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) 2004-2007 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/swap.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/lm_interface.h>
  23 #include <linux/swap.h>
  24
  25 #include "gfs2.h"
  26 #include "incore.h"
  27 #include "bmap.h"
  28 #include "glock.h"
  29 #include "inode.h"
  30 #include "log.h"
  31 #include "meta_io.h"
  32 #include "ops_address.h"
  33 #include "quota.h"
  34 #include "trans.h"
  35 #include "rgrp.h"
  36 #include "super.h"
  37 #include "util.h"
  38 #include "glops.h"
  39
  40
  41 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  42                                    unsigned int from, unsigned int to)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  47         unsigned int start, end;
  48
  49         for (bh = head, start = 0; bh != head || !start;
  50              bh = bh->b_this_page, start = end) {
  51                 end = start + bsize;
  52                 if (end <= from || start >= to)
  53                         continue;
  54                 if (gfs2_is_jdata(ip))
  55                         set_buffer_uptodate(bh);
  56                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  57         }
  58 }
  59
  60 /**
  61  * gfs2_get_block - Fills in a buffer head with details about a block
  62  * @inode: The inode
  63  * @lblock: The block number to look up
  64  * @bh_result: The buffer head to return the result in
  65  * @create: Non-zero if we may add block to the file
  66  *
  67  * Returns: errno
  68  */
  69
  70 int gfs2_get_block(struct inode *inode, sector_t lblock,
  71                    struct buffer_head *bh_result, int create)
  72 {
  73         return gfs2_block_map(inode, lblock, create, bh_result);
  74 }
  75
  76 /**
  77  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  78  * @inode: The inode
  79  * @lblock: The block number to look up
  80  * @bh_result: The buffer head to return the result in
  81  * @create: Non-zero if we may add block to the file
  82  *
  83  * Returns: errno
  84  */
  85
  86 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  87                                   struct buffer_head *bh_result, int create)
  88 {
  89         int error;
  90
  91         error = gfs2_block_map(inode, lblock, 0, bh_result);
  92         if (error)
  93                 return error;
  94         if (!buffer_mapped(bh_result))
  95                 return -EIO;
  96         return 0;
  97 }
  98
  99 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
 100                                  struct buffer_head *bh_result, int create)
 101 {
 102         return gfs2_block_map(inode, lblock, 0, bh_result);
 103 }
 104
 105 /**
 106  * gfs2_writepage - Write complete page
 107  * @page: Page to write
 108  *
 109  * Returns: errno
 110  *
 111  * Some of this is copied from block_write_full_page() although we still
 112  * call it to do most of the work.
 113  */
 114
 115 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 116 {
 117         struct inode *inode = page->mapping->host;
 118         struct gfs2_inode *ip = GFS2_I(inode);
 119         struct gfs2_sbd *sdp = GFS2_SB(inode);
 120         loff_t i_size = i_size_read(inode);
 121         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 122         unsigned offset;
 123         int error;
 124         int done_trans = 0;
 125
 126         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl))) {
 127                 unlock_page(page);
 128                 return -EIO;
 129         }
 130         if (current->journal_info)
 131                 goto out_ignore;
 132
 133         /* Is the page fully outside i_size? (truncate in progress) */
 134         offset = i_size & (PAGE_CACHE_SIZE-1);
 135         if (page->index > end_index || (page->index == end_index && !offset)) {
 136                 page->mapping->a_ops->invalidatepage(page, 0);
 137                 unlock_page(page);
 138                 return 0; /* don't care */
 139         }
 140
 141         if (PageChecked(page)) {
 142                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 143                 if (error)
 144                         goto out_ignore;
 145                 ClearPageChecked(page);
 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         return error;
 157
 158 out_ignore:
 159         redirty_page_for_writepage(wbc, page);
 160         unlock_page(page);
 161         return 0;
 162 }
 163
 164 /**
 165  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
 166  * @mapping: The mapping to write
 167  * @wbc: Write-back control
 168  *
 169  * For the data=writeback case 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 static int gfs2_writeback_writepages(struct address_space *mapping,
 174                                      struct writeback_control *wbc)
 175 {
 176         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 177 }
 178
 179 /**
 180  * stuffed_readpage - Fill in a Linux page with stuffed file data
 181  * @ip: the inode
 182  * @page: the page
 183  *
 184  * Returns: errno
 185  */
 186
 187 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 188 {
 189         struct buffer_head *dibh;
 190         void *kaddr;
 191         int error;
 192
 193         /*
 194          * Due to the order of unstuffing files and ->nopage(), we can be
 195          * asked for a zero page in the case of a stuffed file being extended,
 196          * so we need to supply one here. It doesn't happen often.
