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  * nilfs_folio_buffers_clean() returns false if the folio has dirty buffers.
 139  * Otherwise, it returns true.
 140  */
 141 bool nilfs_folio_buffers_clean(struct folio *folio)
 142 {
 143         struct buffer_head *bh, *head;
 144
 145         bh = head = folio_buffers(folio);
 146         do {
 147                 if (buffer_dirty(bh))
 148                         return false;
 149                 bh = bh->b_this_page;
 150         } while (bh != head);
 151         return true;
 152 }
 153
 154 void nilfs_folio_bug(struct folio *folio)
 155 {
 156         struct buffer_head *bh, *head;
 157         struct address_space *m;
 158         unsigned long ino;
 159
 160         if (unlikely(!folio)) {
 161                 printk(KERN_CRIT "NILFS_FOLIO_BUG(NULL)\n");
 162                 return;
 163         }
 164
 165         m = folio->mapping;
 166         ino = m ? m->host->i_ino : 0;
 167
 168         printk(KERN_CRIT "NILFS_FOLIO_BUG(%p): cnt=%d index#=%llu flags=0x%lx "
 169                "mapping=%p ino=%lu\n",
 170                folio, folio_ref_count(folio),
 171                (unsigned long long)folio->index, folio->flags, m, ino);
 172
 173         head = folio_buffers(folio);
 174         if (head) {
 175                 int i = 0;
 176
 177                 bh = head;
 178                 do {
 179                         printk(KERN_CRIT
 180                                " BH[%d] %p: cnt=%d block#=%llu state=0x%lx\n",
 181                                i++, bh, atomic_read(&bh->b_count),
 182                                (unsigned long long)bh->b_blocknr, bh->b_state);
 183                         bh = bh->b_this_page;
 184                 } while (bh != head);
 185         }
 186 }
 187
 188 /**
 189  * nilfs_copy_folio -- copy the folio with buffers
 190  * @dst: destination folio
 191  * @src: source folio
 192  * @copy_dirty: flag whether to copy dirty states on the folio's buffer heads.
 193  *
 194  * This function is for both data folios and btnode folios.  The dirty flag
 195  * should be treated by caller.  The folio must not be under i/o.
 196  * Both src and dst folio must be locked
 197  */
 198 static void nilfs_copy_folio(struct folio *dst, struct folio *src,
 199                 bool copy_dirty)
 200 {
 201         struct buffer_head *dbh, *dbufs, *sbh;
 202         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
 203
 204         BUG_ON(folio_test_writeback(dst));
 205
 206         sbh = folio_buffers(src);
 207         dbh = folio_buffers(dst);
 208         if (!dbh)
 209                 dbh = create_empty_buffers(dst, sbh->b_size, 0);
 210
 211         if (copy_dirty)
 212                 mask |= BIT(BH_Dirty);
 213
 214         dbufs = dbh;
 215         do {
 216                 lock_buffer(sbh);
 217                 lock_buffer(dbh);
 218                 dbh->b_state = sbh->b_state & mask;
 219                 dbh->b_blocknr = sbh->b_blocknr;
 220                 dbh->b_bdev = sbh->b_bdev;
 221                 sbh = sbh->b_this_page;
 222                 dbh = dbh->b_this_page;
 223         } while (dbh != dbufs);
 224
 225         folio_copy(dst, src);
 226
 227         if (folio_test_uptodate(src) && !folio_test_uptodate(dst))
 228                 folio_mark_uptodate(dst);
 229         else if (!folio_test_uptodate(src) && folio_test_uptodate(dst))
 230                 folio_clear_uptodate(dst);
 231         if (folio_test_mappedtodisk(src) && !folio_test_mappedtodisk(dst))
 232                 folio_set_mappedtodisk(dst);
 233         else if (!folio_test_mappedtodisk(src) && folio_test_mappedtodisk(dst))
 234                 folio_clear_mappedtodisk(dst);
 235
 236         do {
 237                 unlock_buffer(sbh);
 238                 unlock_buffer(dbh);
 239                 sbh = sbh->b_this_page;
 240                 dbh = dbh->b_this_page;
 241         } while (dbh != dbufs);
 242 }
 243
 244 int nilfs_copy_dirty_pages(struct address_space *dmap,
 245                            struct address_space *smap)
 246 {
 247         struct folio_batch fbatch;
 248         unsigned int i;
 249         pgoff_t index = 0;
 250         int err = 0;
 251
 252         folio_batch_init(&fbatch);
 253 repeat:
 254         if (!filemap_get_folios_tag(smap, &index, (pgoff_t)-1,
 255                                 PAGECACHE_TAG_DIRTY, &fbatch))
 256                 return 0;
 257
 258         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 259                 struct folio *folio = fbatch.folios[i], *dfolio;
 260
 261                 folio_lock(folio);
 262                 if (unlikely(!folio_test_dirty(folio)))
 263                         NILFS_FOLIO_BUG(folio, "inconsistent dirty state");
 264
 265                 dfolio = filemap_grab_folio(dmap, folio->index);
 266                 if (IS_ERR(dfolio)) {
 267                         /* No empty page is added to the page cache */
 268                         folio_unlock(folio);
 269                         err = PTR_ERR(dfolio);
 270                         break;
 271                 }
 272                 if (unlikely(!folio_buffers(folio)))
 273                         NILFS_FOLIO_BUG(folio,
 274                                        "found empty page in dat page cache");
 275
 276                 nilfs_copy_folio(dfolio, folio, true);
 277                 filemap_dirty_folio(folio_mapping(dfolio), dfolio);
 278
 279                 folio_unlock(dfolio);
 280                 folio_put(dfolio);
 281                 folio_unlock(folio);
 282         }
 283         folio_batch_release(&fbatch);
 284         cond_resched();
 285
 286         if (likely(!err))
 287                 goto repeat;
 288         return err;
 289 }
 290
 291 /**
 292  * nilfs_copy_back_pages -- copy back pages to original cache from shadow cache
 293  * @dmap: destination page cache
 294  * @smap: source page cache
 295  *
 296  * No pages must be added to the cache during this process.
