fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31
  32 #include <cluster/masklog.h>
  33
  34 #include "ocfs2.h"
  35
  36 #include "alloc.h"
  37 #include "aops.h"
  38 #include "dlmglue.h"
  39 #include "extent_map.h"
  40 #include "file.h"
  41 #include "inode.h"
  42 #include "journal.h"
  43 #include "suballoc.h"
  44 #include "super.h"
  45 #include "symlink.h"
  46 #include "refcounttree.h"
  47 #include "ocfs2_trace.h"
  48
  49 #include "buffer_head_io.h"
  50
  51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  52                                    struct buffer_head *bh_result, int create)
  53 {
  54         int err = -EIO;
  55         int status;
  56         struct ocfs2_dinode *fe = NULL;
  57         struct buffer_head *bh = NULL;
  58         struct buffer_head *buffer_cache_bh = NULL;
  59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  60         void *kaddr;
  61
  62         trace_ocfs2_symlink_get_block(
  63                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  64                         (unsigned long long)iblock, bh_result, create);
  65
  66         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  67
  68         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  69                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  70                      (unsigned long long)iblock);
  71                 goto bail;
  72         }
  73
  74         status = ocfs2_read_inode_block(inode, &bh);
  75         if (status < 0) {
  76                 mlog_errno(status);
  77                 goto bail;
  78         }
  79         fe = (struct ocfs2_dinode *) bh->b_data;
  80
  81         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  82                                                     le32_to_cpu(fe->i_clusters))) {
  83                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  84                      "%llu\n", (unsigned long long)iblock);
  85                 goto bail;
  86         }
  87
  88         /* We don't use the page cache to create symlink data, so if
  89          * need be, copy it over from the buffer cache. */
  90         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  91                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  92                             iblock;
  93                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  94                 if (!buffer_cache_bh) {
  95                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  96                         goto bail;
  97                 }
  98
  99                 /* we haven't locked out transactions, so a commit
 100                  * could've happened. Since we've got a reference on
 101                  * the bh, even if it commits while we're doing the
 102                  * copy, the data is still good. */
 103                 if (buffer_jbd(buffer_cache_bh)
 104                     && ocfs2_inode_is_new(inode)) {
 105                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 106                         if (!kaddr) {
 107                                 mlog(ML_ERROR, "couldn't kmap!\n");
 108                                 goto bail;
 109                         }
 110                         memcpy(kaddr + (bh_result->b_size * iblock),
 111                                buffer_cache_bh->b_data,
 112                                bh_result->b_size);
 113                         kunmap_atomic(kaddr, KM_USER0);
 114                         set_buffer_uptodate(bh_result);
 115                 }
 116                 brelse(buffer_cache_bh);
 117         }
 118
 119         map_bh(bh_result, inode->i_sb,
 120                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 121
 122         err = 0;
 123
 124 bail:
 125         brelse(bh);
 126
 127         return err;
 128 }
 129
 130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 131                     struct buffer_head *bh_result, int create)
 132 {
 133         int err = 0;
 134         unsigned int ext_flags;
 135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 136         u64 p_blkno, count, past_eof;
 137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 138
 139         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 140                               (unsigned long long)iblock, bh_result, create);
 141
 142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 144                      inode, inode->i_ino);
 145
 146         if (S_ISLNK(inode->i_mode)) {
 147                 /* this always does I/O for some reason. */
 148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 149                 goto bail;
 150         }
 151
 152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 153                                           &ext_flags);
 154         if (err) {
 155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 157                      (unsigned long long)p_blkno);
 158                 goto bail;
 159         }
 160
 161         if (max_blocks < count)
 162                 count = max_blocks;
 163
 164         /*
 165          * ocfs2 never allocates in this function - the only time we
 166          * need to use BH_New is when we're extending i_size on a file
 167          * system which doesn't support holes, in which case BH_New
 168          * allows __block_write_begin() to zero.
 169          *
 170          * If we see this on a sparse file system, then a truncate has
 171          * raced us and removed the cluster. In this case, we clear
 172          * the buffers dirty and uptodate bits and let the buffer code
 173          * ignore it as a hole.
 174          */
 175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 176                 clear_buffer_dirty(bh_result);
 177                 clear_buffer_uptodate(bh_result);
 178                 goto bail;
 179         }
 180
 181         /* Treat the unwritten extent as a hole for zeroing purposes. */
 182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 183                 map_bh(bh_result, inode->i_sb, p_blkno);
 184
 185         bh_result->b_size = count << inode->i_blkbits;
 186
 187         if (!ocfs2_sparse_alloc(osb)) {
 188                 if (p_blkno == 0) {
 189                         err = -EIO;
 190                         mlog(ML_ERROR,
 191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 192                              (unsigned long long)iblock,
 193                              (unsigned long long)p_blkno,
 194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 196                         dump_stack();
 197                         goto bail;
 198                 }
 199         }
 200
 201         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 202
 203         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 204                                   (unsigned long long)past_eof);
 205         if (create && (iblock >= past_eof))
 206                 set_buffer_new(bh_result);
 207
 208 bail:
 209         if (err < 0)
 210                 err = -EIO;
 211
 212         return err;
 213 }
 214
 215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 216                            struct buffer_head *di_bh)
 217 {
 218         void *kaddr;
 219         loff_t size;
 220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 221
 222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 225                 return -EROFS;
 226         }
 227
 228         size = i_size_read(inode);
 229
 230         if (size > PAGE_CACHE_SIZE ||
 231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 232                 ocfs2_error(inode->i_sb,
 233                             "Inode %llu has with inline data has bad size: %Lu",
 234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 235                             (unsigned long long)size);
 236                 return -EROFS;
 237         }
 238
 239         kaddr = kmap_atomic(page, KM_USER0);
 240         if (size)
 241                 memcpy(kaddr, di->id2.i_data.id_data, size);
 242         /* Clear the remaining part of the page */
 243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 244         flush_dcache_page(page);
 245         kunmap_atomic(kaddr, KM_USER0);
 246
 247         SetPageUptodate(page);
 248
 249         return 0;
 250 }
 251
 252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 253 {
 254         int ret;
 255         struct buffer_head *di_bh = NULL;
 256
 257         BUG_ON(!PageLocked(page));
 258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 259
 260         ret = ocfs2_read_inode_block(inode, &di_bh);
 261         if (ret) {
 262                 mlog_errno(ret);
 263                 goto out;
 264         }
 265
 266         ret = ocfs2_read_inline_data(inode, page, di_bh);
 267 out:
 268         unlock_page(page);
 269
 270         brelse(di_bh);
 271         return ret;
 272 }
 273
 274 static int ocfs2_readpage(struct file *file, struct page *page)
 275 {
 276         struct inode *inode = page->mapping->host;
 277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 279         int ret, unlock = 1;
 280
 281         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 282                              (page ? page->index : 0));
 283
 284         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 285         if (ret != 0) {
 286                 if (ret == AOP_TRUNCATED_PAGE)
 287                         unlock = 0;
 288                 mlog_errno(ret);
 289                 goto out;
 290         }
 291
 292         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 293                 /*
 294                  * Unlock the page and cycle ip_alloc_sem so that we don't
 295                  * busyloop waiting for ip_alloc_sem to unlock
 296                  */
 297                 ret = AOP_TRUNCATED_PAGE;
 298                 unlock_page(page);
 299                 unlock = 0;
 300                 down_read(&oi->ip_alloc_sem);
 301                 up_read(&oi->ip_alloc_sem);
 302                 goto out_inode_unlock;
 303         }
 304
 305         /*
 306          * i_size might have just been updated as we grabed the meta lock.  We
 307          * might now be discovering a truncate that hit on another node.
