fs/gfs2/aops.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 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/backing-dev.h>
  23
  24 #include "gfs2.h"
  25 #include "incore.h"
  26 #include "bmap.h"
  27 #include "glock.h"
  28 #include "inode.h"
  29 #include "log.h"
  30 #include "meta_io.h"
  31 #include "quota.h"
  32 #include "trans.h"
  33 #include "rgrp.h"
  34 #include "super.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                 if (gfs2_is_jdata(ip))
  53                         set_buffer_uptodate(bh);
  54                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  55         }
  56 }
  57
  58 /**
  59  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  60  * @inode: The inode
  61  * @lblock: The block number to look up
  62  * @bh_result: The buffer head to return the result in
  63  * @create: Non-zero if we may add block to the file
  64  *
  65  * Returns: errno
  66  */
  67
  68 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  69                                   struct buffer_head *bh_result, int create)
  70 {
  71         int error;
  72
  73         error = gfs2_block_map(inode, lblock, bh_result, 0);
  74         if (error)
  75                 return error;
  76         if (!buffer_mapped(bh_result))
  77                 return -EIO;
  78         return 0;
  79 }
  80
  81 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  82                                  struct buffer_head *bh_result, int create)
  83 {
  84         return gfs2_block_map(inode, lblock, bh_result, 0);
  85 }
  86
  87 /**
  88  * gfs2_writepage_common - Common bits of writepage
  89  * @page: The page to be written
  90  * @wbc: The writeback control
  91  *
  92  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  93  */
  94
  95 static int gfs2_writepage_common(struct page *page,
  96                                  struct writeback_control *wbc)
  97 {
  98         struct inode *inode = page->mapping->host;
  99         struct gfs2_inode *ip = GFS2_I(inode);
 100         struct gfs2_sbd *sdp = GFS2_SB(inode);
 101         loff_t i_size = i_size_read(inode);
 102         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 103         unsigned offset;
 104
 105         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 106                 goto out;
 107         if (current->journal_info)
 108                 goto redirty;
 109         /* Is the page fully outside i_size? (truncate in progress) */
 110         offset = i_size & (PAGE_CACHE_SIZE-1);
 111         if (page->index > end_index || (page->index == end_index && !offset)) {
 112                 page->mapping->a_ops->invalidatepage(page, 0);
 113                 goto out;
 114         }
 115         return 1;
 116 redirty:
 117         redirty_page_for_writepage(wbc, page);
 118 out:
 119         unlock_page(page);
 120         return 0;
 121 }
 122
 123 /**
 124  * gfs2_writeback_writepage - Write page for writeback mappings
 125  * @page: The page
 126  * @wbc: The writeback control
 127  *
 128  */
 129
 130 static int gfs2_writeback_writepage(struct page *page,
 131                                     struct writeback_control *wbc)
 132 {
 133         int ret;
 134
 135         ret = gfs2_writepage_common(page, wbc);
 136         if (ret <= 0)
 137                 return ret;
 138
 139         ret = mpage_writepage(page, gfs2_get_block_noalloc, wbc);
 140         if (ret == -EAGAIN)
 141                 ret = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 142         return ret;
 143 }
 144
 145 /**
 146  * gfs2_ordered_writepage - Write page for ordered data files
 147  * @page: The page to write
 148  * @wbc: The writeback control
 149  *
 150  */
 151
 152 static int gfs2_ordered_writepage(struct page *page,
 153                                   struct writeback_control *wbc)
 154 {
 155         struct inode *inode = page->mapping->host;
 156         struct gfs2_inode *ip = GFS2_I(inode);
 157         int ret;
 158
 159         ret = gfs2_writepage_common(page, wbc);
 160         if (ret <= 0)
 161                 return ret;
 162
 163         if (!page_has_buffers(page)) {
 164                 create_empty_buffers(page, inode->i_sb->s_blocksize,
 165                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
 166         }
 167         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
 168         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 169 }
 170
 171 /**
 172  * __gfs2_jdata_writepage - The core of jdata writepage
 173  * @page: The page to write
 174  * @wbc: The writeback control
 175  *
 176  * This is shared between writepage and writepages and implements the
 177  * core of the writepage operation. If a transaction is required then
 178  * PageChecked will have been set and the transaction will have
 179  * already been started before this is called.
