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 Value: On success, 0 is returned. On error, one of the following
 195  * negative error code is returned.
 196  *
 197  * %-ENOMEM - Insufficient memory available.
 198  *
 199  * %-ENOSPC - No space left on device
 200  */
 201 int nilfs_transaction_begin(struct super_block *sb,
 202                             struct nilfs_transaction_info *ti,
 203                             int vacancy_check)
 204 {
 205         struct the_nilfs *nilfs;
 206         int ret = nilfs_prepare_segment_lock(sb, ti);
 207         struct nilfs_transaction_info *trace_ti;
 208
 209         if (unlikely(ret < 0))
 210                 return ret;
 211         if (ret > 0) {
 212                 trace_ti = current->journal_info;
 213
 214                 trace_nilfs2_transaction_transition(sb, trace_ti,
 215                                     trace_ti->ti_count, trace_ti->ti_flags,
 216                                     TRACE_NILFS2_TRANSACTION_BEGIN);
 217                 return 0;
 218         }
 219
 220         sb_start_intwrite(sb);
 221
 222         nilfs = sb->s_fs_info;
 223         down_read(&nilfs->ns_segctor_sem);
 224         if (vacancy_check && nilfs_near_disk_full(nilfs)) {
 225                 up_read(&nilfs->ns_segctor_sem);
 226                 ret = -ENOSPC;
 227                 goto failed;
 228         }
 229
 230         trace_ti = current->journal_info;
 231         trace_nilfs2_transaction_transition(sb, trace_ti, trace_ti->ti_count,
 232                                             trace_ti->ti_flags,
 233                                             TRACE_NILFS2_TRANSACTION_BEGIN);
 234         return 0;
 235
 236  failed:
 237         ti = current->journal_info;
 238         current->journal_info = ti->ti_save;
 239         if (ti->ti_flags & NILFS_TI_DYNAMIC_ALLOC)
 240                 kmem_cache_free(nilfs_transaction_cachep, ti);
 241         sb_end_intwrite(sb);
 242         return ret;
 243 }
 244
 245 /**
 246  * nilfs_transaction_commit - commit indivisible file operations.
 247  * @sb: super block
 248  *
 249  * nilfs_transaction_commit() releases the read semaphore which is
 250  * acquired by nilfs_transaction_begin(). This is only performed
 251  * in outermost call of this function.  If a commit flag is set,
 252  * nilfs_transaction_commit() sets a timer to start the segment
 253  * constructor.  If a sync flag is set, it starts construction
 254  * directly.
 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 static int nilfs_segctor_reset_segment_buffer(struct nilfs_sc_info *sci)
 412 {
 413         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 414         struct buffer_head *sumbh;
 415         unsigned int sumbytes;
 416         unsigned int flags = 0;
 417         int err;
 418
 419         if (nilfs_doing_gc())
 420                 flags = NILFS_SS_GC;
 421         err = nilfs_segbuf_reset(segbuf, flags, sci->sc_seg_ctime, sci->sc_cno);
 422         if (unlikely(err))
 423                 return err;
 424
 425         sumbh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
 426         sumbytes = segbuf->sb_sum.sumbytes;
 427         sci->sc_finfo_ptr.bh = sumbh;  sci->sc_finfo_ptr.offset = sumbytes;
 428         sci->sc_binfo_ptr.bh = sumbh;  sci->sc_binfo_ptr.offset = sumbytes;
 429         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 430         return 0;
 431 }
 432
 433 /**
 434  * nilfs_segctor_zeropad_segsum - zero pad the rest of the segment summary area
 435  * @sci: segment constructor object
 436  *
 437  * nilfs_segctor_zeropad_segsum() zero-fills unallocated space at the end of
 438  * the current segment summary block.
 439  */
 440 static void nilfs_segctor_zeropad_segsum(struct nilfs_sc_info *sci)
 441 {
 442         struct nilfs_segsum_pointer *ssp;
 443
 444         ssp = sci->sc_blk_cnt > 0 ? &sci->sc_binfo_ptr : &sci->sc_finfo_ptr;
 445         if (ssp->offset < ssp->bh->b_size)
 446                 memset(ssp->bh->b_data + ssp->offset, 0,
 447                        ssp->bh->b_size - ssp->offset);
 448 }
 449
 450 static int nilfs_segctor_feed_segment(struct nilfs_sc_info *sci)
 451 {
 452         sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
 453         if (NILFS_SEGBUF_IS_LAST(sci->sc_curseg, &sci->sc_segbufs))
 454                 return -E2BIG; /*
 455                                 * The current segment is filled up
 456                                 * (internal code)
 457                                 */
 458         nilfs_segctor_zeropad_segsum(sci);
 459         sci->sc_curseg = NILFS_NEXT_SEGBUF(sci->sc_curseg);
 460         return nilfs_segctor_reset_segment_buffer(sci);
 461 }
 462
 463 static int nilfs_segctor_add_super_root(struct nilfs_sc_info *sci)
 464 {
 465         struct nilfs_segment_buffer *segbuf = sci->sc_curseg;
 466         int err;
 467
 468         if (segbuf->sb_sum.nblocks >= segbuf->sb_rest_blocks) {
 469                 err = nilfs_segctor_feed_segment(sci);
 470                 if (err)
 471                         return err;
 472                 segbuf = sci->sc_curseg;
 473         }
 474         err = nilfs_segbuf_extend_payload(segbuf, &segbuf->sb_super_root);
 475         if (likely(!err))
 476                 segbuf->sb_sum.flags |= NILFS_SS_SR;
 477         return err;
 478 }
 479
 480 /*
 481  * Functions for making segment summary and payloads
 482  */
 483 static int nilfs_segctor_segsum_block_required(
 484         struct nilfs_sc_info *sci, const struct nilfs_segsum_pointer *ssp,
 485         unsigned int binfo_size)
 486 {
 487         unsigned int blocksize = sci->sc_super->s_blocksize;
 488         /* Size of finfo and binfo is enough small against blocksize */
 489
 490         return ssp->offset + binfo_size +
 491                 (!sci->sc_blk_cnt ? sizeof(struct nilfs_finfo) : 0) >
 492                 blocksize;
 493 }
 494
 495 static void nilfs_segctor_begin_finfo(struct nilfs_sc_info *sci,
 496                                       struct inode *inode)
 497 {
 498         sci->sc_curseg->sb_sum.nfinfo++;
 499         sci->sc_binfo_ptr = sci->sc_finfo_ptr;
 500         nilfs_segctor_map_segsum_entry(
 501                 sci, &sci->sc_binfo_ptr, sizeof(struct nilfs_finfo));
 502
 503         if (NILFS_I(inode)->i_root &&
 504             !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags))
 505                 set_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
 506         /* skip finfo */
 507 }
 508
 509 static void nilfs_segctor_end_finfo(struct nilfs_sc_info *sci,
 510                                     struct inode *inode)
 511 {
 512         struct nilfs_finfo *finfo;
 513         struct nilfs_inode_info *ii;
 514         struct nilfs_segment_buffer *segbuf;
 515         __u64 cno;
 516
 517         if (sci->sc_blk_cnt == 0)
 518                 return;
 519
 520         ii = NILFS_I(inode);
 521
 522         if (ii->i_type & NILFS_I_TYPE_GC)
 523                 cno = ii->i_cno;
 524         else if (NILFS_ROOT_METADATA_FILE(inode->i_ino))
 525                 cno = 0;
 526         else
 527                 cno = sci->sc_cno;
 528
 529         finfo = nilfs_segctor_map_segsum_entry(sci, &sci->sc_finfo_ptr,
 530                                                  sizeof(*finfo));
 531         finfo->fi_ino = cpu_to_le64(inode->i_ino);
 532         finfo->fi_nblocks = cpu_to_le32(sci->sc_blk_cnt);
 533         finfo->fi_ndatablk = cpu_to_le32(sci->sc_datablk_cnt);
 534         finfo->fi_cno = cpu_to_le64(cno);
 535
 536         segbuf = sci->sc_curseg;
 537         segbuf->sb_sum.sumbytes = sci->sc_binfo_ptr.offset +
 538                 sci->sc_super->s_blocksize * (segbuf->sb_sum.nsumblk - 1);
 539         sci->sc_finfo_ptr = sci->sc_binfo_ptr;
 540         sci->sc_blk_cnt = sci->sc_datablk_cnt = 0;
 541 }
 542
 543 static int nilfs_segctor_add_file_block(struct nilfs_sc_info *sci,
 544                                         struct buffer_head *bh,
 545                                         struct inode *inode,
 546                                         unsigned int binfo_size)
 547 {
 548         struct nilfs_segment_buffer *segbuf;
 549         int required, err = 0;
 550
 551  retry:
 552         segbuf = sci->sc_curseg;
 553         required = nilfs_segctor_segsum_block_required(
 554                 sci, &sci->sc_binfo_ptr, binfo_size);
 555         if (segbuf->sb_sum.nblocks + required + 1 > segbuf->sb_rest_blocks) {
 556                 nilfs_segctor_end_finfo(sci, inode);
 557                 err = nilfs_segctor_feed_segment(sci);
 558                 if (err)
 559                         return err;
 560                 goto retry;
 561         }
 562         if (unlikely(required)) {
 563                 nilfs_segctor_zeropad_segsum(sci);
 564                 err = nilfs_segbuf_extend_segsum(segbuf);
 565                 if (unlikely(err))
 566                         goto failed;
 567         }
 568         if (sci->sc_blk_cnt == 0)
 569                 nilfs_segctor_begin_finfo(sci, inode);
 570
 571         nilfs_segctor_map_segsum_entry(sci, &sci->sc_binfo_ptr, binfo_size);
 572         /* Substitution to vblocknr is delayed until update_blocknr() */
 573         nilfs_segbuf_add_file_buffer(segbuf, bh);
 574         sci->sc_blk_cnt++;
 575  failed:
 576         return err;
 577 }
 578
 579 /*
 580  * Callback functions that enumerate, mark, and collect dirty blocks
 581  */
 582 static int nilfs_collect_file_data(struct nilfs_sc_info *sci,
 583                                    struct buffer_head *bh, struct inode *inode)
 584 {
 585         int err;
 586
 587         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 588         if (err < 0)
 589                 return err;
 590
 591         err = nilfs_segctor_add_file_block(sci, bh, inode,
 592                                            sizeof(struct nilfs_binfo_v));
 593         if (!err)
 594                 sci->sc_datablk_cnt++;
 595         return err;
 596 }
 597
 598 static int nilfs_collect_file_node(struct nilfs_sc_info *sci,
 599                                    struct buffer_head *bh,
 600                                    struct inode *inode)
 601 {
 602         return nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 603 }
 604
 605 static int nilfs_collect_file_bmap(struct nilfs_sc_info *sci,
 606                                    struct buffer_head *bh,
 607                                    struct inode *inode)
 608 {
 609         WARN_ON(!buffer_dirty(bh));
 610         return nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 611 }
 612
 613 static void nilfs_write_file_data_binfo(struct nilfs_sc_info *sci,
 614                                         struct nilfs_segsum_pointer *ssp,
 615                                         union nilfs_binfo *binfo)
 616 {
 617         struct nilfs_binfo_v *binfo_v = nilfs_segctor_map_segsum_entry(
 618                 sci, ssp, sizeof(*binfo_v));
 619         *binfo_v = binfo->bi_v;
 620 }
 621
 622 static void nilfs_write_file_node_binfo(struct nilfs_sc_info *sci,
 623                                         struct nilfs_segsum_pointer *ssp,
 624                                         union nilfs_binfo *binfo)
 625 {
 626         __le64 *vblocknr = nilfs_segctor_map_segsum_entry(
 627                 sci, ssp, sizeof(*vblocknr));
 628         *vblocknr = binfo->bi_v.bi_vblocknr;
 629 }
 630
 631 static const struct nilfs_sc_operations nilfs_sc_file_ops = {
 632         .collect_data = nilfs_collect_file_data,
 633         .collect_node = nilfs_collect_file_node,
 634         .collect_bmap = nilfs_collect_file_bmap,
 635         .write_data_binfo = nilfs_write_file_data_binfo,
 636         .write_node_binfo = nilfs_write_file_node_binfo,
 637 };
 638
 639 static int nilfs_collect_dat_data(struct nilfs_sc_info *sci,
 640                                   struct buffer_head *bh, struct inode *inode)
 641 {
 642         int err;
 643
 644         err = nilfs_bmap_propagate(NILFS_I(inode)->i_bmap, bh);
 645         if (err < 0)
 646                 return err;
 647
 648         err = nilfs_segctor_add_file_block(sci, bh, inode, sizeof(__le64));
 649         if (!err)
 650                 sci->sc_datablk_cnt++;
 651         return err;
 652 }
 653
 654 static int nilfs_collect_dat_bmap(struct nilfs_sc_info *sci,
 655                                   struct buffer_head *bh, struct inode *inode)
 656 {
 657         WARN_ON(!buffer_dirty(bh));
 658         return nilfs_segctor_add_file_block(sci, bh, inode,
 659                                             sizeof(struct nilfs_binfo_dat));
 660 }
 661
 662 static void nilfs_write_dat_data_binfo(struct nilfs_sc_info *sci,
 663                                        struct nilfs_segsum_pointer *ssp,
 664                                        union nilfs_binfo *binfo)
 665 {
 666         __le64 *blkoff = nilfs_segctor_map_segsum_entry(sci, ssp,
 667                                                           sizeof(*blkoff));
 668         *blkoff = binfo->bi_dat.bi_blkoff;
 669 }
 670
 671 static void nilfs_write_dat_node_binfo(struct nilfs_sc_info *sci,
 672                                        struct nilfs_segsum_pointer *ssp,
 673                                        union nilfs_binfo *binfo)
 674 {
 675         struct nilfs_binfo_dat *binfo_dat =
 676                 nilfs_segctor_map_segsum_entry(sci, ssp, sizeof(*binfo_dat));
 677         *binfo_dat = binfo->bi_dat;
 678 }
 679
 680 static const struct nilfs_sc_operations nilfs_sc_dat_ops = {
 681         .collect_data = nilfs_collect_dat_data,
 682         .collect_node = nilfs_collect_file_node,
 683         .collect_bmap = nilfs_collect_dat_bmap,
 684         .write_data_binfo = nilfs_write_dat_data_binfo,
 685         .write_node_binfo = nilfs_write_dat_node_binfo,
 686 };
 687
 688 static const struct nilfs_sc_operations nilfs_sc_dsync_ops = {
 689         .collect_data = nilfs_collect_file_data,
 690         .collect_node = NULL,
 691         .collect_bmap = NULL,
 692         .write_data_binfo = nilfs_write_file_data_binfo,
 693         .write_node_binfo = NULL,
 694 };
 695
 696 static size_t nilfs_lookup_dirty_data_buffers(struct inode *inode,
 697                                               struct list_head *listp,
 698                                               size_t nlimit,
 699                                               loff_t start, loff_t end)
 700 {
 701         struct address_space *mapping = inode->i_mapping;
 702         struct folio_batch fbatch;
 703         pgoff_t index = 0, last = ULONG_MAX;
 704         size_t ndirties = 0;
 705         int i;
 706
 707         if (unlikely(start != 0 || end != LLONG_MAX)) {
 708                 /*
 709                  * A valid range is given for sync-ing data pages. The
 710                  * range is rounded to per-page; extra dirty buffers
 711                  * may be included if blocksize < pagesize.
