fs/nilfs2/mdt.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * mdt.c - meta data file for NILFS
   4  *
   5  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   6  *
   7  * Written by Ryusuke Konishi.
   8  */
   9
  10 #include <linux/buffer_head.h>
  11 #include <linux/mpage.h>
  12 #include <linux/mm.h>
  13 #include <linux/writeback.h>
  14 #include <linux/backing-dev.h>
  15 #include <linux/swap.h>
  16 #include <linux/slab.h>
  17 #include "nilfs.h"
  18 #include "btnode.h"
  19 #include "segment.h"
  20 #include "page.h"
  21 #include "mdt.h"
  22 #include "alloc.h"              /* nilfs_palloc_destroy_cache() */
  23
  24 #include <trace/events/nilfs2.h>
  25
  26 #define NILFS_MDT_MAX_RA_BLOCKS         (16 - 1)
  27
  28
  29 static int
  30 nilfs_mdt_insert_new_block(struct inode *inode, unsigned long block,
  31                            struct buffer_head *bh,
  32                            void (*init_block)(struct inode *,
  33                                               struct buffer_head *, void *))
  34 {
  35         struct nilfs_inode_info *ii = NILFS_I(inode);
  36         void *kaddr;
  37         int ret;
  38
  39         /* Caller exclude read accesses using page lock */
  40
  41         /* set_buffer_new(bh); */
  42         bh->b_blocknr = 0;
  43
  44         ret = nilfs_bmap_insert(ii->i_bmap, block, (unsigned long)bh);
  45         if (unlikely(ret))
  46                 return ret;
  47
  48         set_buffer_mapped(bh);
  49
  50         kaddr = kmap_atomic(bh->b_page);
  51         memset(kaddr + bh_offset(bh), 0, i_blocksize(inode));
  52         if (init_block)
  53                 init_block(inode, bh, kaddr);
  54         flush_dcache_page(bh->b_page);
  55         kunmap_atomic(kaddr);
  56
  57         set_buffer_uptodate(bh);
  58         mark_buffer_dirty(bh);
  59         nilfs_mdt_mark_dirty(inode);
  60
  61         trace_nilfs2_mdt_insert_new_block(inode, inode->i_ino, block);
  62
  63         return 0;
  64 }
  65
  66 static int nilfs_mdt_create_block(struct inode *inode, unsigned long block,
  67                                   struct buffer_head **out_bh,
  68                                   void (*init_block)(struct inode *,
  69                                                      struct buffer_head *,
  70                                                      void *))
  71 {
  72         struct super_block *sb = inode->i_sb;
  73         struct nilfs_transaction_info ti;
  74         struct buffer_head *bh;
  75         int err;
  76
  77         nilfs_transaction_begin(sb, &ti, 0);
  78
  79         err = -ENOMEM;
  80         bh = nilfs_grab_buffer(inode, inode->i_mapping, block, 0);
  81         if (unlikely(!bh))
  82                 goto failed_unlock;
  83
  84         err = -EEXIST;
  85         if (buffer_uptodate(bh))
  86                 goto failed_bh;
  87
  88         wait_on_buffer(bh);
  89         if (buffer_uptodate(bh))
  90                 goto failed_bh;
  91
  92         bh->b_bdev = sb->s_bdev;
  93         err = nilfs_mdt_insert_new_block(inode, block, bh, init_block);
  94         if (likely(!err)) {
  95                 get_bh(bh);
  96                 *out_bh = bh;
  97         }
  98
  99  failed_bh:
 100         unlock_page(bh->b_page);
 101         put_page(bh->b_page);
 102         brelse(bh);
 103
 104  failed_unlock:
 105         if (likely(!err))
 106                 err = nilfs_transaction_commit(sb);
 107         else
 108                 nilfs_transaction_abort(sb);
 109
 110         return err;
 111 }
 112
 113 static int
 114 nilfs_mdt_submit_block(struct inode *inode, unsigned long blkoff,
 115                        int mode, int mode_flags, struct buffer_head **out_bh)
 116 {
 117         struct buffer_head *bh;
 118         __u64 blknum = 0;
 119         int ret = -ENOMEM;
 120
