR(t)) {
2664                 int err = PTR_ERR(t);
2665
2666                 printk(KERN_ERR "NILFS: error %d creating segctord thread\n",
2667                        err);
2668                 return err;
2669         }
2670         wait_event(sci->sc_wait_task, sci->sc_task != NULL);
2671         return 0;
2672 }
2673
2674 static void nilfs_segctor_kill_thread(struct nilfs_sc_info *sci)
2675 {
2676         sci->sc_state |= NILFS_SEGCTOR_QUIT;
2677
2678         while (sci->sc_task) {
2679                 wake_up(&sci->sc_wait_daemon);
2680                 spin_unlock(&sci->sc_state_lock);
2681                 wait_event(sci->sc_wait_task, sci->sc_task == NULL);
2682                 spin_lock(&sci->sc_state_lock);
2683         }
2684 }
2685
2686 /*
2687  * Setup & clean-up functions
2688  */
2689 static struct nilfs_sc_info *nilfs_segctor_new(struct nilfs_sb_info *sbi,
2690                                                struct nilfs_root *root)
2691 {
2692         struct nilfs_sc_info *sci;
2693
2694         sci = kzalloc(sizeof(*sci), GFP_KERNEL);
2695         if (!sci)
2696                 return NULL;
2697
2698         sci->sc_sbi = sbi;
2699         sci->sc_super = sbi->s_super;
2700
2701         nilfs_get_root(root);
2702         sci->sc_root = root;
2703
2704         init_waitqueue_head(&sci->sc_wait_request);
2705         init_waitqueue_head(&sci->sc_wait_daemon);
2706         init_waitqueue_head(&sci->sc_wait_task);
2707         spin_lock_init(&sci->sc_state_lock);
2708         INIT_LIST_HEAD(&sci->sc_dirty_files);
2709         INIT_LIST_HEAD(&sci->sc_segbufs);
2710         INIT_LIST_HEAD(&sci->sc_write_logs);
2711         INIT_LIST_HEAD(&sci->sc_gc_inodes);
2712         INIT_LIST_HEAD(&sci->sc_copied_buffers);
2713         init_timer(&sci->sc_timer);
2714
2715         sci->sc_interval = HZ * NILFS_SC_DEFAULT_TIMEOUT;
2716         sci->sc_mjcp_freq = HZ * NILFS_SC_DEFAULT_SR_FREQ;
2717         sci->sc_watermark = NILFS_SC_DEFAULT_WATERMARK;
2718
2719         if (sbi->s_interval)
2720                 sci->sc_interval = sbi->s_interval;
2721         if (sbi->s_watermark)
2722                 sci->sc_watermark = sbi->s_watermark;
2723         return sci;
2724 }
2725
2726 static void nilfs_segctor_write_out(struct nilfs_sc_info *sci)
2727 {
2728         int ret, retrycount = NILFS_SC_CLEANUP_RETRY;
2729
2730         /* The segctord thread was stopped and its timer was removed.
2731            But some tasks remain. */
2732         do {
2733                 struct nilfs_sb_info *sbi = sci->sc_sbi;
2734                 struct nilfs_transaction_info ti;
2735
2736                 nilfs_transaction_lock(sbi, &ti, 0);
2737                 ret = nilfs_segctor_construct(sci, SC_LSEG_SR);
2738                 nilfs_transaction_unlock(sbi);
2739
2740         } while (ret && retrycount-- > 0);
2741 }
2742
2743 /**
2744  * nilfs_segctor_destroy - destroy the segment constructor.
2745  * @sci: nilfs_sc_info
2746  *
2747  * nilfs_segctor_destroy() kills the segctord thread and frees
2748  * the nilfs_sc_info struct.
2749  * Caller must hold the segment semaphore.
