fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37 #include <linux/falloc.h>
  38 #include <linux/quotaops.h>
  39
  40 #define MLOG_MASK_PREFIX ML_INODE
  41 #include <cluster/masklog.h>
  42
  43 #include "ocfs2.h"
  44
  45 #include "alloc.h"
  46 #include "aops.h"
  47 #include "dir.h"
  48 #include "dlmglue.h"
  49 #include "extent_map.h"
  50 #include "file.h"
  51 #include "sysfile.h"
  52 #include "inode.h"
  53 #include "ioctl.h"
  54 #include "journal.h"
  55 #include "locks.h"
  56 #include "mmap.h"
  57 #include "suballoc.h"
  58 #include "super.h"
  59 #include "xattr.h"
  60 #include "acl.h"
  61 #include "quota.h"
  62 #include "refcounttree.h"
  63
  64 #include "buffer_head_io.h"
  65
  66 static int ocfs2_sync_inode(struct inode *inode)
  67 {
  68         filemap_fdatawrite(inode->i_mapping);
  69         return sync_mapping_buffers(inode->i_mapping);
  70 }
  71
  72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  73 {
  74         struct ocfs2_file_private *fp;
  75
  76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  77         if (!fp)
  78                 return -ENOMEM;
  79
  80         fp->fp_file = file;
  81         mutex_init(&fp->fp_mutex);
  82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  83         file->private_data = fp;
  84
  85         return 0;
  86 }
  87
  88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  89 {
  90         struct ocfs2_file_private *fp = file->private_data;
  91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  92
  93         if (fp) {
  94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  95                 ocfs2_lock_res_free(&fp->fp_flock);
  96                 kfree(fp);
  97                 file->private_data = NULL;
  98         }
  99 }
 100
 101 static int ocfs2_file_open(struct inode *inode, struct file *file)
 102 {
 103         int status;
 104         int mode = file->f_flags;
 105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 106
 107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 109
 110         if (file->f_mode & FMODE_WRITE)
 111                 dquot_initialize(inode);
 112
 113         spin_lock(&oi->ip_lock);
 114
 115         /* Check that the inode hasn't been wiped from disk by another
 116          * node. If it hasn't then we're safe as long as we hold the
 117          * spin lock until our increment of open count. */
 118         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 119                 spin_unlock(&oi->ip_lock);
 120
 121                 status = -ENOENT;
 122                 goto leave;
 123         }
 124
 125         if (mode & O_DIRECT)
 126                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 127
 128         oi->ip_open_count++;
 129         spin_unlock(&oi->ip_lock);
 130
 131         status = ocfs2_init_file_private(inode, file);
 132         if (status) {
 133                 /*
 134                  * We want to set open count back if we're failing the
 135                  * open.
 136                  */
 137                 spin_lock(&oi->ip_lock);
 138                 oi->ip_open_count--;
 139                 spin_unlock(&oi->ip_lock);
 140         }
 141
 142 leave:
 143         mlog_exit(status);
 144         return status;
 145 }
 146
 147 static int ocfs2_file_release(struct inode *inode, struct file *file)
 148 {
 149         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 150
 151         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 152                        file->f_path.dentry->d_name.len,
 153                        file->f_path.dentry->d_name.name);
 154
 155         spin_lock(&oi->ip_lock);
 156         if (!--oi->ip_open_count)
 157                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 158         spin_unlock(&oi->ip_lock);
 159
 160         ocfs2_free_file_private(inode, file);
 161
 162         mlog_exit(0);
 163
 164         return 0;
 165 }
 166
 167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 168 {
 169         return ocfs2_init_file_private(inode, file);
 170 }
 171
 172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 173 {
 174         ocfs2_free_file_private(inode, file);
 175         return 0;
 176 }
 177
 178 static int ocfs2_sync_file(struct file *file,
 179                            struct dentry *dentry,
 180                            int datasync)
 181 {
 182         int err = 0;
 183         journal_t *journal;
 184         struct inode *inode = dentry->d_inode;
 185         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 186
 187         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 188                    dentry->d_name.len, dentry->d_name.name);
 189
 190         err = ocfs2_sync_inode(dentry->d_inode);
 191         if (err)
 192                 goto bail;
 193
 194         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
 195                 goto bail;
 196
 197         journal = osb->journal->j_journal;
 198         err = jbd2_journal_force_commit(journal);
 199
 200 bail:
 201         mlog_exit(err);
 202
 203         return (err < 0) ? -EIO : 0;
 204 }
 205
 206 int ocfs2_should_update_atime(struct inode *inode,
 207                               struct vfsmount *vfsmnt)
 208 {
 209         struct timespec now;
 210         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 211
 212         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 213                 return 0;
 214
 215         if ((inode->i_flags & S_NOATIME) ||
 216             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 217                 return 0;
 218
 219         /*
 220          * We can be called with no vfsmnt structure - NFSD will
 221          * sometimes do this.
 222          *
 223          * Note that our action here is different than touch_atime() -
 224          * if we can't tell whether this is a noatime mount, then we
 225          * don't know whether to trust the value of s_atime_quantum.
 226          */
 227         if (vfsmnt == NULL)
 228                 return 0;
 229
 230         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 231             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 232                 return 0;
 233
 234         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 235                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 236                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 237                         return 1;
 238
 239                 return 0;
 240         }
 241
 242         now = CURRENT_TIME;
 243         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 244                 return 0;
 245         else
 246                 return 1;
 247 }
 248
 249 int ocfs2_update_inode_atime(struct inode *inode,
 250                              struct buffer_head *bh)
 251 {
 252         int ret;
 253         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 254         handle_t *handle;
 255         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 256
 257         mlog_entry_void();
 258
 259         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 260         if (IS_ERR(handle)) {
 261                 ret = PTR_ERR(handle);
 262                 mlog_errno(ret);
 263                 goto out;
 264         }
 265
 266         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 267                                       OCFS2_JOURNAL_ACCESS_WRITE);
 268         if (ret) {
 269                 mlog_errno(ret);
 270                 goto out_commit;
 271         }
 272
 273         /*
 274          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 275          * have i_mutex to guard against concurrent changes to other
 276          * inode fields.
