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         ret = mpage_writepage(page, gfs2_get_block_noalloc, wbc);
 140         if (ret == -EAGAIN)
 141                 ret = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 142         return ret;
 143 }
 144
 145 /**
 146  * gfs2_ordered_writepage - Write page for ordered data files
 147  * @page: The page to write
 148  * @wbc: The writeback control
 149  *
 150  */
 151
 152 static int gfs2_ordered_writepage(struct page *page,
 153                                   struct writeback_control *wbc)
 154 {
 155         struct inode *inode = page->mapping->host;
 156         struct gfs2_inode *ip = GFS2_I(inode);
 157         int ret;
 158
 159         ret = gfs2_writepage_common(page, wbc);
 160         if (ret <= 0)
 161                 return ret;
 162
 163         if (!page_has_buffers(page)) {
 164                 create_empty_buffers(page, inode->i_sb->s_blocksize,
 165                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
 166         }
 167         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
 168         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 169 }
 170
 171 /**
 172  * __gfs2_jdata_writepage - The core of jdata writepage
 173  * @page: The page to write
 174  * @wbc: The writeback control
 175  *
 176  * This is shared between writepage and writepages and implements the
 177  * core of the writepage operation. If a transaction is required then
 178  * PageChecked will have been set and the transaction will have
 179  * already been started before this is called.
 180  */
 181
 182 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         struct gfs2_inode *ip = GFS2_I(inode);
 186         struct gfs2_sbd *sdp = GFS2_SB(inode);
 187
 188         if (PageChecked(page)) {
 189                 ClearPageChecked(page);
 190                 if (!page_has_buffers(page)) {
 191                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 192                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 193                 }
 194                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 195         }
 196         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 197 }
 198
 199 /**
 200  * gfs2_jdata_writepage - Write complete page
 201  * @page: Page to write
 202  *
 203  * Returns: errno
 204  *
 205  */
 206
 207 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 208 {
 209         struct inode *inode = page->mapping->host;
 210         struct gfs2_sbd *sdp = GFS2_SB(inode);
 211         int ret;
 212         int done_trans = 0;
 213
 214         if (PageChecked(page)) {
 215                 if (wbc->sync_mode != WB_SYNC_ALL)
 216                         goto out_ignore;
 217                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 218                 if (ret)
 219                         goto out_ignore;
 220                 done_trans = 1;
 221         }
 222         ret = gfs2_writepage_common(page, wbc);
 223         if (ret > 0)
 224                 ret = __gfs2_jdata_writepage(page, wbc);
 225         if (done_trans)
 226                 gfs2_trans_end(sdp);
 227         return ret;
 228
 229 out_ignore:
 230         redirty_page_for_writepage(wbc, page);
 231         unlock_page(page);
 232         return 0;
 233 }
 234
 235 /**
 236  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
 237  * @mapping: The mapping to write
 238  * @wbc: Write-back control
 239  *
 240  * For the data=writeback case we can already ignore buffer heads
 241  * and write whole extents at once. This is a big reduction in the
 242  * number of I/O requests we send and the bmap calls we make in this case.