 180  */
 181
 182 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         struct gfs2_inode *ip = GFS2_I(inode);
 186         struct gfs2_sbd *sdp = GFS2_SB(inode);
 187
 188         if (PageChecked(page)) {
 189                 ClearPageChecked(page);
 190                 if (!page_has_buffers(page)) {
 191                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 192                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 193                 }
 194                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 195         }
 196         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 197 }
 198
 199 /**
 200  * gfs2_jdata_writepage - Write complete page
 201  * @page: Page to write
 202  *
 203  * Returns: errno
 204  *
 205  */
 206
 207 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 208 {
 209         struct inode *inode = page->mapping->host;
 210         struct gfs2_sbd *sdp = GFS2_SB(inode);
 211         int ret;
 212         int done_trans = 0;
 213
 214         if (PageChecked(page)) {
 215                 if (wbc->sync_mode != WB_SYNC_ALL)
 216                         goto out_ignore;
 217                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 218                 if (ret)
 219                         goto out_ignore;
 220                 done_trans = 1;
 221         }
 222         ret = gfs2_writepage_common(page, wbc);
 223         if (ret > 0)
 224                 ret = __gfs2_jdata_writepage(page, wbc);
 225         if (done_trans)
 226                 gfs2_trans_end(sdp);
 227         return ret;
 228
 229 out_ignore:
 230         redirty_page_for_writepage(wbc, page);
 231         unlock_page(page);
 232         return 0;
 233 }
 234
 235 /**
 236  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
 237  * @mapping: The mapping to write
 238  * @wbc: Write-back control
 239  *
 240  * For the data=writeback case we can already ignore buffer heads
 241  * and write whole extents at once. This is a big reduction in the
 242  * number of I/O requests we send and the bmap calls we make in this case.
 243  */
 244 static int gfs2_writeback_writepages(struct address_space *mapping,
 245                                      struct writeback_control *wbc)
 246 {
 247         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 248 }
 249
 250 /**
 251  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 252  * @mapping: The mapping
 253  * @wbc: The writeback control
 254  * @writepage: The writepage function to call for each page
 255  * @pvec: The vector of pages
 256  * @nr_pages: The number of pages to write
 257  *
 258  * Returns: non-zero if loop should terminate, zero otherwise
 259  */
 260
 261 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 262                                     struct writeback_control *wbc,
 263                                     struct pagevec *pvec,
 264                                     int nr_pages, pgoff_t end)
 265 {
 266         struct inode *inode = mapping->host;
 267         struct gfs2_sbd *sdp = GFS2_SB(inode);
 268         loff_t i_size = i_size_read(inode);
 269         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 270         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
 271         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
 272         int i;
 273         int ret;
 274
 275         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 276         if (ret < 0)
 277                 return ret;
 278
 279         for(i = 0; i < nr_pages; i++) {
 280                 struct page *page = pvec->pages[i];
 281
 282                 lock_page(page);
 283
 284                 if (unlikely(page->mapping != mapping)) {
 285                         unlock_page(page);
 286                         continue;
 287                 }
 288
 289                 if (!wbc->range_cyclic && page->index > end) {
 290                         ret = 1;
 291                         unlock_page(page);
 292                         continue;
 293                 }
 294
 295                 if (wbc->sync_mode != WB_SYNC_NONE)
 296                         wait_on_page_writeback(page);
 297
 298                 if (PageWriteback(page) ||
 299                     !clear_page_dirty_for_io(page)) {
 300                         unlock_page(page);
 301                         continue;
 302                 }
 303
 304                 /* Is the page fully outside i_size? (truncate in progress) */
 305                 if (page->index > end_index || (page->index == end_index && !offset)) {
 306                         page->mapping->a_ops->invalidatepage(page, 0);
 307                         unlock_page(page);
 308                         continue;
 309                 }
 310
 311                 ret = __gfs2_jdata_writepage(page, wbc);
 312
 313                 if (ret || (--(wbc->nr_to_write) <= 0))
 314                         ret = 1;
 315         }
 316         gfs2_trans_end(sdp);
 317         return ret;
 318 }
 319
 320 /**
 321  * gfs2_write_cache_jdata - Like write_cache_pages but different
 322  * @mapping: The mapping to write
 323  * @wbc: The writeback control
 324  * @writepage: The writepage function to call
 325  * @data: The data to pass to writepage
 326  *
 327  * The reason that we use our own function here is that we need to
 328  * start transactions before we grab page locks. This allows us
 329  * to get the ordering right.