 277          */
 278         inode->i_atime = CURRENT_TIME;
 279         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 280         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 281
 282         ret = ocfs2_journal_dirty(handle, bh);
 283         if (ret < 0)
 284                 mlog_errno(ret);
 285
 286 out_commit:
 287         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 288 out:
 289         mlog_exit(ret);
 290         return ret;
 291 }
 292
 293 static int ocfs2_set_inode_size(handle_t *handle,
 294                                 struct inode *inode,
 295                                 struct buffer_head *fe_bh,
 296                                 u64 new_i_size)
 297 {
 298         int status;
 299
 300         mlog_entry_void();
 301         i_size_write(inode, new_i_size);
 302         inode->i_blocks = ocfs2_inode_sector_count(inode);
 303         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 304
 305         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 306         if (status < 0) {
 307                 mlog_errno(status);
 308                 goto bail;
 309         }
 310
 311 bail:
 312         mlog_exit(status);
 313         return status;
 314 }
 315
 316 int ocfs2_simple_size_update(struct inode *inode,
 317                              struct buffer_head *di_bh,
 318                              u64 new_i_size)
 319 {
 320         int ret;
 321         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 322         handle_t *handle = NULL;
 323
 324         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 325         if (IS_ERR(handle)) {
 326                 ret = PTR_ERR(handle);
 327                 mlog_errno(ret);
 328                 goto out;
 329         }
 330
 331         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 332                                    new_i_size);
 333         if (ret < 0)
 334                 mlog_errno(ret);
 335
 336         ocfs2_commit_trans(osb, handle);
 337 out:
 338         return ret;
 339 }
 340
 341 static int ocfs2_cow_file_pos(struct inode *inode,
 342                               struct buffer_head *fe_bh,
 343                               u64 offset)
 344 {
 345         int status;
 346         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 347         unsigned int num_clusters = 0;
 348         unsigned int ext_flags = 0;
 349
 350         /*
 351          * If the new offset is aligned to the range of the cluster, there is
 352          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 353          * CoW either.
 354          */
 355         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 356                 return 0;
 357
 358         status = ocfs2_get_clusters(inode, cpos, &phys,
 359                                     &num_clusters, &ext_flags);
 360         if (status) {
 361                 mlog_errno(status);
 362                 goto out;
 363         }
 364
 365         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 366                 goto out;
 367
 368         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 369
 370 out:
 371         return status;
 372 }
 373
 374 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 375                                      struct inode *inode,
 376                                      struct buffer_head *fe_bh,
 377                                      u64 new_i_size)
 378 {
 379         int status;
 380         handle_t *handle;
 381         struct ocfs2_dinode *di;
 382         u64 cluster_bytes;
 383
 384         mlog_entry_void();
 385
 386         /*
 387          * We need to CoW the cluster contains the offset if it is reflinked
 388          * since we will call ocfs2_zero_range_for_truncate later which will
 389          * write "0" from offset to the end of the cluster.
 390          */
 391         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 392         if (status) {
 393                 mlog_errno(status);
 394                 return status;
 395         }
 396
 397         /* TODO: This needs to actually orphan the inode in this
 398          * transaction. */
 399
 400         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 401         if (IS_ERR(handle)) {
 402                 status = PTR_ERR(handle);
 403                 mlog_errno(status);
 404                 goto out;
 405         }
 406
 407         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 408                                          OCFS2_JOURNAL_ACCESS_WRITE);
 409         if (status < 0) {
 410                 mlog_errno(status);
 411                 goto out_commit;
 412         }
 413
 414         /*
 415          * Do this before setting i_size.
 416          */
 417         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 418         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 419                                                cluster_bytes);
 420         if (status) {
 421                 mlog_errno(status);
 422                 goto out_commit;
 423         }
 424
 425         i_size_write(inode, new_i_size);
 426         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 427
 428         di = (struct ocfs2_dinode *) fe_bh->b_data;
 429         di->i_size = cpu_to_le64(new_i_size);
 430         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 431         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 432
 433         status = ocfs2_journal_dirty(handle, fe_bh);
 434         if (status < 0)
 435                 mlog_errno(status);
 436
 437 out_commit:
 438         ocfs2_commit_trans(osb, handle);
 439 out:
 440
 441         mlog_exit(status);
 442         return status;
 443 }
 444
 445 static int ocfs2_truncate_file(struct inode *inode,
 446                                struct buffer_head *di_bh,
 447                                u64 new_i_size)
 448 {
 449         int status = 0;
 450         struct ocfs2_dinode *fe = NULL;
 451         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 452         struct ocfs2_truncate_context *tc = NULL;
 453
 454         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 455                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 456                    (unsigned long long)new_i_size);
 457
 458         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 459          * already validated it */
 460         fe = (struct ocfs2_dinode *) di_bh->b_data;
 461
 462         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 463                         "Inode %llu, inode i_size = %lld != di "
 464                         "i_size = %llu, i_flags = 0x%x\n",
 465                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 466                         i_size_read(inode),
 467                         (unsigned long long)le64_to_cpu(fe->i_size),
 468                         le32_to_cpu(fe->i_flags));
 469
 470         if (new_i_size > le64_to_cpu(fe->i_size)) {
 471                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 472                      (unsigned long long)le64_to_cpu(fe->i_size),
 473                      (unsigned long long)new_i_size);
 474                 status = -EINVAL;
 475                 mlog_errno(status);
 476                 goto bail;
 477         }
 478
 479         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 480              (unsigned long long)le64_to_cpu(fe->i_blkno),
 481              (unsigned long long)le64_to_cpu(fe->i_size),
 482              (unsigned long long)new_i_size);
 483
 484         /* lets handle the simple truncate cases before doing any more
 485          * cluster locking. */
 486         if (new_i_size == le64_to_cpu(fe->i_size))
 487                 goto bail;
 488
 489         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 490
 491         /*
 492          * The inode lock forced other nodes to sync and drop their
 493          * pages, which (correctly) happens even if we have a truncate
 494          * without allocation change - ocfs2 cluster sizes can be much
 495          * greater than page size, so we have to truncate them
 496          * anyway.
 497          */
 498         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 499         truncate_inode_pages(inode->i_mapping, new_i_size);
 500
 501         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 502                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 503                                                i_size_read(inode), 1);
 504                 if (status)
 505                         mlog_errno(status);
 506
 507                 goto bail_unlock_sem;
 508         }
 509
 510         /* alright, we're going to need to do a full blown alloc size
 511          * change. Orphan the inode so that recovery can complete the
 512          * truncate if necessary. This does the task of marking
 513          * i_size. */
 514         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 515         if (status < 0) {
 516                 mlog_errno(status);
 517                 goto bail_unlock_sem;
 518         }
 519
 520         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 521         if (status < 0) {
 522                 mlog_errno(status);
 523                 goto bail_unlock_sem;
 524         }
 525
 526         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 527         if (status < 0) {
 528                 mlog_errno(status);
 529                 goto bail_unlock_sem;
 530         }
 531
 532         /* TODO: orphan dir cleanup here. */
 533 bail_unlock_sem:
 534         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 535
 536 bail:
 537         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 538                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 539
 540         mlog_exit(status);
 541         return status;
 542 }
 543
 544 /*
 545  * extend file allocation only here.
 546  * we'll update all the disk stuff, and oip->alloc_size
 547  *
 548  * expect stuff to be locked, a transaction started and enough data /
 549  * metadata reservations in the contexts.
 550  *
 551  * Will return -EAGAIN, and a reason if a restart is needed.
 552  * If passed in, *reason will always be set, even in error.
