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 len)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  47         unsigned int to = from + len;
  48         unsigned int start, end;
  49
  50         for (bh = head, start = 0; bh != head || !start;
  51              bh = bh->b_this_page, start = end) {
  52                 end = start + bsize;
  53                 if (end <= from)
  54                         continue;
  55                 if (start >= to)
  56                         break;
  57                 set_buffer_uptodate(bh);
  58                 gfs2_trans_add_data(ip->i_gl, bh);
  59         }
  60 }
  61
  62 /**
  63  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  64  * @inode: The inode
  65  * @lblock: The block number to look up
  66  * @bh_result: The buffer head to return the result in
  67  * @create: Non-zero if we may add block to the file
  68  *
  69  * Returns: errno
  70  */
  71
  72 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  73                                   struct buffer_head *bh_result, int create)
  74 {
  75         int error;
  76
  77         error = gfs2_block_map(inode, lblock, bh_result, 0);
  78         if (error)
  79                 return error;
  80         if (!buffer_mapped(bh_result))
  81                 return -EIO;
  82         return 0;
  83 }
  84
  85 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  86                                  struct buffer_head *bh_result, int create)
  87 {
  88         return gfs2_block_map(inode, lblock, bh_result, 0);
  89 }
  90
  91 /**
  92  * gfs2_writepage_common - Common bits of writepage
  93  * @page: The page to be written
  94  * @wbc: The writeback control
  95  *
  96  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  97  */
  98
  99 static int gfs2_writepage_common(struct page *page,
 100                                  struct writeback_control *wbc)
 101 {
 102         struct inode *inode = page->mapping->host;
 103         struct gfs2_inode *ip = GFS2_I(inode);
 104         struct gfs2_sbd *sdp = GFS2_SB(inode);
 105         loff_t i_size = i_size_read(inode);
 106         pgoff_t end_index = i_size >> PAGE_SHIFT;
 107         unsigned offset;
 108
 109         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 110                 goto out;
 111         if (current->journal_info)
 112                 goto redirty;
 113         /* Is the page fully outside i_size? (truncate in progress) */
 114         offset = i_size & (PAGE_SIZE-1);
 115         if (page->index > end_index || (page->index == end_index && !offset)) {
 116                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 117                 goto out;
 118         }
 119         return 1;
 120 redirty:
 121         redirty_page_for_writepage(wbc, page);
 122 out:
 123         unlock_page(page);
 124         return 0;
 125 }
 126
 127 /**
 128  * gfs2_writepage - Write page for writeback mappings
 129  * @page: The page
 130  * @wbc: The writeback control
 131  *
 132  */
 133
 134 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 135 {
 136         int ret;
 137
 138         ret = gfs2_writepage_common(page, wbc);
 139         if (ret <= 0)
 140                 return ret;
 141
 142         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 143 }
 144
 145 /* This is the same as calling block_write_full_page, but it also
 146  * writes pages outside of i_size
 147  */
 148 static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 149                                 struct writeback_control *wbc)
 150 {
 151         struct inode * const inode = page->mapping->host;
 152         loff_t i_size = i_size_read(inode);
 153         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 154         unsigned offset;
 155
 156         /*
 157          * The page straddles i_size.  It must be zeroed out on each and every
 158          * writepage invocation because it may be mmapped.  "A file is mapped
 159          * in multiples of the page size.  For a file that is not a multiple of
 160          * the  page size, the remaining memory is zeroed when mapped, and
 161          * writes to that region are not written out to the file."
 162          */
 163         offset = i_size & (PAGE_SIZE-1);
 164         if (page->index == end_index && offset)
 165                 zero_user_segment(page, offset, PAGE_SIZE);
 166
 167         return __block_write_full_page(inode, page, get_block, wbc,
 168                                        end_buffer_async_write);
 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                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 193                 }
 194                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize);
 195         }
 196         return gfs2_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  * @wbc: The writeback control
 203  *
 204  * Returns: errno
 205  *
 206  */
 207
 208 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 209 {
 210         struct inode *inode = page->mapping->host;
 211         struct gfs2_inode *ip = GFS2_I(inode);
 212         struct gfs2_sbd *sdp = GFS2_SB(inode);
 213         int ret;
 214
 215         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 216                 goto out;
 217         if (PageChecked(page) || current->journal_info)
 218                 goto out_ignore;
 219         ret = __gfs2_jdata_writepage(page, wbc);
 220         return ret;
 221
 222 out_ignore:
 223         redirty_page_for_writepage(wbc, page);
 224 out:
 225         unlock_page(page);
 226         return 0;
 227 }
 228
 229 /**
 230  * gfs2_writepages - Write a bunch of dirty pages back to disk
 231  * @mapping: The mapping to write
 232  * @wbc: Write-back control
 233  *
 234  * Used for both ordered and writeback modes.
