fs/nilfs2/page.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Buffer/page management specific to NILFS
   4  *
   5  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   6  *
   7  * Written by Ryusuke Konishi and Seiji Kihara.
   8  */
   9
  10 #include <linux/pagemap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/swap.h>
  13 #include <linux/bitops.h>
  14 #include <linux/page-flags.h>
  15 #include <linux/list.h>
  16 #include <linux/highmem.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/gfp.h>
  19 #include "nilfs.h"
  20 #include "page.h"
  21 #include "mdt.h"
  22
  23
  24 #define NILFS_BUFFER_INHERENT_BITS                                      \
  25         (BIT(BH_Uptodate) | BIT(BH_Mapped) | BIT(BH_NILFS_Node) |       \
  26          BIT(BH_NILFS_Volatile) | BIT(BH_NILFS_Checked))
  27
  28 static struct buffer_head *__nilfs_get_folio_block(struct folio *folio,
  29                 unsigned long block, pgoff_t index, int blkbits,
  30                 unsigned long b_state)
  31
  32 {
  33         unsigned long first_block;
  34         struct buffer_head *bh = folio_buffers(folio);
  35
  36         if (!bh)
  37                 bh = create_empty_buffers(folio, 1 << blkbits, b_state);
  38
  39         first_block = (unsigned long)index << (PAGE_SHIFT - blkbits);
  40         bh = get_nth_bh(bh, block - first_block);
  41
  42         wait_on_buffer(bh);
  43         return bh;
  44 }
  45
  46 struct buffer_head *nilfs_grab_buffer(struct inode *inode,
  47                                       struct address_space *mapping,
  48                                       unsigned long blkoff,
  49                                       unsigned long b_state)
  50 {
  51         int blkbits = inode->i_blkbits;
  52         pgoff_t index = blkoff >> (PAGE_SHIFT - blkbits);
  53         struct folio *folio;
  54         struct buffer_head *bh;
  55
  56         folio = filemap_grab_folio(mapping, index);
  57         if (IS_ERR(folio))
  58                 return NULL;
  59
  60         bh = __nilfs_get_folio_block(folio, blkoff, index, blkbits, b_state);
  61         if (unlikely(!bh)) {
  62                 folio_unlock(folio);
  63                 folio_put(folio);
  64                 return NULL;
  65         }
  66         bh->b_bdev = inode->i_sb->s_bdev;
  67         return bh;
  68 }
  69
  70 /**
  71  * nilfs_forget_buffer - discard dirty state
  72  * @bh: buffer head of the buffer to be discarded
  73  */
  74 void nilfs_forget_buffer(struct buffer_head *bh)
  75 {
  76         struct folio *folio = bh->b_folio;
  77         const unsigned long clear_bits =
  78                 (BIT(BH_Uptodate) | BIT(BH_Dirty) | BIT(BH_Mapped) |
  79                  BIT(BH_Async_Write) | BIT(BH_NILFS_Volatile) |
  80                  BIT(BH_NILFS_Checked) | BIT(BH_NILFS_Redirected) |
  81                  BIT(BH_Delay));
  82
  83         lock_buffer(bh);
  84         set_mask_bits(&bh->b_state, clear_bits, 0);
  85         if (nilfs_folio_buffers_clean(folio))
  86                 __nilfs_clear_folio_dirty(folio);
  87
  88         bh->b_blocknr = -1;
  89         folio_clear_uptodate(folio);
  90         folio_clear_mappedtodisk(folio);
  91         unlock_buffer(bh);
  92         brelse(bh);
  93 }
  94
  95 /**
  96  * nilfs_copy_buffer -- copy buffer data and flags
  97  * @dbh: destination buffer
  98  * @sbh: source buffer
  99  */
 100 void nilfs_copy_buffer(struct buffer_head *dbh, struct buffer_head *sbh)
 101 {
 102         void *saddr, *daddr;
 103         unsigned long bits;
 104         struct folio *sfolio = sbh->b_folio, *dfolio = dbh->b_folio;
 105         struct buffer_head *bh;
 106
 107         saddr = kmap_local_folio(sfolio, bh_offset(sbh));
 108         daddr = kmap_local_folio(dfolio, bh_offset(dbh));
 109         memcpy(daddr, saddr, sbh->b_size);
 110         kunmap_local(daddr);
 111         kunmap_local(saddr);
 112
 113         dbh->b_state = sbh->b_state & NILFS_BUFFER_INHERENT_BITS;
 114         dbh->b_blocknr = sbh->b_blocknr;
 115         dbh->b_bdev = sbh->b_bdev;
 116
 117         bh = dbh;
 118         bits = sbh->b_state & (BIT(BH_Uptodate) | BIT(BH_Mapped));
 119         while ((bh = bh->b_this_page) != dbh) {
 120                 lock_buffer(bh);
 121                 bits &= bh->b_state;
 122                 unlock_buffer(bh);
 123         }
 124         if (bits & BIT(BH_Uptodate))
 125                 folio_mark_uptodate(dfolio);
 126         else
 127                 folio_clear_uptodate(dfolio);
 128         if (bits & BIT(BH_Mapped))
 129                 folio_set_mappedtodisk(dfolio);
 130         else
 131                 folio_clear_mappedtodisk(dfolio);
 132 }
 133
 134 /**
 135  * nilfs_folio_buffers_clean - Check if a folio has dirty buffers or not.
