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
  39 #define MLOG_MASK_PREFIX ML_INODE
  40 #include <cluster/masklog.h>
  41
  42 #include "ocfs2.h"
  43
  44 #include "alloc.h"
  45 #include "aops.h"
  46 #include "dir.h"
  47 #include "dlmglue.h"
  48 #include "extent_map.h"
  49 #include "file.h"
  50 #include "sysfile.h"
  51 #include "inode.h"
  52 #include "ioctl.h"
  53 #include "journal.h"
  54 #include "locks.h"
  55 #include "mmap.h"
  56 #include "suballoc.h"
  57 #include "super.h"
  58
  59 #include "buffer_head_io.h"
  60
  61 static int ocfs2_sync_inode(struct inode *inode)
  62 {
  63         filemap_fdatawrite(inode->i_mapping);
  64         return sync_mapping_buffers(inode->i_mapping);
  65 }
  66
  67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  68 {
  69         struct ocfs2_file_private *fp;
  70
  71         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  72         if (!fp)
  73                 return -ENOMEM;
  74
  75         fp->fp_file = file;
  76         mutex_init(&fp->fp_mutex);
  77         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  78         file->private_data = fp;
  79
  80         return 0;
  81 }
  82
  83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  84 {
  85         struct ocfs2_file_private *fp = file->private_data;
  86         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  87
  88         if (fp) {
  89                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  90                 ocfs2_lock_res_free(&fp->fp_flock);
  91                 kfree(fp);
  92                 file->private_data = NULL;
  93         }
  94 }
  95
  96 static int ocfs2_file_open(struct inode *inode, struct file *file)
  97 {
  98         int status;
  99         int mode = file->f_flags;
 100         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 101
 102         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 103                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 104
 105         spin_lock(&oi->ip_lock);
 106
 107         /* Check that the inode hasn't been wiped from disk by another
 108          * node. If it hasn't then we're safe as long as we hold the
 109          * spin lock until our increment of open count. */
 110         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 111                 spin_unlock(&oi->ip_lock);
 112
 113                 status = -ENOENT;
 114                 goto leave;
 115         }
 116
 117         if (mode & O_DIRECT)
 118                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 119
 120         oi->ip_open_count++;
 121         spin_unlock(&oi->ip_lock);
 122
 123         status = ocfs2_init_file_private(inode, file);
 124         if (status) {
 125                 /*
 126                  * We want to set open count back if we're failing the
 127                  * open.
 128                  */
 129                 spin_lock(&oi->ip_lock);
 130                 oi->ip_open_count--;
 131                 spin_unlock(&oi->ip_lock);
 132         }
 133
 134 leave:
 135         mlog_exit(status);
 136         return status;
 137 }
 138
 139 static int ocfs2_file_release(struct inode *inode, struct file *file)
 140 {
 141         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 142
 143         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 144                        file->f_path.dentry->d_name.len,
 145                        file->f_path.dentry->d_name.name);
 146
 147         spin_lock(&oi->ip_lock);
 148         if (!--oi->ip_open_count)
 149                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 150         spin_unlock(&oi->ip_lock);
 151
 152         ocfs2_free_file_private(inode, file);
 153
 154         mlog_exit(0);
 155
 156         return 0;
 157 }
 158
 159 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 160 {
 161         return ocfs2_init_file_private(inode, file);
 162 }
 163
 164 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 165 {
 166         ocfs2_free_file_private(inode, file);
 167         return 0;
 168 }
 169
 170 static int ocfs2_sync_file(struct file *file,
 171                            struct dentry *dentry,
 172                            int datasync)
 173 {
 174         int err = 0;
 175         journal_t *journal;
 176         struct inode *inode = dentry->d_inode;
 177         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 178
 179         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 180                    dentry->d_name.len, dentry->d_name.name);
 181
 182         err = ocfs2_sync_inode(dentry->d_inode);
 183         if (err)
 184                 goto bail;
 185
 186         journal = osb->journal->j_journal;
 187         err = journal_force_commit(journal);
 188
 189 bail:
 190         mlog_exit(err);
 191
 192         return (err < 0) ? -EIO : 0;
 193 }
 194
 195 int ocfs2_should_update_atime(struct inode *inode,
 196                               struct vfsmount *vfsmnt)
 197 {
 198         struct timespec now;
 199         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 200
 201         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 202                 return 0;
 203
 204         if ((inode->i_flags & S_NOATIME) ||
 205             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 206                 return 0;
 207
 208         /*
 209          * We can be called with no vfsmnt structure - NFSD will
 210          * sometimes do this.
 211          *
 212          * Note that our action here is different than touch_atime() -
 213          * if we can't tell whether this is a noatime mount, then we
 214          * don't know whether to trust the value of s_atime_quantum.
 215          */
 216         if (vfsmnt == NULL)
 217                 return 0;
 218
 219         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 220             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 221                 return 0;
 222
 223         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 224                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 225                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 226                         return 1;
 227
 228                 return 0;
 229         }
 230
 231         now = CURRENT_TIME;
 232         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 233                 return 0;
 234         else
 235                 return 1;
 236 }
 237
 238 int ocfs2_update_inode_atime(struct inode *inode,
 239                              struct buffer_head *bh)
 240 {
 241         int ret;
 242         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 243         handle_t *handle;
 244         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 245
 246         mlog_entry_void();
 247
 248         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 249         if (handle == NULL) {
 250                 ret = -ENOMEM;
 251                 mlog_errno(ret);
 252                 goto out;
 253         }
 254
 255         ret = ocfs2_journal_access(handle, inode, bh,
 256                                    OCFS2_JOURNAL_ACCESS_WRITE);
 257         if (ret) {
 258                 mlog_errno(ret);
 259                 goto out_commit;
 260         }
 261
 262         /*
 263          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 264          * have i_mutex to guard against concurrent changes to other
 265          * inode fields.
 266          */
 267         inode->i_atime = CURRENT_TIME;
 268         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 269         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 270
 271         ret = ocfs2_journal_dirty(handle, bh);
 272         if (ret < 0)
 273                 mlog_errno(ret);
 274
 275 out_commit:
 276         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 277 out:
 278         mlog_exit(ret);
 279         return ret;
 280 }
 281
 282 static int ocfs2_set_inode_size(handle_t *handle,
 283                                 struct inode *inode,
 284                                 struct buffer_head *fe_bh,
 285                                 u64 new_i_size)
 286 {
 287         int status;
 288
 289         mlog_entry_void();
 290         i_size_write(inode, new_i_size);
 291         inode->i_blocks = ocfs2_inode_sector_count(inode);
 292         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 293
 294         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 295         if (status < 0) {
 296                 mlog_errno(status);
 297                 goto bail;
 298         }
 299
 300 bail:
 301         mlog_exit(status);
 302         return status;
 303 }
 304
 305 static int ocfs2_simple_size_update(struct inode *inode,
 306                                     struct buffer_head *di_bh,
 307                                     u64 new_i_size)
 308 {
 309         int ret;
 310         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 311         handle_t *handle = NULL;
 312
 313         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 314         if (handle == NULL) {
 315                 ret = -ENOMEM;
 316                 mlog_errno(ret);
 317                 goto out;
 318         }
 319
 320         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 321                                    new_i_size);
 322         if (ret < 0)
 323                 mlog_errno(ret);
 324
 325         ocfs2_commit_trans(osb, handle);
 326 out:
 327         return ret;
 328 }
 329
 330 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 331                                      struct inode *inode,
 332                                      struct buffer_head *fe_bh,
 333                                      u64 new_i_size)
 334 {
 335         int status;
 336         handle_t *handle;
 337         struct ocfs2_dinode *di;
 338         u64 cluster_bytes;
 339
 340         mlog_entry_void();
 341
 342         /* TODO: This needs to actually orphan the inode in this
 343          * transaction. */
 344
 345         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 346         if (IS_ERR(handle)) {
 347                 status = PTR_ERR(handle);
 348                 mlog_errno(status);
 349                 goto out;
 350         }
 351
 352         status = ocfs2_journal_access(handle, inode, fe_bh,
 353                                       OCFS2_JOURNAL_ACCESS_WRITE);
 354         if (status < 0) {
 355                 mlog_errno(status);
 356                 goto out_commit;
 357         }
 358
 359         /*
 360          * Do this before setting i_size.