 243  */
 244 static int gfs2_writeback_writepages(struct address_space *mapping,
 245                                      struct writeback_control *wbc)
 246 {
 247         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 248 }
 249
 250 /**
 251  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 252  * @mapping: The mapping
 253  * @wbc: The writeback control
 254  * @writepage: The writepage function to call for each page
 255  * @pvec: The vector of pages
 256  * @nr_pages: The number of pages to write
 257  *
 258  * Returns: non-zero if loop should terminate, zero otherwise
 259  */
 260
 261 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 262                                     struct writeback_control *wbc,
 263                                     struct pagevec *pvec,
 264                                     int nr_pages, pgoff_t end)
 265 {
 266         struct inode *inode = mapping->host;
 267         struct gfs2_sbd *sdp = GFS2_SB(inode);
 268         loff_t i_size = i_size_read(inode);
 269         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 270         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
 271         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
 272         int i;
 273         int ret;
 274
 275         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 276         if (ret < 0)
 277                 return ret;
 278
 279         for(i = 0; i < nr_pages; i++) {
 280                 struct page *page = pvec->pages[i];
 281
 282                 lock_page(page);
 283
 284                 if (unlikely(page->mapping != mapping)) {
 285                         unlock_page(page);
 286                         continue;
 287                 }
 288
 289                 if (!wbc->range_cyclic && page->index > end) {
 290                         ret = 1;
 291                         unlock_page(page);
 292                         continue;
 293                 }
 294
 295                 if (wbc->sync_mode != WB_SYNC_NONE)
 296                         wait_on_page_writeback(page);
 297
 298                 if (PageWriteback(page) ||
 299                     !clear_page_dirty_for_io(page)) {
 300                         unlock_page(page);
 301                         continue;
 302                 }
 303
 304                 /* Is the page fully outside i_size? (truncate in progress) */
 305                 if (page->index > end_index || (page->index == end_index && !offset)) {
 306                         page->mapping->a_ops->invalidatepage(page, 0);
 307                         unlock_page(page);
 308                         continue;
 309                 }
 310
 311                 ret = __gfs2_jdata_writepage(page, wbc);
 312
 313                 if (ret || (--(wbc->nr_to_write) <= 0))
 314                         ret = 1;
 315         }
 316         gfs2_trans_end(sdp);
 317         return ret;
 318 }
 319
 320 /**
 321  * gfs2_write_cache_jdata - Like write_cache_pages but different
 322  * @mapping: The mapping to write
 323  * @wbc: The writeback control
 324  * @writepage: The writepage function to call
 325  * @data: The data to pass to writepage
 326  *
 327  * The reason that we use our own function here is that we need to
 328  * start transactions before we grab page locks. This allows us
 329  * to get the ordering right.
 330  */
 331
 332 static int gfs2_write_cache_jdata(struct address_space *mapping,
 333                                   struct writeback_control *wbc)
 334 {
 335         int ret = 0;
 336         int done = 0;
 337         struct pagevec pvec;
 338         int nr_pages;
 339         pgoff_t index;
 340         pgoff_t end;
 341         int scanned = 0;
 342         int range_whole = 0;
 343
 344         pagevec_init(&pvec, 0);
 345         if (wbc->range_cyclic) {
 346                 index = mapping->writeback_index; /* Start from prev offset */
 347                 end = -1;
 348         } else {
 349                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
 350                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 351                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 352                         range_whole = 1;
 353                 scanned = 1;
 354         }
 355
 356 retry:
 357          while (!done && (index <= end) &&
 358                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 359                                                PAGECACHE_TAG_DIRTY,
 360                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 361                 scanned = 1;
 362                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
 363                 if (ret)
 364                         done = 1;
 365                 if (ret > 0)
 366                         ret = 0;
 367
 368                 pagevec_release(&pvec);
 369                 cond_resched();
 370         }
 371
 372         if (!scanned && !done) {
 373                 /*
 374                  * We hit the last page and there is more work to be done: wrap
 375                  * back to the start of the file
 376                  */
 377                 scanned = 1;
 378                 index = 0;
 379                 goto retry;
 380         }
 381
 382         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 383                 mapping->writeback_index = index;
 384         return ret;
 385 }
 386
 387
 388 /**
 389  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 390  * @mapping: The mapping to write
 391  * @wbc: The writeback control
 392  *
 393  */
 394
 395 static int gfs2_jdata_writepages(struct address_space *mapping,
 396                                  struct writeback_control *wbc)
 397 {
 398         struct gfs2_inode *ip = GFS2_I(mapping->host);
 399         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 400         int ret;
 401
 402         ret = gfs2_write_cache_jdata(mapping, wbc);
 403         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 404                 gfs2_log_flush(sdp, ip->i_gl);
 405                 ret = gfs2_write_cache_jdata(mapping, wbc);
 406         }
 407         return ret;
 408 }
 409
 410 /**
 411  * stuffed_readpage - Fill in a Linux page with stuffed file data
 412  * @ip: the inode
 413  * @page: the page
 414  *
 415  * Returns: errno
 416  */
 417
 418 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 419 {
 420         struct buffer_head *dibh;
 421         u64 dsize = i_size_read(&ip->i_inode);
 422         void *kaddr;
 423         int error;
 424
 425         /*
 426          * Due to the order of unstuffing files and ->fault(), we can be
 427          * asked for a zero page in the case of a stuffed file being extended,
 428          * so we need to supply one here. It doesn't happen often.
