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, vm_ops->fault) 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                 block_end = block_start + bsize;
 744                 if (block_end <= from)
 745                         goto next_bh;
 746                 if (block_start >= to)
 747                         break;
 748
 749                 zero_user_page(page, block_start, bh->b_size, KM_USER0);
 750                 set_buffer_uptodate(bh);
 751                 mark_buffer_dirty(bh);
 752
 753 next_bh:
 754                 block_start = block_end;
 755                 bh = bh->b_this_page;
 756         } while (bh != head);
 757
 758         return ret;
 759 }
 760
 761 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 762 #define OCFS2_MAX_CTXT_PAGES    1
 763 #else
 764 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 765 #endif
 766
 767 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 768
 769 /*
 770  * Describe the state of a single cluster to be written to.
 771  */
 772 struct ocfs2_write_cluster_desc {
 773         u32             c_cpos;
 774         u32             c_phys;
 775         /*
 776          * Give this a unique field because c_phys eventually gets
 777          * filled.
 778          */
 779         unsigned        c_new;
 780         unsigned        c_unwritten;
 781 };
 782
 783 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 784 {
 785         return d->c_new || d->c_unwritten;
 786 }
 787
 788 struct ocfs2_write_ctxt {
 789         /* Logical cluster position / len of write */
 790         u32                             w_cpos;
 791         u32                             w_clen;
 792
 793         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 794
 795         /*
 796          * This is true if page_size > cluster_size.
 797          *
 798          * It triggers a set of special cases during write which might
 799          * have to deal with allocating writes to partial pages.
 800          */
 801         unsigned int                    w_large_pages;
 802
 803         /*
 804          * Pages involved in this write.
 805          *
 806          * w_target_page is the page being written to by the user.
 807          *
 808          * w_pages is an array of pages which always contains
 809          * w_target_page, and in the case of an allocating write with
 810          * page_size < cluster size, it will contain zero'd and mapped
 811          * pages adjacent to w_target_page which need to be written
 812          * out in so that future reads from that region will get
 813          * zero's.
 814          */
 815         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 816         unsigned int                    w_num_pages;
 817         struct page                     *w_target_page;
 818
 819         /*
 820          * ocfs2_write_end() uses this to know what the real range to
 821          * write in the target should be.
 822          */
 823         unsigned int                    w_target_from;
 824         unsigned int                    w_target_to;
 825
 826         /*
 827          * We could use journal_current_handle() but this is cleaner,
 828          * IMHO -Mark
 829          */
 830         handle_t                        *w_handle;
 831
 832         struct buffer_head              *w_di_bh;
 833
 834         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 835 };
 836
 837 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 838 {
 839         int i;
 840
 841         for(i = 0; i < wc->w_num_pages; i++) {
 842                 if (wc->w_pages[i] == NULL)
 843                         continue;
 844
 845                 unlock_page(wc->w_pages[i]);
 846                 mark_page_accessed(wc->w_pages[i]);
 847                 page_cache_release(wc->w_pages[i]);
 848         }
 849
 850         brelse(wc->w_di_bh);
 851         kfree(wc);
 852 }
 853
 854 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 855                                   struct ocfs2_super *osb, loff_t pos,
 856                                   unsigned len, struct buffer_head *di_bh)
 857 {
 858         u32 cend;
 859         struct ocfs2_write_ctxt *wc;
 860
 861         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 862         if (!wc)
 863                 return -ENOMEM;
 864
 865         wc->w_cpos = pos >> osb->s_clustersize_bits;
 866         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 867         wc->w_clen = cend - wc->w_cpos + 1;
 868         get_bh(di_bh);
 869         wc->w_di_bh = di_bh;
 870
 871         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 872                 wc->w_large_pages = 1;
 873         else
 874                 wc->w_large_pages = 0;
 875
 876         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 877
 878         *wcp = wc;
 879
 880         return 0;
 881 }
 882
 883 /*
 884  * If a page has any new buffers, zero them out here, and mark them uptodate
 885  * and dirty so they'll be written out (in order to prevent uninitialised
 886  * block data from leaking). And clear the new bit.
