fs/gfs2/ops_address.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/lm_interface.h>
  23 #include <linux/backing-dev.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 "ops_address.h"
  33 #include "quota.h"
  34 #include "trans.h"
  35 #include "rgrp.h"
  36 #include "super.h"
  37 #include "util.h"
  38 #include "glops.h"
  39
  40
  41 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  42                                    unsigned int from, unsigned int to)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  47         unsigned int start, end;
  48
  49         for (bh = head, start = 0; bh != head || !start;
  50              bh = bh->b_this_page, start = end) {
  51                 end = start + bsize;
  52                 if (end <= from || start >= to)
  53                         continue;
  54                 if (gfs2_is_jdata(ip))
  55                         set_buffer_uptodate(bh);
  56                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  57         }
  58 }
  59
  60 /**
  61  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  62  * @inode: The inode
  63  * @lblock: The block number to look up
  64  * @bh_result: The buffer head to return the result in
  65  * @create: Non-zero if we may add block to the file
  66  *
  67  * Returns: errno
  68  */
  69
  70 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  71                                   struct buffer_head *bh_result, int create)
  72 {
  73         int error;
  74
  75         error = gfs2_block_map(inode, lblock, bh_result, 0);
  76         if (error)
  77                 return error;
  78         if (!buffer_mapped(bh_result))
  79                 return -EIO;
  80         return 0;
  81 }
  82
  83 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  84                                  struct buffer_head *bh_result, int create)
  85 {
  86         return gfs2_block_map(inode, lblock, bh_result, 0);
  87 }
  88
  89 /**
  90  * gfs2_writepage_common - Common bits of writepage
  91  * @page: The page to be written
  92  * @wbc: The writeback control
  93  *
  94  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  95  */
  96
  97 static int gfs2_writepage_common(struct page *page,
  98                                  struct writeback_control *wbc)
  99 {
 100         struct inode *inode = page->mapping->host;
 101         struct gfs2_inode *ip = GFS2_I(inode);
 102         struct gfs2_sbd *sdp = GFS2_SB(inode);
 103         loff_t i_size = i_size_read(inode);
 104         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 105         unsigned offset;
 106
 107         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 108                 goto out;
 109         if (current->journal_info)
 110                 goto redirty;
 111         /* Is the page fully outside i_size? (truncate in progress) */
 112         offset = i_size & (PAGE_CACHE_SIZE-1);
 113         if (page->index > end_index || (page->index == end_index && !offset)) {
 114                 page->mapping->a_ops->invalidatepage(page, 0);
 115                 goto out;
 116         }
 117         return 1;
 118 redirty:
 119         redirty_page_for_writepage(wbc, page);
 120 out:
 121         unlock_page(page);
 122         return 0;
 123 }
 124
 125 /**
 126  * gfs2_writeback_writepage - Write page for writeback mappings
 127  * @page: The page
 128  * @wbc: The writeback control
 129  *
 130  */
 131
 132 static int gfs2_writeback_writepage(struct page *page,
 133                                     struct writeback_control *wbc)
 134 {
 135         int ret;
 136
 137         ret = gfs2_writepage_common(page, wbc);
 138         if (ret <= 0)
 139                 return ret;
 140
 141         ret = mpage_writepage(page, gfs2_get_block_noalloc, wbc);
 142         if (ret == -EAGAIN)
 143                 ret = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 144         return ret;
 145 }
 146
 147 /**
 148  * gfs2_ordered_writepage - Write page for ordered data files
 149  * @page: The page to write
 150  * @wbc: The writeback control
 151  *
 152  */
 153
 154 static int gfs2_ordered_writepage(struct page *page,
 155                                   struct writeback_control *wbc)
 156 {
 157         struct inode *inode = page->mapping->host;
 158         struct gfs2_inode *ip = GFS2_I(inode);
 159         int ret;
 160
 161         ret = gfs2_writepage_common(page, wbc);
 162         if (ret <= 0)
 163                 return ret;
 164
 165         if (!page_has_buffers(page)) {
 166                 create_empty_buffers(page, inode->i_sb->s_blocksize,
 167                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
 168         }
 169         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
 170         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 171 }
 172
 173 /**
 174  * __gfs2_jdata_writepage - The core of jdata writepage
 175  * @page: The page to write
 176  * @wbc: The writeback control
 177  *
 178  * This is shared between writepage and writepages and implements the
 179  * core of the writepage operation. If a transaction is required then
 180  * PageChecked will have been set and the transaction will have
 181  * already been started before this is called.