 308          * block_read_full_page->get_block freaks out if it is asked to read
 309          * beyond the end of a file, so we check here.  Callers
 310          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 311          * and notice that the page they just read isn't needed.
 312          *
 313          * XXX sys_readahead() seems to get that wrong?
 314          */
 315         if (start >= i_size_read(inode)) {
 316                 zero_user(page, 0, PAGE_SIZE);
 317                 SetPageUptodate(page);
 318                 ret = 0;
 319                 goto out_alloc;
 320         }
 321
 322         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 323                 ret = ocfs2_readpage_inline(inode, page);
 324         else
 325                 ret = block_read_full_page(page, ocfs2_get_block);
 326         unlock = 0;
 327
 328 out_alloc:
 329         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 330 out_inode_unlock:
 331         ocfs2_inode_unlock(inode, 0);
 332 out:
 333         if (unlock)
 334                 unlock_page(page);
 335         return ret;
 336 }
 337
 338 /*
 339  * This is used only for read-ahead. Failures or difficult to handle
 340  * situations are safe to ignore.
 341  *
 342  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 343  * are quite large (243 extents on 4k blocks), so most inodes don't
 344  * grow out to a tree. If need be, detecting boundary extents could
 345  * trivially be added in a future version of ocfs2_get_block().
 346  */
 347 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 348                            struct list_head *pages, unsigned nr_pages)
 349 {
 350         int ret, err = -EIO;
 351         struct inode *inode = mapping->host;
 352         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 353         loff_t start;
 354         struct page *last;
 355
 356         /*
 357          * Use the nonblocking flag for the dlm code to avoid page
 358          * lock inversion, but don't bother with retrying.
 359          */
 360         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 361         if (ret)
 362                 return err;
 363
 364         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 365                 ocfs2_inode_unlock(inode, 0);
 366                 return err;
 367         }
 368
 369         /*
 370          * Don't bother with inline-data. There isn't anything
 371          * to read-ahead in that case anyway...
 372          */
 373         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 374                 goto out_unlock;
 375
 376         /*
 377          * Check whether a remote node truncated this file - we just
 378          * drop out in that case as it's not worth handling here.
 379          */
 380         last = list_entry(pages->prev, struct page, lru);
 381         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 382         if (start >= i_size_read(inode))
 383                 goto out_unlock;
 384
 385         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 386
 387 out_unlock:
 388         up_read(&oi->ip_alloc_sem);
 389         ocfs2_inode_unlock(inode, 0);
 390
 391         return err;
 392 }
 393
 394 /* Note: Because we don't support holes, our allocation has
 395  * already happened (allocation writes zeros to the file data)
 396  * so we don't have to worry about ordered writes in
 397  * ocfs2_writepage.
 398  *
 399  * ->writepage is called during the process of invalidating the page cache
 400  * during blocked lock processing.  It can't block on any cluster locks
 401  * to during block mapping.  It's relying on the fact that the block
 402  * mapping can't have disappeared under the dirty pages that it is
 403  * being asked to write back.
 404  */
 405 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 406 {
 407         trace_ocfs2_writepage(
 408                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 409                 page->index);
 410
 411         return block_write_full_page(page, ocfs2_get_block, wbc);
 412 }
 413
 414 /* Taken from ext3. We don't necessarily need the full blown
 415  * functionality yet, but IMHO it's better to cut and paste the whole
 416  * thing so we can avoid introducing our own bugs (and easily pick up
 417  * their fixes when they happen) --Mark */
 418 int walk_page_buffers(  handle_t *handle,
 419                         struct buffer_head *head,
 420                         unsigned from,
 421                         unsigned to,
 422                         int *partial,
 423                         int (*fn)(      handle_t *handle,
 424                                         struct buffer_head *bh))
 425 {
 426         struct buffer_head *bh;
 427         unsigned block_start, block_end;
 428         unsigned blocksize = head->b_size;
 429         int err, ret = 0;
 430         struct buffer_head *next;
 431
 432         for (   bh = head, block_start = 0;
 433                 ret == 0 && (bh != head || !block_start);
 434                 block_start = block_end, bh = next)
 435         {
 436                 next = bh->b_this_page;
 437                 block_end = block_start + blocksize;
 438                 if (block_end <= from || block_start >= to) {
 439                         if (partial && !buffer_uptodate(bh))
 440                                 *partial = 1;
 441                         continue;
 442                 }
 443                 err = (*fn)(handle, bh);
 444                 if (!ret)
 445                         ret = err;
 446         }
 447         return ret;
 448 }
 449
 450 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 451 {
 452         sector_t status;
 453         u64 p_blkno = 0;
 454         int err = 0;
 455         struct inode *inode = mapping->host;
 456
 457         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 458                          (unsigned long long)block);
 459
 460         /* We don't need to lock journal system files, since they aren't
 461          * accessed concurrently from multiple nodes.
 462          */
 463         if (!INODE_JOURNAL(inode)) {
 464                 err = ocfs2_inode_lock(inode, NULL, 0);
 465                 if (err) {
 466                         if (err != -ENOENT)
 467                                 mlog_errno(err);
 468                         goto bail;
 469                 }
 470                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 471         }
 472
 473         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 474                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 475                                                   NULL);
 476
 477         if (!INODE_JOURNAL(inode)) {
 478                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 479                 ocfs2_inode_unlock(inode, 0);
 480         }
 481
 482         if (err) {
 483                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 484                      (unsigned long long)block);
 485                 mlog_errno(err);
 486                 goto bail;
 487         }
 488
 489 bail:
 490         status = err ? 0 : p_blkno;
 491
 492         return status;
 493 }
 494
 495 /*
 496  * TODO: Make this into a generic get_blocks function.
 497  *
 498  * From do_direct_io in direct-io.c:
 499  *  "So what we do is to permit the ->get_blocks function to populate
 500  *   bh.b_size with the size of IO which is permitted at this offset and
 501  *   this i_blkbits."
 502  *
 503  * This function is called directly from get_more_blocks in direct-io.c.
 504  *
 505  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 506  *                                      fs_count, map_bh, dio->rw == WRITE);
 507  *
 508  * Note that we never bother to allocate blocks here, and thus ignore the
 509  * create argument.
 510  */
 511 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 512                                      struct buffer_head *bh_result, int create)
 513 {
 514         int ret;
 515         u64 p_blkno, inode_blocks, contig_blocks;
 516         unsigned int ext_flags;
 517         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 518         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 519
 520         /* This function won't even be called if the request isn't all
 521          * nicely aligned and of the right size, so there's no need
 522          * for us to check any of that. */
 523
 524         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 525
 526         /* This figures out the size of the next contiguous block, and
 527          * our logical offset */
 528         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 529                                           &contig_blocks, &ext_flags);
 530         if (ret) {
 531                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 532                      (unsigned long long)iblock);
 533                 ret = -EIO;
 534                 goto bail;
 535         }
 536
 537         /* We should already CoW the refcounted extent in case of create. */
 538         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 539
 540         /*
 541          * get_more_blocks() expects us to describe a hole by clearing
 542          * the mapped bit on bh_result().
 543          *
 544          * Consider an unwritten extent as a hole.
