Pileus Git - ~andy/linux/blob - 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/aio.h>
  24
  25 #include "gfs2.h"
  26 #include "incore.h"
  27 #include "bmap.h"
  28 #include "glock.h"
  29 #include "inode.h"
  30 #include "log.h"
  31 #include "meta_io.h"
  32 #include "quota.h"
  33 #include "trans.h"
  34 #include "rgrp.h"
  35 #include "super.h"
  36 #include "util.h"
  37 #include "glops.h"
  38
  39
  40 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  41                                    unsigned int from, unsigned int to)
  42 {
  43         struct buffer_head *head = page_buffers(page);
  44         unsigned int bsize = head->b_size;
  45         struct buffer_head *bh;
  46         unsigned int start, end;
  47
  48         for (bh = head, start = 0; bh != head || !start;
  49              bh = bh->b_this_page, start = end) {
  50                 end = start + bsize;
  51                 if (end <= from || start >= to)
  52                         continue;
  53                 if (gfs2_is_jdata(ip))
  54                         set_buffer_uptodate(bh);
  55                 gfs2_trans_add_data(ip->i_gl, bh);
  56         }
  57 }
  58
  59 /**
  60  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  61  * @inode: The inode
  62  * @lblock: The block number to look up
  63  * @bh_result: The buffer head to return the result in
  64  * @create: Non-zero if we may add block to the file
  65  *
  66  * Returns: errno
  67  */
  68
  69 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  70                                   struct buffer_head *bh_result, int create)
  71 {
  72         int error;
  73
  74         error = gfs2_block_map(inode, lblock, bh_result, 0);
  75         if (error)
  76                 return error;
  77         if (!buffer_mapped(bh_result))
  78                 return -EIO;
  79         return 0;
  80 }
  81
  82 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  83                                  struct buffer_head *bh_result, int create)
  84 {
  85         return gfs2_block_map(inode, lblock, bh_result, 0);
  86 }
  87
  88 /**
  89  * gfs2_writepage_common - Common bits of writepage
  90  * @page: The page to be written
  91  * @wbc: The writeback control
  92  *
  93  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  94  */
  95
  96 static int gfs2_writepage_common(struct page *page,
  97                                  struct writeback_control *wbc)
  98 {
  99         struct inode *inode = page->mapping->host;
 100         struct gfs2_inode *ip = GFS2_I(inode);
 101         struct gfs2_sbd *sdp = GFS2_SB(inode);
 102         loff_t i_size = i_size_read(inode);
 103         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 104         unsigned offset;
 105
 106         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 107                 goto out;
 108         if (current->journal_info)
 109                 goto redirty;
 110         /* Is the page fully outside i_size? (truncate in progress) */
 111         offset = i_size & (PAGE_CACHE_SIZE-1);
 112         if (page->index > end_index || (page->index == end_index && !offset)) {
 113                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
 114                 goto out;
 115         }
 116         return 1;
 117 redirty:
 118         redirty_page_for_writepage(wbc, page);
 119 out:
 120         unlock_page(page);
 121         return 0;
 122 }
 123
 124 /**
 125  * gfs2_writepage - Write page for writeback mappings
 126  * @page: The page
 127  * @wbc: The writeback control
 128  *
 129  */
 130
 131 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 132 {
 133         int ret;
 134
 135         ret = gfs2_writepage_common(page, wbc);
 136         if (ret <= 0)
 137                 return ret;
 138
 139         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 140 }
 141
 142 /**
 143  * __gfs2_jdata_writepage - The core of jdata writepage
 144  * @page: The page to write
 145  * @wbc: The writeback control
 146  *
 147  * This is shared between writepage and writepages and implements the
 148  * core of the writepage operation. If a transaction is required then
 149  * PageChecked will have been set and the transaction will have
 150  * already been started before this is called.