 553  */
 554 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 555                          struct inode *inode,
 556                          u32 *logical_offset,
 557                          u32 clusters_to_add,
 558                          int mark_unwritten,
 559                          struct buffer_head *fe_bh,
 560                          handle_t *handle,
 561                          struct ocfs2_alloc_context *data_ac,
 562                          struct ocfs2_alloc_context *meta_ac,
 563                          enum ocfs2_alloc_restarted *reason_ret)
 564 {
 565         int ret;
 566         struct ocfs2_extent_tree et;
 567
 568         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 569         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 570                                           clusters_to_add, mark_unwritten,
 571                                           data_ac, meta_ac, reason_ret);
 572
 573         return ret;
 574 }
 575
 576 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 577                                      u32 clusters_to_add, int mark_unwritten)
 578 {
 579         int status = 0;
 580         int restart_func = 0;
 581         int credits;
 582         u32 prev_clusters;
 583         struct buffer_head *bh = NULL;
 584         struct ocfs2_dinode *fe = NULL;
 585         handle_t *handle = NULL;
 586         struct ocfs2_alloc_context *data_ac = NULL;
 587         struct ocfs2_alloc_context *meta_ac = NULL;
 588         enum ocfs2_alloc_restarted why;
 589         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 590         struct ocfs2_extent_tree et;
 591         int did_quota = 0;
 592
 593         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 594
 595         /*
 596          * This function only exists for file systems which don't
 597          * support holes.
 598          */
 599         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 600
 601         status = ocfs2_read_inode_block(inode, &bh);
 602         if (status < 0) {
 603                 mlog_errno(status);
 604                 goto leave;
 605         }
 606         fe = (struct ocfs2_dinode *) bh->b_data;
 607
 608 restart_all:
 609         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 610
 611         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 612              "clusters_to_add = %u\n",
 613              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 614              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 615              clusters_to_add);
 616         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 617         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 618                                        &data_ac, &meta_ac);
 619         if (status) {
 620                 mlog_errno(status);
 621                 goto leave;
 622         }
 623
 624         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 625                                             clusters_to_add);
 626         handle = ocfs2_start_trans(osb, credits);
 627         if (IS_ERR(handle)) {
 628                 status = PTR_ERR(handle);
 629                 handle = NULL;
 630                 mlog_errno(status);
 631                 goto leave;
 632         }
 633
 634 restarted_transaction:
 635         status = dquot_alloc_space_nodirty(inode,
 636                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 637         if (status)
 638                 goto leave;
 639         did_quota = 1;
 640
 641         /* reserve a write to the file entry early on - that we if we
 642          * run out of credits in the allocation path, we can still
 643          * update i_size. */
 644         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 645                                          OCFS2_JOURNAL_ACCESS_WRITE);
 646         if (status < 0) {
 647                 mlog_errno(status);
 648                 goto leave;
 649         }
 650
 651         prev_clusters = OCFS2_I(inode)->ip_clusters;
 652
 653         status = ocfs2_add_inode_data(osb,
 654                                       inode,
 655                                       &logical_start,
 656                                       clusters_to_add,
 657                                       mark_unwritten,
 658                                       bh,
 659                                       handle,
 660                                       data_ac,
 661                                       meta_ac,
 662                                       &why);
 663         if ((status < 0) && (status != -EAGAIN)) {
 664                 if (status != -ENOSPC)
 665                         mlog_errno(status);
 666                 goto leave;
 667         }
 668
 669         status = ocfs2_journal_dirty(handle, bh);
 670         if (status < 0) {
 671                 mlog_errno(status);
 672                 goto leave;
 673         }
 674
 675         spin_lock(&OCFS2_I(inode)->ip_lock);
 676         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 677         spin_unlock(&OCFS2_I(inode)->ip_lock);
 678         /* Release unused quota reservation */
 679         dquot_free_space(inode,
 680                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 681         did_quota = 0;
 682
 683         if (why != RESTART_NONE && clusters_to_add) {
 684                 if (why == RESTART_META) {
 685                         mlog(0, "restarting function.\n");
 686                         restart_func = 1;
 687                         status = 0;
 688                 } else {
 689                         BUG_ON(why != RESTART_TRANS);
 690
 691                         mlog(0, "restarting transaction.\n");
 692                         /* TODO: This can be more intelligent. */
 693                         credits = ocfs2_calc_extend_credits(osb->sb,
 694                                                             &fe->id2.i_list,
 695                                                             clusters_to_add);
 696                         status = ocfs2_extend_trans(handle, credits);
 697                         if (status < 0) {
 698                                 /* handle still has to be committed at
 699                                  * this point. */
 700                                 status = -ENOMEM;
 701                                 mlog_errno(status);
 702                                 goto leave;
 703                         }
 704                         goto restarted_transaction;
 705                 }
 706         }
 707
 708         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 709              le32_to_cpu(fe->i_clusters),
 710              (unsigned long long)le64_to_cpu(fe->i_size));
 711         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 712              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 713
 714 leave:
 715         if (status < 0 && did_quota)
 716                 dquot_free_space(inode,
 717                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 718         if (handle) {
 719                 ocfs2_commit_trans(osb, handle);
 720                 handle = NULL;
 721         }
 722         if (data_ac) {
 723                 ocfs2_free_alloc_context(data_ac);
 724                 data_ac = NULL;
 725         }
 726         if (meta_ac) {
 727                 ocfs2_free_alloc_context(meta_ac);
 728                 meta_ac = NULL;
 729         }
 730         if ((!status) && restart_func) {
 731                 restart_func = 0;
 732                 goto restart_all;
 733         }
 734         brelse(bh);
 735         bh = NULL;
 736
 737         mlog_exit(status);
 738         return status;
 739 }
 740
 741 /* Some parts of this taken from generic_cont_expand, which turned out
 742  * to be too fragile to do exactly what we need without us having to
 743  * worry about recursive locking in ->write_begin() and ->write_end(). */
 744 static int ocfs2_write_zero_page(struct inode *inode,
 745                                  u64 size)
 746 {
 747         struct address_space *mapping = inode->i_mapping;
 748         struct page *page;
 749         unsigned long index;
 750         unsigned int offset;
 751         handle_t *handle = NULL;
 752         int ret;
 753
 754         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 755         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 756         ** skip the prepare.  make sure we never send an offset for the start
 757         ** of a block
 758         */
 759         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 760                 offset++;
 761         }
 762         index = size >> PAGE_CACHE_SHIFT;
 763
 764         page = grab_cache_page(mapping, index);
 765         if (!page) {
 766                 ret = -ENOMEM;
 767                 mlog_errno(ret);
 768                 goto out;
 769         }
 770
 771         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 772         if (ret < 0) {
 773                 mlog_errno(ret);
 774                 goto out_unlock;
 775         }
 776
 777         if (ocfs2_should_order_data(inode)) {
 778                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 779                                                      offset);
 780                 if (IS_ERR(handle)) {
 781                         ret = PTR_ERR(handle);
 782                         handle = NULL;
 783                         goto out_unlock;
 784                 }
 785         }
 786
 787         /* must not update i_size! */
 788         ret = block_commit_write(page, offset, offset);
 789         if (ret < 0)
 790                 mlog_errno(ret);
 791         else
 792                 ret = 0;
 793
 794         if (handle)
 795                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 796 out_unlock:
 797         unlock_page(page);
 798         page_cache_release(page);
 799 out:
 800         return ret;
 801 }
 802
 803 static int ocfs2_zero_extend(struct inode *inode,
 804                              u64 zero_to_size)
 805 {
 806         int ret = 0;
 807         u64 start_off;
 808         struct super_block *sb = inode->i_sb;
 809
 810         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 811         while (start_off < zero_to_size) {
 812                 ret = ocfs2_write_zero_page(inode, start_off);
 813                 if (ret < 0) {
 814                         mlog_errno(ret);
 815                         goto out;
 816                 }
 817
 818                 start_off += sb->s_blocksize;
 819
 820                 /*
 821                  * Very large extends have the potential to lock up
 822                  * the cpu for extended periods of time.
