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                 err = -ENOMEM;
  84                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  85                      "%llu\n", (unsigned long long)iblock);
  86                 goto bail;
  87         }
  88
  89         /* We don't use the page cache to create symlink data, so if
  90          * need be, copy it over from the buffer cache. */
  91         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  92                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  93                             iblock;
  94                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  95                 if (!buffer_cache_bh) {
  96                         err = -ENOMEM;
  97                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  98                         goto bail;
  99                 }
 100
 101                 /* we haven't locked out transactions, so a commit
 102                  * could've happened. Since we've got a reference on
 103                  * the bh, even if it commits while we're doing the
 104                  * copy, the data is still good. */
 105                 if (buffer_jbd(buffer_cache_bh)
 106                     && ocfs2_inode_is_new(inode)) {
 107                         kaddr = kmap_atomic(bh_result->b_page);
 108                         if (!kaddr) {
 109                                 mlog(ML_ERROR, "couldn't kmap!\n");
 110                                 goto bail;
 111                         }
 112                         memcpy(kaddr + (bh_result->b_size * iblock),
 113                                buffer_cache_bh->b_data,
 114                                bh_result->b_size);
 115                         kunmap_atomic(kaddr);
 116                         set_buffer_uptodate(bh_result);
 117                 }
 118                 brelse(buffer_cache_bh);
 119         }
 120
 121         map_bh(bh_result, inode->i_sb,
 122                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 123
 124         err = 0;
 125
 126 bail:
 127         brelse(bh);
 128
 129         return err;
 130 }
 131
 132 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 133                     struct buffer_head *bh_result, int create)
 134 {
 135         int err = 0;
 136         unsigned int ext_flags;
 137         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 138         u64 p_blkno, count, past_eof;
 139         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 140
 141         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 142                               (unsigned long long)iblock, bh_result, create);
 143
 144         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 145                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 146                      inode, inode->i_ino);
 147
 148         if (S_ISLNK(inode->i_mode)) {
 149                 /* this always does I/O for some reason. */
 150                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 151                 goto bail;
 152         }
 153
 154         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 155                                           &ext_flags);
 156         if (err) {
 157                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 158                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 159                      (unsigned long long)p_blkno);
 160                 goto bail;
 161         }
 162
 163         if (max_blocks < count)
 164                 count = max_blocks;
 165
 166         /*
 167          * ocfs2 never allocates in this function - the only time we
 168          * need to use BH_New is when we're extending i_size on a file
 169          * system which doesn't support holes, in which case BH_New
 170          * allows __block_write_begin() to zero.
 171          *
 172          * If we see this on a sparse file system, then a truncate has
 173          * raced us and removed the cluster. In this case, we clear
 174          * the buffers dirty and uptodate bits and let the buffer code
 175          * ignore it as a hole.
 176          */
 177         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 178                 clear_buffer_dirty(bh_result);
 179                 clear_buffer_uptodate(bh_result);
 180                 goto bail;
 181         }
 182
 183         /* Treat the unwritten extent as a hole for zeroing purposes. */
 184         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 185                 map_bh(bh_result, inode->i_sb, p_blkno);
 186
 187         bh_result->b_size = count << inode->i_blkbits;
 188
 189         if (!ocfs2_sparse_alloc(osb)) {
 190                 if (p_blkno == 0) {
 191                         err = -EIO;
 192                         mlog(ML_ERROR,
 193                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 194                              (unsigned long long)iblock,
 195                              (unsigned long long)p_blkno,
 196                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 197                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 198                         dump_stack();
 199                         goto bail;
 200                 }
 201         }
 202
 203         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 204
 205         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 206                                   (unsigned long long)past_eof);
 207         if (create && (iblock >= past_eof))
 208                 set_buffer_new(bh_result);
 209
 210 bail:
 211         if (err < 0)
 212                 err = -EIO;
 213
 214         return err;
 215 }
 216
 217 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 218                            struct buffer_head *di_bh)
 219 {
 220         void *kaddr;
 221         loff_t size;
 222         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 223
 224         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 225                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 226                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 227                 return -EROFS;
 228         }
 229
 230         size = i_size_read(inode);
 231
 232         if (size > PAGE_CACHE_SIZE ||
 233             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 234                 ocfs2_error(inode->i_sb,
 235                             "Inode %llu has with inline data has bad size: %Lu",
 236                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 237                             (unsigned long long)size);
 238                 return -EROFS;
 239         }
 240
 241         kaddr = kmap_atomic(page);
 242         if (size)
 243                 memcpy(kaddr, di->id2.i_data.id_data, size);
 244         /* Clear the remaining part of the page */
 245         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 246         flush_dcache_page(page);
 247         kunmap_atomic(kaddr);
 248
 249         SetPageUptodate(page);
 250
 251         return 0;
 252 }
 253
 254 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 255 {
 256         int ret;
 257         struct buffer_head *di_bh = NULL;
 258
 259         BUG_ON(!PageLocked(page));
 260         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 261
 262         ret = ocfs2_read_inode_block(inode, &di_bh);
 263         if (ret) {
 264                 mlog_errno(ret);
 265                 goto out;
 266         }
 267
 268         ret = ocfs2_read_inline_data(inode, page, di_bh);
 269 out:
 270         unlock_page(page);
 271
 272         brelse(di_bh);
 273         return ret;
 274 }
 275
 276 static int ocfs2_readpage(struct file *file, struct page *page)
 277 {
 278         struct inode *inode = page->mapping->host;
 279         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 280         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 281         int ret, unlock = 1;
 282
 283         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 284                              (page ? page->index : 0));
 285
 286         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 287         if (ret != 0) {
 288                 if (ret == AOP_TRUNCATED_PAGE)
 289                         unlock = 0;
 290                 mlog_errno(ret);
 291                 goto out;
 292         }
 293
 294         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 295                 /*
 296                  * Unlock the page and cycle ip_alloc_sem so that we don't
 297                  * busyloop waiting for ip_alloc_sem to unlock
 298                  */
 299                 ret = AOP_TRUNCATED_PAGE;
 300                 unlock_page(page);
 301                 unlock = 0;
 302                 down_read(&oi->ip_alloc_sem);
 303                 up_read(&oi->ip_alloc_sem);
 304                 goto out_inode_unlock;
 305         }
 306
 307         /*
 308          * i_size might have just been updated as we grabed the meta lock.  We
 309          * might now be discovering a truncate that hit on another node.
 310          * block_read_full_page->get_block freaks out if it is asked to read
 311          * beyond the end of a file, so we check here.  Callers
 312          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 313          * and notice that the page they just read isn't needed.
 314          *
 315          * XXX sys_readahead() seems to get that wrong?
 316          */
 317         if (start >= i_size_read(inode)) {
 318                 zero_user(page, 0, PAGE_SIZE);
 319                 SetPageUptodate(page);
 320                 ret = 0;
 321                 goto out_alloc;
 322         }
 323
 324         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 325                 ret = ocfs2_readpage_inline(inode, page);
 326         else
 327                 ret = block_read_full_page(page, ocfs2_get_block);
 328         unlock = 0;
 329
 330 out_alloc:
 331         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 332 out_inode_unlock:
 333         ocfs2_inode_unlock(inode, 0);
 334 out:
 335         if (unlock)
 336                 unlock_page(page);
 337         return ret;
 338 }
 339
 340 /*
 341  * This is used only for read-ahead. Failures or difficult to handle
 342  * situations are safe to ignore.