 235  */
 236 static int gfs2_writepages(struct address_space *mapping,
 237                            struct writeback_control *wbc)
 238 {
 239         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 240         int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 241
 242         /*
 243          * Even if we didn't write any pages here, we might still be holding
 244          * dirty pages in the ail. We forcibly flush the ail because we don't
 245          * want balance_dirty_pages() to loop indefinitely trying to write out
 246          * pages held in the ail that it can't find.
 247          */
 248         if (ret == 0)
 249                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 250
 251         return ret;
 252 }
 253
 254 /**
 255  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 256  * @mapping: The mapping
 257  * @wbc: The writeback control
 258  * @pvec: The vector of pages
 259  * @nr_pages: The number of pages to write
 260  * @done_index: Page index
 261  *
 262  * Returns: non-zero if loop should terminate, zero otherwise
 263  */
 264
 265 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 266                                     struct writeback_control *wbc,
 267                                     struct pagevec *pvec,
 268                                     int nr_pages,
 269                                     pgoff_t *done_index)
 270 {
 271         struct inode *inode = mapping->host;
 272         struct gfs2_sbd *sdp = GFS2_SB(inode);
 273         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 274         int i;
 275         int ret;
 276
 277         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 278         if (ret < 0)
 279                 return ret;
 280
 281         for(i = 0; i < nr_pages; i++) {
 282                 struct page *page = pvec->pages[i];
 283
 284                 *done_index = page->index;
 285
 286                 lock_page(page);
 287
 288                 if (unlikely(page->mapping != mapping)) {
 289 continue_unlock:
 290                         unlock_page(page);
 291                         continue;
 292                 }
 293
 294                 if (!PageDirty(page)) {
 295                         /* someone wrote it for us */
 296                         goto continue_unlock;
 297                 }
 298
 299                 if (PageWriteback(page)) {
 300                         if (wbc->sync_mode != WB_SYNC_NONE)
 301                                 wait_on_page_writeback(page);
 302                         else
 303                                 goto continue_unlock;
 304                 }
 305
 306                 BUG_ON(PageWriteback(page));
 307                 if (!clear_page_dirty_for_io(page))
 308                         goto continue_unlock;
 309
 310                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 311
 312                 ret = __gfs2_jdata_writepage(page, wbc);
 313                 if (unlikely(ret)) {
 314                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 315                                 unlock_page(page);
 316                                 ret = 0;
 317                         } else {
 318
 319                                 /*
 320                                  * done_index is set past this page,
 321                                  * so media errors will not choke
 322                                  * background writeout for the entire
 323                                  * file. This has consequences for
 324                                  * range_cyclic semantics (ie. it may
 325                                  * not be suitable for data integrity
 326                                  * writeout).
 327                                  */
 328                                 *done_index = page->index + 1;
 329                                 ret = 1;
 330                                 break;
 331                         }
 332                 }
 333
 334                 /*
 335                  * We stop writing back only if we are not doing
 336                  * integrity sync. In case of integrity sync we have to
 337                  * keep going until we have written all the pages
 338                  * we tagged for writeback prior to entering this loop.
 339                  */
 340                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 341                         ret = 1;
 342                         break;
 343                 }
 344
 345         }
 346         gfs2_trans_end(sdp);
 347         return ret;
 348 }
 349
 350 /**
 351  * gfs2_write_cache_jdata - Like write_cache_pages but different
 352  * @mapping: The mapping to write
 353  * @wbc: The writeback control
 354  *
 355  * The reason that we use our own function here is that we need to
 356  * start transactions before we grab page locks. This allows us
 357  * to get the ordering right.
