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             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, 0);
 418         if (error)
 419                 goto out;
 420
 421         if (gfs2_is_stuffed(ip)) {
 422                 if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 423                         error = gfs2_unstuff_dinode(ip, page);
 424                         if (error == 0)
 425                                 goto prepare_write;
 426                 } else if (!PageUptodate(page))
 427                         error = stuffed_readpage(ip, page);
 428                 goto out;
 429         }
 430
 431 prepare_write:
 432         error = block_prepare_write(page, from, to, gfs2_get_block);
 433
 434 out:
 435         if (error) {
 436                 gfs2_trans_end(sdp);
 437                 if (alloc_required) {
 438                         gfs2_inplace_release(ip);
 439 out_qunlock:
 440                         gfs2_quota_unlock(ip);
 441 out_alloc_put:
 442                         gfs2_alloc_put(ip);
 443                 }
 444 out_unlock:
 445                 gfs2_glock_dq_m(1, &ip->i_gh);
 446 out_uninit:
 447                 gfs2_holder_uninit(&ip->i_gh);
 448         }
 449
 450         return error;
 451 }
 452
 453 /**
 454  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 455  * @inode: the rindex inode
 456  */
 457 static void adjust_fs_space(struct inode *inode)
 458 {
 459         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 460         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 461         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 462         u64 fs_total, new_free;
 463
 464         /* Total up the file system space, according to the latest rindex. */
 465         fs_total = gfs2_ri_total(sdp);
 466
 467         spin_lock(&sdp->sd_statfs_spin);
 468         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 469                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 470         else
 471                 new_free = 0;
 472         spin_unlock(&sdp->sd_statfs_spin);
 473         fs_warn(sdp, "File system extended by %llu blocks.\n",
 474                 (unsigned long long)new_free);
 475         gfs2_statfs_change(sdp, new_free, new_free, 0);
 476 }
 477
 478 /**
 479  * gfs2_commit_write - Commit write to a file
 480  * @file: The file to write to
 481  * @page: The page containing the data
 482  * @from: From (byte range within page)
 483  * @to: To (byte range within page)
 484  *
 485  * Returns: errno
 486  */
 487
 488 static int gfs2_commit_write(struct file *file, struct page *page,
 489                              unsigned from, unsigned to)
 490 {
 491         struct inode *inode = page->mapping->host;
 492         struct gfs2_inode *ip = GFS2_I(inode);
 493         struct gfs2_sbd *sdp = GFS2_SB(inode);
 494         int error = -EOPNOTSUPP;
 495         struct buffer_head *dibh;
 496         struct gfs2_alloc *al = &ip->i_alloc;
 497         struct gfs2_dinode *di;
 498
 499         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
 500                 goto fail_nounlock;
 501
 502         error = gfs2_meta_inode_buffer(ip, &dibh);
 503         if (error)
 504                 goto fail_endtrans;
 505
 506         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 507         di = (struct gfs2_dinode *)dibh->b_data;
 508
 509         if (gfs2_is_stuffed(ip)) {
 510                 u64 file_size;
 511                 void *kaddr;
 512
 513                 file_size = ((u64)page->index << PAGE_CACHE_SHIFT) + to;
 514
 515                 kaddr = kmap_atomic(page, KM_USER0);
 516                 memcpy(dibh->b_data + sizeof(struct gfs2_dinode) + from,
 517                        kaddr + from, to - from);
 518                 kunmap_atomic(kaddr, KM_USER0);
 519
 520                 SetPageUptodate(page);
 521
 522                 if (inode->i_size < file_size) {
 523                         i_size_write(inode, file_size);
 524                         mark_inode_dirty(inode);
 525                 }
 526         } else {
 527                 if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED ||
 528                     gfs2_is_jdata(ip))
 529                         gfs2_page_add_databufs(ip, page, from, to);
 530                 error = generic_commit_write(file, page, from, to);
 531                 if (error)
 532                         goto fail;
 533         }
 534
 535         if (ip->i_di.di_size < inode->i_size) {
 536                 ip->i_di.di_size = inode->i_size;
 537                 di->di_size = cpu_to_be64(inode->i_size);
 538         }
 539
 540         if (inode == sdp->sd_rindex)
 541                 adjust_fs_space(inode);
 542
 543         brelse(dibh);
 544         gfs2_trans_end(sdp);
 545         if (al->al_requested) {
 546                 gfs2_inplace_release(ip);
 547                 gfs2_quota_unlock(ip);
 548                 gfs2_alloc_put(ip);
 549         }
 550         unlock_page(page);
 551         gfs2_glock_dq_m(1, &ip->i_gh);
 552         lock_page(page);
 553         gfs2_holder_uninit(&ip->i_gh);
 554         return 0;
 555
 556 fail:
 557         brelse(dibh);
 558 fail_endtrans:
 559         gfs2_trans_end(sdp);
 560         if (al->al_requested) {
 561                 gfs2_inplace_release(ip);
 562                 gfs2_quota_unlock(ip);
 563                 gfs2_alloc_put(ip);
 564         }
 565         unlock_page(page);
 566         gfs2_glock_dq_m(1, &ip->i_gh);
 567         lock_page(page);
 568         gfs2_holder_uninit(&ip->i_gh);
