fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29
  30 #define MLOG_MASK_PREFIX ML_FILE_IO
  31 #include <cluster/masklog.h>
  32
  33 #include "ocfs2.h"
  34
  35 #include "alloc.h"
  36 #include "aops.h"
  37 #include "dlmglue.h"
  38 #include "extent_map.h"
  39 #include "file.h"
  40 #include "inode.h"
  41 #include "journal.h"
  42 #include "suballoc.h"
  43 #include "super.h"
  44 #include "symlink.h"
  45
  46 #include "buffer_head_io.h"
  47
  48 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  49                                    struct buffer_head *bh_result, int create)
  50 {
  51         int err = -EIO;
  52         int status;
  53         struct ocfs2_dinode *fe = NULL;
  54         struct buffer_head *bh = NULL;
  55         struct buffer_head *buffer_cache_bh = NULL;
  56         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  57         void *kaddr;
  58
  59         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  60                    (unsigned long long)iblock, bh_result, create);
  61
  62         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  63
  64         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  65                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  66                      (unsigned long long)iblock);
  67                 goto bail;
  68         }
  69
  70         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
  71                                   OCFS2_I(inode)->ip_blkno,
  72                                   &bh, OCFS2_BH_CACHED, inode);
  73         if (status < 0) {
  74                 mlog_errno(status);
  75                 goto bail;
  76         }
  77         fe = (struct ocfs2_dinode *) bh->b_data;
  78
  79         if (!OCFS2_IS_VALID_DINODE(fe)) {
  80                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
  81                      (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
  82                      fe->i_signature);
  83                 goto bail;
  84         }
  85
  86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  87                                                     le32_to_cpu(fe->i_clusters))) {
  88                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  89                      "%llu\n", (unsigned long long)iblock);
  90                 goto bail;
  91         }
  92
  93         /* We don't use the page cache to create symlink data, so if
  94          * need be, copy it over from the buffer cache. */
  95         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  96                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  97                             iblock;
  98                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  99                 if (!buffer_cache_bh) {
 100                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 101                         goto bail;
 102                 }
 103
 104                 /* we haven't locked out transactions, so a commit
 105                  * could've happened. Since we've got a reference on
 106                  * the bh, even if it commits while we're doing the
 107                  * copy, the data is still good. */
 108                 if (buffer_jbd(buffer_cache_bh)
 109                     && ocfs2_inode_is_new(inode)) {
 110                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 111                         if (!kaddr) {
 112                                 mlog(ML_ERROR, "couldn't kmap!\n");
 113                                 goto bail;
 114                         }
 115                         memcpy(kaddr + (bh_result->b_size * iblock),
 116                                buffer_cache_bh->b_data,
 117                                bh_result->b_size);
 118                         kunmap_atomic(kaddr, KM_USER0);
 119                         set_buffer_uptodate(bh_result);
 120                 }
 121                 brelse(buffer_cache_bh);
 122         }
 123
 124         map_bh(bh_result, inode->i_sb,
 125                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 126
 127         err = 0;
 128
 129 bail:
 130         if (bh)
 131                 brelse(bh);
 132
 133         mlog_exit(err);
 134         return err;
 135 }
 136
 137 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
 138                            struct buffer_head *bh_result, int create)
 139 {
 140         int err = 0;
 141         unsigned int ext_flags;
 142         u64 p_blkno, past_eof;
 143         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 144
 145         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 146                    (unsigned long long)iblock, bh_result, create);
 147
 148         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 149                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 150                      inode, inode->i_ino);
 151
 152         if (S_ISLNK(inode->i_mode)) {
 153                 /* this always does I/O for some reason. */
 154                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 155                 goto bail;
 156         }
 157
 158         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL,
 159                                           &ext_flags);
 160         if (err) {
 161                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 162                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 163                      (unsigned long long)p_blkno);
 164                 goto bail;
 165         }
 166
 167         /*
 168          * ocfs2 never allocates in this function - the only time we
 169          * need to use BH_New is when we're extending i_size on a file
 170          * system which doesn't support holes, in which case BH_New
 171          * allows block_prepare_write() to zero.
 172          */
 173         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
 174                         "ino %lu, iblock %llu\n", inode->i_ino,
 175                         (unsigned long long)iblock);
 176
 177         /* Treat the unwritten extent as a hole for zeroing purposes. */
 178         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 179                 map_bh(bh_result, inode->i_sb, p_blkno);
 180
 181         if (!ocfs2_sparse_alloc(osb)) {
 182                 if (p_blkno == 0) {
 183                         err = -EIO;
 184                         mlog(ML_ERROR,
 185                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 186                              (unsigned long long)iblock,
 187                              (unsigned long long)p_blkno,
 188                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 189                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 190                         dump_stack();
 191                 }
 192
 193                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 194                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 195                      (unsigned long long)past_eof);
 196
 197                 if (create && (iblock >= past_eof))
 198                         set_buffer_new(bh_result);
 199         }
 200
 201 bail:
 202         if (err < 0)
 203                 err = -EIO;
 204
 205         mlog_exit(err);
 206         return err;
 207 }
 208
 209 static int ocfs2_readpage(struct file *file, struct page *page)
 210 {
 211         struct inode *inode = page->mapping->host;
 212         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 213         int ret, unlock = 1;
 214
 215         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 216
 217         ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
 218         if (ret != 0) {
 219                 if (ret == AOP_TRUNCATED_PAGE)
 220                         unlock = 0;
 221                 mlog_errno(ret);
 222                 goto out;
 223         }
 224
 225         if (down_read_trylock(&OCFS2_I(inode)->ip_alloc_sem) == 0) {
 226                 ret = AOP_TRUNCATED_PAGE;
 227                 goto out_meta_unlock;
 228         }
 229
 230         /*
 231          * i_size might have just been updated as we grabed the meta lock.  We
 232          * might now be discovering a truncate that hit on another node.
 233          * block_read_full_page->get_block freaks out if it is asked to read
 234          * beyond the end of a file, so we check here.  Callers
 235          * (generic_file_read, fault->nopage) are clever enough to check i_size
 236          * and notice that the page they just read isn't needed.
