fs/gfs2/aops.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/pagevec.h>
  17 #include <linux/mpage.h>
  18 #include <linux/fs.h>
  19 #include <linux/writeback.h>
  20 #include <linux/swap.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/uio.h>
  24 #include <trace/events/writeback.h>
  25 #include <linux/sched/signal.h>
  26
  27 #include "gfs2.h"
  28 #include "incore.h"
  29 #include "bmap.h"
  30 #include "glock.h"
  31 #include "inode.h"
  32 #include "log.h"
  33 #include "meta_io.h"
  34 #include "quota.h"
  35 #include "trans.h"
  36 #include "rgrp.h"
  37 #include "super.h"
  38 #include "util.h"
  39 #include "glops.h"
  40 #include "aops.h"
  41
  42
  43 void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  44                             unsigned int from, unsigned int len)
  45 {
  46         struct buffer_head *head = page_buffers(page);
  47         unsigned int bsize = head->b_size;
  48         struct buffer_head *bh;
  49         unsigned int to = from + len;
  50         unsigned int start, end;
  51
  52         for (bh = head, start = 0; bh != head || !start;
  53              bh = bh->b_this_page, start = end) {
  54                 end = start + bsize;
  55                 if (end <= from)
  56                         continue;
  57                 if (start >= to)
  58                         break;
  59                 set_buffer_uptodate(bh);
  60                 gfs2_trans_add_data(ip->i_gl, bh);
  61         }
  62 }
  63
  64 /**
  65  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  66  * @inode: The inode
  67  * @lblock: The block number to look up
  68  * @bh_result: The buffer head to return the result in
  69  * @create: Non-zero if we may add block to the file
  70  *
  71  * Returns: errno
  72  */
  73
  74 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  75                                   struct buffer_head *bh_result, int create)
  76 {
  77         int error;
  78
  79         error = gfs2_block_map(inode, lblock, bh_result, 0);
  80         if (error)
  81                 return error;
  82         if (!buffer_mapped(bh_result))
  83                 return -EIO;
  84         return 0;
  85 }
  86
  87 /**
  88  * gfs2_writepage_common - Common bits of writepage
  89  * @page: The page to be written
  90  * @wbc: The writeback control
  91  *
  92  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  93  */
  94
  95 static int gfs2_writepage_common(struct page *page,
  96                                  struct writeback_control *wbc)
  97 {
  98         struct inode *inode = page->mapping->host;
  99         struct gfs2_inode *ip = GFS2_I(inode);
 100         struct gfs2_sbd *sdp = GFS2_SB(inode);
 101         loff_t i_size = i_size_read(inode);
 102         pgoff_t end_index = i_size >> PAGE_SHIFT;
 103         unsigned offset;
 104
 105         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 106                 goto out;
 107         if (current->journal_info)
 108                 goto redirty;
 109         /* Is the page fully outside i_size? (truncate in progress) */
 110         offset = i_size & (PAGE_SIZE-1);
 111         if (page->index > end_index || (page->index == end_index && !offset)) {
 112                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 113                 goto out;
 114         }
 115         return 1;
 116 redirty:
 117         redirty_page_for_writepage(wbc, page);
 118 out:
 119         unlock_page(page);
 120         return 0;
 121 }
 122
 123 /**
 124  * gfs2_writepage - Write page for writeback mappings
 125  * @page: The page
 126  * @wbc: The writeback control
 127  *
 128  */
 129
 130 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 131 {
 132         int ret;
 133
 134         ret = gfs2_writepage_common(page, wbc);
 135         if (ret <= 0)
 136                 return ret;
 137
 138         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 139 }
 140
 141 /* This is the same as calling block_write_full_page, but it also
 142  * writes pages outside of i_size
 143  */
 144 static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 145                                 struct writeback_control *wbc)
 146 {
 147         struct inode * const inode = page->mapping->host;
 148         loff_t i_size = i_size_read(inode);
 149         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 150         unsigned offset;
 151
 152         /*
 153          * The page straddles i_size.  It must be zeroed out on each and every
 154          * writepage invocation because it may be mmapped.  "A file is mapped
 155          * in multiples of the page size.  For a file that is not a multiple of
 156          * the  page size, the remaining memory is zeroed when mapped, and
 157          * writes to that region are not written out to the file."
