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