fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37 #include <linux/falloc.h>
  38 #include <linux/quotaops.h>
  39
  40 #define MLOG_MASK_PREFIX ML_INODE
  41 #include <cluster/masklog.h>
  42
  43 #include "ocfs2.h"
  44
  45 #include "alloc.h"
  46 #include "aops.h"
  47 #include "dir.h"
  48 #include "dlmglue.h"
  49 #include "extent_map.h"
  50 #include "file.h"
  51 #include "sysfile.h"
  52 #include "inode.h"
  53 #include "ioctl.h"
  54 #include "journal.h"
  55 #include "locks.h"
  56 #include "mmap.h"
  57 #include "suballoc.h"
  58 #include "super.h"
  59 #include "xattr.h"
  60 #include "acl.h"
  61 #include "quota.h"
  62 #include "refcounttree.h"
  63
  64 #include "buffer_head_io.h"
  65
  66 static int ocfs2_sync_inode(struct inode *inode)
  67 {
  68         filemap_fdatawrite(inode->i_mapping);
  69         return sync_mapping_buffers(inode->i_mapping);
  70 }
  71
  72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  73 {
  74         struct ocfs2_file_private *fp;
  75
  76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  77         if (!fp)
  78                 return -ENOMEM;
  79
  80         fp->fp_file = file;
  81         mutex_init(&fp->fp_mutex);
  82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  83         file->private_data = fp;
  84
  85         return 0;
  86 }
  87
  88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  89 {
  90         struct ocfs2_file_private *fp = file->private_data;
  91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  92
  93         if (fp) {
  94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  95                 ocfs2_lock_res_free(&fp->fp_flock);
  96                 kfree(fp);
  97                 file->private_data = NULL;
  98         }
  99 }
 100
 101 static int ocfs2_file_open(struct inode *inode, struct file *file)
 102 {
 103         int status;
 104         int mode = file->f_flags;
 105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 106
 107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 109
 110         spin_lock(&oi->ip_lock);
 111
 112         /* Check that the inode hasn't been wiped from disk by another
 113          * node. If it hasn't then we're safe as long as we hold the
 114          * spin lock until our increment of open count. */
 115         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 116                 spin_unlock(&oi->ip_lock);
 117
 118                 status = -ENOENT;
 119                 goto leave;
 120         }
 121
 122         if (mode & O_DIRECT)
 123                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 124
 125         oi->ip_open_count++;
 126         spin_unlock(&oi->ip_lock);
 127
 128         status = ocfs2_init_file_private(inode, file);
 129         if (status) {
 130                 /*
 131                  * We want to set open count back if we're failing the
 132                  * open.
 133                  */
 134                 spin_lock(&oi->ip_lock);
 135                 oi->ip_open_count--;
 136                 spin_unlock(&oi->ip_lock);
 137         }
 138
 139 leave:
 140         mlog_exit(status);
 141         return status;
 142 }
 143
 144 static int ocfs2_file_release(struct inode *inode, struct file *file)
 145 {
 146         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 147
 148         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 149                        file->f_path.dentry->d_name.len,
 150                        file->f_path.dentry->d_name.name);
 151
 152         spin_lock(&oi->ip_lock);
 153         if (!--oi->ip_open_count)
 154                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 155         spin_unlock(&oi->ip_lock);
 156
 157         ocfs2_free_file_private(inode, file);
 158
 159         mlog_exit(0);
 160
 161         return 0;
 162 }
 163
 164 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 165 {
 166         return ocfs2_init_file_private(inode, file);
 167 }
 168
 169 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 170 {
 171         ocfs2_free_file_private(inode, file);
 172         return 0;
 173 }
 174
 175 static int ocfs2_sync_file(struct file *file,
 176                            struct dentry *dentry,
 177                            int datasync)
 178 {
 179         int err = 0;
 180         journal_t *journal;
 181         struct inode *inode = dentry->d_inode;
 182         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 183
 184         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 185                    dentry->d_name.len, dentry->d_name.name);
 186
 187         err = ocfs2_sync_inode(dentry->d_inode);
 188         if (err)
 189                 goto bail;
 190
 191         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
 192                 goto bail;
 193
 194         journal = osb->journal->j_journal;
 195         err = jbd2_journal_force_commit(journal);
 196
 197 bail:
 198         mlog_exit(err);
 199
 200         return (err < 0) ? -EIO : 0;
 201 }
 202
 203 int ocfs2_should_update_atime(struct inode *inode,
 204                               struct vfsmount *vfsmnt)
 205 {
 206         struct timespec now;
 207         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 208
 209         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 210                 return 0;
 211
 212         if ((inode->i_flags & S_NOATIME) ||
 213             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 214                 return 0;
 215
 216         /*
 217          * We can be called with no vfsmnt structure - NFSD will
 218          * sometimes do this.
 219          *
 220          * Note that our action here is different than touch_atime() -
 221          * if we can't tell whether this is a noatime mount, then we
 222          * don't know whether to trust the value of s_atime_quantum.
 223          */
 224         if (vfsmnt == NULL)
 225                 return 0;
 226
 227         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 228             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 229                 return 0;
 230
 231         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 232                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 233                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 234                         return 1;
 235
 236                 return 0;
 237         }
 238
 239         now = CURRENT_TIME;
 240         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 241                 return 0;
 242         else
 243                 return 1;
 244 }
 245
 246 int ocfs2_update_inode_atime(struct inode *inode,
 247                              struct buffer_head *bh)
 248 {
 249         int ret;
 250         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 251         handle_t *handle;
 252         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 253
 254         mlog_entry_void();
 255
 256         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 257         if (IS_ERR(handle)) {
 258                 ret = PTR_ERR(handle);
 259                 mlog_errno(ret);
 260                 goto out;
 261         }
 262
 263         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 264                                       OCFS2_JOURNAL_ACCESS_WRITE);
 265         if (ret) {
 266                 mlog_errno(ret);
 267                 goto out_commit;
 268         }
 269
 270         /*
 271          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 272          * have i_mutex to guard against concurrent changes to other
 273          * inode fields.
 274          */
 275         inode->i_atime = CURRENT_TIME;
 276         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 277         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 278
 279         ret = ocfs2_journal_dirty(handle, bh);
 280         if (ret < 0)
 281                 mlog_errno(ret);
 282
 283 out_commit:
 284         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 285 out:
 286         mlog_exit(ret);
 287         return ret;
 288 }
 289
 290 static int ocfs2_set_inode_size(handle_t *handle,
 291                                 struct inode *inode,
 292                                 struct buffer_head *fe_bh,
 293                                 u64 new_i_size)
 294 {
 295         int status;
 296
 297         mlog_entry_void();
 298         i_size_write(inode, new_i_size);
 299         inode->i_blocks = ocfs2_inode_sector_count(inode);
 300         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 301
 302         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 303         if (status < 0) {
 304                 mlog_errno(status);
 305                 goto bail;
 306         }
 307
 308 bail:
 309         mlog_exit(status);
 310         return status;
 311 }
 312
 313 int ocfs2_simple_size_update(struct inode *inode,
 314                              struct buffer_head *di_bh,
 315                              u64 new_i_size)
 316 {
 317         int ret;
 318         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 319         handle_t *handle = NULL;
 320
 321         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 322         if (IS_ERR(handle)) {
 323                 ret = PTR_ERR(handle);
 324                 mlog_errno(ret);
 325                 goto out;
 326         }
 327
 328         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 329                                    new_i_size);
 330         if (ret < 0)
 331                 mlog_errno(ret);
 332
 333         ocfs2_commit_trans(osb, handle);
 334 out:
 335         return ret;
 336 }
 337
 338 static int ocfs2_cow_file_pos(struct inode *inode,
 339                               struct buffer_head *fe_bh,
 340                               u64 offset)
 341 {
 342         int status;
 343         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 344         unsigned int num_clusters = 0;
 345         unsigned int ext_flags = 0;
 346
 347         /*
 348          * If the new offset is aligned to the range of the cluster, there is
 349          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 350          * CoW either.
