SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/gfs2/ops_address.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2006 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 "util.h"
  36 #include "glops.h"
  37
  38
  39 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  40                                    unsigned int from, unsigned int to)
  41 {
  42         struct buffer_head *head = page_buffers(page);
  43         unsigned int bsize = head->b_size;
  44         struct buffer_head *bh;
  45         unsigned int start, end;
  46
  47         for (bh = head, start = 0; bh != head || !start;
  48              bh = bh->b_this_page, start = end) {
  49                 end = start + bsize;
  50                 if (end <= from || start >= to)
  51                         continue;
  52                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  53         }
  54 }
  55
  56 /**
  57  * gfs2_get_block - Fills in a buffer head with details about a block
  58  * @inode: The inode
  59  * @lblock: The block number to look up
  60  * @bh_result: The buffer head to return the result in
  61  * @create: Non-zero if we may add block to the file
  62  *
  63  * Returns: errno
  64  */
  65
  66 int gfs2_get_block(struct inode *inode, sector_t lblock,
  67                    struct buffer_head *bh_result, int create)
  68 {
  69         return gfs2_block_map(inode, lblock, create, bh_result);
  70 }
  71
  72 /**
  73  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  74  * @inode: The inode
  75  * @lblock: The block number to look up
  76  * @bh_result: The buffer head to return the result in
  77  * @create: Non-zero if we may add block to the file
  78  *
  79  * Returns: errno
  80  */
  81
  82 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  83                                   struct buffer_head *bh_result, int create)
  84 {
  85         int error;
  86
  87         error = gfs2_block_map(inode, lblock, 0, bh_result);
  88         if (error)
  89                 return error;
  90         if (bh_result->b_blocknr == 0)
  91                 return -EIO;
  92         return 0;
  93 }
  94
  95 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  96                                  struct buffer_head *bh_result, int create)
  97 {
  98         return gfs2_block_map(inode, lblock, 0, bh_result);
  99 }
 100
 101 /**
 102  * gfs2_writepage - Write complete page
 103  * @page: Page to write
 104  *
 105  * Returns: errno
 106  *
 107  * Some of this is copied from block_write_full_page() although we still
 108  * call it to do most of the work.
 109  */
 110
 111 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 112 {
 113         struct inode *inode = page->mapping->host;
 114         struct gfs2_inode *ip = GFS2_I(inode);
 115         struct gfs2_sbd *sdp = GFS2_SB(inode);
 116         loff_t i_size = i_size_read(inode);
 117         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 118         unsigned offset;
 119         int error;
 120         int done_trans = 0;
 121
 122         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl))) {
 123                 unlock_page(page);
 124                 return -EIO;
 125         }
 126         if (current->journal_info)
 127                 goto out_ignore;
 128
 129         /* Is the page fully outside i_size? (truncate in progress) */
 130         offset = i_size & (PAGE_CACHE_SIZE-1);
 131         if (page->index > end_index || (page->index == end_index && !offset)) {
 132                 page->mapping->a_ops->invalidatepage(page, 0);
 133                 unlock_page(page);
 134                 return 0; /* don't care */
 135         }
 136
 137         if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip)) {
 138                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 139                 if (error)
 140                         goto out_ignore;
 141                 if (!page_has_buffers(page)) {
 142                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 143                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 144                 }
 145                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 146                 done_trans = 1;
 147         }
 148         error = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 149         if (done_trans)
 150                 gfs2_trans_end(sdp);
 151         gfs2_meta_cache_flush(ip);
 152         return error;
 153
 154 out_ignore:
 155         redirty_page_for_writepage(wbc, page);
 156         unlock_page(page);
 157         return 0;
 158 }
 159
 160 /**
 161  * gfs2_writepages - Write a bunch of dirty pages back to disk
 162  * @mapping: The mapping to write
 163  * @wbc: Write-back control
 164  *
 165  * For journaled files and/or ordered writes this just falls back to the
 166  * kernel's default writepages path for now. We will probably want to change
 167  * that eventually (i.e. when we look at allocate on flush).
 168  *
 169  * For the data=writeback case though we can already ignore buffer heads
 170  * and write whole extents at once. This is a big reduction in the
 171  * number of I/O requests we send and the bmap calls we make in this case.