 429          */
 430         if (unlikely(page->index)) {
 431                 zero_user(page, 0, PAGE_CACHE_SIZE);
 432                 SetPageUptodate(page);
 433                 return 0;
 434         }
 435
 436         error = gfs2_meta_inode_buffer(ip, &dibh);
 437         if (error)
 438                 return error;
 439
 440         kaddr = kmap_atomic(page, KM_USER0);
 441         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 442                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 443         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 444         memset(kaddr + dsize, 0, PAGE_CACHE_SIZE - dsize);
 445         kunmap_atomic(kaddr, KM_USER0);
 446         flush_dcache_page(page);
 447         brelse(dibh);
 448         SetPageUptodate(page);
 449
 450         return 0;
 451 }
 452
 453
 454 /**
 455  * __gfs2_readpage - readpage
 456  * @file: The file to read a page for
 457  * @page: The page to read
 458  *
 459  * This is the core of gfs2's readpage. Its used by the internal file
 460  * reading code as in that case we already hold the glock. Also its
 461  * called by gfs2_readpage() once the required lock has been granted.
 462  *
 463  */
 464
 465 static int __gfs2_readpage(void *file, struct page *page)
 466 {
 467         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 468         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 469         int error;
 470
 471         if (gfs2_is_stuffed(ip)) {
 472                 error = stuffed_readpage(ip, page);
 473                 unlock_page(page);
 474         } else {
 475                 error = mpage_readpage(page, gfs2_block_map);
 476         }
 477
 478         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 479                 return -EIO;
 480
 481         return error;
 482 }
 483
 484 /**
 485  * gfs2_readpage - read a page of a file
 486  * @file: The file to read
 487  * @page: The page of the file
 488  *
 489  * This deals with the locking required. We have to unlock and
 490  * relock the page in order to get the locking in the right
 491  * order.
 492  */
 493
 494 static int gfs2_readpage(struct file *file, struct page *page)
 495 {
 496         struct address_space *mapping = page->mapping;
 497         struct gfs2_inode *ip = GFS2_I(mapping->host);
 498         struct gfs2_holder gh;
 499         int error;
 500
 501         unlock_page(page);
 502         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 503         error = gfs2_glock_nq(&gh);
 504         if (unlikely(error))
 505                 goto out;
 506         error = AOP_TRUNCATED_PAGE;
 507         lock_page(page);
 508         if (page->mapping == mapping && !PageUptodate(page))
 509                 error = __gfs2_readpage(file, page);
 510         else
 511                 unlock_page(page);
 512         gfs2_glock_dq(&gh);
 513 out:
 514         gfs2_holder_uninit(&gh);
 515         if (error && error != AOP_TRUNCATED_PAGE)
 516                 lock_page(page);
 517         return error;
 518 }
 519
 520 /**
 521  * gfs2_internal_read - read an internal file
 522  * @ip: The gfs2 inode
 523  * @ra_state: The readahead state (or NULL for no readahead)
 524  * @buf: The buffer to fill
 525  * @pos: The file position
 526  * @size: The amount to read
 527  *
 528  */
 529
 530 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
 531                        char *buf, loff_t *pos, unsigned size)
 532 {
 533         struct address_space *mapping = ip->i_inode.i_mapping;
 534         unsigned long index = *pos / PAGE_CACHE_SIZE;
 535         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 536         unsigned copied = 0;
 537         unsigned amt;
 538         struct page *page;
 539         void *p;
 540
 541         do {
 542                 amt = size - copied;
 543                 if (offset + size > PAGE_CACHE_SIZE)
 544                         amt = PAGE_CACHE_SIZE - offset;
 545                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 546                 if (IS_ERR(page))
 547                         return PTR_ERR(page);
 548                 p = kmap_atomic(page, KM_USER0);
 549                 memcpy(buf + copied, p + offset, amt);
 550                 kunmap_atomic(p, KM_USER0);
 551                 mark_page_accessed(page);
 552                 page_cache_release(page);
 553                 copied += amt;
 554                 index++;
 555                 offset = 0;
 556         } while(copied < size);
 557         (*pos) += size;
 558         return size;
 559 }
 560
 561 /**
 562  * gfs2_readpages - Read a bunch of pages at once
 563  *
 564  * Some notes:
 565  * 1. This is only for readahead, so we can simply ignore any things
 566  *    which are slightly inconvenient (such as locking conflicts between
 567  *    the page lock and the glock) and return having done no I/O. Its
 568  *    obviously not something we'd want to do on too regular a basis.
