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
  24 #include "gfs2.h"
  25 #include "incore.h"
  26 #include "bmap.h"
  27 #include "glock.h"
  28 #include "inode.h"
  29 #include "log.h"
  30 #include "meta_io.h"
  31 #include "quota.h"
  32 #include "trans.h"
  33 #include "rgrp.h"
  34 #include "super.h"
  35 #include "util.h"
  36 #include "glops.h"
  37
  38
  39 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  40                                    unsigned int from, unsigned int to)
  41 {
  42         struct buffer_head *head = page_buffers(page);
  43         unsigned int bsize = head->b_size;
  44         struct buffer_head *bh;
  45         unsigned int start, end;
  46
  47         for (bh = head, start = 0; bh != head || !start;
  48              bh = bh->b_this_page, start = end) {
  49                 end = start + bsize;
  50                 if (end <= from || start >= to)
  51                         continue;
  52                 if (gfs2_is_jdata(ip))
  53                         set_buffer_uptodate(bh);
  54                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  55         }
  56 }
  57
  58 /**
  59  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  60  * @inode: The inode
  61  * @lblock: The block number to look up
  62  * @bh_result: The buffer head to return the result in
  63  * @create: Non-zero if we may add block to the file
  64  *
  65  * Returns: errno
  66  */
  67
  68 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  69                                   struct buffer_head *bh_result, int create)
  70 {
  71         int error;
  72
  73         error = gfs2_block_map(inode, lblock, bh_result, 0);
  74         if (error)
  75                 return error;
  76         if (!buffer_mapped(bh_result))
  77                 return -EIO;
  78         return 0;
  79 }
  80
  81 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  82                                  struct buffer_head *bh_result, int create)
  83 {
  84         return gfs2_block_map(inode, lblock, bh_result, 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_CACHE_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_CACHE_SIZE-1);
 111         if (page->index > end_index || (page->index == end_index && !offset)) {
 112                 page->mapping->a_ops->invalidatepage(page, 0);
 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_writeback_writepage - Write page for writeback mappings
 125  * @page: The page
 126  * @wbc: The writeback control
 127  *
 128  */
 129
 130 static int gfs2_writeback_writepage(struct page *page,
 131                                     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_ordered_writepage - Write page for ordered data files
 144  * @page: The page to write
 145  * @wbc: The writeback control
 146  *
 147  */
 148
 149 static int gfs2_ordered_writepage(struct page *page,
 150                                   struct writeback_control *wbc)
 151 {
 152         struct inode *inode = page->mapping->host;
 153         struct gfs2_inode *ip = GFS2_I(inode);
 154         int ret;
 155
 156         ret = gfs2_writepage_common(page, wbc);
 157         if (ret <= 0)
 158                 return ret;
 159
 160         if (!page_has_buffers(page)) {
 161                 create_empty_buffers(page, inode->i_sb->s_blocksize,
 162                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
 163         }
 164         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
 165         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 166 }
 167
 168 /**
 169  * __gfs2_jdata_writepage - The core of jdata writepage
 170  * @page: The page to write
 171  * @wbc: The writeback control
 172  *
 173  * This is shared between writepage and writepages and implements the
 174  * core of the writepage operation. If a transaction is required then
 175  * PageChecked will have been set and the transaction will have
 176  * already been started before this is called.
 177  */
 178
 179 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 180 {
 181         struct inode *inode = page->mapping->host;
 182         struct gfs2_inode *ip = GFS2_I(inode);
 183         struct gfs2_sbd *sdp = GFS2_SB(inode);
 184
 185         if (PageChecked(page)) {
 186                 ClearPageChecked(page);
 187                 if (!page_has_buffers(page)) {
 188                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 189                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 190                 }
 191                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 192         }
 193         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 194 }
 195
 196 /**
 197  * gfs2_jdata_writepage - Write complete page
 198  * @page: Page to write
 199  *
 200  * Returns: errno
 201  *
 202  */
 203
 204 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 205 {
 206         struct inode *inode = page->mapping->host;
 207         struct gfs2_sbd *sdp = GFS2_SB(inode);
 208         int ret;
 209         int done_trans = 0;
 210
 211         if (PageChecked(page)) {
 212                 if (wbc->sync_mode != WB_SYNC_ALL)
 213                         goto out_ignore;
 214                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 215                 if (ret)
 216                         goto out_ignore;
 217                 done_trans = 1;
 218         }
 219         ret = gfs2_writepage_common(page, wbc);
 220         if (ret > 0)
 221                 ret = __gfs2_jdata_writepage(page, wbc);
 222         if (done_trans)
 223                 gfs2_trans_end(sdp);
 224         return ret;
 225
 226 out_ignore:
 227         redirty_page_for_writepage(wbc, page);
 228         unlock_page(page);
 229         return 0;
 230 }
 231
 232 /**
 233  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
 234  * @mapping: The mapping to write
 235  * @wbc: Write-back control
 236  *
 237  * For the data=writeback case we can already ignore buffer heads
 238  * and write whole extents at once. This is a big reduction in the
 239  * number of I/O requests we send and the bmap calls we make in this case.
