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         wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
 575
 576         /* this io's submitter should not have unlocked this before we could */
 577         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 578
 579         if (ocfs2_iocb_is_sem_locked(iocb))
 580                 ocfs2_iocb_clear_sem_locked(iocb);
 581
 582         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 583                 ocfs2_iocb_clear_unaligned_aio(iocb);
 584
 585                 if (atomic_dec_and_test(&OCFS2_I(inode)->ip_unaligned_aio) &&
 586                     waitqueue_active(wq)) {
 587                         wake_up_all(wq);
 588                 }
 589         }
 590
 591         ocfs2_iocb_clear_rw_locked(iocb);
 592
 593         level = ocfs2_iocb_rw_locked_level(iocb);
 594         ocfs2_rw_unlock(inode, level);
 595 }
 596
 597 /*
 598  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 599  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 600  * do journalled data.
 601  */
 602 static void ocfs2_invalidatepage(struct page *page, unsigned int offset,
 603                                  unsigned int length)
 604 {
 605         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 606
 607         jbd2_journal_invalidatepage(journal, page, offset, length);
 608 }
 609
 610 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 611 {
 612         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 613
 614         if (!page_has_buffers(page))
 615                 return 0;
 616         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 617 }
 618
 619 static ssize_t ocfs2_direct_IO(int rw,
 620                                struct kiocb *iocb,
 621                                const struct iovec *iov,
 622                                loff_t offset,
 623                                unsigned long nr_segs)
 624 {
 625         struct file *file = iocb->ki_filp;
 626         struct inode *inode = file_inode(file)->i_mapping->host;
 627
 628         /*
 629          * Fallback to buffered I/O if we see an inode without
 630          * extents.
 631          */
 632         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 633                 return 0;
 634
 635         /* Fallback to buffered I/O if we are appending. */
 636         if (i_size_read(inode) <= offset)
 637                 return 0;
 638
 639         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 640                                     iov, offset, nr_segs,
 641                                     ocfs2_direct_IO_get_blocks,
 642                                     ocfs2_dio_end_io, NULL, 0);
 643 }
 644
 645 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 646                                             u32 cpos,
 647                                             unsigned int *start,
 648                                             unsigned int *end)
 649 {
 650         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 651
 652         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 653                 unsigned int cpp;
 654
 655                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 656
 657                 cluster_start = cpos % cpp;
 658                 cluster_start = cluster_start << osb->s_clustersize_bits;
 659
 660                 cluster_end = cluster_start + osb->s_clustersize;
 661         }
 662
 663         BUG_ON(cluster_start > PAGE_SIZE);
 664         BUG_ON(cluster_end > PAGE_SIZE);
 665
 666         if (start)
 667                 *start = cluster_start;
 668         if (end)
 669                 *end = cluster_end;
 670 }
 671
 672 /*
 673  * 'from' and 'to' are the region in the page to avoid zeroing.
 674  *
 675  * If pagesize > clustersize, this function will avoid zeroing outside
 676  * of the cluster boundary.
 677  *
 678  * from == to == 0 is code for "zero the entire cluster region"
 679  */
 680 static void ocfs2_clear_page_regions(struct page *page,
 681                                      struct ocfs2_super *osb, u32 cpos,
 682                                      unsigned from, unsigned to)
 683 {
 684         void *kaddr;
 685         unsigned int cluster_start, cluster_end;
 686
 687         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 688
 689         kaddr = kmap_atomic(page);
 690
 691         if (from || to) {
 692                 if (from > cluster_start)
 693                         memset(kaddr + cluster_start, 0, from - cluster_start);
 694                 if (to < cluster_end)
 695                         memset(kaddr + to, 0, cluster_end - to);
 696         } else {
 697                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 698         }
 699
 700         kunmap_atomic(kaddr);
 701 }
 702
 703 /*
 704  * Nonsparse file systems fully allocate before we get to the write
 705  * code. This prevents ocfs2_write() from tagging the write as an
 706  * allocating one, which means ocfs2_map_page_blocks() might try to
 707  * read-in the blocks at the tail of our file. Avoid reading them by
 708  * testing i_size against each block offset.
 709  */
 710 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 711                                  unsigned int block_start)
 712 {
 713         u64 offset = page_offset(page) + block_start;
 714
 715         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 716                 return 1;
 717
 718         if (i_size_read(inode) > offset)
 719                 return 1;
 720
 721         return 0;
 722 }
 723
 724 /*
 725  * Some of this taken from __block_write_begin(). We already have our
 726  * mapping by now though, and the entire write will be allocating or
 727  * it won't, so not much need to use BH_New.
 728  *
 729  * This will also skip zeroing, which is handled externally.
 730  */
 731 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 732                           struct inode *inode, unsigned int from,
 733                           unsigned int to, int new)
 734 {
 735         int ret = 0;
 736         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 737         unsigned int block_end, block_start;
 738         unsigned int bsize = 1 << inode->i_blkbits;
 739
 740         if (!page_has_buffers(page))
 741                 create_empty_buffers(page, bsize, 0);
 742
 743         head = page_buffers(page);
 744         for (bh = head, block_start = 0; bh != head || !block_start;
 745              bh = bh->b_this_page, block_start += bsize) {
 746                 block_end = block_start + bsize;
 747
 748                 clear_buffer_new(bh);
 749
 750                 /*
 751                  * Ignore blocks outside of our i/o range -
 752                  * they may belong to unallocated clusters.
 753                  */
 754                 if (block_start >= to || block_end <= from) {
 755                         if (PageUptodate(page))
 756                                 set_buffer_uptodate(bh);
 757                         continue;
 758                 }
 759
 760                 /*
 761                  * For an allocating write with cluster size >= page
 762                  * size, we always write the entire page.