 197          */
 198         if (unlikely(page->index)) {
 199                 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
 200                 return 0;
 201         }
 202
 203         error = gfs2_meta_inode_buffer(ip, &dibh);
 204         if (error)
 205                 return error;
 206
 207         kaddr = kmap_atomic(page, KM_USER0);
 208         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 209                ip->i_di.di_size);
 210         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
 211         kunmap_atomic(kaddr, KM_USER0);
 212         flush_dcache_page(page);
 213         brelse(dibh);
 214         SetPageUptodate(page);
 215
 216         return 0;
 217 }
 218
 219
 220 /**
 221  * __gfs2_readpage - readpage
 222  * @file: The file to read a page for
 223  * @page: The page to read
 224  *
 225  * This is the core of gfs2's readpage. Its used by the internal file
 226  * reading code as in that case we already hold the glock. Also its
 227  * called by gfs2_readpage() once the required lock has been granted.
 228  *
 229  */
 230
 231 static int __gfs2_readpage(void *file, struct page *page)
 232 {
 233         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 234         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 235         int error;
 236
 237         if (gfs2_is_stuffed(ip)) {
 238                 error = stuffed_readpage(ip, page);
 239                 unlock_page(page);
 240         } else {
 241                 error = mpage_readpage(page, gfs2_get_block);
 242         }
 243
 244         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 245                 return -EIO;
 246
 247         return error;
 248 }
 249
 250 /**
 251  * gfs2_readpage - read a page of a file
 252  * @file: The file to read
 253  * @page: The page of the file
 254  *
 255  * This deals with the locking required. We use a trylock in order to
 256  * avoid the page lock / glock ordering problems returning AOP_TRUNCATED_PAGE
 257  * in the event that we are unable to get the lock.
 258  */
 259
 260 static int gfs2_readpage(struct file *file, struct page *page)
 261 {
 262         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 263         struct gfs2_holder gh;
 264         int error;
 265
 266         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
 267         error = gfs2_glock_nq_atime(&gh);
 268         if (unlikely(error)) {
 269                 unlock_page(page);
 270                 goto out;
 271         }
 272         error = __gfs2_readpage(file, page);
 273         gfs2_glock_dq(&gh);
 274 out:
 275         gfs2_holder_uninit(&gh);
 276         if (error == GLR_TRYFAILED) {
 277                 yield();
 278                 return AOP_TRUNCATED_PAGE;
 279         }
 280         return error;
 281 }
 282
 283 /**
 284  * gfs2_internal_read - read an internal file
 285  * @ip: The gfs2 inode
 286  * @ra_state: The readahead state (or NULL for no readahead)
 287  * @buf: The buffer to fill
 288  * @pos: The file position
 289  * @size: The amount to read
 290  *
 291  */
 292
 293 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
 294                        char *buf, loff_t *pos, unsigned size)
 295 {
 296         struct address_space *mapping = ip->i_inode.i_mapping;
 297         unsigned long index = *pos / PAGE_CACHE_SIZE;
 298         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 299         unsigned copied = 0;
 300         unsigned amt;
 301         struct page *page;
 302         void *p;
 303
 304         do {
 305                 amt = size - copied;
 306                 if (offset + size > PAGE_CACHE_SIZE)
 307                         amt = PAGE_CACHE_SIZE - offset;
 308                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 309                 if (IS_ERR(page))
 310                         return PTR_ERR(page);
 311                 p = kmap_atomic(page, KM_USER0);
 312                 memcpy(buf + copied, p + offset, amt);
 313                 kunmap_atomic(p, KM_USER0);
 314                 mark_page_accessed(page);
 315                 page_cache_release(page);
 316                 copied += amt;
 317                 index++;
 318                 offset = 0;
 319         } while(copied < size);
 320         (*pos) += size;
 321         return size;
 322 }
 323
 324 /**
 325  * gfs2_readpages - Read a bunch of pages at once
 326  *
 327  * Some notes:
 328  * 1. This is only for readahead, so we can simply ignore any things
 329  *    which are slightly inconvenient (such as locking conflicts between
 330  *    the page lock and the glock) and return having done no I/O. Its
 331  *    obviously not something we'd want to do on too regular a basis.
