fs/gfs2/aops.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/pagevec.h>
  17 #include <linux/mpage.h>
  18 #include <linux/fs.h>
  19 #include <linux/writeback.h>
  20 #include <linux/swap.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/uio.h>
  24 #include <trace/events/writeback.h>
  25
  26 #include "gfs2.h"
  27 #include "incore.h"
  28 #include "bmap.h"
  29 #include "glock.h"
  30 #include "inode.h"
  31 #include "log.h"
  32 #include "meta_io.h"
  33 #include "quota.h"
  34 #include "trans.h"
  35 #include "rgrp.h"
  36 #include "super.h"
  37 #include "util.h"
  38 #include "glops.h"
  39
  40
  41 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  42                                    unsigned int from, unsigned int to)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  47         unsigned int start, end;
  48
  49         for (bh = head, start = 0; bh != head || !start;
  50              bh = bh->b_this_page, start = end) {
  51                 end = start + bsize;
  52                 if (end <= from || start >= to)
  53                         continue;
  54                 if (gfs2_is_jdata(ip))
  55                         set_buffer_uptodate(bh);
  56                 gfs2_trans_add_data(ip->i_gl, bh);
  57         }
  58 }
  59
  60 /**
  61  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  62  * @inode: The inode
  63  * @lblock: The block number to look up
  64  * @bh_result: The buffer head to return the result in
  65  * @create: Non-zero if we may add block to the file
  66  *
  67  * Returns: errno
  68  */
  69
  70 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  71                                   struct buffer_head *bh_result, int create)
  72 {
  73         int error;
  74
  75         error = gfs2_block_map(inode, lblock, bh_result, 0);
  76         if (error)
  77                 return error;
  78         if (!buffer_mapped(bh_result))
  79                 return -EIO;
  80         return 0;
  81 }
  82
  83 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  84                                  struct buffer_head *bh_result, int create)
  85 {
  86         return gfs2_block_map(inode, lblock, bh_result, 0);
  87 }
  88
  89 /**
  90  * gfs2_writepage_common - Common bits of writepage
  91  * @page: The page to be written
  92  * @wbc: The writeback control
  93  *
  94  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  95  */
  96
  97 static int gfs2_writepage_common(struct page *page,
  98                                  struct writeback_control *wbc)
  99 {
 100         struct inode *inode = page->mapping->host;
 101         struct gfs2_inode *ip = GFS2_I(inode);
 102         struct gfs2_sbd *sdp = GFS2_SB(inode);
 103         loff_t i_size = i_size_read(inode);
 104         pgoff_t end_index = i_size >> PAGE_SHIFT;
 105         unsigned offset;
 106
 107         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 108                 goto out;
 109         if (current->journal_info)
 110                 goto redirty;
 111         /* Is the page fully outside i_size? (truncate in progress) */
 112         offset = i_size & (PAGE_SIZE-1);
 113         if (page->index > end_index || (page->index == end_index && !offset)) {
 114                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 115                 goto out;
 116         }
 117         return 1;
 118 redirty:
 119         redirty_page_for_writepage(wbc, page);
 120 out:
 121         unlock_page(page);
 122         return 0;
 123 }
 124
 125 /**
 126  * gfs2_writepage - Write page for writeback mappings
 127  * @page: The page
 128  * @wbc: The writeback control
 129  *
 130  */
 131
 132 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 133 {
 134         int ret;
 135
 136         ret = gfs2_writepage_common(page, wbc);
 137         if (ret <= 0)
 138                 return ret;
 139
 140         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 141 }
 142
 143 /* This is the same as calling block_write_full_page, but it also
 144  * writes pages outside of i_size
 145  */
 146 int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 147                          struct writeback_control *wbc)
 148 {
 149         struct inode * const inode = page->mapping->host;
 150         loff_t i_size = i_size_read(inode);
 151         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 152         unsigned offset;
 153
 154         /*
 155          * The page straddles i_size.  It must be zeroed out on each and every
 156          * writepage invocation because it may be mmapped.  "A file is mapped
 157          * in multiples of the page size.  For a file that is not a multiple of
 158          * the  page size, the remaining memory is zeroed when mapped, and
 159          * writes to that region are not written out to the file."
 160          */
 161         offset = i_size & (PAGE_SIZE-1);
 162         if (page->index == end_index && offset)
 163                 zero_user_segment(page, offset, PAGE_SIZE);
 164
 165         return __block_write_full_page(inode, page, get_block, wbc,
 166                                        end_buffer_async_write);
 167 }
 168
 169 /**
 170  * __gfs2_jdata_writepage - The core of jdata writepage
 171  * @page: The page to write
 172  * @wbc: The writeback control
 173  *
 174  * This is shared between writepage and writepages and implements the
 175  * core of the writepage operation. If a transaction is required then
 176  * PageChecked will have been set and the transaction will have
 177  * already been started before this is called.