 351          */
 352         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 353                 return 0;
 354
 355         status = ocfs2_get_clusters(inode, cpos, &phys,
 356                                     &num_clusters, &ext_flags);
 357         if (status) {
 358                 mlog_errno(status);
 359                 goto out;
 360         }
 361
 362         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 363                 goto out;
 364
 365         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 366
 367 out:
 368         return status;
 369 }
 370
 371 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 372                                      struct inode *inode,
 373                                      struct buffer_head *fe_bh,
 374                                      u64 new_i_size)
 375 {
 376         int status;
 377         handle_t *handle;
 378         struct ocfs2_dinode *di;
 379         u64 cluster_bytes;
 380
 381         mlog_entry_void();
 382
 383         /*
 384          * We need to CoW the cluster contains the offset if it is reflinked
 385          * since we will call ocfs2_zero_range_for_truncate later which will
 386          * write "0" from offset to the end of the cluster.
 387          */
 388         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 389         if (status) {
 390                 mlog_errno(status);
 391                 return status;
 392         }
 393
 394         /* TODO: This needs to actually orphan the inode in this
 395          * transaction. */
 396
 397         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 398         if (IS_ERR(handle)) {
 399                 status = PTR_ERR(handle);
 400                 mlog_errno(status);
 401                 goto out;
 402         }
 403
 404         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 405                                          OCFS2_JOURNAL_ACCESS_WRITE);
 406         if (status < 0) {
 407                 mlog_errno(status);
 408                 goto out_commit;
 409         }
 410
 411         /*
 412          * Do this before setting i_size.
 413          */
 414         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 415         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 416                                                cluster_bytes);
 417         if (status) {
 418                 mlog_errno(status);
 419                 goto out_commit;
 420         }
 421
 422         i_size_write(inode, new_i_size);
 423         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 424
 425         di = (struct ocfs2_dinode *) fe_bh->b_data;
 426         di->i_size = cpu_to_le64(new_i_size);
 427         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 428         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 429
 430         status = ocfs2_journal_dirty(handle, fe_bh);
 431         if (status < 0)
 432                 mlog_errno(status);
 433
 434 out_commit:
 435         ocfs2_commit_trans(osb, handle);
 436 out:
 437
 438         mlog_exit(status);
 439         return status;
 440 }
 441
 442 static int ocfs2_truncate_file(struct inode *inode,
 443                                struct buffer_head *di_bh,
 444                                u64 new_i_size)
 445 {
 446         int status = 0;
 447         struct ocfs2_dinode *fe = NULL;
 448         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 449         struct ocfs2_truncate_context *tc = NULL;
 450
 451         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 452                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 453                    (unsigned long long)new_i_size);
 454
 455         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 456          * already validated it */
 457         fe = (struct ocfs2_dinode *) di_bh->b_data;
 458
 459         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 460                         "Inode %llu, inode i_size = %lld != di "
 461                         "i_size = %llu, i_flags = 0x%x\n",
 462                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 463                         i_size_read(inode),
 464                         (unsigned long long)le64_to_cpu(fe->i_size),
 465                         le32_to_cpu(fe->i_flags));
 466
 467         if (new_i_size > le64_to_cpu(fe->i_size)) {
 468                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 469                      (unsigned long long)le64_to_cpu(fe->i_size),
 470                      (unsigned long long)new_i_size);
 471                 status = -EINVAL;
 472                 mlog_errno(status);
 473                 goto bail;
 474         }
 475
 476         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 477              (unsigned long long)le64_to_cpu(fe->i_blkno),
 478              (unsigned long long)le64_to_cpu(fe->i_size),
 479              (unsigned long long)new_i_size);
 480
 481         /* lets handle the simple truncate cases before doing any more
 482          * cluster locking. */
 483         if (new_i_size == le64_to_cpu(fe->i_size))
 484                 goto bail;
 485
 486         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 487
 488         /*
 489          * The inode lock forced other nodes to sync and drop their
 490          * pages, which (correctly) happens even if we have a truncate
 491          * without allocation change - ocfs2 cluster sizes can be much
 492          * greater than page size, so we have to truncate them
 493          * anyway.
 494          */
 495         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 496         truncate_inode_pages(inode->i_mapping, new_i_size);
 497
 498         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 499                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 500                                                i_size_read(inode), 1);
 501                 if (status)
 502                         mlog_errno(status);
 503
 504                 goto bail_unlock_sem;
 505         }
 506
 507         /* alright, we're going to need to do a full blown alloc size
 508          * change. Orphan the inode so that recovery can complete the
 509          * truncate if necessary. This does the task of marking
 510          * i_size. */
 511         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 512         if (status < 0) {
 513                 mlog_errno(status);
 514                 goto bail_unlock_sem;
 515         }
 516
 517         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 518         if (status < 0) {
 519                 mlog_errno(status);
 520                 goto bail_unlock_sem;
 521         }
 522
 523         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 524         if (status < 0) {
 525                 mlog_errno(status);
 526                 goto bail_unlock_sem;
 527         }
 528
 529         /* TODO: orphan dir cleanup here. */
 530 bail_unlock_sem:
 531         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 532
 533 bail:
 534         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 535                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 536
 537         mlog_exit(status);
 538         return status;
 539 }
 540
 541 /*
 542  * extend file allocation only here.
 543  * we'll update all the disk stuff, and oip->alloc_size
 544  *
 545  * expect stuff to be locked, a transaction started and enough data /
 546  * metadata reservations in the contexts.
 547  *
 548  * Will return -EAGAIN, and a reason if a restart is needed.
 549  * If passed in, *reason will always be set, even in error.
 550  */
 551 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 552                          struct inode *inode,
 553                          u32 *logical_offset,
 554                          u32 clusters_to_add,
 555                          int mark_unwritten,
 556                          struct buffer_head *fe_bh,
 557                          handle_t *handle,
 558                          struct ocfs2_alloc_context *data_ac,
 559                          struct ocfs2_alloc_context *meta_ac,
 560                          enum ocfs2_alloc_restarted *reason_ret)
 561 {
 562         int ret;
 563         struct ocfs2_extent_tree et;
 564
 565         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 566         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 567                                           clusters_to_add, mark_unwritten,
 568                                           data_ac, meta_ac, reason_ret);
 569
 570         return ret;
 571 }
 572
 573 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 574                                      u32 clusters_to_add, int mark_unwritten)
 575 {
 576         int status = 0;
 577         int restart_func = 0;
 578         int credits;
 579         u32 prev_clusters;
 580         struct buffer_head *bh = NULL;
 581         struct ocfs2_dinode *fe = NULL;
 582         handle_t *handle = NULL;
 583         struct ocfs2_alloc_context *data_ac = NULL;
 584         struct ocfs2_alloc_context *meta_ac = NULL;
 585         enum ocfs2_alloc_restarted why;
 586         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 587         struct ocfs2_extent_tree et;
 588         int did_quota = 0;
 589
 590         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 591
 592         /*
 593          * This function only exists for file systems which don't
 594          * support holes.
