git.oblomov.eu Git - linux-2.6/blob - fs/gfs2/ops_address.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2007 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/gfs2_ondisk.h>
  21 #include <linux/lm_interface.h>
  22
  23 #include "gfs2.h"
  24 #include "incore.h"
  25 #include "bmap.h"
  26 #include "glock.h"
  27 #include "inode.h"
  28 #include "log.h"
  29 #include "meta_io.h"
  30 #include "ops_address.h"
  31 #include "quota.h"
  32 #include "trans.h"
  33 #include "rgrp.h"
  34 #include "ops_file.h"
  35 #include "super.h"
  36 #include "util.h"
  37 #include "glops.h"
  38
  39
  40 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  41                                    unsigned int from, unsigned int to)
  42 {
  43         struct buffer_head *head = page_buffers(page);
  44         unsigned int bsize = head->b_size;
  45         struct buffer_head *bh;
  46         unsigned int start, end;
  47
  48         for (bh = head, start = 0; bh != head || !start;
  49              bh = bh->b_this_page, start = end) {
  50                 end = start + bsize;
  51                 if (end <= from || start >= to)
  52                         continue;
  53                 if (gfs2_is_jdata(ip))
  54                         set_buffer_uptodate(bh);
  55                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  56         }
  57 }
  58
  59 /**
  60  * gfs2_get_block - Fills in a buffer head with details about a block
  61  * @inode: The inode
  62  * @lblock: The block number to look up
  63  * @bh_result: The buffer head to return the result in
  64  * @create: Non-zero if we may add block to the file
  65  *
  66  * Returns: errno
  67  */
  68
  69 int gfs2_get_block(struct inode *inode, sector_t lblock,
  70                    struct buffer_head *bh_result, int create)
  71 {
  72         return gfs2_block_map(inode, lblock, create, bh_result);
  73 }
  74
  75 /**
  76  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  77  * @inode: The inode
  78  * @lblock: The block number to look up
  79  * @bh_result: The buffer head to return the result in
  80  * @create: Non-zero if we may add block to the file
  81  *
  82  * Returns: errno
  83  */
  84
  85 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  86                                   struct buffer_head *bh_result, int create)
  87 {
  88         int error;
  89
  90         error = gfs2_block_map(inode, lblock, 0, bh_result);
  91         if (error)
  92                 return error;
  93         if (!buffer_mapped(bh_result))
  94                 return -EIO;
  95         return 0;
  96 }
  97
  98 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  99                                  struct buffer_head *bh_result, int create)
 100 {
 101         return gfs2_block_map(inode, lblock, 0, bh_result);
 102 }
 103
 104 /**
 105  * gfs2_writepage - Write complete page
 106  * @page: Page to write
 107  *
 108  * Returns: errno
 109  *
 110  * Some of this is copied from block_write_full_page() although we still
 111  * call it to do most of the work.
 112  */
 113
 114 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 115 {
 116         struct inode *inode = page->mapping->host;
 117         struct gfs2_inode *ip = GFS2_I(inode);
 118         struct gfs2_sbd *sdp = GFS2_SB(inode);
 119         loff_t i_size = i_size_read(inode);
 120         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 121         unsigned offset;
 122         int error;
 123         int done_trans = 0;
 124
 125         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl))) {
 126                 unlock_page(page);
 127                 return -EIO;
 128         }
 129         if (current->journal_info)
 130                 goto out_ignore;
 131
 132         /* Is the page fully outside i_size? (truncate in progress) */
 133         offset = i_size & (PAGE_CACHE_SIZE-1);
 134         if (page->index > end_index || (page->index == end_index && !offset)) {
 135                 page->mapping->a_ops->invalidatepage(page, 0);
 136                 unlock_page(page);
 137                 return 0; /* don't care */
 138         }
 139
 140         if ((sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip)) &&
 141             PageChecked(page)) {
 142                 ClearPageChecked(page);
 143                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 144                 if (error)
 145                         goto out_ignore;
 146                 if (!page_has_buffers(page)) {
 147                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 148                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 149                 }
 150                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 151                 done_trans = 1;
 152         }
 153         error = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 154         if (done_trans)
 155                 gfs2_trans_end(sdp);
 156         gfs2_meta_cache_flush(ip);
 157         return error;
 158
 159 out_ignore:
 160         redirty_page_for_writepage(wbc, page);
 161         unlock_page(page);
 162         return 0;
 163 }
 164
 165 /**
 166  * gfs2_writepages - Write a bunch of dirty pages back to disk
 167  * @mapping: The mapping to write
 168  * @wbc: Write-back control
 169  *
 170  * For journaled files and/or ordered writes this just falls back to the
 171  * kernel's default writepages path for now. We will probably want to change
 172  * that eventually (i.e. when we look at allocate on flush).