 151  */
 152
 153 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 154 {
 155         struct inode *inode = page->mapping->host;
 156         struct gfs2_inode *ip = GFS2_I(inode);
 157         struct gfs2_sbd *sdp = GFS2_SB(inode);
 158
 159         if (PageChecked(page)) {
 160                 ClearPageChecked(page);
 161                 if (!page_has_buffers(page)) {
 162                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 163                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 164                 }
 165                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 166         }
 167         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 168 }
 169
 170 /**
 171  * gfs2_jdata_writepage - Write complete page
 172  * @page: Page to write
 173  *
 174  * Returns: errno
 175  *
 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_sbd *sdp = GFS2_SB(inode);
 182         int ret;
 183         int done_trans = 0;
 184
 185         if (PageChecked(page)) {
 186                 if (wbc->sync_mode != WB_SYNC_ALL)
 187                         goto out_ignore;
 188                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 189                 if (ret)
 190                         goto out_ignore;
 191                 done_trans = 1;
 192         }
 193         ret = gfs2_writepage_common(page, wbc);
 194         if (ret > 0)
 195                 ret = __gfs2_jdata_writepage(page, wbc);
 196         if (done_trans)
 197                 gfs2_trans_end(sdp);
 198         return ret;
 199
 200 out_ignore:
 201         redirty_page_for_writepage(wbc, page);
 202         unlock_page(page);
 203         return 0;
 204 }
 205
 206 /**
 207  * gfs2_writepages - Write a bunch of dirty pages back to disk
 208  * @mapping: The mapping to write
 209  * @wbc: Write-back control
 210  *
 211  * Used for both ordered and writeback modes.
 212  */
 213 static int gfs2_writepages(struct address_space *mapping,
 214                            struct writeback_control *wbc)
 215 {
 216         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 217 }
 218
 219 /**
 220  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 221  * @mapping: The mapping
 222  * @wbc: The writeback control
 223  * @writepage: The writepage function to call for each page
 224  * @pvec: The vector of pages
 225  * @nr_pages: The number of pages to write
 226  *
 227  * Returns: non-zero if loop should terminate, zero otherwise
 228  */
 229
 230 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 231                                     struct writeback_control *wbc,
 232                                     struct pagevec *pvec,
 233                                     int nr_pages, pgoff_t end)
 234 {
 235         struct inode *inode = mapping->host;
 236         struct gfs2_sbd *sdp = GFS2_SB(inode);
 237         loff_t i_size = i_size_read(inode);
 238         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 239         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
 240         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
 241         int i;
 242         int ret;
 243
 244         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 245         if (ret < 0)
 246                 return ret;
 247
 248         for(i = 0; i < nr_pages; i++) {
 249                 struct page *page = pvec->pages[i];
 250
 251                 lock_page(page);
 252
 253                 if (unlikely(page->mapping != mapping)) {
 254                         unlock_page(page);
 255                         continue;
 256                 }
 257
 258                 if (!wbc->range_cyclic && page->index > end) {
 259                         ret = 1;
 260                         unlock_page(page);
 261                         continue;
 262                 }
 263
 264                 if (wbc->sync_mode != WB_SYNC_NONE)
 265                         wait_on_page_writeback(page);
 266
 267                 if (PageWriteback(page) ||
 268                     !clear_page_dirty_for_io(page)) {
 269                         unlock_page(page);
 270                         continue;
 271                 }
 272
 273                 /* Is the page fully outside i_size? (truncate in progress) */
 274                 if (page->index > end_index || (page->index == end_index && !offset)) {
 275                         page->mapping->a_ops->invalidatepage(page, 0,
 276                                                              PAGE_CACHE_SIZE);
 277                         unlock_page(page);
 278                         continue;
 279                 }
 280
 281                 ret = __gfs2_jdata_writepage(page, wbc);
 282
 283                 if (ret || (--(wbc->nr_to_write) <= 0))
 284                         ret = 1;
 285         }
 286         gfs2_trans_end(sdp);
 287         return ret;
 288 }
 289
 290 /**
 291  * gfs2_write_cache_jdata - Like write_cache_pages but different
 292  * @mapping: The mapping to write
 293  * @wbc: The writeback control
 294  * @writepage: The writepage function to call
 295  * @data: The data to pass to writepage
 296  *
 297  * The reason that we use our own function here is that we need to
 298  * start transactions before we grab page locks. This allows us
 299  * to get the ordering right.