 823                  */
 824                 cond_resched();
 825         }
 826
 827 out:
 828         return ret;
 829 }
 830
 831 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 832 {
 833         int ret;
 834         u32 clusters_to_add;
 835         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 836
 837         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 838         if (clusters_to_add < oi->ip_clusters)
 839                 clusters_to_add = 0;
 840         else
 841                 clusters_to_add -= oi->ip_clusters;
 842
 843         if (clusters_to_add) {
 844                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 845                                                 clusters_to_add, 0);
 846                 if (ret) {
 847                         mlog_errno(ret);
 848                         goto out;
 849                 }
 850         }
 851
 852         /*
 853          * Call this even if we don't add any clusters to the tree. We
 854          * still need to zero the area between the old i_size and the
 855          * new i_size.
 856          */
 857         ret = ocfs2_zero_extend(inode, zero_to);
 858         if (ret < 0)
 859                 mlog_errno(ret);
 860
 861 out:
 862         return ret;
 863 }
 864
 865 static int ocfs2_extend_file(struct inode *inode,
 866                              struct buffer_head *di_bh,
 867                              u64 new_i_size)
 868 {
 869         int ret = 0;
 870         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 871
 872         BUG_ON(!di_bh);
 873
 874         /* setattr sometimes calls us like this. */
 875         if (new_i_size == 0)
 876                 goto out;
 877
 878         if (i_size_read(inode) == new_i_size)
 879                 goto out;
 880         BUG_ON(new_i_size < i_size_read(inode));
 881
 882         /*
 883          * Fall through for converting inline data, even if the fs
 884          * supports sparse files.
 885          *
 886          * The check for inline data here is legal - nobody can add
 887          * the feature since we have i_mutex. We must check it again
 888          * after acquiring ip_alloc_sem though, as paths like mmap
 889          * might have raced us to converting the inode to extents.
 890          */
 891         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 892             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 893                 goto out_update_size;
 894
 895         /*
 896          * The alloc sem blocks people in read/write from reading our
 897          * allocation until we're done changing it. We depend on
 898          * i_mutex to block other extend/truncate calls while we're
 899          * here.
 900          */
 901         down_write(&oi->ip_alloc_sem);
 902
 903         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 904                 /*
 905                  * We can optimize small extends by keeping the inodes
 906                  * inline data.
 907                  */
 908                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 909                         up_write(&oi->ip_alloc_sem);
 910                         goto out_update_size;
 911                 }
 912
 913                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 914                 if (ret) {
 915                         up_write(&oi->ip_alloc_sem);
 916
 917                         mlog_errno(ret);
 918                         goto out;
 919                 }
 920         }
 921
 922         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 923                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 924
 925         up_write(&oi->ip_alloc_sem);
 926
 927         if (ret < 0) {
 928                 mlog_errno(ret);
 929                 goto out;
 930         }
 931
 932 out_update_size:
 933         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 934         if (ret < 0)
 935                 mlog_errno(ret);
 936
 937 out:
 938         return ret;
 939 }
 940
 941 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 942 {
 943         int status = 0, size_change;
 944         struct inode *inode = dentry->d_inode;
 945         struct super_block *sb = inode->i_sb;
 946         struct ocfs2_super *osb = OCFS2_SB(sb);
 947         struct buffer_head *bh = NULL;
 948         handle_t *handle = NULL;
 949         int qtype;
 950         struct dquot *transfer_from[MAXQUOTAS] = { };
 951         struct dquot *transfer_to[MAXQUOTAS] = { };
 952
 953         mlog_entry("(0x%p, '%.*s')\n", dentry,
 954                    dentry->d_name.len, dentry->d_name.name);
 955
 956         /* ensuring we don't even attempt to truncate a symlink */
 957         if (S_ISLNK(inode->i_mode))
 958                 attr->ia_valid &= ~ATTR_SIZE;
 959
 960         if (attr->ia_valid & ATTR_MODE)
 961                 mlog(0, "mode change: %d\n", attr->ia_mode);
 962         if (attr->ia_valid & ATTR_UID)
 963                 mlog(0, "uid change: %d\n", attr->ia_uid);
 964         if (attr->ia_valid & ATTR_GID)
 965                 mlog(0, "gid change: %d\n", attr->ia_gid);
 966         if (attr->ia_valid & ATTR_SIZE)
 967                 mlog(0, "size change...\n");
 968         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 969                 mlog(0, "time change...\n");
 970
 971 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 972                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 973         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 974                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 975                 return 0;
 976         }
 977
 978         status = inode_change_ok(inode, attr);
 979         if (status)
 980                 return status;
 981
 982         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 983         if (size_change) {
 984                 dquot_initialize(inode);
 985
 986                 status = ocfs2_rw_lock(inode, 1);
 987                 if (status < 0) {
 988                         mlog_errno(status);
 989                         goto bail;
 990                 }
 991         }
 992
 993         status = ocfs2_inode_lock(inode, &bh, 1);
 994         if (status < 0) {
 995                 if (status != -ENOENT)
 996                         mlog_errno(status);
 997                 goto bail_unlock_rw;
 998         }
 999
1000         if (size_change && attr->ia_size != i_size_read(inode)) {
1001                 status = inode_newsize_ok(inode, attr->ia_size);
1002                 if (status)
1003                         goto bail_unlock;
1004
1005                 if (i_size_read(inode) > attr->ia_size) {
1006                         if (ocfs2_should_order_data(inode)) {
1007                                 status = ocfs2_begin_ordered_truncate(inode,
1008                                                                       attr->ia_size);
1009                                 if (status)
1010                                         goto bail_unlock;
1011                         }
1012                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1013                 } else
1014                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1015                 if (status < 0) {
1016                         if (status != -ENOSPC)
1017                                 mlog_errno(status);
1018                         status = -ENOSPC;
1019                         goto bail_unlock;
1020                 }
1021         }
1022
1023         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1024             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1025                 /*
1026                  * Gather pointers to quota structures so that allocation /
1027                  * freeing of quota structures happens here and not inside
1028                  * dquot_transfer() where we have problems with lock ordering
1029                  */
1030                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1031                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1032                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1033                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1034                                                       USRQUOTA);
1035                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1036                                                         USRQUOTA);
1037                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1038                                 status = -ESRCH;
1039                                 goto bail_unlock;
1040                         }
1041                 }
1042                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1043                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1044                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1045                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1046                                                       GRPQUOTA);
1047                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1048                                                         GRPQUOTA);
1049                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1050                                 status = -ESRCH;
1051                                 goto bail_unlock;
1052                         }
1053                 }
1054                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1055                                            2 * ocfs2_quota_trans_credits(sb));
1056                 if (IS_ERR(handle)) {
1057                         status = PTR_ERR(handle);
1058                         mlog_errno(status);
1059                         goto bail_unlock;
1060                 }
1061                 status = dquot_transfer(inode, attr);
1062                 if (status < 0)
1063                         goto bail_commit;
1064         } else {
1065                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1066                 if (IS_ERR(handle)) {
1067                         status = PTR_ERR(handle);
1068                         mlog_errno(status);
1069                         goto bail_unlock;
1070                 }
1071         }
1072
1073         /*
1074          * This will intentionally not wind up calling vmtruncate(),
1075          * since all the work for a size change has been done above.