 712                  */
 713                 index = start >> PAGE_SHIFT;
 714                 last = end >> PAGE_SHIFT;
 715         }
 716         folio_batch_init(&fbatch);
 717  repeat:
 718         if (unlikely(index > last) ||
 719               !filemap_get_folios_tag(mapping, &index, last,
 720                       PAGECACHE_TAG_DIRTY, &fbatch))
 721                 return ndirties;
 722
 723         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 724                 struct buffer_head *bh, *head;
 725                 struct folio *folio = fbatch.folios[i];
 726
 727                 folio_lock(folio);
 728                 if (unlikely(folio->mapping != mapping)) {
 729                         /* Exclude folios removed from the address space */
 730                         folio_unlock(folio);
 731                         continue;
 732                 }
 733                 head = folio_buffers(folio);
 734                 if (!head)
 735                         head = create_empty_buffers(folio,
 736                                         i_blocksize(inode), 0);
 737                 folio_unlock(folio);
 738
 739                 bh = head;
 740                 do {
 741                         if (!buffer_dirty(bh) || buffer_async_write(bh))
 742                                 continue;
 743                         get_bh(bh);
 744                         list_add_tail(&bh->b_assoc_buffers, listp);
 745                         ndirties++;
 746                         if (unlikely(ndirties >= nlimit)) {
 747                                 folio_batch_release(&fbatch);
 748                                 cond_resched();
 749                                 return ndirties;
 750                         }
 751                 } while (bh = bh->b_this_page, bh != head);
 752         }
 753         folio_batch_release(&fbatch);
 754         cond_resched();
 755         goto repeat;
 756 }
 757
 758 static void nilfs_lookup_dirty_node_buffers(struct inode *inode,
 759                                             struct list_head *listp)
 760 {
 761         struct nilfs_inode_info *ii = NILFS_I(inode);
 762         struct inode *btnc_inode = ii->i_assoc_inode;
 763         struct folio_batch fbatch;
 764         struct buffer_head *bh, *head;
 765         unsigned int i;
 766         pgoff_t index = 0;
 767
 768         if (!btnc_inode)
 769                 return;
 770         folio_batch_init(&fbatch);
 771
 772         while (filemap_get_folios_tag(btnc_inode->i_mapping, &index,
 773                                 (pgoff_t)-1, PAGECACHE_TAG_DIRTY, &fbatch)) {
 774                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 775                         bh = head = folio_buffers(fbatch.folios[i]);
 776                         do {
 777                                 if (buffer_dirty(bh) &&
 778                                                 !buffer_async_write(bh)) {
 779                                         get_bh(bh);
 780                                         list_add_tail(&bh->b_assoc_buffers,
 781                                                       listp);
 782                                 }
 783                                 bh = bh->b_this_page;
 784                         } while (bh != head);
 785                 }
 786                 folio_batch_release(&fbatch);
 787                 cond_resched();
 788         }
 789 }
 790
 791 static void nilfs_dispose_list(struct the_nilfs *nilfs,
 792                                struct list_head *head, int force)
 793 {
 794         struct nilfs_inode_info *ii, *n;
 795         struct nilfs_inode_info *ivec[SC_N_INODEVEC], **pii;
 796         unsigned int nv = 0;
 797
 798         while (!list_empty(head)) {
 799                 spin_lock(&nilfs->ns_inode_lock);
 800                 list_for_each_entry_safe(ii, n, head, i_dirty) {
 801                         list_del_init(&ii->i_dirty);
 802                         if (force) {
 803                                 if (unlikely(ii->i_bh)) {
 804                                         brelse(ii->i_bh);
 805                                         ii->i_bh = NULL;
 806                                 }
 807                         } else if (test_bit(NILFS_I_DIRTY, &ii->i_state)) {
 808                                 set_bit(NILFS_I_QUEUED, &ii->i_state);
 809                                 list_add_tail(&ii->i_dirty,
 810                                               &nilfs->ns_dirty_files);
 811                                 continue;
 812                         }
 813                         ivec[nv++] = ii;
 814                         if (nv == SC_N_INODEVEC)
 815                                 break;
 816                 }
 817                 spin_unlock(&nilfs->ns_inode_lock);
 818
 819                 for (pii = ivec; nv > 0; pii++, nv--)
 820                         iput(&(*pii)->vfs_inode);
 821         }
 822 }
 823
 824 static void nilfs_iput_work_func(struct work_struct *work)
 825 {
 826         struct nilfs_sc_info *sci = container_of(work, struct nilfs_sc_info,
 827                                                  sc_iput_work);
 828         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 829
 830         nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 0);
 831 }
 832
 833 static int nilfs_test_metadata_dirty(struct the_nilfs *nilfs,
 834                                      struct nilfs_root *root)
 835 {
 836         int ret = 0;
 837
 838         if (nilfs_mdt_fetch_dirty(root->ifile))
 839                 ret++;
 840         if (nilfs_mdt_fetch_dirty(nilfs->ns_cpfile))
 841                 ret++;
 842         if (nilfs_mdt_fetch_dirty(nilfs->ns_sufile))
 843                 ret++;
 844         if ((ret || nilfs_doing_gc()) && nilfs_mdt_fetch_dirty(nilfs->ns_dat))
 845                 ret++;
 846         return ret;
 847 }
 848
 849 static int nilfs_segctor_clean(struct nilfs_sc_info *sci)
 850 {
 851         return list_empty(&sci->sc_dirty_files) &&
 852                 !test_bit(NILFS_SC_DIRTY, &sci->sc_flags) &&
 853                 sci->sc_nfreesegs == 0 &&
 854                 (!nilfs_doing_gc() || list_empty(&sci->sc_gc_inodes));
 855 }
 856
 857 static int nilfs_segctor_confirm(struct nilfs_sc_info *sci)
 858 {
 859         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 860         int ret = 0;
 861
 862         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
 863                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
 864
 865         spin_lock(&nilfs->ns_inode_lock);
 866         if (list_empty(&nilfs->ns_dirty_files) && nilfs_segctor_clean(sci))
 867                 ret++;
 868
 869         spin_unlock(&nilfs->ns_inode_lock);
 870         return ret;
 871 }
 872
 873 static void nilfs_segctor_clear_metadata_dirty(struct nilfs_sc_info *sci)
 874 {
 875         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
 876
 877         nilfs_mdt_clear_dirty(sci->sc_root->ifile);
 878         nilfs_mdt_clear_dirty(nilfs->ns_cpfile);
 879         nilfs_mdt_clear_dirty(nilfs->ns_sufile);
 880         nilfs_mdt_clear_dirty(nilfs->ns_dat);
 881 }
 882
 883 static void nilfs_fill_in_file_bmap(struct inode *ifile,
 884                                     struct nilfs_inode_info *ii)
 885
 886 {
 887         struct buffer_head *ibh;
 888         struct nilfs_inode *raw_inode;
 889
 890         if (test_bit(NILFS_I_BMAP, &ii->i_state)) {
 891                 ibh = ii->i_bh;
 892                 BUG_ON(!ibh);
 893                 raw_inode = nilfs_ifile_map_inode(ifile, ii->vfs_inode.i_ino,
 894                                                   ibh);
 895                 nilfs_bmap_write(ii->i_bmap, raw_inode);
 896                 nilfs_ifile_unmap_inode(raw_inode);
 897         }
 898 }
 899
 900 static void nilfs_segctor_fill_in_file_bmap(struct nilfs_sc_info *sci)
 901 {
 902         struct nilfs_inode_info *ii;
 903
 904         list_for_each_entry(ii, &sci->sc_dirty_files, i_dirty) {
 905                 nilfs_fill_in_file_bmap(sci->sc_root->ifile, ii);
 906                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
 907         }
 908 }
 909
 910 /**
 911  * nilfs_write_root_mdt_inode - export root metadata inode information to
 912  *                              the on-disk inode
 913  * @inode:     inode object of the root metadata file
 914  * @raw_inode: on-disk inode
 915  *
 916  * nilfs_write_root_mdt_inode() writes inode information and bmap data of
 917  * @inode to the inode area of the metadata file allocated on the super root
 918  * block created to finalize the log.  Since super root blocks are configured
 919  * each time, this function zero-fills the unused area of @raw_inode.
 920  */
 921 static void nilfs_write_root_mdt_inode(struct inode *inode,
 922                                        struct nilfs_inode *raw_inode)
 923 {
 924         struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
 925
 926         nilfs_write_inode_common(inode, raw_inode);
 927
 928         /* zero-fill unused portion of raw_inode */
 929         raw_inode->i_xattr = 0;
 930         raw_inode->i_pad = 0;
 931         memset((void *)raw_inode + sizeof(*raw_inode), 0,
 932                nilfs->ns_inode_size - sizeof(*raw_inode));
 933
 934         nilfs_bmap_write(NILFS_I(inode)->i_bmap, raw_inode);
 935 }
 936
 937 static void nilfs_segctor_fill_in_super_root(struct nilfs_sc_info *sci,
 938                                              struct the_nilfs *nilfs)
 939 {
 940         struct buffer_head *bh_sr;
 941         struct nilfs_super_root *raw_sr;
 942         unsigned int isz, srsz;
 943
 944         bh_sr = NILFS_LAST_SEGBUF(&sci->sc_segbufs)->sb_super_root;
 945
 946         lock_buffer(bh_sr);
 947         raw_sr = (struct nilfs_super_root *)bh_sr->b_data;
 948         isz = nilfs->ns_inode_size;
 949         srsz = NILFS_SR_BYTES(isz);
 950
 951         raw_sr->sr_sum = 0;  /* Ensure initialization within this update */
 952         raw_sr->sr_bytes = cpu_to_le16(srsz);
 953         raw_sr->sr_nongc_ctime
 954                 = cpu_to_le64(nilfs_doing_gc() ?
