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 #define MLOG_MASK_PREFIX ML_FILE_IO
  33 #include <cluster/masklog.h>
  34
  35 #include "ocfs2.h"
  36
  37 #include "alloc.h"
  38 #include "aops.h"
  39 #include "dlmglue.h"
  40 #include "extent_map.h"
  41 #include "file.h"
  42 #include "inode.h"
  43 #include "journal.h"
  44 #include "suballoc.h"
  45 #include "super.h"
  46 #include "symlink.h"
  47
  48 #include "buffer_head_io.h"
  49
  50 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  51                                    struct buffer_head *bh_result, int create)
  52 {
  53         int err = -EIO;
  54         int status;
  55         struct ocfs2_dinode *fe = NULL;
  56         struct buffer_head *bh = NULL;
  57         struct buffer_head *buffer_cache_bh = NULL;
  58         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  59         void *kaddr;
  60
  61         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  62                    (unsigned long long)iblock, bh_result, create);
  63
  64         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  65
  66         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  67                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  68                      (unsigned long long)iblock);
  69                 goto bail;
  70         }
  71
  72         status = ocfs2_read_inode_block(inode, &bh);
  73         if (status < 0) {
  74                 mlog_errno(status);
  75                 goto bail;
  76         }
  77         fe = (struct ocfs2_dinode *) bh->b_data;
  78
  79         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  80                                                     le32_to_cpu(fe->i_clusters))) {
  81                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  82                      "%llu\n", (unsigned long long)iblock);
  83                 goto bail;
  84         }
  85
  86         /* We don't use the page cache to create symlink data, so if
  87          * need be, copy it over from the buffer cache. */
  88         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  89                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  90                             iblock;
  91                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  92                 if (!buffer_cache_bh) {
  93                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  94                         goto bail;
  95                 }
  96
  97                 /* we haven't locked out transactions, so a commit
  98                  * could've happened. Since we've got a reference on
  99                  * the bh, even if it commits while we're doing the
 100                  * copy, the data is still good. */
 101                 if (buffer_jbd(buffer_cache_bh)
 102                     && ocfs2_inode_is_new(inode)) {
 103                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 104                         if (!kaddr) {
 105                                 mlog(ML_ERROR, "couldn't kmap!\n");
 106                                 goto bail;
 107                         }
 108                         memcpy(kaddr + (bh_result->b_size * iblock),
 109                                buffer_cache_bh->b_data,
 110                                bh_result->b_size);
 111                         kunmap_atomic(kaddr, KM_USER0);
 112                         set_buffer_uptodate(bh_result);
 113                 }
 114                 brelse(buffer_cache_bh);
 115         }
 116
 117         map_bh(bh_result, inode->i_sb,
 118                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 119
 120         err = 0;
 121
 122 bail:
 123         brelse(bh);
 124
 125         mlog_exit(err);
 126         return err;
 127 }
 128
 129 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
 130                            struct buffer_head *bh_result, int create)
 131 {
 132         int err = 0;
 133         unsigned int ext_flags;
 134         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 135         u64 p_blkno, count, past_eof;
 136         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 137
 138         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 139                    (unsigned long long)iblock, bh_result, create);
 140
 141         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 142                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 143                      inode, inode->i_ino);
 144
 145         if (S_ISLNK(inode->i_mode)) {
 146                 /* this always does I/O for some reason. */
 147                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 148                 goto bail;
 149         }
 150
 151         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 152                                           &ext_flags);
 153         if (err) {
 154                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 155                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 156                      (unsigned long long)p_blkno);
 157                 goto bail;
 158         }
 159
 160         if (max_blocks < count)
 161                 count = max_blocks;
 162
 163         /*
 164          * ocfs2 never allocates in this function - the only time we
 165          * need to use BH_New is when we're extending i_size on a file
 166          * system which doesn't support holes, in which case BH_New
 167          * allows block_prepare_write() to zero.
 168          *
 169          * If we see this on a sparse file system, then a truncate has
 170          * raced us and removed the cluster. In this case, we clear
 171          * the buffers dirty and uptodate bits and let the buffer code
 172          * ignore it as a hole.
 173          */
 174         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 175                 clear_buffer_dirty(bh_result);
 176                 clear_buffer_uptodate(bh_result);
 177                 goto bail;
 178         }
 179
 180         /* Treat the unwritten extent as a hole for zeroing purposes. */
 181         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 182                 map_bh(bh_result, inode->i_sb, p_blkno);
 183
 184         bh_result->b_size = count << inode->i_blkbits;
 185
 186         if (!ocfs2_sparse_alloc(osb)) {
 187                 if (p_blkno == 0) {
 188                         err = -EIO;
 189                         mlog(ML_ERROR,
 190                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 191                              (unsigned long long)iblock,
 192                              (unsigned long long)p_blkno,
 193                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 194                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 195                         dump_stack();
 196                 }
 197
 198                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 199                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 200                      (unsigned long long)past_eof);
 201
 202                 if (create && (iblock >= past_eof))
 203                         set_buffer_new(bh_result);
 204         }
 205
 206 bail:
 207         if (err < 0)
 208                 err = -EIO;
 209
 210         mlog_exit(err);
 211         return err;
 212 }
 213
 214 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 215                            struct buffer_head *di_bh)
 216 {
 217         void *kaddr;
 218         loff_t size;
 219         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 220
 221         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 222                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 223                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 224                 return -EROFS;
 225         }
 226
 227         size = i_size_read(inode);
 228
 229         if (size > PAGE_CACHE_SIZE ||
 230             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 231                 ocfs2_error(inode->i_sb,
 232                             "Inode %llu has with inline data has bad size: %Lu",
 233                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 234                             (unsigned long long)size);
 235                 return -EROFS;
 236         }
 237
 238         kaddr = kmap_atomic(page, KM_USER0);
 239         if (size)
 240                 memcpy(kaddr, di->id2.i_data.id_data, size);
 241         /* Clear the remaining part of the page */
 242         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 243         flush_dcache_page(page);
 244         kunmap_atomic(kaddr, KM_USER0);
 245
 246         SetPageUptodate(page);
 247
 248         return 0;
 249 }
 250
 251 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 252 {
 253         int ret;
 254         struct buffer_head *di_bh = NULL;
 255
 256         BUG_ON(!PageLocked(page));
 257         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 258
 259         ret = ocfs2_read_inode_block(inode, &di_bh);
 260         if (ret) {
 261                 mlog_errno(ret);
 262                 goto out;
 263         }
 264
 265         ret = ocfs2_read_inline_data(inode, page, di_bh);
 266 out:
 267         unlock_page(page);
 268
 269         brelse(di_bh);
 270         return ret;
 271 }
 272
 273 static int ocfs2_readpage(struct file *file, struct page *page)
 274 {
 275         struct inode *inode = page->mapping->host;
 276         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 277         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 278         int ret, unlock = 1;
 279
 280         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 281
 282         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 283         if (ret != 0) {
 284                 if (ret == AOP_TRUNCATED_PAGE)
 285                         unlock = 0;
 286                 mlog_errno(ret);
 287                 goto out;
 288         }
 289
 290         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 291                 ret = AOP_TRUNCATED_PAGE;
 292                 goto out_inode_unlock;
 293         }
 294
 295         /*
 296          * i_size might have just been updated as we grabed the meta lock.  We
 297          * might now be discovering a truncate that hit on another node.
