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 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31
  32 #include <cluster/masklog.h>
  33
  34 #include "ocfs2.h"
  35
  36 #include "alloc.h"
  37 #include "aops.h"
  38 #include "dlmglue.h"
  39 #include "extent_map.h"
  40 #include "file.h"
  41 #include "inode.h"
  42 #include "journal.h"
  43 #include "suballoc.h"
  44 #include "super.h"
  45 #include "symlink.h"
  46 #include "refcounttree.h"
  47 #include "ocfs2_trace.h"
  48
  49 #include "buffer_head_io.h"
  50
  51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  52                                    struct buffer_head *bh_result, int create)
  53 {
  54         int err = -EIO;
  55         int status;
  56         struct ocfs2_dinode *fe = NULL;
  57         struct buffer_head *bh = NULL;
  58         struct buffer_head *buffer_cache_bh = NULL;
  59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  60         void *kaddr;
  61
  62         trace_ocfs2_symlink_get_block(
  63                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  64                         (unsigned long long)iblock, bh_result, create);
  65
  66         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  67
  68         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  69                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  70                      (unsigned long long)iblock);
  71                 goto bail;
  72         }
  73
  74         status = ocfs2_read_inode_block(inode, &bh);
  75         if (status < 0) {
  76                 mlog_errno(status);
  77                 goto bail;
  78         }
  79         fe = (struct ocfs2_dinode *) bh->b_data;
  80
  81         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  82                                                     le32_to_cpu(fe->i_clusters))) {
  83                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  84                      "%llu\n", (unsigned long long)iblock);
  85                 goto bail;
  86         }
  87
  88         /* We don't use the page cache to create symlink data, so if
  89          * need be, copy it over from the buffer cache. */
  90         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  91                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  92                             iblock;
  93                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  94                 if (!buffer_cache_bh) {
  95                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  96                         goto bail;
  97                 }
  98
  99                 /* we haven't locked out transactions, so a commit
 100                  * could've happened. Since we've got a reference on
 101                  * the bh, even if it commits while we're doing the
 102                  * copy, the data is still good. */
 103                 if (buffer_jbd(buffer_cache_bh)
 104                     && ocfs2_inode_is_new(inode)) {
 105                         kaddr = kmap_atomic(bh_result->b_page);
 106                         if (!kaddr) {
 107                                 mlog(ML_ERROR, "couldn't kmap!\n");
 108                                 goto bail;
 109                         }
 110                         memcpy(kaddr + (bh_result->b_size * iblock),
 111                                buffer_cache_bh->b_data,
 112                                bh_result->b_size);
 113                         kunmap_atomic(kaddr);
 114                         set_buffer_uptodate(bh_result);
 115                 }
 116                 brelse(buffer_cache_bh);
 117         }
 118
 119         map_bh(bh_result, inode->i_sb,
 120                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 121
 122         err = 0;
 123
 124 bail:
 125         brelse(bh);
 126
 127         return err;
 128 }
 129
 130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 131                     struct buffer_head *bh_result, int create)
 132 {
 133         int err = 0;
 134         unsigned int ext_flags;
 135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 136         u64 p_blkno, count, past_eof;
 137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 138
 139         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 140                               (unsigned long long)iblock, bh_result, create);
 141
 142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 144                      inode, inode->i_ino);
 145
 146         if (S_ISLNK(inode->i_mode)) {
 147                 /* this always does I/O for some reason. */
 148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 149                 goto bail;
 150         }
 151
 152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 153                                           &ext_flags);
 154         if (err) {
 155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 157                      (unsigned long long)p_blkno);
 158                 goto bail;
 159         }
 160
 161         if (max_blocks < count)
 162                 count = max_blocks;
 163
 164         /*
 165          * ocfs2 never allocates in this function - the only time we
 166          * need to use BH_New is when we're extending i_size on a file
 167          * system which doesn't support holes, in which case BH_New
 168          * allows __block_write_begin() to zero.
 169          *
 170          * If we see this on a sparse file system, then a truncate has
 171          * raced us and removed the cluster. In this case, we clear
 172          * the buffers dirty and uptodate bits and let the buffer code
 173          * ignore it as a hole.
 174          */
 175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 176                 clear_buffer_dirty(bh_result);
 177                 clear_buffer_uptodate(bh_result);
 178                 goto bail;
 179         }
 180
 181         /* Treat the unwritten extent as a hole for zeroing purposes. */
 182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 183                 map_bh(bh_result, inode->i_sb, p_blkno);
 184
 185         bh_result->b_size = count << inode->i_blkbits;
 186
 187         if (!ocfs2_sparse_alloc(osb)) {
 188                 if (p_blkno == 0) {
 189                         err = -EIO;
 190                         mlog(ML_ERROR,
 191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 192                              (unsigned long long)iblock,
 193                              (unsigned long long)p_blkno,
 194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 196                         dump_stack();
 197                         goto bail;
 198                 }
 199         }
 200
 201         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 202
 203         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 204                                   (unsigned long long)past_eof);
 205         if (create && (iblock >= past_eof))
 206                 set_buffer_new(bh_result);
 207
 208 bail:
 209         if (err < 0)
 210                 err = -EIO;
 211
 212         return err;
 213 }
 214
 215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 216                            struct buffer_head *di_bh)
 217 {
 218         void *kaddr;
 219         loff_t size;
 220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 221
 222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 225                 return -EROFS;
 226         }
 227
 228         size = i_size_read(inode);
 229
 230         if (size > PAGE_CACHE_SIZE ||
 231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 232                 ocfs2_error(inode->i_sb,
 233                             "Inode %llu has with inline data has bad size: %Lu",
 234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 235                             (unsigned long long)size);
 236                 return -EROFS;
 237         }
 238
 239         kaddr = kmap_atomic(page);
 240         if (size)
 241                 memcpy(kaddr, di->id2.i_data.id_data, size);
 242         /* Clear the remaining part of the page */
 243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 244         flush_dcache_page(page);
 245         kunmap_atomic(kaddr);
 246
 247         SetPageUptodate(page);
 248
 249         return 0;
 250 }
 251
 252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 253 {
 254         int ret;
 255         struct buffer_head *di_bh = NULL;
 256
 257         BUG_ON(!PageLocked(page));
 258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 259
 260         ret = ocfs2_read_inode_block(inode, &di_bh);
 261         if (ret) {
 262                 mlog_errno(ret);
 263                 goto out;
 264         }
 265
 266         ret = ocfs2_read_inline_data(inode, page, di_bh);
 267 out:
 268         unlock_page(page);
 269
 270         brelse(di_bh);
 271         return ret;
 272 }
 273
 274 static int ocfs2_readpage(struct file *file, struct page *page)
 275 {
 276         struct inode *inode = page->mapping->host;
 277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 279         int ret, unlock = 1;
 280
 281         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 282                              (page ? page->index : 0));
 283
 284         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 285         if (ret != 0) {
 286                 if (ret == AOP_TRUNCATED_PAGE)
 287                         unlock = 0;
 288                 mlog_errno(ret);
 289                 goto out;
 290         }
 291
 292         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 293                 /*
 294                  * Unlock the page and cycle ip_alloc_sem so that we don't
 295                  * busyloop waiting for ip_alloc_sem to unlock
 296                  */
 297                 ret = AOP_TRUNCATED_PAGE;
 298                 unlock_page(page);
 299                 unlock = 0;
 300                 down_read(&oi->ip_alloc_sem);
 301                 up_read(&oi->ip_alloc_sem);
 302                 goto out_inode_unlock;
 303         }
 304
 305         /*
 306          * i_size might have just been updated as we grabed the meta lock.  We
 307          * might now be discovering a truncate that hit on another node.