 121         bh = nilfs_grab_buffer(inode, inode->i_mapping, blkoff, 0);
 122         if (unlikely(!bh))
 123                 goto failed;
 124
 125         ret = -EEXIST; /* internal code */
 126         if (buffer_uptodate(bh))
 127                 goto out;
 128
 129         if (mode_flags & REQ_RAHEAD) {
 130                 if (!trylock_buffer(bh)) {
 131                         ret = -EBUSY;
 132                         goto failed_bh;
 133                 }
 134         } else /* mode == READ */
 135                 lock_buffer(bh);
 136
 137         if (buffer_uptodate(bh)) {
 138                 unlock_buffer(bh);
 139                 goto out;
 140         }
 141
 142         ret = nilfs_bmap_lookup(NILFS_I(inode)->i_bmap, blkoff, &blknum);
 143         if (unlikely(ret)) {
 144                 unlock_buffer(bh);
 145                 goto failed_bh;
 146         }
 147         map_bh(bh, inode->i_sb, (sector_t)blknum);
 148
 149         bh->b_end_io = end_buffer_read_sync;
 150         get_bh(bh);
 151         submit_bh(mode, mode_flags, bh);
 152         ret = 0;
 153
 154         trace_nilfs2_mdt_submit_block(inode, inode->i_ino, blkoff, mode);
 155  out:
 156         get_bh(bh);
 157         *out_bh = bh;
 158
 159  failed_bh:
 160         unlock_page(bh->b_page);
 161         put_page(bh->b_page);
 162         brelse(bh);
 163  failed:
 164         return ret;
 165 }
 166
 167 static int nilfs_mdt_read_block(struct inode *inode, unsigned long block,
 168                                 int readahead, struct buffer_head **out_bh)
 169 {
 170         struct buffer_head *first_bh, *bh;
 171         unsigned long blkoff;
 172         int i, nr_ra_blocks = NILFS_MDT_MAX_RA_BLOCKS;
 173         int err;
 174
 175         err = nilfs_mdt_submit_block(inode, block, REQ_OP_READ, 0, &first_bh);
 176         if (err == -EEXIST) /* internal code */
 177                 goto out;
 178
 179         if (unlikely(err))
 180                 goto failed;
 181
 182         if (readahead) {
 183                 blkoff = block + 1;
 184                 for (i = 0; i < nr_ra_blocks; i++, blkoff++) {
 185                         err = nilfs_mdt_submit_block(inode, blkoff, REQ_OP_READ,
 186                                                      REQ_RAHEAD, &bh);
 187                         if (likely(!err || err == -EEXIST))
 188                                 brelse(bh);
 189                         else if (err != -EBUSY)
 190                                 break;
 191                                 /* abort readahead if bmap lookup failed */
 192                         if (!buffer_locked(first_bh))
 193                                 goto out_no_wait;
 194                 }
 195         }
 196
 197         wait_on_buffer(first_bh);
 198
 199  out_no_wait:
 200         err = -EIO;
 201         if (!buffer_uptodate(first_bh)) {
 202                 nilfs_msg(inode->i_sb, KERN_ERR,
 203                           "I/O error reading meta-data file (ino=%lu, block-offset=%lu)",
 204                           inode->i_ino, block);
 205                 goto failed_bh;
 206         }
 207  out:
 208         *out_bh = first_bh;
 209         return 0;
 210
 211  failed_bh:
 212         brelse(first_bh);
 213  failed:
 214         return err;
 215 }
 216
 217 /**
 218  * nilfs_mdt_get_block - read or create a buffer on meta data file.
 219  * @inode: inode of the meta data file
 220  * @blkoff: block offset
 221  * @create: create flag
 222  * @init_block: initializer used for newly allocated block
 223  * @out_bh: output of a pointer to the buffer_head
 224  *
 225  * nilfs_mdt_get_block() looks up the specified buffer and tries to create
 226  * a new buffer if @create is not zero.  On success, the returned buffer is
 227  * assured to be either existing or formatted using a buffer lock on success.
 228  * @out_bh is substituted only when zero is returned.