 173  *
 174  * For the data=writeback case though we can already ignore buffer heads
 175  * and write whole extents at once. This is a big reduction in the
 176  * number of I/O requests we send and the bmap calls we make in this case.
 177  */
 178 static int gfs2_writepages(struct address_space *mapping,
 179                            struct writeback_control *wbc)
 180 {
 181         struct inode *inode = mapping->host;
 182         struct gfs2_inode *ip = GFS2_I(inode);
 183         struct gfs2_sbd *sdp = GFS2_SB(inode);
 184
 185         if (sdp->sd_args.ar_data == GFS2_DATA_WRITEBACK && !gfs2_is_jdata(ip))
 186                 return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 187
 188         return generic_writepages(mapping, wbc);
 189 }
 190
 191 /**
 192  * stuffed_readpage - Fill in a Linux page with stuffed file data
 193  * @ip: the inode
 194  * @page: the page
 195  *
 196  * Returns: errno
 197  */
 198
 199 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 200 {
 201         struct buffer_head *dibh;
 202         void *kaddr;
 203         int error;
 204
 205         /*
 206          * Due to the order of unstuffing files and ->nopage(), we can be
 207          * asked for a zero page in the case of a stuffed file being extended,
 208          * so we need to supply one here. It doesn't happen often.
 209          */
 210         if (unlikely(page->index)) {
 211                 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
 212                 return 0;
 213         }
 214
 215         error = gfs2_meta_inode_buffer(ip, &dibh);
 216         if (error)
 217                 return error;
 218
 219         kaddr = kmap_atomic(page, KM_USER0);
 220         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 221                ip->i_di.di_size);
 222         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
 223         kunmap_atomic(kaddr, KM_USER0);
 224         flush_dcache_page(page);
 225         brelse(dibh);
 226         SetPageUptodate(page);
 227
 228         return 0;
 229 }
 230
 231
 232 /**
 233  * gfs2_readpage - readpage with locking
 234  * @file: The file to read a page for. N.B. This may be NULL if we are
 235  * reading an internal file.
 236  * @page: The page to read
 237  *
 238  * Returns: errno
 239  */
 240
 241 static int gfs2_readpage(struct file *file, struct page *page)
 242 {
 243         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 244         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 245         struct gfs2_file *gf = NULL;
 246         struct gfs2_holder gh;
 247         int error;
 248         int do_unlock = 0;
 249
 250         if (likely(file != &gfs2_internal_file_sentinel)) {
 251                 if (file) {
 252                         gf = file->private_data;
 253                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 254                                 /* gfs2_sharewrite_fault has grabbed the ip->i_gl already */
 255                                 goto skip_lock;
 256                 }
 257                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
 258                 do_unlock = 1;
 259                 error = gfs2_glock_nq_atime(&gh);
 260                 if (unlikely(error))
 261                         goto out_unlock;
 262         }
 263
 264 skip_lock:
 265         if (gfs2_is_stuffed(ip)) {
 266                 error = stuffed_readpage(ip, page);
 267                 unlock_page(page);
 268         } else
 269                 error = mpage_readpage(page, gfs2_get_block);
 270
 271         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 272                 error = -EIO;
 273
 274         if (do_unlock) {
 275                 gfs2_glock_dq_m(1, &gh);
 276                 gfs2_holder_uninit(&gh);
 277         }
 278 out:
 279         return error;
 280 out_unlock:
 281         unlock_page(page);
 282         if (error == GLR_TRYFAILED) {
 283                 error = AOP_TRUNCATED_PAGE;
 284                 yield();
 285         }
 286         if (do_unlock)
 287                 gfs2_holder_uninit(&gh);
 288         goto out;
 289 }
 290
 291 /**
 292  * gfs2_readpages - Read a bunch of pages at once
 293  *
 294  * Some notes:
 295  * 1. This is only for readahead, so we can simply ignore any things
 296  *    which are slightly inconvenient (such as locking conflicts between
 297  *    the page lock and the glock) and return having done no I/O. Its
 298  *    obviously not something we'd want to do on too regular a basis.