 361          */
 362         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 363         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 364                                                cluster_bytes);
 365         if (status) {
 366                 mlog_errno(status);
 367                 goto out_commit;
 368         }
 369
 370         i_size_write(inode, new_i_size);
 371         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 372
 373         di = (struct ocfs2_dinode *) fe_bh->b_data;
 374         di->i_size = cpu_to_le64(new_i_size);
 375         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 376         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 377
 378         status = ocfs2_journal_dirty(handle, fe_bh);
 379         if (status < 0)
 380                 mlog_errno(status);
 381
 382 out_commit:
 383         ocfs2_commit_trans(osb, handle);
 384 out:
 385
 386         mlog_exit(status);
 387         return status;
 388 }
 389
 390 static int ocfs2_truncate_file(struct inode *inode,
 391                                struct buffer_head *di_bh,
 392                                u64 new_i_size)
 393 {
 394         int status = 0;
 395         struct ocfs2_dinode *fe = NULL;
 396         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 397         struct ocfs2_truncate_context *tc = NULL;
 398
 399         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 400                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 401                    (unsigned long long)new_i_size);
 402
 403         fe = (struct ocfs2_dinode *) di_bh->b_data;
 404         if (!OCFS2_IS_VALID_DINODE(fe)) {
 405                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 406                 status = -EIO;
 407                 goto bail;
 408         }
 409
 410         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 411                         "Inode %llu, inode i_size = %lld != di "
 412                         "i_size = %llu, i_flags = 0x%x\n",
 413                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 414                         i_size_read(inode),
 415                         (unsigned long long)le64_to_cpu(fe->i_size),
 416                         le32_to_cpu(fe->i_flags));
 417
 418         if (new_i_size > le64_to_cpu(fe->i_size)) {
 419                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 420                      (unsigned long long)le64_to_cpu(fe->i_size),
 421                      (unsigned long long)new_i_size);
 422                 status = -EINVAL;
 423                 mlog_errno(status);
 424                 goto bail;
 425         }
 426
 427         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 428              (unsigned long long)le64_to_cpu(fe->i_blkno),
 429              (unsigned long long)le64_to_cpu(fe->i_size),
 430              (unsigned long long)new_i_size);
 431
 432         /* lets handle the simple truncate cases before doing any more
 433          * cluster locking. */
 434         if (new_i_size == le64_to_cpu(fe->i_size))
 435                 goto bail;
 436
 437         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 438
 439         /*
 440          * The inode lock forced other nodes to sync and drop their
 441          * pages, which (correctly) happens even if we have a truncate
 442          * without allocation change - ocfs2 cluster sizes can be much
 443          * greater than page size, so we have to truncate them
 444          * anyway.
 445          */
 446         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 447         truncate_inode_pages(inode->i_mapping, new_i_size);
 448
 449         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 450                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 451                                                i_size_read(inode), 1);
 452                 if (status)
 453                         mlog_errno(status);
 454
 455                 goto bail_unlock_sem;
 456         }
 457
 458         /* alright, we're going to need to do a full blown alloc size
 459          * change. Orphan the inode so that recovery can complete the
 460          * truncate if necessary. This does the task of marking
 461          * i_size. */
 462         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 463         if (status < 0) {
 464                 mlog_errno(status);
 465                 goto bail_unlock_sem;
 466         }
 467
 468         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 469         if (status < 0) {
 470                 mlog_errno(status);
 471                 goto bail_unlock_sem;
 472         }
 473
 474         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 475         if (status < 0) {
 476                 mlog_errno(status);
 477                 goto bail_unlock_sem;
 478         }
 479
 480         /* TODO: orphan dir cleanup here. */
 481 bail_unlock_sem:
 482         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 483
 484 bail:
 485
 486         mlog_exit(status);
 487         return status;
 488 }
 489
 490 /*
 491  * extend allocation only here.
 492  * we'll update all the disk stuff, and oip->alloc_size
 493  *
 494  * expect stuff to be locked, a transaction started and enough data /
 495  * metadata reservations in the contexts.
 496  *
 497  * Will return -EAGAIN, and a reason if a restart is needed.
 498  * If passed in, *reason will always be set, even in error.
 499  */
 500 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 501                                struct inode *inode,
 502                                u32 *logical_offset,
 503                                u32 clusters_to_add,
 504                                int mark_unwritten,
 505                                struct buffer_head *fe_bh,
 506                                handle_t *handle,
 507                                struct ocfs2_alloc_context *data_ac,
 508                                struct ocfs2_alloc_context *meta_ac,
 509                                enum ocfs2_alloc_restarted *reason_ret)
 510 {
 511         int status = 0;
 512         int free_extents;
 513         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
 514         enum ocfs2_alloc_restarted reason = RESTART_NONE;
 515         u32 bit_off, num_bits;
 516         u64 block;
 517         u8 flags = 0;
 518
 519         BUG_ON(!clusters_to_add);
 520
 521         if (mark_unwritten)
 522                 flags = OCFS2_EXT_UNWRITTEN;
 523
 524         free_extents = ocfs2_num_free_extents(osb, inode, fe);
 525         if (free_extents < 0) {
 526                 status = free_extents;
 527                 mlog_errno(status);
 528                 goto leave;
 529         }
 530
 531         /* there are two cases which could cause us to EAGAIN in the
 532          * we-need-more-metadata case:
 533          * 1) we haven't reserved *any*
 534          * 2) we are so fragmented, we've needed to add metadata too
 535          *    many times. */
 536         if (!free_extents && !meta_ac) {
 537                 mlog(0, "we haven't reserved any metadata!\n");
 538                 status = -EAGAIN;
 539                 reason = RESTART_META;
 540                 goto leave;
 541         } else if ((!free_extents)
 542                    && (ocfs2_alloc_context_bits_left(meta_ac)
 543                        < ocfs2_extend_meta_needed(fe))) {
 544                 mlog(0, "filesystem is really fragmented...\n");
 545                 status = -EAGAIN;
 546                 reason = RESTART_META;
 547                 goto leave;
 548         }
 549
 550         status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
 551                                         clusters_to_add, &bit_off, &num_bits);
 552         if (status < 0) {
 553                 if (status != -ENOSPC)
 554                         mlog_errno(status);
 555                 goto leave;
 556         }
 557
 558         BUG_ON(num_bits > clusters_to_add);
 559
 560         /* reserve our write early -- insert_extent may update the inode */
 561         status = ocfs2_journal_access(handle, inode, fe_bh,
 562                                       OCFS2_JOURNAL_ACCESS_WRITE);
 563         if (status < 0) {
 564                 mlog_errno(status);
 565                 goto leave;
 566         }
 567
 568         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
 569         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
 570              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
 571         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
 572                                      *logical_offset, block, num_bits,
 573                                      flags, meta_ac);
 574         if (status < 0) {
 575                 mlog_errno(status);
 576                 goto leave;
 577         }
 578
 579         status = ocfs2_journal_dirty(handle, fe_bh);
 580         if (status < 0) {
 581                 mlog_errno(status);
 582                 goto leave;
 583         }
 584
 585         clusters_to_add -= num_bits;
 586         *logical_offset += num_bits;
 587
 588         if (clusters_to_add) {
 589                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
 590                      "%u\n", fe->i_clusters, clusters_to_add);
 591                 status = -EAGAIN;
 592                 reason = RESTART_TRANS;
 593         }
 594
 595 leave:
 596         mlog_exit(status);
 597         if (reason_ret)
 598                 *reason_ret = reason;
 599         return status;
 600 }
 601
 602 /*
 603  * For a given allocation, determine which allocators will need to be
 604  * accessed, and lock them, reserving the appropriate number of bits.
 605  *
 606  * Sparse file systems call this from ocfs2_write_begin_nolock()
 607  * and ocfs2_allocate_unwritten_extents().
 608  *
 609  * File systems which don't support holes call this from
 610  * ocfs2_extend_allocation().