 298          * block_read_full_page->get_block freaks out if it is asked to read
 299          * beyond the end of a file, so we check here.  Callers
 300          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 301          * and notice that the page they just read isn't needed.
 302          *
 303          * XXX sys_readahead() seems to get that wrong?
 304          */
 305         if (start >= i_size_read(inode)) {
 306                 zero_user(page, 0, PAGE_SIZE);
 307                 SetPageUptodate(page);
 308                 ret = 0;
 309                 goto out_alloc;
 310         }
 311
 312         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 313                 ret = ocfs2_readpage_inline(inode, page);
 314         else
 315                 ret = block_read_full_page(page, ocfs2_get_block);
 316         unlock = 0;
 317
 318 out_alloc:
 319         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 320 out_inode_unlock:
 321         ocfs2_inode_unlock(inode, 0);
 322 out:
 323         if (unlock)
 324                 unlock_page(page);
 325         mlog_exit(ret);
 326         return ret;
 327 }
 328
 329 /*
 330  * This is used only for read-ahead. Failures or difficult to handle
 331  * situations are safe to ignore.
 332  *
 333  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 334  * are quite large (243 extents on 4k blocks), so most inodes don't
 335  * grow out to a tree. If need be, detecting boundary extents could
 336  * trivially be added in a future version of ocfs2_get_block().
 337  */
 338 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 339                            struct list_head *pages, unsigned nr_pages)
 340 {
 341         int ret, err = -EIO;
 342         struct inode *inode = mapping->host;
 343         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 344         loff_t start;
 345         struct page *last;
 346
 347         /*
 348          * Use the nonblocking flag for the dlm code to avoid page
 349          * lock inversion, but don't bother with retrying.
 350          */
 351         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 352         if (ret)
 353                 return err;
 354
 355         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 356                 ocfs2_inode_unlock(inode, 0);
 357                 return err;
 358         }
 359
 360         /*
 361          * Don't bother with inline-data. There isn't anything
 362          * to read-ahead in that case anyway...
 363          */
 364         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 365                 goto out_unlock;
 366
 367         /*
 368          * Check whether a remote node truncated this file - we just
 369          * drop out in that case as it's not worth handling here.
 370          */
 371         last = list_entry(pages->prev, struct page, lru);
 372         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 373         if (start >= i_size_read(inode))
 374                 goto out_unlock;
 375
 376         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 377
 378 out_unlock:
 379         up_read(&oi->ip_alloc_sem);
 380         ocfs2_inode_unlock(inode, 0);
 381
 382         return err;
 383 }
 384
 385 /* Note: Because we don't support holes, our allocation has
 386  * already happened (allocation writes zeros to the file data)
 387  * so we don't have to worry about ordered writes in
 388  * ocfs2_writepage.
 389  *
 390  * ->writepage is called during the process of invalidating the page cache
 391  * during blocked lock processing.  It can't block on any cluster locks
 392  * to during block mapping.  It's relying on the fact that the block
 393  * mapping can't have disappeared under the dirty pages that it is
 394  * being asked to write back.
 395  */
 396 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 397 {
 398         int ret;
 399
 400         mlog_entry("(0x%p)\n", page);
 401
 402         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 403
 404         mlog_exit(ret);
 405
 406         return ret;
 407 }
 408
 409 /*
 410  * This is called from ocfs2_write_zero_page() which has handled it's
 411  * own cluster locking and has ensured allocation exists for those
 412  * blocks to be written.
 413  */
 414 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 415                                unsigned from, unsigned to)
 416 {
 417         int ret;
 418
 419         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 420
 421         return ret;
 422 }
 423
 424 /* Taken from ext3. We don't necessarily need the full blown
 425  * functionality yet, but IMHO it's better to cut and paste the whole
 426  * thing so we can avoid introducing our own bugs (and easily pick up
 427  * their fixes when they happen) --Mark */
 428 int walk_page_buffers(  handle_t *handle,
 429                         struct buffer_head *head,
 430                         unsigned from,
 431                         unsigned to,
 432                         int *partial,
 433                         int (*fn)(      handle_t *handle,
 434                                         struct buffer_head *bh))
 435 {
 436         struct buffer_head *bh;
 437         unsigned block_start, block_end;
 438         unsigned blocksize = head->b_size;
 439         int err, ret = 0;
 440         struct buffer_head *next;
 441
 442         for (   bh = head, block_start = 0;
 443                 ret == 0 && (bh != head || !block_start);
 444                 block_start = block_end, bh = next)
 445         {
 446                 next = bh->b_this_page;
 447                 block_end = block_start + blocksize;
 448                 if (block_end <= from || block_start >= to) {
 449                         if (partial && !buffer_uptodate(bh))
 450                                 *partial = 1;
 451                         continue;
 452                 }
 453                 err = (*fn)(handle, bh);
 454                 if (!ret)
 455                         ret = err;
 456         }
 457         return ret;
 458 }
 459
 460 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 461                                                          struct page *page,
 462                                                          unsigned from,
 463                                                          unsigned to)
 464 {
 465         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 466         handle_t *handle;
 467         int ret = 0;
 468
 469         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 470         if (IS_ERR(handle)) {
 471                 ret = -ENOMEM;
 472                 mlog_errno(ret);
 473                 goto out;
 474         }
 475
 476         if (ocfs2_should_order_data(inode)) {
 477                 ret = ocfs2_jbd2_file_inode(handle, inode);
 478                 if (ret < 0)
 479                         mlog_errno(ret);
 480         }
 481 out:
 482         if (ret) {
 483                 if (!IS_ERR(handle))
 484                         ocfs2_commit_trans(osb, handle);
 485                 handle = ERR_PTR(ret);
 486         }
 487         return handle;
 488 }
 489
 490 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 491 {
 492         sector_t status;
 493         u64 p_blkno = 0;
 494         int err = 0;
 495         struct inode *inode = mapping->host;
 496
 497         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 498
 499         /* We don't need to lock journal system files, since they aren't
 500          * accessed concurrently from multiple nodes.
 501          */
 502         if (!INODE_JOURNAL(inode)) {
 503                 err = ocfs2_inode_lock(inode, NULL, 0);
 504                 if (err) {
 505                         if (err != -ENOENT)
 506                                 mlog_errno(err);
 507                         goto bail;
 508                 }
 509                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 510         }
 511
 512         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 513                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 514                                                   NULL);
 515
 516         if (!INODE_JOURNAL(inode)) {
 517                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 518                 ocfs2_inode_unlock(inode, 0);
 519         }
 520
 521         if (err) {
 522                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 523                      (unsigned long long)block);
 524                 mlog_errno(err);
 525                 goto bail;
 526         }
 527
 528 bail:
 529         status = err ? 0 : p_blkno;
 530
 531         mlog_exit((int)status);
 532
 533         return status;
 534 }
 535
 536 /*
 537  * TODO: Make this into a generic get_blocks function.