 178  */
 179
 180 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 181 {
 182         struct inode *inode = page->mapping->host;
 183         struct gfs2_inode *ip = GFS2_I(inode);
 184         struct gfs2_sbd *sdp = GFS2_SB(inode);
 185
 186         if (PageChecked(page)) {
 187                 ClearPageChecked(page);
 188                 if (!page_has_buffers(page)) {
 189                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 190                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 191                 }
 192                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 193         }
 194         return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
 195 }
 196
 197 /**
 198  * gfs2_jdata_writepage - Write complete page
 199  * @page: Page to write
 200  * @wbc: The writeback control
 201  *
 202  * Returns: errno
 203  *
 204  */
 205
 206 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 207 {
 208         struct inode *inode = page->mapping->host;
 209         struct gfs2_inode *ip = GFS2_I(inode);
 210         struct gfs2_sbd *sdp = GFS2_SB(inode);
 211         int ret;
 212
 213         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 214                 goto out;
 215         if (PageChecked(page) || current->journal_info)
 216                 goto out_ignore;
 217         ret = __gfs2_jdata_writepage(page, wbc);
 218         return ret;
 219
 220 out_ignore:
 221         redirty_page_for_writepage(wbc, page);
 222 out:
 223         unlock_page(page);
 224         return 0;
 225 }
 226
 227 /**
 228  * gfs2_writepages - Write a bunch of dirty pages back to disk
 229  * @mapping: The mapping to write
 230  * @wbc: Write-back control
 231  *
 232  * Used for both ordered and writeback modes.
 233  */
 234 static int gfs2_writepages(struct address_space *mapping,
 235                            struct writeback_control *wbc)
 236 {
 237         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 238 }
 239
 240 /**
 241  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 242  * @mapping: The mapping
 243  * @wbc: The writeback control
 244  * @pvec: The vector of pages
 245  * @nr_pages: The number of pages to write
 246  * @end: End position
 247  * @done_index: Page index
 248  *
 249  * Returns: non-zero if loop should terminate, zero otherwise
 250  */
 251
 252 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 253                                     struct writeback_control *wbc,
 254                                     struct pagevec *pvec,
 255                                     int nr_pages, pgoff_t end,
 256                                     pgoff_t *done_index)
 257 {
 258         struct inode *inode = mapping->host;
 259         struct gfs2_sbd *sdp = GFS2_SB(inode);
 260         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 261         int i;
 262         int ret;
 263
 264         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 265         if (ret < 0)
 266                 return ret;
 267
 268         for(i = 0; i < nr_pages; i++) {
 269                 struct page *page = pvec->pages[i];
 270
 271                 /*
 272                  * At this point, the page may be truncated or
 273                  * invalidated (changing page->mapping to NULL), or
 274                  * even swizzled back from swapper_space to tmpfs file
 275                  * mapping. However, page->index will not change
 276                  * because we have a reference on the page.
 277                  */
 278                 if (page->index > end) {
 279                         /*
 280                          * can't be range_cyclic (1st pass) because
 281                          * end == -1 in that case.
 282                          */
 283                         ret = 1;
 284                         break;
 285                 }
 286
 287                 *done_index = page->index;
 288
 289                 lock_page(page);
 290
 291                 if (unlikely(page->mapping != mapping)) {
 292 continue_unlock:
 293                         unlock_page(page);
 294                         continue;
 295                 }
 296
 297                 if (!PageDirty(page)) {
 298                         /* someone wrote it for us */
 299                         goto continue_unlock;
 300                 }
 301
 302                 if (PageWriteback(page)) {
 303                         if (wbc->sync_mode != WB_SYNC_NONE)
 304                                 wait_on_page_writeback(page);
 305                         else
 306                                 goto continue_unlock;
 307                 }
 308
 309                 BUG_ON(PageWriteback(page));
 310                 if (!clear_page_dirty_for_io(page))
 311                         goto continue_unlock;
 312
 313                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 314
 315                 ret = __gfs2_jdata_writepage(page, wbc);
 316                 if (unlikely(ret)) {
 317                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 318                                 unlock_page(page);
 319                                 ret = 0;
 320                         } else {
 321
 322                                 /*
 323                                  * done_index is set past this page,
 324                                  * so media errors will not choke
 325                                  * background writeout for the entire
 326                                  * file. This has consequences for
 327                                  * range_cyclic semantics (ie. it may
 328                                  * not be suitable for data integrity
 329                                  * writeout).
 330                                  */
 331                                 *done_index = page->index + 1;
 332                                 ret = 1;
 333                                 break;
 334                         }
 335                 }
 336
 337                 /*
 338                  * We stop writing back only if we are not doing
 339                  * integrity sync. In case of integrity sync we have to
 340                  * keep going until we have written all the pages
 341                  * we tagged for writeback prior to entering this loop.
 342                  */
 343                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 344                         ret = 1;
 345                         break;
 346                 }
 347
 348         }
 349         gfs2_trans_end(sdp);
 350         return ret;
 351 }
 352
 353 /**
 354  * gfs2_write_cache_jdata - Like write_cache_pages but different
 355  * @mapping: The mapping to write
 356  * @wbc: The writeback control
 357  *
 358  * The reason that we use our own function here is that we need to
 359  * start transactions before we grab page locks. This allows us
 360  * to get the ordering right.