 887  */
 888 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 889 {
 890         unsigned int block_start, block_end;
 891         struct buffer_head *head, *bh;
 892
 893         BUG_ON(!PageLocked(page));
 894         if (!page_has_buffers(page))
 895                 return;
 896
 897         bh = head = page_buffers(page);
 898         block_start = 0;
 899         do {
 900                 block_end = block_start + bh->b_size;
 901
 902                 if (buffer_new(bh)) {
 903                         if (block_end > from && block_start < to) {
 904                                 if (!PageUptodate(page)) {
 905                                         unsigned start, end;
 906
 907                                         start = max(from, block_start);
 908                                         end = min(to, block_end);
 909
 910                                         zero_user_page(page, start, end - start, KM_USER0);
 911                                         set_buffer_uptodate(bh);
 912                                 }
 913
 914                                 clear_buffer_new(bh);
 915                                 mark_buffer_dirty(bh);
 916                         }
 917                 }
 918
 919                 block_start = block_end;
 920                 bh = bh->b_this_page;
 921         } while (bh != head);
 922 }
 923
 924 /*
 925  * Only called when we have a failure during allocating write to write
 926  * zero's to the newly allocated region.
 927  */
 928 static void ocfs2_write_failure(struct inode *inode,
 929                                 struct ocfs2_write_ctxt *wc,
 930                                 loff_t user_pos, unsigned user_len)
 931 {
 932         int i;
 933         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
 934                 to = user_pos + user_len;
 935         struct page *tmppage;
 936
 937         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 938
 939         for(i = 0; i < wc->w_num_pages; i++) {
 940                 tmppage = wc->w_pages[i];
 941
 942                 if (ocfs2_should_order_data(inode))
 943                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
 944                                           from, to, NULL,
 945                                           ocfs2_journal_dirty_data);
 946
 947                 block_commit_write(tmppage, from, to);
 948         }
 949 }
 950
 951 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 952                                         struct ocfs2_write_ctxt *wc,
 953                                         struct page *page, u32 cpos,
 954                                         loff_t user_pos, unsigned user_len,
 955                                         int new)
 956 {
 957         int ret;
 958         unsigned int map_from = 0, map_to = 0;
 959         unsigned int cluster_start, cluster_end;
 960         unsigned int user_data_from = 0, user_data_to = 0;
 961
 962         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 963                                         &cluster_start, &cluster_end);
 964
 965         if (page == wc->w_target_page) {
 966                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
 967                 map_to = map_from + user_len;
 968
 969                 if (new)
 970                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 971                                                     cluster_start, cluster_end,
 972                                                     new);
 973                 else
 974                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 975                                                     map_from, map_to, new);
 976                 if (ret) {
 977                         mlog_errno(ret);
 978                         goto out;
 979                 }
 980
 981                 user_data_from = map_from;
 982                 user_data_to = map_to;
 983                 if (new) {
 984                         map_from = cluster_start;
 985                         map_to = cluster_end;
 986                 }
 987         } else {
 988                 /*
 989                  * If we haven't allocated the new page yet, we
 990                  * shouldn't be writing it out without copying user
 991                  * data. This is likely a math error from the caller.
 992                  */
 993                 BUG_ON(!new);
 994
 995                 map_from = cluster_start;
 996                 map_to = cluster_end;
 997
 998                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 999                                             cluster_start, cluster_end, new);
1000                 if (ret) {
1001                         mlog_errno(ret);
1002                         goto out;
1003                 }
1004         }
1005
1006         /*
1007          * Parts of newly allocated pages need to be zero'd.