 358  */
 359
 360 static int gfs2_write_cache_jdata(struct address_space *mapping,
 361                                   struct writeback_control *wbc)
 362 {
 363         int ret = 0;
 364         int done = 0;
 365         struct pagevec pvec;
 366         int nr_pages;
 367         pgoff_t uninitialized_var(writeback_index);
 368         pgoff_t index;
 369         pgoff_t end;
 370         pgoff_t done_index;
 371         int cycled;
 372         int range_whole = 0;
 373         int tag;
 374
 375         pagevec_init(&pvec);
 376         if (wbc->range_cyclic) {
 377                 writeback_index = mapping->writeback_index; /* prev offset */
 378                 index = writeback_index;
 379                 if (index == 0)
 380                         cycled = 1;
 381                 else
 382                         cycled = 0;
 383                 end = -1;
 384         } else {
 385                 index = wbc->range_start >> PAGE_SHIFT;
 386                 end = wbc->range_end >> PAGE_SHIFT;
 387                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 388                         range_whole = 1;
 389                 cycled = 1; /* ignore range_cyclic tests */
 390         }
 391         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 392                 tag = PAGECACHE_TAG_TOWRITE;
 393         else
 394                 tag = PAGECACHE_TAG_DIRTY;
 395
 396 retry:
 397         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 398                 tag_pages_for_writeback(mapping, index, end);
 399         done_index = index;
 400         while (!done && (index <= end)) {
 401                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
 402                                 tag);
 403                 if (nr_pages == 0)
 404                         break;
 405
 406                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
 407                 if (ret)
 408                         done = 1;
 409                 if (ret > 0)
 410                         ret = 0;
 411                 pagevec_release(&pvec);
 412                 cond_resched();
 413         }
 414
 415         if (!cycled && !done) {
 416                 /*
 417                  * range_cyclic:
 418                  * We hit the last page and there is more work to be done: wrap
 419                  * back to the start of the file
 420                  */
 421                 cycled = 1;
 422                 index = 0;
 423                 end = writeback_index - 1;
 424                 goto retry;
 425         }
 426
 427         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 428                 mapping->writeback_index = done_index;
 429
 430         return ret;
 431 }
 432
 433
 434 /**
 435  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 436  * @mapping: The mapping to write
 437  * @wbc: The writeback control
 438  *
 439  */
 440
 441 static int gfs2_jdata_writepages(struct address_space *mapping,
 442                                  struct writeback_control *wbc)
 443 {
 444         struct gfs2_inode *ip = GFS2_I(mapping->host);
 445         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 446         int ret;
 447
 448         ret = gfs2_write_cache_jdata(mapping, wbc);
 449         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 450                 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
 451                                GFS2_LFC_JDATA_WPAGES);
 452                 ret = gfs2_write_cache_jdata(mapping, wbc);
 453         }
 454         return ret;
 455 }
 456
 457 /**
 458  * stuffed_readpage - Fill in a Linux page with stuffed file data
 459  * @ip: the inode
 460  * @page: the page
 461  *
 462  * Returns: errno
 463  */
 464
 465 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 466 {
 467         struct buffer_head *dibh;
 468         u64 dsize = i_size_read(&ip->i_inode);
 469         void *kaddr;
 470         int error;
 471
 472         /*
 473          * Due to the order of unstuffing files and ->fault(), we can be
 474          * asked for a zero page in the case of a stuffed file being extended,
 475          * so we need to supply one here. It doesn't happen often.
 476          */
 477         if (unlikely(page->index)) {
 478                 zero_user(page, 0, PAGE_SIZE);
 479                 SetPageUptodate(page);
 480                 return 0;
 481         }
 482
 483         error = gfs2_meta_inode_buffer(ip, &dibh);
 484         if (error)
 485                 return error;
 486
 487         kaddr = kmap_atomic(page);
 488         if (dsize > gfs2_max_stuffed_size(ip))
 489                 dsize = gfs2_max_stuffed_size(ip);
 490         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 491         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 492         kunmap_atomic(kaddr);
 493         flush_dcache_page(page);
 494         brelse(dibh);
 495         SetPageUptodate(page);
 496
 497         return 0;
 498 }
 499
 500
 501 /**
 502  * __gfs2_readpage - readpage
 503  * @file: The file to read a page for
 504  * @page: The page to read
 505  *
 506  * This is the core of gfs2's readpage. It's used by the internal file
 507  * reading code as in that case we already hold the glock. Also it's
 508  * called by gfs2_readpage() once the required lock has been granted.
 509  */
 510
 511 static int __gfs2_readpage(void *file, struct page *page)
 512 {
 513         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 514         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 515         int error;
 516
 517         if (gfs2_is_stuffed(ip)) {
 518                 error = stuffed_readpage(ip, page);
 519                 unlock_page(page);
 520         } else {
 521                 error = mpage_readpage(page, gfs2_block_map);
 522         }
 523
 524         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 525                 return -EIO;
 526
 527         return error;
 528 }
 529
 530 /**
 531  * gfs2_readpage - read a page of a file
 532  * @file: The file to read
 533  * @page: The page of the file
 534  *
 535  * This deals with the locking required. We have to unlock and
 536  * relock the page in order to get the locking in the right
 537  * order.