 237          *
 238          * XXX sys_readahead() seems to get that wrong?
 239          */
 240         if (start >= i_size_read(inode)) {
 241                 zero_user_page(page, 0, PAGE_SIZE, KM_USER0);
 242                 SetPageUptodate(page);
 243                 ret = 0;
 244                 goto out_alloc;
 245         }
 246
 247         ret = ocfs2_data_lock_with_page(inode, 0, page);
 248         if (ret != 0) {
 249                 if (ret == AOP_TRUNCATED_PAGE)
 250                         unlock = 0;
 251                 mlog_errno(ret);
 252                 goto out_alloc;
 253         }
 254
 255         ret = block_read_full_page(page, ocfs2_get_block);
 256         unlock = 0;
 257
 258         ocfs2_data_unlock(inode, 0);
 259 out_alloc:
 260         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 261 out_meta_unlock:
 262         ocfs2_meta_unlock(inode, 0);
 263 out:
 264         if (unlock)
 265                 unlock_page(page);
 266         mlog_exit(ret);
 267         return ret;
 268 }
 269
 270 /* Note: Because we don't support holes, our allocation has
 271  * already happened (allocation writes zeros to the file data)
 272  * so we don't have to worry about ordered writes in
 273  * ocfs2_writepage.
 274  *
 275  * ->writepage is called during the process of invalidating the page cache
 276  * during blocked lock processing.  It can't block on any cluster locks
 277  * to during block mapping.  It's relying on the fact that the block
 278  * mapping can't have disappeared under the dirty pages that it is
 279  * being asked to write back.
 280  */
 281 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 282 {
 283         int ret;
 284
 285         mlog_entry("(0x%p)\n", page);
 286
 287         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 288
 289         mlog_exit(ret);
 290
 291         return ret;
 292 }
 293
 294 /*
 295  * This is called from ocfs2_write_zero_page() which has handled it's
 296  * own cluster locking and has ensured allocation exists for those
 297  * blocks to be written.
 298  */
 299 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 300                                unsigned from, unsigned to)
 301 {
 302         int ret;
 303
 304         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 305
 306         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 307
 308         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 309
 310         return ret;
 311 }
 312
 313 /* Taken from ext3. We don't necessarily need the full blown
 314  * functionality yet, but IMHO it's better to cut and paste the whole
 315  * thing so we can avoid introducing our own bugs (and easily pick up
 316  * their fixes when they happen) --Mark */
 317 int walk_page_buffers(  handle_t *handle,
 318                         struct buffer_head *head,
 319                         unsigned from,
 320                         unsigned to,
 321                         int *partial,
 322                         int (*fn)(      handle_t *handle,
 323                                         struct buffer_head *bh))
 324 {
 325         struct buffer_head *bh;
 326         unsigned block_start, block_end;
 327         unsigned blocksize = head->b_size;
 328         int err, ret = 0;
 329         struct buffer_head *next;
 330
 331         for (   bh = head, block_start = 0;
 332                 ret == 0 && (bh != head || !block_start);
 333                 block_start = block_end, bh = next)
 334         {
 335                 next = bh->b_this_page;
 336                 block_end = block_start + blocksize;
 337                 if (block_end <= from || block_start >= to) {
 338                         if (partial && !buffer_uptodate(bh))
 339                                 *partial = 1;
 340                         continue;
 341                 }
 342                 err = (*fn)(handle, bh);
 343                 if (!ret)
 344                         ret = err;
 345         }
 346         return ret;
 347 }
 348
 349 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 350                                                          struct page *page,
 351                                                          unsigned from,
 352                                                          unsigned to)
 353 {
 354         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 355         handle_t *handle = NULL;
 356         int ret = 0;
 357
 358         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 359         if (!handle) {
 360                 ret = -ENOMEM;
 361                 mlog_errno(ret);
 362                 goto out;
 363         }
 364
 365         if (ocfs2_should_order_data(inode)) {
 366                 ret = walk_page_buffers(handle,
 367                                         page_buffers(page),
 368                                         from, to, NULL,
 369                                         ocfs2_journal_dirty_data);
 370                 if (ret < 0)
 371                         mlog_errno(ret);
 372         }
 373 out:
 374         if (ret) {
 375                 if (handle)
 376                         ocfs2_commit_trans(osb, handle);
 377                 handle = ERR_PTR(ret);
 378         }
 379         return handle;
 380 }
 381
 382 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 383 {
 384         sector_t status;
 385         u64 p_blkno = 0;
 386         int err = 0;
 387         struct inode *inode = mapping->host;
 388
 389         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 390
 391         /* We don't need to lock journal system files, since they aren't
 392          * accessed concurrently from multiple nodes.
 393          */
 394         if (!INODE_JOURNAL(inode)) {
 395                 err = ocfs2_meta_lock(inode, NULL, 0);
 396                 if (err) {
 397                         if (err != -ENOENT)
 398                                 mlog_errno(err);
 399                         goto bail;
 400                 }
 401                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 402         }
 403
 404         err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL);
 405
 406         if (!INODE_JOURNAL(inode)) {
 407                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 408                 ocfs2_meta_unlock(inode, 0);
 409         }
 410
 411         if (err) {
 412                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 413                      (unsigned long long)block);
 414                 mlog_errno(err);
 415                 goto bail;
 416         }
 417
 418
 419 bail:
 420         status = err ? 0 : p_blkno;
 421
 422         mlog_exit((int)status);
 423
 424         return status;
 425 }
 426
 427 /*
 428  * TODO: Make this into a generic get_blocks function.
 429  *
 430  * From do_direct_io in direct-io.c:
 431  *  "So what we do is to permit the ->get_blocks function to populate
 432  *   bh.b_size with the size of IO which is permitted at this offset and
 433  *   this i_blkbits."
 434  *
 435  * This function is called directly from get_more_blocks in direct-io.c.