 361  */
 362
 363 static int gfs2_write_cache_jdata(struct address_space *mapping,
 364                                   struct writeback_control *wbc)
 365 {
 366         int ret = 0;
 367         int done = 0;
 368         struct pagevec pvec;
 369         int nr_pages;
 370         pgoff_t uninitialized_var(writeback_index);
 371         pgoff_t index;
 372         pgoff_t end;
 373         pgoff_t done_index;
 374         int cycled;
 375         int range_whole = 0;
 376         int tag;
 377
 378         pagevec_init(&pvec, 0);
 379         if (wbc->range_cyclic) {
 380                 writeback_index = mapping->writeback_index; /* prev offset */
 381                 index = writeback_index;
 382                 if (index == 0)
 383                         cycled = 1;
 384                 else
 385                         cycled = 0;
 386                 end = -1;
 387         } else {
 388                 index = wbc->range_start >> PAGE_SHIFT;
 389                 end = wbc->range_end >> PAGE_SHIFT;
 390                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 391                         range_whole = 1;
 392                 cycled = 1; /* ignore range_cyclic tests */
 393         }
 394         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 395                 tag = PAGECACHE_TAG_TOWRITE;
 396         else
 397                 tag = PAGECACHE_TAG_DIRTY;
 398
 399 retry:
 400         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 401                 tag_pages_for_writeback(mapping, index, end);
 402         done_index = index;
 403         while (!done && (index <= end)) {
 404                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
 405                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
 406                 if (nr_pages == 0)
 407                         break;
 408
 409                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end, &done_index);
 410                 if (ret)
 411                         done = 1;
 412                 if (ret > 0)
 413                         ret = 0;
 414                 pagevec_release(&pvec);
 415                 cond_resched();
 416         }
 417
 418         if (!cycled && !done) {
 419                 /*
 420                  * range_cyclic:
 421                  * We hit the last page and there is more work to be done: wrap
 422                  * back to the start of the file
 423                  */
 424                 cycled = 1;
 425                 index = 0;
 426                 end = writeback_index - 1;
 427                 goto retry;
 428         }
 429
 430         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 431                 mapping->writeback_index = done_index;
 432
 433         return ret;
 434 }
 435
 436
 437 /**
 438  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 439  * @mapping: The mapping to write
 440  * @wbc: The writeback control
 441  *
 442  */
 443
 444 static int gfs2_jdata_writepages(struct address_space *mapping,
 445                                  struct writeback_control *wbc)
 446 {
 447         struct gfs2_inode *ip = GFS2_I(mapping->host);
 448         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 449         int ret;
 450
 451         ret = gfs2_write_cache_jdata(mapping, wbc);
 452         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 453                 gfs2_log_flush(sdp, ip->i_gl, NORMAL_FLUSH);
 454                 ret = gfs2_write_cache_jdata(mapping, wbc);
 455         }
 456         return ret;
 457 }
 458
 459 /**
 460  * stuffed_readpage - Fill in a Linux page with stuffed file data
 461  * @ip: the inode
 462  * @page: the page
 463  *
 464  * Returns: errno
 465  */
 466
 467 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 468 {
 469         struct buffer_head *dibh;
 470         u64 dsize = i_size_read(&ip->i_inode);
 471         void *kaddr;
 472         int error;
 473
 474         /*
 475          * Due to the order of unstuffing files and ->fault(), we can be
 476          * asked for a zero page in the case of a stuffed file being extended,
 477          * so we need to supply one here. It doesn't happen often.
 478          */
 479         if (unlikely(page->index)) {
 480                 zero_user(page, 0, PAGE_SIZE);
 481                 SetPageUptodate(page);
 482                 return 0;
 483         }
 484
 485         error = gfs2_meta_inode_buffer(ip, &dibh);
 486         if (error)
 487                 return error;
 488
 489         kaddr = kmap_atomic(page);
 490         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 491                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 492         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 493         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 494         kunmap_atomic(kaddr);
 495         flush_dcache_page(page);
 496         brelse(dibh);
 497         SetPageUptodate(page);
 498
 499         return 0;
 500 }
 501
 502
 503 /**
 504  * __gfs2_readpage - readpage
 505  * @file: The file to read a page for
 506  * @page: The page to read
 507  *
 508  * This is the core of gfs2's readpage. Its used by the internal file
 509  * reading code as in that case we already hold the glock. Also its
 510  * called by gfs2_readpage() once the required lock has been granted.