 182  */
 183
 184 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 185 {
 186         struct inode *inode = page->mapping->host;
 187         struct gfs2_inode *ip = GFS2_I(inode);
 188         struct gfs2_sbd *sdp = GFS2_SB(inode);
 189
 190         if (PageChecked(page)) {
 191                 ClearPageChecked(page);
 192                 if (!page_has_buffers(page)) {
 193                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 194                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 195                 }
 196                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 197         }
 198         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 199 }
 200
 201 /**
 202  * gfs2_jdata_writepage - Write complete page
 203  * @page: Page to write
 204  *
 205  * Returns: errno
 206  *
 207  */
 208
 209 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 210 {
 211         struct inode *inode = page->mapping->host;
 212         struct gfs2_sbd *sdp = GFS2_SB(inode);
 213         int ret;
 214         int done_trans = 0;
 215
 216         if (PageChecked(page)) {
 217                 if (wbc->sync_mode != WB_SYNC_ALL)
 218                         goto out_ignore;
 219                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 220                 if (ret)
 221                         goto out_ignore;
 222                 done_trans = 1;
 223         }
 224         ret = gfs2_writepage_common(page, wbc);
 225         if (ret > 0)
 226                 ret = __gfs2_jdata_writepage(page, wbc);
 227         if (done_trans)
 228                 gfs2_trans_end(sdp);
 229         return ret;
 230
 231 out_ignore:
 232         redirty_page_for_writepage(wbc, page);
 233         unlock_page(page);
 234         return 0;
 235 }
 236
 237 /**
 238  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
 239  * @mapping: The mapping to write
 240  * @wbc: Write-back control
 241  *
 242  * For the data=writeback case we can already ignore buffer heads
 243  * and write whole extents at once. This is a big reduction in the
 244  * number of I/O requests we send and the bmap calls we make in this case.
 245  */
 246 static int gfs2_writeback_writepages(struct address_space *mapping,
 247                                      struct writeback_control *wbc)
 248 {
 249         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 250 }
 251
 252 /**
 253  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 254  * @mapping: The mapping
 255  * @wbc: The writeback control
 256  * @writepage: The writepage function to call for each page
 257  * @pvec: The vector of pages
 258  * @nr_pages: The number of pages to write
 259  *
 260  * Returns: non-zero if loop should terminate, zero otherwise
 261  */
 262
 263 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 264                                     struct writeback_control *wbc,
 265                                     struct pagevec *pvec,
 266                                     int nr_pages, pgoff_t end)
 267 {
 268         struct inode *inode = mapping->host;
 269         struct gfs2_sbd *sdp = GFS2_SB(inode);
 270         loff_t i_size = i_size_read(inode);
 271         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 272         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
 273         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
 274         struct backing_dev_info *bdi = mapping->backing_dev_info;
 275         int i;
 276         int ret;
 277
 278         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 279         if (ret < 0)
 280                 return ret;
 281
 282         for(i = 0; i < nr_pages; i++) {
 283                 struct page *page = pvec->pages[i];
 284
 285                 lock_page(page);
 286
 287                 if (unlikely(page->mapping != mapping)) {
 288                         unlock_page(page);
 289                         continue;
 290                 }
 291
 292                 if (!wbc->range_cyclic && page->index > end) {
 293                         ret = 1;
 294                         unlock_page(page);
 295                         continue;
 296                 }
 297
 298                 if (wbc->sync_mode != WB_SYNC_NONE)
 299                         wait_on_page_writeback(page);
 300
 301                 if (PageWriteback(page) ||
 302                     !clear_page_dirty_for_io(page)) {
 303                         unlock_page(page);
 304                         continue;
 305                 }
 306
 307                 /* Is the page fully outside i_size? (truncate in progress) */
 308                 if (page->index > end_index || (page->index == end_index && !offset)) {
 309                         page->mapping->a_ops->invalidatepage(page, 0);
 310                         unlock_page(page);
 311                         continue;
 312                 }
 313
 314                 ret = __gfs2_jdata_writepage(page, wbc);
 315
 316                 if (ret || (--(wbc->nr_to_write) <= 0))
 317                         ret = 1;
 318                 if (wbc->nonblocking && bdi_write_congested(bdi)) {
 319                         wbc->encountered_congestion = 1;
 320                         ret = 1;
 321                 }
 322
 323         }
 324         gfs2_trans_end(sdp);
 325         return ret;
 326 }
 327
 328 /**
 329  * gfs2_write_cache_jdata - Like write_cache_pages but different
 330  * @mapping: The mapping to write
 331  * @wbc: The writeback control
 332  * @writepage: The writepage function to call
 333  * @data: The data to pass to writepage
 334  *
 335  * The reason that we use our own function here is that we need to
 336  * start transactions before we grab page locks. This allows us
 337  * to get the ordering right.