 436  *
 437  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 438  *                                      fs_count, map_bh, dio->rw == WRITE);
 439  */
 440 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 441                                      struct buffer_head *bh_result, int create)
 442 {
 443         int ret;
 444         u64 p_blkno, inode_blocks, contig_blocks;
 445         unsigned int ext_flags;
 446         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 447         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 448
 449         /* This function won't even be called if the request isn't all
 450          * nicely aligned and of the right size, so there's no need
 451          * for us to check any of that. */
 452
 453         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 454
 455         /*
 456          * Any write past EOF is not allowed because we'd be extending.
 457          */
 458         if (create && (iblock + max_blocks) > inode_blocks) {
 459                 ret = -EIO;
 460                 goto bail;
 461         }
 462
 463         /* This figures out the size of the next contiguous block, and
 464          * our logical offset */
 465         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 466                                           &contig_blocks, &ext_flags);
 467         if (ret) {
 468                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 469                      (unsigned long long)iblock);
 470                 ret = -EIO;
 471                 goto bail;
 472         }
 473
 474         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
 475                 ocfs2_error(inode->i_sb,
 476                             "Inode %llu has a hole at block %llu\n",
 477                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 478                             (unsigned long long)iblock);
 479                 ret = -EROFS;
 480                 goto bail;
 481         }
 482
 483         /*
 484          * get_more_blocks() expects us to describe a hole by clearing
 485          * the mapped bit on bh_result().
 486          *
 487          * Consider an unwritten extent as a hole.
 488          */
 489         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 490                 map_bh(bh_result, inode->i_sb, p_blkno);
 491         else {
 492                 /*
 493                  * ocfs2_prepare_inode_for_write() should have caught
 494                  * the case where we'd be filling a hole and triggered
 495                  * a buffered write instead.
 496                  */
 497                 if (create) {
 498                         ret = -EIO;
 499                         mlog_errno(ret);
 500                         goto bail;
 501                 }
 502
 503                 clear_buffer_mapped(bh_result);
 504         }
 505
 506         /* make sure we don't map more than max_blocks blocks here as
 507            that's all the kernel will handle at this point. */
 508         if (max_blocks < contig_blocks)
 509                 contig_blocks = max_blocks;
 510         bh_result->b_size = contig_blocks << blocksize_bits;
 511 bail:
 512         return ret;
 513 }
 514
 515 /*
 516  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 517  * particularly interested in the aio/dio case.  Like the core uses
 518  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 519  * truncation on another.
 520  */
 521 static void ocfs2_dio_end_io(struct kiocb *iocb,
 522                              loff_t offset,
 523                              ssize_t bytes,
 524                              void *private)
 525 {
 526         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 527         int level;
 528
 529         /* this io's submitter should not have unlocked this before we could */
 530         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 531
 532         ocfs2_iocb_clear_rw_locked(iocb);
 533
 534         level = ocfs2_iocb_rw_locked_level(iocb);
 535         if (!level)
 536                 up_read(&inode->i_alloc_sem);
 537         ocfs2_rw_unlock(inode, level);
 538 }
 539
 540 /*
 541  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 542  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 543  * do journalled data.
 544  */
 545 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 546 {
 547         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 548
 549         journal_invalidatepage(journal, page, offset);
 550 }
 551
 552 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 553 {
 554         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 555
 556         if (!page_has_buffers(page))
 557                 return 0;
 558         return journal_try_to_free_buffers(journal, page, wait);
 559 }
 560
 561 static ssize_t ocfs2_direct_IO(int rw,
 562                                struct kiocb *iocb,
 563                                const struct iovec *iov,
 564                                loff_t offset,
 565                                unsigned long nr_segs)
 566 {
 567         struct file *file = iocb->ki_filp;
 568         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 569         int ret;
 570
 571         mlog_entry_void();
 572
 573         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 574                 /*
 575                  * We get PR data locks even for O_DIRECT.  This
 576                  * allows concurrent O_DIRECT I/O but doesn't let
 577                  * O_DIRECT with extending and buffered zeroing writes
 578                  * race.  If they did race then the buffered zeroing
 579                  * could be written back after the O_DIRECT I/O.  It's
 580                  * one thing to tell people not to mix buffered and
 581                  * O_DIRECT writes, but expecting them to understand
 582                  * that file extension is also an implicit buffered
 583                  * write is too much.  By getting the PR we force
 584                  * writeback of the buffered zeroing before
 585                  * proceeding.
 586                  */
 587                 ret = ocfs2_data_lock(inode, 0);
 588                 if (ret < 0) {
 589                         mlog_errno(ret);
 590                         goto out;
 591                 }
 592                 ocfs2_data_unlock(inode, 0);
 593         }
 594
 595         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 596                                             inode->i_sb->s_bdev, iov, offset,
 597                                             nr_segs,
 598                                             ocfs2_direct_IO_get_blocks,
 599                                             ocfs2_dio_end_io);
 600 out:
 601         mlog_exit(ret);
 602         return ret;
 603 }
 604
 605 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 606                                             u32 cpos,
 607                                             unsigned int *start,
 608                                             unsigned int *end)
 609 {
 610         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 611
 612         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 613                 unsigned int cpp;
 614
 615                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 616
 617                 cluster_start = cpos % cpp;
 618                 cluster_start = cluster_start << osb->s_clustersize_bits;
 619
 620                 cluster_end = cluster_start + osb->s_clustersize;
 621         }
 622
 623         BUG_ON(cluster_start > PAGE_SIZE);
 624         BUG_ON(cluster_end > PAGE_SIZE);
 625
 626         if (start)
 627                 *start = cluster_start;
 628         if (end)
 629                 *end = cluster_end;
 630 }
 631
 632 /*
 633  * 'from' and 'to' are the region in the page to avoid zeroing.
 634  *
 635  * If pagesize > clustersize, this function will avoid zeroing outside
 636  * of the cluster boundary.
