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         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 305
 306         return ret;
 307 }
 308
 309 /* Taken from ext3. We don't necessarily need the full blown
 310  * functionality yet, but IMHO it's better to cut and paste the whole
 311  * thing so we can avoid introducing our own bugs (and easily pick up
 312  * their fixes when they happen) --Mark */
 313 int walk_page_buffers(  handle_t *handle,
 314                         struct buffer_head *head,
 315                         unsigned from,
 316                         unsigned to,
 317                         int *partial,
 318                         int (*fn)(      handle_t *handle,
 319                                         struct buffer_head *bh))
 320 {
 321         struct buffer_head *bh;
 322         unsigned block_start, block_end;
 323         unsigned blocksize = head->b_size;
 324         int err, ret = 0;
 325         struct buffer_head *next;
 326
 327         for (   bh = head, block_start = 0;
 328                 ret == 0 && (bh != head || !block_start);
 329                 block_start = block_end, bh = next)
 330         {
 331                 next = bh->b_this_page;
 332                 block_end = block_start + blocksize;
 333                 if (block_end <= from || block_start >= to) {
 334                         if (partial && !buffer_uptodate(bh))
 335                                 *partial = 1;
 336                         continue;
 337                 }
 338                 err = (*fn)(handle, bh);
 339                 if (!ret)
 340                         ret = err;
 341         }
 342         return ret;
 343 }
 344
 345 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 346                                                          struct page *page,
 347                                                          unsigned from,
 348                                                          unsigned to)
 349 {
 350         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 351         handle_t *handle = NULL;
 352         int ret = 0;
 353
 354         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 355         if (!handle) {
 356                 ret = -ENOMEM;
 357                 mlog_errno(ret);
 358                 goto out;
 359         }
 360
 361         if (ocfs2_should_order_data(inode)) {
 362                 ret = walk_page_buffers(handle,
 363                                         page_buffers(page),
 364                                         from, to, NULL,
 365                                         ocfs2_journal_dirty_data);
 366                 if (ret < 0)
 367                         mlog_errno(ret);
 368         }
 369 out:
 370         if (ret) {
 371                 if (handle)
 372                         ocfs2_commit_trans(osb, handle);
 373                 handle = ERR_PTR(ret);
 374         }
 375         return handle;
 376 }
 377
 378 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 379 {
 380         sector_t status;
 381         u64 p_blkno = 0;
 382         int err = 0;
 383         struct inode *inode = mapping->host;
 384
 385         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 386
 387         /* We don't need to lock journal system files, since they aren't
 388          * accessed concurrently from multiple nodes.
 389          */
 390         if (!INODE_JOURNAL(inode)) {
 391                 err = ocfs2_meta_lock(inode, NULL, 0);
 392                 if (err) {
 393                         if (err != -ENOENT)
 394                                 mlog_errno(err);
 395                         goto bail;
 396                 }
 397                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 398         }
 399
 400         err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL);
 401
 402         if (!INODE_JOURNAL(inode)) {
 403                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 404                 ocfs2_meta_unlock(inode, 0);
 405         }
 406
 407         if (err) {
 408                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 409                      (unsigned long long)block);
 410                 mlog_errno(err);
 411                 goto bail;
 412         }
 413
 414
 415 bail:
 416         status = err ? 0 : p_blkno;
 417
 418         mlog_exit((int)status);
 419
 420         return status;
 421 }
 422
 423 /*
 424  * TODO: Make this into a generic get_blocks function.
 425  *
 426  * From do_direct_io in direct-io.c:
 427  *  "So what we do is to permit the ->get_blocks function to populate
 428  *   bh.b_size with the size of IO which is permitted at this offset and
 429  *   this i_blkbits."
 430  *
 431  * This function is called directly from get_more_blocks in direct-io.c.
 432  *
 433  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 434  *                                      fs_count, map_bh, dio->rw == WRITE);
 435  */
 436 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 437                                      struct buffer_head *bh_result, int create)
 438 {
 439         int ret;
 440         u64 p_blkno, inode_blocks, contig_blocks;
 441         unsigned int ext_flags;
 442         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 443         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 444
 445         /* This function won't even be called if the request isn't all
 446          * nicely aligned and of the right size, so there's no need
 447          * for us to check any of that. */
 448
 449         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 450
 451         /*
 452          * Any write past EOF is not allowed because we'd be extending.
 453          */
 454         if (create && (iblock + max_blocks) > inode_blocks) {
 455                 ret = -EIO;
 456                 goto bail;
 457         }
 458
 459         /* This figures out the size of the next contiguous block, and
 460          * our logical offset */
 461         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 462                                           &contig_blocks, &ext_flags);
 463         if (ret) {
 464                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 465                      (unsigned long long)iblock);
 466                 ret = -EIO;
 467                 goto bail;
 468         }
 469
 470         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
 471                 ocfs2_error(inode->i_sb,
 472                             "Inode %llu has a hole at block %llu\n",
 473                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 474                             (unsigned long long)iblock);
 475                 ret = -EROFS;
 476                 goto bail;
 477         }
 478
 479         /*
 480          * get_more_blocks() expects us to describe a hole by clearing
 481          * the mapped bit on bh_result().
