err.no 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 (bh_result->b_blocknr == 0)
  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                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 142                 if (error)
 143                         goto out_ignore;
 144                 if (!page_has_buffers(page)) {
 145                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 146                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 147                 }
 148                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 149                 done_trans = 1;
 150         }
 151         error = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 152         if (done_trans)
 153                 gfs2_trans_end(sdp);
 154         gfs2_meta_cache_flush(ip);
 155         return error;
 156
 157 out_ignore:
 158         redirty_page_for_writepage(wbc, page);
 159         unlock_page(page);
 160         return 0;
 161 }
 162
 163 /**
 164  * gfs2_writepages - Write a bunch of dirty pages back to disk
 165  * @mapping: The mapping to write
 166  * @wbc: Write-back control
 167  *
 168  * For journaled files and/or ordered writes this just falls back to the
 169  * kernel's default writepages path for now. We will probably want to change
 170  * that eventually (i.e. when we look at allocate on flush).
 171  *
 172  * For the data=writeback case though we can already ignore buffer heads
 173  * and write whole extents at once. This is a big reduction in the
 174  * number of I/O requests we send and the bmap calls we make in this case.
 175  */
 176 static int gfs2_writepages(struct address_space *mapping,
 177                            struct writeback_control *wbc)
 178 {
 179         struct inode *inode = mapping->host;
 180         struct gfs2_inode *ip = GFS2_I(inode);
 181         struct gfs2_sbd *sdp = GFS2_SB(inode);
 182
 183         if (sdp->sd_args.ar_data == GFS2_DATA_WRITEBACK && !gfs2_is_jdata(ip))
 184                 return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 185
 186         return generic_writepages(mapping, wbc);
 187 }
 188
 189 /**
 190  * stuffed_readpage - Fill in a Linux page with stuffed file data
 191  * @ip: the inode
 192  * @page: the page
 193  *
 194  * Returns: errno
 195  */
 196
 197 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 198 {
 199         struct buffer_head *dibh;
 200         void *kaddr;
 201         int error;
 202
 203         /*
 204          * Due to the order of unstuffing files and ->nopage(), we can be
 205          * asked for a zero page in the case of a stuffed file being extended,
 206          * so we need to supply one here. It doesn't happen often.
 207          */
 208         if (unlikely(page->index)) {
 209                 kaddr = kmap_atomic(page, KM_USER0);
 210                 memset(kaddr, 0, PAGE_CACHE_SIZE);
 211                 kunmap_atomic(kaddr, KM_USER0);
 212                 flush_dcache_page(page);
 213                 SetPageUptodate(page);
 214                 return 0;
 215         }
 216
 217         error = gfs2_meta_inode_buffer(ip, &dibh);
 218         if (error)
 219                 return error;
 220
 221         kaddr = kmap_atomic(page, KM_USER0);
 222         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 223                ip->i_di.di_size);
 224         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
 225         kunmap_atomic(kaddr, KM_USER0);
 226         flush_dcache_page(page);
 227         brelse(dibh);
 228         SetPageUptodate(page);
 229
 230         return 0;
 231 }
 232
 233
 234 /**
 235  * gfs2_readpage - readpage with locking
 236  * @file: The file to read a page for. N.B. This may be NULL if we are
 237  * reading an internal file.
 238  * @page: The page to read
 239  *
 240  * Returns: errno
 241  */
 242
 243 static int gfs2_readpage(struct file *file, struct page *page)
 244 {
 245         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 246         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 247         struct gfs2_file *gf = NULL;
 248         struct gfs2_holder gh;
 249         int error;
 250         int do_unlock = 0;
 251
 252         if (likely(file != &gfs2_internal_file_sentinel)) {
 253                 if (file) {
 254                         gf = file->private_data;
 255                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 256                                 /* gfs2_sharewrite_nopage has grabbed the ip->i_gl already */
 257                                 goto skip_lock;
 258                 }
 259                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
 260                 do_unlock = 1;
 261                 error = gfs2_glock_nq_atime(&gh);
 262                 if (unlikely(error))
 263                         goto out_unlock;
 264         }
 265
 266 skip_lock:
 267         if (gfs2_is_stuffed(ip)) {
 268                 error = stuffed_readpage(ip, page);
 269                 unlock_page(page);
 270         } else
 271                 error = mpage_readpage(page, gfs2_get_block);
 272
 273         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 274                 error = -EIO;
 275
 276         if (do_unlock) {
 277                 gfs2_glock_dq_m(1, &gh);
 278                 gfs2_holder_uninit(&gh);
 279         }
 280 out:
 281         return error;
 282 out_unlock:
 283         unlock_page(page);
 284         if (error == GLR_TRYFAILED) {
 285                 error = AOP_TRUNCATED_PAGE;
 286                 yield();
 287         }
 288         if (do_unlock)
 289                 gfs2_holder_uninit(&gh);
 290         goto out;
 291 }
 292
 293 /**
 294  * gfs2_readpages - Read a bunch of pages at once
 295  *
 296  * Some notes:
 297  * 1. This is only for readahead, so we can simply ignore any things
 298  *    which are slightly inconvenient (such as locking conflicts between
 299  *    the page lock and the glock) and return having done no I/O. Its
 300  *    obviously not something we'd want to do on too regular a basis.
