fs/gfs2/aops.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #include <linux/sched.h>
   8 #include <linux/slab.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/completion.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/pagevec.h>
  14 #include <linux/mpage.h>
  15 #include <linux/fs.h>
  16 #include <linux/writeback.h>
  17 #include <linux/swap.h>
  18 #include <linux/gfs2_ondisk.h>
  19 #include <linux/backing-dev.h>
  20 #include <linux/uio.h>
  21 #include <trace/events/writeback.h>
  22 #include <linux/sched/signal.h>
  23
  24 #include "gfs2.h"
  25 #include "incore.h"
  26 #include "bmap.h"
  27 #include "glock.h"
  28 #include "inode.h"
  29 #include "log.h"
  30 #include "meta_io.h"
  31 #include "quota.h"
  32 #include "trans.h"
  33 #include "rgrp.h"
  34 #include "super.h"
  35 #include "util.h"
  36 #include "glops.h"
  37 #include "aops.h"
  38
  39
  40 void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  41                             unsigned int from, unsigned int len)
  42 {
  43         struct buffer_head *head = page_buffers(page);
  44         unsigned int bsize = head->b_size;
  45         struct buffer_head *bh;
  46         unsigned int to = from + len;
  47         unsigned int start, end;
  48
  49         for (bh = head, start = 0; bh != head || !start;
  50              bh = bh->b_this_page, start = end) {
  51                 end = start + bsize;
  52                 if (end <= from)
  53                         continue;
  54                 if (start >= to)
  55                         break;
  56                 set_buffer_uptodate(bh);
  57                 gfs2_trans_add_data(ip->i_gl, bh);
  58         }
  59 }
  60
  61 /**
  62  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  63  * @inode: The inode
  64  * @lblock: The block number to look up
  65  * @bh_result: The buffer head to return the result in
  66  * @create: Non-zero if we may add block to the file
  67  *
  68  * Returns: errno
  69  */
  70
  71 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  72                                   struct buffer_head *bh_result, int create)
  73 {
  74         int error;
  75
  76         error = gfs2_block_map(inode, lblock, bh_result, 0);
  77         if (error)
  78                 return error;
  79         if (!buffer_mapped(bh_result))
  80                 return -ENODATA;
  81         return 0;
  82 }
  83
  84 /**
  85  * gfs2_writepage - Write page for writeback mappings
  86  * @page: The page
  87  * @wbc: The writeback control
  88  */
  89 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
  90 {
  91         struct inode *inode = page->mapping->host;
  92         struct gfs2_inode *ip = GFS2_I(inode);
  93         struct gfs2_sbd *sdp = GFS2_SB(inode);
  94         struct iomap_writepage_ctx wpc = { };
  95
  96         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
  97                 goto out;
  98         if (current->journal_info)
  99                 goto redirty;
 100         return iomap_writepage(page, wbc, &wpc, &gfs2_writeback_ops);
 101
 102 redirty:
 103         redirty_page_for_writepage(wbc, page);
 104 out:
 105         unlock_page(page);
 106         return 0;
 107 }
 108
 109 /**
 110  * gfs2_write_jdata_page - gfs2 jdata-specific version of block_write_full_page
 111  * @page: The page to write
 112  * @wbc: The writeback control
 113  *
 114  * This is the same as calling block_write_full_page, but it also
 115  * writes pages outside of i_size
 116  */
 117 static int gfs2_write_jdata_page(struct page *page,
 118                                  struct writeback_control *wbc)
 119 {
 120         struct inode * const inode = page->mapping->host;
 121         loff_t i_size = i_size_read(inode);
 122         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 123         unsigned offset;
 124
 125         /*
 126          * The page straddles i_size.  It must be zeroed out on each and every
 127          * writepage invocation because it may be mmapped.  "A file is mapped
 128          * in multiples of the page size.  For a file that is not a multiple of
 129          * the  page size, the remaining memory is zeroed when mapped, and
 130          * writes to that region are not written out to the file."
 131          */
 132         offset = i_size & (PAGE_SIZE - 1);
 133         if (page->index == end_index && offset)
 134                 zero_user_segment(page, offset, PAGE_SIZE);
 135
 136         return __block_write_full_page(inode, page, gfs2_get_block_noalloc, wbc,
 137                                        end_buffer_async_write);
 138 }
 139
 140 /**
 141  * __gfs2_jdata_writepage - The core of jdata writepage
 142  * @page: The page to write
 143  * @wbc: The writeback control
 144  *
 145  * This is shared between writepage and writepages and implements the
 146  * core of the writepage operation. If a transaction is required then
 147  * PageChecked will have been set and the transaction will have
 148  * already been started before this is called.