 482          *
 483          * Consider an unwritten extent as a hole.
 484          */
 485         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 486                 map_bh(bh_result, inode->i_sb, p_blkno);
 487         else {
 488                 /*
 489                  * ocfs2_prepare_inode_for_write() should have caught
 490                  * the case where we'd be filling a hole and triggered
 491                  * a buffered write instead.
 492                  */
 493                 if (create) {
 494                         ret = -EIO;
 495                         mlog_errno(ret);
 496                         goto bail;
 497                 }
 498
 499                 clear_buffer_mapped(bh_result);
 500         }
 501
 502         /* make sure we don't map more than max_blocks blocks here as
 503            that's all the kernel will handle at this point. */
 504         if (max_blocks < contig_blocks)
 505                 contig_blocks = max_blocks;
 506         bh_result->b_size = contig_blocks << blocksize_bits;
 507 bail:
 508         return ret;
 509 }
 510
 511 /*
 512  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 513  * particularly interested in the aio/dio case.  Like the core uses
 514  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 515  * truncation on another.
 516  */
 517 static void ocfs2_dio_end_io(struct kiocb *iocb,
 518                              loff_t offset,
 519                              ssize_t bytes,
 520                              void *private)
 521 {
 522         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 523         int level;
 524
 525         /* this io's submitter should not have unlocked this before we could */
 526         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 527
 528         ocfs2_iocb_clear_rw_locked(iocb);
 529
 530         level = ocfs2_iocb_rw_locked_level(iocb);
 531         if (!level)
 532                 up_read(&inode->i_alloc_sem);
 533         ocfs2_rw_unlock(inode, level);
 534 }
 535
 536 /*
 537  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 538  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 539  * do journalled data.
 540  */
 541 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 542 {
 543         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 544
 545         journal_invalidatepage(journal, page, offset);
 546 }
 547
 548 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 549 {
 550         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 551
 552         if (!page_has_buffers(page))
 553                 return 0;
 554         return journal_try_to_free_buffers(journal, page, wait);
 555 }
 556
 557 static ssize_t ocfs2_direct_IO(int rw,
 558                                struct kiocb *iocb,
 559                                const struct iovec *iov,
 560                                loff_t offset,
 561                                unsigned long nr_segs)
 562 {
 563         struct file *file = iocb->ki_filp;
 564         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 565         int ret;
 566
 567         mlog_entry_void();
 568
 569         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 570                 /*
 571                  * We get PR data locks even for O_DIRECT.  This
 572                  * allows concurrent O_DIRECT I/O but doesn't let
 573                  * O_DIRECT with extending and buffered zeroing writes
 574                  * race.  If they did race then the buffered zeroing
 575                  * could be written back after the O_DIRECT I/O.  It's
 576                  * one thing to tell people not to mix buffered and
 577                  * O_DIRECT writes, but expecting them to understand
 578                  * that file extension is also an implicit buffered
 579                  * write is too much.  By getting the PR we force
 580                  * writeback of the buffered zeroing before
 581                  * proceeding.
 582                  */
 583                 ret = ocfs2_data_lock(inode, 0);
 584                 if (ret < 0) {
 585                         mlog_errno(ret);
 586                         goto out;
 587                 }
 588                 ocfs2_data_unlock(inode, 0);
 589         }
 590
 591         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 592                                             inode->i_sb->s_bdev, iov, offset,
 593                                             nr_segs,
 594                                             ocfs2_direct_IO_get_blocks,
 595                                             ocfs2_dio_end_io);
 596 out:
 597         mlog_exit(ret);
 598         return ret;
 599 }
 600
 601 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 602                                             u32 cpos,
 603                                             unsigned int *start,
 604                                             unsigned int *end)
 605 {
 606         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 607
 608         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 609                 unsigned int cpp;
 610
 611                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 612
 613                 cluster_start = cpos % cpp;
 614                 cluster_start = cluster_start << osb->s_clustersize_bits;
 615
 616                 cluster_end = cluster_start + osb->s_clustersize;
 617         }
 618
 619         BUG_ON(cluster_start > PAGE_SIZE);
 620         BUG_ON(cluster_end > PAGE_SIZE);
 621
 622         if (start)
 623                 *start = cluster_start;
 624         if (end)
 625                 *end = cluster_end;
 626 }
 627
 628 /*
 629  * 'from' and 'to' are the region in the page to avoid zeroing.
 630  *
 631  * If pagesize > clustersize, this function will avoid zeroing outside
 632  * of the cluster boundary.
 633  *
 634  * from == to == 0 is code for "zero the entire cluster region"
 635  */
 636 static void ocfs2_clear_page_regions(struct page *page,
 637                                      struct ocfs2_super *osb, u32 cpos,
 638                                      unsigned from, unsigned to)
 639 {
 640         void *kaddr;
 641         unsigned int cluster_start, cluster_end;
 642
 643         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 644
 645         kaddr = kmap_atomic(page, KM_USER0);
 646
 647         if (from || to) {
 648                 if (from > cluster_start)
 649                         memset(kaddr + cluster_start, 0, from - cluster_start);
 650                 if (to < cluster_end)
 651                         memset(kaddr + to, 0, cluster_end - to);
 652         } else {
 653                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 654         }
 655
 656         kunmap_atomic(kaddr, KM_USER0);
 657 }
 658
 659 /*
 660  * Some of this taken from block_prepare_write(). We already have our
 661  * mapping by now though, and the entire write will be allocating or
 662  * it won't, so not much need to use BH_New.