 343  *
 344  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 345  * are quite large (243 extents on 4k blocks), so most inodes don't
 346  * grow out to a tree. If need be, detecting boundary extents could
 347  * trivially be added in a future version of ocfs2_get_block().
 348  */
 349 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 350                            struct list_head *pages, unsigned nr_pages)
 351 {
 352         int ret, err = -EIO;
 353         struct inode *inode = mapping->host;
 354         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 355         loff_t start;
 356         struct page *last;
 357
 358         /*
 359          * Use the nonblocking flag for the dlm code to avoid page
 360          * lock inversion, but don't bother with retrying.
 361          */
 362         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 363         if (ret)
 364                 return err;
 365
 366         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 367                 ocfs2_inode_unlock(inode, 0);
 368                 return err;
 369         }
 370
 371         /*
 372          * Don't bother with inline-data. There isn't anything
 373          * to read-ahead in that case anyway...
 374          */
 375         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 376                 goto out_unlock;
 377
 378         /*
 379          * Check whether a remote node truncated this file - we just
 380          * drop out in that case as it's not worth handling here.
 381          */
 382         last = list_entry(pages->prev, struct page, lru);
 383         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 384         if (start >= i_size_read(inode))
 385                 goto out_unlock;
 386
 387         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 388
 389 out_unlock:
 390         up_read(&oi->ip_alloc_sem);
 391         ocfs2_inode_unlock(inode, 0);
 392
 393         return err;
 394 }
 395
 396 /* Note: Because we don't support holes, our allocation has
 397  * already happened (allocation writes zeros to the file data)
 398  * so we don't have to worry about ordered writes in
 399  * ocfs2_writepage.
 400  *
 401  * ->writepage is called during the process of invalidating the page cache
 402  * during blocked lock processing.  It can't block on any cluster locks
 403  * to during block mapping.  It's relying on the fact that the block
 404  * mapping can't have disappeared under the dirty pages that it is
 405  * being asked to write back.
 406  */
 407 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 408 {
 409         trace_ocfs2_writepage(
 410                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 411                 page->index);
 412
 413         return block_write_full_page(page, ocfs2_get_block, wbc);
 414 }
 415
 416 /* Taken from ext3. We don't necessarily need the full blown
 417  * functionality yet, but IMHO it's better to cut and paste the whole
 418  * thing so we can avoid introducing our own bugs (and easily pick up
 419  * their fixes when they happen) --Mark */
 420 int walk_page_buffers(  handle_t *handle,
 421                         struct buffer_head *head,
 422                         unsigned from,
 423                         unsigned to,
 424                         int *partial,
 425                         int (*fn)(      handle_t *handle,
 426                                         struct buffer_head *bh))
 427 {
 428         struct buffer_head *bh;
 429         unsigned block_start, block_end;
 430         unsigned blocksize = head->b_size;
 431         int err, ret = 0;
 432         struct buffer_head *next;
 433
 434         for (   bh = head, block_start = 0;
 435                 ret == 0 && (bh != head || !block_start);
 436                 block_start = block_end, bh = next)
 437         {
 438                 next = bh->b_this_page;
 439                 block_end = block_start + blocksize;
 440                 if (block_end <= from || block_start >= to) {
 441                         if (partial && !buffer_uptodate(bh))
 442                                 *partial = 1;
 443                         continue;
 444                 }
 445                 err = (*fn)(handle, bh);
 446                 if (!ret)
 447                         ret = err;
 448         }
 449         return ret;
 450 }
 451
 452 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 453 {
 454         sector_t status;
 455         u64 p_blkno = 0;
 456         int err = 0;
 457         struct inode *inode = mapping->host;
 458
 459         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 460                          (unsigned long long)block);
 461
 462         /* We don't need to lock journal system files, since they aren't
 463          * accessed concurrently from multiple nodes.
 464          */
 465         if (!INODE_JOURNAL(inode)) {
 466                 err = ocfs2_inode_lock(inode, NULL, 0);
 467                 if (err) {
 468                         if (err != -ENOENT)
 469                                 mlog_errno(err);
 470                         goto bail;
 471                 }
 472                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 473         }
 474
 475         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 476                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 477                                                   NULL);
 478
 479         if (!INODE_JOURNAL(inode)) {
 480                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 481                 ocfs2_inode_unlock(inode, 0);
 482         }
 483
 484         if (err) {
 485                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 486                      (unsigned long long)block);
 487                 mlog_errno(err);
 488                 goto bail;
 489         }
 490
 491 bail:
 492         status = err ? 0 : p_blkno;
 493
 494         return status;
 495 }
 496
 497 /*
 498  * TODO: Make this into a generic get_blocks function.
 499  *
 500  * From do_direct_io in direct-io.c:
 501  *  "So what we do is to permit the ->get_blocks function to populate
 502  *   bh.b_size with the size of IO which is permitted at this offset and
 503  *   this i_blkbits."
 504  *
 505  * This function is called directly from get_more_blocks in direct-io.c.
 506  *
 507  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 508  *                                      fs_count, map_bh, dio->rw == WRITE);
 509  *
 510  * Note that we never bother to allocate blocks here, and thus ignore the
 511  * create argument.
 512  */
 513 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 514                                      struct buffer_head *bh_result, int create)
 515 {
 516         int ret;
 517         u64 p_blkno, inode_blocks, contig_blocks;
 518         unsigned int ext_flags;
 519         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 520         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 521
 522         /* This function won't even be called if the request isn't all
 523          * nicely aligned and of the right size, so there's no need
 524          * for us to check any of that. */
 525
 526         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 527
 528         /* This figures out the size of the next contiguous block, and
 529          * our logical offset */
 530         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 531                                           &contig_blocks, &ext_flags);
 532         if (ret) {
 533                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 534                      (unsigned long long)iblock);
 535                 ret = -EIO;
 536                 goto bail;
 537         }
 538
 539         /* We should already CoW the refcounted extent in case of create. */
 540         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 541
 542         /*
 543          * get_more_blocks() expects us to describe a hole by clearing
 544          * the mapped bit on bh_result().
 545          *
 546          * Consider an unwritten extent as a hole.
 547          */
 548         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 549                 map_bh(bh_result, inode->i_sb, p_blkno);
 550         else
 551                 clear_buffer_mapped(bh_result);
 552
 553         /* make sure we don't map more than max_blocks blocks here as
 554            that's all the kernel will handle at this point. */
 555         if (max_blocks < contig_blocks)
 556                 contig_blocks = max_blocks;
 557         bh_result->b_size = contig_blocks << blocksize_bits;
 558 bail:
 559         return ret;
 560 }
 561
 562 /*
 563  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 564  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 565  * to protect io on one node from truncation on another.