 511  *
 512  */
 513
 514 static int __gfs2_readpage(void *file, struct page *page)
 515 {
 516         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 517         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 518         int error;
 519
 520         if (gfs2_is_stuffed(ip)) {
 521                 error = stuffed_readpage(ip, page);
 522                 unlock_page(page);
 523         } else {
 524                 error = mpage_readpage(page, gfs2_block_map);
 525         }
 526
 527         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 528                 return -EIO;
 529
 530         return error;
 531 }
 532
 533 /**
 534  * gfs2_readpage - read a page of a file
 535  * @file: The file to read
 536  * @page: The page of the file
 537  *
 538  * This deals with the locking required. We have to unlock and
 539  * relock the page in order to get the locking in the right
 540  * order.
 541  */
 542
 543 static int gfs2_readpage(struct file *file, struct page *page)
 544 {
 545         struct address_space *mapping = page->mapping;
 546         struct gfs2_inode *ip = GFS2_I(mapping->host);
 547         struct gfs2_holder gh;
 548         int error;
 549
 550         unlock_page(page);
 551         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 552         error = gfs2_glock_nq(&gh);
 553         if (unlikely(error))
 554                 goto out;
 555         error = AOP_TRUNCATED_PAGE;
 556         lock_page(page);
 557         if (page->mapping == mapping && !PageUptodate(page))
 558                 error = __gfs2_readpage(file, page);
 559         else
 560                 unlock_page(page);
 561         gfs2_glock_dq(&gh);
 562 out:
 563         gfs2_holder_uninit(&gh);
 564         if (error && error != AOP_TRUNCATED_PAGE)
 565                 lock_page(page);
 566         return error;
 567 }
 568
 569 /**
 570  * gfs2_internal_read - read an internal file
 571  * @ip: The gfs2 inode
 572  * @buf: The buffer to fill
 573  * @pos: The file position
 574  * @size: The amount to read
 575  *
 576  */
 577
 578 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 579                        unsigned size)
 580 {
 581         struct address_space *mapping = ip->i_inode.i_mapping;
 582         unsigned long index = *pos / PAGE_SIZE;
 583         unsigned offset = *pos & (PAGE_SIZE - 1);
 584         unsigned copied = 0;
 585         unsigned amt;
 586         struct page *page;
 587         void *p;
 588
 589         do {
 590                 amt = size - copied;
 591                 if (offset + size > PAGE_SIZE)
 592                         amt = PAGE_SIZE - offset;
 593                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 594                 if (IS_ERR(page))
 595                         return PTR_ERR(page);
 596                 p = kmap_atomic(page);
 597                 memcpy(buf + copied, p + offset, amt);
 598                 kunmap_atomic(p);
 599                 put_page(page);
 600                 copied += amt;
 601                 index++;
 602                 offset = 0;
 603         } while(copied < size);
 604         (*pos) += size;
 605         return size;
 606 }
 607
 608 /**
 609  * gfs2_readpages - Read a bunch of pages at once
 610  * @file: The file to read from
 611  * @mapping: Address space info
 612  * @pages: List of pages to read
 613  * @nr_pages: Number of pages to read
 614  *
 615  * Some notes:
 616  * 1. This is only for readahead, so we can simply ignore any things
 617  *    which are slightly inconvenient (such as locking conflicts between
 618  *    the page lock and the glock) and return having done no I/O. Its
 619  *    obviously not something we'd want to do on too regular a basis.
 620  *    Any I/O we ignore at this time will be done via readpage later.
 621  * 2. We don't handle stuffed files here we let readpage do the honours.
 622  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 623  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 624  */
 625
 626 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 627                           struct list_head *pages, unsigned nr_pages)
 628 {
 629         struct inode *inode = mapping->host;
 630         struct gfs2_inode *ip = GFS2_I(inode);
 631         struct gfs2_sbd *sdp = GFS2_SB(inode);
 632         struct gfs2_holder gh;
 633         int ret;
 634
 635         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 636         ret = gfs2_glock_nq(&gh);
 637         if (unlikely(ret))
 638                 goto out_uninit;
 639         if (!gfs2_is_stuffed(ip))
 640                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 641         gfs2_glock_dq(&gh);
 642 out_uninit:
 643         gfs2_holder_uninit(&gh);
 644         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 645                 ret = -EIO;
 646         return ret;
 647 }
 648
 649 /**
 650  * gfs2_write_begin - Begin to write to a file
 651  * @file: The file to write to
 652  * @mapping: The mapping in which to write
 653  * @pos: The file offset at which to start writing
 654  * @len: Length of the write
 655  * @flags: Various flags
 656  * @pagep: Pointer to return the page
 657  * @fsdata: Pointer to return fs data (unused by GFS2)
 658  *
 659  * Returns: errno
 660  */
 661
 662 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 663                             loff_t pos, unsigned len, unsigned flags,
 664                             struct page **pagep, void **fsdata)
 665 {
 666         struct gfs2_inode *ip = GFS2_I(mapping->host);
 667         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 668         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 669         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 670         unsigned requested = 0;
 671         int alloc_required;
 672         int error = 0;