 149  */
 150
 151 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 152 {
 153         struct inode *inode = page->mapping->host;
 154         struct gfs2_inode *ip = GFS2_I(inode);
 155         struct gfs2_sbd *sdp = GFS2_SB(inode);
 156
 157         if (PageChecked(page)) {
 158                 ClearPageChecked(page);
 159                 if (!page_has_buffers(page)) {
 160                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 161                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 162                 }
 163                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize);
 164         }
 165         return gfs2_write_jdata_page(page, wbc);
 166 }
 167
 168 /**
 169  * gfs2_jdata_writepage - Write complete page
 170  * @page: Page to write
 171  * @wbc: The writeback control
 172  *
 173  * Returns: errno
 174  *
 175  */
 176
 177 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 178 {
 179         struct inode *inode = page->mapping->host;
 180         struct gfs2_inode *ip = GFS2_I(inode);
 181         struct gfs2_sbd *sdp = GFS2_SB(inode);
 182
 183         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 184                 goto out;
 185         if (PageChecked(page) || current->journal_info)
 186                 goto out_ignore;
 187         return __gfs2_jdata_writepage(page, wbc);
 188
 189 out_ignore:
 190         redirty_page_for_writepage(wbc, page);
 191 out:
 192         unlock_page(page);
 193         return 0;
 194 }
 195
 196 /**
 197  * gfs2_writepages - Write a bunch of dirty pages back to disk
 198  * @mapping: The mapping to write
 199  * @wbc: Write-back control
 200  *
 201  * Used for both ordered and writeback modes.
 202  */
 203 static int gfs2_writepages(struct address_space *mapping,
 204                            struct writeback_control *wbc)
 205 {
 206         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 207         struct iomap_writepage_ctx wpc = { };
 208         int ret;
 209
 210         /*
 211          * Even if we didn't write any pages here, we might still be holding
 212          * dirty pages in the ail. We forcibly flush the ail because we don't
 213          * want balance_dirty_pages() to loop indefinitely trying to write out
 214          * pages held in the ail that it can't find.
 215          */
 216         ret = iomap_writepages(mapping, wbc, &wpc, &gfs2_writeback_ops);
 217         if (ret == 0)
 218                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 219         return ret;
 220 }
 221
 222 /**
 223  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 224  * @mapping: The mapping
 225  * @wbc: The writeback control
 226  * @pvec: The vector of pages
 227  * @nr_pages: The number of pages to write
 228  * @done_index: Page index
 229  *
 230  * Returns: non-zero if loop should terminate, zero otherwise
 231  */
 232
 233 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 234                                     struct writeback_control *wbc,
 235                                     struct pagevec *pvec,
 236                                     int nr_pages,
 237                                     pgoff_t *done_index)
 238 {
 239         struct inode *inode = mapping->host;
 240         struct gfs2_sbd *sdp = GFS2_SB(inode);
 241         unsigned nrblocks = nr_pages * (PAGE_SIZE >> inode->i_blkbits);
 242         int i;
 243         int ret;
 244
 245         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 246         if (ret < 0)
 247                 return ret;
 248
 249         for(i = 0; i < nr_pages; i++) {
 250                 struct page *page = pvec->pages[i];
 251
 252                 *done_index = page->index;
 253
 254                 lock_page(page);
 255
 256                 if (unlikely(page->mapping != mapping)) {
 257 continue_unlock:
 258                         unlock_page(page);
 259                         continue;
 260                 }
 261
 262                 if (!PageDirty(page)) {
 263                         /* someone wrote it for us */
 264                         goto continue_unlock;
 265                 }
 266
 267                 if (PageWriteback(page)) {
 268                         if (wbc->sync_mode != WB_SYNC_NONE)
 269                                 wait_on_page_writeback(page);
 270                         else
 271                                 goto continue_unlock;
 272                 }
 273
 274                 BUG_ON(PageWriteback(page));
 275                 if (!clear_page_dirty_for_io(page))
 276                         goto continue_unlock;
 277
 278                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 279
 280                 ret = __gfs2_jdata_writepage(page, wbc);
 281                 if (unlikely(ret)) {
 282                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 283                                 unlock_page(page);
 284                                 ret = 0;
 285                         } else {
 286
 287                                 /*
 288                                  * done_index is set past this page,
 289                                  * so media errors will not choke
 290                                  * background writeout for the entire
 291                                  * file. This has consequences for
 292                                  * range_cyclic semantics (ie. it may
 293                                  * not be suitable for data integrity
 294                                  * writeout).