 300  */
 301
 302 static int gfs2_write_cache_jdata(struct address_space *mapping,
 303                                   struct writeback_control *wbc)
 304 {
 305         int ret = 0;
 306         int done = 0;
 307         struct pagevec pvec;
 308         int nr_pages;
 309         pgoff_t index;
 310         pgoff_t end;
 311         int scanned = 0;
 312         int range_whole = 0;
 313
 314         pagevec_init(&pvec, 0);
 315         if (wbc->range_cyclic) {
 316                 index = mapping->writeback_index; /* Start from prev offset */
 317                 end = -1;
 318         } else {
 319                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
 320                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 321                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 322                         range_whole = 1;
 323                 scanned = 1;
 324         }
 325
 326 retry:
 327          while (!done && (index <= end) &&
 328                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 329                                                PAGECACHE_TAG_DIRTY,
 330                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 331                 scanned = 1;
 332                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
 333                 if (ret)
 334                         done = 1;
 335                 if (ret > 0)
 336                         ret = 0;
 337
 338                 pagevec_release(&pvec);
 339                 cond_resched();
 340         }
 341
 342         if (!scanned && !done) {
 343                 /*
 344                  * We hit the last page and there is more work to be done: wrap
 345                  * back to the start of the file
 346                  */
 347                 scanned = 1;
 348                 index = 0;
 349                 goto retry;
 350         }
 351
 352         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 353                 mapping->writeback_index = index;
 354         return ret;
 355 }
 356
 357
 358 /**
 359  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 360  * @mapping: The mapping to write
 361  * @wbc: The writeback control
 362  *
 363  */
 364
 365 static int gfs2_jdata_writepages(struct address_space *mapping,
 366                                  struct writeback_control *wbc)
 367 {
 368         struct gfs2_inode *ip = GFS2_I(mapping->host);
 369         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 370         int ret;
 371
 372         ret = gfs2_write_cache_jdata(mapping, wbc);
 373         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 374                 gfs2_log_flush(sdp, ip->i_gl);
 375                 ret = gfs2_write_cache_jdata(mapping, wbc);
 376         }
 377         return ret;
 378 }
 379
 380 /**
 381  * stuffed_readpage - Fill in a Linux page with stuffed file data
 382  * @ip: the inode
 383  * @page: the page
 384  *
 385  * Returns: errno
 386  */
 387
 388 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 389 {
 390         struct buffer_head *dibh;
 391         u64 dsize = i_size_read(&ip->i_inode);
 392         void *kaddr;
 393         int error;
 394
 395         /*
 396          * Due to the order of unstuffing files and ->fault(), we can be
 397          * asked for a zero page in the case of a stuffed file being extended,
 398          * so we need to supply one here. It doesn't happen often.
 399          */
 400         if (unlikely(page->index)) {
 401                 zero_user(page, 0, PAGE_CACHE_SIZE);
 402                 SetPageUptodate(page);
 403                 return 0;
 404         }
 405
 406         error = gfs2_meta_inode_buffer(ip, &dibh);
 407         if (error)
 408                 return error;
 409
 410         kaddr = kmap_atomic(page);
 411         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 412                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 413         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 414         memset(kaddr + dsize, 0, PAGE_CACHE_SIZE - dsize);
 415         kunmap_atomic(kaddr);
 416         flush_dcache_page(page);
 417         brelse(dibh);
 418         SetPageUptodate(page);
 419
 420         return 0;
 421 }
 422
 423
 424 /**
 425  * __gfs2_readpage - readpage
 426  * @file: The file to read a page for
 427  * @page: The page to read
 428  *
 429  * This is the core of gfs2's readpage. Its used by the internal file
 430  * reading code as in that case we already hold the glock. Also its
 431  * called by gfs2_readpage() once the required lock has been granted.
 432  *
 433  */
 434
 435 static int __gfs2_readpage(void *file, struct page *page)
 436 {
 437         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 438         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 439         int error;
 440
 441         if (gfs2_is_stuffed(ip)) {
 442                 error = stuffed_readpage(ip, page);
 443                 unlock_page(page);
 444         } else {
 445                 error = mpage_readpage(page, gfs2_block_map);
 446         }
 447
 448         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 449                 return -EIO;
 450
 451         return error;
 452 }
 453
 454 /**
 455  * gfs2_readpage - read a page of a file
 456  * @file: The file to read
 457  * @page: The page of the file
 458  *
 459  * This deals with the locking required. We have to unlock and
 460  * relock the page in order to get the locking in the right
 461  * order.