 332  *    Any I/O we ignore at this time will be done via readpage later.
 333  * 2. We don't handle stuffed files here we let readpage do the honours.
 334  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 335  * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
 336  */
 337
 338 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 339                           struct list_head *pages, unsigned nr_pages)
 340 {
 341         struct inode *inode = mapping->host;
 342         struct gfs2_inode *ip = GFS2_I(inode);
 343         struct gfs2_sbd *sdp = GFS2_SB(inode);
 344         struct gfs2_holder gh;
 345         int ret;
 346
 347         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME, &gh);
 348         ret = gfs2_glock_nq_atime(&gh);
 349         if (unlikely(ret))
 350                 goto out_uninit;
 351         if (!gfs2_is_stuffed(ip))
 352                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
 353         gfs2_glock_dq(&gh);
 354 out_uninit:
 355         gfs2_holder_uninit(&gh);
 356         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 357                 ret = -EIO;
 358         return ret;
 359 }
 360
 361 /**
 362  * gfs2_write_begin - Begin to write to a file
 363  * @file: The file to write to
 364  * @mapping: The mapping in which to write
 365  * @pos: The file offset at which to start writing
 366  * @len: Length of the write
 367  * @flags: Various flags
 368  * @pagep: Pointer to return the page
 369  * @fsdata: Pointer to return fs data (unused by GFS2)
 370  *
 371  * Returns: errno
 372  */
 373
 374 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 375                             loff_t pos, unsigned len, unsigned flags,
 376                             struct page **pagep, void **fsdata)
 377 {
 378         struct gfs2_inode *ip = GFS2_I(mapping->host);
 379         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 380         unsigned int data_blocks, ind_blocks, rblocks;
 381         int alloc_required;
 382         int error = 0;
 383         struct gfs2_alloc *al;
 384         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 385         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 386         unsigned to = from + len;
 387         struct page *page;
 388
 389         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME, &ip->i_gh);
 390         error = gfs2_glock_nq_atime(&ip->i_gh);
 391         if (unlikely(error))
 392                 goto out_uninit;
 393
 394         error = -ENOMEM;
 395         page = __grab_cache_page(mapping, index);
 396         *pagep = page;
 397         if (!page)
 398                 goto out_unlock;
 399
 400         gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 401
 402         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
 403         if (error)
 404                 goto out_putpage;
 405
 406
 407         ip->i_alloc.al_requested = 0;
 408         if (alloc_required) {
 409                 al = gfs2_alloc_get(ip);
 410
 411                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 412                 if (error)
 413                         goto out_alloc_put;
 414
 415                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
 416                 if (error)
 417                         goto out_qunlock;
 418
 419                 al->al_requested = data_blocks + ind_blocks;