 595          */
 596         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 597
 598         status = ocfs2_read_inode_block(inode, &bh);
 599         if (status < 0) {
 600                 mlog_errno(status);
 601                 goto leave;
 602         }
 603         fe = (struct ocfs2_dinode *) bh->b_data;
 604
 605 restart_all:
 606         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 607
 608         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 609              "clusters_to_add = %u\n",
 610              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 611              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 612              clusters_to_add);
 613         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 614         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 615                                        &data_ac, &meta_ac);
 616         if (status) {
 617                 mlog_errno(status);
 618                 goto leave;
 619         }
 620
 621         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 622                                             clusters_to_add);
 623         handle = ocfs2_start_trans(osb, credits);
 624         if (IS_ERR(handle)) {
 625                 status = PTR_ERR(handle);
 626                 handle = NULL;
 627                 mlog_errno(status);
 628                 goto leave;
 629         }
 630
 631 restarted_transaction:
 632         if (vfs_dq_alloc_space_nodirty(inode, ocfs2_clusters_to_bytes(osb->sb,
 633             clusters_to_add))) {
 634                 status = -EDQUOT;
 635                 goto leave;
 636         }
 637         did_quota = 1;
 638
 639         /* reserve a write to the file entry early on - that we if we
 640          * run out of credits in the allocation path, we can still
 641          * update i_size. */
 642         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 643                                          OCFS2_JOURNAL_ACCESS_WRITE);
 644         if (status < 0) {
 645                 mlog_errno(status);
 646                 goto leave;
 647         }
 648
 649         prev_clusters = OCFS2_I(inode)->ip_clusters;
 650
 651         status = ocfs2_add_inode_data(osb,
 652                                       inode,
 653                                       &logical_start,
 654                                       clusters_to_add,
 655                                       mark_unwritten,
 656                                       bh,
 657                                       handle,
 658                                       data_ac,
 659                                       meta_ac,
 660                                       &why);
 661         if ((status < 0) && (status != -EAGAIN)) {
 662                 if (status != -ENOSPC)
 663                         mlog_errno(status);
 664                 goto leave;
 665         }
 666
 667         status = ocfs2_journal_dirty(handle, bh);
 668         if (status < 0) {
 669                 mlog_errno(status);
 670                 goto leave;
 671         }
 672
 673         spin_lock(&OCFS2_I(inode)->ip_lock);
 674         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 675         spin_unlock(&OCFS2_I(inode)->ip_lock);
 676         /* Release unused quota reservation */
 677         vfs_dq_free_space(inode,
 678                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 679         did_quota = 0;
 680
 681         if (why != RESTART_NONE && clusters_to_add) {
 682                 if (why == RESTART_META) {
 683                         mlog(0, "restarting function.\n");
 684                         restart_func = 1;
 685                 } else {
 686                         BUG_ON(why != RESTART_TRANS);
 687
 688                         mlog(0, "restarting transaction.\n");
 689                         /* TODO: This can be more intelligent. */
 690                         credits = ocfs2_calc_extend_credits(osb->sb,
 691                                                             &fe->id2.i_list,
 692                                                             clusters_to_add);
 693                         status = ocfs2_extend_trans(handle, credits);
 694                         if (status < 0) {
 695                                 /* handle still has to be committed at
 696                                  * this point. */
 697                                 status = -ENOMEM;
 698                                 mlog_errno(status);
 699                                 goto leave;
 700                         }
 701                         goto restarted_transaction;
 702                 }
 703         }
 704
 705         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 706              le32_to_cpu(fe->i_clusters),
 707              (unsigned long long)le64_to_cpu(fe->i_size));
 708         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 709              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 710
 711 leave:
 712         if (status < 0 && did_quota)
 713                 vfs_dq_free_space(inode,
 714                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 715         if (handle) {
 716                 ocfs2_commit_trans(osb, handle);
 717                 handle = NULL;
 718         }
 719         if (data_ac) {
 720                 ocfs2_free_alloc_context(data_ac);
 721                 data_ac = NULL;
 722         }
 723         if (meta_ac) {
 724                 ocfs2_free_alloc_context(meta_ac);
 725                 meta_ac = NULL;
 726         }
 727         if ((!status) && restart_func) {
 728                 restart_func = 0;
 729                 goto restart_all;
 730         }
 731         brelse(bh);
 732         bh = NULL;
 733
 734         mlog_exit(status);
 735         return status;
 736 }
 737
 738 /* Some parts of this taken from generic_cont_expand, which turned out
 739  * to be too fragile to do exactly what we need without us having to
 740  * worry about recursive locking in ->write_begin() and ->write_end(). */
 741 static int ocfs2_write_zero_page(struct inode *inode,
 742                                  u64 size)
 743 {
 744         struct address_space *mapping = inode->i_mapping;
 745         struct page *page;
 746         unsigned long index;
 747         unsigned int offset;
 748         handle_t *handle = NULL;
 749         int ret;
 750
 751         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 752         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 753         ** skip the prepare.  make sure we never send an offset for the start
 754         ** of a block
 755         */
 756         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 757                 offset++;
 758         }
 759         index = size >> PAGE_CACHE_SHIFT;
 760
 761         page = grab_cache_page(mapping, index);
 762         if (!page) {
 763                 ret = -ENOMEM;
 764                 mlog_errno(ret);
 765                 goto out;
 766         }
 767
 768         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 769         if (ret < 0) {
 770                 mlog_errno(ret);
 771                 goto out_unlock;
 772         }
 773
 774         if (ocfs2_should_order_data(inode)) {
 775                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 776                                                      offset);
 777                 if (IS_ERR(handle)) {
 778                         ret = PTR_ERR(handle);
 779                         handle = NULL;
 780                         goto out_unlock;
 781                 }
 782         }
 783
 784         /* must not update i_size! */
 785         ret = block_commit_write(page, offset, offset);
 786         if (ret < 0)
 787                 mlog_errno(ret);
 788         else
 789                 ret = 0;
 790
 791         if (handle)
 792                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 793 out_unlock:
 794         unlock_page(page);
 795         page_cache_release(page);
 796 out:
 797         return ret;
 798 }
 799
 800 static int ocfs2_zero_extend(struct inode *inode,
 801                              u64 zero_to_size)
 802 {
 803         int ret = 0;
 804         u64 start_off;
 805         struct super_block *sb = inode->i_sb;
 806
 807         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 808         while (start_off < zero_to_size) {
 809                 ret = ocfs2_write_zero_page(inode, start_off);
 810                 if (ret < 0) {
 811                         mlog_errno(ret);
 812                         goto out;
 813                 }
 814
 815                 start_off += sb->s_blocksize;
 816
 817                 /*
 818                  * Very large extends have the potential to lock up
 819                  * the cpu for extended periods of time.
 820                  */
 821                 cond_resched();
 822         }
 823
 824 out:
 825         return ret;
 826 }
 827
 828 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 829 {
 830         int ret;
 831         u32 clusters_to_add;
 832         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 833
 834         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 835         if (clusters_to_add < oi->ip_clusters)
 836                 clusters_to_add = 0;
 837         else
 838                 clusters_to_add -= oi->ip_clusters;
 839
 840         if (clusters_to_add) {
 841                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 842                                                 clusters_to_add, 0);
 843                 if (ret) {
 844                         mlog_errno(ret);
 845                         goto out;
 846                 }
 847         }
 848
 849         /*
 850          * Call this even if we don't add any clusters to the tree. We
 851          * still need to zero the area between the old i_size and the
 852          * new i_size.
 853          */
 854         ret = ocfs2_zero_extend(inode, zero_to);
 855         if (ret < 0)
 856                 mlog_errno(ret);
 857
 858 out:
 859         return ret;
 860 }
 861
 862 static int ocfs2_extend_file(struct inode *inode,
 863                              struct buffer_head *di_bh,
 864                              u64 new_i_size)
 865 {
 866         int ret = 0;
 867         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 868
 869         BUG_ON(!di_bh);
 870
 871         /* setattr sometimes calls us like this. */
 872         if (new_i_size == 0)
 873                 goto out;
 874
 875         if (i_size_read(inode) == new_i_size)
 876                 goto out;
 877         BUG_ON(new_i_size < i_size_read(inode));
 878
 879         /*
 880          * Fall through for converting inline data, even if the fs
 881          * supports sparse files.
 882          *
 883          * The check for inline data here is legal - nobody can add
 884          * the feature since we have i_mutex. We must check it again
 885          * after acquiring ip_alloc_sem though, as paths like mmap
 886          * might have raced us to converting the inode to extents.
 887          */
 888         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 889             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 890                 goto out_update_size;
 891
 892         /*
 893          * The alloc sem blocks people in read/write from reading our
 894          * allocation until we're done changing it. We depend on
 895          * i_mutex to block other extend/truncate calls while we're
 896          * here.
 897          */
 898         down_write(&oi->ip_alloc_sem);
 899
 900         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 901                 /*
 902                  * We can optimize small extends by keeping the inodes
 903                  * inline data.