 673         pgoff_t index = pos >> PAGE_SHIFT;
 674         unsigned from = pos & (PAGE_SIZE - 1);
 675         struct page *page;
 676
 677         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 678         error = gfs2_glock_nq(&ip->i_gh);
 679         if (unlikely(error))
 680                 goto out_uninit;
 681         if (&ip->i_inode == sdp->sd_rindex) {
 682                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 683                                            GL_NOCACHE, &m_ip->i_gh);
 684                 if (unlikely(error)) {
 685                         gfs2_glock_dq(&ip->i_gh);
 686                         goto out_uninit;
 687                 }
 688         }
 689
 690         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 691
 692         if (alloc_required || gfs2_is_jdata(ip))
 693                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 694
 695         if (alloc_required) {
 696                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 697                 requested = data_blocks + ind_blocks;
 698                 ap.target = requested;
 699                 error = gfs2_quota_lock_check(ip, &ap);
 700                 if (error)
 701                         goto out_unlock;
 702
 703                 error = gfs2_inplace_reserve(ip, &ap);
 704                 if (error)
 705                         goto out_qunlock;
 706         }
 707
 708         rblocks = RES_DINODE + ind_blocks;
 709         if (gfs2_is_jdata(ip))
 710                 rblocks += data_blocks ? data_blocks : 1;
 711         if (ind_blocks || data_blocks)
 712                 rblocks += RES_STATFS + RES_QUOTA;
 713         if (&ip->i_inode == sdp->sd_rindex)
 714                 rblocks += 2 * RES_STATFS;
 715         if (alloc_required)
 716                 rblocks += gfs2_rg_blocks(ip, requested);
 717
 718         error = gfs2_trans_begin(sdp, rblocks,
 719                                  PAGE_SIZE/sdp->sd_sb.sb_bsize);
 720         if (error)
 721                 goto out_trans_fail;
 722
 723         error = -ENOMEM;
 724         flags |= AOP_FLAG_NOFS;
 725         page = grab_cache_page_write_begin(mapping, index, flags);
 726         *pagep = page;
 727         if (unlikely(!page))
 728                 goto out_endtrans;
 729
 730         if (gfs2_is_stuffed(ip)) {
 731                 error = 0;
 732                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 733                         error = gfs2_unstuff_dinode(ip, page);
 734                         if (error == 0)
 735                                 goto prepare_write;
 736                 } else if (!PageUptodate(page)) {
 737                         error = stuffed_readpage(ip, page);
 738                 }
 739                 goto out;
 740         }
 741
 742 prepare_write:
 743         error = __block_write_begin(page, from, len, gfs2_block_map);
 744 out:
 745         if (error == 0)
 746                 return 0;
 747
 748         unlock_page(page);
 749         put_page(page);
 750
 751         gfs2_trans_end(sdp);
 752         if (pos + len > ip->i_inode.i_size)
 753                 gfs2_trim_blocks(&ip->i_inode);
 754         goto out_trans_fail;
 755
 756 out_endtrans:
 757         gfs2_trans_end(sdp);
 758 out_trans_fail:
 759         if (alloc_required) {
 760                 gfs2_inplace_release(ip);
 761 out_qunlock:
 762                 gfs2_quota_unlock(ip);
 763         }
 764 out_unlock:
 765         if (&ip->i_inode == sdp->sd_rindex) {
 766                 gfs2_glock_dq(&m_ip->i_gh);
 767                 gfs2_holder_uninit(&m_ip->i_gh);
 768         }
 769         gfs2_glock_dq(&ip->i_gh);
 770 out_uninit:
 771         gfs2_holder_uninit(&ip->i_gh);
 772         return error;
 773 }
 774
 775 /**
 776  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 777  * @inode: the rindex inode
 778  */
 779 static void adjust_fs_space(struct inode *inode)
 780 {
 781         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 782         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 783         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 784         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 785         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 786         struct buffer_head *m_bh, *l_bh;
 787         u64 fs_total, new_free;
 788
 789         /* Total up the file system space, according to the latest rindex. */
 790         fs_total = gfs2_ri_total(sdp);
 791         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 792                 return;
 793
 794         spin_lock(&sdp->sd_statfs_spin);
 795         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 796                               sizeof(struct gfs2_dinode));
 797         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 798                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 799         else
 800                 new_free = 0;
 801         spin_unlock(&sdp->sd_statfs_spin);
 802         fs_warn(sdp, "File system extended by %llu blocks.\n",
 803                 (unsigned long long)new_free);
 804         gfs2_statfs_change(sdp, new_free, new_free, 0);
 805
 806         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 807                 goto out;
 808         update_statfs(sdp, m_bh, l_bh);
 809         brelse(l_bh);
 810 out:
 811         brelse(m_bh);
 812 }
 813
 814 /**
 815  * gfs2_stuffed_write_end - Write end for stuffed files
 816  * @inode: The inode
 817  * @dibh: The buffer_head containing the on-disk inode
 818  * @pos: The file position
 819  * @len: The length of the write
 820  * @copied: How much was actually copied by the VFS
 821  * @page: The page
 822  *
 823  * This copies the data from the page into the inode block after
 824  * the inode data structure itself.