 297  * This must be ensured by the caller.
 298  */
 299 void nilfs_copy_back_pages(struct address_space *dmap,
 300                            struct address_space *smap)
 301 {
 302         struct folio_batch fbatch;
 303         unsigned int i, n;
 304         pgoff_t start = 0;
 305
 306         folio_batch_init(&fbatch);
 307 repeat:
 308         n = filemap_get_folios(smap, &start, ~0UL, &fbatch);
 309         if (!n)
 310                 return;
 311
 312         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 313                 struct folio *folio = fbatch.folios[i], *dfolio;
 314                 pgoff_t index = folio->index;
 315
 316                 folio_lock(folio);
 317                 dfolio = filemap_lock_folio(dmap, index);
 318                 if (!IS_ERR(dfolio)) {
 319                         /* overwrite existing folio in the destination cache */
 320                         WARN_ON(folio_test_dirty(dfolio));
 321                         nilfs_copy_folio(dfolio, folio, false);
 322                         folio_unlock(dfolio);
 323                         folio_put(dfolio);
 324                         /* Do we not need to remove folio from smap here? */
 325                 } else {
 326                         struct folio *f;
 327
 328                         /* move the folio to the destination cache */
 329                         xa_lock_irq(&smap->i_pages);
 330                         f = __xa_erase(&smap->i_pages, index);
 331                         WARN_ON(folio != f);
 332                         smap->nrpages--;
 333                         xa_unlock_irq(&smap->i_pages);
 334
 335                         xa_lock_irq(&dmap->i_pages);
 336                         f = __xa_store(&dmap->i_pages, index, folio, GFP_NOFS);
 337                         if (unlikely(f)) {
 338                                 /* Probably -ENOMEM */
 339                                 folio->mapping = NULL;
 340                                 folio_put(folio);
 341                         } else {
 342                                 folio->mapping = dmap;
 343                                 dmap->nrpages++;
 344                                 if (folio_test_dirty(folio))
 345                                         __xa_set_mark(&dmap->i_pages, index,
 346                                                         PAGECACHE_TAG_DIRTY);
 347                         }
 348                         xa_unlock_irq(&dmap->i_pages);
 349                 }
 350                 folio_unlock(folio);
 351         }
 352         folio_batch_release(&fbatch);
 353         cond_resched();
 354
 355         goto repeat;
 356 }
 357
 358 /**
 359  * nilfs_clear_dirty_pages - discard dirty pages in address space
 360  * @mapping: address space with dirty pages for discarding
 361  */
 362 void nilfs_clear_dirty_pages(struct address_space *mapping)
 363 {
 364         struct folio_batch fbatch;
 365         unsigned int i;
 366         pgoff_t index = 0;
 367
 368         folio_batch_init(&fbatch);
 369
 370         while (filemap_get_folios_tag(mapping, &index, (pgoff_t)-1,
 371                                 PAGECACHE_TAG_DIRTY, &fbatch)) {
 372                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 373                         struct folio *folio = fbatch.folios[i];
 374
 375                         folio_lock(folio);
 376
 377                         /*
 378                          * This folio may have been removed from the address
 379                          * space by truncation or invalidation when the lock
 380                          * was acquired.  Skip processing in that case.