1076          * Otherwise, we could get into problems with truncate as
1077          * ip_alloc_sem is used there to protect against i_size
1078          * changes.
1079          */
1080         status = inode_setattr(inode, attr);
1081         if (status < 0) {
1082                 mlog_errno(status);
1083                 goto bail_commit;
1084         }
1085
1086         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1087         if (status < 0)
1088                 mlog_errno(status);
1089
1090 bail_commit:
1091         ocfs2_commit_trans(osb, handle);
1092 bail_unlock:
1093         ocfs2_inode_unlock(inode, 1);
1094 bail_unlock_rw:
1095         if (size_change)
1096                 ocfs2_rw_unlock(inode, 1);
1097 bail:
1098         brelse(bh);
1099
1100         /* Release quota pointers in case we acquired them */
1101         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1102                 dqput(transfer_to[qtype]);
1103                 dqput(transfer_from[qtype]);
1104         }
1105
1106         if (!status && attr->ia_valid & ATTR_MODE) {
1107                 status = ocfs2_acl_chmod(inode);
1108                 if (status < 0)
1109                         mlog_errno(status);
1110         }
1111
1112         mlog_exit(status);
1113         return status;
1114 }
1115
1116 int ocfs2_getattr(struct vfsmount *mnt,
1117                   struct dentry *dentry,
1118                   struct kstat *stat)
1119 {
1120         struct inode *inode = dentry->d_inode;
1121         struct super_block *sb = dentry->d_inode->i_sb;
1122         struct ocfs2_super *osb = sb->s_fs_info;
1123         int err;
1124
1125         mlog_entry_void();
1126
1127         err = ocfs2_inode_revalidate(dentry);
1128         if (err) {
1129                 if (err != -ENOENT)
1130                         mlog_errno(err);
1131                 goto bail;
1132         }
1133
1134         generic_fillattr(inode, stat);
1135
1136         /* We set the blksize from the cluster size for performance */
1137         stat->blksize = osb->s_clustersize;
1138
1139 bail:
1140         mlog_exit(err);
1141
1142         return err;
1143 }
1144
1145 int ocfs2_permission(struct inode *inode, int mask)
1146 {
1147         int ret;
1148
1149         mlog_entry_void();
1150
1151         ret = ocfs2_inode_lock(inode, NULL, 0);
1152         if (ret) {
1153                 if (ret != -ENOENT)
1154                         mlog_errno(ret);
1155                 goto out;
1156         }
1157
1158         ret = generic_permission(inode, mask, ocfs2_check_acl);
1159
1160         ocfs2_inode_unlock(inode, 0);
1161 out:
1162         mlog_exit(ret);
1163         return ret;
1164 }
1165
1166 static int __ocfs2_write_remove_suid(struct inode *inode,
1167                                      struct buffer_head *bh)
1168 {
1169         int ret;
1170         handle_t *handle;
1171         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1172         struct ocfs2_dinode *di;
1173
1174         mlog_entry("(Inode %llu, mode 0%o)\n",
1175                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1176
1177         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1178         if (IS_ERR(handle)) {
1179                 ret = PTR_ERR(handle);
1180                 mlog_errno(ret);
1181                 goto out;
1182         }
1183
1184         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1185                                       OCFS2_JOURNAL_ACCESS_WRITE);
1186         if (ret < 0) {
1187                 mlog_errno(ret);
1188                 goto out_trans;
1189         }
1190
1191         inode->i_mode &= ~S_ISUID;
1192         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1193                 inode->i_mode &= ~S_ISGID;
1194
1195         di = (struct ocfs2_dinode *) bh->b_data;
1196         di->i_mode = cpu_to_le16(inode->i_mode);
1197
1198         ret = ocfs2_journal_dirty(handle, bh);
1199         if (ret < 0)
1200                 mlog_errno(ret);
1201
1202 out_trans:
1203         ocfs2_commit_trans(osb, handle);
1204 out:
1205         mlog_exit(ret);
1206         return ret;
1207 }
1208
1209 /*
1210  * Will look for holes and unwritten extents in the range starting at
1211  * pos for count bytes (inclusive).
1212  */
1213 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1214                                        size_t count)
1215 {
1216         int ret = 0;
1217         unsigned int extent_flags;
1218         u32 cpos, clusters, extent_len, phys_cpos;
1219         struct super_block *sb = inode->i_sb;
1220
1221         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1222         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1223
1224         while (clusters) {
1225                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1226                                          &extent_flags);
1227                 if (ret < 0) {
1228                         mlog_errno(ret);
1229                         goto out;
1230                 }
1231
1232                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1233                         ret = 1;
1234                         break;
1235                 }
1236
1237                 if (extent_len > clusters)
1238                         extent_len = clusters;
1239
1240                 clusters -= extent_len;
1241                 cpos += extent_len;
1242         }
1243 out:
1244         return ret;
1245 }
1246
1247 static int ocfs2_write_remove_suid(struct inode *inode)
1248 {
1249         int ret;
1250         struct buffer_head *bh = NULL;
1251
1252         ret = ocfs2_read_inode_block(inode, &bh);
1253         if (ret < 0) {
1254                 mlog_errno(ret);
1255                 goto out;
1256         }
1257
1258         ret =  __ocfs2_write_remove_suid(inode, bh);
1259 out:
1260         brelse(bh);
1261         return ret;
1262 }
1263
1264 /*
1265  * Allocate enough extents to cover the region starting at byte offset
1266  * start for len bytes. Existing extents are skipped, any extents
1267  * added are marked as "unwritten".
1268  */
1269 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1270                                             u64 start, u64 len)
1271 {
1272         int ret;
1273         u32 cpos, phys_cpos, clusters, alloc_size;
1274         u64 end = start + len;
1275         struct buffer_head *di_bh = NULL;
1276
1277         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1278                 ret = ocfs2_read_inode_block(inode, &di_bh);
1279                 if (ret) {
1280                         mlog_errno(ret);
1281                         goto out;
1282                 }
1283
1284                 /*
1285                  * Nothing to do if the requested reservation range
1286                  * fits within the inode.