 158          */
 159         offset = i_size & (PAGE_SIZE-1);
 160         if (page->index == end_index && offset)
 161                 zero_user_segment(page, offset, PAGE_SIZE);
 162
 163         return __block_write_full_page(inode, page, get_block, wbc,
 164                                        end_buffer_async_write);
 165 }
 166
 167 /**
 168  * __gfs2_jdata_writepage - The core of jdata writepage
 169  * @page: The page to write
 170  * @wbc: The writeback control
 171  *
 172  * This is shared between writepage and writepages and implements the
 173  * core of the writepage operation. If a transaction is required then
 174  * PageChecked will have been set and the transaction will have
 175  * already been started before this is called.
 176  */
 177
 178 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 179 {
 180         struct inode *inode = page->mapping->host;
 181         struct gfs2_inode *ip = GFS2_I(inode);
 182
 183         if (PageChecked(page)) {
 184                 ClearPageChecked(page);
 185                 if (!page_has_buffers(page)) {
 186                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 187                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 188                 }
 189                 gfs2_page_add_databufs(ip, page, 0, PAGE_SIZE);
 190         }
 191         return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
 192 }
 193
 194 /**
 195  * gfs2_jdata_writepage - Write complete page
 196  * @page: Page to write
 197  * @wbc: The writeback control
 198  *
 199  * Returns: errno
 200  *
 201  */
 202
 203 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 204 {
 205         struct inode *inode = page->mapping->host;
 206         struct gfs2_inode *ip = GFS2_I(inode);
 207         struct gfs2_sbd *sdp = GFS2_SB(inode);
 208         int ret;
 209
 210         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 211                 goto out;
 212         if (PageChecked(page) || current->journal_info)
 213                 goto out_ignore;
 214         ret = __gfs2_jdata_writepage(page, wbc);
 215         return ret;
 216
 217 out_ignore:
 218         redirty_page_for_writepage(wbc, page);
 219 out:
 220         unlock_page(page);
 221         return 0;
 222 }
 223
 224 /**
 225  * gfs2_writepages - Write a bunch of dirty pages back to disk
 226  * @mapping: The mapping to write
 227  * @wbc: Write-back control
 228  *
 229  * Used for both ordered and writeback modes.
 230  */
 231 static int gfs2_writepages(struct address_space *mapping,
 232                            struct writeback_control *wbc)
 233 {
 234         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 235         int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 236
 237         /*
 238          * Even if we didn't write any pages here, we might still be holding
 239          * dirty pages in the ail. We forcibly flush the ail because we don't
 240          * want balance_dirty_pages() to loop indefinitely trying to write out
 241          * pages held in the ail that it can't find.
 242          */
 243         if (ret == 0)
 244                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 245
 246         return ret;
 247 }
 248
 249 /**
 250  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 251  * @mapping: The mapping
 252  * @wbc: The writeback control
 253  * @pvec: The vector of pages
 254  * @nr_pages: The number of pages to write
 255  * @done_index: Page index
 256  *
 257  * Returns: non-zero if loop should terminate, zero otherwise
 258  */
 259
 260 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 261                                     struct writeback_control *wbc,
 262                                     struct pagevec *pvec,
 263                                     int nr_pages,
 264                                     pgoff_t *done_index)
 265 {
 266         struct inode *inode = mapping->host;
 267         struct gfs2_sbd *sdp = GFS2_SB(inode);
 268         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 269         int i;
 270         int ret;
 271
 272         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 273         if (ret < 0)
 274                 return ret;
 275
 276         for(i = 0; i < nr_pages; i++) {
 277                 struct page *page = pvec->pages[i];
 278
 279                 *done_index = page->index;
 280
 281                 lock_page(page);
 282
 283                 if (unlikely(page->mapping != mapping)) {
 284 continue_unlock:
 285                         unlock_page(page);
 286                         continue;
 287                 }
 288
 289                 if (!PageDirty(page)) {
 290                         /* someone wrote it for us */
 291                         goto continue_unlock;
 292                 }
 293
 294                 if (PageWriteback(page)) {
 295                         if (wbc->sync_mode != WB_SYNC_NONE)
 296                                 wait_on_page_writeback(page);
 297                         else
 298                                 goto continue_unlock;
 299                 }
 300
 301                 BUG_ON(PageWriteback(page));
 302                 if (!clear_page_dirty_for_io(page))
 303                         goto continue_unlock;
 304
 305                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 306
 307                 ret = __gfs2_jdata_writepage(page, wbc);
 308                 if (unlikely(ret)) {
 309                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 310                                 unlock_page(page);
 311                                 ret = 0;
 312                         } else {
 313
 314                                 /*
 315                                  * done_index is set past this page,
 316                                  * so media errors will not choke
 317                                  * background writeout for the entire
 318                                  * file. This has consequences for
 319                                  * range_cyclic semantics (ie. it may
 320                                  * not be suitable for data integrity
 321                                  * writeout).