 136  * @folio: Folio to be checked.
 137  *
 138  * Return: false if the folio has dirty buffers, true otherwise.
 139  */
 140 bool nilfs_folio_buffers_clean(struct folio *folio)
 141 {
 142         struct buffer_head *bh, *head;
 143
 144         bh = head = folio_buffers(folio);
 145         do {
 146                 if (buffer_dirty(bh))
 147                         return false;
 148                 bh = bh->b_this_page;
 149         } while (bh != head);
 150         return true;
 151 }
 152
 153 void nilfs_folio_bug(struct folio *folio)
 154 {
 155         struct buffer_head *bh, *head;
 156         struct address_space *m;
 157         unsigned long ino;
 158
 159         if (unlikely(!folio)) {
 160                 printk(KERN_CRIT "NILFS_FOLIO_BUG(NULL)\n");
 161                 return;
 162         }
 163
 164         m = folio->mapping;
 165         ino = m ? m->host->i_ino : 0;
 166
 167         printk(KERN_CRIT "NILFS_FOLIO_BUG(%p): cnt=%d index#=%llu flags=0x%lx "
 168                "mapping=%p ino=%lu\n",
 169                folio, folio_ref_count(folio),
 170                (unsigned long long)folio->index, folio->flags, m, ino);
 171
 172         head = folio_buffers(folio);
 173         if (head) {
 174                 int i = 0;
 175
 176                 bh = head;
 177                 do {
 178                         printk(KERN_CRIT
 179                                " BH[%d] %p: cnt=%d block#=%llu state=0x%lx\n",
 180                                i++, bh, atomic_read(&bh->b_count),
 181                                (unsigned long long)bh->b_blocknr, bh->b_state);
 182                         bh = bh->b_this_page;
 183                 } while (bh != head);
 184         }
 185 }
 186
 187 /**
 188  * nilfs_copy_folio -- copy the folio with buffers
 189  * @dst: destination folio
 190  * @src: source folio
 191  * @copy_dirty: flag whether to copy dirty states on the folio's buffer heads.
 192  *
 193  * This function is for both data folios and btnode folios.  The dirty flag
 194  * should be treated by caller.  The folio must not be under i/o.