 545          */
 546         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 547                 map_bh(bh_result, inode->i_sb, p_blkno);
 548         else
 549                 clear_buffer_mapped(bh_result);
 550
 551         /* make sure we don't map more than max_blocks blocks here as
 552            that's all the kernel will handle at this point. */
 553         if (max_blocks < contig_blocks)
 554                 contig_blocks = max_blocks;
 555         bh_result->b_size = contig_blocks << blocksize_bits;
 556 bail:
 557         return ret;
 558 }
 559
 560 /*
 561  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 562  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 563  * to protect io on one node from truncation on another.
 564  */
 565 static void ocfs2_dio_end_io(struct kiocb *iocb,
 566                              loff_t offset,
 567                              ssize_t bytes,
 568                              void *private,
 569                              int ret,
 570                              bool is_async)
 571 {
 572         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 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         inode_dio_done(inode);
 597         if (is_async)
 598                 aio_complete(iocb, ret, 0);
 599 }
 600
 601 /*
 602  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 603  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 604  * do journalled data.
 605  */
 606 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 607 {
 608         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 609
 610         jbd2_journal_invalidatepage(journal, page, offset);
 611 }
 612
 613 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 614 {
 615         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 616
 617         if (!page_has_buffers(page))
 618                 return 0;
 619         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 620 }
 621
 622 static ssize_t ocfs2_direct_IO(int rw,
 623                                struct kiocb *iocb,
 624                                const struct iovec *iov,
 625                                loff_t offset,
 626                                unsigned long nr_segs)
 627 {
 628         struct file *file = iocb->ki_filp;
 629         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 630
 631         /*
 632          * Fallback to buffered I/O if we see an inode without
 633          * extents.
 634          */
 635         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 636                 return 0;
 637
 638         /* Fallback to buffered I/O if we are appending. */
 639         if (i_size_read(inode) <= offset)
 640                 return 0;
 641
 642         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 643                                     iov, offset, nr_segs,
 644                                     ocfs2_direct_IO_get_blocks,
 645                                     ocfs2_dio_end_io, NULL, 0);
 646 }
 647
 648 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 649                                             u32 cpos,
 650                                             unsigned int *start,
 651                                             unsigned int *end)
 652 {
 653         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 654
 655         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 656                 unsigned int cpp;
 657
 658                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 659
 660                 cluster_start = cpos % cpp;
 661                 cluster_start = cluster_start << osb->s_clustersize_bits;
 662
 663                 cluster_end = cluster_start + osb->s_clustersize;
 664         }
 665
 666         BUG_ON(cluster_start > PAGE_SIZE);
 667         BUG_ON(cluster_end > PAGE_SIZE);
 668
 669         if (start)
 670                 *start = cluster_start;
 671         if (end)
 672                 *end = cluster_end;
 673 }
 674
 675 /*
 676  * 'from' and 'to' are the region in the page to avoid zeroing.
 677  *
 678  * If pagesize > clustersize, this function will avoid zeroing outside
 679  * of the cluster boundary.
 680  *
 681  * from == to == 0 is code for "zero the entire cluster region"
 682  */
 683 static void ocfs2_clear_page_regions(struct page *page,
 684                                      struct ocfs2_super *osb, u32 cpos,
 685                                      unsigned from, unsigned to)
 686 {
 687         void *kaddr;
 688         unsigned int cluster_start, cluster_end;
 689
 690         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 691
 692         kaddr = kmap_atomic(page, KM_USER0);
 693
 694         if (from || to) {
 695                 if (from > cluster_start)
 696                         memset(kaddr + cluster_start, 0, from - cluster_start);
 697                 if (to < cluster_end)
 698                         memset(kaddr + to, 0, cluster_end - to);
 699         } else {
 700                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 701         }
 702
 703         kunmap_atomic(kaddr, KM_USER0);
 704 }
 705
 706 /*
 707  * Nonsparse file systems fully allocate before we get to the write
 708  * code. This prevents ocfs2_write() from tagging the write as an
 709  * allocating one, which means ocfs2_map_page_blocks() might try to
 710  * read-in the blocks at the tail of our file. Avoid reading them by
 711  * testing i_size against each block offset.
 712  */
 713 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 714                                  unsigned int block_start)
 715 {
 716         u64 offset = page_offset(page) + block_start;
 717
 718         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 719                 return 1;
 720
 721         if (i_size_read(inode) > offset)
 722                 return 1;
 723
 724         return 0;
 725 }
 726
 727 /*
 728  * Some of this taken from __block_write_begin(). We already have our
 729  * mapping by now though, and the entire write will be allocating or
 730  * it won't, so not much need to use BH_New.
 731  *
 732  * This will also skip zeroing, which is handled externally.
 733  */
 734 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 735                           struct inode *inode, unsigned int from,
 736                           unsigned int to, int new)
 737 {
 738         int ret = 0;
 739         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 740         unsigned int block_end, block_start;
 741         unsigned int bsize = 1 << inode->i_blkbits;
 742
 743         if (!page_has_buffers(page))
 744                 create_empty_buffers(page, bsize, 0);
 745
 746         head = page_buffers(page);
 747         for (bh = head, block_start = 0; bh != head || !block_start;
 748              bh = bh->b_this_page, block_start += bsize) {
 749                 block_end = block_start + bsize;
 750
 751                 clear_buffer_new(bh);
 752
 753                 /*
 754                  * Ignore blocks outside of our i/o range -
 755                  * they may belong to unallocated clusters.
 756                  */
 757                 if (block_start >= to || block_end <= from) {
 758                         if (PageUptodate(page))
 759                                 set_buffer_uptodate(bh);
 760                         continue;
 761                 }
 762
 763                 /*
 764                  * For an allocating write with cluster size >= page
 765                  * size, we always write the entire page.
 766                  */
 767                 if (new)
 768                         set_buffer_new(bh);
 769
 770                 if (!buffer_mapped(bh)) {
 771                         map_bh(bh, inode->i_sb, *p_blkno);
 772                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 773                 }
 774
 775                 if (PageUptodate(page)) {
 776                         if (!buffer_uptodate(bh))
 777                                 set_buffer_uptodate(bh);
 778                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 779                            !buffer_new(bh) &&
 780                            ocfs2_should_read_blk(inode, page, block_start) &&
 781                            (block_start < from || block_end > to)) {
 782                         ll_rw_block(READ, 1, &bh);
 783                         *wait_bh++=bh;
 784                 }
 785
 786                 *p_blkno = *p_blkno + 1;
 787         }
 788
 789         /*
 790          * If we issued read requests - let them complete.
 791          */
 792         while(wait_bh > wait) {
 793                 wait_on_buffer(*--wait_bh);
 794                 if (!buffer_uptodate(*wait_bh))
 795                         ret = -EIO;
 796         }
 797
 798         if (ret == 0 || !new)
 799                 return ret;
 800
 801         /*
 802          * If we get -EIO above, zero out any newly allocated blocks
 803          * to avoid exposing stale data.
 804          */
 805         bh = head;
 806         block_start = 0;
 807         do {
 808                 block_end = block_start + bsize;
 809                 if (block_end <= from)
 810                         goto next_bh;
 811                 if (block_start >= to)
 812                         break;
 813
 814                 zero_user(page, block_start, bh->b_size);
 815                 set_buffer_uptodate(bh);
 816                 mark_buffer_dirty(bh);
 817
 818 next_bh:
 819                 block_start = block_end;
 820                 bh = bh->b_this_page;
 821         } while (bh != head);
 822
 823         return ret;
 824 }
 825
 826 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 827 #define OCFS2_MAX_CTXT_PAGES    1
 828 #else
 829 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 830 #endif
 831
 832 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 833
 834 /*
 835  * Describe the state of a single cluster to be written to.