 322                                  */
 323                                 *done_index = page->index + 1;
 324                                 ret = 1;
 325                                 break;
 326                         }
 327                 }
 328
 329                 /*
 330                  * We stop writing back only if we are not doing
 331                  * integrity sync. In case of integrity sync we have to
 332                  * keep going until we have written all the pages
 333                  * we tagged for writeback prior to entering this loop.
 334                  */
 335                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 336                         ret = 1;
 337                         break;
 338                 }
 339
 340         }
 341         gfs2_trans_end(sdp);
 342         return ret;
 343 }
 344
 345 /**
 346  * gfs2_write_cache_jdata - Like write_cache_pages but different
 347  * @mapping: The mapping to write
 348  * @wbc: The writeback control
 349  *
 350  * The reason that we use our own function here is that we need to
 351  * start transactions before we grab page locks. This allows us
 352  * to get the ordering right.
 353  */
 354
 355 static int gfs2_write_cache_jdata(struct address_space *mapping,
 356                                   struct writeback_control *wbc)
 357 {
 358         int ret = 0;
 359         int done = 0;
 360         struct pagevec pvec;
 361         int nr_pages;
 362         pgoff_t writeback_index;
 363         pgoff_t index;
 364         pgoff_t end;
 365         pgoff_t done_index;
 366         int cycled;
 367         int range_whole = 0;
 368         int tag;
 369
 370         pagevec_init(&pvec);
 371         if (wbc->range_cyclic) {
 372                 writeback_index = mapping->writeback_index; /* prev offset */
 373                 index = writeback_index;
 374                 if (index == 0)
 375                         cycled = 1;
 376                 else
 377                         cycled = 0;
 378                 end = -1;
 379         } else {
 380                 index = wbc->range_start >> PAGE_SHIFT;
 381                 end = wbc->range_end >> PAGE_SHIFT;
 382                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 383                         range_whole = 1;
 384                 cycled = 1; /* ignore range_cyclic tests */
 385         }
 386         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 387                 tag = PAGECACHE_TAG_TOWRITE;
 388         else
 389                 tag = PAGECACHE_TAG_DIRTY;
 390
 391 retry:
 392         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 393                 tag_pages_for_writeback(mapping, index, end);
 394         done_index = index;
 395         while (!done && (index <= end)) {
 396                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
 397                                 tag);
 398                 if (nr_pages == 0)
 399                         break;
 400
 401                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
 402                 if (ret)
 403                         done = 1;
 404                 if (ret > 0)
 405                         ret = 0;
 406                 pagevec_release(&pvec);
 407                 cond_resched();
 408         }
 409
 410         if (!cycled && !done) {
 411                 /*
 412                  * range_cyclic:
 413                  * We hit the last page and there is more work to be done: wrap
 414                  * back to the start of the file
 415                  */
 416                 cycled = 1;
 417                 index = 0;
 418                 end = writeback_index - 1;
 419                 goto retry;
 420         }
 421
 422         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 423                 mapping->writeback_index = done_index;
 424
 425         return ret;
 426 }
 427
 428
 429 /**
 430  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 431  * @mapping: The mapping to write
 432  * @wbc: The writeback control
 433  *
 434  */
 435
 436 static int gfs2_jdata_writepages(struct address_space *mapping,
 437                                  struct writeback_control *wbc)
 438 {
 439         struct gfs2_inode *ip = GFS2_I(mapping->host);
 440         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 441         int ret;
 442
 443         ret = gfs2_write_cache_jdata(mapping, wbc);
 444         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 445                 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
 446                                GFS2_LFC_JDATA_WPAGES);
 447                 ret = gfs2_write_cache_jdata(mapping, wbc);
 448         }
 449         return ret;
 450 }
 451
 452 /**
 453  * stuffed_readpage - Fill in a Linux page with stuffed file data
 454  * @ip: the inode
 455  * @page: the page
 456  *
 457  * Returns: errno
 458  */
 459
 460 int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 461 {
 462         struct buffer_head *dibh;
 463         u64 dsize = i_size_read(&ip->i_inode);
 464         void *kaddr;
 465         int error;
 466
 467         /*
 468          * Due to the order of unstuffing files and ->fault(), we can be
 469          * asked for a zero page in the case of a stuffed file being extended,
 470          * so we need to supply one here. It doesn't happen often.