 637  *
 638  * from == to == 0 is code for "zero the entire cluster region"
 639  */
 640 static void ocfs2_clear_page_regions(struct page *page,
 641                                      struct ocfs2_super *osb, u32 cpos,
 642                                      unsigned from, unsigned to)
 643 {
 644         void *kaddr;
 645         unsigned int cluster_start, cluster_end;
 646
 647         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 648
 649         kaddr = kmap_atomic(page, KM_USER0);
 650
 651         if (from || to) {
 652                 if (from > cluster_start)
 653                         memset(kaddr + cluster_start, 0, from - cluster_start);
 654                 if (to < cluster_end)
 655                         memset(kaddr + to, 0, cluster_end - to);
 656         } else {
 657                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 658         }
 659
 660         kunmap_atomic(kaddr, KM_USER0);
 661 }
 662
 663 /*
 664  * Some of this taken from block_prepare_write(). We already have our
 665  * mapping by now though, and the entire write will be allocating or
 666  * it won't, so not much need to use BH_New.
 667  *
 668  * This will also skip zeroing, which is handled externally.
 669  */
 670 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 671                           struct inode *inode, unsigned int from,
 672                           unsigned int to, int new)
 673 {
 674         int ret = 0;
 675         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 676         unsigned int block_end, block_start;
 677         unsigned int bsize = 1 << inode->i_blkbits;
 678
 679         if (!page_has_buffers(page))
 680                 create_empty_buffers(page, bsize, 0);
 681
 682         head = page_buffers(page);
 683         for (bh = head, block_start = 0; bh != head || !block_start;
 684              bh = bh->b_this_page, block_start += bsize) {
 685                 block_end = block_start + bsize;
 686
 687                 clear_buffer_new(bh);
 688
 689                 /*
 690                  * Ignore blocks outside of our i/o range -
 691                  * they may belong to unallocated clusters.
 692                  */
 693                 if (block_start >= to || block_end <= from) {
 694                         if (PageUptodate(page))
 695                                 set_buffer_uptodate(bh);
 696                         continue;
 697                 }
 698
 699                 /*
 700                  * For an allocating write with cluster size >= page
 701                  * size, we always write the entire page.
 702                  */
 703                 if (new)
 704                         set_buffer_new(bh);
 705
 706                 if (!buffer_mapped(bh)) {
 707                         map_bh(bh, inode->i_sb, *p_blkno);
 708                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 709                 }
 710
 711                 if (PageUptodate(page)) {
 712                         if (!buffer_uptodate(bh))
 713                                 set_buffer_uptodate(bh);
 714                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 715                            !buffer_new(bh) &&
 716                            (block_start < from || block_end > to)) {
 717                         ll_rw_block(READ, 1, &bh);
 718                         *wait_bh++=bh;
 719                 }
 720
 721                 *p_blkno = *p_blkno + 1;
 722         }
 723
 724         /*
 725          * If we issued read requests - let them complete.
 726          */
 727         while(wait_bh > wait) {
 728                 wait_on_buffer(*--wait_bh);
 729                 if (!buffer_uptodate(*wait_bh))
 730                         ret = -EIO;
 731         }
 732
 733         if (ret == 0 || !new)
 734                 return ret;
 735
 736         /*
 737          * If we get -EIO above, zero out any newly allocated blocks
 738          * to avoid exposing stale data.
 739          */
 740         bh = head;
 741         block_start = 0;
 742         do {
 743                 void *kaddr;
 744
 745                 block_end = block_start + bsize;
 746                 if (block_end <= from)
 747                         goto next_bh;
 748                 if (block_start >= to)
 749                         break;
 750
 751                 kaddr = kmap_atomic(page, KM_USER0);
 752                 memset(kaddr+block_start, 0, bh->b_size);
 753                 flush_dcache_page(page);
 754                 kunmap_atomic(kaddr, KM_USER0);
 755                 set_buffer_uptodate(bh);
 756                 mark_buffer_dirty(bh);
 757
 758 next_bh:
 759                 block_start = block_end;
 760                 bh = bh->b_this_page;
 761         } while (bh != head);
 762
 763         return ret;
 764 }
 765
 766 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 767 #define OCFS2_MAX_CTXT_PAGES    1
 768 #else
 769 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 770 #endif
 771
 772 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 773
 774 /*
 775  * Describe the state of a single cluster to be written to.
 776  */
 777 struct ocfs2_write_cluster_desc {
 778         u32             c_cpos;
 779         u32             c_phys;
 780         /*
 781          * Give this a unique field because c_phys eventually gets
 782          * filled.
 783          */
 784         unsigned        c_new;
 785         unsigned        c_unwritten;
 786 };
 787
 788 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 789 {
 790         return d->c_new || d->c_unwritten;
 791 }
 792
 793 struct ocfs2_write_ctxt {
 794         /* Logical cluster position / len of write */
 795         u32                             w_cpos;
 796         u32                             w_clen;
 797
 798         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 799
 800         /*
 801          * This is true if page_size > cluster_size.
 802          *
 803          * It triggers a set of special cases during write which might
 804          * have to deal with allocating writes to partial pages.
 805          */
 806         unsigned int                    w_large_pages;
 807
 808         /*
 809          * Pages involved in this write.
 810          *
 811          * w_target_page is the page being written to by the user.
 812          *
 813          * w_pages is an array of pages which always contains
 814          * w_target_page, and in the case of an allocating write with
 815          * page_size < cluster size, it will contain zero'd and mapped
 816          * pages adjacent to w_target_page which need to be written
 817          * out in so that future reads from that region will get
 818          * zero's.
 819          */
 820         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 821         unsigned int                    w_num_pages;
 822         struct page                     *w_target_page;
 823
 824         /*
 825          * ocfs2_write_end() uses this to know what the real range to
 826          * write in the target should be.