 308          * block_read_full_page->get_block freaks out if it is asked to read
 309          * beyond the end of a file, so we check here.  Callers
 310          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 311          * and notice that the page they just read isn't needed.
 312          *
 313          * XXX sys_readahead() seems to get that wrong?
 314          */
 315         if (start >= i_size_read(inode)) {
 316                 zero_user(page, 0, PAGE_SIZE);
 317                 SetPageUptodate(page);
 318                 ret = 0;
 319                 goto out_alloc;
 320         }
 321
 322         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 323                 ret = ocfs2_readpage_inline(inode, page);
 324         else
 325                 ret = block_read_full_page(page, ocfs2_get_block);
 326         unlock = 0;
 327
 328 out_alloc:
 329         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 330 out_inode_unlock:
 331         ocfs2_inode_unlock(inode, 0);
 332 out:
 333         if (unlock)
 334                 unlock_page(page);
 335         return ret;
 336 }
 337
 338 /*
 339  * This is used only for read-ahead. Failures or difficult to handle
 340  * situations are safe to ignore.
 341  *
 342  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 343  * are quite large (243 extents on 4k blocks), so most inodes don't
 344  * grow out to a tree. If need be, detecting boundary extents could
 345  * trivially be added in a future version of ocfs2_get_block().
 346  */
 347 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 348                            struct list_head *pages, unsigned nr_pages)
 349 {
 350         int ret, err = -EIO;
 351         struct inode *inode = mapping->host;
 352         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 353         loff_t start;
 354         struct page *last;
 355
 356         /*
 357          * Use the nonblocking flag for the dlm code to avoid page
 358          * lock inversion, but don't bother with retrying.
 359          */
 360         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 361         if (ret)
 362                 return err;
 363
 364         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 365                 ocfs2_inode_unlock(inode, 0);
 366                 return err;
 367         }
 368
 369         /*
 370          * Don't bother with inline-data. There isn't anything
 371          * to read-ahead in that case anyway...
 372          */
 373         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 374                 goto out_unlock;
 375
 376         /*
 377          * Check whether a remote node truncated this file - we just
 378          * drop out in that case as it's not worth handling here.
 379          */
 380         last = list_entry(pages->prev, struct page, lru);
 381         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 382         if (start >= i_size_read(inode))
 383                 goto out_unlock;
 384
 385         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 386
 387 out_unlock:
 388         up_read(&oi->ip_alloc_sem);
 389         ocfs2_inode_unlock(inode, 0);
 390
 391         return err;
 392 }
 393
 394 /* Note: Because we don't support holes, our allocation has
 395  * already happened (allocation writes zeros to the file data)
 396  * so we don't have to worry about ordered writes in
 397  * ocfs2_writepage.
 398  *
 399  * ->writepage is called during the process of invalidating the page cache
 400  * during blocked lock processing.  It can't block on any cluster locks
 401  * to during block mapping.  It's relying on the fact that the block
 402  * mapping can't have disappeared under the dirty pages that it is
 403  * being asked to write back.
 404  */
 405 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 406 {
 407         trace_ocfs2_writepage(
 408                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 409                 page->index);
 410
 411         return block_write_full_page(page, ocfs2_get_block, wbc);
 412 }
 413
 414 /* Taken from ext3. We don't necessarily need the full blown
 415  * functionality yet, but IMHO it's better to cut and paste the whole
 416  * thing so we can avoid introducing our own bugs (and easily pick up
 417  * their fixes when they happen) --Mark */
 418 int walk_page_buffers(  handle_t *handle,
 419                         struct buffer_head *head,
 420                         unsigned from,
 421                         unsigned to,
 422                         int *partial,
 423                         int (*fn)(      handle_t *handle,
 424                                         struct buffer_head *bh))
 425 {
 426         struct buffer_head *bh;
 427         unsigned block_start, block_end;
 428         unsigned blocksize = head->b_size;
 429         int err, ret = 0;
 430         struct buffer_head *next;
 431
 432         for (   bh = head, block_start = 0;
 433                 ret == 0 && (bh != head || !block_start);
 434                 block_start = block_end, bh = next)
 435         {
 436                 next = bh->b_this_page;
 437                 block_end = block_start + blocksize;
 438                 if (block_end <= from || block_start >= to) {
 439                         if (partial && !buffer_uptodate(bh))
 440                                 *partial = 1;
 441                         continue;
 442                 }
 443                 err = (*fn)(handle, bh);
 444                 if (!ret)
 445                         ret = err;
 446         }
 447         return ret;
 448 }
 449
 450 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 451 {
 452         sector_t status;
 453         u64 p_blkno = 0;
 454         int err = 0;
 455         struct inode *inode = mapping->host;
 456
 457         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 458                          (unsigned long long)block);
 459
 460         /* We don't need to lock journal system files, since they aren't
 461          * accessed concurrently from multiple nodes.
 462          */
 463         if (!INODE_JOURNAL(inode)) {
 464                 err = ocfs2_inode_lock(inode, NULL, 0);
 465                 if (err) {
 466                         if (err != -ENOENT)
 467                                 mlog_errno(err);
 468                         goto bail;
 469                 }
 470                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 471         }
 472
 473         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 474                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 475                                                   NULL);
 476
 477         if (!INODE_JOURNAL(inode)) {
 478                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 479                 ocfs2_inode_unlock(inode, 0);
 480         }
 481
 482         if (err) {
 483                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 484                      (unsigned long long)block);
 485                 mlog_errno(err);
 486                 goto bail;
 487         }
 488
 489 bail:
 490         status = err ? 0 : p_blkno;
 491
 492         return status;
 493 }
 494
 495 /*
 496  * TODO: Make this into a generic get_blocks function.
 497  *
 498  * From do_direct_io in direct-io.c:
 499  *  "So what we do is to permit the ->get_blocks function to populate
 500  *   bh.b_size with the size of IO which is permitted at this offset and
 501  *   this i_blkbits."
 502  *
 503  * This function is called directly from get_more_blocks in direct-io.c.
 504  *
 505  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 506  *                                      fs_count, map_bh, dio->rw == WRITE);
 507  *
 508  * Note that we never bother to allocate blocks here, and thus ignore the
 509  * create argument.
 510  */
 511 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 512                                      struct buffer_head *bh_result, int create)
 513 {
 514         int ret;
 515         u64 p_blkno, inode_blocks, contig_blocks;
 516         unsigned int ext_flags;
 517         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 518         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 519
 520         /* This function won't even be called if the request isn't all
 521          * nicely aligned and of the right size, so there's no need
 522          * for us to check any of that. */
 523
 524         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 525
 526         /* This figures out the size of the next contiguous block, and
 527          * our logical offset */
 528         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 529                                           &contig_blocks, &ext_flags);
 530         if (ret) {
 531                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 532                      (unsigned long long)iblock);
 533                 ret = -EIO;
 534                 goto bail;
 535         }
 536
 537         /* We should already CoW the refcounted extent in case of create. */
 538         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 539
 540         /*
 541          * get_more_blocks() expects us to describe a hole by clearing
 542          * the mapped bit on bh_result().
 543          *
 544          * Consider an unwritten extent as a hole.
 545          */
 546         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 547                 map_bh(bh_result, inode->i_sb, p_blkno);
 548         else
 549                 clear_buffer_mapped(bh_result);
 550
 551         /* make sure we don't map more than max_blocks blocks here as
 552            that's all the kernel will handle at this point. */
 553         if (max_blocks < contig_blocks)
 554                 contig_blocks = max_blocks;
 555         bh_result->b_size = contig_blocks << blocksize_bits;
 556 bail:
 557         return ret;
 558 }
 559
 560 /*
 561  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 562  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 563  * to protect io on one node from truncation on another.