 299  *    Any I/O we ignore at this time will be done via readpage later.
 300  * 2. We don't handle stuffed files here we let readpage do the honours.
 301  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 302  * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
 303  * 5. We use LM_FLAG_TRY_1CB here, effectively we then have lock-ahead as
 304  *    well as read-ahead.
 305  */
 306 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 307                           struct list_head *pages, unsigned nr_pages)
 308 {
 309         struct inode *inode = mapping->host;
 310         struct gfs2_inode *ip = GFS2_I(inode);
 311         struct gfs2_sbd *sdp = GFS2_SB(inode);
 312         struct gfs2_holder gh;
 313         int ret = 0;
 314         int do_unlock = 0;
 315
 316         if (likely(file != &gfs2_internal_file_sentinel)) {
 317                 if (file) {
 318                         struct gfs2_file *gf = file->private_data;
 319                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 320                                 goto skip_lock;
 321                 }
 322                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED,
 323                                  LM_FLAG_TRY_1CB|GL_ATIME, &gh);
 324                 do_unlock = 1;
 325                 ret = gfs2_glock_nq_atime(&gh);
 326                 if (ret == GLR_TRYFAILED)
 327                         goto out_noerror;
 328                 if (unlikely(ret))
 329                         goto out_unlock;
 330         }
 331 skip_lock:
 332         if (!gfs2_is_stuffed(ip))
 333                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
 334
 335         if (do_unlock) {
 336                 gfs2_glock_dq_m(1, &gh);
 337                 gfs2_holder_uninit(&gh);
 338         }
 339 out:
 340         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 341                 ret = -EIO;
 342         return ret;
 343 out_noerror:
 344         ret = 0;
 345 out_unlock:
 346         if (do_unlock)
 347                 gfs2_holder_uninit(&gh);
 348         goto out;
 349 }
 350
 351 /**
 352  * gfs2_prepare_write - Prepare to write a page to a file
 353  * @file: The file to write to
 354  * @page: The page which is to be prepared for writing
 355  * @from: From (byte range within page)
 356  * @to: To (byte range within page)
 357  *
 358  * Returns: errno
 359  */
 360
 361 static int gfs2_prepare_write(struct file *file, struct page *page,
 362                               unsigned from, unsigned to)
 363 {
 364         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 365         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 366         unsigned int data_blocks, ind_blocks, rblocks;
 367         int alloc_required;
 368         int error = 0;
 369         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + from;
 370         loff_t end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
 371         struct gfs2_alloc *al;
 372         unsigned int write_len = to - from;
 373
 374
 375         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME|LM_FLAG_TRY_1CB, &ip->i_gh);
 376         error = gfs2_glock_nq_atime(&ip->i_gh);
 377         if (unlikely(error)) {
 378                 if (error == GLR_TRYFAILED) {
 379                         unlock_page(page);
 380                         error = AOP_TRUNCATED_PAGE;
 381                         yield();
 382                 }
 383                 goto out_uninit;
 384         }
 385
 386         gfs2_write_calc_reserv(ip, write_len, &data_blocks, &ind_blocks);
 387
 388         error = gfs2_write_alloc_required(ip, pos, write_len, &alloc_required);
 389         if (error)
 390                 goto out_unlock;
 391
 392
 393         ip->i_alloc.al_requested = 0;
 394         if (alloc_required) {
 395                 al = gfs2_alloc_get(ip);
 396
 397                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 398                 if (error)
 399                         goto out_alloc_put;
 400
 401                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
 402                 if (error)
 403                         goto out_qunlock;
 404
 405                 al->al_requested = data_blocks + ind_blocks;
 406                 error = gfs2_inplace_reserve(ip);
 407                 if (error)
 408                         goto out_qunlock;
 409         }
 410
 411         rblocks = RES_DINODE + ind_blocks;
 412         if (gfs2_is_jdata(ip))
 413                 rblocks += data_blocks ? data_blocks : 1;
 414         if (ind_blocks || data_blocks)
 415                 rblocks += RES_STATFS + RES_QUOTA;