 566  */
 567 static void ocfs2_dio_end_io(struct kiocb *iocb,
 568                              loff_t offset,
 569                              ssize_t bytes,
 570                              void *private)
 571 {
 572         struct inode *inode = file_inode(iocb->ki_filp);
 573         int level;
 574
 575         /* this io's submitter should not have unlocked this before we could */
 576         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 577
 578         if (ocfs2_iocb_is_sem_locked(iocb))
 579                 ocfs2_iocb_clear_sem_locked(iocb);
 580
 581         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 582                 ocfs2_iocb_clear_unaligned_aio(iocb);
 583
 584                 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio);
 585         }
 586
 587         ocfs2_iocb_clear_rw_locked(iocb);
 588
 589         level = ocfs2_iocb_rw_locked_level(iocb);
 590         ocfs2_rw_unlock(inode, level);
 591 }
 592
 593 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 594 {
 595         if (!page_has_buffers(page))
 596                 return 0;
 597         return try_to_free_buffers(page);
 598 }
 599
 600 static ssize_t ocfs2_direct_IO(int rw,
 601                                struct kiocb *iocb,
 602                                struct iov_iter *iter,
 603                                loff_t offset)
 604 {
 605         struct file *file = iocb->ki_filp;
 606         struct inode *inode = file_inode(file)->i_mapping->host;
 607
 608         /*
 609          * Fallback to buffered I/O if we see an inode without
 610          * extents.
 611          */
 612         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 613                 return 0;
 614
 615         /* Fallback to buffered I/O if we are appending. */
 616         if (i_size_read(inode) <= offset)
 617                 return 0;
 618
 619         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 620                                     iter, offset,
 621                                     ocfs2_direct_IO_get_blocks,
 622                                     ocfs2_dio_end_io, NULL, 0);
 623 }
 624
 625 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 626                                             u32 cpos,
 627                                             unsigned int *start,
 628                                             unsigned int *end)
 629 {
 630         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 631
 632         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 633                 unsigned int cpp;
 634
 635                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 636
 637                 cluster_start = cpos % cpp;
 638                 cluster_start = cluster_start << osb->s_clustersize_bits;
 639
 640                 cluster_end = cluster_start + osb->s_clustersize;
 641         }
 642
 643         BUG_ON(cluster_start > PAGE_SIZE);
 644         BUG_ON(cluster_end > PAGE_SIZE);
 645
 646         if (start)
 647                 *start = cluster_start;
 648         if (end)
 649                 *end = cluster_end;
 650 }
 651
 652 /*
 653  * 'from' and 'to' are the region in the page to avoid zeroing.
 654  *
 655  * If pagesize > clustersize, this function will avoid zeroing outside
 656  * of the cluster boundary.
 657  *
 658  * from == to == 0 is code for "zero the entire cluster region"
 659  */
 660 static void ocfs2_clear_page_regions(struct page *page,
 661                                      struct ocfs2_super *osb, u32 cpos,
 662                                      unsigned from, unsigned to)
 663 {
 664         void *kaddr;
 665         unsigned int cluster_start, cluster_end;
 666
 667         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 668
 669         kaddr = kmap_atomic(page);
 670
 671         if (from || to) {
 672                 if (from > cluster_start)
 673                         memset(kaddr + cluster_start, 0, from - cluster_start);
 674                 if (to < cluster_end)
 675                         memset(kaddr + to, 0, cluster_end - to);
 676         } else {
 677                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 678         }
 679
 680         kunmap_atomic(kaddr);
 681 }
 682
 683 /*
 684  * Nonsparse file systems fully allocate before we get to the write
 685  * code. This prevents ocfs2_write() from tagging the write as an
 686  * allocating one, which means ocfs2_map_page_blocks() might try to
 687  * read-in the blocks at the tail of our file. Avoid reading them by
 688  * testing i_size against each block offset.
 689  */
 690 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 691                                  unsigned int block_start)
 692 {
 693         u64 offset = page_offset(page) + block_start;
 694
 695         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 696                 return 1;
 697
 698         if (i_size_read(inode) > offset)
 699                 return 1;
 700
 701         return 0;
 702 }
 703
 704 /*
 705  * Some of this taken from __block_write_begin(). We already have our
 706  * mapping by now though, and the entire write will be allocating or
 707  * it won't, so not much need to use BH_New.
 708  *
 709  * This will also skip zeroing, which is handled externally.
 710  */
 711 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 712                           struct inode *inode, unsigned int from,
 713                           unsigned int to, int new)
 714 {
 715         int ret = 0;
 716         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 717         unsigned int block_end, block_start;
 718         unsigned int bsize = 1 << inode->i_blkbits;
 719
 720         if (!page_has_buffers(page))
 721                 create_empty_buffers(page, bsize, 0);
 722
 723         head = page_buffers(page);
 724         for (bh = head, block_start = 0; bh != head || !block_start;
 725              bh = bh->b_this_page, block_start += bsize) {
 726                 block_end = block_start + bsize;
 727
 728                 clear_buffer_new(bh);
 729
 730                 /*
 731                  * Ignore blocks outside of our i/o range -
 732                  * they may belong to unallocated clusters.
 733                  */
 734                 if (block_start >= to || block_end <= from) {
 735                         if (PageUptodate(page))
 736                                 set_buffer_uptodate(bh);
 737                         continue;
 738                 }
 739
 740                 /*
 741                  * For an allocating write with cluster size >= page
 742                  * size, we always write the entire page.
 743                  */
 744                 if (new)
 745                         set_buffer_new(bh);
 746
 747                 if (!buffer_mapped(bh)) {
 748                         map_bh(bh, inode->i_sb, *p_blkno);
 749                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 750                 }
 751
 752                 if (PageUptodate(page)) {
 753                         if (!buffer_uptodate(bh))
 754                                 set_buffer_uptodate(bh);
 755                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 756                            !buffer_new(bh) &&
 757                            ocfs2_should_read_blk(inode, page, block_start) &&
 758                            (block_start < from || block_end > to)) {
 759                         ll_rw_block(READ, 1, &bh);
 760                         *wait_bh++=bh;
 761                 }
 762
 763                 *p_blkno = *p_blkno + 1;
 764         }
 765
 766         /*
 767          * If we issued read requests - let them complete.
 768          */
 769         while(wait_bh > wait) {
 770                 wait_on_buffer(*--wait_bh);
 771                 if (!buffer_uptodate(*wait_bh))
 772                         ret = -EIO;
 773         }
 774
 775         if (ret == 0 || !new)
 776                 return ret;
 777
 778         /*
 779          * If we get -EIO above, zero out any newly allocated blocks
 780          * to avoid exposing stale data.
 781          */
 782         bh = head;
 783         block_start = 0;
 784         do {
 785                 block_end = block_start + bsize;
 786                 if (block_end <= from)
 787                         goto next_bh;
 788                 if (block_start >= to)
 789                         break;
 790
 791                 zero_user(page, block_start, bh->b_size);
 792                 set_buffer_uptodate(bh);
 793                 mark_buffer_dirty(bh);
 794
 795 next_bh:
 796                 block_start = block_end;
 797                 bh = bh->b_this_page;
 798         } while (bh != head);
 799
 800         return ret;
 801 }
 802
 803 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 804 #define OCFS2_MAX_CTXT_PAGES    1
 805 #else
 806 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 807 #endif
 808
 809 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 810
 811 /*
 812  * Describe the state of a single cluster to be written to.
 813  */
 814 struct ocfs2_write_cluster_desc {
 815         u32             c_cpos;
 816         u32             c_phys;
 817         /*
 818          * Give this a unique field because c_phys eventually gets
 819          * filled.
 820          */
 821         unsigned        c_new;
 822         unsigned        c_unwritten;
 823         unsigned        c_needs_zero;
 824 };
 825
 826 struct ocfs2_write_ctxt {
 827         /* Logical cluster position / len of write */
 828         u32                             w_cpos;
 829         u32                             w_clen;
 830
 831         /* First cluster allocated in a nonsparse extend */
 832         u32                             w_first_new_cpos;
 833
 834         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 835
 836         /*
 837          * This is true if page_size > cluster_size.