 663  *
 664  * This will also skip zeroing, which is handled externally.
 665  */
 666 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 667                           struct inode *inode, unsigned int from,
 668                           unsigned int to, int new)
 669 {
 670         int ret = 0;
 671         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 672         unsigned int block_end, block_start;
 673         unsigned int bsize = 1 << inode->i_blkbits;
 674
 675         if (!page_has_buffers(page))
 676                 create_empty_buffers(page, bsize, 0);
 677
 678         head = page_buffers(page);
 679         for (bh = head, block_start = 0; bh != head || !block_start;
 680              bh = bh->b_this_page, block_start += bsize) {
 681                 block_end = block_start + bsize;
 682
 683                 clear_buffer_new(bh);
 684
 685                 /*
 686                  * Ignore blocks outside of our i/o range -
 687                  * they may belong to unallocated clusters.
 688                  */
 689                 if (block_start >= to || block_end <= from) {
 690                         if (PageUptodate(page))
 691                                 set_buffer_uptodate(bh);
 692                         continue;
 693                 }
 694
 695                 /*
 696                  * For an allocating write with cluster size >= page
 697                  * size, we always write the entire page.
 698                  */
 699                 if (new)
 700                         set_buffer_new(bh);
 701
 702                 if (!buffer_mapped(bh)) {
 703                         map_bh(bh, inode->i_sb, *p_blkno);
 704                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 705                 }
 706
 707                 if (PageUptodate(page)) {
 708                         if (!buffer_uptodate(bh))
 709                                 set_buffer_uptodate(bh);
 710                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 711                            !buffer_new(bh) &&
 712                            (block_start < from || block_end > to)) {
 713                         ll_rw_block(READ, 1, &bh);
 714                         *wait_bh++=bh;
 715                 }
 716
 717                 *p_blkno = *p_blkno + 1;
 718         }
 719
 720         /*
 721          * If we issued read requests - let them complete.
 722          */
 723         while(wait_bh > wait) {
 724                 wait_on_buffer(*--wait_bh);
 725                 if (!buffer_uptodate(*wait_bh))
 726                         ret = -EIO;
 727         }
 728
 729         if (ret == 0 || !new)
 730                 return ret;
 731
 732         /*
 733          * If we get -EIO above, zero out any newly allocated blocks
 734          * to avoid exposing stale data.
 735          */
 736         bh = head;
 737         block_start = 0;
 738         do {
 739                 block_end = block_start + bsize;
 740                 if (block_end <= from)
 741                         goto next_bh;
 742                 if (block_start >= to)
 743                         break;
 744
 745                 zero_user_page(page, block_start, bh->b_size, KM_USER0);
 746                 set_buffer_uptodate(bh);
 747                 mark_buffer_dirty(bh);
 748
 749 next_bh:
 750                 block_start = block_end;
 751                 bh = bh->b_this_page;
 752         } while (bh != head);
 753
 754         return ret;
 755 }
 756
 757 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 758 #define OCFS2_MAX_CTXT_PAGES    1
 759 #else
 760 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 761 #endif
 762
 763 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 764
 765 /*
 766  * Describe the state of a single cluster to be written to.
 767  */
 768 struct ocfs2_write_cluster_desc {
 769         u32             c_cpos;
 770         u32             c_phys;
 771         /*
 772          * Give this a unique field because c_phys eventually gets
 773          * filled.
 774          */
 775         unsigned        c_new;
 776         unsigned        c_unwritten;
 777 };
 778
 779 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 780 {
 781         return d->c_new || d->c_unwritten;
 782 }
 783
 784 struct ocfs2_write_ctxt {
 785         /* Logical cluster position / len of write */
 786         u32                             w_cpos;
 787         u32                             w_clen;
 788
 789         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 790
 791         /*
 792          * This is true if page_size > cluster_size.
 793          *
 794          * It triggers a set of special cases during write which might
 795          * have to deal with allocating writes to partial pages.
 796          */
 797         unsigned int                    w_large_pages;
 798
 799         /*
 800          * Pages involved in this write.
 801          *
 802          * w_target_page is the page being written to by the user.
 803          *
 804          * w_pages is an array of pages which always contains
 805          * w_target_page, and in the case of an allocating write with
 806          * page_size < cluster size, it will contain zero'd and mapped
 807          * pages adjacent to w_target_page which need to be written
 808          * out in so that future reads from that region will get
 809          * zero's.
 810          */
 811         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 812         unsigned int                    w_num_pages;
 813         struct page                     *w_target_page;
 814
 815         /*
 816          * ocfs2_write_end() uses this to know what the real range to
 817          * write in the target should be.
