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                                const struct iovec *iov,
 620                                loff_t offset,
 621                                unsigned long nr_segs)
 622 {
 623         struct file *file = iocb->ki_filp;
 624         struct inode *inode = file_inode(file)->i_mapping->host;
 625
 626         /*
 627          * Fallback to buffered I/O if we see an inode without
 628          * extents.
 629          */
 630         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 631                 return 0;
 632
 633         /* Fallback to buffered I/O if we are appending. */
 634         if (i_size_read(inode) <= offset)
 635                 return 0;
 636
 637         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 638                                     iov, offset, nr_segs,
 639                                     ocfs2_direct_IO_get_blocks,
 640                                     ocfs2_dio_end_io, NULL, 0);
 641 }
 642
 643 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 644                                             u32 cpos,
 645                                             unsigned int *start,
 646                                             unsigned int *end)
 647 {
 648         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 649
 650         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 651                 unsigned int cpp;
 652
 653                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 654
 655                 cluster_start = cpos % cpp;
 656                 cluster_start = cluster_start << osb->s_clustersize_bits;
 657
 658                 cluster_end = cluster_start + osb->s_clustersize;
 659         }
 660
 661         BUG_ON(cluster_start > PAGE_SIZE);
 662         BUG_ON(cluster_end > PAGE_SIZE);
 663
 664         if (start)
 665                 *start = cluster_start;
 666         if (end)
 667                 *end = cluster_end;
 668 }
 669
 670 /*
 671  * 'from' and 'to' are the region in the page to avoid zeroing.
 672  *
 673  * If pagesize > clustersize, this function will avoid zeroing outside
 674  * of the cluster boundary.
 675  *
 676  * from == to == 0 is code for "zero the entire cluster region"
 677  */
 678 static void ocfs2_clear_page_regions(struct page *page,
 679                                      struct ocfs2_super *osb, u32 cpos,
 680                                      unsigned from, unsigned to)
 681 {
 682         void *kaddr;
 683         unsigned int cluster_start, cluster_end;
 684
 685         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 686
 687         kaddr = kmap_atomic(page);
 688
 689         if (from || to) {
 690                 if (from > cluster_start)
 691                         memset(kaddr + cluster_start, 0, from - cluster_start);
 692                 if (to < cluster_end)
 693                         memset(kaddr + to, 0, cluster_end - to);
 694         } else {
 695                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 696         }
 697
 698         kunmap_atomic(kaddr);
 699 }
 700
 701 /*
 702  * Nonsparse file systems fully allocate before we get to the write
 703  * code. This prevents ocfs2_write() from tagging the write as an
 704  * allocating one, which means ocfs2_map_page_blocks() might try to
 705  * read-in the blocks at the tail of our file. Avoid reading them by
 706  * testing i_size against each block offset.
 707  */
 708 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 709                                  unsigned int block_start)
 710 {
 711         u64 offset = page_offset(page) + block_start;
 712
 713         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 714                 return 1;
 715
 716         if (i_size_read(inode) > offset)
 717                 return 1;
 718
 719         return 0;
 720 }
 721
 722 /*
 723  * Some of this taken from __block_write_begin(). We already have our
 724  * mapping by now though, and the entire write will be allocating or
 725  * it won't, so not much need to use BH_New.
 726  *
 727  * This will also skip zeroing, which is handled externally.
 728  */
 729 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 730                           struct inode *inode, unsigned int from,
 731                           unsigned int to, int new)
 732 {
 733         int ret = 0;
 734         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 735         unsigned int block_end, block_start;
 736         unsigned int bsize = 1 << inode->i_blkbits;
 737
 738         if (!page_has_buffers(page))
 739                 create_empty_buffers(page, bsize, 0);
 740
 741         head = page_buffers(page);
 742         for (bh = head, block_start = 0; bh != head || !block_start;
 743              bh = bh->b_this_page, block_start += bsize) {
 744                 block_end = block_start + bsize;
 745
 746                 clear_buffer_new(bh);
 747
 748                 /*
 749                  * Ignore blocks outside of our i/o range -
 750                  * they may belong to unallocated clusters.
 751                  */
 752                 if (block_start >= to || block_end <= from) {
 753                         if (PageUptodate(page))
 754                                 set_buffer_uptodate(bh);
 755                         continue;
 756                 }
 757
 758                 /*
 759                  * For an allocating write with cluster size >= page
 760                  * size, we always write the entire page.
 761                  */
 762                 if (new)
 763                         set_buffer_new(bh);
 764
 765                 if (!buffer_mapped(bh)) {
 766                         map_bh(bh, inode->i_sb, *p_blkno);
 767                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 768                 }
 769
 770                 if (PageUptodate(page)) {
 771                         if (!buffer_uptodate(bh))
 772                                 set_buffer_uptodate(bh);
 773                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 774                            !buffer_new(bh) &&
 775                            ocfs2_should_read_blk(inode, page, block_start) &&
 776                            (block_start < from || block_end > to)) {
 777                         ll_rw_block(READ, 1, &bh);
 778                         *wait_bh++=bh;
 779                 }
 780
 781                 *p_blkno = *p_blkno + 1;
 782         }
 783
 784         /*
 785          * If we issued read requests - let them complete.
 786          */
 787         while(wait_bh > wait) {
 788                 wait_on_buffer(*--wait_bh);
 789                 if (!buffer_uptodate(*wait_bh))
 790                         ret = -EIO;
 791         }
 792
 793         if (ret == 0 || !new)
 794                 return ret;
 795
 796         /*
 797          * If we get -EIO above, zero out any newly allocated blocks
 798          * to avoid exposing stale data.