 763                  */
 764                 if (new)
 765                         set_buffer_new(bh);
 766
 767                 if (!buffer_mapped(bh)) {
 768                         map_bh(bh, inode->i_sb, *p_blkno);
 769                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 770                 }
 771
 772                 if (PageUptodate(page)) {
 773                         if (!buffer_uptodate(bh))
 774                                 set_buffer_uptodate(bh);
 775                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 776                            !buffer_new(bh) &&
 777                            ocfs2_should_read_blk(inode, page, block_start) &&
 778                            (block_start < from || block_end > to)) {
 779                         ll_rw_block(READ, 1, &bh);
 780                         *wait_bh++=bh;
 781                 }
 782
 783                 *p_blkno = *p_blkno + 1;
 784         }
 785
 786         /*
 787          * If we issued read requests - let them complete.
 788          */
 789         while(wait_bh > wait) {
 790                 wait_on_buffer(*--wait_bh);
 791                 if (!buffer_uptodate(*wait_bh))
 792                         ret = -EIO;
 793         }
 794
 795         if (ret == 0 || !new)
 796                 return ret;
 797
 798         /*
 799          * If we get -EIO above, zero out any newly allocated blocks
 800          * to avoid exposing stale data.
 801          */
 802         bh = head;
 803         block_start = 0;
 804         do {
 805                 block_end = block_start + bsize;
 806                 if (block_end <= from)
 807                         goto next_bh;
 808                 if (block_start >= to)
 809                         break;
 810
 811                 zero_user(page, block_start, bh->b_size);
 812                 set_buffer_uptodate(bh);
 813                 mark_buffer_dirty(bh);
 814
 815 next_bh:
 816                 block_start = block_end;
 817                 bh = bh->b_this_page;
 818         } while (bh != head);
 819
 820         return ret;
 821 }
 822
 823 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 824 #define OCFS2_MAX_CTXT_PAGES    1
 825 #else
 826 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 827 #endif
 828
 829 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 830
 831 /*
 832  * Describe the state of a single cluster to be written to.
 833  */
 834 struct ocfs2_write_cluster_desc {
 835         u32             c_cpos;
 836         u32             c_phys;
 837         /*
 838          * Give this a unique field because c_phys eventually gets
 839          * filled.
 840          */
 841         unsigned        c_new;
 842         unsigned        c_unwritten;
 843         unsigned        c_needs_zero;
 844 };
 845
 846 struct ocfs2_write_ctxt {
 847         /* Logical cluster position / len of write */
 848         u32                             w_cpos;
 849         u32                             w_clen;
 850
 851         /* First cluster allocated in a nonsparse extend */
 852         u32                             w_first_new_cpos;
 853
 854         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 855
 856         /*
 857          * This is true if page_size > cluster_size.
 858          *
 859          * It triggers a set of special cases during write which might
 860          * have to deal with allocating writes to partial pages.
 861          */
 862         unsigned int                    w_large_pages;
 863
 864         /*
 865          * Pages involved in this write.
 866          *
 867          * w_target_page is the page being written to by the user.
 868          *
 869          * w_pages is an array of pages which always contains
 870          * w_target_page, and in the case of an allocating write with
 871          * page_size < cluster size, it will contain zero'd and mapped
 872          * pages adjacent to w_target_page which need to be written
 873          * out in so that future reads from that region will get
 874          * zero's.
 875          */
 876         unsigned int                    w_num_pages;
 877         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 878         struct page                     *w_target_page;
 879
 880         /*
 881          * w_target_locked is used for page_mkwrite path indicating no unlocking
 882          * against w_target_page in ocfs2_write_end_nolock.
 883          */
 884         unsigned int                    w_target_locked:1;
 885
 886         /*
 887          * ocfs2_write_end() uses this to know what the real range to
 888          * write in the target should be.
 889          */
 890         unsigned int                    w_target_from;
 891         unsigned int                    w_target_to;
 892
 893         /*
 894          * We could use journal_current_handle() but this is cleaner,
 895          * IMHO -Mark
 896          */
 897         handle_t                        *w_handle;
 898
 899         struct buffer_head              *w_di_bh;
 900
 901         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 902 };
 903
 904 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 905 {
 906         int i;
 907
 908         for(i = 0; i < num_pages; i++) {
 909                 if (pages[i]) {
 910                         unlock_page(pages[i]);
 911                         mark_page_accessed(pages[i]);
 912                         page_cache_release(pages[i]);
 913                 }
 914         }
 915 }
 916
 917 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 918 {
 919         int i;
 920
 921         /*
 922          * w_target_locked is only set to true in the page_mkwrite() case.
 923          * The intent is to allow us to lock the target page from write_begin()
 924          * to write_end(). The caller must hold a ref on w_target_page.
 925          */
 926         if (wc->w_target_locked) {
 927                 BUG_ON(!wc->w_target_page);
 928                 for (i = 0; i < wc->w_num_pages; i++) {
 929                         if (wc->w_target_page == wc->w_pages[i]) {
 930                                 wc->w_pages[i] = NULL;
 931                                 break;
 932                         }
 933                 }
 934                 mark_page_accessed(wc->w_target_page);
 935                 page_cache_release(wc->w_target_page);
 936         }
 937         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 938
 939         brelse(wc->w_di_bh);
 940         kfree(wc);
 941 }
 942
 943 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 944                                   struct ocfs2_super *osb, loff_t pos,
 945                                   unsigned len, struct buffer_head *di_bh)
 946 {
 947         u32 cend;
 948         struct ocfs2_write_ctxt *wc;
 949
 950         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 951         if (!wc)
 952                 return -ENOMEM;
 953
 954         wc->w_cpos = pos >> osb->s_clustersize_bits;
 955         wc->w_first_new_cpos = UINT_MAX;
 956         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 957         wc->w_clen = cend - wc->w_cpos + 1;
 958         get_bh(di_bh);
 959         wc->w_di_bh = di_bh;
 960
 961         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 962                 wc->w_large_pages = 1;
 963         else
 964                 wc->w_large_pages = 0;
 965
 966         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 967
 968         *wcp = wc;
 969
 970         return 0;
 971 }
 972
 973 /*
 974  * If a page has any new buffers, zero them out here, and mark them uptodate
 975  * and dirty so they'll be written out (in order to prevent uninitialised
 976  * block data from leaking). And clear the new bit.