 564  */
 565 static void ocfs2_dio_end_io(struct kiocb *iocb,
 566                              loff_t offset,
 567                              ssize_t bytes,
 568                              void *private)
 569 {
 570         struct inode *inode = file_inode(iocb->ki_filp);
 571         int level;
 572         wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
 573
 574         /* this io's submitter should not have unlocked this before we could */
 575         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 576
 577         if (ocfs2_iocb_is_sem_locked(iocb))
 578                 ocfs2_iocb_clear_sem_locked(iocb);
 579
 580         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 581                 ocfs2_iocb_clear_unaligned_aio(iocb);
 582
 583                 if (atomic_dec_and_test(&OCFS2_I(inode)->ip_unaligned_aio) &&
 584                     waitqueue_active(wq)) {
 585                         wake_up_all(wq);
 586                 }
 587         }
 588
 589         ocfs2_iocb_clear_rw_locked(iocb);
 590
 591         level = ocfs2_iocb_rw_locked_level(iocb);
 592         ocfs2_rw_unlock(inode, level);
 593 }
 594
 595 /*
 596  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 597  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 598  * do journalled data.
 599  */
 600 static void ocfs2_invalidatepage(struct page *page, unsigned int offset,
 601                                  unsigned int length)
 602 {
 603         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 604
 605         jbd2_journal_invalidatepage(journal, page, offset, length);
 606 }
 607
 608 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 609 {
 610         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 611
 612         if (!page_has_buffers(page))
 613                 return 0;
 614         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 615 }
 616
 617 static ssize_t ocfs2_direct_IO(int rw,
 618                                struct kiocb *iocb,
 619                                struct iov_iter *iter,
 620                                loff_t offset)
 621 {
 622         struct file *file = iocb->ki_filp;
 623         struct inode *inode = file_inode(file)->i_mapping->host;
 624
 625         /*
 626          * Fallback to buffered I/O if we see an inode without
 627          * extents.
 628          */
 629         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 630                 return 0;
 631
 632         /* Fallback to buffered I/O if we are appending. */
 633         if (i_size_read(inode) <= offset)
 634                 return 0;
 635
 636         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 637                                     iter, offset, ocfs2_direct_IO_get_blocks,
 638                                     ocfs2_dio_end_io, NULL, 0);
 639 }
 640
 641 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 642                                             u32 cpos,
 643                                             unsigned int *start,
 644                                             unsigned int *end)
 645 {
 646         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 647
 648         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 649                 unsigned int cpp;
 650
 651                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 652
 653                 cluster_start = cpos % cpp;
 654                 cluster_start = cluster_start << osb->s_clustersize_bits;
 655
 656                 cluster_end = cluster_start + osb->s_clustersize;
 657         }
 658
 659         BUG_ON(cluster_start > PAGE_SIZE);
 660         BUG_ON(cluster_end > PAGE_SIZE);
 661
 662         if (start)
 663                 *start = cluster_start;
 664         if (end)
 665                 *end = cluster_end;
 666 }
 667
 668 /*
 669  * 'from' and 'to' are the region in the page to avoid zeroing.
 670  *
 671  * If pagesize > clustersize, this function will avoid zeroing outside
 672  * of the cluster boundary.
 673  *
 674  * from == to == 0 is code for "zero the entire cluster region"
 675  */
 676 static void ocfs2_clear_page_regions(struct page *page,
 677                                      struct ocfs2_super *osb, u32 cpos,
 678                                      unsigned from, unsigned to)
 679 {
 680         void *kaddr;
 681         unsigned int cluster_start, cluster_end;
 682
 683         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 684
 685         kaddr = kmap_atomic(page);
 686
 687         if (from || to) {
 688                 if (from > cluster_start)
 689                         memset(kaddr + cluster_start, 0, from - cluster_start);
 690                 if (to < cluster_end)
 691                         memset(kaddr + to, 0, cluster_end - to);
 692         } else {
 693                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 694         }
 695
 696         kunmap_atomic(kaddr);
 697 }
 698
 699 /*
 700  * Nonsparse file systems fully allocate before we get to the write
 701  * code. This prevents ocfs2_write() from tagging the write as an
 702  * allocating one, which means ocfs2_map_page_blocks() might try to
 703  * read-in the blocks at the tail of our file. Avoid reading them by
 704  * testing i_size against each block offset.
 705  */
 706 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 707                                  unsigned int block_start)
 708 {
 709         u64 offset = page_offset(page) + block_start;
 710
 711         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 712                 return 1;
 713
 714         if (i_size_read(inode) > offset)
 715                 return 1;
 716
 717         return 0;
 718 }
 719
 720 /*
 721  * Some of this taken from __block_write_begin(). We already have our
 722  * mapping by now though, and the entire write will be allocating or
 723  * it won't, so not much need to use BH_New.
 724  *
 725  * This will also skip zeroing, which is handled externally.
 726  */
 727 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 728                           struct inode *inode, unsigned int from,
 729                           unsigned int to, int new)
 730 {
 731         int ret = 0;
 732         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 733         unsigned int block_end, block_start;
 734         unsigned int bsize = 1 << inode->i_blkbits;
 735
 736         if (!page_has_buffers(page))
 737                 create_empty_buffers(page, bsize, 0);
 738
 739         head = page_buffers(page);
 740         for (bh = head, block_start = 0; bh != head || !block_start;
 741              bh = bh->b_this_page, block_start += bsize) {
 742                 block_end = block_start + bsize;
 743
 744                 clear_buffer_new(bh);
 745
 746                 /*
 747                  * Ignore blocks outside of our i/o range -
 748                  * they may belong to unallocated clusters.
 749                  */
 750                 if (block_start >= to || block_end <= from) {
 751                         if (PageUptodate(page))
 752                                 set_buffer_uptodate(bh);
 753                         continue;
 754                 }
 755
 756                 /*
 757                  * For an allocating write with cluster size >= page
 758                  * size, we always write the entire page.
 759                  */
 760                 if (new)
 761                         set_buffer_new(bh);
 762
 763                 if (!buffer_mapped(bh)) {
 764                         map_bh(bh, inode->i_sb, *p_blkno);
 765                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 766                 }
 767
 768                 if (PageUptodate(page)) {
 769                         if (!buffer_uptodate(bh))
 770                                 set_buffer_uptodate(bh);
 771                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 772                            !buffer_new(bh) &&
 773                            ocfs2_should_read_blk(inode, page, block_start) &&
 774                            (block_start < from || block_end > to)) {
 775                         ll_rw_block(READ, 1, &bh);
 776                         *wait_bh++=bh;
 777                 }
 778
 779                 *p_blkno = *p_blkno + 1;
 780         }
 781
 782         /*
 783          * If we issued read requests - let them complete.
 784          */
 785         while(wait_bh > wait) {
 786                 wait_on_buffer(*--wait_bh);
 787                 if (!buffer_uptodate(*wait_bh))
 788                         ret = -EIO;
 789         }
 790
 791         if (ret == 0 || !new)
 792                 return ret;
 793
 794         /*
 795          * If we get -EIO above, zero out any newly allocated blocks
 796          * to avoid exposing stale data.