 420                 error = gfs2_inplace_reserve(ip);
 421                 if (error)
 422                         goto out_qunlock;
 423         }
 424
 425         rblocks = RES_DINODE + ind_blocks;
 426         if (gfs2_is_jdata(ip))
 427                 rblocks += data_blocks ? data_blocks : 1;
 428         if (ind_blocks || data_blocks)
 429                 rblocks += RES_STATFS + RES_QUOTA;
 430
 431         error = gfs2_trans_begin(sdp, rblocks,
 432                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 433         if (error)
 434                 goto out_trans_fail;
 435
 436         if (gfs2_is_stuffed(ip)) {
 437                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 438                         error = gfs2_unstuff_dinode(ip, page);
 439                         if (error == 0)
 440                                 goto prepare_write;
 441                 } else if (!PageUptodate(page))
 442                         error = stuffed_readpage(ip, page);
 443                 goto out;
 444         }
 445
 446 prepare_write:
 447         error = block_prepare_write(page, from, to, gfs2_get_block);
 448
 449 out:
 450         if (error) {
 451                 gfs2_trans_end(sdp);
 452 out_trans_fail:
 453                 if (alloc_required) {
 454                         gfs2_inplace_release(ip);
 455 out_qunlock:
 456                         gfs2_quota_unlock(ip);
 457 out_alloc_put:
 458                         gfs2_alloc_put(ip);
 459                 }
 460 out_putpage:
 461                 page_cache_release(page);
 462                 if (pos + len > ip->i_inode.i_size)
 463                         vmtruncate(&ip->i_inode, ip->i_inode.i_size);
 464 out_unlock:
 465                 gfs2_glock_dq_m(1, &ip->i_gh);
 466 out_uninit:
 467                 gfs2_holder_uninit(&ip->i_gh);
 468         }
 469
 470         return error;
 471 }
 472
 473 /**
 474  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 475  * @inode: the rindex inode
 476  */
 477 static void adjust_fs_space(struct inode *inode)
 478 {
 479         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 480         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 481         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 482         u64 fs_total, new_free;
 483
 484         /* Total up the file system space, according to the latest rindex. */
 485         fs_total = gfs2_ri_total(sdp);
 486
 487         spin_lock(&sdp->sd_statfs_spin);
 488         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 489                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 490         else
 491                 new_free = 0;
 492         spin_unlock(&sdp->sd_statfs_spin);
 493         fs_warn(sdp, "File system extended by %llu blocks.\n",
 494                 (unsigned long long)new_free);
 495         gfs2_statfs_change(sdp, new_free, new_free, 0);
 496 }
 497
 498 /**
 499  * gfs2_stuffed_write_end - Write end for stuffed files
 500  * @inode: The inode
 501  * @dibh: The buffer_head containing the on-disk inode
 502  * @pos: The file position
 503  * @len: The length of the write
 504  * @copied: How much was actually copied by the VFS
 505  * @page: The page
 506  *
 507  * This copies the data from the page into the inode block after
 508  * the inode data structure itself.