 338  */
 339
 340 static int gfs2_write_cache_jdata(struct address_space *mapping,
 341                                   struct writeback_control *wbc)
 342 {
 343         struct backing_dev_info *bdi = mapping->backing_dev_info;
 344         int ret = 0;
 345         int done = 0;
 346         struct pagevec pvec;
 347         int nr_pages;
 348         pgoff_t index;
 349         pgoff_t end;
 350         int scanned = 0;
 351         int range_whole = 0;
 352
 353         if (wbc->nonblocking && bdi_write_congested(bdi)) {
 354                 wbc->encountered_congestion = 1;
 355                 return 0;
 356         }
 357
 358         pagevec_init(&pvec, 0);
 359         if (wbc->range_cyclic) {
 360                 index = mapping->writeback_index; /* Start from prev offset */
 361                 end = -1;
 362         } else {
 363                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
 364                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 365                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 366                         range_whole = 1;
 367                 scanned = 1;
 368         }
 369
 370 retry:
 371          while (!done && (index <= end) &&
 372                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 373                                                PAGECACHE_TAG_DIRTY,
 374                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 375                 scanned = 1;
 376                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
 377                 if (ret)
 378                         done = 1;
 379                 if (ret > 0)
 380                         ret = 0;
 381
 382                 pagevec_release(&pvec);
 383                 cond_resched();
 384         }
 385
 386         if (!scanned && !done) {
 387                 /*
 388                  * We hit the last page and there is more work to be done: wrap
 389                  * back to the start of the file
 390                  */
 391                 scanned = 1;
 392                 index = 0;
 393                 goto retry;
 394         }
 395
 396         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 397                 mapping->writeback_index = index;
 398         return ret;
 399 }
 400
 401
 402 /**
 403  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 404  * @mapping: The mapping to write
 405  * @wbc: The writeback control
 406  *
 407  */
 408
 409 static int gfs2_jdata_writepages(struct address_space *mapping,
 410                                  struct writeback_control *wbc)
 411 {
 412         struct gfs2_inode *ip = GFS2_I(mapping->host);
 413         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 414         int ret;
 415
 416         ret = gfs2_write_cache_jdata(mapping, wbc);
 417         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 418                 gfs2_log_flush(sdp, ip->i_gl);
 419                 ret = gfs2_write_cache_jdata(mapping, wbc);
 420         }
 421         return ret;
 422 }
 423
 424 /**
 425  * stuffed_readpage - Fill in a Linux page with stuffed file data
 426  * @ip: the inode
 427  * @page: the page
 428  *
 429  * Returns: errno
 430  */
 431
 432 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 433 {
 434         struct buffer_head *dibh;
 435         void *kaddr;
 436         int error;
 437
 438         /*
 439          * Due to the order of unstuffing files and ->fault(), we can be
 440          * asked for a zero page in the case of a stuffed file being extended,
 441          * so we need to supply one here. It doesn't happen often.