 818          */
 819         unsigned int                    w_target_from;
 820         unsigned int                    w_target_to;
 821
 822         /*
 823          * We could use journal_current_handle() but this is cleaner,
 824          * IMHO -Mark
 825          */
 826         handle_t                        *w_handle;
 827
 828         struct buffer_head              *w_di_bh;
 829
 830         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 831 };
 832
 833 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 834 {
 835         int i;
 836
 837         for(i = 0; i < wc->w_num_pages; i++) {
 838                 if (wc->w_pages[i] == NULL)
 839                         continue;
 840
 841                 unlock_page(wc->w_pages[i]);
 842                 mark_page_accessed(wc->w_pages[i]);
 843                 page_cache_release(wc->w_pages[i]);
 844         }
 845
 846         brelse(wc->w_di_bh);
 847         kfree(wc);
 848 }
 849
 850 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 851                                   struct ocfs2_super *osb, loff_t pos,
 852                                   unsigned len, struct buffer_head *di_bh)
 853 {
 854         u32 cend;
 855         struct ocfs2_write_ctxt *wc;
 856
 857         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 858         if (!wc)
 859                 return -ENOMEM;
 860
 861         wc->w_cpos = pos >> osb->s_clustersize_bits;
 862         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 863         wc->w_clen = cend - wc->w_cpos + 1;
 864         get_bh(di_bh);
 865         wc->w_di_bh = di_bh;
 866
 867         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 868                 wc->w_large_pages = 1;
 869         else
 870                 wc->w_large_pages = 0;
 871
 872         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 873
 874         *wcp = wc;
 875
 876         return 0;
 877 }
 878
 879 /*
 880  * If a page has any new buffers, zero them out here, and mark them uptodate
 881  * and dirty so they'll be written out (in order to prevent uninitialised
 882  * block data from leaking). And clear the new bit.
 883  */
 884 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 885 {
 886         unsigned int block_start, block_end;
 887         struct buffer_head *head, *bh;
 888
 889         BUG_ON(!PageLocked(page));
 890         if (!page_has_buffers(page))
 891                 return;
 892
 893         bh = head = page_buffers(page);
 894         block_start = 0;
 895         do {
 896                 block_end = block_start + bh->b_size;
 897
 898                 if (buffer_new(bh)) {
 899                         if (block_end > from && block_start < to) {
 900                                 if (!PageUptodate(page)) {
 901                                         unsigned start, end;
 902
 903                                         start = max(from, block_start);
 904                                         end = min(to, block_end);
 905
 906                                         zero_user_page(page, start, end - start, KM_USER0);
 907                                         set_buffer_uptodate(bh);
 908                                 }
 909
 910                                 clear_buffer_new(bh);
 911                                 mark_buffer_dirty(bh);
 912                         }
 913                 }
 914
 915                 block_start = block_end;
 916                 bh = bh->b_this_page;
 917         } while (bh != head);
 918 }
 919
 920 /*
 921  * Only called when we have a failure during allocating write to write
 922  * zero's to the newly allocated region.
 923  */
 924 static void ocfs2_write_failure(struct inode *inode,
 925                                 struct ocfs2_write_ctxt *wc,
 926                                 loff_t user_pos, unsigned user_len)
 927 {
 928         int i;
 929         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
 930                 to = user_pos + user_len;
 931         struct page *tmppage;
 932
 933         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 934
 935         for(i = 0; i < wc->w_num_pages; i++) {
 936                 tmppage = wc->w_pages[i];
 937
 938                 if (ocfs2_should_order_data(inode))
 939                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
 940                                           from, to, NULL,
 941                                           ocfs2_journal_dirty_data);
 942
 943                 block_commit_write(tmppage, from, to);
 944         }
 945 }
 946
 947 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 948                                         struct ocfs2_write_ctxt *wc,
 949                                         struct page *page, u32 cpos,
 950                                         loff_t user_pos, unsigned user_len,
 951                                         int new)
 952 {
 953         int ret;
 954         unsigned int map_from = 0, map_to = 0;
 955         unsigned int cluster_start, cluster_end;
 956         unsigned int user_data_from = 0, user_data_to = 0;
 957
 958         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 959                                         &cluster_start, &cluster_end);
 960
 961         if (page == wc->w_target_page) {
 962                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
 963                 map_to = map_from + user_len;
 964
 965                 if (new)
 966                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 967                                                     cluster_start, cluster_end,
 968                                                     new);
 969                 else
 970                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 971                                                     map_from, map_to, new);
 972                 if (ret) {
 973                         mlog_errno(ret);
 974                         goto out;
 975                 }
 976
 977                 user_data_from = map_from;
 978                 user_data_to = map_to;
 979                 if (new) {
 980                         map_from = cluster_start;
 981                         map_to = cluster_end;
 982                 }
 983         } else {
 984                 /*
 985                  * If we haven't allocated the new page yet, we
 986                  * shouldn't be writing it out without copying user
 987                  * data. This is likely a math error from the caller.
 988                  */
 989                 BUG_ON(!new);
 990
 991                 map_from = cluster_start;
 992                 map_to = cluster_end;
 993
 994                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 995                                             cluster_start, cluster_end, new);
 996                 if (ret) {
 997                         mlog_errno(ret);
 998                         goto out;
 999                 }
1000         }
1001
1002         /*
1003          * Parts of newly allocated pages need to be zero'd.