 904                  */
 905                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 906                         up_write(&oi->ip_alloc_sem);
 907                         goto out_update_size;
 908                 }
 909
 910                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 911                 if (ret) {
 912                         up_write(&oi->ip_alloc_sem);
 913
 914                         mlog_errno(ret);
 915                         goto out;
 916                 }
 917         }
 918
 919         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 920                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 921
 922         up_write(&oi->ip_alloc_sem);
 923
 924         if (ret < 0) {
 925                 mlog_errno(ret);
 926                 goto out;
 927         }
 928
 929 out_update_size:
 930         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 931         if (ret < 0)
 932                 mlog_errno(ret);
 933
 934 out:
 935         return ret;
 936 }
 937
 938 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 939 {
 940         int status = 0, size_change;
 941         struct inode *inode = dentry->d_inode;
 942         struct super_block *sb = inode->i_sb;
 943         struct ocfs2_super *osb = OCFS2_SB(sb);
 944         struct buffer_head *bh = NULL;
 945         handle_t *handle = NULL;
 946         int qtype;
 947         struct dquot *transfer_from[MAXQUOTAS] = { };
 948         struct dquot *transfer_to[MAXQUOTAS] = { };
 949
 950         mlog_entry("(0x%p, '%.*s')\n", dentry,
 951                    dentry->d_name.len, dentry->d_name.name);
 952
 953         /* ensuring we don't even attempt to truncate a symlink */
 954         if (S_ISLNK(inode->i_mode))
 955                 attr->ia_valid &= ~ATTR_SIZE;
 956
 957         if (attr->ia_valid & ATTR_MODE)
 958                 mlog(0, "mode change: %d\n", attr->ia_mode);
 959         if (attr->ia_valid & ATTR_UID)
 960                 mlog(0, "uid change: %d\n", attr->ia_uid);
 961         if (attr->ia_valid & ATTR_GID)
 962                 mlog(0, "gid change: %d\n", attr->ia_gid);
 963         if (attr->ia_valid & ATTR_SIZE)
 964                 mlog(0, "size change...\n");
 965         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 966                 mlog(0, "time change...\n");
 967
 968 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 969                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 970         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 971                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 972                 return 0;
 973         }
 974
 975         status = inode_change_ok(inode, attr);
 976         if (status)
 977                 return status;
 978
 979         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 980         if (size_change) {
 981                 status = ocfs2_rw_lock(inode, 1);
 982                 if (status < 0) {
 983                         mlog_errno(status);
 984                         goto bail;
 985                 }
 986         }
 987
 988         status = ocfs2_inode_lock(inode, &bh, 1);
 989         if (status < 0) {
 990                 if (status != -ENOENT)
 991                         mlog_errno(status);
 992                 goto bail_unlock_rw;
 993         }
 994
 995         if (size_change && attr->ia_size != i_size_read(inode)) {
 996                 if (attr->ia_size > sb->s_maxbytes) {
 997                         status = -EFBIG;
 998                         goto bail_unlock;
 999                 }
1000
1001                 if (i_size_read(inode) > attr->ia_size) {
1002                         if (ocfs2_should_order_data(inode)) {
1003                                 status = ocfs2_begin_ordered_truncate(inode,
1004                                                                       attr->ia_size);
1005                                 if (status)
1006                                         goto bail_unlock;
1007                         }
1008                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1009                 } else
1010                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1011                 if (status < 0) {
1012                         if (status != -ENOSPC)
1013                                 mlog_errno(status);
1014                         status = -ENOSPC;
1015                         goto bail_unlock;
1016                 }
1017         }
1018
1019         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1020             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1021                 /*
1022                  * Gather pointers to quota structures so that allocation /
1023                  * freeing of quota structures happens here and not inside
1024                  * vfs_dq_transfer() where we have problems with lock ordering
1025                  */
1026                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1027                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1028                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1029                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1030                                                       USRQUOTA);
1031                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1032                                                         USRQUOTA);
1033                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1034                                 status = -ESRCH;
1035                                 goto bail_unlock;
1036                         }
1037                 }
1038                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1039                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1040                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1041                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1042                                                       GRPQUOTA);
1043                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1044                                                         GRPQUOTA);
1045                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1046                                 status = -ESRCH;
1047                                 goto bail_unlock;
1048                         }
1049                 }
1050                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1051                                            2 * ocfs2_quota_trans_credits(sb));
1052                 if (IS_ERR(handle)) {
1053                         status = PTR_ERR(handle);
1054                         mlog_errno(status);
1055                         goto bail_unlock;
1056                 }
1057                 status = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
1058                 if (status < 0)
1059                         goto bail_commit;
1060         } else {
1061                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1062                 if (IS_ERR(handle)) {
1063                         status = PTR_ERR(handle);
1064                         mlog_errno(status);
1065                         goto bail_unlock;
1066                 }
1067         }
1068
1069         /*
1070          * This will intentionally not wind up calling vmtruncate(),
1071          * since all the work for a size change has been done above.
1072          * Otherwise, we could get into problems with truncate as
1073          * ip_alloc_sem is used there to protect against i_size
1074          * changes.
1075          */
1076         status = inode_setattr(inode, attr);
1077         if (status < 0) {
1078                 mlog_errno(status);
1079                 goto bail_commit;
1080         }
1081
1082         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1083         if (status < 0)
1084                 mlog_errno(status);
1085
1086 bail_commit:
1087         ocfs2_commit_trans(osb, handle);
1088 bail_unlock:
1089         ocfs2_inode_unlock(inode, 1);
1090 bail_unlock_rw:
1091         if (size_change)
1092                 ocfs2_rw_unlock(inode, 1);
1093 bail:
1094         brelse(bh);
1095
1096         /* Release quota pointers in case we acquired them */
1097         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1098                 dqput(transfer_to[qtype]);
1099                 dqput(transfer_from[qtype]);
1100         }
1101
1102         if (!status && attr->ia_valid & ATTR_MODE) {
1103                 status = ocfs2_acl_chmod(inode);
1104                 if (status < 0)
1105                         mlog_errno(status);
1106         }
1107
1108         mlog_exit(status);
1109         return status;
1110 }
1111
1112 int ocfs2_getattr(struct vfsmount *mnt,
1113                   struct dentry *dentry,
1114                   struct kstat *stat)
1115 {
1116         struct inode *inode = dentry->d_inode;
1117         struct super_block *sb = dentry->d_inode->i_sb;
1118         struct ocfs2_super *osb = sb->s_fs_info;
1119         int err;
1120
1121         mlog_entry_void();
1122
1123         err = ocfs2_inode_revalidate(dentry);
1124         if (err) {
1125                 if (err != -ENOENT)
1126                         mlog_errno(err);
1127                 goto bail;
1128         }
1129
1130         generic_fillattr(inode, stat);
1131
1132         /* We set the blksize from the cluster size for performance */
1133         stat->blksize = osb->s_clustersize;
1134
1135 bail:
1136         mlog_exit(err);
1137
1138         return err;
1139 }
1140
1141 int ocfs2_permission(struct inode *inode, int mask)
1142 {
1143         int ret;
1144
1145         mlog_entry_void();
1146
1147         ret = ocfs2_inode_lock(inode, NULL, 0);
1148         if (ret) {
1149                 if (ret != -ENOENT)
1150                         mlog_errno(ret);
1151                 goto out;
1152         }
1153
1154         ret = generic_permission(inode, mask, ocfs2_check_acl);
1155
1156         ocfs2_inode_unlock(inode, 0);
1157 out:
1158         mlog_exit(ret);
1159         return ret;
1160 }
1161
1162 static int __ocfs2_write_remove_suid(struct inode *inode,
1163                                      struct buffer_head *bh)
1164 {
1165         int ret;
1166         handle_t *handle;
1167         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1168         struct ocfs2_dinode *di;
1169
1170         mlog_entry("(Inode %llu, mode 0%o)\n",
1171                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1172
1173         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1174         if (IS_ERR(handle)) {
1175                 ret = PTR_ERR(handle);
1176                 mlog_errno(ret);
1177                 goto out;
1178         }
1179
1180         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1181                                       OCFS2_JOURNAL_ACCESS_WRITE);
1182         if (ret < 0) {
1183                 mlog_errno(ret);
1184                 goto out_trans;
1185         }
1186
1187         inode->i_mode &= ~S_ISUID;
1188         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1189                 inode->i_mode &= ~S_ISGID;
1190
1191         di = (struct ocfs2_dinode *) bh->b_data;
1192         di->i_mode = cpu_to_le16(inode->i_mode);
1193
1194         ret = ocfs2_journal_dirty(handle, bh);
1195         if (ret < 0)
1196                 mlog_errno(ret);
1197
1198 out_trans:
1199         ocfs2_commit_trans(osb, handle);
1200 out:
1201         mlog_exit(ret);
1202         return ret;
1203 }
1204
1205 /*
1206  * Will look for holes and unwritten extents in the range starting at
1207  * pos for count bytes (inclusive).