 462  */
 463
 464 static int gfs2_readpage(struct file *file, struct page *page)
 465 {
 466         struct address_space *mapping = page->mapping;
 467         struct gfs2_inode *ip = GFS2_I(mapping->host);
 468         struct gfs2_holder gh;
 469         int error;
 470
 471         unlock_page(page);
 472         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 473         error = gfs2_glock_nq(&gh);
 474         if (unlikely(error))
 475                 goto out;
 476         error = AOP_TRUNCATED_PAGE;
 477         lock_page(page);
 478         if (page->mapping == mapping && !PageUptodate(page))
 479                 error = __gfs2_readpage(file, page);
 480         else
 481                 unlock_page(page);
 482         gfs2_glock_dq(&gh);
 483 out:
 484         gfs2_holder_uninit(&gh);
 485         if (error && error != AOP_TRUNCATED_PAGE)
 486                 lock_page(page);
 487         return error;
 488 }
 489
 490 /**
 491  * gfs2_internal_read - read an internal file
 492  * @ip: The gfs2 inode
 493  * @buf: The buffer to fill
 494  * @pos: The file position
 495  * @size: The amount to read
 496  *
 497  */
 498
 499 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 500                        unsigned size)
 501 {
 502         struct address_space *mapping = ip->i_inode.i_mapping;
 503         unsigned long index = *pos / PAGE_CACHE_SIZE;
 504         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 505         unsigned copied = 0;
 506         unsigned amt;
 507         struct page *page;
 508         void *p;
 509
 510         do {
 511                 amt = size - copied;
 512                 if (offset + size > PAGE_CACHE_SIZE)
 513                         amt = PAGE_CACHE_SIZE - offset;
 514                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 515                 if (IS_ERR(page))
 516                         return PTR_ERR(page);
 517                 p = kmap_atomic(page);
 518                 memcpy(buf + copied, p + offset, amt);
 519                 kunmap_atomic(p);
 520                 mark_page_accessed(page);
 521                 page_cache_release(page);
 522                 copied += amt;
 523                 index++;
 524                 offset = 0;
 525         } while(copied < size);
 526         (*pos) += size;
 527         return size;
 528 }
 529
 530 /**
 531  * gfs2_readpages - Read a bunch of pages at once
 532  *
 533  * Some notes:
 534  * 1. This is only for readahead, so we can simply ignore any things
 535  *    which are slightly inconvenient (such as locking conflicts between
 536  *    the page lock and the glock) and return having done no I/O. Its
 537  *    obviously not something we'd want to do on too regular a basis.
 538  *    Any I/O we ignore at this time will be done via readpage later.
 539  * 2. We don't handle stuffed files here we let readpage do the honours.
 540  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 541  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 542  */
 543
 544 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 545                           struct list_head *pages, unsigned nr_pages)
 546 {
 547         struct inode *inode = mapping->host;
 548         struct gfs2_inode *ip = GFS2_I(inode);
 549         struct gfs2_sbd *sdp = GFS2_SB(inode);
 550         struct gfs2_holder gh;
 551         int ret;
 552
 553         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 554         ret = gfs2_glock_nq(&gh);
 555         if (unlikely(ret))
 556                 goto out_uninit;
 557         if (!gfs2_is_stuffed(ip))
 558                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 559         gfs2_glock_dq(&gh);
 560 out_uninit:
 561         gfs2_holder_uninit(&gh);
 562         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 563                 ret = -EIO;
 564         return ret;
 565 }
 566
 567 /**
 568  * gfs2_write_begin - Begin to write to a file
 569  * @file: The file to write to
 570  * @mapping: The mapping in which to write
 571  * @pos: The file offset at which to start writing
 572  * @len: Length of the write
 573  * @flags: Various flags
 574  * @pagep: Pointer to return the page
 575  * @fsdata: Pointer to return fs data (unused by GFS2)
 576  *
 577  * Returns: errno
 578  */
 579
 580 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 581                             loff_t pos, unsigned len, unsigned flags,
 582                             struct page **pagep, void **fsdata)
 583 {
 584         struct gfs2_inode *ip = GFS2_I(mapping->host);
 585         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 586         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 587         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 588         unsigned requested = 0;
 589         int alloc_required;
 590         int error = 0;
 591         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 592         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 593         struct page *page;
 594
 595         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 596         error = gfs2_glock_nq(&ip->i_gh);
 597         if (unlikely(error))
 598                 goto out_uninit;
 599         if (&ip->i_inode == sdp->sd_rindex) {
 600                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 601                                            GL_NOCACHE, &m_ip->i_gh);