 538  */
 539
 540 static int gfs2_readpage(struct file *file, struct page *page)
 541 {
 542         struct address_space *mapping = page->mapping;
 543         struct gfs2_inode *ip = GFS2_I(mapping->host);
 544         struct gfs2_holder gh;
 545         int error;
 546
 547         unlock_page(page);
 548         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 549         error = gfs2_glock_nq(&gh);
 550         if (unlikely(error))
 551                 goto out;
 552         error = AOP_TRUNCATED_PAGE;
 553         lock_page(page);
 554         if (page->mapping == mapping && !PageUptodate(page))
 555                 error = __gfs2_readpage(file, page);
 556         else
 557                 unlock_page(page);
 558         gfs2_glock_dq(&gh);
 559 out:
 560         gfs2_holder_uninit(&gh);
 561         if (error && error != AOP_TRUNCATED_PAGE)
 562                 lock_page(page);
 563         return error;
 564 }
 565
 566 /**
 567  * gfs2_internal_read - read an internal file
 568  * @ip: The gfs2 inode
 569  * @buf: The buffer to fill
 570  * @pos: The file position
 571  * @size: The amount to read
 572  *
 573  */
 574
 575 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 576                        unsigned size)
 577 {
 578         struct address_space *mapping = ip->i_inode.i_mapping;
 579         unsigned long index = *pos / PAGE_SIZE;
 580         unsigned offset = *pos & (PAGE_SIZE - 1);
 581         unsigned copied = 0;
 582         unsigned amt;
 583         struct page *page;
 584         void *p;
 585
 586         do {
 587                 amt = size - copied;
 588                 if (offset + size > PAGE_SIZE)
 589                         amt = PAGE_SIZE - offset;
 590                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 591                 if (IS_ERR(page))
 592                         return PTR_ERR(page);
 593                 p = kmap_atomic(page);
 594                 memcpy(buf + copied, p + offset, amt);
 595                 kunmap_atomic(p);
 596                 put_page(page);
 597                 copied += amt;
 598                 index++;
 599                 offset = 0;
 600         } while(copied < size);
 601         (*pos) += size;
 602         return size;
 603 }
 604
 605 /**
 606  * gfs2_readpages - Read a bunch of pages at once
 607  * @file: The file to read from
 608  * @mapping: Address space info
 609  * @pages: List of pages to read
 610  * @nr_pages: Number of pages to read
 611  *
 612  * Some notes:
 613  * 1. This is only for readahead, so we can simply ignore any things
 614  *    which are slightly inconvenient (such as locking conflicts between
 615  *    the page lock and the glock) and return having done no I/O. Its
 616  *    obviously not something we'd want to do on too regular a basis.
 617  *    Any I/O we ignore at this time will be done via readpage later.
 618  * 2. We don't handle stuffed files here we let readpage do the honours.
 619  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 620  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 621  */
 622
 623 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 624                           struct list_head *pages, unsigned nr_pages)
 625 {
 626         struct inode *inode = mapping->host;
 627         struct gfs2_inode *ip = GFS2_I(inode);
 628         struct gfs2_sbd *sdp = GFS2_SB(inode);
 629         struct gfs2_holder gh;
 630         int ret;
 631
 632         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 633         ret = gfs2_glock_nq(&gh);
 634         if (unlikely(ret))
 635                 goto out_uninit;
 636         if (!gfs2_is_stuffed(ip))
 637                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 638         gfs2_glock_dq(&gh);
 639 out_uninit:
 640         gfs2_holder_uninit(&gh);
 641         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 642                 ret = -EIO;
 643         return ret;
 644 }
 645
 646 /**
 647  * gfs2_write_begin - Begin to write to a file
 648  * @file: The file to write to
 649  * @mapping: The mapping in which to write
 650  * @pos: The file offset at which to start writing
 651  * @len: Length of the write
 652  * @flags: Various flags
 653  * @pagep: Pointer to return the page
 654  * @fsdata: Pointer to return fs data (unused by GFS2)
 655  *
 656  * Returns: errno
 657  */
 658
 659 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 660                             loff_t pos, unsigned len, unsigned flags,
 661                             struct page **pagep, void **fsdata)
 662 {
 663         struct gfs2_inode *ip = GFS2_I(mapping->host);
 664         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 665         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 666         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 667         unsigned requested = 0;
 668         int alloc_required;
 669         int error = 0;
 670         pgoff_t index = pos >> PAGE_SHIFT;
 671         unsigned from = pos & (PAGE_SIZE - 1);