 799          */
 800         bh = head;
 801         block_start = 0;
 802         do {
 803                 block_end = block_start + bsize;
 804                 if (block_end <= from)
 805                         goto next_bh;
 806                 if (block_start >= to)
 807                         break;
 808
 809                 zero_user(page, block_start, bh->b_size);
 810                 set_buffer_uptodate(bh);
 811                 mark_buffer_dirty(bh);
 812
 813 next_bh:
 814                 block_start = block_end;
 815                 bh = bh->b_this_page;
 816         } while (bh != head);
 817
 818         return ret;
 819 }
 820
 821 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 822 #define OCFS2_MAX_CTXT_PAGES    1
 823 #else
 824 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 825 #endif
 826
 827 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 828
 829 /*
 830  * Describe the state of a single cluster to be written to.
 831  */
 832 struct ocfs2_write_cluster_desc {
 833         u32             c_cpos;
 834         u32             c_phys;
 835         /*
 836          * Give this a unique field because c_phys eventually gets
 837          * filled.
 838          */
 839         unsigned        c_new;
 840         unsigned        c_unwritten;
 841         unsigned        c_needs_zero;
 842 };
 843
 844 struct ocfs2_write_ctxt {
 845         /* Logical cluster position / len of write */
 846         u32                             w_cpos;
 847         u32                             w_clen;
 848
 849         /* First cluster allocated in a nonsparse extend */
 850         u32                             w_first_new_cpos;
 851
 852         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 853
 854         /*
 855          * This is true if page_size > cluster_size.
 856          *
 857          * It triggers a set of special cases during write which might
 858          * have to deal with allocating writes to partial pages.
 859          */
 860         unsigned int                    w_large_pages;
 861
 862         /*
 863          * Pages involved in this write.
 864          *
 865          * w_target_page is the page being written to by the user.
 866          *
 867          * w_pages is an array of pages which always contains
 868          * w_target_page, and in the case of an allocating write with
 869          * page_size < cluster size, it will contain zero'd and mapped
 870          * pages adjacent to w_target_page which need to be written
 871          * out in so that future reads from that region will get
 872          * zero's.
 873          */
 874         unsigned int                    w_num_pages;
 875         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 876         struct page                     *w_target_page;
 877
 878         /*
 879          * w_target_locked is used for page_mkwrite path indicating no unlocking
 880          * against w_target_page in ocfs2_write_end_nolock.
 881          */
 882         unsigned int                    w_target_locked:1;
 883
 884         /*
 885          * ocfs2_write_end() uses this to know what the real range to
 886          * write in the target should be.
 887          */
 888         unsigned int                    w_target_from;
 889         unsigned int                    w_target_to;
 890
 891         /*
 892          * We could use journal_current_handle() but this is cleaner,
 893          * IMHO -Mark
 894          */
 895         handle_t                        *w_handle;
 896
 897         struct buffer_head              *w_di_bh;
 898
 899         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 900 };
 901
 902 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 903 {
 904         int i;
 905
 906         for(i = 0; i < num_pages; i++) {
 907                 if (pages[i]) {
 908                         unlock_page(pages[i]);
 909                         mark_page_accessed(pages[i]);
 910                         page_cache_release(pages[i]);
 911                 }
 912         }
 913 }
 914
 915 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 916 {
 917         int i;
 918
 919         /*
 920          * w_target_locked is only set to true in the page_mkwrite() case.
 921          * The intent is to allow us to lock the target page from write_begin()
 922          * to write_end(). The caller must hold a ref on w_target_page.
 923          */
 924         if (wc->w_target_locked) {
 925                 BUG_ON(!wc->w_target_page);
 926                 for (i = 0; i < wc->w_num_pages; i++) {
 927                         if (wc->w_target_page == wc->w_pages[i]) {
 928                                 wc->w_pages[i] = NULL;
 929                                 break;
 930                         }
 931                 }
 932                 mark_page_accessed(wc->w_target_page);
 933                 page_cache_release(wc->w_target_page);
 934         }
 935         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 936
 937         brelse(wc->w_di_bh);
 938         kfree(wc);
 939 }
 940
 941 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 942                                   struct ocfs2_super *osb, loff_t pos,
 943                                   unsigned len, struct buffer_head *di_bh)
 944 {
 945         u32 cend;
 946         struct ocfs2_write_ctxt *wc;
 947
 948         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 949         if (!wc)
 950                 return -ENOMEM;
 951
 952         wc->w_cpos = pos >> osb->s_clustersize_bits;
 953         wc->w_first_new_cpos = UINT_MAX;
 954         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 955         wc->w_clen = cend - wc->w_cpos + 1;
 956         get_bh(di_bh);
 957         wc->w_di_bh = di_bh;
 958
 959         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 960                 wc->w_large_pages = 1;
 961         else
 962                 wc->w_large_pages = 0;
 963
 964         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 965
 966         *wcp = wc;
 967
 968         return 0;
 969 }
 970
 971 /*
 972  * If a page has any new buffers, zero them out here, and mark them uptodate
 973  * and dirty so they'll be written out (in order to prevent uninitialised
 974  * block data from leaking). And clear the new bit.
