fs/nilfs2/segment.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * NILFS segment constructor.
   4  *
   5  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   6  *
   7  * Written by Ryusuke Konishi.
   8  *
   9  */
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/writeback.h>
  14 #include <linux/bitops.h>
  15 #include <linux/bio.h>
  16 #include <linux/completion.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/backing-dev.h>
  19 #include <linux/freezer.h>
  20 #include <linux/kthread.h>
  21 #include <linux/crc32.h>
  22 #include <linux/pagevec.h>
  23 #include <linux/slab.h>
  24 #include <linux/sched/signal.h>
  25
  26 #include "nilfs.h"
  27 #include "btnode.h"
  28 #include "page.h"
  29 #include "segment.h"
  30 #include "sufile.h"
  31 #include "cpfile.h"
  32 #include "ifile.h"
  33 #include "segbuf.h"
  34
  35
  36 /*
  37  * Segment constructor
  38  */
  39 #define SC_N_INODEVEC   16   /* Size of locally allocated inode vector */
  40
  41 #define SC_MAX_SEGDELTA 64   /*
  42                               * Upper limit of the number of segments
  43                               * appended in collection retry loop
  44                               */
  45
  46 /* Construction mode */
  47 enum {
  48         SC_LSEG_SR = 1, /* Make a logical segment having a super root */
  49         SC_LSEG_DSYNC,  /*
  50                          * Flush data blocks of a given file and make
  51                          * a logical segment without a super root.
  52                          */
  53         SC_FLUSH_FILE,  /*
  54                          * Flush data files, leads to segment writes without
  55                          * creating a checkpoint.
  56                          */
  57         SC_FLUSH_DAT,   /*
  58                          * Flush DAT file.  This also creates segments
  59                          * without a checkpoint.
  60                          */
  61 };
  62
  63 /* Stage numbers of dirty block collection */
  64 enum {
  65         NILFS_ST_INIT = 0,
  66         NILFS_ST_GC,            /* Collecting dirty blocks for GC */
  67         NILFS_ST_FILE,
  68         NILFS_ST_IFILE,
  69         NILFS_ST_CPFILE,
  70         NILFS_ST_SUFILE,
  71         NILFS_ST_DAT,
  72         NILFS_ST_SR,            /* Super root */
  73         NILFS_ST_DSYNC,         /* Data sync blocks */
  74         NILFS_ST_DONE,
  75 };
  76
  77 #define CREATE_TRACE_POINTS
  78 #include <trace/events/nilfs2.h>
  79
  80 /*
  81  * nilfs_sc_cstage_inc(), nilfs_sc_cstage_set(), nilfs_sc_cstage_get() are
  82  * wrapper functions of stage count (nilfs_sc_info->sc_stage.scnt). Users of
  83  * the variable must use them because transition of stage count must involve
  84  * trace events (trace_nilfs2_collection_stage_transition).
  85  *
  86  * nilfs_sc_cstage_get() isn't required for the above purpose because it doesn't
  87  * produce tracepoint events. It is provided just for making the intention
  88  * clear.
  89  */
  90 static inline void nilfs_sc_cstage_inc(struct nilfs_sc_info *sci)
  91 {
  92         sci->sc_stage.scnt++;
  93         trace_nilfs2_collection_stage_transition(sci);
  94 }
  95
  96 static inline void nilfs_sc_cstage_set(struct nilfs_sc_info *sci, int next_scnt)
  97 {
  98         sci->sc_stage.scnt = next_scnt;
  99         trace_nilfs2_collection_stage_transition(sci);
 100 }
 101
 102 static inline int nilfs_sc_cstage_get(struct nilfs_sc_info *sci)
 103 {
 104         return sci->sc_stage.scnt;
 105 }
 106
 107 /* State flags of collection */
 108 #define NILFS_CF_NODE           0x0001  /* Collecting node blocks */
 109 #define NILFS_CF_IFILE_STARTED  0x0002  /* IFILE stage has started */
 110 #define NILFS_CF_SUFREED        0x0004  /* segment usages has been freed */
 111 #define NILFS_CF_HISTORY_MASK   (NILFS_CF_IFILE_STARTED | NILFS_CF_SUFREED)
 112
 113 /* Operations depending on the construction mode and file type */
 114 struct nilfs_sc_operations {
 115         int (*collect_data)(struct nilfs_sc_info *, struct buffer_head *,
 116                             struct inode *);
 117         int (*collect_node)(struct nilfs_sc_info *, struct buffer_head *,
 118                             struct inode *);
 119         int (*collect_bmap)(struct nilfs_sc_info *, struct buffer_head *,
 120                             struct inode *);
 121         void (*write_data_binfo)(struct nilfs_sc_info *,
 122                                  struct nilfs_segsum_pointer *,
 123                                  union nilfs_binfo *);
 124         void (*write_node_binfo)(struct nilfs_sc_info *,
 125                                  struct nilfs_segsum_pointer *,
 126                                  union nilfs_binfo *);
 127 };
 128
 129 /*
 130  * Other definitions
 131  */
 132 static void nilfs_segctor_start_timer(struct nilfs_sc_info *);
 133 static void nilfs_segctor_do_flush(struct nilfs_sc_info *, int);
 134 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *);
 135 static void nilfs_dispose_list(struct the_nilfs *, struct list_head *, int);
 136
 137 #define nilfs_cnt32_ge(a, b)   \
 138         (typecheck(__u32, a) && typecheck(__u32, b) && \
 139          ((__s32)((a) - (b)) >= 0))
 140
 141 static int nilfs_prepare_segment_lock(struct super_block *sb,
 142                                       struct nilfs_transaction_info *ti)
 143 {
 144         struct nilfs_transaction_info *cur_ti = current->journal_info;
 145         void *save = NULL;
 146
 147         if (cur_ti) {
 148                 if (cur_ti->ti_magic == NILFS_TI_MAGIC)
 149                         return ++cur_ti->ti_count;
 150
 151                 /*
 152                  * If journal_info field is occupied by other FS,
 153                  * it is saved and will be restored on
 154                  * nilfs_transaction_commit().
 155                  */
 156                 nilfs_warn(sb, "journal info from a different FS");
 157                 save = current->journal_info;
 158         }
 159         if (!ti) {
 160                 ti = kmem_cache_alloc(nilfs_transaction_cachep, GFP_NOFS);
 161                 if (!ti)
 162                         return -ENOMEM;
 163                 ti->ti_flags = NILFS_TI_DYNAMIC_ALLOC;
 164         } else {
 165                 ti->ti_flags = 0;
 166         }
 167         ti->ti_count = 0;
 168         ti->ti_save = save;
 169         ti->ti_magic = NILFS_TI_MAGIC;
 170         current->journal_info = ti;
 171         return 0;
 172 }
 173
 174 /**
 175  * nilfs_transaction_begin - start indivisible file operations.
 176  * @sb: super block
 177  * @ti: nilfs_transaction_info
 178  * @vacancy_check: flags for vacancy rate checks
 179  *
 180  * nilfs_transaction_begin() acquires a reader/writer semaphore, called
 181  * the segment semaphore, to make a segment construction and write tasks
 182  * exclusive.  The function is used with nilfs_transaction_commit() in pairs.
 183  * The region enclosed by these two functions can be nested.  To avoid a
 184  * deadlock, the semaphore is only acquired or released in the outermost call.
 185  *
 186  * This function allocates a nilfs_transaction_info struct to keep context
 187  * information on it.  It is initialized and hooked onto the current task in
 188  * the outermost call.  If a pre-allocated struct is given to @ti, it is used
 189  * instead; otherwise a new struct is assigned from a slab.
 190  *
 191  * When @vacancy_check flag is set, this function will check the amount of
 192  * free space, and will wait for the GC to reclaim disk space if low capacity.
 193  *
 194  * Return: 0 on success, or one of the following negative error codes on
 195  * failure:
 196  * * %-ENOMEM   - Insufficient memory available.
 197  * * %-ENOSPC   - No space left on device (if checking free space).
 198  */
 199 int nilfs_transaction_begin(struct super_block *sb,
 200                             struct nilfs_transaction_info *ti,
 201                             int vacancy_check)
 202 {
 203         struct the_nilfs *nilfs;
 204         int ret = nilfs_prepare_segment_lock(sb, ti);
 205         struct nilfs_transaction_info *trace_ti;
 206
 207         if (unlikely(ret < 0))
 208                 return ret;
 209         if (ret > 0) {
 210                 trace_ti = current->journal_info;
 211
 212                 trace_nilfs2_transaction_transition(sb, trace_ti,
 213                                     trace_ti->ti_count, trace_ti->ti_flags,
 214                                     TRACE_NILFS2_TRANSACTION_BEGIN);
 215                 return 0;
 216         }
 217
 218         sb_start_intwrite(sb);
 219
 220         nilfs = sb->s_fs_info;
 221         down_read(&nilfs->ns_segctor_sem);
 222         if (vacancy_check && nilfs_near_disk_full(nilfs)) {
 223                 up_read(&nilfs->ns_segctor_sem);
 224                 ret = -ENOSPC;
 225                 goto failed;
 226         }
 227
 228         trace_ti = current->journal_info;
 229         trace_nilfs2_transaction_transition(sb, trace_ti, trace_ti->ti_count,
 230                                             trace_ti->ti_flags,
 231                                             TRACE_NILFS2_TRANSACTION_BEGIN);
 232         return 0;
 233
 234  failed:
 235         ti = current->journal_info;
 236         current->journal_info = ti->ti_save;
 237         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 238                 kmem_cache_free(nilfs_transaction_cachep, ti);
 239         sb_end_intwrite(sb);
 240         return ret;
 241 }
 242
 243 /**
 244  * nilfs_transaction_commit - commit indivisible file operations.
 245  * @sb: super block
 246  *
 247  * nilfs_transaction_commit() releases the read semaphore which is
 248  * acquired by nilfs_transaction_begin(). This is only performed
 249  * in outermost call of this function.  If a commit flag is set,
 250  * nilfs_transaction_commit() sets a timer to start the segment
 251  * constructor.  If a sync flag is set, it starts construction
 252  * directly.
 253  *
 254  * Return: 0 on success, or a negative error code on failure.
 255  */
 256 int nilfs_transaction_commit(struct super_block *sb)
 257 {
 258         struct nilfs_transaction_info *ti = current->journal_info;
 259         struct the_nilfs *nilfs = sb->s_fs_info;
 260         int err = 0;
 261
 262         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 263         ti->ti_flags |= NILFS_TI_COMMIT;
 264         if (ti->ti_count > 0) {
 265                 ti->ti_count--;
 266                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 267                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_COMMIT);
 268                 return 0;
 269         }
 270         if (nilfs->ns_writer) {
 271                 struct nilfs_sc_info *sci = nilfs->ns_writer;
 272
 273                 if (ti->ti_flags & NILFS_TI_COMMIT)
 274                         nilfs_segctor_start_timer(sci);
 275                 if (atomic_read(&nilfs->ns_ndirtyblks) > sci->sc_watermark)
 276                         nilfs_segctor_do_flush(sci, 0);
 277         }
 278         up_read(&nilfs->ns_segctor_sem);
 279         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 280                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_COMMIT);
 281
 282         current->journal_info = ti->ti_save;
 283
 284         if (ti->ti_flags & NILFS_TI_SYNC)
 285                 err = nilfs_construct_segment(sb);
 286         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 287                 kmem_cache_free(nilfs_transaction_cachep, ti);
 288         sb_end_intwrite(sb);
 289         return err;
 290 }
 291
 292 void nilfs_transaction_abort(struct super_block *sb)
 293 {
 294         struct nilfs_transaction_info *ti = current->journal_info;
 295         struct the_nilfs *nilfs = sb->s_fs_info;
 296
 297         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 298         if (ti->ti_count > 0) {
 299                 ti->ti_count--;
 300                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 301                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_ABORT);
 302                 return;
 303         }
 304         up_read(&nilfs->ns_segctor_sem);
 305
 306         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 307                     ti->ti_flags, TRACE_NILFS2_TRANSACTION_ABORT);
 308
 309         current->journal_info = ti->ti_save;
 310         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 311                 kmem_cache_free(nilfs_transaction_cachep, ti);
 312         sb_end_intwrite(sb);
 313 }
 314
 315 void nilfs_relax_pressure_in_lock(struct super_block *sb)
 316 {
 317         struct the_nilfs *nilfs = sb->s_fs_info;
 318         struct nilfs_sc_info *sci = nilfs->ns_writer;
 319
 320         if (sb_rdonly(sb) || unlikely(!sci) || !sci->sc_flush_request)
 321                 return;
 322
 323         set_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
 324         up_read(&nilfs->ns_segctor_sem);
 325
 326         down_write(&nilfs->ns_segctor_sem);
 327         if (sci->sc_flush_request &&
 328             test_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags)) {
 329                 struct nilfs_transaction_info *ti = current->journal_info;
 330
 331                 ti->ti_flags |= NILFS_TI_WRITER;
 332                 nilfs_segctor_do_immediate_flush(sci);
 333                 ti->ti_flags &= ~NILFS_TI_WRITER;
 334         }
 335         downgrade_write(&nilfs->ns_segctor_sem);
 336 }
 337
 338 static void nilfs_transaction_lock(struct super_block *sb,
 339                                    struct nilfs_transaction_info *ti,
 340                                    int gcflag)
 341 {
 342         struct nilfs_transaction_info *cur_ti = current->journal_info;
 343         struct the_nilfs *nilfs = sb->s_fs_info;
 344         struct nilfs_sc_info *sci = nilfs->ns_writer;
 345
 346         WARN_ON(cur_ti);
 347         ti->ti_flags = NILFS_TI_WRITER;
 348         ti->ti_count = 0;
 349         ti->ti_save = cur_ti;
 350         ti->ti_magic = NILFS_TI_MAGIC;
 351         current->journal_info = ti;
 352
 353         for (;;) {
 354                 trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 355                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_TRYLOCK);
 356
 357                 down_write(&nilfs->ns_segctor_sem);
 358                 if (!test_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags))
 359                         break;
 360
 361                 nilfs_segctor_do_immediate_flush(sci);
 362
 363                 up_write(&nilfs->ns_segctor_sem);
 364                 cond_resched();
 365         }
 366         if (gcflag)
 367                 ti->ti_flags |= NILFS_TI_GC;
 368
 369         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 370                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_LOCK);
 371 }
 372
 373 static void nilfs_transaction_unlock(struct super_block *sb)
 374 {
 375         struct nilfs_transaction_info *ti = current->journal_info;
 376         struct the_nilfs *nilfs = sb->s_fs_info;
 377
 378         BUG_ON(ti == NULL || ti->ti_magic != NILFS_TI_MAGIC);
 379         BUG_ON(ti->ti_count > 0);
 380
 381         up_write(&nilfs->ns_segctor_sem);
 382         current->journal_info = ti->ti_save;
 383
 384         trace_nilfs2_transaction_transition(sb, ti, ti->ti_count,
 385                             ti->ti_flags, TRACE_NILFS2_TRANSACTION_UNLOCK);
 386 }
 387
 388 static void *nilfs_segctor_map_segsum_entry(struct nilfs_sc_info *sci,
 389                                             struct nilfs_segsum_pointer *ssp,
 390                                             unsigned int bytes)
 391 {
 392         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 393         unsigned int blocksize = sci->sc_super->s_blocksize;
 394         void *p;
 395
 396         if (unlikely(ssp->offset + bytes > blocksize)) {
 397                 ssp->offset = 0;
 398                 BUG_ON(NILFS_SEGBUF_BH_IS_LAST(ssp->bh,
 399                                                &segbuf->sb_segsum_buffers));
 400                 ssp->bh = NILFS_SEGBUF_NEXT_BH(ssp->bh);
 401         }
 402         p = ssp->bh->b_data + ssp->offset;
 403         ssp->offset += bytes;
 404         return p;
 405 }
 406
 407 /**
 408  * nilfs_segctor_reset_segment_buffer - reset the current segment buffer
 409  * @sci: nilfs_sc_info
 410  *
 411  * Return: 0 on success, or a negative error code on failure.
