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                 if (gfs2_is_jdata(ip))
  58                         set_buffer_uptodate(bh);
  59                 gfs2_trans_add_data(ip->i_gl, bh);
  60         }
  61 }
  62
  63 /**
  64  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  65  * @inode: The inode
  66  * @lblock: The block number to look up
  67  * @bh_result: The buffer head to return the result in
  68  * @create: Non-zero if we may add block to the file
  69  *
  70  * Returns: errno
  71  */
  72
  73 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  74                                   struct buffer_head *bh_result, int create)
  75 {
  76         int error;
  77
  78         error = gfs2_block_map(inode, lblock, bh_result, 0);
  79         if (error)
  80                 return error;
  81         if (!buffer_mapped(bh_result))
  82                 return -EIO;
  83         return 0;
  84 }
  85
  86 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  87                                  struct buffer_head *bh_result, int create)
  88 {
  89         return gfs2_block_map(inode, lblock, bh_result, 0);
  90 }
  91
  92 /**
  93  * gfs2_writepage_common - Common bits of writepage
  94  * @page: The page to be written
  95  * @wbc: The writeback control
  96  *
  97  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  98  */
  99
 100 static int gfs2_writepage_common(struct page *page,
 101                                  struct writeback_control *wbc)
 102 {
 103         struct inode *inode = page->mapping->host;
 104         struct gfs2_inode *ip = GFS2_I(inode);
 105         struct gfs2_sbd *sdp = GFS2_SB(inode);
 106         loff_t i_size = i_size_read(inode);
 107         pgoff_t end_index = i_size >> PAGE_SHIFT;
 108         unsigned offset;
 109
 110         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 111                 goto out;
 112         if (current->journal_info)
 113                 goto redirty;
 114         /* Is the page fully outside i_size? (truncate in progress) */
 115         offset = i_size & (PAGE_SIZE-1);
 116         if (page->index > end_index || (page->index == end_index && !offset)) {
 117                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 118                 goto out;
 119         }
 120         return 1;
 121 redirty:
 122         redirty_page_for_writepage(wbc, page);
 123 out:
 124         unlock_page(page);
 125         return 0;
 126 }
 127
 128 /**
 129  * gfs2_writepage - Write page for writeback mappings
 130  * @page: The page
 131  * @wbc: The writeback control
 132  *
 133  */
 134
 135 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 136 {
 137         int ret;
 138
 139         ret = gfs2_writepage_common(page, wbc);
 140         if (ret <= 0)
 141                 return ret;
 142
 143         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 144 }
 145
 146 /* This is the same as calling block_write_full_page, but it also
 147  * writes pages outside of i_size
 148  */
 149 static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 150                                 struct writeback_control *wbc)
 151 {
 152         struct inode * const inode = page->mapping->host;
 153         loff_t i_size = i_size_read(inode);
 154         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 155         unsigned offset;
 156
 157         /*
 158          * The page straddles i_size.  It must be zeroed out on each and every
 159          * writepage invocation because it may be mmapped.  "A file is mapped
 160          * in multiples of the page size.  For a file that is not a multiple of
 161          * the  page size, the remaining memory is zeroed when mapped, and
 162          * writes to that region are not written out to the file."
 163          */
 164         offset = i_size & (PAGE_SIZE-1);
 165         if (page->index == end_index && offset)
 166                 zero_user_segment(page, offset, PAGE_SIZE);
 167
 168         return __block_write_full_page(inode, page, get_block, wbc,
 169                                        end_buffer_async_write);
 170 }
 171
 172 /**
 173  * __gfs2_jdata_writepage - The core of jdata writepage
 174  * @page: The page to write
 175  * @wbc: The writeback control
 176  *
 177  * This is shared between writepage and writepages and implements the
 178  * core of the writepage operation. If a transaction is required then
 179  * PageChecked will have been set and the transaction will have
 180  * already been started before this is called.
 181  */
 182
 183 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 184 {
 185         struct inode *inode = page->mapping->host;
 186         struct gfs2_inode *ip = GFS2_I(inode);
 187         struct gfs2_sbd *sdp = GFS2_SB(inode);
 188
 189         if (PageChecked(page)) {
 190                 ClearPageChecked(page);
 191                 if (!page_has_buffers(page)) {
 192                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 193                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 194                 }
 195                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize);
 196         }
 197         return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
 198 }
 199
 200 /**
 201  * gfs2_jdata_writepage - Write complete page
 202  * @page: Page to write
 203  * @wbc: The writeback control
 204  *
 205  * Returns: errno
 206  *
 207  */
 208
 209 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 210 {
 211         struct inode *inode = page->mapping->host;
 212         struct gfs2_inode *ip = GFS2_I(inode);
 213         struct gfs2_sbd *sdp = GFS2_SB(inode);
 214         int ret;
 215
 216         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 217                 goto out;
 218         if (PageChecked(page) || current->journal_info)
 219                 goto out_ignore;
 220         ret = __gfs2_jdata_writepage(page, wbc);
 221         return ret;
 222
 223 out_ignore:
 224         redirty_page_for_writepage(wbc, page);
 225 out:
 226         unlock_page(page);
 227         return 0;
 228 }
 229
 230 /**
 231  * gfs2_writepages - Write a bunch of dirty pages back to disk
 232  * @mapping: The mapping to write
 233  * @wbc: Write-back control
 234  *
 235  * Used for both ordered and writeback modes.