 611  */
 612 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
 613                           u32 clusters_to_add, u32 extents_to_split,
 614                           struct ocfs2_alloc_context **data_ac,
 615                           struct ocfs2_alloc_context **meta_ac)
 616 {
 617         int ret = 0, num_free_extents;
 618         unsigned int max_recs_needed = clusters_to_add + 2 * extents_to_split;
 619         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 620
 621         *meta_ac = NULL;
 622         if (data_ac)
 623                 *data_ac = NULL;
 624
 625         BUG_ON(clusters_to_add != 0 && data_ac == NULL);
 626
 627         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 628              "clusters_to_add = %u, extents_to_split = %u\n",
 629              (unsigned long long)OCFS2_I(inode)->ip_blkno, (long long)i_size_read(inode),
 630              le32_to_cpu(di->i_clusters), clusters_to_add, extents_to_split);
 631
 632         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
 633         if (num_free_extents < 0) {
 634                 ret = num_free_extents;
 635                 mlog_errno(ret);
 636                 goto out;
 637         }
 638
 639         /*
 640          * Sparse allocation file systems need to be more conservative
 641          * with reserving room for expansion - the actual allocation
 642          * happens while we've got a journal handle open so re-taking
 643          * a cluster lock (because we ran out of room for another
 644          * extent) will violate ordering rules.
 645          *
 646          * Most of the time we'll only be seeing this 1 cluster at a time
 647          * anyway.
 648          *
 649          * Always lock for any unwritten extents - we might want to
 650          * add blocks during a split.
 651          */
 652         if (!num_free_extents ||
 653             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed)) {
 654                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
 655                 if (ret < 0) {
 656                         if (ret != -ENOSPC)
 657                                 mlog_errno(ret);
 658                         goto out;
 659                 }
 660         }
 661
 662         if (clusters_to_add == 0)
 663                 goto out;
 664
 665         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
 666         if (ret < 0) {
 667                 if (ret != -ENOSPC)
 668                         mlog_errno(ret);
 669                 goto out;
 670         }
 671
 672 out:
 673         if (ret) {
 674                 if (*meta_ac) {
 675                         ocfs2_free_alloc_context(*meta_ac);
 676                         *meta_ac = NULL;
 677                 }
 678
 679                 /*
 680                  * We cannot have an error and a non null *data_ac.
 681                  */
 682         }
 683
 684         return ret;
 685 }
 686
 687 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 688                                      u32 clusters_to_add, int mark_unwritten)
 689 {
 690         int status = 0;
 691         int restart_func = 0;
 692         int credits;
 693         u32 prev_clusters;
 694         struct buffer_head *bh = NULL;
 695         struct ocfs2_dinode *fe = NULL;
 696         handle_t *handle = NULL;
 697         struct ocfs2_alloc_context *data_ac = NULL;
 698         struct ocfs2_alloc_context *meta_ac = NULL;
 699         enum ocfs2_alloc_restarted why;
 700         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 701
 702         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 703
 704         /*
 705          * This function only exists for file systems which don't
 706          * support holes.
 707          */
 708         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 709
 710         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
 711                                   OCFS2_BH_CACHED, inode);
 712         if (status < 0) {
 713                 mlog_errno(status);
 714                 goto leave;
 715         }
 716
 717         fe = (struct ocfs2_dinode *) bh->b_data;
 718         if (!OCFS2_IS_VALID_DINODE(fe)) {
 719                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 720                 status = -EIO;
 721                 goto leave;
 722         }
 723
 724 restart_all:
 725         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 726
 727         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, 0, &data_ac,
 728                                        &meta_ac);
 729         if (status) {
 730                 mlog_errno(status);
 731                 goto leave;
 732         }
 733
 734         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
 735         handle = ocfs2_start_trans(osb, credits);
 736         if (IS_ERR(handle)) {
 737                 status = PTR_ERR(handle);
 738                 handle = NULL;
 739                 mlog_errno(status);
 740                 goto leave;
 741         }
 742
 743 restarted_transaction:
 744         /* reserve a write to the file entry early on - that we if we
 745          * run out of credits in the allocation path, we can still
 746          * update i_size. */
 747         status = ocfs2_journal_access(handle, inode, bh,
 748                                       OCFS2_JOURNAL_ACCESS_WRITE);
 749         if (status < 0) {
 750                 mlog_errno(status);
 751                 goto leave;
 752         }
 753
 754         prev_clusters = OCFS2_I(inode)->ip_clusters;
 755
 756         status = ocfs2_do_extend_allocation(osb,
 757                                             inode,
 758                                             &logical_start,
 759                                             clusters_to_add,
 760                                             mark_unwritten,
 761                                             bh,
 762                                             handle,
 763                                             data_ac,
 764                                             meta_ac,
 765                                             &why);
 766         if ((status < 0) && (status != -EAGAIN)) {
 767                 if (status != -ENOSPC)
 768                         mlog_errno(status);
 769                 goto leave;
 770         }
 771
 772         status = ocfs2_journal_dirty(handle, bh);
 773         if (status < 0) {
 774                 mlog_errno(status);
 775                 goto leave;
 776         }
 777
 778         spin_lock(&OCFS2_I(inode)->ip_lock);
 779         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 780         spin_unlock(&OCFS2_I(inode)->ip_lock);
 781
 782         if (why != RESTART_NONE && clusters_to_add) {
 783                 if (why == RESTART_META) {
 784                         mlog(0, "restarting function.\n");
 785                         restart_func = 1;
 786                 } else {
 787                         BUG_ON(why != RESTART_TRANS);
 788
 789                         mlog(0, "restarting transaction.\n");
 790                         /* TODO: This can be more intelligent. */
 791                         credits = ocfs2_calc_extend_credits(osb->sb,
 792                                                             fe,
 793                                                             clusters_to_add);
 794                         status = ocfs2_extend_trans(handle, credits);
 795                         if (status < 0) {
 796                                 /* handle still has to be committed at
 797                                  * this point. */
 798                                 status = -ENOMEM;
 799                                 mlog_errno(status);
 800                                 goto leave;
 801                         }
 802                         goto restarted_transaction;
 803                 }
 804         }
 805
 806         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 807              le32_to_cpu(fe->i_clusters),
 808              (unsigned long long)le64_to_cpu(fe->i_size));
 809         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 810              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 811
 812 leave:
 813         if (handle) {
 814                 ocfs2_commit_trans(osb, handle);
 815                 handle = NULL;
 816         }
 817         if (data_ac) {
 818                 ocfs2_free_alloc_context(data_ac);
 819                 data_ac = NULL;
 820         }
 821         if (meta_ac) {
 822                 ocfs2_free_alloc_context(meta_ac);
 823                 meta_ac = NULL;
 824         }
 825         if ((!status) && restart_func) {
 826                 restart_func = 0;
 827                 goto restart_all;
 828         }
 829         if (bh) {
 830                 brelse(bh);
 831                 bh = NULL;
 832         }
 833
 834         mlog_exit(status);
 835         return status;
 836 }
 837
 838 /* Some parts of this taken from generic_cont_expand, which turned out
 839  * to be too fragile to do exactly what we need without us having to
 840  * worry about recursive locking in ->prepare_write() and
 841  * ->commit_write(). */
 842 static int ocfs2_write_zero_page(struct inode *inode,
 843                                  u64 size)
 844 {
 845         struct address_space *mapping = inode->i_mapping;
 846         struct page *page;
 847         unsigned long index;
 848         unsigned int offset;
 849         handle_t *handle = NULL;
 850         int ret;
 851
 852         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 853         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 854         ** skip the prepare.  make sure we never send an offset for the start
 855         ** of a block
 856         */
 857         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 858                 offset++;
 859         }
 860         index = size >> PAGE_CACHE_SHIFT;
 861
 862         page = grab_cache_page(mapping, index);
 863         if (!page) {
 864                 ret = -ENOMEM;
 865                 mlog_errno(ret);
 866                 goto out;
 867         }
 868
 869         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 870         if (ret < 0) {
 871                 mlog_errno(ret);
 872                 goto out_unlock;
 873         }
 874
 875         if (ocfs2_should_order_data(inode)) {
 876                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 877                                                      offset);
 878                 if (IS_ERR(handle)) {
 879                         ret = PTR_ERR(handle);
 880                         handle = NULL;
 881                         goto out_unlock;
 882                 }
 883         }
 884
 885         /* must not update i_size! */
 886         ret = block_commit_write(page, offset, offset);
 887         if (ret < 0)
 888                 mlog_errno(ret);
 889         else
 890                 ret = 0;
 891
 892         if (handle)
 893                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 894 out_unlock:
 895         unlock_page(page);
 896         page_cache_release(page);
 897 out:
 898         return ret;
 899 }
 900
 901 static int ocfs2_zero_extend(struct inode *inode,
 902                              u64 zero_to_size)
 903 {
 904         int ret = 0;
 905         u64 start_off;
 906         struct super_block *sb = inode->i_sb;
 907
 908         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 909         while (start_off < zero_to_size) {
 910                 ret = ocfs2_write_zero_page(inode, start_off);
 911                 if (ret < 0) {
 912                         mlog_errno(ret);
 913                         goto out;
 914                 }
 915
 916                 start_off += sb->s_blocksize;
 917
 918                 /*
 919                  * Very large extends have the potential to lock up
 920                  * the cpu for extended periods of time.