 412  */
 413 static int nilfs_segctor_reset_segment_buffer(struct nilfs_sc_info *sci)
 414 {
 415         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 416         struct buffer_head *sumbh;
 417         unsigned int sumbytes;
 418         unsigned int flags = 0;
 419         int err;
 420
 421         if (nilfs_doing_gc())
 422                 flags = NILFS_SS_GC;
 423         err = nilfs_segbuf_reset(segbuf, flags, sci->sc_seg_ctime, sci->sc_cno);
 424         if (unlikely(err))
 425                 return err;
 426
 427         sumbh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
 428         sumbytes = segbuf->sb_sum.sumbytes;
 429         sci->sc_finfo_ptr.bh = sumbh;  sci->sc_finfo_ptr.offset = sumbytes;
 430         sci->sc_binfo_ptr.bh = sumbh;  sci->sc_binfo_ptr.offset = sumbytes;
 431         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 432         return 0;
 433 }
 434
 435 /**
 436  * nilfs_segctor_zeropad_segsum - zero pad the rest of the segment summary area
 437  * @sci: segment constructor object
 438  *
 439  * nilfs_segctor_zeropad_segsum() zero-fills unallocated space at the end of
 440  * the current segment summary block.
 441  */
 442 static void nilfs_segctor_zeropad_segsum(struct nilfs_sc_info *sci)
 443 {
 444         struct nilfs_segsum_pointer *ssp;
 445
 446         ssp = sci->sc_blk_cnt > 0 ? &sci->sc_binfo_ptr : &sci->sc_finfo_ptr;
 447         if (ssp->offset < ssp->bh->b_size)
 448                 memset(ssp->bh->b_data + ssp->offset, 0,
 449                        ssp->bh->b_size - ssp->offset);
 450 }
 451
 452 static int nilfs_segctor_feed_segment(struct nilfs_sc_info *sci)
 453 {
 454         sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
 455         if (NILFS_SEGBUF_IS_LAST(sci->sc_curseg, &sci->sc_segbufs))
 456                 return -E2BIG; /*
 457                                 * The current segment is filled up
 458                                 * (internal code)
 459                                 */
 460         nilfs_segctor_zeropad_segsum(sci);
 461         sci->sc_curseg = NILFS_NEXT_SEGBUF(sci->sc_curseg);
 462         return nilfs_segctor_reset_segment_buffer(sci);
 463 }
 464
 465 static int nilfs_segctor_add_super_root(struct nilfs_sc_info *sci)
 466 {
 467         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 468         int err;
 469
 470         if (segbuf->sb_sum.nblocks >= segbuf->sb_rest_blocks) {
 471                 err = nilfs_segctor_feed_segment(sci);
 472                 if (err)
 473                         return err;
 474                 segbuf = sci->sc_curseg;
 475         }
 476         err = nilfs_segbuf_extend_payload(segbuf, &segbuf->sb_super_root);
 477         if (likely(!err))
 478                 segbuf->sb_sum.flags |= NILFS_SS_SR;
 479         return err;
 480 }
 481
 482 /*
 483  * Functions for making segment summary and payloads
 484  */
 485 static int nilfs_segctor_segsum_block_required(
 486         struct nilfs_sc_info *sci, const struct nilfs_segsum_pointer *ssp,
 487         unsigned int binfo_size)
 488 {
 489         unsigned int blocksize = sci->sc_super->s_blocksize;
 490         /* Size of finfo and binfo is enough small against blocksize */
 491
 492         return ssp->offset + binfo_size +
 493                 (!sci->sc_blk_cnt ? sizeof(struct nilfs_finfo) : 0) >
 494                 blocksize;
 495 }
 496
 497 static void nilfs_segctor_begin_finfo(struct nilfs_sc_info *sci,
 498                                       struct inode *inode)
 499 {
 500         sci->sc_curseg->sb_sum.nfinfo++;
 501         sci->sc_binfo_ptr = sci->sc_finfo_ptr;
 502         nilfs_segctor_map_segsum_entry(
 503                 sci, &sci->sc_binfo_ptr, sizeof(struct nilfs_finfo));
 504
 505         if (NILFS_I(inode)->i_root &&
 506             !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 507                 set_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
 508         /* skip finfo */
 509 }
 510
 511 static void nilfs_segctor_end_finfo(struct nilfs_sc_info *sci,
 512                                     struct inode *inode)
 513 {
 514         struct nilfs_finfo *finfo;
 515         struct nilfs_inode_info *ii;
 516         struct nilfs_segment_buffer *segbuf;
 517         __u64 cno;
 518
 519         if (sci->sc_blk_cnt == 0)
 520                 return;
 521
 522         ii = NILFS_I(inode);
 523
 524         if (ii->i_type & NILFS_I_TYPE_GC)
 525                 cno = ii->i_cno;
 526         else if (NILFS_ROOT_METADATA_FILE(inode->i_ino))
 527                 cno = 0;
 528         else
 529                 cno = sci->sc_cno;
 530
 531         finfo = nilfs_segctor_map_segsum_entry(sci, &sci->sc_finfo_ptr,
 532                                                  sizeof(*finfo));
 533         finfo->fi_ino = cpu_to_le64(inode->i_ino);
 534         finfo->fi_nblocks = cpu_to_le32(sci->sc_blk_cnt);
 535         finfo->fi_ndatablk = cpu_to_le32(sci->sc_datablk_cnt);
 536         finfo->fi_cno = cpu_to_le64(cno);
 537
 538         segbuf = sci->sc_curseg;
 539         segbuf->sb_sum.sumbytes = sci->sc_binfo_ptr.offset +
 540                 sci->sc_super->s_blocksize * (segbuf->sb_sum.nsumblk - 1);
 541         sci->sc_finfo_ptr = sci->sc_binfo_ptr;
 542         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 543 }
 544
 545 static int nilfs_segctor_add_file_block(struct nilfs_sc_info *sci,
 546                                         struct buffer_head *bh,
 547                                         struct inode *inode,
 548                                         unsigned int binfo_size)
 549 {
 550         struct nilfs_segment_buffer *segbuf;
 551         int required, err = 0;
 552
 553  retry:
 554         segbuf = sci->sc_curseg;
 555         required = nilfs_segctor_segsum_block_required(
 556                 sci, &sci->sc_binfo_ptr, binfo_size);
 557         if (segbuf->sb_sum.nblocks + required + 1 > segbuf->sb_rest_blocks) {
 558                 nilfs_segctor_end_finfo(sci, inode);
 559                 err = nilfs_segctor_feed_segment(sci);
 560                 if (err)
 561                         return err;
 562                 goto retry;
 563         }
 564         if (unlikely(required)) {
 565                 nilfs_segctor_zeropad_segsum(sci);
 566                 err = nilfs_segbuf_extend_segsum(segbuf);
 567                 if (unlikely(err))
 568                         goto failed;
 569         }
 570         if (sci->sc_blk_cnt == 0)
 571                 nilfs_segctor_begin_finfo(sci, inode);
 572
 573         nilfs_segctor_map_segsum_entry(sci, &sci->sc_binfo_ptr, binfo_size);
 574         /* Substitution to vblocknr is delayed until update_blocknr() */
 575         nilfs_segbuf_add_file_buffer(segbuf, bh);
 576         sci->sc_blk_cnt++;
 577  failed:
 578         return err;
 579 }
 580
 581 /*
 582  * Callback functions that enumerate, mark, and collect dirty blocks
 583  */
 584 static int nilfs_collect_file_data(struct nilfs_sc_info *sci,
 585                                    struct buffer_head *bh, struct inode *inode)
 586 {
 587         int err;
 588
 589         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 590         if (err < 0)
 591                 return err;
 592
 593         err = nilfs_segctor_add_file_block(sci, bh, inode,
 594                                            sizeof(struct nilfs_binfo_v));
 595         if (!err)
 596                 sci->sc_datablk_cnt++;
 597         return err;
 598 }
 599
 600 static int nilfs_collect_file_node(struct nilfs_sc_info *sci,
 601                                    struct buffer_head *bh,
 602                                    struct inode *inode)
 603 {
 604         return nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 605 }
 606
 607 static int nilfs_collect_file_bmap(struct nilfs_sc_info *sci,
 608                                    struct buffer_head *bh,
 609                                    struct inode *inode)
 610 {
 611         WARN_ON(!buffer_dirty(bh));
 612         return nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 613 }
 614
 615 static void nilfs_write_file_data_binfo(struct nilfs_sc_info *sci,
 616                                         struct nilfs_segsum_pointer *ssp,
 617                                         union nilfs_binfo *binfo)
 618 {
 619         struct nilfs_binfo_v *binfo_v = nilfs_segctor_map_segsum_entry(
 620                 sci, ssp, sizeof(*binfo_v));
 621         *binfo_v = binfo->bi_v;
 622 }
 623
 624 static void nilfs_write_file_node_binfo(struct nilfs_sc_info *sci,
 625                                         struct nilfs_segsum_pointer *ssp,
 626                                         union nilfs_binfo *binfo)
 627 {
 628         __le64 *vblocknr = nilfs_segctor_map_segsum_entry(
 629                 sci, ssp, sizeof(*vblocknr));
 630         *vblocknr = binfo->bi_v.bi_vblocknr;
 631 }
 632
 633 static const struct nilfs_sc_operations nilfs_sc_file_ops = {
 634         .collect_data = nilfs_collect_file_data,
 635         .collect_node = nilfs_collect_file_node,
 636         .collect_bmap = nilfs_collect_file_bmap,
 637         .write_data_binfo = nilfs_write_file_data_binfo,
 638         .write_node_binfo = nilfs_write_file_node_binfo,
 639 };
 640
 641 static int nilfs_collect_dat_data(struct nilfs_sc_info *sci,
 642                                   struct buffer_head *bh, struct inode *inode)
 643 {
 644         int err;
 645
 646         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 647         if (err < 0)
 648                 return err;
 649
 650         err = nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 651         if (!err)
 652                 sci->sc_datablk_cnt++;
 653         return err;
 654 }
 655
 656 static int nilfs_collect_dat_bmap(struct nilfs_sc_info *sci,
 657                                   struct buffer_head *bh, struct inode *inode)
 658 {
 659         WARN_ON(!buffer_dirty(bh));
 660         return nilfs_segctor_add_file_block(sci, bh, inode,
 661                                             sizeof(struct nilfs_binfo_dat));
 662 }
 663
 664 static void nilfs_write_dat_data_binfo(struct nilfs_sc_info *sci,
 665                                        struct nilfs_segsum_pointer *ssp,
 666                                        union nilfs_binfo *binfo)
 667 {
 668         __le64 *blkoff = nilfs_segctor_map_segsum_entry(sci, ssp,
 669                                                           sizeof(*blkoff));
 670         *blkoff = binfo->bi_dat.bi_blkoff;
 671 }
 672
 673 static void nilfs_write_dat_node_binfo(struct nilfs_sc_info *sci,
 674                                        struct nilfs_segsum_pointer *ssp,
 675                                        union nilfs_binfo *binfo)
 676 {
 677         struct nilfs_binfo_dat *binfo_dat =
 678                 nilfs_segctor_map_segsum_entry(sci, ssp, sizeof(*binfo_dat));
 679         *binfo_dat = binfo->bi_dat;
 680 }
 681
 682 static const struct nilfs_sc_operations nilfs_sc_dat_ops = {
 683         .collect_data = nilfs_collect_dat_data,
 684         .collect_node = nilfs_collect_file_node,
 685         .collect_bmap = nilfs_collect_dat_bmap,
 686         .write_data_binfo = nilfs_write_dat_data_binfo,
 687         .write_node_binfo = nilfs_write_dat_node_binfo,
 688 };
 689
 690 static const struct nilfs_sc_operations nilfs_sc_dsync_ops = {
 691         .collect_data = nilfs_collect_file_data,
 692         .collect_node = NULL,
 693         .collect_bmap = NULL,
 694         .write_data_binfo = nilfs_write_file_data_binfo,
 695         .write_node_binfo = NULL,
 696 };
 697
 698 static size_t nilfs_lookup_dirty_data_buffers(struct inode *inode,
 699                                               struct list_head *listp,
 700                                               size_t nlimit,
 701                                               loff_t start, loff_t end)
 702 {
 703         struct address_space *mapping = inode->i_mapping;
 704         struct folio_batch fbatch;
 705         pgoff_t index = 0, last = ULONG_MAX;
 706         size_t ndirties = 0;
 707         int i;
 708
 709         if (unlikely(start != 0 || end != LLONG_MAX)) {
 710                 /*
 711                  * A valid range is given for sync-ing data pages. The
 712                  * range is rounded to per-page; extra dirty buffers
 713                  * may be included if blocksize < pagesize.