 977  */
 978 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 979 {
 980         unsigned int block_start, block_end;
 981         struct buffer_head *head, *bh;
 982
 983         BUG_ON(!PageLocked(page));
 984         if (!page_has_buffers(page))
 985                 return;
 986
 987         bh = head = page_buffers(page);
 988         block_start = 0;
 989         do {
 990                 block_end = block_start + bh->b_size;
 991
 992                 if (buffer_new(bh)) {
 993                         if (block_end > from && block_start < to) {
 994                                 if (!PageUptodate(page)) {
 995                                         unsigned start, end;
 996
 997                                         start = max(from, block_start);
 998                                         end = min(to, block_end);
 999
1000                                         zero_user_segment(page, start, end);
1001                                         set_buffer_uptodate(bh);
1002                                 }
1003
1004                                 clear_buffer_new(bh);
1005                                 mark_buffer_dirty(bh);
1006                         }
1007                 }
1008
1009                 block_start = block_end;
1010                 bh = bh->b_this_page;
1011         } while (bh != head);
1012 }
1013
1014 /*
1015  * Only called when we have a failure during allocating write to write
1016  * zero's to the newly allocated region.
1017  */
1018 static void ocfs2_write_failure(struct inode *inode,
1019                                 struct ocfs2_write_ctxt *wc,
1020                                 loff_t user_pos, unsigned user_len)
1021 {
1022         int i;
1023         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1024                 to = user_pos + user_len;
1025         struct page *tmppage;
1026
1027         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1028
1029         for(i = 0; i < wc->w_num_pages; i++) {
1030                 tmppage = wc->w_pages[i];
1031
1032                 if (page_has_buffers(tmppage)) {
1033                         if (ocfs2_should_order_data(inode))
1034                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1035
1036                         block_commit_write(tmppage, from, to);
1037                 }
1038         }
1039 }
1040
1041 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1042                                         struct ocfs2_write_ctxt *wc,
1043                                         struct page *page, u32 cpos,
1044                                         loff_t user_pos, unsigned user_len,
1045                                         int new)
1046 {
1047         int ret;
1048         unsigned int map_from = 0, map_to = 0;
1049         unsigned int cluster_start, cluster_end;
1050         unsigned int user_data_from = 0, user_data_to = 0;
1051
1052         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1053                                         &cluster_start, &cluster_end);
1054
1055         /* treat the write as new if the a hole/lseek spanned across
1056          * the page boundary.
1057          */
1058         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1059                         (page_offset(page) <= user_pos));
1060
1061         if (page == wc->w_target_page) {
1062                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1063                 map_to = map_from + user_len;
1064
1065                 if (new)
1066                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1067                                                     cluster_start, cluster_end,
1068                                                     new);
1069                 else
1070                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1071                                                     map_from, map_to, new);
1072                 if (ret) {
1073                         mlog_errno(ret);
1074                         goto out;
1075                 }
1076
1077                 user_data_from = map_from;
1078                 user_data_to = map_to;
1079                 if (new) {
1080                         map_from = cluster_start;
1081                         map_to = cluster_end;
1082                 }
1083         } else {
1084                 /*
1085                  * If we haven't allocated the new page yet, we
1086                  * shouldn't be writing it out without copying user
1087                  * data. This is likely a math error from the caller.
1088                  */
1089                 BUG_ON(!new);
1090
1091                 map_from = cluster_start;
1092                 map_to = cluster_end;
1093
1094                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1095                                             cluster_start, cluster_end, new);
1096                 if (ret) {
1097                         mlog_errno(ret);
1098                         goto out;
1099                 }
1100         }
1101
1102         /*
1103          * Parts of newly allocated pages need to be zero'd.
1104          *
1105          * Above, we have also rewritten 'to' and 'from' - as far as
1106          * the rest of the function is concerned, the entire cluster
1107          * range inside of a page needs to be written.
1108          *
1109          * We can skip this if the page is up to date - it's already
1110          * been zero'd from being read in as a hole.
1111          */
1112         if (new && !PageUptodate(page))
1113                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1114                                          cpos, user_data_from, user_data_to);
1115
1116         flush_dcache_page(page);
1117
1118 out:
1119         return ret;
1120 }
1121
1122 /*
1123  * This function will only grab one clusters worth of pages.
1124  */
1125 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1126                                       struct ocfs2_write_ctxt *wc,
1127                                       u32 cpos, loff_t user_pos,
1128                                       unsigned user_len, int new,
1129                                       struct page *mmap_page)
1130 {
1131         int ret = 0, i;
1132         unsigned long start, target_index, end_index, index;
1133         struct inode *inode = mapping->host;
1134         loff_t last_byte;
1135
1136         target_index = user_pos >> PAGE_CACHE_SHIFT;
1137
1138         /*
1139          * Figure out how many pages we'll be manipulating here. For
1140          * non allocating write, we just change the one
1141          * page. Otherwise, we'll need a whole clusters worth.  If we're
1142          * writing past i_size, we only need enough pages to cover the
1143          * last page of the write.