 825  *
 826  * Returns: errno
 827  */
 828 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 829                                   loff_t pos, unsigned len, unsigned copied,
 830                                   struct page *page)
 831 {
 832         struct gfs2_inode *ip = GFS2_I(inode);
 833         struct gfs2_sbd *sdp = GFS2_SB(inode);
 834         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 835         u64 to = pos + copied;
 836         void *kaddr;
 837         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 838
 839         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 840         kaddr = kmap_atomic(page);
 841         memcpy(buf + pos, kaddr + pos, copied);
 842         memset(kaddr + pos + copied, 0, len - copied);
 843         flush_dcache_page(page);
 844         kunmap_atomic(kaddr);
 845
 846         if (!PageUptodate(page))
 847                 SetPageUptodate(page);
 848         unlock_page(page);
 849         put_page(page);
 850
 851         if (copied) {
 852                 if (inode->i_size < to)
 853                         i_size_write(inode, to);
 854                 mark_inode_dirty(inode);
 855         }
 856
 857         if (inode == sdp->sd_rindex) {
 858                 adjust_fs_space(inode);
 859                 sdp->sd_rindex_uptodate = 0;
 860         }
 861
 862         brelse(dibh);
 863         gfs2_trans_end(sdp);
 864         if (inode == sdp->sd_rindex) {
 865                 gfs2_glock_dq(&m_ip->i_gh);
 866                 gfs2_holder_uninit(&m_ip->i_gh);
 867         }
 868         gfs2_glock_dq(&ip->i_gh);
 869         gfs2_holder_uninit(&ip->i_gh);
 870         return copied;
 871 }
 872
 873 /**
 874  * gfs2_write_end
 875  * @file: The file to write to
 876  * @mapping: The address space to write to
 877  * @pos: The file position
 878  * @len: The length of the data
 879  * @copied: How much was actually copied by the VFS
 880  * @page: The page that has been written
 881  * @fsdata: The fsdata (unused in GFS2)
 882  *
 883  * The main write_end function for GFS2. We have a separate one for
 884  * stuffed files as they are slightly different, otherwise we just
 885  * put our locking around the VFS provided functions.
 886  *
 887  * Returns: errno
 888  */
 889
 890 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 891                           loff_t pos, unsigned len, unsigned copied,
 892                           struct page *page, void *fsdata)
 893 {
 894         struct inode *inode = page->mapping->host;
 895         struct gfs2_inode *ip = GFS2_I(inode);
 896         struct gfs2_sbd *sdp = GFS2_SB(inode);
 897         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 898         struct buffer_head *dibh;
 899         unsigned int from = pos & (PAGE_SIZE - 1);
 900         unsigned int to = from + len;
 901         int ret;
 902         struct gfs2_trans *tr = current->journal_info;
 903         BUG_ON(!tr);
 904
 905         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 906
 907         ret = gfs2_meta_inode_buffer(ip, &dibh);
 908         if (unlikely(ret)) {
 909                 unlock_page(page);
 910                 put_page(page);
 911                 goto failed;
 912         }
 913
 914         if (gfs2_is_stuffed(ip))
 915                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 916
 917         if (!gfs2_is_writeback(ip))
 918                 gfs2_page_add_databufs(ip, page, from, to);
 919
 920         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 921         if (tr->tr_num_buf_new)
 922                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 923         else
 924                 gfs2_trans_add_meta(ip->i_gl, dibh);
 925
 926
 927         if (inode == sdp->sd_rindex) {
 928                 adjust_fs_space(inode);
 929                 sdp->sd_rindex_uptodate = 0;
 930         }
 931
 932         brelse(dibh);
 933 failed:
 934         gfs2_trans_end(sdp);
 935         gfs2_inplace_release(ip);
 936         if (ip->i_qadata && ip->i_qadata->qa_qd_num)
 937                 gfs2_quota_unlock(ip);
 938         if (inode == sdp->sd_rindex) {
 939                 gfs2_glock_dq(&m_ip->i_gh);
 940                 gfs2_holder_uninit(&m_ip->i_gh);
 941         }
 942         gfs2_glock_dq(&ip->i_gh);
 943         gfs2_holder_uninit(&ip->i_gh);
 944         return ret;
 945 }
 946
 947 /**
 948  * gfs2_set_page_dirty - Page dirtying function
 949  * @page: The page to dirty
 950  *
 951  * Returns: 1 if it dirtyed the page, or 0 otherwise
 952  */
 953
 954 static int gfs2_set_page_dirty(struct page *page)
 955 {
 956         SetPageChecked(page);
 957         return __set_page_dirty_buffers(page);
 958 }
 959
 960 /**
 961  * gfs2_bmap - Block map function
 962  * @mapping: Address space info
 963  * @lblock: The block to map
 964  *
 965  * Returns: The disk address for the block or 0 on hole or error
 966  */
 967
 968 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 969 {
 970         struct gfs2_inode *ip = GFS2_I(mapping->host);
 971         struct gfs2_holder i_gh;
 972         sector_t dblock = 0;
 973         int error;
 974
 975         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 976         if (error)
 977                 return 0;
 978
 979         if (!gfs2_is_stuffed(ip))
 980                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 981
 982         gfs2_glock_dq_uninit(&i_gh);
 983
 984         return dblock;
 985 }
 986
 987 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 988 {
 989         struct gfs2_bufdata *bd;
 990
 991         lock_buffer(bh);
 992         gfs2_log_lock(sdp);
 993         clear_buffer_dirty(bh);