 714                  */
 715                 index = start >> PAGE_SHIFT;
 716                 last = end >> PAGE_SHIFT;
 717         }
 718         folio_batch_init(&fbatch);
 719  repeat:
 720         if (unlikely(index > last) ||
 721               !filemap_get_folios_tag(mapping, &index, last,
 722                       PAGECACHE_TAG_DIRTY, &fbatch))
 723                 return ndirties;
 724
 725         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 726                 struct buffer_head *bh, *head;
 727                 struct folio *folio = fbatch.folios[i];
 728
 729                 folio_lock(folio);
 730                 if (unlikely(folio->mapping != mapping)) {
 731                         /* Exclude folios removed from the address space */
 732                         folio_unlock(folio);
 733                         continue;
 734                 }
 735                 head = folio_buffers(folio);
 736                 if (!head)
 737                         head = create_empty_buffers(folio,
 738                                         i_blocksize(inode), 0);
 739
 740                 bh = head;
 741                 do {
 742                         if (!buffer_dirty(bh) || buffer_async_write(bh))
 743                                 continue;
 744                         get_bh(bh);
 745                         list_add_tail(&bh->b_assoc_buffers, listp);
 746                         ndirties++;
 747                         if (unlikely(ndirties >= nlimit)) {
 748                                 folio_unlock(folio);
 749                                 folio_batch_release(&fbatch);
 750                                 cond_resched();
 751                                 return ndirties;
 752                         }
 753                 } while (bh = bh->b_this_page, bh != head);
 754
 755                 folio_unlock(folio);
 756         }
 757         folio_batch_release(&fbatch);
 758         cond_resched();
 759         goto repeat;
 760 }
 761
 762 static void nilfs_lookup_dirty_node_buffers(struct inode *inode,
 763                                             struct list_head *listp)
 764 {
 765         struct nilfs_inode_info *ii = NILFS_I(inode);
 766         struct inode *btnc_inode = ii->i_assoc_inode;
 767         struct folio_batch fbatch;
 768         struct buffer_head *bh, *head;
 769         unsigned int i;
 770         pgoff_t index = 0;
 771
 772         if (!btnc_inode)
 773                 return;
 774         folio_batch_init(&fbatch);
 775
 776         while (filemap_get_folios_tag(btnc_inode->i_mapping, &index,
 777                                 (pgoff_t)-1, PAGECACHE_TAG_DIRTY, &fbatch)) {
 778                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 779                         bh = head = folio_buffers(fbatch.folios[i]);
 780                         do {
 781                                 if (buffer_dirty(bh) &&
 782                                                 !buffer_async_write(bh)) {
 783                                         get_bh(bh);
 784                                         list_add_tail(&bh->b_assoc_buffers,
 785                                                       listp);
 786                                 }
 787                                 bh = bh->b_this_page;
 788                         } while (bh != head);
 789                 }
 790                 folio_batch_release(&fbatch);
 791                 cond_resched();
 792         }
 793 }
 794
 795 static void nilfs_dispose_list(struct the_nilfs *nilfs,
 796                                struct list_head *head, int force)
 797 {
 798         struct nilfs_inode_info *ii, *n;
 799         struct nilfs_inode_info *ivec[SC_N_INODEVEC], **pii;
 800         unsigned int nv = 0;
 801
 802         while (!list_empty(head)) {
 803                 spin_lock(&nilfs->ns_inode_lock);
 804                 list_for_each_entry_safe(ii, n, head, i_dirty) {
 805                         list_del_init(&ii->i_dirty);
 806                         if (force) {
 807                                 if (unlikely(ii->i_bh)) {
 808                                         brelse(ii->i_bh);
 809                                         ii->i_bh = NULL;
 810                                 }
 811                         } else if (test_bit(NILFS_I_DIRTY, &ii->i_state)) {
 812                                 set_bit(NILFS_I_QUEUED, &ii->i_state);
 813                                 list_add_tail(&ii->i_dirty,
 814                                               &nilfs->ns_dirty_files);
 815                                 continue;
 816                         }
 817                         ivec[nv++] = ii;
 818                         if (nv == SC_N_INODEVEC)
 819                                 break;
 820                 }
 821                 spin_unlock(&nilfs->ns_inode_lock);
 822
 823                 for (pii = ivec; nv > 0; pii++, nv--)
 824                         iput(&(*pii)->vfs_inode);
 825         }
 826 }
 827
 828 static void nilfs_iput_work_func(struct work_struct *work)
 829 {
 830         struct nilfs_sc_info *sci = container_of(work, struct nilfs_sc_info,
 831                                                  sc_iput_work);
 832         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 833
 834         nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 0);
 835 }
 836
 837 static int nilfs_test_metadata_dirty(struct the_nilfs *nilfs,
 838                                      struct nilfs_root *root)
 839 {
 840         int ret = 0;
 841
 842         if (nilfs_mdt_fetch_dirty(root->ifile))
 843                 ret++;
 844         if (nilfs_mdt_fetch_dirty(nilfs->ns_cpfile))
 845                 ret++;
 846         if (nilfs_mdt_fetch_dirty(nilfs->ns_sufile))
 847                 ret++;
 848         if ((ret || nilfs_doing_gc()) && nilfs_mdt_fetch_dirty(nilfs->ns_dat))
 849                 ret++;
 850         return ret;
 851 }
 852
 853 static int nilfs_segctor_clean(struct nilfs_sc_info *sci)
 854 {
 855         return list_empty(&sci->sc_dirty_files) &&
 856                 !test_bit(NILFS_SC_DIRTY, &sci->sc_flags) &&
 857                 sci->sc_nfreesegs == 0 &&
 858                 (!nilfs_doing_gc() || list_empty(&sci->sc_gc_inodes));
 859 }
 860
 861 static int nilfs_segctor_confirm(struct nilfs_sc_info *sci)
 862 {
 863         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 864         int ret = 0;
 865
 866         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
 867                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
 868
 869         spin_lock(&nilfs->ns_inode_lock);
 870         if (list_empty(&nilfs->ns_dirty_files) && nilfs_segctor_clean(sci))
 871                 ret++;
 872
 873         spin_unlock(&nilfs->ns_inode_lock);
 874         return ret;
 875 }
 876
 877 static void nilfs_segctor_clear_metadata_dirty(struct nilfs_sc_info *sci)
 878 {
 879         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 880
 881         nilfs_mdt_clear_dirty(sci->sc_root->ifile);
 882         nilfs_mdt_clear_dirty(nilfs->ns_cpfile);
 883         nilfs_mdt_clear_dirty(nilfs->ns_sufile);
 884         nilfs_mdt_clear_dirty(nilfs->ns_dat);
 885 }
 886
 887 static void nilfs_fill_in_file_bmap(struct inode *ifile,
 888                                     struct nilfs_inode_info *ii)
 889
 890 {
 891         struct buffer_head *ibh;
 892         struct nilfs_inode *raw_inode;
 893
 894         if (test_bit(NILFS_I_BMAP, &ii->i_state)) {
 895                 ibh = ii->i_bh;
 896                 BUG_ON(!ibh);
 897                 raw_inode = nilfs_ifile_map_inode(ifile, ii->vfs_inode.i_ino,
 898                                                   ibh);
 899                 nilfs_bmap_write(ii->i_bmap, raw_inode);
 900                 nilfs_ifile_unmap_inode(raw_inode);
 901         }
 902 }
 903
 904 static void nilfs_segctor_fill_in_file_bmap(struct nilfs_sc_info *sci)
 905 {
 906         struct nilfs_inode_info *ii;
 907
 908         list_for_each_entry(ii, &sci->sc_dirty_files, i_dirty) {
 909                 nilfs_fill_in_file_bmap(sci->sc_root->ifile, ii);
 910                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
 911         }
 912 }
 913
 914 /**
 915  * nilfs_write_root_mdt_inode - export root metadata inode information to
 916  *                              the on-disk inode
 917  * @inode:     inode object of the root metadata file
 918  * @raw_inode: on-disk inode
 919  *
 920  * nilfs_write_root_mdt_inode() writes inode information and bmap data of
 921  * @inode to the inode area of the metadata file allocated on the super root
 922  * block created to finalize the log.  Since super root blocks are configured
 923  * each time, this function zero-fills the unused area of @raw_inode.
 924  */
 925 static void nilfs_write_root_mdt_inode(struct inode *inode,
 926                                        struct nilfs_inode *raw_inode)
 927 {
 928         struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
 929
 930         nilfs_write_inode_common(inode, raw_inode);
 931
 932         /* zero-fill unused portion of raw_inode */
 933         raw_inode->i_xattr = 0;
 934         raw_inode->i_pad = 0;
 935         memset((void *)raw_inode + sizeof(*raw_inode), 0,
 936                nilfs->ns_inode_size - sizeof(*raw_inode));
 937
 938         nilfs_bmap_write(NILFS_I(inode)->i_bmap, raw_inode);
 939 }
 940
 941 static void nilfs_segctor_fill_in_super_root(struct nilfs_sc_info *sci,
 942                                              struct the_nilfs *nilfs)
 943 {
 944         struct buffer_head *bh_sr;
 945         struct nilfs_super_root *raw_sr;
 946         unsigned int isz, srsz;
 947
 948         bh_sr = NILFS_LAST_SEGBUF(&sci->sc_segbufs)->sb_super_root;
 949
 950         lock_buffer(bh_sr);
 951         raw_sr = (struct nilfs_super_root *)bh_sr->b_data;
 952         isz = nilfs->ns_inode_size;
 953         srsz = NILFS_SR_BYTES(isz);
 954
 955         raw_sr->sr_sum = 0;  /* Ensure initialization within this update */
 956         raw_sr->sr_bytes = cpu_to_le16(srsz);
 957         raw_sr->sr_nongc_ctime
 958                 = cpu_to_le64(nilfs_doing_gc() ?
 959                               nilfs->ns_nongc_ctime : sci->sc_seg_ctime);
 960         raw_sr->sr_flags = 0;
 961
 962         nilfs_write_root_mdt_inode(nilfs->ns_dat, (void *)raw_sr +
 963                                    NILFS_SR_DAT_OFFSET(isz));
 964         nilfs_write_root_mdt_inode(nilfs->ns_cpfile, (void *)raw_sr +
 965                                    NILFS_SR_CPFILE_OFFSET(isz));
 966         nilfs_write_root_mdt_inode(nilfs->ns_sufile, (void *)raw_sr +
 967                                    NILFS_SR_SUFILE_OFFSET(isz));
 968
 969         memset((void *)raw_sr + srsz, 0, nilfs->ns_blocksize - srsz);
 970         set_buffer_uptodate(bh_sr);
 971         unlock_buffer(bh_sr);
 972 }
 973
 974 static void nilfs_redirty_inodes(struct list_head *head)
 975 {
 976         struct nilfs_inode_info *ii;
 977
 978         list_for_each_entry(ii, head, i_dirty) {
 979                 if (test_bit(NILFS_I_COLLECTED, &ii->i_state))
 980                         clear_bit(NILFS_I_COLLECTED, &ii->i_state);
 981         }
 982 }
 983
 984 static void nilfs_drop_collected_inodes(struct list_head *head)
 985 {
 986         struct nilfs_inode_info *ii;
 987
 988         list_for_each_entry(ii, head, i_dirty) {
 989                 if (!test_and_clear_bit(NILFS_I_COLLECTED, &ii->i_state))
 990                         continue;
 991
 992                 clear_bit(NILFS_I_INODE_SYNC, &ii->i_state);
 993                 set_bit(NILFS_I_UPDATED, &ii->i_state);
 994         }
 995 }
 996
 997 static int nilfs_segctor_apply_buffers(struct nilfs_sc_info *sci,
 998                                        struct inode *inode,
 999                                        struct list_head *listp,
1000                                        int (*collect)(struct nilfs_sc_info *,
1001                                                       struct buffer_head *,
1002                                                       struct inode *))
1003 {
1004         struct buffer_head *bh, *n;
1005         int err = 0;
1006
1007         if (collect) {
1008                 list_for_each_entry_safe(bh, n, listp, b_assoc_buffers) {
1009                         list_del_init(&bh->b_assoc_buffers);
1010                         err = collect(sci, bh, inode);
1011                         brelse(bh);
1012                         if (unlikely(err))
1013                                 goto dispose_buffers;
1014                 }
1015                 return 0;
1016         }
1017
1018  dispose_buffers:
1019         while (!list_empty(listp)) {
1020                 bh = list_first_entry(listp, struct buffer_head,
1021                                       b_assoc_buffers);
1022                 list_del_init(&bh->b_assoc_buffers);
1023                 brelse(bh);
1024         }
1025         return err;
1026 }
1027
1028 static size_t nilfs_segctor_buffer_rest(struct nilfs_sc_info *sci)
1029 {
1030         /* Remaining number of blocks within segment buffer */
1031         return sci->sc_segbuf_nblocks -
1032                 (sci->sc_nblk_this_inc + sci->sc_curseg->sb_sum.nblocks);
1033 }
1034
1035 static int nilfs_segctor_scan_file(struct nilfs_sc_info *sci,
1036                                    struct inode *inode,
1037                                    const struct nilfs_sc_operations *sc_ops)
1038 {
1039         LIST_HEAD(data_buffers);
1040         LIST_HEAD(node_buffers);
1041         int err;
1042
1043         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1044                 size_t n, rest = nilfs_segctor_buffer_rest(sci);
1045
1046                 n = nilfs_lookup_dirty_data_buffers(
1047                         inode, &data_buffers, rest + 1, 0, LLONG_MAX);
1048                 if (n > rest) {
1049                         err = nilfs_segctor_apply_buffers(
1050                                 sci, inode, &data_buffers,
1051                                 sc_ops->collect_data);
1052                         BUG_ON(!err); /* always receive -E2BIG or true error */
1053                         goto break_or_fail;
1054                 }
1055         }
1056         nilfs_lookup_dirty_node_buffers(inode, &node_buffers);
1057
1058         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1059                 err = nilfs_segctor_apply_buffers(
1060                         sci, inode, &data_buffers, sc_ops->collect_data);
1061                 if (unlikely(err)) {
1062                         /* dispose node list */
1063                         nilfs_segctor_apply_buffers(
1064                                 sci, inode, &node_buffers, NULL);
1065                         goto break_or_fail;
1066                 }
1067                 sci->sc_stage.flags |= NILFS_CF_NODE;
1068         }
1069         /* Collect node */
1070         err = nilfs_segctor_apply_buffers(
1071                 sci, inode, &node_buffers, sc_ops->collect_node);
1072         if (unlikely(err))
1073                 goto break_or_fail;
1074
1075         nilfs_bmap_lookup_dirty_buffers(NILFS_I(inode)->i_bmap, &node_buffers);
1076         err = nilfs_segctor_apply_buffers(
1077                 sci, inode, &node_buffers, sc_ops->collect_bmap);
1078         if (unlikely(err))
1079                 goto break_or_fail;
1080
1081         nilfs_segctor_end_finfo(sci, inode);
1082         sci->sc_stage.flags &= ~NILFS_CF_NODE;
1083
1084  break_or_fail:
1085         return err;
1086 }
1087
1088 static int nilfs_segctor_scan_file_dsync(struct nilfs_sc_info *sci,
1089                                          struct inode *inode)
1090 {
1091         LIST_HEAD(data_buffers);
1092         size_t n, rest = nilfs_segctor_buffer_rest(sci);
1093         int err;
1094
1095         n = nilfs_lookup_dirty_data_buffers(inode, &data_buffers, rest + 1,
1096                                             sci->sc_dsync_start,
1097                                             sci->sc_dsync_end);
1098
1099         err = nilfs_segctor_apply_buffers(sci, inode, &data_buffers,
1100                                           nilfs_collect_file_data);
1101         if (!err) {
1102                 nilfs_segctor_end_finfo(sci, inode);
1103                 BUG_ON(n > rest);
1104                 /* always receive -E2BIG or true error if n > rest */
1105         }
1106         return err;
1107 }
1108
1109 /**
1110  * nilfs_free_segments - free the segments given by an array of segment numbers
1111  * @nilfs:   nilfs object
1112  * @segnumv: array of segment numbers to be freed
1113  * @nsegs:   number of segments to be freed in @segnumv
1114  *
1115  * nilfs_free_segments() wraps nilfs_sufile_freev() and
1116  * nilfs_sufile_cancel_freev(), and edits the segment usage metadata file
1117  * (sufile) to free all segments given by @segnumv and @nsegs at once.  If
1118  * it fails midway, it cancels the changes so that none of the segments are
1119  * freed.  If @nsegs is 0, this function does nothing.
