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
  26 #include "gfs2.h"
  27 #include "incore.h"
  28 #include "bmap.h"
  29 #include "glock.h"
  30 #include "inode.h"
  31 #include "log.h"
  32 #include "meta_io.h"
  33 #include "quota.h"
  34 #include "trans.h"
  35 #include "rgrp.h"
  36 #include "super.h"
  37 #include "util.h"
  38 #include "glops.h"
  39
  40
  41 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  42                                    unsigned int from, unsigned int to)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  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 || start >= to)
  53                         continue;
  54                 if (gfs2_is_jdata(ip))
  55                         set_buffer_uptodate(bh);
  56                 gfs2_trans_add_data(ip->i_gl, bh);
  57         }
  58 }
  59
  60 /**
  61  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  62  * @inode: The inode
  63  * @lblock: The block number to look up
  64  * @bh_result: The buffer head to return the result in
  65  * @create: Non-zero if we may add block to the file
  66  *
  67  * Returns: errno
  68  */
  69
  70 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  71                                   struct buffer_head *bh_result, int create)
  72 {
  73         int error;
  74
  75         error = gfs2_block_map(inode, lblock, bh_result, 0);
  76         if (error)
  77                 return error;
  78         if (!buffer_mapped(bh_result))
  79                 return -EIO;
  80         return 0;
  81 }
  82
  83 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  84                                  struct buffer_head *bh_result, int create)
  85 {
  86         return gfs2_block_map(inode, lblock, bh_result, 0);
  87 }
  88
  89 /**
  90  * gfs2_writepage_common - Common bits of writepage
  91  * @page: The page to be written
  92  * @wbc: The writeback control
  93  *
  94  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  95  */
  96
  97 static int gfs2_writepage_common(struct page *page,
  98                                  struct writeback_control *wbc)
  99 {
 100         struct inode *inode = page->mapping->host;
 101         struct gfs2_inode *ip = GFS2_I(inode);
 102         struct gfs2_sbd *sdp = GFS2_SB(inode);
 103         loff_t i_size = i_size_read(inode);
 104         pgoff_t end_index = i_size >> PAGE_SHIFT;
 105         unsigned offset;
 106
 107         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 108                 goto out;
 109         if (current->journal_info)
 110                 goto redirty;
 111         /* Is the page fully outside i_size? (truncate in progress) */
 112         offset = i_size & (PAGE_SIZE-1);
 113         if (page->index > end_index || (page->index == end_index && !offset)) {
 114                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 115                 goto out;
 116         }
 117         return 1;
 118 redirty:
 119         redirty_page_for_writepage(wbc, page);
 120 out:
 121         unlock_page(page);
 122         return 0;
 123 }
 124
 125 /**
 126  * gfs2_writepage - Write page for writeback mappings
 127  * @page: The page
 128  * @wbc: The writeback control
 129  *
 130  */
 131
 132 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 133 {
 134         int ret;
 135
 136         ret = gfs2_writepage_common(page, wbc);
 137         if (ret <= 0)
 138                 return ret;
 139
 140         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 141 }
 142
 143 /* This is the same as calling block_write_full_page, but it also
 144  * writes pages outside of i_size
 145  */
 146 static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 147                                 struct writeback_control *wbc)
 148 {
 149         struct inode * const inode = page->mapping->host;
 150         loff_t i_size = i_size_read(inode);
 151         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 152         unsigned offset;
 153
 154         /*
 155          * The page straddles i_size.  It must be zeroed out on each and every
 156          * writepage invocation because it may be mmapped.  "A file is mapped
 157          * in multiples of the page size.  For a file that is not a multiple of
 158          * the  page size, the remaining memory is zeroed when mapped, and
 159          * writes to that region are not written out to the file."
 160          */
 161         offset = i_size & (PAGE_SIZE-1);
 162         if (page->index == end_index && offset)
 163                 zero_user_segment(page, offset, PAGE_SIZE);
 164
 165         return __block_write_full_page(inode, page, get_block, wbc,
 166                                        end_buffer_async_write);
 167 }
 168
 169 /**
 170  * __gfs2_jdata_writepage - The core of jdata writepage
 171  * @page: The page to write
 172  * @wbc: The writeback control
 173  *
 174  * This is shared between writepage and writepages and implements the
 175  * core of the writepage operation. If a transaction is required then
 176  * PageChecked will have been set and the transaction will have
 177  * already been started before this is called.