 172  */
 173 int gfs2_writepages(struct address_space *mapping, struct writeback_control *wbc)
 174 {
 175         struct inode *inode = mapping->host;
 176         struct gfs2_inode *ip = GFS2_I(inode);
 177         struct gfs2_sbd *sdp = GFS2_SB(inode);
 178
 179         if (sdp->sd_args.ar_data == GFS2_DATA_WRITEBACK && !gfs2_is_jdata(ip))
 180                 return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 181
 182         return generic_writepages(mapping, wbc);
 183 }
 184
 185 /**
 186  * stuffed_readpage - Fill in a Linux page with stuffed file data
 187  * @ip: the inode
 188  * @page: the page
 189  *
 190  * Returns: errno
 191  */
 192
 193 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 194 {
 195         struct buffer_head *dibh;
 196         void *kaddr;
 197         int error;
 198
 199         BUG_ON(page->index);
 200
 201         error = gfs2_meta_inode_buffer(ip, &dibh);
 202         if (error)
 203                 return error;
 204
 205         kaddr = kmap_atomic(page, KM_USER0);
 206         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 207                ip->i_di.di_size);
 208         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
 209         kunmap_atomic(kaddr, KM_USER0);
 210
 211         brelse(dibh);
 212
 213         SetPageUptodate(page);
 214
 215         return 0;
 216 }
 217
 218
 219 /**
 220  * gfs2_readpage - readpage with locking
 221  * @file: The file to read a page for. N.B. This may be NULL if we are
 222  * reading an internal file.
 223  * @page: The page to read
 224  *
 225  * Returns: errno
 226  */
 227
 228 static int gfs2_readpage(struct file *file, struct page *page)
 229 {
 230         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 231         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 232         struct gfs2_file *gf = NULL;
 233         struct gfs2_holder gh;
 234         int error;
 235         int do_unlock = 0;
 236
 237         if (likely(file != &gfs2_internal_file_sentinel)) {
 238                 if (file) {
 239                         gf = file->private_data;
 240                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 241                                 /* gfs2_sharewrite_nopage has grabbed the ip->i_gl already */
 242                                 goto skip_lock;
 243                 }
 244                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
 245                 do_unlock = 1;
 246                 error = gfs2_glock_nq_atime(&gh);
 247                 if (unlikely(error))
 248                         goto out_unlock;
 249         }
 250
 251 skip_lock:
 252         if (gfs2_is_stuffed(ip)) {
 253                 error = stuffed_readpage(ip, page);
 254                 unlock_page(page);
 255         } else
 256                 error = mpage_readpage(page, gfs2_get_block);
 257
 258         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 259                 error = -EIO;
 260
 261         if (do_unlock) {
 262                 gfs2_glock_dq_m(1, &gh);
 263                 gfs2_holder_uninit(&gh);
 264         }
 265 out:
 266         return error;
 267 out_unlock:
 268         if (error == GLR_TRYFAILED)
 269                 error = AOP_TRUNCATED_PAGE;
 270         unlock_page(page);
 271         if (do_unlock)
 272                 gfs2_holder_uninit(&gh);
 273         goto out;
 274 }
 275
 276 /**
 277  * gfs2_readpages - Read a bunch of pages at once
 278  *
 279  * Some notes:
 280  * 1. This is only for readahead, so we can simply ignore any things
 281  *    which are slightly inconvenient (such as locking conflicts between
 282  *    the page lock and the glock) and return having done no I/O. Its
 283  *    obviously not something we'd want to do on too regular a basis.
 284  *    Any I/O we ignore at this time will be done via readpage later.
 285  * 2. We don't handle stuffed files here we let readpage do the honours.
 286  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 287  * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
 288  * 5. We use LM_FLAG_TRY_1CB here, effectively we then have lock-ahead as
 289  *    well as read-ahead.