 797          */
 798         bh = head;
 799         block_start = 0;
 800         do {
 801                 block_end = block_start + bsize;
 802                 if (block_end <= from)
 803                         goto next_bh;
 804                 if (block_start >= to)
 805                         break;
 806
 807                 zero_user(page, block_start, bh->b_size);
 808                 set_buffer_uptodate(bh);
 809                 mark_buffer_dirty(bh);
 810
 811 next_bh:
 812                 block_start = block_end;
 813                 bh = bh->b_this_page;
 814         } while (bh != head);
 815
 816         return ret;
 817 }
 818
 819 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 820 #define OCFS2_MAX_CTXT_PAGES    1
 821 #else
 822 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 823 #endif
 824
 825 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 826
 827 /*
 828  * Describe the state of a single cluster to be written to.
 829  */
 830 struct ocfs2_write_cluster_desc {
 831         u32             c_cpos;
 832         u32             c_phys;
 833         /*
 834          * Give this a unique field because c_phys eventually gets
 835          * filled.
 836          */
 837         unsigned        c_new;
 838         unsigned        c_unwritten;
 839         unsigned        c_needs_zero;
 840 };
 841
 842 struct ocfs2_write_ctxt {
 843         /* Logical cluster position / len of write */
 844         u32                             w_cpos;
 845         u32                             w_clen;
 846
 847         /* First cluster allocated in a nonsparse extend */
 848         u32                             w_first_new_cpos;
 849
 850         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 851
 852         /*
 853          * This is true if page_size > cluster_size.
 854          *
 855          * It triggers a set of special cases during write which might
 856          * have to deal with allocating writes to partial pages.
 857          */
 858         unsigned int                    w_large_pages;
 859
 860         /*
 861          * Pages involved in this write.
 862          *
 863          * w_target_page is the page being written to by the user.
 864          *
 865          * w_pages is an array of pages which always contains
 866          * w_target_page, and in the case of an allocating write with
 867          * page_size < cluster size, it will contain zero'd and mapped
 868          * pages adjacent to w_target_page which need to be written
 869          * out in so that future reads from that region will get
 870          * zero's.
 871          */
 872         unsigned int                    w_num_pages;
 873         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 874         struct page                     *w_target_page;
 875
 876         /*
 877          * w_target_locked is used for page_mkwrite path indicating no unlocking
 878          * against w_target_page in ocfs2_write_end_nolock.
 879          */
 880         unsigned int                    w_target_locked:1;
 881
 882         /*
 883          * ocfs2_write_end() uses this to know what the real range to
 884          * write in the target should be.
 885          */
 886         unsigned int                    w_target_from;
 887         unsigned int                    w_target_to;
 888
 889         /*
 890          * We could use journal_current_handle() but this is cleaner,
 891          * IMHO -Mark
 892          */
 893         handle_t                        *w_handle;
 894
 895         struct buffer_head              *w_di_bh;
 896
 897         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 898 };
 899
 900 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 901 {
 902         int i;
 903
 904         for(i = 0; i < num_pages; i++) {
 905                 if (pages[i]) {
 906                         unlock_page(pages[i]);
 907                         mark_page_accessed(pages[i]);
 908                         page_cache_release(pages[i]);
 909                 }
 910         }
 911 }
 912
 913 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 914 {
 915         int i;
 916
 917         /*
 918          * w_target_locked is only set to true in the page_mkwrite() case.
 919          * The intent is to allow us to lock the target page from write_begin()
 920          * to write_end(). The caller must hold a ref on w_target_page.
 921          */
 922         if (wc->w_target_locked) {
 923                 BUG_ON(!wc->w_target_page);
 924                 for (i = 0; i < wc->w_num_pages; i++) {
 925                         if (wc->w_target_page == wc->w_pages[i]) {
 926                                 wc->w_pages[i] = NULL;
 927                                 break;
 928                         }
 929                 }
 930                 mark_page_accessed(wc->w_target_page);
 931                 page_cache_release(wc->w_target_page);
 932         }
 933         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 934
 935         brelse(wc->w_di_bh);
 936         kfree(wc);
 937 }
 938
 939 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 940                                   struct ocfs2_super *osb, loff_t pos,
 941                                   unsigned len, struct buffer_head *di_bh)
 942 {
 943         u32 cend;
 944         struct ocfs2_write_ctxt *wc;
 945
 946         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 947         if (!wc)
 948                 return -ENOMEM;
 949
 950         wc->w_cpos = pos >> osb->s_clustersize_bits;
 951         wc->w_first_new_cpos = UINT_MAX;
 952         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 953         wc->w_clen = cend - wc->w_cpos + 1;
 954         get_bh(di_bh);
 955         wc->w_di_bh = di_bh;
 956
 957         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 958                 wc->w_large_pages = 1;
 959         else
 960                 wc->w_large_pages = 0;
 961
 962         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 963
 964         *wcp = wc;
 965
 966         return 0;
 967 }
 968
 969 /*
 970  * If a page has any new buffers, zero them out here, and mark them uptodate
 971  * and dirty so they'll be written out (in order to prevent uninitialised
 972  * block data from leaking). And clear the new bit.
 973  */
 974 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 975 {
 976         unsigned int block_start, block_end;
 977         struct buffer_head *head, *bh;
 978
 979         BUG_ON(!PageLocked(page));
 980         if (!page_has_buffers(page))
 981                 return;
 982
 983         bh = head = page_buffers(page);
 984         block_start = 0;
 985         do {
 986                 block_end = block_start + bh->b_size;
 987
 988                 if (buffer_new(bh)) {
 989                         if (block_end > from && block_start < to) {
 990                                 if (!PageUptodate(page)) {
 991                                         unsigned start, end;
 992
 993                                         start = max(from, block_start);
 994                                         end = min(to, block_end);
 995
 996                                         zero_user_segment(page, start, end);
 997                                         set_buffer_uptodate(bh);
 998                                 }
 999
1000                                 clear_buffer_new(bh);
1001                                 mark_buffer_dirty(bh);
1002                         }
1003                 }
1004
1005                 block_start = block_end;
1006                 bh = bh->b_this_page;
1007         } while (bh != head);
1008 }
1009
1010 /*
1011  * Only called when we have a failure during allocating write to write
1012  * zero's to the newly allocated region.
1013  */
1014 static void ocfs2_write_failure(struct inode *inode,
1015                                 struct ocfs2_write_ctxt *wc,
1016                                 loff_t user_pos, unsigned user_len)
1017 {
1018         int i;
1019         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1020                 to = user_pos + user_len;
1021         struct page *tmppage;
1022
1023         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1024
1025         for(i = 0; i < wc->w_num_pages; i++) {
1026                 tmppage = wc->w_pages[i];
1027
1028                 if (page_has_buffers(tmppage)) {
1029                         if (ocfs2_should_order_data(inode))
1030                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1031
1032                         block_commit_write(tmppage, from, to);
1033                 }
1034         }
1035 }
1036
1037 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1038                                         struct ocfs2_write_ctxt *wc,
1039                                         struct page *page, u32 cpos,
1040                                         loff_t user_pos, unsigned user_len,
1041                                         int new)
1042 {
1043         int ret;
1044         unsigned int map_from = 0, map_to = 0;
1045         unsigned int cluster_start, cluster_end;
1046         unsigned int user_data_from = 0, user_data_to = 0;
1047
1048         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1049                                         &cluster_start, &cluster_end);
1050
1051         /* treat the write as new if the a hole/lseek spanned across
1052          * the page boundary.