1008          *
1009          * Above, we have also rewritten 'to' and 'from' - as far as
1010          * the rest of the function is concerned, the entire cluster
1011          * range inside of a page needs to be written.
1012          *
1013          * We can skip this if the page is up to date - it's already
1014          * been zero'd from being read in as a hole.
1015          */
1016         if (new && !PageUptodate(page))
1017                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1018                                          cpos, user_data_from, user_data_to);
1019
1020         flush_dcache_page(page);
1021
1022 out:
1023         return ret;
1024 }
1025
1026 /*
1027  * This function will only grab one clusters worth of pages.
1028  */
1029 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1030                                       struct ocfs2_write_ctxt *wc,
1031                                       u32 cpos, loff_t user_pos, int new,
1032                                       struct page *mmap_page)
1033 {
1034         int ret = 0, i;
1035         unsigned long start, target_index, index;
1036         struct inode *inode = mapping->host;
1037
1038         target_index = user_pos >> PAGE_CACHE_SHIFT;
1039
1040         /*
1041          * Figure out how many pages we'll be manipulating here. For
1042          * non allocating write, we just change the one
1043          * page. Otherwise, we'll need a whole clusters worth.
1044          */
1045         if (new) {
1046                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1047                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1048         } else {
1049                 wc->w_num_pages = 1;
1050                 start = target_index;
1051         }
1052
1053         for(i = 0; i < wc->w_num_pages; i++) {
1054                 index = start + i;
1055
1056                 if (index == target_index && mmap_page) {
1057                         /*
1058                          * ocfs2_pagemkwrite() is a little different
1059                          * and wants us to directly use the page
1060                          * passed in.
1061                          */
1062                         lock_page(mmap_page);
1063
1064                         if (mmap_page->mapping != mapping) {
1065                                 unlock_page(mmap_page);
1066                                 /*
1067                                  * Sanity check - the locking in
1068                                  * ocfs2_pagemkwrite() should ensure
1069                                  * that this code doesn't trigger.
1070                                  */
1071                                 ret = -EINVAL;
1072                                 mlog_errno(ret);
1073                                 goto out;
1074                         }
1075
1076                         page_cache_get(mmap_page);
1077                         wc->w_pages[i] = mmap_page;
1078                 } else {
1079                         wc->w_pages[i] = find_or_create_page(mapping, index,
1080                                                              GFP_NOFS);
1081                         if (!wc->w_pages[i]) {
1082                                 ret = -ENOMEM;
1083                                 mlog_errno(ret);
1084                                 goto out;
1085                         }
1086                 }
1087
1088                 if (index == target_index)
1089                         wc->w_target_page = wc->w_pages[i];
1090         }
1091 out:
1092         return ret;
1093 }
1094
1095 /*
1096  * Prepare a single cluster for write one cluster into the file.
1097  */
1098 static int ocfs2_write_cluster(struct address_space *mapping,
1099                                u32 phys, unsigned int unwritten,
1100                                struct ocfs2_alloc_context *data_ac,
1101                                struct ocfs2_alloc_context *meta_ac,
1102                                struct ocfs2_write_ctxt *wc, u32 cpos,
1103                                loff_t user_pos, unsigned user_len)
1104 {
1105         int ret, i, new, should_zero = 0;
1106         u64 v_blkno, p_blkno;
1107         struct inode *inode = mapping->host;
1108
1109         new = phys == 0 ? 1 : 0;
1110         if (new || unwritten)
1111                 should_zero = 1;
1112
1113         if (new) {
1114                 u32 tmp_pos;
1115
1116                 /*
1117                  * This is safe to call with the page locks - it won't take
1118                  * any additional semaphores or cluster locks.
1119                  */
1120                 tmp_pos = cpos;
1121                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1122                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1123                                                  wc->w_handle, data_ac,
1124                                                  meta_ac, NULL);
1125                 /*
1126                  * This shouldn't happen because we must have already
1127                  * calculated the correct meta data allocation required. The
1128                  * internal tree allocation code should know how to increase
1129                  * transaction credits itself.
