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, KM_USER0);
 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, KM_USER0);
 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, KM_USER0);
 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, KM_USER0);
 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.  Like the core uses
 563  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 564  * truncation on another.
 565  */
 566 static void ocfs2_dio_end_io(struct kiocb *iocb,
 567                              loff_t offset,
 568                              ssize_t bytes,
 569                              void *private,
 570                              int ret,
 571                              bool is_async)
 572 {
 573         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 574         int level;
 575         wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
 576
 577         /* this io's submitter should not have unlocked this before we could */
 578         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 579
 580         if (ocfs2_iocb_is_sem_locked(iocb)) {
 581                 up_read(&inode->i_alloc_sem);
 582                 ocfs2_iocb_clear_sem_locked(iocb);
 583         }
 584
 585         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 586                 ocfs2_iocb_clear_unaligned_aio(iocb);
 587
 588                 if (atomic_dec_and_test(&OCFS2_I(inode)->ip_unaligned_aio) &&
 589                     waitqueue_active(wq)) {
 590                         wake_up_all(wq);
 591                 }
 592         }
 593
 594         ocfs2_iocb_clear_rw_locked(iocb);
 595
 596         level = ocfs2_iocb_rw_locked_level(iocb);
 597         ocfs2_rw_unlock(inode, level);
 598
 599         if (is_async)
 600                 aio_complete(iocb, ret, 0);
 601 }
 602
 603 /*
 604  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 605  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 606  * do journalled data.
 607  */
 608 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 609 {
 610         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 611
 612         jbd2_journal_invalidatepage(journal, page, offset);
 613 }
 614
 615 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 616 {
 617         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 618
 619         if (!page_has_buffers(page))
 620                 return 0;
 621         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 622 }
 623
 624 static ssize_t ocfs2_direct_IO(int rw,
 625                                struct kiocb *iocb,
 626                                const struct iovec *iov,
 627                                loff_t offset,
 628                                unsigned long nr_segs)
 629 {
 630         struct file *file = iocb->ki_filp;
 631         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 632
 633         /*
 634          * Fallback to buffered I/O if we see an inode without
 635          * extents.
 636          */
 637         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 638                 return 0;
 639
 640         /* Fallback to buffered I/O if we are appending. */
 641         if (i_size_read(inode) <= offset)
 642                 return 0;
 643
 644         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 645                                     iov, offset, nr_segs,
 646                                     ocfs2_direct_IO_get_blocks,
 647                                     ocfs2_dio_end_io, NULL, 0);
 648 }
 649
 650 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 651                                             u32 cpos,
 652                                             unsigned int *start,
 653                                             unsigned int *end)
 654 {
 655         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 656
 657         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 658                 unsigned int cpp;
 659
 660                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 661
 662                 cluster_start = cpos % cpp;
 663                 cluster_start = cluster_start << osb->s_clustersize_bits;
 664
 665                 cluster_end = cluster_start + osb->s_clustersize;
 666         }
 667
 668         BUG_ON(cluster_start > PAGE_SIZE);
 669         BUG_ON(cluster_end > PAGE_SIZE);
 670
 671         if (start)
 672                 *start = cluster_start;
 673         if (end)
 674                 *end = cluster_end;
 675 }
 676
 677 /*
 678  * 'from' and 'to' are the region in the page to avoid zeroing.
 679  *
 680  * If pagesize > clustersize, this function will avoid zeroing outside
 681  * of the cluster boundary.
 682  *
 683  * from == to == 0 is code for "zero the entire cluster region"
 684  */
 685 static void ocfs2_clear_page_regions(struct page *page,
 686                                      struct ocfs2_super *osb, u32 cpos,
 687                                      unsigned from, unsigned to)
 688 {
 689         void *kaddr;
 690         unsigned int cluster_start, cluster_end;
 691
 692         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 693
 694         kaddr = kmap_atomic(page, KM_USER0);
 695
 696         if (from || to) {
 697                 if (from > cluster_start)
 698                         memset(kaddr + cluster_start, 0, from - cluster_start);
 699                 if (to < cluster_end)
 700                         memset(kaddr + to, 0, cluster_end - to);
 701         } else {
 702                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 703         }
 704
 705         kunmap_atomic(kaddr, KM_USER0);
 706 }
 707
 708 /*
 709  * Nonsparse file systems fully allocate before we get to the write
 710  * code. This prevents ocfs2_write() from tagging the write as an
 711  * allocating one, which means ocfs2_map_page_blocks() might try to
 712  * read-in the blocks at the tail of our file. Avoid reading them by
 713  * testing i_size against each block offset.
 714  */
 715 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 716                                  unsigned int block_start)
 717 {
 718         u64 offset = page_offset(page) + block_start;
 719
 720         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 721                 return 1;
 722
 723         if (i_size_read(inode) > offset)
 724                 return 1;
 725
 726         return 0;
 727 }
 728
 729 /*
 730  * Some of this taken from __block_write_begin(). We already have our
 731  * mapping by now though, and the entire write will be allocating or
 732  * it won't, so not much need to use BH_New.
 733  *
 734  * This will also skip zeroing, which is handled externally.
 735  */
 736 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 737                           struct inode *inode, unsigned int from,
 738                           unsigned int to, int new)
 739 {
 740         int ret = 0;
 741         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 742         unsigned int block_end, block_start;
 743         unsigned int bsize = 1 << inode->i_blkbits;
 744
 745         if (!page_has_buffers(page))
 746                 create_empty_buffers(page, bsize, 0);
 747
 748         head = page_buffers(page);
 749         for (bh = head, block_start = 0; bh != head || !block_start;
 750              bh = bh->b_this_page, block_start += bsize) {
 751                 block_end = block_start + bsize;
 752
 753                 clear_buffer_new(bh);
 754
 755                 /*
 756                  * Ignore blocks outside of our i/o range -
 757                  * they may belong to unallocated clusters.