 290  */
 291 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 292                           struct list_head *pages, unsigned nr_pages)
 293 {
 294         struct inode *inode = mapping->host;
 295         struct gfs2_inode *ip = GFS2_I(inode);
 296         struct gfs2_sbd *sdp = GFS2_SB(inode);
 297         struct gfs2_holder gh;
 298         int ret = 0;
 299         int do_unlock = 0;
 300
 301         if (likely(file != &gfs2_internal_file_sentinel)) {
 302                 if (file) {
 303                         struct gfs2_file *gf = file->private_data;
 304                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 305                                 goto skip_lock;
 306                 }
 307                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED,
 308                                  LM_FLAG_TRY_1CB|GL_ATIME, &gh);
 309                 do_unlock = 1;
 310                 ret = gfs2_glock_nq_atime(&gh);
 311                 if (ret == GLR_TRYFAILED)
 312                         goto out_noerror;
 313                 if (unlikely(ret))
 314                         goto out_unlock;
 315         }
 316 skip_lock:
 317         if (!gfs2_is_stuffed(ip))
 318                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
 319
 320         if (do_unlock) {
 321                 gfs2_glock_dq_m(1, &gh);
 322                 gfs2_holder_uninit(&gh);
 323         }
 324 out:
 325         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 326                 ret = -EIO;
 327         return ret;
 328 out_noerror:
 329         ret = 0;
 330 out_unlock:
 331         if (do_unlock)
 332                 gfs2_holder_uninit(&gh);
 333         goto out;
 334 }
 335
 336 /**
 337  * gfs2_prepare_write - Prepare to write a page to a file
 338  * @file: The file to write to
 339  * @page: The page which is to be prepared for writing
 340  * @from: From (byte range within page)
 341  * @to: To (byte range within page)
 342  *
 343  * Returns: errno
 344  */
 345
 346 static int gfs2_prepare_write(struct file *file, struct page *page,
 347                               unsigned from, unsigned to)
 348 {
 349         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 350         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 351         unsigned int data_blocks, ind_blocks, rblocks;
 352         int alloc_required;
 353         int error = 0;
 354         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + from;
 355         loff_t end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
 356         struct gfs2_alloc *al;
 357         unsigned int write_len = to - from;
 358
 359
 360         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME|LM_FLAG_TRY_1CB, &ip->i_gh);
 361         error = gfs2_glock_nq_atime(&ip->i_gh);
 362         if (unlikely(error)) {
 363                 if (error == GLR_TRYFAILED) {
 364                         unlock_page(page);
 365                         error = AOP_TRUNCATED_PAGE;
 366                 }
 367                 goto out_uninit;
 368         }
 369
 370         gfs2_write_calc_reserv(ip, write_len, &data_blocks, &ind_blocks);
 371
 372         error = gfs2_write_alloc_required(ip, pos, write_len, &alloc_required);
 373         if (error)
 374                 goto out_unlock;
 375
 376
 377         ip->i_alloc.al_requested = 0;
 378         if (alloc_required) {
 379                 al = gfs2_alloc_get(ip);
 380
 381                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 382                 if (error)
 383                         goto out_alloc_put;
 384
 385                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
 386                 if (error)
 387                         goto out_qunlock;
 388
 389                 al->al_requested = data_blocks + ind_blocks;
 390                 error = gfs2_inplace_reserve(ip);
 391                 if (error)
 392                         goto out_qunlock;
 393         }
 394
 395         rblocks = RES_DINODE + ind_blocks;
 396         if (gfs2_is_jdata(ip))
 397                 rblocks += data_blocks ? data_blocks : 1;
 398         if (ind_blocks || data_blocks)
 399                 rblocks += RES_STATFS + RES_QUOTA;
 400
 401         error = gfs2_trans_begin(sdp, rblocks, 0);
 402         if (error)
 403                 goto out;
 404
 405         if (gfs2_is_stuffed(ip)) {
 406                 if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 407                         error = gfs2_unstuff_dinode(ip, page);