1130                  *
1131                  * If need be, we could handle -EAGAIN for a
1132                  * RESTART_TRANS here.
1133                  */
1134                 mlog_bug_on_msg(ret == -EAGAIN,
1135                                 "Inode %llu: EAGAIN return during allocation.\n",
1136                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1137                 if (ret < 0) {
1138                         mlog_errno(ret);
1139                         goto out;
1140                 }
1141         } else if (unwritten) {
1142                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1143                                                 wc->w_handle, cpos, 1, phys,
1144                                                 meta_ac, &wc->w_dealloc);
1145                 if (ret < 0) {
1146                         mlog_errno(ret);
1147                         goto out;
1148                 }
1149         }
1150
1151         if (should_zero)
1152                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1153         else
1154                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1155
1156         /*
1157          * The only reason this should fail is due to an inability to
1158          * find the extent added.
1159          */
1160         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1161                                           NULL);
1162         if (ret < 0) {
1163                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1164                             "at logical block %llu",
1165                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1166                             (unsigned long long)v_blkno);
1167                 goto out;
1168         }
1169
1170         BUG_ON(p_blkno == 0);
1171
1172         for(i = 0; i < wc->w_num_pages; i++) {
1173                 int tmpret;
1174
1175                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1176                                                       wc->w_pages[i], cpos,
1177                                                       user_pos, user_len,
1178                                                       should_zero);
1179                 if (tmpret) {
1180                         mlog_errno(tmpret);
1181                         if (ret == 0)
1182                                 tmpret = ret;
1183                 }
1184         }
1185
1186         /*
1187          * We only have cleanup to do in case of allocating write.
1188          */
1189         if (ret && new)
1190                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1191
1192 out:
1193
1194         return ret;
1195 }
1196
1197 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1198                                        struct ocfs2_alloc_context *data_ac,
1199                                        struct ocfs2_alloc_context *meta_ac,
1200                                        struct ocfs2_write_ctxt *wc,
1201                                        loff_t pos, unsigned len)
1202 {
1203         int ret, i;
1204         loff_t cluster_off;
1205         unsigned int local_len = len;
1206         struct ocfs2_write_cluster_desc *desc;
1207         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1208
1209         for (i = 0; i < wc->w_clen; i++) {
1210                 desc = &wc->w_desc[i];
1211
1212                 /*
1213                  * We have to make sure that the total write passed in
1214                  * doesn't extend past a single cluster.
1215                  */
1216                 local_len = len;
1217                 cluster_off = pos & (osb->s_clustersize - 1);
1218                 if ((cluster_off + local_len) > osb->s_clustersize)
1219                         local_len = osb->s_clustersize - cluster_off;
1220
1221                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1222                                           desc->c_unwritten, data_ac, meta_ac,
1223                                           wc, desc->c_cpos, pos, local_len);
1224                 if (ret) {
1225                         mlog_errno(ret);
1226                         goto out;
1227                 }
1228
1229                 len -= local_len;
1230                 pos += local_len;
1231         }
1232
1233         ret = 0;
1234 out:
1235         return ret;
1236 }
1237
1238 /*
1239  * ocfs2_write_end() wants to know which parts of the target page it
1240  * should complete the write on. It's easiest to compute them ahead of
1241  * time when a more complete view of the write is available.
1242  */
1243 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1244                                         struct ocfs2_write_ctxt *wc,
1245                                         loff_t pos, unsigned len, int alloc)
1246 {
1247         struct ocfs2_write_cluster_desc *desc;
1248
1249         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1250         wc->w_target_to = wc->w_target_from + len;
1251
1252         if (alloc == 0)
1253                 return;
1254
1255         /*
1256          * Allocating write - we may have different boundaries based
1257          * on page size and cluster size.