 827          */
 828         unsigned int                    w_target_from;
 829         unsigned int                    w_target_to;
 830
 831         /*
 832          * We could use journal_current_handle() but this is cleaner,
 833          * IMHO -Mark
 834          */
 835         handle_t                        *w_handle;
 836
 837         struct buffer_head              *w_di_bh;
 838
 839         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 840 };
 841
 842 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 843 {
 844         int i;
 845
 846         for(i = 0; i < wc->w_num_pages; i++) {
 847                 if (wc->w_pages[i] == NULL)
 848                         continue;
 849
 850                 unlock_page(wc->w_pages[i]);
 851                 mark_page_accessed(wc->w_pages[i]);
 852                 page_cache_release(wc->w_pages[i]);
 853         }
 854
 855         brelse(wc->w_di_bh);
 856         kfree(wc);
 857 }
 858
 859 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 860                                   struct ocfs2_super *osb, loff_t pos,
 861                                   unsigned len, struct buffer_head *di_bh)
 862 {
 863         struct ocfs2_write_ctxt *wc;
 864
 865         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 866         if (!wc)
 867                 return -ENOMEM;
 868
 869         wc->w_cpos = pos >> osb->s_clustersize_bits;
 870         wc->w_clen = ocfs2_clusters_for_bytes(osb->sb, len);
 871         get_bh(di_bh);
 872         wc->w_di_bh = di_bh;
 873
 874         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 875                 wc->w_large_pages = 1;
 876         else
 877                 wc->w_large_pages = 0;
 878
 879         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 880
 881         *wcp = wc;
 882
 883         return 0;
 884 }
 885
 886 /*
 887  * If a page has any new buffers, zero them out here, and mark them uptodate
 888  * and dirty so they'll be written out (in order to prevent uninitialised
 889  * block data from leaking). And clear the new bit.
 890  */
 891 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 892 {
 893         unsigned int block_start, block_end;
 894         struct buffer_head *head, *bh;
 895
 896         BUG_ON(!PageLocked(page));
 897         if (!page_has_buffers(page))
 898                 return;
 899
 900         bh = head = page_buffers(page);
 901         block_start = 0;
 902         do {
 903                 block_end = block_start + bh->b_size;
 904
 905                 if (buffer_new(bh)) {
 906                         if (block_end > from && block_start < to) {
 907                                 if (!PageUptodate(page)) {
 908                                         unsigned start, end;
 909                                         void *kaddr;
 910
 911                                         start = max(from, block_start);
 912                                         end = min(to, block_end);
 913
 914                                         kaddr = kmap_atomic(page, KM_USER0);
 915                                         memset(kaddr+start, 0, end - start);
 916                                         flush_dcache_page(page);
 917                                         kunmap_atomic(kaddr, KM_USER0);
 918                                         set_buffer_uptodate(bh);
 919                                 }
 920
 921                                 clear_buffer_new(bh);
 922                                 mark_buffer_dirty(bh);
 923                         }
 924                 }
 925
 926                 block_start = block_end;
 927                 bh = bh->b_this_page;
 928         } while (bh != head);
 929 }
 930
 931 /*
 932  * Only called when we have a failure during allocating write to write
 933  * zero's to the newly allocated region.
 934  */
 935 static void ocfs2_write_failure(struct inode *inode,
 936                                 struct ocfs2_write_ctxt *wc,
 937                                 loff_t user_pos, unsigned user_len)
 938 {
 939         int i;
 940         unsigned from, to;
 941         struct page *tmppage;
 942
 943         ocfs2_zero_new_buffers(wc->w_target_page, user_pos, user_len);
 944
 945         if (wc->w_large_pages) {
 946                 from = wc->w_target_from;
 947                 to = wc->w_target_to;
 948         } else {
 949                 from = 0;
 950                 to = PAGE_CACHE_SIZE;
 951         }
 952
 953         for(i = 0; i < wc->w_num_pages; i++) {
 954                 tmppage = wc->w_pages[i];
 955
 956                 if (ocfs2_should_order_data(inode))
 957                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
 958                                           from, to, NULL,
 959                                           ocfs2_journal_dirty_data);
 960
 961                 block_commit_write(tmppage, from, to);
 962         }
 963 }
 964
 965 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 966                                         struct ocfs2_write_ctxt *wc,
 967                                         struct page *page, u32 cpos,
 968                                         loff_t user_pos, unsigned user_len,
 969                                         int new)
 970 {
 971         int ret;
 972         unsigned int map_from = 0, map_to = 0;
 973         unsigned int cluster_start, cluster_end;
 974         unsigned int user_data_from = 0, user_data_to = 0;
 975
 976         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 977                                         &cluster_start, &cluster_end);
 978
 979         if (page == wc->w_target_page) {
 980                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
 981                 map_to = map_from + user_len;
 982
 983                 if (new)
 984                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 985                                                     cluster_start, cluster_end,
 986                                                     new);
 987                 else
 988                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 989                                                     map_from, map_to, new);
 990                 if (ret) {
 991                         mlog_errno(ret);
 992                         goto out;
 993                 }
 994
 995                 user_data_from = map_from;
 996                 user_data_to = map_to;
 997                 if (new) {
 998                         map_from = cluster_start;
 999                         map_to = cluster_end;
1000                 }
1001
1002                 wc->w_target_from = map_from;
1003                 wc->w_target_to = map_to;
1004         } else {
1005                 /*
1006                  * If we haven't allocated the new page yet, we
1007                  * shouldn't be writing it out without copying user
1008                  * data. This is likely a math error from the caller.
1009                  */
1010                 BUG_ON(!new);
1011
1012                 map_from = cluster_start;
1013                 map_to = cluster_end;
1014
1015                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1016                                             cluster_start, cluster_end, new);
1017                 if (ret) {
1018                         mlog_errno(ret);
1019                         goto out;
1020                 }
1021         }
1022
1023         /*
1024          * Parts of newly allocated pages need to be zero'd.
1025          *
1026          * Above, we have also rewritten 'to' and 'from' - as far as
1027          * the rest of the function is concerned, the entire cluster
1028          * range inside of a page needs to be written.
1029          *
1030          * We can skip this if the page is up to date - it's already
1031          * been zero'd from being read in as a hole.
1032          */
1033         if (new && !PageUptodate(page))
1034                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1035                                          cpos, user_data_from, user_data_to);
1036
1037         flush_dcache_page(page);
1038
1039 out:
1040         return ret;
1041 }
1042
1043 /*
1044  * This function will only grab one clusters worth of pages.