 295                                  */
 296                                 *done_index = page->index + 1;
 297                                 ret = 1;
 298                                 break;
 299                         }
 300                 }
 301
 302                 /*
 303                  * We stop writing back only if we are not doing
 304                  * integrity sync. In case of integrity sync we have to
 305                  * keep going until we have written all the pages
 306                  * we tagged for writeback prior to entering this loop.
 307                  */
 308                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 309                         ret = 1;
 310                         break;
 311                 }
 312
 313         }
 314         gfs2_trans_end(sdp);
 315         return ret;
 316 }
 317
 318 /**
 319  * gfs2_write_cache_jdata - Like write_cache_pages but different
 320  * @mapping: The mapping to write
 321  * @wbc: The writeback control
 322  *
 323  * The reason that we use our own function here is that we need to
 324  * start transactions before we grab page locks. This allows us
 325  * to get the ordering right.
 326  */
 327
 328 static int gfs2_write_cache_jdata(struct address_space *mapping,
 329                                   struct writeback_control *wbc)
 330 {
 331         int ret = 0;
 332         int done = 0;
 333         struct pagevec pvec;
 334         int nr_pages;
 335         pgoff_t writeback_index;
 336         pgoff_t index;
 337         pgoff_t end;
 338         pgoff_t done_index;
 339         int cycled;
 340         int range_whole = 0;
 341         xa_mark_t tag;
 342
 343         pagevec_init(&pvec);
 344         if (wbc->range_cyclic) {
 345                 writeback_index = mapping->writeback_index; /* prev offset */
 346                 index = writeback_index;
 347                 if (index == 0)
 348                         cycled = 1;
 349                 else
 350                         cycled = 0;
 351                 end = -1;
 352         } else {
 353                 index = wbc->range_start >> PAGE_SHIFT;
 354                 end = wbc->range_end >> PAGE_SHIFT;
 355                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 356                         range_whole = 1;
 357                 cycled = 1; /* ignore range_cyclic tests */
 358         }
 359         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 360                 tag = PAGECACHE_TAG_TOWRITE;
 361         else
 362                 tag = PAGECACHE_TAG_DIRTY;
 363
 364 retry:
 365         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 366                 tag_pages_for_writeback(mapping, index, end);
 367         done_index = index;
 368         while (!done && (index <= end)) {
 369                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
 370                                 tag);
 371                 if (nr_pages == 0)
 372                         break;
 373
 374                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
 375                 if (ret)
 376                         done = 1;
 377                 if (ret > 0)
 378                         ret = 0;
 379                 pagevec_release(&pvec);
 380                 cond_resched();
 381         }
 382
 383         if (!cycled && !done) {
 384                 /*
 385                  * range_cyclic:
 386                  * We hit the last page and there is more work to be done: wrap
 387                  * back to the start of the file
 388                  */
 389                 cycled = 1;
 390                 index = 0;
 391                 end = writeback_index - 1;
 392                 goto retry;
 393         }
 394
 395         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 396                 mapping->writeback_index = done_index;
 397
 398         return ret;
 399 }
 400
 401
 402 /**
 403  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 404  * @mapping: The mapping to write
 405  * @wbc: The writeback control
 406  *
 407  */
 408
 409 static int gfs2_jdata_writepages(struct address_space *mapping,
 410                                  struct writeback_control *wbc)
 411 {
 412         struct gfs2_inode *ip = GFS2_I(mapping->host);
 413         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 414         int ret;
 415
 416         ret = gfs2_write_cache_jdata(mapping, wbc);
 417         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 418                 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
 419                                GFS2_LFC_JDATA_WPAGES);
 420                 ret = gfs2_write_cache_jdata(mapping, wbc);
 421         }
 422         return ret;
 423 }
 424
 425 /**
 426  * stuffed_readpage - Fill in a Linux page with stuffed file data
 427  * @ip: the inode
 428  * @page: the page
 429  *
 430  * Returns: errno
 431  */
 432 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 433 {
 434         struct buffer_head *dibh;
 435         u64 dsize = i_size_read(&ip->i_inode);
 436         void *kaddr;
 437         int error;
 438
 439         /*
 440          * Due to the order of unstuffing files and ->fault(), we can be
 441          * asked for a zero page in the case of a stuffed file being extended,
 442          * so we need to supply one here. It doesn't happen often.