 178  */
 179
 180 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 181 {
 182         struct inode *inode = page->mapping->host;
 183         struct gfs2_inode *ip = GFS2_I(inode);
 184         struct gfs2_sbd *sdp = GFS2_SB(inode);
 185
 186         if (PageChecked(page)) {
 187                 ClearPageChecked(page);
 188                 if (!page_has_buffers(page)) {
 189                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 190                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 191                 }
 192                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 193         }
 194         return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
 195 }
 196
 197 /**
 198  * gfs2_jdata_writepage - Write complete page
 199  * @page: Page to write
 200  * @wbc: The writeback control
 201  *
 202  * Returns: errno
 203  *
 204  */
 205
 206 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 207 {
 208         struct inode *inode = page->mapping->host;
 209         struct gfs2_inode *ip = GFS2_I(inode);
 210         struct gfs2_sbd *sdp = GFS2_SB(inode);
 211         int ret;
 212
 213         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 214                 goto out;
 215         if (PageChecked(page) || current->journal_info)
 216                 goto out_ignore;
 217         ret = __gfs2_jdata_writepage(page, wbc);
 218         return ret;
 219
 220 out_ignore:
 221         redirty_page_for_writepage(wbc, page);
 222 out:
 223         unlock_page(page);
 224         return 0;
 225 }
 226
 227 /**
 228  * gfs2_writepages - Write a bunch of dirty pages back to disk
 229  * @mapping: The mapping to write
 230  * @wbc: Write-back control
 231  *
 232  * Used for both ordered and writeback modes.
 233  */
 234 static int gfs2_writepages(struct address_space *mapping,
 235                            struct writeback_control *wbc)
 236 {
 237         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 238         int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 239
 240         /*
 241          * Even if we didn't write any pages here, we might still be holding
 242          * dirty pages in the ail. We forcibly flush the ail because we don't
 243          * want balance_dirty_pages() to loop indefinitely trying to write out
 244          * pages held in the ail that it can't find.
 245          */
 246         if (ret == 0)
 247                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 248
 249         return ret;
 250 }
 251
 252 /**
 253  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 254  * @mapping: The mapping
 255  * @wbc: The writeback control
 256  * @pvec: The vector of pages
 257  * @nr_pages: The number of pages to write
 258  * @end: End position
 259  * @done_index: Page index
 260  *
 261  * Returns: non-zero if loop should terminate, zero otherwise
 262  */
 263
 264 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 265                                     struct writeback_control *wbc,
 266                                     struct pagevec *pvec,
 267                                     int nr_pages, pgoff_t end,
 268                                     pgoff_t *done_index)
 269 {
 270         struct inode *inode = mapping->host;
 271         struct gfs2_sbd *sdp = GFS2_SB(inode);
 272         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 273         int i;
 274         int ret;
 275
 276         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 277         if (ret < 0)
 278                 return ret;
 279
 280         for(i = 0; i < nr_pages; i++) {
 281                 struct page *page = pvec->pages[i];
 282
 283                 *done_index = page->index;
 284
 285                 lock_page(page);
 286
 287                 if (unlikely(page->mapping != mapping)) {
 288 continue_unlock:
 289                         unlock_page(page);
 290                         continue;
 291                 }
 292
 293                 if (!PageDirty(page)) {
 294                         /* someone wrote it for us */
 295                         goto continue_unlock;
 296                 }
 297
 298                 if (PageWriteback(page)) {
 299                         if (wbc->sync_mode != WB_SYNC_NONE)
 300                                 wait_on_page_writeback(page);
 301                         else
 302                                 goto continue_unlock;
 303                 }
 304
 305                 BUG_ON(PageWriteback(page));
 306                 if (!clear_page_dirty_for_io(page))
 307                         goto continue_unlock;
 308
 309                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 310
 311                 ret = __gfs2_jdata_writepage(page, wbc);
 312                 if (unlikely(ret)) {
 313                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 314                                 unlock_page(page);
 315                                 ret = 0;
 316                         } else {
 317
 318                                 /*
 319                                  * done_index is set past this page,
 320                                  * so media errors will not choke
 321                                  * background writeout for the entire
 322                                  * file. This has consequences for
 323                                  * range_cyclic semantics (ie. it may
 324                                  * not be suitable for data integrity
 325                                  * writeout).
 326                                  */
 327                                 *done_index = page->index + 1;
 328                                 ret = 1;
 329                                 break;
 330                         }
 331                 }
 332
 333                 /*
 334                  * We stop writing back only if we are not doing
 335                  * integrity sync. In case of integrity sync we have to
 336                  * keep going until we have written all the pages
 337                  * we tagged for writeback prior to entering this loop.
 338                  */
 339                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 340                         ret = 1;
 341                         break;
 342                 }
 343
 344         }
 345         gfs2_trans_end(sdp);
 346         return ret;
 347 }
 348
 349 /**
 350  * gfs2_write_cache_jdata - Like write_cache_pages but different
 351  * @mapping: The mapping to write
 352  * @wbc: The writeback control
 353  *
 354  * The reason that we use our own function here is that we need to
 355  * start transactions before we grab page locks. This allows us
 356  * to get the ordering right.