 416
 417         error = gfs2_trans_begin(sdp, rblocks,
 418                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 419         if (error)
 420                 goto out_trans_fail;
 421
 422         if (gfs2_is_stuffed(ip)) {
 423                 if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 424                         error = gfs2_unstuff_dinode(ip, page);
 425                         if (error == 0)
 426                                 goto prepare_write;
 427                 } else if (!PageUptodate(page))
 428                         error = stuffed_readpage(ip, page);
 429                 goto out;
 430         }
 431
 432 prepare_write:
 433         error = block_prepare_write(page, from, to, gfs2_get_block);
 434
 435 out:
 436         if (error) {
 437                 gfs2_trans_end(sdp);
 438 out_trans_fail:
 439                 if (alloc_required) {
 440                         gfs2_inplace_release(ip);
 441 out_qunlock:
 442                         gfs2_quota_unlock(ip);
 443 out_alloc_put:
 444                         gfs2_alloc_put(ip);
 445                 }
 446 out_unlock:
 447                 gfs2_glock_dq_m(1, &ip->i_gh);
 448 out_uninit:
 449                 gfs2_holder_uninit(&ip->i_gh);
 450         }
 451
 452         return error;
 453 }
 454
 455 /**
 456  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 457  * @inode: the rindex inode
 458  */
 459 static void adjust_fs_space(struct inode *inode)
 460 {
 461         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 462         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 463         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 464         u64 fs_total, new_free;
 465
 466         /* Total up the file system space, according to the latest rindex. */
 467         fs_total = gfs2_ri_total(sdp);
 468
 469         spin_lock(&sdp->sd_statfs_spin);
 470         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 471                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 472         else
 473                 new_free = 0;
 474         spin_unlock(&sdp->sd_statfs_spin);
 475         fs_warn(sdp, "File system extended by %llu blocks.\n",
 476                 (unsigned long long)new_free);
 477         gfs2_statfs_change(sdp, new_free, new_free, 0);
 478 }
 479
 480 /**
 481  * gfs2_commit_write - Commit write to a file
 482  * @file: The file to write to
 483  * @page: The page containing the data
 484  * @from: From (byte range within page)
 485  * @to: To (byte range within page)
 486  *
 487  * Returns: errno
 488  */
 489
 490 static int gfs2_commit_write(struct file *file, struct page *page,
 491                              unsigned from, unsigned to)
 492 {
 493         struct inode *inode = page->mapping->host;
 494         struct gfs2_inode *ip = GFS2_I(inode);
 495         struct gfs2_sbd *sdp = GFS2_SB(inode);
 496         int error = -EOPNOTSUPP;
 497         struct buffer_head *dibh;
 498         struct gfs2_alloc *al = &ip->i_alloc;
 499         struct gfs2_dinode *di;
 500
 501         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
 502                 goto fail_nounlock;
 503
 504         error = gfs2_meta_inode_buffer(ip, &dibh);
 505         if (error)
 506                 goto fail_endtrans;
 507
 508         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 509         di = (struct gfs2_dinode *)dibh->b_data;
 510
 511         if (gfs2_is_stuffed(ip)) {
 512                 u64 file_size;
 513                 void *kaddr;
 514
 515                 file_size = ((u64)page->index << PAGE_CACHE_SHIFT) + to;
 516
 517                 kaddr = kmap_atomic(page, KM_USER0);
 518                 memcpy(dibh->b_data + sizeof(struct gfs2_dinode) + from,
 519                        kaddr + from, to - from);
 520                 kunmap_atomic(kaddr, KM_USER0);
 521
 522                 SetPageUptodate(page);
 523
 524                 if (inode->i_size < file_size) {
 525                         i_size_write(inode, file_size);
 526                         mark_inode_dirty(inode);
 527                 }
 528         } else {
 529                 if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED ||
 530                     gfs2_is_jdata(ip))
 531                         gfs2_page_add_databufs(ip, page, from, to);
 532                 error = generic_commit_write(file, page, from, to);
 533                 if (error)
 534                         goto fail;
 535         }
 536
 537         if (ip->i_di.di_size < inode->i_size) {
 538                 ip->i_di.di_size = inode->i_size;