 442          */
 443         if (unlikely(page->index)) {
 444                 zero_user(page, 0, PAGE_CACHE_SIZE);
 445                 return 0;
 446         }
 447
 448         error = gfs2_meta_inode_buffer(ip, &dibh);
 449         if (error)
 450                 return error;
 451
 452         kaddr = kmap_atomic(page, KM_USER0);
 453         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 454                ip->i_disksize);
 455         memset(kaddr + ip->i_disksize, 0, PAGE_CACHE_SIZE - ip->i_disksize);
 456         kunmap_atomic(kaddr, KM_USER0);
 457         flush_dcache_page(page);
 458         brelse(dibh);
 459         SetPageUptodate(page);
 460
 461         return 0;
 462 }
 463
 464
 465 /**
 466  * __gfs2_readpage - readpage
 467  * @file: The file to read a page for
 468  * @page: The page to read
 469  *
 470  * This is the core of gfs2's readpage. Its used by the internal file
 471  * reading code as in that case we already hold the glock. Also its
 472  * called by gfs2_readpage() once the required lock has been granted.
 473  *
 474  */
 475
 476 static int __gfs2_readpage(void *file, struct page *page)
 477 {
 478         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 479         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 480         int error;
 481
 482         if (gfs2_is_stuffed(ip)) {
 483                 error = stuffed_readpage(ip, page);
 484                 unlock_page(page);
 485         } else {
 486                 error = mpage_readpage(page, gfs2_block_map);
 487         }
 488
 489         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 490                 return -EIO;
 491
 492         return error;
 493 }
 494
 495 /**
 496  * gfs2_readpage - read a page of a file
 497  * @file: The file to read
 498  * @page: The page of the file
 499  *
 500  * This deals with the locking required. We have to unlock and
 501  * relock the page in order to get the locking in the right
 502  * order.
 503  */
 504
 505 static int gfs2_readpage(struct file *file, struct page *page)
 506 {
 507         struct address_space *mapping = page->mapping;
 508         struct gfs2_inode *ip = GFS2_I(mapping->host);
 509         struct gfs2_holder gh;
 510         int error;
 511
 512         unlock_page(page);
 513         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 514         error = gfs2_glock_nq(&gh);
 515         if (unlikely(error))
 516                 goto out;
 517         error = AOP_TRUNCATED_PAGE;
 518         lock_page(page);
 519         if (page->mapping == mapping && !PageUptodate(page))
 520                 error = __gfs2_readpage(file, page);
 521         else
 522                 unlock_page(page);
 523         gfs2_glock_dq(&gh);
 524 out:
 525         gfs2_holder_uninit(&gh);
 526         if (error && error != AOP_TRUNCATED_PAGE)
 527                 lock_page(page);
 528         return error;
 529 }
 530
 531 /**
 532  * gfs2_internal_read - read an internal file
 533  * @ip: The gfs2 inode
 534  * @ra_state: The readahead state (or NULL for no readahead)
 535  * @buf: The buffer to fill
 536  * @pos: The file position
 537  * @size: The amount to read
 538  *
 539  */
 540
 541 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
 542                        char *buf, loff_t *pos, unsigned size)
 543 {
 544         struct address_space *mapping = ip->i_inode.i_mapping;
 545         unsigned long index = *pos / PAGE_CACHE_SIZE;
 546         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 547         unsigned copied = 0;
 548         unsigned amt;
 549         struct page *page;
 550         void *p;
 551
 552         do {
 553                 amt = size - copied;
 554                 if (offset + size > PAGE_CACHE_SIZE)
 555                         amt = PAGE_CACHE_SIZE - offset;
 556                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 557                 if (IS_ERR(page))
 558                         return PTR_ERR(page);
 559                 p = kmap_atomic(page, KM_USER0);
 560                 memcpy(buf + copied, p + offset, amt);
 561                 kunmap_atomic(p, KM_USER0);
 562                 mark_page_accessed(page);
 563                 page_cache_release(page);
 564                 copied += amt;
 565                 index++;
 566                 offset = 0;
 567         } while(copied < size);
 568         (*pos) += size;
 569         return size;
 570 }
 571
 572 /**
 573  * gfs2_readpages - Read a bunch of pages at once
 574  *
 575  * Some notes:
 576  * 1. This is only for readahead, so we can simply ignore any things
 577  *    which are slightly inconvenient (such as locking conflicts between
 578  *    the page lock and the glock) and return having done no I/O. Its
 579  *    obviously not something we'd want to do on too regular a basis.