1045  */
1046 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1047                                       struct ocfs2_write_ctxt *wc,
1048                                       u32 cpos, loff_t user_pos, int new,
1049                                       struct page *mmap_page)
1050 {
1051         int ret = 0, i;
1052         unsigned long start, target_index, index;
1053         struct inode *inode = mapping->host;
1054
1055         target_index = user_pos >> PAGE_CACHE_SHIFT;
1056
1057         /*
1058          * Figure out how many pages we'll be manipulating here. For
1059          * non allocating write, we just change the one
1060          * page. Otherwise, we'll need a whole clusters worth.
1061          */
1062         if (new) {
1063                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1064                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1065         } else {
1066                 wc->w_num_pages = 1;
1067                 start = target_index;
1068         }
1069
1070         for(i = 0; i < wc->w_num_pages; i++) {
1071                 index = start + i;
1072
1073                 if (index == target_index && mmap_page) {
1074                         /*
1075                          * ocfs2_pagemkwrite() is a little different
1076                          * and wants us to directly use the page
1077                          * passed in.
1078                          */
1079                         lock_page(mmap_page);
1080
1081                         if (mmap_page->mapping != mapping) {
1082                                 unlock_page(mmap_page);
1083                                 /*
1084                                  * Sanity check - the locking in
1085                                  * ocfs2_pagemkwrite() should ensure
1086                                  * that this code doesn't trigger.
1087                                  */
1088                                 ret = -EINVAL;
1089                                 mlog_errno(ret);
1090                                 goto out;
1091                         }
1092
1093                         page_cache_get(mmap_page);
1094                         wc->w_pages[i] = mmap_page;
1095                 } else {
1096                         wc->w_pages[i] = find_or_create_page(mapping, index,
1097                                                              GFP_NOFS);
1098                         if (!wc->w_pages[i]) {
1099                                 ret = -ENOMEM;
1100                                 mlog_errno(ret);
1101                                 goto out;
1102                         }
1103                 }
1104
1105                 if (index == target_index)
1106                         wc->w_target_page = wc->w_pages[i];
1107         }
1108 out:
1109         return ret;
1110 }
1111
1112 /*
1113  * Prepare a single cluster for write one cluster into the file.
1114  */
1115 static int ocfs2_write_cluster(struct address_space *mapping,
1116                                u32 phys, unsigned int unwritten,
1117                                struct ocfs2_alloc_context *data_ac,
1118                                struct ocfs2_alloc_context *meta_ac,
1119                                struct ocfs2_write_ctxt *wc, u32 cpos,
1120                                loff_t user_pos, unsigned user_len)
1121 {
1122         int ret, i, new, should_zero = 0;
1123         u64 v_blkno, p_blkno;
1124         struct inode *inode = mapping->host;
1125
1126         new = phys == 0 ? 1 : 0;
1127         if (new || unwritten)
1128                 should_zero = 1;
1129
1130         if (new) {
1131                 u32 tmp_pos;
1132
1133                 /*
1134                  * This is safe to call with the page locks - it won't take
1135                  * any additional semaphores or cluster locks.
1136                  */
1137                 tmp_pos = cpos;
1138                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1139                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1140                                                  wc->w_handle, data_ac,
1141                                                  meta_ac, NULL);
1142                 /*
1143                  * This shouldn't happen because we must have already
1144                  * calculated the correct meta data allocation required. The
1145                  * internal tree allocation code should know how to increase
1146                  * transaction credits itself.
1147                  *
1148                  * If need be, we could handle -EAGAIN for a
1149                  * RESTART_TRANS here.
1150                  */
1151                 mlog_bug_on_msg(ret == -EAGAIN,
1152                                 "Inode %llu: EAGAIN return during allocation.\n",
1153                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1154                 if (ret < 0) {
1155                         mlog_errno(ret);
1156                         goto out;
1157                 }
1158         } else if (unwritten) {
1159                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1160                                                 wc->w_handle, cpos, 1, phys,
1161                                                 meta_ac, &wc->w_dealloc);
1162                 if (ret < 0) {
1163                         mlog_errno(ret);
1164                         goto out;
1165                 }
1166         }
1167
1168         if (should_zero)
1169                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1170         else
1171                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1172
1173         /*
1174          * The only reason this should fail is due to an inability to
1175          * find the extent added.
1176          */
1177         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1178                                           NULL);
1179         if (ret < 0) {
1180                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1181                             "at logical block %llu",
1182                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1183                             (unsigned long long)v_blkno);
1184                 goto out;
1185         }
1186
1187         BUG_ON(p_blkno == 0);
1188
1189         for(i = 0; i < wc->w_num_pages; i++) {
1190                 int tmpret;
1191
1192                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1193                                                       wc->w_pages[i], cpos,
1194                                                       user_pos, user_len,
1195                                                       should_zero);
1196                 if (tmpret) {
1197                         mlog_errno(tmpret);
1198                         if (ret == 0)
1199                                 tmpret = ret;
1200                 }
1201         }
1202
1203         /*
1204          * We only have cleanup to do in case of allocating write.
1205          */
1206         if (ret && new)
1207                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1208
1209 out:
1210
1211         return ret;
1212 }
1213
1214 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1215                                        struct ocfs2_alloc_context *data_ac,
1216                                        struct ocfs2_alloc_context *meta_ac,
1217                                        struct ocfs2_write_ctxt *wc,
1218                                        loff_t pos, unsigned len)
1219 {
1220         int ret, i;
1221         struct ocfs2_write_cluster_desc *desc;
1222
1223         for (i = 0; i < wc->w_clen; i++) {
1224                 desc = &wc->w_desc[i];
1225
1226                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1227                                           desc->c_unwritten, data_ac, meta_ac,
1228                                           wc, desc->c_cpos, pos, len);
1229                 if (ret) {
1230                         mlog_errno(ret);
1231                         goto out;
1232                 }
1233         }
1234
1235         ret = 0;
1236 out:
1237         return ret;
1238 }
1239
1240 /*
1241  * ocfs2_write_end() wants to know which parts of the target page it
1242  * should complete the write on. It's easiest to compute them ahead of
1243  * time when a more complete view of the write is available.