1120  *
1121  * The freeing of segments is not finalized until the writing of a log with
1122  * a super root block containing this sufile change is complete, and it can
1123  * be canceled with nilfs_sufile_cancel_freev() until then.
1124  *
1125  * Return: 0 on success, or one of the following negative error codes on
1126  * failure:
1127  * * %-EINVAL   - Invalid segment number.
1128  * * %-EIO      - I/O error (including metadata corruption).
1129  * * %-ENOMEM   - Insufficient memory available.
1130  */
1131 static int nilfs_free_segments(struct the_nilfs *nilfs, __u64 *segnumv,
1132                                size_t nsegs)
1133 {
1134         size_t ndone;
1135         int ret;
1136
1137         if (!nsegs)
1138                 return 0;
1139
1140         ret = nilfs_sufile_freev(nilfs->ns_sufile, segnumv, nsegs, &ndone);
1141         if (unlikely(ret)) {
1142                 nilfs_sufile_cancel_freev(nilfs->ns_sufile, segnumv, ndone,
1143                                           NULL);
1144                 /*
1145                  * If a segment usage of the segments to be freed is in a
1146                  * hole block, nilfs_sufile_freev() will return -ENOENT.
1147                  * In this case, -EINVAL should be returned to the caller
1148                  * since there is something wrong with the given segment
1149                  * number array.  This error can only occur during GC, so
1150                  * there is no need to worry about it propagating to other
1151                  * callers (such as fsync).
1152                  */
1153                 if (ret == -ENOENT) {
1154                         nilfs_err(nilfs->ns_sb,
1155                                   "The segment usage entry %llu to be freed is invalid (in a hole)",
1156                                   (unsigned long long)segnumv[ndone]);
1157                         ret = -EINVAL;
1158                 }
1159         }
1160         return ret;
1161 }
1162
1163 static int nilfs_segctor_collect_blocks(struct nilfs_sc_info *sci, int mode)
1164 {
1165         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1166         struct list_head *head;
1167         struct nilfs_inode_info *ii;
1168         int err = 0;
1169
1170         switch (nilfs_sc_cstage_get(sci)) {
1171         case NILFS_ST_INIT:
1172                 /* Pre-processes */
1173                 sci->sc_stage.flags = 0;
1174
1175                 if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags)) {
1176                         sci->sc_nblk_inc = 0;
1177                         sci->sc_curseg->sb_sum.flags = NILFS_SS_LOGBGN;
1178                         if (mode == SC_LSEG_DSYNC) {
1179                                 nilfs_sc_cstage_set(sci, NILFS_ST_DSYNC);
1180                                 goto dsync_mode;
1181                         }
1182                 }
1183
1184                 sci->sc_stage.dirty_file_ptr = NULL;
1185                 sci->sc_stage.gc_inode_ptr = NULL;
1186                 if (mode == SC_FLUSH_DAT) {
1187                         nilfs_sc_cstage_set(sci, NILFS_ST_DAT);
1188                         goto dat_stage;
1189                 }
1190                 nilfs_sc_cstage_inc(sci);
1191                 fallthrough;
1192         case NILFS_ST_GC:
1193                 if (nilfs_doing_gc()) {
1194                         head = &sci->sc_gc_inodes;
1195                         ii = list_prepare_entry(sci->sc_stage.gc_inode_ptr,
1196                                                 head, i_dirty);
1197                         list_for_each_entry_continue(ii, head, i_dirty) {
1198                                 err = nilfs_segctor_scan_file(
1199                                         sci, &ii->vfs_inode,
1200                                         &nilfs_sc_file_ops);
1201                                 if (unlikely(err)) {
1202                                         sci->sc_stage.gc_inode_ptr = list_entry(
1203                                                 ii->i_dirty.prev,
1204                                                 struct nilfs_inode_info,
1205                                                 i_dirty);
1206                                         goto break_or_fail;
1207                                 }
1208                                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
1209                         }
1210                         sci->sc_stage.gc_inode_ptr = NULL;
1211                 }
1212                 nilfs_sc_cstage_inc(sci);
1213                 fallthrough;
1214         case NILFS_ST_FILE:
1215                 head = &sci->sc_dirty_files;
1216                 ii = list_prepare_entry(sci->sc_stage.dirty_file_ptr, head,
1217                                         i_dirty);
1218                 list_for_each_entry_continue(ii, head, i_dirty) {
1219                         clear_bit(NILFS_I_DIRTY, &ii->i_state);
1220
1221                         err = nilfs_segctor_scan_file(sci, &ii->vfs_inode,
1222                                                       &nilfs_sc_file_ops);
1223                         if (unlikely(err)) {
1224                                 sci->sc_stage.dirty_file_ptr =
1225                                         list_entry(ii->i_dirty.prev,
1226                                                    struct nilfs_inode_info,
1227                                                    i_dirty);
1228                                 goto break_or_fail;
1229                         }
1230                         /* sci->sc_stage.dirty_file_ptr = NILFS_I(inode); */
1231                         /* XXX: required ? */
1232                 }
1233                 sci->sc_stage.dirty_file_ptr = NULL;
1234                 if (mode == SC_FLUSH_FILE) {
1235                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1236                         return 0;
1237                 }
1238                 nilfs_sc_cstage_inc(sci);
1239                 sci->sc_stage.flags |= NILFS_CF_IFILE_STARTED;
1240                 fallthrough;
1241         case NILFS_ST_IFILE:
1242                 err = nilfs_segctor_scan_file(sci, sci->sc_root->ifile,
1243                                               &nilfs_sc_file_ops);
1244                 if (unlikely(err))
1245                         break;
1246                 nilfs_sc_cstage_inc(sci);
1247                 /* Creating a checkpoint */
1248                 err = nilfs_cpfile_create_checkpoint(nilfs->ns_cpfile,
1249                                                      nilfs->ns_cno);
1250                 if (unlikely(err))
1251                         break;
1252                 fallthrough;
1253         case NILFS_ST_CPFILE:
1254                 err = nilfs_segctor_scan_file(sci, nilfs->ns_cpfile,
1255                                               &nilfs_sc_file_ops);
1256                 if (unlikely(err))
1257                         break;
1258                 nilfs_sc_cstage_inc(sci);
1259                 fallthrough;
1260         case NILFS_ST_SUFILE:
1261                 err = nilfs_free_segments(nilfs, sci->sc_freesegs,
1262                                           sci->sc_nfreesegs);
1263                 if (unlikely(err))
1264                         break;
1265                 sci->sc_stage.flags |= NILFS_CF_SUFREED;
1266
1267                 err = nilfs_segctor_scan_file(sci, nilfs->ns_sufile,
1268                                               &nilfs_sc_file_ops);
1269                 if (unlikely(err))
1270                         break;
1271                 nilfs_sc_cstage_inc(sci);
1272                 fallthrough;
1273         case NILFS_ST_DAT:
1274  dat_stage:
1275                 err = nilfs_segctor_scan_file(sci, nilfs->ns_dat,
1276                                               &nilfs_sc_dat_ops);
1277                 if (unlikely(err))
1278                         break;
1279                 if (mode == SC_FLUSH_DAT) {
1280                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1281                         return 0;
1282                 }
1283                 nilfs_sc_cstage_inc(sci);
1284                 fallthrough;
1285         case NILFS_ST_SR:
1286                 if (mode == SC_LSEG_SR) {
1287                         /* Appending a super root */
1288                         err = nilfs_segctor_add_super_root(sci);
1289                         if (unlikely(err))
1290                                 break;
1291                 }
1292                 /* End of a logical segment */
1293                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1294                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1295                 return 0;
1296         case NILFS_ST_DSYNC:
1297  dsync_mode:
1298                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_SYNDT;
1299                 ii = sci->sc_dsync_inode;
1300                 if (!test_bit(NILFS_I_BUSY, &ii->i_state))
1301                         break;
1302
1303                 err = nilfs_segctor_scan_file_dsync(sci, &ii->vfs_inode);
1304                 if (unlikely(err))
1305                         break;
1306                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1307                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1308                 return 0;
1309         case NILFS_ST_DONE:
1310                 return 0;
1311         default:
1312                 BUG();
1313         }
1314
1315  break_or_fail:
1316         return err;
1317 }
1318
1319 /**
1320  * nilfs_segctor_begin_construction - setup segment buffer to make a new log
1321  * @sci: nilfs_sc_info
1322  * @nilfs: nilfs object
1323  *
1324  * Return: 0 on success, or a negative error code on failure.
1325  */
1326 static int nilfs_segctor_begin_construction(struct nilfs_sc_info *sci,
1327                                             struct the_nilfs *nilfs)
1328 {
1329         struct nilfs_segment_buffer *segbuf, *prev;
1330         __u64 nextnum;
1331         int err, alloc = 0;
1332
1333         segbuf = nilfs_segbuf_new(sci->sc_super);
1334         if (unlikely(!segbuf))
1335                 return -ENOMEM;
1336
1337         if (list_empty(&sci->sc_write_logs)) {
1338                 nilfs_segbuf_map(segbuf, nilfs->ns_segnum,
1339                                  nilfs->ns_pseg_offset, nilfs);
1340                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1341                         nilfs_shift_to_next_segment(nilfs);
1342                         nilfs_segbuf_map(segbuf, nilfs->ns_segnum, 0, nilfs);
1343                 }
1344
1345                 segbuf->sb_sum.seg_seq = nilfs->ns_seg_seq;
1346                 nextnum = nilfs->ns_nextnum;
1347
1348                 if (nilfs->ns_segnum == nilfs->ns_nextnum)
1349                         /* Start from the head of a new full segment */
1350                         alloc++;
1351         } else {
1352                 /* Continue logs */
1353                 prev = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1354                 nilfs_segbuf_map_cont(segbuf, prev);
1355                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq;
1356                 nextnum = prev->sb_nextnum;
1357
1358                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1359                         nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1360                         segbuf->sb_sum.seg_seq++;
1361                         alloc++;
1362                 }
1363         }
1364
1365         err = nilfs_sufile_mark_dirty(nilfs->ns_sufile, segbuf->sb_segnum);
1366         if (err)
1367                 goto failed;
1368
1369         if (alloc) {
1370                 err = nilfs_sufile_alloc(nilfs->ns_sufile, &nextnum);
1371                 if (err)
1372                         goto failed;
1373         }
1374         nilfs_segbuf_set_next_segnum(segbuf, nextnum, nilfs);
1375
1376         BUG_ON(!list_empty(&sci->sc_segbufs));
1377         list_add_tail(&segbuf->sb_list, &sci->sc_segbufs);
1378         sci->sc_segbuf_nblocks = segbuf->sb_rest_blocks;
1379         return 0;
1380
1381  failed:
1382         nilfs_segbuf_free(segbuf);
1383         return err;
1384 }
1385
1386 static int nilfs_segctor_extend_segments(struct nilfs_sc_info *sci,
1387                                          struct the_nilfs *nilfs, int nadd)
1388 {
1389         struct nilfs_segment_buffer *segbuf, *prev;
1390         struct inode *sufile = nilfs->ns_sufile;
1391         __u64 nextnextnum;
1392         LIST_HEAD(list);
1393         int err, ret, i;
1394
1395         prev = NILFS_LAST_SEGBUF(&sci->sc_segbufs);
1396         /*
1397          * Since the segment specified with nextnum might be allocated during
1398          * the previous construction, the buffer including its segusage may
1399          * not be dirty.  The following call ensures that the buffer is dirty
1400          * and will pin the buffer on memory until the sufile is written.