 602                 if (unlikely(error)) {
 603                         gfs2_glock_dq(&ip->i_gh);
 604                         goto out_uninit;
 605                 }
 606         }
 607
 608         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 609
 610         if (alloc_required || gfs2_is_jdata(ip))
 611                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 612
 613         if (alloc_required) {
 614                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 615                 error = gfs2_quota_lock_check(ip);
 616                 if (error)
 617                         goto out_unlock;
 618
 619                 requested = data_blocks + ind_blocks;
 620                 ap.target = requested;
 621                 error = gfs2_inplace_reserve(ip, &ap);
 622                 if (error)
 623                         goto out_qunlock;
 624         }
 625
 626         rblocks = RES_DINODE + ind_blocks;
 627         if (gfs2_is_jdata(ip))
 628                 rblocks += data_blocks ? data_blocks : 1;
 629         if (ind_blocks || data_blocks)
 630                 rblocks += RES_STATFS + RES_QUOTA;
 631         if (&ip->i_inode == sdp->sd_rindex)
 632                 rblocks += 2 * RES_STATFS;
 633         if (alloc_required)
 634                 rblocks += gfs2_rg_blocks(ip, requested);
 635
 636         error = gfs2_trans_begin(sdp, rblocks,
 637                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 638         if (error)
 639                 goto out_trans_fail;
 640
 641         error = -ENOMEM;
 642         flags |= AOP_FLAG_NOFS;
 643         page = grab_cache_page_write_begin(mapping, index, flags);
 644         *pagep = page;
 645         if (unlikely(!page))
 646                 goto out_endtrans;
 647
 648         if (gfs2_is_stuffed(ip)) {
 649                 error = 0;
 650                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 651                         error = gfs2_unstuff_dinode(ip, page);
 652                         if (error == 0)
 653                                 goto prepare_write;
 654                 } else if (!PageUptodate(page)) {
 655                         error = stuffed_readpage(ip, page);
 656                 }
 657                 goto out;
 658         }
 659
 660 prepare_write:
 661         error = __block_write_begin(page, from, len, gfs2_block_map);
 662 out:
 663         if (error == 0)
 664                 return 0;
 665
 666         unlock_page(page);
 667         page_cache_release(page);
 668
 669         gfs2_trans_end(sdp);
 670         if (pos + len > ip->i_inode.i_size)
 671                 gfs2_trim_blocks(&ip->i_inode);
 672         goto out_trans_fail;
 673
 674 out_endtrans:
 675         gfs2_trans_end(sdp);
 676 out_trans_fail:
 677         if (alloc_required) {
 678                 gfs2_inplace_release(ip);
 679 out_qunlock:
 680                 gfs2_quota_unlock(ip);
 681         }
 682 out_unlock:
 683         if (&ip->i_inode == sdp->sd_rindex) {
 684                 gfs2_glock_dq(&m_ip->i_gh);
 685                 gfs2_holder_uninit(&m_ip->i_gh);
 686         }
 687         gfs2_glock_dq(&ip->i_gh);
 688 out_uninit:
 689         gfs2_holder_uninit(&ip->i_gh);
 690         return error;
 691 }
 692
 693 /**
 694  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 695  * @inode: the rindex inode
 696  */
 697 static void adjust_fs_space(struct inode *inode)
 698 {
 699         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 700         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 701         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 702         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 703         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 704         struct buffer_head *m_bh, *l_bh;
 705         u64 fs_total, new_free;
 706
 707         /* Total up the file system space, according to the latest rindex. */
 708         fs_total = gfs2_ri_total(sdp);
 709         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 710                 return;
 711
 712         spin_lock(&sdp->sd_statfs_spin);
 713         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 714                               sizeof(struct gfs2_dinode));
 715         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 716                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 717         else
 718                 new_free = 0;
 719         spin_unlock(&sdp->sd_statfs_spin);
 720         fs_warn(sdp, "File system extended by %llu blocks.\n",
 721                 (unsigned long long)new_free);
 722         gfs2_statfs_change(sdp, new_free, new_free, 0);
 723
 724         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 725                 goto out;
 726         update_statfs(sdp, m_bh, l_bh);
 727         brelse(l_bh);
 728 out:
 729         brelse(m_bh);
 730 }
 731
 732 /**
 733  * gfs2_stuffed_write_end - Write end for stuffed files
 734  * @inode: The inode
 735  * @dibh: The buffer_head containing the on-disk inode
 736  * @pos: The file position
 737  * @len: The length of the write
 738  * @copied: How much was actually copied by the VFS
 739  * @page: The page
 740  *
 741  * This copies the data from the page into the inode block after
 742  * the inode data structure itself.