1244  */
1245 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1246                                         struct ocfs2_write_ctxt *wc,
1247                                         loff_t pos, unsigned len, int alloc)
1248 {
1249         struct ocfs2_write_cluster_desc *desc;
1250
1251         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1252         wc->w_target_to = wc->w_target_from + len;
1253
1254         if (alloc == 0)
1255                 return;
1256
1257         /*
1258          * Allocating write - we may have different boundaries based
1259          * on page size and cluster size.
1260          *
1261          * NOTE: We can no longer compute one value from the other as
1262          * the actual write length and user provided length may be
1263          * different.
1264          */
1265
1266         if (wc->w_large_pages) {
1267                 /*
1268                  * We only care about the 1st and last cluster within
1269                  * our range and whether they should be zero'd or not. Either
1270                  * value may be extended out to the start/end of a
1271                  * newly allocated cluster.
1272                  */
1273                 desc = &wc->w_desc[0];
1274                 if (ocfs2_should_zero_cluster(desc))
1275                         ocfs2_figure_cluster_boundaries(osb,
1276                                                         desc->c_cpos,
1277                                                         &wc->w_target_from,
1278                                                         NULL);
1279
1280                 desc = &wc->w_desc[wc->w_clen - 1];
1281                 if (ocfs2_should_zero_cluster(desc))
1282                         ocfs2_figure_cluster_boundaries(osb,
1283                                                         desc->c_cpos,
1284                                                         NULL,
1285                                                         &wc->w_target_to);
1286         } else {
1287                 wc->w_target_from = 0;
1288                 wc->w_target_to = PAGE_CACHE_SIZE;
1289         }
1290 }
1291
1292 /*
1293  * Populate each single-cluster write descriptor in the write context
1294  * with information about the i/o to be done.
1295  *
1296  * Returns the number of clusters that will have to be allocated, as
1297  * well as a worst case estimate of the number of extent records that
1298  * would have to be created during a write to an unwritten region.
1299  */
1300 static int ocfs2_populate_write_desc(struct inode *inode,
1301                                      struct ocfs2_write_ctxt *wc,
1302                                      unsigned int *clusters_to_alloc,
1303                                      unsigned int *extents_to_split)
1304 {
1305         int ret;
1306         struct ocfs2_write_cluster_desc *desc;
1307         unsigned int num_clusters = 0;
1308         unsigned int ext_flags = 0;
1309         u32 phys = 0;
1310         int i;
1311
1312         *clusters_to_alloc = 0;
1313         *extents_to_split = 0;
1314
1315         for (i = 0; i < wc->w_clen; i++) {
1316                 desc = &wc->w_desc[i];
1317                 desc->c_cpos = wc->w_cpos + i;
1318
1319                 if (num_clusters == 0) {
1320                         /*
1321                          * Need to look up the next extent record.
1322                          */
1323                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1324                                                  &num_clusters, &ext_flags);
1325                         if (ret) {
1326                                 mlog_errno(ret);
1327                                 goto out;
1328                         }
1329
1330                         /*
1331                          * Assume worst case - that we're writing in
1332                          * the middle of the extent.
1333                          *
1334                          * We can assume that the write proceeds from
1335                          * left to right, in which case the extent
1336                          * insert code is smart enough to coalesce the
1337                          * next splits into the previous records created.
1338                          */
1339                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1340                                 *extents_to_split = *extents_to_split + 2;
1341                 } else if (phys) {
1342                         /*
1343                          * Only increment phys if it doesn't describe
1344                          * a hole.
1345                          */
1346                         phys++;
1347                 }
1348
1349                 desc->c_phys = phys;
1350                 if (phys == 0) {
1351                         desc->c_new = 1;
1352                         *clusters_to_alloc = *clusters_to_alloc + 1;
1353                 }
1354                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1355                         desc->c_unwritten = 1;
1356
1357                 num_clusters--;
1358         }
1359
1360         ret = 0;
1361 out:
1362         return ret;
1363 }
1364
1365 int ocfs2_write_begin_nolock(struct address_space *mapping,
1366                              loff_t pos, unsigned len, unsigned flags,
1367                              struct page **pagep, void **fsdata,
1368                              struct buffer_head *di_bh, struct page *mmap_page)
1369 {
1370         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1371         unsigned int clusters_to_alloc, extents_to_split;
1372         struct ocfs2_write_ctxt *wc;
1373         struct inode *inode = mapping->host;
1374         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1375         struct ocfs2_dinode *di;
1376         struct ocfs2_alloc_context *data_ac = NULL;
1377         struct ocfs2_alloc_context *meta_ac = NULL;
1378         handle_t *handle;
1379
1380         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1381         if (ret) {
1382                 mlog_errno(ret);
1383                 return ret;
1384         }
1385
1386         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1387                                         &extents_to_split);
1388         if (ret) {
1389                 mlog_errno(ret);
1390                 goto out;
1391         }
1392
1393         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1394
1395         /*
1396          * We set w_target_from, w_target_to here so that
1397          * ocfs2_write_end() knows which range in the target page to
1398          * write out. An allocation requires that we write the entire
1399          * cluster range.
1400          */
1401         if (clusters_to_alloc || extents_to_split) {
1402                 /*
1403                  * XXX: We are stretching the limits of
1404                  * ocfs2_lock_allocators(). It greatly over-estimates
1405                  * the work to be done.
1406                  */
1407                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1408                                             extents_to_split, &data_ac, &meta_ac);
1409                 if (ret) {
1410                         mlog_errno(ret);
1411                         goto out;
1412                 }
1413
1414                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1415                                                     clusters_to_alloc);
1416
1417         }
1418
1419         ocfs2_set_target_boundaries(osb, wc, pos, len,
1420                                     clusters_to_alloc + extents_to_split);
1421
1422         handle = ocfs2_start_trans(osb, credits);
1423         if (IS_ERR(handle)) {
1424                 ret = PTR_ERR(handle);
1425                 mlog_errno(ret);
1426                 goto out;
1427         }
1428
1429         wc->w_handle = handle;
1430
1431         /*
1432          * We don't want this to fail in ocfs2_write_end(), so do it
1433          * here.