 240  */
 241 static int gfs2_writeback_writepages(struct address_space *mapping,
 242                                      struct writeback_control *wbc)
 243 {
 244         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 245 }
 246
 247 /**
 248  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 249  * @mapping: The mapping
 250  * @wbc: The writeback control
 251  * @writepage: The writepage function to call for each page
 252  * @pvec: The vector of pages
 253  * @nr_pages: The number of pages to write
 254  *
 255  * Returns: non-zero if loop should terminate, zero otherwise
 256  */
 257
 258 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 259                                     struct writeback_control *wbc,
 260                                     struct pagevec *pvec,
 261                                     int nr_pages, pgoff_t end)
 262 {
 263         struct inode *inode = mapping->host;
 264         struct gfs2_sbd *sdp = GFS2_SB(inode);
 265         loff_t i_size = i_size_read(inode);
 266         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 267         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
 268         unsigned nrblocks = nr_pages * (PAGE_CACHE_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                 lock_page(page);
 280
 281                 if (unlikely(page->mapping != mapping)) {
 282                         unlock_page(page);
 283                         continue;
 284                 }
 285
 286                 if (!wbc->range_cyclic && page->index > end) {
 287                         ret = 1;
 288                         unlock_page(page);
 289                         continue;
 290                 }
 291
 292                 if (wbc->sync_mode != WB_SYNC_NONE)
 293                         wait_on_page_writeback(page);
 294
 295                 if (PageWriteback(page) ||
 296                     !clear_page_dirty_for_io(page)) {
 297                         unlock_page(page);
 298                         continue;
 299                 }
 300
 301                 /* Is the page fully outside i_size? (truncate in progress) */
 302                 if (page->index > end_index || (page->index == end_index && !offset)) {
 303                         page->mapping->a_ops->invalidatepage(page, 0);
 304                         unlock_page(page);
 305                         continue;
 306                 }
 307
 308                 ret = __gfs2_jdata_writepage(page, wbc);
 309
 310                 if (ret || (--(wbc->nr_to_write) <= 0))
 311                         ret = 1;
 312         }
 313         gfs2_trans_end(sdp);
 314         return ret;
 315 }
 316
 317 /**
 318  * gfs2_write_cache_jdata - Like write_cache_pages but different
 319  * @mapping: The mapping to write
 320  * @wbc: The writeback control
 321  * @writepage: The writepage function to call
 322  * @data: The data to pass to writepage
 323  *
 324  * The reason that we use our own function here is that we need to
 325  * start transactions before we grab page locks. This allows us
 326  * to get the ordering right.