 975  */
 976 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 977 {
 978         unsigned int block_start, block_end;
 979         struct buffer_head *head, *bh;
 980
 981         BUG_ON(!PageLocked(page));
 982         if (!page_has_buffers(page))
 983                 return;
 984
 985         bh = head = page_buffers(page);
 986         block_start = 0;
 987         do {
 988                 block_end = block_start + bh->b_size;
 989
 990                 if (buffer_new(bh)) {
 991                         if (block_end > from && block_start < to) {
 992                                 if (!PageUptodate(page)) {
 993                                         unsigned start, end;
 994
 995                                         start = max(from, block_start);
 996                                         end = min(to, block_end);
 997
 998                                         zero_user_segment(page, start, end);
 999                                         set_buffer_uptodate(bh);
1000                                 }
1001
1002                                 clear_buffer_new(bh);
1003                                 mark_buffer_dirty(bh);
1004                         }
1005                 }
1006
1007                 block_start = block_end;
1008                 bh = bh->b_this_page;
1009         } while (bh != head);
1010 }
1011
1012 /*
1013  * Only called when we have a failure during allocating write to write
1014  * zero's to the newly allocated region.
1015  */
1016 static void ocfs2_write_failure(struct inode *inode,
1017                                 struct ocfs2_write_ctxt *wc,
1018                                 loff_t user_pos, unsigned user_len)
1019 {
1020         int i;
1021         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1022                 to = user_pos + user_len;
1023         struct page *tmppage;
1024
1025         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1026
1027         for(i = 0; i < wc->w_num_pages; i++) {
1028                 tmppage = wc->w_pages[i];
1029
1030                 if (page_has_buffers(tmppage)) {
1031                         if (ocfs2_should_order_data(inode))
1032                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1033
1034                         block_commit_write(tmppage, from, to);
1035                 }
1036         }
1037 }
1038
1039 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1040                                         struct ocfs2_write_ctxt *wc,
1041                                         struct page *page, u32 cpos,
1042                                         loff_t user_pos, unsigned user_len,
1043                                         int new)
1044 {
1045         int ret;
1046         unsigned int map_from = 0, map_to = 0;
1047         unsigned int cluster_start, cluster_end;
1048         unsigned int user_data_from = 0, user_data_to = 0;
1049
1050         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1051                                         &cluster_start, &cluster_end);
1052
1053         /* treat the write as new if the a hole/lseek spanned across
1054          * the page boundary.
1055          */
1056         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1057                         (page_offset(page) <= user_pos));
1058
1059         if (page == wc->w_target_page) {
1060                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1061                 map_to = map_from + user_len;
1062
1063                 if (new)
1064                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1065                                                     cluster_start, cluster_end,
1066                                                     new);
1067                 else
1068                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1069                                                     map_from, map_to, new);
1070                 if (ret) {
1071                         mlog_errno(ret);
1072                         goto out;
1073                 }
1074
1075                 user_data_from = map_from;
1076                 user_data_to = map_to;
1077                 if (new) {
1078                         map_from = cluster_start;
1079                         map_to = cluster_end;
1080                 }
1081         } else {
1082                 /*
1083                  * If we haven't allocated the new page yet, we
1084                  * shouldn't be writing it out without copying user
1085                  * data. This is likely a math error from the caller.
1086                  */
1087                 BUG_ON(!new);
1088
1089                 map_from = cluster_start;
1090                 map_to = cluster_end;
1091
1092                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1093                                             cluster_start, cluster_end, new);
1094                 if (ret) {
1095                         mlog_errno(ret);
1096                         goto out;
1097                 }
1098         }
1099
1100         /*
1101          * Parts of newly allocated pages need to be zero'd.
1102          *
1103          * Above, we have also rewritten 'to' and 'from' - as far as
1104          * the rest of the function is concerned, the entire cluster
1105          * range inside of a page needs to be written.
1106          *
1107          * We can skip this if the page is up to date - it's already
1108          * been zero'd from being read in as a hole.
1109          */
1110         if (new && !PageUptodate(page))
1111                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1112                                          cpos, user_data_from, user_data_to);
1113
1114         flush_dcache_page(page);
1115
1116 out:
1117         return ret;
1118 }
1119
1120 /*
1121  * This function will only grab one clusters worth of pages.
1122  */
1123 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1124                                       struct ocfs2_write_ctxt *wc,
1125                                       u32 cpos, loff_t user_pos,
1126                                       unsigned user_len, int new,
1127                                       struct page *mmap_page)
1128 {
1129         int ret = 0, i;
1130         unsigned long start, target_index, end_index, index;
1131         struct inode *inode = mapping->host;
1132         loff_t last_byte;
1133
1134         target_index = user_pos >> PAGE_CACHE_SHIFT;
1135
1136         /*
1137          * Figure out how many pages we'll be manipulating here. For
1138          * non allocating write, we just change the one
1139          * page. Otherwise, we'll need a whole clusters worth.  If we're
1140          * writing past i_size, we only need enough pages to cover the
1141          * last page of the write.
1142          */
1143         if (new) {
1144                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1145                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1146                 /*
1147                  * We need the index *past* the last page we could possibly
1148                  * touch.  This is the page past the end of the write or
1149                  * i_size, whichever is greater.
1150                  */
1151                 last_byte = max(user_pos + user_len, i_size_read(inode));
1152                 BUG_ON(last_byte < 1);
1153                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1154                 if ((start + wc->w_num_pages) > end_index)
1155                         wc->w_num_pages = end_index - start;
1156         } else {
1157                 wc->w_num_pages = 1;
1158                 start = target_index;
1159         }
1160
1161         for(i = 0; i < wc->w_num_pages; i++) {
1162                 index = start + i;
1163
1164                 if (index == target_index && mmap_page) {
1165                         /*
1166                          * ocfs2_pagemkwrite() is a little different
1167                          * and wants us to directly use the page
1168                          * passed in.