 743  *
 744  * Returns: errno
 745  */
 746 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 747                                   loff_t pos, unsigned len, unsigned copied,
 748                                   struct page *page)
 749 {
 750         struct gfs2_inode *ip = GFS2_I(inode);
 751         struct gfs2_sbd *sdp = GFS2_SB(inode);
 752         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 753         u64 to = pos + copied;
 754         void *kaddr;
 755         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 756
 757         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 758         kaddr = kmap_atomic(page);
 759         memcpy(buf + pos, kaddr + pos, copied);
 760         memset(kaddr + pos + copied, 0, len - copied);
 761         flush_dcache_page(page);
 762         kunmap_atomic(kaddr);
 763
 764         if (!PageUptodate(page))
 765                 SetPageUptodate(page);
 766         unlock_page(page);
 767         page_cache_release(page);
 768
 769         if (copied) {
 770                 if (inode->i_size < to)
 771                         i_size_write(inode, to);
 772                 mark_inode_dirty(inode);
 773         }
 774
 775         if (inode == sdp->sd_rindex) {
 776                 adjust_fs_space(inode);
 777                 sdp->sd_rindex_uptodate = 0;
 778         }
 779
 780         brelse(dibh);
 781         gfs2_trans_end(sdp);
 782         if (inode == sdp->sd_rindex) {
 783                 gfs2_glock_dq(&m_ip->i_gh);
 784                 gfs2_holder_uninit(&m_ip->i_gh);
 785         }
 786         gfs2_glock_dq(&ip->i_gh);
 787         gfs2_holder_uninit(&ip->i_gh);
 788         return copied;
 789 }
 790
 791 /**
 792  * gfs2_write_end
 793  * @file: The file to write to
 794  * @mapping: The address space to write to
 795  * @pos: The file position
 796  * @len: The length of the data
 797  * @copied:
 798  * @page: The page that has been written
 799  * @fsdata: The fsdata (unused in GFS2)
 800  *
 801  * The main write_end function for GFS2. We have a separate one for
 802  * stuffed files as they are slightly different, otherwise we just
 803  * put our locking around the VFS provided functions.
 804  *
 805  * Returns: errno
 806  */
 807
 808 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 809                           loff_t pos, unsigned len, unsigned copied,
 810                           struct page *page, void *fsdata)
 811 {
 812         struct inode *inode = page->mapping->host;
 813         struct gfs2_inode *ip = GFS2_I(inode);
 814         struct gfs2_sbd *sdp = GFS2_SB(inode);
 815         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 816         struct buffer_head *dibh;
 817         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 818         unsigned int to = from + len;
 819         int ret;
 820         struct gfs2_trans *tr = current->journal_info;
 821         BUG_ON(!tr);
 822
 823         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 824
 825         ret = gfs2_meta_inode_buffer(ip, &dibh);
 826         if (unlikely(ret)) {
 827                 unlock_page(page);
 828                 page_cache_release(page);
 829                 goto failed;
 830         }
 831
 832         if (gfs2_is_stuffed(ip))
 833                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 834
 835         if (!gfs2_is_writeback(ip))
 836                 gfs2_page_add_databufs(ip, page, from, to);
 837
 838         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 839         if (tr->tr_num_buf_new)
 840                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 841         else
 842                 gfs2_trans_add_meta(ip->i_gl, dibh);
 843
 844
 845         if (inode == sdp->sd_rindex) {
 846                 adjust_fs_space(inode);
 847                 sdp->sd_rindex_uptodate = 0;
 848         }
 849
 850         brelse(dibh);
 851 failed:
 852         gfs2_trans_end(sdp);
 853         gfs2_inplace_release(ip);
 854         if (ip->i_res->rs_qa_qd_num)
 855                 gfs2_quota_unlock(ip);
 856         if (inode == sdp->sd_rindex) {
 857                 gfs2_glock_dq(&m_ip->i_gh);
 858                 gfs2_holder_uninit(&m_ip->i_gh);
 859         }
 860         gfs2_glock_dq(&ip->i_gh);
 861         gfs2_holder_uninit(&ip->i_gh);
 862         return ret;
 863 }
 864
 865 /**
 866  * gfs2_set_page_dirty - Page dirtying function
 867  * @page: The page to dirty
 868  *
 869  * Returns: 1 if it dirtyed the page, or 0 otherwise
 870  */
 871
 872 static int gfs2_set_page_dirty(struct page *page)
 873 {
 874         SetPageChecked(page);
 875         return __set_page_dirty_buffers(page);
 876 }
 877
 878 /**
 879  * gfs2_bmap - Block map function
 880  * @mapping: Address space info
 881  * @lblock: The block to map
 882  *
 883  * Returns: The disk address for the block or 0 on hole or error
 884  */
 885
 886 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 887 {
 888         struct gfs2_inode *ip = GFS2_I(mapping->host);
 889         struct gfs2_holder i_gh;
 890         sector_t dblock = 0;
 891         int error;
 892
 893         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 894         if (error)
 895                 return 0;
 896
 897         if (!gfs2_is_stuffed(ip))
 898                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 899
 900         gfs2_glock_dq_uninit(&i_gh);
 901
 902         return dblock;
 903 }
 904
 905 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 906 {
 907         struct gfs2_bufdata *bd;
 908
 909         lock_buffer(bh);
 910         gfs2_log_lock(sdp);
 911         clear_buffer_dirty(bh);
 912         bd = bh->b_private;
 913         if (bd) {