 921                  */
 922                 cond_resched();
 923         }
 924
 925 out:
 926         return ret;
 927 }
 928
 929 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 930 {
 931         int ret;
 932         u32 clusters_to_add;
 933         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 934
 935         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 936         if (clusters_to_add < oi->ip_clusters)
 937                 clusters_to_add = 0;
 938         else
 939                 clusters_to_add -= oi->ip_clusters;
 940
 941         if (clusters_to_add) {
 942                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 943                                                 clusters_to_add, 0);
 944                 if (ret) {
 945                         mlog_errno(ret);
 946                         goto out;
 947                 }
 948         }
 949
 950         /*
 951          * Call this even if we don't add any clusters to the tree. We
 952          * still need to zero the area between the old i_size and the
 953          * new i_size.
 954          */
 955         ret = ocfs2_zero_extend(inode, zero_to);
 956         if (ret < 0)
 957                 mlog_errno(ret);
 958
 959 out:
 960         return ret;
 961 }
 962
 963 static int ocfs2_extend_file(struct inode *inode,
 964                              struct buffer_head *di_bh,
 965                              u64 new_i_size)
 966 {
 967         int ret = 0;
 968         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 969
 970         BUG_ON(!di_bh);
 971
 972         /* setattr sometimes calls us like this. */
 973         if (new_i_size == 0)
 974                 goto out;
 975
 976         if (i_size_read(inode) == new_i_size)
 977                 goto out;
 978         BUG_ON(new_i_size < i_size_read(inode));
 979
 980         /*
 981          * Fall through for converting inline data, even if the fs
 982          * supports sparse files.
 983          *
 984          * The check for inline data here is legal - nobody can add
 985          * the feature since we have i_mutex. We must check it again
 986          * after acquiring ip_alloc_sem though, as paths like mmap
 987          * might have raced us to converting the inode to extents.
 988          */
 989         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 990             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 991                 goto out_update_size;
 992
 993         /*
 994          * The alloc sem blocks people in read/write from reading our
 995          * allocation until we're done changing it. We depend on
 996          * i_mutex to block other extend/truncate calls while we're
 997          * here.
 998          */
 999         down_write(&oi->ip_alloc_sem);
1000
1001         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1002                 /*
1003                  * We can optimize small extends by keeping the inodes
1004                  * inline data.
1005                  */
1006                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1007                         up_write(&oi->ip_alloc_sem);
1008                         goto out_update_size;
1009                 }
1010
1011                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1012                 if (ret) {
1013                         up_write(&oi->ip_alloc_sem);
1014
1015                         mlog_errno(ret);
1016                         goto out;
1017                 }
1018         }
1019
1020         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1021                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
1022
1023         up_write(&oi->ip_alloc_sem);
1024
1025         if (ret < 0) {
1026                 mlog_errno(ret);
1027                 goto out;
1028         }
1029
1030 out_update_size:
1031         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1032         if (ret < 0)
1033                 mlog_errno(ret);
1034
1035 out:
1036         return ret;
1037 }
1038
1039 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1040 {
1041         int status = 0, size_change;
1042         struct inode *inode = dentry->d_inode;
1043         struct super_block *sb = inode->i_sb;
1044         struct ocfs2_super *osb = OCFS2_SB(sb);
1045         struct buffer_head *bh = NULL;
1046         handle_t *handle = NULL;
1047
1048         mlog_entry("(0x%p, '%.*s')\n", dentry,
1049                    dentry->d_name.len, dentry->d_name.name);
1050
1051         /* ensuring we don't even attempt to truncate a symlink */
1052         if (S_ISLNK(inode->i_mode))
1053                 attr->ia_valid &= ~ATTR_SIZE;
1054
1055         if (attr->ia_valid & ATTR_MODE)
1056                 mlog(0, "mode change: %d\n", attr->ia_mode);
1057         if (attr->ia_valid & ATTR_UID)
1058                 mlog(0, "uid change: %d\n", attr->ia_uid);
1059         if (attr->ia_valid & ATTR_GID)
1060                 mlog(0, "gid change: %d\n", attr->ia_gid);
1061         if (attr->ia_valid & ATTR_SIZE)
1062                 mlog(0, "size change...\n");
1063         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1064                 mlog(0, "time change...\n");
1065
1066 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1067                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1068         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1069                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1070                 return 0;
1071         }
1072
1073         status = inode_change_ok(inode, attr);
1074         if (status)
1075                 return status;
1076
1077         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1078         if (size_change) {
1079                 status = ocfs2_rw_lock(inode, 1);
1080                 if (status < 0) {
1081                         mlog_errno(status);
1082                         goto bail;
1083                 }
1084         }
1085
1086         status = ocfs2_inode_lock(inode, &bh, 1);
1087         if (status < 0) {
1088                 if (status != -ENOENT)
1089                         mlog_errno(status);
1090                 goto bail_unlock_rw;
1091         }
1092
1093         if (size_change && attr->ia_size != i_size_read(inode)) {
1094                 if (attr->ia_size > sb->s_maxbytes) {
1095                         status = -EFBIG;
1096                         goto bail_unlock;
1097                 }
1098
1099                 if (i_size_read(inode) > attr->ia_size)
1100                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1101                 else
1102                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1103                 if (status < 0) {
1104                         if (status != -ENOSPC)
1105                                 mlog_errno(status);
1106                         status = -ENOSPC;
1107                         goto bail_unlock;
1108                 }
1109         }
1110
1111         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1112         if (IS_ERR(handle)) {
1113                 status = PTR_ERR(handle);
1114                 mlog_errno(status);
1115                 goto bail_unlock;
1116         }
1117
1118         /*
1119          * This will intentionally not wind up calling vmtruncate(),
1120          * since all the work for a size change has been done above.
1121          * Otherwise, we could get into problems with truncate as
1122          * ip_alloc_sem is used there to protect against i_size
1123          * changes.