 471          */
 472         if (unlikely(page->index)) {
 473                 zero_user(page, 0, PAGE_SIZE);
 474                 SetPageUptodate(page);
 475                 return 0;
 476         }
 477
 478         error = gfs2_meta_inode_buffer(ip, &dibh);
 479         if (error)
 480                 return error;
 481
 482         kaddr = kmap_atomic(page);
 483         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 484         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 485         kunmap_atomic(kaddr);
 486         flush_dcache_page(page);
 487         brelse(dibh);
 488         SetPageUptodate(page);
 489
 490         return 0;
 491 }
 492
 493
 494 /**
 495  * __gfs2_readpage - readpage
 496  * @file: The file to read a page for
 497  * @page: The page to read
 498  *
 499  * This is the core of gfs2's readpage. It's used by the internal file
 500  * reading code as in that case we already hold the glock. Also it's
 501  * called by gfs2_readpage() once the required lock has been granted.
 502  */
 503
 504 static int __gfs2_readpage(void *file, struct page *page)
 505 {
 506         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 507         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 508
 509         int error;
 510
 511         if (i_blocksize(page->mapping->host) == PAGE_SIZE &&
 512             !page_has_buffers(page)) {
 513                 error = iomap_readpage(page, &gfs2_iomap_ops);
 514         } else if (gfs2_is_stuffed(ip)) {
 515                 error = stuffed_readpage(ip, page);
 516                 unlock_page(page);
 517         } else {
 518                 error = mpage_readpage(page, gfs2_block_map);
 519         }
 520
 521         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 522                 return -EIO;
 523
 524         return error;
 525 }
 526
 527 /**
 528  * gfs2_readpage - read a page of a file
 529  * @file: The file to read
 530  * @page: The page of the file
 531  *
 532  * This deals with the locking required. We have to unlock and
 533  * relock the page in order to get the locking in the right
 534  * order.