 327  */
 328
 329 static int gfs2_write_cache_jdata(struct address_space *mapping,
 330                                   struct writeback_control *wbc)
 331 {
 332         int ret = 0;
 333         int done = 0;
 334         struct pagevec pvec;
 335         int nr_pages;
 336         pgoff_t index;
 337         pgoff_t end;
 338         int scanned = 0;
 339         int range_whole = 0;
 340
 341         pagevec_init(&pvec, 0);
 342         if (wbc->range_cyclic) {
 343                 index = mapping->writeback_index; /* Start from prev offset */
 344                 end = -1;
 345         } else {
 346                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
 347                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 348                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 349                         range_whole = 1;
 350                 scanned = 1;
 351         }
 352
 353 retry:
 354          while (!done && (index <= end) &&
 355                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 356                                                PAGECACHE_TAG_DIRTY,
 357                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 358                 scanned = 1;
 359                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
 360                 if (ret)
 361                         done = 1;
 362                 if (ret > 0)
 363                         ret = 0;
 364
 365                 pagevec_release(&pvec);
 366                 cond_resched();
 367         }
 368
 369         if (!scanned && !done) {
 370                 /*
 371                  * We hit the last page and there is more work to be done: wrap
 372                  * back to the start of the file
 373                  */
 374                 scanned = 1;
 375                 index = 0;
 376                 goto retry;
 377         }
 378
 379         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 380                 mapping->writeback_index = index;
 381         return ret;
 382 }
 383
 384
 385 /**
 386  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 387  * @mapping: The mapping to write
 388  * @wbc: The writeback control
 389  *
 390  */
 391
 392 static int gfs2_jdata_writepages(struct address_space *mapping,
 393                                  struct writeback_control *wbc)
 394 {
 395         struct gfs2_inode *ip = GFS2_I(mapping->host);
 396         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 397         int ret;
 398
 399         ret = gfs2_write_cache_jdata(mapping, wbc);
 400         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 401                 gfs2_log_flush(sdp, ip->i_gl);
 402                 ret = gfs2_write_cache_jdata(mapping, wbc);
 403         }
 404         return ret;
 405 }
 406
 407 /**
 408  * stuffed_readpage - Fill in a Linux page with stuffed file data
 409  * @ip: the inode
 410  * @page: the page
 411  *
 412  * Returns: errno
 413  */
 414
 415 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 416 {
 417         struct buffer_head *dibh;
 418         u64 dsize = i_size_read(&ip->i_inode);
 419         void *kaddr;
 420         int error;
 421
 422         /*
 423          * Due to the order of unstuffing files and ->fault(), we can be
 424          * asked for a zero page in the case of a stuffed file being extended,
 425          * so we need to supply one here. It doesn't happen often.
 426          */
 427         if (unlikely(page->index)) {
 428                 zero_user(page, 0, PAGE_CACHE_SIZE);
 429                 SetPageUptodate(page);
 430                 return 0;
 431         }
 432
 433         error = gfs2_meta_inode_buffer(ip, &dibh);
 434         if (error)
 435                 return error;
 436
 437         kaddr = kmap_atomic(page, KM_USER0);
 438         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 439                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 440         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 441         memset(kaddr + dsize, 0, PAGE_CACHE_SIZE - dsize);
 442         kunmap_atomic(kaddr, KM_USER0);
 443         flush_dcache_page(page);
 444         brelse(dibh);
 445         SetPageUptodate(page);
 446
 447         return 0;
 448 }
 449
 450
 451 /**
 452  * __gfs2_readpage - readpage
 453  * @file: The file to read a page for
 454  * @page: The page to read
 455  *
 456  * This is the core of gfs2's readpage. Its used by the internal file
 457  * reading code as in that case we already hold the glock. Also its
 458  * called by gfs2_readpage() once the required lock has been granted.
 459  *
 460  */
 461
 462 static int __gfs2_readpage(void *file, struct page *page)
 463 {
 464         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 465         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 466         int error;
 467
 468         if (gfs2_is_stuffed(ip)) {
 469                 error = stuffed_readpage(ip, page);
 470                 unlock_page(page);
 471         } else {
 472                 error = mpage_readpage(page, gfs2_block_map);
 473         }
 474
 475         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 476                 return -EIO;
 477
 478         return error;
 479 }
 480
 481 /**
 482  * gfs2_readpage - read a page of a file
 483  * @file: The file to read
 484  * @page: The page of the file
 485  *
 486  * This deals with the locking required. We have to unlock and
 487  * relock the page in order to get the locking in the right
 488  * order.