 758                  */
 759                 if (block_start >= to || block_end <= from) {
 760                         if (PageUptodate(page))
 761                                 set_buffer_uptodate(bh);
 762                         continue;
 763                 }
 764
 765                 /*
 766                  * For an allocating write with cluster size >= page
 767                  * size, we always write the entire page.
 768                  */
 769                 if (new)
 770                         set_buffer_new(bh);
 771
 772                 if (!buffer_mapped(bh)) {
 773                         map_bh(bh, inode->i_sb, *p_blkno);
 774                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 775                 }
 776
 777                 if (PageUptodate(page)) {
 778                         if (!buffer_uptodate(bh))
 779                                 set_buffer_uptodate(bh);
 780                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 781                            !buffer_new(bh) &&
 782                            ocfs2_should_read_blk(inode, page, block_start) &&
 783                            (block_start < from || block_end > to)) {
 784                         ll_rw_block(READ, 1, &bh);
 785                         *wait_bh++=bh;
 786                 }
 787
 788                 *p_blkno = *p_blkno + 1;
 789         }
 790
 791         /*
 792          * If we issued read requests - let them complete.
 793          */
 794         while(wait_bh > wait) {
 795                 wait_on_buffer(*--wait_bh);
 796                 if (!buffer_uptodate(*wait_bh))
 797                         ret = -EIO;
 798         }
 799
 800         if (ret == 0 || !new)
 801                 return ret;
 802
 803         /*
 804          * If we get -EIO above, zero out any newly allocated blocks
 805          * to avoid exposing stale data.
 806          */
 807         bh = head;
 808         block_start = 0;
 809         do {
 810                 block_end = block_start + bsize;
 811                 if (block_end <= from)
 812                         goto next_bh;
 813                 if (block_start >= to)
 814                         break;
 815
 816                 zero_user(page, block_start, bh->b_size);
 817                 set_buffer_uptodate(bh);
 818                 mark_buffer_dirty(bh);
 819
 820 next_bh:
 821                 block_start = block_end;
 822                 bh = bh->b_this_page;
 823         } while (bh != head);
 824
 825         return ret;
 826 }
 827
 828 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 829 #define OCFS2_MAX_CTXT_PAGES    1
 830 #else
 831 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 832 #endif
 833
 834 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 835
 836 /*
 837  * Describe the state of a single cluster to be written to.
 838  */
 839 struct ocfs2_write_cluster_desc {
 840         u32             c_cpos;
 841         u32             c_phys;
 842         /*
 843          * Give this a unique field because c_phys eventually gets
 844          * filled.
 845          */
 846         unsigned        c_new;
 847         unsigned        c_unwritten;
 848         unsigned        c_needs_zero;
 849 };
 850
 851 struct ocfs2_write_ctxt {
 852         /* Logical cluster position / len of write */
 853         u32                             w_cpos;
 854         u32                             w_clen;
 855
 856         /* First cluster allocated in a nonsparse extend */
 857         u32                             w_first_new_cpos;
 858
 859         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 860
 861         /*
 862          * This is true if page_size > cluster_size.
 863          *
 864          * It triggers a set of special cases during write which might
 865          * have to deal with allocating writes to partial pages.
 866          */
 867         unsigned int                    w_large_pages;
 868
 869         /*
 870          * Pages involved in this write.
 871          *
 872          * w_target_page is the page being written to by the user.
 873          *
 874          * w_pages is an array of pages which always contains
 875          * w_target_page, and in the case of an allocating write with
 876          * page_size < cluster size, it will contain zero'd and mapped
 877          * pages adjacent to w_target_page which need to be written
 878          * out in so that future reads from that region will get
 879          * zero's.
 880          */
 881         unsigned int                    w_num_pages;
 882         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 883         struct page                     *w_target_page;
 884
 885         /*
 886          * ocfs2_write_end() uses this to know what the real range to
 887          * write in the target should be.
 888          */
 889         unsigned int                    w_target_from;
 890         unsigned int                    w_target_to;
 891
 892         /*
 893          * We could use journal_current_handle() but this is cleaner,
 894          * IMHO -Mark
 895          */
 896         handle_t                        *w_handle;
 897
 898         struct buffer_head              *w_di_bh;
 899
 900         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 901 };
 902
 903 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 904 {
 905         int i;
 906
 907         for(i = 0; i < num_pages; i++) {
 908                 if (pages[i]) {
 909                         unlock_page(pages[i]);
 910                         mark_page_accessed(pages[i]);
 911                         page_cache_release(pages[i]);
 912                 }
 913         }
 914 }
 915
 916 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 917 {
 918         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 919
 920         brelse(wc->w_di_bh);
 921         kfree(wc);
 922 }
 923
 924 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 925                                   struct ocfs2_super *osb, loff_t pos,
 926                                   unsigned len, struct buffer_head *di_bh)
 927 {
 928         u32 cend;
 929         struct ocfs2_write_ctxt *wc;
 930
 931         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 932         if (!wc)
 933                 return -ENOMEM;
 934
 935         wc->w_cpos = pos >> osb->s_clustersize_bits;
 936         wc->w_first_new_cpos = UINT_MAX;
 937         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 938         wc->w_clen = cend - wc->w_cpos + 1;
 939         get_bh(di_bh);
 940         wc->w_di_bh = di_bh;
 941
 942         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 943                 wc->w_large_pages = 1;
 944         else
 945                 wc->w_large_pages = 0;
 946
 947         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 948
 949         *wcp = wc;
 950
 951         return 0;
 952 }
 953
 954 /*
 955  * If a page has any new buffers, zero them out here, and mark them uptodate
 956  * and dirty so they'll be written out (in order to prevent uninitialised
 957  * block data from leaking). And clear the new bit.