 538  *
 539  * From do_direct_io in direct-io.c:
 540  *  "So what we do is to permit the ->get_blocks function to populate
 541  *   bh.b_size with the size of IO which is permitted at this offset and
 542  *   this i_blkbits."
 543  *
 544  * This function is called directly from get_more_blocks in direct-io.c.
 545  *
 546  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 547  *                                      fs_count, map_bh, dio->rw == WRITE);
 548  */
 549 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 550                                      struct buffer_head *bh_result, int create)
 551 {
 552         int ret;
 553         u64 p_blkno, inode_blocks, contig_blocks;
 554         unsigned int ext_flags;
 555         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 556         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 557
 558         /* This function won't even be called if the request isn't all
 559          * nicely aligned and of the right size, so there's no need
 560          * for us to check any of that. */
 561
 562         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 563
 564         /*
 565          * Any write past EOF is not allowed because we'd be extending.
 566          */
 567         if (create && (iblock + max_blocks) > inode_blocks) {
 568                 ret = -EIO;
 569                 goto bail;
 570         }
 571
 572         /* This figures out the size of the next contiguous block, and
 573          * our logical offset */
 574         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 575                                           &contig_blocks, &ext_flags);
 576         if (ret) {
 577                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 578                      (unsigned long long)iblock);
 579                 ret = -EIO;
 580                 goto bail;
 581         }
 582
 583         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
 584                 ocfs2_error(inode->i_sb,
 585                             "Inode %llu has a hole at block %llu\n",
 586                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 587                             (unsigned long long)iblock);
 588                 ret = -EROFS;
 589                 goto bail;
 590         }
 591
 592         /*
 593          * get_more_blocks() expects us to describe a hole by clearing
 594          * the mapped bit on bh_result().
 595          *
 596          * Consider an unwritten extent as a hole.
 597          */
 598         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 599                 map_bh(bh_result, inode->i_sb, p_blkno);
 600         else {
 601                 /*
 602                  * ocfs2_prepare_inode_for_write() should have caught
 603                  * the case where we'd be filling a hole and triggered
 604                  * a buffered write instead.
 605                  */
 606                 if (create) {
 607                         ret = -EIO;
 608                         mlog_errno(ret);
 609                         goto bail;
 610                 }
 611
 612                 clear_buffer_mapped(bh_result);
 613         }
 614
 615         /* make sure we don't map more than max_blocks blocks here as
 616            that's all the kernel will handle at this point. */
 617         if (max_blocks < contig_blocks)
 618                 contig_blocks = max_blocks;
 619         bh_result->b_size = contig_blocks << blocksize_bits;
 620 bail:
 621         return ret;
 622 }
 623
 624 /*
 625  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 626  * particularly interested in the aio/dio case.  Like the core uses
 627  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 628  * truncation on another.
 629  */
 630 static void ocfs2_dio_end_io(struct kiocb *iocb,
 631                              loff_t offset,
 632                              ssize_t bytes,
 633                              void *private)
 634 {
 635         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 636         int level;
 637
 638         /* this io's submitter should not have unlocked this before we could */
 639         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 640
 641         ocfs2_iocb_clear_rw_locked(iocb);
 642
 643         level = ocfs2_iocb_rw_locked_level(iocb);
 644         if (!level)
 645                 up_read(&inode->i_alloc_sem);
 646         ocfs2_rw_unlock(inode, level);
 647 }
 648
 649 /*
 650  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 651  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 652  * do journalled data.
 653  */
 654 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 655 {
 656         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 657
 658         jbd2_journal_invalidatepage(journal, page, offset);
 659 }
 660
 661 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 662 {
 663         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 664
 665         if (!page_has_buffers(page))
 666                 return 0;
 667         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 668 }
 669
 670 static ssize_t ocfs2_direct_IO(int rw,
 671                                struct kiocb *iocb,
 672                                const struct iovec *iov,
 673                                loff_t offset,
 674                                unsigned long nr_segs)
 675 {
 676         struct file *file = iocb->ki_filp;
 677         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 678         int ret;
 679
 680         mlog_entry_void();
 681
 682         /*
 683          * Fallback to buffered I/O if we see an inode without
 684          * extents.
 685          */
 686         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 687                 return 0;
 688
 689         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 690                                             inode->i_sb->s_bdev, iov, offset,
 691                                             nr_segs,
 692                                             ocfs2_direct_IO_get_blocks,
 693                                             ocfs2_dio_end_io);
 694
 695         mlog_exit(ret);
 696         return ret;
 697 }
 698
 699 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 700                                             u32 cpos,
 701                                             unsigned int *start,
 702                                             unsigned int *end)
 703 {
 704         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 705
 706         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 707                 unsigned int cpp;
 708
 709                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 710
 711                 cluster_start = cpos % cpp;
 712                 cluster_start = cluster_start << osb->s_clustersize_bits;
 713
 714                 cluster_end = cluster_start + osb->s_clustersize;
 715         }
 716
 717         BUG_ON(cluster_start > PAGE_SIZE);
 718         BUG_ON(cluster_end > PAGE_SIZE);
 719
 720         if (start)
 721                 *start = cluster_start;
 722         if (end)
 723                 *end = cluster_end;
 724 }
 725
 726 /*
 727  * 'from' and 'to' are the region in the page to avoid zeroing.
 728  *
 729  * If pagesize > clustersize, this function will avoid zeroing outside
 730  * of the cluster boundary.
 731  *
 732  * from == to == 0 is code for "zero the entire cluster region"
 733  */
 734 static void ocfs2_clear_page_regions(struct page *page,
 735                                      struct ocfs2_super *osb, u32 cpos,
 736                                      unsigned from, unsigned to)
 737 {
 738         void *kaddr;
 739         unsigned int cluster_start, cluster_end;
 740
 741         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 742
 743         kaddr = kmap_atomic(page, KM_USER0);
 744
 745         if (from || to) {
 746                 if (from > cluster_start)
 747                         memset(kaddr + cluster_start, 0, from - cluster_start);
 748                 if (to < cluster_end)
 749                         memset(kaddr + to, 0, cluster_end - to);
 750         } else {
 751                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 752         }
 753
 754         kunmap_atomic(kaddr, KM_USER0);
 755 }
 756
 757 /*
 758  * Nonsparse file systems fully allocate before we get to the write
 759  * code. This prevents ocfs2_write() from tagging the write as an
 760  * allocating one, which means ocfs2_map_page_blocks() might try to
 761  * read-in the blocks at the tail of our file. Avoid reading them by
 762  * testing i_size against each block offset.
 763  */
 764 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 765                                  unsigned int block_start)
 766 {
 767         u64 offset = page_offset(page) + block_start;
 768
 769         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 770                 return 1;
 771
 772         if (i_size_read(inode) > offset)
 773                 return 1;
 774
 775         return 0;
 776 }
 777
 778 /*
 779  * Some of this taken from block_prepare_write(). We already have our
 780  * mapping by now though, and the entire write will be allocating or
 781  * it won't, so not much need to use BH_New.