 236  */
 237 static int gfs2_writepages(struct address_space *mapping,
 238                            struct writeback_control *wbc)
 239 {
 240         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 241         int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 242
 243         /*
 244          * Even if we didn't write any pages here, we might still be holding
 245          * dirty pages in the ail. We forcibly flush the ail because we don't
 246          * want balance_dirty_pages() to loop indefinitely trying to write out
 247          * pages held in the ail that it can't find.
 248          */
 249         if (ret == 0)
 250                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 251
 252         return ret;
 253 }
 254
 255 /**
 256  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 257  * @mapping: The mapping
 258  * @wbc: The writeback control
 259  * @pvec: The vector of pages
 260  * @nr_pages: The number of pages to write
 261  * @done_index: Page index
 262  *
 263  * Returns: non-zero if loop should terminate, zero otherwise
 264  */
 265
 266 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 267                                     struct writeback_control *wbc,
 268                                     struct pagevec *pvec,
 269                                     int nr_pages,
 270                                     pgoff_t *done_index)
 271 {
 272         struct inode *inode = mapping->host;
 273         struct gfs2_sbd *sdp = GFS2_SB(inode);
 274         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 275         int i;
 276         int ret;
 277
 278         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 279         if (ret < 0)
 280                 return ret;
 281
 282         for(i = 0; i < nr_pages; i++) {
 283                 struct page *page = pvec->pages[i];
 284
 285                 *done_index = page->index;
 286
 287                 lock_page(page);
 288
 289                 if (unlikely(page->mapping != mapping)) {
 290 continue_unlock:
 291                         unlock_page(page);
 292                         continue;
 293                 }
 294
 295                 if (!PageDirty(page)) {
 296                         /* someone wrote it for us */
 297                         goto continue_unlock;
 298                 }
 299
 300                 if (PageWriteback(page)) {
 301                         if (wbc->sync_mode != WB_SYNC_NONE)
 302                                 wait_on_page_writeback(page);
 303                         else
 304                                 goto continue_unlock;
 305                 }
 306
 307                 BUG_ON(PageWriteback(page));
 308                 if (!clear_page_dirty_for_io(page))
 309                         goto continue_unlock;
 310
 311                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 312
 313                 ret = __gfs2_jdata_writepage(page, wbc);
 314                 if (unlikely(ret)) {
 315                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 316                                 unlock_page(page);
 317                                 ret = 0;
 318                         } else {
 319
 320                                 /*
 321                                  * done_index is set past this page,
 322                                  * so media errors will not choke
 323                                  * background writeout for the entire
 324                                  * file. This has consequences for
 325                                  * range_cyclic semantics (ie. it may
 326                                  * not be suitable for data integrity
 327                                  * writeout).
 328                                  */
 329                                 *done_index = page->index + 1;
 330                                 ret = 1;
 331                                 break;
 332                         }
 333                 }
 334
 335                 /*
 336                  * We stop writing back only if we are not doing
 337                  * integrity sync. In case of integrity sync we have to
 338                  * keep going until we have written all the pages
 339                  * we tagged for writeback prior to entering this loop.
 340                  */
 341                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 342                         ret = 1;
 343                         break;
 344                 }
 345
 346         }
 347         gfs2_trans_end(sdp);
 348         return ret;
 349 }
 350
 351 /**
 352  * gfs2_write_cache_jdata - Like write_cache_pages but different
 353  * @mapping: The mapping to write
 354  * @wbc: The writeback control
 355  *
 356  * The reason that we use our own function here is that we need to
 357  * start transactions before we grab page locks. This allows us
 358  * to get the ordering right.