1124          */
1125         status = inode_setattr(inode, attr);
1126         if (status < 0) {
1127                 mlog_errno(status);
1128                 goto bail_commit;
1129         }
1130
1131         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1132         if (status < 0)
1133                 mlog_errno(status);
1134
1135 bail_commit:
1136         ocfs2_commit_trans(osb, handle);
1137 bail_unlock:
1138         ocfs2_inode_unlock(inode, 1);
1139 bail_unlock_rw:
1140         if (size_change)
1141                 ocfs2_rw_unlock(inode, 1);
1142 bail:
1143         if (bh)
1144                 brelse(bh);
1145
1146         mlog_exit(status);
1147         return status;
1148 }
1149
1150 int ocfs2_getattr(struct vfsmount *mnt,
1151                   struct dentry *dentry,
1152                   struct kstat *stat)
1153 {
1154         struct inode *inode = dentry->d_inode;
1155         struct super_block *sb = dentry->d_inode->i_sb;
1156         struct ocfs2_super *osb = sb->s_fs_info;
1157         int err;
1158
1159         mlog_entry_void();
1160
1161         err = ocfs2_inode_revalidate(dentry);
1162         if (err) {
1163                 if (err != -ENOENT)
1164                         mlog_errno(err);
1165                 goto bail;
1166         }
1167
1168         generic_fillattr(inode, stat);
1169
1170         /* We set the blksize from the cluster size for performance */
1171         stat->blksize = osb->s_clustersize;
1172
1173 bail:
1174         mlog_exit(err);
1175
1176         return err;
1177 }
1178
1179 int ocfs2_permission(struct inode *inode, int mask)
1180 {
1181         int ret;
1182
1183         mlog_entry_void();
1184
1185         ret = ocfs2_inode_lock(inode, NULL, 0);
1186         if (ret) {
1187                 if (ret != -ENOENT)
1188                         mlog_errno(ret);
1189                 goto out;
1190         }
1191
1192         ret = generic_permission(inode, mask, NULL);
1193
1194         ocfs2_inode_unlock(inode, 0);
1195 out:
1196         mlog_exit(ret);
1197         return ret;
1198 }
1199
1200 static int __ocfs2_write_remove_suid(struct inode *inode,
1201                                      struct buffer_head *bh)
1202 {
1203         int ret;
1204         handle_t *handle;
1205         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1206         struct ocfs2_dinode *di;
1207
1208         mlog_entry("(Inode %llu, mode 0%o)\n",
1209                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1210
1211         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1212         if (handle == NULL) {
1213                 ret = -ENOMEM;
1214                 mlog_errno(ret);
1215                 goto out;
1216         }
1217
1218         ret = ocfs2_journal_access(handle, inode, bh,
1219                                    OCFS2_JOURNAL_ACCESS_WRITE);
1220         if (ret < 0) {
1221                 mlog_errno(ret);
1222                 goto out_trans;
1223         }
1224
1225         inode->i_mode &= ~S_ISUID;
1226         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1227                 inode->i_mode &= ~S_ISGID;
1228
1229         di = (struct ocfs2_dinode *) bh->b_data;
1230         di->i_mode = cpu_to_le16(inode->i_mode);
1231
1232         ret = ocfs2_journal_dirty(handle, bh);
1233         if (ret < 0)
1234                 mlog_errno(ret);
1235
1236 out_trans:
1237         ocfs2_commit_trans(osb, handle);
1238 out:
1239         mlog_exit(ret);
1240         return ret;
1241 }
1242
1243 /*
1244  * Will look for holes and unwritten extents in the range starting at
1245  * pos for count bytes (inclusive).
1246  */
1247 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1248                                        size_t count)
1249 {
1250         int ret = 0;
1251         unsigned int extent_flags;
1252         u32 cpos, clusters, extent_len, phys_cpos;
1253         struct super_block *sb = inode->i_sb;
1254
1255         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1256         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1257
1258         while (clusters) {
1259                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1260                                          &extent_flags);
1261                 if (ret < 0) {
1262                         mlog_errno(ret);
1263                         goto out;
1264                 }
1265
1266                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1267                         ret = 1;
1268                         break;
1269                 }
1270
1271                 if (extent_len > clusters)
1272                         extent_len = clusters;
1273
1274                 clusters -= extent_len;
1275                 cpos += extent_len;
1276         }
1277 out:
1278         return ret;
1279 }
1280
1281 static int ocfs2_write_remove_suid(struct inode *inode)
1282 {
1283         int ret;
1284         struct buffer_head *bh = NULL;
1285         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1286
1287         ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1288                                oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1289         if (ret < 0) {
1290                 mlog_errno(ret);
1291                 goto out;
1292         }
1293
1294         ret =  __ocfs2_write_remove_suid(inode, bh);
1295 out:
1296         brelse(bh);
1297         return ret;
1298 }
1299
1300 /*
1301  * Allocate enough extents to cover the region starting at byte offset
1302  * start for len bytes. Existing extents are skipped, any extents
1303  * added are marked as "unwritten".
1304  */
1305 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1306                                             u64 start, u64 len)
1307 {
1308         int ret;
1309         u32 cpos, phys_cpos, clusters, alloc_size;
1310         u64 end = start + len;
1311         struct buffer_head *di_bh = NULL;
1312
1313         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1314                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1315                                        OCFS2_I(inode)->ip_blkno, &di_bh,
1316                                        OCFS2_BH_CACHED, inode);
1317                 if (ret) {
1318                         mlog_errno(ret);
1319                         goto out;
1320                 }
1321
1322                 /*
1323                  * Nothing to do if the requested reservation range
1324                  * fits within the inode.
1325                  */
1326                 if (ocfs2_size_fits_inline_data(di_bh, end))
1327                         goto out;
1328
1329                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1330                 if (ret) {
1331                         mlog_errno(ret);
1332                         goto out;
1333                 }
1334         }
1335
1336         /*
1337          * We consider both start and len to be inclusive.
1338          */
1339         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1340         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1341         clusters -= cpos;
1342
1343         while (clusters) {
1344                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1345                                          &alloc_size, NULL);
1346                 if (ret) {
1347                         mlog_errno(ret);
1348                         goto out;
1349                 }
1350
1351                 /*
1352                  * Hole or existing extent len can be arbitrary, so
1353                  * cap it to our own allocation request.
1354                  */
1355                 if (alloc_size > clusters)
1356                         alloc_size = clusters;
1357
1358                 if (phys_cpos) {
1359                         /*
1360                          * We already have an allocation at this
1361                          * region so we can safely skip it.
1362                          */
1363                         goto next;
1364                 }
1365
1366                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1367                 if (ret) {
1368                         if (ret != -ENOSPC)
1369                                 mlog_errno(ret);
1370                         goto out;
1371                 }
1372
1373 next:
1374                 cpos += alloc_size;
1375                 clusters -= alloc_size;
1376         }
1377
1378         ret = 0;
1379 out:
1380
1381         brelse(di_bh);
1382         return ret;
1383 }
1384
1385 static int __ocfs2_remove_inode_range(struct inode *inode,
1386                                       struct buffer_head *di_bh,
1387                                       u32 cpos, u32 phys_cpos, u32 len,
1388                                       struct ocfs2_cached_dealloc_ctxt *dealloc)
1389 {
1390         int ret;
1391         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
1392         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1393         struct inode *tl_inode = osb->osb_tl_inode;
1394         handle_t *handle;
1395         struct ocfs2_alloc_context *meta_ac = NULL;
1396         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1397
1398         ret = ocfs2_lock_allocators(inode, di, 0, 1, NULL, &meta_ac);
1399         if (ret) {
1400                 mlog_errno(ret);
1401                 return ret;
1402         }
1403
1404         mutex_lock(&tl_inode->i_mutex);
1405
1406         if (ocfs2_truncate_log_needs_flush(osb)) {
1407                 ret = __ocfs2_flush_truncate_log(osb);
1408                 if (ret < 0) {
1409                         mlog_errno(ret);
1410                         goto out;
1411                 }
1412         }
1413
1414         handle = ocfs2_start_trans(osb, OCFS2_REMOVE_EXTENT_CREDITS);
1415         if (handle == NULL) {
1416                 ret = -ENOMEM;
1417                 mlog_errno(ret);
1418                 goto out;
1419         }
1420
1421         ret = ocfs2_journal_access(handle, inode, di_bh,
1422                                    OCFS2_JOURNAL_ACCESS_WRITE);
1423         if (ret) {
1424                 mlog_errno(ret);
1425                 goto out;
1426         }
1427
1428         ret = ocfs2_remove_extent(inode, di_bh, cpos, len, handle, meta_ac,
1429                                   dealloc);
1430         if (ret) {
1431                 mlog_errno(ret);
1432                 goto out_commit;
1433         }
1434
1435         OCFS2_I(inode)->ip_clusters -= len;
1436         di->i_clusters = cpu_to_le32(OCFS2_I(inode)->ip_clusters);
1437
1438         ret = ocfs2_journal_dirty(handle, di_bh);
1439         if (ret) {
1440                 mlog_errno(ret);
1441                 goto out_commit;
1442         }
1443
1444         ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
1445         if (ret)
1446                 mlog_errno(ret);
1447
1448 out_commit:
1449         ocfs2_commit_trans(osb, handle);
1450 out:
1451         mutex_unlock(&tl_inode->i_mutex);
1452
1453         if (meta_ac)
1454                 ocfs2_free_alloc_context(meta_ac);
1455
1456         return ret;
1457 }
1458
1459 /*
1460  * Truncate a byte range, avoiding pages within partial clusters. This
1461  * preserves those pages for the zeroing code to write to.