 357  */
 358
 359 static int gfs2_write_cache_jdata(struct address_space *mapping,
 360                                   struct writeback_control *wbc)
 361 {
 362         int ret = 0;
 363         int done = 0;
 364         struct pagevec pvec;
 365         int nr_pages;
 366         pgoff_t uninitialized_var(writeback_index);
 367         pgoff_t index;
 368         pgoff_t end;
 369         pgoff_t done_index;
 370         int cycled;
 371         int range_whole = 0;
 372         int tag;
 373
 374         pagevec_init(&pvec, 0);
 375         if (wbc->range_cyclic) {
 376                 writeback_index = mapping->writeback_index; /* prev offset */
 377                 index = writeback_index;
 378                 if (index == 0)
 379                         cycled = 1;
 380                 else
 381                         cycled = 0;
 382                 end = -1;
 383         } else {
 384                 index = wbc->range_start >> PAGE_SHIFT;
 385                 end = wbc->range_end >> PAGE_SHIFT;
 386                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 387                         range_whole = 1;
 388                 cycled = 1; /* ignore range_cyclic tests */
 389         }
 390         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 391                 tag = PAGECACHE_TAG_TOWRITE;
 392         else
 393                 tag = PAGECACHE_TAG_DIRTY;
 394
 395 retry:
 396         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 397                 tag_pages_for_writeback(mapping, index, end);
 398         done_index = index;
 399         while (!done && (index <= end)) {
 400                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
 401                                 tag, PAGEVEC_SIZE);
 402                 if (nr_pages == 0)
 403                         break;
 404
 405                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end, &done_index);
 406                 if (ret)
 407                         done = 1;
 408                 if (ret > 0)
 409                         ret = 0;
 410                 pagevec_release(&pvec);
 411                 cond_resched();
 412         }
 413
 414         if (!cycled && !done) {
 415                 /*
 416                  * range_cyclic:
 417                  * We hit the last page and there is more work to be done: wrap
 418                  * back to the start of the file
 419                  */
 420                 cycled = 1;
 421                 index = 0;
 422                 end = writeback_index - 1;
 423                 goto retry;
 424         }
 425
 426         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 427                 mapping->writeback_index = done_index;
 428
 429         return ret;
 430 }
 431
 432
 433 /**
 434  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 435  * @mapping: The mapping to write
 436  * @wbc: The writeback control
 437  *
 438  */
 439
 440 static int gfs2_jdata_writepages(struct address_space *mapping,
 441                                  struct writeback_control *wbc)
 442 {
 443         struct gfs2_inode *ip = GFS2_I(mapping->host);
 444         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 445         int ret;
 446
 447         ret = gfs2_write_cache_jdata(mapping, wbc);
 448         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 449                 gfs2_log_flush(sdp, ip->i_gl, NORMAL_FLUSH);
 450                 ret = gfs2_write_cache_jdata(mapping, wbc);
 451         }
 452         return ret;
 453 }
 454
 455 /**
 456  * stuffed_readpage - Fill in a Linux page with stuffed file data
 457  * @ip: the inode
 458  * @page: the page
 459  *
 460  * Returns: errno
 461  */
 462
 463 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 464 {
 465         struct buffer_head *dibh;
 466         u64 dsize = i_size_read(&ip->i_inode);
 467         void *kaddr;
 468         int error;
 469
 470         /*
 471          * Due to the order of unstuffing files and ->fault(), we can be
 472          * asked for a zero page in the case of a stuffed file being extended,
 473          * so we need to supply one here. It doesn't happen often.
 474          */
 475         if (unlikely(page->index)) {
 476                 zero_user(page, 0, PAGE_SIZE);
 477                 SetPageUptodate(page);
 478                 return 0;
 479         }
 480
 481         error = gfs2_meta_inode_buffer(ip, &dibh);
 482         if (error)
 483                 return error;
 484
 485         kaddr = kmap_atomic(page);
 486         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 487                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 488         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 489         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 490         kunmap_atomic(kaddr);
 491         flush_dcache_page(page);
 492         brelse(dibh);
 493         SetPageUptodate(page);
 494
 495         return 0;
 496 }
 497
 498
 499 /**
 500  * __gfs2_readpage - readpage
 501  * @file: The file to read a page for
 502  * @page: The page to read
 503  *
 504  * This is the core of gfs2's readpage. Its used by the internal file
 505  * reading code as in that case we already hold the glock. Also its
 506  * called by gfs2_readpage() once the required lock has been granted.
 507  *
 508  */
 509
 510 static int __gfs2_readpage(void *file, struct page *page)
 511 {
 512         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 513         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 514         int error;
 515
 516         if (gfs2_is_stuffed(ip)) {
 517                 error = stuffed_readpage(ip, page);
 518                 unlock_page(page);
 519         } else {
 520                 error = mpage_readpage(page, gfs2_block_map);
 521         }
 522
 523         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 524                 return -EIO;
 525
 526         return error;
 527 }
 528
 529 /**
 530  * gfs2_readpage - read a page of a file
 531  * @file: The file to read
 532  * @page: The page of the file
 533  *
 534  * This deals with the locking required. We have to unlock and
 535  * relock the page in order to get the locking in the right
 536  * order.