 359  */
 360
 361 static int gfs2_write_cache_jdata(struct address_space *mapping,
 362                                   struct writeback_control *wbc)
 363 {
 364         int ret = 0;
 365         int done = 0;
 366         struct pagevec pvec;
 367         int nr_pages;
 368         pgoff_t uninitialized_var(writeback_index);
 369         pgoff_t index;
 370         pgoff_t end;
 371         pgoff_t done_index;
 372         int cycled;
 373         int range_whole = 0;
 374         int tag;
 375
 376         pagevec_init(&pvec);
 377         if (wbc->range_cyclic) {
 378                 writeback_index = mapping->writeback_index; /* prev offset */
 379                 index = writeback_index;
 380                 if (index == 0)
 381                         cycled = 1;
 382                 else
 383                         cycled = 0;
 384                 end = -1;
 385         } else {
 386                 index = wbc->range_start >> PAGE_SHIFT;
 387                 end = wbc->range_end >> PAGE_SHIFT;
 388                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 389                         range_whole = 1;
 390                 cycled = 1; /* ignore range_cyclic tests */
 391         }
 392         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 393                 tag = PAGECACHE_TAG_TOWRITE;
 394         else
 395                 tag = PAGECACHE_TAG_DIRTY;
 396
 397 retry:
 398         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 399                 tag_pages_for_writeback(mapping, index, end);
 400         done_index = index;
 401         while (!done && (index <= end)) {
 402                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
 403                                 tag);
 404                 if (nr_pages == 0)
 405                         break;
 406
 407                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
 408                 if (ret)
 409                         done = 1;
 410                 if (ret > 0)
 411                         ret = 0;
 412                 pagevec_release(&pvec);
 413                 cond_resched();
 414         }
 415
 416         if (!cycled && !done) {
 417                 /*
 418                  * range_cyclic:
 419                  * We hit the last page and there is more work to be done: wrap
 420                  * back to the start of the file
 421                  */
 422                 cycled = 1;
 423                 index = 0;
 424                 end = writeback_index - 1;
 425                 goto retry;
 426         }
 427
 428         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 429                 mapping->writeback_index = done_index;
 430
 431         return ret;
 432 }
 433
 434
 435 /**
 436  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 437  * @mapping: The mapping to write
 438  * @wbc: The writeback control
 439  *
 440  */
 441
 442 static int gfs2_jdata_writepages(struct address_space *mapping,
 443                                  struct writeback_control *wbc)
 444 {
 445         struct gfs2_inode *ip = GFS2_I(mapping->host);
 446         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 447         int ret;
 448
 449         ret = gfs2_write_cache_jdata(mapping, wbc);
 450         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 451                 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
 452                                GFS2_LFC_JDATA_WPAGES);
 453                 ret = gfs2_write_cache_jdata(mapping, wbc);
 454         }
 455         return ret;
 456 }
 457
 458 /**
 459  * stuffed_readpage - Fill in a Linux page with stuffed file data
 460  * @ip: the inode
 461  * @page: the page
 462  *
 463  * Returns: errno
 464  */
 465
 466 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 467 {
 468         struct buffer_head *dibh;
 469         u64 dsize = i_size_read(&ip->i_inode);
 470         void *kaddr;
 471         int error;
 472
 473         /*
 474          * Due to the order of unstuffing files and ->fault(), we can be
 475          * asked for a zero page in the case of a stuffed file being extended,
 476          * so we need to supply one here. It doesn't happen often.
 477          */
 478         if (unlikely(page->index)) {
 479                 zero_user(page, 0, PAGE_SIZE);
 480                 SetPageUptodate(page);
 481                 return 0;
 482         }
 483
 484         error = gfs2_meta_inode_buffer(ip, &dibh);
 485         if (error)
 486                 return error;
 487
 488         kaddr = kmap_atomic(page);
 489         if (dsize > gfs2_max_stuffed_size(ip))
 490                 dsize = gfs2_max_stuffed_size(ip);
 491         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 492         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 493         kunmap_atomic(kaddr);
 494         flush_dcache_page(page);
 495         brelse(dibh);
 496         SetPageUptodate(page);
 497
 498         return 0;
 499 }
 500
 501
 502 /**
 503  * __gfs2_readpage - readpage
 504  * @file: The file to read a page for
 505  * @page: The page to read
 506  *
 507  * This is the core of gfs2's readpage. It's used by the internal file
 508  * reading code as in that case we already hold the glock. Also it's
 509  * called by gfs2_readpage() once the required lock has been granted.
 510  */
 511
 512 static int __gfs2_readpage(void *file, struct page *page)
 513 {
 514         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 515         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 516         int error;
 517
 518         if (gfs2_is_stuffed(ip)) {
 519                 error = stuffed_readpage(ip, page);
 520                 unlock_page(page);
 521         } else {
 522                 error = mpage_readpage(page, gfs2_block_map);
 523         }
 524
 525         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 526                 return -EIO;
 527
 528         return error;
 529 }
 530
 531 /**
 532  * gfs2_readpage - read a page of a file
 533  * @file: The file to read
 534  * @page: The page of the file
 535  *
 536  * This deals with the locking required. We have to unlock and
 537  * relock the page in order to get the locking in the right
 538  * order.