 330  */
 331
 332 static int gfs2_write_cache_jdata(struct address_space *mapping,
 333                                   struct writeback_control *wbc)
 334 {
 335         int ret = 0;
 336         int done = 0;
 337         struct pagevec pvec;
 338         int nr_pages;
 339         pgoff_t index;
 340         pgoff_t end;
 341         int scanned = 0;
 342         int range_whole = 0;
 343
 344         pagevec_init(&pvec, 0);
 345         if (wbc->range_cyclic) {
 346                 index = mapping->writeback_index; /* Start from prev offset */
 347                 end = -1;
 348         } else {
 349                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
 350                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 351                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 352                         range_whole = 1;
 353                 scanned = 1;
 354         }
 355
 356 retry:
 357          while (!done && (index <= end) &&
 358                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 359                                                PAGECACHE_TAG_DIRTY,
 360                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
 361                 scanned = 1;
 362                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
 363                 if (ret)
 364                         done = 1;
 365                 if (ret > 0)
 366                         ret = 0;
 367
 368                 pagevec_release(&pvec);
 369                 cond_resched();
 370         }
 371
 372         if (!scanned && !done) {
 373                 /*
 374                  * We hit the last page and there is more work to be done: wrap
 375                  * back to the start of the file
 376                  */
 377                 scanned = 1;
 378                 index = 0;
 379                 goto retry;
 380         }
 381
 382         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 383                 mapping->writeback_index = index;
 384         return ret;
 385 }
 386
 387
 388 /**
 389  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 390  * @mapping: The mapping to write
 391  * @wbc: The writeback control
 392  *
 393  */
 394
 395 static int gfs2_jdata_writepages(struct address_space *mapping,
 396                                  struct writeback_control *wbc)
 397 {
 398         struct gfs2_inode *ip = GFS2_I(mapping->host);
 399         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 400         int ret;
 401
 402         ret = gfs2_write_cache_jdata(mapping, wbc);
 403         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 404                 gfs2_log_flush(sdp, ip->i_gl);
 405                 ret = gfs2_write_cache_jdata(mapping, wbc);
 406         }
 407         return ret;
 408 }
 409
 410 /**
 411  * stuffed_readpage - Fill in a Linux page with stuffed file data
 412  * @ip: the inode
 413  * @page: the page
 414  *
 415  * Returns: errno
 416  */
 417
 418 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 419 {
 420         struct buffer_head *dibh;
 421         void *kaddr;
 422         int error;
 423
 424         /*
 425          * Due to the order of unstuffing files and ->fault(), we can be
 426          * asked for a zero page in the case of a stuffed file being extended,
 427          * so we need to supply one here. It doesn't happen often.
 428          */
 429         if (unlikely(page->index)) {
 430                 zero_user(page, 0, PAGE_CACHE_SIZE);
 431                 SetPageUptodate(page);
 432                 return 0;
 433         }
 434
 435         error = gfs2_meta_inode_buffer(ip, &dibh);
 436         if (error)
 437                 return error;
 438
 439         kaddr = kmap_atomic(page, KM_USER0);
 440         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 441                ip->i_disksize);
 442         memset(kaddr + ip->i_disksize, 0, PAGE_CACHE_SIZE - ip->i_disksize);
 443         kunmap_atomic(kaddr, KM_USER0);
 444         flush_dcache_page(page);
 445         brelse(dibh);
 446         SetPageUptodate(page);
 447
 448         return 0;
 449 }
 450
 451
 452 /**
 453  * __gfs2_readpage - readpage
 454  * @file: The file to read a page for
 455  * @page: The page to read
 456  *
 457  * This is the core of gfs2's readpage. Its used by the internal file
 458  * reading code as in that case we already hold the glock. Also its
 459  * called by gfs2_readpage() once the required lock has been granted.
 460  *
 461  */
 462
 463 static int __gfs2_readpage(void *file, struct page *page)
 464 {
 465         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 466         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 467         int error;
 468
 469         if (gfs2_is_stuffed(ip)) {
 470                 error = stuffed_readpage(ip, page);
 471                 unlock_page(page);
 472         } else {
 473                 error = mpage_readpage(page, gfs2_block_map);
 474         }
 475
 476         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 477                 return -EIO;
 478
 479         return error;
 480 }
 481
 482 /**
 483  * gfs2_readpage - read a page of a file
 484  * @file: The file to read
 485  * @page: The page of the file
 486  *
 487  * This deals with the locking required. We have to unlock and
 488  * relock the page in order to get the locking in the right
 489  * order.