 836  */
 837 struct ocfs2_write_cluster_desc {
 838         u32             c_cpos;
 839         u32             c_phys;
 840         /*
 841          * Give this a unique field because c_phys eventually gets
 842          * filled.
 843          */
 844         unsigned        c_new;
 845         unsigned        c_unwritten;
 846         unsigned        c_needs_zero;
 847 };
 848
 849 struct ocfs2_write_ctxt {
 850         /* Logical cluster position / len of write */
 851         u32                             w_cpos;
 852         u32                             w_clen;
 853
 854         /* First cluster allocated in a nonsparse extend */
 855         u32                             w_first_new_cpos;
 856
 857         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 858
 859         /*
 860          * This is true if page_size > cluster_size.
 861          *
 862          * It triggers a set of special cases during write which might
 863          * have to deal with allocating writes to partial pages.
 864          */
 865         unsigned int                    w_large_pages;
 866
 867         /*
 868          * Pages involved in this write.
 869          *
 870          * w_target_page is the page being written to by the user.
 871          *
 872          * w_pages is an array of pages which always contains
 873          * w_target_page, and in the case of an allocating write with
 874          * page_size < cluster size, it will contain zero'd and mapped
 875          * pages adjacent to w_target_page which need to be written
 876          * out in so that future reads from that region will get
 877          * zero's.
 878          */
 879         unsigned int                    w_num_pages;
 880         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 881         struct page                     *w_target_page;
 882
 883         /*
 884          * w_target_locked is used for page_mkwrite path indicating no unlocking
 885          * against w_target_page in ocfs2_write_end_nolock.
 886          */
 887         unsigned int                    w_target_locked:1;
 888
 889         /*
 890          * ocfs2_write_end() uses this to know what the real range to
 891          * write in the target should be.
 892          */
 893         unsigned int                    w_target_from;
 894         unsigned int                    w_target_to;
 895
 896         /*
 897          * We could use journal_current_handle() but this is cleaner,
 898          * IMHO -Mark
 899          */
 900         handle_t                        *w_handle;
 901
 902         struct buffer_head              *w_di_bh;
 903
 904         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 905 };
 906
 907 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 908 {
 909         int i;
 910
 911         for(i = 0; i < num_pages; i++) {
 912                 if (pages[i]) {
 913                         unlock_page(pages[i]);
 914                         mark_page_accessed(pages[i]);
 915                         page_cache_release(pages[i]);
 916                 }
 917         }
 918 }
 919
 920 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 921 {
 922         int i;
 923
 924         /*
 925          * w_target_locked is only set to true in the page_mkwrite() case.
 926          * The intent is to allow us to lock the target page from write_begin()
 927          * to write_end(). The caller must hold a ref on w_target_page.
 928          */
 929         if (wc->w_target_locked) {
 930                 BUG_ON(!wc->w_target_page);
 931                 for (i = 0; i < wc->w_num_pages; i++) {
 932                         if (wc->w_target_page == wc->w_pages[i]) {
 933                                 wc->w_pages[i] = NULL;
 934                                 break;
 935                         }
 936                 }
 937                 mark_page_accessed(wc->w_target_page);
 938                 page_cache_release(wc->w_target_page);
 939         }
 940         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 941 }
 942
 943 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 944 {
 945         ocfs2_unlock_pages(wc);
 946         brelse(wc->w_di_bh);
 947         kfree(wc);
 948 }
 949
 950 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 951                                   struct ocfs2_super *osb, loff_t pos,
 952                                   unsigned len, struct buffer_head *di_bh)
 953 {
 954         u32 cend;
 955         struct ocfs2_write_ctxt *wc;
 956
 957         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 958         if (!wc)
 959                 return -ENOMEM;
 960
 961         wc->w_cpos = pos >> osb->s_clustersize_bits;
 962         wc->w_first_new_cpos = UINT_MAX;
 963         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 964         wc->w_clen = cend - wc->w_cpos + 1;
 965         get_bh(di_bh);
 966         wc->w_di_bh = di_bh;
 967
 968         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 969                 wc->w_large_pages = 1;
 970         else
 971                 wc->w_large_pages = 0;
 972
 973         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 974
 975         *wcp = wc;
 976
 977         return 0;
 978 }
 979
 980 /*
 981  * If a page has any new buffers, zero them out here, and mark them uptodate
 982  * and dirty so they'll be written out (in order to prevent uninitialised
 983  * block data from leaking). And clear the new bit.
 984  */
 985 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 986 {
 987         unsigned int block_start, block_end;
 988         struct buffer_head *head, *bh;
 989
 990         BUG_ON(!PageLocked(page));
 991         if (!page_has_buffers(page))
 992                 return;
 993
 994         bh = head = page_buffers(page);
 995         block_start = 0;
 996         do {
 997                 block_end = block_start + bh->b_size;
 998
 999                 if (buffer_new(bh)) {
1000                         if (block_end > from && block_start < to) {
1001                                 if (!PageUptodate(page)) {
1002                                         unsigned start, end;
1003
1004                                         start = max(from, block_start);
1005                                         end = min(to, block_end);
1006
1007                                         zero_user_segment(page, start, end);
1008                                         set_buffer_uptodate(bh);
1009                                 }
1010
1011                                 clear_buffer_new(bh);
1012                                 mark_buffer_dirty(bh);
1013                         }
1014                 }
1015
1016                 block_start = block_end;
1017                 bh = bh->b_this_page;
1018         } while (bh != head);
1019 }
1020
1021 /*
1022  * Only called when we have a failure during allocating write to write
1023  * zero's to the newly allocated region.
1024  */
1025 static void ocfs2_write_failure(struct inode *inode,
1026                                 struct ocfs2_write_ctxt *wc,
1027                                 loff_t user_pos, unsigned user_len)
1028 {
1029         int i;
1030         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1031                 to = user_pos + user_len;
1032         struct page *tmppage;
1033
1034         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1035
1036         for(i = 0; i < wc->w_num_pages; i++) {
1037                 tmppage = wc->w_pages[i];
1038
1039                 if (page_has_buffers(tmppage)) {
1040                         if (ocfs2_should_order_data(inode))
1041                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1042
1043                         block_commit_write(tmppage, from, to);
1044                 }
1045         }
1046 }
1047
1048 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1049                                         struct ocfs2_write_ctxt *wc,
1050                                         struct page *page, u32 cpos,
1051                                         loff_t user_pos, unsigned user_len,
1052                                         int new)
1053 {
1054         int ret;
1055         unsigned int map_from = 0, map_to = 0;
1056         unsigned int cluster_start, cluster_end;
1057         unsigned int user_data_from = 0, user_data_to = 0;
1058
1059         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1060                                         &cluster_start, &cluster_end);
1061
1062         /* treat the write as new if the a hole/lseek spanned across
1063          * the page boundary.