 955                               nilfs->ns_nongc_ctime : sci->sc_seg_ctime);
 956         raw_sr->sr_flags = 0;
 957
 958         nilfs_write_root_mdt_inode(nilfs->ns_dat, (void *)raw_sr +
 959                                    NILFS_SR_DAT_OFFSET(isz));
 960         nilfs_write_root_mdt_inode(nilfs->ns_cpfile, (void *)raw_sr +
 961                                    NILFS_SR_CPFILE_OFFSET(isz));
 962         nilfs_write_root_mdt_inode(nilfs->ns_sufile, (void *)raw_sr +
 963                                    NILFS_SR_SUFILE_OFFSET(isz));
 964
 965         memset((void *)raw_sr + srsz, 0, nilfs->ns_blocksize - srsz);
 966         set_buffer_uptodate(bh_sr);
 967         unlock_buffer(bh_sr);
 968 }
 969
 970 static void nilfs_redirty_inodes(struct list_head *head)
 971 {
 972         struct nilfs_inode_info *ii;
 973
 974         list_for_each_entry(ii, head, i_dirty) {
 975                 if (test_bit(NILFS_I_COLLECTED, &ii->i_state))
 976                         clear_bit(NILFS_I_COLLECTED, &ii->i_state);
 977         }
 978 }
 979
 980 static void nilfs_drop_collected_inodes(struct list_head *head)
 981 {
 982         struct nilfs_inode_info *ii;
 983
 984         list_for_each_entry(ii, head, i_dirty) {
 985                 if (!test_and_clear_bit(NILFS_I_COLLECTED, &ii->i_state))
 986                         continue;
 987
 988                 clear_bit(NILFS_I_INODE_SYNC, &ii->i_state);
 989                 set_bit(NILFS_I_UPDATED, &ii->i_state);
 990         }
 991 }
 992
 993 static int nilfs_segctor_apply_buffers(struct nilfs_sc_info *sci,
 994                                        struct inode *inode,
 995                                        struct list_head *listp,
 996                                        int (*collect)(struct nilfs_sc_info *,
 997                                                       struct buffer_head *,
 998                                                       struct inode *))
 999 {
1000         struct buffer_head *bh, *n;
1001         int err = 0;
1002
1003         if (collect) {
1004                 list_for_each_entry_safe(bh, n, listp, b_assoc_buffers) {
1005                         list_del_init(&bh->b_assoc_buffers);
1006                         err = collect(sci, bh, inode);
1007                         brelse(bh);
1008                         if (unlikely(err))
1009                                 goto dispose_buffers;
1010                 }
1011                 return 0;
1012         }
1013
1014  dispose_buffers:
1015         while (!list_empty(listp)) {
1016                 bh = list_first_entry(listp, struct buffer_head,
1017                                       b_assoc_buffers);
1018                 list_del_init(&bh->b_assoc_buffers);
1019                 brelse(bh);
1020         }
1021         return err;
1022 }
1023
1024 static size_t nilfs_segctor_buffer_rest(struct nilfs_sc_info *sci)
1025 {
1026         /* Remaining number of blocks within segment buffer */
1027         return sci->sc_segbuf_nblocks -
1028                 (sci->sc_nblk_this_inc + sci->sc_curseg->sb_sum.nblocks);
1029 }
1030
1031 static int nilfs_segctor_scan_file(struct nilfs_sc_info *sci,
1032                                    struct inode *inode,
1033                                    const struct nilfs_sc_operations *sc_ops)
1034 {
1035         LIST_HEAD(data_buffers);
1036         LIST_HEAD(node_buffers);
1037         int err;
1038
1039         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1040                 size_t n, rest = nilfs_segctor_buffer_rest(sci);
1041
1042                 n = nilfs_lookup_dirty_data_buffers(
1043                         inode, &data_buffers, rest + 1, 0, LLONG_MAX);
1044                 if (n > rest) {
1045                         err = nilfs_segctor_apply_buffers(
1046                                 sci, inode, &data_buffers,
1047                                 sc_ops->collect_data);
1048                         BUG_ON(!err); /* always receive -E2BIG or true error */
1049                         goto break_or_fail;
1050                 }
1051         }
1052         nilfs_lookup_dirty_node_buffers(inode, &node_buffers);
1053
1054         if (!(sci->sc_stage.flags & NILFS_CF_NODE)) {
1055                 err = nilfs_segctor_apply_buffers(
1056                         sci, inode, &data_buffers, sc_ops->collect_data);
1057                 if (unlikely(err)) {
1058                         /* dispose node list */
1059                         nilfs_segctor_apply_buffers(
1060                                 sci, inode, &node_buffers, NULL);
1061                         goto break_or_fail;
1062                 }
1063                 sci->sc_stage.flags |= NILFS_CF_NODE;
1064         }
1065         /* Collect node */
1066         err = nilfs_segctor_apply_buffers(
1067                 sci, inode, &node_buffers, sc_ops->collect_node);
1068         if (unlikely(err))
1069                 goto break_or_fail;
1070
1071         nilfs_bmap_lookup_dirty_buffers(NILFS_I(inode)->i_bmap, &node_buffers);
1072         err = nilfs_segctor_apply_buffers(
1073                 sci, inode, &node_buffers, sc_ops->collect_bmap);
1074         if (unlikely(err))
1075                 goto break_or_fail;
1076
1077         nilfs_segctor_end_finfo(sci, inode);
1078         sci->sc_stage.flags &= ~NILFS_CF_NODE;
1079
1080  break_or_fail:
1081         return err;
1082 }
1083
1084 static int nilfs_segctor_scan_file_dsync(struct nilfs_sc_info *sci,
1085                                          struct inode *inode)
1086 {
1087         LIST_HEAD(data_buffers);
1088         size_t n, rest = nilfs_segctor_buffer_rest(sci);
1089         int err;
1090
1091         n = nilfs_lookup_dirty_data_buffers(inode, &data_buffers, rest + 1,
1092                                             sci->sc_dsync_start,
1093                                             sci->sc_dsync_end);
1094
1095         err = nilfs_segctor_apply_buffers(sci, inode, &data_buffers,
1096                                           nilfs_collect_file_data);
1097         if (!err) {
1098                 nilfs_segctor_end_finfo(sci, inode);
1099                 BUG_ON(n > rest);
1100                 /* always receive -E2BIG or true error if n > rest */
1101         }
1102         return err;
1103 }
1104
1105 /**
1106  * nilfs_free_segments - free the segments given by an array of segment numbers
1107  * @nilfs:   nilfs object
1108  * @segnumv: array of segment numbers to be freed
1109  * @nsegs:   number of segments to be freed in @segnumv
1110  *
1111  * nilfs_free_segments() wraps nilfs_sufile_freev() and
1112  * nilfs_sufile_cancel_freev(), and edits the segment usage metadata file
1113  * (sufile) to free all segments given by @segnumv and @nsegs at once.  If
1114  * it fails midway, it cancels the changes so that none of the segments are
1115  * freed.  If @nsegs is 0, this function does nothing.
1116  *
1117  * The freeing of segments is not finalized until the writing of a log with
1118  * a super root block containing this sufile change is complete, and it can
1119  * be canceled with nilfs_sufile_cancel_freev() until then.
1120  *
1121  * Return: 0 on success, or the following negative error code on failure.
1122  * * %-EINVAL   - Invalid segment number.
1123  * * %-EIO      - I/O error (including metadata corruption).
1124  * * %-ENOMEM   - Insufficient memory available.
1125  */
1126 static int nilfs_free_segments(struct the_nilfs *nilfs, __u64 *segnumv,
1127                                size_t nsegs)
1128 {
1129         size_t ndone;
1130         int ret;
1131
1132         if (!nsegs)
1133                 return 0;
1134
1135         ret = nilfs_sufile_freev(nilfs->ns_sufile, segnumv, nsegs, &ndone);
1136         if (unlikely(ret)) {
1137                 nilfs_sufile_cancel_freev(nilfs->ns_sufile, segnumv, ndone,
1138                                           NULL);
1139                 /*
1140                  * If a segment usage of the segments to be freed is in a
1141                  * hole block, nilfs_sufile_freev() will return -ENOENT.
1142                  * In this case, -EINVAL should be returned to the caller
1143                  * since there is something wrong with the given segment
1144                  * number array.  This error can only occur during GC, so
1145                  * there is no need to worry about it propagating to other
1146                  * callers (such as fsync).
1147                  */
1148                 if (ret == -ENOENT) {
1149                         nilfs_err(nilfs->ns_sb,
1150                                   "The segment usage entry %llu to be freed is invalid (in a hole)",
1151                                   (unsigned long long)segnumv[ndone]);
1152                         ret = -EINVAL;
1153                 }
1154         }
1155         return ret;
1156 }
1157
1158 static int nilfs_segctor_collect_blocks(struct nilfs_sc_info *sci, int mode)
1159 {
1160         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1161         struct list_head *head;
1162         struct nilfs_inode_info *ii;
1163         int err = 0;
1164
1165         switch (nilfs_sc_cstage_get(sci)) {
1166         case NILFS_ST_INIT:
1167                 /* Pre-processes */
1168                 sci->sc_stage.flags = 0;
1169
1170                 if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags)) {
1171                         sci->sc_nblk_inc = 0;
1172                         sci->sc_curseg->sb_sum.flags = NILFS_SS_LOGBGN;
1173                         if (mode == SC_LSEG_DSYNC) {
1174                                 nilfs_sc_cstage_set(sci, NILFS_ST_DSYNC);
1175                                 goto dsync_mode;
1176                         }
1177                 }
1178
1179                 sci->sc_stage.dirty_file_ptr = NULL;
1180                 sci->sc_stage.gc_inode_ptr = NULL;
1181                 if (mode == SC_FLUSH_DAT) {
1182                         nilfs_sc_cstage_set(sci, NILFS_ST_DAT);
1183                         goto dat_stage;
1184                 }
1185                 nilfs_sc_cstage_inc(sci);
1186                 fallthrough;
1187         case NILFS_ST_GC:
1188                 if (nilfs_doing_gc()) {
1189                         head = &sci->sc_gc_inodes;
1190                         ii = list_prepare_entry(sci->sc_stage.gc_inode_ptr,
1191                                                 head, i_dirty);
1192                         list_for_each_entry_continue(ii, head, i_dirty) {
1193                                 err = nilfs_segctor_scan_file(
1194                                         sci, &ii->vfs_inode,
1195                                         &nilfs_sc_file_ops);
1196                                 if (unlikely(err)) {
1197                                         sci->sc_stage.gc_inode_ptr = list_entry(
1198                                                 ii->i_dirty.prev,
1199                                                 struct nilfs_inode_info,
1200                                                 i_dirty);
1201                                         goto break_or_fail;
1202                                 }
1203                                 set_bit(NILFS_I_COLLECTED, &ii->i_state);
1204                         }
1205                         sci->sc_stage.gc_inode_ptr = NULL;
1206                 }
1207                 nilfs_sc_cstage_inc(sci);
1208                 fallthrough;
1209         case NILFS_ST_FILE:
1210                 head = &sci->sc_dirty_files;
1211                 ii = list_prepare_entry(sci->sc_stage.dirty_file_ptr, head,
1212                                         i_dirty);
1213                 list_for_each_entry_continue(ii, head, i_dirty) {
1214                         clear_bit(NILFS_I_DIRTY, &ii->i_state);
1215
1216                         err = nilfs_segctor_scan_file(sci, &ii->vfs_inode,
1217                                                       &nilfs_sc_file_ops);
1218                         if (unlikely(err)) {
1219                                 sci->sc_stage.dirty_file_ptr =
1220                                         list_entry(ii->i_dirty.prev,
1221                                                    struct nilfs_inode_info,
1222                                                    i_dirty);
1223                                 goto break_or_fail;
1224                         }
1225                         /* sci->sc_stage.dirty_file_ptr = NILFS_I(inode); */
1226                         /* XXX: required ? */
1227                 }
1228                 sci->sc_stage.dirty_file_ptr = NULL;
1229                 if (mode == SC_FLUSH_FILE) {
1230                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1231                         return 0;
1232                 }
1233                 nilfs_sc_cstage_inc(sci);
1234                 sci->sc_stage.flags |= NILFS_CF_IFILE_STARTED;
1235                 fallthrough;
1236         case NILFS_ST_IFILE:
1237                 err = nilfs_segctor_scan_file(sci, sci->sc_root->ifile,
1238                                               &nilfs_sc_file_ops);
1239                 if (unlikely(err))
1240                         break;
1241                 nilfs_sc_cstage_inc(sci);
1242                 /* Creating a checkpoint */
1243                 err = nilfs_cpfile_create_checkpoint(nilfs->ns_cpfile,
1244                                                      nilfs->ns_cno);
1245                 if (unlikely(err))
1246                         break;
1247                 fallthrough;
1248         case NILFS_ST_CPFILE:
1249                 err = nilfs_segctor_scan_file(sci, nilfs->ns_cpfile,
1250                                               &nilfs_sc_file_ops);
1251                 if (unlikely(err))
1252                         break;
1253                 nilfs_sc_cstage_inc(sci);
1254                 fallthrough;
1255         case NILFS_ST_SUFILE:
1256                 err = nilfs_free_segments(nilfs, sci->sc_freesegs,
1257                                           sci->sc_nfreesegs);
1258                 if (unlikely(err))
1259                         break;
1260                 sci->sc_stage.flags |= NILFS_CF_SUFREED;
1261
1262                 err = nilfs_segctor_scan_file(sci, nilfs->ns_sufile,
1263                                               &nilfs_sc_file_ops);
1264                 if (unlikely(err))
1265                         break;
1266                 nilfs_sc_cstage_inc(sci);
1267                 fallthrough;
1268         case NILFS_ST_DAT:
1269  dat_stage:
1270                 err = nilfs_segctor_scan_file(sci, nilfs->ns_dat,
1271                                               &nilfs_sc_dat_ops);
1272                 if (unlikely(err))
1273                         break;
1274                 if (mode == SC_FLUSH_DAT) {
1275                         nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1276                         return 0;
1277                 }
1278                 nilfs_sc_cstage_inc(sci);
1279                 fallthrough;
1280         case NILFS_ST_SR:
1281                 if (mode == SC_LSEG_SR) {
1282                         /* Appending a super root */
1283                         err = nilfs_segctor_add_super_root(sci);
1284                         if (unlikely(err))
1285                                 break;
1286                 }
1287                 /* End of a logical segment */
1288                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1289                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1290                 return 0;
1291         case NILFS_ST_DSYNC:
1292  dsync_mode:
1293                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_SYNDT;
1294                 ii = sci->sc_dsync_inode;
1295                 if (!test_bit(NILFS_I_BUSY, &ii->i_state))
1296                         break;
1297
1298                 err = nilfs_segctor_scan_file_dsync(sci, &ii->vfs_inode);
1299                 if (unlikely(err))