 958  */
 959 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 960 {
 961         unsigned int block_start, block_end;
 962         struct buffer_head *head, *bh;
 963
 964         BUG_ON(!PageLocked(page));
 965         if (!page_has_buffers(page))
 966                 return;
 967
 968         bh = head = page_buffers(page);
 969         block_start = 0;
 970         do {
 971                 block_end = block_start + bh->b_size;
 972
 973                 if (buffer_new(bh)) {
 974                         if (block_end > from && block_start < to) {
 975                                 if (!PageUptodate(page)) {
 976                                         unsigned start, end;
 977
 978                                         start = max(from, block_start);
 979                                         end = min(to, block_end);
 980
 981                                         zero_user_segment(page, start, end);
 982                                         set_buffer_uptodate(bh);
 983                                 }
 984
 985                                 clear_buffer_new(bh);
 986                                 mark_buffer_dirty(bh);
 987                         }
 988                 }
 989
 990                 block_start = block_end;
 991                 bh = bh->b_this_page;
 992         } while (bh != head);
 993 }
 994
 995 /*
 996  * Only called when we have a failure during allocating write to write
 997  * zero's to the newly allocated region.
 998  */
 999 static void ocfs2_write_failure(struct inode *inode,
1000                                 struct ocfs2_write_ctxt *wc,
1001                                 loff_t user_pos, unsigned user_len)
1002 {
1003         int i;
1004         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1005                 to = user_pos + user_len;
1006         struct page *tmppage;
1007
1008         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1009
1010         for(i = 0; i < wc->w_num_pages; i++) {
1011                 tmppage = wc->w_pages[i];
1012
1013                 if (page_has_buffers(tmppage)) {
1014                         if (ocfs2_should_order_data(inode))
1015                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1016
1017                         block_commit_write(tmppage, from, to);
1018                 }
1019         }
1020 }
1021
1022 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1023                                         struct ocfs2_write_ctxt *wc,
1024                                         struct page *page, u32 cpos,
1025                                         loff_t user_pos, unsigned user_len,
1026                                         int new)
1027 {
1028         int ret;
1029         unsigned int map_from = 0, map_to = 0;
1030         unsigned int cluster_start, cluster_end;
1031         unsigned int user_data_from = 0, user_data_to = 0;
1032
1033         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1034                                         &cluster_start, &cluster_end);
1035
1036         /* treat the write as new if the a hole/lseek spanned across
1037          * the page boundary.
1038          */
1039         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1040                         (page_offset(page) <= user_pos));
1041
1042         if (page == wc->w_target_page) {
1043                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1044                 map_to = map_from + user_len;
1045
1046                 if (new)
1047                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1048                                                     cluster_start, cluster_end,
1049                                                     new);
1050                 else
1051                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1052                                                     map_from, map_to, new);
1053                 if (ret) {
1054                         mlog_errno(ret);
1055                         goto out;
1056                 }
1057
1058                 user_data_from = map_from;
1059                 user_data_to = map_to;
1060                 if (new) {
1061                         map_from = cluster_start;
1062                         map_to = cluster_end;
1063                 }
1064         } else {
1065                 /*
1066                  * If we haven't allocated the new page yet, we
1067                  * shouldn't be writing it out without copying user
1068                  * data. This is likely a math error from the caller.
1069                  */
1070                 BUG_ON(!new);
1071
1072                 map_from = cluster_start;
1073                 map_to = cluster_end;
1074
1075                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1076                                             cluster_start, cluster_end, new);
1077                 if (ret) {
1078                         mlog_errno(ret);
1079                         goto out;
1080                 }
1081         }
1082
1083         /*
1084          * Parts of newly allocated pages need to be zero'd.
1085          *
1086          * Above, we have also rewritten 'to' and 'from' - as far as
1087          * the rest of the function is concerned, the entire cluster
1088          * range inside of a page needs to be written.
1089          *
1090          * We can skip this if the page is up to date - it's already
1091          * been zero'd from being read in as a hole.
1092          */
1093         if (new && !PageUptodate(page))
1094                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1095                                          cpos, user_data_from, user_data_to);
1096
1097         flush_dcache_page(page);
1098
1099 out:
1100         return ret;
1101 }
1102
1103 /*
1104  * This function will only grab one clusters worth of pages.
1105  */
1106 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1107                                       struct ocfs2_write_ctxt *wc,
1108                                       u32 cpos, loff_t user_pos,
1109                                       unsigned user_len, int new,
1110                                       struct page *mmap_page)
1111 {
1112         int ret = 0, i;
1113         unsigned long start, target_index, end_index, index;
1114         struct inode *inode = mapping->host;
1115         loff_t last_byte;
1116
1117         target_index = user_pos >> PAGE_CACHE_SHIFT;
1118
1119         /*
1120          * Figure out how many pages we'll be manipulating here. For
1121          * non allocating write, we just change the one
1122          * page. Otherwise, we'll need a whole clusters worth.  If we're
1123          * writing past i_size, we only need enough pages to cover the
1124          * last page of the write.