1053          */
1054         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1055                         (page_offset(page) <= user_pos));
1056
1057         if (page == wc->w_target_page) {
1058                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1059                 map_to = map_from + user_len;
1060
1061                 if (new)
1062                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1063                                                     cluster_start, cluster_end,
1064                                                     new);
1065                 else
1066                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1067                                                     map_from, map_to, new);
1068                 if (ret) {
1069                         mlog_errno(ret);
1070                         goto out;
1071                 }
1072
1073                 user_data_from = map_from;
1074                 user_data_to = map_to;
1075                 if (new) {
1076                         map_from = cluster_start;
1077                         map_to = cluster_end;
1078                 }
1079         } else {
1080                 /*
1081                  * If we haven't allocated the new page yet, we
1082                  * shouldn't be writing it out without copying user
1083                  * data. This is likely a math error from the caller.
1084                  */
1085                 BUG_ON(!new);
1086
1087                 map_from = cluster_start;
1088                 map_to = cluster_end;
1089
1090                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1091                                             cluster_start, cluster_end, new);
1092                 if (ret) {
1093                         mlog_errno(ret);
1094                         goto out;
1095                 }
1096         }
1097
1098         /*
1099          * Parts of newly allocated pages need to be zero'd.
1100          *
1101          * Above, we have also rewritten 'to' and 'from' - as far as
1102          * the rest of the function is concerned, the entire cluster
1103          * range inside of a page needs to be written.
1104          *
1105          * We can skip this if the page is up to date - it's already
1106          * been zero'd from being read in as a hole.
1107          */
1108         if (new && !PageUptodate(page))
1109                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1110                                          cpos, user_data_from, user_data_to);
1111
1112         flush_dcache_page(page);
1113
1114 out:
1115         return ret;
1116 }
1117
1118 /*
1119  * This function will only grab one clusters worth of pages.
1120  */
1121 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1122                                       struct ocfs2_write_ctxt *wc,
1123                                       u32 cpos, loff_t user_pos,
1124                                       unsigned user_len, int new,
1125                                       struct page *mmap_page)
1126 {
1127         int ret = 0, i;
1128         unsigned long start, target_index, end_index, index;
1129         struct inode *inode = mapping->host;
1130         loff_t last_byte;
1131
1132         target_index = user_pos >> PAGE_CACHE_SHIFT;
1133
1134         /*
1135          * Figure out how many pages we'll be manipulating here. For
1136          * non allocating write, we just change the one
1137          * page. Otherwise, we'll need a whole clusters worth.  If we're
1138          * writing past i_size, we only need enough pages to cover the
1139          * last page of the write.
1140          */
1141         if (new) {
1142                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1143                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1144                 /*
1145                  * We need the index *past* the last page we could possibly
1146                  * touch.  This is the page past the end of the write or
1147                  * i_size, whichever is greater.
1148                  */
1149                 last_byte = max(user_pos + user_len, i_size_read(inode));
1150                 BUG_ON(last_byte < 1);
1151                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1152                 if ((start + wc->w_num_pages) > end_index)
1153                         wc->w_num_pages = end_index - start;
1154         } else {
1155                 wc->w_num_pages = 1;
1156                 start = target_index;
1157         }
1158
1159         for(i = 0; i < wc->w_num_pages; i++) {
1160                 index = start + i;
1161
1162                 if (index == target_index && mmap_page) {
1163                         /*
1164                          * ocfs2_pagemkwrite() is a little different
1165                          * and wants us to directly use the page
1166                          * passed in.
1167                          */
1168                         lock_page(mmap_page);
1169
1170                         /* Exit and let the caller retry */
1171                         if (mmap_page->mapping != mapping) {
1172                                 WARN_ON(mmap_page->mapping);
1173                                 unlock_page(mmap_page);
1174                                 ret = -EAGAIN;
1175                                 goto out;
1176                         }
1177
1178                         page_cache_get(mmap_page);
1179                         wc->w_pages[i] = mmap_page;
1180                         wc->w_target_locked = true;
1181                 } else {
1182                         wc->w_pages[i] = find_or_create_page(mapping, index,
1183                                                              GFP_NOFS);
1184                         if (!wc->w_pages[i]) {
1185                                 ret = -ENOMEM;
1186                                 mlog_errno(ret);
1187                                 goto out;
1188                         }
1189                 }
1190                 wait_for_stable_page(wc->w_pages[i]);
1191
1192                 if (index == target_index)
1193                         wc->w_target_page = wc->w_pages[i];
1194         }
1195 out:
1196         if (ret)
1197                 wc->w_target_locked = false;
1198         return ret;
1199 }
1200
1201 /*
1202  * Prepare a single cluster for write one cluster into the file.
1203  */
1204 static int ocfs2_write_cluster(struct address_space *mapping,
1205                                u32 phys, unsigned int unwritten,
1206                                unsigned int should_zero,
1207                                struct ocfs2_alloc_context *data_ac,
1208                                struct ocfs2_alloc_context *meta_ac,
1209                                struct ocfs2_write_ctxt *wc, u32 cpos,
1210                                loff_t user_pos, unsigned user_len)
1211 {
1212         int ret, i, new;
1213         u64 v_blkno, p_blkno;
1214         struct inode *inode = mapping->host;
1215         struct ocfs2_extent_tree et;
1216
1217         new = phys == 0 ? 1 : 0;
1218         if (new) {
1219                 u32 tmp_pos;
1220
1221                 /*
1222                  * This is safe to call with the page locks - it won't take
1223                  * any additional semaphores or cluster locks.
1224                  */
1225                 tmp_pos = cpos;
1226                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1227                                            &tmp_pos, 1, 0, wc->w_di_bh,
1228                                            wc->w_handle, data_ac,
1229                                            meta_ac, NULL);
1230                 /*
1231                  * This shouldn't happen because we must have already
1232                  * calculated the correct meta data allocation required. The
1233                  * internal tree allocation code should know how to increase
1234                  * transaction credits itself.
1235                  *
1236                  * If need be, we could handle -EAGAIN for a
1237                  * RESTART_TRANS here.
1238                  */
1239                 mlog_bug_on_msg(ret == -EAGAIN,
1240                                 "Inode %llu: EAGAIN return during allocation.\n",
1241                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1242                 if (ret < 0) {
1243                         mlog_errno(ret);
1244                         goto out;
1245                 }
1246         } else if (unwritten) {
1247                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1248                                               wc->w_di_bh);
1249                 ret = ocfs2_mark_extent_written(inode, &et,
1250                                                 wc->w_handle, cpos, 1, phys,
1251                                                 meta_ac, &wc->w_dealloc);
1252                 if (ret < 0) {
1253                         mlog_errno(ret);
1254                         goto out;
1255                 }
1256         }
1257
1258         if (should_zero)
1259                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1260         else
1261                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1262
1263         /*
1264          * The only reason this should fail is due to an inability to
1265          * find the extent added.
1266          */
1267         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1268                                           NULL);
1269         if (ret < 0) {
1270                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1271                             "at logical block %llu",
1272                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1273                             (unsigned long long)v_blkno);
1274                 goto out;
1275         }
1276
1277         BUG_ON(p_blkno == 0);
1278
1279         for(i = 0; i < wc->w_num_pages; i++) {
1280                 int tmpret;
1281
1282                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1283                                                       wc->w_pages[i], cpos,
1284                                                       user_pos, user_len,
1285                                                       should_zero);
1286                 if (tmpret) {
1287                         mlog_errno(tmpret);
1288                         if (ret == 0)
1289                                 ret = tmpret;
1290                 }
1291         }
1292
1293         /*
1294          * We only have cleanup to do in case of allocating write.