1064          */
1065         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1066                         (page_offset(page) <= user_pos));
1067
1068         if (page == wc->w_target_page) {
1069                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1070                 map_to = map_from + user_len;
1071
1072                 if (new)
1073                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1074                                                     cluster_start, cluster_end,
1075                                                     new);
1076                 else
1077                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1078                                                     map_from, map_to, new);
1079                 if (ret) {
1080                         mlog_errno(ret);
1081                         goto out;
1082                 }
1083
1084                 user_data_from = map_from;
1085                 user_data_to = map_to;
1086                 if (new) {
1087                         map_from = cluster_start;
1088                         map_to = cluster_end;
1089                 }
1090         } else {
1091                 /*
1092                  * If we haven't allocated the new page yet, we
1093                  * shouldn't be writing it out without copying user
1094                  * data. This is likely a math error from the caller.
1095                  */
1096                 BUG_ON(!new);
1097
1098                 map_from = cluster_start;
1099                 map_to = cluster_end;
1100
1101                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1102                                             cluster_start, cluster_end, new);
1103                 if (ret) {
1104                         mlog_errno(ret);
1105                         goto out;
1106                 }
1107         }
1108
1109         /*
1110          * Parts of newly allocated pages need to be zero'd.
1111          *
1112          * Above, we have also rewritten 'to' and 'from' - as far as
1113          * the rest of the function is concerned, the entire cluster
1114          * range inside of a page needs to be written.
1115          *
1116          * We can skip this if the page is up to date - it's already
1117          * been zero'd from being read in as a hole.
1118          */
1119         if (new && !PageUptodate(page))
1120                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1121                                          cpos, user_data_from, user_data_to);
1122
1123         flush_dcache_page(page);
1124
1125 out:
1126         return ret;
1127 }
1128
1129 /*
1130  * This function will only grab one clusters worth of pages.
1131  */
1132 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1133                                       struct ocfs2_write_ctxt *wc,
1134                                       u32 cpos, loff_t user_pos,
1135                                       unsigned user_len, int new,
1136                                       struct page *mmap_page)
1137 {
1138         int ret = 0, i;
1139         unsigned long start, target_index, end_index, index;
1140         struct inode *inode = mapping->host;
1141         loff_t last_byte;
1142
1143         target_index = user_pos >> PAGE_CACHE_SHIFT;
1144
1145         /*
1146          * Figure out how many pages we'll be manipulating here. For
1147          * non allocating write, we just change the one
1148          * page. Otherwise, we'll need a whole clusters worth.  If we're
1149          * writing past i_size, we only need enough pages to cover the
1150          * last page of the write.
1151          */
1152         if (new) {
1153                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1154                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1155                 /*
1156                  * We need the index *past* the last page we could possibly
1157                  * touch.  This is the page past the end of the write or
1158                  * i_size, whichever is greater.
1159                  */
1160                 last_byte = max(user_pos + user_len, i_size_read(inode));
1161                 BUG_ON(last_byte < 1);
1162                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1163                 if ((start + wc->w_num_pages) > end_index)
1164                         wc->w_num_pages = end_index - start;
1165         } else {
1166                 wc->w_num_pages = 1;
1167                 start = target_index;
1168         }
1169
1170         for(i = 0; i < wc->w_num_pages; i++) {
1171                 index = start + i;
1172
1173                 if (index == target_index && mmap_page) {
1174                         /*
1175                          * ocfs2_pagemkwrite() is a little different
1176                          * and wants us to directly use the page
1177                          * passed in.
1178                          */
1179                         lock_page(mmap_page);
1180
1181                         /* Exit and let the caller retry */
1182                         if (mmap_page->mapping != mapping) {
1183                                 WARN_ON(mmap_page->mapping);
1184                                 unlock_page(mmap_page);
1185                                 ret = -EAGAIN;
1186                                 goto out;
1187                         }
1188
1189                         page_cache_get(mmap_page);
1190                         wc->w_pages[i] = mmap_page;
1191                         wc->w_target_locked = true;
1192                 } else {
1193                         wc->w_pages[i] = find_or_create_page(mapping, index,
1194                                                              GFP_NOFS);
1195                         if (!wc->w_pages[i]) {
1196                                 ret = -ENOMEM;
1197                                 mlog_errno(ret);
1198                                 goto out;
1199                         }
1200                 }
1201
1202                 if (index == target_index)
1203                         wc->w_target_page = wc->w_pages[i];
1204         }
1205 out:
1206         if (ret)
1207                 wc->w_target_locked = false;
1208         return ret;
1209 }
1210
1211 /*
1212  * Prepare a single cluster for write one cluster into the file.
1213  */
1214 static int ocfs2_write_cluster(struct address_space *mapping,
1215                                u32 phys, unsigned int unwritten,
1216                                unsigned int should_zero,
1217                                struct ocfs2_alloc_context *data_ac,
1218                                struct ocfs2_alloc_context *meta_ac,
1219                                struct ocfs2_write_ctxt *wc, u32 cpos,
1220                                loff_t user_pos, unsigned user_len)
1221 {
1222         int ret, i, new;
1223         u64 v_blkno, p_blkno;
1224         struct inode *inode = mapping->host;
1225         struct ocfs2_extent_tree et;
1226
1227         new = phys == 0 ? 1 : 0;
1228         if (new) {
1229                 u32 tmp_pos;
1230
1231                 /*
1232                  * This is safe to call with the page locks - it won't take
1233                  * any additional semaphores or cluster locks.
1234                  */
1235                 tmp_pos = cpos;
1236                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1237                                            &tmp_pos, 1, 0, wc->w_di_bh,
1238                                            wc->w_handle, data_ac,
1239                                            meta_ac, NULL);
1240                 /*
1241                  * This shouldn't happen because we must have already
1242                  * calculated the correct meta data allocation required. The
1243                  * internal tree allocation code should know how to increase
1244                  * transaction credits itself.
1245                  *
1246                  * If need be, we could handle -EAGAIN for a
1247                  * RESTART_TRANS here.
1248                  */
1249                 mlog_bug_on_msg(ret == -EAGAIN,
1250                                 "Inode %llu: EAGAIN return during allocation.\n",
1251                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1252                 if (ret < 0) {
1253                         mlog_errno(ret);
1254                         goto out;
1255                 }
1256         } else if (unwritten) {
1257                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1258                                               wc->w_di_bh);
1259                 ret = ocfs2_mark_extent_written(inode, &et,
1260                                                 wc->w_handle, cpos, 1, phys,
1261                                                 meta_ac, &wc->w_dealloc);
1262                 if (ret < 0) {
1263                         mlog_errno(ret);
1264                         goto out;
1265                 }
1266         }
1267
1268         if (should_zero)
1269                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1270         else
1271                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1272
1273         /*
1274          * The only reason this should fail is due to an inability to
1275          * find the extent added.
1276          */
1277         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1278                                           NULL);
1279         if (ret < 0) {
1280                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1281                             "at logical block %llu",
1282                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1283                             (unsigned long long)v_blkno);
1284                 goto out;
1285         }
1286
1287         BUG_ON(p_blkno == 0);
1288
1289         for(i = 0; i < wc->w_num_pages; i++) {
1290                 int tmpret;
1291
1292                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1293                                                       wc->w_pages[i], cpos,
1294                                                       user_pos, user_len,
1295                                                       should_zero);
1296                 if (tmpret) {
1297                         mlog_errno(tmpret);
1298                         if (ret == 0)
1299                                 ret = tmpret;
1300                 }
1301         }
1302
1303         /*
1304          * We only have cleanup to do in case of allocating write.