1169                          */
1170                         lock_page(mmap_page);
1171
1172                         /* Exit and let the caller retry */
1173                         if (mmap_page->mapping != mapping) {
1174                                 WARN_ON(mmap_page->mapping);
1175                                 unlock_page(mmap_page);
1176                                 ret = -EAGAIN;
1177                                 goto out;
1178                         }
1179
1180                         page_cache_get(mmap_page);
1181                         wc->w_pages[i] = mmap_page;
1182                         wc->w_target_locked = true;
1183                 } else {
1184                         wc->w_pages[i] = find_or_create_page(mapping, index,
1185                                                              GFP_NOFS);
1186                         if (!wc->w_pages[i]) {
1187                                 ret = -ENOMEM;
1188                                 mlog_errno(ret);
1189                                 goto out;
1190                         }
1191                 }
1192                 wait_for_stable_page(wc->w_pages[i]);
1193
1194                 if (index == target_index)
1195                         wc->w_target_page = wc->w_pages[i];
1196         }
1197 out:
1198         if (ret)
1199                 wc->w_target_locked = false;
1200         return ret;
1201 }
1202
1203 /*
1204  * Prepare a single cluster for write one cluster into the file.
1205  */
1206 static int ocfs2_write_cluster(struct address_space *mapping,
1207                                u32 phys, unsigned int unwritten,
1208                                unsigned int should_zero,
1209                                struct ocfs2_alloc_context *data_ac,
1210                                struct ocfs2_alloc_context *meta_ac,
1211                                struct ocfs2_write_ctxt *wc, u32 cpos,
1212                                loff_t user_pos, unsigned user_len)
1213 {
1214         int ret, i, new;
1215         u64 v_blkno, p_blkno;
1216         struct inode *inode = mapping->host;
1217         struct ocfs2_extent_tree et;
1218
1219         new = phys == 0 ? 1 : 0;
1220         if (new) {
1221                 u32 tmp_pos;
1222
1223                 /*
1224                  * This is safe to call with the page locks - it won't take
1225                  * any additional semaphores or cluster locks.
1226                  */
1227                 tmp_pos = cpos;
1228                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1229                                            &tmp_pos, 1, 0, wc->w_di_bh,
1230                                            wc->w_handle, data_ac,
1231                                            meta_ac, NULL);
1232                 /*
1233                  * This shouldn't happen because we must have already
1234                  * calculated the correct meta data allocation required. The
1235                  * internal tree allocation code should know how to increase
1236                  * transaction credits itself.
1237                  *
1238                  * If need be, we could handle -EAGAIN for a
1239                  * RESTART_TRANS here.
1240                  */
1241                 mlog_bug_on_msg(ret == -EAGAIN,
1242                                 "Inode %llu: EAGAIN return during allocation.\n",
1243                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1244                 if (ret < 0) {
1245                         mlog_errno(ret);
1246                         goto out;
1247                 }
1248         } else if (unwritten) {
1249                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1250                                               wc->w_di_bh);
1251                 ret = ocfs2_mark_extent_written(inode, &et,
1252                                                 wc->w_handle, cpos, 1, phys,
1253                                                 meta_ac, &wc->w_dealloc);
1254                 if (ret < 0) {
1255                         mlog_errno(ret);
1256                         goto out;
1257                 }
1258         }
1259
1260         if (should_zero)
1261                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1262         else
1263                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1264
1265         /*
1266          * The only reason this should fail is due to an inability to
1267          * find the extent added.
1268          */
1269         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1270                                           NULL);
1271         if (ret < 0) {
1272                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1273                             "at logical block %llu",
1274                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1275                             (unsigned long long)v_blkno);
1276                 goto out;
1277         }
1278
1279         BUG_ON(p_blkno == 0);
1280
1281         for(i = 0; i < wc->w_num_pages; i++) {
1282                 int tmpret;
1283
1284                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1285                                                       wc->w_pages[i], cpos,
1286                                                       user_pos, user_len,
1287                                                       should_zero);
1288                 if (tmpret) {
1289                         mlog_errno(tmpret);
1290                         if (ret == 0)
1291                                 ret = tmpret;
1292                 }
1293         }
1294
1295         /*
1296          * We only have cleanup to do in case of allocating write.
1297          */
1298         if (ret && new)
1299                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1300
1301 out:
1302
1303         return ret;
1304 }
1305
1306 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1307                                        struct ocfs2_alloc_context *data_ac,
1308                                        struct ocfs2_alloc_context *meta_ac,
1309                                        struct ocfs2_write_ctxt *wc,
1310                                        loff_t pos, unsigned len)
1311 {
1312         int ret, i;
1313         loff_t cluster_off;
1314         unsigned int local_len = len;
1315         struct ocfs2_write_cluster_desc *desc;
1316         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1317
1318         for (i = 0; i < wc->w_clen; i++) {
1319                 desc = &wc->w_desc[i];
1320
1321                 /*
1322                  * We have to make sure that the total write passed in
1323                  * doesn't extend past a single cluster.