 914                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 915                         list_del_init(&bd->bd_list);
 916                 else
 917                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 918         }
 919         bh->b_bdev = NULL;
 920         clear_buffer_mapped(bh);
 921         clear_buffer_req(bh);
 922         clear_buffer_new(bh);
 923         gfs2_log_unlock(sdp);
 924         unlock_buffer(bh);
 925 }
 926
 927 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
 928                                 unsigned int length)
 929 {
 930         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 931         unsigned int stop = offset + length;
 932         int partial_page = (offset || length < PAGE_CACHE_SIZE);
 933         struct buffer_head *bh, *head;
 934         unsigned long pos = 0;
 935
 936         BUG_ON(!PageLocked(page));
 937         if (!partial_page)
 938                 ClearPageChecked(page);
 939         if (!page_has_buffers(page))
 940                 goto out;
 941
 942         bh = head = page_buffers(page);
 943         do {
 944                 if (pos + bh->b_size > stop)
 945                         return;
 946
 947                 if (offset <= pos)
 948                         gfs2_discard(sdp, bh);
 949                 pos += bh->b_size;
 950                 bh = bh->b_this_page;
 951         } while (bh != head);
 952 out:
 953         if (!partial_page)
 954                 try_to_release_page(page, 0);
 955 }
 956
 957 /**
 958  * gfs2_ok_for_dio - check that dio is valid on this file
 959  * @ip: The inode
 960  * @rw: READ or WRITE
 961  * @offset: The offset at which we are reading or writing
 962  *
 963  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 964  *          1 (to accept the i/o request)
 965  */
 966 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 967 {
 968         /*
 969          * Should we return an error here? I can't see that O_DIRECT for
 970          * a stuffed file makes any sense. For now we'll silently fall
 971          * back to buffered I/O
 972          */
 973         if (gfs2_is_stuffed(ip))
 974                 return 0;
 975
 976         if (offset >= i_size_read(&ip->i_inode))
 977                 return 0;
 978         return 1;
 979 }
 980
 981
 982
 983 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 984                               const struct iovec *iov, loff_t offset,
 985                               unsigned long nr_segs)
 986 {
 987         struct file *file = iocb->ki_filp;
 988         struct inode *inode = file->f_mapping->host;
 989         struct address_space *mapping = inode->i_mapping;
 990         struct gfs2_inode *ip = GFS2_I(inode);
 991         struct gfs2_holder gh;
 992         int rv;
 993
 994         /*
 995          * Deferred lock, even if its a write, since we do no allocation
 996          * on this path. All we need change is atime, and this lock mode
 997          * ensures that other nodes have flushed their buffered read caches
 998          * (i.e. their page cache entries for this inode). We do not,
 999          * unfortunately have the option of only flushing a range like
1000          * the VFS does.
1001          */
1002         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1003         rv = gfs2_glock_nq(&gh);
1004         if (rv)
1005                 return rv;
1006         rv = gfs2_ok_for_dio(ip, rw, offset);
1007         if (rv != 1)
1008                 goto out; /* dio not valid, fall back to buffered i/o */
1009
1010         /*
1011          * Now since we are holding a deferred (CW) lock at this point, you
1012          * might be wondering why this is ever needed. There is a case however
1013          * where we've granted a deferred local lock against a cached exclusive
1014          * glock. That is ok provided all granted local locks are deferred, but
1015          * it also means that it is possible to encounter pages which are
1016          * cached and possibly also mapped. So here we check for that and sort
1017          * them out ahead of the dio. The glock state machine will take care of
1018          * everything else.
1019          *
1020          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1021          * the first place, mapping->nr_pages will always be zero.
1022          */
1023         if (mapping->nrpages) {
1024                 loff_t lstart = offset & (PAGE_CACHE_SIZE - 1);
1025                 loff_t len = iov_length(iov, nr_segs);
1026                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1027
1028                 rv = 0;
1029                 if (len == 0)
1030                         goto out;
1031                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1032                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1033                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1034                 if (rv)
1035                         goto out;
1036                 if (rw == WRITE)
1037                         truncate_inode_pages_range(mapping, lstart, end);
1038         }
1039
1040         rv = __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1041                                   offset, nr_segs, gfs2_get_block_direct,
1042                                   NULL, NULL, 0);
1043 out:
1044         gfs2_glock_dq(&gh);
1045         gfs2_holder_uninit(&gh);
1046         return rv;
1047 }
1048
1049 /**
1050  * gfs2_releasepage - free the metadata associated with a page
1051  * @page: the page that's being released
1052  * @gfp_mask: passed from Linux VFS, ignored by us
1053  *
1054  * Call try_to_free_buffers() if the buffers in this page can be
1055  * released.