 838          *
 839          * It triggers a set of special cases during write which might
 840          * have to deal with allocating writes to partial pages.
 841          */
 842         unsigned int                    w_large_pages;
 843
 844         /*
 845          * Pages involved in this write.
 846          *
 847          * w_target_page is the page being written to by the user.
 848          *
 849          * w_pages is an array of pages which always contains
 850          * w_target_page, and in the case of an allocating write with
 851          * page_size < cluster size, it will contain zero'd and mapped
 852          * pages adjacent to w_target_page which need to be written
 853          * out in so that future reads from that region will get
 854          * zero's.
 855          */
 856         unsigned int                    w_num_pages;
 857         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 858         struct page                     *w_target_page;
 859
 860         /*
 861          * w_target_locked is used for page_mkwrite path indicating no unlocking
 862          * against w_target_page in ocfs2_write_end_nolock.
 863          */
 864         unsigned int                    w_target_locked:1;
 865
 866         /*
 867          * ocfs2_write_end() uses this to know what the real range to
 868          * write in the target should be.
 869          */
 870         unsigned int                    w_target_from;
 871         unsigned int                    w_target_to;
 872
 873         /*
 874          * We could use journal_current_handle() but this is cleaner,
 875          * IMHO -Mark
 876          */
 877         handle_t                        *w_handle;
 878
 879         struct buffer_head              *w_di_bh;
 880
 881         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 882 };
 883
 884 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 885 {
 886         int i;
 887
 888         for(i = 0; i < num_pages; i++) {
 889                 if (pages[i]) {
 890                         unlock_page(pages[i]);
 891                         mark_page_accessed(pages[i]);
 892                         page_cache_release(pages[i]);
 893                 }
 894         }
 895 }
 896
 897 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 898 {
 899         int i;
 900
 901         /*
 902          * w_target_locked is only set to true in the page_mkwrite() case.
 903          * The intent is to allow us to lock the target page from write_begin()
 904          * to write_end(). The caller must hold a ref on w_target_page.
 905          */
 906         if (wc->w_target_locked) {
 907                 BUG_ON(!wc->w_target_page);
 908                 for (i = 0; i < wc->w_num_pages; i++) {
 909                         if (wc->w_target_page == wc->w_pages[i]) {
 910                                 wc->w_pages[i] = NULL;
 911                                 break;
 912                         }
 913                 }
 914                 mark_page_accessed(wc->w_target_page);
 915                 page_cache_release(wc->w_target_page);
 916         }
 917         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 918 }
 919
 920 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 921 {
 922         ocfs2_unlock_pages(wc);
 923         brelse(wc->w_di_bh);
 924         kfree(wc);
 925 }
 926
 927 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 928                                   struct ocfs2_super *osb, loff_t pos,
 929                                   unsigned len, struct buffer_head *di_bh)
 930 {
 931         u32 cend;
 932         struct ocfs2_write_ctxt *wc;
 933
 934         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 935         if (!wc)
 936                 return -ENOMEM;
 937
 938         wc->w_cpos = pos >> osb->s_clustersize_bits;
 939         wc->w_first_new_cpos = UINT_MAX;
 940         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 941         wc->w_clen = cend - wc->w_cpos + 1;
 942         get_bh(di_bh);
 943         wc->w_di_bh = di_bh;
 944
 945         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 946                 wc->w_large_pages = 1;
 947         else
 948                 wc->w_large_pages = 0;
 949
 950         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 951
 952         *wcp = wc;
 953
 954         return 0;
 955 }
 956
 957 /*
 958  * If a page has any new buffers, zero them out here, and mark them uptodate
 959  * and dirty so they'll be written out (in order to prevent uninitialised
 960  * block data from leaking). And clear the new bit.
 961  */
 962 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 963 {
 964         unsigned int block_start, block_end;
 965         struct buffer_head *head, *bh;
 966
 967         BUG_ON(!PageLocked(page));
 968         if (!page_has_buffers(page))
 969                 return;
 970
 971         bh = head = page_buffers(page);
 972         block_start = 0;
 973         do {
 974                 block_end = block_start + bh->b_size;
 975
 976                 if (buffer_new(bh)) {
 977                         if (block_end > from && block_start < to) {
 978                                 if (!PageUptodate(page)) {
 979                                         unsigned start, end;
 980
 981                                         start = max(from, block_start);
 982                                         end = min(to, block_end);
 983
 984                                         zero_user_segment(page, start, end);
 985                                         set_buffer_uptodate(bh);
 986                                 }
 987
 988                                 clear_buffer_new(bh);
 989                                 mark_buffer_dirty(bh);
 990                         }
 991                 }
 992
 993                 block_start = block_end;
 994                 bh = bh->b_this_page;
 995         } while (bh != head);
 996 }
 997
 998 /*
 999  * Only called when we have a failure during allocating write to write
1000  * zero's to the newly allocated region.
1001  */
1002 static void ocfs2_write_failure(struct inode *inode,
1003                                 struct ocfs2_write_ctxt *wc,
1004                                 loff_t user_pos, unsigned user_len)
1005 {
1006         int i;
1007         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1008                 to = user_pos + user_len;
1009         struct page *tmppage;
1010
1011         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1012
1013         for(i = 0; i < wc->w_num_pages; i++) {
1014                 tmppage = wc->w_pages[i];
1015
1016                 if (page_has_buffers(tmppage)) {
1017                         if (ocfs2_should_order_data(inode))
1018                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1019
1020                         block_commit_write(tmppage, from, to);
1021                 }
1022         }
1023 }
1024
1025 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1026                                         struct ocfs2_write_ctxt *wc,
1027                                         struct page *page, u32 cpos,
1028                                         loff_t user_pos, unsigned user_len,
1029                                         int new)
1030 {
1031         int ret;
1032         unsigned int map_from = 0, map_to = 0;
1033         unsigned int cluster_start, cluster_end;
1034         unsigned int user_data_from = 0, user_data_to = 0;
1035
1036         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1037                                         &cluster_start, &cluster_end);
1038
1039         /* treat the write as new if the a hole/lseek spanned across
1040          * the page boundary.
1041          */
1042         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1043                         (page_offset(page) <= user_pos));
1044
1045         if (page == wc->w_target_page) {
1046                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1047                 map_to = map_from + user_len;
1048
1049                 if (new)
1050                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1051                                                     cluster_start, cluster_end,
1052                                                     new);
1053                 else
1054                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1055                                                     map_from, map_to, new);
1056                 if (ret) {
1057                         mlog_errno(ret);
1058                         goto out;
1059                 }
1060
1061                 user_data_from = map_from;
1062                 user_data_to = map_to;
1063                 if (new) {
1064                         map_from = cluster_start;
1065                         map_to = cluster_end;
1066                 }
1067         } else {
1068                 /*
1069                  * If we haven't allocated the new page yet, we
1070                  * shouldn't be writing it out without copying user
1071                  * data. This is likely a math error from the caller.
1072                  */
1073                 BUG_ON(!new);
1074
1075                 map_from = cluster_start;
1076                 map_to = cluster_end;
1077
1078                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1079                                             cluster_start, cluster_end, new);
1080                 if (ret) {
1081                         mlog_errno(ret);
1082                         goto out;
1083                 }
1084         }
1085
1086         /*
1087          * Parts of newly allocated pages need to be zero'd.