 672         struct page *page;
 673
 674         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 675         error = gfs2_glock_nq(&ip->i_gh);
 676         if (unlikely(error))
 677                 goto out_uninit;
 678         if (&ip->i_inode == sdp->sd_rindex) {
 679                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 680                                            GL_NOCACHE, &m_ip->i_gh);
 681                 if (unlikely(error)) {
 682                         gfs2_glock_dq(&ip->i_gh);
 683                         goto out_uninit;
 684                 }
 685         }
 686
 687         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 688
 689         if (alloc_required || gfs2_is_jdata(ip))
 690                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 691
 692         if (alloc_required) {
 693                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 694                 requested = data_blocks + ind_blocks;
 695                 ap.target = requested;
 696                 error = gfs2_quota_lock_check(ip, &ap);
 697                 if (error)
 698                         goto out_unlock;
 699
 700                 error = gfs2_inplace_reserve(ip, &ap);
 701                 if (error)
 702                         goto out_qunlock;
 703         }
 704
 705         rblocks = RES_DINODE + ind_blocks;
 706         if (gfs2_is_jdata(ip))
 707                 rblocks += data_blocks ? data_blocks : 1;
 708         if (ind_blocks || data_blocks)
 709                 rblocks += RES_STATFS + RES_QUOTA;
 710         if (&ip->i_inode == sdp->sd_rindex)
 711                 rblocks += 2 * RES_STATFS;
 712         if (alloc_required)
 713                 rblocks += gfs2_rg_blocks(ip, requested);
 714
 715         error = gfs2_trans_begin(sdp, rblocks,
 716                                  PAGE_SIZE/sdp->sd_sb.sb_bsize);
 717         if (error)
 718                 goto out_trans_fail;
 719
 720         error = -ENOMEM;
 721         flags |= AOP_FLAG_NOFS;
 722         page = grab_cache_page_write_begin(mapping, index, flags);
 723         *pagep = page;
 724         if (unlikely(!page))
 725                 goto out_endtrans;
 726
 727         if (gfs2_is_stuffed(ip)) {
 728                 error = 0;
 729                 if (pos + len > gfs2_max_stuffed_size(ip)) {
 730                         error = gfs2_unstuff_dinode(ip, page);
 731                         if (error == 0)
 732                                 goto prepare_write;
 733                 } else if (!PageUptodate(page)) {
 734                         error = stuffed_readpage(ip, page);
 735                 }
 736                 goto out;
 737         }
 738
 739 prepare_write:
 740         error = __block_write_begin(page, from, len, gfs2_block_map);
 741 out:
 742         if (error == 0)
 743                 return 0;
 744
 745         unlock_page(page);
 746         put_page(page);
 747
 748         gfs2_trans_end(sdp);
 749         if (alloc_required) {
 750                 gfs2_inplace_release(ip);
 751                 if (pos + len > ip->i_inode.i_size)
 752                         gfs2_trim_blocks(&ip->i_inode);
 753         }
 754         goto out_qunlock;
 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         if (alloc_required)
 763                 gfs2_quota_unlock(ip);
 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  * @copied: How much was actually copied by the VFS
 820  * @page: The page
 821  *
 822  * This copies the data from the page into the inode block after
 823  * the inode data structure itself.
 824  *
 825  * Returns: errno
 826  */
 827 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 828                                   loff_t pos, unsigned copied,
 829                                   struct page *page)
 830 {
 831         struct gfs2_inode *ip = GFS2_I(inode);
 832         u64 to = pos + copied;
 833         void *kaddr;
 834         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 835
 836         BUG_ON(pos + copied > gfs2_max_stuffed_size(ip));
 837
 838         kaddr = kmap_atomic(page);
 839         memcpy(buf + pos, kaddr + pos, copied);
 840         flush_dcache_page(page);
 841         kunmap_atomic(kaddr);
 842
 843         WARN_ON(!PageUptodate(page));
 844         unlock_page(page);
 845         put_page(page);
 846
 847         if (copied) {
 848                 if (inode->i_size < to)
 849                         i_size_write(inode, to);
 850                 mark_inode_dirty(inode);
 851         }
 852         return copied;
 853 }
 854
 855 /**
 856  * gfs2_write_end
 857  * @file: The file to write to
 858  * @mapping: The address space to write to
 859  * @pos: The file position
 860  * @len: The length of the data
 861  * @copied: How much was actually copied by the VFS
 862  * @page: The page that has been written
 863  * @fsdata: The fsdata (unused in GFS2)
 864  *
 865  * The main write_end function for GFS2. We just put our locking around the VFS
 866  * provided functions.