 539  */
 540
 541 static int gfs2_readpage(struct file *file, struct page *page)
 542 {
 543         struct address_space *mapping = page->mapping;
 544         struct gfs2_inode *ip = GFS2_I(mapping->host);
 545         struct gfs2_holder gh;
 546         int error;
 547
 548         unlock_page(page);
 549         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 550         error = gfs2_glock_nq(&gh);
 551         if (unlikely(error))
 552                 goto out;
 553         error = AOP_TRUNCATED_PAGE;
 554         lock_page(page);
 555         if (page->mapping == mapping && !PageUptodate(page))
 556                 error = __gfs2_readpage(file, page);
 557         else
 558                 unlock_page(page);
 559         gfs2_glock_dq(&gh);
 560 out:
 561         gfs2_holder_uninit(&gh);
 562         if (error && error != AOP_TRUNCATED_PAGE)
 563                 lock_page(page);
 564         return error;
 565 }
 566
 567 /**
 568  * gfs2_internal_read - read an internal file
 569  * @ip: The gfs2 inode
 570  * @buf: The buffer to fill
 571  * @pos: The file position
 572  * @size: The amount to read
 573  *
 574  */
 575
 576 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 577                        unsigned size)
 578 {
 579         struct address_space *mapping = ip->i_inode.i_mapping;
 580         unsigned long index = *pos / PAGE_SIZE;
 581         unsigned offset = *pos & (PAGE_SIZE - 1);
 582         unsigned copied = 0;
 583         unsigned amt;
 584         struct page *page;
 585         void *p;
 586
 587         do {
 588                 amt = size - copied;
 589                 if (offset + size > PAGE_SIZE)
 590                         amt = PAGE_SIZE - offset;
 591                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 592                 if (IS_ERR(page))
 593                         return PTR_ERR(page);
 594                 p = kmap_atomic(page);
 595                 memcpy(buf + copied, p + offset, amt);
 596                 kunmap_atomic(p);
 597                 put_page(page);
 598                 copied += amt;
 599                 index++;
 600                 offset = 0;
 601         } while(copied < size);
 602         (*pos) += size;
 603         return size;
 604 }
 605
 606 /**
 607  * gfs2_readpages - Read a bunch of pages at once
 608  * @file: The file to read from
 609  * @mapping: Address space info
 610  * @pages: List of pages to read
 611  * @nr_pages: Number of pages to read
 612  *
 613  * Some notes:
 614  * 1. This is only for readahead, so we can simply ignore any things
 615  *    which are slightly inconvenient (such as locking conflicts between
 616  *    the page lock and the glock) and return having done no I/O. Its
 617  *    obviously not something we'd want to do on too regular a basis.
 618  *    Any I/O we ignore at this time will be done via readpage later.
 619  * 2. We don't handle stuffed files here we let readpage do the honours.
 620  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 621  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 622  */
 623
 624 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 625                           struct list_head *pages, unsigned nr_pages)
 626 {
 627         struct inode *inode = mapping->host;
 628         struct gfs2_inode *ip = GFS2_I(inode);
 629         struct gfs2_sbd *sdp = GFS2_SB(inode);
 630         struct gfs2_holder gh;
 631         int ret;
 632
 633         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 634         ret = gfs2_glock_nq(&gh);
 635         if (unlikely(ret))
 636                 goto out_uninit;
 637         if (!gfs2_is_stuffed(ip))
 638                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 639         gfs2_glock_dq(&gh);
 640 out_uninit:
 641         gfs2_holder_uninit(&gh);
 642         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 643                 ret = -EIO;
 644         return ret;
 645 }
 646
 647 /**
 648  * gfs2_write_begin - Begin to write to a file
 649  * @file: The file to write to
 650  * @mapping: The mapping in which to write
 651  * @pos: The file offset at which to start writing
 652  * @len: Length of the write
 653  * @flags: Various flags
 654  * @pagep: Pointer to return the page
 655  * @fsdata: Pointer to return fs data (unused by GFS2)
 656  *
 657  * Returns: errno
 658  */
 659
 660 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 661                             loff_t pos, unsigned len, unsigned flags,
 662                             struct page **pagep, void **fsdata)
 663 {
 664         struct gfs2_inode *ip = GFS2_I(mapping->host);
 665         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 666         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 667         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 668         unsigned requested = 0;
 669         int alloc_required;
 670         int error = 0;
 671         pgoff_t index = pos >> PAGE_SHIFT;
 672         unsigned from = pos & (PAGE_SIZE - 1);
 673         struct page *page;
 674
 675         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 676         error = gfs2_glock_nq(&ip->i_gh);
 677         if (unlikely(error))
 678                 goto out_uninit;
 679         if (&ip->i_inode == sdp->sd_rindex) {