 782  *
 783  * This will also skip zeroing, which is handled externally.
 784  */
 785 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 786                           struct inode *inode, unsigned int from,
 787                           unsigned int to, int new)
 788 {
 789         int ret = 0;
 790         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 791         unsigned int block_end, block_start;
 792         unsigned int bsize = 1 << inode->i_blkbits;
 793
 794         if (!page_has_buffers(page))
 795                 create_empty_buffers(page, bsize, 0);
 796
 797         head = page_buffers(page);
 798         for (bh = head, block_start = 0; bh != head || !block_start;
 799              bh = bh->b_this_page, block_start += bsize) {
 800                 block_end = block_start + bsize;
 801
 802                 clear_buffer_new(bh);
 803
 804                 /*
 805                  * Ignore blocks outside of our i/o range -
 806                  * they may belong to unallocated clusters.
 807                  */
 808                 if (block_start >= to || block_end <= from) {
 809                         if (PageUptodate(page))
 810                                 set_buffer_uptodate(bh);
 811                         continue;
 812                 }
 813
 814                 /*
 815                  * For an allocating write with cluster size >= page
 816                  * size, we always write the entire page.
 817                  */
 818                 if (new)
 819                         set_buffer_new(bh);
 820
 821                 if (!buffer_mapped(bh)) {
 822                         map_bh(bh, inode->i_sb, *p_blkno);
 823                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 824                 }
 825
 826                 if (PageUptodate(page)) {
 827                         if (!buffer_uptodate(bh))
 828                                 set_buffer_uptodate(bh);
 829                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 830                            !buffer_new(bh) &&
 831                            ocfs2_should_read_blk(inode, page, block_start) &&
 832                            (block_start < from || block_end > to)) {
 833                         ll_rw_block(READ, 1, &bh);
 834                         *wait_bh++=bh;
 835                 }
 836
 837                 *p_blkno = *p_blkno + 1;
 838         }
 839
 840         /*
 841          * If we issued read requests - let them complete.
 842          */
 843         while(wait_bh > wait) {
 844                 wait_on_buffer(*--wait_bh);
 845                 if (!buffer_uptodate(*wait_bh))
 846                         ret = -EIO;
 847         }
 848
 849         if (ret == 0 || !new)
 850                 return ret;
 851
 852         /*
 853          * If we get -EIO above, zero out any newly allocated blocks
 854          * to avoid exposing stale data.
 855          */
 856         bh = head;
 857         block_start = 0;
 858         do {
 859                 block_end = block_start + bsize;
 860                 if (block_end <= from)
 861                         goto next_bh;
 862                 if (block_start >= to)
 863                         break;
 864
 865                 zero_user(page, block_start, bh->b_size);
 866                 set_buffer_uptodate(bh);
 867                 mark_buffer_dirty(bh);
 868
 869 next_bh:
 870                 block_start = block_end;
 871                 bh = bh->b_this_page;
 872         } while (bh != head);
 873
 874         return ret;
 875 }
 876
 877 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 878 #define OCFS2_MAX_CTXT_PAGES    1
 879 #else
 880 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 881 #endif
 882
 883 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 884
 885 /*
 886  * Describe the state of a single cluster to be written to.
 887  */
 888 struct ocfs2_write_cluster_desc {
 889         u32             c_cpos;
 890         u32             c_phys;
 891         /*
 892          * Give this a unique field because c_phys eventually gets
 893          * filled.
 894          */
 895         unsigned        c_new;
 896         unsigned        c_unwritten;
 897 };
 898
 899 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 900 {
 901         return d->c_new || d->c_unwritten;
 902 }
 903
 904 struct ocfs2_write_ctxt {
 905         /* Logical cluster position / len of write */
 906         u32                             w_cpos;
 907         u32                             w_clen;
 908
 909         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 910
 911         /*
 912          * This is true if page_size > cluster_size.
 913          *
 914          * It triggers a set of special cases during write which might
 915          * have to deal with allocating writes to partial pages.
 916          */
 917         unsigned int                    w_large_pages;
 918
 919         /*
 920          * Pages involved in this write.
 921          *
 922          * w_target_page is the page being written to by the user.
 923          *
 924          * w_pages is an array of pages which always contains
 925          * w_target_page, and in the case of an allocating write with
 926          * page_size < cluster size, it will contain zero'd and mapped
 927          * pages adjacent to w_target_page which need to be written
 928          * out in so that future reads from that region will get
 929          * zero's.
 930          */
 931         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 932         unsigned int                    w_num_pages;
 933         struct page                     *w_target_page;
 934
 935         /*
 936          * ocfs2_write_end() uses this to know what the real range to
 937          * write in the target should be.
 938          */
 939         unsigned int                    w_target_from;
 940         unsigned int                    w_target_to;
 941
 942         /*
 943          * We could use journal_current_handle() but this is cleaner,
 944          * IMHO -Mark
 945          */
 946         handle_t                        *w_handle;
 947
 948         struct buffer_head              *w_di_bh;
 949
 950         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 951 };
 952
 953 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 954 {
 955         int i;
 956
 957         for(i = 0; i < num_pages; i++) {
 958                 if (pages[i]) {
 959                         unlock_page(pages[i]);
 960                         mark_page_accessed(pages[i]);
 961                         page_cache_release(pages[i]);
 962                 }
 963         }
 964 }
 965
 966 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 967 {
 968         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 969
 970         brelse(wc->w_di_bh);
 971         kfree(wc);
 972 }
 973
 974 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 975                                   struct ocfs2_super *osb, loff_t pos,
 976                                   unsigned len, struct buffer_head *di_bh)
 977 {
 978         u32 cend;
 979         struct ocfs2_write_ctxt *wc;
 980
 981         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 982         if (!wc)
 983                 return -ENOMEM;
 984
 985         wc->w_cpos = pos >> osb->s_clustersize_bits;
 986         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 987         wc->w_clen = cend - wc->w_cpos + 1;
 988         get_bh(di_bh);
 989         wc->w_di_bh = di_bh;
 990
 991         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 992                 wc->w_large_pages = 1;
 993         else
 994                 wc->w_large_pages = 0;
 995
 996         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 997
 998         *wcp = wc;
 999
1000         return 0;
1001 }
1002
1003 /*
1004  * If a page has any new buffers, zero them out here, and mark them uptodate
1005  * and dirty so they'll be written out (in order to prevent uninitialised
1006  * block data from leaking). And clear the new bit.