1144          */
1145         if (new) {
1146                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1147                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1148                 /*
1149                  * We need the index *past* the last page we could possibly
1150                  * touch.  This is the page past the end of the write or
1151                  * i_size, whichever is greater.
1152                  */
1153                 last_byte = max(user_pos + user_len, i_size_read(inode));
1154                 BUG_ON(last_byte < 1);
1155                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1156                 if ((start + wc->w_num_pages) > end_index)
1157                         wc->w_num_pages = end_index - start;
1158         } else {
1159                 wc->w_num_pages = 1;
1160                 start = target_index;
1161         }
1162
1163         for(i = 0; i < wc->w_num_pages; i++) {
1164                 index = start + i;
1165
1166                 if (index == target_index && mmap_page) {
1167                         /*
1168                          * ocfs2_pagemkwrite() is a little different
1169                          * and wants us to directly use the page
1170                          * passed in.
1171                          */
1172                         lock_page(mmap_page);
1173
1174                         /* Exit and let the caller retry */
1175                         if (mmap_page->mapping != mapping) {
1176                                 WARN_ON(mmap_page->mapping);
1177                                 unlock_page(mmap_page);
1178                                 ret = -EAGAIN;
1179                                 goto out;
1180                         }
1181
1182                         page_cache_get(mmap_page);
1183                         wc->w_pages[i] = mmap_page;
1184                         wc->w_target_locked = true;
1185                 } else {
1186                         wc->w_pages[i] = find_or_create_page(mapping, index,
1187                                                              GFP_NOFS);
1188                         if (!wc->w_pages[i]) {
1189                                 ret = -ENOMEM;
1190                                 mlog_errno(ret);
1191                                 goto out;
1192                         }
1193                 }
1194                 wait_for_stable_page(wc->w_pages[i]);
1195
1196                 if (index == target_index)
1197                         wc->w_target_page = wc->w_pages[i];
1198         }
1199 out:
1200         if (ret)
1201                 wc->w_target_locked = false;
1202         return ret;
1203 }
1204
1205 /*
1206  * Prepare a single cluster for write one cluster into the file.
1207  */
1208 static int ocfs2_write_cluster(struct address_space *mapping,
1209                                u32 phys, unsigned int unwritten,
1210                                unsigned int should_zero,
1211                                struct ocfs2_alloc_context *data_ac,
1212                                struct ocfs2_alloc_context *meta_ac,
1213                                struct ocfs2_write_ctxt *wc, u32 cpos,
1214                                loff_t user_pos, unsigned user_len)
1215 {
1216         int ret, i, new;
1217         u64 v_blkno, p_blkno;
1218         struct inode *inode = mapping->host;
1219         struct ocfs2_extent_tree et;
1220
1221         new = phys == 0 ? 1 : 0;
1222         if (new) {
1223                 u32 tmp_pos;
1224
1225                 /*
1226                  * This is safe to call with the page locks - it won't take
1227                  * any additional semaphores or cluster locks.
1228                  */
1229                 tmp_pos = cpos;
1230                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1231                                            &tmp_pos, 1, 0, wc->w_di_bh,
1232                                            wc->w_handle, data_ac,
1233                                            meta_ac, NULL);
1234                 /*
1235                  * This shouldn't happen because we must have already
1236                  * calculated the correct meta data allocation required. The
1237                  * internal tree allocation code should know how to increase
1238                  * transaction credits itself.
1239                  *
1240                  * If need be, we could handle -EAGAIN for a
1241                  * RESTART_TRANS here.
1242                  */
1243                 mlog_bug_on_msg(ret == -EAGAIN,
1244                                 "Inode %llu: EAGAIN return during allocation.\n",
1245                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1246                 if (ret < 0) {
1247                         mlog_errno(ret);
1248                         goto out;
1249                 }
1250         } else if (unwritten) {
1251                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1252                                               wc->w_di_bh);
1253                 ret = ocfs2_mark_extent_written(inode, &et,
1254                                                 wc->w_handle, cpos, 1, phys,
1255                                                 meta_ac, &wc->w_dealloc);
1256                 if (ret < 0) {
1257                         mlog_errno(ret);
1258                         goto out;
1259                 }
1260         }
1261
1262         if (should_zero)
1263                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1264         else
1265                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1266
1267         /*
1268          * The only reason this should fail is due to an inability to
1269          * find the extent added.
1270          */
1271         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1272                                           NULL);
1273         if (ret < 0) {
1274                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1275                             "at logical block %llu",
1276                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1277                             (unsigned long long)v_blkno);
1278                 goto out;
1279         }
1280
1281         BUG_ON(p_blkno == 0);
1282
1283         for(i = 0; i < wc->w_num_pages; i++) {
1284                 int tmpret;
1285
1286                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1287                                                       wc->w_pages[i], cpos,
1288                                                       user_pos, user_len,
1289                                                       should_zero);
1290                 if (tmpret) {
1291                         mlog_errno(tmpret);
1292                         if (ret == 0)
1293                                 ret = tmpret;
1294                 }
1295         }
1296
1297         /*
1298          * We only have cleanup to do in case of allocating write.
1299          */
1300         if (ret && new)
1301                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1302
1303 out:
1304
1305         return ret;
1306 }
1307
1308 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1309                                        struct ocfs2_alloc_context *data_ac,
1310                                        struct ocfs2_alloc_context *meta_ac,
1311                                        struct ocfs2_write_ctxt *wc,
1312                                        loff_t pos, unsigned len)
1313 {
1314         int ret, i;
1315         loff_t cluster_off;
1316         unsigned int local_len = len;
1317         struct ocfs2_write_cluster_desc *desc;
1318         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1319
1320         for (i = 0; i < wc->w_clen; i++) {
1321                 desc = &wc->w_desc[i];
1322
1323                 /*
1324                  * We have to make sure that the total write passed in
1325                  * doesn't extend past a single cluster.