 537  */
 538
 539 static int gfs2_readpage(struct file *file, struct page *page)
 540 {
 541         struct address_space *mapping = page->mapping;
 542         struct gfs2_inode *ip = GFS2_I(mapping->host);
 543         struct gfs2_holder gh;
 544         int error;
 545
 546         unlock_page(page);
 547         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 548         error = gfs2_glock_nq(&gh);
 549         if (unlikely(error))
 550                 goto out;
 551         error = AOP_TRUNCATED_PAGE;
 552         lock_page(page);
 553         if (page->mapping == mapping && !PageUptodate(page))
 554                 error = __gfs2_readpage(file, page);
 555         else
 556                 unlock_page(page);
 557         gfs2_glock_dq(&gh);
 558 out:
 559         gfs2_holder_uninit(&gh);
 560         if (error && error != AOP_TRUNCATED_PAGE)
 561                 lock_page(page);
 562         return error;
 563 }
 564
 565 /**
 566  * gfs2_internal_read - read an internal file
 567  * @ip: The gfs2 inode
 568  * @buf: The buffer to fill
 569  * @pos: The file position
 570  * @size: The amount to read
 571  *
 572  */
 573
 574 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 575                        unsigned size)
 576 {
 577         struct address_space *mapping = ip->i_inode.i_mapping;
 578         unsigned long index = *pos / PAGE_SIZE;
 579         unsigned offset = *pos & (PAGE_SIZE - 1);
 580         unsigned copied = 0;
 581         unsigned amt;
 582         struct page *page;
 583         void *p;
 584
 585         do {
 586                 amt = size - copied;
 587                 if (offset + size > PAGE_SIZE)
 588                         amt = PAGE_SIZE - offset;
 589                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 590                 if (IS_ERR(page))
 591                         return PTR_ERR(page);
 592                 p = kmap_atomic(page);
 593                 memcpy(buf + copied, p + offset, amt);
 594                 kunmap_atomic(p);
 595                 put_page(page);
 596                 copied += amt;
 597                 index++;
 598                 offset = 0;
 599         } while(copied < size);
 600         (*pos) += size;
 601         return size;
 602 }
 603
 604 /**
 605  * gfs2_readpages - Read a bunch of pages at once
 606  * @file: The file to read from
 607  * @mapping: Address space info
 608  * @pages: List of pages to read
 609  * @nr_pages: Number of pages to read
 610  *
 611  * Some notes:
 612  * 1. This is only for readahead, so we can simply ignore any things
 613  *    which are slightly inconvenient (such as locking conflicts between
 614  *    the page lock and the glock) and return having done no I/O. Its
 615  *    obviously not something we'd want to do on too regular a basis.
 616  *    Any I/O we ignore at this time will be done via readpage later.
 617  * 2. We don't handle stuffed files here we let readpage do the honours.
 618  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 619  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 620  */
 621
 622 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 623                           struct list_head *pages, unsigned nr_pages)
 624 {
 625         struct inode *inode = mapping->host;
 626         struct gfs2_inode *ip = GFS2_I(inode);
 627         struct gfs2_sbd *sdp = GFS2_SB(inode);
 628         struct gfs2_holder gh;
 629         int ret;
 630
 631         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 632         ret = gfs2_glock_nq(&gh);
 633         if (unlikely(ret))
 634                 goto out_uninit;
 635         if (!gfs2_is_stuffed(ip))
 636                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 637         gfs2_glock_dq(&gh);
 638 out_uninit:
 639         gfs2_holder_uninit(&gh);
 640         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 641                 ret = -EIO;
 642         return ret;
 643 }
 644
 645 /**
 646  * gfs2_write_begin - Begin to write to a file
 647  * @file: The file to write to
 648  * @mapping: The mapping in which to write
 649  * @pos: The file offset at which to start writing
 650  * @len: Length of the write
 651  * @flags: Various flags
 652  * @pagep: Pointer to return the page
 653  * @fsdata: Pointer to return fs data (unused by GFS2)
 654  *
 655  * Returns: errno
 656  */
 657
 658 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 659                             loff_t pos, unsigned len, unsigned flags,
 660                             struct page **pagep, void **fsdata)
 661 {
 662         struct gfs2_inode *ip = GFS2_I(mapping->host);
 663         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 664         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 665         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 666         unsigned requested = 0;
 667         int alloc_required;
 668         int error = 0;
 669         pgoff_t index = pos >> PAGE_SHIFT;
 670         unsigned from = pos & (PAGE_SIZE - 1);
 671         struct page *page;
 672
 673         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 674         error = gfs2_glock_nq(&ip->i_gh);
 675         if (unlikely(error))
 676                 goto out_uninit;
 677         if (&ip->i_inode == sdp->sd_rindex) {
 678                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 679                                            GL_NOCACHE, &m_ip->i_gh);