 195  * Both src and dst folio must be locked
 196  */
 197 static void nilfs_copy_folio(struct folio *dst, struct folio *src,
 198                 bool copy_dirty)
 199 {
 200         struct buffer_head *dbh, *dbufs, *sbh;
 201         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
 202
 203         BUG_ON(folio_test_writeback(dst));
 204
 205         sbh = folio_buffers(src);
 206         dbh = folio_buffers(dst);
 207         if (!dbh)
 208                 dbh = create_empty_buffers(dst, sbh->b_size, 0);
 209
 210         if (copy_dirty)
 211                 mask |= BIT(BH_Dirty);
 212
 213         dbufs = dbh;
 214         do {
 215                 lock_buffer(sbh);
 216                 lock_buffer(dbh);
 217                 dbh->b_state = sbh->b_state & mask;
 218                 dbh->b_blocknr = sbh->b_blocknr;
 219                 dbh->b_bdev = sbh->b_bdev;
 220                 sbh = sbh->b_this_page;
 221                 dbh = dbh->b_this_page;
 222         } while (dbh != dbufs);
 223
 224         folio_copy(dst, src);
 225
 226         if (folio_test_uptodate(src) && !folio_test_uptodate(dst))
 227                 folio_mark_uptodate(dst);
 228         else if (!folio_test_uptodate(src) && folio_test_uptodate(dst))
 229                 folio_clear_uptodate(dst);
 230         if (folio_test_mappedtodisk(src) && !folio_test_mappedtodisk(dst))
 231                 folio_set_mappedtodisk(dst);
 232         else if (!folio_test_mappedtodisk(src) && folio_test_mappedtodisk(dst))
 233                 folio_clear_mappedtodisk(dst);
 234
 235         do {
 236                 unlock_buffer(sbh);
 237                 unlock_buffer(dbh);
 238                 sbh = sbh->b_this_page;
 239                 dbh = dbh->b_this_page;
 240         } while (dbh != dbufs);
 241 }
 242
 243 int nilfs_copy_dirty_pages(struct address_space *dmap,
 244                            struct address_space *smap)
 245 {
 246         struct folio_batch fbatch;
 247         unsigned int i;
 248         pgoff_t index = 0;
 249         int err = 0;
 250
 251         folio_batch_init(&fbatch);
 252 repeat:
 253         if (!filemap_get_folios_tag(smap, &index, (pgoff_t)-1,
 254                                 PAGECACHE_TAG_DIRTY, &fbatch))
 255                 return 0;
 256
 257         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 258                 struct folio *folio = fbatch.folios[i], *dfolio;
 259
 260                 folio_lock(folio);
 261                 if (unlikely(!folio_test_dirty(folio)))
 262                         NILFS_FOLIO_BUG(folio, "inconsistent dirty state");
 263
 264                 dfolio = filemap_grab_folio(dmap, folio->index);
 265                 if (IS_ERR(dfolio)) {
 266                         /* No empty page is added to the page cache */
 267                         folio_unlock(folio);
 268                         err = PTR_ERR(dfolio);
 269                         break;
 270                 }
 271                 if (unlikely(!folio_buffers(folio)))
 272                         NILFS_FOLIO_BUG(folio,
 273                                        "found empty page in dat page cache");
 274
 275                 nilfs_copy_folio(dfolio, folio, true);
 276                 filemap_dirty_folio(folio_mapping(dfolio), dfolio);
 277
 278                 folio_unlock(dfolio);
 279                 folio_put(dfolio);
 280                 folio_unlock(folio);
 281         }
 282         folio_batch_release(&fbatch);
 283         cond_resched();
 284
 285         if (likely(!err))
 286                 goto repeat;
 287         return err;
 288 }
 289
 290 /**
 291  * nilfs_copy_back_pages -- copy back pages to original cache from shadow cache
 292  * @dmap: destination page cache
 293  * @smap: source page cache
 294  *
 295  * No pages must be added to the cache during this process.
 296  * This must be ensured by the caller.
 297  */
 298 void nilfs_copy_back_pages(struct address_space *dmap,
 299                            struct address_space *smap)
 300 {
 301         struct folio_batch fbatch;
 302         unsigned int i, n;
 303         pgoff_t start = 0;
 304
 305         folio_batch_init(&fbatch);
 306 repeat:
 307         n = filemap_get_folios(smap, &start, ~0UL, &fbatch);
 308         if (!n)
 309                 return;
 310
 311         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 312                 struct folio *folio = fbatch.folios[i], *dfolio;
 313                 pgoff_t index = folio->index;
 314
 315                 folio_lock(folio);
 316                 dfolio = filemap_lock_folio(dmap, index);
 317                 if (!IS_ERR(dfolio)) {
 318                         /* overwrite existing folio in the destination cache */
 319                         WARN_ON(folio_test_dirty(dfolio));
 320                         nilfs_copy_folio(dfolio, folio, false);
 321                         folio_unlock(dfolio);
 322                         folio_put(dfolio);
 323                         /* Do we not need to remove folio from smap here? */
 324                 } else {
 325                         struct folio *f;
 326
 327                         /* move the folio to the destination cache */
 328                         xa_lock_irq(&smap->i_pages);
 329                         f = __xa_erase(&smap->i_pages, index);
 330                         WARN_ON(folio != f);
 331                         smap->nrpages--;
 332                         xa_unlock_irq(&smap->i_pages);
 333
 334                         xa_lock_irq(&dmap->i_pages);
