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