1324                  */
1325                 local_len = len;
1326                 cluster_off = pos & (osb->s_clustersize - 1);
1327                 if ((cluster_off + local_len) > osb->s_clustersize)
1328                         local_len = osb->s_clustersize - cluster_off;
1329
1330                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1331                                           desc->c_unwritten,
1332                                           desc->c_needs_zero,
1333                                           data_ac, meta_ac,
1334                                           wc, desc->c_cpos, pos, local_len);
1335                 if (ret) {
1336                         mlog_errno(ret);
1337                         goto out;
1338                 }
1339
1340                 len -= local_len;
1341                 pos += local_len;
1342         }
1343
1344         ret = 0;
1345 out:
1346         return ret;
1347 }
1348
1349 /*
1350  * ocfs2_write_end() wants to know which parts of the target page it
1351  * should complete the write on. It's easiest to compute them ahead of
1352  * time when a more complete view of the write is available.
1353  */
1354 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1355                                         struct ocfs2_write_ctxt *wc,
1356                                         loff_t pos, unsigned len, int alloc)
1357 {
1358         struct ocfs2_write_cluster_desc *desc;
1359
1360         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1361         wc->w_target_to = wc->w_target_from + len;
1362
1363         if (alloc == 0)
1364                 return;
1365
1366         /*
1367          * Allocating write - we may have different boundaries based
1368          * on page size and cluster size.
1369          *
1370          * NOTE: We can no longer compute one value from the other as
1371          * the actual write length and user provided length may be
1372          * different.
1373          */
1374
1375         if (wc->w_large_pages) {
1376                 /*
1377                  * We only care about the 1st and last cluster within
1378                  * our range and whether they should be zero'd or not. Either
1379                  * value may be extended out to the start/end of a
1380                  * newly allocated cluster.
1381                  */
1382                 desc = &wc->w_desc[0];
1383                 if (desc->c_needs_zero)
1384                         ocfs2_figure_cluster_boundaries(osb,
1385                                                         desc->c_cpos,
1386                                                         &wc->w_target_from,
1387                                                         NULL);
1388
1389                 desc = &wc->w_desc[wc->w_clen - 1];
1390                 if (desc->c_needs_zero)
1391                         ocfs2_figure_cluster_boundaries(osb,
1392                                                         desc->c_cpos,
1393                                                         NULL,
1394                                                         &wc->w_target_to);
1395         } else {
1396                 wc->w_target_from = 0;
1397                 wc->w_target_to = PAGE_CACHE_SIZE;
1398         }
1399 }
1400
1401 /*
1402  * Populate each single-cluster write descriptor in the write context
1403  * with information about the i/o to be done.
1404  *
1405  * Returns the number of clusters that will have to be allocated, as
1406  * well as a worst case estimate of the number of extent records that
1407  * would have to be created during a write to an unwritten region.
1408  */
1409 static int ocfs2_populate_write_desc(struct inode *inode,
1410                                      struct ocfs2_write_ctxt *wc,
1411                                      unsigned int *clusters_to_alloc,
1412                                      unsigned int *extents_to_split)
1413 {
1414         int ret;
1415         struct ocfs2_write_cluster_desc *desc;
1416         unsigned int num_clusters = 0;
1417         unsigned int ext_flags = 0;
1418         u32 phys = 0;
1419         int i;
1420
1421         *clusters_to_alloc = 0;
1422         *extents_to_split = 0;
1423
1424         for (i = 0; i < wc->w_clen; i++) {
1425                 desc = &wc->w_desc[i];
1426                 desc->c_cpos = wc->w_cpos + i;
1427
1428                 if (num_clusters == 0) {
1429                         /*
1430                          * Need to look up the next extent record.
1431                          */
1432                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1433                                                  &num_clusters, &ext_flags);
1434                         if (ret) {
1435                                 mlog_errno(ret);
1436                                 goto out;
1437                         }
1438
1439                         /* We should already CoW the refcountd extent. */
1440                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1441
1442                         /*
1443                          * Assume worst case - that we're writing in
1444                          * the middle of the extent.
1445                          *
1446                          * We can assume that the write proceeds from
1447                          * left to right, in which case the extent
1448                          * insert code is smart enough to coalesce the
1449                          * next splits into the previous records created.
1450                          */
1451                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1452                                 *extents_to_split = *extents_to_split + 2;
1453                 } else if (phys) {
1454                         /*
1455                          * Only increment phys if it doesn't describe
1456                          * a hole.
1457                          */
1458                         phys++;
1459                 }
1460
1461                 /*
1462                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1463                  * file that got extended.  w_first_new_cpos tells us
1464                  * where the newly allocated clusters are so we can
1465                  * zero them.
1466                  */
1467                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1468                         BUG_ON(phys == 0);
1469                         desc->c_needs_zero = 1;
1470                 }
1471
1472                 desc->c_phys = phys;
1473                 if (phys == 0) {
1474                         desc->c_new = 1;
1475                         desc->c_needs_zero = 1;
1476                         *clusters_to_alloc = *clusters_to_alloc + 1;
1477                 }
1478
1479                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1480                         desc->c_unwritten = 1;
1481                         desc->c_needs_zero = 1;
1482                 }
1483
1484                 num_clusters--;
1485         }
1486
1487         ret = 0;
1488 out:
1489         return ret;
1490 }
1491
1492 static int ocfs2_write_begin_inline(struct address_space *mapping,
1493                                     struct inode *inode,
1494                                     struct ocfs2_write_ctxt *wc)
1495 {
1496         int ret;
1497         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1498         struct page *page;
1499         handle_t *handle;
1500         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1501
1502         page = find_or_create_page(mapping, 0, GFP_NOFS);
1503         if (!page) {
1504                 ret = -ENOMEM;
1505                 mlog_errno(ret);
1506                 goto out;
1507         }
1508         /*
1509          * If we don't set w_num_pages then this page won't get unlocked
1510          * and freed on cleanup of the write context.