 867  *
 868  * Returns: errno
 869  */
 870
 871 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 872                           loff_t pos, unsigned len, unsigned copied,
 873                           struct page *page, void *fsdata)
 874 {
 875         struct inode *inode = page->mapping->host;
 876         struct gfs2_inode *ip = GFS2_I(inode);
 877         struct gfs2_sbd *sdp = GFS2_SB(inode);
 878         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 879         struct buffer_head *dibh;
 880         int ret;
 881         struct gfs2_trans *tr = current->journal_info;
 882         BUG_ON(!tr);
 883
 884         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 885
 886         ret = gfs2_meta_inode_buffer(ip, &dibh);
 887         if (unlikely(ret))
 888                 goto out;
 889
 890         if (gfs2_is_stuffed(ip)) {
 891                 ret = gfs2_stuffed_write_end(inode, dibh, pos, copied, page);
 892                 page = NULL;
 893                 goto out2;
 894         }
 895
 896         if (gfs2_is_jdata(ip))
 897                 gfs2_page_add_databufs(ip, page, pos & ~PAGE_MASK, len);
 898         else
 899                 gfs2_ordered_add_inode(ip);
 900
 901         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 902         page = NULL;
 903         if (tr->tr_num_buf_new)
 904                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 905         else
 906                 gfs2_trans_add_meta(ip->i_gl, dibh);
 907
 908 out2:
 909         if (inode == sdp->sd_rindex) {
 910                 adjust_fs_space(inode);
 911                 sdp->sd_rindex_uptodate = 0;
 912         }
 913
 914         brelse(dibh);
 915 out:
 916         if (page) {
 917                 unlock_page(page);
 918                 put_page(page);
 919         }
 920         gfs2_trans_end(sdp);
 921         gfs2_inplace_release(ip);
 922         if (ip->i_qadata && ip->i_qadata->qa_qd_num)
 923                 gfs2_quota_unlock(ip);
 924         if (inode == sdp->sd_rindex) {
 925                 gfs2_glock_dq(&m_ip->i_gh);
 926                 gfs2_holder_uninit(&m_ip->i_gh);
 927         }
 928         gfs2_glock_dq(&ip->i_gh);
 929         gfs2_holder_uninit(&ip->i_gh);
 930         return ret;
 931 }
 932
 933 /**
 934  * jdata_set_page_dirty - Page dirtying function
 935  * @page: The page to dirty
 936  *
 937  * Returns: 1 if it dirtyed the page, or 0 otherwise
 938  */
 939
 940 static int jdata_set_page_dirty(struct page *page)
 941 {
 942         SetPageChecked(page);
 943         return __set_page_dirty_buffers(page);
 944 }
 945
 946 /**
 947  * gfs2_bmap - Block map function
 948  * @mapping: Address space info
 949  * @lblock: The block to map
 950  *
 951  * Returns: The disk address for the block or 0 on hole or error
 952  */
 953
 954 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 955 {
 956         struct gfs2_inode *ip = GFS2_I(mapping->host);
 957         struct gfs2_holder i_gh;
 958         sector_t dblock = 0;
 959         int error;
 960
 961         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 962         if (error)
 963                 return 0;
 964
 965         if (!gfs2_is_stuffed(ip))
 966                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 967
 968         gfs2_glock_dq_uninit(&i_gh);
 969
 970         return dblock;
 971 }
 972
 973 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 974 {
 975         struct gfs2_bufdata *bd;
 976
 977         lock_buffer(bh);
 978         gfs2_log_lock(sdp);
 979         clear_buffer_dirty(bh);
 980         bd = bh->b_private;
 981         if (bd) {
 982                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 983                         list_del_init(&bd->bd_list);
 984                 else
 985                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
 986         }
 987         bh->b_bdev = NULL;
 988         clear_buffer_mapped(bh);
 989         clear_buffer_req(bh);
 990         clear_buffer_new(bh);
 991         gfs2_log_unlock(sdp);
 992         unlock_buffer(bh);
 993 }
 994
 995 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
 996                                 unsigned int length)
 997 {
 998         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 999         unsigned int stop = offset + length;
1000         int partial_page = (offset || length < PAGE_SIZE);
1001         struct buffer_head *bh, *head;
1002         unsigned long pos = 0;
1003
1004         BUG_ON(!PageLocked(page));
1005         if (!partial_page)
1006                 ClearPageChecked(page);