 994         bd = bh->b_private;
 995         if (bd) {
 996                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 997                         list_del_init(&bd->bd_list);
 998                 else
 999                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
1000         }
1001         bh->b_bdev = NULL;
1002         clear_buffer_mapped(bh);
1003         clear_buffer_req(bh);
1004         clear_buffer_new(bh);
1005         gfs2_log_unlock(sdp);
1006         unlock_buffer(bh);
1007 }
1008
1009 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
1010                                 unsigned int length)
1011 {
1012         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
1013         unsigned int stop = offset + length;
1014         int partial_page = (offset || length < PAGE_SIZE);
1015         struct buffer_head *bh, *head;
1016         unsigned long pos = 0;
1017
1018         BUG_ON(!PageLocked(page));
1019         if (!partial_page)
1020                 ClearPageChecked(page);
1021         if (!page_has_buffers(page))
1022                 goto out;
1023
1024         bh = head = page_buffers(page);
1025         do {
1026                 if (pos + bh->b_size > stop)
1027                         return;
1028
1029                 if (offset <= pos)
1030                         gfs2_discard(sdp, bh);
1031                 pos += bh->b_size;
1032                 bh = bh->b_this_page;
1033         } while (bh != head);
1034 out:
1035         if (!partial_page)
1036                 try_to_release_page(page, 0);
1037 }
1038
1039 /**
1040  * gfs2_ok_for_dio - check that dio is valid on this file
1041  * @ip: The inode
1042  * @offset: The offset at which we are reading or writing
1043  *
1044  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1045  *          1 (to accept the i/o request)
1046  */
1047 static int gfs2_ok_for_dio(struct gfs2_inode *ip, loff_t offset)
1048 {
1049         /*
1050          * Should we return an error here? I can't see that O_DIRECT for
1051          * a stuffed file makes any sense. For now we'll silently fall
1052          * back to buffered I/O
1053          */
1054         if (gfs2_is_stuffed(ip))
1055                 return 0;
1056
1057         if (offset >= i_size_read(&ip->i_inode))
1058                 return 0;
1059         return 1;
1060 }
1061
1062
1063
1064 static ssize_t gfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
1065 {
1066         struct file *file = iocb->ki_filp;
1067         struct inode *inode = file->f_mapping->host;
1068         struct address_space *mapping = inode->i_mapping;
1069         struct gfs2_inode *ip = GFS2_I(inode);
1070         loff_t offset = iocb->ki_pos;
1071         struct gfs2_holder gh;
1072         int rv;
1073
1074         /*
1075          * Deferred lock, even if its a write, since we do no allocation
1076          * on this path. All we need change is atime, and this lock mode
1077          * ensures that other nodes have flushed their buffered read caches
1078          * (i.e. their page cache entries for this inode). We do not,
1079          * unfortunately have the option of only flushing a range like
1080          * the VFS does.
1081          */
1082         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1083         rv = gfs2_glock_nq(&gh);
1084         if (rv)
1085                 goto out_uninit;
1086         rv = gfs2_ok_for_dio(ip, offset);
1087         if (rv != 1)
1088                 goto out; /* dio not valid, fall back to buffered i/o */
1089
1090         /*
1091          * Now since we are holding a deferred (CW) lock at this point, you
1092          * might be wondering why this is ever needed. There is a case however
1093          * where we've granted a deferred local lock against a cached exclusive
1094          * glock. That is ok provided all granted local locks are deferred, but
1095          * it also means that it is possible to encounter pages which are
1096          * cached and possibly also mapped. So here we check for that and sort
1097          * them out ahead of the dio. The glock state machine will take care of
1098          * everything else.
1099          *
1100          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1101          * the first place, mapping->nr_pages will always be zero.
1102          */
1103         if (mapping->nrpages) {
1104                 loff_t lstart = offset & ~(PAGE_SIZE - 1);
1105                 loff_t len = iov_iter_count(iter);
1106                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1107
1108                 rv = 0;
1109                 if (len == 0)
1110                         goto out;
1111                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1112                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1113                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1114                 if (rv)
1115                         goto out;
1116                 if (iov_iter_rw(iter) == WRITE)
1117                         truncate_inode_pages_range(mapping, lstart, end);
1118         }
1119
1120         rv = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
1121                                   gfs2_get_block_direct, NULL, NULL, 0);
1122 out:
1123         gfs2_glock_dq(&gh);
1124 out_uninit:
1125         gfs2_holder_uninit(&gh);
1126         return rv;
1127 }
1128
1129 /**
1130  * gfs2_releasepage - free the metadata associated with a page
1131  * @page: the page that's being released
1132  * @gfp_mask: passed from Linux VFS, ignored by us
1133  *
1134  * Call try_to_free_buffers() if the buffers in this page can be
1135  * released.