 408                         if (error == 0)
 409                                 goto prepare_write;
 410                 } else if (!PageUptodate(page))
 411                         error = stuffed_readpage(ip, page);
 412                 goto out;
 413         }
 414
 415 prepare_write:
 416         error = block_prepare_write(page, from, to, gfs2_get_block);
 417
 418 out:
 419         if (error) {
 420                 gfs2_trans_end(sdp);
 421                 if (alloc_required) {
 422                         gfs2_inplace_release(ip);
 423 out_qunlock:
 424                         gfs2_quota_unlock(ip);
 425 out_alloc_put:
 426                         gfs2_alloc_put(ip);
 427                 }
 428 out_unlock:
 429                 gfs2_glock_dq_m(1, &ip->i_gh);
 430 out_uninit:
 431                 gfs2_holder_uninit(&ip->i_gh);
 432         }
 433
 434         return error;
 435 }
 436
 437 /**
 438  * gfs2_commit_write - Commit write to a file
 439  * @file: The file to write to
 440  * @page: The page containing the data
 441  * @from: From (byte range within page)
 442  * @to: To (byte range within page)
 443  *
 444  * Returns: errno
 445  */
 446
 447 static int gfs2_commit_write(struct file *file, struct page *page,
 448                              unsigned from, unsigned to)
 449 {
 450         struct inode *inode = page->mapping->host;
 451         struct gfs2_inode *ip = GFS2_I(inode);
 452         struct gfs2_sbd *sdp = GFS2_SB(inode);
 453         int error = -EOPNOTSUPP;
 454         struct buffer_head *dibh;
 455         struct gfs2_alloc *al = &ip->i_alloc;
 456         struct gfs2_dinode *di;
 457
 458         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
 459                 goto fail_nounlock;
 460
 461         error = gfs2_meta_inode_buffer(ip, &dibh);
 462         if (error)
 463                 goto fail_endtrans;
 464
 465         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 466         di = (struct gfs2_dinode *)dibh->b_data;
 467
 468         if (gfs2_is_stuffed(ip)) {
 469                 u64 file_size;
 470                 void *kaddr;
 471
 472                 file_size = ((u64)page->index << PAGE_CACHE_SHIFT) + to;
 473
 474                 kaddr = kmap_atomic(page, KM_USER0);
 475                 memcpy(dibh->b_data + sizeof(struct gfs2_dinode) + from,
 476                        kaddr + from, to - from);
 477                 kunmap_atomic(kaddr, KM_USER0);
 478
 479                 SetPageUptodate(page);
 480
 481                 if (inode->i_size < file_size) {
 482                         i_size_write(inode, file_size);
 483                         mark_inode_dirty(inode);
 484                 }
 485         } else {
 486                 if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED ||
 487                     gfs2_is_jdata(ip))
 488                         gfs2_page_add_databufs(ip, page, from, to);
 489                 error = generic_commit_write(file, page, from, to);
 490                 if (error)
 491                         goto fail;
 492         }
 493
 494         if (ip->i_di.di_size < inode->i_size) {
 495                 ip->i_di.di_size = inode->i_size;
 496                 di->di_size = cpu_to_be64(inode->i_size);
 497         }
 498
 499         brelse(dibh);
 500         gfs2_trans_end(sdp);
 501         if (al->al_requested) {
 502                 gfs2_inplace_release(ip);
 503                 gfs2_quota_unlock(ip);
 504                 gfs2_alloc_put(ip);
 505         }
 506         gfs2_glock_dq_m(1, &ip->i_gh);
 507         gfs2_holder_uninit(&ip->i_gh);
 508         return 0;
 509
 510 fail:
 511         brelse(dibh);
 512 fail_endtrans:
 513         gfs2_trans_end(sdp);
 514         if (al->al_requested) {
 515                 gfs2_inplace_release(ip);
 516                 gfs2_quota_unlock(ip);
 517                 gfs2_alloc_put(ip);
 518         }
 519         gfs2_glock_dq_m(1, &ip->i_gh);
 520         gfs2_holder_uninit(&ip->i_gh);
 521 fail_nounlock:
 522         ClearPageUptodate(page);
 523         return error;
 524 }
 525
 526 /**
 527  * gfs2_bmap - Block map function
 528  * @mapping: Address space info
 529  * @lblock: The block to map
 530  *
 531  * Returns: The disk address for the block or 0 on hole or error
 532  */
 533
 534 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 535 {
 536         struct gfs2_inode *ip = GFS2_I(mapping->host);