 569  *    Any I/O we ignore at this time will be done via readpage later.
 570  * 2. We don't handle stuffed files here we let readpage do the honours.
 571  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 572  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 573  */
 574
 575 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 576                           struct list_head *pages, unsigned nr_pages)
 577 {
 578         struct inode *inode = mapping->host;
 579         struct gfs2_inode *ip = GFS2_I(inode);
 580         struct gfs2_sbd *sdp = GFS2_SB(inode);
 581         struct gfs2_holder gh;
 582         int ret;
 583
 584         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 585         ret = gfs2_glock_nq(&gh);
 586         if (unlikely(ret))
 587                 goto out_uninit;
 588         if (!gfs2_is_stuffed(ip))
 589                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 590         gfs2_glock_dq(&gh);
 591 out_uninit:
 592         gfs2_holder_uninit(&gh);
 593         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 594                 ret = -EIO;
 595         return ret;
 596 }
 597
 598 /**
 599  * gfs2_write_begin - Begin to write to a file
 600  * @file: The file to write to
 601  * @mapping: The mapping in which to write
 602  * @pos: The file offset at which to start writing
 603  * @len: Length of the write
 604  * @flags: Various flags
 605  * @pagep: Pointer to return the page
 606  * @fsdata: Pointer to return fs data (unused by GFS2)
 607  *
 608  * Returns: errno
 609  */
 610
 611 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 612                             loff_t pos, unsigned len, unsigned flags,
 613                             struct page **pagep, void **fsdata)
 614 {
 615         struct gfs2_inode *ip = GFS2_I(mapping->host);
 616         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 617         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 618         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 619         int alloc_required;
 620         int error = 0;
 621         struct gfs2_alloc *al;
 622         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 623         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 624         unsigned to = from + len;
 625         struct page *page;
 626
 627         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 628         error = gfs2_glock_nq(&ip->i_gh);
 629         if (unlikely(error))
 630                 goto out_uninit;
 631         if (&ip->i_inode == sdp->sd_rindex) {
 632                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 633                                            GL_NOCACHE, &m_ip->i_gh);
 634                 if (unlikely(error)) {
 635                         gfs2_glock_dq(&ip->i_gh);
 636                         goto out_uninit;
 637                 }
 638         }
 639
 640         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
 641         if (error)
 642                 goto out_unlock;
 643
 644         if (alloc_required || gfs2_is_jdata(ip))
 645                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 646
 647         if (alloc_required) {
 648                 al = gfs2_alloc_get(ip);
 649                 if (!al) {
 650                         error = -ENOMEM;
 651                         goto out_unlock;
 652                 }
 653
 654                 error = gfs2_quota_lock_check(ip);
 655                 if (error)
 656                         goto out_alloc_put;
 657
 658                 al->al_requested = data_blocks + ind_blocks;
 659                 error = gfs2_inplace_reserve(ip);
 660                 if (error)
 661                         goto out_qunlock;
 662         }
 663
 664         rblocks = RES_DINODE + ind_blocks;
 665         if (gfs2_is_jdata(ip))
 666                 rblocks += data_blocks ? data_blocks : 1;
 667         if (ind_blocks || data_blocks)
 668                 rblocks += RES_STATFS + RES_QUOTA;
 669         if (&ip->i_inode == sdp->sd_rindex)
 670                 rblocks += 2 * RES_STATFS;
 671
 672         error = gfs2_trans_begin(sdp, rblocks,
 673                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 674         if (error)
 675                 goto out_trans_fail;
 676
 677         error = -ENOMEM;
 678         flags |= AOP_FLAG_NOFS;
 679         page = grab_cache_page_write_begin(mapping, index, flags);
 680         *pagep = page;
 681         if (unlikely(!page))
 682                 goto out_endtrans;
 683
 684         if (gfs2_is_stuffed(ip)) {
 685                 error = 0;
 686                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 687                         error = gfs2_unstuff_dinode(ip, page);
 688                         if (error == 0)
 689                                 goto prepare_write;
 690                 } else if (!PageUptodate(page)) {
 691                         error = stuffed_readpage(ip, page);
 692                 }
 693                 goto out;
 694         }
 695
 696 prepare_write:
 697         error = block_prepare_write(page, from, to, gfs2_block_map);
 698 out:
 699         if (error == 0)
 700                 return 0;
 701
 702         page_cache_release(page);
 703
 704         /*
 705          * XXX(hch): the call below should probably be replaced with
 706          * a call to the gfs2-specific truncate blocks helper to actually
 707          * release disk blocks..