1007         if (!page_has_buffers(page))
1008                 goto out;
1009
1010         bh = head = page_buffers(page);
1011         do {
1012                 if (pos + bh->b_size > stop)
1013                         return;
1014
1015                 if (offset <= pos)
1016                         gfs2_discard(sdp, bh);
1017                 pos += bh->b_size;
1018                 bh = bh->b_this_page;
1019         } while (bh != head);
1020 out:
1021         if (!partial_page)
1022                 try_to_release_page(page, 0);
1023 }
1024
1025 /**
1026  * gfs2_ok_for_dio - check that dio is valid on this file
1027  * @ip: The inode
1028  * @offset: The offset at which we are reading or writing
1029  *
1030  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1031  *          1 (to accept the i/o request)
1032  */
1033 static int gfs2_ok_for_dio(struct gfs2_inode *ip, loff_t offset)
1034 {
1035         /*
1036          * Should we return an error here? I can't see that O_DIRECT for
1037          * a stuffed file makes any sense. For now we'll silently fall
1038          * back to buffered I/O
1039          */
1040         if (gfs2_is_stuffed(ip))
1041                 return 0;
1042
1043         if (offset >= i_size_read(&ip->i_inode))
1044                 return 0;
1045         return 1;
1046 }
1047
1048
1049
1050 static ssize_t gfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
1051 {
1052         struct file *file = iocb->ki_filp;
1053         struct inode *inode = file->f_mapping->host;
1054         struct address_space *mapping = inode->i_mapping;
1055         struct gfs2_inode *ip = GFS2_I(inode);
1056         loff_t offset = iocb->ki_pos;
1057         struct gfs2_holder gh;
1058         int rv;
1059
1060         /*
1061          * Deferred lock, even if its a write, since we do no allocation
1062          * on this path. All we need change is atime, and this lock mode
1063          * ensures that other nodes have flushed their buffered read caches
1064          * (i.e. their page cache entries for this inode). We do not,
1065          * unfortunately have the option of only flushing a range like
1066          * the VFS does.
1067          */
1068         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1069         rv = gfs2_glock_nq(&gh);
1070         if (rv)
1071                 goto out_uninit;
1072         rv = gfs2_ok_for_dio(ip, offset);
1073         if (rv != 1)
1074                 goto out; /* dio not valid, fall back to buffered i/o */
1075
1076         /*
1077          * Now since we are holding a deferred (CW) lock at this point, you
1078          * might be wondering why this is ever needed. There is a case however
1079          * where we've granted a deferred local lock against a cached exclusive
1080          * glock. That is ok provided all granted local locks are deferred, but
1081          * it also means that it is possible to encounter pages which are
1082          * cached and possibly also mapped. So here we check for that and sort
1083          * them out ahead of the dio. The glock state machine will take care of
1084          * everything else.
1085          *
1086          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1087          * the first place, mapping->nr_pages will always be zero.
1088          */
1089         if (mapping->nrpages) {
1090                 loff_t lstart = offset & ~(PAGE_SIZE - 1);
1091                 loff_t len = iov_iter_count(iter);
1092                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1093
1094                 rv = 0;
1095                 if (len == 0)
1096                         goto out;
1097                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1098                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1099                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1100                 if (rv)
1101                         goto out;
1102                 if (iov_iter_rw(iter) == WRITE)
1103                         truncate_inode_pages_range(mapping, lstart, end);
1104         }
1105
1106         rv = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
1107                                   gfs2_get_block_direct, NULL, NULL, 0);
1108 out:
1109         gfs2_glock_dq(&gh);
1110 out_uninit:
1111         gfs2_holder_uninit(&gh);
1112         return rv;
1113 }
1114
1115 /**
1116  * gfs2_releasepage - free the metadata associated with a page
1117  * @page: the page that's being released
1118  * @gfp_mask: passed from Linux VFS, ignored by us
1119  *
1120  * Call try_to_free_buffers() if the buffers in this page can be
1121  * released.