1305          */
1306         if (ret && new)
1307                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1308
1309 out:
1310
1311         return ret;
1312 }
1313
1314 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1315                                        struct ocfs2_alloc_context *data_ac,
1316                                        struct ocfs2_alloc_context *meta_ac,
1317                                        struct ocfs2_write_ctxt *wc,
1318                                        loff_t pos, unsigned len)
1319 {
1320         int ret, i;
1321         loff_t cluster_off;
1322         unsigned int local_len = len;
1323         struct ocfs2_write_cluster_desc *desc;
1324         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1325
1326         for (i = 0; i < wc->w_clen; i++) {
1327                 desc = &wc->w_desc[i];
1328
1329                 /*
1330                  * We have to make sure that the total write passed in
1331                  * doesn't extend past a single cluster.
1332                  */
1333                 local_len = len;
1334                 cluster_off = pos & (osb->s_clustersize - 1);
1335                 if ((cluster_off + local_len) > osb->s_clustersize)
1336                         local_len = osb->s_clustersize - cluster_off;
1337
1338                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1339                                           desc->c_unwritten,
1340                                           desc->c_needs_zero,
1341                                           data_ac, meta_ac,
1342                                           wc, desc->c_cpos, pos, local_len);
1343                 if (ret) {
1344                         mlog_errno(ret);
1345                         goto out;
1346                 }
1347
1348                 len -= local_len;
1349                 pos += local_len;
1350         }
1351
1352         ret = 0;
1353 out:
1354         return ret;
1355 }
1356
1357 /*
1358  * ocfs2_write_end() wants to know which parts of the target page it
1359  * should complete the write on. It's easiest to compute them ahead of
1360  * time when a more complete view of the write is available.
1361  */
1362 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1363                                         struct ocfs2_write_ctxt *wc,
1364                                         loff_t pos, unsigned len, int alloc)
1365 {
1366         struct ocfs2_write_cluster_desc *desc;
1367
1368         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1369         wc->w_target_to = wc->w_target_from + len;
1370
1371         if (alloc == 0)
1372                 return;
1373
1374         /*
1375          * Allocating write - we may have different boundaries based
1376          * on page size and cluster size.
1377          *
1378          * NOTE: We can no longer compute one value from the other as
1379          * the actual write length and user provided length may be
1380          * different.
1381          */
1382
1383         if (wc->w_large_pages) {
1384                 /*
1385                  * We only care about the 1st and last cluster within
1386                  * our range and whether they should be zero'd or not. Either
1387                  * value may be extended out to the start/end of a
1388                  * newly allocated cluster.
1389                  */
1390                 desc = &wc->w_desc[0];
1391                 if (desc->c_needs_zero)
1392                         ocfs2_figure_cluster_boundaries(osb,
1393                                                         desc->c_cpos,
1394                                                         &wc->w_target_from,
1395                                                         NULL);
1396
1397                 desc = &wc->w_desc[wc->w_clen - 1];
1398                 if (desc->c_needs_zero)
1399                         ocfs2_figure_cluster_boundaries(osb,
1400                                                         desc->c_cpos,
1401                                                         NULL,
1402                                                         &wc->w_target_to);
1403         } else {
1404                 wc->w_target_from = 0;
1405                 wc->w_target_to = PAGE_CACHE_SIZE;
1406         }
1407 }
1408
1409 /*
1410  * Populate each single-cluster write descriptor in the write context
1411  * with information about the i/o to be done.
1412  *
1413  * Returns the number of clusters that will have to be allocated, as
1414  * well as a worst case estimate of the number of extent records that
1415  * would have to be created during a write to an unwritten region.
1416  */
1417 static int ocfs2_populate_write_desc(struct inode *inode,
1418                                      struct ocfs2_write_ctxt *wc,
1419                                      unsigned int *clusters_to_alloc,
1420                                      unsigned int *extents_to_split)
1421 {
1422         int ret;
1423         struct ocfs2_write_cluster_desc *desc;
1424         unsigned int num_clusters = 0;
1425         unsigned int ext_flags = 0;
1426         u32 phys = 0;
1427         int i;
1428
1429         *clusters_to_alloc = 0;
1430         *extents_to_split = 0;
1431
1432         for (i = 0; i < wc->w_clen; i++) {
1433                 desc = &wc->w_desc[i];
1434                 desc->c_cpos = wc->w_cpos + i;
1435
1436                 if (num_clusters == 0) {
1437                         /*
1438                          * Need to look up the next extent record.
1439                          */
1440                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1441                                                  &num_clusters, &ext_flags);
1442                         if (ret) {
1443                                 mlog_errno(ret);
1444                                 goto out;
1445                         }
1446
1447                         /* We should already CoW the refcountd extent. */
1448                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
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                 /*
1470                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1471                  * file that got extended.  w_first_new_cpos tells us
1472                  * where the newly allocated clusters are so we can
1473                  * zero them.
1474                  */
1475                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1476                         BUG_ON(phys == 0);
1477                         desc->c_needs_zero = 1;
1478                 }
1479
1480                 desc->c_phys = phys;
1481                 if (phys == 0) {
1482                         desc->c_new = 1;
1483                         desc->c_needs_zero = 1;
1484                         *clusters_to_alloc = *clusters_to_alloc + 1;
1485                 }
1486
1487                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1488                         desc->c_unwritten = 1;
1489                         desc->c_needs_zero = 1;
1490                 }
1491
1492                 num_clusters--;
1493         }
1494
1495         ret = 0;
1496 out:
1497         return ret;
1498 }
1499
1500 static int ocfs2_write_begin_inline(struct address_space *mapping,
1501                                     struct inode *inode,
1502                                     struct ocfs2_write_ctxt *wc)
1503 {
1504         int ret;
1505         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1506         struct page *page;
1507         handle_t *handle;
1508         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1509
1510         page = find_or_create_page(mapping, 0, GFP_NOFS);
1511         if (!page) {
1512                 ret = -ENOMEM;
1513                 mlog_errno(ret);
1514                 goto out;
1515         }
1516         /*
1517          * If we don't set w_num_pages then this page won't get unlocked
1518          * and freed on cleanup of the write context.
1519          */
1520         wc->w_pages[0] = wc->w_target_page = page;
1521         wc->w_num_pages = 1;
1522
1523         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1524         if (IS_ERR(handle)) {
1525                 ret = PTR_ERR(handle);
1526                 mlog_errno(ret);
1527                 goto out;
1528         }
1529
1530         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1531                                       OCFS2_JOURNAL_ACCESS_WRITE);
1532         if (ret) {
1533                 ocfs2_commit_trans(osb, handle);
1534
1535                 mlog_errno(ret);
1536                 goto out;
1537         }
1538
1539         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1540                 ocfs2_set_inode_data_inline(inode, di);
1541
1542         if (!PageUptodate(page)) {
1543                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1544                 if (ret) {
1545                         ocfs2_commit_trans(osb, handle);
1546
1547                         goto out;
1548                 }
1549         }
1550
1551         wc->w_handle = handle;
1552 out:
1553         return ret;
1554 }
1555
1556 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1557 {
1558         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1559
1560         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1561                 return 1;
1562         return 0;
1563 }
1564
1565 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1566                                           struct inode *inode, loff_t pos,
1567                                           unsigned len, struct page *mmap_page,
1568                                           struct ocfs2_write_ctxt *wc)
1569 {
1570         int ret, written = 0;
1571         loff_t end = pos + len;
1572         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1573         struct ocfs2_dinode *di = NULL;
1574
1575         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1576                                              len, (unsigned long long)pos,
1577                                              oi->ip_dyn_features);
1578
1579         /*
1580          * Handle inodes which already have inline data 1st.