 537         struct gfs2_holder i_gh;
 538         sector_t dblock = 0;
 539         int error;
 540
 541         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 542         if (error)
 543                 return 0;
 544
 545         if (!gfs2_is_stuffed(ip))
 546                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 547
 548         gfs2_glock_dq_uninit(&i_gh);
 549
 550         return dblock;
 551 }
 552
 553 static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
 554 {
 555         struct gfs2_bufdata *bd;
 556
 557         gfs2_log_lock(sdp);
 558         bd = bh->b_private;
 559         if (bd) {
 560                 bd->bd_bh = NULL;
 561                 bh->b_private = NULL;
 562         }
 563         gfs2_log_unlock(sdp);
 564
 565         lock_buffer(bh);
 566         clear_buffer_dirty(bh);
 567         bh->b_bdev = NULL;
 568         clear_buffer_mapped(bh);
 569         clear_buffer_req(bh);
 570         clear_buffer_new(bh);
 571         clear_buffer_delay(bh);
 572         unlock_buffer(bh);
 573 }
 574
 575 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 576 {
 577         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 578         struct buffer_head *head, *bh, *next;
 579         unsigned int curr_off = 0;
 580
 581         BUG_ON(!PageLocked(page));
 582         if (!page_has_buffers(page))
 583                 return;
 584
 585         bh = head = page_buffers(page);
 586         do {
 587                 unsigned int next_off = curr_off + bh->b_size;
 588                 next = bh->b_this_page;
 589
 590                 if (offset <= curr_off)
 591                         discard_buffer(sdp, bh);
 592
 593                 curr_off = next_off;
 594                 bh = next;
 595         } while (bh != head);
 596
 597         if (!offset)
 598                 try_to_release_page(page, 0);
 599
 600         return;
 601 }
 602
 603 /**
 604  * gfs2_ok_for_dio - check that dio is valid on this file
 605  * @ip: The inode
 606  * @rw: READ or WRITE
 607  * @offset: The offset at which we are reading or writing
 608  *
 609  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 610  *          1 (to accept the i/o request)
 611  */
 612 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 613 {
 614         /*
 615          * Should we return an error here? I can't see that O_DIRECT for
 616          * a journaled file makes any sense. For now we'll silently fall
 617          * back to buffered I/O, likewise we do the same for stuffed
 618          * files since they are (a) small and (b) unaligned.
 619          */
 620         if (gfs2_is_jdata(ip))
 621                 return 0;
 622
 623         if (gfs2_is_stuffed(ip))
 624                 return 0;
 625
 626         if (offset > i_size_read(&ip->i_inode))
 627                 return 0;
 628         return 1;
 629 }
 630
 631
 632
 633 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 634                               const struct iovec *iov, loff_t offset,
 635                               unsigned long nr_segs)
 636 {
 637         struct file *file = iocb->ki_filp;
 638         struct inode *inode = file->f_mapping->host;
 639         struct gfs2_inode *ip = GFS2_I(inode);
 640         struct gfs2_holder gh;
 641         int rv;
 642
 643         /*
 644          * Deferred lock, even if its a write, since we do no allocation
 645          * on this path. All we need change is atime, and this lock mode
 646          * ensures that other nodes have flushed their buffered read caches
 647          * (i.e. their page cache entries for this inode). We do not,
 648          * unfortunately have the option of only flushing a range like
 649          * the VFS does.
 650          */
 651         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 652         rv = gfs2_glock_nq_atime(&gh);
 653         if (rv)
 654                 return rv;
 655         rv = gfs2_ok_for_dio(ip, rw, offset);
 656         if (rv != 1)
 657                 goto out; /* dio not valid, fall back to buffered i/o */
 658
 659         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 660                                            iov, offset, nr_segs,
 661                                            gfs2_get_block_direct, NULL);
 662 out:
 663         gfs2_glock_dq_m(1, &gh);
 664         gfs2_holder_uninit(&gh);
 665         return rv;
 666 }
 667
 668 /**
 669  * stuck_releasepage - We're stuck in gfs2_releasepage().  Print stuff out.