 680                 if (unlikely(error)) {
 681                         gfs2_glock_dq(&ip->i_gh);
 682                         goto out_uninit;
 683                 }
 684         }
 685
 686         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 687
 688         if (alloc_required || gfs2_is_jdata(ip))
 689                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 690
 691         if (alloc_required) {
 692                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 693                 requested = data_blocks + ind_blocks;
 694                 ap.target = requested;
 695                 error = gfs2_quota_lock_check(ip, &ap);
 696                 if (error)
 697                         goto out_unlock;
 698
 699                 error = gfs2_inplace_reserve(ip, &ap);
 700                 if (error)
 701                         goto out_qunlock;
 702         }
 703
 704         rblocks = RES_DINODE + ind_blocks;
 705         if (gfs2_is_jdata(ip))
 706                 rblocks += data_blocks ? data_blocks : 1;
 707         if (ind_blocks || data_blocks)
 708                 rblocks += RES_STATFS + RES_QUOTA;
 709         if (&ip->i_inode == sdp->sd_rindex)
 710                 rblocks += 2 * RES_STATFS;
 711         if (alloc_required)
 712                 rblocks += gfs2_rg_blocks(ip, requested);
 713
 714         error = gfs2_trans_begin(sdp, rblocks,
 715                                  PAGE_SIZE/sdp->sd_sb.sb_bsize);
 716         if (error)
 717                 goto out_trans_fail;
 718
 719         error = -ENOMEM;
 720         flags |= AOP_FLAG_NOFS;
 721         page = grab_cache_page_write_begin(mapping, index, flags);
 722         *pagep = page;
 723         if (unlikely(!page))
 724                 goto out_endtrans;
 725
 726         if (gfs2_is_stuffed(ip)) {
 727                 error = 0;
 728                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 729                         error = gfs2_unstuff_dinode(ip, page);
 730                         if (error == 0)
 731                                 goto prepare_write;
 732                 } else if (!PageUptodate(page)) {
 733                         error = stuffed_readpage(ip, page);
 734                 }
 735                 goto out;
 736         }
 737
 738 prepare_write:
 739         error = __block_write_begin(page, from, len, gfs2_block_map);
 740 out:
 741         if (error == 0)
 742                 return 0;
 743
 744         unlock_page(page);
 745         put_page(page);
 746
 747         gfs2_trans_end(sdp);
 748         if (pos + len > ip->i_inode.i_size)
 749                 gfs2_trim_blocks(&ip->i_inode);
 750         goto out_trans_fail;
 751
 752 out_endtrans:
 753         gfs2_trans_end(sdp);
 754 out_trans_fail:
 755         if (alloc_required) {
 756                 gfs2_inplace_release(ip);
 757 out_qunlock:
 758                 gfs2_quota_unlock(ip);
 759         }
 760 out_unlock:
 761         if (&ip->i_inode == sdp->sd_rindex) {
 762                 gfs2_glock_dq(&m_ip->i_gh);
 763                 gfs2_holder_uninit(&m_ip->i_gh);
 764         }
 765         gfs2_glock_dq(&ip->i_gh);
 766 out_uninit:
 767         gfs2_holder_uninit(&ip->i_gh);
 768         return error;
 769 }
 770
 771 /**
 772  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 773  * @inode: the rindex inode
 774  */
 775 static void adjust_fs_space(struct inode *inode)
 776 {
 777         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 778         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 779         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 780         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 781         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 782         struct buffer_head *m_bh, *l_bh;
 783         u64 fs_total, new_free;
 784
 785         /* Total up the file system space, according to the latest rindex. */
 786         fs_total = gfs2_ri_total(sdp);
 787         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 788                 return;
 789
 790         spin_lock(&sdp->sd_statfs_spin);
 791         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 792                               sizeof(struct gfs2_dinode));
 793         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 794                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 795         else
 796                 new_free = 0;
 797         spin_unlock(&sdp->sd_statfs_spin);
 798         fs_warn(sdp, "File system extended by %llu blocks.\n",
 799                 (unsigned long long)new_free);
 800         gfs2_statfs_change(sdp, new_free, new_free, 0);
 801
 802         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 803                 goto out;
 804         update_statfs(sdp, m_bh, l_bh);
 805         brelse(l_bh);
 806 out:
 807         brelse(m_bh);
 808 }
 809
 810 /**
 811  * gfs2_stuffed_write_end - Write end for stuffed files
 812  * @inode: The inode
 813  * @dibh: The buffer_head containing the on-disk inode
 814  * @pos: The file position
 815  * @len: The length of the write
 816  * @copied: How much was actually copied by the VFS
 817  * @page: The page
 818  *
 819  * This copies the data from the page into the inode block after
 820  * the inode data structure itself.