1125          */
1126         if (new) {
1127                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1128                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1129                 /*
1130                  * We need the index *past* the last page we could possibly
1131                  * touch.  This is the page past the end of the write or
1132                  * i_size, whichever is greater.
1133                  */
1134                 last_byte = max(user_pos + user_len, i_size_read(inode));
1135                 BUG_ON(last_byte < 1);
1136                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1137                 if ((start + wc->w_num_pages) > end_index)
1138                         wc->w_num_pages = end_index - start;
1139         } else {
1140                 wc->w_num_pages = 1;
1141                 start = target_index;
1142         }
1143
1144         for(i = 0; i < wc->w_num_pages; i++) {
1145                 index = start + i;
1146
1147                 if (index == target_index && mmap_page) {
1148                         /*
1149                          * ocfs2_pagemkwrite() is a little different
1150                          * and wants us to directly use the page
1151                          * passed in.
1152                          */
1153                         lock_page(mmap_page);
1154
1155                         if (mmap_page->mapping != mapping) {
1156                                 unlock_page(mmap_page);
1157                                 /*
1158                                  * Sanity check - the locking in
1159                                  * ocfs2_pagemkwrite() should ensure
1160                                  * that this code doesn't trigger.
1161                                  */
1162                                 ret = -EINVAL;
1163                                 mlog_errno(ret);
1164                                 goto out;
1165                         }
1166
1167                         page_cache_get(mmap_page);
1168                         wc->w_pages[i] = mmap_page;
1169                 } else {
1170                         wc->w_pages[i] = find_or_create_page(mapping, index,
1171                                                              GFP_NOFS);
1172                         if (!wc->w_pages[i]) {
1173                                 ret = -ENOMEM;
1174                                 mlog_errno(ret);
1175                                 goto out;
1176                         }
1177                 }
1178
1179                 if (index == target_index)
1180                         wc->w_target_page = wc->w_pages[i];
1181         }
1182 out:
1183         return ret;
1184 }
1185
1186 /*
1187  * Prepare a single cluster for write one cluster into the file.
1188  */
1189 static int ocfs2_write_cluster(struct address_space *mapping,
1190                                u32 phys, unsigned int unwritten,
1191                                unsigned int should_zero,
1192                                struct ocfs2_alloc_context *data_ac,
1193                                struct ocfs2_alloc_context *meta_ac,
1194                                struct ocfs2_write_ctxt *wc, u32 cpos,
1195                                loff_t user_pos, unsigned user_len)
1196 {
1197         int ret, i, new;
1198         u64 v_blkno, p_blkno;
1199         struct inode *inode = mapping->host;
1200         struct ocfs2_extent_tree et;
1201
1202         new = phys == 0 ? 1 : 0;
1203         if (new) {
1204                 u32 tmp_pos;
1205
1206                 /*
1207                  * This is safe to call with the page locks - it won't take
1208                  * any additional semaphores or cluster locks.
1209                  */
1210                 tmp_pos = cpos;
1211                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1212                                            &tmp_pos, 1, 0, wc->w_di_bh,
1213                                            wc->w_handle, data_ac,
1214                                            meta_ac, NULL);
1215                 /*
1216                  * This shouldn't happen because we must have already
1217                  * calculated the correct meta data allocation required. The
1218                  * internal tree allocation code should know how to increase
1219                  * transaction credits itself.
1220                  *
1221                  * If need be, we could handle -EAGAIN for a
1222                  * RESTART_TRANS here.
1223                  */
1224                 mlog_bug_on_msg(ret == -EAGAIN,
1225                                 "Inode %llu: EAGAIN return during allocation.\n",
1226                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1227                 if (ret < 0) {
1228                         mlog_errno(ret);
1229                         goto out;
1230                 }
1231         } else if (unwritten) {
1232                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1233                                               wc->w_di_bh);
1234                 ret = ocfs2_mark_extent_written(inode, &et,
1235                                                 wc->w_handle, cpos, 1, phys,
1236                                                 meta_ac, &wc->w_dealloc);
1237                 if (ret < 0) {
1238                         mlog_errno(ret);
1239                         goto out;
1240                 }
1241         }
1242
1243         if (should_zero)
1244                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1245         else
1246                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1247
1248         /*
1249          * The only reason this should fail is due to an inability to
1250          * find the extent added.
1251          */
1252         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1253                                           NULL);
1254         if (ret < 0) {
1255                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1256                             "at logical block %llu",
1257                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1258                             (unsigned long long)v_blkno);
1259                 goto out;
1260         }
1261
1262         BUG_ON(p_blkno == 0);
1263
1264         for(i = 0; i < wc->w_num_pages; i++) {
1265                 int tmpret;
1266
1267                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1268                                                       wc->w_pages[i], cpos,
1269                                                       user_pos, user_len,
1270                                                       should_zero);
1271                 if (tmpret) {
1272                         mlog_errno(tmpret);
1273                         if (ret == 0)
1274                                 ret = tmpret;
1275                 }
1276         }
1277
1278         /*
1279          * We only have cleanup to do in case of allocating write.
1280          */
1281         if (ret && new)
1282                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1283
1284 out:
1285
1286         return ret;
1287 }
1288
1289 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1290                                        struct ocfs2_alloc_context *data_ac,
1291                                        struct ocfs2_alloc_context *meta_ac,
1292                                        struct ocfs2_write_ctxt *wc,
1293                                        loff_t pos, unsigned len)
1294 {
1295         int ret, i;
1296         loff_t cluster_off;
1297         unsigned int local_len = len;
1298         struct ocfs2_write_cluster_desc *desc;
1299         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1300
1301         for (i = 0; i < wc->w_clen; i++) {
1302                 desc = &wc->w_desc[i];
1303
1304                 /*
1305                  * We have to make sure that the total write passed in
1306                  * doesn't extend past a single cluster.