1208  */
1209 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1210                                        size_t count)
1211 {
1212         int ret = 0;
1213         unsigned int extent_flags;
1214         u32 cpos, clusters, extent_len, phys_cpos;
1215         struct super_block *sb = inode->i_sb;
1216
1217         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1218         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1219
1220         while (clusters) {
1221                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1222                                          &extent_flags);
1223                 if (ret < 0) {
1224                         mlog_errno(ret);
1225                         goto out;
1226                 }
1227
1228                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1229                         ret = 1;
1230                         break;
1231                 }
1232
1233                 if (extent_len > clusters)
1234                         extent_len = clusters;
1235
1236                 clusters -= extent_len;
1237                 cpos += extent_len;
1238         }
1239 out:
1240         return ret;
1241 }
1242
1243 static int ocfs2_write_remove_suid(struct inode *inode)
1244 {
1245         int ret;
1246         struct buffer_head *bh = NULL;
1247
1248         ret = ocfs2_read_inode_block(inode, &bh);
1249         if (ret < 0) {
1250                 mlog_errno(ret);
1251                 goto out;
1252         }
1253
1254         ret =  __ocfs2_write_remove_suid(inode, bh);
1255 out:
1256         brelse(bh);
1257         return ret;
1258 }
1259
1260 /*
1261  * Allocate enough extents to cover the region starting at byte offset
1262  * start for len bytes. Existing extents are skipped, any extents
1263  * added are marked as "unwritten".
1264  */
1265 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1266                                             u64 start, u64 len)
1267 {
1268         int ret;
1269         u32 cpos, phys_cpos, clusters, alloc_size;
1270         u64 end = start + len;
1271         struct buffer_head *di_bh = NULL;
1272
1273         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1274                 ret = ocfs2_read_inode_block(inode, &di_bh);
1275                 if (ret) {
1276                         mlog_errno(ret);
1277                         goto out;
1278                 }
1279
1280                 /*
1281                  * Nothing to do if the requested reservation range
1282                  * fits within the inode.
1283                  */
1284                 if (ocfs2_size_fits_inline_data(di_bh, end))
1285                         goto out;
1286
1287                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1288                 if (ret) {
1289                         mlog_errno(ret);
1290                         goto out;
1291                 }
1292         }
1293
1294         /*
1295          * We consider both start and len to be inclusive.
1296          */
1297         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1298         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1299         clusters -= cpos;
1300
1301         while (clusters) {
1302                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1303                                          &alloc_size, NULL);
1304                 if (ret) {
1305                         mlog_errno(ret);
1306                         goto out;
1307                 }
1308
1309                 /*
1310                  * Hole or existing extent len can be arbitrary, so
1311                  * cap it to our own allocation request.
1312                  */
1313                 if (alloc_size > clusters)
1314                         alloc_size = clusters;
1315
1316                 if (phys_cpos) {
1317                         /*
1318                          * We already have an allocation at this
1319                          * region so we can safely skip it.
1320                          */
1321                         goto next;
1322                 }
1323
1324                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1325                 if (ret) {
1326                         if (ret != -ENOSPC)
1327                                 mlog_errno(ret);
1328                         goto out;
1329                 }
1330
1331 next:
1332                 cpos += alloc_size;
1333                 clusters -= alloc_size;
1334         }
1335
1336         ret = 0;
1337 out:
1338
1339         brelse(di_bh);
1340         return ret;
1341 }
1342
1343 /*
1344  * Truncate a byte range, avoiding pages within partial clusters. This
1345  * preserves those pages for the zeroing code to write to.
1346  */
1347 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1348                                          u64 byte_len)
1349 {
1350         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1351         loff_t start, end;
1352         struct address_space *mapping = inode->i_mapping;
1353
1354         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1355         end = byte_start + byte_len;
1356         end = end & ~(osb->s_clustersize - 1);
1357
1358         if (start < end) {
1359                 unmap_mapping_range(mapping, start, end - start, 0);
1360                 truncate_inode_pages_range(mapping, start, end - 1);
1361         }
1362 }
1363
1364 static int ocfs2_zero_partial_clusters(struct inode *inode,
1365                                        u64 start, u64 len)
1366 {
1367         int ret = 0;
1368         u64 tmpend, end = start + len;
1369         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1370         unsigned int csize = osb->s_clustersize;
1371         handle_t *handle;
1372
1373         /*
1374          * The "start" and "end" values are NOT necessarily part of
1375          * the range whose allocation is being deleted. Rather, this
1376          * is what the user passed in with the request. We must zero
1377          * partial clusters here. There's no need to worry about
1378          * physical allocation - the zeroing code knows to skip holes.
1379          */
1380         mlog(0, "byte start: %llu, end: %llu\n",
1381              (unsigned long long)start, (unsigned long long)end);
1382
1383         /*
1384          * If both edges are on a cluster boundary then there's no
1385          * zeroing required as the region is part of the allocation to
1386          * be truncated.
1387          */
1388         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1389                 goto out;
1390
1391         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1392         if (IS_ERR(handle)) {
1393                 ret = PTR_ERR(handle);
1394                 mlog_errno(ret);
1395                 goto out;
1396         }
1397
1398         /*
1399          * We want to get the byte offset of the end of the 1st cluster.
1400          */
1401         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1402         if (tmpend > end)
1403                 tmpend = end;
1404
1405         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1406              (unsigned long long)start, (unsigned long long)tmpend);
1407
1408         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1409         if (ret)
1410                 mlog_errno(ret);
1411
1412         if (tmpend < end) {
1413                 /*
1414                  * This may make start and end equal, but the zeroing
1415                  * code will skip any work in that case so there's no
1416                  * need to catch it up here.
1417                  */
1418                 start = end & ~(osb->s_clustersize - 1);
1419
1420                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1421                      (unsigned long long)start, (unsigned long long)end);
1422
1423                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1424                 if (ret)
1425                         mlog_errno(ret);
1426         }
1427
1428         ocfs2_commit_trans(osb, handle);
1429 out:
1430         return ret;
1431 }
1432
1433 static int ocfs2_remove_inode_range(struct inode *inode,
1434                                     struct buffer_head *di_bh, u64 byte_start,
1435                                     u64 byte_len)
1436 {
1437         int ret = 0;
1438         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1439         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1440         struct ocfs2_cached_dealloc_ctxt dealloc;
1441         struct address_space *mapping = inode->i_mapping;
1442         struct ocfs2_extent_tree et;
1443
1444         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1445         ocfs2_init_dealloc_ctxt(&dealloc);
1446
1447         if (byte_len == 0)
1448                 return 0;
1449
1450         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1451                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1452                                             byte_start + byte_len, 0);
1453                 if (ret) {
1454                         mlog_errno(ret);
1455                         goto out;
1456                 }
1457                 /*
1458                  * There's no need to get fancy with the page cache
1459                  * truncate of an inline-data inode. We're talking
1460                  * about less than a page here, which will be cached
1461                  * in the dinode buffer anyway.