 535  */
 536
 537 static int gfs2_readpage(struct file *file, struct page *page)
 538 {
 539         struct address_space *mapping = page->mapping;
 540         struct gfs2_inode *ip = GFS2_I(mapping->host);
 541         struct gfs2_holder gh;
 542         int error;
 543
 544         unlock_page(page);
 545         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 546         error = gfs2_glock_nq(&gh);
 547         if (unlikely(error))
 548                 goto out;
 549         error = AOP_TRUNCATED_PAGE;
 550         lock_page(page);
 551         if (page->mapping == mapping && !PageUptodate(page))
 552                 error = __gfs2_readpage(file, page);
 553         else
 554                 unlock_page(page);
 555         gfs2_glock_dq(&gh);
 556 out:
 557         gfs2_holder_uninit(&gh);
 558         if (error && error != AOP_TRUNCATED_PAGE)
 559                 lock_page(page);
 560         return error;
 561 }
 562
 563 /**
 564  * gfs2_internal_read - read an internal file
 565  * @ip: The gfs2 inode
 566  * @buf: The buffer to fill
 567  * @pos: The file position
 568  * @size: The amount to read
 569  *
 570  */
 571
 572 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 573                        unsigned size)
 574 {
 575         struct address_space *mapping = ip->i_inode.i_mapping;
 576         unsigned long index = *pos / PAGE_SIZE;
 577         unsigned offset = *pos & (PAGE_SIZE - 1);
 578         unsigned copied = 0;
 579         unsigned amt;
 580         struct page *page;
 581         void *p;
 582
 583         do {
 584                 amt = size - copied;
 585                 if (offset + size > PAGE_SIZE)
 586                         amt = PAGE_SIZE - offset;
 587                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 588                 if (IS_ERR(page))
 589                         return PTR_ERR(page);
 590                 p = kmap_atomic(page);
 591                 memcpy(buf + copied, p + offset, amt);
 592                 kunmap_atomic(p);
 593                 put_page(page);
 594                 copied += amt;
 595                 index++;
 596                 offset = 0;
 597         } while(copied < size);
 598         (*pos) += size;
 599         return size;
 600 }
 601
 602 /**
 603  * gfs2_readpages - Read a bunch of pages at once
 604  * @file: The file to read from
 605  * @mapping: Address space info
 606  * @pages: List of pages to read
 607  * @nr_pages: Number of pages to read
 608  *
 609  * Some notes:
 610  * 1. This is only for readahead, so we can simply ignore any things
 611  *    which are slightly inconvenient (such as locking conflicts between
 612  *    the page lock and the glock) and return having done no I/O. Its
 613  *    obviously not something we'd want to do on too regular a basis.
 614  *    Any I/O we ignore at this time will be done via readpage later.
 615  * 2. We don't handle stuffed files here we let readpage do the honours.
 616  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 617  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 618  */
 619
 620 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 621                           struct list_head *pages, unsigned nr_pages)
 622 {
 623         struct inode *inode = mapping->host;
 624         struct gfs2_inode *ip = GFS2_I(inode);
 625         struct gfs2_sbd *sdp = GFS2_SB(inode);
 626         struct gfs2_holder gh;
 627         int ret;
 628
 629         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 630         ret = gfs2_glock_nq(&gh);
 631         if (unlikely(ret))
 632                 goto out_uninit;
 633         if (!gfs2_is_stuffed(ip))
 634                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 635         gfs2_glock_dq(&gh);
 636 out_uninit:
 637         gfs2_holder_uninit(&gh);
 638         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 639                 ret = -EIO;
 640         return ret;
 641 }
 642
 643 /**
 644  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 645  * @inode: the rindex inode
 646  */
 647 void adjust_fs_space(struct inode *inode)
 648 {
 649         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 650         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 651         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 652         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 653         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 654         struct buffer_head *m_bh, *l_bh;
 655         u64 fs_total, new_free;
 656
 657         /* Total up the file system space, according to the latest rindex. */
 658         fs_total = gfs2_ri_total(sdp);
 659         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 660                 return;
 661
 662         spin_lock(&sdp->sd_statfs_spin);
 663         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 664                               sizeof(struct gfs2_dinode));
 665         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 666                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 667         else
 668                 new_free = 0;
 669         spin_unlock(&sdp->sd_statfs_spin);
 670         fs_warn(sdp, "File system extended by %llu blocks.\n",
 671                 (unsigned long long)new_free);
 672         gfs2_statfs_change(sdp, new_free, new_free, 0);
 673
 674         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 675                 goto out;
 676         update_statfs(sdp, m_bh, l_bh);
 677         brelse(l_bh);
 678 out:
 679         brelse(m_bh);
 680 }
 681
 682 /**
 683  * gfs2_stuffed_write_end - Write end for stuffed files
 684  * @inode: The inode
 685  * @dibh: The buffer_head containing the on-disk inode
 686  * @pos: The file position
 687  * @copied: How much was actually copied by the VFS
 688  * @page: The page
 689  *
 690  * This copies the data from the page into the inode block after
 691  * the inode data structure itself.