 490  */
 491
 492 static int gfs2_readpage(struct file *file, struct page *page)
 493 {
 494         struct address_space *mapping = page->mapping;
 495         struct gfs2_inode *ip = GFS2_I(mapping->host);
 496         struct gfs2_holder gh;
 497         int error;
 498
 499         unlock_page(page);
 500         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 501         error = gfs2_glock_nq(&gh);
 502         if (unlikely(error))
 503                 goto out;
 504         error = AOP_TRUNCATED_PAGE;
 505         lock_page(page);
 506         if (page->mapping == mapping && !PageUptodate(page))
 507                 error = __gfs2_readpage(file, page);
 508         else
 509                 unlock_page(page);
 510         gfs2_glock_dq(&gh);
 511 out:
 512         gfs2_holder_uninit(&gh);
 513         if (error && error != AOP_TRUNCATED_PAGE)
 514                 lock_page(page);
 515         return error;
 516 }
 517
 518 /**
 519  * gfs2_internal_read - read an internal file
 520  * @ip: The gfs2 inode
 521  * @ra_state: The readahead state (or NULL for no readahead)
 522  * @buf: The buffer to fill
 523  * @pos: The file position
 524  * @size: The amount to read
 525  *
 526  */
 527
 528 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
 529                        char *buf, loff_t *pos, unsigned size)
 530 {
 531         struct address_space *mapping = ip->i_inode.i_mapping;
 532         unsigned long index = *pos / PAGE_CACHE_SIZE;
 533         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
 534         unsigned copied = 0;
 535         unsigned amt;
 536         struct page *page;
 537         void *p;
 538
 539         do {
 540                 amt = size - copied;
 541                 if (offset + size > PAGE_CACHE_SIZE)
 542                         amt = PAGE_CACHE_SIZE - offset;
 543                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 544                 if (IS_ERR(page))
 545                         return PTR_ERR(page);
 546                 p = kmap_atomic(page, KM_USER0);
 547                 memcpy(buf + copied, p + offset, amt);
 548                 kunmap_atomic(p, KM_USER0);
 549                 mark_page_accessed(page);
 550                 page_cache_release(page);
 551                 copied += amt;
 552                 index++;
 553                 offset = 0;
 554         } while(copied < size);
 555         (*pos) += size;
 556         return size;
 557 }
 558
 559 /**
 560  * gfs2_readpages - Read a bunch of pages at once
 561  *
 562  * Some notes:
 563  * 1. This is only for readahead, so we can simply ignore any things
 564  *    which are slightly inconvenient (such as locking conflicts between
 565  *    the page lock and the glock) and return having done no I/O. Its
 566  *    obviously not something we'd want to do on too regular a basis.
 567  *    Any I/O we ignore at this time will be done via readpage later.
 568  * 2. We don't handle stuffed files here we let readpage do the honours.
 569  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 570  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 571  */
 572
 573 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 574                           struct list_head *pages, unsigned nr_pages)
 575 {
 576         struct inode *inode = mapping->host;
 577         struct gfs2_inode *ip = GFS2_I(inode);
 578         struct gfs2_sbd *sdp = GFS2_SB(inode);
 579         struct gfs2_holder gh;
 580         int ret;
 581
 582         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 583         ret = gfs2_glock_nq(&gh);
 584         if (unlikely(ret))
 585                 goto out_uninit;
 586         if (!gfs2_is_stuffed(ip))
 587                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 588         gfs2_glock_dq(&gh);
 589 out_uninit:
 590         gfs2_holder_uninit(&gh);
 591         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 592                 ret = -EIO;
 593         return ret;
 594 }
 595
 596 /**
 597  * gfs2_write_begin - Begin to write to a file
 598  * @file: The file to write to
 599  * @mapping: The mapping in which to write
 600  * @pos: The file offset at which to start writing
 601  * @len: Length of the write
 602  * @flags: Various flags
 603  * @pagep: Pointer to return the page
 604  * @fsdata: Pointer to return fs data (unused by GFS2)
 605  *
 606  * Returns: errno
 607  */
 608
 609 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 610                             loff_t pos, unsigned len, unsigned flags,
 611                             struct page **pagep, void **fsdata)
 612 {
 613         struct gfs2_inode *ip = GFS2_I(mapping->host);
 614         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 615         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 616         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 617         int alloc_required;
 618         int error = 0;
 619         struct gfs2_alloc *al;
 620         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 621         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 622         unsigned to = from + len;
 623         struct page *page;
 624
 625         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 626         error = gfs2_glock_nq(&ip->i_gh);
 627         if (unlikely(error))
 628                 goto out_uninit;
 629         if (&ip->i_inode == sdp->sd_rindex) {