1462  */
1463 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1464                                          u64 byte_len)
1465 {
1466         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1467         loff_t start, end;
1468         struct address_space *mapping = inode->i_mapping;
1469
1470         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1471         end = byte_start + byte_len;
1472         end = end & ~(osb->s_clustersize - 1);
1473
1474         if (start < end) {
1475                 unmap_mapping_range(mapping, start, end - start, 0);
1476                 truncate_inode_pages_range(mapping, start, end - 1);
1477         }
1478 }
1479
1480 static int ocfs2_zero_partial_clusters(struct inode *inode,
1481                                        u64 start, u64 len)
1482 {
1483         int ret = 0;
1484         u64 tmpend, end = start + len;
1485         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1486         unsigned int csize = osb->s_clustersize;
1487         handle_t *handle;
1488
1489         /*
1490          * The "start" and "end" values are NOT necessarily part of
1491          * the range whose allocation is being deleted. Rather, this
1492          * is what the user passed in with the request. We must zero
1493          * partial clusters here. There's no need to worry about
1494          * physical allocation - the zeroing code knows to skip holes.
1495          */
1496         mlog(0, "byte start: %llu, end: %llu\n",
1497              (unsigned long long)start, (unsigned long long)end);
1498
1499         /*
1500          * If both edges are on a cluster boundary then there's no
1501          * zeroing required as the region is part of the allocation to
1502          * be truncated.
1503          */
1504         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1505                 goto out;
1506
1507         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1508         if (handle == NULL) {
1509                 ret = -ENOMEM;
1510                 mlog_errno(ret);
1511                 goto out;
1512         }
1513
1514         /*
1515          * We want to get the byte offset of the end of the 1st cluster.
1516          */
1517         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1518         if (tmpend > end)
1519                 tmpend = end;
1520
1521         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1522              (unsigned long long)start, (unsigned long long)tmpend);
1523
1524         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1525         if (ret)
1526                 mlog_errno(ret);
1527
1528         if (tmpend < end) {
1529                 /*
1530                  * This may make start and end equal, but the zeroing
1531                  * code will skip any work in that case so there's no
1532                  * need to catch it up here.
1533                  */
1534                 start = end & ~(osb->s_clustersize - 1);
1535
1536                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1537                      (unsigned long long)start, (unsigned long long)end);
1538
1539                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1540                 if (ret)
1541                         mlog_errno(ret);
1542         }
1543
1544         ocfs2_commit_trans(osb, handle);
1545 out:
1546         return ret;
1547 }
1548
1549 static int ocfs2_remove_inode_range(struct inode *inode,
1550                                     struct buffer_head *di_bh, u64 byte_start,
1551                                     u64 byte_len)
1552 {
1553         int ret = 0;
1554         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1555         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1556         struct ocfs2_cached_dealloc_ctxt dealloc;
1557         struct address_space *mapping = inode->i_mapping;
1558
1559         ocfs2_init_dealloc_ctxt(&dealloc);
1560
1561         if (byte_len == 0)
1562                 return 0;
1563
1564         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1565                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1566                                             byte_start + byte_len, 0);
1567                 if (ret) {
1568                         mlog_errno(ret);
1569                         goto out;
1570                 }
1571                 /*
1572                  * There's no need to get fancy with the page cache
1573                  * truncate of an inline-data inode. We're talking
1574                  * about less than a page here, which will be cached
1575                  * in the dinode buffer anyway.
1576                  */
1577                 unmap_mapping_range(mapping, 0, 0, 0);
1578                 truncate_inode_pages(mapping, 0);
1579                 goto out;
1580         }
1581
1582         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1583         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1584         if (trunc_len >= trunc_start)
1585                 trunc_len -= trunc_start;
1586         else
1587                 trunc_len = 0;
1588
1589         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1590              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1591              (unsigned long long)byte_start,
1592              (unsigned long long)byte_len, trunc_start, trunc_len);
1593
1594         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1595         if (ret) {
1596                 mlog_errno(ret);
1597                 goto out;
1598         }
1599
1600         cpos = trunc_start;
1601         while (trunc_len) {
1602                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1603                                          &alloc_size, NULL);
1604                 if (ret) {
1605                         mlog_errno(ret);
1606                         goto out;
1607                 }
1608
1609                 if (alloc_size > trunc_len)
1610                         alloc_size = trunc_len;
1611
1612                 /* Only do work for non-holes */
1613                 if (phys_cpos != 0) {
1614                         ret = __ocfs2_remove_inode_range(inode, di_bh, cpos,
1615                                                          phys_cpos, alloc_size,
1616                                                          &dealloc);
1617                         if (ret) {
1618                                 mlog_errno(ret);
1619                                 goto out;
1620                         }
1621                 }
1622
1623                 cpos += alloc_size;
1624                 trunc_len -= alloc_size;
1625         }
1626
1627         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1628
1629 out:
1630         ocfs2_schedule_truncate_log_flush(osb, 1);
1631         ocfs2_run_deallocs(osb, &dealloc);
1632
1633         return ret;
1634 }
1635
1636 /*
1637  * Parts of this function taken from xfs_change_file_space()
1638  */
1639 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1640                                      loff_t f_pos, unsigned int cmd,
1641                                      struct ocfs2_space_resv *sr,
1642                                      int change_size)
1643 {
1644         int ret;
1645         s64 llen;
1646         loff_t size;
1647         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1648         struct buffer_head *di_bh = NULL;
1649         handle_t *handle;
1650         unsigned long long max_off = inode->i_sb->s_maxbytes;
1651
1652         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1653                 return -EROFS;
1654
1655         mutex_lock(&inode->i_mutex);
1656
1657         /*
1658          * This prevents concurrent writes on other nodes
1659          */
1660         ret = ocfs2_rw_lock(inode, 1);
1661         if (ret) {
1662                 mlog_errno(ret);
1663                 goto out;
1664         }
1665
1666         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1667         if (ret) {
1668                 mlog_errno(ret);
1669                 goto out_rw_unlock;
1670         }
1671
1672         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1673                 ret = -EPERM;
1674                 goto out_inode_unlock;
1675         }
1676
1677         switch (sr->l_whence) {
1678         case 0: /*SEEK_SET*/
1679                 break;
1680         case 1: /*SEEK_CUR*/
1681                 sr->l_start += f_pos;
1682                 break;
1683         case 2: /*SEEK_END*/
1684                 sr->l_start += i_size_read(inode);
1685                 break;
1686         default:
1687                 ret = -EINVAL;
1688                 goto out_inode_unlock;
1689         }
1690         sr->l_whence = 0;
1691
1692         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1693
1694         if (sr->l_start < 0
1695             || sr->l_start > max_off
1696             || (sr->l_start + llen) < 0
1697             || (sr->l_start + llen) > max_off) {
1698                 ret = -EINVAL;
1699                 goto out_inode_unlock;
1700         }
1701         size = sr->l_start + sr->l_len;
1702
1703         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1704                 if (sr->l_len <= 0) {
1705                         ret = -EINVAL;
1706                         goto out_inode_unlock;
1707                 }
1708         }
1709
1710         if (file && should_remove_suid(file->f_path.dentry)) {
1711                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1712                 if (ret) {
1713                         mlog_errno(ret);
1714                         goto out_inode_unlock;
1715                 }
1716         }
1717
1718         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1719         switch (cmd) {
1720         case OCFS2_IOC_RESVSP:
1721         case OCFS2_IOC_RESVSP64:
1722                 /*
1723                  * This takes unsigned offsets, but the signed ones we
1724                  * pass have been checked against overflow above.