1004          *
1005          * Above, we have also rewritten 'to' and 'from' - as far as
1006          * the rest of the function is concerned, the entire cluster
1007          * range inside of a page needs to be written.
1008          *
1009          * We can skip this if the page is up to date - it's already
1010          * been zero'd from being read in as a hole.
1011          */
1012         if (new && !PageUptodate(page))
1013                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1014                                          cpos, user_data_from, user_data_to);
1015
1016         flush_dcache_page(page);
1017
1018 out:
1019         return ret;
1020 }
1021
1022 /*
1023  * This function will only grab one clusters worth of pages.
1024  */
1025 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1026                                       struct ocfs2_write_ctxt *wc,
1027                                       u32 cpos, loff_t user_pos, int new,
1028                                       struct page *mmap_page)
1029 {
1030         int ret = 0, i;
1031         unsigned long start, target_index, index;
1032         struct inode *inode = mapping->host;
1033
1034         target_index = user_pos >> PAGE_CACHE_SHIFT;
1035
1036         /*
1037          * Figure out how many pages we'll be manipulating here. For
1038          * non allocating write, we just change the one
1039          * page. Otherwise, we'll need a whole clusters worth.
1040          */
1041         if (new) {
1042                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1043                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1044         } else {
1045                 wc->w_num_pages = 1;
1046                 start = target_index;
1047         }
1048
1049         for(i = 0; i < wc->w_num_pages; i++) {
1050                 index = start + i;
1051
1052                 if (index == target_index && mmap_page) {
1053                         /*
1054                          * ocfs2_pagemkwrite() is a little different
1055                          * and wants us to directly use the page
1056                          * passed in.
1057                          */
1058                         lock_page(mmap_page);
1059
1060                         if (mmap_page->mapping != mapping) {
1061                                 unlock_page(mmap_page);
1062                                 /*
1063                                  * Sanity check - the locking in
1064                                  * ocfs2_pagemkwrite() should ensure
1065                                  * that this code doesn't trigger.
1066                                  */
1067                                 ret = -EINVAL;
1068                                 mlog_errno(ret);
1069                                 goto out;
1070                         }
1071
1072                         page_cache_get(mmap_page);
1073                         wc->w_pages[i] = mmap_page;
1074                 } else {
1075                         wc->w_pages[i] = find_or_create_page(mapping, index,
1076                                                              GFP_NOFS);
1077                         if (!wc->w_pages[i]) {
1078                                 ret = -ENOMEM;
1079                                 mlog_errno(ret);
1080                                 goto out;
1081                         }
1082                 }
1083
1084                 if (index == target_index)
1085                         wc->w_target_page = wc->w_pages[i];
1086         }
1087 out:
1088         return ret;
1089 }
1090
1091 /*
1092  * Prepare a single cluster for write one cluster into the file.
1093  */
1094 static int ocfs2_write_cluster(struct address_space *mapping,
1095                                u32 phys, unsigned int unwritten,
1096                                struct ocfs2_alloc_context *data_ac,
1097                                struct ocfs2_alloc_context *meta_ac,
1098                                struct ocfs2_write_ctxt *wc, u32 cpos,
1099                                loff_t user_pos, unsigned user_len)
1100 {
1101         int ret, i, new, should_zero = 0;
1102         u64 v_blkno, p_blkno;
1103         struct inode *inode = mapping->host;
1104
1105         new = phys == 0 ? 1 : 0;
1106         if (new || unwritten)
1107                 should_zero = 1;
1108
1109         if (new) {
1110                 u32 tmp_pos;
1111
1112                 /*
1113                  * This is safe to call with the page locks - it won't take
1114                  * any additional semaphores or cluster locks.
1115                  */
1116                 tmp_pos = cpos;
1117                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1118                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1119                                                  wc->w_handle, data_ac,
1120                                                  meta_ac, NULL);
1121                 /*
1122                  * This shouldn't happen because we must have already
1123                  * calculated the correct meta data allocation required. The
1124                  * internal tree allocation code should know how to increase
1125                  * transaction credits itself.
1126                  *
1127                  * If need be, we could handle -EAGAIN for a
1128                  * RESTART_TRANS here.
1129                  */
1130                 mlog_bug_on_msg(ret == -EAGAIN,
1131                                 "Inode %llu: EAGAIN return during allocation.\n",
1132                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1133                 if (ret < 0) {
1134                         mlog_errno(ret);
1135                         goto out;
1136                 }
1137         } else if (unwritten) {
1138                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1139                                                 wc->w_handle, cpos, 1, phys,
1140                                                 meta_ac, &wc->w_dealloc);
1141                 if (ret < 0) {
1142                         mlog_errno(ret);
1143                         goto out;
1144                 }
1145         }
1146
1147         if (should_zero)
1148                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1149         else
1150                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1151
1152         /*
1153          * The only reason this should fail is due to an inability to
1154          * find the extent added.