 680                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 681                                            GL_NOCACHE, &m_ip->i_gh);
 682                 if (unlikely(error)) {
 683                         gfs2_glock_dq(&ip->i_gh);
 684                         goto out_uninit;
 685                 }
 686         }
 687
 688         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 689
 690         if (alloc_required || gfs2_is_jdata(ip))
 691                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 692
 693         if (alloc_required) {
 694                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 695                 requested = data_blocks + ind_blocks;
 696                 ap.target = requested;
 697                 error = gfs2_quota_lock_check(ip, &ap);
 698                 if (error)
 699                         goto out_unlock;
 700
 701                 error = gfs2_inplace_reserve(ip, &ap);
 702                 if (error)
 703                         goto out_qunlock;
 704         }
 705
 706         rblocks = RES_DINODE + ind_blocks;
 707         if (gfs2_is_jdata(ip))
 708                 rblocks += data_blocks ? data_blocks : 1;
 709         if (ind_blocks || data_blocks)
 710                 rblocks += RES_STATFS + RES_QUOTA;
 711         if (&ip->i_inode == sdp->sd_rindex)
 712                 rblocks += 2 * RES_STATFS;
 713         if (alloc_required)
 714                 rblocks += gfs2_rg_blocks(ip, requested);
 715
 716         error = gfs2_trans_begin(sdp, rblocks,
 717                                  PAGE_SIZE/sdp->sd_sb.sb_bsize);
 718         if (error)
 719                 goto out_trans_fail;
 720
 721         error = -ENOMEM;
 722         flags |= AOP_FLAG_NOFS;
 723         page = grab_cache_page_write_begin(mapping, index, flags);
 724         *pagep = page;
 725         if (unlikely(!page))
 726                 goto out_endtrans;
 727
 728         if (gfs2_is_stuffed(ip)) {
 729                 error = 0;
 730                 if (pos + len > gfs2_max_stuffed_size(ip)) {
 731                         error = gfs2_unstuff_dinode(ip, page);
 732                         if (error == 0)
 733                                 goto prepare_write;
 734                 } else if (!PageUptodate(page)) {
 735                         error = stuffed_readpage(ip, page);
 736                 }
 737                 goto out;
 738         }
 739
 740 prepare_write:
 741         error = __block_write_begin(page, from, len, gfs2_block_map);
 742 out:
 743         if (error == 0)
 744                 return 0;
 745
 746         unlock_page(page);
 747         put_page(page);
 748
 749         gfs2_trans_end(sdp);
 750         if (pos + len > ip->i_inode.i_size)
 751                 gfs2_trim_blocks(&ip->i_inode);
 752         goto out_trans_fail;
 753
 754 out_endtrans:
 755         gfs2_trans_end(sdp);
 756 out_trans_fail:
 757         if (alloc_required) {
 758                 gfs2_inplace_release(ip);
 759 out_qunlock:
 760                 gfs2_quota_unlock(ip);
 761         }
 762 out_unlock:
 763         if (&ip->i_inode == sdp->sd_rindex) {
 764                 gfs2_glock_dq(&m_ip->i_gh);
 765                 gfs2_holder_uninit(&m_ip->i_gh);
 766         }
 767         gfs2_glock_dq(&ip->i_gh);
 768 out_uninit:
 769         gfs2_holder_uninit(&ip->i_gh);
 770         return error;
 771 }
 772
 773 /**
 774  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 775  * @inode: the rindex inode
 776  */
 777 static void adjust_fs_space(struct inode *inode)
 778 {
 779         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 780         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 781         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 782         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 783         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 784         struct buffer_head *m_bh, *l_bh;
 785         u64 fs_total, new_free;
 786
 787         /* Total up the file system space, according to the latest rindex. */
 788         fs_total = gfs2_ri_total(sdp);
 789         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 790                 return;
 791
 792         spin_lock(&sdp->sd_statfs_spin);
 793         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 794                               sizeof(struct gfs2_dinode));
 795         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 796                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 797         else
 798                 new_free = 0;
 799         spin_unlock(&sdp->sd_statfs_spin);
 800         fs_warn(sdp, "File system extended by %llu blocks.\n",
 801                 (unsigned long long)new_free);
 802         gfs2_statfs_change(sdp, new_free, new_free, 0);
 803
 804         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 805                 goto out;
 806         update_statfs(sdp, m_bh, l_bh);
 807         brelse(l_bh);
 808 out:
 809         brelse(m_bh);
 810 }
 811
 812 /**
 813  * gfs2_stuffed_write_end - Write end for stuffed files
 814  * @inode: The inode
 815  * @dibh: The buffer_head containing the on-disk inode
 816  * @pos: The file position
 817  * @len: The length of the write
 818  * @copied: How much was actually copied by the VFS
 819  * @page: The page
 820  *
 821  * This copies the data from the page into the inode block after
 822  * the inode data structure itself.