 443          */
 444         if (unlikely(page->index)) {
 445                 zero_user(page, 0, PAGE_SIZE);
 446                 SetPageUptodate(page);
 447                 return 0;
 448         }
 449
 450         error = gfs2_meta_inode_buffer(ip, &dibh);
 451         if (error)
 452                 return error;
 453
 454         kaddr = kmap_atomic(page);
 455         if (dsize > gfs2_max_stuffed_size(ip))
 456                 dsize = gfs2_max_stuffed_size(ip);
 457         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 458         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 459         kunmap_atomic(kaddr);
 460         flush_dcache_page(page);
 461         brelse(dibh);
 462         SetPageUptodate(page);
 463
 464         return 0;
 465 }
 466
 467
 468 static int __gfs2_readpage(void *file, struct page *page)
 469 {
 470         struct inode *inode = page->mapping->host;
 471         struct gfs2_inode *ip = GFS2_I(inode);
 472         struct gfs2_sbd *sdp = GFS2_SB(inode);
 473         int error;
 474
 475         if (!gfs2_is_jdata(ip) ||
 476             (i_blocksize(inode) == PAGE_SIZE && !page_has_buffers(page))) {
 477                 error = iomap_readpage(page, &gfs2_iomap_ops);
 478         } else if (gfs2_is_stuffed(ip)) {
 479                 error = stuffed_readpage(ip, page);
 480                 unlock_page(page);
 481         } else {
 482                 error = mpage_readpage(page, gfs2_block_map);
 483         }
 484
 485         if (unlikely(gfs2_withdrawn(sdp)))
 486                 return -EIO;
 487
 488         return error;
 489 }
 490
 491 /**
 492  * gfs2_readpage - read a page of a file
 493  * @file: The file to read
 494  * @page: The page of the file
 495  */
 496
 497 static int gfs2_readpage(struct file *file, struct page *page)
 498 {
 499         return __gfs2_readpage(file, page);
 500 }
 501
 502 /**
 503  * gfs2_internal_read - read an internal file
 504  * @ip: The gfs2 inode
 505  * @buf: The buffer to fill
 506  * @pos: The file position
 507  * @size: The amount to read
 508  *
 509  */
 510
 511 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 512                        unsigned size)
 513 {
 514         struct address_space *mapping = ip->i_inode.i_mapping;
 515         unsigned long index = *pos >> PAGE_SHIFT;
 516         unsigned offset = *pos & (PAGE_SIZE - 1);
 517         unsigned copied = 0;
 518         unsigned amt;
 519         struct page *page;
 520         void *p;
 521
 522         do {
 523                 amt = size - copied;
 524                 if (offset + size > PAGE_SIZE)
 525                         amt = PAGE_SIZE - offset;
 526                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 527                 if (IS_ERR(page))
 528                         return PTR_ERR(page);
 529                 p = kmap_atomic(page);
 530                 memcpy(buf + copied, p + offset, amt);
 531                 kunmap_atomic(p);
 532                 put_page(page);
 533                 copied += amt;
 534                 index++;
 535                 offset = 0;
 536         } while(copied < size);
 537         (*pos) += size;
 538         return size;
 539 }
 540
 541 /**
 542  * gfs2_readahead - Read a bunch of pages at once
 543  * @rac: Read-ahead control structure
 544  *
 545  * Some notes:
 546  * 1. This is only for readahead, so we can simply ignore any things
 547  *    which are slightly inconvenient (such as locking conflicts between
 548  *    the page lock and the glock) and return having done no I/O. Its
 549  *    obviously not something we'd want to do on too regular a basis.
 550  *    Any I/O we ignore at this time will be done via readpage later.
 551  * 2. We don't handle stuffed files here we let readpage do the honours.
 552  * 3. mpage_readahead() does most of the heavy lifting in the common case.