 692  *
 693  * Returns: copied bytes or errno
 694  */
 695 int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 696                            loff_t pos, unsigned copied,
 697                            struct page *page)
 698 {
 699         struct gfs2_inode *ip = GFS2_I(inode);
 700         u64 to = pos + copied;
 701         void *kaddr;
 702         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 703
 704         BUG_ON(pos + copied > gfs2_max_stuffed_size(ip));
 705
 706         kaddr = kmap_atomic(page);
 707         memcpy(buf + pos, kaddr + pos, copied);
 708         flush_dcache_page(page);
 709         kunmap_atomic(kaddr);
 710
 711         WARN_ON(!PageUptodate(page));
 712         unlock_page(page);
 713         put_page(page);
 714
 715         if (copied) {
 716                 if (inode->i_size < to)
 717                         i_size_write(inode, to);
 718                 mark_inode_dirty(inode);
 719         }
 720         return copied;
 721 }
 722
 723 /**
 724  * jdata_set_page_dirty - Page dirtying function
 725  * @page: The page to dirty
 726  *
 727  * Returns: 1 if it dirtyed the page, or 0 otherwise
 728  */
 729
 730 static int jdata_set_page_dirty(struct page *page)
 731 {
 732         SetPageChecked(page);
 733         return __set_page_dirty_buffers(page);
 734 }
 735
 736 /**
 737  * gfs2_bmap - Block map function
 738  * @mapping: Address space info
 739  * @lblock: The block to map
 740  *
 741  * Returns: The disk address for the block or 0 on hole or error
 742  */
 743
 744 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 745 {
 746         struct gfs2_inode *ip = GFS2_I(mapping->host);
 747         struct gfs2_holder i_gh;
 748         sector_t dblock = 0;
 749         int error;
 750
 751         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 752         if (error)
 753                 return 0;
 754
 755         if (!gfs2_is_stuffed(ip))
 756                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 757
 758         gfs2_glock_dq_uninit(&i_gh);
 759
 760         return dblock;
 761 }
 762
 763 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 764 {
 765         struct gfs2_bufdata *bd;
 766
 767         lock_buffer(bh);
 768         gfs2_log_lock(sdp);
 769         clear_buffer_dirty(bh);
 770         bd = bh->b_private;
 771         if (bd) {
 772                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 773                         list_del_init(&bd->bd_list);
 774                 else
 775                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
 776         }
 777         bh->b_bdev = NULL;
 778         clear_buffer_mapped(bh);
 779         clear_buffer_req(bh);
 780         clear_buffer_new(bh);
 781         gfs2_log_unlock(sdp);
 782         unlock_buffer(bh);
 783 }
 784
 785 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
 786                                 unsigned int length)
 787 {
 788         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 789         unsigned int stop = offset + length;
 790         int partial_page = (offset || length < PAGE_SIZE);
 791         struct buffer_head *bh, *head;
 792         unsigned long pos = 0;
 793
 794         BUG_ON(!PageLocked(page));
 795         if (!partial_page)
 796                 ClearPageChecked(page);
 797         if (!page_has_buffers(page))
 798                 goto out;
 799
 800         bh = head = page_buffers(page);
 801         do {
 802                 if (pos + bh->b_size > stop)
 803                         return;
 804
 805                 if (offset <= pos)
 806                         gfs2_discard(sdp, bh);
 807                 pos += bh->b_size;
 808                 bh = bh->b_this_page;
 809         } while (bh != head);
 810 out:
 811         if (!partial_page)
 812                 try_to_release_page(page, 0);
 813 }
 814
 815 /**
 816  * gfs2_releasepage - free the metadata associated with a page
 817  * @page: the page that's being released
 818  * @gfp_mask: passed from Linux VFS, ignored by us
 819  *
 820  * Call try_to_free_buffers() if the buffers in this page can be
 821  * released.
 822  *
 823  * Returns: 0
 824  */
 825
 826 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 827 {
 828         struct address_space *mapping = page->mapping;
 829         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 830         struct buffer_head *bh, *head;
 831         struct gfs2_bufdata *bd;
 832
 833         if (!page_has_buffers(page))
 834                 return 0;
 835
 836         /*
 837          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
 838          * clean pages might not have had the dirty bit cleared.  Thus, it can
 839          * send actual dirty pages to ->releasepage() via shrink_active_list().