 489  */
 490
 491 static int gfs2_readpage(struct file *file, struct page *page)
 492 {
 493         struct address_space *mapping = page->mapping;
 494         struct gfs2_inode *ip = GFS2_I(mapping->host);
 495         struct gfs2_holder gh;
 496         int error;
 497
 498         unlock_page(page);
 499         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 500         error = gfs2_glock_nq(&gh);
 501         if (unlikely(error))
 502                 goto out;
 503         error = AOP_TRUNCATED_PAGE;
 504         lock_page(page);
 505         if (page->mapping == mapping && !PageUptodate(page))
 506                 error = __gfs2_readpage(file, page);
 507         else
 508                 unlock_page(page);
 509         gfs2_glock_dq(&gh);
 510 out:
 511         gfs2_holder_uninit(&gh);
 512         if (error && error != AOP_TRUNCATED_PAGE)
 513                 lock_page(page);
 514         return error;
 515 }
 516
 517 /**
 518  * gfs2_internal_read - read an internal file
 519  * @ip: The gfs2 inode
 520  * @ra_state: The readahead state (or NULL for no readahead)
 521  * @buf: The buffer to fill
 522  * @pos: The file position
 523  * @size: The amount to read
 524  *
 525  */
 526
 527 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
 528                        char *buf, loff_t *pos, unsigned size)
 529 {
 530         struct address_space *mapping = ip->i_inode.i_mapping;
 531         unsigned long index = *pos / PAGE_CACHE_SIZE;
 532         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 533         unsigned copied = 0;
 534         unsigned amt;
 535         struct page *page;
 536         void *p;
 537
 538         do {
 539                 amt = size - copied;
 540                 if (offset + size > PAGE_CACHE_SIZE)
 541                         amt = PAGE_CACHE_SIZE - offset;
 542                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 543                 if (IS_ERR(page))
 544                         return PTR_ERR(page);
 545                 p = kmap_atomic(page, KM_USER0);
 546                 memcpy(buf + copied, p + offset, amt);
 547                 kunmap_atomic(p, KM_USER0);
 548                 mark_page_accessed(page);
 549                 page_cache_release(page);
 550                 copied += amt;
 551                 index++;
 552                 offset = 0;
 553         } while(copied < size);
 554         (*pos) += size;
 555         return size;
 556 }
 557
 558 /**
 559  * gfs2_readpages - Read a bunch of pages at once
 560  *
 561  * Some notes:
 562  * 1. This is only for readahead, so we can simply ignore any things
 563  *    which are slightly inconvenient (such as locking conflicts between
 564  *    the page lock and the glock) and return having done no I/O. Its
 565  *    obviously not something we'd want to do on too regular a basis.
 566  *    Any I/O we ignore at this time will be done via readpage later.
 567  * 2. We don't handle stuffed files here we let readpage do the honours.
 568  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 569  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 570  */
 571
 572 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 573                           struct list_head *pages, unsigned nr_pages)
 574 {
 575         struct inode *inode = mapping->host;
 576         struct gfs2_inode *ip = GFS2_I(inode);
 577         struct gfs2_sbd *sdp = GFS2_SB(inode);
 578         struct gfs2_holder gh;
 579         int ret;
 580
 581         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 582         ret = gfs2_glock_nq(&gh);
 583         if (unlikely(ret))
 584                 goto out_uninit;
 585         if (!gfs2_is_stuffed(ip))
 586                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 587         gfs2_glock_dq(&gh);
 588 out_uninit:
 589         gfs2_holder_uninit(&gh);
 590         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 591                 ret = -EIO;
 592         return ret;
 593 }
 594
 595 /**
 596  * gfs2_write_begin - Begin to write to a file
 597  * @file: The file to write to
 598  * @mapping: The mapping in which to write
 599  * @pos: The file offset at which to start writing
 600  * @len: Length of the write
 601  * @flags: Various flags
 602  * @pagep: Pointer to return the page
 603  * @fsdata: Pointer to return fs data (unused by GFS2)
 604  *
 605  * Returns: errno
 606  */
 607
 608 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 609                             loff_t pos, unsigned len, unsigned flags,
 610                             struct page **pagep, void **fsdata)
 611 {
 612         struct gfs2_inode *ip = GFS2_I(mapping->host);
 613         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 614         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 615         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 616         int alloc_required;
 617         int error = 0;
 618         struct gfs2_alloc *al;
 619         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 620         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 621         unsigned to = from + len;
 622         struct page *page;
 623
 624         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 625         error = gfs2_glock_nq(&ip->i_gh);
 626         if (unlikely(error))
 627                 goto out_uninit;
 628         if (&ip->i_inode == sdp->sd_rindex) {
 629                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 630                                            GL_NOCACHE, &m_ip->i_gh);