1511          */
1512         wc->w_pages[0] = wc->w_target_page = page;
1513         wc->w_num_pages = 1;
1514
1515         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1516         if (IS_ERR(handle)) {
1517                 ret = PTR_ERR(handle);
1518                 mlog_errno(ret);
1519                 goto out;
1520         }
1521
1522         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1523                                       OCFS2_JOURNAL_ACCESS_WRITE);
1524         if (ret) {
1525                 ocfs2_commit_trans(osb, handle);
1526
1527                 mlog_errno(ret);
1528                 goto out;
1529         }
1530
1531         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1532                 ocfs2_set_inode_data_inline(inode, di);
1533
1534         if (!PageUptodate(page)) {
1535                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1536                 if (ret) {
1537                         ocfs2_commit_trans(osb, handle);
1538
1539                         goto out;
1540                 }
1541         }
1542
1543         wc->w_handle = handle;
1544 out:
1545         return ret;
1546 }
1547
1548 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1549 {
1550         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1551
1552         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1553                 return 1;
1554         return 0;
1555 }
1556
1557 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1558                                           struct inode *inode, loff_t pos,
1559                                           unsigned len, struct page *mmap_page,
1560                                           struct ocfs2_write_ctxt *wc)
1561 {
1562         int ret, written = 0;
1563         loff_t end = pos + len;
1564         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1565         struct ocfs2_dinode *di = NULL;
1566
1567         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1568                                              len, (unsigned long long)pos,
1569                                              oi->ip_dyn_features);
1570
1571         /*
1572          * Handle inodes which already have inline data 1st.
1573          */
1574         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1575                 if (mmap_page == NULL &&
1576                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1577                         goto do_inline_write;
1578
1579                 /*
1580                  * The write won't fit - we have to give this inode an
1581                  * inline extent list now.
1582                  */
1583                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1584                 if (ret)
1585                         mlog_errno(ret);
1586                 goto out;
1587         }
1588
1589         /*
1590          * Check whether the inode can accept inline data.
1591          */
1592         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1593                 return 0;
1594
1595         /*
1596          * Check whether the write can fit.
1597          */
1598         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1599         if (mmap_page ||
1600             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1601                 return 0;
1602
1603 do_inline_write:
1604         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1605         if (ret) {
1606                 mlog_errno(ret);
1607                 goto out;
1608         }
1609
1610         /*
1611          * This signals to the caller that the data can be written
1612          * inline.
1613          */
1614         written = 1;
1615 out:
1616         return written ? written : ret;
1617 }
1618
1619 /*
1620  * This function only does anything for file systems which can't
1621  * handle sparse files.
1622  *
1623  * What we want to do here is fill in any hole between the current end
1624  * of allocation and the end of our write. That way the rest of the
1625  * write path can treat it as an non-allocating write, which has no
1626  * special case code for sparse/nonsparse files.
1627  */
1628 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1629                                         struct buffer_head *di_bh,
1630                                         loff_t pos, unsigned len,
1631                                         struct ocfs2_write_ctxt *wc)
1632 {
1633         int ret;
1634         loff_t newsize = pos + len;
1635
1636         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1637
1638         if (newsize <= i_size_read(inode))
1639                 return 0;
1640
1641         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1642         if (ret)
1643                 mlog_errno(ret);
1644
1645         wc->w_first_new_cpos =
1646                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1647
1648         return ret;
1649 }
1650
1651 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1652                            loff_t pos)
1653 {
1654         int ret = 0;
1655
1656         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1657         if (pos > i_size_read(inode))
1658                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1659
1660         return ret;
1661 }
1662
1663 /*
1664  * Try to flush truncate logs if we can free enough clusters from it.
1665  * As for return value, "< 0" means error, "0" no space and "1" means
1666  * we have freed enough spaces and let the caller try to allocate again.
1667  */
1668 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1669                                           unsigned int needed)
1670 {
1671         tid_t target;
1672         int ret = 0;
1673         unsigned int truncated_clusters;
1674
1675         mutex_lock(&osb->osb_tl_inode->i_mutex);
1676         truncated_clusters = osb->truncated_clusters;
1677         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1678
1679         /*
1680          * Check whether we can succeed in allocating if we free
1681          * the truncate log.