 301  *    Any I/O we ignore at this time will be done via readpage later.
 302  * 2. We don't handle stuffed files here we let readpage do the honours.
 303  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 304  * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
 305  * 5. We use LM_FLAG_TRY_1CB here, effectively we then have lock-ahead as
 306  *    well as read-ahead.
 307  */
 308 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 309                           struct list_head *pages, unsigned nr_pages)
 310 {
 311         struct inode *inode = mapping->host;
 312         struct gfs2_inode *ip = GFS2_I(inode);
 313         struct gfs2_sbd *sdp = GFS2_SB(inode);
 314         struct gfs2_holder gh;
 315         int ret = 0;
 316         int do_unlock = 0;
 317
 318         if (likely(file != &gfs2_internal_file_sentinel)) {
 319                 if (file) {
 320                         struct gfs2_file *gf = file->private_data;
 321                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 322                                 goto skip_lock;
 323                 }
 324                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED,
 325                                  LM_FLAG_TRY_1CB|GL_ATIME, &gh);
 326                 do_unlock = 1;
 327                 ret = gfs2_glock_nq_atime(&gh);
 328                 if (ret == GLR_TRYFAILED)
 329                         goto out_noerror;
 330                 if (unlikely(ret))
 331                         goto out_unlock;
 332         }
 333 skip_lock:
 334         if (!gfs2_is_stuffed(ip))
 335                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
 336
 337         if (do_unlock) {
 338                 gfs2_glock_dq_m(1, &gh);
 339                 gfs2_holder_uninit(&gh);
 340         }
 341 out:
 342         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 343                 ret = -EIO;
 344         return ret;
 345 out_noerror:
 346         ret = 0;
 347 out_unlock:
 348         if (do_unlock)
 349                 gfs2_holder_uninit(&gh);
 350         goto out;
 351 }
 352
 353 /**
 354  * gfs2_prepare_write - Prepare to write a page to a file
 355  * @file: The file to write to
 356  * @page: The page which is to be prepared for writing
 357  * @from: From (byte range within page)
 358  * @to: To (byte range within page)
 359  *
 360  * Returns: errno
 361  */
 362
 363 static int gfs2_prepare_write(struct file *file, struct page *page,
 364                               unsigned from, unsigned to)
 365 {
 366         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 367         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 368         unsigned int data_blocks, ind_blocks, rblocks;
 369         int alloc_required;
 370         int error = 0;
 371         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + from;
 372         loff_t end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
 373         struct gfs2_alloc *al;
 374         unsigned int write_len = to - from;
 375
 376
 377         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME|LM_FLAG_TRY_1CB, &ip->i_gh);
 378         error = gfs2_glock_nq_atime(&ip->i_gh);
 379         if (unlikely(error)) {
 380                 if (error == GLR_TRYFAILED) {
 381                         unlock_page(page);
 382                         error = AOP_TRUNCATED_PAGE;
 383                         yield();
 384                 }
 385                 goto out_uninit;
 386         }
 387
 388         gfs2_write_calc_reserv(ip, write_len, &data_blocks, &ind_blocks);
 389
 390         error = gfs2_write_alloc_required(ip, pos, write_len, &alloc_required);
 391         if (error)
 392                 goto out_unlock;
 393
 394
 395         ip->i_alloc.al_requested = 0;
 396         if (alloc_required) {
 397                 al = gfs2_alloc_get(ip);
 398
 399                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 400                 if (error)
 401                         goto out_alloc_put;
 402
 403                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
 404                 if (error)
 405                         goto out_qunlock;
 406
 407                 al->al_requested = data_blocks + ind_blocks;
 408                 error = gfs2_inplace_reserve(ip);
 409                 if (error)
 410                         goto out_qunlock;
 411         }
 412
 413         rblocks = RES_DINODE + ind_blocks;
 414         if (gfs2_is_jdata(ip))
 415                 rblocks += data_blocks ? data_blocks : 1;
 416         if (ind_blocks || data_blocks)
 417                 rblocks += RES_STATFS + RES_QUOTA;
 418
 419         error = gfs2_trans_begin(sdp, rblocks, 0);
 420         if (error)
 421                 goto out;
 422
 423         if (gfs2_is_stuffed(ip)) {
 424                 if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 425                         error = gfs2_unstuff_dinode(ip, page);
 426                         if (error == 0)
 427                                 goto prepare_write;
 428                 } else if (!PageUptodate(page))