 539                 di->di_size = cpu_to_be64(inode->i_size);
 540         }
 541
 542         if (inode == sdp->sd_rindex)
 543                 adjust_fs_space(inode);
 544
 545         brelse(dibh);
 546         gfs2_trans_end(sdp);
 547         if (al->al_requested) {
 548                 gfs2_inplace_release(ip);
 549                 gfs2_quota_unlock(ip);
 550                 gfs2_alloc_put(ip);
 551         }
 552         unlock_page(page);
 553         gfs2_glock_dq_m(1, &ip->i_gh);
 554         lock_page(page);
 555         gfs2_holder_uninit(&ip->i_gh);
 556         return 0;
 557
 558 fail:
 559         brelse(dibh);
 560 fail_endtrans:
 561         gfs2_trans_end(sdp);
 562         if (al->al_requested) {
 563                 gfs2_inplace_release(ip);
 564                 gfs2_quota_unlock(ip);
 565                 gfs2_alloc_put(ip);
 566         }
 567         unlock_page(page);
 568         gfs2_glock_dq_m(1, &ip->i_gh);
 569         lock_page(page);
 570         gfs2_holder_uninit(&ip->i_gh);
 571 fail_nounlock:
 572         ClearPageUptodate(page);
 573         return error;
 574 }
 575
 576 /**
 577  * gfs2_set_page_dirty - Page dirtying function
 578  * @page: The page to dirty
 579  *
 580  * Returns: 1 if it dirtyed the page, or 0 otherwise
 581  */
 582
 583 static int gfs2_set_page_dirty(struct page *page)
 584 {
 585         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 586         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 587
 588         if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip))
 589                 SetPageChecked(page);
 590         return __set_page_dirty_buffers(page);
 591 }
 592
 593 /**
 594  * gfs2_bmap - Block map function
 595  * @mapping: Address space info
 596  * @lblock: The block to map
 597  *
 598  * Returns: The disk address for the block or 0 on hole or error
 599  */
 600
 601 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 602 {
 603         struct gfs2_inode *ip = GFS2_I(mapping->host);
 604         struct gfs2_holder i_gh;
 605         sector_t dblock = 0;
 606         int error;
 607
 608         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 609         if (error)
 610                 return 0;
 611
 612         if (!gfs2_is_stuffed(ip))
 613                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 614
 615         gfs2_glock_dq_uninit(&i_gh);
 616
 617         return dblock;
 618 }
 619
 620 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 621 {
 622         struct gfs2_bufdata *bd;
 623
 624         lock_buffer(bh);
 625         gfs2_log_lock(sdp);
 626         clear_buffer_dirty(bh);
 627         bd = bh->b_private;
 628         if (bd) {
 629                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 630                         list_del_init(&bd->bd_le.le_list);
 631                 else
 632                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 633         }
 634         bh->b_bdev = NULL;
 635         clear_buffer_mapped(bh);
 636         clear_buffer_req(bh);
 637         clear_buffer_new(bh);
 638         gfs2_log_unlock(sdp);
 639         unlock_buffer(bh);
 640 }
 641
 642 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 643 {
 644         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 645         struct buffer_head *bh, *head;
 646         unsigned long pos = 0;
 647
 648         BUG_ON(!PageLocked(page));
 649         if (offset == 0)
 650                 ClearPageChecked(page);
 651         if (!page_has_buffers(page))
 652                 goto out;
 653
 654         bh = head = page_buffers(page);
 655         do {
 656                 if (offset <= pos)
 657                         gfs2_discard(sdp, bh);
 658                 pos += bh->b_size;
 659                 bh = bh->b_this_page;
 660         } while (bh != head);
 661 out:
 662         if (offset == 0)
 663                 try_to_release_page(page, 0);
 664 }
 665
 666 /**
 667  * gfs2_ok_for_dio - check that dio is valid on this file
 668  * @ip: The inode
 669  * @rw: READ or WRITE
 670  * @offset: The offset at which we are reading or writing
 671  *
 672  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 673  *          1 (to accept the i/o request)
 674  */
 675 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 676 {
 677         /*
 678          * Should we return an error here? I can't see that O_DIRECT for
 679          * a journaled file makes any sense. For now we'll silently fall
 680          * back to buffered I/O, likewise we do the same for stuffed
 681          * files since they are (a) small and (b) unaligned.