 229  *
 230  * Return Value: On success, it returns 0. On error, the following negative
 231  * error code is returned.
 232  *
 233  * %-ENOMEM - Insufficient memory available.
 234  *
 235  * %-EIO - I/O error
 236  *
 237  * %-ENOENT - the specified block does not exist (hole block)
 238  *
 239  * %-EROFS - Read only filesystem (for create mode)
 240  */
 241 int nilfs_mdt_get_block(struct inode *inode, unsigned long blkoff, int create,
 242                         void (*init_block)(struct inode *,
 243                                            struct buffer_head *, void *),
 244                         struct buffer_head **out_bh)
 245 {
 246         int ret;
 247
 248         /* Should be rewritten with merging nilfs_mdt_read_block() */
 249  retry:
 250         ret = nilfs_mdt_read_block(inode, blkoff, !create, out_bh);
 251         if (!create || ret != -ENOENT)
 252                 return ret;
 253
 254         ret = nilfs_mdt_create_block(inode, blkoff, out_bh, init_block);
 255         if (unlikely(ret == -EEXIST)) {
 256                 /* create = 0; */  /* limit read-create loop retries */
 257                 goto retry;
 258         }
 259         return ret;
 260 }
 261
 262 /**
 263  * nilfs_mdt_find_block - find and get a buffer on meta data file.
 264  * @inode: inode of the meta data file
 265  * @start: start block offset (inclusive)
 266  * @end: end block offset (inclusive)
 267  * @blkoff: block offset
 268  * @out_bh: place to store a pointer to buffer_head struct
 269  *
 270  * nilfs_mdt_find_block() looks up an existing block in range of
 271  * [@start, @end] and stores pointer to a buffer head of the block to
 272  * @out_bh, and block offset to @blkoff, respectively.  @out_bh and
 273  * @blkoff are substituted only when zero is returned.
 274  *
 275  * Return Value: On success, it returns 0. On error, the following negative
 276  * error code is returned.
 277  *
 278  * %-ENOMEM - Insufficient memory available.
 279  *
 280  * %-EIO - I/O error
 281  *
 282  * %-ENOENT - no block was found in the range
 283  */
 284 int nilfs_mdt_find_block(struct inode *inode, unsigned long start,
 285                          unsigned long end, unsigned long *blkoff,
 286                          struct buffer_head **out_bh)
 287 {
 288         __u64 next;
 289         int ret;
 290
 291         if (unlikely(start > end))
 292                 return -ENOENT;
 293
 294         ret = nilfs_mdt_read_block(inode, start, true, out_bh);
 295         if (!ret) {
 296                 *blkoff = start;
 297                 goto out;
 298         }
 299         if (unlikely(ret != -ENOENT || start == ULONG_MAX))
 300                 goto out;
 301
 302         ret = nilfs_bmap_seek_key(NILFS_I(inode)->i_bmap, start + 1, &next);
 303         if (!ret) {
 304                 if (next <= end) {
 305                         ret = nilfs_mdt_read_block(inode, next, true, out_bh);
 306                         if (!ret)
 307                                 *blkoff = next;
 308                 } else {
 309                         ret = -ENOENT;
 310                 }
 311         }
 312 out:
 313         return ret;
 314 }
 315
 316 /**
 317  * nilfs_mdt_delete_block - make a hole on the meta data file.
 318  * @inode: inode of the meta data file
 319  * @block: block offset
 320  *
 321  * Return Value: On success, zero is returned.
 322  * On error, one of the following negative error code is returned.
 323  *
 324  * %-ENOMEM - Insufficient memory available.
 325  *
 326  * %-EIO - I/O error
 327  */
 328 int nilfs_mdt_delete_block(struct inode *inode, unsigned long block)
 329 {
 330         struct nilfs_inode_info *ii = NILFS_I(inode);
 331         int err;
 332
 333         err = nilfs_bmap_delete(ii->i_bmap, block);
 334         if (!err || err == -ENOENT) {
 335                 nilfs_mdt_mark_dirty(inode);
 336                 nilfs_mdt_forget_block(inode, block);
 337         }
 338         return err;
 339 }
 340
 341 /**
 342  * nilfs_mdt_forget_block - discard dirty state and try to remove the page
 343  * @inode: inode of the meta data file
 344  * @block: block offset
 345  *
 346  * nilfs_mdt_forget_block() clears a dirty flag of the specified buffer, and
 347  * tries to release the page including the buffer from a page cache.