1434          */
1435         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1436                                    OCFS2_JOURNAL_ACCESS_WRITE);
1437         if (ret) {
1438                 mlog_errno(ret);
1439                 goto out_commit;
1440         }
1441
1442         /*
1443          * Fill our page array first. That way we've grabbed enough so
1444          * that we can zero and flush if we error after adding the
1445          * extent.
1446          */
1447         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1448                                          clusters_to_alloc + extents_to_split,
1449                                          mmap_page);
1450         if (ret) {
1451                 mlog_errno(ret);
1452                 goto out_commit;
1453         }
1454
1455         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1456                                           len);
1457         if (ret) {
1458                 mlog_errno(ret);
1459                 goto out_commit;
1460         }
1461
1462         if (data_ac)
1463                 ocfs2_free_alloc_context(data_ac);
1464         if (meta_ac)
1465                 ocfs2_free_alloc_context(meta_ac);
1466
1467         *pagep = wc->w_target_page;
1468         *fsdata = wc;
1469         return 0;
1470 out_commit:
1471         ocfs2_commit_trans(osb, handle);
1472
1473 out:
1474         ocfs2_free_write_ctxt(wc);
1475
1476         if (data_ac)
1477                 ocfs2_free_alloc_context(data_ac);
1478         if (meta_ac)
1479                 ocfs2_free_alloc_context(meta_ac);
1480         return ret;
1481 }
1482
1483 int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1484                       loff_t pos, unsigned len, unsigned flags,
1485                       struct page **pagep, void **fsdata)
1486 {
1487         int ret;
1488         struct buffer_head *di_bh = NULL;
1489         struct inode *inode = mapping->host;
1490
1491         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1492         if (ret) {
1493                 mlog_errno(ret);
1494                 return ret;
1495         }
1496
1497         /*
1498          * Take alloc sem here to prevent concurrent lookups. That way
1499          * the mapping, zeroing and tree manipulation within
1500          * ocfs2_write() will be safe against ->readpage(). This
1501          * should also serve to lock out allocation from a shared
1502          * writeable region.
1503          */
1504         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1505
1506         ret = ocfs2_data_lock(inode, 1);
1507         if (ret) {
1508                 mlog_errno(ret);
1509                 goto out_fail;
1510         }
1511
1512         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1513                                        fsdata, di_bh, NULL);
1514         if (ret) {
1515                 mlog_errno(ret);
1516                 goto out_fail_data;
1517         }
1518
1519         brelse(di_bh);
1520
1521         return 0;
1522
1523 out_fail_data:
1524         ocfs2_data_unlock(inode, 1);
1525 out_fail:
1526         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1527
1528         brelse(di_bh);
1529         ocfs2_meta_unlock(inode, 1);
1530
1531         return ret;
1532 }
1533
1534 int ocfs2_write_end_nolock(struct address_space *mapping,
1535                            loff_t pos, unsigned len, unsigned copied,
1536                            struct page *page, void *fsdata)
1537 {
1538         int i;
1539         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1540         struct inode *inode = mapping->host;
1541         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1542         struct ocfs2_write_ctxt *wc = fsdata;
1543         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1544         handle_t *handle = wc->w_handle;
1545         struct page *tmppage;
1546
1547         if (unlikely(copied < len)) {
1548                 if (!PageUptodate(wc->w_target_page))
1549                         copied = 0;
1550
1551                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1552                                        start+len);
1553         }
1554         flush_dcache_page(wc->w_target_page);
1555
1556         for(i = 0; i < wc->w_num_pages; i++) {
1557                 tmppage = wc->w_pages[i];
1558
1559                 if (tmppage == wc->w_target_page) {
1560                         from = wc->w_target_from;
1561                         to = wc->w_target_to;
1562
1563                         BUG_ON(from > PAGE_CACHE_SIZE ||
1564                                to > PAGE_CACHE_SIZE ||
1565                                to < from);
1566                 } else {
1567                         /*
1568                          * Pages adjacent to the target (if any) imply
1569                          * a hole-filling write in which case we want
1570                          * to flush their entire range.
1571                          */
1572                         from = 0;
1573                         to = PAGE_CACHE_SIZE;
1574                 }
1575
1576                 if (ocfs2_should_order_data(inode))
1577                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1578                                           from, to, NULL,
1579                                           ocfs2_journal_dirty_data);
1580
1581                 block_commit_write(tmppage, from, to);
1582         }
1583
1584         pos += copied;
1585         if (pos > inode->i_size) {
1586                 i_size_write(inode, pos);
1587                 mark_inode_dirty(inode);
1588         }
1589         inode->i_blocks = ocfs2_inode_sector_count(inode);
1590         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1591         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1592         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1593         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1594         ocfs2_journal_dirty(handle, wc->w_di_bh);
1595
1596         ocfs2_commit_trans(osb, handle);
1597
1598         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1599
1600         ocfs2_free_write_ctxt(wc);
1601
1602         return copied;
1603 }
1604
1605 int ocfs2_write_end(struct file *file, struct address_space *mapping,
1606                     loff_t pos, unsigned len, unsigned copied,
1607                     struct page *page, void *fsdata)
1608 {
1609         int ret;
1610         struct inode *inode = mapping->host;
1611
1612         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1613
1614         ocfs2_data_unlock(inode, 1);
1615         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1616         ocfs2_meta_unlock(inode, 1);
1617
1618         return ret;
1619 }
1620
1621 const struct address_space_operations ocfs2_aops = {
1622         .readpage       = ocfs2_readpage,
1623         .writepage      = ocfs2_writepage,
1624         .bmap           = ocfs2_bmap,
1625         .sync_page      = block_sync_page,
1626         .direct_IO      = ocfs2_direct_IO,
1627         .invalidatepage = ocfs2_invalidatepage,
1628         .releasepage    = ocfs2_releasepage,
1629         .migratepage    = buffer_migrate_page,
1630 };