 580  *    Any I/O we ignore at this time will be done via readpage later.
 581  * 2. We don't handle stuffed files here we let readpage do the honours.
 582  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 583  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 584  */
 585
 586 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 587                           struct list_head *pages, unsigned nr_pages)
 588 {
 589         struct inode *inode = mapping->host;
 590         struct gfs2_inode *ip = GFS2_I(inode);
 591         struct gfs2_sbd *sdp = GFS2_SB(inode);
 592         struct gfs2_holder gh;
 593         int ret;
 594
 595         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 596         ret = gfs2_glock_nq(&gh);
 597         if (unlikely(ret))
 598                 goto out_uninit;
 599         if (!gfs2_is_stuffed(ip))
 600                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 601         gfs2_glock_dq(&gh);
 602 out_uninit:
 603         gfs2_holder_uninit(&gh);
 604         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 605                 ret = -EIO;
 606         return ret;
 607 }
 608
 609 /**
 610  * gfs2_write_begin - Begin to write to a file
 611  * @file: The file to write to
 612  * @mapping: The mapping in which to write
 613  * @pos: The file offset at which to start writing
 614  * @len: Length of the write
 615  * @flags: Various flags
 616  * @pagep: Pointer to return the page
 617  * @fsdata: Pointer to return fs data (unused by GFS2)
 618  *
 619  * Returns: errno
 620  */
 621
 622 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 623                             loff_t pos, unsigned len, unsigned flags,
 624                             struct page **pagep, void **fsdata)
 625 {
 626         struct gfs2_inode *ip = GFS2_I(mapping->host);
 627         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 628         unsigned int data_blocks, ind_blocks, rblocks;
 629         int alloc_required;
 630         int error = 0;
 631         struct gfs2_alloc *al;
 632         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 633         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 634         unsigned to = from + len;
 635         struct page *page;
 636
 637         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 638         error = gfs2_glock_nq(&ip->i_gh);
 639         if (unlikely(error))
 640                 goto out_uninit;
 641
 642         gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 643         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
 644         if (error)
 645                 goto out_unlock;
 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
 670         error = gfs2_trans_begin(sdp, rblocks,
 671                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 672         if (error)
 673                 goto out_trans_fail;
 674
 675         error = -ENOMEM;
 676         page = grab_cache_page_write_begin(mapping, index, flags);
 677         *pagep = page;
 678         if (unlikely(!page))
 679                 goto out_endtrans;
 680
 681         if (gfs2_is_stuffed(ip)) {
 682                 error = 0;
 683                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 684                         error = gfs2_unstuff_dinode(ip, page);
 685                         if (error == 0)
 686                                 goto prepare_write;
 687                 } else if (!PageUptodate(page)) {
 688                         error = stuffed_readpage(ip, page);
 689                 }
 690                 goto out;
 691         }
 692
 693 prepare_write:
 694         error = block_prepare_write(page, from, to, gfs2_block_map);
 695 out:
 696         if (error == 0)
 697                 return 0;
 698
 699         page_cache_release(page);
 700         if (pos + len > ip->i_inode.i_size)
 701                 vmtruncate(&ip->i_inode, ip->i_inode.i_size);
 702 out_endtrans:
 703         gfs2_trans_end(sdp);
 704 out_trans_fail:
 705         if (alloc_required) {
 706                 gfs2_inplace_release(ip);
 707 out_qunlock:
 708                 gfs2_quota_unlock(ip);
 709 out_alloc_put:
 710                 gfs2_alloc_put(ip);
 711         }
 712 out_unlock:
 713         gfs2_glock_dq(&ip->i_gh);
 714 out_uninit:
 715         gfs2_holder_uninit(&ip->i_gh);
 716         return error;
 717 }
 718
 719 /**
 720  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 721  * @inode: the rindex inode
 722  */
 723 static void adjust_fs_space(struct inode *inode)
 724 {
 725         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 726         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 727         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 728         u64 fs_total, new_free;
 729
 730         /* Total up the file system space, according to the latest rindex. */
 731         fs_total = gfs2_ri_total(sdp);
 732
 733         spin_lock(&sdp->sd_statfs_spin);
 734         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 735                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 736         else
 737                 new_free = 0;
 738         spin_unlock(&sdp->sd_statfs_spin);
 739         fs_warn(sdp, "File system extended by %llu blocks.\n",
 740                 (unsigned long long)new_free);
 741         gfs2_statfs_change(sdp, new_free, new_free, 0);
 742 }
 743
 744 /**
 745  * gfs2_stuffed_write_end - Write end for stuffed files
 746  * @inode: The inode
 747  * @dibh: The buffer_head containing the on-disk inode
 748  * @pos: The file position
 749  * @len: The length of the write
 750  * @copied: How much was actually copied by the VFS
 751  * @page: The page
 752  *
 753  * This copies the data from the page into the inode block after
 754  * the inode data structure itself.