 630                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 631                                            GL_NOCACHE, &m_ip->i_gh);
 632                 if (unlikely(error)) {
 633                         gfs2_glock_dq(&ip->i_gh);
 634                         goto out_uninit;
 635                 }
 636         }
 637
 638         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
 639         if (error)
 640                 goto out_unlock;
 641
 642         if (alloc_required || gfs2_is_jdata(ip))
 643                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 644
 645         if (alloc_required) {
 646                 al = gfs2_alloc_get(ip);
 647                 if (!al) {
 648                         error = -ENOMEM;
 649                         goto out_unlock;
 650                 }
 651
 652                 error = gfs2_quota_lock_check(ip);
 653                 if (error)
 654                         goto out_alloc_put;
 655
 656                 al->al_requested = data_blocks + ind_blocks;
 657                 error = gfs2_inplace_reserve(ip);
 658                 if (error)
 659                         goto out_qunlock;
 660         }
 661
 662         rblocks = RES_DINODE + ind_blocks;
 663         if (gfs2_is_jdata(ip))
 664                 rblocks += data_blocks ? data_blocks : 1;
 665         if (ind_blocks || data_blocks)
 666                 rblocks += RES_STATFS + RES_QUOTA;
 667         if (&ip->i_inode == sdp->sd_rindex)
 668                 rblocks += 2 * RES_STATFS;
 669
 670         error = gfs2_trans_begin(sdp, rblocks,
 671                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
 672         if (error)
 673                 goto out_trans_fail;
 674
 675         error = -ENOMEM;
 676         flags |= AOP_FLAG_NOFS;
 677         page = grab_cache_page_write_begin(mapping, index, flags);
 678         *pagep = page;
 679         if (unlikely(!page))
 680                 goto out_endtrans;
 681
 682         if (gfs2_is_stuffed(ip)) {
 683                 error = 0;
 684                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 685                         error = gfs2_unstuff_dinode(ip, page);
 686                         if (error == 0)
 687                                 goto prepare_write;
 688                 } else if (!PageUptodate(page)) {
 689                         error = stuffed_readpage(ip, page);
 690                 }
 691                 goto out;
 692         }
 693
 694 prepare_write:
 695         error = block_prepare_write(page, from, to, gfs2_block_map);
 696 out:
 697         if (error == 0)
 698                 return 0;
 699
 700         page_cache_release(page);
 701         if (pos + len > ip->i_inode.i_size)
 702                 vmtruncate(&ip->i_inode, ip->i_inode.i_size);
 703 out_endtrans:
 704         gfs2_trans_end(sdp);
 705 out_trans_fail:
 706         if (alloc_required) {
 707                 gfs2_inplace_release(ip);
 708 out_qunlock:
 709                 gfs2_quota_unlock(ip);
 710 out_alloc_put:
 711                 gfs2_alloc_put(ip);
 712         }
 713 out_unlock:
 714         if (&ip->i_inode == sdp->sd_rindex) {
 715                 gfs2_glock_dq(&m_ip->i_gh);
 716                 gfs2_holder_uninit(&m_ip->i_gh);
 717         }
 718         gfs2_glock_dq(&ip->i_gh);
 719 out_uninit:
 720         gfs2_holder_uninit(&ip->i_gh);
 721         return error;
 722 }
 723
 724 /**
 725  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 726  * @inode: the rindex inode
 727  */
 728 static void adjust_fs_space(struct inode *inode)
 729 {
 730         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 731         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 732         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 733         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 734         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 735         struct buffer_head *m_bh, *l_bh;
 736         u64 fs_total, new_free;
 737
 738         /* Total up the file system space, according to the latest rindex. */
 739         fs_total = gfs2_ri_total(sdp);
 740         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 741                 return;
 742
 743         spin_lock(&sdp->sd_statfs_spin);
 744         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 745                               sizeof(struct gfs2_dinode));
 746         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 747                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 748         else
 749                 new_free = 0;
 750         spin_unlock(&sdp->sd_statfs_spin);
 751         fs_warn(sdp, "File system extended by %llu blocks.\n",
 752                 (unsigned long long)new_free);
 753         gfs2_statfs_change(sdp, new_free, new_free, 0);
 754
 755         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 756                 goto out;
 757         update_statfs(sdp, m_bh, l_bh);
 758         brelse(l_bh);
 759 out:
 760         brelse(m_bh);
 761 }
 762
 763 /**
 764  * gfs2_stuffed_write_end - Write end for stuffed files
 765  * @inode: The inode
 766  * @dibh: The buffer_head containing the on-disk inode
 767  * @pos: The file position
 768  * @len: The length of the write
 769  * @copied: How much was actually copied by the VFS
 770  * @page: The page
 771  *
 772  * This copies the data from the page into the inode block after
 773  * the inode data structure itself.