1307                  */
1308                 local_len = len;
1309                 cluster_off = pos & (osb->s_clustersize - 1);
1310                 if ((cluster_off + local_len) > osb->s_clustersize)
1311                         local_len = osb->s_clustersize - cluster_off;
1312
1313                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1314                                           desc->c_unwritten,
1315                                           desc->c_needs_zero,
1316                                           data_ac, meta_ac,
1317                                           wc, desc->c_cpos, pos, local_len);
1318                 if (ret) {
1319                         mlog_errno(ret);
1320                         goto out;
1321                 }
1322
1323                 len -= local_len;
1324                 pos += local_len;
1325         }
1326
1327         ret = 0;
1328 out:
1329         return ret;
1330 }
1331
1332 /*
1333  * ocfs2_write_end() wants to know which parts of the target page it
1334  * should complete the write on. It's easiest to compute them ahead of
1335  * time when a more complete view of the write is available.
1336  */
1337 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1338                                         struct ocfs2_write_ctxt *wc,
1339                                         loff_t pos, unsigned len, int alloc)
1340 {
1341         struct ocfs2_write_cluster_desc *desc;
1342
1343         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1344         wc->w_target_to = wc->w_target_from + len;
1345
1346         if (alloc == 0)
1347                 return;
1348
1349         /*
1350          * Allocating write - we may have different boundaries based
1351          * on page size and cluster size.
1352          *
1353          * NOTE: We can no longer compute one value from the other as
1354          * the actual write length and user provided length may be
1355          * different.
1356          */
1357
1358         if (wc->w_large_pages) {
1359                 /*
1360                  * We only care about the 1st and last cluster within
1361                  * our range and whether they should be zero'd or not. Either
1362                  * value may be extended out to the start/end of a
1363                  * newly allocated cluster.
1364                  */
1365                 desc = &wc->w_desc[0];
1366                 if (desc->c_needs_zero)
1367                         ocfs2_figure_cluster_boundaries(osb,
1368                                                         desc->c_cpos,
1369                                                         &wc->w_target_from,
1370                                                         NULL);
1371
1372                 desc = &wc->w_desc[wc->w_clen - 1];
1373                 if (desc->c_needs_zero)
1374                         ocfs2_figure_cluster_boundaries(osb,
1375                                                         desc->c_cpos,
1376                                                         NULL,
1377                                                         &wc->w_target_to);
1378         } else {
1379                 wc->w_target_from = 0;
1380                 wc->w_target_to = PAGE_CACHE_SIZE;
1381         }
1382 }
1383
1384 /*
1385  * Populate each single-cluster write descriptor in the write context
1386  * with information about the i/o to be done.
1387  *
1388  * Returns the number of clusters that will have to be allocated, as
1389  * well as a worst case estimate of the number of extent records that
1390  * would have to be created during a write to an unwritten region.
1391  */
1392 static int ocfs2_populate_write_desc(struct inode *inode,
1393                                      struct ocfs2_write_ctxt *wc,
1394                                      unsigned int *clusters_to_alloc,
1395                                      unsigned int *extents_to_split)
1396 {
1397         int ret;
1398         struct ocfs2_write_cluster_desc *desc;
1399         unsigned int num_clusters = 0;
1400         unsigned int ext_flags = 0;
1401         u32 phys = 0;
1402         int i;
1403
1404         *clusters_to_alloc = 0;
1405         *extents_to_split = 0;
1406
1407         for (i = 0; i < wc->w_clen; i++) {
1408                 desc = &wc->w_desc[i];
1409                 desc->c_cpos = wc->w_cpos + i;
1410
1411                 if (num_clusters == 0) {
1412                         /*
1413                          * Need to look up the next extent record.
1414                          */
1415                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1416                                                  &num_clusters, &ext_flags);
1417                         if (ret) {
1418                                 mlog_errno(ret);
1419                                 goto out;
1420                         }
1421
1422                         /* We should already CoW the refcountd extent. */
1423                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1424
1425                         /*
1426                          * Assume worst case - that we're writing in
1427                          * the middle of the extent.
1428                          *
1429                          * We can assume that the write proceeds from
1430                          * left to right, in which case the extent
1431                          * insert code is smart enough to coalesce the
1432                          * next splits into the previous records created.
1433                          */
1434                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1435                                 *extents_to_split = *extents_to_split + 2;
1436                 } else if (phys) {
1437                         /*
1438                          * Only increment phys if it doesn't describe
1439                          * a hole.
1440                          */
1441                         phys++;
1442                 }
1443
1444                 /*
1445                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1446                  * file that got extended.  w_first_new_cpos tells us
1447                  * where the newly allocated clusters are so we can
1448                  * zero them.