 823  *
 824  * Returns: errno
 825  */
 826 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 827                                   loff_t pos, unsigned len, unsigned copied,
 828                                   struct page *page)
 829 {
 830         struct gfs2_inode *ip = GFS2_I(inode);
 831         struct gfs2_sbd *sdp = GFS2_SB(inode);
 832         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 833         u64 to = pos + copied;
 834         void *kaddr;
 835         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 836
 837         BUG_ON(pos + len > gfs2_max_stuffed_size(ip));
 838
 839         kaddr = kmap_atomic(page);
 840         memcpy(buf + pos, kaddr + pos, copied);
 841         flush_dcache_page(page);
 842         kunmap_atomic(kaddr);
 843
 844         WARN_ON(!PageUptodate(page));
 845         unlock_page(page);
 846         put_page(page);
 847
 848         if (copied) {
 849                 if (inode->i_size < to)
 850                         i_size_write(inode, to);
 851                 mark_inode_dirty(inode);
 852         }
 853
 854         if (inode == sdp->sd_rindex) {
 855                 adjust_fs_space(inode);
 856                 sdp->sd_rindex_uptodate = 0;
 857         }
 858
 859         brelse(dibh);
 860         gfs2_trans_end(sdp);
 861         if (inode == sdp->sd_rindex) {
 862                 gfs2_glock_dq(&m_ip->i_gh);
 863                 gfs2_holder_uninit(&m_ip->i_gh);
 864         }
 865         gfs2_glock_dq(&ip->i_gh);
 866         gfs2_holder_uninit(&ip->i_gh);
 867         return copied;
 868 }
 869
 870 /**
 871  * gfs2_write_end
 872  * @file: The file to write to
 873  * @mapping: The address space to write to
 874  * @pos: The file position
 875  * @len: The length of the data
 876  * @copied: How much was actually copied by the VFS
 877  * @page: The page that has been written
 878  * @fsdata: The fsdata (unused in GFS2)
 879  *
 880  * The main write_end function for GFS2. We have a separate one for
 881  * stuffed files as they are slightly different, otherwise we just
 882  * put our locking around the VFS provided functions.
 883  *
 884  * Returns: errno
 885  */
 886
 887 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 888                           loff_t pos, unsigned len, unsigned copied,
 889                           struct page *page, void *fsdata)
 890 {
 891         struct inode *inode = page->mapping->host;
 892         struct gfs2_inode *ip = GFS2_I(inode);
 893         struct gfs2_sbd *sdp = GFS2_SB(inode);
 894         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 895         struct buffer_head *dibh;
 896         int ret;
 897         struct gfs2_trans *tr = current->journal_info;
 898         BUG_ON(!tr);
 899
 900         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 901
 902         ret = gfs2_meta_inode_buffer(ip, &dibh);
 903         if (unlikely(ret)) {
 904                 unlock_page(page);
 905                 put_page(page);
 906                 goto failed;
 907         }
 908
 909         if (gfs2_is_stuffed(ip))
 910                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 911
 912         if (!gfs2_is_writeback(ip))
 913                 gfs2_page_add_databufs(ip, page, pos & ~PAGE_MASK, len);
 914
 915         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 916         if (tr->tr_num_buf_new)
 917                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 918         else
 919                 gfs2_trans_add_meta(ip->i_gl, dibh);
 920
 921
 922         if (inode == sdp->sd_rindex) {
 923                 adjust_fs_space(inode);
 924                 sdp->sd_rindex_uptodate = 0;
 925         }
 926
 927         brelse(dibh);
 928 failed:
 929         gfs2_trans_end(sdp);
 930         gfs2_inplace_release(ip);
 931         if (ip->i_qadata && ip->i_qadata->qa_qd_num)
 932                 gfs2_quota_unlock(ip);
 933         if (inode == sdp->sd_rindex) {
 934                 gfs2_glock_dq(&m_ip->i_gh);
 935                 gfs2_holder_uninit(&m_ip->i_gh);
 936         }
 937         gfs2_glock_dq(&ip->i_gh);
 938         gfs2_holder_uninit(&ip->i_gh);
 939         return ret;
 940 }
 941
 942 /**
 943  * jdata_set_page_dirty - Page dirtying function
 944  * @page: The page to dirty
 945  *
 946  * Returns: 1 if it dirtyed the page, or 0 otherwise
 947  */
 948
 949 static int jdata_set_page_dirty(struct page *page)
 950 {
 951         SetPageChecked(page);
 952         return __set_page_dirty_buffers(page);
 953 }
 954
 955 /**
 956  * gfs2_bmap - Block map function
 957  * @mapping: Address space info
 958  * @lblock: The block to map
 959  *
 960  * Returns: The disk address for the block or 0 on hole or error
 961  */
 962
 963 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 964 {
 965         struct gfs2_inode *ip = GFS2_I(mapping->host);
 966         struct gfs2_holder i_gh;
 967         sector_t dblock = 0;
 968         int error;
 969
 970         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 971         if (error)
 972                 return 0;
 973
 974         if (!gfs2_is_stuffed(ip))
 975                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 976
 977         gfs2_glock_dq_uninit(&i_gh);
 978
 979         return dblock;
 980 }
 981
 982 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 983 {
 984         struct gfs2_bufdata *bd;
 985
 986         lock_buffer(bh);
 987         gfs2_log_lock(sdp);
 988         clear_buffer_dirty(bh);
 989         bd = bh->b_private;
 990         if (bd) {