1725                  */
1726                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1727                                                        sr->l_len);
1728                 break;
1729         case OCFS2_IOC_UNRESVSP:
1730         case OCFS2_IOC_UNRESVSP64:
1731                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1732                                                sr->l_len);
1733                 break;
1734         default:
1735                 ret = -EINVAL;
1736         }
1737         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1738         if (ret) {
1739                 mlog_errno(ret);
1740                 goto out_inode_unlock;
1741         }
1742
1743         /*
1744          * We update c/mtime for these changes
1745          */
1746         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1747         if (IS_ERR(handle)) {
1748                 ret = PTR_ERR(handle);
1749                 mlog_errno(ret);
1750                 goto out_inode_unlock;
1751         }
1752
1753         if (change_size && i_size_read(inode) < size)
1754                 i_size_write(inode, size);
1755
1756         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1757         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1758         if (ret < 0)
1759                 mlog_errno(ret);
1760
1761         ocfs2_commit_trans(osb, handle);
1762
1763 out_inode_unlock:
1764         brelse(di_bh);
1765         ocfs2_inode_unlock(inode, 1);
1766 out_rw_unlock:
1767         ocfs2_rw_unlock(inode, 1);
1768
1769 out:
1770         mutex_unlock(&inode->i_mutex);
1771         return ret;
1772 }
1773
1774 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1775                             struct ocfs2_space_resv *sr)
1776 {
1777         struct inode *inode = file->f_path.dentry->d_inode;
1778         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);;
1779
1780         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1781             !ocfs2_writes_unwritten_extents(osb))
1782                 return -ENOTTY;
1783         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1784                  !ocfs2_sparse_alloc(osb))
1785                 return -ENOTTY;
1786
1787         if (!S_ISREG(inode->i_mode))
1788                 return -EINVAL;
1789
1790         if (!(file->f_mode & FMODE_WRITE))
1791                 return -EBADF;
1792
1793         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1794 }
1795
1796 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1797                             loff_t len)
1798 {
1799         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1800         struct ocfs2_space_resv sr;
1801         int change_size = 1;
1802
1803         if (!ocfs2_writes_unwritten_extents(osb))
1804                 return -EOPNOTSUPP;
1805
1806         if (S_ISDIR(inode->i_mode))
1807                 return -ENODEV;
1808
1809         if (mode & FALLOC_FL_KEEP_SIZE)
1810                 change_size = 0;
1811
1812         sr.l_whence = 0;
1813         sr.l_start = (s64)offset;
1814         sr.l_len = (s64)len;
1815
1816         return __ocfs2_change_file_space(NULL, inode, offset,
1817                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1818 }
1819
1820 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1821                                          loff_t *ppos,
1822                                          size_t count,
1823                                          int appending,
1824                                          int *direct_io)
1825 {
1826         int ret = 0, meta_level = 0;
1827         struct inode *inode = dentry->d_inode;
1828         loff_t saved_pos, end;
1829
1830         /*
1831          * We start with a read level meta lock and only jump to an ex
1832          * if we need to make modifications here.
1833          */
1834         for(;;) {
1835                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1836                 if (ret < 0) {
1837                         meta_level = -1;
1838                         mlog_errno(ret);
1839                         goto out;
1840                 }
1841
1842                 /* Clear suid / sgid if necessary. We do this here
1843                  * instead of later in the write path because
1844                  * remove_suid() calls ->setattr without any hint that
1845                  * we may have already done our cluster locking. Since
1846                  * ocfs2_setattr() *must* take cluster locks to
1847                  * proceeed, this will lead us to recursively lock the
1848                  * inode. There's also the dinode i_size state which
1849                  * can be lost via setattr during extending writes (we
1850                  * set inode->i_size at the end of a write. */
1851                 if (should_remove_suid(dentry)) {
1852                         if (meta_level == 0) {
1853                                 ocfs2_inode_unlock(inode, meta_level);
1854                                 meta_level = 1;
1855                                 continue;
1856                         }
1857
1858                         ret = ocfs2_write_remove_suid(inode);
1859                         if (ret < 0) {
1860                                 mlog_errno(ret);
1861                                 goto out_unlock;
1862                         }
1863                 }
1864
1865                 /* work on a copy of ppos until we're sure that we won't have
1866                  * to recalculate it due to relocking. */
1867                 if (appending) {
1868                         saved_pos = i_size_read(inode);
1869                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1870                 } else {
1871                         saved_pos = *ppos;
1872                 }
1873
1874                 end = saved_pos + count;
1875
1876                 /*
1877                  * Skip the O_DIRECT checks if we don't need
1878                  * them.
1879                  */
1880                 if (!direct_io || !(*direct_io))
1881                         break;
1882
1883                 /*
1884                  * There's no sane way to do direct writes to an inode
1885                  * with inline data.
1886                  */
1887                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1888                         *direct_io = 0;
1889                         break;
1890                 }
1891
1892                 /*
1893                  * Allowing concurrent direct writes means
1894                  * i_size changes wouldn't be synchronized, so
1895                  * one node could wind up truncating another
1896                  * nodes writes.
1897                  */
1898                 if (end > i_size_read(inode)) {
1899                         *direct_io = 0;
1900                         break;
1901                 }
1902
1903                 /*
1904                  * We don't fill holes during direct io, so
1905                  * check for them here. If any are found, the
1906                  * caller will have to retake some cluster
1907                  * locks and initiate the io as buffered.
1908                  */
1909                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1910                 if (ret == 1) {
1911                         *direct_io = 0;
1912                         ret = 0;
1913                 } else if (ret < 0)
1914                         mlog_errno(ret);
1915                 break;
1916         }
1917
1918         if (appending)
1919                 *ppos = saved_pos;
1920
1921 out_unlock:
1922         ocfs2_inode_unlock(inode, meta_level);
1923
1924 out:
1925         return ret;
1926 }
1927
1928 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1929                                     const struct iovec *iov,
1930                                     unsigned long nr_segs,
1931                                     loff_t pos)
1932 {
1933         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1934         int can_do_direct;
1935         ssize_t written = 0;
1936         size_t ocount;          /* original count */
1937         size_t count;           /* after file limit checks */
1938         loff_t old_size, *ppos = &iocb->ki_pos;
1939         u32 old_clusters;
1940         struct file *file = iocb->ki_filp;
1941         struct inode *inode = file->f_path.dentry->d_inode;
1942         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1943
1944         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1945                    (unsigned int)nr_segs,
1946                    file->f_path.dentry->d_name.len,
1947                    file->f_path.dentry->d_name.name);
1948
1949         if (iocb->ki_left == 0)
1950                 return 0;
1951
1952         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1953
1954         appending = file->f_flags & O_APPEND ? 1 : 0;
1955         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1956
1957         mutex_lock(&inode->i_mutex);
1958
1959 relock:
1960         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1961         if (direct_io) {
1962                 down_read(&inode->i_alloc_sem);
1963                 have_alloc_sem = 1;
1964         }
1965
1966         /* concurrent O_DIRECT writes are allowed */
1967         rw_level = !direct_io;
1968         ret = ocfs2_rw_lock(inode, rw_level);
1969         if (ret < 0) {
1970                 mlog_errno(ret);
1971                 goto out_sems;
1972         }
1973
1974         can_do_direct = direct_io;
1975         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1976                                             iocb->ki_left, appending,
1977                                             &can_do_direct);
1978         if (ret < 0) {
1979                 mlog_errno(ret);
1980                 goto out;
1981         }
1982
1983         /*
1984          * We can't complete the direct I/O as requested, fall back to
1985          * buffered I/O.
1986          */
1987         if (direct_io && !can_do_direct) {
1988                 ocfs2_rw_unlock(inode, rw_level);
1989                 up_read(&inode->i_alloc_sem);
1990
1991                 have_alloc_sem = 0;
1992                 rw_level = -1;
1993
1994                 direct_io = 0;
1995                 goto relock;
1996         }
1997
1998         /*
1999          * To later detect whether a journal commit for sync writes is
2000          * necessary, we sample i_size, and cluster count here.