1449                  */
1450                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1451                         BUG_ON(phys == 0);
1452                         desc->c_needs_zero = 1;
1453                 }
1454
1455                 desc->c_phys = phys;
1456                 if (phys == 0) {
1457                         desc->c_new = 1;
1458                         desc->c_needs_zero = 1;
1459                         *clusters_to_alloc = *clusters_to_alloc + 1;
1460                 }
1461
1462                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1463                         desc->c_unwritten = 1;
1464                         desc->c_needs_zero = 1;
1465                 }
1466
1467                 num_clusters--;
1468         }
1469
1470         ret = 0;
1471 out:
1472         return ret;
1473 }
1474
1475 static int ocfs2_write_begin_inline(struct address_space *mapping,
1476                                     struct inode *inode,
1477                                     struct ocfs2_write_ctxt *wc)
1478 {
1479         int ret;
1480         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1481         struct page *page;
1482         handle_t *handle;
1483         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1484
1485         page = find_or_create_page(mapping, 0, GFP_NOFS);
1486         if (!page) {
1487                 ret = -ENOMEM;
1488                 mlog_errno(ret);
1489                 goto out;
1490         }
1491         /*
1492          * If we don't set w_num_pages then this page won't get unlocked
1493          * and freed on cleanup of the write context.
1494          */
1495         wc->w_pages[0] = wc->w_target_page = page;
1496         wc->w_num_pages = 1;
1497
1498         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1499         if (IS_ERR(handle)) {
1500                 ret = PTR_ERR(handle);
1501                 mlog_errno(ret);
1502                 goto out;
1503         }
1504
1505         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1506                                       OCFS2_JOURNAL_ACCESS_WRITE);
1507         if (ret) {
1508                 ocfs2_commit_trans(osb, handle);
1509
1510                 mlog_errno(ret);
1511                 goto out;
1512         }
1513
1514         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1515                 ocfs2_set_inode_data_inline(inode, di);
1516
1517         if (!PageUptodate(page)) {
1518                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1519                 if (ret) {
1520                         ocfs2_commit_trans(osb, handle);
1521
1522                         goto out;
1523                 }
1524         }
1525
1526         wc->w_handle = handle;
1527 out:
1528         return ret;
1529 }
1530
1531 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1532 {
1533         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1534
1535         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1536                 return 1;
1537         return 0;
1538 }
1539
1540 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1541                                           struct inode *inode, loff_t pos,
1542                                           unsigned len, struct page *mmap_page,
1543                                           struct ocfs2_write_ctxt *wc)
1544 {
1545         int ret, written = 0;
1546         loff_t end = pos + len;
1547         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1548         struct ocfs2_dinode *di = NULL;
1549
1550         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1551                                              len, (unsigned long long)pos,
1552                                              oi->ip_dyn_features);
1553
1554         /*
1555          * Handle inodes which already have inline data 1st.
1556          */
1557         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1558                 if (mmap_page == NULL &&
1559                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1560                         goto do_inline_write;
1561
1562                 /*
1563                  * The write won't fit - we have to give this inode an
1564                  * inline extent list now.
1565                  */
1566                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1567                 if (ret)
1568                         mlog_errno(ret);
1569                 goto out;
1570         }
1571
1572         /*
1573          * Check whether the inode can accept inline data.
1574          */
1575         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1576                 return 0;
1577
1578         /*
1579          * Check whether the write can fit.
1580          */
1581         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1582         if (mmap_page ||
1583             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1584                 return 0;
1585
1586 do_inline_write:
1587         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1588         if (ret) {
1589                 mlog_errno(ret);
1590                 goto out;
1591         }
1592
1593         /*
1594          * This signals to the caller that the data can be written
1595          * inline.
1596          */
1597         written = 1;
1598 out:
1599         return written ? written : ret;
1600 }
1601
1602 /*
1603  * This function only does anything for file systems which can't
1604  * handle sparse files.
1605  *
1606  * What we want to do here is fill in any hole between the current end
1607  * of allocation and the end of our write. That way the rest of the
1608  * write path can treat it as an non-allocating write, which has no
1609  * special case code for sparse/nonsparse files.
1610  */
1611 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1612                                         struct buffer_head *di_bh,
1613                                         loff_t pos, unsigned len,
1614                                         struct ocfs2_write_ctxt *wc)
1615 {
1616         int ret;
1617         loff_t newsize = pos + len;
1618
1619         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1620
1621         if (newsize <= i_size_read(inode))
1622                 return 0;
1623
1624         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1625         if (ret)
1626                 mlog_errno(ret);
1627
1628         wc->w_first_new_cpos =
1629                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1630
1631         return ret;
1632 }
1633
1634 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1635                            loff_t pos)
1636 {
1637         int ret = 0;
1638
1639         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1640         if (pos > i_size_read(inode))
1641                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1642
1643         return ret;
1644 }
1645
1646 /*
1647  * Try to flush truncate logs if we can free enough clusters from it.
1648  * As for return value, "< 0" means error, "0" no space and "1" means
1649  * we have freed enough spaces and let the caller try to allocate again.
1650  */
1651 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1652                                           unsigned int needed)
1653 {
1654         tid_t target;
1655         int ret = 0;
1656         unsigned int truncated_clusters;
1657
1658         mutex_lock(&osb->osb_tl_inode->i_mutex);
1659         truncated_clusters = osb->truncated_clusters;
1660         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1661
1662         /*
1663          * Check whether we can succeed in allocating if we free
1664          * the truncate log.