 991                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
 992                         list_del_init(&bd->bd_list);
 993                 else
 994                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
 995         }
 996         bh->b_bdev = NULL;
 997         clear_buffer_mapped(bh);
 998         clear_buffer_req(bh);
 999         clear_buffer_new(bh);
1000         gfs2_log_unlock(sdp);
1001         unlock_buffer(bh);
1002 }
1003
1004 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
1005                                 unsigned int length)
1006 {
1007         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
1008         unsigned int stop = offset + length;
1009         int partial_page = (offset || length < PAGE_SIZE);
1010         struct buffer_head *bh, *head;
1011         unsigned long pos = 0;
1012
1013         BUG_ON(!PageLocked(page));
1014         if (!partial_page)
1015                 ClearPageChecked(page);
1016         if (!page_has_buffers(page))
1017                 goto out;
1018
1019         bh = head = page_buffers(page);
1020         do {
1021                 if (pos + bh->b_size > stop)
1022                         return;
1023
1024                 if (offset <= pos)
1025                         gfs2_discard(sdp, bh);
1026                 pos += bh->b_size;
1027                 bh = bh->b_this_page;
1028         } while (bh != head);
1029 out:
1030         if (!partial_page)
1031                 try_to_release_page(page, 0);
1032 }
1033
1034 /**
1035  * gfs2_ok_for_dio - check that dio is valid on this file
1036  * @ip: The inode
1037  * @offset: The offset at which we are reading or writing
1038  *
1039  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1040  *          1 (to accept the i/o request)
1041  */
1042 static int gfs2_ok_for_dio(struct gfs2_inode *ip, loff_t offset)
1043 {
1044         /*
1045          * Should we return an error here? I can't see that O_DIRECT for
1046          * a stuffed file makes any sense. For now we'll silently fall
1047          * back to buffered I/O
1048          */
1049         if (gfs2_is_stuffed(ip))
1050                 return 0;
1051
1052         if (offset >= i_size_read(&ip->i_inode))
1053                 return 0;
1054         return 1;
1055 }
1056
1057
1058
1059 static ssize_t gfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
1060 {
1061         struct file *file = iocb->ki_filp;
1062         struct inode *inode = file->f_mapping->host;
1063         struct address_space *mapping = inode->i_mapping;
1064         struct gfs2_inode *ip = GFS2_I(inode);
1065         loff_t offset = iocb->ki_pos;
1066         struct gfs2_holder gh;
1067         int rv;
1068
1069         /*
1070          * Deferred lock, even if its a write, since we do no allocation
1071          * on this path. All we need change is atime, and this lock mode
1072          * ensures that other nodes have flushed their buffered read caches
1073          * (i.e. their page cache entries for this inode). We do not,
1074          * unfortunately have the option of only flushing a range like
1075          * the VFS does.
1076          */
1077         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1078         rv = gfs2_glock_nq(&gh);
1079         if (rv)
1080                 goto out_uninit;
1081         rv = gfs2_ok_for_dio(ip, offset);
1082         if (rv != 1)
1083                 goto out; /* dio not valid, fall back to buffered i/o */
1084
1085         /*
1086          * Now since we are holding a deferred (CW) lock at this point, you
1087          * might be wondering why this is ever needed. There is a case however
1088          * where we've granted a deferred local lock against a cached exclusive
1089          * glock. That is ok provided all granted local locks are deferred, but
1090          * it also means that it is possible to encounter pages which are
1091          * cached and possibly also mapped. So here we check for that and sort
1092          * them out ahead of the dio. The glock state machine will take care of
1093          * everything else.
1094          *
1095          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1096          * the first place, mapping->nr_pages will always be zero.
1097          */
1098         if (mapping->nrpages) {
1099                 loff_t lstart = offset & ~(PAGE_SIZE - 1);
1100                 loff_t len = iov_iter_count(iter);
1101                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1102
1103                 rv = 0;
1104                 if (len == 0)
1105                         goto out;
1106                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1107                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1108                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1109                 if (rv)
1110                         goto out;
1111                 if (iov_iter_rw(iter) == WRITE)
1112                         truncate_inode_pages_range(mapping, lstart, end);
1113         }
1114
1115         rv = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
1116                                   gfs2_get_block_direct, NULL, NULL, 0);
1117 out:
1118         gfs2_glock_dq(&gh);
1119 out_uninit:
1120         gfs2_holder_uninit(&gh);
1121         return rv;
1122 }
1123
1124 /**
1125  * gfs2_releasepage - free the metadata associated with a page
1126  * @page: the page that's being released
1127  * @gfp_mask: passed from Linux VFS, ignored by us
1128  *
1129  * Call try_to_free_buffers() if the buffers in this page can be
1130  * released.