1287                  */
1288                 if (ocfs2_size_fits_inline_data(di_bh, end))
1289                         goto out;
1290
1291                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1292                 if (ret) {
1293                         mlog_errno(ret);
1294                         goto out;
1295                 }
1296         }
1297
1298         /*
1299          * We consider both start and len to be inclusive.
1300          */
1301         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1302         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1303         clusters -= cpos;
1304
1305         while (clusters) {
1306                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1307                                          &alloc_size, NULL);
1308                 if (ret) {
1309                         mlog_errno(ret);
1310                         goto out;
1311                 }
1312
1313                 /*
1314                  * Hole or existing extent len can be arbitrary, so
1315                  * cap it to our own allocation request.
1316                  */
1317                 if (alloc_size > clusters)
1318                         alloc_size = clusters;
1319
1320                 if (phys_cpos) {
1321                         /*
1322                          * We already have an allocation at this
1323                          * region so we can safely skip it.
1324                          */
1325                         goto next;
1326                 }
1327
1328                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1329                 if (ret) {
1330                         if (ret != -ENOSPC)
1331                                 mlog_errno(ret);
1332                         goto out;
1333                 }
1334
1335 next:
1336                 cpos += alloc_size;
1337                 clusters -= alloc_size;
1338         }
1339
1340         ret = 0;
1341 out:
1342
1343         brelse(di_bh);
1344         return ret;
1345 }
1346
1347 /*
1348  * Truncate a byte range, avoiding pages within partial clusters. This
1349  * preserves those pages for the zeroing code to write to.
1350  */
1351 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1352                                          u64 byte_len)
1353 {
1354         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1355         loff_t start, end;
1356         struct address_space *mapping = inode->i_mapping;
1357
1358         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1359         end = byte_start + byte_len;
1360         end = end & ~(osb->s_clustersize - 1);
1361
1362         if (start < end) {
1363                 unmap_mapping_range(mapping, start, end - start, 0);
1364                 truncate_inode_pages_range(mapping, start, end - 1);
1365         }
1366 }
1367
1368 static int ocfs2_zero_partial_clusters(struct inode *inode,
1369                                        u64 start, u64 len)
1370 {
1371         int ret = 0;
1372         u64 tmpend, end = start + len;
1373         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1374         unsigned int csize = osb->s_clustersize;
1375         handle_t *handle;
1376
1377         /*
1378          * The "start" and "end" values are NOT necessarily part of
1379          * the range whose allocation is being deleted. Rather, this
1380          * is what the user passed in with the request. We must zero
1381          * partial clusters here. There's no need to worry about
1382          * physical allocation - the zeroing code knows to skip holes.
1383          */
1384         mlog(0, "byte start: %llu, end: %llu\n",
1385              (unsigned long long)start, (unsigned long long)end);
1386
1387         /*
1388          * If both edges are on a cluster boundary then there's no
1389          * zeroing required as the region is part of the allocation to
1390          * be truncated.
1391          */
1392         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1393                 goto out;
1394
1395         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1396         if (IS_ERR(handle)) {
1397                 ret = PTR_ERR(handle);
1398                 mlog_errno(ret);
1399                 goto out;
1400         }
1401
1402         /*
1403          * We want to get the byte offset of the end of the 1st cluster.
1404          */
1405         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1406         if (tmpend > end)
1407                 tmpend = end;
1408
1409         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1410              (unsigned long long)start, (unsigned long long)tmpend);
1411
1412         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1413         if (ret)
1414                 mlog_errno(ret);
1415
1416         if (tmpend < end) {
1417                 /*
1418                  * This may make start and end equal, but the zeroing
1419                  * code will skip any work in that case so there's no
1420                  * need to catch it up here.
1421                  */
1422                 start = end & ~(osb->s_clustersize - 1);
1423
1424                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1425                      (unsigned long long)start, (unsigned long long)end);
1426
1427                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1428                 if (ret)
1429                         mlog_errno(ret);
1430         }
1431
1432         ocfs2_commit_trans(osb, handle);
1433 out:
1434         return ret;
1435 }
1436
1437 static int ocfs2_remove_inode_range(struct inode *inode,
1438                                     struct buffer_head *di_bh, u64 byte_start,
1439                                     u64 byte_len)
1440 {
1441         int ret = 0;
1442         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1443         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1444         struct ocfs2_cached_dealloc_ctxt dealloc;
1445         struct address_space *mapping = inode->i_mapping;
1446         struct ocfs2_extent_tree et;
1447
1448         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1449         ocfs2_init_dealloc_ctxt(&dealloc);
1450
1451         if (byte_len == 0)
1452                 return 0;
1453
1454         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1455                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1456                                             byte_start + byte_len, 0);
1457                 if (ret) {
1458                         mlog_errno(ret);
1459                         goto out;
1460                 }
1461                 /*
1462                  * There's no need to get fancy with the page cache
1463                  * truncate of an inline-data inode. We're talking
1464                  * about less than a page here, which will be cached
1465                  * in the dinode buffer anyway.
1466                  */
1467                 unmap_mapping_range(mapping, 0, 0, 0);
1468                 truncate_inode_pages(mapping, 0);
1469                 goto out;
1470         }
1471
1472         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1473         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1474         if (trunc_len >= trunc_start)
1475                 trunc_len -= trunc_start;
1476         else
1477                 trunc_len = 0;
1478
1479         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1480              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1481              (unsigned long long)byte_start,
1482              (unsigned long long)byte_len, trunc_start, trunc_len);
1483
1484         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1485         if (ret) {
1486                 mlog_errno(ret);
1487                 goto out;
1488         }
1489
1490         cpos = trunc_start;
1491         while (trunc_len) {
1492                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1493                                          &alloc_size, NULL);
1494                 if (ret) {
1495                         mlog_errno(ret);
1496                         goto out;
1497                 }
1498
1499                 if (alloc_size > trunc_len)
1500                         alloc_size = trunc_len;
1501
1502                 /* Only do work for non-holes */
1503                 if (phys_cpos != 0) {
1504                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1505                                                        phys_cpos, alloc_size,
1506                                                        &dealloc);
1507                         if (ret) {
1508                                 mlog_errno(ret);
1509                                 goto out;
1510                         }
1511                 }
1512
1513                 cpos += alloc_size;
1514                 trunc_len -= alloc_size;
1515         }
1516
1517         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1518
1519 out:
1520         ocfs2_schedule_truncate_log_flush(osb, 1);
1521         ocfs2_run_deallocs(osb, &dealloc);
1522
1523         return ret;
1524 }
1525
1526 /*
1527  * Parts of this function taken from xfs_change_file_space()
1528  */
1529 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1530                                      loff_t f_pos, unsigned int cmd,
1531                                      struct ocfs2_space_resv *sr,
1532                                      int change_size)
1533 {
1534         int ret;
1535         s64 llen;
1536         loff_t size;
1537         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1538         struct buffer_head *di_bh = NULL;
1539         handle_t *handle;
1540         unsigned long long max_off = inode->i_sb->s_maxbytes;
1541
1542         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1543                 return -EROFS;
1544
1545         mutex_lock(&inode->i_mutex);
1546
1547         /*
1548          * This prevents concurrent writes on other nodes
1549          */
1550         ret = ocfs2_rw_lock(inode, 1);
1551         if (ret) {
1552                 mlog_errno(ret);
1553                 goto out;
1554         }
1555
1556         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1557         if (ret) {
1558                 mlog_errno(ret);
1559                 goto out_rw_unlock;
1560         }
1561
1562         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1563                 ret = -EPERM;
1564                 goto out_inode_unlock;
1565         }
1566
1567         switch (sr->l_whence) {
1568         case 0: /*SEEK_SET*/
1569                 break;
1570         case 1: /*SEEK_CUR*/
1571                 sr->l_start += f_pos;
1572                 break;
1573         case 2: /*SEEK_END*/
1574                 sr->l_start += i_size_read(inode);
1575                 break;
1576         default:
1577                 ret = -EINVAL;
1578                 goto out_inode_unlock;
1579         }
1580         sr->l_whence = 0;
1581
1582         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1583
1584         if (sr->l_start < 0
1585             || sr->l_start > max_off
1586             || (sr->l_start + llen) < 0
1587             || (sr->l_start + llen) > max_off) {
1588                 ret = -EINVAL;
1589                 goto out_inode_unlock;
1590         }
1591         size = sr->l_start + sr->l_len;
1592
1593         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1594                 if (sr->l_len <= 0) {
1595                         ret = -EINVAL;
1596                         goto out_inode_unlock;
1597                 }
1598         }
1599
1600         if (file && should_remove_suid(file->f_path.dentry)) {
1601                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1602                 if (ret) {
1603                         mlog_errno(ret);
1604                         goto out_inode_unlock;
1605                 }
1606         }
1607
1608         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1609         switch (cmd) {
1610         case OCFS2_IOC_RESVSP:
1611         case OCFS2_IOC_RESVSP64:
1612                 /*
1613                  * This takes unsigned offsets, but the signed ones we
1614                  * pass have been checked against overflow above.