1300                         break;
1301                 sci->sc_curseg->sb_sum.flags |= NILFS_SS_LOGEND;
1302                 nilfs_sc_cstage_set(sci, NILFS_ST_DONE);
1303                 return 0;
1304         case NILFS_ST_DONE:
1305                 return 0;
1306         default:
1307                 BUG();
1308         }
1309
1310  break_or_fail:
1311         return err;
1312 }
1313
1314 /**
1315  * nilfs_segctor_begin_construction - setup segment buffer to make a new log
1316  * @sci: nilfs_sc_info
1317  * @nilfs: nilfs object
1318  */
1319 static int nilfs_segctor_begin_construction(struct nilfs_sc_info *sci,
1320                                             struct the_nilfs *nilfs)
1321 {
1322         struct nilfs_segment_buffer *segbuf, *prev;
1323         __u64 nextnum;
1324         int err, alloc = 0;
1325
1326         segbuf = nilfs_segbuf_new(sci->sc_super);
1327         if (unlikely(!segbuf))
1328                 return -ENOMEM;
1329
1330         if (list_empty(&sci->sc_write_logs)) {
1331                 nilfs_segbuf_map(segbuf, nilfs->ns_segnum,
1332                                  nilfs->ns_pseg_offset, nilfs);
1333                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1334                         nilfs_shift_to_next_segment(nilfs);
1335                         nilfs_segbuf_map(segbuf, nilfs->ns_segnum, 0, nilfs);
1336                 }
1337
1338                 segbuf->sb_sum.seg_seq = nilfs->ns_seg_seq;
1339                 nextnum = nilfs->ns_nextnum;
1340
1341                 if (nilfs->ns_segnum == nilfs->ns_nextnum)
1342                         /* Start from the head of a new full segment */
1343                         alloc++;
1344         } else {
1345                 /* Continue logs */
1346                 prev = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1347                 nilfs_segbuf_map_cont(segbuf, prev);
1348                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq;
1349                 nextnum = prev->sb_nextnum;
1350
1351                 if (segbuf->sb_rest_blocks < NILFS_PSEG_MIN_BLOCKS) {
1352                         nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1353                         segbuf->sb_sum.seg_seq++;
1354                         alloc++;
1355                 }
1356         }
1357
1358         err = nilfs_sufile_mark_dirty(nilfs->ns_sufile, segbuf->sb_segnum);
1359         if (err)
1360                 goto failed;
1361
1362         if (alloc) {
1363                 err = nilfs_sufile_alloc(nilfs->ns_sufile, &nextnum);
1364                 if (err)
1365                         goto failed;
1366         }
1367         nilfs_segbuf_set_next_segnum(segbuf, nextnum, nilfs);
1368
1369         BUG_ON(!list_empty(&sci->sc_segbufs));
1370         list_add_tail(&segbuf->sb_list, &sci->sc_segbufs);
1371         sci->sc_segbuf_nblocks = segbuf->sb_rest_blocks;
1372         return 0;
1373
1374  failed:
1375         nilfs_segbuf_free(segbuf);
1376         return err;
1377 }
1378
1379 static int nilfs_segctor_extend_segments(struct nilfs_sc_info *sci,
1380                                          struct the_nilfs *nilfs, int nadd)
1381 {
1382         struct nilfs_segment_buffer *segbuf, *prev;
1383         struct inode *sufile = nilfs->ns_sufile;
1384         __u64 nextnextnum;
1385         LIST_HEAD(list);
1386         int err, ret, i;
1387
1388         prev = NILFS_LAST_SEGBUF(&sci->sc_segbufs);
1389         /*
1390          * Since the segment specified with nextnum might be allocated during
1391          * the previous construction, the buffer including its segusage may
1392          * not be dirty.  The following call ensures that the buffer is dirty
1393          * and will pin the buffer on memory until the sufile is written.
1394          */
1395         err = nilfs_sufile_mark_dirty(sufile, prev->sb_nextnum);
1396         if (unlikely(err))
1397                 return err;
1398
1399         for (i = 0; i < nadd; i++) {
1400                 /* extend segment info */
1401                 err = -ENOMEM;
1402                 segbuf = nilfs_segbuf_new(sci->sc_super);
1403                 if (unlikely(!segbuf))
1404                         goto failed;
1405
1406                 /* map this buffer to region of segment on-disk */
1407                 nilfs_segbuf_map(segbuf, prev->sb_nextnum, 0, nilfs);
1408                 sci->sc_segbuf_nblocks += segbuf->sb_rest_blocks;
1409
1410                 /* allocate the next next full segment */
1411                 err = nilfs_sufile_alloc(sufile, &nextnextnum);
1412                 if (unlikely(err))
1413                         goto failed_segbuf;
1414
1415                 segbuf->sb_sum.seg_seq = prev->sb_sum.seg_seq + 1;
1416                 nilfs_segbuf_set_next_segnum(segbuf, nextnextnum, nilfs);
1417
1418                 list_add_tail(&segbuf->sb_list, &list);
1419                 prev = segbuf;
1420         }
1421         list_splice_tail(&list, &sci->sc_segbufs);
1422         return 0;
1423
1424  failed_segbuf:
1425         nilfs_segbuf_free(segbuf);
1426  failed:
1427         list_for_each_entry(segbuf, &list, sb_list) {
1428                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1429                 WARN_ON(ret); /* never fails */
1430         }
1431         nilfs_destroy_logs(&list);
1432         return err;
1433 }
1434
1435 static void nilfs_free_incomplete_logs(struct list_head *logs,
1436                                        struct the_nilfs *nilfs)
1437 {
1438         struct nilfs_segment_buffer *segbuf, *prev;
1439         struct inode *sufile = nilfs->ns_sufile;
1440         int ret;
1441
1442         segbuf = NILFS_FIRST_SEGBUF(logs);
1443         if (nilfs->ns_nextnum != segbuf->sb_nextnum) {
1444                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1445                 WARN_ON(ret); /* never fails */
1446         }
1447         if (atomic_read(&segbuf->sb_err)) {
1448                 /* Case 1: The first segment failed */
1449                 if (segbuf->sb_pseg_start != segbuf->sb_fseg_start)
1450                         /*
1451                          * Case 1a:  Partial segment appended into an existing
1452                          * segment
1453                          */
1454                         nilfs_terminate_segment(nilfs, segbuf->sb_fseg_start,
1455                                                 segbuf->sb_fseg_end);
1456                 else /* Case 1b:  New full segment */
1457                         set_nilfs_discontinued(nilfs);
1458         }
1459
1460         prev = segbuf;
1461         list_for_each_entry_continue(segbuf, logs, sb_list) {
1462                 if (prev->sb_nextnum != segbuf->sb_nextnum) {
1463                         ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1464                         WARN_ON(ret); /* never fails */
1465                 }
1466                 if (atomic_read(&segbuf->sb_err) &&
1467                     segbuf->sb_segnum != nilfs->ns_nextnum)
1468                         /* Case 2: extended segment (!= next) failed */
1469                         nilfs_sufile_set_error(sufile, segbuf->sb_segnum);
1470                 prev = segbuf;
1471         }
1472 }
1473
1474 static void nilfs_segctor_update_segusage(struct nilfs_sc_info *sci,
1475                                           struct inode *sufile)
1476 {
1477         struct nilfs_segment_buffer *segbuf;
1478         unsigned long live_blocks;
1479         int ret;
1480
1481         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1482                 live_blocks = segbuf->sb_sum.nblocks +
1483                         (segbuf->sb_pseg_start - segbuf->sb_fseg_start);
1484                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1485                                                      live_blocks,
1486                                                      sci->sc_seg_ctime);
1487                 WARN_ON(ret); /* always succeed because the segusage is dirty */
1488         }
1489 }
1490
1491 static void nilfs_cancel_segusage(struct list_head *logs, struct inode *sufile)
1492 {
1493         struct nilfs_segment_buffer *segbuf;
1494         int ret;
1495
1496         segbuf = NILFS_FIRST_SEGBUF(logs);
1497         ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1498                                              segbuf->sb_pseg_start -
1499                                              segbuf->sb_fseg_start, 0);
1500         WARN_ON(ret); /* always succeed because the segusage is dirty */
1501
1502         list_for_each_entry_continue(segbuf, logs, sb_list) {
1503                 ret = nilfs_sufile_set_segment_usage(sufile, segbuf->sb_segnum,
1504                                                      0, 0);
1505                 WARN_ON(ret); /* always succeed */
1506         }
1507 }
1508
1509 static void nilfs_segctor_truncate_segments(struct nilfs_sc_info *sci,
1510                                             struct nilfs_segment_buffer *last,
1511                                             struct inode *sufile)
1512 {
1513         struct nilfs_segment_buffer *segbuf = last;
1514         int ret;
1515
1516         list_for_each_entry_continue(segbuf, &sci->sc_segbufs, sb_list) {
1517                 sci->sc_segbuf_nblocks -= segbuf->sb_rest_blocks;
1518                 ret = nilfs_sufile_free(sufile, segbuf->sb_nextnum);
1519                 WARN_ON(ret);
1520         }
1521         nilfs_truncate_logs(&sci->sc_segbufs, last);
1522 }
1523
1524
1525 static int nilfs_segctor_collect(struct nilfs_sc_info *sci,
1526                                  struct the_nilfs *nilfs, int mode)
1527 {
1528         struct nilfs_cstage prev_stage = sci->sc_stage;
1529         int err, nadd = 1;
1530
1531         /* Collection retry loop */
1532         for (;;) {
1533                 sci->sc_nblk_this_inc = 0;
1534                 sci->sc_curseg = NILFS_FIRST_SEGBUF(&sci->sc_segbufs);
1535
1536                 err = nilfs_segctor_reset_segment_buffer(sci);
1537                 if (unlikely(err))
1538                         goto failed;
1539
1540                 err = nilfs_segctor_collect_blocks(sci, mode);
1541                 sci->sc_nblk_this_inc += sci->sc_curseg->sb_sum.nblocks;
1542                 if (!err)
1543                         break;
1544
1545                 if (unlikely(err != -E2BIG))
1546                         goto failed;
1547
1548                 /* The current segment is filled up */
1549                 if (mode != SC_LSEG_SR ||
1550                     nilfs_sc_cstage_get(sci) < NILFS_ST_CPFILE)
1551                         break;
1552
1553                 nilfs_clear_logs(&sci->sc_segbufs);
1554
1555                 if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1556                         err = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1557                                                         sci->sc_freesegs,
1558                                                         sci->sc_nfreesegs,
1559                                                         NULL);
1560                         WARN_ON(err); /* do not happen */
1561                         sci->sc_stage.flags &= ~NILFS_CF_SUFREED;
1562                 }
1563
1564                 err = nilfs_segctor_extend_segments(sci, nilfs, nadd);
1565                 if (unlikely(err))
1566                         return err;
1567
1568                 nadd = min_t(int, nadd << 1, SC_MAX_SEGDELTA);
1569                 sci->sc_stage = prev_stage;
1570         }
1571         nilfs_segctor_zeropad_segsum(sci);
1572         nilfs_segctor_truncate_segments(sci, sci->sc_curseg, nilfs->ns_sufile);
1573         return 0;
1574
1575  failed:
1576         return err;
1577 }
1578
1579 static void nilfs_list_replace_buffer(struct buffer_head *old_bh,
1580                                       struct buffer_head *new_bh)
1581 {
1582         BUG_ON(!list_empty(&new_bh->b_assoc_buffers));
1583
1584         list_replace_init(&old_bh->b_assoc_buffers, &new_bh->b_assoc_buffers);
1585         /* The caller must release old_bh */
1586 }
1587
1588 static int
1589 nilfs_segctor_update_payload_blocknr(struct nilfs_sc_info *sci,
1590                                      struct nilfs_segment_buffer *segbuf,
1591                                      int mode)
1592 {
1593         struct inode *inode = NULL;
1594         sector_t blocknr;
1595         unsigned long nfinfo = segbuf->sb_sum.nfinfo;
1596         unsigned long nblocks = 0, ndatablk = 0;
1597         const struct nilfs_sc_operations *sc_op = NULL;
1598         struct nilfs_segsum_pointer ssp;
1599         struct nilfs_finfo *finfo = NULL;
1600         union nilfs_binfo binfo;
1601         struct buffer_head *bh, *bh_org;
1602         ino_t ino = 0;
1603         int err = 0;
1604
1605         if (!nfinfo)
1606                 goto out;
1607
1608         blocknr = segbuf->sb_pseg_start + segbuf->sb_sum.nsumblk;
1609         ssp.bh = NILFS_SEGBUF_FIRST_BH(&segbuf->sb_segsum_buffers);
1610         ssp.offset = sizeof(struct nilfs_segment_summary);
1611
1612         list_for_each_entry(bh, &segbuf->sb_payload_buffers, b_assoc_buffers) {
1613                 if (bh == segbuf->sb_super_root)
1614                         break;
1615                 if (!finfo) {
1616                         finfo = nilfs_segctor_map_segsum_entry(
1617                                 sci, &ssp, sizeof(*finfo));
1618                         ino = le64_to_cpu(finfo->fi_ino);
1619                         nblocks = le32_to_cpu(finfo->fi_nblocks);
1620                         ndatablk = le32_to_cpu(finfo->fi_ndatablk);
1621
1622                         inode = bh->b_folio->mapping->host;
1623
1624                         if (mode == SC_LSEG_DSYNC)
1625                                 sc_op = &nilfs_sc_dsync_ops;
1626                         else if (ino == NILFS_DAT_INO)
1627                                 sc_op = &nilfs_sc_dat_ops;
1628                         else /* file blocks */
1629                                 sc_op = &nilfs_sc_file_ops;
1630                 }
1631                 bh_org = bh;
1632                 get_bh(bh_org);
1633                 err = nilfs_bmap_assign(NILFS_I(inode)->i_bmap, &bh, blocknr,
1634                                         &binfo);
1635                 if (bh != bh_org)
1636                         nilfs_list_replace_buffer(bh_org, bh);
1637                 brelse(bh_org);
1638                 if (unlikely(err))
1639                         goto failed_bmap;
1640
1641                 if (ndatablk > 0)
1642                         sc_op->write_data_binfo(sci, &ssp, &binfo);
1643                 else
1644                         sc_op->write_node_binfo(sci, &ssp, &binfo);
1645
1646                 blocknr++;
1647                 if (--nblocks == 0) {
1648                         finfo = NULL;
1649                         if (--nfinfo == 0)
1650                                 break;
1651                 } else if (ndatablk > 0)
1652                         ndatablk--;
1653         }
1654  out:
1655         return 0;
1656
1657  failed_bmap:
1658         return err;
1659 }
1660
1661 static int nilfs_segctor_assign(struct nilfs_sc_info *sci, int mode)
1662 {
1663         struct nilfs_segment_buffer *segbuf;
1664         int err;
1665
1666         list_for_each_entry(segbuf, &sci->sc_segbufs, sb_list) {
1667                 err = nilfs_segctor_update_payload_blocknr(sci, segbuf, mode);
1668                 if (unlikely(err))
1669                         return err;
1670                 nilfs_segbuf_fill_in_segsum(segbuf);
1671         }
1672         return 0;
1673 }
1674
1675 static void nilfs_begin_folio_io(struct folio *folio)
1676 {
1677         if (!folio || folio_test_writeback(folio))
1678                 /*
1679                  * For split b-tree node pages, this function may be called
1680                  * twice.  We ignore the 2nd or later calls by this check.