1581          */
1582         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1583                 if (mmap_page == NULL &&
1584                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1585                         goto do_inline_write;
1586
1587                 /*
1588                  * The write won't fit - we have to give this inode an
1589                  * inline extent list now.
1590                  */
1591                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1592                 if (ret)
1593                         mlog_errno(ret);
1594                 goto out;
1595         }
1596
1597         /*
1598          * Check whether the inode can accept inline data.
1599          */
1600         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1601                 return 0;
1602
1603         /*
1604          * Check whether the write can fit.
1605          */
1606         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1607         if (mmap_page ||
1608             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1609                 return 0;
1610
1611 do_inline_write:
1612         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1613         if (ret) {
1614                 mlog_errno(ret);
1615                 goto out;
1616         }
1617
1618         /*
1619          * This signals to the caller that the data can be written
1620          * inline.
1621          */
1622         written = 1;
1623 out:
1624         return written ? written : ret;
1625 }
1626
1627 /*
1628  * This function only does anything for file systems which can't
1629  * handle sparse files.
1630  *
1631  * What we want to do here is fill in any hole between the current end
1632  * of allocation and the end of our write. That way the rest of the
1633  * write path can treat it as an non-allocating write, which has no
1634  * special case code for sparse/nonsparse files.
1635  */
1636 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1637                                         struct buffer_head *di_bh,
1638                                         loff_t pos, unsigned len,
1639                                         struct ocfs2_write_ctxt *wc)
1640 {
1641         int ret;
1642         loff_t newsize = pos + len;
1643
1644         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1645
1646         if (newsize <= i_size_read(inode))
1647                 return 0;
1648
1649         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1650         if (ret)
1651                 mlog_errno(ret);
1652
1653         wc->w_first_new_cpos =
1654                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1655
1656         return ret;
1657 }
1658
1659 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1660                            loff_t pos)
1661 {
1662         int ret = 0;
1663
1664         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1665         if (pos > i_size_read(inode))
1666                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1667
1668         return ret;
1669 }
1670
1671 /*
1672  * Try to flush truncate logs if we can free enough clusters from it.
1673  * As for return value, "< 0" means error, "0" no space and "1" means
1674  * we have freed enough spaces and let the caller try to allocate again.
1675  */
1676 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1677                                           unsigned int needed)
1678 {
1679         tid_t target;
1680         int ret = 0;
1681         unsigned int truncated_clusters;
1682
1683         mutex_lock(&osb->osb_tl_inode->i_mutex);
1684         truncated_clusters = osb->truncated_clusters;
1685         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1686
1687         /*
1688          * Check whether we can succeed in allocating if we free
1689          * the truncate log.
1690          */
1691         if (truncated_clusters < needed)
1692                 goto out;
1693
1694         ret = ocfs2_flush_truncate_log(osb);
1695         if (ret) {
1696                 mlog_errno(ret);
1697                 goto out;
1698         }
1699
1700         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1701                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1702                 ret = 1;
1703         }
1704 out:
1705         return ret;
1706 }
1707
1708 int ocfs2_write_begin_nolock(struct file *filp,
1709                              struct address_space *mapping,
1710                              loff_t pos, unsigned len, unsigned flags,
1711                              struct page **pagep, void **fsdata,
1712                              struct buffer_head *di_bh, struct page *mmap_page)
1713 {
1714         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1715         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1716         struct ocfs2_write_ctxt *wc;
1717         struct inode *inode = mapping->host;
1718         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1719         struct ocfs2_dinode *di;
1720         struct ocfs2_alloc_context *data_ac = NULL;
1721         struct ocfs2_alloc_context *meta_ac = NULL;
1722         handle_t *handle;
1723         struct ocfs2_extent_tree et;
1724         int try_free = 1, ret1;
1725
1726 try_again:
1727         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1728         if (ret) {
1729                 mlog_errno(ret);
1730                 return ret;
1731         }
1732
1733         if (ocfs2_supports_inline_data(osb)) {
1734                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1735                                                      mmap_page, wc);
1736                 if (ret == 1) {
1737                         ret = 0;
1738                         goto success;
1739                 }
1740                 if (ret < 0) {
1741                         mlog_errno(ret);
1742                         goto out;
1743                 }
1744         }
1745
1746         if (ocfs2_sparse_alloc(osb))
1747                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1748         else
1749                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1750                                                    wc);
1751         if (ret) {
1752                 mlog_errno(ret);
1753                 goto out;
1754         }
1755
1756         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1757         if (ret < 0) {
1758                 mlog_errno(ret);
1759                 goto out;
1760         } else if (ret == 1) {
1761                 clusters_need = wc->w_clen;
1762                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1763                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1764                 if (ret) {
1765                         mlog_errno(ret);
1766                         goto out;
1767                 }
1768         }
1769
1770         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1771                                         &extents_to_split);
1772         if (ret) {
1773                 mlog_errno(ret);
1774                 goto out;
1775         }
1776         clusters_need += clusters_to_alloc;
1777
1778         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1779
1780         trace_ocfs2_write_begin_nolock(
1781                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1782                         (long long)i_size_read(inode),
1783                         le32_to_cpu(di->i_clusters),
1784                         pos, len, flags, mmap_page,
1785                         clusters_to_alloc, extents_to_split);
1786
1787         /*
1788          * We set w_target_from, w_target_to here so that
1789          * ocfs2_write_end() knows which range in the target page to
1790          * write out. An allocation requires that we write the entire
1791          * cluster range.
1792          */
1793         if (clusters_to_alloc || extents_to_split) {
1794                 /*
1795                  * XXX: We are stretching the limits of
1796                  * ocfs2_lock_allocators(). It greatly over-estimates
1797                  * the work to be done.
1798                  */
1799                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1800                                               wc->w_di_bh);
1801                 ret = ocfs2_lock_allocators(inode, &et,
1802                                             clusters_to_alloc, extents_to_split,
1803                                             &data_ac, &meta_ac);
1804                 if (ret) {
1805                         mlog_errno(ret);
1806                         goto out;
1807                 }
1808
1809                 if (data_ac)
1810                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1811
1812                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1813                                                     &di->id2.i_list,
1814                                                     clusters_to_alloc);
1815
1816         }
1817
1818         /*
1819          * We have to zero sparse allocated clusters, unwritten extent clusters,
1820          * and non-sparse clusters we just extended.  For non-sparse writes,
1821          * we know zeros will only be needed in the first and/or last cluster.
1822          */
1823         if (clusters_to_alloc || extents_to_split ||
1824             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1825                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1826                 cluster_of_pages = 1;
1827         else
1828                 cluster_of_pages = 0;
1829
1830         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1831
1832         handle = ocfs2_start_trans(osb, credits);
1833         if (IS_ERR(handle)) {
1834                 ret = PTR_ERR(handle);
1835                 mlog_errno(ret);
1836                 goto out;
1837         }
1838
1839         wc->w_handle = handle;
1840
1841         if (clusters_to_alloc) {
1842                 ret = dquot_alloc_space_nodirty(inode,
1843                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1844                 if (ret)
1845                         goto out_commit;
1846         }
1847         /*
1848          * We don't want this to fail in ocfs2_write_end(), so do it
1849          * here.
1850          */
1851         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1852                                       OCFS2_JOURNAL_ACCESS_WRITE);
1853         if (ret) {
1854                 mlog_errno(ret);
1855                 goto out_quota;
1856         }
1857
1858         /*
1859          * Fill our page array first. That way we've grabbed enough so
1860          * that we can zero and flush if we error after adding the
1861          * extent.