1007  */
1008 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1009 {
1010         unsigned int block_start, block_end;
1011         struct buffer_head *head, *bh;
1012
1013         BUG_ON(!PageLocked(page));
1014         if (!page_has_buffers(page))
1015                 return;
1016
1017         bh = head = page_buffers(page);
1018         block_start = 0;
1019         do {
1020                 block_end = block_start + bh->b_size;
1021
1022                 if (buffer_new(bh)) {
1023                         if (block_end > from && block_start < to) {
1024                                 if (!PageUptodate(page)) {
1025                                         unsigned start, end;
1026
1027                                         start = max(from, block_start);
1028                                         end = min(to, block_end);
1029
1030                                         zero_user_segment(page, start, end);
1031                                         set_buffer_uptodate(bh);
1032                                 }
1033
1034                                 clear_buffer_new(bh);
1035                                 mark_buffer_dirty(bh);
1036                         }
1037                 }
1038
1039                 block_start = block_end;
1040                 bh = bh->b_this_page;
1041         } while (bh != head);
1042 }
1043
1044 /*
1045  * Only called when we have a failure during allocating write to write
1046  * zero's to the newly allocated region.
1047  */
1048 static void ocfs2_write_failure(struct inode *inode,
1049                                 struct ocfs2_write_ctxt *wc,
1050                                 loff_t user_pos, unsigned user_len)
1051 {
1052         int i;
1053         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1054                 to = user_pos + user_len;
1055         struct page *tmppage;
1056
1057         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1058
1059         for(i = 0; i < wc->w_num_pages; i++) {
1060                 tmppage = wc->w_pages[i];
1061
1062                 if (page_has_buffers(tmppage)) {
1063                         if (ocfs2_should_order_data(inode))
1064                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1065
1066                         block_commit_write(tmppage, from, to);
1067                 }
1068         }
1069 }
1070
1071 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1072                                         struct ocfs2_write_ctxt *wc,
1073                                         struct page *page, u32 cpos,
1074                                         loff_t user_pos, unsigned user_len,
1075                                         int new)
1076 {
1077         int ret;
1078         unsigned int map_from = 0, map_to = 0;
1079         unsigned int cluster_start, cluster_end;
1080         unsigned int user_data_from = 0, user_data_to = 0;
1081
1082         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1083                                         &cluster_start, &cluster_end);
1084
1085         if (page == wc->w_target_page) {
1086                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1087                 map_to = map_from + user_len;
1088
1089                 if (new)
1090                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1091                                                     cluster_start, cluster_end,
1092                                                     new);
1093                 else
1094                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1095                                                     map_from, map_to, new);
1096                 if (ret) {
1097                         mlog_errno(ret);
1098                         goto out;
1099                 }
1100
1101                 user_data_from = map_from;
1102                 user_data_to = map_to;
1103                 if (new) {
1104                         map_from = cluster_start;
1105                         map_to = cluster_end;
1106                 }
1107         } else {
1108                 /*
1109                  * If we haven't allocated the new page yet, we
1110                  * shouldn't be writing it out without copying user
1111                  * data. This is likely a math error from the caller.
1112                  */
1113                 BUG_ON(!new);
1114
1115                 map_from = cluster_start;
1116                 map_to = cluster_end;
1117
1118                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1119                                             cluster_start, cluster_end, new);
1120                 if (ret) {
1121                         mlog_errno(ret);
1122                         goto out;
1123                 }
1124         }
1125
1126         /*
1127          * Parts of newly allocated pages need to be zero'd.
1128          *
1129          * Above, we have also rewritten 'to' and 'from' - as far as
1130          * the rest of the function is concerned, the entire cluster
1131          * range inside of a page needs to be written.
1132          *
1133          * We can skip this if the page is up to date - it's already
1134          * been zero'd from being read in as a hole.
1135          */
1136         if (new && !PageUptodate(page))
1137                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1138                                          cpos, user_data_from, user_data_to);
1139
1140         flush_dcache_page(page);
1141
1142 out:
1143         return ret;
1144 }
1145
1146 /*
1147  * This function will only grab one clusters worth of pages.
1148  */
1149 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1150                                       struct ocfs2_write_ctxt *wc,
1151                                       u32 cpos, loff_t user_pos, int new,
1152                                       struct page *mmap_page)
1153 {
1154         int ret = 0, i;
1155         unsigned long start, target_index, index;
1156         struct inode *inode = mapping->host;
1157
1158         target_index = user_pos >> PAGE_CACHE_SHIFT;
1159
1160         /*
1161          * Figure out how many pages we'll be manipulating here. For
1162          * non allocating write, we just change the one
1163          * page. Otherwise, we'll need a whole clusters worth.
1164          */
1165         if (new) {
1166                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1167                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1168         } else {
1169                 wc->w_num_pages = 1;
1170                 start = target_index;
1171         }
1172
1173         for(i = 0; i < wc->w_num_pages; i++) {
1174                 index = start + i;
1175
1176                 if (index == target_index && mmap_page) {
1177                         /*
1178                          * ocfs2_pagemkwrite() is a little different
1179                          * and wants us to directly use the page
1180                          * passed in.
1181                          */
1182                         lock_page(mmap_page);
1183
1184                         if (mmap_page->mapping != mapping) {
1185                                 unlock_page(mmap_page);
1186                                 /*
1187                                  * Sanity check - the locking in
1188                                  * ocfs2_pagemkwrite() should ensure
1189                                  * that this code doesn't trigger.
1190                                  */
1191                                 ret = -EINVAL;
1192                                 mlog_errno(ret);
1193                                 goto out;
1194                         }
1195
1196                         page_cache_get(mmap_page);
1197                         wc->w_pages[i] = mmap_page;
1198                 } else {
1199                         wc->w_pages[i] = find_or_create_page(mapping, index,
1200                                                              GFP_NOFS);
1201                         if (!wc->w_pages[i]) {
1202                                 ret = -ENOMEM;
1203                                 mlog_errno(ret);
1204                                 goto out;
1205                         }
1206                 }
1207
1208                 if (index == target_index)
1209                         wc->w_target_page = wc->w_pages[i];
1210         }
1211 out:
1212         return ret;
1213 }
1214
1215 /*
1216  * Prepare a single cluster for write one cluster into the file.
1217  */
1218 static int ocfs2_write_cluster(struct address_space *mapping,
1219                                u32 phys, unsigned int unwritten,
1220                                struct ocfs2_alloc_context *data_ac,
1221                                struct ocfs2_alloc_context *meta_ac,
1222                                struct ocfs2_write_ctxt *wc, u32 cpos,
1223                                loff_t user_pos, unsigned user_len)
1224 {
1225         int ret, i, new, should_zero = 0;
1226         u64 v_blkno, p_blkno;
1227         struct inode *inode = mapping->host;
1228         struct ocfs2_extent_tree et;
1229
1230         new = phys == 0 ? 1 : 0;
1231         if (new || unwritten)
1232                 should_zero = 1;
1233
1234         if (new) {
1235                 u32 tmp_pos;
1236
1237                 /*
1238                  * This is safe to call with the page locks - it won't take
1239                  * any additional semaphores or cluster locks.
1240                  */
1241                 tmp_pos = cpos;
1242                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1243                                            &tmp_pos, 1, 0, wc->w_di_bh,
1244                                            wc->w_handle, data_ac,
1245                                            meta_ac, NULL);
1246                 /*
1247                  * This shouldn't happen because we must have already
1248                  * calculated the correct meta data allocation required. The
1249                  * internal tree allocation code should know how to increase
1250                  * transaction credits itself.