1326                  */
1327                 local_len = len;
1328                 cluster_off = pos & (osb->s_clustersize - 1);
1329                 if ((cluster_off + local_len) > osb->s_clustersize)
1330                         local_len = osb->s_clustersize - cluster_off;
1331
1332                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1333                                           desc->c_unwritten,
1334                                           desc->c_needs_zero,
1335                                           data_ac, meta_ac,
1336                                           wc, desc->c_cpos, pos, local_len);
1337                 if (ret) {
1338                         mlog_errno(ret);
1339                         goto out;
1340                 }
1341
1342                 len -= local_len;
1343                 pos += local_len;
1344         }
1345
1346         ret = 0;
1347 out:
1348         return ret;
1349 }
1350
1351 /*
1352  * ocfs2_write_end() wants to know which parts of the target page it
1353  * should complete the write on. It's easiest to compute them ahead of
1354  * time when a more complete view of the write is available.
1355  */
1356 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1357                                         struct ocfs2_write_ctxt *wc,
1358                                         loff_t pos, unsigned len, int alloc)
1359 {
1360         struct ocfs2_write_cluster_desc *desc;
1361
1362         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1363         wc->w_target_to = wc->w_target_from + len;
1364
1365         if (alloc == 0)
1366                 return;
1367
1368         /*
1369          * Allocating write - we may have different boundaries based
1370          * on page size and cluster size.
1371          *
1372          * NOTE: We can no longer compute one value from the other as
1373          * the actual write length and user provided length may be
1374          * different.
1375          */
1376
1377         if (wc->w_large_pages) {
1378                 /*
1379                  * We only care about the 1st and last cluster within
1380                  * our range and whether they should be zero'd or not. Either
1381                  * value may be extended out to the start/end of a
1382                  * newly allocated cluster.
1383                  */
1384                 desc = &wc->w_desc[0];
1385                 if (desc->c_needs_zero)
1386                         ocfs2_figure_cluster_boundaries(osb,
1387                                                         desc->c_cpos,
1388                                                         &wc->w_target_from,
1389                                                         NULL);
1390
1391                 desc = &wc->w_desc[wc->w_clen - 1];
1392                 if (desc->c_needs_zero)
1393                         ocfs2_figure_cluster_boundaries(osb,
1394                                                         desc->c_cpos,
1395                                                         NULL,
1396                                                         &wc->w_target_to);
1397         } else {
1398                 wc->w_target_from = 0;
1399                 wc->w_target_to = PAGE_CACHE_SIZE;
1400         }
1401 }
1402
1403 /*
1404  * Populate each single-cluster write descriptor in the write context
1405  * with information about the i/o to be done.
1406  *
1407  * Returns the number of clusters that will have to be allocated, as
1408  * well as a worst case estimate of the number of extent records that
1409  * would have to be created during a write to an unwritten region.
1410  */
1411 static int ocfs2_populate_write_desc(struct inode *inode,
1412                                      struct ocfs2_write_ctxt *wc,
1413                                      unsigned int *clusters_to_alloc,
1414                                      unsigned int *extents_to_split)
1415 {
1416         int ret;
1417         struct ocfs2_write_cluster_desc *desc;
1418         unsigned int num_clusters = 0;
1419         unsigned int ext_flags = 0;
1420         u32 phys = 0;
1421         int i;
1422
1423         *clusters_to_alloc = 0;
1424         *extents_to_split = 0;
1425
1426         for (i = 0; i < wc->w_clen; i++) {
1427                 desc = &wc->w_desc[i];
1428                 desc->c_cpos = wc->w_cpos + i;
1429
1430                 if (num_clusters == 0) {
1431                         /*
1432                          * Need to look up the next extent record.
1433                          */
1434                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1435                                                  &num_clusters, &ext_flags);
1436                         if (ret) {
1437                                 mlog_errno(ret);
1438                                 goto out;
1439                         }
1440
1441                         /* We should already CoW the refcountd extent. */
1442                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1443
1444                         /*
1445                          * Assume worst case - that we're writing in
1446                          * the middle of the extent.
1447                          *
1448                          * We can assume that the write proceeds from
1449                          * left to right, in which case the extent
1450                          * insert code is smart enough to coalesce the
1451                          * next splits into the previous records created.
1452                          */
1453                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1454                                 *extents_to_split = *extents_to_split + 2;
1455                 } else if (phys) {
1456                         /*
1457                          * Only increment phys if it doesn't describe
1458                          * a hole.
1459                          */
1460                         phys++;
1461                 }
1462
1463                 /*
1464                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1465                  * file that got extended.  w_first_new_cpos tells us
1466                  * where the newly allocated clusters are so we can
1467                  * zero them.