1682          */
1683         if (truncated_clusters < needed)
1684                 goto out;
1685
1686         ret = ocfs2_flush_truncate_log(osb);
1687         if (ret) {
1688                 mlog_errno(ret);
1689                 goto out;
1690         }
1691
1692         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1693                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1694                 ret = 1;
1695         }
1696 out:
1697         return ret;
1698 }
1699
1700 int ocfs2_write_begin_nolock(struct file *filp,
1701                              struct address_space *mapping,
1702                              loff_t pos, unsigned len, unsigned flags,
1703                              struct page **pagep, void **fsdata,
1704                              struct buffer_head *di_bh, struct page *mmap_page)
1705 {
1706         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1707         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1708         struct ocfs2_write_ctxt *wc;
1709         struct inode *inode = mapping->host;
1710         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1711         struct ocfs2_dinode *di;
1712         struct ocfs2_alloc_context *data_ac = NULL;
1713         struct ocfs2_alloc_context *meta_ac = NULL;
1714         handle_t *handle;
1715         struct ocfs2_extent_tree et;
1716         int try_free = 1, ret1;
1717
1718 try_again:
1719         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1720         if (ret) {
1721                 mlog_errno(ret);
1722                 return ret;
1723         }
1724
1725         if (ocfs2_supports_inline_data(osb)) {
1726                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1727                                                      mmap_page, wc);
1728                 if (ret == 1) {
1729                         ret = 0;
1730                         goto success;
1731                 }
1732                 if (ret < 0) {
1733                         mlog_errno(ret);
1734                         goto out;
1735                 }
1736         }
1737
1738         if (ocfs2_sparse_alloc(osb))
1739                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1740         else
1741                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1742                                                    wc);
1743         if (ret) {
1744                 mlog_errno(ret);
1745                 goto out;
1746         }
1747
1748         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1749         if (ret < 0) {
1750                 mlog_errno(ret);
1751                 goto out;
1752         } else if (ret == 1) {
1753                 clusters_need = wc->w_clen;
1754                 ret = ocfs2_refcount_cow(inode, di_bh,
1755                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1756                 if (ret) {
1757                         mlog_errno(ret);
1758                         goto out;
1759                 }
1760         }
1761
1762         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1763                                         &extents_to_split);
1764         if (ret) {
1765                 mlog_errno(ret);
1766                 goto out;
1767         }
1768         clusters_need += clusters_to_alloc;
1769
1770         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1771
1772         trace_ocfs2_write_begin_nolock(
1773                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1774                         (long long)i_size_read(inode),
1775                         le32_to_cpu(di->i_clusters),
1776                         pos, len, flags, mmap_page,
1777                         clusters_to_alloc, extents_to_split);
1778
1779         /*
1780          * We set w_target_from, w_target_to here so that
1781          * ocfs2_write_end() knows which range in the target page to
1782          * write out. An allocation requires that we write the entire
1783          * cluster range.
1784          */
1785         if (clusters_to_alloc || extents_to_split) {
1786                 /*
1787                  * XXX: We are stretching the limits of
1788                  * ocfs2_lock_allocators(). It greatly over-estimates
1789                  * the work to be done.
1790                  */
1791                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1792                                               wc->w_di_bh);
1793                 ret = ocfs2_lock_allocators(inode, &et,
1794                                             clusters_to_alloc, extents_to_split,
1795                                             &data_ac, &meta_ac);
1796                 if (ret) {
1797                         mlog_errno(ret);
1798                         goto out;
1799                 }
1800
1801                 if (data_ac)
1802                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1803
1804                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1805                                                     &di->id2.i_list,
1806                                                     clusters_to_alloc);
1807
1808         }
1809
1810         /*
1811          * We have to zero sparse allocated clusters, unwritten extent clusters,
1812          * and non-sparse clusters we just extended.  For non-sparse writes,
1813          * we know zeros will only be needed in the first and/or last cluster.
1814          */
1815         if (clusters_to_alloc || extents_to_split ||
1816             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1817                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1818                 cluster_of_pages = 1;
1819         else
1820                 cluster_of_pages = 0;
1821
1822         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1823
1824         handle = ocfs2_start_trans(osb, credits);
1825         if (IS_ERR(handle)) {
1826                 ret = PTR_ERR(handle);
1827                 mlog_errno(ret);
1828                 goto out;
1829         }
1830
1831         wc->w_handle = handle;
1832
1833         if (clusters_to_alloc) {
1834                 ret = dquot_alloc_space_nodirty(inode,
1835                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1836                 if (ret)
1837                         goto out_commit;
1838         }
1839         /*
1840          * We don't want this to fail in ocfs2_write_end(), so do it
1841          * here.
1842          */
1843         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1844                                       OCFS2_JOURNAL_ACCESS_WRITE);
1845         if (ret) {
1846                 mlog_errno(ret);
1847                 goto out_quota;
1848         }
1849
1850         /*
1851          * Fill our page array first. That way we've grabbed enough so
1852          * that we can zero and flush if we error after adding the
1853          * extent.
1854          */
1855         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1856                                          cluster_of_pages, mmap_page);
1857         if (ret && ret != -EAGAIN) {
1858                 mlog_errno(ret);
1859                 goto out_quota;
1860         }
1861
1862         /*
1863          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1864          * the target page. In this case, we exit with no error and no target
1865          * page. This will trigger the caller, page_mkwrite(), to re-try
1866          * the operation.
1867          */
1868         if (ret == -EAGAIN) {
1869                 BUG_ON(wc->w_target_page);
1870                 ret = 0;
1871                 goto out_quota;
1872         }
1873
1874         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1875                                           len);
1876         if (ret) {
1877                 mlog_errno(ret);
1878                 goto out_quota;
1879         }
1880
1881         if (data_ac)
1882                 ocfs2_free_alloc_context(data_ac);
1883         if (meta_ac)
1884                 ocfs2_free_alloc_context(meta_ac);
1885
1886 success:
1887         *pagep = wc->w_target_page;
1888         *fsdata = wc;
1889         return 0;
1890 out_quota:
1891         if (clusters_to_alloc)
1892                 dquot_free_space(inode,
1893                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1894 out_commit:
1895         ocfs2_commit_trans(osb, handle);
1896
1897 out:
1898         ocfs2_free_write_ctxt(wc);
1899
1900         if (data_ac)
1901                 ocfs2_free_alloc_context(data_ac);
1902         if (meta_ac)
1903                 ocfs2_free_alloc_context(meta_ac);
1904
1905         if (ret == -ENOSPC && try_free) {
1906                 /*
1907                  * Try to free some truncate log so that we can have enough
1908                  * clusters to allocate.