 631                 if (unlikely(error)) {
 632                         gfs2_glock_dq(&ip->i_gh);
 633                         goto out_uninit;
 634                 }
 635         }
 636
 637         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 638
 639         if (alloc_required || gfs2_is_jdata(ip))
 640                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 641
 642         if (alloc_required) {
 643                 al = gfs2_alloc_get(ip);
 644                 if (!al) {
 645                         error = -ENOMEM;
 646                         goto out_unlock;
 647                 }
 648
 649                 error = gfs2_quota_lock_check(ip);
 650                 if (error)
 651                         goto out_alloc_put;
 652
 653                 al->al_requested = data_blocks + ind_blocks;
 654                 error = gfs2_inplace_reserve(ip);
 655                 if (error)
 656                         goto out_qunlock;
 657         }
 658
 659         rblocks = RES_DINODE + ind_blocks;
 660         if (gfs2_is_jdata(ip))
 661                 rblocks += data_blocks ? data_blocks : 1;
 662         if (ind_blocks || data_blocks)
 663                 rblocks += RES_STATFS + RES_QUOTA;
 664         if (&ip->i_inode == sdp->sd_rindex)
 665                 rblocks += 2 * RES_STATFS;
 666
 667         error = gfs2_trans_begin(sdp, rblocks,
 668                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 669         if (error)
 670                 goto out_trans_fail;
 671
 672         error = -ENOMEM;
 673         flags |= AOP_FLAG_NOFS;
 674         page = grab_cache_page_write_begin(mapping, index, flags);
 675         *pagep = page;
 676         if (unlikely(!page))
 677                 goto out_endtrans;
 678
 679         if (gfs2_is_stuffed(ip)) {
 680                 error = 0;
 681                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 682                         error = gfs2_unstuff_dinode(ip, page);
 683                         if (error == 0)
 684                                 goto prepare_write;
 685                 } else if (!PageUptodate(page)) {
 686                         error = stuffed_readpage(ip, page);
 687                 }
 688                 goto out;
 689         }
 690
 691 prepare_write:
 692         error = block_prepare_write(page, from, to, gfs2_block_map);
 693 out:
 694         if (error == 0)
 695                 return 0;
 696
 697         page_cache_release(page);
 698
 699         /*
 700          * XXX(hch): the call below should probably be replaced with
 701          * a call to the gfs2-specific truncate blocks helper to actually
 702          * release disk blocks..
 703          */
 704         if (pos + len > ip->i_inode.i_size)
 705                 simple_setsize(&ip->i_inode, ip->i_inode.i_size);
 706 out_endtrans:
 707         gfs2_trans_end(sdp);
 708 out_trans_fail:
 709         if (alloc_required) {
 710                 gfs2_inplace_release(ip);
 711 out_qunlock:
 712                 gfs2_quota_unlock(ip);
 713 out_alloc_put:
 714                 gfs2_alloc_put(ip);
 715         }
 716 out_unlock:
 717         if (&ip->i_inode == sdp->sd_rindex) {
 718                 gfs2_glock_dq(&m_ip->i_gh);
 719                 gfs2_holder_uninit(&m_ip->i_gh);
 720         }
 721         gfs2_glock_dq(&ip->i_gh);
 722 out_uninit:
 723         gfs2_holder_uninit(&ip->i_gh);
 724         return error;
 725 }
 726
 727 /**
 728  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 729  * @inode: the rindex inode
 730  */
 731 static void adjust_fs_space(struct inode *inode)
 732 {
 733         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 734         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 735         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 736         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 737         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 738         struct buffer_head *m_bh, *l_bh;
 739         u64 fs_total, new_free;
 740
 741         /* Total up the file system space, according to the latest rindex. */
 742         fs_total = gfs2_ri_total(sdp);
 743         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 744                 return;
 745
 746         spin_lock(&sdp->sd_statfs_spin);
 747         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 748                               sizeof(struct gfs2_dinode));
 749         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 750                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 751         else
 752                 new_free = 0;
 753         spin_unlock(&sdp->sd_statfs_spin);
 754         fs_warn(sdp, "File system extended by %llu blocks.\n",
 755                 (unsigned long long)new_free);
 756         gfs2_statfs_change(sdp, new_free, new_free, 0);
 757
 758         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 759                 goto out;
 760         update_statfs(sdp, m_bh, l_bh);
 761         brelse(l_bh);
 762 out:
 763         brelse(m_bh);
 764 }
 765
 766 /**
 767  * gfs2_stuffed_write_end - Write end for stuffed files
 768  * @inode: The inode
 769  * @dibh: The buffer_head containing the on-disk inode
 770  * @pos: The file position
 771  * @len: The length of the write
 772  * @copied: How much was actually copied by the VFS
 773  * @page: The page
 774  *
 775  * This copies the data from the page into the inode block after
 776  * the inode data structure itself.