 821  *
 822  * Returns: errno
 823  */
 824 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 825                                   loff_t pos, unsigned len, unsigned copied,
 826                                   struct page *page)
 827 {
 828         struct gfs2_inode *ip = GFS2_I(inode);
 829         struct gfs2_sbd *sdp = GFS2_SB(inode);
 830         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 831         u64 to = pos + copied;
 832         void *kaddr;
 833         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 834
 835         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 836         kaddr = kmap_atomic(page);
 837         memcpy(buf + pos, kaddr + pos, copied);
 838         flush_dcache_page(page);
 839         kunmap_atomic(kaddr);
 840
 841         WARN_ON(!PageUptodate(page));
 842         unlock_page(page);
 843         put_page(page);
 844
 845         if (copied) {
 846                 if (inode->i_size < to)
 847                         i_size_write(inode, to);
 848                 mark_inode_dirty(inode);
 849         }
 850
 851         if (inode == sdp->sd_rindex) {
 852                 adjust_fs_space(inode);
 853                 sdp->sd_rindex_uptodate = 0;
 854         }
 855
 856         brelse(dibh);
 857         gfs2_trans_end(sdp);
 858         if (inode == sdp->sd_rindex) {
 859                 gfs2_glock_dq(&m_ip->i_gh);
 860                 gfs2_holder_uninit(&m_ip->i_gh);
 861         }
 862         gfs2_glock_dq(&ip->i_gh);
 863         gfs2_holder_uninit(&ip->i_gh);
 864         return copied;
 865 }
 866
 867 /**
 868  * gfs2_write_end
 869  * @file: The file to write to
 870  * @mapping: The address space to write to
 871  * @pos: The file position
 872  * @len: The length of the data
 873  * @copied: How much was actually copied by the VFS
 874  * @page: The page that has been written
 875  * @fsdata: The fsdata (unused in GFS2)
 876  *
 877  * The main write_end function for GFS2. We have a separate one for
 878  * stuffed files as they are slightly different, otherwise we just
 879  * put our locking around the VFS provided functions.
 880  *
 881  * Returns: errno
 882  */
 883
 884 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 885                           loff_t pos, unsigned len, unsigned copied,
 886                           struct page *page, void *fsdata)
 887 {
 888         struct inode *inode = page->mapping->host;
 889         struct gfs2_inode *ip = GFS2_I(inode);
 890         struct gfs2_sbd *sdp = GFS2_SB(inode);
 891         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 892         struct buffer_head *dibh;
 893         unsigned int from = pos & (PAGE_SIZE - 1);
 894         unsigned int to = from + len;
 895         int ret;
 896         struct gfs2_trans *tr = current->journal_info;
 897         BUG_ON(!tr);
 898
 899         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 900
 901         ret = gfs2_meta_inode_buffer(ip, &dibh);
 902         if (unlikely(ret)) {
 903                 unlock_page(page);
 904                 put_page(page);
 905                 goto failed;
 906         }
 907
 908         if (gfs2_is_stuffed(ip))
 909                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 910
 911         if (!gfs2_is_writeback(ip))
 912                 gfs2_page_add_databufs(ip, page, from, to);
 913
 914         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 915         if (tr->tr_num_buf_new)
 916                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 917         else
 918                 gfs2_trans_add_meta(ip->i_gl, dibh);
 919
 920
 921         if (inode == sdp->sd_rindex) {
 922                 adjust_fs_space(inode);
 923                 sdp->sd_rindex_uptodate = 0;
 924         }
 925
 926         brelse(dibh);
 927 failed:
 928         gfs2_trans_end(sdp);
 929         gfs2_inplace_release(ip);
 930         if (ip->i_qadata && ip->i_qadata->qa_qd_num)
 931                 gfs2_quota_unlock(ip);
 932         if (inode == sdp->sd_rindex) {
 933                 gfs2_glock_dq(&m_ip->i_gh);
 934                 gfs2_holder_uninit(&m_ip->i_gh);
 935         }
 936         gfs2_glock_dq(&ip->i_gh);
 937         gfs2_holder_uninit(&ip->i_gh);
 938         return ret;
 939 }
 940
 941 /**
 942  * gfs2_set_page_dirty - Page dirtying function
 943  * @page: The page to dirty
 944  *
 945  * Returns: 1 if it dirtyed the page, or 0 otherwise
 946  */
 947
 948 static int gfs2_set_page_dirty(struct page *page)
 949 {
 950         SetPageChecked(page);
 951         return __set_page_dirty_buffers(page);
 952 }
 953
 954 /**
 955  * gfs2_bmap - Block map function
 956  * @mapping: Address space info
 957  * @lblock: The block to map
 958  *
 959  * Returns: The disk address for the block or 0 on hole or error
 960  */
 961
 962 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 963 {
 964         struct gfs2_inode *ip = GFS2_I(mapping->host);
 965         struct gfs2_holder i_gh;
 966         sector_t dblock = 0;
 967         int error;
 968
 969         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 970         if (error)
 971                 return 0;
 972
 973         if (!gfs2_is_stuffed(ip))
 974                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 975
 976         gfs2_glock_dq_uninit(&i_gh);
 977
 978         return dblock;
 979 }
 980
 981 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 982 {
 983         struct gfs2_bufdata *bd;
 984
 985         lock_buffer(bh);
 986         gfs2_log_lock(sdp);
 987         clear_buffer_dirty(bh);
 988         bd = bh->b_private;