 569 fail_nounlock:
 570         ClearPageUptodate(page);
 571         return error;
 572 }
 573
 574 /**
 575  * gfs2_set_page_dirty - Page dirtying function
 576  * @page: The page to dirty
 577  *
 578  * Returns: 1 if it dirtyed the page, or 0 otherwise
 579  */
 580
 581 static int gfs2_set_page_dirty(struct page *page)
 582 {
 583         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 584         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 585
 586         if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip))
 587                 SetPageChecked(page);
 588         return __set_page_dirty_buffers(page);
 589 }
 590
 591 /**
 592  * gfs2_bmap - Block map function
 593  * @mapping: Address space info
 594  * @lblock: The block to map
 595  *
 596  * Returns: The disk address for the block or 0 on hole or error
 597  */
 598
 599 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 600 {
 601         struct gfs2_inode *ip = GFS2_I(mapping->host);
 602         struct gfs2_holder i_gh;
 603         sector_t dblock = 0;
 604         int error;
 605
 606         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 607         if (error)
 608                 return 0;
 609
 610         if (!gfs2_is_stuffed(ip))
 611                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 612
 613         gfs2_glock_dq_uninit(&i_gh);
 614
 615         return dblock;
 616 }
 617
 618 static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
 619 {
 620         struct gfs2_bufdata *bd;
 621
 622         gfs2_log_lock(sdp);
 623         bd = bh->b_private;
 624         if (bd) {
 625                 bd->bd_bh = NULL;
 626                 bh->b_private = NULL;
 627                 if (!bd->bd_ail && list_empty(&bd->bd_le.le_list))
 628                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 629         }
 630         gfs2_log_unlock(sdp);
 631
 632         lock_buffer(bh);
 633         clear_buffer_dirty(bh);
 634         bh->b_bdev = NULL;
 635         clear_buffer_mapped(bh);
 636         clear_buffer_req(bh);
 637         clear_buffer_new(bh);
 638         clear_buffer_delay(bh);
 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 *head, *bh, *next;
 646         unsigned int curr_off = 0;
 647
 648         BUG_ON(!PageLocked(page));
 649         if (offset == 0)
 650                 ClearPageChecked(page);
 651         if (!page_has_buffers(page))
 652                 return;
 653
 654         bh = head = page_buffers(page);
 655         do {
 656                 unsigned int next_off = curr_off + bh->b_size;
 657                 next = bh->b_this_page;
 658
 659                 if (offset <= curr_off)
 660                         discard_buffer(sdp, bh);
 661
 662                 curr_off = next_off;
 663                 bh = next;
 664         } while (bh != head);
 665
 666         if (!offset)
 667                 try_to_release_page(page, 0);
 668
 669         return;
 670 }
 671
 672 /**
 673  * gfs2_ok_for_dio - check that dio is valid on this file
 674  * @ip: The inode
 675  * @rw: READ or WRITE
 676  * @offset: The offset at which we are reading or writing
 677  *
 678  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 679  *          1 (to accept the i/o request)
 680  */
 681 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 682 {
 683         /*
 684          * Should we return an error here? I can't see that O_DIRECT for
 685          * a journaled file makes any sense. For now we'll silently fall
 686          * back to buffered I/O, likewise we do the same for stuffed
 687          * files since they are (a) small and (b) unaligned.
 688          */
 689         if (gfs2_is_jdata(ip))
 690                 return 0;
 691
 692         if (gfs2_is_stuffed(ip))
 693                 return 0;
 694
 695         if (offset > i_size_read(&ip->i_inode))
 696                 return 0;
 697         return 1;
 698 }
 699
 700
 701
 702 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 703                               const struct iovec *iov, loff_t offset,
 704                               unsigned long nr_segs)
 705 {
 706         struct file *file = iocb->ki_filp;
 707         struct inode *inode = file->f_mapping->host;
 708         struct gfs2_inode *ip = GFS2_I(inode);
 709         struct gfs2_holder gh;
 710         int rv;
 711
 712         /*
 713          * Deferred lock, even if its a write, since we do no allocation
 714          * on this path. All we need change is atime, and this lock mode
 715          * ensures that other nodes have flushed their buffered read caches
 716          * (i.e. their page cache entries for this inode). We do not,
 717          * unfortunately have the option of only flushing a range like
 718          * the VFS does.
 719          */
 720         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 721         rv = gfs2_glock_nq_atime(&gh);
 722         if (rv)
 723                 return rv;
 724         rv = gfs2_ok_for_dio(ip, rw, offset);
 725         if (rv != 1)
 726                 goto out; /* dio not valid, fall back to buffered i/o */
 727
 728         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 729                                            iov, offset, nr_segs,
 730                                            gfs2_get_block_direct, NULL);
 731 out:
 732         gfs2_glock_dq_m(1, &gh);
 733         gfs2_holder_uninit(&gh);
 734         return rv;
 735 }
 736
 737 /**
 738  * stuck_releasepage - We're stuck in gfs2_releasepage().  Print stuff out.