1295          */
1296         if (ret && new)
1297                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1298
1299 out:
1300
1301         return ret;
1302 }
1303
1304 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1305                                        struct ocfs2_alloc_context *data_ac,
1306                                        struct ocfs2_alloc_context *meta_ac,
1307                                        struct ocfs2_write_ctxt *wc,
1308                                        loff_t pos, unsigned len)
1309 {
1310         int ret, i;
1311         loff_t cluster_off;
1312         unsigned int local_len = len;
1313         struct ocfs2_write_cluster_desc *desc;
1314         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1315
1316         for (i = 0; i < wc->w_clen; i++) {
1317                 desc = &wc->w_desc[i];
1318
1319                 /*
1320                  * We have to make sure that the total write passed in
1321                  * doesn't extend past a single cluster.
1322                  */
1323                 local_len = len;
1324                 cluster_off = pos & (osb->s_clustersize - 1);
1325                 if ((cluster_off + local_len) > osb->s_clustersize)
1326                         local_len = osb->s_clustersize - cluster_off;
1327
1328                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1329                                           desc->c_unwritten,
1330                                           desc->c_needs_zero,
1331                                           data_ac, meta_ac,
1332                                           wc, desc->c_cpos, pos, local_len);
1333                 if (ret) {
1334                         mlog_errno(ret);
1335                         goto out;
1336                 }
1337
1338                 len -= local_len;
1339                 pos += local_len;
1340         }
1341
1342         ret = 0;
1343 out:
1344         return ret;
1345 }
1346
1347 /*
1348  * ocfs2_write_end() wants to know which parts of the target page it
1349  * should complete the write on. It's easiest to compute them ahead of
1350  * time when a more complete view of the write is available.
1351  */
1352 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1353                                         struct ocfs2_write_ctxt *wc,
1354                                         loff_t pos, unsigned len, int alloc)
1355 {
1356         struct ocfs2_write_cluster_desc *desc;
1357
1358         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1359         wc->w_target_to = wc->w_target_from + len;
1360
1361         if (alloc == 0)
1362                 return;
1363
1364         /*
1365          * Allocating write - we may have different boundaries based
1366          * on page size and cluster size.
1367          *
1368          * NOTE: We can no longer compute one value from the other as
1369          * the actual write length and user provided length may be
1370          * different.
1371          */
1372
1373         if (wc->w_large_pages) {
1374                 /*
1375                  * We only care about the 1st and last cluster within
1376                  * our range and whether they should be zero'd or not. Either
1377                  * value may be extended out to the start/end of a
1378                  * newly allocated cluster.
1379                  */
1380                 desc = &wc->w_desc[0];
1381                 if (desc->c_needs_zero)
1382                         ocfs2_figure_cluster_boundaries(osb,
1383                                                         desc->c_cpos,
1384                                                         &wc->w_target_from,
1385                                                         NULL);
1386
1387                 desc = &wc->w_desc[wc->w_clen - 1];
1388                 if (desc->c_needs_zero)
1389                         ocfs2_figure_cluster_boundaries(osb,
1390                                                         desc->c_cpos,
1391                                                         NULL,
1392                                                         &wc->w_target_to);
1393         } else {
1394                 wc->w_target_from = 0;
1395                 wc->w_target_to = PAGE_CACHE_SIZE;
1396         }
1397 }
1398
1399 /*
1400  * Populate each single-cluster write descriptor in the write context
1401  * with information about the i/o to be done.
1402  *
1403  * Returns the number of clusters that will have to be allocated, as
1404  * well as a worst case estimate of the number of extent records that
1405  * would have to be created during a write to an unwritten region.
1406  */
1407 static int ocfs2_populate_write_desc(struct inode *inode,
1408                                      struct ocfs2_write_ctxt *wc,
1409                                      unsigned int *clusters_to_alloc,
1410                                      unsigned int *extents_to_split)
1411 {
1412         int ret;
1413         struct ocfs2_write_cluster_desc *desc;
1414         unsigned int num_clusters = 0;
1415         unsigned int ext_flags = 0;
1416         u32 phys = 0;
1417         int i;
1418
1419         *clusters_to_alloc = 0;
1420         *extents_to_split = 0;
1421
1422         for (i = 0; i < wc->w_clen; i++) {
1423                 desc = &wc->w_desc[i];
1424                 desc->c_cpos = wc->w_cpos + i;
1425
1426                 if (num_clusters == 0) {
1427                         /*
1428                          * Need to look up the next extent record.
1429                          */
1430                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1431                                                  &num_clusters, &ext_flags);
1432                         if (ret) {
1433                                 mlog_errno(ret);
1434                                 goto out;
1435                         }
1436
1437                         /* We should already CoW the refcountd extent. */
1438                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1439
1440                         /*
1441                          * Assume worst case - that we're writing in
1442                          * the middle of the extent.
1443                          *
1444                          * We can assume that the write proceeds from
1445                          * left to right, in which case the extent
1446                          * insert code is smart enough to coalesce the
1447                          * next splits into the previous records created.
1448                          */
1449                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1450                                 *extents_to_split = *extents_to_split + 2;
1451                 } else if (phys) {
1452                         /*
1453                          * Only increment phys if it doesn't describe
1454                          * a hole.
1455                          */
1456                         phys++;
1457                 }
1458
1459                 /*
1460                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1461                  * file that got extended.  w_first_new_cpos tells us
1462                  * where the newly allocated clusters are so we can
1463                  * zero them.
1464                  */
1465                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1466                         BUG_ON(phys == 0);
1467                         desc->c_needs_zero = 1;
1468                 }
1469
1470                 desc->c_phys = phys;
1471                 if (phys == 0) {
1472                         desc->c_new = 1;
1473                         desc->c_needs_zero = 1;
1474                         *clusters_to_alloc = *clusters_to_alloc + 1;
1475                 }
1476
1477                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1478                         desc->c_unwritten = 1;
1479                         desc->c_needs_zero = 1;
1480                 }
1481
1482                 num_clusters--;
1483         }
1484
1485         ret = 0;
1486 out:
1487         return ret;
1488 }
1489
1490 static int ocfs2_write_begin_inline(struct address_space *mapping,
1491                                     struct inode *inode,
1492                                     struct ocfs2_write_ctxt *wc)
1493 {
1494         int ret;
1495         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1496         struct page *page;
1497         handle_t *handle;
1498         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1499
1500         page = find_or_create_page(mapping, 0, GFP_NOFS);
1501         if (!page) {
1502                 ret = -ENOMEM;
1503                 mlog_errno(ret);
1504                 goto out;
1505         }
1506         /*
1507          * If we don't set w_num_pages then this page won't get unlocked
1508          * and freed on cleanup of the write context.
1509          */
1510         wc->w_pages[0] = wc->w_target_page = page;
1511         wc->w_num_pages = 1;
1512
1513         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1514         if (IS_ERR(handle)) {
1515                 ret = PTR_ERR(handle);
1516                 mlog_errno(ret);
1517                 goto out;
1518         }
1519
1520         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1521                                       OCFS2_JOURNAL_ACCESS_WRITE);
1522         if (ret) {
1523                 ocfs2_commit_trans(osb, handle);
1524
1525                 mlog_errno(ret);
1526                 goto out;
1527         }
1528
1529         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1530                 ocfs2_set_inode_data_inline(inode, di);
1531
1532         if (!PageUptodate(page)) {
1533                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1534                 if (ret) {
1535                         ocfs2_commit_trans(osb, handle);
1536
1537                         goto out;
1538                 }
1539         }
1540
1541         wc->w_handle = handle;
1542 out:
1543         return ret;
1544 }
1545
1546 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1547 {
1548         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1549
1550         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1551                 return 1;
1552         return 0;
1553 }
1554
1555 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1556                                           struct inode *inode, loff_t pos,
1557                                           unsigned len, struct page *mmap_page,
1558                                           struct ocfs2_write_ctxt *wc)
1559 {
1560         int ret, written = 0;
1561         loff_t end = pos + len;
1562         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1563         struct ocfs2_dinode *di = NULL;
1564
1565         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1566                                              len, (unsigned long long)pos,
1567                                              oi->ip_dyn_features);
1568
1569         /*
1570          * Handle inodes which already have inline data 1st.