1681                  */
1682                 return;
1683
1684         folio_lock(folio);
1685         folio_clear_dirty_for_io(folio);
1686         folio_start_writeback(folio);
1687         folio_unlock(folio);
1688 }
1689
1690 /**
1691  * nilfs_prepare_write_logs - prepare to write logs
1692  * @logs: logs to prepare for writing
1693  * @seed: checksum seed value
1694  *
1695  * nilfs_prepare_write_logs() adds checksums and prepares the block
1696  * buffers/folios for writing logs.  In order to stabilize folios of
1697  * memory-mapped file blocks by putting them in writeback state before
1698  * calculating the checksums, first prepare to write payload blocks other
1699  * than segment summary and super root blocks in which the checksums will
1700  * be embedded.
1701  */
1702 static void nilfs_prepare_write_logs(struct list_head *logs, u32 seed)
1703 {
1704         struct nilfs_segment_buffer *segbuf;
1705         struct folio *bd_folio = NULL, *fs_folio = NULL;
1706         struct buffer_head *bh;
1707
1708         /* Prepare to write payload blocks */
1709         list_for_each_entry(segbuf, logs, sb_list) {
1710                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1711                                     b_assoc_buffers) {
1712                         if (bh == segbuf->sb_super_root)
1713                                 break;
1714                         set_buffer_async_write(bh);
1715                         if (bh->b_folio != fs_folio) {
1716                                 nilfs_begin_folio_io(fs_folio);
1717                                 fs_folio = bh->b_folio;
1718                         }
1719                 }
1720         }
1721         nilfs_begin_folio_io(fs_folio);
1722
1723         nilfs_add_checksums_on_logs(logs, seed);
1724
1725         /* Prepare to write segment summary blocks */
1726         list_for_each_entry(segbuf, logs, sb_list) {
1727                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1728                                     b_assoc_buffers) {
1729                         mark_buffer_dirty(bh);
1730                         if (bh->b_folio == bd_folio)
1731                                 continue;
1732                         if (bd_folio) {
1733                                 folio_lock(bd_folio);
1734                                 folio_wait_writeback(bd_folio);
1735                                 folio_clear_dirty_for_io(bd_folio);
1736                                 folio_start_writeback(bd_folio);
1737                                 folio_unlock(bd_folio);
1738                         }
1739                         bd_folio = bh->b_folio;
1740                 }
1741         }
1742
1743         /* Prepare to write super root block */
1744         bh = NILFS_LAST_SEGBUF(logs)->sb_super_root;
1745         if (bh) {
1746                 mark_buffer_dirty(bh);
1747                 if (bh->b_folio != bd_folio) {
1748                         folio_lock(bd_folio);
1749                         folio_wait_writeback(bd_folio);
1750                         folio_clear_dirty_for_io(bd_folio);
1751                         folio_start_writeback(bd_folio);
1752                         folio_unlock(bd_folio);
1753                         bd_folio = bh->b_folio;
1754                 }
1755         }
1756
1757         if (bd_folio) {
1758                 folio_lock(bd_folio);
1759                 folio_wait_writeback(bd_folio);
1760                 folio_clear_dirty_for_io(bd_folio);
1761                 folio_start_writeback(bd_folio);
1762                 folio_unlock(bd_folio);
1763         }
1764 }
1765
1766 static int nilfs_segctor_write(struct nilfs_sc_info *sci,
1767                                struct the_nilfs *nilfs)
1768 {
1769         int ret;
1770
1771         ret = nilfs_write_logs(&sci->sc_segbufs, nilfs);
1772         list_splice_tail_init(&sci->sc_segbufs, &sci->sc_write_logs);
1773         return ret;
1774 }
1775
1776 static void nilfs_end_folio_io(struct folio *folio, int err)
1777 {
1778         if (!folio)
1779                 return;
1780
1781         if (buffer_nilfs_node(folio_buffers(folio)) &&
1782                         !folio_test_writeback(folio)) {
1783                 /*
1784                  * For b-tree node pages, this function may be called twice
1785                  * or more because they might be split in a segment.
1786                  */
1787                 if (folio_test_dirty(folio)) {
1788                         /*
1789                          * For pages holding split b-tree node buffers, dirty
1790                          * flag on the buffers may be cleared discretely.
1791                          * In that case, the page is once redirtied for
1792                          * remaining buffers, and it must be cancelled if
1793                          * all the buffers get cleaned later.
1794                          */
1795                         folio_lock(folio);
1796                         if (nilfs_folio_buffers_clean(folio))
1797                                 __nilfs_clear_folio_dirty(folio);
1798                         folio_unlock(folio);
1799                 }
1800                 return;
1801         }
1802
1803         if (err || !nilfs_folio_buffers_clean(folio))
1804                 filemap_dirty_folio(folio->mapping, folio);
1805
1806         folio_end_writeback(folio);
1807 }
1808
1809 static void nilfs_abort_logs(struct list_head *logs, int err)
1810 {
1811         struct nilfs_segment_buffer *segbuf;
1812         struct folio *bd_folio = NULL, *fs_folio = NULL;
1813         struct buffer_head *bh;
1814
1815         if (list_empty(logs))
1816                 return;
1817
1818         list_for_each_entry(segbuf, logs, sb_list) {
1819                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1820                                     b_assoc_buffers) {
1821                         clear_buffer_uptodate(bh);
1822                         if (bh->b_folio != bd_folio) {
1823                                 if (bd_folio)
1824                                         folio_end_writeback(bd_folio);
1825                                 bd_folio = bh->b_folio;
1826                         }
1827                 }
1828
1829                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1830                                     b_assoc_buffers) {
1831                         if (bh == segbuf->sb_super_root) {
1832                                 clear_buffer_uptodate(bh);
1833                                 if (bh->b_folio != bd_folio) {
1834                                         folio_end_writeback(bd_folio);
1835                                         bd_folio = bh->b_folio;
1836                                 }
1837                                 break;
1838                         }
1839                         clear_buffer_async_write(bh);
1840                         if (bh->b_folio != fs_folio) {
1841                                 nilfs_end_folio_io(fs_folio, err);
1842                                 fs_folio = bh->b_folio;
1843                         }
1844                 }
1845         }
1846         if (bd_folio)
1847                 folio_end_writeback(bd_folio);
1848
1849         nilfs_end_folio_io(fs_folio, err);
1850 }
1851
1852 static void nilfs_segctor_abort_construction(struct nilfs_sc_info *sci,
1853                                              struct the_nilfs *nilfs, int err)
1854 {
1855         LIST_HEAD(logs);
1856         int ret;
1857
1858         list_splice_tail_init(&sci->sc_write_logs, &logs);
1859         ret = nilfs_wait_on_logs(&logs);
1860         nilfs_abort_logs(&logs, ret ? : err);
1861
1862         list_splice_tail_init(&sci->sc_segbufs, &logs);
1863         if (list_empty(&logs))
1864                 return; /* if the first segment buffer preparation failed */
1865
1866         nilfs_cancel_segusage(&logs, nilfs->ns_sufile);
1867         nilfs_free_incomplete_logs(&logs, nilfs);
1868
1869         if (sci->sc_stage.flags & NILFS_CF_SUFREED) {
1870                 ret = nilfs_sufile_cancel_freev(nilfs->ns_sufile,
1871                                                 sci->sc_freesegs,
1872                                                 sci->sc_nfreesegs,
1873                                                 NULL);
1874                 WARN_ON(ret); /* do not happen */
1875         }
1876
1877         nilfs_destroy_logs(&logs);
1878 }
1879
1880 static void nilfs_set_next_segment(struct the_nilfs *nilfs,
1881                                    struct nilfs_segment_buffer *segbuf)
1882 {
1883         nilfs->ns_segnum = segbuf->sb_segnum;
1884         nilfs->ns_nextnum = segbuf->sb_nextnum;
1885         nilfs->ns_pseg_offset = segbuf->sb_pseg_start - segbuf->sb_fseg_start
1886                 + segbuf->sb_sum.nblocks;
1887         nilfs->ns_seg_seq = segbuf->sb_sum.seg_seq;
1888         nilfs->ns_ctime = segbuf->sb_sum.ctime;
1889 }
1890
1891 static void nilfs_segctor_complete_write(struct nilfs_sc_info *sci)
1892 {
1893         struct nilfs_segment_buffer *segbuf;
1894         struct folio *bd_folio = NULL, *fs_folio = NULL;
1895         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
1896         int update_sr = false;
1897
1898         list_for_each_entry(segbuf, &sci->sc_write_logs, sb_list) {
1899                 struct buffer_head *bh;
1900
1901                 list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
1902                                     b_assoc_buffers) {
1903                         set_buffer_uptodate(bh);
1904                         clear_buffer_dirty(bh);
1905                         if (bh->b_folio != bd_folio) {
1906                                 if (bd_folio)
1907                                         folio_end_writeback(bd_folio);
1908                                 bd_folio = bh->b_folio;
1909                         }
1910                 }
1911                 /*
1912                  * We assume that the buffers which belong to the same folio
1913                  * continue over the buffer list.
1914                  * Under this assumption, the last BHs of folios is
1915                  * identifiable by the discontinuity of bh->b_folio
1916                  * (folio != fs_folio).
1917                  *
1918                  * For B-tree node blocks, however, this assumption is not
1919                  * guaranteed.  The cleanup code of B-tree node folios needs
1920                  * special care.
1921                  */
1922                 list_for_each_entry(bh, &segbuf->sb_payload_buffers,
1923                                     b_assoc_buffers) {
1924                         const unsigned long set_bits = BIT(BH_Uptodate);
1925                         const unsigned long clear_bits =
1926                                 (BIT(BH_Dirty) | BIT(BH_Async_Write) |
1927                                  BIT(BH_Delay) | BIT(BH_NILFS_Volatile) |
1928                                  BIT(BH_NILFS_Redirected));
1929
1930                         if (bh == segbuf->sb_super_root) {
1931                                 set_buffer_uptodate(bh);
1932                                 clear_buffer_dirty(bh);
1933                                 if (bh->b_folio != bd_folio) {
1934                                         folio_end_writeback(bd_folio);
1935                                         bd_folio = bh->b_folio;
1936                                 }
1937                                 update_sr = true;
1938                                 break;
1939                         }
1940                         set_mask_bits(&bh->b_state, clear_bits, set_bits);
1941                         if (bh->b_folio != fs_folio) {
1942                                 nilfs_end_folio_io(fs_folio, 0);
1943                                 fs_folio = bh->b_folio;
1944                         }
1945                 }
1946
1947                 if (!nilfs_segbuf_simplex(segbuf)) {
1948                         if (segbuf->sb_sum.flags & NILFS_SS_LOGBGN) {
1949                                 set_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1950                                 sci->sc_lseg_stime = jiffies;
1951                         }
1952                         if (segbuf->sb_sum.flags & NILFS_SS_LOGEND)
1953                                 clear_bit(NILFS_SC_UNCLOSED, &sci->sc_flags);
1954                 }
1955         }
1956         /*
1957          * Since folios may continue over multiple segment buffers,
1958          * end of the last folio must be checked outside of the loop.