1131  *
1132  * Returns: 0
1133  */
1134
1135 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1136 {
1137         struct address_space *mapping = page->mapping;
1138         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1139         struct buffer_head *bh, *head;
1140         struct gfs2_bufdata *bd;
1141
1142         if (!page_has_buffers(page))
1143                 return 0;
1144
1145         /*
1146          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
1147          * clean pages might not have had the dirty bit cleared.  Thus, it can
1148          * send actual dirty pages to ->releasepage() via shrink_active_list().
1149          *
1150          * As a workaround, we skip pages that contain dirty buffers below.
1151          * Once ->releasepage isn't called on dirty pages anymore, we can warn
1152          * on dirty buffers like we used to here again.
1153          */
1154
1155         gfs2_log_lock(sdp);
1156         spin_lock(&sdp->sd_ail_lock);
1157         head = bh = page_buffers(page);
1158         do {
1159                 if (atomic_read(&bh->b_count))
1160                         goto cannot_release;
1161                 bd = bh->b_private;
1162                 if (bd && bd->bd_tr)
1163                         goto cannot_release;
1164                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
1165                         goto cannot_release;
1166                 bh = bh->b_this_page;
1167         } while(bh != head);
1168         spin_unlock(&sdp->sd_ail_lock);
1169
1170         head = bh = page_buffers(page);
1171         do {
1172                 bd = bh->b_private;
1173                 if (bd) {
1174                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1175                         if (!list_empty(&bd->bd_list))
1176                                 list_del_init(&bd->bd_list);
1177                         bd->bd_bh = NULL;
1178                         bh->b_private = NULL;
1179                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1180                 }
1181
1182                 bh = bh->b_this_page;
1183         } while (bh != head);
1184         gfs2_log_unlock(sdp);
1185
1186         return try_to_free_buffers(page);
1187
1188 cannot_release:
1189         spin_unlock(&sdp->sd_ail_lock);
1190         gfs2_log_unlock(sdp);
1191         return 0;
1192 }
1193
1194 static const struct address_space_operations gfs2_writeback_aops = {
1195         .writepage = gfs2_writepage,
1196         .writepages = gfs2_writepages,
1197         .readpage = gfs2_readpage,
1198         .readpages = gfs2_readpages,
1199         .write_begin = gfs2_write_begin,
1200         .write_end = gfs2_write_end,
1201         .bmap = gfs2_bmap,
1202         .invalidatepage = gfs2_invalidatepage,
1203         .releasepage = gfs2_releasepage,
1204         .direct_IO = gfs2_direct_IO,
1205         .migratepage = buffer_migrate_page,
1206         .is_partially_uptodate = block_is_partially_uptodate,
1207         .error_remove_page = generic_error_remove_page,
1208 };
1209
1210 static const struct address_space_operations gfs2_ordered_aops = {
1211         .writepage = gfs2_writepage,
1212         .writepages = gfs2_writepages,
1213         .readpage = gfs2_readpage,
1214         .readpages = gfs2_readpages,
1215         .write_begin = gfs2_write_begin,
1216         .write_end = gfs2_write_end,
1217         .set_page_dirty = __set_page_dirty_buffers,
1218         .bmap = gfs2_bmap,
1219         .invalidatepage = gfs2_invalidatepage,
1220         .releasepage = gfs2_releasepage,
1221         .direct_IO = gfs2_direct_IO,
1222         .migratepage = buffer_migrate_page,
1223         .is_partially_uptodate = block_is_partially_uptodate,
1224         .error_remove_page = generic_error_remove_page,
1225 };
1226
1227 static const struct address_space_operations gfs2_jdata_aops = {
1228         .writepage = gfs2_jdata_writepage,
1229         .writepages = gfs2_jdata_writepages,
1230         .readpage = gfs2_readpage,
1231         .readpages = gfs2_readpages,
1232         .write_begin = gfs2_write_begin,
1233         .write_end = gfs2_write_end,
1234         .set_page_dirty = jdata_set_page_dirty,
1235         .bmap = gfs2_bmap,
1236         .invalidatepage = gfs2_invalidatepage,
1237         .releasepage = gfs2_releasepage,
1238         .is_partially_uptodate = block_is_partially_uptodate,
1239         .error_remove_page = generic_error_remove_page,
1240 };
1241
1242 void gfs2_set_aops(struct inode *inode)
1243 {
1244         struct gfs2_inode *ip = GFS2_I(inode);
1245
1246         if (gfs2_is_writeback(ip))
1247                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1248         else if (gfs2_is_ordered(ip))
1249                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1250         else if (gfs2_is_jdata(ip))
1251                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1252         else
1253                 BUG();
1254 }
1255