1462                  */
1463                 unmap_mapping_range(mapping, 0, 0, 0);
1464                 truncate_inode_pages(mapping, 0);
1465                 goto out;
1466         }
1467
1468         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1469         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1470         if (trunc_len >= trunc_start)
1471                 trunc_len -= trunc_start;
1472         else
1473                 trunc_len = 0;
1474
1475         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1476              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1477              (unsigned long long)byte_start,
1478              (unsigned long long)byte_len, trunc_start, trunc_len);
1479
1480         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1481         if (ret) {
1482                 mlog_errno(ret);
1483                 goto out;
1484         }
1485
1486         cpos = trunc_start;
1487         while (trunc_len) {
1488                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1489                                          &alloc_size, NULL);
1490                 if (ret) {
1491                         mlog_errno(ret);
1492                         goto out;
1493                 }
1494
1495                 if (alloc_size > trunc_len)
1496                         alloc_size = trunc_len;
1497
1498                 /* Only do work for non-holes */
1499                 if (phys_cpos != 0) {
1500                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1501                                                        phys_cpos, alloc_size,
1502                                                        &dealloc);
1503                         if (ret) {
1504                                 mlog_errno(ret);
1505                                 goto out;
1506                         }
1507                 }
1508
1509                 cpos += alloc_size;
1510                 trunc_len -= alloc_size;
1511         }
1512
1513         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1514
1515 out:
1516         ocfs2_schedule_truncate_log_flush(osb, 1);
1517         ocfs2_run_deallocs(osb, &dealloc);
1518
1519         return ret;
1520 }
1521
1522 /*
1523  * Parts of this function taken from xfs_change_file_space()
1524  */
1525 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1526                                      loff_t f_pos, unsigned int cmd,
1527                                      struct ocfs2_space_resv *sr,
1528                                      int change_size)
1529 {
1530         int ret;
1531         s64 llen;
1532         loff_t size;
1533         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1534         struct buffer_head *di_bh = NULL;
1535         handle_t *handle;
1536         unsigned long long max_off = inode->i_sb->s_maxbytes;
1537
1538         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1539                 return -EROFS;
1540
1541         mutex_lock(&inode->i_mutex);
1542
1543         /*
1544          * This prevents concurrent writes on other nodes
1545          */
1546         ret = ocfs2_rw_lock(inode, 1);
1547         if (ret) {
1548                 mlog_errno(ret);
1549                 goto out;
1550         }
1551
1552         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1553         if (ret) {
1554                 mlog_errno(ret);
1555                 goto out_rw_unlock;
1556         }
1557
1558         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1559                 ret = -EPERM;
1560                 goto out_inode_unlock;
1561         }
1562
1563         switch (sr->l_whence) {
1564         case 0: /*SEEK_SET*/
1565                 break;
1566         case 1: /*SEEK_CUR*/
1567                 sr->l_start += f_pos;
1568                 break;
1569         case 2: /*SEEK_END*/
1570                 sr->l_start += i_size_read(inode);
1571                 break;
1572         default:
1573                 ret = -EINVAL;
1574                 goto out_inode_unlock;
1575         }
1576         sr->l_whence = 0;
1577
1578         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1579
1580         if (sr->l_start < 0
1581             || sr->l_start > max_off
1582             || (sr->l_start + llen) < 0
1583             || (sr->l_start + llen) > max_off) {
1584                 ret = -EINVAL;
1585                 goto out_inode_unlock;
1586         }
1587         size = sr->l_start + sr->l_len;
1588
1589         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1590                 if (sr->l_len <= 0) {
1591                         ret = -EINVAL;
1592                         goto out_inode_unlock;
1593                 }
1594         }
1595
1596         if (file && should_remove_suid(file->f_path.dentry)) {
1597                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1598                 if (ret) {
1599                         mlog_errno(ret);
1600                         goto out_inode_unlock;
1601                 }
1602         }
1603
1604         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1605         switch (cmd) {
1606         case OCFS2_IOC_RESVSP:
1607         case OCFS2_IOC_RESVSP64:
1608                 /*
1609                  * This takes unsigned offsets, but the signed ones we
1610                  * pass have been checked against overflow above.
1611                  */
1612                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1613                                                        sr->l_len);
1614                 break;
1615         case OCFS2_IOC_UNRESVSP:
1616         case OCFS2_IOC_UNRESVSP64:
1617                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1618                                                sr->l_len);
1619                 break;
1620         default:
1621                 ret = -EINVAL;
1622         }
1623         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1624         if (ret) {
1625                 mlog_errno(ret);
1626                 goto out_inode_unlock;
1627         }
1628
1629         /*
1630          * We update c/mtime for these changes
1631          */
1632         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1633         if (IS_ERR(handle)) {
1634                 ret = PTR_ERR(handle);
1635                 mlog_errno(ret);
1636                 goto out_inode_unlock;
1637         }
1638
1639         if (change_size && i_size_read(inode) < size)
1640                 i_size_write(inode, size);
1641
1642         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1643         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1644         if (ret < 0)
1645                 mlog_errno(ret);
1646
1647         ocfs2_commit_trans(osb, handle);
1648
1649 out_inode_unlock:
1650         brelse(di_bh);
1651         ocfs2_inode_unlock(inode, 1);
1652 out_rw_unlock:
1653         ocfs2_rw_unlock(inode, 1);
1654
1655 out:
1656         mutex_unlock(&inode->i_mutex);
1657         return ret;
1658 }
1659
1660 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1661                             struct ocfs2_space_resv *sr)
1662 {
1663         struct inode *inode = file->f_path.dentry->d_inode;
1664         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1665
1666         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1667             !ocfs2_writes_unwritten_extents(osb))
1668                 return -ENOTTY;
1669         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1670                  !ocfs2_sparse_alloc(osb))
1671                 return -ENOTTY;
1672
1673         if (!S_ISREG(inode->i_mode))
1674                 return -EINVAL;
1675
1676         if (!(file->f_mode & FMODE_WRITE))
1677                 return -EBADF;
1678
1679         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1680 }
1681
1682 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1683                             loff_t len)
1684 {
1685         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1686         struct ocfs2_space_resv sr;
1687         int change_size = 1;
1688
1689         if (!ocfs2_writes_unwritten_extents(osb))
1690                 return -EOPNOTSUPP;
1691
1692         if (S_ISDIR(inode->i_mode))
1693                 return -ENODEV;
1694
1695         if (mode & FALLOC_FL_KEEP_SIZE)
1696                 change_size = 0;
1697
1698         sr.l_whence = 0;
1699         sr.l_start = (s64)offset;
1700         sr.l_len = (s64)len;
1701
1702         return __ocfs2_change_file_space(NULL, inode, offset,
1703                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1704 }
1705
1706 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1707                                    size_t count)
1708 {
1709         int ret = 0;
1710         unsigned int extent_flags;
1711         u32 cpos, clusters, extent_len, phys_cpos;
1712         struct super_block *sb = inode->i_sb;
1713
1714         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1715             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1716             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1717                 return 0;
1718
1719         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1720         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1721
1722         while (clusters) {
1723                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1724                                          &extent_flags);
1725                 if (ret < 0) {
1726                         mlog_errno(ret);
1727                         goto out;
1728                 }
1729
1730                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1731                         ret = 1;
1732                         break;
1733                 }
1734
1735                 if (extent_len > clusters)
1736                         extent_len = clusters;
1737
1738                 clusters -= extent_len;
1739                 cpos += extent_len;
1740         }
1741 out:
1742         return ret;
1743 }
1744
1745 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1746                                             loff_t pos, size_t count,
1747                                             int *meta_level)
1748 {
1749         int ret;
1750         struct buffer_head *di_bh = NULL;
1751         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1752         u32 clusters =
1753                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1754
1755         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1756         if (ret) {
1757                 mlog_errno(ret);
1758                 goto out;
1759         }
1760
1761         *meta_level = 1;
1762
1763         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1764         if (ret)
1765                 mlog_errno(ret);
1766 out:
1767         brelse(di_bh);
1768         return ret;
1769 }
1770
1771 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1772                                          loff_t *ppos,
1773                                          size_t count,
1774                                          int appending,
1775                                          int *direct_io)
1776 {
1777         int ret = 0, meta_level = 0;
1778         struct inode *inode = dentry->d_inode;
1779         loff_t saved_pos, end;
1780
1781         /*
1782          * We start with a read level meta lock and only jump to an ex
1783          * if we need to make modifications here.
1784          */
1785         for(;;) {
1786                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1787                 if (ret < 0) {
1788                         meta_level = -1;
1789                         mlog_errno(ret);
1790                         goto out;
1791                 }
1792
1793                 /* Clear suid / sgid if necessary. We do this here
1794                  * instead of later in the write path because
1795                  * remove_suid() calls ->setattr without any hint that
1796                  * we may have already done our cluster locking. Since
1797                  * ocfs2_setattr() *must* take cluster locks to
1798                  * proceeed, this will lead us to recursively lock the
1799                  * inode. There's also the dinode i_size state which
1800                  * can be lost via setattr during extending writes (we
1801                  * set inode->i_size at the end of a write. */
1802                 if (should_remove_suid(dentry)) {
1803                         if (meta_level == 0) {
1804                                 ocfs2_inode_unlock(inode, meta_level);
1805                                 meta_level = 1;
1806                                 continue;
1807                         }
1808
1809                         ret = ocfs2_write_remove_suid(inode);
1810                         if (ret < 0) {
1811                                 mlog_errno(ret);
1812                                 goto out_unlock;
1813                         }
1814                 }
1815
1816                 /* work on a copy of ppos until we're sure that we won't have
1817                  * to recalculate it due to relocking. */
1818                 if (appending) {
1819                         saved_pos = i_size_read(inode);
1820                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1821                 } else {
1822                         saved_pos = *ppos;
1823                 }
1824
1825                 end = saved_pos + count;
1826
1827                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1828                 if (ret == 1) {
1829                         ocfs2_inode_unlock(inode, meta_level);
1830                         meta_level = -1;
1831
1832                         ret = ocfs2_prepare_inode_for_refcount(inode,
1833                                                                saved_pos,
1834                                                                count,
1835                                                                &meta_level);
1836                 }
1837
1838                 if (ret < 0) {
1839                         mlog_errno(ret);
1840                         goto out_unlock;
1841                 }
1842
1843                 /*
1844                  * Skip the O_DIRECT checks if we don't need
1845                  * them.