 429                         error = stuffed_readpage(ip, page);
 430                 goto out;
 431         }
 432
 433 prepare_write:
 434         error = block_prepare_write(page, from, to, gfs2_get_block);
 435
 436 out:
 437         if (error) {
 438                 gfs2_trans_end(sdp);
 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_bmap - Block map function
 578  * @mapping: Address space info
 579  * @lblock: The block to map
 580  *
 581  * Returns: The disk address for the block or 0 on hole or error
 582  */
 583
 584 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 585 {
 586         struct gfs2_inode *ip = GFS2_I(mapping->host);
 587         struct gfs2_holder i_gh;
 588         sector_t dblock = 0;
 589         int error;
 590
 591         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 592         if (error)
 593                 return 0;
 594
 595         if (!gfs2_is_stuffed(ip))
 596                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 597
 598         gfs2_glock_dq_uninit(&i_gh);
 599
 600         return dblock;
 601 }
 602
 603 static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
 604 {
 605         struct gfs2_bufdata *bd;
 606
 607         gfs2_log_lock(sdp);
 608         bd = bh->b_private;
 609         if (bd) {
 610                 bd->bd_bh = NULL;
 611                 bh->b_private = NULL;
 612         }
 613         gfs2_log_unlock(sdp);
 614
 615         lock_buffer(bh);
 616         clear_buffer_dirty(bh);
 617         bh->b_bdev = NULL;
 618         clear_buffer_mapped(bh);
 619         clear_buffer_req(bh);
 620         clear_buffer_new(bh);
 621         clear_buffer_delay(bh);
 622         unlock_buffer(bh);
 623 }
 624
 625 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 626 {
 627         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 628         struct buffer_head *head, *bh, *next;
 629         unsigned int curr_off = 0;
 630
 631         BUG_ON(!PageLocked(page));
 632         if (!page_has_buffers(page))
 633                 return;
 634
 635         bh = head = page_buffers(page);
 636         do {
 637                 unsigned int next_off = curr_off + bh->b_size;
 638                 next = bh->b_this_page;
 639
 640                 if (offset <= curr_off)
 641                         discard_buffer(sdp, bh);
 642
 643                 curr_off = next_off;
 644                 bh = next;
 645         } while (bh != head);
 646
 647         if (!offset)
 648                 try_to_release_page(page, 0);
 649
 650         return;
 651 }
 652
 653 /**
 654  * gfs2_ok_for_dio - check that dio is valid on this file
 655  * @ip: The inode
 656  * @rw: READ or WRITE
 657  * @offset: The offset at which we are reading or writing
 658  *
 659  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 660  *          1 (to accept the i/o request)
 661  */
 662 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 663 {
 664         /*
 665          * Should we return an error here? I can't see that O_DIRECT for
 666          * a journaled file makes any sense. For now we'll silently fall
 667          * back to buffered I/O, likewise we do the same for stuffed
 668          * files since they are (a) small and (b) unaligned.
 669          */
 670         if (gfs2_is_jdata(ip))
 671                 return 0;
 672
 673         if (gfs2_is_stuffed(ip))
 674                 return 0;
 675
 676         if (offset > i_size_read(&ip->i_inode))
 677                 return 0;
 678         return 1;
 679 }
 680
 681
 682
 683 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 684                               const struct iovec *iov, loff_t offset,
 685                               unsigned long nr_segs)
 686 {
 687         struct file *file = iocb->ki_filp;
 688         struct inode *inode = file->f_mapping->host;
 689         struct gfs2_inode *ip = GFS2_I(inode);
 690         struct gfs2_holder gh;
 691         int rv;
 692
 693         /*
 694          * Deferred lock, even if its a write, since we do no allocation
 695          * on this path. All we need change is atime, and this lock mode
 696          * ensures that other nodes have flushed their buffered read caches
 697          * (i.e. their page cache entries for this inode). We do not,
 698          * unfortunately have the option of only flushing a range like
 699          * the VFS does.
 700          */
 701         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 702         rv = gfs2_glock_nq_atime(&gh);
 703         if (rv)
 704                 return rv;
 705         rv = gfs2_ok_for_dio(ip, rw, offset);
 706         if (rv != 1)
 707                 goto out; /* dio not valid, fall back to buffered i/o */
 708
 709         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 710                                            iov, offset, nr_segs,
 711                                            gfs2_get_block_direct, NULL);
 712 out:
 713         gfs2_glock_dq_m(1, &gh);
 714         gfs2_holder_uninit(&gh);
 715         return rv;
 716 }
 717
 718 /**
 719  * stuck_releasepage - We're stuck in gfs2_releasepage().  Print stuff out.