 777  *
 778  * Returns: errno
 779  */
 780 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 781                                   loff_t pos, unsigned len, unsigned copied,
 782                                   struct page *page)
 783 {
 784         struct gfs2_inode *ip = GFS2_I(inode);
 785         struct gfs2_sbd *sdp = GFS2_SB(inode);
 786         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 787         u64 to = pos + copied;
 788         void *kaddr;
 789         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 790         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
 791
 792         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 793         kaddr = kmap_atomic(page, KM_USER0);
 794         memcpy(buf + pos, kaddr + pos, copied);
 795         memset(kaddr + pos + copied, 0, len - copied);
 796         flush_dcache_page(page);
 797         kunmap_atomic(kaddr, KM_USER0);
 798
 799         if (!PageUptodate(page))
 800                 SetPageUptodate(page);
 801         unlock_page(page);
 802         page_cache_release(page);
 803
 804         if (copied) {
 805                 if (inode->i_size < to) {
 806                         i_size_write(inode, to);
 807                         ip->i_disksize = inode->i_size;
 808                 }
 809                 gfs2_dinode_out(ip, di);
 810                 mark_inode_dirty(inode);
 811         }
 812
 813         if (inode == sdp->sd_rindex) {
 814                 adjust_fs_space(inode);
 815                 ip->i_gh.gh_flags |= GL_NOCACHE;
 816         }
 817
 818         brelse(dibh);
 819         gfs2_trans_end(sdp);
 820         if (inode == sdp->sd_rindex) {
 821                 gfs2_glock_dq(&m_ip->i_gh);
 822                 gfs2_holder_uninit(&m_ip->i_gh);
 823         }
 824         gfs2_glock_dq(&ip->i_gh);
 825         gfs2_holder_uninit(&ip->i_gh);
 826         return copied;
 827 }
 828
 829 /**
 830  * gfs2_write_end
 831  * @file: The file to write to
 832  * @mapping: The address space to write to
 833  * @pos: The file position
 834  * @len: The length of the data
 835  * @copied:
 836  * @page: The page that has been written
 837  * @fsdata: The fsdata (unused in GFS2)
 838  *
 839  * The main write_end function for GFS2. We have a separate one for
 840  * stuffed files as they are slightly different, otherwise we just
 841  * put our locking around the VFS provided functions.
 842  *
 843  * Returns: errno
 844  */
 845
 846 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 847                           loff_t pos, unsigned len, unsigned copied,
 848                           struct page *page, void *fsdata)
 849 {
 850         struct inode *inode = page->mapping->host;
 851         struct gfs2_inode *ip = GFS2_I(inode);
 852         struct gfs2_sbd *sdp = GFS2_SB(inode);
 853         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 854         struct buffer_head *dibh;
 855         struct gfs2_alloc *al = ip->i_alloc;
 856         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 857         unsigned int to = from + len;
 858         int ret;
 859
 860         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 861
 862         ret = gfs2_meta_inode_buffer(ip, &dibh);
 863         if (unlikely(ret)) {
 864                 unlock_page(page);
 865                 page_cache_release(page);
 866                 goto failed;
 867         }
 868
 869         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 870
 871         if (gfs2_is_stuffed(ip))
 872                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 873
 874         if (!gfs2_is_writeback(ip))
 875                 gfs2_page_add_databufs(ip, page, from, to);
 876
 877         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 878         if (ret > 0) {
 879                 if (inode->i_size > ip->i_disksize)
 880                         ip->i_disksize = inode->i_size;
 881                 gfs2_dinode_out(ip, dibh->b_data);
 882                 mark_inode_dirty(inode);
 883         }
 884
 885         if (inode == sdp->sd_rindex) {
 886                 adjust_fs_space(inode);
 887                 ip->i_gh.gh_flags |= GL_NOCACHE;
 888         }
 889
 890         brelse(dibh);
 891         gfs2_trans_end(sdp);
 892 failed:
 893         if (al) {
 894                 gfs2_inplace_release(ip);
 895                 gfs2_quota_unlock(ip);
 896                 gfs2_alloc_put(ip);
 897         }
 898         if (inode == sdp->sd_rindex) {
 899                 gfs2_glock_dq(&m_ip->i_gh);
 900                 gfs2_holder_uninit(&m_ip->i_gh);
 901         }
 902         gfs2_glock_dq(&ip->i_gh);
 903         gfs2_holder_uninit(&ip->i_gh);
 904         return ret;
 905 }
 906
 907 /**
 908  * gfs2_set_page_dirty - Page dirtying function
 909  * @page: The page to dirty
 910  *
 911  * Returns: 1 if it dirtyed the page, or 0 otherwise
 912  */
 913
 914 static int gfs2_set_page_dirty(struct page *page)
 915 {