 670  * @bh: the buffer we're stuck on
 671  *
 672  */
 673
 674 static void stuck_releasepage(struct buffer_head *bh)
 675 {
 676         struct inode *inode = bh->b_page->mapping->host;
 677         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 678         struct gfs2_bufdata *bd = bh->b_private;
 679         struct gfs2_glock *gl;
 680 static unsigned limit = 0;
 681
 682         if (limit > 3)
 683                 return;
 684         limit++;
 685
 686         fs_warn(sdp, "stuck in gfs2_releasepage() %p\n", inode);
 687         fs_warn(sdp, "blkno = %llu, bh->b_count = %d\n",
 688                 (unsigned long long)bh->b_blocknr, atomic_read(&bh->b_count));
 689         fs_warn(sdp, "pinned = %u\n", buffer_pinned(bh));
 690         fs_warn(sdp, "bh->b_private = %s\n", (bd) ? "!NULL" : "NULL");
 691
 692         if (!bd)
 693                 return;
 694
 695         gl = bd->bd_gl;
 696
 697         fs_warn(sdp, "gl = (%u, %llu)\n",
 698                 gl->gl_name.ln_type, (unsigned long long)gl->gl_name.ln_number);
 699
 700         fs_warn(sdp, "bd_list_tr = %s, bd_le.le_list = %s\n",
 701                 (list_empty(&bd->bd_list_tr)) ? "no" : "yes",
 702                 (list_empty(&bd->bd_le.le_list)) ? "no" : "yes");
 703
 704         if (gl->gl_ops == &gfs2_inode_glops) {
 705                 struct gfs2_inode *ip = gl->gl_object;
 706                 unsigned int x;
 707
 708                 if (!ip)
 709                         return;
 710
 711                 fs_warn(sdp, "ip = %llu %llu\n",
 712                         (unsigned long long)ip->i_num.no_formal_ino,
 713                         (unsigned long long)ip->i_num.no_addr);
 714
 715                 for (x = 0; x < GFS2_MAX_META_HEIGHT; x++)
 716                         fs_warn(sdp, "ip->i_cache[%u] = %s\n",
 717                                 x, (ip->i_cache[x]) ? "!NULL" : "NULL");
 718         }
 719 }
 720
 721 /**
 722  * gfs2_releasepage - free the metadata associated with a page
 723  * @page: the page that's being released
 724  * @gfp_mask: passed from Linux VFS, ignored by us
 725  *
 726  * Call try_to_free_buffers() if the buffers in this page can be
 727  * released.
 728  *
 729  * Returns: 0
 730  */
 731
 732 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 733 {
 734         struct inode *aspace = page->mapping->host;
 735         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 736         struct buffer_head *bh, *head;
 737         struct gfs2_bufdata *bd;
 738         unsigned long t = jiffies + gfs2_tune_get(sdp, gt_stall_secs) * HZ;
 739
 740         if (!page_has_buffers(page))
 741                 goto out;
 742
 743         head = bh = page_buffers(page);
 744         do {
 745                 while (atomic_read(&bh->b_count)) {
 746                         if (!atomic_read(&aspace->i_writecount))
 747                                 return 0;
 748
 749                         if (!(gfp_mask & __GFP_WAIT))
 750                                 return 0;
 751
 752                         if (time_after_eq(jiffies, t)) {
 753                                 stuck_releasepage(bh);
 754                                 /* should we withdraw here? */
 755                                 return 0;
 756                         }
 757
 758                         yield();
 759                 }
 760
 761                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 762                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 763
 764                 gfs2_log_lock(sdp);
 765                 bd = bh->b_private;
 766                 if (bd) {
 767                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 768                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 769                         gfs2_assert_warn(sdp, !bd->bd_ail);
 770                         bd->bd_bh = NULL;
 771                         if (!list_empty(&bd->bd_le.le_list))
 772                                 bd = NULL;
 773                         bh->b_private = NULL;
 774                 }
 775                 gfs2_log_unlock(sdp);
 776                 if (bd)
 777                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 778
 779                 bh = bh->b_this_page;
 780         } while (bh != head);
 781
 782 out:
 783         return try_to_free_buffers(page);
 784 }
 785
 786 const struct address_space_operations gfs2_file_aops = {
 787         .writepage = gfs2_writepage,
 788         .writepages = gfs2_writepages,
 789         .readpage = gfs2_readpage,
 790         .readpages = gfs2_readpages,
 791         .sync_page = block_sync_page,
 792         .prepare_write = gfs2_prepare_write,
 793         .commit_write = gfs2_commit_write,
 794         .bmap = gfs2_bmap,
 795         .invalidatepage = gfs2_invalidatepage,
 796         .releasepage = gfs2_releasepage,
 797         .direct_IO = gfs2_direct_IO,
 798 };
 799