 553  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 554  */
 555
 556 static void gfs2_readahead(struct readahead_control *rac)
 557 {
 558         struct inode *inode = rac->mapping->host;
 559         struct gfs2_inode *ip = GFS2_I(inode);
 560
 561         if (gfs2_is_stuffed(ip))
 562                 ;
 563         else if (gfs2_is_jdata(ip))
 564                 mpage_readahead(rac, gfs2_block_map);
 565         else
 566                 iomap_readahead(rac, &gfs2_iomap_ops);
 567 }
 568
 569 /**
 570  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 571  * @inode: the rindex inode
 572  */
 573 void adjust_fs_space(struct inode *inode)
 574 {
 575         struct gfs2_sbd *sdp = GFS2_SB(inode);
 576         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 577         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 578         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 579         struct buffer_head *m_bh;
 580         u64 fs_total, new_free;
 581
 582         if (gfs2_trans_begin(sdp, 2 * RES_STATFS, 0) != 0)
 583                 return;
 584
 585         /* Total up the file system space, according to the latest rindex. */
 586         fs_total = gfs2_ri_total(sdp);
 587         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 588                 goto out;
 589
 590         spin_lock(&sdp->sd_statfs_spin);
 591         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 592                               sizeof(struct gfs2_dinode));
 593         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 594                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 595         else
 596                 new_free = 0;
 597         spin_unlock(&sdp->sd_statfs_spin);
 598         fs_warn(sdp, "File system extended by %llu blocks.\n",
 599                 (unsigned long long)new_free);
 600         gfs2_statfs_change(sdp, new_free, new_free, 0);
 601
 602         update_statfs(sdp, m_bh);
 603         brelse(m_bh);
 604 out:
 605         sdp->sd_rindex_uptodate = 0;
 606         gfs2_trans_end(sdp);
 607 }
 608
 609 /**
 610  * jdata_set_page_dirty - Page dirtying function
 611  * @page: The page to dirty
 612  *
 613  * Returns: 1 if it dirtyed the page, or 0 otherwise
 614  */
 615
 616 static int jdata_set_page_dirty(struct page *page)
 617 {
 618         if (current->journal_info)
 619                 SetPageChecked(page);
 620         return __set_page_dirty_buffers(page);
 621 }
 622
 623 /**
 624  * gfs2_bmap - Block map function
 625  * @mapping: Address space info
 626  * @lblock: The block to map
 627  *
 628  * Returns: The disk address for the block or 0 on hole or error
 629  */
 630
 631 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 632 {
 633         struct gfs2_inode *ip = GFS2_I(mapping->host);
 634         struct gfs2_holder i_gh;
 635         sector_t dblock = 0;
 636         int error;
 637
 638         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 639         if (error)
 640                 return 0;
 641
 642         if (!gfs2_is_stuffed(ip))
 643                 dblock = iomap_bmap(mapping, lblock, &gfs2_iomap_ops);
 644
 645         gfs2_glock_dq_uninit(&i_gh);
 646
 647         return dblock;
 648 }
 649
 650 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 651 {
 652         struct gfs2_bufdata *bd;
 653
 654         lock_buffer(bh);
 655         gfs2_log_lock(sdp);
 656         clear_buffer_dirty(bh);
 657         bd = bh->b_private;
 658         if (bd) {
 659                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 660                         list_del_init(&bd->bd_list);
 661                 else {
 662                         spin_lock(&sdp->sd_ail_lock);
 663                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
 664                         spin_unlock(&sdp->sd_ail_lock);
 665                 }
 666         }
 667         bh->b_bdev = NULL;
 668         clear_buffer_mapped(bh);
 669         clear_buffer_req(bh);
 670         clear_buffer_new(bh);
 671         gfs2_log_unlock(sdp);
 672         unlock_buffer(bh);
 673 }
 674
 675 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
 676                                 unsigned int length)
 677 {
 678         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 679         unsigned int stop = offset + length;
 680         int partial_page = (offset || length < PAGE_SIZE);
 681         struct buffer_head *bh, *head;
 682         unsigned long pos = 0;
 683
 684         BUG_ON(!PageLocked(page));
 685         if (!partial_page)
 686                 ClearPageChecked(page);
 687         if (!page_has_buffers(page))
 688                 goto out;
 689
 690         bh = head = page_buffers(page);
 691         do {
 692                 if (pos + bh->b_size > stop)
 693                         return;
 694
 695                 if (offset <= pos)
 696                         gfs2_discard(sdp, bh);
 697                 pos += bh->b_size;
 698                 bh = bh->b_this_page;
 699         } while (bh != head);
 700 out:
 701         if (!partial_page)
 702                 try_to_release_page(page, 0);
 703 }
 704
 705 /**
 706  * gfs2_releasepage - free the metadata associated with a page
 707  * @page: the page that's being released
 708  * @gfp_mask: passed from Linux VFS, ignored by us
 709  *
 710  * Calls try_to_free_buffers() to free the buffers and put the page if the
 711  * buffers can be released.