1846                  */
1847                 if (!direct_io || !(*direct_io))
1848                         break;
1849
1850                 /*
1851                  * There's no sane way to do direct writes to an inode
1852                  * with inline data.
1853                  */
1854                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1855                         *direct_io = 0;
1856                         break;
1857                 }
1858
1859                 /*
1860                  * Allowing concurrent direct writes means
1861                  * i_size changes wouldn't be synchronized, so
1862                  * one node could wind up truncating another
1863                  * nodes writes.
1864                  */
1865                 if (end > i_size_read(inode)) {
1866                         *direct_io = 0;
1867                         break;
1868                 }
1869
1870                 /*
1871                  * We don't fill holes during direct io, so
1872                  * check for them here. If any are found, the
1873                  * caller will have to retake some cluster
1874                  * locks and initiate the io as buffered.
1875                  */
1876                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1877                 if (ret == 1) {
1878                         *direct_io = 0;
1879                         ret = 0;
1880                 } else if (ret < 0)
1881                         mlog_errno(ret);
1882                 break;
1883         }
1884
1885         if (appending)
1886                 *ppos = saved_pos;
1887
1888 out_unlock:
1889         if (meta_level >= 0)
1890                 ocfs2_inode_unlock(inode, meta_level);
1891
1892 out:
1893         return ret;
1894 }
1895
1896 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1897                                     const struct iovec *iov,
1898                                     unsigned long nr_segs,
1899                                     loff_t pos)
1900 {
1901         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1902         int can_do_direct;
1903         ssize_t written = 0;
1904         size_t ocount;          /* original count */
1905         size_t count;           /* after file limit checks */
1906         loff_t old_size, *ppos = &iocb->ki_pos;
1907         u32 old_clusters;
1908         struct file *file = iocb->ki_filp;
1909         struct inode *inode = file->f_path.dentry->d_inode;
1910         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1911
1912         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1913                    (unsigned int)nr_segs,
1914                    file->f_path.dentry->d_name.len,
1915                    file->f_path.dentry->d_name.name);
1916
1917         if (iocb->ki_left == 0)
1918                 return 0;
1919
1920         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1921
1922         appending = file->f_flags & O_APPEND ? 1 : 0;
1923         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1924
1925         mutex_lock(&inode->i_mutex);
1926
1927 relock:
1928         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1929         if (direct_io) {
1930                 down_read(&inode->i_alloc_sem);
1931                 have_alloc_sem = 1;
1932         }
1933
1934         /* concurrent O_DIRECT writes are allowed */
1935         rw_level = !direct_io;
1936         ret = ocfs2_rw_lock(inode, rw_level);
1937         if (ret < 0) {
1938                 mlog_errno(ret);
1939                 goto out_sems;
1940         }
1941
1942         can_do_direct = direct_io;
1943         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1944                                             iocb->ki_left, appending,
1945                                             &can_do_direct);
1946         if (ret < 0) {
1947                 mlog_errno(ret);
1948                 goto out;
1949         }
1950
1951         /*
1952          * We can't complete the direct I/O as requested, fall back to
1953          * buffered I/O.
1954          */
1955         if (direct_io && !can_do_direct) {
1956                 ocfs2_rw_unlock(inode, rw_level);
1957                 up_read(&inode->i_alloc_sem);
1958
1959                 have_alloc_sem = 0;
1960                 rw_level = -1;
1961
1962                 direct_io = 0;
1963                 goto relock;
1964         }
1965
1966         /*
1967          * To later detect whether a journal commit for sync writes is
1968          * necessary, we sample i_size, and cluster count here.
1969          */
1970         old_size = i_size_read(inode);
1971         old_clusters = OCFS2_I(inode)->ip_clusters;
1972
1973         /* communicate with ocfs2_dio_end_io */
1974         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1975
1976         if (direct_io) {
1977                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1978                                              VERIFY_READ);
1979                 if (ret)
1980                         goto out_dio;
1981
1982                 count = ocount;
1983                 ret = generic_write_checks(file, ppos, &count,
1984                                            S_ISBLK(inode->i_mode));
1985                 if (ret)
1986                         goto out_dio;
1987
1988                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1989                                                     ppos, count, ocount);
1990                 if (written < 0) {
1991                         /*
1992                          * direct write may have instantiated a few
1993                          * blocks outside i_size. Trim these off again.
1994                          * Don't need i_size_read because we hold i_mutex.
1995                          */
1996                         if (*ppos + count > inode->i_size)
1997                                 vmtruncate(inode, inode->i_size);
1998                         ret = written;
1999                         goto out_dio;
2000                 }
2001         } else {
2002                 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2003         }
2004
2005 out_dio:
2006         /* buffered aio wouldn't have proper lock coverage today */
2007         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2008
2009         if ((file->f_flags & O_DSYNC && !direct_io) || IS_SYNC(inode)) {
2010                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2011                                                pos + count - 1);
2012                 if (ret < 0)
2013                         written = ret;
2014
2015                 if (!ret && (old_size != i_size_read(inode) ||
2016                     old_clusters != OCFS2_I(inode)->ip_clusters)) {
2017                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2018                         if (ret < 0)
2019                                 written = ret;
2020                 }
2021
2022                 if (!ret)
2023                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2024                                                       pos + count - 1);
2025         }
2026
2027         /*
2028          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2029          * function pointer which is called when o_direct io completes so that
2030          * it can unlock our rw lock.  (it's the clustered equivalent of
2031          * i_alloc_sem; protects truncate from racing with pending ios).
2032          * Unfortunately there are error cases which call end_io and others
2033          * that don't.  so we don't have to unlock the rw_lock if either an
2034          * async dio is going to do it in the future or an end_io after an
2035          * error has already done it.