1571          */
1572         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1573                 if (mmap_page == NULL &&
1574                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1575                         goto do_inline_write;
1576
1577                 /*
1578                  * The write won't fit - we have to give this inode an
1579                  * inline extent list now.
1580                  */
1581                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1582                 if (ret)
1583                         mlog_errno(ret);
1584                 goto out;
1585         }
1586
1587         /*
1588          * Check whether the inode can accept inline data.
1589          */
1590         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1591                 return 0;
1592
1593         /*
1594          * Check whether the write can fit.
1595          */
1596         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1597         if (mmap_page ||
1598             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1599                 return 0;
1600
1601 do_inline_write:
1602         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1603         if (ret) {
1604                 mlog_errno(ret);
1605                 goto out;
1606         }
1607
1608         /*
1609          * This signals to the caller that the data can be written
1610          * inline.
1611          */
1612         written = 1;
1613 out:
1614         return written ? written : ret;
1615 }
1616
1617 /*
1618  * This function only does anything for file systems which can't
1619  * handle sparse files.
1620  *
1621  * What we want to do here is fill in any hole between the current end
1622  * of allocation and the end of our write. That way the rest of the
1623  * write path can treat it as an non-allocating write, which has no
1624  * special case code for sparse/nonsparse files.
1625  */
1626 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1627                                         struct buffer_head *di_bh,
1628                                         loff_t pos, unsigned len,
1629                                         struct ocfs2_write_ctxt *wc)
1630 {
1631         int ret;
1632         loff_t newsize = pos + len;
1633
1634         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1635
1636         if (newsize <= i_size_read(inode))
1637                 return 0;
1638
1639         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1640         if (ret)
1641                 mlog_errno(ret);
1642
1643         wc->w_first_new_cpos =
1644                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1645
1646         return ret;
1647 }
1648
1649 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1650                            loff_t pos)
1651 {
1652         int ret = 0;
1653
1654         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1655         if (pos > i_size_read(inode))
1656                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1657
1658         return ret;
1659 }
1660
1661 /*
1662  * Try to flush truncate logs if we can free enough clusters from it.
1663  * As for return value, "< 0" means error, "0" no space and "1" means
1664  * we have freed enough spaces and let the caller try to allocate again.
1665  */
1666 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1667                                           unsigned int needed)
1668 {
1669         tid_t target;
1670         int ret = 0;
1671         unsigned int truncated_clusters;
1672
1673         mutex_lock(&osb->osb_tl_inode->i_mutex);
1674         truncated_clusters = osb->truncated_clusters;
1675         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1676
1677         /*
1678          * Check whether we can succeed in allocating if we free
1679          * the truncate log.
1680          */
1681         if (truncated_clusters < needed)
1682                 goto out;
1683
1684         ret = ocfs2_flush_truncate_log(osb);
1685         if (ret) {
1686                 mlog_errno(ret);
1687                 goto out;
1688         }
1689
1690         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1691                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1692                 ret = 1;
1693         }
1694 out:
1695         return ret;
1696 }
1697
1698 int ocfs2_write_begin_nolock(struct file *filp,
1699                              struct address_space *mapping,
1700                              loff_t pos, unsigned len, unsigned flags,
1701                              struct page **pagep, void **fsdata,
1702                              struct buffer_head *di_bh, struct page *mmap_page)
1703 {
1704         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1705         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1706         struct ocfs2_write_ctxt *wc;
1707         struct inode *inode = mapping->host;
1708         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1709         struct ocfs2_dinode *di;
1710         struct ocfs2_alloc_context *data_ac = NULL;
1711         struct ocfs2_alloc_context *meta_ac = NULL;
1712         handle_t *handle;
1713         struct ocfs2_extent_tree et;
1714         int try_free = 1, ret1;
1715
1716 try_again:
1717         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1718         if (ret) {
1719                 mlog_errno(ret);
1720                 return ret;
1721         }
1722
1723         if (ocfs2_supports_inline_data(osb)) {
1724                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1725                                                      mmap_page, wc);
1726                 if (ret == 1) {
1727                         ret = 0;
1728                         goto success;
1729                 }
1730                 if (ret < 0) {
1731                         mlog_errno(ret);
1732                         goto out;
1733                 }
1734         }
1735
1736         if (ocfs2_sparse_alloc(osb))
1737                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1738         else
1739                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1740                                                    wc);
1741         if (ret) {
1742                 mlog_errno(ret);
1743                 goto out;
1744         }
1745
1746         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1747         if (ret < 0) {
1748                 mlog_errno(ret);
1749                 goto out;
1750         } else if (ret == 1) {
1751                 clusters_need = wc->w_clen;
1752                 ret = ocfs2_refcount_cow(inode, di_bh,
1753                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1754                 if (ret) {
1755                         mlog_errno(ret);
1756                         goto out;
1757                 }
1758         }
1759
1760         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1761                                         &extents_to_split);
1762         if (ret) {
1763                 mlog_errno(ret);
1764                 goto out;
1765         }
1766         clusters_need += clusters_to_alloc;
1767
1768         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1769
1770         trace_ocfs2_write_begin_nolock(
1771                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1772                         (long long)i_size_read(inode),
1773                         le32_to_cpu(di->i_clusters),
1774                         pos, len, flags, mmap_page,
1775                         clusters_to_alloc, extents_to_split);
1776
1777         /*
1778          * We set w_target_from, w_target_to here so that
1779          * ocfs2_write_end() knows which range in the target page to
1780          * write out. An allocation requires that we write the entire
1781          * cluster range.
1782          */
1783         if (clusters_to_alloc || extents_to_split) {
1784                 /*
1785                  * XXX: We are stretching the limits of
1786                  * ocfs2_lock_allocators(). It greatly over-estimates
1787                  * the work to be done.
1788                  */
1789                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1790                                               wc->w_di_bh);
1791                 ret = ocfs2_lock_allocators(inode, &et,
1792                                             clusters_to_alloc, extents_to_split,
1793                                             &data_ac, &meta_ac);
1794                 if (ret) {
1795                         mlog_errno(ret);
1796                         goto out;
1797                 }
1798
1799                 if (data_ac)
1800                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1801
1802                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1803                                                     &di->id2.i_list,
1804                                                     clusters_to_alloc);
1805
1806         }
1807
1808         /*
1809          * We have to zero sparse allocated clusters, unwritten extent clusters,
1810          * and non-sparse clusters we just extended.  For non-sparse writes,
1811          * we know zeros will only be needed in the first and/or last cluster.
1812          */
1813         if (clusters_to_alloc || extents_to_split ||
1814             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1815                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1816                 cluster_of_pages = 1;
1817         else
1818                 cluster_of_pages = 0;
1819
1820         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1821
1822         handle = ocfs2_start_trans(osb, credits);
1823         if (IS_ERR(handle)) {
1824                 ret = PTR_ERR(handle);
1825                 mlog_errno(ret);
1826                 goto out;
1827         }
1828
1829         wc->w_handle = handle;
1830
1831         if (clusters_to_alloc) {
1832                 ret = dquot_alloc_space_nodirty(inode,
1833                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1834                 if (ret)
1835                         goto out_commit;
1836         }
1837         /*
1838          * We don't want this to fail in ocfs2_write_end(), so do it
1839          * here.