 682          */
 683         if (gfs2_is_jdata(ip))
 684                 return 0;
 685
 686         if (gfs2_is_stuffed(ip))
 687                 return 0;
 688
 689         if (offset > i_size_read(&ip->i_inode))
 690                 return 0;
 691         return 1;
 692 }
 693
 694
 695
 696 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 697                               const struct iovec *iov, loff_t offset,
 698                               unsigned long nr_segs)
 699 {
 700         struct file *file = iocb->ki_filp;
 701         struct inode *inode = file->f_mapping->host;
 702         struct gfs2_inode *ip = GFS2_I(inode);
 703         struct gfs2_holder gh;
 704         int rv;
 705
 706         /*
 707          * Deferred lock, even if its a write, since we do no allocation
 708          * on this path. All we need change is atime, and this lock mode
 709          * ensures that other nodes have flushed their buffered read caches
 710          * (i.e. their page cache entries for this inode). We do not,
 711          * unfortunately have the option of only flushing a range like
 712          * the VFS does.
 713          */
 714         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 715         rv = gfs2_glock_nq_atime(&gh);
 716         if (rv)
 717                 return rv;
 718         rv = gfs2_ok_for_dio(ip, rw, offset);
 719         if (rv != 1)
 720                 goto out; /* dio not valid, fall back to buffered i/o */
 721
 722         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 723                                            iov, offset, nr_segs,
 724                                            gfs2_get_block_direct, NULL);
 725 out:
 726         gfs2_glock_dq_m(1, &gh);
 727         gfs2_holder_uninit(&gh);
 728         return rv;
 729 }
 730
 731 /**
 732  * gfs2_releasepage - free the metadata associated with a page
 733  * @page: the page that's being released
 734  * @gfp_mask: passed from Linux VFS, ignored by us
 735  *
 736  * Call try_to_free_buffers() if the buffers in this page can be
 737  * released.
 738  *
 739  * Returns: 0
 740  */
 741
 742 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 743 {
 744         struct inode *aspace = page->mapping->host;
 745         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 746         struct buffer_head *bh, *head;
 747         struct gfs2_bufdata *bd;
 748
 749         if (!page_has_buffers(page))
 750                 return 0;
 751
 752         gfs2_log_lock(sdp);
 753         head = bh = page_buffers(page);
 754         do {
 755                 if (atomic_read(&bh->b_count))
 756                         goto cannot_release;
 757                 bd = bh->b_private;
 758                 if (bd && bd->bd_ail)
 759                         goto cannot_release;
 760                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 761                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 762                 bh = bh->b_this_page;
 763         } while(bh != head);
 764         gfs2_log_unlock(sdp);
 765
 766         head = bh = page_buffers(page);
 767         do {
 768                 gfs2_log_lock(sdp);
 769                 bd = bh->b_private;
 770                 if (bd) {
 771                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 772                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 773                         if (!list_empty(&bd->bd_le.le_list)) {
 774                                 if (!buffer_pinned(bh))
 775                                         list_del_init(&bd->bd_le.le_list);
 776                                 else
 777                                         bd = NULL;
 778                         }
 779                         if (bd)
 780                                 bd->bd_bh = NULL;
 781                         bh->b_private = NULL;
 782                 }
 783                 gfs2_log_unlock(sdp);
 784                 if (bd)
 785                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 786
 787                 bh = bh->b_this_page;
 788         } while (bh != head);
 789
 790         return try_to_free_buffers(page);
 791 cannot_release:
 792         gfs2_log_unlock(sdp);
 793         return 0;
 794 }
 795
 796 const struct address_space_operations gfs2_file_aops = {
 797         .writepage = gfs2_writepage,
 798         .writepages = gfs2_writepages,
 799         .readpage = gfs2_readpage,
 800         .readpages = gfs2_readpages,
 801         .sync_page = block_sync_page,
 802         .prepare_write = gfs2_prepare_write,
 803         .commit_write = gfs2_commit_write,
 804         .set_page_dirty = gfs2_set_page_dirty,
 805         .bmap = gfs2_bmap,
 806         .invalidatepage = gfs2_invalidatepage,
 807         .releasepage = gfs2_releasepage,
 808         .direct_IO = gfs2_direct_IO,
 809 };
 810