1088          *
1089          * Above, we have also rewritten 'to' and 'from' - as far as
1090          * the rest of the function is concerned, the entire cluster
1091          * range inside of a page needs to be written.
1092          *
1093          * We can skip this if the page is up to date - it's already
1094          * been zero'd from being read in as a hole.
1095          */
1096         if (new && !PageUptodate(page))
1097                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1098                                          cpos, user_data_from, user_data_to);
1099
1100         flush_dcache_page(page);
1101
1102 out:
1103         return ret;
1104 }
1105
1106 /*
1107  * This function will only grab one clusters worth of pages.
1108  */
1109 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1110                                       struct ocfs2_write_ctxt *wc,
1111                                       u32 cpos, loff_t user_pos,
1112                                       unsigned user_len, int new,
1113                                       struct page *mmap_page)
1114 {
1115         int ret = 0, i;
1116         unsigned long start, target_index, end_index, index;
1117         struct inode *inode = mapping->host;
1118         loff_t last_byte;
1119
1120         target_index = user_pos >> PAGE_CACHE_SHIFT;
1121
1122         /*
1123          * Figure out how many pages we'll be manipulating here. For
1124          * non allocating write, we just change the one
1125          * page. Otherwise, we'll need a whole clusters worth.  If we're
1126          * writing past i_size, we only need enough pages to cover the
1127          * last page of the write.
1128          */
1129         if (new) {
1130                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1131                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1132                 /*
1133                  * We need the index *past* the last page we could possibly
1134                  * touch.  This is the page past the end of the write or
1135                  * i_size, whichever is greater.
1136                  */
1137                 last_byte = max(user_pos + user_len, i_size_read(inode));
1138                 BUG_ON(last_byte < 1);
1139                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1140                 if ((start + wc->w_num_pages) > end_index)
1141                         wc->w_num_pages = end_index - start;
1142         } else {
1143                 wc->w_num_pages = 1;
1144                 start = target_index;
1145         }
1146
1147         for(i = 0; i < wc->w_num_pages; i++) {
1148                 index = start + i;
1149
1150                 if (index == target_index && mmap_page) {
1151                         /*
1152                          * ocfs2_pagemkwrite() is a little different
1153                          * and wants us to directly use the page
1154                          * passed in.
1155                          */
1156                         lock_page(mmap_page);
1157
1158                         /* Exit and let the caller retry */
1159                         if (mmap_page->mapping != mapping) {
1160                                 WARN_ON(mmap_page->mapping);
1161                                 unlock_page(mmap_page);
1162                                 ret = -EAGAIN;
1163                                 goto out;
1164                         }
1165
1166                         page_cache_get(mmap_page);
1167                         wc->w_pages[i] = mmap_page;
1168                         wc->w_target_locked = true;
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                 wait_for_stable_page(wc->w_pages[i]);
1179
1180                 if (index == target_index)
1181                         wc->w_target_page = wc->w_pages[i];
1182         }
1183 out:
1184         if (ret)
1185                 wc->w_target_locked = false;
1186         return ret;
1187 }
1188
1189 /*
1190  * Prepare a single cluster for write one cluster into the file.
1191  */
1192 static int ocfs2_write_cluster(struct address_space *mapping,
1193                                u32 phys, unsigned int unwritten,
1194                                unsigned int should_zero,
1195                                struct ocfs2_alloc_context *data_ac,
1196                                struct ocfs2_alloc_context *meta_ac,
1197                                struct ocfs2_write_ctxt *wc, u32 cpos,
1198                                loff_t user_pos, unsigned user_len)
1199 {
1200         int ret, i, new;
1201         u64 v_blkno, p_blkno;
1202         struct inode *inode = mapping->host;
1203         struct ocfs2_extent_tree et;
1204
1205         new = phys == 0 ? 1 : 0;
1206         if (new) {
1207                 u32 tmp_pos;
1208
1209                 /*
1210                  * This is safe to call with the page locks - it won't take
1211                  * any additional semaphores or cluster locks.
1212                  */
1213                 tmp_pos = cpos;
1214                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1215                                            &tmp_pos, 1, 0, wc->w_di_bh,
1216                                            wc->w_handle, data_ac,
1217                                            meta_ac, NULL);
1218                 /*
1219                  * This shouldn't happen because we must have already
1220                  * calculated the correct meta data allocation required. The
1221                  * internal tree allocation code should know how to increase
1222                  * transaction credits itself.
1223                  *
1224                  * If need be, we could handle -EAGAIN for a
1225                  * RESTART_TRANS here.
1226                  */
1227                 mlog_bug_on_msg(ret == -EAGAIN,
1228                                 "Inode %llu: EAGAIN return during allocation.\n",
1229                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1230                 if (ret < 0) {
1231                         mlog_errno(ret);
1232                         goto out;
1233                 }
1234         } else if (unwritten) {
1235                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1236                                               wc->w_di_bh);
1237                 ret = ocfs2_mark_extent_written(inode, &et,
1238                                                 wc->w_handle, cpos, 1, phys,
1239                                                 meta_ac, &wc->w_dealloc);
1240                 if (ret < 0) {
1241                         mlog_errno(ret);
1242                         goto out;
1243                 }
1244         }
1245
1246         if (should_zero)
1247                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1248         else
1249                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1250
1251         /*
1252          * The only reason this should fail is due to an inability to
1253          * find the extent added.
1254          */
1255         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1256                                           NULL);
1257         if (ret < 0) {
1258                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1259                             "at logical block %llu",
1260                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1261                             (unsigned long long)v_blkno);
1262                 goto out;
1263         }
1264
1265         BUG_ON(p_blkno == 0);
1266
1267         for(i = 0; i < wc->w_num_pages; i++) {
1268                 int tmpret;
1269
1270                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1271                                                       wc->w_pages[i], cpos,
1272                                                       user_pos, user_len,
1273                                                       should_zero);
1274                 if (tmpret) {
1275                         mlog_errno(tmpret);
1276                         if (ret == 0)
1277                                 ret = tmpret;
1278                 }
1279         }
1280
1281         /*
1282          * We only have cleanup to do in case of allocating write.
1283          */
1284         if (ret && new)
1285                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1286
1287 out:
1288
1289         return ret;
1290 }
1291
1292 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1293                                        struct ocfs2_alloc_context *data_ac,
1294                                        struct ocfs2_alloc_context *meta_ac,
1295                                        struct ocfs2_write_ctxt *wc,
1296                                        loff_t pos, unsigned len)
1297 {
1298         int ret, i;
1299         loff_t cluster_off;
1300         unsigned int local_len = len;
1301         struct ocfs2_write_cluster_desc *desc;
1302         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1303
1304         for (i = 0; i < wc->w_clen; i++) {
1305                 desc = &wc->w_desc[i];
1306
1307                 /*
1308                  * We have to make sure that the total write passed in
1309                  * doesn't extend past a single cluster.