1468                  */
1469                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1470                         BUG_ON(phys == 0);
1471                         desc->c_needs_zero = 1;
1472                 }
1473
1474                 desc->c_phys = phys;
1475                 if (phys == 0) {
1476                         desc->c_new = 1;
1477                         desc->c_needs_zero = 1;
1478                         *clusters_to_alloc = *clusters_to_alloc + 1;
1479                 }
1480
1481                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1482                         desc->c_unwritten = 1;
1483                         desc->c_needs_zero = 1;
1484                 }
1485
1486                 num_clusters--;
1487         }
1488
1489         ret = 0;
1490 out:
1491         return ret;
1492 }
1493
1494 static int ocfs2_write_begin_inline(struct address_space *mapping,
1495                                     struct inode *inode,
1496                                     struct ocfs2_write_ctxt *wc)
1497 {
1498         int ret;
1499         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1500         struct page *page;
1501         handle_t *handle;
1502         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1503
1504         page = find_or_create_page(mapping, 0, GFP_NOFS);
1505         if (!page) {
1506                 ret = -ENOMEM;
1507                 mlog_errno(ret);
1508                 goto out;
1509         }
1510         /*
1511          * If we don't set w_num_pages then this page won't get unlocked
1512          * and freed on cleanup of the write context.
1513          */
1514         wc->w_pages[0] = wc->w_target_page = page;
1515         wc->w_num_pages = 1;
1516
1517         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1518         if (IS_ERR(handle)) {
1519                 ret = PTR_ERR(handle);
1520                 mlog_errno(ret);
1521                 goto out;
1522         }
1523
1524         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1525                                       OCFS2_JOURNAL_ACCESS_WRITE);
1526         if (ret) {
1527                 ocfs2_commit_trans(osb, handle);
1528
1529                 mlog_errno(ret);
1530                 goto out;
1531         }
1532
1533         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1534                 ocfs2_set_inode_data_inline(inode, di);
1535
1536         if (!PageUptodate(page)) {
1537                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1538                 if (ret) {
1539                         ocfs2_commit_trans(osb, handle);
1540
1541                         goto out;
1542                 }
1543         }
1544
1545         wc->w_handle = handle;
1546 out:
1547         return ret;
1548 }
1549
1550 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1551 {
1552         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1553
1554         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1555                 return 1;
1556         return 0;
1557 }
1558
1559 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1560                                           struct inode *inode, loff_t pos,
1561                                           unsigned len, struct page *mmap_page,
1562                                           struct ocfs2_write_ctxt *wc)
1563 {
1564         int ret, written = 0;
1565         loff_t end = pos + len;
1566         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1567         struct ocfs2_dinode *di = NULL;
1568
1569         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1570                                              len, (unsigned long long)pos,
1571                                              oi->ip_dyn_features);
1572
1573         /*
1574          * Handle inodes which already have inline data 1st.
1575          */
1576         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1577                 if (mmap_page == NULL &&
1578                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1579                         goto do_inline_write;
1580
1581                 /*
1582                  * The write won't fit - we have to give this inode an
1583                  * inline extent list now.
1584                  */
1585                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1586                 if (ret)
1587                         mlog_errno(ret);
1588                 goto out;
1589         }
1590
1591         /*
1592          * Check whether the inode can accept inline data.
1593          */
1594         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1595                 return 0;
1596
1597         /*
1598          * Check whether the write can fit.
1599          */
1600         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1601         if (mmap_page ||
1602             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1603                 return 0;
1604
1605 do_inline_write:
1606         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1607         if (ret) {
1608                 mlog_errno(ret);
1609                 goto out;
1610         }
1611
1612         /*
1613          * This signals to the caller that the data can be written
1614          * inline.
1615          */
1616         written = 1;
1617 out:
1618         return written ? written : ret;
1619 }
1620
1621 /*
1622  * This function only does anything for file systems which can't
1623  * handle sparse files.
1624  *
1625  * What we want to do here is fill in any hole between the current end
1626  * of allocation and the end of our write. That way the rest of the
1627  * write path can treat it as an non-allocating write, which has no
1628  * special case code for sparse/nonsparse files.
1629  */
1630 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1631                                         struct buffer_head *di_bh,
1632                                         loff_t pos, unsigned len,
1633                                         struct ocfs2_write_ctxt *wc)
1634 {
1635         int ret;
1636         loff_t newsize = pos + len;
1637
1638         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1639
1640         if (newsize <= i_size_read(inode))
1641                 return 0;
1642
1643         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1644         if (ret)
1645                 mlog_errno(ret);
1646
1647         wc->w_first_new_cpos =
1648                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1649
1650         return ret;
1651 }
1652
1653 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1654                            loff_t pos)
1655 {
1656         int ret = 0;
1657
1658         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1659         if (pos > i_size_read(inode))
1660                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1661
1662         return ret;
1663 }
1664
1665 /*
1666  * Try to flush truncate logs if we can free enough clusters from it.
1667  * As for return value, "< 0" means error, "0" no space and "1" means
1668  * we have freed enough spaces and let the caller try to allocate again.
1669  */
1670 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1671                                           unsigned int needed)
1672 {
1673         tid_t target;
1674         int ret = 0;
1675         unsigned int truncated_clusters;
1676
1677         mutex_lock(&osb->osb_tl_inode->i_mutex);
1678         truncated_clusters = osb->truncated_clusters;
1679         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1680
1681         /*
1682          * Check whether we can succeed in allocating if we free
1683          * the truncate log.