2036          */
2037         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2038                 rw_level = -1;
2039                 have_alloc_sem = 0;
2040         }
2041
2042 out:
2043         if (rw_level != -1)
2044                 ocfs2_rw_unlock(inode, rw_level);
2045
2046 out_sems:
2047         if (have_alloc_sem)
2048                 up_read(&inode->i_alloc_sem);
2049
2050         mutex_unlock(&inode->i_mutex);
2051
2052         if (written)
2053                 ret = written;
2054         mlog_exit(ret);
2055         return ret;
2056 }
2057
2058 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2059                                 struct file *out,
2060                                 struct splice_desc *sd)
2061 {
2062         int ret;
2063
2064         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2065                                             sd->total_len, 0, NULL);
2066         if (ret < 0) {
2067                 mlog_errno(ret);
2068                 return ret;
2069         }
2070
2071         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2072 }
2073
2074 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2075                                        struct file *out,
2076                                        loff_t *ppos,
2077                                        size_t len,
2078                                        unsigned int flags)
2079 {
2080         int ret;
2081         struct address_space *mapping = out->f_mapping;
2082         struct inode *inode = mapping->host;
2083         struct splice_desc sd = {
2084                 .total_len = len,
2085                 .flags = flags,
2086                 .pos = *ppos,
2087                 .u.file = out,
2088         };
2089
2090         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2091                    (unsigned int)len,
2092                    out->f_path.dentry->d_name.len,
2093                    out->f_path.dentry->d_name.name);
2094
2095         if (pipe->inode)
2096                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2097
2098         splice_from_pipe_begin(&sd);
2099         do {
2100                 ret = splice_from_pipe_next(pipe, &sd);
2101                 if (ret <= 0)
2102                         break;
2103
2104                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2105                 ret = ocfs2_rw_lock(inode, 1);
2106                 if (ret < 0)
2107                         mlog_errno(ret);
2108                 else {
2109                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2110                         ocfs2_rw_unlock(inode, 1);
2111                 }
2112                 mutex_unlock(&inode->i_mutex);
2113         } while (ret > 0);
2114         splice_from_pipe_end(pipe, &sd);
2115
2116         if (pipe->inode)
2117                 mutex_unlock(&pipe->inode->i_mutex);
2118
2119         if (sd.num_spliced)
2120                 ret = sd.num_spliced;
2121
2122         if (ret > 0) {
2123                 unsigned long nr_pages;
2124                 int err;
2125
2126                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2127
2128                 err = generic_write_sync(out, *ppos, ret);
2129                 if (err)
2130                         ret = err;
2131                 else
2132                         *ppos += ret;
2133
2134                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2135         }
2136
2137         mlog_exit(ret);
2138         return ret;
2139 }
2140
2141 static ssize_t ocfs2_file_splice_read(struct file *in,
2142                                       loff_t *ppos,
2143                                       struct pipe_inode_info *pipe,
2144                                       size_t len,
2145                                       unsigned int flags)
2146 {
2147         int ret = 0, lock_level = 0;
2148         struct inode *inode = in->f_path.dentry->d_inode;
2149
2150         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2151                    (unsigned int)len,
2152                    in->f_path.dentry->d_name.len,
2153                    in->f_path.dentry->d_name.name);
2154
2155         /*
2156          * See the comment in ocfs2_file_aio_read()
2157          */
2158         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2159         if (ret < 0) {
2160                 mlog_errno(ret);
2161                 goto bail;
2162         }
2163         ocfs2_inode_unlock(inode, lock_level);
2164
2165         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2166
2167 bail:
2168         mlog_exit(ret);
2169         return ret;
2170 }
2171
2172 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2173                                    const struct iovec *iov,
2174                                    unsigned long nr_segs,
2175                                    loff_t pos)
2176 {
2177         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2178         struct file *filp = iocb->ki_filp;
2179         struct inode *inode = filp->f_path.dentry->d_inode;
2180
2181         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2182                    (unsigned int)nr_segs,
2183                    filp->f_path.dentry->d_name.len,
2184                    filp->f_path.dentry->d_name.name);
2185
2186         if (!inode) {
2187                 ret = -EINVAL;
2188                 mlog_errno(ret);
2189                 goto bail;
2190         }
2191
2192         /*
2193          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2194          * need locks to protect pending reads from racing with truncate.
2195          */
2196         if (filp->f_flags & O_DIRECT) {
2197                 down_read(&inode->i_alloc_sem);
2198                 have_alloc_sem = 1;
2199
2200                 ret = ocfs2_rw_lock(inode, 0);
2201                 if (ret < 0) {
2202                         mlog_errno(ret);
2203                         goto bail;
2204                 }
2205                 rw_level = 0;
2206                 /* communicate with ocfs2_dio_end_io */
2207                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2208         }
2209
2210         /*
2211          * We're fine letting folks race truncates and extending
2212          * writes with read across the cluster, just like they can
2213          * locally. Hence no rw_lock during read.
2214          *
2215          * Take and drop the meta data lock to update inode fields
2216          * like i_size. This allows the checks down below
2217          * generic_file_aio_read() a chance of actually working.
2218          */
2219         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2220         if (ret < 0) {
2221                 mlog_errno(ret);
2222                 goto bail;
2223         }
2224         ocfs2_inode_unlock(inode, lock_level);
2225
2226         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2227         if (ret == -EINVAL)
2228                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2229
2230         /* buffered aio wouldn't have proper lock coverage today */
2231         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2232
2233         /* see ocfs2_file_aio_write */
2234         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2235                 rw_level = -1;
2236                 have_alloc_sem = 0;
2237         }
2238
2239 bail:
2240         if (have_alloc_sem)
2241                 up_read(&inode->i_alloc_sem);
2242         if (rw_level != -1)
2243                 ocfs2_rw_unlock(inode, rw_level);
2244         mlog_exit(ret);
2245
2246         return ret;
2247 }
2248
2249 const struct inode_operations ocfs2_file_iops = {
2250         .setattr        = ocfs2_setattr,
2251         .getattr        = ocfs2_getattr,
2252         .permission     = ocfs2_permission,
2253         .setxattr       = generic_setxattr,
2254         .getxattr       = generic_getxattr,
2255         .listxattr      = ocfs2_listxattr,
2256         .removexattr    = generic_removexattr,
2257         .fallocate      = ocfs2_fallocate,
2258         .fiemap         = ocfs2_fiemap,
2259 };
2260
2261 const struct inode_operations ocfs2_special_file_iops = {
2262         .setattr        = ocfs2_setattr,
2263         .getattr        = ocfs2_getattr,
2264         .permission     = ocfs2_permission,
2265 };
2266
2267 /*
2268  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2269  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2270  */
2271 const struct file_operations ocfs2_fops = {
2272         .llseek         = generic_file_llseek,
2273         .read           = do_sync_read,
2274         .write          = do_sync_write,
2275         .mmap           = ocfs2_mmap,
2276         .fsync          = ocfs2_sync_file,
2277         .release        = ocfs2_file_release,
2278         .open           = ocfs2_file_open,
2279         .aio_read       = ocfs2_file_aio_read,
2280         .aio_write      = ocfs2_file_aio_write,
2281         .unlocked_ioctl = ocfs2_ioctl,
2282 #ifdef CONFIG_COMPAT
2283         .compat_ioctl   = ocfs2_compat_ioctl,
2284 #endif
2285         .lock           = ocfs2_lock,
2286         .flock          = ocfs2_flock,
2287         .splice_read    = ocfs2_file_splice_read,
2288         .splice_write   = ocfs2_file_splice_write,
2289 };
2290
2291 const struct file_operations ocfs2_dops = {
2292         .llseek         = generic_file_llseek,
2293         .read           = generic_read_dir,
2294         .readdir        = ocfs2_readdir,
2295         .fsync          = ocfs2_sync_file,
2296         .release        = ocfs2_dir_release,
2297         .open           = ocfs2_dir_open,
2298         .unlocked_ioctl = ocfs2_ioctl,
2299 #ifdef CONFIG_COMPAT
2300         .compat_ioctl   = ocfs2_compat_ioctl,
2301 #endif
2302         .lock           = ocfs2_lock,
2303         .flock          = ocfs2_flock,
2304 };
2305
2306 /*
2307  * POSIX-lockless variants of our file_operations.
2308  *
2309  * These will be used if the underlying cluster stack does not support
2310  * posix file locking, if the user passes the "localflocks" mount
2311  * option, or if we have a local-only fs.
2312  *
2313  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2314  * so we still want it in the case of no stack support for
2315  * plocks. Internally, it will do the right thing when asked to ignore
2316  * the cluster.
2317  */
2318 const struct file_operations ocfs2_fops_no_plocks = {
2319         .llseek         = generic_file_llseek,
2320         .read           = do_sync_read,
2321         .write          = do_sync_write,
2322         .mmap           = ocfs2_mmap,
2323         .fsync          = ocfs2_sync_file,
2324         .release        = ocfs2_file_release,
2325         .open           = ocfs2_file_open,
2326         .aio_read       = ocfs2_file_aio_read,
2327         .aio_write      = ocfs2_file_aio_write,
2328         .unlocked_ioctl = ocfs2_ioctl,
2329 #ifdef CONFIG_COMPAT
2330         .compat_ioctl   = ocfs2_compat_ioctl,
2331 #endif
2332         .flock          = ocfs2_flock,
2333         .splice_read    = ocfs2_file_splice_read,
2334         .splice_write   = ocfs2_file_splice_write,
2335 };
2336
2337 const struct file_operations ocfs2_dops_no_plocks = {
2338         .llseek         = generic_file_llseek,
2339         .read           = generic_read_dir,
2340         .readdir        = ocfs2_readdir,
2341         .fsync          = ocfs2_sync_file,
2342         .release        = ocfs2_dir_release,
2343         .open           = ocfs2_dir_open,
2344         .unlocked_ioctl = ocfs2_ioctl,
2345 #ifdef CONFIG_COMPAT
2346         .compat_ioctl   = ocfs2_compat_ioctl,
2347 #endif
2348         .flock          = ocfs2_flock,
2349 };