 348  *
 349  * Return Value: On success, 0 is returned. On error, one of the following
 350  * negative error code is returned.
 351  *
 352  * %-EBUSY - page has an active buffer.
 353  *
 354  * %-ENOENT - page cache has no page addressed by the offset.
 355  */
 356 int nilfs_mdt_forget_block(struct inode *inode, unsigned long block)
 357 {
 358         pgoff_t index = (pgoff_t)block >>
 359                 (PAGE_SHIFT - inode->i_blkbits);
 360         struct page *page;
 361         unsigned long first_block;
 362         int ret = 0;
 363         int still_dirty;
 364
 365         page = find_lock_page(inode->i_mapping, index);
 366         if (!page)
 367                 return -ENOENT;
 368
 369         wait_on_page_writeback(page);
 370
 371         first_block = (unsigned long)index <<
 372                 (PAGE_SHIFT - inode->i_blkbits);
 373         if (page_has_buffers(page)) {
 374                 struct buffer_head *bh;
 375
 376                 bh = nilfs_page_get_nth_block(page, block - first_block);
 377                 nilfs_forget_buffer(bh);
 378         }
 379         still_dirty = PageDirty(page);
 380         unlock_page(page);
 381         put_page(page);
 382
 383         if (still_dirty ||
 384             invalidate_inode_pages2_range(inode->i_mapping, index, index) != 0)
 385                 ret = -EBUSY;
 386         return ret;
 387 }
 388
 389 int nilfs_mdt_fetch_dirty(struct inode *inode)
 390 {
 391         struct nilfs_inode_info *ii = NILFS_I(inode);
 392
 393         if (nilfs_bmap_test_and_clear_dirty(ii->i_bmap)) {
 394                 set_bit(NILFS_I_DIRTY, &ii->i_state);
 395                 return 1;
 396         }
 397         return test_bit(NILFS_I_DIRTY, &ii->i_state);
 398 }
 399
 400 static int
 401 nilfs_mdt_write_page(struct page *page, struct writeback_control *wbc)
 402 {
 403         struct inode *inode = page->mapping->host;
 404         struct super_block *sb;
 405         int err = 0;
 406
 407         if (inode && sb_rdonly(inode->i_sb)) {
 408                 /*
 409                  * It means that filesystem was remounted in read-only
 410                  * mode because of error or metadata corruption. But we
 411                  * have dirty pages that try to be flushed in background.
 412                  * So, here we simply discard this dirty page.
 413                  */
 414                 nilfs_clear_dirty_page(page, false);
 415                 unlock_page(page);
 416                 return -EROFS;
 417         }
 418
 419         redirty_page_for_writepage(wbc, page);
 420         unlock_page(page);
 421
 422         if (!inode)
 423                 return 0;
 424
 425         sb = inode->i_sb;
 426
 427         if (wbc->sync_mode == WB_SYNC_ALL)
 428                 err = nilfs_construct_segment(sb);
 429         else if (wbc->for_reclaim)
 430                 nilfs_flush_segment(sb, inode->i_ino);
 431
 432         return err;
 433 }
 434
 435
 436 static const struct address_space_operations def_mdt_aops = {
 437         .writepage              = nilfs_mdt_write_page,
 438 };
 439
 440 static const struct inode_operations def_mdt_iops;
 441 static const struct file_operations def_mdt_fops;
 442
 443
 444 int nilfs_mdt_init(struct inode *inode, gfp_t gfp_mask, size_t objsz)
 445 {
 446         struct nilfs_mdt_info *mi;
 447
 448         mi = kzalloc(max(sizeof(*mi), objsz), GFP_NOFS);
 449         if (!mi)
 450                 return -ENOMEM;
 451
 452         init_rwsem(&mi->mi_sem);