1959          */
1960         if (bd_folio)
1961                 folio_end_writeback(bd_folio);
1962
1963         nilfs_end_folio_io(fs_folio, 0);
1964
1965         nilfs_drop_collected_inodes(&sci->sc_dirty_files);
1966
1967         if (nilfs_doing_gc())
1968                 nilfs_drop_collected_inodes(&sci->sc_gc_inodes);
1969         else
1970                 nilfs->ns_nongc_ctime = sci->sc_seg_ctime;
1971
1972         sci->sc_nblk_inc += sci->sc_nblk_this_inc;
1973
1974         segbuf = NILFS_LAST_SEGBUF(&sci->sc_write_logs);
1975         nilfs_set_next_segment(nilfs, segbuf);
1976
1977         if (update_sr) {
1978                 nilfs->ns_flushed_device = 0;
1979                 nilfs_set_last_segment(nilfs, segbuf->sb_pseg_start,
1980                                        segbuf->sb_sum.seg_seq, nilfs->ns_cno++);
1981
1982                 clear_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags);
1983                 clear_bit(NILFS_SC_DIRTY, &sci->sc_flags);
1984                 set_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1985                 nilfs_segctor_clear_metadata_dirty(sci);
1986         } else
1987                 clear_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags);
1988 }
1989
1990 static int nilfs_segctor_wait(struct nilfs_sc_info *sci)
1991 {
1992         int ret;
1993
1994         ret = nilfs_wait_on_logs(&sci->sc_write_logs);
1995         if (!ret) {
1996                 nilfs_segctor_complete_write(sci);
1997                 nilfs_destroy_logs(&sci->sc_write_logs);
1998         }
1999         return ret;
2000 }
2001
2002 static int nilfs_segctor_collect_dirty_files(struct nilfs_sc_info *sci,
2003                                              struct the_nilfs *nilfs)
2004 {
2005         struct nilfs_inode_info *ii, *n;
2006         struct inode *ifile = sci->sc_root->ifile;
2007
2008         spin_lock(&nilfs->ns_inode_lock);
2009  retry:
2010         list_for_each_entry_safe(ii, n, &nilfs->ns_dirty_files, i_dirty) {
2011                 if (!ii->i_bh) {
2012                         struct buffer_head *ibh;
2013                         int err;
2014
2015                         spin_unlock(&nilfs->ns_inode_lock);
2016                         err = nilfs_ifile_get_inode_block(
2017                                 ifile, ii->vfs_inode.i_ino, &ibh);
2018                         if (unlikely(err)) {
2019                                 nilfs_warn(sci->sc_super,
2020                                            "log writer: error %d getting inode block (ino=%lu)",
2021                                            err, ii->vfs_inode.i_ino);
2022                                 return err;
2023                         }
2024                         spin_lock(&nilfs->ns_inode_lock);
2025                         if (likely(!ii->i_bh))
2026                                 ii->i_bh = ibh;
2027                         else
2028                                 brelse(ibh);
2029                         goto retry;
2030                 }
2031
2032                 // Always redirty the buffer to avoid race condition
2033                 mark_buffer_dirty(ii->i_bh);
2034                 nilfs_mdt_mark_dirty(ifile);
2035
2036                 clear_bit(NILFS_I_QUEUED, &ii->i_state);
2037                 set_bit(NILFS_I_BUSY, &ii->i_state);
2038                 list_move_tail(&ii->i_dirty, &sci->sc_dirty_files);
2039         }
2040         spin_unlock(&nilfs->ns_inode_lock);
2041
2042         return 0;
2043 }
2044
2045 static void nilfs_segctor_drop_written_files(struct nilfs_sc_info *sci,
2046                                              struct the_nilfs *nilfs)
2047 {
2048         struct nilfs_inode_info *ii, *n;
2049         int during_mount = !(sci->sc_super->s_flags & SB_ACTIVE);
2050         int defer_iput = false;
2051
2052         spin_lock(&nilfs->ns_inode_lock);
2053         list_for_each_entry_safe(ii, n, &sci->sc_dirty_files, i_dirty) {
2054                 if (!test_and_clear_bit(NILFS_I_UPDATED, &ii->i_state) ||
2055                     test_bit(NILFS_I_DIRTY, &ii->i_state))
2056                         continue;
2057
2058                 clear_bit(NILFS_I_BUSY, &ii->i_state);
2059                 brelse(ii->i_bh);
2060                 ii->i_bh = NULL;
2061                 list_del_init(&ii->i_dirty);
2062                 if (!ii->vfs_inode.i_nlink || during_mount) {
2063                         /*
2064                          * Defer calling iput() to avoid deadlocks if
2065                          * i_nlink == 0 or mount is not yet finished.
2066                          */
2067                         list_add_tail(&ii->i_dirty, &sci->sc_iput_queue);
2068                         defer_iput = true;
2069                 } else {
2070                         spin_unlock(&nilfs->ns_inode_lock);
2071                         iput(&ii->vfs_inode);
2072                         spin_lock(&nilfs->ns_inode_lock);
2073                 }
2074         }
2075         spin_unlock(&nilfs->ns_inode_lock);
2076
2077         if (defer_iput)
2078                 schedule_work(&sci->sc_iput_work);
2079 }
2080
2081 /*
2082  * Main procedure of segment constructor
2083  */
2084 static int nilfs_segctor_do_construct(struct nilfs_sc_info *sci, int mode)
2085 {
2086         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2087         int err;
2088
2089         if (sb_rdonly(sci->sc_super))
2090                 return -EROFS;
2091
2092         nilfs_sc_cstage_set(sci, NILFS_ST_INIT);
2093         sci->sc_cno = nilfs->ns_cno;
2094
2095         err = nilfs_segctor_collect_dirty_files(sci, nilfs);
2096         if (unlikely(err))
2097                 goto out;
2098
2099         if (nilfs_test_metadata_dirty(nilfs, sci->sc_root))
2100                 set_bit(NILFS_SC_DIRTY, &sci->sc_flags);
2101
2102         if (nilfs_segctor_clean(sci))
2103                 goto out;
2104
2105         do {
2106                 sci->sc_stage.flags &= ~NILFS_CF_HISTORY_MASK;
2107
2108                 err = nilfs_segctor_begin_construction(sci, nilfs);
2109                 if (unlikely(err))
2110                         goto failed;
2111
2112                 /* Update time stamp */
2113                 sci->sc_seg_ctime = ktime_get_real_seconds();
2114
2115                 err = nilfs_segctor_collect(sci, nilfs, mode);
2116                 if (unlikely(err))
2117                         goto failed;
2118
2119                 /* Avoid empty segment */
2120                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE &&
2121                     nilfs_segbuf_empty(sci->sc_curseg)) {
2122                         nilfs_segctor_abort_construction(sci, nilfs, 1);
2123                         goto out;
2124                 }
2125
2126                 err = nilfs_segctor_assign(sci, mode);
2127                 if (unlikely(err))
2128                         goto failed;
2129
2130                 if (sci->sc_stage.flags & NILFS_CF_IFILE_STARTED)
2131                         nilfs_segctor_fill_in_file_bmap(sci);
2132
2133                 if (mode == SC_LSEG_SR &&
2134                     nilfs_sc_cstage_get(sci) >= NILFS_ST_CPFILE) {
2135                         err = nilfs_cpfile_finalize_checkpoint(
2136                                 nilfs->ns_cpfile, nilfs->ns_cno, sci->sc_root,
2137                                 sci->sc_nblk_inc + sci->sc_nblk_this_inc,
2138                                 sci->sc_seg_ctime,
2139                                 !test_bit(NILFS_SC_HAVE_DELTA, &sci->sc_flags));
2140                         if (unlikely(err))
2141                                 goto failed_to_write;
2142
2143                         nilfs_segctor_fill_in_super_root(sci, nilfs);
2144                 }
2145                 nilfs_segctor_update_segusage(sci, nilfs->ns_sufile);
2146
2147                 /* Write partial segments */
2148                 nilfs_prepare_write_logs(&sci->sc_segbufs, nilfs->ns_crc_seed);
2149
2150                 err = nilfs_segctor_write(sci, nilfs);
2151                 if (unlikely(err))
2152                         goto failed_to_write;
2153
2154                 if (nilfs_sc_cstage_get(sci) == NILFS_ST_DONE ||
2155                     nilfs->ns_blocksize_bits != PAGE_SHIFT) {
2156                         /*
2157                          * At this point, we avoid double buffering
2158                          * for blocksize < pagesize because page dirty
2159                          * flag is turned off during write and dirty
2160                          * buffers are not properly collected for
2161                          * pages crossing over segments.
2162                          */
2163                         err = nilfs_segctor_wait(sci);
2164                         if (err)
2165                                 goto failed_to_write;
2166                 }
2167         } while (nilfs_sc_cstage_get(sci) != NILFS_ST_DONE);
2168
2169  out:
2170         nilfs_segctor_drop_written_files(sci, nilfs);
2171         return err;
2172
2173  failed_to_write:
2174  failed:
2175         if (mode == SC_LSEG_SR && nilfs_sc_cstage_get(sci) >= NILFS_ST_IFILE)
2176                 nilfs_redirty_inodes(&sci->sc_dirty_files);
2177         if (nilfs_doing_gc())
2178                 nilfs_redirty_inodes(&sci->sc_gc_inodes);
2179         nilfs_segctor_abort_construction(sci, nilfs, err);
2180         goto out;
2181 }
2182
2183 /**
2184  * nilfs_segctor_start_timer - set timer of background write
2185  * @sci: nilfs_sc_info
2186  *
2187  * If the timer has already been set, it ignores the new request.
2188  * This function MUST be called within a section locking the segment
2189  * semaphore.
2190  */
2191 static void nilfs_segctor_start_timer(struct nilfs_sc_info *sci)
2192 {
2193         spin_lock(&sci->sc_state_lock);
2194         if (!(sci->sc_state & NILFS_SEGCTOR_COMMIT)) {
2195                 if (sci->sc_task) {
2196                         sci->sc_timer.expires = jiffies + sci->sc_interval;
2197                         add_timer(&sci->sc_timer);
2198                 }
2199                 sci->sc_state |= NILFS_SEGCTOR_COMMIT;
2200         }
2201         spin_unlock(&sci->sc_state_lock);
2202 }
2203
2204 static void nilfs_segctor_do_flush(struct nilfs_sc_info *sci, int bn)
2205 {
2206         spin_lock(&sci->sc_state_lock);
2207         if (!(sci->sc_flush_request & BIT(bn))) {
2208                 unsigned long prev_req = sci->sc_flush_request;
2209
2210                 sci->sc_flush_request |= BIT(bn);
2211                 if (!prev_req)
2212                         wake_up(&sci->sc_wait_daemon);
2213         }
2214         spin_unlock(&sci->sc_state_lock);
2215 }
2216
2217 /**
2218  * nilfs_flush_segment - trigger a segment construction for resource control
2219  * @sb: super block
2220  * @ino: inode number of the file to be flushed out.
2221  */
2222 void nilfs_flush_segment(struct super_block *sb, ino_t ino)
2223 {
2224         struct the_nilfs *nilfs = sb->s_fs_info;
2225         struct nilfs_sc_info *sci = nilfs->ns_writer;
2226
2227         if (!sci || nilfs_doing_construction())
2228                 return;
2229         nilfs_segctor_do_flush(sci, NILFS_MDT_INODE(sb, ino) ? ino : 0);
2230                                         /* assign bit 0 to data files */
2231 }
2232
2233 struct nilfs_segctor_wait_request {
2234         wait_queue_entry_t      wq;
2235         __u32           seq;
2236         int             err;
2237         atomic_t        done;
2238 };
2239
2240 static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
2241 {
2242         struct nilfs_segctor_wait_request wait_req;
2243         int err = 0;
2244
2245         init_wait(&wait_req.wq);
2246         wait_req.err = 0;
2247         atomic_set(&wait_req.done, 0);
2248         init_waitqueue_entry(&wait_req.wq, current);
2249
2250         /*
2251          * To prevent a race issue where completion notifications from the
2252          * log writer thread are missed, increment the request sequence count
2253          * "sc_seq_request" and insert a wait queue entry using the current
2254          * sequence number into the "sc_wait_request" queue at the same time
2255          * within the lock section of "sc_state_lock".