1310                  */
1311                 local_len = len;
1312                 cluster_off = pos & (osb->s_clustersize - 1);
1313                 if ((cluster_off + local_len) > osb->s_clustersize)
1314                         local_len = osb->s_clustersize - cluster_off;
1315
1316                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1317                                           desc->c_unwritten,
1318                                           desc->c_needs_zero,
1319                                           data_ac, meta_ac,
1320                                           wc, desc->c_cpos, pos, local_len);
1321                 if (ret) {
1322                         mlog_errno(ret);
1323                         goto out;
1324                 }
1325
1326                 len -= local_len;
1327                 pos += local_len;
1328         }
1329
1330         ret = 0;
1331 out:
1332         return ret;
1333 }
1334
1335 /*
1336  * ocfs2_write_end() wants to know which parts of the target page it
1337  * should complete the write on. It's easiest to compute them ahead of
1338  * time when a more complete view of the write is available.
1339  */
1340 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1341                                         struct ocfs2_write_ctxt *wc,
1342                                         loff_t pos, unsigned len, int alloc)
1343 {
1344         struct ocfs2_write_cluster_desc *desc;
1345
1346         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1347         wc->w_target_to = wc->w_target_from + len;
1348
1349         if (alloc == 0)
1350                 return;
1351
1352         /*
1353          * Allocating write - we may have different boundaries based
1354          * on page size and cluster size.
1355          *
1356          * NOTE: We can no longer compute one value from the other as
1357          * the actual write length and user provided length may be
1358          * different.
1359          */
1360
1361         if (wc->w_large_pages) {
1362                 /*
1363                  * We only care about the 1st and last cluster within
1364                  * our range and whether they should be zero'd or not. Either
1365                  * value may be extended out to the start/end of a
1366                  * newly allocated cluster.
1367                  */
1368                 desc = &wc->w_desc[0];
1369                 if (desc->c_needs_zero)
1370                         ocfs2_figure_cluster_boundaries(osb,
1371                                                         desc->c_cpos,
1372                                                         &wc->w_target_from,
1373                                                         NULL);
1374
1375                 desc = &wc->w_desc[wc->w_clen - 1];
1376                 if (desc->c_needs_zero)
1377                         ocfs2_figure_cluster_boundaries(osb,
1378                                                         desc->c_cpos,
1379                                                         NULL,
1380                                                         &wc->w_target_to);
1381         } else {
1382                 wc->w_target_from = 0;
1383                 wc->w_target_to = PAGE_CACHE_SIZE;
1384         }
1385 }
1386
1387 /*
1388  * Populate each single-cluster write descriptor in the write context
1389  * with information about the i/o to be done.
1390  *
1391  * Returns the number of clusters that will have to be allocated, as
1392  * well as a worst case estimate of the number of extent records that
1393  * would have to be created during a write to an unwritten region.
1394  */
1395 static int ocfs2_populate_write_desc(struct inode *inode,
1396                                      struct ocfs2_write_ctxt *wc,
1397                                      unsigned int *clusters_to_alloc,
1398                                      unsigned int *extents_to_split)
1399 {
1400         int ret;
1401         struct ocfs2_write_cluster_desc *desc;
1402         unsigned int num_clusters = 0;
1403         unsigned int ext_flags = 0;
1404         u32 phys = 0;
1405         int i;
1406
1407         *clusters_to_alloc = 0;
1408         *extents_to_split = 0;
1409
1410         for (i = 0; i < wc->w_clen; i++) {
1411                 desc = &wc->w_desc[i];
1412                 desc->c_cpos = wc->w_cpos + i;
1413
1414                 if (num_clusters == 0) {
1415                         /*
1416                          * Need to look up the next extent record.
1417                          */
1418                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1419                                                  &num_clusters, &ext_flags);
1420                         if (ret) {
1421                                 mlog_errno(ret);
1422                                 goto out;
1423                         }
1424
1425                         /* We should already CoW the refcountd extent. */
1426                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1427
1428                         /*
1429                          * Assume worst case - that we're writing in
1430                          * the middle of the extent.
1431                          *
1432                          * We can assume that the write proceeds from
1433                          * left to right, in which case the extent
1434                          * insert code is smart enough to coalesce the
1435                          * next splits into the previous records created.
1436                          */
1437                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1438                                 *extents_to_split = *extents_to_split + 2;
1439                 } else if (phys) {
1440                         /*
1441                          * Only increment phys if it doesn't describe
1442                          * a hole.
1443                          */
1444                         phys++;
1445                 }
1446
1447                 /*
1448                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1449                  * file that got extended.  w_first_new_cpos tells us
1450                  * where the newly allocated clusters are so we can
1451                  * zero them.
1452                  */
1453                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1454                         BUG_ON(phys == 0);
1455                         desc->c_needs_zero = 1;
1456                 }
1457
1458                 desc->c_phys = phys;
1459                 if (phys == 0) {
1460                         desc->c_new = 1;
1461                         desc->c_needs_zero = 1;
1462                         *clusters_to_alloc = *clusters_to_alloc + 1;
1463                 }
1464
1465                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1466                         desc->c_unwritten = 1;
1467                         desc->c_needs_zero = 1;
1468                 }
1469
1470                 num_clusters--;
1471         }
1472
1473         ret = 0;
1474 out:
1475         return ret;
1476 }
1477
1478 static int ocfs2_write_begin_inline(struct address_space *mapping,
1479                                     struct inode *inode,
1480                                     struct ocfs2_write_ctxt *wc)
1481 {
1482         int ret;
1483         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1484         struct page *page;
1485         handle_t *handle;
1486         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1487
1488         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1489         if (IS_ERR(handle)) {
1490                 ret = PTR_ERR(handle);
1491                 mlog_errno(ret);
1492                 goto out;
1493         }
1494
1495         page = find_or_create_page(mapping, 0, GFP_NOFS);
1496         if (!page) {
1497                 ocfs2_commit_trans(osb, handle);
1498                 ret = -ENOMEM;
1499                 mlog_errno(ret);
1500                 goto out;
1501         }
1502         /*
1503          * If we don't set w_num_pages then this page won't get unlocked
1504          * and freed on cleanup of the write context.
1505          */
1506         wc->w_pages[0] = wc->w_target_page = page;
1507         wc->w_num_pages = 1;
1508
1509         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1510                                       OCFS2_JOURNAL_ACCESS_WRITE);
1511         if (ret) {
1512                 ocfs2_commit_trans(osb, handle);
1513
1514                 mlog_errno(ret);
1515                 goto out;
1516         }
1517
1518         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1519                 ocfs2_set_inode_data_inline(inode, di);
1520
1521         if (!PageUptodate(page)) {
1522                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1523                 if (ret) {
1524                         ocfs2_commit_trans(osb, handle);
1525
1526                         goto out;
1527                 }
1528         }
1529
1530         wc->w_handle = handle;
1531 out:
1532         return ret;
1533 }
1534
1535 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1536 {
1537         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1538
1539         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1540                 return 1;
1541         return 0;
1542 }
1543
1544 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1545                                           struct inode *inode, loff_t pos,
1546                                           unsigned len, struct page *mmap_page,
1547                                           struct ocfs2_write_ctxt *wc)
1548 {
1549         int ret, written = 0;
1550         loff_t end = pos + len;
1551         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1552         struct ocfs2_dinode *di = NULL;
1553
1554         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1555                                              len, (unsigned long long)pos,
1556                                              oi->ip_dyn_features);
1557
1558         /*
1559          * Handle inodes which already have inline data 1st.
1560          */
1561         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1562                 if (mmap_page == NULL &&
1563                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1564                         goto do_inline_write;
1565
1566                 /*
1567                  * The write won't fit - we have to give this inode an
1568                  * inline extent list now.