1258          *
1259          * NOTE: We can no longer compute one value from the other as
1260          * the actual write length and user provided length may be
1261          * different.
1262          */
1263
1264         if (wc->w_large_pages) {
1265                 /*
1266                  * We only care about the 1st and last cluster within
1267                  * our range and whether they should be zero'd or not. Either
1268                  * value may be extended out to the start/end of a
1269                  * newly allocated cluster.
1270                  */
1271                 desc = &wc->w_desc[0];
1272                 if (ocfs2_should_zero_cluster(desc))
1273                         ocfs2_figure_cluster_boundaries(osb,
1274                                                         desc->c_cpos,
1275                                                         &wc->w_target_from,
1276                                                         NULL);
1277
1278                 desc = &wc->w_desc[wc->w_clen - 1];
1279                 if (ocfs2_should_zero_cluster(desc))
1280                         ocfs2_figure_cluster_boundaries(osb,
1281                                                         desc->c_cpos,
1282                                                         NULL,
1283                                                         &wc->w_target_to);
1284         } else {
1285                 wc->w_target_from = 0;
1286                 wc->w_target_to = PAGE_CACHE_SIZE;
1287         }
1288 }
1289
1290 /*
1291  * Populate each single-cluster write descriptor in the write context
1292  * with information about the i/o to be done.
1293  *
1294  * Returns the number of clusters that will have to be allocated, as
1295  * well as a worst case estimate of the number of extent records that
1296  * would have to be created during a write to an unwritten region.
1297  */
1298 static int ocfs2_populate_write_desc(struct inode *inode,
1299                                      struct ocfs2_write_ctxt *wc,
1300                                      unsigned int *clusters_to_alloc,
1301                                      unsigned int *extents_to_split)
1302 {
1303         int ret;
1304         struct ocfs2_write_cluster_desc *desc;
1305         unsigned int num_clusters = 0;
1306         unsigned int ext_flags = 0;
1307         u32 phys = 0;
1308         int i;
1309
1310         *clusters_to_alloc = 0;
1311         *extents_to_split = 0;
1312
1313         for (i = 0; i < wc->w_clen; i++) {
1314                 desc = &wc->w_desc[i];
1315                 desc->c_cpos = wc->w_cpos + i;
1316
1317                 if (num_clusters == 0) {
1318                         /*
1319                          * Need to look up the next extent record.
1320                          */
1321                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1322                                                  &num_clusters, &ext_flags);
1323                         if (ret) {
1324                                 mlog_errno(ret);
1325                                 goto out;
1326                         }
1327
1328                         /*
1329                          * Assume worst case - that we're writing in
1330                          * the middle of the extent.
1331                          *
1332                          * We can assume that the write proceeds from
1333                          * left to right, in which case the extent
1334                          * insert code is smart enough to coalesce the
1335                          * next splits into the previous records created.
1336                          */
1337                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1338                                 *extents_to_split = *extents_to_split + 2;
1339                 } else if (phys) {
1340                         /*
1341                          * Only increment phys if it doesn't describe
1342                          * a hole.
1343                          */
1344                         phys++;
1345                 }
1346
1347                 desc->c_phys = phys;
1348                 if (phys == 0) {
1349                         desc->c_new = 1;
1350                         *clusters_to_alloc = *clusters_to_alloc + 1;
1351                 }
1352                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1353                         desc->c_unwritten = 1;
1354
1355                 num_clusters--;
1356         }
1357
1358         ret = 0;
1359 out:
1360         return ret;
1361 }
1362
1363 int ocfs2_write_begin_nolock(struct address_space *mapping,
1364                              loff_t pos, unsigned len, unsigned flags,
1365                              struct page **pagep, void **fsdata,
1366                              struct buffer_head *di_bh, struct page *mmap_page)
1367 {
1368         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1369         unsigned int clusters_to_alloc, extents_to_split;
1370         struct ocfs2_write_ctxt *wc;
1371         struct inode *inode = mapping->host;
1372         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1373         struct ocfs2_dinode *di;
1374         struct ocfs2_alloc_context *data_ac = NULL;
1375         struct ocfs2_alloc_context *meta_ac = NULL;
1376         handle_t *handle;
1377
1378         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1379         if (ret) {
1380                 mlog_errno(ret);
1381                 return ret;
1382         }
1383
1384         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1385                                         &extents_to_split);
1386         if (ret) {
1387                 mlog_errno(ret);
1388                 goto out;
1389         }
1390
1391         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1392
1393         /*
1394          * We set w_target_from, w_target_to here so that
1395          * ocfs2_write_end() knows which range in the target page to
1396          * write out. An allocation requires that we write the entire
1397          * cluster range.