 720  * @bh: the buffer we're stuck on
 721  *
 722  */
 723
 724 static void stuck_releasepage(struct buffer_head *bh)
 725 {
 726         struct inode *inode = bh->b_page->mapping->host;
 727         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 728         struct gfs2_bufdata *bd = bh->b_private;
 729         struct gfs2_glock *gl;
 730 static unsigned limit = 0;
 731
 732         if (limit > 3)
 733                 return;
 734         limit++;
 735
 736         fs_warn(sdp, "stuck in gfs2_releasepage() %p\n", inode);
 737         fs_warn(sdp, "blkno = %llu, bh->b_count = %d\n",
 738                 (unsigned long long)bh->b_blocknr, atomic_read(&bh->b_count));
 739         fs_warn(sdp, "pinned = %u\n", buffer_pinned(bh));
 740         fs_warn(sdp, "bh->b_private = %s\n", (bd) ? "!NULL" : "NULL");
 741
 742         if (!bd)
 743                 return;
 744
 745         gl = bd->bd_gl;
 746
 747         fs_warn(sdp, "gl = (%u, %llu)\n",
 748                 gl->gl_name.ln_type, (unsigned long long)gl->gl_name.ln_number);
 749
 750         fs_warn(sdp, "bd_list_tr = %s, bd_le.le_list = %s\n",
 751                 (list_empty(&bd->bd_list_tr)) ? "no" : "yes",
 752                 (list_empty(&bd->bd_le.le_list)) ? "no" : "yes");
 753
 754         if (gl->gl_ops == &gfs2_inode_glops) {
 755                 struct gfs2_inode *ip = gl->gl_object;
 756                 unsigned int x;
 757
 758                 if (!ip)
 759                         return;
 760
 761                 fs_warn(sdp, "ip = %llu %llu\n",
 762                         (unsigned long long)ip->i_no_formal_ino,
 763                         (unsigned long long)ip->i_no_addr);
 764
 765                 for (x = 0; x < GFS2_MAX_META_HEIGHT; x++)
 766                         fs_warn(sdp, "ip->i_cache[%u] = %s\n",
 767                                 x, (ip->i_cache[x]) ? "!NULL" : "NULL");
 768         }
 769 }
 770
 771 /**
 772  * gfs2_releasepage - free the metadata associated with a page
 773  * @page: the page that's being released
 774  * @gfp_mask: passed from Linux VFS, ignored by us
 775  *
 776  * Call try_to_free_buffers() if the buffers in this page can be
 777  * released.
 778  *
 779  * Returns: 0
 780  */
 781
 782 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 783 {
 784         struct inode *aspace = page->mapping->host;
 785         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 786         struct buffer_head *bh, *head;
 787         struct gfs2_bufdata *bd;
 788         unsigned long t = jiffies + gfs2_tune_get(sdp, gt_stall_secs) * HZ;
 789
 790         if (!page_has_buffers(page))
 791                 goto out;
 792
 793         head = bh = page_buffers(page);
 794         do {
 795                 while (atomic_read(&bh->b_count)) {
 796                         if (!atomic_read(&aspace->i_writecount))
 797                                 return 0;
 798
 799                         if (!(gfp_mask & __GFP_WAIT))
 800                                 return 0;
 801
 802                         if (time_after_eq(jiffies, t)) {
 803                                 stuck_releasepage(bh);
 804                                 /* should we withdraw here? */
 805                                 return 0;
 806                         }
 807
 808                         yield();
 809                 }
 810
 811                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 812                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 813
 814                 gfs2_log_lock(sdp);
 815                 bd = bh->b_private;
 816                 if (bd) {
 817                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 818                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 819                         gfs2_assert_warn(sdp, !bd->bd_ail);
 820                         bd->bd_bh = NULL;
 821                         if (!list_empty(&bd->bd_le.le_list))
 822                                 bd = NULL;
 823                         bh->b_private = NULL;
 824                 }
 825                 gfs2_log_unlock(sdp);
 826                 if (bd)
 827                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 828
 829                 bh = bh->b_this_page;
 830         } while (bh != head);
 831
 832 out:
 833         return try_to_free_buffers(page);
 834 }
 835
 836 const struct address_space_operations gfs2_file_aops = {
 837         .writepage = gfs2_writepage,
 838         .writepages = gfs2_writepages,
 839         .readpage = gfs2_readpage,
 840         .readpages = gfs2_readpages,
 841         .sync_page = block_sync_page,
 842         .prepare_write = gfs2_prepare_write,
 843         .commit_write = gfs2_commit_write,
 844         .bmap = gfs2_bmap,
 845         .invalidatepage = gfs2_invalidatepage,
 846         .releasepage = gfs2_releasepage,
 847         .direct_IO = gfs2_direct_IO,
 848 };
 849