1684          */
1685         if (truncated_clusters < needed)
1686                 goto out;
1687
1688         ret = ocfs2_flush_truncate_log(osb);
1689         if (ret) {
1690                 mlog_errno(ret);
1691                 goto out;
1692         }
1693
1694         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1695                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1696                 ret = 1;
1697         }
1698 out:
1699         return ret;
1700 }
1701
1702 int ocfs2_write_begin_nolock(struct file *filp,
1703                              struct address_space *mapping,
1704                              loff_t pos, unsigned len, unsigned flags,
1705                              struct page **pagep, void **fsdata,
1706                              struct buffer_head *di_bh, struct page *mmap_page)
1707 {
1708         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1709         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1710         struct ocfs2_write_ctxt *wc;
1711         struct inode *inode = mapping->host;
1712         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1713         struct ocfs2_dinode *di;
1714         struct ocfs2_alloc_context *data_ac = NULL;
1715         struct ocfs2_alloc_context *meta_ac = NULL;
1716         handle_t *handle;
1717         struct ocfs2_extent_tree et;
1718         int try_free = 1, ret1;
1719
1720 try_again:
1721         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1722         if (ret) {
1723                 mlog_errno(ret);
1724                 return ret;
1725         }
1726
1727         if (ocfs2_supports_inline_data(osb)) {
1728                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1729                                                      mmap_page, wc);
1730                 if (ret == 1) {
1731                         ret = 0;
1732                         goto success;
1733                 }
1734                 if (ret < 0) {
1735                         mlog_errno(ret);
1736                         goto out;
1737                 }
1738         }
1739
1740         if (ocfs2_sparse_alloc(osb))
1741                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1742         else
1743                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1744                                                    wc);
1745         if (ret) {
1746                 mlog_errno(ret);
1747                 goto out;
1748         }
1749
1750         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1751         if (ret < 0) {
1752                 mlog_errno(ret);
1753                 goto out;
1754         } else if (ret == 1) {
1755                 clusters_need = wc->w_clen;
1756                 ret = ocfs2_refcount_cow(inode, di_bh,
1757                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1758                 if (ret) {
1759                         mlog_errno(ret);
1760                         goto out;
1761                 }
1762         }
1763
1764         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1765                                         &extents_to_split);
1766         if (ret) {
1767                 mlog_errno(ret);
1768                 goto out;
1769         }
1770         clusters_need += clusters_to_alloc;
1771
1772         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1773
1774         trace_ocfs2_write_begin_nolock(
1775                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1776                         (long long)i_size_read(inode),
1777                         le32_to_cpu(di->i_clusters),
1778                         pos, len, flags, mmap_page,
1779                         clusters_to_alloc, extents_to_split);
1780
1781         /*
1782          * We set w_target_from, w_target_to here so that
1783          * ocfs2_write_end() knows which range in the target page to
1784          * write out. An allocation requires that we write the entire
1785          * cluster range.
1786          */
1787         if (clusters_to_alloc || extents_to_split) {
1788                 /*
1789                  * XXX: We are stretching the limits of
1790                  * ocfs2_lock_allocators(). It greatly over-estimates
1791                  * the work to be done.
1792                  */
1793                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1794                                               wc->w_di_bh);
1795                 ret = ocfs2_lock_allocators(inode, &et,
1796                                             clusters_to_alloc, extents_to_split,
1797                                             &data_ac, &meta_ac);
1798                 if (ret) {
1799                         mlog_errno(ret);
1800                         goto out;
1801                 }
1802
1803                 if (data_ac)
1804                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1805
1806                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1807                                                     &di->id2.i_list);
1808
1809         }
1810
1811         /*
1812          * We have to zero sparse allocated clusters, unwritten extent clusters,
1813          * and non-sparse clusters we just extended.  For non-sparse writes,
1814          * we know zeros will only be needed in the first and/or last cluster.
1815          */
1816         if (clusters_to_alloc || extents_to_split ||
1817             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1818                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1819                 cluster_of_pages = 1;
1820         else
1821                 cluster_of_pages = 0;
1822
1823         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1824
1825         handle = ocfs2_start_trans(osb, credits);
1826         if (IS_ERR(handle)) {
1827                 ret = PTR_ERR(handle);
1828                 mlog_errno(ret);
1829                 goto out;
1830         }
1831
1832         wc->w_handle = handle;
1833
1834         if (clusters_to_alloc) {
1835                 ret = dquot_alloc_space_nodirty(inode,
1836                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1837                 if (ret)
1838                         goto out_commit;
1839         }
1840         /*
1841          * We don't want this to fail in ocfs2_write_end(), so do it
1842          * here.
1843          */
1844         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1845                                       OCFS2_JOURNAL_ACCESS_WRITE);
1846         if (ret) {
1847                 mlog_errno(ret);
1848                 goto out_quota;
1849         }
1850
1851         /*
1852          * Fill our page array first. That way we've grabbed enough so
1853          * that we can zero and flush if we error after adding the
1854          * extent.
1855          */
1856         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1857                                          cluster_of_pages, mmap_page);
1858         if (ret && ret != -EAGAIN) {
1859                 mlog_errno(ret);
1860                 goto out_quota;
1861         }
1862
1863         /*
1864          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1865          * the target page. In this case, we exit with no error and no target
1866          * page. This will trigger the caller, page_mkwrite(), to re-try
1867          * the operation.
1868          */
1869         if (ret == -EAGAIN) {
1870                 BUG_ON(wc->w_target_page);
1871                 ret = 0;
1872                 goto out_quota;
1873         }
1874
1875         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1876                                           len);
1877         if (ret) {
1878                 mlog_errno(ret);
1879                 goto out_quota;
1880         }
1881
1882         if (data_ac)
1883                 ocfs2_free_alloc_context(data_ac);
1884         if (meta_ac)
1885                 ocfs2_free_alloc_context(meta_ac);
1886
1887 success:
1888         *pagep = wc->w_target_page;
1889         *fsdata = wc;
1890         return 0;
1891 out_quota:
1892         if (clusters_to_alloc)
1893                 dquot_free_space(inode,
1894                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1895 out_commit:
1896         ocfs2_commit_trans(osb, handle);
1897
1898 out:
1899         ocfs2_free_write_ctxt(wc);
1900
1901         if (data_ac)
1902                 ocfs2_free_alloc_context(data_ac);
1903         if (meta_ac)
1904                 ocfs2_free_alloc_context(meta_ac);
1905
1906         if (ret == -ENOSPC && try_free) {
1907                 /*
1908                  * Try to free some truncate log so that we can have enough
1909                  * clusters to allocate.