1862          */
1863         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1864                                          cluster_of_pages, mmap_page);
1865         if (ret && ret != -EAGAIN) {
1866                 mlog_errno(ret);
1867                 goto out_quota;
1868         }
1869
1870         /*
1871          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1872          * the target page. In this case, we exit with no error and no target
1873          * page. This will trigger the caller, page_mkwrite(), to re-try
1874          * the operation.
1875          */
1876         if (ret == -EAGAIN) {
1877                 BUG_ON(wc->w_target_page);
1878                 ret = 0;
1879                 goto out_quota;
1880         }
1881
1882         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1883                                           len);
1884         if (ret) {
1885                 mlog_errno(ret);
1886                 goto out_quota;
1887         }
1888
1889         if (data_ac)
1890                 ocfs2_free_alloc_context(data_ac);
1891         if (meta_ac)
1892                 ocfs2_free_alloc_context(meta_ac);
1893
1894 success:
1895         *pagep = wc->w_target_page;
1896         *fsdata = wc;
1897         return 0;
1898 out_quota:
1899         if (clusters_to_alloc)
1900                 dquot_free_space(inode,
1901                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1902 out_commit:
1903         ocfs2_commit_trans(osb, handle);
1904
1905 out:
1906         ocfs2_free_write_ctxt(wc);
1907
1908         if (data_ac)
1909                 ocfs2_free_alloc_context(data_ac);
1910         if (meta_ac)
1911                 ocfs2_free_alloc_context(meta_ac);
1912
1913         if (ret == -ENOSPC && try_free) {
1914                 /*
1915                  * Try to free some truncate log so that we can have enough
1916                  * clusters to allocate.
1917                  */
1918                 try_free = 0;
1919
1920                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1921                 if (ret1 == 1)
1922                         goto try_again;
1923
1924                 if (ret1 < 0)
1925                         mlog_errno(ret1);
1926         }
1927
1928         return ret;
1929 }
1930
1931 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1932                              loff_t pos, unsigned len, unsigned flags,
1933                              struct page **pagep, void **fsdata)
1934 {
1935         int ret;
1936         struct buffer_head *di_bh = NULL;
1937         struct inode *inode = mapping->host;
1938
1939         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1940         if (ret) {
1941                 mlog_errno(ret);
1942                 return ret;
1943         }
1944
1945         /*
1946          * Take alloc sem here to prevent concurrent lookups. That way
1947          * the mapping, zeroing and tree manipulation within
1948          * ocfs2_write() will be safe against ->readpage(). This
1949          * should also serve to lock out allocation from a shared
1950          * writeable region.
1951          */
1952         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1953
1954         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1955                                        fsdata, di_bh, NULL);
1956         if (ret) {
1957                 mlog_errno(ret);
1958                 goto out_fail;
1959         }
1960
1961         brelse(di_bh);
1962
1963         return 0;
1964
1965 out_fail:
1966         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1967
1968         brelse(di_bh);
1969         ocfs2_inode_unlock(inode, 1);
1970
1971         return ret;
1972 }
1973
1974 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1975                                    unsigned len, unsigned *copied,
1976                                    struct ocfs2_dinode *di,
1977                                    struct ocfs2_write_ctxt *wc)
1978 {
1979         void *kaddr;
1980
1981         if (unlikely(*copied < len)) {
1982                 if (!PageUptodate(wc->w_target_page)) {
1983                         *copied = 0;
1984                         return;
1985                 }
1986         }
1987
1988         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1989         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1990         kunmap_atomic(kaddr, KM_USER0);
1991
1992         trace_ocfs2_write_end_inline(
1993              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1994              (unsigned long long)pos, *copied,
1995              le16_to_cpu(di->id2.i_data.id_count),
1996              le16_to_cpu(di->i_dyn_features));
1997 }
1998
1999 int ocfs2_write_end_nolock(struct address_space *mapping,
2000                            loff_t pos, unsigned len, unsigned copied,
2001                            struct page *page, void *fsdata)
2002 {
2003         int i;
2004         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2005         struct inode *inode = mapping->host;
2006         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2007         struct ocfs2_write_ctxt *wc = fsdata;
2008         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2009         handle_t *handle = wc->w_handle;
2010         struct page *tmppage;
2011
2012         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2013                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2014                 goto out_write_size;
2015         }
2016
2017         if (unlikely(copied < len)) {
2018                 if (!PageUptodate(wc->w_target_page))
2019                         copied = 0;
2020
2021                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2022                                        start+len);
2023         }
2024         flush_dcache_page(wc->w_target_page);
2025
2026         for(i = 0; i < wc->w_num_pages; i++) {
2027                 tmppage = wc->w_pages[i];
2028
2029                 if (tmppage == wc->w_target_page) {
2030                         from = wc->w_target_from;
2031                         to = wc->w_target_to;
2032
2033                         BUG_ON(from > PAGE_CACHE_SIZE ||
2034                                to > PAGE_CACHE_SIZE ||
2035                                to < from);
2036                 } else {
2037                         /*
2038                          * Pages adjacent to the target (if any) imply
2039                          * a hole-filling write in which case we want
2040                          * to flush their entire range.
2041                          */
2042                         from = 0;
2043                         to = PAGE_CACHE_SIZE;
2044                 }
2045
2046                 if (page_has_buffers(tmppage)) {
2047                         if (ocfs2_should_order_data(inode))
2048                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2049                         block_commit_write(tmppage, from, to);
2050                 }
2051         }
2052
2053 out_write_size:
2054         pos += copied;
2055         if (pos > inode->i_size) {
2056                 i_size_write(inode, pos);
2057                 mark_inode_dirty(inode);
2058         }
2059         inode->i_blocks = ocfs2_inode_sector_count(inode);
2060         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2061         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2062         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2063         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2064         ocfs2_journal_dirty(handle, wc->w_di_bh);
2065
2066         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2067          * lock, or it will cause a deadlock since journal commit threads holds
2068          * this lock and will ask for the page lock when flushing the data.
2069          * put it here to preserve the unlock order.
2070          */
2071         ocfs2_unlock_pages(wc);
2072
2073         ocfs2_commit_trans(osb, handle);
2074
2075         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2076
2077         brelse(wc->w_di_bh);
2078         kfree(wc);
2079
2080         return copied;
2081 }
2082
2083 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2084                            loff_t pos, unsigned len, unsigned copied,
2085                            struct page *page, void *fsdata)
2086 {
2087         int ret;
2088         struct inode *inode = mapping->host;
2089
2090         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2091
2092         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2093         ocfs2_inode_unlock(inode, 1);
2094
2095         return ret;
2096 }
2097
2098 const struct address_space_operations ocfs2_aops = {
2099         .readpage               = ocfs2_readpage,
2100         .readpages              = ocfs2_readpages,
2101         .writepage              = ocfs2_writepage,
2102         .write_begin            = ocfs2_write_begin,
2103         .write_end              = ocfs2_write_end,
2104         .bmap                   = ocfs2_bmap,
2105         .direct_IO              = ocfs2_direct_IO,
2106         .invalidatepage         = ocfs2_invalidatepage,
2107         .releasepage            = ocfs2_releasepage,
2108         .migratepage            = buffer_migrate_page,
2109         .is_partially_uptodate  = block_is_partially_uptodate,
2110         .error_remove_page      = generic_error_remove_page,
2111 };