 774  *
 775  * Returns: errno
 776  */
 777 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 778                                   loff_t pos, unsigned len, unsigned copied,
 779                                   struct page *page)
 780 {
 781         struct gfs2_inode *ip = GFS2_I(inode);
 782         struct gfs2_sbd *sdp = GFS2_SB(inode);
 783         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 784         u64 to = pos + copied;
 785         void *kaddr;
 786         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 787         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
 788
 789         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 790         kaddr = kmap_atomic(page, KM_USER0);
 791         memcpy(buf + pos, kaddr + pos, copied);
 792         memset(kaddr + pos + copied, 0, len - copied);
 793         flush_dcache_page(page);
 794         kunmap_atomic(kaddr, KM_USER0);
 795
 796         if (!PageUptodate(page))
 797                 SetPageUptodate(page);
 798         unlock_page(page);
 799         page_cache_release(page);
 800
 801         if (copied) {
 802                 if (inode->i_size < to) {
 803                         i_size_write(inode, to);
 804                         ip->i_disksize = inode->i_size;
 805                 }
 806                 gfs2_dinode_out(ip, di);
 807                 mark_inode_dirty(inode);
 808         }
 809
 810         if (inode == sdp->sd_rindex) {
 811                 adjust_fs_space(inode);
 812                 ip->i_gh.gh_flags |= GL_NOCACHE;
 813         }
 814
 815         brelse(dibh);
 816         gfs2_trans_end(sdp);
 817         if (inode == sdp->sd_rindex) {
 818                 gfs2_glock_dq(&m_ip->i_gh);
 819                 gfs2_holder_uninit(&m_ip->i_gh);
 820         }
 821         gfs2_glock_dq(&ip->i_gh);
 822         gfs2_holder_uninit(&ip->i_gh);
 823         return copied;
 824 }
 825
 826 /**
 827  * gfs2_write_end
 828  * @file: The file to write to
 829  * @mapping: The address space to write to
 830  * @pos: The file position
 831  * @len: The length of the data
 832  * @copied:
 833  * @page: The page that has been written
 834  * @fsdata: The fsdata (unused in GFS2)
 835  *
 836  * The main write_end function for GFS2. We have a separate one for
 837  * stuffed files as they are slightly different, otherwise we just
 838  * put our locking around the VFS provided functions.
 839  *
 840  * Returns: errno
 841  */
 842
 843 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 844                           loff_t pos, unsigned len, unsigned copied,
 845                           struct page *page, void *fsdata)
 846 {
 847         struct inode *inode = page->mapping->host;
 848         struct gfs2_inode *ip = GFS2_I(inode);
 849         struct gfs2_sbd *sdp = GFS2_SB(inode);
 850         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 851         struct buffer_head *dibh;
 852         struct gfs2_alloc *al = ip->i_alloc;
 853         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
 854         unsigned int to = from + len;
 855         int ret;
 856
 857         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 858
 859         ret = gfs2_meta_inode_buffer(ip, &dibh);
 860         if (unlikely(ret)) {
 861                 unlock_page(page);
 862                 page_cache_release(page);
 863                 goto failed;
 864         }
 865
 866         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 867
 868         if (gfs2_is_stuffed(ip))
 869                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 870
 871         if (!gfs2_is_writeback(ip))
 872                 gfs2_page_add_databufs(ip, page, from, to);
 873
 874         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 875         if (ret > 0) {
 876                 if (inode->i_size > ip->i_disksize)
 877                         ip->i_disksize = inode->i_size;
 878                 gfs2_dinode_out(ip, dibh->b_data);
 879                 mark_inode_dirty(inode);
 880         }
 881
 882         if (inode == sdp->sd_rindex) {
 883                 adjust_fs_space(inode);
 884                 ip->i_gh.gh_flags |= GL_NOCACHE;
 885         }
 886
 887         brelse(dibh);
 888         gfs2_trans_end(sdp);
 889 failed:
 890         if (al) {
 891                 gfs2_inplace_release(ip);
 892                 gfs2_quota_unlock(ip);
 893                 gfs2_alloc_put(ip);
 894         }
 895         if (inode == sdp->sd_rindex) {
 896                 gfs2_glock_dq(&m_ip->i_gh);
 897                 gfs2_holder_uninit(&m_ip->i_gh);
 898         }
 899         gfs2_glock_dq(&ip->i_gh);
 900         gfs2_holder_uninit(&ip->i_gh);
 901         return ret;
 902 }
 903
 904 /**
 905  * gfs2_set_page_dirty - Page dirtying function
 906  * @page: The page to dirty
 907  *
 908  * Returns: 1 if it dirtyed the page, or 0 otherwise
 909  */
 910
 911 static int gfs2_set_page_dirty(struct page *page)
 912 {