1136  *
1137  * Returns: 0
1138  */
1139
1140 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1141 {
1142         struct address_space *mapping = page->mapping;
1143         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1144         struct buffer_head *bh, *head;
1145         struct gfs2_bufdata *bd;
1146
1147         if (!page_has_buffers(page))
1148                 return 0;
1149
1150         /*
1151          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
1152          * clean pages might not have had the dirty bit cleared.  Thus, it can
1153          * send actual dirty pages to ->releasepage() via shrink_active_list().
1154          *
1155          * As a workaround, we skip pages that contain dirty buffers below.
1156          * Once ->releasepage isn't called on dirty pages anymore, we can warn
1157          * on dirty buffers like we used to here again.
1158          */
1159
1160         gfs2_log_lock(sdp);
1161         spin_lock(&sdp->sd_ail_lock);
1162         head = bh = page_buffers(page);
1163         do {
1164                 if (atomic_read(&bh->b_count))
1165                         goto cannot_release;
1166                 bd = bh->b_private;
1167                 if (bd && bd->bd_tr)
1168                         goto cannot_release;
1169                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
1170                         goto cannot_release;
1171                 bh = bh->b_this_page;
1172         } while(bh != head);
1173         spin_unlock(&sdp->sd_ail_lock);
1174
1175         head = bh = page_buffers(page);
1176         do {
1177                 bd = bh->b_private;
1178                 if (bd) {
1179                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1180                         if (!list_empty(&bd->bd_list))
1181                                 list_del_init(&bd->bd_list);
1182                         bd->bd_bh = NULL;
1183                         bh->b_private = NULL;
1184                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1185                 }
1186
1187                 bh = bh->b_this_page;
1188         } while (bh != head);
1189         gfs2_log_unlock(sdp);
1190
1191         return try_to_free_buffers(page);
1192
1193 cannot_release:
1194         spin_unlock(&sdp->sd_ail_lock);
1195         gfs2_log_unlock(sdp);
1196         return 0;
1197 }
1198
1199 static const struct address_space_operations gfs2_writeback_aops = {
1200         .writepage = gfs2_writepage,
1201         .writepages = gfs2_writepages,
1202         .readpage = gfs2_readpage,
1203         .readpages = gfs2_readpages,
1204         .write_begin = gfs2_write_begin,
1205         .write_end = gfs2_write_end,
1206         .bmap = gfs2_bmap,
1207         .invalidatepage = gfs2_invalidatepage,
1208         .releasepage = gfs2_releasepage,
1209         .direct_IO = gfs2_direct_IO,
1210         .migratepage = buffer_migrate_page,
1211         .is_partially_uptodate = block_is_partially_uptodate,
1212         .error_remove_page = generic_error_remove_page,
1213 };
1214
1215 static const struct address_space_operations gfs2_ordered_aops = {
1216         .writepage = gfs2_writepage,
1217         .writepages = gfs2_writepages,
1218         .readpage = gfs2_readpage,
1219         .readpages = gfs2_readpages,
1220         .write_begin = gfs2_write_begin,
1221         .write_end = gfs2_write_end,
1222         .set_page_dirty = gfs2_set_page_dirty,
1223         .bmap = gfs2_bmap,
1224         .invalidatepage = gfs2_invalidatepage,
1225         .releasepage = gfs2_releasepage,
1226         .direct_IO = gfs2_direct_IO,
1227         .migratepage = buffer_migrate_page,
1228         .is_partially_uptodate = block_is_partially_uptodate,
1229         .error_remove_page = generic_error_remove_page,
1230 };
1231
1232 static const struct address_space_operations gfs2_jdata_aops = {
1233         .writepage = gfs2_jdata_writepage,
1234         .writepages = gfs2_jdata_writepages,
1235         .readpage = gfs2_readpage,
1236         .readpages = gfs2_readpages,
1237         .write_begin = gfs2_write_begin,
1238         .write_end = gfs2_write_end,
1239         .set_page_dirty = gfs2_set_page_dirty,
1240         .bmap = gfs2_bmap,
1241         .invalidatepage = gfs2_invalidatepage,
1242         .releasepage = gfs2_releasepage,
1243         .is_partially_uptodate = block_is_partially_uptodate,
1244         .error_remove_page = generic_error_remove_page,
1245 };
1246
1247 void gfs2_set_aops(struct inode *inode)
1248 {
1249         struct gfs2_inode *ip = GFS2_I(inode);
1250
1251         if (gfs2_is_writeback(ip))
1252                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1253         else if (gfs2_is_ordered(ip))
1254                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1255         else if (gfs2_is_jdata(ip))
1256                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1257         else
1258                 BUG();
1259 }
1260