2750  */
2751 static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
2752 {
2753         struct nilfs_sb_info *sbi = sci->sc_sbi;
2754         int flag;
2755
2756         up_write(&sbi->s_nilfs->ns_segctor_sem);
2757
2758         spin_lock(&sci->sc_state_lock);
2759         nilfs_segctor_kill_thread(sci);
2760         flag = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) || sci->sc_flush_request
2761                 || sci->sc_seq_request != sci->sc_seq_done);
2762         spin_unlock(&sci->sc_state_lock);
2763
2764         if (flag || !nilfs_segctor_confirm(sci))
2765                 nilfs_segctor_write_out(sci);
2766
2767         WARN_ON(!list_empty(&sci->sc_copied_buffers));
2768
2769         if (!list_empty(&sci->sc_dirty_files)) {
2770                 nilfs_warning(sbi->s_super, __func__,
2771                               "dirty file(s) after the final construction\n");
2772                 nilfs_dispose_list(sbi, &sci->sc_dirty_files, 1);
2773         }
2774
2775         WARN_ON(!list_empty(&sci->sc_segbufs));
2776         WARN_ON(!list_empty(&sci->sc_write_logs));
2777
2778         nilfs_put_root(sci->sc_root);
2779
2780         down_write(&sbi->s_nilfs->ns_segctor_sem);
2781
2782         del_timer_sync(&sci->sc_timer);
2783         kfree(sci);
2784 }
2785
2786 /**
2787  * nilfs_attach_segment_constructor - attach a segment constructor
2788  * @sbi: nilfs_sb_info
2789  * @root: root object of the current filesystem tree
2790  *
2791  * nilfs_attach_segment_constructor() allocates a struct nilfs_sc_info,
2792  * initializes it, and starts the segment constructor.
2793  *
2794  * Return Value: On success, 0 is returned. On error, one of the following
2795  * negative error code is returned.
2796  *
2797  * %-ENOMEM - Insufficient memory available.
2798  */
2799 int nilfs_attach_segment_constructor(struct nilfs_sb_info *sbi,
2800                                      struct nilfs_root *root)
2801 {
2802         struct the_nilfs *nilfs = sbi->s_nilfs;
2803         int err;
2804
2805         if (NILFS_SC(sbi)) {
2806                 /*
2807                  * This happens if the filesystem was remounted
2808                  * read/write after nilfs_error degenerated it into a
2809                  * read-only mount.
2810                  */
2811                 nilfs_detach_segment_constructor(sbi);
2812         }
2813
2814         sbi->s_sc_info = nilfs_segctor_new(sbi, root);
2815         if (!sbi->s_sc_info)
2816                 return -ENOMEM;
2817
2818         nilfs_attach_writer(nilfs, sbi);
2819         err = nilfs_segctor_start_thread(NILFS_SC(sbi));
2820         if (err) {
2821                 nilfs_detach_writer(nilfs, sbi);
2822                 kfree(sbi->s_sc_info);
2823                 sbi->s_sc_info = NULL;
2824         }
2825         return err;
2826 }
2827
2828 /**
2829  * nilfs_detach_segment_constructor - destroy the segment constructor
2830  * @sbi: nilfs_sb_info
2831  *
2832  * nilfs_detach_segment_constructor() kills the segment constructor daemon,
2833  * frees the struct nilfs_sc_info, and destroy the dirty file list.
2834  */
2835 void nilfs_detach_segment_constructor(struct nilfs_sb_info *sbi)
2836 {
2837         struct the_nilfs *nilfs = sbi->s_nilfs;
2838         LIST_HEAD(garbage_list);
2839
2840         down_write(&nilfs->ns_segctor_sem);
2841         if (NILFS_SC(sbi)) {
2842                 nilfs_segctor_destroy(NILFS_SC(sbi));
2843                 sbi->s_sc_info = NULL;
2844         }
2845
2846         /* Force to free the list of dirty files */
2847         spin_lock(&sbi->s_inode_lock);
2848         if (!list_empty(&sbi->s_dirty_files)) {
2849                 list_splice_init(&sbi->s_dirty_files, &garbage_list);
2850                 nilfs_warning(sbi->s_super, __func__,
2851                               "Non empty dirty list after the last "
2852                               "segment construction\n");
2853         }
2854         spin_unlock(&sbi->s_inode_lock);
2855         up_write(&nilfs->ns_segctor_sem);
2856
2857         nilfs_dispose_list(sbi, &garbage_list, 1);
2858         nilfs_detach_writer(nilfs, sbi);
2859 }