1615                  */
1616                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1617                                                        sr->l_len);
1618                 break;
1619         case OCFS2_IOC_UNRESVSP:
1620         case OCFS2_IOC_UNRESVSP64:
1621                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1622                                                sr->l_len);
1623                 break;
1624         default:
1625                 ret = -EINVAL;
1626         }
1627         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1628         if (ret) {
1629                 mlog_errno(ret);
1630                 goto out_inode_unlock;
1631         }
1632
1633         /*
1634          * We update c/mtime for these changes
1635          */
1636         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1637         if (IS_ERR(handle)) {
1638                 ret = PTR_ERR(handle);
1639                 mlog_errno(ret);
1640                 goto out_inode_unlock;
1641         }
1642
1643         if (change_size && i_size_read(inode) < size)
1644                 i_size_write(inode, size);
1645
1646         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1647         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1648         if (ret < 0)
1649                 mlog_errno(ret);
1650
1651         ocfs2_commit_trans(osb, handle);
1652
1653 out_inode_unlock:
1654         brelse(di_bh);
1655         ocfs2_inode_unlock(inode, 1);
1656 out_rw_unlock:
1657         ocfs2_rw_unlock(inode, 1);
1658
1659 out:
1660         mutex_unlock(&inode->i_mutex);
1661         return ret;
1662 }
1663
1664 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1665                             struct ocfs2_space_resv *sr)
1666 {
1667         struct inode *inode = file->f_path.dentry->d_inode;
1668         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1669
1670         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1671             !ocfs2_writes_unwritten_extents(osb))
1672                 return -ENOTTY;
1673         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1674                  !ocfs2_sparse_alloc(osb))
1675                 return -ENOTTY;
1676
1677         if (!S_ISREG(inode->i_mode))
1678                 return -EINVAL;
1679
1680         if (!(file->f_mode & FMODE_WRITE))
1681                 return -EBADF;
1682
1683         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1684 }
1685
1686 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1687                             loff_t len)
1688 {
1689         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1690         struct ocfs2_space_resv sr;
1691         int change_size = 1;
1692
1693         if (!ocfs2_writes_unwritten_extents(osb))
1694                 return -EOPNOTSUPP;
1695
1696         if (S_ISDIR(inode->i_mode))
1697                 return -ENODEV;
1698
1699         if (mode & FALLOC_FL_KEEP_SIZE)
1700                 change_size = 0;
1701
1702         sr.l_whence = 0;
1703         sr.l_start = (s64)offset;
1704         sr.l_len = (s64)len;
1705
1706         return __ocfs2_change_file_space(NULL, inode, offset,
1707                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1708 }
1709
1710 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1711                                    size_t count)
1712 {
1713         int ret = 0;
1714         unsigned int extent_flags;
1715         u32 cpos, clusters, extent_len, phys_cpos;
1716         struct super_block *sb = inode->i_sb;
1717
1718         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1719             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1720             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1721                 return 0;
1722
1723         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1724         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1725
1726         while (clusters) {
1727                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1728                                          &extent_flags);
1729                 if (ret < 0) {
1730                         mlog_errno(ret);
1731                         goto out;
1732                 }
1733
1734                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1735                         ret = 1;
1736                         break;
1737                 }
1738
1739                 if (extent_len > clusters)
1740                         extent_len = clusters;
1741
1742                 clusters -= extent_len;
1743                 cpos += extent_len;
1744         }
1745 out:
1746         return ret;
1747 }
1748
1749 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1750                                             loff_t pos, size_t count,
1751                                             int *meta_level)
1752 {
1753         int ret;
1754         struct buffer_head *di_bh = NULL;
1755         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1756         u32 clusters =
1757                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1758
1759         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1760         if (ret) {
1761                 mlog_errno(ret);
1762                 goto out;
1763         }
1764
1765         *meta_level = 1;
1766
1767         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1768         if (ret)
1769                 mlog_errno(ret);
1770 out:
1771         brelse(di_bh);
1772         return ret;
1773 }
1774
1775 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1776                                          loff_t *ppos,
1777                                          size_t count,
1778                                          int appending,
1779                                          int *direct_io,
1780                                          int *has_refcount)
1781 {
1782         int ret = 0, meta_level = 0;
1783         struct inode *inode = dentry->d_inode;
1784         loff_t saved_pos, end;
1785
1786         /*
1787          * We start with a read level meta lock and only jump to an ex
1788          * if we need to make modifications here.