 381                          */
 382                         if (likely(folio->mapping == mapping))
 383                                 nilfs_clear_folio_dirty(folio);
 384
 385                         folio_unlock(folio);
 386                 }
 387                 folio_batch_release(&fbatch);
 388                 cond_resched();
 389         }
 390 }
 391
 392 /**
 393  * nilfs_clear_folio_dirty - discard dirty folio
 394  * @folio: dirty folio that will be discarded
 395  */
 396 void nilfs_clear_folio_dirty(struct folio *folio)
 397 {
 398         struct buffer_head *bh, *head;
 399
 400         BUG_ON(!folio_test_locked(folio));
 401
 402         folio_clear_uptodate(folio);
 403         folio_clear_mappedtodisk(folio);
 404         folio_clear_checked(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
 414                 bh = head;
 415                 do {
 416                         lock_buffer(bh);
 417                         set_mask_bits(&bh->b_state, clear_bits, 0);
 418                         unlock_buffer(bh);
 419                 } while (bh = bh->b_this_page, bh != head);
 420         }
 421
 422         __nilfs_clear_folio_dirty(folio);
 423 }
 424
 425 unsigned int nilfs_page_count_clean_buffers(struct folio *folio,
 426                                             unsigned int from, unsigned int to)
 427 {
 428         unsigned int block_start, block_end;
 429         struct buffer_head *bh, *head;
 430         unsigned int nc = 0;
 431
 432         for (bh = head = folio_buffers(folio), block_start = 0;
 433              bh != head || !block_start;
 434              block_start = block_end, bh = bh->b_this_page) {
 435                 block_end = block_start + bh->b_size;
 436                 if (block_end > from && block_start < to && !buffer_dirty(bh))
 437                         nc++;
 438         }
 439         return nc;
 440 }
 441
 442 /*
 443  * NILFS2 needs clear_page_dirty() in the following two cases:
 444  *
 445  * 1) For B-tree node pages and data pages of DAT file, NILFS2 clears dirty
 446  *    flag of pages when it copies back pages from shadow cache to the
 447  *    original cache.
 448  *
 449  * 2) Some B-tree operations like insertion or deletion may dispose buffers
 450  *    in dirty state, and this needs to cancel the dirty state of their pages.
 451  */
 452 void __nilfs_clear_folio_dirty(struct folio *folio)
 453 {
 454         struct address_space *mapping = folio->mapping;
 455
 456         if (mapping) {
 457                 xa_lock_irq(&mapping->i_pages);
 458                 if (folio_test_dirty(folio)) {
 459                         __xa_clear_mark(&mapping->i_pages, folio->index,
 460                                              PAGECACHE_TAG_DIRTY);
 461                         xa_unlock_irq(&mapping->i_pages);
 462                         folio_clear_dirty_for_io(folio);
 463                         return;
 464                 }
 465                 xa_unlock_irq(&mapping->i_pages);
 466                 return;
 467         }
 468         folio_clear_dirty(folio);
 469 }
 470
 471 /**
 472  * nilfs_find_uncommitted_extent - find extent of uncommitted data
 473  * @inode: inode
 474  * @start_blk: start block offset (in)
 475  * @blkoff: start offset of the found extent (out)
 476  *
 477  * This function searches an extent of buffers marked "delayed" which
 478  * starts from a block offset equal to or larger than @start_blk.  If
 479  * such an extent was found, this will store the start offset in
 480  * @blkoff and return its length in blocks.  Otherwise, zero is
 481  * returned.
 482  */
 483 unsigned long nilfs_find_uncommitted_extent(struct inode *inode,
 484                                             sector_t start_blk,
 485                                             sector_t *blkoff)
 486 {
 487         unsigned int i, nr_folios;
 488         pgoff_t index;
 489         unsigned long length = 0;
 490         struct folio_batch fbatch;
 491         struct folio *folio;
 492
 493         if (inode->i_mapping->nrpages == 0)
 494                 return 0;
 495
 496         index = start_blk >> (PAGE_SHIFT - inode->i_blkbits);
 497
 498         folio_batch_init(&fbatch);
 499
 500 repeat:
 501         nr_folios = filemap_get_folios_contig(inode->i_mapping, &index, ULONG_MAX,
 502                         &fbatch);
 503         if (nr_folios == 0)
 504                 return length;
 505
 506         i = 0;
 507         do {
 508                 folio = fbatch.folios[i];
 509
 510                 folio_lock(folio);
 511                 if (folio_buffers(folio)) {
 512                         struct buffer_head *bh, *head;
 513                         sector_t b;
 514
 515                         b = folio->index << (PAGE_SHIFT - inode->i_blkbits);
 516                         bh = head = folio_buffers(folio);
 517                         do {
 518                                 if (b < start_blk)
 519                                         continue;
 520                                 if (buffer_delay(bh)) {
 521                                         if (length == 0)
 522                                                 *blkoff = b;
 523                                         length++;
 524                                 } else if (length > 0) {
 525                                         goto out_locked;
 526                                 }
 527                         } while (++b, bh = bh->b_this_page, bh != head);
 528                 } else {
 529                         if (length > 0)
 530                                 goto out_locked;
 531                 }
 532                 folio_unlock(folio);
 533
 534         } while (++i < nr_folios);
 535
 536         folio_batch_release(&fbatch);
 537         cond_resched();
 538         goto repeat;
 539
 540 out_locked:
 541         folio_unlock(folio);
 542         folio_batch_release(&fbatch);
 543         return length;
 544 }