2256          */
2257         spin_lock(&sci->sc_state_lock);
2258         wait_req.seq = ++sci->sc_seq_request;
2259         add_wait_queue(&sci->sc_wait_request, &wait_req.wq);
2260         spin_unlock(&sci->sc_state_lock);
2261
2262         wake_up(&sci->sc_wait_daemon);
2263
2264         for (;;) {
2265                 set_current_state(TASK_INTERRUPTIBLE);
2266
2267                 /*
2268                  * Synchronize only while the log writer thread is alive.
2269                  * Leave flushing out after the log writer thread exits to
2270                  * the cleanup work in nilfs_segctor_destroy().
2271                  */
2272                 if (!sci->sc_task)
2273                         break;
2274
2275                 if (atomic_read(&wait_req.done)) {
2276                         err = wait_req.err;
2277                         break;
2278                 }
2279                 if (!signal_pending(current)) {
2280                         schedule();
2281                         continue;
2282                 }
2283                 err = -ERESTARTSYS;
2284                 break;
2285         }
2286         finish_wait(&sci->sc_wait_request, &wait_req.wq);
2287         return err;
2288 }
2289
2290 static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err, bool force)
2291 {
2292         struct nilfs_segctor_wait_request *wrq, *n;
2293         unsigned long flags;
2294
2295         spin_lock_irqsave(&sci->sc_wait_request.lock, flags);
2296         list_for_each_entry_safe(wrq, n, &sci->sc_wait_request.head, wq.entry) {
2297                 if (!atomic_read(&wrq->done) &&
2298                     (force || nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq))) {
2299                         wrq->err = err;
2300                         atomic_set(&wrq->done, 1);
2301                 }
2302                 if (atomic_read(&wrq->done)) {
2303                         wrq->wq.func(&wrq->wq,
2304                                      TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2305                                      0, NULL);
2306                 }
2307         }
2308         spin_unlock_irqrestore(&sci->sc_wait_request.lock, flags);
2309 }
2310
2311 /**
2312  * nilfs_construct_segment - construct a logical segment
2313  * @sb: super block
2314  *
2315  * Return Value: On success, 0 is returned. On errors, one of the following
2316  * negative error code is returned.
2317  *
2318  * %-EROFS - Read only filesystem.
2319  *
2320  * %-EIO - I/O error
2321  *
2322  * %-ENOSPC - No space left on device (only in a panic state).
2323  *
2324  * %-ERESTARTSYS - Interrupted.
2325  *
2326  * %-ENOMEM - Insufficient memory available.
2327  */
2328 int nilfs_construct_segment(struct super_block *sb)
2329 {
2330         struct the_nilfs *nilfs = sb->s_fs_info;
2331         struct nilfs_sc_info *sci = nilfs->ns_writer;
2332         struct nilfs_transaction_info *ti;
2333
2334         if (sb_rdonly(sb) || unlikely(!sci))
2335                 return -EROFS;
2336
2337         /* A call inside transactions causes a deadlock. */
2338         BUG_ON((ti = current->journal_info) && ti->ti_magic == NILFS_TI_MAGIC);
2339
2340         return nilfs_segctor_sync(sci);
2341 }
2342
2343 /**
2344  * nilfs_construct_dsync_segment - construct a data-only logical segment
2345  * @sb: super block
2346  * @inode: inode whose data blocks should be written out
2347  * @start: start byte offset
2348  * @end: end byte offset (inclusive)
2349  *
2350  * Return Value: On success, 0 is returned. On errors, one of the following
2351  * negative error code is returned.
2352  *
2353  * %-EROFS - Read only filesystem.
2354  *
2355  * %-EIO - I/O error
2356  *
2357  * %-ENOSPC - No space left on device (only in a panic state).
2358  *
2359  * %-ERESTARTSYS - Interrupted.
2360  *
2361  * %-ENOMEM - Insufficient memory available.
2362  */
2363 int nilfs_construct_dsync_segment(struct super_block *sb, struct inode *inode,
2364                                   loff_t start, loff_t end)
2365 {
2366         struct the_nilfs *nilfs = sb->s_fs_info;
2367         struct nilfs_sc_info *sci = nilfs->ns_writer;
2368         struct nilfs_inode_info *ii;
2369         struct nilfs_transaction_info ti;
2370         int err = 0;
2371
2372         if (sb_rdonly(sb) || unlikely(!sci))
2373                 return -EROFS;
2374
2375         nilfs_transaction_lock(sb, &ti, 0);
2376
2377         ii = NILFS_I(inode);
2378         if (test_bit(NILFS_I_INODE_SYNC, &ii->i_state) ||
2379             nilfs_test_opt(nilfs, STRICT_ORDER) ||
2380             test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2381             nilfs_discontinued(nilfs)) {
2382                 nilfs_transaction_unlock(sb);
2383                 err = nilfs_segctor_sync(sci);
2384                 return err;
2385         }
2386
2387         spin_lock(&nilfs->ns_inode_lock);
2388         if (!test_bit(NILFS_I_QUEUED, &ii->i_state) &&
2389             !test_bit(NILFS_I_BUSY, &ii->i_state)) {
2390                 spin_unlock(&nilfs->ns_inode_lock);
2391                 nilfs_transaction_unlock(sb);
2392                 return 0;
2393         }
2394         spin_unlock(&nilfs->ns_inode_lock);
2395         sci->sc_dsync_inode = ii;
2396         sci->sc_dsync_start = start;
2397         sci->sc_dsync_end = end;
2398
2399         err = nilfs_segctor_do_construct(sci, SC_LSEG_DSYNC);
2400         if (!err)
2401                 nilfs->ns_flushed_device = 0;
2402
2403         nilfs_transaction_unlock(sb);
2404         return err;
2405 }
2406
2407 #define FLUSH_FILE_BIT  (0x1) /* data file only */
2408 #define FLUSH_DAT_BIT   BIT(NILFS_DAT_INO) /* DAT only */
2409
2410 /**
2411  * nilfs_segctor_accept - record accepted sequence count of log-write requests
2412  * @sci: segment constructor object
2413  */
2414 static void nilfs_segctor_accept(struct nilfs_sc_info *sci)
2415 {
2416         bool thread_is_alive;
2417
2418         spin_lock(&sci->sc_state_lock);
2419         sci->sc_seq_accepted = sci->sc_seq_request;
2420         thread_is_alive = (bool)sci->sc_task;
2421         spin_unlock(&sci->sc_state_lock);
2422
2423         /*
2424          * This function does not race with the log writer thread's
2425          * termination.  Therefore, deleting sc_timer, which should not be
2426          * done after the log writer thread exits, can be done safely outside
2427          * the area protected by sc_state_lock.
2428          */
2429         if (thread_is_alive)
2430                 del_timer_sync(&sci->sc_timer);
2431 }
2432
2433 /**
2434  * nilfs_segctor_notify - notify the result of request to caller threads
2435  * @sci: segment constructor object
2436  * @mode: mode of log forming
2437  * @err: error code to be notified
2438  */
2439 static void nilfs_segctor_notify(struct nilfs_sc_info *sci, int mode, int err)
2440 {
2441         /* Clear requests (even when the construction failed) */
2442         spin_lock(&sci->sc_state_lock);
2443
2444         if (mode == SC_LSEG_SR) {
2445                 sci->sc_state &= ~NILFS_SEGCTOR_COMMIT;
2446                 sci->sc_seq_done = sci->sc_seq_accepted;
2447                 nilfs_segctor_wakeup(sci, err, false);
2448                 sci->sc_flush_request = 0;
2449         } else {
2450                 if (mode == SC_FLUSH_FILE)
2451                         sci->sc_flush_request &= ~FLUSH_FILE_BIT;
2452                 else if (mode == SC_FLUSH_DAT)
2453                         sci->sc_flush_request &= ~FLUSH_DAT_BIT;
2454
2455                 /* re-enable timer if checkpoint creation was not done */
2456                 if ((sci->sc_state & NILFS_SEGCTOR_COMMIT) && sci->sc_task &&
2457                     time_before(jiffies, sci->sc_timer.expires))
2458                         add_timer(&sci->sc_timer);
2459         }
2460         spin_unlock(&sci->sc_state_lock);
2461 }
2462
2463 /**
2464  * nilfs_segctor_construct - form logs and write them to disk
2465  * @sci: segment constructor object
2466  * @mode: mode of log forming
2467  */
2468 static int nilfs_segctor_construct(struct nilfs_sc_info *sci, int mode)
2469 {
2470         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2471         struct nilfs_super_block **sbp;
2472         int err = 0;
2473
2474         nilfs_segctor_accept(sci);
2475
2476         if (nilfs_discontinued(nilfs))
2477                 mode = SC_LSEG_SR;
2478         if (!nilfs_segctor_confirm(sci))
2479                 err = nilfs_segctor_do_construct(sci, mode);
2480
2481         if (likely(!err)) {
2482                 if (mode != SC_FLUSH_DAT)
2483                         atomic_set(&nilfs->ns_ndirtyblks, 0);
2484                 if (test_bit(NILFS_SC_SUPER_ROOT, &sci->sc_flags) &&
2485                     nilfs_discontinued(nilfs)) {
2486                         down_write(&nilfs->ns_sem);
2487                         err = -EIO;
2488                         sbp = nilfs_prepare_super(sci->sc_super,
2489                                                   nilfs_sb_will_flip(nilfs));
2490                         if (likely(sbp)) {
2491                                 nilfs_set_log_cursor(sbp[0], nilfs);
2492                                 err = nilfs_commit_super(sci->sc_super,
2493                                                          NILFS_SB_COMMIT);
2494                         }
2495                         up_write(&nilfs->ns_sem);
2496                 }
2497         }
2498
2499         nilfs_segctor_notify(sci, mode, err);
2500         return err;
2501 }
2502
2503 static void nilfs_construction_timeout(struct timer_list *t)
2504 {
2505         struct nilfs_sc_info *sci = from_timer(sci, t, sc_timer);
2506
2507         wake_up_process(sci->sc_task);
2508 }
2509
2510 static void
2511 nilfs_remove_written_gcinodes(struct the_nilfs *nilfs, struct list_head *head)
2512 {
2513         struct nilfs_inode_info *ii, *n;
2514
2515         list_for_each_entry_safe(ii, n, head, i_dirty) {
2516                 if (!test_bit(NILFS_I_UPDATED, &ii->i_state))
2517                         continue;
2518                 list_del_init(&ii->i_dirty);
2519                 truncate_inode_pages(&ii->vfs_inode.i_data, 0);
2520                 nilfs_btnode_cache_clear(ii->i_assoc_inode->i_mapping);
2521                 iput(&ii->vfs_inode);
2522         }
2523 }
2524
2525 int nilfs_clean_segments(struct super_block *sb, struct nilfs_argv *argv,
2526                          void **kbufs)
2527 {
2528         struct the_nilfs *nilfs = sb->s_fs_info;
2529         struct nilfs_sc_info *sci = nilfs->ns_writer;
2530         struct nilfs_transaction_info ti;
2531         int err;
2532
2533         if (unlikely(!sci))
2534                 return -EROFS;
2535
2536         nilfs_transaction_lock(sb, &ti, 1);
2537
2538         err = nilfs_mdt_save_to_shadow_map(nilfs->ns_dat);
2539         if (unlikely(err))
2540                 goto out_unlock;
2541
2542         err = nilfs_ioctl_prepare_clean_segments(nilfs, argv, kbufs);
2543         if (unlikely(err)) {
2544                 nilfs_mdt_restore_from_shadow_map(nilfs->ns_dat);
2545                 goto out_unlock;
2546         }
2547
2548         sci->sc_freesegs = kbufs[4];
2549         sci->sc_nfreesegs = argv[4].v_nmembs;
2550         list_splice_tail_init(&nilfs->ns_gc_inodes, &sci->sc_gc_inodes);
2551
2552         for (;;) {
2553                 err = nilfs_segctor_construct(sci, SC_LSEG_SR);
2554                 nilfs_remove_written_gcinodes(nilfs, &sci->sc_gc_inodes);
2555
2556                 if (likely(!err))
2557                         break;
2558
2559                 nilfs_warn(sb, "error %d cleaning segments", err);
2560                 set_current_state(TASK_INTERRUPTIBLE);
2561                 schedule_timeout(sci->sc_interval);
2562         }
2563         if (nilfs_test_opt(nilfs, DISCARD)) {
2564                 int ret = nilfs_discard_segments(nilfs, sci->sc_freesegs,
2565                                                  sci->sc_nfreesegs);
2566                 if (ret) {
2567                         nilfs_warn(sb,
2568                                    "error %d on discard request, turning discards off for the device",
2569                                    ret);
2570                         nilfs_clear_opt(nilfs, DISCARD);
2571                 }
2572         }
2573
2574  out_unlock:
2575         sci->sc_freesegs = NULL;
2576         sci->sc_nfreesegs = 0;
2577         nilfs_mdt_clear_shadow_map(nilfs->ns_dat);
2578         nilfs_transaction_unlock(sb);
2579         return err;
2580 }
2581
2582 static void nilfs_segctor_thread_construct(struct nilfs_sc_info *sci, int mode)
2583 {
2584         struct nilfs_transaction_info ti;
2585
2586         nilfs_transaction_lock(sci->sc_super, &ti, 0);
2587         nilfs_segctor_construct(sci, mode);
2588
2589         /*
2590          * Unclosed segment should be retried.  We do this using sc_timer.