 335                         f = __xa_store(&dmap->i_pages, index, folio, GFP_NOFS);
 336                         if (unlikely(f)) {
 337                                 /* Probably -ENOMEM */
 338                                 folio->mapping = NULL;
 339                                 folio_put(folio);
 340                         } else {
 341                                 folio->mapping = dmap;
 342                                 dmap->nrpages++;
 343                                 if (folio_test_dirty(folio))
 344                                         __xa_set_mark(&dmap->i_pages, index,
 345                                                         PAGECACHE_TAG_DIRTY);
 346                         }
 347                         xa_unlock_irq(&dmap->i_pages);
 348                 }
 349                 folio_unlock(folio);
 350         }
 351         folio_batch_release(&fbatch);
 352         cond_resched();
 353
 354         goto repeat;
 355 }
 356
 357 /**
 358  * nilfs_clear_dirty_pages - discard dirty pages in address space
 359  * @mapping: address space with dirty pages for discarding
 360  */
 361 void nilfs_clear_dirty_pages(struct address_space *mapping)
 362 {
 363         struct folio_batch fbatch;
 364         unsigned int i;
 365         pgoff_t index = 0;
 366
 367         folio_batch_init(&fbatch);
 368
 369         while (filemap_get_folios_tag(mapping, &index, (pgoff_t)-1,
 370                                 PAGECACHE_TAG_DIRTY, &fbatch)) {
 371                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 372                         struct folio *folio = fbatch.folios[i];
 373
 374                         folio_lock(folio);
 375
 376                         /*
 377                          * This folio may have been removed from the address
 378                          * space by truncation or invalidation when the lock
 379                          * was acquired.  Skip processing in that case.
 380                          */
 381                         if (likely(folio->mapping == mapping))
 382                                 nilfs_clear_folio_dirty(folio);
 383
 384                         folio_unlock(folio);
 385                 }
 386                 folio_batch_release(&fbatch);
 387                 cond_resched();
 388         }
 389 }
 390
 391 /**
 392  * nilfs_clear_folio_dirty - discard dirty folio
 393  * @folio: dirty folio that will be discarded
 394  *
 395  * nilfs_clear_folio_dirty() clears working states including dirty state for
 396  * the folio and its buffers.  If the folio has buffers, clear only if it is
 397  * confirmed that none of the buffer heads are busy (none have valid
 398  * references and none are locked).
 399  */
 400 void nilfs_clear_folio_dirty(struct folio *folio)
 401 {
 402         struct buffer_head *bh, *head;
 403
 404         BUG_ON(!folio_test_locked(folio));
 405
 406         head = folio_buffers(folio);
 407         if (head) {
 408                 const unsigned long clear_bits =
 409                         (BIT(BH_Uptodate) | BIT(BH_Dirty) | BIT(BH_Mapped) |
 410                          BIT(BH_Async_Write) | BIT(BH_NILFS_Volatile) |
 411                          BIT(BH_NILFS_Checked) | BIT(BH_NILFS_Redirected) |
 412                          BIT(BH_Delay));
 413                 bool busy, invalidated = false;
 414
 415 recheck_buffers:
 416                 busy = false;
 417                 bh = head;
 418                 do {
 419                         if (atomic_read(&bh->b_count) | buffer_locked(bh)) {
 420                                 busy = true;
 421                                 break;
 422                         }
 423                 } while (bh = bh->b_this_page, bh != head);
 424
 425                 if (busy) {
 426                         if (invalidated)
 427                                 return;
 428                         invalidate_bh_lrus();
 429                         invalidated = true;
 430                         goto recheck_buffers;
 431                 }
 432
 433                 bh = head;
 434                 do {
 435                         lock_buffer(bh);
 436                         set_mask_bits(&bh->b_state, clear_bits, 0);
 437                         unlock_buffer(bh);
 438                 } while (bh = bh->b_this_page, bh != head);
 439         }
 440
 441         folio_clear_uptodate(folio);
 442         folio_clear_mappedtodisk(folio);
 443         folio_clear_checked(folio);
 444         __nilfs_clear_folio_dirty(folio);
 445 }
 446
 447 unsigned int nilfs_page_count_clean_buffers(struct folio *folio,
 448                                             unsigned int from, unsigned int to)
 449 {
 450         unsigned int block_start, block_end;
 451         struct buffer_head *bh, *head;
 452         unsigned int nc = 0;
 453
 454         for (bh = head = folio_buffers(folio), block_start = 0;
 455              bh != head || !block_start;
 456              block_start = block_end, bh = bh->b_this_page) {
 457                 block_end = block_start + bh->b_size;
 458                 if (block_end > from && block_start < to && !buffer_dirty(bh))
 459                         nc++;
 460         }
 461         return nc;
 462 }
 463
 464 /*
 465  * NILFS2 needs clear_page_dirty() in the following two cases:
 466  *
 467  * 1) For B-tree node pages and data pages of DAT file, NILFS2 clears dirty
 468  *    flag of pages when it copies back pages from shadow cache to the
 469  *    original cache.