1840          */
1841         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1842                                       OCFS2_JOURNAL_ACCESS_WRITE);
1843         if (ret) {
1844                 mlog_errno(ret);
1845                 goto out_quota;
1846         }
1847
1848         /*
1849          * Fill our page array first. That way we've grabbed enough so
1850          * that we can zero and flush if we error after adding the
1851          * extent.
1852          */
1853         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1854                                          cluster_of_pages, mmap_page);
1855         if (ret && ret != -EAGAIN) {
1856                 mlog_errno(ret);
1857                 goto out_quota;
1858         }
1859
1860         /*
1861          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1862          * the target page. In this case, we exit with no error and no target
1863          * page. This will trigger the caller, page_mkwrite(), to re-try
1864          * the operation.
1865          */
1866         if (ret == -EAGAIN) {
1867                 BUG_ON(wc->w_target_page);
1868                 ret = 0;
1869                 goto out_quota;
1870         }
1871
1872         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1873                                           len);
1874         if (ret) {
1875                 mlog_errno(ret);
1876                 goto out_quota;
1877         }
1878
1879         if (data_ac)
1880                 ocfs2_free_alloc_context(data_ac);
1881         if (meta_ac)
1882                 ocfs2_free_alloc_context(meta_ac);
1883
1884 success:
1885         *pagep = wc->w_target_page;
1886         *fsdata = wc;
1887         return 0;
1888 out_quota:
1889         if (clusters_to_alloc)
1890                 dquot_free_space(inode,
1891                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1892 out_commit:
1893         ocfs2_commit_trans(osb, handle);
1894
1895 out:
1896         ocfs2_free_write_ctxt(wc);
1897
1898         if (data_ac)
1899                 ocfs2_free_alloc_context(data_ac);
1900         if (meta_ac)
1901                 ocfs2_free_alloc_context(meta_ac);
1902
1903         if (ret == -ENOSPC && try_free) {
1904                 /*
1905                  * Try to free some truncate log so that we can have enough
1906                  * clusters to allocate.
1907                  */
1908                 try_free = 0;
1909
1910                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1911                 if (ret1 == 1)
1912                         goto try_again;
1913
1914                 if (ret1 < 0)
1915                         mlog_errno(ret1);
1916         }
1917
1918         return ret;
1919 }
1920
1921 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1922                              loff_t pos, unsigned len, unsigned flags,
1923                              struct page **pagep, void **fsdata)
1924 {
1925         int ret;
1926         struct buffer_head *di_bh = NULL;
1927         struct inode *inode = mapping->host;
1928
1929         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1930         if (ret) {
1931                 mlog_errno(ret);
1932                 return ret;
1933         }
1934
1935         /*
1936          * Take alloc sem here to prevent concurrent lookups. That way
1937          * the mapping, zeroing and tree manipulation within
1938          * ocfs2_write() will be safe against ->readpage(). This
1939          * should also serve to lock out allocation from a shared
1940          * writeable region.
1941          */
1942         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1943
1944         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1945                                        fsdata, di_bh, NULL);
1946         if (ret) {
1947                 mlog_errno(ret);
1948                 goto out_fail;
1949         }
1950
1951         brelse(di_bh);
1952
1953         return 0;
1954
1955 out_fail:
1956         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1957
1958         brelse(di_bh);
1959         ocfs2_inode_unlock(inode, 1);
1960
1961         return ret;
1962 }
1963
1964 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1965                                    unsigned len, unsigned *copied,
1966                                    struct ocfs2_dinode *di,
1967                                    struct ocfs2_write_ctxt *wc)
1968 {
1969         void *kaddr;
1970
1971         if (unlikely(*copied < len)) {
1972                 if (!PageUptodate(wc->w_target_page)) {
1973                         *copied = 0;
1974                         return;
1975                 }
1976         }
1977
1978         kaddr = kmap_atomic(wc->w_target_page);
1979         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1980         kunmap_atomic(kaddr);
1981
1982         trace_ocfs2_write_end_inline(
1983              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1984              (unsigned long long)pos, *copied,
1985              le16_to_cpu(di->id2.i_data.id_count),
1986              le16_to_cpu(di->i_dyn_features));
1987 }
1988
1989 int ocfs2_write_end_nolock(struct address_space *mapping,
1990                            loff_t pos, unsigned len, unsigned copied,
1991                            struct page *page, void *fsdata)
1992 {
1993         int i;
1994         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1995         struct inode *inode = mapping->host;
1996         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1997         struct ocfs2_write_ctxt *wc = fsdata;
1998         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1999         handle_t *handle = wc->w_handle;
2000         struct page *tmppage;
2001
2002         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2003                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2004                 goto out_write_size;
2005         }
2006
2007         if (unlikely(copied < len)) {
2008                 if (!PageUptodate(wc->w_target_page))
2009                         copied = 0;
2010
2011                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2012                                        start+len);
2013         }
2014         flush_dcache_page(wc->w_target_page);
2015
2016         for(i = 0; i < wc->w_num_pages; i++) {
2017                 tmppage = wc->w_pages[i];
2018
2019                 if (tmppage == wc->w_target_page) {
2020                         from = wc->w_target_from;
2021                         to = wc->w_target_to;
2022
2023                         BUG_ON(from > PAGE_CACHE_SIZE ||
2024                                to > PAGE_CACHE_SIZE ||
2025                                to < from);
2026                 } else {
2027                         /*
2028                          * Pages adjacent to the target (if any) imply
2029                          * a hole-filling write in which case we want
2030                          * to flush their entire range.
2031                          */
2032                         from = 0;
2033                         to = PAGE_CACHE_SIZE;
2034                 }
2035
2036                 if (page_has_buffers(tmppage)) {
2037                         if (ocfs2_should_order_data(inode))
2038                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2039                         block_commit_write(tmppage, from, to);
2040                 }
2041         }
2042
2043 out_write_size:
2044         pos += copied;
2045         if (pos > i_size_read(inode)) {
2046                 i_size_write(inode, pos);
2047                 mark_inode_dirty(inode);
2048         }
2049         inode->i_blocks = ocfs2_inode_sector_count(inode);
2050         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2051         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2052         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2053         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2054         ocfs2_journal_dirty(handle, wc->w_di_bh);
2055
2056         ocfs2_commit_trans(osb, handle);
2057
2058         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2059
2060         ocfs2_free_write_ctxt(wc);
2061
2062         return copied;
2063 }
2064
2065 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2066                            loff_t pos, unsigned len, unsigned copied,
2067                            struct page *page, void *fsdata)
2068 {
2069         int ret;
2070         struct inode *inode = mapping->host;
2071
2072         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2073
2074         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2075         ocfs2_inode_unlock(inode, 1);
2076
2077         return ret;
2078 }
2079
2080 const struct address_space_operations ocfs2_aops = {
2081         .readpage               = ocfs2_readpage,
2082         .readpages              = ocfs2_readpages,
2083         .writepage              = ocfs2_writepage,
2084         .write_begin            = ocfs2_write_begin,
2085         .write_end              = ocfs2_write_end,
2086         .bmap                   = ocfs2_bmap,
2087         .direct_IO              = ocfs2_direct_IO,
2088         .invalidatepage         = ocfs2_invalidatepage,
2089         .releasepage            = ocfs2_releasepage,
2090         .migratepage            = buffer_migrate_page,
2091         .is_partially_uptodate  = block_is_partially_uptodate,
2092         .error_remove_page      = generic_error_remove_page,
2093 };