2001          */
2002         old_size = i_size_read(inode);
2003         old_clusters = OCFS2_I(inode)->ip_clusters;
2004
2005         /* communicate with ocfs2_dio_end_io */
2006         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2007
2008         if (direct_io) {
2009                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
2010                                              VERIFY_READ);
2011                 if (ret)
2012                         goto out_dio;
2013
2014                 ret = generic_write_checks(file, ppos, &count,
2015                                            S_ISBLK(inode->i_mode));
2016                 if (ret)
2017                         goto out_dio;
2018
2019                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2020                                                     ppos, count, ocount);
2021                 if (written < 0) {
2022                         ret = written;
2023                         goto out_dio;
2024                 }
2025         } else {
2026                 written = generic_file_aio_write_nolock(iocb, iov, nr_segs,
2027                                                         *ppos);
2028         }
2029
2030 out_dio:
2031         /* buffered aio wouldn't have proper lock coverage today */
2032         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2033
2034         if ((file->f_flags & O_SYNC && !direct_io) || IS_SYNC(inode)) {
2035                 /*
2036                  * The generic write paths have handled getting data
2037                  * to disk, but since we don't make use of the dirty
2038                  * inode list, a manual journal commit is necessary
2039                  * here.
2040                  */
2041                 if (old_size != i_size_read(inode) ||
2042                     old_clusters != OCFS2_I(inode)->ip_clusters) {
2043                         ret = journal_force_commit(osb->journal->j_journal);
2044                         if (ret < 0)
2045                                 written = ret;
2046                 }
2047         }
2048
2049         /*
2050          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2051          * function pointer which is called when o_direct io completes so that
2052          * it can unlock our rw lock.  (it's the clustered equivalent of
2053          * i_alloc_sem; protects truncate from racing with pending ios).
2054          * Unfortunately there are error cases which call end_io and others
2055          * that don't.  so we don't have to unlock the rw_lock if either an
2056          * async dio is going to do it in the future or an end_io after an
2057          * error has already done it.
2058          */
2059         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2060                 rw_level = -1;
2061                 have_alloc_sem = 0;
2062         }
2063
2064 out:
2065         if (rw_level != -1)
2066                 ocfs2_rw_unlock(inode, rw_level);
2067
2068 out_sems:
2069         if (have_alloc_sem)
2070                 up_read(&inode->i_alloc_sem);
2071
2072         mutex_unlock(&inode->i_mutex);
2073
2074         mlog_exit(ret);
2075         return written ? written : ret;
2076 }
2077
2078 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2079                                        struct file *out,
2080                                        loff_t *ppos,
2081                                        size_t len,
2082                                        unsigned int flags)
2083 {
2084         int ret;
2085         struct inode *inode = out->f_path.dentry->d_inode;
2086
2087         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2088                    (unsigned int)len,
2089                    out->f_path.dentry->d_name.len,
2090                    out->f_path.dentry->d_name.name);
2091
2092         inode_double_lock(inode, pipe->inode);
2093
2094         ret = ocfs2_rw_lock(inode, 1);
2095         if (ret < 0) {
2096                 mlog_errno(ret);
2097                 goto out;
2098         }
2099
2100         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
2101                                             NULL);
2102         if (ret < 0) {
2103                 mlog_errno(ret);
2104                 goto out_unlock;
2105         }
2106
2107         ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags);
2108
2109 out_unlock:
2110         ocfs2_rw_unlock(inode, 1);
2111 out:
2112         inode_double_unlock(inode, pipe->inode);
2113
2114         mlog_exit(ret);
2115         return ret;
2116 }
2117
2118 static ssize_t ocfs2_file_splice_read(struct file *in,
2119                                       loff_t *ppos,
2120                                       struct pipe_inode_info *pipe,
2121                                       size_t len,
2122                                       unsigned int flags)
2123 {
2124         int ret = 0;
2125         struct inode *inode = in->f_path.dentry->d_inode;
2126
2127         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2128                    (unsigned int)len,
2129                    in->f_path.dentry->d_name.len,
2130                    in->f_path.dentry->d_name.name);
2131
2132         /*
2133          * See the comment in ocfs2_file_aio_read()
2134          */
2135         ret = ocfs2_inode_lock(inode, NULL, 0);
2136         if (ret < 0) {
2137                 mlog_errno(ret);
2138                 goto bail;
2139         }
2140         ocfs2_inode_unlock(inode, 0);
2141
2142         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2143
2144 bail:
2145         mlog_exit(ret);
2146         return ret;
2147 }
2148
2149 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2150                                    const struct iovec *iov,
2151                                    unsigned long nr_segs,
2152                                    loff_t pos)
2153 {
2154         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2155         struct file *filp = iocb->ki_filp;
2156         struct inode *inode = filp->f_path.dentry->d_inode;
2157
2158         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2159                    (unsigned int)nr_segs,
2160                    filp->f_path.dentry->d_name.len,
2161                    filp->f_path.dentry->d_name.name);
2162
2163         if (!inode) {
2164                 ret = -EINVAL;
2165                 mlog_errno(ret);
2166                 goto bail;
2167         }
2168
2169         /*
2170          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2171          * need locks to protect pending reads from racing with truncate.
2172          */
2173         if (filp->f_flags & O_DIRECT) {
2174                 down_read(&inode->i_alloc_sem);
2175                 have_alloc_sem = 1;
2176
2177                 ret = ocfs2_rw_lock(inode, 0);
2178                 if (ret < 0) {
2179                         mlog_errno(ret);
2180                         goto bail;
2181                 }
2182                 rw_level = 0;
2183                 /* communicate with ocfs2_dio_end_io */
2184                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2185         }
2186
2187         /*
2188          * We're fine letting folks race truncates and extending
2189          * writes with read across the cluster, just like they can
2190          * locally. Hence no rw_lock during read.
2191          *
2192          * Take and drop the meta data lock to update inode fields
2193          * like i_size. This allows the checks down below
2194          * generic_file_aio_read() a chance of actually working.
2195          */
2196         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2197         if (ret < 0) {
2198                 mlog_errno(ret);
2199                 goto bail;
2200         }
2201         ocfs2_inode_unlock(inode, lock_level);
2202
2203         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2204         if (ret == -EINVAL)
2205                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2206
2207         /* buffered aio wouldn't have proper lock coverage today */
2208         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2209
2210         /* see ocfs2_file_aio_write */
2211         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2212                 rw_level = -1;
2213                 have_alloc_sem = 0;
2214         }
2215
2216 bail:
2217         if (have_alloc_sem)
2218                 up_read(&inode->i_alloc_sem);
2219         if (rw_level != -1)
2220                 ocfs2_rw_unlock(inode, rw_level);
2221         mlog_exit(ret);
2222
2223         return ret;
2224 }
2225
2226 const struct inode_operations ocfs2_file_iops = {
2227         .setattr        = ocfs2_setattr,
2228         .getattr        = ocfs2_getattr,
2229         .permission     = ocfs2_permission,
2230         .fallocate      = ocfs2_fallocate,
2231 };
2232
2233 const struct inode_operations ocfs2_special_file_iops = {
2234         .setattr        = ocfs2_setattr,
2235         .getattr        = ocfs2_getattr,
2236         .permission     = ocfs2_permission,
2237 };
2238
2239 const struct file_operations ocfs2_fops = {
2240         .llseek         = generic_file_llseek,
2241         .read           = do_sync_read,
2242         .write          = do_sync_write,
2243         .mmap           = ocfs2_mmap,
2244         .fsync          = ocfs2_sync_file,
2245         .release        = ocfs2_file_release,
2246         .open           = ocfs2_file_open,
2247         .aio_read       = ocfs2_file_aio_read,
2248         .aio_write      = ocfs2_file_aio_write,
2249         .unlocked_ioctl = ocfs2_ioctl,
2250 #ifdef CONFIG_COMPAT
2251         .compat_ioctl   = ocfs2_compat_ioctl,
2252 #endif
2253         .flock          = ocfs2_flock,
2254         .splice_read    = ocfs2_file_splice_read,
2255         .splice_write   = ocfs2_file_splice_write,
2256 };
2257
2258 const struct file_operations ocfs2_dops = {
2259         .llseek         = generic_file_llseek,
2260         .read           = generic_read_dir,
2261         .readdir        = ocfs2_readdir,
2262         .fsync          = ocfs2_sync_file,
2263         .release        = ocfs2_dir_release,
2264         .open           = ocfs2_dir_open,
2265         .unlocked_ioctl = ocfs2_ioctl,
2266 #ifdef CONFIG_COMPAT
2267         .compat_ioctl   = ocfs2_compat_ioctl,
2268 #endif
2269         .flock          = ocfs2_flock,
2270 };