1401          */
1402         err = nilfs_sufile_mark_dirty(sufile, prev->sb_nextnum);
1403         if (unlikely(err))
1404                 return err;
1405
1406         for (i = 0; i < nadd; i++) {
1407                 /* extend segment info */
1408                 err = -ENOMEM;
1409                 segbuf = nilfs_segbuf_new(sci->sc_super);
1410                 if (unlikely(!segbuf))
1411                         goto failed;
1412
1413                 /* map this buffer to region of segment on-disk */
1414                 nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1415                 sci->sc_segbuf_nblocks += segbuf->sb_rest_blocks;
1416
1417                 /* allocate the next next full segment */
1418                 err = nilfs_sufile_alloc(sufile, &nextnextnum);
1419                 if (unlikely(err))
1420                         goto failed_segbuf;
1421
1422                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq + 1;
1423                 nilfs_segbuf_set_next_segnum(segbuf, nextnextnum, nilfs);
1424
1425                 list_add_tail(&segbuf->sb_list, &list);
1426                 prev = segbuf;
1427         }
1428         list_splice_tail(&list, &sci->sc_segbufs);
1429         return 0;
1430
1431  failed_segbuf:
1432         nilfs_segbuf_free(segbuf);
1433  failed:
1434         list_for_each_entry(segbuf, &list, sb_list) {
1435                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1436                 WARN_ON(ret); /* never fails */
1437         }
1438         nilfs_destroy_logs(&list);
1439         return err;
1440 }
1441
1442 static void nilfs_free_incomplete_logs(struct list_head *logs,
1443                                        struct the_nilfs *nilfs)
1444 {
1445         struct nilfs_segment_buffer *segbuf, *prev;
1446         struct inode *sufile = nilfs->ns_sufile;
1447         int ret;
1448
1449         segbuf = NILFS_FIRST_SEGBUF(logs);
1450         if (nilfs->ns_nextnum != segbuf->sb_nextnum) {
1451                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1452                 WARN_ON(ret); /* never fails */
1453         }
1454         if (atomic_read(&segbuf->sb_err)) {
1455                 /* Case 1: The first segment failed */
1456                 if (segbuf->sb_pseg_start != segbuf->sb_fseg_start)
1457                         /*
1458                          * Case 1a:  Partial segment appended into an existing
1459                          * segment
1460                          */
1461                         nilfs_terminate_segment(nilfs, segbuf->sb_fseg_start,
1462                                                 segbuf->sb_fseg_end);
1463                 else /* Case 1b:  New full segment */
1464                         set_nilfs_discontinued(nilfs);
1465         }
1466
1467         prev = segbuf;
1468         list_for_each_entry_continue(segbuf, logs, sb_list) {
1469                 if (prev->sb_nextnum != segbuf->sb_nextnum) {
1470                         ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1471                         WARN_ON(ret); /* never fails */
1472                 }
1473                 if (atomic_read(&segbuf->sb_err) &&
1474                     segbuf->sb_segnum != nilfs->ns_nextnum)
1475                         /* Case 2: extended segment (!= next) failed */
1476                         nilfs_sufile_set_error(sufile, segbuf->sb_segnum);
1477                 prev = segbuf;
1478         }
1479 }
1480
1481 static void nilfs_segctor_update_segusage(struct nilfs_sc_info *sci,
1482                                           struct inode *sufile)
1483 {
1484         struct nilfs_segment_buffer *segbuf;
1485         unsigned long live_blocks;
1486         int ret;
1487
1488         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1489                 live_blocks = segbuf->sb_sum.nblocks +
1490                         (segbuf->sb_pseg_start - segbuf->sb_fseg_start);
1491                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1492                                                      live_blocks,
1493                                                      sci->sc_seg_ctime);
1494                 WARN_ON(ret); /* always succeed because the segusage is dirty */
1495         }
1496 }
1497
1498 static void nilfs_cancel_segusage(struct list_head *logs, struct inode *sufile)
1499 {
1500         struct nilfs_segment_buffer *segbuf;
1501         int ret;
1502
1503         segbuf = NILFS_FIRST_SEGBUF(logs);
1504         ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1505                                              segbuf->sb_pseg_start -
1506                                              segbuf->sb_fseg_start, 0);
1507         WARN_ON(ret); /* always succeed because the segusage is dirty */
1508
1509         list_for_each_entry_continue(segbuf, logs, sb_list) {
1510                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1511                                                      0, 0);
1512                 WARN_ON(ret); /* always succeed */
1513         }
1514 }
1515
1516 static void nilfs_segctor_truncate_segments(struct nilfs_sc_info *sci,
1517                                             struct nilfs_segment_buffer *last,
1518                                             struct inode *sufile)
1519 {
1520         struct nilfs_segment_buffer *segbuf = last;
1521         int ret;
1522
1523         list_for_each_entry_continue(segbuf, &sci->sc_segbufs, sb_list) {
1524                 sci->sc_segbuf_nblocks -= segbuf->sb_rest_blocks;
1525                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1526                 WARN_ON(ret);
1527         }
1528         nilfs_truncate_logs(&sci->sc_segbufs, last);
1529 }
1530
1531
1532 static int nilfs_segctor_collect(struct nilfs_sc_info *sci,
1533                                  struct the_nilfs *nilfs, int mode)
1534 {
1535         struct nilfs_cstage prev_stage = sci->sc_stage;
1536         int err, nadd = 1;
1537
1538         /* Collection retry loop */
1539         for (;;) {
1540                 sci->sc_nblk_this_inc = 0;
1541                 sci->sc_curseg = NILFS_FIRST_SEGBUF(&sci->sc_segbufs);
1542
1543                 err = nilfs_segctor_reset_segment_buffer(sci);
1544                 if (unlikely(err))
1545                         goto failed;
1546
1547                 err = nilfs_segctor_collect_blocks(sci, mode);
1548                 sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
1549                 if (!err)
1550                         break;
1551
1552                 if (unlikely(err != -E2BIG))
1553                         goto failed;
1554
1555                 /* The current segment is filled up */
1556                 if (mode != SC_LSEG_SR ||
1557                     nilfs_sc_cstage_get(sci) < NILFS_ST_CPFILE)
1558                         break;
1559
1560                 nilfs_clear_logs(&sci->sc_segbufs);
1561
1562                 if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1563                         err = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1564                                                         sci->sc_freesegs,
1565                                                         sci->sc_nfreesegs,
1566                                                         NULL);
1567                         WARN_ON(err); /* do not happen */
1568                         sci->sc_stage.flags &= ~NILFS_CF_SUFREED;
1569                 }
1570
1571                 err = nilfs_segctor_extend_segments(sci, nilfs, nadd);
1572                 if (unlikely(err))
1573                         return err;
1574
1575                 nadd = min_t(int, nadd << 1, SC_MAX_SEGDELTA);
1576                 sci->sc_stage = prev_stage;
1577         }
1578         nilfs_segctor_zeropad_segsum(sci);
1579         nilfs_segctor_truncate_segments(sci, sci->sc_curseg, nilfs->ns_sufile);
1580         return 0;
1581
1582  failed:
1583         return err;
1584 }
1585
1586 static void nilfs_list_replace_buffer(struct buffer_head *old_bh,
1587                                       struct buffer_head *new_bh)
1588 {
1589         BUG_ON(!list_empty(&new_bh->b_assoc_buffers));
1590
1591         list_replace_init(&old_bh->b_assoc_buffers, &new_bh->b_assoc_buffers);
1592         /* The caller must release old_bh */
1593 }
1594
1595 static int
1596 nilfs_segctor_update_payload_blocknr(struct nilfs_sc_info *sci,
1597                                      struct nilfs_segment_buffer *segbuf,
1598                                      int mode)
1599 {
1600         struct inode *inode = NULL;
1601         sector_t blocknr;
1602         unsigned long nfinfo = segbuf->sb_sum.nfinfo;
1603         unsigned long nblocks = 0, ndatablk = 0;
1604         const struct nilfs_sc_operations *sc_op = NULL;
1605         struct nilfs_segsum_pointer ssp;
1606         struct nilfs_finfo *finfo = NULL;
1607         union nilfs_binfo binfo;
1608         struct buffer_head *bh, *bh_org;
1609         ino_t ino = 0;
1610         int err = 0;
1611
1612         if (!nfinfo)
1613                 goto out;
1614
1615         blocknr = segbuf->sb_pseg_start + segbuf->sb_sum.nsumblk;
1616         ssp.bh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
1617         ssp.offset = sizeof(struct nilfs_segment_summary);
1618
1619         list_for_each_entry(bh, &segbuf->sb_payload_buffers, b_assoc_buffers) {
1620                 if (bh == segbuf->sb_super_root)
1621                         break;
1622                 if (!finfo) {
1623                         finfo = nilfs_segctor_map_segsum_entry(
1624                                 sci, &ssp, sizeof(*finfo));
1625                         ino = le64_to_cpu(finfo->fi_ino);
1626                         nblocks = le32_to_cpu(finfo->fi_nblocks);
1627                         ndatablk = le32_to_cpu(finfo->fi_ndatablk);
1628
1629                         inode = bh->b_folio->mapping->host;
1630
1631                         if (mode == SC_LSEG_DSYNC)
1632                                 sc_op = &nilfs_sc_dsync_ops;
1633                         else if (ino == NILFS_DAT_INO)
1634                                 sc_op = &nilfs_sc_dat_ops;
1635                         else /* file blocks */
1636                                 sc_op = &nilfs_sc_file_ops;
1637                 }
1638                 bh_org = bh;
1639                 get_bh(bh_org);
1640                 err = nilfs_bmap_assign(NILFS_I(inode)->i_bmap, &bh, blocknr,
1641                                         &binfo);
1642                 if (bh != bh_org)
1643                         nilfs_list_replace_buffer(bh_org, bh);
1644                 brelse(bh_org);
1645                 if (unlikely(err))
1646                         goto failed_bmap;
1647
1648                 if (ndatablk > 0)
1649                         sc_op->write_data_binfo(sci, &ssp, &binfo);
1650                 else
1651                         sc_op->write_node_binfo(sci, &ssp, &binfo);
1652
1653                 blocknr++;
1654                 if (--nblocks == 0) {
1655                         finfo = NULL;
1656                         if (--nfinfo == 0)
1657                                 break;
1658                 } else if (ndatablk > 0)
1659                         ndatablk--;
1660         }
1661  out:
1662         return 0;
1663
1664  failed_bmap:
1665         return err;
1666 }
1667
1668 static int nilfs_segctor_assign(struct nilfs_sc_info *sci, int mode)
1669 {
1670         struct nilfs_segment_buffer *segbuf;
1671         int err;
1672
1673         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1674                 err = nilfs_segctor_update_payload_blocknr(sci, segbuf, mode);
1675                 if (unlikely(err))
1676                         return err;
1677                 nilfs_segbuf_fill_in_segsum(segbuf);
1678         }
1679         return 0;
1680 }
1681
1682 static void nilfs_begin_folio_io(struct folio *folio)
1683 {
1684         if (!folio || folio_test_writeback(folio))
1685                 /*
1686                  * For split b-tree node pages, this function may be called
1687                  * twice.  We ignore the 2nd or later calls by this check.
1688                  */
1689                 return;
1690
1691         folio_lock(folio);
1692         folio_clear_dirty_for_io(folio);
1693         folio_start_writeback(folio);
1694         folio_unlock(folio);
1695 }
1696
1697 /**
1698  * nilfs_prepare_write_logs - prepare to write logs
1699  * @logs: logs to prepare for writing
1700  * @seed: checksum seed value
1701  *
1702  * nilfs_prepare_write_logs() adds checksums and prepares the block
1703  * buffers/folios for writing logs.  In order to stabilize folios of
1704  * memory-mapped file blocks by putting them in writeback state before
1705  * calculating the checksums, first prepare to write payload blocks other
1706  * than segment summary and super root blocks in which the checksums will
1707  * be embedded.
1708  */
1709 static void nilfs_prepare_write_logs(struct list_head *logs, u32 seed)
1710 {
1711         struct nilfs_segment_buffer *segbuf;
1712         struct folio *bd_folio = NULL, *fs_folio = NULL;
1713         struct buffer_head *bh;
1714
1715         /* Prepare to write payload blocks */
1716         list_for_each_entry(segbuf, logs, sb_list) {
1717                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1718                                     b_assoc_buffers) {
1719                         if (bh == segbuf->sb_super_root)
1720                                 break;
1721                         set_buffer_async_write(bh);
1722                         if (bh->b_folio != fs_folio) {
1723                                 nilfs_begin_folio_io(fs_folio);
1724                                 fs_folio = bh->b_folio;
1725                         }
1726                 }
1727         }
1728         nilfs_begin_folio_io(fs_folio);
1729
1730         nilfs_add_checksums_on_logs(logs, seed);
1731
1732         /* Prepare to write segment summary blocks */
1733         list_for_each_entry(segbuf, logs, sb_list) {
1734                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1735                                     b_assoc_buffers) {
1736                         mark_buffer_dirty(bh);
1737                         if (bh->b_folio == bd_folio)
1738                                 continue;
1739                         if (bd_folio) {
1740                                 folio_lock(bd_folio);
1741                                 folio_wait_writeback(bd_folio);
1742                                 folio_clear_dirty_for_io(bd_folio);
1743                                 folio_start_writeback(bd_folio);
1744                                 folio_unlock(bd_folio);
1745                         }
1746                         bd_folio = bh->b_folio;
1747                 }
1748         }
1749
1750         /* Prepare to write super root block */
1751         bh = NILFS_LAST_SEGBUF(logs)->sb_super_root;
1752         if (bh) {
1753                 mark_buffer_dirty(bh);
1754                 if (bh->b_folio != bd_folio) {
1755                         folio_lock(bd_folio);
1756                         folio_wait_writeback(bd_folio);
1757                         folio_clear_dirty_for_io(bd_folio);
1758                         folio_start_writeback(bd_folio);
1759                         folio_unlock(bd_folio);
1760                         bd_folio = bh->b_folio;
1761                 }
1762         }
1763
1764         if (bd_folio) {
1765                 folio_lock(bd_folio);
1766                 folio_wait_writeback(bd_folio);
1767                 folio_clear_dirty_for_io(bd_folio);
1768                 folio_start_writeback(bd_folio);
1769                 folio_unlock(bd_folio);
1770         }
1771 }
1772
1773 static int nilfs_segctor_write(struct nilfs_sc_info *sci,
1774                                struct the_nilfs *nilfs)
1775 {
1776         int ret;
1777
1778         ret = nilfs_write_logs(&sci->sc_segbufs, nilfs);
1779         list_splice_tail_init(&sci->sc_segbufs, &sci->sc_write_logs);
1780         return ret;
1781 }
1782
1783 static void nilfs_end_folio_io(struct folio *folio, int err)
1784 {
1785         if (!folio)
1786                 return;
1787
1788         if (buffer_nilfs_node(folio_buffers(folio)) &&
1789                         !folio_test_writeback(folio)) {
1790                 /*
1791                  * For b-tree node pages, this function may be called twice
1792                  * or more because they might be split in a segment.
1793                  */
1794                 if (folio_test_dirty(folio)) {
1795                         /*
1796                          * For pages holding split b-tree node buffers, dirty
1797                          * flag on the buffers may be cleared discretely.
1798                          * In that case, the page is once redirtied for
1799                          * remaining buffers, and it must be cancelled if
1800                          * all the buffers get cleaned later.