1665          */
1666         if (truncated_clusters < needed)
1667                 goto out;
1668
1669         ret = ocfs2_flush_truncate_log(osb);
1670         if (ret) {
1671                 mlog_errno(ret);
1672                 goto out;
1673         }
1674
1675         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1676                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1677                 ret = 1;
1678         }
1679 out:
1680         return ret;
1681 }
1682
1683 int ocfs2_write_begin_nolock(struct file *filp,
1684                              struct address_space *mapping,
1685                              loff_t pos, unsigned len, unsigned flags,
1686                              struct page **pagep, void **fsdata,
1687                              struct buffer_head *di_bh, struct page *mmap_page)
1688 {
1689         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1690         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1691         struct ocfs2_write_ctxt *wc;
1692         struct inode *inode = mapping->host;
1693         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1694         struct ocfs2_dinode *di;
1695         struct ocfs2_alloc_context *data_ac = NULL;
1696         struct ocfs2_alloc_context *meta_ac = NULL;
1697         handle_t *handle;
1698         struct ocfs2_extent_tree et;
1699         int try_free = 1, ret1;
1700
1701 try_again:
1702         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1703         if (ret) {
1704                 mlog_errno(ret);
1705                 return ret;
1706         }
1707
1708         if (ocfs2_supports_inline_data(osb)) {
1709                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1710                                                      mmap_page, wc);
1711                 if (ret == 1) {
1712                         ret = 0;
1713                         goto success;
1714                 }
1715                 if (ret < 0) {
1716                         mlog_errno(ret);
1717                         goto out;
1718                 }
1719         }
1720
1721         if (ocfs2_sparse_alloc(osb))
1722                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1723         else
1724                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1725                                                    wc);
1726         if (ret) {
1727                 mlog_errno(ret);
1728                 goto out;
1729         }
1730
1731         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1732         if (ret < 0) {
1733                 mlog_errno(ret);
1734                 goto out;
1735         } else if (ret == 1) {
1736                 clusters_need = wc->w_clen;
1737                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1738                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1739                 if (ret) {
1740                         mlog_errno(ret);
1741                         goto out;
1742                 }
1743         }
1744
1745         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1746                                         &extents_to_split);
1747         if (ret) {
1748                 mlog_errno(ret);
1749                 goto out;
1750         }
1751         clusters_need += clusters_to_alloc;
1752
1753         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1754
1755         trace_ocfs2_write_begin_nolock(
1756                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1757                         (long long)i_size_read(inode),
1758                         le32_to_cpu(di->i_clusters),
1759                         pos, len, flags, mmap_page,
1760                         clusters_to_alloc, extents_to_split);
1761
1762         /*
1763          * We set w_target_from, w_target_to here so that
1764          * ocfs2_write_end() knows which range in the target page to
1765          * write out. An allocation requires that we write the entire
1766          * cluster range.
1767          */
1768         if (clusters_to_alloc || extents_to_split) {
1769                 /*
1770                  * XXX: We are stretching the limits of
1771                  * ocfs2_lock_allocators(). It greatly over-estimates
1772                  * the work to be done.
1773                  */
1774                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1775                                               wc->w_di_bh);
1776                 ret = ocfs2_lock_allocators(inode, &et,
1777                                             clusters_to_alloc, extents_to_split,
1778                                             &data_ac, &meta_ac);
1779                 if (ret) {
1780                         mlog_errno(ret);
1781                         goto out;
1782                 }
1783
1784                 if (data_ac)
1785                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1786
1787                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1788                                                     &di->id2.i_list,
1789                                                     clusters_to_alloc);
1790
1791         }
1792
1793         /*
1794          * We have to zero sparse allocated clusters, unwritten extent clusters,
1795          * and non-sparse clusters we just extended.  For non-sparse writes,
1796          * we know zeros will only be needed in the first and/or last cluster.
1797          */
1798         if (clusters_to_alloc || extents_to_split ||
1799             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1800                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1801                 cluster_of_pages = 1;
1802         else
1803                 cluster_of_pages = 0;
1804
1805         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1806
1807         handle = ocfs2_start_trans(osb, credits);
1808         if (IS_ERR(handle)) {
1809                 ret = PTR_ERR(handle);
1810                 mlog_errno(ret);
1811                 goto out;
1812         }
1813
1814         wc->w_handle = handle;
1815
1816         if (clusters_to_alloc) {
1817                 ret = dquot_alloc_space_nodirty(inode,
1818                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1819                 if (ret)
1820                         goto out_commit;
1821         }
1822         /*
1823          * We don't want this to fail in ocfs2_write_end(), so do it
1824          * here.
1825          */
1826         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1827                                       OCFS2_JOURNAL_ACCESS_WRITE);
1828         if (ret) {
1829                 mlog_errno(ret);
1830                 goto out_quota;
1831         }
1832
1833         /*
1834          * Fill our page array first. That way we've grabbed enough so
1835          * that we can zero and flush if we error after adding the
1836          * extent.
1837          */
1838         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1839                                          cluster_of_pages, mmap_page);
1840         if (ret) {
1841                 mlog_errno(ret);
1842                 goto out_quota;
1843         }
1844
1845         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1846                                           len);
1847         if (ret) {
1848                 mlog_errno(ret);
1849                 goto out_quota;
1850         }
1851
1852         if (data_ac)
1853                 ocfs2_free_alloc_context(data_ac);
1854         if (meta_ac)
1855                 ocfs2_free_alloc_context(meta_ac);
1856
1857 success:
1858         *pagep = wc->w_target_page;
1859         *fsdata = wc;
1860         return 0;
1861 out_quota:
1862         if (clusters_to_alloc)
1863                 dquot_free_space(inode,
1864                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1865 out_commit:
1866         ocfs2_commit_trans(osb, handle);
1867
1868 out:
1869         ocfs2_free_write_ctxt(wc);
1870
1871         if (data_ac)
1872                 ocfs2_free_alloc_context(data_ac);
1873         if (meta_ac)
1874                 ocfs2_free_alloc_context(meta_ac);
1875
1876         if (ret == -ENOSPC && try_free) {
1877                 /*
1878                  * Try to free some truncate log so that we can have enough
1879                  * clusters to allocate.