1910                  */
1911                 try_free = 0;
1912
1913                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1914                 if (ret1 == 1)
1915                         goto try_again;
1916
1917                 if (ret1 < 0)
1918                         mlog_errno(ret1);
1919         }
1920
1921         return ret;
1922 }
1923
1924 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1925                              loff_t pos, unsigned len, unsigned flags,
1926                              struct page **pagep, void **fsdata)
1927 {
1928         int ret;
1929         struct buffer_head *di_bh = NULL;
1930         struct inode *inode = mapping->host;
1931
1932         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1933         if (ret) {
1934                 mlog_errno(ret);
1935                 return ret;
1936         }
1937
1938         /*
1939          * Take alloc sem here to prevent concurrent lookups. That way
1940          * the mapping, zeroing and tree manipulation within
1941          * ocfs2_write() will be safe against ->readpage(). This
1942          * should also serve to lock out allocation from a shared
1943          * writeable region.
1944          */
1945         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1946
1947         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1948                                        fsdata, di_bh, NULL);
1949         if (ret) {
1950                 mlog_errno(ret);
1951                 goto out_fail;
1952         }
1953
1954         brelse(di_bh);
1955
1956         return 0;
1957
1958 out_fail:
1959         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1960
1961         brelse(di_bh);
1962         ocfs2_inode_unlock(inode, 1);
1963
1964         return ret;
1965 }
1966
1967 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1968                                    unsigned len, unsigned *copied,
1969                                    struct ocfs2_dinode *di,
1970                                    struct ocfs2_write_ctxt *wc)
1971 {
1972         void *kaddr;
1973
1974         if (unlikely(*copied < len)) {
1975                 if (!PageUptodate(wc->w_target_page)) {
1976                         *copied = 0;
1977                         return;
1978                 }
1979         }
1980
1981         kaddr = kmap_atomic(wc->w_target_page);
1982         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1983         kunmap_atomic(kaddr);
1984
1985         trace_ocfs2_write_end_inline(
1986              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1987              (unsigned long long)pos, *copied,
1988              le16_to_cpu(di->id2.i_data.id_count),
1989              le16_to_cpu(di->i_dyn_features));
1990 }
1991
1992 int ocfs2_write_end_nolock(struct address_space *mapping,
1993                            loff_t pos, unsigned len, unsigned copied,
1994                            struct page *page, void *fsdata)
1995 {
1996         int i;
1997         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1998         struct inode *inode = mapping->host;
1999         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2000         struct ocfs2_write_ctxt *wc = fsdata;
2001         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2002         handle_t *handle = wc->w_handle;
2003         struct page *tmppage;
2004
2005         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2006                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2007                 goto out_write_size;
2008         }
2009
2010         if (unlikely(copied < len)) {
2011                 if (!PageUptodate(wc->w_target_page))
2012                         copied = 0;
2013
2014                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2015                                        start+len);
2016         }
2017         flush_dcache_page(wc->w_target_page);
2018
2019         for(i = 0; i < wc->w_num_pages; i++) {
2020                 tmppage = wc->w_pages[i];
2021
2022                 if (tmppage == wc->w_target_page) {
2023                         from = wc->w_target_from;
2024                         to = wc->w_target_to;
2025
2026                         BUG_ON(from > PAGE_CACHE_SIZE ||
2027                                to > PAGE_CACHE_SIZE ||
2028                                to < from);
2029                 } else {
2030                         /*
2031                          * Pages adjacent to the target (if any) imply
2032                          * a hole-filling write in which case we want
2033                          * to flush their entire range.
2034                          */
2035                         from = 0;
2036                         to = PAGE_CACHE_SIZE;
2037                 }
2038
2039                 if (page_has_buffers(tmppage)) {
2040                         if (ocfs2_should_order_data(inode))
2041                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2042                         block_commit_write(tmppage, from, to);
2043                 }
2044         }
2045
2046 out_write_size:
2047         pos += copied;
2048         if (pos > i_size_read(inode)) {
2049                 i_size_write(inode, pos);
2050                 mark_inode_dirty(inode);
2051         }
2052         inode->i_blocks = ocfs2_inode_sector_count(inode);
2053         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2054         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2055         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2056         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2057         ocfs2_journal_dirty(handle, wc->w_di_bh);
2058
2059         ocfs2_commit_trans(osb, handle);
2060
2061         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2062
2063         ocfs2_free_write_ctxt(wc);
2064
2065         return copied;
2066 }
2067
2068 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2069                            loff_t pos, unsigned len, unsigned copied,
2070                            struct page *page, void *fsdata)
2071 {
2072         int ret;
2073         struct inode *inode = mapping->host;
2074
2075         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2076
2077         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2078         ocfs2_inode_unlock(inode, 1);
2079
2080         return ret;
2081 }
2082
2083 const struct address_space_operations ocfs2_aops = {
2084         .readpage               = ocfs2_readpage,
2085         .readpages              = ocfs2_readpages,
2086         .writepage              = ocfs2_writepage,
2087         .write_begin            = ocfs2_write_begin,
2088         .write_end              = ocfs2_write_end,
2089         .bmap                   = ocfs2_bmap,
2090         .direct_IO              = ocfs2_direct_IO,
2091         .invalidatepage         = ocfs2_invalidatepage,
2092         .releasepage            = ocfs2_releasepage,
2093         .migratepage            = buffer_migrate_page,
2094         .is_partially_uptodate  = block_is_partially_uptodate,
2095         .error_remove_page      = generic_error_remove_page,
2096 };