 739  * @bh: the buffer we're stuck on
 740  *
 741  */
 742
 743 static void stuck_releasepage(struct buffer_head *bh)
 744 {
 745         struct inode *inode = bh->b_page->mapping->host;
 746         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 747         struct gfs2_bufdata *bd = bh->b_private;
 748         struct gfs2_glock *gl;
 749 static unsigned limit = 0;
 750
 751         if (limit > 3)
 752                 return;
 753         limit++;
 754
 755         fs_warn(sdp, "stuck in gfs2_releasepage() %p\n", inode);
 756         fs_warn(sdp, "blkno = %llu, bh->b_count = %d\n",
 757                 (unsigned long long)bh->b_blocknr, atomic_read(&bh->b_count));
 758         fs_warn(sdp, "pinned = %u\n", buffer_pinned(bh));
 759         fs_warn(sdp, "bh->b_private = %s\n", (bd) ? "!NULL" : "NULL");
 760
 761         if (!bd)
 762                 return;
 763
 764         gl = bd->bd_gl;
 765
 766         fs_warn(sdp, "gl = (%u, %llu)\n",
 767                 gl->gl_name.ln_type, (unsigned long long)gl->gl_name.ln_number);
 768
 769         fs_warn(sdp, "bd_list_tr = %s, bd_le.le_list = %s\n",
 770                 (list_empty(&bd->bd_list_tr)) ? "no" : "yes",
 771                 (list_empty(&bd->bd_le.le_list)) ? "no" : "yes");
 772
 773         if (gl->gl_ops == &gfs2_inode_glops) {
 774                 struct gfs2_inode *ip = gl->gl_object;
 775                 unsigned int x;
 776
 777                 if (!ip)
 778                         return;
 779
 780                 fs_warn(sdp, "ip = %llu %llu\n",
 781                         (unsigned long long)ip->i_no_formal_ino,
 782                         (unsigned long long)ip->i_no_addr);
 783
 784                 for (x = 0; x < GFS2_MAX_META_HEIGHT; x++)
 785                         fs_warn(sdp, "ip->i_cache[%u] = %s\n",
 786                                 x, (ip->i_cache[x]) ? "!NULL" : "NULL");
 787         }
 788 }
 789
 790 /**
 791  * gfs2_releasepage - free the metadata associated with a page
 792  * @page: the page that's being released
 793  * @gfp_mask: passed from Linux VFS, ignored by us
 794  *
 795  * Call try_to_free_buffers() if the buffers in this page can be
 796  * released.
 797  *
 798  * Returns: 0
 799  */
 800
 801 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 802 {
 803         struct inode *aspace = page->mapping->host;
 804         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 805         struct buffer_head *bh, *head;
 806         struct gfs2_bufdata *bd;
 807         unsigned long t = jiffies + gfs2_tune_get(sdp, gt_stall_secs) * HZ;
 808
 809         if (!page_has_buffers(page))
 810                 goto out;
 811
 812         head = bh = page_buffers(page);
 813         do {
 814                 while (atomic_read(&bh->b_count)) {
 815                         if (!atomic_read(&aspace->i_writecount))
 816                                 return 0;
 817
 818                         if (!(gfp_mask & __GFP_WAIT))
 819                                 return 0;
 820
 821                         if (time_after_eq(jiffies, t)) {
 822                                 stuck_releasepage(bh);
 823                                 /* should we withdraw here? */
 824                                 return 0;
 825                         }
 826
 827                         yield();
 828                 }
 829
 830                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 831                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 832
 833                 gfs2_log_lock(sdp);
 834                 bd = bh->b_private;
 835                 if (bd) {
 836                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 837                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 838                         gfs2_assert_warn(sdp, !bd->bd_ail);
 839                         bd->bd_bh = NULL;
 840                         if (!list_empty(&bd->bd_le.le_list))
 841                                 bd = NULL;
 842                         bh->b_private = NULL;
 843                 }
 844                 gfs2_log_unlock(sdp);
 845                 if (bd)
 846                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 847
 848                 bh = bh->b_this_page;
 849         } while (bh != head);
 850
 851 out:
 852         return try_to_free_buffers(page);
 853 }
 854
 855 const struct address_space_operations gfs2_file_aops = {
 856         .writepage = gfs2_writepage,
 857         .writepages = gfs2_writepages,
 858         .readpage = gfs2_readpage,
 859         .readpages = gfs2_readpages,
 860         .sync_page = block_sync_page,
 861         .prepare_write = gfs2_prepare_write,
 862         .commit_write = gfs2_commit_write,
 863         .set_page_dirty = gfs2_set_page_dirty,
 864         .bmap = gfs2_bmap,
 865         .invalidatepage = gfs2_invalidatepage,
 866         .releasepage = gfs2_releasepage,
 867         .direct_IO = gfs2_direct_IO,
 868 };
 869