1155          */
1156         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1157                                           NULL);
1158         if (ret < 0) {
1159                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1160                             "at logical block %llu",
1161                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1162                             (unsigned long long)v_blkno);
1163                 goto out;
1164         }
1165
1166         BUG_ON(p_blkno == 0);
1167
1168         for(i = 0; i < wc->w_num_pages; i++) {
1169                 int tmpret;
1170
1171                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1172                                                       wc->w_pages[i], cpos,
1173                                                       user_pos, user_len,
1174                                                       should_zero);
1175                 if (tmpret) {
1176                         mlog_errno(tmpret);
1177                         if (ret == 0)
1178                                 tmpret = ret;
1179                 }
1180         }
1181
1182         /*
1183          * We only have cleanup to do in case of allocating write.
1184          */
1185         if (ret && new)
1186                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1187
1188 out:
1189
1190         return ret;
1191 }
1192
1193 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1194                                        struct ocfs2_alloc_context *data_ac,
1195                                        struct ocfs2_alloc_context *meta_ac,
1196                                        struct ocfs2_write_ctxt *wc,
1197                                        loff_t pos, unsigned len)
1198 {
1199         int ret, i;
1200         loff_t cluster_off;
1201         unsigned int local_len = len;
1202         struct ocfs2_write_cluster_desc *desc;
1203         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1204
1205         for (i = 0; i < wc->w_clen; i++) {
1206                 desc = &wc->w_desc[i];
1207
1208                 /*
1209                  * We have to make sure that the total write passed in
1210                  * doesn't extend past a single cluster.
1211                  */
1212                 local_len = len;
1213                 cluster_off = pos & (osb->s_clustersize - 1);
1214                 if ((cluster_off + local_len) > osb->s_clustersize)
1215                         local_len = osb->s_clustersize - cluster_off;
1216
1217                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1218                                           desc->c_unwritten, data_ac, meta_ac,
1219                                           wc, desc->c_cpos, pos, local_len);
1220                 if (ret) {
1221                         mlog_errno(ret);
1222                         goto out;
1223                 }
1224
1225                 len -= local_len;
1226                 pos += local_len;
1227         }
1228
1229         ret = 0;
1230 out:
1231         return ret;
1232 }
1233
1234 /*
1235  * ocfs2_write_end() wants to know which parts of the target page it
1236  * should complete the write on. It's easiest to compute them ahead of
1237  * time when a more complete view of the write is available.
1238  */
1239 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1240                                         struct ocfs2_write_ctxt *wc,
1241                                         loff_t pos, unsigned len, int alloc)
1242 {
1243         struct ocfs2_write_cluster_desc *desc;
1244
1245         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1246         wc->w_target_to = wc->w_target_from + len;
1247
1248         if (alloc == 0)
1249                 return;
1250
1251         /*
1252          * Allocating write - we may have different boundaries based
1253          * on page size and cluster size.
1254          *
1255          * NOTE: We can no longer compute one value from the other as
1256          * the actual write length and user provided length may be
1257          * different.
1258          */
1259
1260         if (wc->w_large_pages) {
1261                 /*
1262                  * We only care about the 1st and last cluster within
1263                  * our range and whether they should be zero'd or not. Either
1264                  * value may be extended out to the start/end of a
1265                  * newly allocated cluster.
1266                  */
1267                 desc = &wc->w_desc[0];
1268                 if (ocfs2_should_zero_cluster(desc))
1269                         ocfs2_figure_cluster_boundaries(osb,
1270                                                         desc->c_cpos,
1271                                                         &wc->w_target_from,
1272                                                         NULL);
1273
1274                 desc = &wc->w_desc[wc->w_clen - 1];
1275                 if (ocfs2_should_zero_cluster(desc))
1276                         ocfs2_figure_cluster_boundaries(osb,
1277                                                         desc->c_cpos,
1278                                                         NULL,
1279                                                         &wc->w_target_to);
1280         } else {
1281                 wc->w_target_from = 0;
1282                 wc->w_target_to = PAGE_CACHE_SIZE;
1283         }
1284 }
1285
1286 /*
1287  * Populate each single-cluster write descriptor in the write context
1288  * with information about the i/o to be done.
1289  *
1290  * Returns the number of clusters that will have to be allocated, as
1291  * well as a worst case estimate of the number of extent records that
1292  * would have to be created during a write to an unwritten region.
1293  */
1294 static int ocfs2_populate_write_desc(struct inode *inode,
1295                                      struct ocfs2_write_ctxt *wc,
1296                                      unsigned int *clusters_to_alloc,
1297                                      unsigned int *extents_to_split)
1298 {
1299         int ret;
1300         struct ocfs2_write_cluster_desc *desc;
1301         unsigned int num_clusters = 0;
1302         unsigned int ext_flags = 0;
1303         u32 phys = 0;
1304         int i;
1305
1306         *clusters_to_alloc = 0;
1307         *extents_to_split = 0;
1308
1309         for (i = 0; i < wc->w_clen; i++) {
1310                 desc = &wc->w_desc[i];
1311                 desc->c_cpos = wc->w_cpos + i;
1312
1313                 if (num_clusters == 0) {
1314                         /*
1315                          * Need to look up the next extent record.