 916         SetPageChecked(page);
 917         return __set_page_dirty_buffers(page);
 918 }
 919
 920 /**
 921  * gfs2_bmap - Block map function
 922  * @mapping: Address space info
 923  * @lblock: The block to map
 924  *
 925  * Returns: The disk address for the block or 0 on hole or error
 926  */
 927
 928 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 929 {
 930         struct gfs2_inode *ip = GFS2_I(mapping->host);
 931         struct gfs2_holder i_gh;
 932         sector_t dblock = 0;
 933         int error;
 934
 935         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 936         if (error)
 937                 return 0;
 938
 939         if (!gfs2_is_stuffed(ip))
 940                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 941
 942         gfs2_glock_dq_uninit(&i_gh);
 943
 944         return dblock;
 945 }
 946
 947 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 948 {
 949         struct gfs2_bufdata *bd;
 950
 951         lock_buffer(bh);
 952         gfs2_log_lock(sdp);
 953         clear_buffer_dirty(bh);
 954         bd = bh->b_private;
 955         if (bd) {
 956                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 957                         list_del_init(&bd->bd_le.le_list);
 958                 else
 959                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 960         }
 961         bh->b_bdev = NULL;
 962         clear_buffer_mapped(bh);
 963         clear_buffer_req(bh);
 964         clear_buffer_new(bh);
 965         gfs2_log_unlock(sdp);
 966         unlock_buffer(bh);
 967 }
 968
 969 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 970 {
 971         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 972         struct buffer_head *bh, *head;
 973         unsigned long pos = 0;
 974
 975         BUG_ON(!PageLocked(page));
 976         if (offset == 0)
 977                 ClearPageChecked(page);
 978         if (!page_has_buffers(page))
 979                 goto out;
 980
 981         bh = head = page_buffers(page);
 982         do {
 983                 if (offset <= pos)
 984                         gfs2_discard(sdp, bh);
 985                 pos += bh->b_size;
 986                 bh = bh->b_this_page;
 987         } while (bh != head);
 988 out:
 989         if (offset == 0)
 990                 try_to_release_page(page, 0);
 991 }
 992
 993 /**
 994  * gfs2_ok_for_dio - check that dio is valid on this file
 995  * @ip: The inode
 996  * @rw: READ or WRITE
 997  * @offset: The offset at which we are reading or writing
 998  *
 999  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1000  *          1 (to accept the i/o request)
1001  */
1002 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
1003 {
1004         /*
1005          * Should we return an error here? I can't see that O_DIRECT for
1006          * a stuffed file makes any sense. For now we'll silently fall
1007          * back to buffered I/O
1008          */
1009         if (gfs2_is_stuffed(ip))
1010                 return 0;
1011
1012         if (offset >= i_size_read(&ip->i_inode))
1013                 return 0;
1014         return 1;
1015 }
1016
1017
1018
1019 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
1020                               const struct iovec *iov, loff_t offset,
1021                               unsigned long nr_segs)
1022 {
1023         struct file *file = iocb->ki_filp;
1024         struct inode *inode = file->f_mapping->host;
1025         struct gfs2_inode *ip = GFS2_I(inode);
1026         struct gfs2_holder gh;
1027         int rv;
1028
1029         /*
1030          * Deferred lock, even if its a write, since we do no allocation
1031          * on this path. All we need change is atime, and this lock mode
1032          * ensures that other nodes have flushed their buffered read caches
1033          * (i.e. their page cache entries for this inode). We do not,
1034          * unfortunately have the option of only flushing a range like
1035          * the VFS does.
1036          */
1037         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1038         rv = gfs2_glock_nq(&gh);
1039         if (rv)
1040                 return rv;
1041         rv = gfs2_ok_for_dio(ip, rw, offset);
1042         if (rv != 1)
1043                 goto out; /* dio not valid, fall back to buffered i/o */
1044
1045         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1046                                            iov, offset, nr_segs,
1047                                            gfs2_get_block_direct, NULL);
1048 out:
1049         gfs2_glock_dq_m(1, &gh);
1050         gfs2_holder_uninit(&gh);
1051         return rv;
1052 }
1053
1054 /**
1055  * gfs2_releasepage - free the metadata associated with a page
1056  * @page: the page that's being released
1057  * @gfp_mask: passed from Linux VFS, ignored by us
1058  *
1059  * Call try_to_free_buffers() if the buffers in this page can be
1060  * released.