 840          *
 841          * As a workaround, we skip pages that contain dirty buffers below.
 842          * Once ->releasepage isn't called on dirty pages anymore, we can warn
 843          * on dirty buffers like we used to here again.
 844          */
 845
 846         gfs2_log_lock(sdp);
 847         spin_lock(&sdp->sd_ail_lock);
 848         head = bh = page_buffers(page);
 849         do {
 850                 if (atomic_read(&bh->b_count))
 851                         goto cannot_release;
 852                 bd = bh->b_private;
 853                 if (bd && bd->bd_tr)
 854                         goto cannot_release;
 855                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
 856                         goto cannot_release;
 857                 bh = bh->b_this_page;
 858         } while(bh != head);
 859         spin_unlock(&sdp->sd_ail_lock);
 860
 861         head = bh = page_buffers(page);
 862         do {
 863                 bd = bh->b_private;
 864                 if (bd) {
 865                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 866                         if (!list_empty(&bd->bd_list))
 867                                 list_del_init(&bd->bd_list);
 868                         bd->bd_bh = NULL;
 869                         bh->b_private = NULL;
 870                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 871                 }
 872
 873                 bh = bh->b_this_page;
 874         } while (bh != head);
 875         gfs2_log_unlock(sdp);
 876
 877         return try_to_free_buffers(page);
 878
 879 cannot_release:
 880         spin_unlock(&sdp->sd_ail_lock);
 881         gfs2_log_unlock(sdp);
 882         return 0;
 883 }
 884
 885 static const struct address_space_operations gfs2_writeback_aops = {
 886         .writepage = gfs2_writepage,
 887         .writepages = gfs2_writepages,
 888         .readpage = gfs2_readpage,
 889         .readpages = gfs2_readpages,
 890         .bmap = gfs2_bmap,
 891         .invalidatepage = gfs2_invalidatepage,
 892         .releasepage = gfs2_releasepage,
 893         .direct_IO = noop_direct_IO,
 894         .migratepage = buffer_migrate_page,
 895         .is_partially_uptodate = block_is_partially_uptodate,
 896         .error_remove_page = generic_error_remove_page,
 897 };
 898
 899 static const struct address_space_operations gfs2_ordered_aops = {
 900         .writepage = gfs2_writepage,
 901         .writepages = gfs2_writepages,
 902         .readpage = gfs2_readpage,
 903         .readpages = gfs2_readpages,
 904         .set_page_dirty = __set_page_dirty_buffers,
 905         .bmap = gfs2_bmap,
 906         .invalidatepage = gfs2_invalidatepage,
 907         .releasepage = gfs2_releasepage,
 908         .direct_IO = noop_direct_IO,
 909         .migratepage = buffer_migrate_page,
 910         .is_partially_uptodate = block_is_partially_uptodate,
 911         .error_remove_page = generic_error_remove_page,
 912 };
 913
 914 static const struct address_space_operations gfs2_jdata_aops = {
 915         .writepage = gfs2_jdata_writepage,
 916         .writepages = gfs2_jdata_writepages,
 917         .readpage = gfs2_readpage,
 918         .readpages = gfs2_readpages,
 919         .set_page_dirty = jdata_set_page_dirty,
 920         .bmap = gfs2_bmap,
 921         .invalidatepage = gfs2_invalidatepage,
 922         .releasepage = gfs2_releasepage,
 923         .is_partially_uptodate = block_is_partially_uptodate,
 924         .error_remove_page = generic_error_remove_page,
 925 };
 926
 927 void gfs2_set_aops(struct inode *inode)
 928 {
 929         struct gfs2_inode *ip = GFS2_I(inode);
 930
 931         if (gfs2_is_writeback(ip))
 932                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
 933         else if (gfs2_is_ordered(ip))
 934                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
 935         else if (gfs2_is_jdata(ip))
 936                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
 937         else
 938                 BUG();
 939 }
 940