1801                          */
1802                         folio_lock(folio);
1803                         if (nilfs_folio_buffers_clean(folio))
1804                                 __nilfs_clear_folio_dirty(folio);
1805                         folio_unlock(folio);
1806                 }
1807                 return;
1808         }
1809
1810         if (err || !nilfs_folio_buffers_clean(folio))
1811                 filemap_dirty_folio(folio->mapping, folio);
1812
1813         folio_end_writeback(folio);
1814 }
1815
1816 static void nilfs_abort_logs(struct list_head *logs, int err)
1817 {
1818         struct nilfs_segment_buffer *segbuf;
1819         struct folio *bd_folio = NULL, *fs_folio = NULL;
1820         struct buffer_head *bh;
1821
1822         if (list_empty(logs))
1823                 return;
1824
1825         list_for_each_entry(segbuf, logs, sb_list) {
1826                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1827                                     b_assoc_buffers) {
1828                         clear_buffer_uptodate(bh);
1829                         if (bh->b_folio != bd_folio) {
1830                                 if (bd_folio)
1831                                         folio_end_writeback(bd_folio);
1832                                 bd_folio = bh->b_folio;
1833                         }
1834                 }
1835
1836                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1837                                     b_assoc_buffers) {
1838                         if (bh == segbuf->sb_super_root) {
1839                                 clear_buffer_uptodate(bh);
1840                                 if (bh->b_folio != bd_folio) {
1841                                         folio_end_writeback(bd_folio);
1842                                         bd_folio = bh->b_folio;
1843                                 }
1844                                 break;
1845                         }
1846                         clear_buffer_async_write(bh);
1847                         if (bh->b_folio != fs_folio) {
1848                                 nilfs_end_folio_io(fs_folio, err);
1849                                 fs_folio = bh->b_folio;
1850                         }
1851                 }
1852         }
1853         if (bd_folio)
1854                 folio_end_writeback(bd_folio);
1855
1856         nilfs_end_folio_io(fs_folio, err);
1857 }
1858
1859 static void nilfs_segctor_abort_construction(struct nilfs_sc_info *sci,
1860                                              struct the_nilfs *nilfs, int err)
1861 {
1862         LIST_HEAD(logs);
1863         int ret;
1864
1865         list_splice_tail_init(&sci->sc_write_logs, &logs);
1866         ret = nilfs_wait_on_logs(&logs);
1867         nilfs_abort_logs(&logs, ret ? : err);
1868
1869         list_splice_tail_init(&sci->sc_segbufs, &logs);
1870         if (list_empty(&logs))
1871                 return; /* if the first segment buffer preparation failed */
1872
1873         nilfs_cancel_segusage(&logs, nilfs->ns_sufile);
1874         nilfs_free_incomplete_logs(&logs, nilfs);
1875
1876         if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1877                 ret = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1878                                                 sci->sc_freesegs,
1879                                                 sci->sc_nfreesegs,
1880                                                 NULL);
1881                 WARN_ON(ret); /* do not happen */
1882         }
1883
1884         nilfs_destroy_logs(&logs);
1885 }
1886
1887 static void nilfs_set_next_segment(struct the_nilfs *nilfs,
1888                                    struct nilfs_segment_buffer *segbuf)
1889 {
1890         nilfs->ns_segnum = segbuf->sb_segnum;
1891         nilfs->ns_nextnum = segbuf->sb_nextnum;
1892         nilfs->ns_pseg_offset = segbuf->sb_pseg_start - segbuf->sb_fseg_start
1893                 + segbuf->sb_sum.nblocks;
1894         nilfs->ns_seg_seq = segbuf->sb_sum.seg_seq;
1895         nilfs->ns_ctime = segbuf->sb_sum.ctime;
1896 }
1897
1898 static void nilfs_segctor_complete_write(struct nilfs_sc_info *sci)
1899 {
1900         struct nilfs_segment_buffer *segbuf;
1901         struct folio *bd_folio = NULL, *fs_folio = NULL;
1902         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1903         int update_sr = false;
1904
1905         list_for_each_entry(segbuf, &sci->sc_write_logs, sb_list) {
1906                 struct buffer_head *bh;
1907
1908                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1909                                     b_assoc_buffers) {
1910                         set_buffer_uptodate(bh);
1911                         clear_buffer_dirty(bh);
1912                         if (bh->b_folio != bd_folio) {
1913                                 if (bd_folio)
1914                                         folio_end_writeback(bd_folio);
1915                                 bd_folio = bh->b_folio;
1916                         }
1917                 }
1918                 /*
1919                  * We assume that the buffers which belong to the same folio
1920                  * continue over the buffer list.
1921                  * Under this assumption, the last BHs of folios is
1922                  * identifiable by the discontinuity of bh->b_folio
1923                  * (folio != fs_folio).
1924                  *
1925                  * For B-tree node blocks, however, this assumption is not
1926                  * guaranteed.  The cleanup code of B-tree node folios needs
1927                  * special care.
1928                  */
1929                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1930                                     b_assoc_buffers) {
1931                         const unsigned long set_bits = BIT(BH_Uptodate);
1932                         const unsigned long clear_bits =
1933                                 (BIT(BH_Dirty) | BIT(BH_Async_Write) |
1934                                  BIT(BH_Delay) | BIT(BH_NILFS_Volatile) |
1935                                  BIT(BH_NILFS_Redirected));
1936
1937                         if (bh == segbuf->sb_super_root) {
1938                                 set_buffer_uptodate(bh);
1939                                 clear_buffer_dirty(bh);
1940                                 if (bh->b_folio != bd_folio) {
1941                                         folio_end_writeback(bd_folio);
1942                                         bd_folio = bh->b_folio;
1943                                 }
1944                                 update_sr = true;
1945                                 break;
1946                         }
1947                         set_mask_bits(&bh->b_state, clear_bits, set_bits);
1948                         if (bh->b_folio != fs_folio) {
1949                                 nilfs_end_folio_io(fs_folio, 0);
1950                                 fs_folio = bh->b_folio;
1951                         }
1952                 }
1953
1954                 if (!nilfs_segbuf_simplex(segbuf)) {
1955                         if (segbuf->sb_sum.flags & NILFS_SS_LOGBGN) {
1956                                 set_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1957                                 sci->sc_lseg_stime = jiffies;
1958                         }
1959                         if (segbuf->sb_sum.flags & NILFS_SS_LOGEND)
1960                                 clear_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1961                 }
1962         }
1963         /*
1964          * Since folios may continue over multiple segment buffers,
1965          * end of the last folio must be checked outside of the loop.
1966          */
1967         if (bd_folio)
1968                 folio_end_writeback(bd_folio);
1969
1970         nilfs_end_folio_io(fs_folio, 0);
1971
1972         nilfs_drop_collected_inodes(&sci->sc_dirty_files);
1973
1974         if (nilfs_doing_gc())
1975                 nilfs_drop_collected_inodes(&sci->sc_gc_inodes);
1976         else
1977                 nilfs->ns_nongc_ctime = sci->sc_seg_ctime;
1978
1979         sci->sc_nblk_inc += sci->sc_nblk_this_inc;
1980
1981         segbuf = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1982         nilfs_set_next_segment(nilfs, segbuf);
1983
1984         if (update_sr) {
1985                 nilfs->ns_flushed_device = 0;
1986                 nilfs_set_last_segment(nilfs, segbuf->sb_pseg_start,
1987                                        segbuf->sb_sum.seg_seq, nilfs->ns_cno++);
1988
1989                 clear_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
1990                 clear_bit(NILFS_SC_DIRTY, &sci->sc_flags);
1991                 set_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1992                 nilfs_segctor_clear_metadata_dirty(sci);
1993         } else
1994                 clear_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1995 }
1996
1997 static int nilfs_segctor_wait(struct nilfs_sc_info *sci)
1998 {
1999         int ret;
2000
2001         ret = nilfs_wait_on_logs(&sci->sc_write_logs);
2002         if (!ret) {
2003                 nilfs_segctor_complete_write(sci);
2004                 nilfs_destroy_logs(&sci->sc_write_logs);
2005         }
2006         return ret;
2007 }
2008
2009 static int nilfs_segctor_collect_dirty_files(struct nilfs_sc_info *sci,
2010                                              struct the_nilfs *nilfs)
2011 {
2012         struct nilfs_inode_info *ii, *n;
2013         struct inode *ifile = sci->sc_root->ifile;
2014
2015         spin_lock(&nilfs->ns_inode_lock);
2016  retry:
2017         list_for_each_entry_safe(ii, n, &nilfs->ns_dirty_files, i_dirty) {
2018                 if (!ii->i_bh) {
2019                         struct buffer_head *ibh;
2020                         int err;
2021
2022                         spin_unlock(&nilfs->ns_inode_lock);
2023                         err = nilfs_ifile_get_inode_block(
2024                                 ifile, ii->vfs_inode.i_ino, &ibh);
2025                         if (unlikely(err)) {
2026                                 nilfs_warn(sci->sc_super,
2027                                            "log writer: error %d getting inode block (ino=%lu)",
2028                                            err, ii->vfs_inode.i_ino);
2029                                 return err;
2030                         }
2031                         spin_lock(&nilfs->ns_inode_lock);
2032                         if (likely(!ii->i_bh))
2033                                 ii->i_bh = ibh;
2034                         else
2035                                 brelse(ibh);
2036                         goto retry;
2037                 }
2038
2039                 // Always redirty the buffer to avoid race condition
2040                 mark_buffer_dirty(ii->i_bh);
2041                 nilfs_mdt_mark_dirty(ifile);
2042
2043                 clear_bit(NILFS_I_QUEUED, &ii->i_state);
2044                 set_bit(NILFS_I_BUSY, &ii->i_state);
2045                 list_move_tail(&ii->i_dirty, &sci->sc_dirty_files);
2046         }
2047         spin_unlock(&nilfs->ns_inode_lock);
2048
2049         return 0;
2050 }
2051
2052 static void nilfs_segctor_drop_written_files(struct nilfs_sc_info *sci,
2053                                              struct the_nilfs *nilfs)
2054 {
2055         struct nilfs_inode_info *ii, *n;
2056         int during_mount = !(sci->sc_super->s_flags & SB_ACTIVE);
2057         int defer_iput = false;
2058
2059         spin_lock(&nilfs->ns_inode_lock);
2060         list_for_each_entry_safe(ii, n, &sci->sc_dirty_files, i_dirty) {
2061                 if (!test_and_clear_bit(NILFS_I_UPDATED, &ii->i_state) ||
2062                     test_bit(NILFS_I_DIRTY, &ii->i_state))
2063                         continue;
2064
2065                 clear_bit(NILFS_I_BUSY, &ii->i_state);
2066                 brelse(ii->i_bh);
2067                 ii->i_bh = NULL;
2068                 list_del_init(&ii->i_dirty);
2069                 if (!ii->vfs_inode.i_nlink || during_mount) {
2070                         /*
2071                          * Defer calling iput() to avoid deadlocks if
2072                          * i_nlink == 0 or mount is not yet finished.
2073                          */
2074                         list_add_tail(&ii->i_dirty, &sci->sc_iput_queue);
2075                         defer_iput = true;
2076                 } else {
2077                         spin_unlock(&nilfs->ns_inode_lock);
2078                         iput(&ii->vfs_inode);
2079                         spin_lock(&nilfs->ns_inode_lock);
2080                 }
2081         }
2082         spin_unlock(&nilfs->ns_inode_lock);
2083
2084         if (defer_iput)
2085                 schedule_work(&sci->sc_iput_work);
2086 }
2087
2088 /*
2089  * Main procedure of segment constructor
2090  */
2091 static int nilfs_segctor_do_construct(struct nilfs_sc_info *sci, int mode)
2092 {
2093         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2094         int err;
2095
2096         if (sb_rdonly(sci->sc_super))
2097                 return -EROFS;
2098
2099         nilfs_sc_cstage_set(sci, NILFS_ST_INIT);
2100         sci->sc_cno = nilfs->ns_cno;
2101
2102         err = nilfs_segctor_collect_dirty_files(sci, nilfs);
2103         if (unlikely(err))
2104                 goto out;
2105
2106         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
2107                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
2108
2109         if (nilfs_segctor_clean(sci))
2110                 goto out;
2111
2112         do {
2113                 sci->sc_stage.flags &= ~NILFS_CF_HISTORY_MASK;
2114
2115                 err = nilfs_segctor_begin_construction(sci, nilfs);
2116                 if (unlikely(err))
2117                         goto failed;
2118
2119                 /* Update time stamp */
2120                 sci->sc_seg_ctime = ktime_get_real_seconds();
2121
2122                 err = nilfs_segctor_collect(sci, nilfs, mode);
2123                 if (unlikely(err))
2124                         goto failed;
2125
2126                 /* Avoid empty segment */
2127                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE &&
2128                     nilfs_segbuf_empty(sci->sc_curseg)) {
2129                         nilfs_segctor_abort_construction(sci, nilfs, 1);
2130                         goto out;
2131                 }
2132
2133                 err = nilfs_segctor_assign(sci, mode);
2134                 if (unlikely(err))
2135                         goto failed;
2136
2137                 if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2138                         nilfs_segctor_fill_in_file_bmap(sci);
2139
2140                 if (mode == SC_LSEG_SR &&
2141                     nilfs_sc_cstage_get(sci) >= NILFS_ST_CPFILE) {
2142                         err = nilfs_cpfile_finalize_checkpoint(
2143                                 nilfs->ns_cpfile, nilfs->ns_cno, sci->sc_root,
2144                                 sci->sc_nblk_inc + sci->sc_nblk_this_inc,
2145                                 sci->sc_seg_ctime,
2146                                 !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags));
2147                         if (unlikely(err))
2148                                 goto failed_to_write;
2149
2150                         nilfs_segctor_fill_in_super_root(sci, nilfs);
2151                 }
2152                 nilfs_segctor_update_segusage(sci, nilfs->ns_sufile);
2153
2154                 /* Write partial segments */
2155                 nilfs_prepare_write_logs(&sci->sc_segbufs, nilfs->ns_crc_seed);
2156
2157                 err = nilfs_segctor_write(sci, nilfs);
2158                 if (unlikely(err))
2159                         goto failed_to_write;
2160
2161                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE ||
2162                     nilfs->ns_blocksize_bits != PAGE_SHIFT) {
2163                         /*
2164                          * At this point, we avoid double buffering
2165                          * for blocksize < pagesize because page dirty
2166                          * flag is turned off during write and dirty
2167                          * buffers are not properly collected for
2168                          * pages crossing over segments.
