fs/gfs2/aops.c

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