1061  *
1062  * Returns: 0
1063  */
1064
1065 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1066 {
1067         struct address_space *mapping = page->mapping;
1068         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1069         struct buffer_head *bh, *head;
1070         struct gfs2_bufdata *bd;
1071
1072         if (!page_has_buffers(page))
1073                 return 0;
1074
1075         gfs2_log_lock(sdp);
1076         head = bh = page_buffers(page);
1077         do {
1078                 if (atomic_read(&bh->b_count))
1079                         goto cannot_release;
1080                 bd = bh->b_private;
1081                 if (bd && bd->bd_ail)
1082                         goto cannot_release;
1083                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1084                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1085                 bh = bh->b_this_page;
1086         } while(bh != head);
1087         gfs2_log_unlock(sdp);
1088
1089         head = bh = page_buffers(page);
1090         do {
1091                 gfs2_log_lock(sdp);
1092                 bd = bh->b_private;
1093                 if (bd) {
1094                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1095                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1096                         if (!list_empty(&bd->bd_le.le_list)) {
1097                                 if (!buffer_pinned(bh))
1098                                         list_del_init(&bd->bd_le.le_list);
1099                                 else
1100                                         bd = NULL;
1101                         }
1102                         if (bd)
1103                                 bd->bd_bh = NULL;
1104                         bh->b_private = NULL;
1105                 }
1106                 gfs2_log_unlock(sdp);
1107                 if (bd)
1108                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1109
1110                 bh = bh->b_this_page;
1111         } while (bh != head);
1112
1113         return try_to_free_buffers(page);
1114 cannot_release:
1115         gfs2_log_unlock(sdp);
1116         return 0;
1117 }
1118
1119 static const struct address_space_operations gfs2_writeback_aops = {
1120         .writepage = gfs2_writeback_writepage,
1121         .writepages = gfs2_writeback_writepages,
1122         .readpage = gfs2_readpage,
1123         .readpages = gfs2_readpages,
1124         .sync_page = block_sync_page,
1125         .write_begin = gfs2_write_begin,
1126         .write_end = gfs2_write_end,
1127         .bmap = gfs2_bmap,
1128         .invalidatepage = gfs2_invalidatepage,
1129         .releasepage = gfs2_releasepage,
1130         .direct_IO = gfs2_direct_IO,
1131         .migratepage = buffer_migrate_page,
1132         .is_partially_uptodate = block_is_partially_uptodate,
1133         .error_remove_page = generic_error_remove_page,
1134 };
1135
1136 static const struct address_space_operations gfs2_ordered_aops = {
1137         .writepage = gfs2_ordered_writepage,
1138         .readpage = gfs2_readpage,
1139         .readpages = gfs2_readpages,
1140         .sync_page = block_sync_page,
1141         .write_begin = gfs2_write_begin,
1142         .write_end = gfs2_write_end,
1143         .set_page_dirty = gfs2_set_page_dirty,
1144         .bmap = gfs2_bmap,
1145         .invalidatepage = gfs2_invalidatepage,
1146         .releasepage = gfs2_releasepage,
1147         .direct_IO = gfs2_direct_IO,
1148         .migratepage = buffer_migrate_page,
1149         .is_partially_uptodate = block_is_partially_uptodate,
1150         .error_remove_page = generic_error_remove_page,
1151 };
1152
1153 static const struct address_space_operations gfs2_jdata_aops = {
1154         .writepage = gfs2_jdata_writepage,
1155         .writepages = gfs2_jdata_writepages,
1156         .readpage = gfs2_readpage,
1157         .readpages = gfs2_readpages,
1158         .sync_page = block_sync_page,
1159         .write_begin = gfs2_write_begin,
1160         .write_end = gfs2_write_end,
1161         .set_page_dirty = gfs2_set_page_dirty,
1162         .bmap = gfs2_bmap,
1163         .invalidatepage = gfs2_invalidatepage,
1164         .releasepage = gfs2_releasepage,
1165         .is_partially_uptodate = block_is_partially_uptodate,
1166         .error_remove_page = generic_error_remove_page,
1167 };
1168
1169 void gfs2_set_aops(struct inode *inode)
1170 {
1171         struct gfs2_inode *ip = GFS2_I(inode);
1172
1173         if (gfs2_is_writeback(ip))
1174                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1175         else if (gfs2_is_ordered(ip))
1176                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1177         else if (gfs2_is_jdata(ip))
1178                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1179         else
1180                 BUG();
1181 }
1182