 755  *
 756  * Returns: errno
 757  */
 758 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 759                                   loff_t pos, unsigned len, unsigned copied,
 760                                   struct page *page)
 761 {
 762         struct gfs2_inode *ip = GFS2_I(inode);
 763         struct gfs2_sbd *sdp = GFS2_SB(inode);
 764         u64 to = pos + copied;
 765         void *kaddr;
 766         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 767         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
 768
 769         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 770         kaddr = kmap_atomic(page, KM_USER0);
 771         memcpy(buf + pos, kaddr + pos, copied);
 772         memset(kaddr + pos + copied, 0, len - copied);
 773         flush_dcache_page(page);
 774         kunmap_atomic(kaddr, KM_USER0);
 775
 776         if (!PageUptodate(page))
 777                 SetPageUptodate(page);
 778         unlock_page(page);
 779         page_cache_release(page);
 780
 781         if (inode->i_size < to) {
 782                 i_size_write(inode, to);
 783                 ip->i_disksize = inode->i_size;
 784                 di->di_size = cpu_to_be64(inode->i_size);
 785                 mark_inode_dirty(inode);
 786         }
 787
 788         if (inode == sdp->sd_rindex)
 789                 adjust_fs_space(inode);
 790
 791         brelse(dibh);
 792         gfs2_trans_end(sdp);
 793         gfs2_glock_dq(&ip->i_gh);
 794         gfs2_holder_uninit(&ip->i_gh);
 795         return copied;
 796 }
 797
 798 /**
 799  * gfs2_write_end
 800  * @file: The file to write to
 801  * @mapping: The address space to write to
 802  * @pos: The file position
 803  * @len: The length of the data
 804  * @copied:
 805  * @page: The page that has been written
 806  * @fsdata: The fsdata (unused in GFS2)
 807  *
 808  * The main write_end function for GFS2. We have a separate one for
 809  * stuffed files as they are slightly different, otherwise we just
 810  * put our locking around the VFS provided functions.
 811  *
 812  * Returns: errno
 813  */
 814
 815 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 816                           loff_t pos, unsigned len, unsigned copied,
 817                           struct page *page, void *fsdata)
 818 {
 819         struct inode *inode = page->mapping->host;
 820         struct gfs2_inode *ip = GFS2_I(inode);
 821         struct gfs2_sbd *sdp = GFS2_SB(inode);
 822         struct buffer_head *dibh;
 823         struct gfs2_alloc *al = ip->i_alloc;
 824         struct gfs2_dinode *di;
 825         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 826         unsigned int to = from + len;
 827         int ret;
 828
 829         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 830
 831         ret = gfs2_meta_inode_buffer(ip, &dibh);
 832         if (unlikely(ret)) {
 833                 unlock_page(page);
 834                 page_cache_release(page);
 835                 goto failed;
 836         }
 837
 838         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 839
 840         if (gfs2_is_stuffed(ip))
 841                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 842
 843         if (!gfs2_is_writeback(ip))
 844                 gfs2_page_add_databufs(ip, page, from, to);
 845
 846         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 847
 848         if (likely(ret >= 0) && (inode->i_size > ip->i_disksize)) {
 849                 di = (struct gfs2_dinode *)dibh->b_data;
 850                 ip->i_disksize = inode->i_size;
 851                 di->di_size = cpu_to_be64(inode->i_size);
 852                 mark_inode_dirty(inode);
 853         }
 854
 855         if (inode == sdp->sd_rindex)
 856                 adjust_fs_space(inode);
 857
 858         brelse(dibh);
 859         gfs2_trans_end(sdp);
 860 failed:
 861         if (al) {
 862                 gfs2_inplace_release(ip);
 863                 gfs2_quota_unlock(ip);
 864                 gfs2_alloc_put(ip);