 712  *
 713  * Returns: 1 if the page was put or else 0
 714  */
 715
 716 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 717 {
 718         struct address_space *mapping = page->mapping;
 719         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 720         struct buffer_head *bh, *head;
 721         struct gfs2_bufdata *bd;
 722
 723         if (!page_has_buffers(page))
 724                 return 0;
 725
 726         /*
 727          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
 728          * clean pages might not have had the dirty bit cleared.  Thus, it can
 729          * send actual dirty pages to ->releasepage() via shrink_active_list().
 730          *
 731          * As a workaround, we skip pages that contain dirty buffers below.
 732          * Once ->releasepage isn't called on dirty pages anymore, we can warn
 733          * on dirty buffers like we used to here again.
 734          */
 735
 736         gfs2_log_lock(sdp);
 737         head = bh = page_buffers(page);
 738         do {
 739                 if (atomic_read(&bh->b_count))
 740                         goto cannot_release;
 741                 bd = bh->b_private;
 742                 if (bd && bd->bd_tr)
 743                         goto cannot_release;
 744                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
 745                         goto cannot_release;
 746                 bh = bh->b_this_page;
 747         } while(bh != head);
 748
 749         head = bh = page_buffers(page);
 750         do {
 751                 bd = bh->b_private;
 752                 if (bd) {
 753                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 754                         bd->bd_bh = NULL;
 755                         bh->b_private = NULL;
 756                         /*
 757                          * The bd may still be queued as a revoke, in which
 758                          * case we must not dequeue nor free it.
 759                          */
 760                         if (!bd->bd_blkno && !list_empty(&bd->bd_list))
 761                                 list_del_init(&bd->bd_list);
 762                         if (list_empty(&bd->bd_list))
 763                                 kmem_cache_free(gfs2_bufdata_cachep, bd);
 764                 }
 765
 766                 bh = bh->b_this_page;
 767         } while (bh != head);
 768         gfs2_log_unlock(sdp);
 769
 770         return try_to_free_buffers(page);
 771
 772 cannot_release:
 773         gfs2_log_unlock(sdp);
 774         return 0;
 775 }
 776
 777 static const struct address_space_operations gfs2_aops = {
 778         .writepage = gfs2_writepage,
 779         .writepages = gfs2_writepages,
 780         .readpage = gfs2_readpage,
 781         .readahead = gfs2_readahead,
 782         .set_page_dirty = __set_page_dirty_nobuffers,
 783         .releasepage = iomap_releasepage,
 784         .invalidatepage = iomap_invalidatepage,
 785         .bmap = gfs2_bmap,
 786         .direct_IO = noop_direct_IO,
 787         .migratepage = iomap_migrate_page,
 788         .is_partially_uptodate = iomap_is_partially_uptodate,
 789         .error_remove_page = generic_error_remove_page,
 790 };
 791
 792 static const struct address_space_operations gfs2_jdata_aops = {
 793         .writepage = gfs2_jdata_writepage,
 794         .writepages = gfs2_jdata_writepages,
 795         .readpage = gfs2_readpage,
 796         .readahead = gfs2_readahead,
 797         .set_page_dirty = jdata_set_page_dirty,
 798         .bmap = gfs2_bmap,
 799         .invalidatepage = gfs2_invalidatepage,
 800         .releasepage = gfs2_releasepage,
 801         .is_partially_uptodate = block_is_partially_uptodate,
 802         .error_remove_page = generic_error_remove_page,
 803 };
 804
 805 void gfs2_set_aops(struct inode *inode)
 806 {
 807         if (gfs2_is_jdata(GFS2_I(inode)))
 808                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
 809         else
 810                 inode->i_mapping->a_ops = &gfs2_aops;
 811 }