1569                  */
1570                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1571                 if (ret)
1572                         mlog_errno(ret);
1573                 goto out;
1574         }
1575
1576         /*
1577          * Check whether the inode can accept inline data.
1578          */
1579         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1580                 return 0;
1581
1582         /*
1583          * Check whether the write can fit.
1584          */
1585         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1586         if (mmap_page ||
1587             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1588                 return 0;
1589
1590 do_inline_write:
1591         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1592         if (ret) {
1593                 mlog_errno(ret);
1594                 goto out;
1595         }
1596
1597         /*
1598          * This signals to the caller that the data can be written
1599          * inline.
1600          */
1601         written = 1;
1602 out:
1603         return written ? written : ret;
1604 }
1605
1606 /*
1607  * This function only does anything for file systems which can't
1608  * handle sparse files.
1609  *
1610  * What we want to do here is fill in any hole between the current end
1611  * of allocation and the end of our write. That way the rest of the
1612  * write path can treat it as an non-allocating write, which has no
1613  * special case code for sparse/nonsparse files.
1614  */
1615 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1616                                         struct buffer_head *di_bh,
1617                                         loff_t pos, unsigned len,
1618                                         struct ocfs2_write_ctxt *wc)
1619 {
1620         int ret;
1621         loff_t newsize = pos + len;
1622
1623         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1624
1625         if (newsize <= i_size_read(inode))
1626                 return 0;
1627
1628         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1629         if (ret)
1630                 mlog_errno(ret);
1631
1632         wc->w_first_new_cpos =
1633                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1634
1635         return ret;
1636 }
1637
1638 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1639                            loff_t pos)
1640 {
1641         int ret = 0;
1642
1643         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1644         if (pos > i_size_read(inode))
1645                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1646
1647         return ret;
1648 }
1649
1650 /*
1651  * Try to flush truncate logs if we can free enough clusters from it.
1652  * As for return value, "< 0" means error, "0" no space and "1" means
1653  * we have freed enough spaces and let the caller try to allocate again.
1654  */
1655 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1656                                           unsigned int needed)
1657 {
1658         tid_t target;
1659         int ret = 0;
1660         unsigned int truncated_clusters;
1661
1662         mutex_lock(&osb->osb_tl_inode->i_mutex);
1663         truncated_clusters = osb->truncated_clusters;
1664         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1665
1666         /*
1667          * Check whether we can succeed in allocating if we free
1668          * the truncate log.
1669          */
1670         if (truncated_clusters < needed)
1671                 goto out;
1672
1673         ret = ocfs2_flush_truncate_log(osb);
1674         if (ret) {
1675                 mlog_errno(ret);
1676                 goto out;
1677         }
1678
1679         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1680                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1681                 ret = 1;
1682         }
1683 out:
1684         return ret;
1685 }
1686
1687 int ocfs2_write_begin_nolock(struct file *filp,
1688                              struct address_space *mapping,
1689                              loff_t pos, unsigned len, unsigned flags,
1690                              struct page **pagep, void **fsdata,
1691                              struct buffer_head *di_bh, struct page *mmap_page)
1692 {
1693         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1694         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1695         struct ocfs2_write_ctxt *wc;
1696         struct inode *inode = mapping->host;
1697         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1698         struct ocfs2_dinode *di;
1699         struct ocfs2_alloc_context *data_ac = NULL;
1700         struct ocfs2_alloc_context *meta_ac = NULL;
1701         handle_t *handle;
1702         struct ocfs2_extent_tree et;
1703         int try_free = 1, ret1;
1704
1705 try_again:
1706         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1707         if (ret) {
1708                 mlog_errno(ret);
1709                 return ret;
1710         }
1711
1712         if (ocfs2_supports_inline_data(osb)) {
1713                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1714                                                      mmap_page, wc);
1715                 if (ret == 1) {
1716                         ret = 0;
1717                         goto success;
1718                 }
1719                 if (ret < 0) {
1720                         mlog_errno(ret);
1721                         goto out;
1722                 }
1723         }
1724
1725         if (ocfs2_sparse_alloc(osb))
1726                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1727         else
1728                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1729                                                    wc);
1730         if (ret) {
1731                 mlog_errno(ret);
1732                 goto out;
1733         }
1734
1735         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1736         if (ret < 0) {
1737                 mlog_errno(ret);
1738                 goto out;
1739         } else if (ret == 1) {
1740                 clusters_need = wc->w_clen;
1741                 ret = ocfs2_refcount_cow(inode, di_bh,
1742                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1743                 if (ret) {
1744                         mlog_errno(ret);
1745                         goto out;
1746                 }
1747         }
1748
1749         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1750                                         &extents_to_split);
1751         if (ret) {
1752                 mlog_errno(ret);
1753                 goto out;
1754         }
1755         clusters_need += clusters_to_alloc;
1756
1757         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1758
1759         trace_ocfs2_write_begin_nolock(
1760                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1761                         (long long)i_size_read(inode),
1762                         le32_to_cpu(di->i_clusters),
1763                         pos, len, flags, mmap_page,
1764                         clusters_to_alloc, extents_to_split);
1765
1766         /*
1767          * We set w_target_from, w_target_to here so that
1768          * ocfs2_write_end() knows which range in the target page to
1769          * write out. An allocation requires that we write the entire
1770          * cluster range.
1771          */
1772         if (clusters_to_alloc || extents_to_split) {
1773                 /*
1774                  * XXX: We are stretching the limits of
1775                  * ocfs2_lock_allocators(). It greatly over-estimates
1776                  * the work to be done.
1777                  */
1778                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1779                                               wc->w_di_bh);
1780                 ret = ocfs2_lock_allocators(inode, &et,
1781                                             clusters_to_alloc, extents_to_split,
1782                                             &data_ac, &meta_ac);
1783                 if (ret) {
1784                         mlog_errno(ret);
1785                         goto out;
1786                 }
1787
1788                 if (data_ac)
1789                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1790
1791                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1792                                                     &di->id2.i_list);
1793
1794         }
1795
1796         /*
1797          * We have to zero sparse allocated clusters, unwritten extent clusters,
1798          * and non-sparse clusters we just extended.  For non-sparse writes,
1799          * we know zeros will only be needed in the first and/or last cluster.
1800          */
1801         if (clusters_to_alloc || extents_to_split ||
1802             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1803                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1804                 cluster_of_pages = 1;
1805         else
1806                 cluster_of_pages = 0;
1807
1808         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1809
1810         handle = ocfs2_start_trans(osb, credits);
1811         if (IS_ERR(handle)) {
1812                 ret = PTR_ERR(handle);
1813                 mlog_errno(ret);
1814                 goto out;
1815         }
1816
1817         wc->w_handle = handle;
1818
1819         if (clusters_to_alloc) {
1820                 ret = dquot_alloc_space_nodirty(inode,
1821                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1822                 if (ret)
1823                         goto out_commit;
1824         }
1825         /*
1826          * We don't want this to fail in ocfs2_write_end(), so do it
1827          * here.
1828          */
1829         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1830                                       OCFS2_JOURNAL_ACCESS_WRITE);
1831         if (ret) {
1832                 mlog_errno(ret);
1833                 goto out_quota;
1834         }
1835
1836         /*
1837          * Fill our page array first. That way we've grabbed enough so
1838          * that we can zero and flush if we error after adding the
1839          * extent.