1316                          */
1317                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1318                                                  &num_clusters, &ext_flags);
1319                         if (ret) {
1320                                 mlog_errno(ret);
1321                                 goto out;
1322                         }
1323
1324                         /*
1325                          * Assume worst case - that we're writing in
1326                          * the middle of the extent.
1327                          *
1328                          * We can assume that the write proceeds from
1329                          * left to right, in which case the extent
1330                          * insert code is smart enough to coalesce the
1331                          * next splits into the previous records created.
1332                          */
1333                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1334                                 *extents_to_split = *extents_to_split + 2;
1335                 } else if (phys) {
1336                         /*
1337                          * Only increment phys if it doesn't describe
1338                          * a hole.
1339                          */
1340                         phys++;
1341                 }
1342
1343                 desc->c_phys = phys;
1344                 if (phys == 0) {
1345                         desc->c_new = 1;
1346                         *clusters_to_alloc = *clusters_to_alloc + 1;
1347                 }
1348                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1349                         desc->c_unwritten = 1;
1350
1351                 num_clusters--;
1352         }
1353
1354         ret = 0;
1355 out:
1356         return ret;
1357 }
1358
1359 /*
1360  * This function only does anything for file systems which can't
1361  * handle sparse files.
1362  *
1363  * What we want to do here is fill in any hole between the current end
1364  * of allocation and the end of our write. That way the rest of the
1365  * write path can treat it as an non-allocating write, which has no
1366  * special case code for sparse/nonsparse files.
1367  */
1368 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1369                                         unsigned len,
1370                                         struct ocfs2_write_ctxt *wc)
1371 {
1372         int ret;
1373         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1374         loff_t newsize = pos + len;
1375
1376         if (ocfs2_sparse_alloc(osb))
1377                 return 0;
1378
1379         if (newsize <= i_size_read(inode))
1380                 return 0;
1381
1382         ret = ocfs2_extend_no_holes(inode, newsize, newsize - len);
1383         if (ret)
1384                 mlog_errno(ret);
1385
1386         return ret;
1387 }
1388
1389 int ocfs2_write_begin_nolock(struct address_space *mapping,
1390                              loff_t pos, unsigned len, unsigned flags,
1391                              struct page **pagep, void **fsdata,
1392                              struct buffer_head *di_bh, struct page *mmap_page)
1393 {
1394         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1395         unsigned int clusters_to_alloc, extents_to_split;
1396         struct ocfs2_write_ctxt *wc;
1397         struct inode *inode = mapping->host;
1398         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1399         struct ocfs2_dinode *di;
1400         struct ocfs2_alloc_context *data_ac = NULL;
1401         struct ocfs2_alloc_context *meta_ac = NULL;
1402         handle_t *handle;
1403
1404         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1405         if (ret) {
1406                 mlog_errno(ret);
1407                 return ret;
1408         }
1409
1410         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1411         if (ret) {
1412                 mlog_errno(ret);
1413                 goto out;
1414         }
1415
1416         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1417                                         &extents_to_split);
1418         if (ret) {
1419                 mlog_errno(ret);
1420                 goto out;
1421         }
1422
1423         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1424
1425         /*
1426          * We set w_target_from, w_target_to here so that
1427          * ocfs2_write_end() knows which range in the target page to
1428          * write out. An allocation requires that we write the entire
1429          * cluster range.
1430          */
1431         if (clusters_to_alloc || extents_to_split) {
1432                 /*
1433                  * XXX: We are stretching the limits of
1434                  * ocfs2_lock_allocators(). It greatly over-estimates
1435                  * the work to be done.
1436                  */
1437                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1438                                             extents_to_split, &data_ac, &meta_ac);
1439                 if (ret) {
1440                         mlog_errno(ret);
1441                         goto out;
1442                 }
1443
1444                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1445                                                     clusters_to_alloc);
1446
1447         }
1448
1449         ocfs2_set_target_boundaries(osb, wc, pos, len,
1450                                     clusters_to_alloc + extents_to_split);
1451
1452         handle = ocfs2_start_trans(osb, credits);
1453         if (IS_ERR(handle)) {
1454                 ret = PTR_ERR(handle);
1455                 mlog_errno(ret);
1456                 goto out;
1457         }
1458
1459         wc->w_handle = handle;
1460
1461         /*
1462          * We don't want this to fail in ocfs2_write_end(), so do it
1463          * here.
1464          */
1465         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1466                                    OCFS2_JOURNAL_ACCESS_WRITE);
1467         if (ret) {
1468                 mlog_errno(ret);
1469                 goto out_commit;
1470         }
1471
1472         /*
1473          * Fill our page array first. That way we've grabbed enough so
1474          * that we can zero and flush if we error after adding the
1475          * extent.