1122  *
1123  * Returns: 0
1124  */
1125
1126 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1127 {
1128         struct address_space *mapping = page->mapping;
1129         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1130         struct buffer_head *bh, *head;
1131         struct gfs2_bufdata *bd;
1132
1133         if (!page_has_buffers(page))
1134                 return 0;
1135
1136         /*
1137          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
1138          * clean pages might not have had the dirty bit cleared.  Thus, it can
1139          * send actual dirty pages to ->releasepage() via shrink_active_list().
1140          *
1141          * As a workaround, we skip pages that contain dirty buffers below.
1142          * Once ->releasepage isn't called on dirty pages anymore, we can warn
1143          * on dirty buffers like we used to here again.
1144          */
1145
1146         gfs2_log_lock(sdp);
1147         spin_lock(&sdp->sd_ail_lock);
1148         head = bh = page_buffers(page);
1149         do {
1150                 if (atomic_read(&bh->b_count))
1151                         goto cannot_release;
1152                 bd = bh->b_private;
1153                 if (bd && bd->bd_tr)
1154                         goto cannot_release;
1155                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
1156                         goto cannot_release;
1157                 bh = bh->b_this_page;
1158         } while(bh != head);
1159         spin_unlock(&sdp->sd_ail_lock);
1160
1161         head = bh = page_buffers(page);
1162         do {
1163                 bd = bh->b_private;
1164                 if (bd) {
1165                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1166                         if (!list_empty(&bd->bd_list))
1167                                 list_del_init(&bd->bd_list);
1168                         bd->bd_bh = NULL;
1169                         bh->b_private = NULL;
1170                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1171                 }
1172
1173                 bh = bh->b_this_page;
1174         } while (bh != head);
1175         gfs2_log_unlock(sdp);
1176
1177         return try_to_free_buffers(page);
1178
1179 cannot_release:
1180         spin_unlock(&sdp->sd_ail_lock);
1181         gfs2_log_unlock(sdp);
1182         return 0;
1183 }
1184
1185 static const struct address_space_operations gfs2_writeback_aops = {
1186         .writepage = gfs2_writepage,
1187         .writepages = gfs2_writepages,
1188         .readpage = gfs2_readpage,
1189         .readpages = gfs2_readpages,
1190         .write_begin = gfs2_write_begin,
1191         .write_end = gfs2_write_end,
1192         .bmap = gfs2_bmap,
1193         .invalidatepage = gfs2_invalidatepage,
1194         .releasepage = gfs2_releasepage,
1195         .direct_IO = gfs2_direct_IO,
1196         .migratepage = buffer_migrate_page,
1197         .is_partially_uptodate = block_is_partially_uptodate,
1198         .error_remove_page = generic_error_remove_page,
1199 };
1200
1201 static const struct address_space_operations gfs2_ordered_aops = {
1202         .writepage = gfs2_writepage,
1203         .writepages = gfs2_writepages,
1204         .readpage = gfs2_readpage,
1205         .readpages = gfs2_readpages,
1206         .write_begin = gfs2_write_begin,
1207         .write_end = gfs2_write_end,
1208         .set_page_dirty = __set_page_dirty_buffers,
1209         .bmap = gfs2_bmap,
1210         .invalidatepage = gfs2_invalidatepage,
1211         .releasepage = gfs2_releasepage,
1212         .direct_IO = gfs2_direct_IO,
1213         .migratepage = buffer_migrate_page,
1214         .is_partially_uptodate = block_is_partially_uptodate,
1215         .error_remove_page = generic_error_remove_page,
1216 };
1217
1218 static const struct address_space_operations gfs2_jdata_aops = {
1219         .writepage = gfs2_jdata_writepage,
1220         .writepages = gfs2_jdata_writepages,
1221         .readpage = gfs2_readpage,
1222         .readpages = gfs2_readpages,
1223         .write_begin = gfs2_write_begin,
1224         .write_end = gfs2_write_end,
1225         .set_page_dirty = jdata_set_page_dirty,
1226         .bmap = gfs2_bmap,
1227         .invalidatepage = gfs2_invalidatepage,
1228         .releasepage = gfs2_releasepage,
1229         .is_partially_uptodate = block_is_partially_uptodate,
1230         .error_remove_page = generic_error_remove_page,
1231 };
1232
1233 void gfs2_set_aops(struct inode *inode)
1234 {
1235         struct gfs2_inode *ip = GFS2_I(inode);
1236
1237         if (gfs2_is_writeback(ip))
1238                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1239         else if (gfs2_is_ordered(ip))
1240                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1241         else if (gfs2_is_jdata(ip))
1242                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1243         else
1244                 BUG();
1245 }
1246