1840          */
1841         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1842                                          cluster_of_pages, mmap_page);
1843         if (ret && ret != -EAGAIN) {
1844                 mlog_errno(ret);
1845                 goto out_quota;
1846         }
1847
1848         /*
1849          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1850          * the target page. In this case, we exit with no error and no target
1851          * page. This will trigger the caller, page_mkwrite(), to re-try
1852          * the operation.
1853          */
1854         if (ret == -EAGAIN) {
1855                 BUG_ON(wc->w_target_page);
1856                 ret = 0;
1857                 goto out_quota;
1858         }
1859
1860         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1861                                           len);
1862         if (ret) {
1863                 mlog_errno(ret);
1864                 goto out_quota;
1865         }
1866
1867         if (data_ac)
1868                 ocfs2_free_alloc_context(data_ac);
1869         if (meta_ac)
1870                 ocfs2_free_alloc_context(meta_ac);
1871
1872 success:
1873         *pagep = wc->w_target_page;
1874         *fsdata = wc;
1875         return 0;
1876 out_quota:
1877         if (clusters_to_alloc)
1878                 dquot_free_space(inode,
1879                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1880 out_commit:
1881         ocfs2_commit_trans(osb, handle);
1882
1883 out:
1884         ocfs2_free_write_ctxt(wc);
1885
1886         if (data_ac) {
1887                 ocfs2_free_alloc_context(data_ac);
1888                 data_ac = NULL;
1889         }
1890         if (meta_ac) {
1891                 ocfs2_free_alloc_context(meta_ac);
1892                 meta_ac = NULL;
1893         }
1894
1895         if (ret == -ENOSPC && try_free) {
1896                 /*
1897                  * Try to free some truncate log so that we can have enough
1898                  * clusters to allocate.
1899                  */
1900                 try_free = 0;
1901
1902                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1903                 if (ret1 == 1)
1904                         goto try_again;
1905
1906                 if (ret1 < 0)
1907                         mlog_errno(ret1);
1908         }
1909
1910         return ret;
1911 }
1912
1913 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1914                              loff_t pos, unsigned len, unsigned flags,
1915                              struct page **pagep, void **fsdata)
1916 {
1917         int ret;
1918         struct buffer_head *di_bh = NULL;
1919         struct inode *inode = mapping->host;
1920
1921         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1922         if (ret) {
1923                 mlog_errno(ret);
1924                 return ret;
1925         }
1926
1927         /*
1928          * Take alloc sem here to prevent concurrent lookups. That way
1929          * the mapping, zeroing and tree manipulation within
1930          * ocfs2_write() will be safe against ->readpage(). This
1931          * should also serve to lock out allocation from a shared
1932          * writeable region.
1933          */
1934         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1935
1936         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1937                                        fsdata, di_bh, NULL);
1938         if (ret) {
1939                 mlog_errno(ret);
1940                 goto out_fail;
1941         }
1942
1943         brelse(di_bh);
1944
1945         return 0;
1946
1947 out_fail:
1948         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1949
1950         brelse(di_bh);
1951         ocfs2_inode_unlock(inode, 1);
1952
1953         return ret;
1954 }
1955
1956 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1957                                    unsigned len, unsigned *copied,
1958                                    struct ocfs2_dinode *di,
1959                                    struct ocfs2_write_ctxt *wc)
1960 {
1961         void *kaddr;
1962
1963         if (unlikely(*copied < len)) {
1964                 if (!PageUptodate(wc->w_target_page)) {
1965                         *copied = 0;
1966                         return;
1967                 }
1968         }
1969
1970         kaddr = kmap_atomic(wc->w_target_page);
1971         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1972         kunmap_atomic(kaddr);
1973
1974         trace_ocfs2_write_end_inline(
1975              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1976              (unsigned long long)pos, *copied,
1977              le16_to_cpu(di->id2.i_data.id_count),
1978              le16_to_cpu(di->i_dyn_features));
1979 }
1980
1981 int ocfs2_write_end_nolock(struct address_space *mapping,
1982                            loff_t pos, unsigned len, unsigned copied,
1983                            struct page *page, void *fsdata)
1984 {
1985         int i;
1986         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1987         struct inode *inode = mapping->host;
1988         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1989         struct ocfs2_write_ctxt *wc = fsdata;
1990         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1991         handle_t *handle = wc->w_handle;
1992         struct page *tmppage;
1993
1994         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1995                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1996                 goto out_write_size;
1997         }
1998
1999         if (unlikely(copied < len)) {
2000                 if (!PageUptodate(wc->w_target_page))
2001                         copied = 0;
2002
2003                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2004                                        start+len);
2005         }
2006         flush_dcache_page(wc->w_target_page);
2007
2008         for(i = 0; i < wc->w_num_pages; i++) {
2009                 tmppage = wc->w_pages[i];
2010
2011                 if (tmppage == wc->w_target_page) {
2012                         from = wc->w_target_from;
2013                         to = wc->w_target_to;
2014
2015                         BUG_ON(from > PAGE_CACHE_SIZE ||
2016                                to > PAGE_CACHE_SIZE ||
2017                                to < from);
2018                 } else {
2019                         /*
2020                          * Pages adjacent to the target (if any) imply
2021                          * a hole-filling write in which case we want
2022                          * to flush their entire range.
2023                          */
2024                         from = 0;
2025                         to = PAGE_CACHE_SIZE;
2026                 }
2027
2028                 if (page_has_buffers(tmppage)) {
2029                         if (ocfs2_should_order_data(inode))
2030                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2031                         block_commit_write(tmppage, from, to);
2032                 }
2033         }
2034
2035 out_write_size:
2036         pos += copied;
2037         if (pos > i_size_read(inode)) {
2038                 i_size_write(inode, pos);
2039                 mark_inode_dirty(inode);
2040         }
2041         inode->i_blocks = ocfs2_inode_sector_count(inode);
2042         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2043         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2044         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2045         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2046         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2047         ocfs2_journal_dirty(handle, wc->w_di_bh);
2048
2049         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2050          * lock, or it will cause a deadlock since journal commit threads holds
2051          * this lock and will ask for the page lock when flushing the data.
2052          * put it here to preserve the unlock order.
2053          */
2054         ocfs2_unlock_pages(wc);
2055
2056         ocfs2_commit_trans(osb, handle);
2057
2058         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2059
2060         brelse(wc->w_di_bh);
2061         kfree(wc);
2062
2063         return copied;
2064 }
2065
2066 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2067                            loff_t pos, unsigned len, unsigned copied,
2068                            struct page *page, void *fsdata)
2069 {
2070         int ret;
2071         struct inode *inode = mapping->host;
2072
2073         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2074
2075         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2076         ocfs2_inode_unlock(inode, 1);
2077
2078         return ret;
2079 }
2080
2081 const struct address_space_operations ocfs2_aops = {
2082         .readpage               = ocfs2_readpage,
2083         .readpages              = ocfs2_readpages,
2084         .writepage              = ocfs2_writepage,
2085         .write_begin            = ocfs2_write_begin,
2086         .write_end              = ocfs2_write_end,
2087         .bmap                   = ocfs2_bmap,
2088         .direct_IO              = ocfs2_direct_IO,
2089         .invalidatepage         = block_invalidatepage,
2090         .releasepage            = ocfs2_releasepage,
2091         .migratepage            = buffer_migrate_page,
2092         .is_partially_uptodate  = block_is_partially_uptodate,
2093         .error_remove_page      = generic_error_remove_page,
2094 };