1880                  */
1881                 try_free = 0;
1882
1883                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1884                 if (ret1 == 1)
1885                         goto try_again;
1886
1887                 if (ret1 < 0)
1888                         mlog_errno(ret1);
1889         }
1890
1891         return ret;
1892 }
1893
1894 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1895                              loff_t pos, unsigned len, unsigned flags,
1896                              struct page **pagep, void **fsdata)
1897 {
1898         int ret;
1899         struct buffer_head *di_bh = NULL;
1900         struct inode *inode = mapping->host;
1901
1902         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1903         if (ret) {
1904                 mlog_errno(ret);
1905                 return ret;
1906         }
1907
1908         /*
1909          * Take alloc sem here to prevent concurrent lookups. That way
1910          * the mapping, zeroing and tree manipulation within
1911          * ocfs2_write() will be safe against ->readpage(). This
1912          * should also serve to lock out allocation from a shared
1913          * writeable region.
1914          */
1915         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1916
1917         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1918                                        fsdata, di_bh, NULL);
1919         if (ret) {
1920                 mlog_errno(ret);
1921                 goto out_fail;
1922         }
1923
1924         brelse(di_bh);
1925
1926         return 0;
1927
1928 out_fail:
1929         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1930
1931         brelse(di_bh);
1932         ocfs2_inode_unlock(inode, 1);
1933
1934         return ret;
1935 }
1936
1937 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1938                                    unsigned len, unsigned *copied,
1939                                    struct ocfs2_dinode *di,
1940                                    struct ocfs2_write_ctxt *wc)
1941 {
1942         void *kaddr;
1943
1944         if (unlikely(*copied < len)) {
1945                 if (!PageUptodate(wc->w_target_page)) {
1946                         *copied = 0;
1947                         return;
1948                 }
1949         }
1950
1951         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1952         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1953         kunmap_atomic(kaddr, KM_USER0);
1954
1955         trace_ocfs2_write_end_inline(
1956              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1957              (unsigned long long)pos, *copied,
1958              le16_to_cpu(di->id2.i_data.id_count),
1959              le16_to_cpu(di->i_dyn_features));
1960 }
1961
1962 int ocfs2_write_end_nolock(struct address_space *mapping,
1963                            loff_t pos, unsigned len, unsigned copied,
1964                            struct page *page, void *fsdata)
1965 {
1966         int i;
1967         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1968         struct inode *inode = mapping->host;
1969         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1970         struct ocfs2_write_ctxt *wc = fsdata;
1971         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1972         handle_t *handle = wc->w_handle;
1973         struct page *tmppage;
1974
1975         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1976                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1977                 goto out_write_size;
1978         }
1979
1980         if (unlikely(copied < len)) {
1981                 if (!PageUptodate(wc->w_target_page))
1982                         copied = 0;
1983
1984                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1985                                        start+len);
1986         }
1987         flush_dcache_page(wc->w_target_page);
1988
1989         for(i = 0; i < wc->w_num_pages; i++) {
1990                 tmppage = wc->w_pages[i];
1991
1992                 if (tmppage == wc->w_target_page) {
1993                         from = wc->w_target_from;
1994                         to = wc->w_target_to;
1995
1996                         BUG_ON(from > PAGE_CACHE_SIZE ||
1997                                to > PAGE_CACHE_SIZE ||
1998                                to < from);
1999                 } else {
2000                         /*
2001                          * Pages adjacent to the target (if any) imply
2002                          * a hole-filling write in which case we want
2003                          * to flush their entire range.
2004                          */
2005                         from = 0;
2006                         to = PAGE_CACHE_SIZE;
2007                 }
2008
2009                 if (page_has_buffers(tmppage)) {
2010                         if (ocfs2_should_order_data(inode))
2011                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2012                         block_commit_write(tmppage, from, to);
2013                 }
2014         }
2015
2016 out_write_size:
2017         pos += copied;
2018         if (pos > inode->i_size) {
2019                 i_size_write(inode, pos);
2020                 mark_inode_dirty(inode);
2021         }
2022         inode->i_blocks = ocfs2_inode_sector_count(inode);
2023         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2024         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2025         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2026         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2027         ocfs2_journal_dirty(handle, wc->w_di_bh);
2028
2029         ocfs2_commit_trans(osb, handle);
2030
2031         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2032
2033         ocfs2_free_write_ctxt(wc);
2034
2035         return copied;
2036 }
2037
2038 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2039                            loff_t pos, unsigned len, unsigned copied,
2040                            struct page *page, void *fsdata)
2041 {
2042         int ret;
2043         struct inode *inode = mapping->host;
2044
2045         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2046
2047         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2048         ocfs2_inode_unlock(inode, 1);
2049
2050         return ret;
2051 }
2052
2053 const struct address_space_operations ocfs2_aops = {
2054         .readpage               = ocfs2_readpage,
2055         .readpages              = ocfs2_readpages,
2056         .writepage              = ocfs2_writepage,
2057         .write_begin            = ocfs2_write_begin,
2058         .write_end              = ocfs2_write_end,
2059         .bmap                   = ocfs2_bmap,
2060         .direct_IO              = ocfs2_direct_IO,
2061         .invalidatepage         = ocfs2_invalidatepage,
2062         .releasepage            = ocfs2_releasepage,
2063         .migratepage            = buffer_migrate_page,
2064         .is_partially_uptodate  = block_is_partially_uptodate,
2065         .error_remove_page      = generic_error_remove_page,
2066 };