1251                  *
1252                  * If need be, we could handle -EAGAIN for a
1253                  * RESTART_TRANS here.
1254                  */
1255                 mlog_bug_on_msg(ret == -EAGAIN,
1256                                 "Inode %llu: EAGAIN return during allocation.\n",
1257                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1258                 if (ret < 0) {
1259                         mlog_errno(ret);
1260                         goto out;
1261                 }
1262         } else if (unwritten) {
1263                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1264                 ret = ocfs2_mark_extent_written(inode, &et,
1265                                                 wc->w_handle, cpos, 1, phys,
1266                                                 meta_ac, &wc->w_dealloc);
1267                 if (ret < 0) {
1268                         mlog_errno(ret);
1269                         goto out;
1270                 }
1271         }
1272
1273         if (should_zero)
1274                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1275         else
1276                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1277
1278         /*
1279          * The only reason this should fail is due to an inability to
1280          * find the extent added.
1281          */
1282         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1283                                           NULL);
1284         if (ret < 0) {
1285                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1286                             "at logical block %llu",
1287                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1288                             (unsigned long long)v_blkno);
1289                 goto out;
1290         }
1291
1292         BUG_ON(p_blkno == 0);
1293
1294         for(i = 0; i < wc->w_num_pages; i++) {
1295                 int tmpret;
1296
1297                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1298                                                       wc->w_pages[i], cpos,
1299                                                       user_pos, user_len,
1300                                                       should_zero);
1301                 if (tmpret) {
1302                         mlog_errno(tmpret);
1303                         if (ret == 0)
1304                                 tmpret = ret;
1305                 }
1306         }
1307
1308         /*
1309          * We only have cleanup to do in case of allocating write.
1310          */
1311         if (ret && new)
1312                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1313
1314 out:
1315
1316         return ret;
1317 }
1318
1319 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1320                                        struct ocfs2_alloc_context *data_ac,
1321                                        struct ocfs2_alloc_context *meta_ac,
1322                                        struct ocfs2_write_ctxt *wc,
1323                                        loff_t pos, unsigned len)
1324 {
1325         int ret, i;
1326         loff_t cluster_off;
1327         unsigned int local_len = len;
1328         struct ocfs2_write_cluster_desc *desc;
1329         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1330
1331         for (i = 0; i < wc->w_clen; i++) {
1332                 desc = &wc->w_desc[i];
1333
1334                 /*
1335                  * We have to make sure that the total write passed in
1336                  * doesn't extend past a single cluster.
1337                  */
1338                 local_len = len;
1339                 cluster_off = pos & (osb->s_clustersize - 1);
1340                 if ((cluster_off + local_len) > osb->s_clustersize)
1341                         local_len = osb->s_clustersize - cluster_off;
1342
1343                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1344                                           desc->c_unwritten, data_ac, meta_ac,
1345                                           wc, desc->c_cpos, pos, local_len);
1346                 if (ret) {
1347                         mlog_errno(ret);
1348                         goto out;
1349                 }
1350
1351                 len -= local_len;
1352                 pos += local_len;
1353         }
1354
1355         ret = 0;
1356 out:
1357         return ret;
1358 }
1359
1360 /*
1361  * ocfs2_write_end() wants to know which parts of the target page it
1362  * should complete the write on. It's easiest to compute them ahead of
1363  * time when a more complete view of the write is available.
1364  */
1365 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1366                                         struct ocfs2_write_ctxt *wc,
1367                                         loff_t pos, unsigned len, int alloc)
1368 {
1369         struct ocfs2_write_cluster_desc *desc;
1370
1371         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1372         wc->w_target_to = wc->w_target_from + len;
1373
1374         if (alloc == 0)
1375                 return;
1376
1377         /*
1378          * Allocating write - we may have different boundaries based
1379          * on page size and cluster size.
1380          *
1381          * NOTE: We can no longer compute one value from the other as
1382          * the actual write length and user provided length may be
1383          * different.
1384          */
1385
1386         if (wc->w_large_pages) {
1387                 /*
1388                  * We only care about the 1st and last cluster within
1389                  * our range and whether they should be zero'd or not. Either
1390                  * value may be extended out to the start/end of a
1391                  * newly allocated cluster.
1392                  */
1393                 desc = &wc->w_desc[0];
1394                 if (ocfs2_should_zero_cluster(desc))
1395                         ocfs2_figure_cluster_boundaries(osb,
1396                                                         desc->c_cpos,
1397                                                         &wc->w_target_from,
1398                                                         NULL);
1399
1400                 desc = &wc->w_desc[wc->w_clen - 1];
1401                 if (ocfs2_should_zero_cluster(desc))
1402                         ocfs2_figure_cluster_boundaries(osb,
1403                                                         desc->c_cpos,
1404                                                         NULL,
1405                                                         &wc->w_target_to);
1406         } else {
1407                 wc->w_target_from = 0;
1408                 wc->w_target_to = PAGE_CACHE_SIZE;
1409         }
1410 }
1411
1412 /*
1413  * Populate each single-cluster write descriptor in the write context
1414  * with information about the i/o to be done.
1415  *
1416  * Returns the number of clusters that will have to be allocated, as
1417  * well as a worst case estimate of the number of extent records that
1418  * would have to be created during a write to an unwritten region.
1419  */
1420 static int ocfs2_populate_write_desc(struct inode *inode,
1421                                      struct ocfs2_write_ctxt *wc,
1422                                      unsigned int *clusters_to_alloc,
1423                                      unsigned int *extents_to_split)
1424 {
1425         int ret;
1426         struct ocfs2_write_cluster_desc *desc;
1427         unsigned int num_clusters = 0;
1428         unsigned int ext_flags = 0;
1429         u32 phys = 0;
1430         int i;
1431
1432         *clusters_to_alloc = 0;
1433         *extents_to_split = 0;
1434
1435         for (i = 0; i < wc->w_clen; i++) {
1436                 desc = &wc->w_desc[i];
1437                 desc->c_cpos = wc->w_cpos + i;
1438
1439                 if (num_clusters == 0) {
1440                         /*
1441                          * Need to look up the next extent record.
1442                          */
1443                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1444                                                  &num_clusters, &ext_flags);
1445                         if (ret) {
1446                                 mlog_errno(ret);
1447                                 goto out;
1448                         }
1449
1450                         /*
1451                          * Assume worst case - that we're writing in
1452                          * the middle of the extent.
1453                          *
1454                          * We can assume that the write proceeds from
1455                          * left to right, in which case the extent
1456                          * insert code is smart enough to coalesce the
1457                          * next splits into the previous records created.
1458                          */
1459                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1460                                 *extents_to_split = *extents_to_split + 2;
1461                 } else if (phys) {
1462                         /*
1463                          * Only increment phys if it doesn't describe
1464                          * a hole.