 708          */
 709         if (pos + len > ip->i_inode.i_size)
 710                 simple_setsize(&ip->i_inode, ip->i_inode.i_size);
 711 out_endtrans:
 712         gfs2_trans_end(sdp);
 713 out_trans_fail:
 714         if (alloc_required) {
 715                 gfs2_inplace_release(ip);
 716 out_qunlock:
 717                 gfs2_quota_unlock(ip);
 718 out_alloc_put:
 719                 gfs2_alloc_put(ip);
 720         }
 721 out_unlock:
 722         if (&ip->i_inode == sdp->sd_rindex) {
 723                 gfs2_glock_dq(&m_ip->i_gh);
 724                 gfs2_holder_uninit(&m_ip->i_gh);
 725         }
 726         gfs2_glock_dq(&ip->i_gh);
 727 out_uninit:
 728         gfs2_holder_uninit(&ip->i_gh);
 729         return error;
 730 }
 731
 732 /**
 733  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 734  * @inode: the rindex inode
 735  */
 736 static void adjust_fs_space(struct inode *inode)
 737 {
 738         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 739         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 740         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 741         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 742         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 743         struct buffer_head *m_bh, *l_bh;
 744         u64 fs_total, new_free;
 745
 746         /* Total up the file system space, according to the latest rindex. */
 747         fs_total = gfs2_ri_total(sdp);
 748         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 749                 return;
 750
 751         spin_lock(&sdp->sd_statfs_spin);
 752         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 753                               sizeof(struct gfs2_dinode));
 754         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 755                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 756         else
 757                 new_free = 0;
 758         spin_unlock(&sdp->sd_statfs_spin);
 759         fs_warn(sdp, "File system extended by %llu blocks.\n",
 760                 (unsigned long long)new_free);
 761         gfs2_statfs_change(sdp, new_free, new_free, 0);
 762
 763         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 764                 goto out;
 765         update_statfs(sdp, m_bh, l_bh);
 766         brelse(l_bh);
 767 out:
 768         brelse(m_bh);
 769 }
 770
 771 /**
 772  * gfs2_stuffed_write_end - Write end for stuffed files
 773  * @inode: The inode
 774  * @dibh: The buffer_head containing the on-disk inode
 775  * @pos: The file position
 776  * @len: The length of the write
 777  * @copied: How much was actually copied by the VFS
 778  * @page: The page
 779  *
 780  * This copies the data from the page into the inode block after
 781  * the inode data structure itself.