1398          */
1399         if (clusters_to_alloc || extents_to_split) {
1400                 /*
1401                  * XXX: We are stretching the limits of
1402                  * ocfs2_lock_allocators(). It greatly over-estimates
1403                  * the work to be done.
1404                  */
1405                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1406                                             extents_to_split, &data_ac, &meta_ac);
1407                 if (ret) {
1408                         mlog_errno(ret);
1409                         goto out;
1410                 }
1411
1412                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1413                                                     clusters_to_alloc);
1414
1415         }
1416
1417         ocfs2_set_target_boundaries(osb, wc, pos, len,
1418                                     clusters_to_alloc + extents_to_split);
1419
1420         handle = ocfs2_start_trans(osb, credits);
1421         if (IS_ERR(handle)) {
1422                 ret = PTR_ERR(handle);
1423                 mlog_errno(ret);
1424                 goto out;
1425         }
1426
1427         wc->w_handle = handle;
1428
1429         /*
1430          * We don't want this to fail in ocfs2_write_end(), so do it
1431          * here.
1432          */
1433         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1434                                    OCFS2_JOURNAL_ACCESS_WRITE);
1435         if (ret) {
1436                 mlog_errno(ret);
1437                 goto out_commit;
1438         }
1439
1440         /*
1441          * Fill our page array first. That way we've grabbed enough so
1442          * that we can zero and flush if we error after adding the
1443          * extent.
1444          */
1445         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1446                                          clusters_to_alloc + extents_to_split,
1447                                          mmap_page);
1448         if (ret) {
1449                 mlog_errno(ret);
1450                 goto out_commit;
1451         }
1452
1453         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1454                                           len);
1455         if (ret) {
1456                 mlog_errno(ret);
1457                 goto out_commit;
1458         }
1459
1460         if (data_ac)
1461                 ocfs2_free_alloc_context(data_ac);
1462         if (meta_ac)
1463                 ocfs2_free_alloc_context(meta_ac);
1464
1465         *pagep = wc->w_target_page;
1466         *fsdata = wc;
1467         return 0;
1468 out_commit:
1469         ocfs2_commit_trans(osb, handle);
1470
1471 out:
1472         ocfs2_free_write_ctxt(wc);
1473
1474         if (data_ac)
1475                 ocfs2_free_alloc_context(data_ac);
1476         if (meta_ac)
1477                 ocfs2_free_alloc_context(meta_ac);
1478         return ret;
1479 }
1480
1481 int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1482                       loff_t pos, unsigned len, unsigned flags,
1483                       struct page **pagep, void **fsdata)
1484 {
1485         int ret;
1486         struct buffer_head *di_bh = NULL;
1487         struct inode *inode = mapping->host;
1488
1489         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1490         if (ret) {
1491                 mlog_errno(ret);
1492                 return ret;
1493         }
1494
1495         /*
1496          * Take alloc sem here to prevent concurrent lookups. That way
1497          * the mapping, zeroing and tree manipulation within
1498          * ocfs2_write() will be safe against ->readpage(). This
1499          * should also serve to lock out allocation from a shared
1500          * writeable region.