 453         inode->i_private = mi;
 454
 455         inode->i_mode = S_IFREG;
 456         mapping_set_gfp_mask(inode->i_mapping, gfp_mask);
 457
 458         inode->i_op = &def_mdt_iops;
 459         inode->i_fop = &def_mdt_fops;
 460         inode->i_mapping->a_ops = &def_mdt_aops;
 461
 462         return 0;
 463 }
 464
 465 /**
 466  * nilfs_mdt_clear - do cleanup for the metadata file
 467  * @inode: inode of the metadata file
 468  */
 469 void nilfs_mdt_clear(struct inode *inode)
 470 {
 471         struct nilfs_mdt_info *mdi = NILFS_MDT(inode);
 472
 473         if (mdi->mi_palloc_cache)
 474                 nilfs_palloc_destroy_cache(inode);
 475 }
 476
 477 /**
 478  * nilfs_mdt_destroy - release resources used by the metadata file
 479  * @inode: inode of the metadata file
 480  */
 481 void nilfs_mdt_destroy(struct inode *inode)
 482 {
 483         struct nilfs_mdt_info *mdi = NILFS_MDT(inode);
 484
 485         kfree(mdi->mi_bgl); /* kfree(NULL) is safe */
 486         kfree(mdi);
 487 }
 488
 489 void nilfs_mdt_set_entry_size(struct inode *inode, unsigned int entry_size,
 490                               unsigned int header_size)
 491 {
 492         struct nilfs_mdt_info *mi = NILFS_MDT(inode);
 493
 494         mi->mi_entry_size = entry_size;
 495         mi->mi_entries_per_block = i_blocksize(inode) / entry_size;
 496         mi->mi_first_entry_offset = DIV_ROUND_UP(header_size, entry_size);
 497 }
 498
 499 /**
 500  * nilfs_mdt_setup_shadow_map - setup shadow map and bind it to metadata file
 501  * @inode: inode of the metadata file
 502  * @shadow: shadow mapping
 503  */
 504 int nilfs_mdt_setup_shadow_map(struct inode *inode,
 505                                struct nilfs_shadow_map *shadow)
 506 {
 507         struct nilfs_mdt_info *mi = NILFS_MDT(inode);
 508
 509         INIT_LIST_HEAD(&shadow->frozen_buffers);
 510         address_space_init_once(&shadow->frozen_data);
 511         nilfs_mapping_init(&shadow->frozen_data, inode);
 512         address_space_init_once(&shadow->frozen_btnodes);
 513         nilfs_mapping_init(&shadow->frozen_btnodes, inode);
 514         mi->mi_shadow = shadow;
 515         return 0;
 516 }
 517
 518 /**
 519  * nilfs_mdt_save_to_shadow_map - copy bmap and dirty pages to shadow map
 520  * @inode: inode of the metadata file
 521  */
 522 int nilfs_mdt_save_to_shadow_map(struct inode *inode)
 523 {
 524         struct nilfs_mdt_info *mi = NILFS_MDT(inode);
 525         struct nilfs_inode_info *ii = NILFS_I(inode);
 526         struct nilfs_shadow_map *shadow = mi->mi_shadow;
 527         int ret;
 528
 529         ret = nilfs_copy_dirty_pages(&shadow->frozen_data, inode->i_mapping);
 530         if (ret)
 531                 goto out;
 532
 533         ret = nilfs_copy_dirty_pages(&shadow->frozen_btnodes,
 534                                      &ii->i_btnode_cache);
 535         if (ret)
 536                 goto out;
 537
 538         nilfs_bmap_save(ii->i_bmap, &shadow->bmap_store);
 539  out:
 540         return ret;
 541 }
 542
 543 int nilfs_mdt_freeze_buffer(struct inode *inode, struct buffer_head *bh)
 544 {
 545         struct nilfs_shadow_map *shadow = NILFS_MDT(inode)->mi_shadow;
 546         struct buffer_head *bh_frozen;
 547         struct page *page;
 548         int blkbits = inode->i_blkbits;
 549
 550         page = grab_cache_page(&shadow->frozen_data, bh->b_page->index);
 551         if (!page)
 552                 return -ENOMEM;
 553
 554         if (!page_has_buffers(page))