 782  *
 783  * Returns: errno
 784  */
 785 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 786                                   loff_t pos, unsigned len, unsigned copied,
 787                                   struct page *page)
 788 {
 789         struct gfs2_inode *ip = GFS2_I(inode);
 790         struct gfs2_sbd *sdp = GFS2_SB(inode);
 791         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 792         u64 to = pos + copied;
 793         void *kaddr;
 794         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 795         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
 796
 797         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 798         kaddr = kmap_atomic(page, KM_USER0);
 799         memcpy(buf + pos, kaddr + pos, copied);
 800         memset(kaddr + pos + copied, 0, len - copied);
 801         flush_dcache_page(page);
 802         kunmap_atomic(kaddr, KM_USER0);
 803
 804         if (!PageUptodate(page))
 805                 SetPageUptodate(page);
 806         unlock_page(page);
 807         page_cache_release(page);
 808
 809         if (copied) {
 810                 if (inode->i_size < to) {
 811                         i_size_write(inode, to);
 812                         ip->i_disksize = inode->i_size;
 813                 }
 814                 gfs2_dinode_out(ip, di);
 815                 mark_inode_dirty(inode);
 816         }
 817
 818         if (inode == sdp->sd_rindex) {
 819                 adjust_fs_space(inode);
 820                 ip->i_gh.gh_flags |= GL_NOCACHE;
 821         }
 822
 823         brelse(dibh);
 824         gfs2_trans_end(sdp);
 825         if (inode == sdp->sd_rindex) {
 826                 gfs2_glock_dq(&m_ip->i_gh);
 827                 gfs2_holder_uninit(&m_ip->i_gh);
 828         }
 829         gfs2_glock_dq(&ip->i_gh);
 830         gfs2_holder_uninit(&ip->i_gh);
 831         return copied;
 832 }
 833
 834 /**
 835  * gfs2_write_end
 836  * @file: The file to write to
 837  * @mapping: The address space to write to
 838  * @pos: The file position
 839  * @len: The length of the data
 840  * @copied:
 841  * @page: The page that has been written
 842  * @fsdata: The fsdata (unused in GFS2)
 843  *
 844  * The main write_end function for GFS2. We have a separate one for
 845  * stuffed files as they are slightly different, otherwise we just
 846  * put our locking around the VFS provided functions.
 847  *
 848  * Returns: errno
 849  */
 850
 851 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 852                           loff_t pos, unsigned len, unsigned copied,
 853                           struct page *page, void *fsdata)
 854 {
 855         struct inode *inode = page->mapping->host;
 856         struct gfs2_inode *ip = GFS2_I(inode);
 857         struct gfs2_sbd *sdp = GFS2_SB(inode);
 858         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 859         struct buffer_head *dibh;
 860         struct gfs2_alloc *al = ip->i_alloc;
 861         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 862         unsigned int to = from + len;
 863         int ret;
 864
 865         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 866
 867         ret = gfs2_meta_inode_buffer(ip, &dibh);
 868         if (unlikely(ret)) {
 869                 unlock_page(page);
 870                 page_cache_release(page);
 871                 goto failed;
 872         }
 873
 874         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 875
 876         if (gfs2_is_stuffed(ip))
 877                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 878
 879         if (!gfs2_is_writeback(ip))
 880                 gfs2_page_add_databufs(ip, page, from, to);
 881
 882         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 883         if (ret > 0) {
 884                 if (inode->i_size > ip->i_disksize)
 885                         ip->i_disksize = inode->i_size;
 886                 gfs2_dinode_out(ip, dibh->b_data);
 887                 mark_inode_dirty(inode);
 888         }
 889
 890         if (inode == sdp->sd_rindex) {
 891                 adjust_fs_space(inode);
 892                 ip->i_gh.gh_flags |= GL_NOCACHE;
 893         }
 894
 895         brelse(dibh);
 896         gfs2_trans_end(sdp);
 897 failed:
 898         if (al) {
 899                 gfs2_inplace_release(ip);
 900                 gfs2_quota_unlock(ip);
 901                 gfs2_alloc_put(ip);
 902         }
 903         if (inode == sdp->sd_rindex) {
 904                 gfs2_glock_dq(&m_ip->i_gh);
 905                 gfs2_holder_uninit(&m_ip->i_gh);
 906         }
 907         gfs2_glock_dq(&ip->i_gh);
 908         gfs2_holder_uninit(&ip->i_gh);
 909         return ret;
 910 }
 911
 912 /**
 913  * gfs2_set_page_dirty - Page dirtying function
 914  * @page: The page to dirty
 915  *
 916  * Returns: 1 if it dirtyed the page, or 0 otherwise
 917  */
 918
 919 static int gfs2_set_page_dirty(struct page *page)
 920 {