 509  *
 510  * Returns: errno
 511  */
 512 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 513                                   loff_t pos, unsigned len, unsigned copied,
 514                                   struct page *page)
 515 {
 516         struct gfs2_inode *ip = GFS2_I(inode);
 517         struct gfs2_sbd *sdp = GFS2_SB(inode);
 518         u64 to = pos + copied;
 519         void *kaddr;
 520         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 521         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
 522
 523         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 524         kaddr = kmap_atomic(page, KM_USER0);
 525         memcpy(buf + pos, kaddr + pos, copied);
 526         memset(kaddr + pos + copied, 0, len - copied);
 527         flush_dcache_page(page);
 528         kunmap_atomic(kaddr, KM_USER0);
 529
 530         if (!PageUptodate(page))
 531                 SetPageUptodate(page);
 532         unlock_page(page);
 533         page_cache_release(page);
 534
 535         if (inode->i_size < to) {
 536                 i_size_write(inode, to);
 537                 ip->i_di.di_size = inode->i_size;
 538                 di->di_size = cpu_to_be64(inode->i_size);
 539                 mark_inode_dirty(inode);
 540         }
 541
 542         if (inode == sdp->sd_rindex)
 543                 adjust_fs_space(inode);
 544
 545         brelse(dibh);
 546         gfs2_trans_end(sdp);
 547         gfs2_glock_dq(&ip->i_gh);
 548         gfs2_holder_uninit(&ip->i_gh);
 549         return copied;
 550 }
 551
 552 /**
 553  * gfs2_write_end
 554  * @file: The file to write to
 555  * @mapping: The address space to write to
 556  * @pos: The file position
 557  * @len: The length of the data
 558  * @copied:
 559  * @page: The page that has been written
 560  * @fsdata: The fsdata (unused in GFS2)
 561  *
 562  * The main write_end function for GFS2. We have a separate one for
 563  * stuffed files as they are slightly different, otherwise we just
 564  * put our locking around the VFS provided functions.
 565  *
 566  * Returns: errno
 567  */
 568
 569 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 570                           loff_t pos, unsigned len, unsigned copied,
 571                           struct page *page, void *fsdata)
 572 {
 573         struct inode *inode = page->mapping->host;
 574         struct gfs2_inode *ip = GFS2_I(inode);
 575         struct gfs2_sbd *sdp = GFS2_SB(inode);
 576         struct buffer_head *dibh;
 577         struct gfs2_alloc *al = &ip->i_alloc;
 578         struct gfs2_dinode *di;
 579         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 580         unsigned int to = from + len;
 581         int ret;
 582
 583         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == 0);
 584
 585         ret = gfs2_meta_inode_buffer(ip, &dibh);
 586         if (unlikely(ret)) {
 587                 unlock_page(page);
 588                 page_cache_release(page);
 589                 goto failed;
 590         }
 591
 592         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 593
 594         if (gfs2_is_stuffed(ip))
 595                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 596
 597         if (!gfs2_is_writeback(ip))
 598                 gfs2_page_add_databufs(ip, page, from, to);
 599
 600         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 601
 602         if (likely(ret >= 0)) {
 603                 copied = ret;
 604                 if  ((pos + copied) > inode->i_size) {
 605                         di = (struct gfs2_dinode *)dibh->b_data;
 606                         ip->i_di.di_size = inode->i_size;
 607                         di->di_size = cpu_to_be64(inode->i_size);
 608                         mark_inode_dirty(inode);
 609                 }
 610         }
 611
 612         if (inode == sdp->sd_rindex)
 613                 adjust_fs_space(inode);
 614
 615         brelse(dibh);
 616         gfs2_trans_end(sdp);
 617 failed:
 618         if (al->al_requested) {
 619                 gfs2_inplace_release(ip);
 620                 gfs2_quota_unlock(ip);
 621                 gfs2_alloc_put(ip);
 622         }
 623         gfs2_glock_dq(&ip->i_gh);
 624         gfs2_holder_uninit(&ip->i_gh);
 625         return ret;
 626 }
 627
 628 /**
 629  * gfs2_set_page_dirty - Page dirtying function
 630  * @page: The page to dirty
 631  *
 632  * Returns: 1 if it dirtyed the page, or 0 otherwise
 633  */
 634
 635 static int gfs2_set_page_dirty(struct page *page)
 636 {
 637         SetPageChecked(page);
 638         return __set_page_dirty_buffers(page);
 639 }
 640
 641 /**
 642  * gfs2_bmap - Block map function
 643  * @mapping: Address space info
 644  * @lblock: The block to map