1465                          */
1466                         phys++;
1467                 }
1468
1469                 desc->c_phys = phys;
1470                 if (phys == 0) {
1471                         desc->c_new = 1;
1472                         *clusters_to_alloc = *clusters_to_alloc + 1;
1473                 }
1474                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1475                         desc->c_unwritten = 1;
1476
1477                 num_clusters--;
1478         }
1479
1480         ret = 0;
1481 out:
1482         return ret;
1483 }
1484
1485 static int ocfs2_write_begin_inline(struct address_space *mapping,
1486                                     struct inode *inode,
1487                                     struct ocfs2_write_ctxt *wc)
1488 {
1489         int ret;
1490         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1491         struct page *page;
1492         handle_t *handle;
1493         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1494
1495         page = find_or_create_page(mapping, 0, GFP_NOFS);
1496         if (!page) {
1497                 ret = -ENOMEM;
1498                 mlog_errno(ret);
1499                 goto out;
1500         }
1501         /*
1502          * If we don't set w_num_pages then this page won't get unlocked
1503          * and freed on cleanup of the write context.
1504          */
1505         wc->w_pages[0] = wc->w_target_page = page;
1506         wc->w_num_pages = 1;
1507
1508         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1509         if (IS_ERR(handle)) {
1510                 ret = PTR_ERR(handle);
1511                 mlog_errno(ret);
1512                 goto out;
1513         }
1514
1515         ret = ocfs2_journal_access_di(handle, inode, wc->w_di_bh,
1516                                       OCFS2_JOURNAL_ACCESS_WRITE);
1517         if (ret) {
1518                 ocfs2_commit_trans(osb, handle);
1519
1520                 mlog_errno(ret);
1521                 goto out;
1522         }
1523
1524         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1525                 ocfs2_set_inode_data_inline(inode, di);
1526
1527         if (!PageUptodate(page)) {
1528                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1529                 if (ret) {
1530                         ocfs2_commit_trans(osb, handle);
1531
1532                         goto out;
1533                 }
1534         }
1535
1536         wc->w_handle = handle;
1537 out:
1538         return ret;
1539 }
1540
1541 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1542 {
1543         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1544
1545         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1546                 return 1;
1547         return 0;
1548 }
1549
1550 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1551                                           struct inode *inode, loff_t pos,
1552                                           unsigned len, struct page *mmap_page,
1553                                           struct ocfs2_write_ctxt *wc)
1554 {
1555         int ret, written = 0;
1556         loff_t end = pos + len;
1557         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1558         struct ocfs2_dinode *di = NULL;
1559
1560         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1561              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1562              oi->ip_dyn_features);
1563
1564         /*
1565          * Handle inodes which already have inline data 1st.
1566          */
1567         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1568                 if (mmap_page == NULL &&
1569                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1570                         goto do_inline_write;
1571
1572                 /*
1573                  * The write won't fit - we have to give this inode an
1574                  * inline extent list now.
1575                  */
1576                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1577                 if (ret)
1578                         mlog_errno(ret);
1579                 goto out;
1580         }
1581
1582         /*
1583          * Check whether the inode can accept inline data.
1584          */
1585         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1586                 return 0;
1587
1588         /*
1589          * Check whether the write can fit.
1590          */
1591         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1592         if (mmap_page ||
1593             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1594                 return 0;
1595
1596 do_inline_write:
1597         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1598         if (ret) {
1599                 mlog_errno(ret);
1600                 goto out;
1601         }
1602
1603         /*
1604          * This signals to the caller that the data can be written
1605          * inline.
1606          */
1607         written = 1;
1608 out:
1609         return written ? written : ret;
1610 }
1611
1612 /*
1613  * This function only does anything for file systems which can't
1614  * handle sparse files.
1615  *
1616  * What we want to do here is fill in any hole between the current end
1617  * of allocation and the end of our write. That way the rest of the
1618  * write path can treat it as an non-allocating write, which has no
1619  * special case code for sparse/nonsparse files.
1620  */
1621 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1622                                         unsigned len,
1623                                         struct ocfs2_write_ctxt *wc)
1624 {
1625         int ret;
1626         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1627         loff_t newsize = pos + len;
1628
1629         if (ocfs2_sparse_alloc(osb))
1630                 return 0;
1631
1632         if (newsize <= i_size_read(inode))
1633                 return 0;
1634
1635         ret = ocfs2_extend_no_holes(inode, newsize, newsize - len);
1636         if (ret)
1637                 mlog_errno(ret);
1638
1639         return ret;
1640 }
1641
1642 int ocfs2_write_begin_nolock(struct address_space *mapping,
1643                              loff_t pos, unsigned len, unsigned flags,
1644                              struct page **pagep, void **fsdata,
1645                              struct buffer_head *di_bh, struct page *mmap_page)
1646 {
1647         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1648         unsigned int clusters_to_alloc, extents_to_split;
1649         struct ocfs2_write_ctxt *wc;
1650         struct inode *inode = mapping->host;
1651         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1652         struct ocfs2_dinode *di;
1653         struct ocfs2_alloc_context *data_ac = NULL;
1654         struct ocfs2_alloc_context *meta_ac = NULL;
1655         handle_t *handle;
1656         struct ocfs2_extent_tree et;
1657
1658         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1659         if (ret) {
1660                 mlog_errno(ret);
1661                 return ret;
1662         }
1663
1664         if (ocfs2_supports_inline_data(osb)) {
1665                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1666                                                      mmap_page, wc);
1667                 if (ret == 1) {
1668                         ret = 0;
1669                         goto success;
1670                 }
1671                 if (ret < 0) {
1672                         mlog_errno(ret);
1673                         goto out;
1674                 }
1675         }
1676
1677         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1678         if (ret) {
1679                 mlog_errno(ret);
1680                 goto out;
1681         }
1682
1683         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1684                                         &extents_to_split);
1685         if (ret) {
1686                 mlog_errno(ret);
1687                 goto out;
1688         }
1689
1690         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1691
1692         /*
1693          * We set w_target_from, w_target_to here so that
1694          * ocfs2_write_end() knows which range in the target page to
1695          * write out. An allocation requires that we write the entire
1696          * cluster range.
1697          */
1698         if (clusters_to_alloc || extents_to_split) {
1699                 /*
1700                  * XXX: We are stretching the limits of
1701                  * ocfs2_lock_allocators(). It greatly over-estimates
1702                  * the work to be done.
1703                  */
1704                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1705                      " clusters_to_add = %u, extents_to_split = %u\n",
1706                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1707                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1708                      clusters_to_alloc, extents_to_split);
1709
1710                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1711                 ret = ocfs2_lock_allocators(inode, &et,
1712                                             clusters_to_alloc, extents_to_split,
1713                                             &data_ac, &meta_ac);
1714                 if (ret) {
1715                         mlog_errno(ret);
1716                         goto out;
1717                 }
1718
1719                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1720                                                     &di->id2.i_list,
1721                                                     clusters_to_alloc);
1722
1723         }
1724
1725         ocfs2_set_target_boundaries(osb, wc, pos, len,
1726                                     clusters_to_alloc + extents_to_split);
1727
1728         handle = ocfs2_start_trans(osb, credits);
1729         if (IS_ERR(handle)) {
1730                 ret = PTR_ERR(handle);
1731                 mlog_errno(ret);
1732                 goto out;
1733         }
1734
1735         wc->w_handle = handle;
1736
1737         if (clusters_to_alloc && vfs_dq_alloc_space_nodirty(inode,
1738                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc))) {
1739                 ret = -EDQUOT;
1740                 goto out_commit;
1741         }
1742         /*
1743          * We don't want this to fail in ocfs2_write_end(), so do it
1744          * here.