1501          */
1502         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1503
1504         ret = ocfs2_data_lock(inode, 1);
1505         if (ret) {
1506                 mlog_errno(ret);
1507                 goto out_fail;
1508         }
1509
1510         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1511                                        fsdata, di_bh, NULL);
1512         if (ret) {
1513                 mlog_errno(ret);
1514                 goto out_fail_data;
1515         }
1516
1517         brelse(di_bh);
1518
1519         return 0;
1520
1521 out_fail_data:
1522         ocfs2_data_unlock(inode, 1);
1523 out_fail:
1524         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1525
1526         brelse(di_bh);
1527         ocfs2_meta_unlock(inode, 1);
1528
1529         return ret;
1530 }
1531
1532 int ocfs2_write_end_nolock(struct address_space *mapping,
1533                            loff_t pos, unsigned len, unsigned copied,
1534                            struct page *page, void *fsdata)
1535 {
1536         int i;
1537         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1538         struct inode *inode = mapping->host;
1539         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1540         struct ocfs2_write_ctxt *wc = fsdata;
1541         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1542         handle_t *handle = wc->w_handle;
1543         struct page *tmppage;
1544
1545         if (unlikely(copied < len)) {
1546                 if (!PageUptodate(wc->w_target_page))
1547                         copied = 0;
1548
1549                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1550                                        start+len);
1551         }
1552         flush_dcache_page(wc->w_target_page);
1553
1554         for(i = 0; i < wc->w_num_pages; i++) {
1555                 tmppage = wc->w_pages[i];
1556
1557                 if (tmppage == wc->w_target_page) {
1558                         from = wc->w_target_from;
1559                         to = wc->w_target_to;
1560
1561                         BUG_ON(from > PAGE_CACHE_SIZE ||
1562                                to > PAGE_CACHE_SIZE ||
1563                                to < from);
1564                 } else {
1565                         /*
1566                          * Pages adjacent to the target (if any) imply
1567                          * a hole-filling write in which case we want
1568                          * to flush their entire range.
1569                          */
1570                         from = 0;
1571                         to = PAGE_CACHE_SIZE;
1572                 }
1573
1574                 if (ocfs2_should_order_data(inode))
1575                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1576                                           from, to, NULL,
1577                                           ocfs2_journal_dirty_data);
1578
1579                 block_commit_write(tmppage, from, to);
1580         }
1581
1582         pos += copied;
1583         if (pos > inode->i_size) {
1584                 i_size_write(inode, pos);
1585                 mark_inode_dirty(inode);
1586         }
1587         inode->i_blocks = ocfs2_inode_sector_count(inode);
1588         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1589         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1590         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1591         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1592         ocfs2_journal_dirty(handle, wc->w_di_bh);
1593
1594         ocfs2_commit_trans(osb, handle);
1595
1596         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1597
1598         ocfs2_free_write_ctxt(wc);
1599
1600         return copied;
1601 }
1602
1603 int ocfs2_write_end(struct file *file, struct address_space *mapping,
1604                     loff_t pos, unsigned len, unsigned copied,
1605                     struct page *page, void *fsdata)
1606 {
1607         int ret;
1608         struct inode *inode = mapping->host;
1609
1610         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1611
1612         ocfs2_data_unlock(inode, 1);
1613         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1614         ocfs2_meta_unlock(inode, 1);
1615
1616         return ret;
1617 }
1618
1619 const struct address_space_operations ocfs2_aops = {
1620         .readpage       = ocfs2_readpage,
1621         .writepage      = ocfs2_writepage,
1622         .bmap           = ocfs2_bmap,
1623         .sync_page      = block_sync_page,
1624         .direct_IO      = ocfs2_direct_IO,
1625         .invalidatepage = ocfs2_invalidatepage,
1626         .releasepage    = ocfs2_releasepage,
1627         .migratepage    = buffer_migrate_page,
1628 };