 921         SetPageChecked(page);
 922         return __set_page_dirty_buffers(page);
 923 }
 924
 925 /**
 926  * gfs2_bmap - Block map function
 927  * @mapping: Address space info
 928  * @lblock: The block to map
 929  *
 930  * Returns: The disk address for the block or 0 on hole or error
 931  */
 932
 933 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 934 {
 935         struct gfs2_inode *ip = GFS2_I(mapping->host);
 936         struct gfs2_holder i_gh;
 937         sector_t dblock = 0;
 938         int error;
 939
 940         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 941         if (error)
 942                 return 0;
 943
 944         if (!gfs2_is_stuffed(ip))
 945                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 946
 947         gfs2_glock_dq_uninit(&i_gh);
 948
 949         return dblock;
 950 }
 951
 952 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 953 {
 954         struct gfs2_bufdata *bd;
 955
 956         lock_buffer(bh);
 957         gfs2_log_lock(sdp);
 958         clear_buffer_dirty(bh);
 959         bd = bh->b_private;
 960         if (bd) {
 961                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 962                         list_del_init(&bd->bd_le.le_list);
 963                 else
 964                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 965         }
 966         bh->b_bdev = NULL;
 967         clear_buffer_mapped(bh);
 968         clear_buffer_req(bh);
 969         clear_buffer_new(bh);
 970         gfs2_log_unlock(sdp);
 971         unlock_buffer(bh);
 972 }
 973
 974 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 975 {
 976         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 977         struct buffer_head *bh, *head;
 978         unsigned long pos = 0;
 979
 980         BUG_ON(!PageLocked(page));
 981         if (offset == 0)
 982                 ClearPageChecked(page);
 983         if (!page_has_buffers(page))
 984                 goto out;
 985
 986         bh = head = page_buffers(page);
 987         do {
 988                 if (offset <= pos)
 989                         gfs2_discard(sdp, bh);
 990                 pos += bh->b_size;
 991                 bh = bh->b_this_page;
 992         } while (bh != head);
 993 out:
 994         if (offset == 0)
 995                 try_to_release_page(page, 0);
 996 }
 997
 998 /**
 999  * gfs2_ok_for_dio - check that dio is valid on this file
1000  * @ip: The inode
1001  * @rw: READ or WRITE
1002  * @offset: The offset at which we are reading or writing
1003  *
1004  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1005  *          1 (to accept the i/o request)
1006  */
1007 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
1008 {
1009         /*
1010          * Should we return an error here? I can't see that O_DIRECT for
1011          * a stuffed file makes any sense. For now we'll silently fall
1012          * back to buffered I/O
1013          */
1014         if (gfs2_is_stuffed(ip))
1015                 return 0;
1016
1017         if (offset >= i_size_read(&ip->i_inode))
1018                 return 0;
1019         return 1;
1020 }
1021
1022
1023
1024 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
1025                               const struct iovec *iov, loff_t offset,
1026                               unsigned long nr_segs)
1027 {
1028         struct file *file = iocb->ki_filp;
1029         struct inode *inode = file->f_mapping->host;
1030         struct gfs2_inode *ip = GFS2_I(inode);
1031         struct gfs2_holder gh;
1032         int rv;
1033
1034         /*
1035          * Deferred lock, even if its a write, since we do no allocation
1036          * on this path. All we need change is atime, and this lock mode
1037          * ensures that other nodes have flushed their buffered read caches
1038          * (i.e. their page cache entries for this inode). We do not,
1039          * unfortunately have the option of only flushing a range like
1040          * the VFS does.
1041          */
1042         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1043         rv = gfs2_glock_nq(&gh);
1044         if (rv)
1045                 return rv;
1046         rv = gfs2_ok_for_dio(ip, rw, offset);
1047         if (rv != 1)
1048                 goto out; /* dio not valid, fall back to buffered i/o */
1049
1050         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1051                                            iov, offset, nr_segs,
1052                                            gfs2_get_block_direct, NULL);
1053 out:
1054         gfs2_glock_dq_m(1, &gh);
1055         gfs2_holder_uninit(&gh);
1056         return rv;
1057 }
1058
1059 /**
1060  * gfs2_releasepage - free the metadata associated with a page
1061  * @page: the page that's being released
1062  * @gfp_mask: passed from Linux VFS, ignored by us
1063  *
1064  * Call try_to_free_buffers() if the buffers in this page can be
1065  * released.