1476          */
1477         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1478                                          clusters_to_alloc + extents_to_split,
1479                                          mmap_page);
1480         if (ret) {
1481                 mlog_errno(ret);
1482                 goto out_commit;
1483         }
1484
1485         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1486                                           len);
1487         if (ret) {
1488                 mlog_errno(ret);
1489                 goto out_commit;
1490         }
1491
1492         if (data_ac)
1493                 ocfs2_free_alloc_context(data_ac);
1494         if (meta_ac)
1495                 ocfs2_free_alloc_context(meta_ac);
1496
1497         *pagep = wc->w_target_page;
1498         *fsdata = wc;
1499         return 0;
1500 out_commit:
1501         ocfs2_commit_trans(osb, handle);
1502
1503 out:
1504         ocfs2_free_write_ctxt(wc);
1505
1506         if (data_ac)
1507                 ocfs2_free_alloc_context(data_ac);
1508         if (meta_ac)
1509                 ocfs2_free_alloc_context(meta_ac);
1510         return ret;
1511 }
1512
1513 int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1514                       loff_t pos, unsigned len, unsigned flags,
1515                       struct page **pagep, void **fsdata)
1516 {
1517         int ret;
1518         struct buffer_head *di_bh = NULL;
1519         struct inode *inode = mapping->host;
1520
1521         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1522         if (ret) {
1523                 mlog_errno(ret);
1524                 return ret;
1525         }
1526
1527         /*
1528          * Take alloc sem here to prevent concurrent lookups. That way
1529          * the mapping, zeroing and tree manipulation within
1530          * ocfs2_write() will be safe against ->readpage(). This
1531          * should also serve to lock out allocation from a shared
1532          * writeable region.
1533          */
1534         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1535
1536         ret = ocfs2_data_lock(inode, 1);
1537         if (ret) {
1538                 mlog_errno(ret);
1539                 goto out_fail;
1540         }
1541
1542         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1543                                        fsdata, di_bh, NULL);
1544         if (ret) {
1545                 mlog_errno(ret);
1546                 goto out_fail_data;
1547         }
1548
1549         brelse(di_bh);
1550
1551         return 0;
1552
1553 out_fail_data:
1554         ocfs2_data_unlock(inode, 1);
1555 out_fail:
1556         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1557
1558         brelse(di_bh);
1559         ocfs2_meta_unlock(inode, 1);
1560
1561         return ret;
1562 }
1563
1564 int ocfs2_write_end_nolock(struct address_space *mapping,
1565                            loff_t pos, unsigned len, unsigned copied,
1566                            struct page *page, void *fsdata)
1567 {
1568         int i;
1569         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1570         struct inode *inode = mapping->host;
1571         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1572         struct ocfs2_write_ctxt *wc = fsdata;
1573         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1574         handle_t *handle = wc->w_handle;
1575         struct page *tmppage;
1576
1577         if (unlikely(copied < len)) {
1578                 if (!PageUptodate(wc->w_target_page))
1579                         copied = 0;
1580
1581                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1582                                        start+len);
1583         }
1584         flush_dcache_page(wc->w_target_page);
1585
1586         for(i = 0; i < wc->w_num_pages; i++) {
1587                 tmppage = wc->w_pages[i];
1588
1589                 if (tmppage == wc->w_target_page) {
1590                         from = wc->w_target_from;
1591                         to = wc->w_target_to;
1592
1593                         BUG_ON(from > PAGE_CACHE_SIZE ||
1594                                to > PAGE_CACHE_SIZE ||
1595                                to < from);
1596                 } else {
1597                         /*
1598                          * Pages adjacent to the target (if any) imply
1599                          * a hole-filling write in which case we want
1600                          * to flush their entire range.
1601                          */
1602                         from = 0;
1603                         to = PAGE_CACHE_SIZE;
1604                 }
1605
1606                 if (ocfs2_should_order_data(inode))
1607                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1608                                           from, to, NULL,
1609                                           ocfs2_journal_dirty_data);
1610
1611                 block_commit_write(tmppage, from, to);
1612         }
1613
1614         pos += copied;
1615         if (pos > inode->i_size) {
1616                 i_size_write(inode, pos);
1617                 mark_inode_dirty(inode);
1618         }
1619         inode->i_blocks = ocfs2_inode_sector_count(inode);
1620         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1621         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1622         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1623         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1624         ocfs2_journal_dirty(handle, wc->w_di_bh);
1625
1626         ocfs2_commit_trans(osb, handle);
1627
1628         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1629
1630         ocfs2_free_write_ctxt(wc);
1631
1632         return copied;
1633 }
1634
1635 int ocfs2_write_end(struct file *file, struct address_space *mapping,
1636                     loff_t pos, unsigned len, unsigned copied,
1637                     struct page *page, void *fsdata)
1638 {
1639         int ret;
1640         struct inode *inode = mapping->host;
1641
1642         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1643
1644         ocfs2_data_unlock(inode, 1);
1645         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1646         ocfs2_meta_unlock(inode, 1);
1647
1648         return ret;
1649 }
1650
1651 const struct address_space_operations ocfs2_aops = {
1652         .readpage       = ocfs2_readpage,
1653         .writepage      = ocfs2_writepage,
1654         .bmap           = ocfs2_bmap,
1655         .sync_page      = block_sync_page,
1656         .direct_IO      = ocfs2_direct_IO,
1657         .invalidatepage = ocfs2_invalidatepage,
1658         .releasepage    = ocfs2_releasepage,
1659         .migratepage    = buffer_migrate_page,
1660 };