 645  *
 646  * Returns: The disk address for the block or 0 on hole or error
 647  */
 648
 649 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 650 {
 651         struct gfs2_inode *ip = GFS2_I(mapping->host);
 652         struct gfs2_holder i_gh;
 653         sector_t dblock = 0;
 654         int error;
 655
 656         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 657         if (error)
 658                 return 0;
 659
 660         if (!gfs2_is_stuffed(ip))
 661                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 662
 663         gfs2_glock_dq_uninit(&i_gh);
 664
 665         return dblock;
 666 }
 667
 668 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 669 {
 670         struct gfs2_bufdata *bd;
 671
 672         lock_buffer(bh);
 673         gfs2_log_lock(sdp);
 674         clear_buffer_dirty(bh);
 675         bd = bh->b_private;
 676         if (bd) {
 677                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 678                         list_del_init(&bd->bd_le.le_list);
 679                 else
 680                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 681         }
 682         bh->b_bdev = NULL;
 683         clear_buffer_mapped(bh);
 684         clear_buffer_req(bh);
 685         clear_buffer_new(bh);
 686         gfs2_log_unlock(sdp);
 687         unlock_buffer(bh);
 688 }
 689
 690 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 691 {
 692         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 693         struct buffer_head *bh, *head;
 694         unsigned long pos = 0;
 695
 696         BUG_ON(!PageLocked(page));
 697         if (offset == 0)
 698                 ClearPageChecked(page);
 699         if (!page_has_buffers(page))
 700                 goto out;
 701
 702         bh = head = page_buffers(page);
 703         do {
 704                 if (offset <= pos)
 705                         gfs2_discard(sdp, bh);
 706                 pos += bh->b_size;
 707                 bh = bh->b_this_page;
 708         } while (bh != head);
 709 out:
 710         if (offset == 0)
 711                 try_to_release_page(page, 0);
 712 }
 713
 714 /**
 715  * gfs2_ok_for_dio - check that dio is valid on this file
 716  * @ip: The inode
 717  * @rw: READ or WRITE
 718  * @offset: The offset at which we are reading or writing
 719  *
 720  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 721  *          1 (to accept the i/o request)
 722  */
 723 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 724 {
 725         /*
 726          * Should we return an error here? I can't see that O_DIRECT for
 727          * a stuffed file makes any sense. For now we'll silently fall
 728          * back to buffered I/O
 729          */
 730         if (gfs2_is_stuffed(ip))
 731                 return 0;
 732
 733         if (offset > i_size_read(&ip->i_inode))
 734                 return 0;
 735         return 1;
 736 }
 737
 738
 739
 740 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 741                               const struct iovec *iov, loff_t offset,
 742                               unsigned long nr_segs)
 743 {
 744         struct file *file = iocb->ki_filp;
 745         struct inode *inode = file->f_mapping->host;
 746         struct gfs2_inode *ip = GFS2_I(inode);
 747         struct gfs2_holder gh;
 748         int rv;
 749
 750         /*
 751          * Deferred lock, even if its a write, since we do no allocation
 752          * on this path. All we need change is atime, and this lock mode
 753          * ensures that other nodes have flushed their buffered read caches
 754          * (i.e. their page cache entries for this inode). We do not,
 755          * unfortunately have the option of only flushing a range like
 756          * the VFS does.
 757          */
 758         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 759         rv = gfs2_glock_nq_atime(&gh);
 760         if (rv)
 761                 return rv;
 762         rv = gfs2_ok_for_dio(ip, rw, offset);
 763         if (rv != 1)
 764                 goto out; /* dio not valid, fall back to buffered i/o */
 765
 766         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 767                                            iov, offset, nr_segs,
 768                                            gfs2_get_block_direct, NULL);
 769 out:
 770         gfs2_glock_dq_m(1, &gh);
 771         gfs2_holder_uninit(&gh);
 772         return rv;
 773 }
 774
 775 /**
 776  * gfs2_releasepage - free the metadata associated with a page
 777  * @page: the page that's being released
 778  * @gfp_mask: passed from Linux VFS, ignored by us
 779  *
 780  * Call try_to_free_buffers() if the buffers in this page can be
 781  * released.