2169                          */
2170                         err = nilfs_segctor_wait(sci);
2171                         if (err)
2172                                 goto failed_to_write;
2173                 }
2174         } while (nilfs_sc_cstage_get(sci) != NILFS_ST_DONE);
2175
2176  out:
2177         nilfs_segctor_drop_written_files(sci, nilfs);
2178         return err;
2179
2180  failed_to_write:
2181  failed:
2182         if (mode == SC_LSEG_SR && nilfs_sc_cstage_get(sci) >= NILFS_ST_IFILE)
2183                 nilfs_redirty_inodes(&sci->sc_dirty_files);
2184         if (nilfs_doing_gc())
2185                 nilfs_redirty_inodes(&sci->sc_gc_inodes);
2186         nilfs_segctor_abort_construction(sci, nilfs, err);
2187         goto out;
2188 }
2189
2190 /**
2191  * nilfs_segctor_start_timer - set timer of background write
2192  * @sci: nilfs_sc_info
2193  *
2194  * If the timer has already been set, it ignores the new request.
2195  * This function MUST be called within a section locking the segment
2196  * semaphore.
2197  */
2198 static void nilfs_segctor_start_timer(struct nilfs_sc_info *sci)
2199 {
2200         spin_lock(&sci->sc_state_lock);
2201         if (!(sci->sc_state & NILFS_SEGCTOR_COMMIT)) {
2202                 if (sci->sc_task) {
2203                         sci->sc_timer.expires = jiffies + sci->sc_interval;
2204                         add_timer(&sci->sc_timer);
2205                 }
2206                 sci->sc_state |= NILFS_SEGCTOR_COMMIT;
2207         }
2208         spin_unlock(&sci->sc_state_lock);
2209 }
2210
2211 static void nilfs_segctor_do_flush(struct nilfs_sc_info *sci, int bn)
2212 {
2213         spin_lock(&sci->sc_state_lock);
2214         if (!(sci->sc_flush_request & BIT(bn))) {
2215                 unsigned long prev_req = sci->sc_flush_request;
2216
2217                 sci->sc_flush_request |= BIT(bn);
2218                 if (!prev_req)
2219                         wake_up(&sci->sc_wait_daemon);
2220         }
2221         spin_unlock(&sci->sc_state_lock);
2222 }
2223
2224 struct nilfs_segctor_wait_request {
2225         wait_queue_entry_t      wq;
2226         __u32           seq;
2227         int             err;
2228         atomic_t        done;
2229 };
2230
2231 static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
2232 {
2233         struct nilfs_segctor_wait_request wait_req;
2234         int err = 0;
2235
2236         init_wait(&wait_req.wq);
2237         wait_req.err = 0;
2238         atomic_set(&wait_req.done, 0);
2239         init_waitqueue_entry(&wait_req.wq, current);
2240
2241         /*
2242          * To prevent a race issue where completion notifications from the
2243          * log writer thread are missed, increment the request sequence count
2244          * "sc_seq_request" and insert a wait queue entry using the current
2245          * sequence number into the "sc_wait_request" queue at the same time
2246          * within the lock section of "sc_state_lock".
2247          */
2248         spin_lock(&sci->sc_state_lock);
2249         wait_req.seq = ++sci->sc_seq_request;
2250         add_wait_queue(&sci->sc_wait_request, &wait_req.wq);
2251         spin_unlock(&sci->sc_state_lock);
2252
2253         wake_up(&sci->sc_wait_daemon);
2254
2255         for (;;) {
2256                 set_current_state(TASK_INTERRUPTIBLE);
2257
2258                 /*
2259                  * Synchronize only while the log writer thread is alive.
2260                  * Leave flushing out after the log writer thread exits to
2261                  * the cleanup work in nilfs_segctor_destroy().
2262                  */
2263                 if (!sci->sc_task)
2264                         break;
2265
2266                 if (atomic_read(&wait_req.done)) {
2267                         err = wait_req.err;
2268                         break;
2269                 }
2270                 if (!signal_pending(current)) {
2271                         schedule();
2272                         continue;
2273                 }
2274                 err = -ERESTARTSYS;
2275                 break;
2276         }
2277         finish_wait(&sci->sc_wait_request, &wait_req.wq);
2278         return err;
2279 }
2280
2281 static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err, bool force)
2282 {
2283         struct nilfs_segctor_wait_request *wrq, *n;
2284         unsigned long flags;
2285
2286         spin_lock_irqsave(&sci->sc_wait_request.lock, flags);
2287         list_for_each_entry_safe(wrq, n, &sci->sc_wait_request.head, wq.entry) {
2288                 if (!atomic_read(&wrq->done) &&
2289                     (force || nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq))) {
2290                         wrq->err = err;
2291                         atomic_set(&wrq->done, 1);
2292                 }
2293                 if (atomic_read(&wrq->done)) {
2294                         wrq->wq.func(&wrq->wq,
2295                                      TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2296                                      0, NULL);
2297                 }
2298         }
2299         spin_unlock_irqrestore(&sci->sc_wait_request.lock, flags);
2300 }
2301
2302 /**
2303  * nilfs_construct_segment - construct a logical segment
2304  * @sb: super block
2305  *
2306  * Return: 0 on success, or one of the following negative error codes on
2307  * failure:
2308  * * %-EIO              - I/O error (including metadata corruption).
2309  * * %-ENOMEM           - Insufficient memory available.
2310  * * %-ENOSPC           - No space left on device (only in a panic state).
2311  * * %-ERESTARTSYS      - Interrupted.
2312  * * %-EROFS            - Read only filesystem.
2313  */
2314 int nilfs_construct_segment(struct super_block *sb)
2315 {
2316         struct the_nilfs *nilfs = sb->s_fs_info;
2317         struct nilfs_sc_info *sci = nilfs->ns_writer;
2318         struct nilfs_transaction_info *ti;
2319
2320         if (sb_rdonly(sb) || unlikely(!sci))
2321                 return -EROFS;
2322
2323         /* A call inside transactions causes a deadlock. */
2324         BUG_ON((ti = current->journal_info) && ti->ti_magic == NILFS_TI_MAGIC);
2325
2326         return nilfs_segctor_sync(sci);
2327 }
2328
2329 /**
2330  * nilfs_construct_dsync_segment - construct a data-only logical segment
2331  * @sb: super block
2332  * @inode: inode whose data blocks should be written out
2333  * @start: start byte offset
2334  * @end: end byte offset (inclusive)
2335  *
2336  * Return: 0 on success, or one of the following negative error codes on
2337  * failure:
2338  * * %-EIO              - I/O error (including metadata corruption).
2339  * * %-ENOMEM           - Insufficient memory available.
2340  * * %-ENOSPC           - No space left on device (only in a panic state).
2341  * * %-ERESTARTSYS      - Interrupted.
2342  * * %-EROFS            - Read only filesystem.
2343  */
2344 int nilfs_construct_dsync_segment(struct super_block *sb, struct inode *inode,
2345                                   loff_t start, loff_t end)
2346 {
2347         struct the_nilfs *nilfs = sb->s_fs_info;
2348         struct nilfs_sc_info *sci = nilfs->ns_writer;
2349         struct nilfs_inode_info *ii;
2350         struct nilfs_transaction_info ti;
2351         int err = 0;
2352
2353         if (sb_rdonly(sb) || unlikely(!sci))
2354                 return -EROFS;
2355
2356         nilfs_transaction_lock(sb, &ti, 0);
2357
2358         ii = NILFS_I(inode);
2359         if (test_bit(NILFS_I_INODE_SYNC, &ii->i_state) ||
2360             nilfs_test_opt(nilfs, STRICT_ORDER) ||
2361             test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2362             nilfs_discontinued(nilfs)) {
2363                 nilfs_transaction_unlock(sb);
2364                 err = nilfs_segctor_sync(sci);
2365                 return err;
2366         }
2367
2368         spin_lock(&nilfs->ns_inode_lock);
2369         if (!test_bit(NILFS_I_QUEUED, &ii->i_state) &&
2370             !test_bit(NILFS_I_BUSY, &ii->i_state)) {
2371                 spin_unlock(&nilfs->ns_inode_lock);
2372                 nilfs_transaction_unlock(sb);
2373                 return 0;
2374         }
2375         spin_unlock(&nilfs->ns_inode_lock);
2376         sci->sc_dsync_inode = ii;
2377         sci->sc_dsync_start = start;
2378         sci->sc_dsync_end = end;
2379
2380         err = nilfs_segctor_do_construct(sci, SC_LSEG_DSYNC);
2381         if (!err)
2382                 nilfs->ns_flushed_device = 0;
2383
2384         nilfs_transaction_unlock(sb);
2385         return err;
2386 }
2387
2388 #define FLUSH_FILE_BIT  (0x1) /* data file only */
2389 #define FLUSH_DAT_BIT   BIT(NILFS_DAT_INO) /* DAT only */
2390
2391 /**
2392  * nilfs_segctor_accept - record accepted sequence count of log-write requests
2393  * @sci: segment constructor object
2394  */
2395 static void nilfs_segctor_accept(struct nilfs_sc_info *sci)
2396 {
2397         bool thread_is_alive;
2398
2399         spin_lock(&sci->sc_state_lock);
2400         sci->sc_seq_accepted = sci->sc_seq_request;
2401         thread_is_alive = (bool)sci->sc_task;
2402         spin_unlock(&sci->sc_state_lock);
2403
2404         /*
2405          * This function does not race with the log writer thread's
2406          * termination.  Therefore, deleting sc_timer, which should not be
2407          * done after the log writer thread exits, can be done safely outside
2408          * the area protected by sc_state_lock.
2409          */
2410         if (thread_is_alive)
2411                 timer_delete_sync(&sci->sc_timer);
2412 }
2413
2414 /**
2415  * nilfs_segctor_notify - notify the result of request to caller threads
2416  * @sci: segment constructor object
2417  * @mode: mode of log forming
2418  * @err: error code to be notified
2419  */
2420 static void nilfs_segctor_notify(struct nilfs_sc_info *sci, int mode, int err)
2421 {
2422         /* Clear requests (even when the construction failed) */
2423         spin_lock(&sci->sc_state_lock);
2424
2425         if (mode == SC_LSEG_SR) {
2426                 sci->sc_state &= ~NILFS_SEGCTOR_COMMIT;
2427                 sci->sc_seq_done = sci->sc_seq_accepted;
2428                 nilfs_segctor_wakeup(sci, err, false);
2429                 sci->sc_flush_request = 0;
2430         } else {
2431                 if (mode == SC_FLUSH_FILE)
2432                         sci->sc_flush_request &= ~FLUSH_FILE_BIT;
2433                 else if (mode == SC_FLUSH_DAT)
2434                         sci->sc_flush_request &= ~FLUSH_DAT_BIT;
2435
2436                 /* re-enable timer if checkpoint creation was not done */
2437                 if ((sci->sc_state & NILFS_SEGCTOR_COMMIT) && sci->sc_task &&
2438                     time_before(jiffies, sci->sc_timer.expires))
2439                         add_timer(&sci->sc_timer);
2440         }
2441         spin_unlock(&sci->sc_state_lock);
2442 }
2443
2444 /**
2445  * nilfs_segctor_construct - form logs and write them to disk
2446  * @sci: segment constructor object
2447  * @mode: mode of log forming
2448  *
2449  * Return: 0 on success, or a negative error code on failure.
2450  */
2451 static int nilfs_segctor_construct(struct nilfs_sc_info *sci, int mode)
2452 {
2453         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2454         struct nilfs_super_block **sbp;
2455         int err = 0;
2456
2457         nilfs_segctor_accept(sci);
2458
2459         if (nilfs_discontinued(nilfs))
2460                 mode = SC_LSEG_SR;
2461         if (!nilfs_segctor_confirm(sci))
2462                 err = nilfs_segctor_do_construct(sci, mode);
2463
2464         if (likely(!err)) {
2465                 if (mode != SC_FLUSH_DAT)
2466                         atomic_set(&nilfs->ns_ndirtyblks, 0);
2467                 if (test_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags) &&
2468                     nilfs_discontinued(nilfs)) {
2469                         down_write(&nilfs->ns_sem);
2470                         err = -EIO;
2471                         sbp = nilfs_prepare_super(sci->sc_super,
2472                                                   nilfs_sb_will_flip(nilfs));
2473                         if (likely(sbp)) {
2474                                 nilfs_set_log_cursor(sbp[0], nilfs);
2475                                 err = nilfs_commit_super(sci->sc_super,
2476                                                          NILFS_SB_COMMIT);
2477                         }
2478                         up_write(&nilfs->ns_sem);
2479                 }
2480         }
2481
2482         nilfs_segctor_notify(sci, mode, err);
2483         return err;
2484 }
2485
2486 static void nilfs_construction_timeout(struct timer_list *t)
2487 {
2488         struct nilfs_sc_info *sci = timer_container_of(sci, t, sc_timer);
2489
2490         wake_up_process(sci->sc_task);
2491 }
2492
2493 static void
2494 nilfs_remove_written_gcinodes(struct the_nilfs *nilfs, struct list_head *head)
2495 {
2496         struct nilfs_inode_info *ii, *n;
2497
2498         list_for_each_entry_safe(ii, n, head, i_dirty) {
2499                 if (!test_bit(NILFS_I_UPDATED, &ii->i_state))
2500                         continue;
2501                 list_del_init(&ii->i_dirty);
2502                 truncate_inode_pages(&ii->vfs_inode.i_data, 0);
2503                 nilfs_btnode_cache_clear(ii->i_assoc_inode->i_mapping);
2504                 iput(&ii->vfs_inode);
2505         }
2506 }
2507
2508 int nilfs_clean_segments(struct super_block *sb, struct nilfs_argv *argv,
2509                          void **kbufs)
2510 {
2511         struct the_nilfs *nilfs = sb->s_fs_info;
2512         struct nilfs_sc_info *sci = nilfs->ns_writer;
2513         struct nilfs_transaction_info ti;
2514         int err;
2515
2516         if (unlikely(!sci))
2517                 return -EROFS;
2518
2519         nilfs_transaction_lock(sb, &ti, 1);
2520
2521         err = nilfs_mdt_save_to_shadow_map(nilfs->ns_dat);
2522         if (unlikely(err))
2523                 goto out_unlock;
2524
2525         err = nilfs_ioctl_prepare_clean_segments(nilfs, argv, kbufs);
2526         if (unlikely(err)) {
2527                 nilfs_mdt_restore_from_shadow_map(nilfs->ns_dat);
2528                 goto out_unlock;
2529         }
2530
2531         sci->sc_freesegs = kbufs[4];
2532         sci->sc_nfreesegs = argv[4].v_nmembs;
2533         list_splice_tail_init(&nilfs->ns_gc_inodes, &sci->sc_gc_inodes);
2534
2535         for (;;) {
2536                 err = nilfs_segctor_construct(sci, SC_LSEG_SR);
2537                 nilfs_remove_written_gcinodes(nilfs, &sci->sc_gc_inodes);
2538
2539                 if (likely(!err))
2540                         break;
2541
2542                 nilfs_warn(sb, "error %d cleaning segments", err);
2543                 set_current_state(TASK_INTERRUPTIBLE);
2544                 schedule_timeout(sci->sc_interval);
2545         }
2546         if (nilfs_test_opt(nilfs, DISCARD)) {
2547                 int ret = nilfs_discard_segments(nilfs, sci->sc_freesegs,
2548                                                  sci->sc_nfreesegs);
2549                 if (ret) {
2550                         nilfs_warn(sb,
2551                                    "error %d on discard request, turning discards off for the device",
2552                                    ret);
2553                         nilfs_clear_opt(nilfs, DISCARD);
2554                 }
2555         }
2556
2557  out_unlock:
2558         sci->sc_freesegs = NULL;
2559         sci->sc_nfreesegs = 0;
2560         nilfs_mdt_clear_shadow_map(nilfs->ns_dat);
2561         nilfs_transaction_unlock(sb);
2562         return err;
2563 }
2564
2565 static void nilfs_segctor_thread_construct(struct nilfs_sc_info *sci, int mode)
2566 {
2567         struct nilfs_transaction_info ti;
2568
2569         nilfs_transaction_lock(sci->sc_super, &ti, 0);
2570         nilfs_segctor_construct(sci, mode);
2571
2572         /*
2573          * Unclosed segment should be retried.  We do this using sc_timer.