 470  *
 471  * 2) Some B-tree operations like insertion or deletion may dispose buffers
 472  *    in dirty state, and this needs to cancel the dirty state of their pages.
 473  */
 474 void __nilfs_clear_folio_dirty(struct folio *folio)
 475 {
 476         struct address_space *mapping = folio->mapping;
 477
 478         if (mapping) {
 479                 xa_lock_irq(&mapping->i_pages);
 480                 if (folio_test_dirty(folio)) {
 481                         __xa_clear_mark(&mapping->i_pages, folio->index,
 482                                              PAGECACHE_TAG_DIRTY);
 483                         xa_unlock_irq(&mapping->i_pages);
 484                         folio_clear_dirty_for_io(folio);
 485                         return;
 486                 }
 487                 xa_unlock_irq(&mapping->i_pages);
 488                 return;
 489         }
 490         folio_clear_dirty(folio);
 491 }
 492
 493 /**
 494  * nilfs_find_uncommitted_extent - find extent of uncommitted data
 495  * @inode: inode
 496  * @start_blk: start block offset (in)
 497  * @blkoff: start offset of the found extent (out)
 498  *
 499  * This function searches an extent of buffers marked "delayed" which
 500  * starts from a block offset equal to or larger than @start_blk.  If
 501  * such an extent was found, this will store the start offset in
 502  * @blkoff and return its length in blocks.
 503  *
 504  * Return: Length in blocks of found extent, 0 otherwise.
 505  */
 506 unsigned long nilfs_find_uncommitted_extent(struct inode *inode,
 507                                             sector_t start_blk,
 508                                             sector_t *blkoff)
 509 {
 510         unsigned int i, nr_folios;
 511         pgoff_t index;
 512         unsigned long length = 0;
 513         struct folio_batch fbatch;
 514         struct folio *folio;
 515
 516         if (inode->i_mapping->nrpages == 0)
 517                 return 0;
 518
 519         index = start_blk >> (PAGE_SHIFT - inode->i_blkbits);
 520
 521         folio_batch_init(&fbatch);
 522
 523 repeat:
 524         nr_folios = filemap_get_folios_contig(inode->i_mapping, &index, ULONG_MAX,
 525                         &fbatch);
 526         if (nr_folios == 0)
 527                 return length;
 528
 529         i = 0;
 530         do {
 531                 folio = fbatch.folios[i];
 532
 533                 folio_lock(folio);
 534                 if (folio_buffers(folio)) {
 535                         struct buffer_head *bh, *head;
 536                         sector_t b;
 537
 538                         b = folio->index << (PAGE_SHIFT - inode->i_blkbits);
 539                         bh = head = folio_buffers(folio);
 540                         do {
 541                                 if (b < start_blk)
 542                                         continue;
 543                                 if (buffer_delay(bh)) {
 544                                         if (length == 0)
 545                                                 *blkoff = b;
 546                                         length++;
 547                                 } else if (length > 0) {
 548                                         goto out_locked;
 549                                 }
 550                         } while (++b, bh = bh->b_this_page, bh != head);
 551                 } else {
 552                         if (length > 0)
 553                                 goto out_locked;
 554                 }
 555                 folio_unlock(folio);
 556
 557         } while (++i < nr_folios);
 558
 559         folio_batch_release(&fbatch);
 560         cond_resched();
 561         goto repeat;
 562
 563 out_locked:
 564         folio_unlock(folio);
 565         folio_batch_release(&fbatch);
 566         return length;
 567 }