1909                  */
1910                 try_free = 0;
1911
1912                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1913                 if (ret1 == 1)
1914                         goto try_again;
1915
1916                 if (ret1 < 0)
1917                         mlog_errno(ret1);
1918         }
1919
1920         return ret;
1921 }
1922
1923 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1924                              loff_t pos, unsigned len, unsigned flags,
1925                              struct page **pagep, void **fsdata)
1926 {
1927         int ret;
1928         struct buffer_head *di_bh = NULL;
1929         struct inode *inode = mapping->host;
1930
1931         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1932         if (ret) {
1933                 mlog_errno(ret);
1934                 return ret;
1935         }
1936
1937         /*
1938          * Take alloc sem here to prevent concurrent lookups. That way
1939          * the mapping, zeroing and tree manipulation within
1940          * ocfs2_write() will be safe against ->readpage(). This
1941          * should also serve to lock out allocation from a shared
1942          * writeable region.
1943          */
1944         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1945
1946         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1947                                        fsdata, di_bh, NULL);
1948         if (ret) {
1949                 mlog_errno(ret);
1950                 goto out_fail;
1951         }
1952
1953         brelse(di_bh);
1954
1955         return 0;
1956
1957 out_fail:
1958         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1959
1960         brelse(di_bh);
1961         ocfs2_inode_unlock(inode, 1);
1962
1963         return ret;
1964 }
1965
1966 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1967                                    unsigned len, unsigned *copied,
1968                                    struct ocfs2_dinode *di,
1969                                    struct ocfs2_write_ctxt *wc)
1970 {
1971         void *kaddr;
1972
1973         if (unlikely(*copied < len)) {
1974                 if (!PageUptodate(wc->w_target_page)) {
1975                         *copied = 0;
1976                         return;
1977                 }
1978         }
1979
1980         kaddr = kmap_atomic(wc->w_target_page);
1981         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1982         kunmap_atomic(kaddr);
1983
1984         trace_ocfs2_write_end_inline(
1985              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1986              (unsigned long long)pos, *copied,
1987              le16_to_cpu(di->id2.i_data.id_count),
1988              le16_to_cpu(di->i_dyn_features));
1989 }
1990
1991 int ocfs2_write_end_nolock(struct address_space *mapping,
1992                            loff_t pos, unsigned len, unsigned copied,
1993                            struct page *page, void *fsdata)
1994 {
1995         int i;
1996         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1997         struct inode *inode = mapping->host;
1998         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1999         struct ocfs2_write_ctxt *wc = fsdata;
2000         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2001         handle_t *handle = wc->w_handle;
2002         struct page *tmppage;
2003
2004         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2005                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2006                 goto out_write_size;
2007         }
2008
2009         if (unlikely(copied < len)) {
2010                 if (!PageUptodate(wc->w_target_page))
2011                         copied = 0;
2012
2013                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2014                                        start+len);
2015         }
2016         flush_dcache_page(wc->w_target_page);
2017
2018         for(i = 0; i < wc->w_num_pages; i++) {
2019                 tmppage = wc->w_pages[i];
2020
2021                 if (tmppage == wc->w_target_page) {
2022                         from = wc->w_target_from;
2023                         to = wc->w_target_to;
2024
2025                         BUG_ON(from > PAGE_CACHE_SIZE ||
2026                                to > PAGE_CACHE_SIZE ||
2027                                to < from);
2028                 } else {
2029                         /*
2030                          * Pages adjacent to the target (if any) imply
2031                          * a hole-filling write in which case we want
2032                          * to flush their entire range.
2033                          */
2034                         from = 0;
2035                         to = PAGE_CACHE_SIZE;
2036                 }
2037
2038                 if (page_has_buffers(tmppage)) {
2039                         if (ocfs2_should_order_data(inode))
2040                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2041                         block_commit_write(tmppage, from, to);
2042                 }
2043         }
2044
2045 out_write_size:
2046         pos += copied;
2047         if (pos > inode->i_size) {
2048                 i_size_write(inode, pos);
2049                 mark_inode_dirty(inode);
2050         }
2051         inode->i_blocks = ocfs2_inode_sector_count(inode);
2052         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2053         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2054         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2055         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2056         ocfs2_journal_dirty(handle, wc->w_di_bh);
2057
2058         ocfs2_commit_trans(osb, handle);
2059
2060         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2061
2062         ocfs2_free_write_ctxt(wc);
2063
2064         return copied;
2065 }
2066
2067 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2068                            loff_t pos, unsigned len, unsigned copied,
2069                            struct page *page, void *fsdata)
2070 {
2071         int ret;
2072         struct inode *inode = mapping->host;
2073
2074         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2075
2076         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2077         ocfs2_inode_unlock(inode, 1);
2078
2079         return ret;
2080 }
2081
2082 const struct address_space_operations ocfs2_aops = {
2083         .readpage               = ocfs2_readpage,
2084         .readpages              = ocfs2_readpages,
2085         .writepage              = ocfs2_writepage,
2086         .write_begin            = ocfs2_write_begin,
2087         .write_end              = ocfs2_write_end,
2088         .bmap                   = ocfs2_bmap,
2089         .direct_IO              = ocfs2_direct_IO,
2090         .invalidatepage         = ocfs2_invalidatepage,
2091         .releasepage            = ocfs2_releasepage,
2092         .migratepage            = buffer_migrate_page,
2093         .is_partially_uptodate  = block_is_partially_uptodate,
2094         .error_remove_page      = generic_error_remove_page,
2095 };