2574          * Timeout of sc_timer will invoke complete construction which leads
2575          * to close the current logical segment.
2576          */
2577         if (test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags))
2578                 nilfs_segctor_start_timer(sci);
2579
2580         nilfs_transaction_unlock(sci->sc_super);
2581 }
2582
2583 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *sci)
2584 {
2585         int mode = 0;
2586
2587         spin_lock(&sci->sc_state_lock);
2588         mode = (sci->sc_flush_request & FLUSH_DAT_BIT) ?
2589                 SC_FLUSH_DAT : SC_FLUSH_FILE;
2590         spin_unlock(&sci->sc_state_lock);
2591
2592         if (mode) {
2593                 nilfs_segctor_do_construct(sci, mode);
2594
2595                 spin_lock(&sci->sc_state_lock);
2596                 sci->sc_flush_request &= (mode == SC_FLUSH_FILE) ?
2597                         ~FLUSH_FILE_BIT : ~FLUSH_DAT_BIT;
2598                 spin_unlock(&sci->sc_state_lock);
2599         }
2600         clear_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
2601 }
2602
2603 static int nilfs_segctor_flush_mode(struct nilfs_sc_info *sci)
2604 {
2605         if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2606             time_before(jiffies, sci->sc_lseg_stime + sci->sc_mjcp_freq)) {
2607                 if (!(sci->sc_flush_request & ~FLUSH_FILE_BIT))
2608                         return SC_FLUSH_FILE;
2609                 else if (!(sci->sc_flush_request & ~FLUSH_DAT_BIT))
2610                         return SC_FLUSH_DAT;
2611         }
2612         return SC_LSEG_SR;
2613 }
2614
2615 /**
2616  * nilfs_log_write_required - determine whether log writing is required
2617  * @sci:   nilfs_sc_info struct
2618  * @modep: location for storing log writing mode
2619  *
2620  * Return: true if log writing is required, false otherwise.  If log writing
2621  * is required, the mode is stored in the location pointed to by @modep.
2622  */
2623 static bool nilfs_log_write_required(struct nilfs_sc_info *sci, int *modep)
2624 {
2625         bool timedout, ret = true;
2626
2627         spin_lock(&sci->sc_state_lock);
2628         timedout = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2629                    time_after_eq(jiffies, sci->sc_timer.expires));
2630         if (timedout || sci->sc_seq_request != sci->sc_seq_done)
2631                 *modep = SC_LSEG_SR;
2632         else if (sci->sc_flush_request)
2633                 *modep = nilfs_segctor_flush_mode(sci);
2634         else
2635                 ret = false;
2636
2637         spin_unlock(&sci->sc_state_lock);
2638         return ret;
2639 }
2640
2641 /**
2642  * nilfs_segctor_thread - main loop of the log writer thread
2643  * @arg: pointer to a struct nilfs_sc_info.
2644  *
2645  * nilfs_segctor_thread() is the main loop function of the log writer kernel
2646  * thread, which determines whether log writing is necessary, and if so,
2647  * performs the log write in the background, or waits if not.  It is also
2648  * used to decide the background writeback of the superblock.
2649  *
2650  * Return: Always 0.
2651  */
2652 static int nilfs_segctor_thread(void *arg)
2653 {
2654         struct nilfs_sc_info *sci = (struct nilfs_sc_info *)arg;
2655         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2656
2657         nilfs_info(sci->sc_super,
2658                    "segctord starting. Construction interval = %lu seconds, CP frequency < %lu seconds",
2659                    sci->sc_interval / HZ, sci->sc_mjcp_freq / HZ);
2660
2661         set_freezable();
2662
2663         while (!kthread_should_stop()) {
2664                 DEFINE_WAIT(wait);
2665                 bool should_write;
2666                 int mode;
2667
2668                 if (freezing(current)) {
2669                         try_to_freeze();
2670                         continue;
2671                 }
2672
2673                 prepare_to_wait(&sci->sc_wait_daemon, &wait,
2674                                 TASK_INTERRUPTIBLE);
2675                 should_write = nilfs_log_write_required(sci, &mode);
2676                 if (!should_write)
2677                         schedule();
2678                 finish_wait(&sci->sc_wait_daemon, &wait);
2679
2680                 if (nilfs_sb_dirty(nilfs) && nilfs_sb_need_update(nilfs))
2681                         set_nilfs_discontinued(nilfs);
2682
2683                 if (should_write)
2684                         nilfs_segctor_thread_construct(sci, mode);
2685         }
2686
2687         /* end sync. */
2688         spin_lock(&sci->sc_state_lock);
2689         sci->sc_task = NULL;
2690         timer_shutdown_sync(&sci->sc_timer);
2691         spin_unlock(&sci->sc_state_lock);
2692         return 0;
2693 }
2694
2695 /*
2696  * Setup & clean-up functions
2697  */
2698 static struct nilfs_sc_info *nilfs_segctor_new(struct super_block *sb,
2699                                                struct nilfs_root *root)
2700 {
2701         struct the_nilfs *nilfs = sb->s_fs_info;
2702         struct nilfs_sc_info *sci;
2703
2704         sci = kzalloc(sizeof(*sci), GFP_KERNEL);
2705         if (!sci)
2706                 return NULL;
2707
2708         sci->sc_super = sb;
2709
2710         nilfs_get_root(root);
2711         sci->sc_root = root;
2712
2713         init_waitqueue_head(&sci->sc_wait_request);
2714         init_waitqueue_head(&sci->sc_wait_daemon);
2715         spin_lock_init(&sci->sc_state_lock);
2716         INIT_LIST_HEAD(&sci->sc_dirty_files);
2717         INIT_LIST_HEAD(&sci->sc_segbufs);
2718         INIT_LIST_HEAD(&sci->sc_write_logs);
2719         INIT_LIST_HEAD(&sci->sc_gc_inodes);
2720         INIT_LIST_HEAD(&sci->sc_iput_queue);
2721         INIT_WORK(&sci->sc_iput_work, nilfs_iput_work_func);
2722
2723         sci->sc_interval = HZ * NILFS_SC_DEFAULT_TIMEOUT;
2724         sci->sc_mjcp_freq = HZ * NILFS_SC_DEFAULT_SR_FREQ;
2725         sci->sc_watermark = NILFS_SC_DEFAULT_WATERMARK;
2726
2727         if (nilfs->ns_interval)
2728                 sci->sc_interval = HZ * nilfs->ns_interval;
2729         if (nilfs->ns_watermark)
2730                 sci->sc_watermark = nilfs->ns_watermark;
2731         return sci;
2732 }
2733
2734 static void nilfs_segctor_write_out(struct nilfs_sc_info *sci)
2735 {
2736         int ret, retrycount = NILFS_SC_CLEANUP_RETRY;
2737
2738         /*
2739          * The segctord thread was stopped and its timer was removed.
2740          * But some tasks remain.
2741          */
2742         do {
2743                 struct nilfs_transaction_info ti;
2744
2745                 nilfs_transaction_lock(sci->sc_super, &ti, 0);
2746                 ret = nilfs_segctor_construct(sci, SC_LSEG_SR);
2747                 nilfs_transaction_unlock(sci->sc_super);
2748
2749                 flush_work(&sci->sc_iput_work);
2750
2751         } while (ret && ret != -EROFS && retrycount-- > 0);
2752 }
2753
2754 /**
2755  * nilfs_segctor_destroy - destroy the segment constructor.
2756  * @sci: nilfs_sc_info
2757  *
2758  * nilfs_segctor_destroy() kills the segctord thread and frees
2759  * the nilfs_sc_info struct.
2760  * Caller must hold the segment semaphore.
2761  */
2762 static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
2763 {
2764         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2765         int flag;
2766
2767         up_write(&nilfs->ns_segctor_sem);
2768
2769         if (sci->sc_task) {
2770                 wake_up(&sci->sc_wait_daemon);
2771                 kthread_stop(sci->sc_task);
2772         }
2773
2774         spin_lock(&sci->sc_state_lock);
2775         flag = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) || sci->sc_flush_request
2776                 || sci->sc_seq_request != sci->sc_seq_done);
2777         spin_unlock(&sci->sc_state_lock);
2778
2779         /*
2780          * Forcibly wake up tasks waiting in nilfs_segctor_sync(), which can
2781          * be called from delayed iput() via nilfs_evict_inode() and can race
2782          * with the above log writer thread termination.
2783          */
2784         nilfs_segctor_wakeup(sci, 0, true);
2785
2786         if (flush_work(&sci->sc_iput_work))
2787                 flag = true;
2788
2789         if (flag || !nilfs_segctor_confirm(sci))
2790                 nilfs_segctor_write_out(sci);
2791
2792         if (!list_empty(&sci->sc_dirty_files)) {
2793                 nilfs_warn(sci->sc_super,
2794                            "disposed unprocessed dirty file(s) when stopping log writer");
2795                 nilfs_dispose_list(nilfs, &sci->sc_dirty_files, 1);
2796         }
2797
2798         if (!list_empty(&sci->sc_iput_queue)) {
2799                 nilfs_warn(sci->sc_super,
2800                            "disposed unprocessed inode(s) in iput queue when stopping log writer");
2801                 nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 1);
2802         }
2803
2804         WARN_ON(!list_empty(&sci->sc_segbufs));
2805         WARN_ON(!list_empty(&sci->sc_write_logs));
2806
2807         nilfs_put_root(sci->sc_root);
2808
2809         down_write(&nilfs->ns_segctor_sem);
2810
2811         kfree(sci);
2812 }
2813
2814 /**
2815  * nilfs_attach_log_writer - attach log writer
2816  * @sb: super block instance
2817  * @root: root object of the current filesystem tree
2818  *
2819  * This allocates a log writer object, initializes it, and starts the
2820  * log writer.
2821  *
2822  * Return: 0 on success, or one of the following negative error codes on
2823  * failure:
2824  * * %-EINTR    - Log writer thread creation failed due to interruption.
2825  * * %-ENOMEM   - Insufficient memory available.
2826  */
2827 int nilfs_attach_log_writer(struct super_block *sb, struct nilfs_root *root)
2828 {
2829         struct the_nilfs *nilfs = sb->s_fs_info;
2830         struct nilfs_sc_info *sci;
2831         struct task_struct *t;
2832         int err;
2833
2834         if (nilfs->ns_writer) {
2835                 /*
2836                  * This happens if the filesystem is made read-only by
2837                  * __nilfs_error or nilfs_remount and then remounted
2838                  * read/write.  In these cases, reuse the existing
2839                  * writer.
2840                  */
2841                 return 0;
2842         }
2843
2844         sci = nilfs_segctor_new(sb, root);
2845         if (unlikely(!sci))
2846                 return -ENOMEM;
2847
2848         nilfs->ns_writer = sci;
2849         t = kthread_create(nilfs_segctor_thread, sci, "segctord");
2850         if (IS_ERR(t)) {
2851                 err = PTR_ERR(t);
2852                 nilfs_err(sb, "error %d creating segctord thread", err);
2853                 nilfs_detach_log_writer(sb);
2854                 return err;
2855         }
2856         sci->sc_task = t;
2857         timer_setup(&sci->sc_timer, nilfs_construction_timeout, 0);
2858
2859         wake_up_process(sci->sc_task);
2860         return 0;
2861 }
2862
2863 /**
2864  * nilfs_detach_log_writer - destroy log writer
2865  * @sb: super block instance
2866  *
2867  * This kills log writer daemon, frees the log writer object, and
2868  * destroys list of dirty files.
2869  */
2870 void nilfs_detach_log_writer(struct super_block *sb)
2871 {
2872         struct the_nilfs *nilfs = sb->s_fs_info;
2873         LIST_HEAD(garbage_list);
2874
2875         down_write(&nilfs->ns_segctor_sem);
2876         if (nilfs->ns_writer) {
2877                 nilfs_segctor_destroy(nilfs->ns_writer);
2878                 nilfs->ns_writer = NULL;
2879         }
2880         set_nilfs_purging(nilfs);
2881
2882         /* Force to free the list of dirty files */
2883         spin_lock(&nilfs->ns_inode_lock);
2884         if (!list_empty(&nilfs->ns_dirty_files)) {
2885                 list_splice_init(&nilfs->ns_dirty_files, &garbage_list);
2886                 nilfs_warn(sb,
2887                            "disposed unprocessed dirty file(s) when detaching log writer");
2888         }
2889         spin_unlock(&nilfs->ns_inode_lock);
2890         up_write(&nilfs->ns_segctor_sem);
2891
2892         nilfs_dispose_list(nilfs, &garbage_list, 1);
2893         clear_nilfs_purging(nilfs);
2894 }