 782  *
 783  * Returns: 0
 784  */
 785
 786 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 787 {
 788         struct inode *aspace = page->mapping->host;
 789         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 790         struct buffer_head *bh, *head;
 791         struct gfs2_bufdata *bd;
 792
 793         if (!page_has_buffers(page))
 794                 return 0;
 795
 796         gfs2_log_lock(sdp);
 797         head = bh = page_buffers(page);
 798         do {
 799                 if (atomic_read(&bh->b_count))
 800                         goto cannot_release;
 801                 bd = bh->b_private;
 802                 if (bd && bd->bd_ail)
 803                         goto cannot_release;
 804                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 805                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 806                 bh = bh->b_this_page;
 807         } while(bh != head);
 808         gfs2_log_unlock(sdp);
 809
 810         head = bh = page_buffers(page);
 811         do {
 812                 gfs2_log_lock(sdp);
 813                 bd = bh->b_private;
 814                 if (bd) {
 815                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 816                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 817                         if (!list_empty(&bd->bd_le.le_list)) {
 818                                 if (!buffer_pinned(bh))
 819                                         list_del_init(&bd->bd_le.le_list);
 820                                 else
 821                                         bd = NULL;
 822                         }
 823                         if (bd)
 824                                 bd->bd_bh = NULL;
 825                         bh->b_private = NULL;
 826                 }
 827                 gfs2_log_unlock(sdp);
 828                 if (bd)
 829                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 830
 831                 bh = bh->b_this_page;
 832         } while (bh != head);
 833
 834         return try_to_free_buffers(page);
 835 cannot_release:
 836         gfs2_log_unlock(sdp);
 837         return 0;
 838 }
 839
 840 static const struct address_space_operations gfs2_writeback_aops = {
 841         .writepage = gfs2_writepage,
 842         .writepages = gfs2_writeback_writepages,
 843         .readpage = gfs2_readpage,
 844         .readpages = gfs2_readpages,
 845         .sync_page = block_sync_page,
 846         .write_begin = gfs2_write_begin,
 847         .write_end = gfs2_write_end,
 848         .bmap = gfs2_bmap,
 849         .invalidatepage = gfs2_invalidatepage,
 850         .releasepage = gfs2_releasepage,
 851         .direct_IO = gfs2_direct_IO,
 852 };
 853
 854 static const struct address_space_operations gfs2_ordered_aops = {
 855         .writepage = gfs2_writepage,
 856         .readpage = gfs2_readpage,
 857         .readpages = gfs2_readpages,
 858         .sync_page = block_sync_page,
 859         .write_begin = gfs2_write_begin,
 860         .write_end = gfs2_write_end,
 861         .set_page_dirty = gfs2_set_page_dirty,
 862         .bmap = gfs2_bmap,
 863         .invalidatepage = gfs2_invalidatepage,
 864         .releasepage = gfs2_releasepage,
 865         .direct_IO = gfs2_direct_IO,
 866 };
 867
 868 static const struct address_space_operations gfs2_jdata_aops = {
 869         .writepage = gfs2_writepage,
 870         .readpage = gfs2_readpage,
 871         .readpages = gfs2_readpages,
 872         .sync_page = block_sync_page,
 873         .write_begin = gfs2_write_begin,
 874         .write_end = gfs2_write_end,
 875         .set_page_dirty = gfs2_set_page_dirty,
 876         .bmap = gfs2_bmap,
 877         .invalidatepage = gfs2_invalidatepage,
 878         .releasepage = gfs2_releasepage,
 879 };
 880
 881 void gfs2_set_aops(struct inode *inode)
 882 {
 883         struct gfs2_inode *ip = GFS2_I(inode);
 884
 885         if (gfs2_is_writeback(ip))
 886                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
 887         else if (gfs2_is_ordered(ip))
 888                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
 889         else if (gfs2_is_jdata(ip))
 890                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
 891         else
 892                 BUG();
 893 }
 894