 865         }
 866         gfs2_glock_dq(&ip->i_gh);
 867         gfs2_holder_uninit(&ip->i_gh);
 868         return ret;
 869 }
 870
 871 /**
 872  * gfs2_set_page_dirty - Page dirtying function
 873  * @page: The page to dirty
 874  *
 875  * Returns: 1 if it dirtyed the page, or 0 otherwise
 876  */
 877
 878 static int gfs2_set_page_dirty(struct page *page)
 879 {
 880         SetPageChecked(page);
 881         return __set_page_dirty_buffers(page);
 882 }
 883
 884 /**
 885  * gfs2_bmap - Block map function
 886  * @mapping: Address space info
 887  * @lblock: The block to map
 888  *
 889  * Returns: The disk address for the block or 0 on hole or error
 890  */
 891
 892 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 893 {
 894         struct gfs2_inode *ip = GFS2_I(mapping->host);
 895         struct gfs2_holder i_gh;
 896         sector_t dblock = 0;
 897         int error;
 898
 899         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 900         if (error)
 901                 return 0;
 902
 903         if (!gfs2_is_stuffed(ip))
 904                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 905
 906         gfs2_glock_dq_uninit(&i_gh);
 907
 908         return dblock;
 909 }
 910
 911 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 912 {
 913         struct gfs2_bufdata *bd;
 914
 915         lock_buffer(bh);
 916         gfs2_log_lock(sdp);
 917         clear_buffer_dirty(bh);
 918         bd = bh->b_private;
 919         if (bd) {
 920                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 921                         list_del_init(&bd->bd_le.le_list);
 922                 else
 923                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 924         }
 925         bh->b_bdev = NULL;
 926         clear_buffer_mapped(bh);
 927         clear_buffer_req(bh);
 928         clear_buffer_new(bh);
 929         gfs2_log_unlock(sdp);
 930         unlock_buffer(bh);
 931 }
 932
 933 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 934 {
 935         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 936         struct buffer_head *bh, *head;
 937         unsigned long pos = 0;
 938
 939         BUG_ON(!PageLocked(page));
 940         if (offset == 0)
 941                 ClearPageChecked(page);
 942         if (!page_has_buffers(page))
 943                 goto out;
 944
 945         bh = head = page_buffers(page);
 946         do {
 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 (offset == 0)
 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 gfs2_inode *ip = GFS2_I(inode);
 990         struct gfs2_holder gh;
 991         int rv;
 992
 993         /*
 994          * Deferred lock, even if its a write, since we do no allocation
 995          * on this path. All we need change is atime, and this lock mode
 996          * ensures that other nodes have flushed their buffered read caches
 997          * (i.e. their page cache entries for this inode). We do not,
 998          * unfortunately have the option of only flushing a range like
 999          * the VFS does.
1000          */
1001         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1002         rv = gfs2_glock_nq(&gh);
1003         if (rv)
1004                 return rv;
1005         rv = gfs2_ok_for_dio(ip, rw, offset);
1006         if (rv != 1)
1007                 goto out; /* dio not valid, fall back to buffered i/o */
1008
1009         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1010                                            iov, offset, nr_segs,
1011                                            gfs2_get_block_direct, NULL);
1012 out:
1013         gfs2_glock_dq_m(1, &gh);
1014         gfs2_holder_uninit(&gh);
1015         return rv;
1016 }
1017
1018 /**
1019  * gfs2_releasepage - free the metadata associated with a page
1020  * @page: the page that's being released
1021  * @gfp_mask: passed from Linux VFS, ignored by us
1022  *
1023  * Call try_to_free_buffers() if the buffers in this page can be
1024  * released.