 989         if (bd) {
 990                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 991                         list_del_init(&bd->bd_list);
 992                 else
 993                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
 994         }
 995         bh->b_bdev = NULL;
 996         clear_buffer_mapped(bh);
 997         clear_buffer_req(bh);
 998         clear_buffer_new(bh);
 999         gfs2_log_unlock(sdp);
1000         unlock_buffer(bh);
1001 }
1002
1003 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
1004                                 unsigned int length)
1005 {
1006         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
1007         unsigned int stop = offset + length;
1008         int partial_page = (offset || length < PAGE_SIZE);
1009         struct buffer_head *bh, *head;
1010         unsigned long pos = 0;
1011
1012         BUG_ON(!PageLocked(page));
1013         if (!partial_page)
1014                 ClearPageChecked(page);
1015         if (!page_has_buffers(page))
1016                 goto out;
1017
1018         bh = head = page_buffers(page);
1019         do {
1020                 if (pos + bh->b_size > stop)
1021                         return;
1022
1023                 if (offset <= pos)
1024                         gfs2_discard(sdp, bh);
1025                 pos += bh->b_size;
1026                 bh = bh->b_this_page;
1027         } while (bh != head);
1028 out:
1029         if (!partial_page)
1030                 try_to_release_page(page, 0);
1031 }
1032
1033 /**
1034  * gfs2_ok_for_dio - check that dio is valid on this file
1035  * @ip: The inode
1036  * @offset: The offset at which we are reading or writing
1037  *
1038  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1039  *          1 (to accept the i/o request)
1040  */
1041 static int gfs2_ok_for_dio(struct gfs2_inode *ip, loff_t offset)
1042 {
1043         /*
1044          * Should we return an error here? I can't see that O_DIRECT for
1045          * a stuffed file makes any sense. For now we'll silently fall
1046          * back to buffered I/O
1047          */
1048         if (gfs2_is_stuffed(ip))
1049                 return 0;
1050
1051         if (offset >= i_size_read(&ip->i_inode))
1052                 return 0;
1053         return 1;
1054 }
1055
1056
1057
1058 static ssize_t gfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
1059 {
1060         struct file *file = iocb->ki_filp;
1061         struct inode *inode = file->f_mapping->host;
1062         struct address_space *mapping = inode->i_mapping;
1063         struct gfs2_inode *ip = GFS2_I(inode);
1064         loff_t offset = iocb->ki_pos;
1065         struct gfs2_holder gh;
1066         int rv;
1067
1068         /*
1069          * Deferred lock, even if its a write, since we do no allocation
1070          * on this path. All we need change is atime, and this lock mode
1071          * ensures that other nodes have flushed their buffered read caches
1072          * (i.e. their page cache entries for this inode). We do not,
1073          * unfortunately have the option of only flushing a range like
1074          * the VFS does.
1075          */
1076         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1077         rv = gfs2_glock_nq(&gh);
1078         if (rv)
1079                 goto out_uninit;
1080         rv = gfs2_ok_for_dio(ip, offset);
1081         if (rv != 1)
1082                 goto out; /* dio not valid, fall back to buffered i/o */
1083
1084         /*
1085          * Now since we are holding a deferred (CW) lock at this point, you
1086          * might be wondering why this is ever needed. There is a case however
1087          * where we've granted a deferred local lock against a cached exclusive
1088          * glock. That is ok provided all granted local locks are deferred, but
1089          * it also means that it is possible to encounter pages which are
1090          * cached and possibly also mapped. So here we check for that and sort
1091          * them out ahead of the dio. The glock state machine will take care of
1092          * everything else.
1093          *
1094          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1095          * the first place, mapping->nr_pages will always be zero.
1096          */
1097         if (mapping->nrpages) {
1098                 loff_t lstart = offset & ~(PAGE_SIZE - 1);
1099                 loff_t len = iov_iter_count(iter);
1100                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1101
1102                 rv = 0;
1103                 if (len == 0)
1104                         goto out;
1105                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1106                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1107                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1108                 if (rv)
1109                         goto out;
1110                 if (iov_iter_rw(iter) == WRITE)
1111                         truncate_inode_pages_range(mapping, lstart, end);
1112         }
1113
1114         rv = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
1115                                   gfs2_get_block_direct, NULL, NULL, 0);
1116 out:
1117         gfs2_glock_dq(&gh);
1118 out_uninit:
1119         gfs2_holder_uninit(&gh);
1120         return rv;
1121 }
1122
1123 /**
1124  * gfs2_releasepage - free the metadata associated with a page
1125  * @page: the page that's being released
1126  * @gfp_mask: passed from Linux VFS, ignored by us
1127  *
1128  * Call try_to_free_buffers() if the buffers in this page can be
1129  * released.