1745          */
1746         ret = ocfs2_journal_access_di(handle, inode, wc->w_di_bh,
1747                                       OCFS2_JOURNAL_ACCESS_WRITE);
1748         if (ret) {
1749                 mlog_errno(ret);
1750                 goto out_quota;
1751         }
1752
1753         /*
1754          * Fill our page array first. That way we've grabbed enough so
1755          * that we can zero and flush if we error after adding the
1756          * extent.
1757          */
1758         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1759                                          clusters_to_alloc + extents_to_split,
1760                                          mmap_page);
1761         if (ret) {
1762                 mlog_errno(ret);
1763                 goto out_quota;
1764         }
1765
1766         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1767                                           len);
1768         if (ret) {
1769                 mlog_errno(ret);
1770                 goto out_quota;
1771         }
1772
1773         if (data_ac)
1774                 ocfs2_free_alloc_context(data_ac);
1775         if (meta_ac)
1776                 ocfs2_free_alloc_context(meta_ac);
1777
1778 success:
1779         *pagep = wc->w_target_page;
1780         *fsdata = wc;
1781         return 0;
1782 out_quota:
1783         if (clusters_to_alloc)
1784                 vfs_dq_free_space(inode,
1785                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1786 out_commit:
1787         ocfs2_commit_trans(osb, handle);
1788
1789 out:
1790         ocfs2_free_write_ctxt(wc);
1791
1792         if (data_ac)
1793                 ocfs2_free_alloc_context(data_ac);
1794         if (meta_ac)
1795                 ocfs2_free_alloc_context(meta_ac);
1796         return ret;
1797 }
1798
1799 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1800                              loff_t pos, unsigned len, unsigned flags,
1801                              struct page **pagep, void **fsdata)
1802 {
1803         int ret;
1804         struct buffer_head *di_bh = NULL;
1805         struct inode *inode = mapping->host;
1806
1807         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1808         if (ret) {
1809                 mlog_errno(ret);
1810                 return ret;
1811         }
1812
1813         /*
1814          * Take alloc sem here to prevent concurrent lookups. That way
1815          * the mapping, zeroing and tree manipulation within
1816          * ocfs2_write() will be safe against ->readpage(). This
1817          * should also serve to lock out allocation from a shared
1818          * writeable region.
1819          */
1820         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1821
1822         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1823                                        fsdata, di_bh, NULL);
1824         if (ret) {
1825                 mlog_errno(ret);
1826                 goto out_fail;
1827         }
1828
1829         brelse(di_bh);
1830
1831         return 0;
1832
1833 out_fail:
1834         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1835
1836         brelse(di_bh);
1837         ocfs2_inode_unlock(inode, 1);
1838
1839         return ret;
1840 }
1841
1842 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1843                                    unsigned len, unsigned *copied,
1844                                    struct ocfs2_dinode *di,
1845                                    struct ocfs2_write_ctxt *wc)
1846 {
1847         void *kaddr;
1848
1849         if (unlikely(*copied < len)) {
1850                 if (!PageUptodate(wc->w_target_page)) {
1851                         *copied = 0;
1852                         return;
1853                 }
1854         }
1855
1856         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1857         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1858         kunmap_atomic(kaddr, KM_USER0);
1859
1860         mlog(0, "Data written to inode at offset %llu. "
1861              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1862              (unsigned long long)pos, *copied,
1863              le16_to_cpu(di->id2.i_data.id_count),
1864              le16_to_cpu(di->i_dyn_features));
1865 }
1866
1867 int ocfs2_write_end_nolock(struct address_space *mapping,
1868                            loff_t pos, unsigned len, unsigned copied,
1869                            struct page *page, void *fsdata)
1870 {
1871         int i;
1872         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1873         struct inode *inode = mapping->host;
1874         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1875         struct ocfs2_write_ctxt *wc = fsdata;
1876         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1877         handle_t *handle = wc->w_handle;
1878         struct page *tmppage;
1879
1880         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1881                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1882                 goto out_write_size;
1883         }
1884
1885         if (unlikely(copied < len)) {
1886                 if (!PageUptodate(wc->w_target_page))
1887                         copied = 0;
1888
1889                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1890                                        start+len);
1891         }
1892         flush_dcache_page(wc->w_target_page);
1893
1894         for(i = 0; i < wc->w_num_pages; i++) {
1895                 tmppage = wc->w_pages[i];
1896
1897                 if (tmppage == wc->w_target_page) {
1898                         from = wc->w_target_from;
1899                         to = wc->w_target_to;
1900
1901                         BUG_ON(from > PAGE_CACHE_SIZE ||
1902                                to > PAGE_CACHE_SIZE ||
1903                                to < from);
1904                 } else {
1905                         /*
1906                          * Pages adjacent to the target (if any) imply
1907                          * a hole-filling write in which case we want
1908                          * to flush their entire range.
1909                          */
1910                         from = 0;
1911                         to = PAGE_CACHE_SIZE;
1912                 }
1913
1914                 if (page_has_buffers(tmppage)) {
1915                         if (ocfs2_should_order_data(inode))
1916                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1917                         block_commit_write(tmppage, from, to);
1918                 }
1919         }
1920
1921 out_write_size:
1922         pos += copied;
1923         if (pos > inode->i_size) {
1924                 i_size_write(inode, pos);
1925                 mark_inode_dirty(inode);
1926         }
1927         inode->i_blocks = ocfs2_inode_sector_count(inode);
1928         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1929         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1930         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1931         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1932         ocfs2_journal_dirty(handle, wc->w_di_bh);
1933
1934         ocfs2_commit_trans(osb, handle);
1935
1936         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1937
1938         ocfs2_free_write_ctxt(wc);
1939
1940         return copied;
1941 }
1942
1943 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1944                            loff_t pos, unsigned len, unsigned copied,
1945                            struct page *page, void *fsdata)
1946 {
1947         int ret;
1948         struct inode *inode = mapping->host;
1949
1950         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1951
1952         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1953         ocfs2_inode_unlock(inode, 1);
1954
1955         return ret;
1956 }
1957
1958 const struct address_space_operations ocfs2_aops = {
1959         .readpage               = ocfs2_readpage,
1960         .readpages              = ocfs2_readpages,
1961         .writepage              = ocfs2_writepage,
1962         .write_begin            = ocfs2_write_begin,
1963         .write_end              = ocfs2_write_end,
1964         .bmap                   = ocfs2_bmap,
1965         .sync_page              = block_sync_page,
1966         .direct_IO              = ocfs2_direct_IO,
1967         .invalidatepage         = ocfs2_invalidatepage,
1968         .releasepage            = ocfs2_releasepage,
1969         .migratepage            = buffer_migrate_page,
1970         .is_partially_uptodate  = block_is_partially_uptodate,
1971 };