 913         SetPageChecked(page);
 914         return __set_page_dirty_buffers(page);
 915 }
 916
 917 /**
 918  * gfs2_bmap - Block map function
 919  * @mapping: Address space info
 920  * @lblock: The block to map
 921  *
 922  * Returns: The disk address for the block or 0 on hole or error
 923  */
 924
 925 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 926 {
 927         struct gfs2_inode *ip = GFS2_I(mapping->host);
 928         struct gfs2_holder i_gh;
 929         sector_t dblock = 0;
 930         int error;
 931
 932         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 933         if (error)
 934                 return 0;
 935
 936         if (!gfs2_is_stuffed(ip))
 937                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 938
 939         gfs2_glock_dq_uninit(&i_gh);
 940
 941         return dblock;
 942 }
 943
 944 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 945 {
 946         struct gfs2_bufdata *bd;
 947
 948         lock_buffer(bh);
 949         gfs2_log_lock(sdp);
 950         clear_buffer_dirty(bh);
 951         bd = bh->b_private;
 952         if (bd) {
 953                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
 954                         list_del_init(&bd->bd_le.le_list);
 955                 else
 956                         gfs2_remove_from_journal(bh, current->journal_info, 0);
 957         }
 958         bh->b_bdev = NULL;
 959         clear_buffer_mapped(bh);
 960         clear_buffer_req(bh);
 961         clear_buffer_new(bh);
 962         gfs2_log_unlock(sdp);
 963         unlock_buffer(bh);
 964 }
 965
 966 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 967 {
 968         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 969         struct buffer_head *bh, *head;
 970         unsigned long pos = 0;
 971
 972         BUG_ON(!PageLocked(page));
 973         if (offset == 0)
 974                 ClearPageChecked(page);
 975         if (!page_has_buffers(page))
 976                 goto out;
 977
 978         bh = head = page_buffers(page);
 979         do {
 980                 if (offset <= pos)
 981                         gfs2_discard(sdp, bh);
 982                 pos += bh->b_size;
 983                 bh = bh->b_this_page;
 984         } while (bh != head);
 985 out:
 986         if (offset == 0)
 987                 try_to_release_page(page, 0);
 988 }
 989
 990 /**
 991  * gfs2_ok_for_dio - check that dio is valid on this file
 992  * @ip: The inode
 993  * @rw: READ or WRITE
 994  * @offset: The offset at which we are reading or writing
 995  *
 996  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 997  *          1 (to accept the i/o request)
 998  */
 999 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
1000 {
1001         /*
1002          * Should we return an error here? I can't see that O_DIRECT for
1003          * a stuffed file makes any sense. For now we'll silently fall
1004          * back to buffered I/O
1005          */
1006         if (gfs2_is_stuffed(ip))
1007                 return 0;
1008
1009         if (offset >= i_size_read(&ip->i_inode))
1010                 return 0;
1011         return 1;
1012 }
1013
1014
1015
1016 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
1017                               const struct iovec *iov, loff_t offset,
1018                               unsigned long nr_segs)
1019 {
1020         struct file *file = iocb->ki_filp;
1021         struct inode *inode = file->f_mapping->host;
1022         struct gfs2_inode *ip = GFS2_I(inode);
1023         struct gfs2_holder gh;
1024         int rv;
1025
1026         /*
1027          * Deferred lock, even if its a write, since we do no allocation
1028          * on this path. All we need change is atime, and this lock mode
1029          * ensures that other nodes have flushed their buffered read caches
1030          * (i.e. their page cache entries for this inode). We do not,
1031          * unfortunately have the option of only flushing a range like
1032          * the VFS does.
1033          */
1034         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1035         rv = gfs2_glock_nq(&gh);
1036         if (rv)
1037                 return rv;
1038         rv = gfs2_ok_for_dio(ip, rw, offset);
1039         if (rv != 1)
1040                 goto out; /* dio not valid, fall back to buffered i/o */
1041
1042         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1043                                            iov, offset, nr_segs,
1044                                            gfs2_get_block_direct, NULL);
1045 out:
1046         gfs2_glock_dq_m(1, &gh);
1047         gfs2_holder_uninit(&gh);
1048         return rv;
1049 }
1050
1051 /**
1052  * gfs2_releasepage - free the metadata associated with a page
1053  * @page: the page that's being released
1054  * @gfp_mask: passed from Linux VFS, ignored by us
1055  *
1056  * Call try_to_free_buffers() if the buffers in this page can be
1057  * released.