2591          * Timeout of sc_timer will invoke complete construction which leads
2592          * to close the current logical segment.
2593          */
2594         if (test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags))
2595                 nilfs_segctor_start_timer(sci);
2596
2597         nilfs_transaction_unlock(sci->sc_super);
2598 }
2599
2600 static void nilfs_segctor_do_immediate_flush(struct nilfs_sc_info *sci)
2601 {
2602         int mode = 0;
2603
2604         spin_lock(&sci->sc_state_lock);
2605         mode = (sci->sc_flush_request & FLUSH_DAT_BIT) ?
2606                 SC_FLUSH_DAT : SC_FLUSH_FILE;
2607         spin_unlock(&sci->sc_state_lock);
2608
2609         if (mode) {
2610                 nilfs_segctor_do_construct(sci, mode);
2611
2612                 spin_lock(&sci->sc_state_lock);
2613                 sci->sc_flush_request &= (mode == SC_FLUSH_FILE) ?
2614                         ~FLUSH_FILE_BIT : ~FLUSH_DAT_BIT;
2615                 spin_unlock(&sci->sc_state_lock);
2616         }
2617         clear_bit(NILFS_SC_PRIOR_FLUSH, &sci->sc_flags);
2618 }
2619
2620 static int nilfs_segctor_flush_mode(struct nilfs_sc_info *sci)
2621 {
2622         if (!test_bit(NILFS_SC_UNCLOSED, &sci->sc_flags) ||
2623             time_before(jiffies, sci->sc_lseg_stime + sci->sc_mjcp_freq)) {
2624                 if (!(sci->sc_flush_request & ~FLUSH_FILE_BIT))
2625                         return SC_FLUSH_FILE;
2626                 else if (!(sci->sc_flush_request & ~FLUSH_DAT_BIT))
2627                         return SC_FLUSH_DAT;
2628         }
2629         return SC_LSEG_SR;
2630 }
2631
2632 /**
2633  * nilfs_log_write_required - determine whether log writing is required
2634  * @sci:   nilfs_sc_info struct
2635  * @modep: location for storing log writing mode
2636  *
2637  * Return: true if log writing is required, false otherwise.  If log writing
2638  * is required, the mode is stored in the location pointed to by @modep.
2639  */
2640 static bool nilfs_log_write_required(struct nilfs_sc_info *sci, int *modep)
2641 {
2642         bool timedout, ret = true;
2643
2644         spin_lock(&sci->sc_state_lock);
2645         timedout = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) &&
2646                    time_after_eq(jiffies, sci->sc_timer.expires));
2647         if (timedout || sci->sc_seq_request != sci->sc_seq_done)
2648                 *modep = SC_LSEG_SR;
2649         else if (sci->sc_flush_request)
2650                 *modep = nilfs_segctor_flush_mode(sci);
2651         else
2652                 ret = false;
2653
2654         spin_unlock(&sci->sc_state_lock);
2655         return ret;
2656 }
2657
2658 /**
2659  * nilfs_segctor_thread - main loop of the log writer thread
2660  * @arg: pointer to a struct nilfs_sc_info.
2661  *
2662  * nilfs_segctor_thread() is the main loop function of the log writer kernel
2663  * thread, which determines whether log writing is necessary, and if so,
2664  * performs the log write in the background, or waits if not.  It is also
2665  * used to decide the background writeback of the superblock.
2666  *
2667  * Return: Always 0.
2668  */
2669 static int nilfs_segctor_thread(void *arg)
2670 {
2671         struct nilfs_sc_info *sci = (struct nilfs_sc_info *)arg;
2672         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2673
2674         nilfs_info(sci->sc_super,
2675                    "segctord starting. Construction interval = %lu seconds, CP frequency < %lu seconds",
2676                    sci->sc_interval / HZ, sci->sc_mjcp_freq / HZ);
2677
2678         set_freezable();
2679
2680         while (!kthread_should_stop()) {
2681                 DEFINE_WAIT(wait);
2682                 bool should_write;
2683                 int mode;
2684
2685                 if (freezing(current)) {
2686                         try_to_freeze();
2687                         continue;
2688                 }
2689
2690                 prepare_to_wait(&sci->sc_wait_daemon, &wait,
2691                                 TASK_INTERRUPTIBLE);
2692                 should_write = nilfs_log_write_required(sci, &mode);
2693                 if (!should_write)
2694                         schedule();
2695                 finish_wait(&sci->sc_wait_daemon, &wait);
2696
2697                 if (nilfs_sb_dirty(nilfs) && nilfs_sb_need_update(nilfs))
2698                         set_nilfs_discontinued(nilfs);
2699
2700                 if (should_write)
2701                         nilfs_segctor_thread_construct(sci, mode);
2702         }
2703
2704         /* end sync. */
2705         spin_lock(&sci->sc_state_lock);
2706         sci->sc_task = NULL;
2707         timer_shutdown_sync(&sci->sc_timer);
2708         spin_unlock(&sci->sc_state_lock);
2709         return 0;
2710 }
2711
2712 /*
2713  * Setup & clean-up functions
2714  */
2715 static struct nilfs_sc_info *nilfs_segctor_new(struct super_block *sb,
2716                                                struct nilfs_root *root)
2717 {
2718         struct the_nilfs *nilfs = sb->s_fs_info;
2719         struct nilfs_sc_info *sci;
2720
2721         sci = kzalloc(sizeof(*sci), GFP_KERNEL);
2722         if (!sci)
2723                 return NULL;
2724
2725         sci->sc_super = sb;
2726
2727         nilfs_get_root(root);
2728         sci->sc_root = root;
2729
2730         init_waitqueue_head(&sci->sc_wait_request);
2731         init_waitqueue_head(&sci->sc_wait_daemon);
2732         spin_lock_init(&sci->sc_state_lock);
2733         INIT_LIST_HEAD(&sci->sc_dirty_files);
2734         INIT_LIST_HEAD(&sci->sc_segbufs);
2735         INIT_LIST_HEAD(&sci->sc_write_logs);
2736         INIT_LIST_HEAD(&sci->sc_gc_inodes);
2737         INIT_LIST_HEAD(&sci->sc_iput_queue);
2738         INIT_WORK(&sci->sc_iput_work, nilfs_iput_work_func);
2739
2740         sci->sc_interval = HZ * NILFS_SC_DEFAULT_TIMEOUT;
2741         sci->sc_mjcp_freq = HZ * NILFS_SC_DEFAULT_SR_FREQ;
2742         sci->sc_watermark = NILFS_SC_DEFAULT_WATERMARK;
2743
2744         if (nilfs->ns_interval)
2745                 sci->sc_interval = HZ * nilfs->ns_interval;
2746         if (nilfs->ns_watermark)
2747                 sci->sc_watermark = nilfs->ns_watermark;
2748         return sci;
2749 }
2750
2751 static void nilfs_segctor_write_out(struct nilfs_sc_info *sci)
2752 {
2753         int ret, retrycount = NILFS_SC_CLEANUP_RETRY;
2754
2755         /*
2756          * The segctord thread was stopped and its timer was removed.
2757          * But some tasks remain.
2758          */
2759         do {
2760                 struct nilfs_transaction_info ti;
2761
2762                 nilfs_transaction_lock(sci->sc_super, &ti, 0);
2763                 ret = nilfs_segctor_construct(sci, SC_LSEG_SR);
2764                 nilfs_transaction_unlock(sci->sc_super);
2765
2766                 flush_work(&sci->sc_iput_work);
2767
2768         } while (ret && ret != -EROFS && retrycount-- > 0);
2769 }
2770
2771 /**
2772  * nilfs_segctor_destroy - destroy the segment constructor.
2773  * @sci: nilfs_sc_info
2774  *
2775  * nilfs_segctor_destroy() kills the segctord thread and frees
2776  * the nilfs_sc_info struct.
2777  * Caller must hold the segment semaphore.
2778  */
2779 static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
2780 {
2781         struct the_nilfs *nilfs = sci->sc_super->s_fs_info;
2782         int flag;
2783
2784         up_write(&nilfs->ns_segctor_sem);
2785
2786         if (sci->sc_task) {
2787                 wake_up(&sci->sc_wait_daemon);
2788                 kthread_stop(sci->sc_task);
2789         }
2790
2791         spin_lock(&sci->sc_state_lock);
2792         flag = ((sci->sc_state & NILFS_SEGCTOR_COMMIT) || sci->sc_flush_request
2793                 || sci->sc_seq_request != sci->sc_seq_done);
2794         spin_unlock(&sci->sc_state_lock);
2795
2796         /*
2797          * Forcibly wake up tasks waiting in nilfs_segctor_sync(), which can
2798          * be called from delayed iput() via nilfs_evict_inode() and can race
2799          * with the above log writer thread termination.
2800          */
2801         nilfs_segctor_wakeup(sci, 0, true);
2802
2803         if (flush_work(&sci->sc_iput_work))
2804                 flag = true;
2805
2806         if (flag || !nilfs_segctor_confirm(sci))
2807                 nilfs_segctor_write_out(sci);
2808
2809         if (!list_empty(&sci->sc_dirty_files)) {
2810                 nilfs_warn(sci->sc_super,
2811                            "disposed unprocessed dirty file(s) when stopping log writer");
2812                 nilfs_dispose_list(nilfs, &sci->sc_dirty_files, 1);
2813         }
2814
2815         if (!list_empty(&sci->sc_iput_queue)) {
2816                 nilfs_warn(sci->sc_super,
2817                            "disposed unprocessed inode(s) in iput queue when stopping log writer");
2818                 nilfs_dispose_list(nilfs, &sci->sc_iput_queue, 1);
2819         }
2820
2821         WARN_ON(!list_empty(&sci->sc_segbufs));
2822         WARN_ON(!list_empty(&sci->sc_write_logs));
2823
2824         nilfs_put_root(sci->sc_root);
2825
2826         down_write(&nilfs->ns_segctor_sem);
2827
2828         kfree(sci);
2829 }
2830
2831 /**
2832  * nilfs_attach_log_writer - attach log writer
2833  * @sb: super block instance
2834  * @root: root object of the current filesystem tree
2835  *
2836  * This allocates a log writer object, initializes it, and starts the
2837  * log writer.
2838  *
2839  * Return: 0 on success, or the following negative error code on failure.
2840  * * %-EINTR    - Log writer thread creation failed due to interruption.
2841  * * %-ENOMEM   - Insufficient memory available.
2842  */
2843 int nilfs_attach_log_writer(struct super_block *sb, struct nilfs_root *root)
2844 {
2845         struct the_nilfs *nilfs = sb->s_fs_info;
2846         struct nilfs_sc_info *sci;
2847         struct task_struct *t;
2848         int err;
2849
2850         if (nilfs->ns_writer) {
2851                 /*
2852                  * This happens if the filesystem is made read-only by
2853                  * __nilfs_error or nilfs_remount and then remounted
2854                  * read/write.  In these cases, reuse the existing
2855                  * writer.
2856                  */
2857                 return 0;
2858         }
2859
2860         sci = nilfs_segctor_new(sb, root);
2861         if (unlikely(!sci))
2862                 return -ENOMEM;
2863
2864         nilfs->ns_writer = sci;
2865         t = kthread_create(nilfs_segctor_thread, sci, "segctord");
2866         if (IS_ERR(t)) {
2867                 err = PTR_ERR(t);
2868                 nilfs_err(sb, "error %d creating segctord thread", err);
2869                 nilfs_detach_log_writer(sb);
2870                 return err;
2871         }
2872         sci->sc_task = t;
2873         timer_setup(&sci->sc_timer, nilfs_construction_timeout, 0);
2874
2875         wake_up_process(sci->sc_task);
2876         return 0;
2877 }
2878
2879 /**
2880  * nilfs_detach_log_writer - destroy log writer
2881  * @sb: super block instance
2882  *
2883  * This kills log writer daemon, frees the log writer object, and
2884  * destroys list of dirty files.
2885  */
2886 void nilfs_detach_log_writer(struct super_block *sb)
2887 {
2888         struct the_nilfs *nilfs = sb->s_fs_info;
2889         LIST_HEAD(garbage_list);
2890
2891         down_write(&nilfs->ns_segctor_sem);
2892         if (nilfs->ns_writer) {
2893                 nilfs_segctor_destroy(nilfs->ns_writer);
2894                 nilfs->ns_writer = NULL;
2895         }
2896         set_nilfs_purging(nilfs);
2897
2898         /* Force to free the list of dirty files */
2899         spin_lock(&nilfs->ns_inode_lock);
2900         if (!list_empty(&nilfs->ns_dirty_files)) {
2901                 list_splice_init(&nilfs->ns_dirty_files, &garbage_list);
2902                 nilfs_warn(sb,
2903                            "disposed unprocessed dirty file(s) when detaching log writer");
2904         }
2905         spin_unlock(&nilfs->ns_inode_lock);
2906         up_write(&nilfs->ns_segctor_sem);
2907
2908         nilfs_dispose_list(nilfs, &garbage_list, 1);
2909         clear_nilfs_purging(nilfs);
2910 }