1066  *
1067  * Returns: 0
1068  */
1069
1070 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1071 {
1072         struct address_space *mapping = page->mapping;
1073         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1074         struct buffer_head *bh, *head;
1075         struct gfs2_bufdata *bd;
1076
1077         if (!page_has_buffers(page))
1078                 return 0;
1079
1080         gfs2_log_lock(sdp);
1081         head = bh = page_buffers(page);
1082         do {
1083                 if (atomic_read(&bh->b_count))
1084                         goto cannot_release;
1085                 bd = bh->b_private;
1086                 if (bd && bd->bd_ail)
1087                         goto cannot_release;
1088                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1089                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1090                 bh = bh->b_this_page;
1091         } while(bh != head);
1092         gfs2_log_unlock(sdp);
1093
1094         head = bh = page_buffers(page);
1095         do {
1096                 gfs2_log_lock(sdp);
1097                 bd = bh->b_private;
1098                 if (bd) {
1099                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1100                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1101                         if (!list_empty(&bd->bd_le.le_list)) {
1102                                 if (!buffer_pinned(bh))
1103                                         list_del_init(&bd->bd_le.le_list);
1104                                 else
1105                                         bd = NULL;
1106                         }
1107                         if (bd)
1108                                 bd->bd_bh = NULL;
1109                         bh->b_private = NULL;
1110                 }
1111                 gfs2_log_unlock(sdp);
1112                 if (bd)
1113                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1114
1115                 bh = bh->b_this_page;
1116         } while (bh != head);
1117
1118         return try_to_free_buffers(page);
1119 cannot_release:
1120         gfs2_log_unlock(sdp);
1121         return 0;
1122 }
1123
1124 static const struct address_space_operations gfs2_writeback_aops = {
1125         .writepage = gfs2_writeback_writepage,
1126         .writepages = gfs2_writeback_writepages,
1127         .readpage = gfs2_readpage,
1128         .readpages = gfs2_readpages,
1129         .sync_page = block_sync_page,
1130         .write_begin = gfs2_write_begin,
1131         .write_end = gfs2_write_end,
1132         .bmap = gfs2_bmap,
1133         .invalidatepage = gfs2_invalidatepage,
1134         .releasepage = gfs2_releasepage,
1135         .direct_IO = gfs2_direct_IO,
1136         .migratepage = buffer_migrate_page,
1137         .is_partially_uptodate = block_is_partially_uptodate,
1138         .error_remove_page = generic_error_remove_page,
1139 };
1140
1141 static const struct address_space_operations gfs2_ordered_aops = {
1142         .writepage = gfs2_ordered_writepage,
1143         .readpage = gfs2_readpage,
1144         .readpages = gfs2_readpages,
1145         .sync_page = block_sync_page,
1146         .write_begin = gfs2_write_begin,
1147         .write_end = gfs2_write_end,
1148         .set_page_dirty = gfs2_set_page_dirty,
1149         .bmap = gfs2_bmap,
1150         .invalidatepage = gfs2_invalidatepage,
1151         .releasepage = gfs2_releasepage,
1152         .direct_IO = gfs2_direct_IO,
1153         .migratepage = buffer_migrate_page,
1154         .is_partially_uptodate = block_is_partially_uptodate,
1155         .error_remove_page = generic_error_remove_page,
1156 };
1157
1158 static const struct address_space_operations gfs2_jdata_aops = {
1159         .writepage = gfs2_jdata_writepage,
1160         .writepages = gfs2_jdata_writepages,
1161         .readpage = gfs2_readpage,
1162         .readpages = gfs2_readpages,
1163         .sync_page = block_sync_page,
1164         .write_begin = gfs2_write_begin,
1165         .write_end = gfs2_write_end,
1166         .set_page_dirty = gfs2_set_page_dirty,
1167         .bmap = gfs2_bmap,
1168         .invalidatepage = gfs2_invalidatepage,
1169         .releasepage = gfs2_releasepage,
1170         .is_partially_uptodate = block_is_partially_uptodate,
1171         .error_remove_page = generic_error_remove_page,
1172 };
1173
1174 void gfs2_set_aops(struct inode *inode)
1175 {
1176         struct gfs2_inode *ip = GFS2_I(inode);
1177
1178         if (gfs2_is_writeback(ip))
1179                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1180         else if (gfs2_is_ordered(ip))
1181                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1182         else if (gfs2_is_jdata(ip))
1183                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1184         else
1185                 BUG();
1186 }
1187