1058  *
1059  * Returns: 0
1060  */
1061
1062 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1063 {
1064         struct address_space *mapping = page->mapping;
1065         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1066         struct buffer_head *bh, *head;
1067         struct gfs2_bufdata *bd;
1068
1069         if (!page_has_buffers(page))
1070                 return 0;
1071
1072         gfs2_log_lock(sdp);
1073         head = bh = page_buffers(page);
1074         do {
1075                 if (atomic_read(&bh->b_count))
1076                         goto cannot_release;
1077                 bd = bh->b_private;
1078                 if (bd && bd->bd_ail)
1079                         goto cannot_release;
1080                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1081                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1082                 bh = bh->b_this_page;
1083         } while(bh != head);
1084         gfs2_log_unlock(sdp);
1085
1086         head = bh = page_buffers(page);
1087         do {
1088                 gfs2_log_lock(sdp);
1089                 bd = bh->b_private;
1090                 if (bd) {
1091                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1092                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1093                         if (!list_empty(&bd->bd_le.le_list)) {
1094                                 if (!buffer_pinned(bh))
1095                                         list_del_init(&bd->bd_le.le_list);
1096                                 else
1097                                         bd = NULL;
1098                         }
1099                         if (bd)
1100                                 bd->bd_bh = NULL;
1101                         bh->b_private = NULL;
1102                 }
1103                 gfs2_log_unlock(sdp);
1104                 if (bd)
1105                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1106
1107                 bh = bh->b_this_page;
1108         } while (bh != head);
1109
1110         return try_to_free_buffers(page);
1111 cannot_release:
1112         gfs2_log_unlock(sdp);
1113         return 0;
1114 }
1115
1116 static const struct address_space_operations gfs2_writeback_aops = {
1117         .writepage = gfs2_writeback_writepage,
1118         .writepages = gfs2_writeback_writepages,
1119         .readpage = gfs2_readpage,
1120         .readpages = gfs2_readpages,
1121         .sync_page = block_sync_page,
1122         .write_begin = gfs2_write_begin,
1123         .write_end = gfs2_write_end,
1124         .bmap = gfs2_bmap,
1125         .invalidatepage = gfs2_invalidatepage,
1126         .releasepage = gfs2_releasepage,
1127         .direct_IO = gfs2_direct_IO,
1128         .migratepage = buffer_migrate_page,
1129         .is_partially_uptodate = block_is_partially_uptodate,
1130         .error_remove_page = generic_error_remove_page,
1131 };
1132
1133 static const struct address_space_operations gfs2_ordered_aops = {
1134         .writepage = gfs2_ordered_writepage,
1135         .readpage = gfs2_readpage,
1136         .readpages = gfs2_readpages,
1137         .sync_page = block_sync_page,
1138         .write_begin = gfs2_write_begin,
1139         .write_end = gfs2_write_end,
1140         .set_page_dirty = gfs2_set_page_dirty,
1141         .bmap = gfs2_bmap,
1142         .invalidatepage = gfs2_invalidatepage,
1143         .releasepage = gfs2_releasepage,
1144         .direct_IO = gfs2_direct_IO,
1145         .migratepage = buffer_migrate_page,
1146         .is_partially_uptodate = block_is_partially_uptodate,
1147         .error_remove_page = generic_error_remove_page,
1148 };
1149
1150 static const struct address_space_operations gfs2_jdata_aops = {
1151         .writepage = gfs2_jdata_writepage,
1152         .writepages = gfs2_jdata_writepages,
1153         .readpage = gfs2_readpage,
1154         .readpages = gfs2_readpages,
1155         .sync_page = block_sync_page,
1156         .write_begin = gfs2_write_begin,
1157         .write_end = gfs2_write_end,
1158         .set_page_dirty = gfs2_set_page_dirty,
1159         .bmap = gfs2_bmap,
1160         .invalidatepage = gfs2_invalidatepage,
1161         .releasepage = gfs2_releasepage,
1162         .is_partially_uptodate = block_is_partially_uptodate,
1163         .error_remove_page = generic_error_remove_page,
1164 };
1165
1166 void gfs2_set_aops(struct inode *inode)
1167 {
1168         struct gfs2_inode *ip = GFS2_I(inode);
1169
1170         if (gfs2_is_writeback(ip))
1171                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1172         else if (gfs2_is_ordered(ip))
1173                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1174         else if (gfs2_is_jdata(ip))
1175                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1176         else
1177                 BUG();
1178 }
1179