1130  *
1131  * Returns: 0
1132  */
1133
1134 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1135 {
1136         struct address_space *mapping = page->mapping;
1137         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1138         struct buffer_head *bh, *head;
1139         struct gfs2_bufdata *bd;
1140
1141         if (!page_has_buffers(page))
1142                 return 0;
1143
1144         /*
1145          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
1146          * clean pages might not have had the dirty bit cleared.  Thus, it can
1147          * send actual dirty pages to ->releasepage() via shrink_active_list().
1148          *
1149          * As a workaround, we skip pages that contain dirty buffers below.
1150          * Once ->releasepage isn't called on dirty pages anymore, we can warn
1151          * on dirty buffers like we used to here again.
1152          */
1153
1154         gfs2_log_lock(sdp);
1155         spin_lock(&sdp->sd_ail_lock);
1156         head = bh = page_buffers(page);
1157         do {
1158                 if (atomic_read(&bh->b_count))
1159                         goto cannot_release;
1160                 bd = bh->b_private;
1161                 if (bd && bd->bd_tr)
1162                         goto cannot_release;
1163                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
1164                         goto cannot_release;
1165                 bh = bh->b_this_page;
1166         } while(bh != head);
1167         spin_unlock(&sdp->sd_ail_lock);
1168
1169         head = bh = page_buffers(page);
1170         do {
1171                 bd = bh->b_private;
1172                 if (bd) {
1173                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1174                         if (!list_empty(&bd->bd_list))
1175                                 list_del_init(&bd->bd_list);
1176                         bd->bd_bh = NULL;
1177                         bh->b_private = NULL;
1178                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1179                 }
1180
1181                 bh = bh->b_this_page;
1182         } while (bh != head);
1183         gfs2_log_unlock(sdp);
1184
1185         return try_to_free_buffers(page);
1186
1187 cannot_release:
1188         spin_unlock(&sdp->sd_ail_lock);
1189         gfs2_log_unlock(sdp);
1190         return 0;
1191 }
1192
1193 static const struct address_space_operations gfs2_writeback_aops = {
1194         .writepage = gfs2_writepage,
1195         .writepages = gfs2_writepages,
1196         .readpage = gfs2_readpage,
1197         .readpages = gfs2_readpages,
1198         .write_begin = gfs2_write_begin,
1199         .write_end = gfs2_write_end,
1200         .bmap = gfs2_bmap,
1201         .invalidatepage = gfs2_invalidatepage,
1202         .releasepage = gfs2_releasepage,
1203         .direct_IO = gfs2_direct_IO,
1204         .migratepage = buffer_migrate_page,
1205         .is_partially_uptodate = block_is_partially_uptodate,
1206         .error_remove_page = generic_error_remove_page,
1207 };
1208
1209 static const struct address_space_operations gfs2_ordered_aops = {
1210         .writepage = gfs2_writepage,
1211         .writepages = gfs2_writepages,
1212         .readpage = gfs2_readpage,
1213         .readpages = gfs2_readpages,
1214         .write_begin = gfs2_write_begin,
1215         .write_end = gfs2_write_end,
1216         .set_page_dirty = gfs2_set_page_dirty,
1217         .bmap = gfs2_bmap,
1218         .invalidatepage = gfs2_invalidatepage,
1219         .releasepage = gfs2_releasepage,
1220         .direct_IO = gfs2_direct_IO,
1221         .migratepage = buffer_migrate_page,
1222         .is_partially_uptodate = block_is_partially_uptodate,
1223         .error_remove_page = generic_error_remove_page,
1224 };
1225
1226 static const struct address_space_operations gfs2_jdata_aops = {
1227         .writepage = gfs2_jdata_writepage,
1228         .writepages = gfs2_jdata_writepages,
1229         .readpage = gfs2_readpage,
1230         .readpages = gfs2_readpages,
1231         .write_begin = gfs2_write_begin,
1232         .write_end = gfs2_write_end,
1233         .set_page_dirty = gfs2_set_page_dirty,
1234         .bmap = gfs2_bmap,
1235         .invalidatepage = gfs2_invalidatepage,
1236         .releasepage = gfs2_releasepage,
1237         .is_partially_uptodate = block_is_partially_uptodate,
1238         .error_remove_page = generic_error_remove_page,
1239 };
1240
1241 void gfs2_set_aops(struct inode *inode)
1242 {
1243         struct gfs2_inode *ip = GFS2_I(inode);
1244
1245         if (gfs2_is_writeback(ip))
1246                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1247         else if (gfs2_is_ordered(ip))
1248                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1249         else if (gfs2_is_jdata(ip))
1250                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1251         else
1252                 BUG();
1253 }
1254