 555                 create_empty_buffers(page, 1 << blkbits, 0);
 556
 557         bh_frozen = nilfs_page_get_nth_block(page, bh_offset(bh) >> blkbits);
 558
 559         if (!buffer_uptodate(bh_frozen))
 560                 nilfs_copy_buffer(bh_frozen, bh);
 561         if (list_empty(&bh_frozen->b_assoc_buffers)) {
 562                 list_add_tail(&bh_frozen->b_assoc_buffers,
 563                               &shadow->frozen_buffers);
 564                 set_buffer_nilfs_redirected(bh);
 565         } else {
 566                 brelse(bh_frozen); /* already frozen */
 567         }
 568
 569         unlock_page(page);
 570         put_page(page);
 571         return 0;
 572 }
 573
 574 struct buffer_head *
 575 nilfs_mdt_get_frozen_buffer(struct inode *inode, struct buffer_head *bh)
 576 {
 577         struct nilfs_shadow_map *shadow = NILFS_MDT(inode)->mi_shadow;
 578         struct buffer_head *bh_frozen = NULL;
 579         struct page *page;
 580         int n;
 581
 582         page = find_lock_page(&shadow->frozen_data, bh->b_page->index);
 583         if (page) {
 584                 if (page_has_buffers(page)) {
 585                         n = bh_offset(bh) >> inode->i_blkbits;
 586                         bh_frozen = nilfs_page_get_nth_block(page, n);
 587                 }
 588                 unlock_page(page);
 589                 put_page(page);
 590         }
 591         return bh_frozen;
 592 }
 593
 594 static void nilfs_release_frozen_buffers(struct nilfs_shadow_map *shadow)
 595 {
 596         struct list_head *head = &shadow->frozen_buffers;
 597         struct buffer_head *bh;
 598
 599         while (!list_empty(head)) {
 600                 bh = list_first_entry(head, struct buffer_head,
 601                                       b_assoc_buffers);
 602                 list_del_init(&bh->b_assoc_buffers);
 603                 brelse(bh); /* drop ref-count to make it releasable */
 604         }
 605 }
 606
 607 /**
 608  * nilfs_mdt_restore_from_shadow_map - restore dirty pages and bmap state
 609  * @inode: inode of the metadata file
 610  */
 611 void nilfs_mdt_restore_from_shadow_map(struct inode *inode)
 612 {
 613         struct nilfs_mdt_info *mi = NILFS_MDT(inode);
 614         struct nilfs_inode_info *ii = NILFS_I(inode);
 615         struct nilfs_shadow_map *shadow = mi->mi_shadow;
 616
 617         down_write(&mi->mi_sem);
 618
 619         if (mi->mi_palloc_cache)
 620                 nilfs_palloc_clear_cache(inode);
 621
 622         nilfs_clear_dirty_pages(inode->i_mapping, true);
 623         nilfs_copy_back_pages(inode->i_mapping, &shadow->frozen_data);
 624
 625         nilfs_clear_dirty_pages(&ii->i_btnode_cache, true);
 626         nilfs_copy_back_pages(&ii->i_btnode_cache, &shadow->frozen_btnodes);
 627
 628         nilfs_bmap_restore(ii->i_bmap, &shadow->bmap_store);
 629
 630         up_write(&mi->mi_sem);
 631 }
 632
 633 /**
 634  * nilfs_mdt_clear_shadow_map - truncate pages in shadow map caches
 635  * @inode: inode of the metadata file
 636  */
 637 void nilfs_mdt_clear_shadow_map(struct inode *inode)
 638 {
 639         struct nilfs_mdt_info *mi = NILFS_MDT(inode);
 640         struct nilfs_shadow_map *shadow = mi->mi_shadow;
 641
 642         down_write(&mi->mi_sem);
 643         nilfs_release_frozen_buffers(shadow);
 644         truncate_inode_pages(&shadow->frozen_data, 0);
 645         truncate_inode_pages(&shadow->frozen_btnodes, 0);
 646         up_write(&mi->mi_sem);
 647 }