1056  *
1057  * Returns: 0
1058  */
1059
1060 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1061 {
1062         struct address_space *mapping = page->mapping;
1063         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1064         struct buffer_head *bh, *head;
1065         struct gfs2_bufdata *bd;
1066
1067         if (!page_has_buffers(page))
1068                 return 0;
1069
1070         gfs2_log_lock(sdp);
1071         spin_lock(&sdp->sd_ail_lock);
1072         head = bh = page_buffers(page);
1073         do {
1074                 if (atomic_read(&bh->b_count))
1075                         goto cannot_release;
1076                 bd = bh->b_private;
1077                 if (bd && bd->bd_tr)
1078                         goto cannot_release;
1079                 if (buffer_pinned(bh) || buffer_dirty(bh))
1080                         goto not_possible;
1081                 bh = bh->b_this_page;
1082         } while(bh != head);
1083         spin_unlock(&sdp->sd_ail_lock);
1084
1085         head = bh = page_buffers(page);
1086         do {
1087                 bd = bh->b_private;
1088                 if (bd) {
1089                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1090                         if (!list_empty(&bd->bd_list))
1091                                 list_del_init(&bd->bd_list);
1092                         bd->bd_bh = NULL;
1093                         bh->b_private = NULL;
1094                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1095                 }
1096
1097                 bh = bh->b_this_page;
1098         } while (bh != head);
1099         gfs2_log_unlock(sdp);
1100
1101         return try_to_free_buffers(page);
1102
1103 not_possible: /* Should never happen */
1104         WARN_ON(buffer_dirty(bh));
1105         WARN_ON(buffer_pinned(bh));
1106 cannot_release:
1107         spin_unlock(&sdp->sd_ail_lock);
1108         gfs2_log_unlock(sdp);
1109         return 0;
1110 }
1111
1112 static const struct address_space_operations gfs2_writeback_aops = {
1113         .writepage = gfs2_writepage,
1114         .writepages = gfs2_writepages,
1115         .readpage = gfs2_readpage,
1116         .readpages = gfs2_readpages,
1117         .write_begin = gfs2_write_begin,
1118         .write_end = gfs2_write_end,
1119         .bmap = gfs2_bmap,
1120         .invalidatepage = gfs2_invalidatepage,
1121         .releasepage = gfs2_releasepage,
1122         .direct_IO = gfs2_direct_IO,
1123         .migratepage = buffer_migrate_page,
1124         .is_partially_uptodate = block_is_partially_uptodate,
1125         .error_remove_page = generic_error_remove_page,
1126 };
1127
1128 static const struct address_space_operations gfs2_ordered_aops = {
1129         .writepage = gfs2_writepage,
1130         .writepages = gfs2_writepages,
1131         .readpage = gfs2_readpage,
1132         .readpages = gfs2_readpages,
1133         .write_begin = gfs2_write_begin,
1134         .write_end = gfs2_write_end,
1135         .set_page_dirty = gfs2_set_page_dirty,
1136         .bmap = gfs2_bmap,
1137         .invalidatepage = gfs2_invalidatepage,
1138         .releasepage = gfs2_releasepage,
1139         .direct_IO = gfs2_direct_IO,
1140         .migratepage = buffer_migrate_page,
1141         .is_partially_uptodate = block_is_partially_uptodate,
1142         .error_remove_page = generic_error_remove_page,
1143 };
1144
1145 static const struct address_space_operations gfs2_jdata_aops = {
1146         .writepage = gfs2_jdata_writepage,
1147         .writepages = gfs2_jdata_writepages,
1148         .readpage = gfs2_readpage,
1149         .readpages = gfs2_readpages,
1150         .write_begin = gfs2_write_begin,
1151         .write_end = gfs2_write_end,
1152         .set_page_dirty = gfs2_set_page_dirty,
1153         .bmap = gfs2_bmap,
1154         .invalidatepage = gfs2_invalidatepage,
1155         .releasepage = gfs2_releasepage,
1156         .is_partially_uptodate = block_is_partially_uptodate,
1157         .error_remove_page = generic_error_remove_page,
1158 };
1159
1160 void gfs2_set_aops(struct inode *inode)
1161 {
1162         struct gfs2_inode *ip = GFS2_I(inode);
1163
1164         if (gfs2_is_writeback(ip))
1165                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1166         else if (gfs2_is_ordered(ip))
1167                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1168         else if (gfs2_is_jdata(ip))
1169                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1170         else
1171                 BUG();
1172 }
1173