1025  *
1026  * Returns: 0
1027  */
1028
1029 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1030 {
1031         struct inode *aspace = page->mapping->host;
1032         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
1033         struct buffer_head *bh, *head;
1034         struct gfs2_bufdata *bd;
1035
1036         if (!page_has_buffers(page))
1037                 return 0;
1038
1039         gfs2_log_lock(sdp);
1040         head = bh = page_buffers(page);
1041         do {
1042                 if (atomic_read(&bh->b_count))
1043                         goto cannot_release;
1044                 bd = bh->b_private;
1045                 if (bd && bd->bd_ail)
1046                         goto cannot_release;
1047                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1048                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1049                 bh = bh->b_this_page;
1050         } while(bh != head);
1051         gfs2_log_unlock(sdp);
1052
1053         head = bh = page_buffers(page);
1054         do {
1055                 gfs2_log_lock(sdp);
1056                 bd = bh->b_private;
1057                 if (bd) {
1058                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1059                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1060                         if (!list_empty(&bd->bd_le.le_list)) {
1061                                 if (!buffer_pinned(bh))
1062                                         list_del_init(&bd->bd_le.le_list);
1063                                 else
1064                                         bd = NULL;
1065                         }
1066                         if (bd)
1067                                 bd->bd_bh = NULL;
1068                         bh->b_private = NULL;
1069                 }
1070                 gfs2_log_unlock(sdp);
1071                 if (bd)
1072                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1073
1074                 bh = bh->b_this_page;
1075         } while (bh != head);
1076
1077         return try_to_free_buffers(page);
1078 cannot_release:
1079         gfs2_log_unlock(sdp);
1080         return 0;
1081 }
1082
1083 static const struct address_space_operations gfs2_writeback_aops = {
1084         .writepage = gfs2_writeback_writepage,
1085         .writepages = gfs2_writeback_writepages,
1086         .readpage = gfs2_readpage,
1087         .readpages = gfs2_readpages,
1088         .sync_page = block_sync_page,
1089         .write_begin = gfs2_write_begin,
1090         .write_end = gfs2_write_end,
1091         .bmap = gfs2_bmap,
1092         .invalidatepage = gfs2_invalidatepage,
1093         .releasepage = gfs2_releasepage,
1094         .direct_IO = gfs2_direct_IO,
1095         .migratepage = buffer_migrate_page,
1096 };
1097
1098 static const struct address_space_operations gfs2_ordered_aops = {
1099         .writepage = gfs2_ordered_writepage,
1100         .readpage = gfs2_readpage,
1101         .readpages = gfs2_readpages,
1102         .sync_page = block_sync_page,
1103         .write_begin = gfs2_write_begin,
1104         .write_end = gfs2_write_end,
1105         .set_page_dirty = gfs2_set_page_dirty,
1106         .bmap = gfs2_bmap,
1107         .invalidatepage = gfs2_invalidatepage,
1108         .releasepage = gfs2_releasepage,
1109         .direct_IO = gfs2_direct_IO,
1110         .migratepage = buffer_migrate_page,
1111 };
1112
1113 static const struct address_space_operations gfs2_jdata_aops = {
1114         .writepage = gfs2_jdata_writepage,
1115         .writepages = gfs2_jdata_writepages,
1116         .readpage = gfs2_readpage,
1117         .readpages = gfs2_readpages,
1118         .sync_page = block_sync_page,
1119         .write_begin = gfs2_write_begin,
1120         .write_end = gfs2_write_end,
1121         .set_page_dirty = gfs2_set_page_dirty,
1122         .bmap = gfs2_bmap,
1123         .invalidatepage = gfs2_invalidatepage,
1124         .releasepage = gfs2_releasepage,
1125 };
1126
1127 void gfs2_set_aops(struct inode *inode)
1128 {
1129         struct gfs2_inode *ip = GFS2_I(inode);
1130
1131         if (gfs2_is_writeback(ip))
1132                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1133         else if (gfs2_is_ordered(ip))
1134                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1135         else if (gfs2_is_jdata(ip))
1136                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1137         else
1138                 BUG();
1139 }
1140