1789          */
1790         for(;;) {
1791                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1792                 if (ret < 0) {
1793                         meta_level = -1;
1794                         mlog_errno(ret);
1795                         goto out;
1796                 }
1797
1798                 /* Clear suid / sgid if necessary. We do this here
1799                  * instead of later in the write path because
1800                  * remove_suid() calls ->setattr without any hint that
1801                  * we may have already done our cluster locking. Since
1802                  * ocfs2_setattr() *must* take cluster locks to
1803                  * proceeed, this will lead us to recursively lock the
1804                  * inode. There's also the dinode i_size state which
1805                  * can be lost via setattr during extending writes (we
1806                  * set inode->i_size at the end of a write. */
1807                 if (should_remove_suid(dentry)) {
1808                         if (meta_level == 0) {
1809                                 ocfs2_inode_unlock(inode, meta_level);
1810                                 meta_level = 1;
1811                                 continue;
1812                         }
1813
1814                         ret = ocfs2_write_remove_suid(inode);
1815                         if (ret < 0) {
1816                                 mlog_errno(ret);
1817                                 goto out_unlock;
1818                         }
1819                 }
1820
1821                 /* work on a copy of ppos until we're sure that we won't have
1822                  * to recalculate it due to relocking. */
1823                 if (appending) {
1824                         saved_pos = i_size_read(inode);
1825                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1826                 } else {
1827                         saved_pos = *ppos;
1828                 }
1829
1830                 end = saved_pos + count;
1831
1832                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1833                 if (ret == 1) {
1834                         ocfs2_inode_unlock(inode, meta_level);
1835                         meta_level = -1;
1836
1837                         ret = ocfs2_prepare_inode_for_refcount(inode,
1838                                                                saved_pos,
1839                                                                count,
1840                                                                &meta_level);
1841                         if (has_refcount)
1842                                 *has_refcount = 1;
1843                         if (direct_io)
1844                                 *direct_io = 0;
1845                 }
1846
1847                 if (ret < 0) {
1848                         mlog_errno(ret);
1849                         goto out_unlock;
1850                 }
1851
1852                 /*
1853                  * Skip the O_DIRECT checks if we don't need
1854                  * them.
1855                  */
1856                 if (!direct_io || !(*direct_io))
1857                         break;
1858
1859                 /*
1860                  * There's no sane way to do direct writes to an inode
1861                  * with inline data.
1862                  */
1863                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1864                         *direct_io = 0;
1865                         break;
1866                 }
1867
1868                 /*
1869                  * Allowing concurrent direct writes means
1870                  * i_size changes wouldn't be synchronized, so
1871                  * one node could wind up truncating another
1872                  * nodes writes.
1873                  */
1874                 if (end > i_size_read(inode)) {
1875                         *direct_io = 0;
1876                         break;
1877                 }
1878
1879                 /*
1880                  * We don't fill holes during direct io, so
1881                  * check for them here. If any are found, the
1882                  * caller will have to retake some cluster
1883                  * locks and initiate the io as buffered.
1884                  */
1885                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1886                 if (ret == 1) {
1887                         *direct_io = 0;
1888                         ret = 0;
1889                 } else if (ret < 0)
1890                         mlog_errno(ret);
1891                 break;
1892         }
1893
1894         if (appending)
1895                 *ppos = saved_pos;
1896
1897 out_unlock:
1898         if (meta_level >= 0)
1899                 ocfs2_inode_unlock(inode, meta_level);
1900
1901 out:
1902         return ret;
1903 }
1904
1905 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1906                                     const struct iovec *iov,
1907                                     unsigned long nr_segs,
1908                                     loff_t pos)
1909 {
1910         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1911         int can_do_direct, has_refcount = 0;
1912         ssize_t written = 0;
1913         size_t ocount;          /* original count */
1914         size_t count;           /* after file limit checks */
1915         loff_t old_size, *ppos = &iocb->ki_pos;
1916         u32 old_clusters;
1917         struct file *file = iocb->ki_filp;
1918         struct inode *inode = file->f_path.dentry->d_inode;
1919         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1920
1921         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1922                    (unsigned int)nr_segs,
1923                    file->f_path.dentry->d_name.len,
1924                    file->f_path.dentry->d_name.name);
1925
1926         if (iocb->ki_left == 0)
1927                 return 0;
1928
1929         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1930
1931         appending = file->f_flags & O_APPEND ? 1 : 0;
1932         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1933
1934         mutex_lock(&inode->i_mutex);
1935
1936 relock:
1937         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1938         if (direct_io) {
1939                 down_read(&inode->i_alloc_sem);
1940                 have_alloc_sem = 1;
1941         }
1942
1943         /* concurrent O_DIRECT writes are allowed */
1944         rw_level = !direct_io;
1945         ret = ocfs2_rw_lock(inode, rw_level);
1946         if (ret < 0) {
1947                 mlog_errno(ret);
1948                 goto out_sems;
1949         }
1950
1951         can_do_direct = direct_io;
1952         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1953                                             iocb->ki_left, appending,
1954                                             &can_do_direct, &has_refcount);
1955         if (ret < 0) {
1956                 mlog_errno(ret);
1957                 goto out;
1958         }
1959
1960         /*
1961          * We can't complete the direct I/O as requested, fall back to
1962          * buffered I/O.
1963          */
1964         if (direct_io && !can_do_direct) {
1965                 ocfs2_rw_unlock(inode, rw_level);
1966                 up_read(&inode->i_alloc_sem);
1967
1968                 have_alloc_sem = 0;
1969                 rw_level = -1;
1970
1971                 direct_io = 0;
1972                 goto relock;
1973         }
1974
1975         /*
1976          * To later detect whether a journal commit for sync writes is
1977          * necessary, we sample i_size, and cluster count here.
1978          */
1979         old_size = i_size_read(inode);
1980         old_clusters = OCFS2_I(inode)->ip_clusters;
1981
1982         /* communicate with ocfs2_dio_end_io */
1983         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1984
1985         ret = generic_segment_checks(iov, &nr_segs, &ocount,
1986                                      VERIFY_READ);
1987         if (ret)
1988                 goto out_dio;
1989
1990         count = ocount;
1991         ret = generic_write_checks(file, ppos, &count,
1992                                    S_ISBLK(inode->i_mode));
1993         if (ret)
1994                 goto out_dio;
1995
1996         if (direct_io) {
1997                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1998                                                     ppos, count, ocount);
1999                 if (written < 0) {
2000                         /*
2001                          * direct write may have instantiated a few
2002                          * blocks outside i_size. Trim these off again.
2003                          * Don't need i_size_read because we hold i_mutex.
2004                          */
2005                         if (*ppos + count > inode->i_size)
2006                                 vmtruncate(inode, inode->i_size);
2007                         ret = written;
2008                         goto out_dio;
2009                 }
2010         } else {
2011                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2012                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2013                                                       ppos, count, 0);
2014                 current->backing_dev_info = NULL;
2015         }
2016
2017 out_dio:
2018         /* buffered aio wouldn't have proper lock coverage today */
2019         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2020
2021         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2022             ((file->f_flags & O_DIRECT) && has_refcount)) {
2023                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2024                                                pos + count - 1);
2025                 if (ret < 0)
2026                         written = ret;
2027
2028                 if (!ret && ((old_size != i_size_read(inode)) ||
2029                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2030                              has_refcount)) {
2031                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2032                         if (ret < 0)
2033                                 written = ret;
2034                 }
2035
2036                 if (!ret)
2037                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2038                                                       pos + count - 1);
2039         }
2040
2041         /*
2042          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2043          * function pointer which is called when o_direct io completes so that
2044          * it can unlock our rw lock.  (it's the clustered equivalent of
2045          * i_alloc_sem; protects truncate from racing with pending ios).
2046          * Unfortunately there are error cases which call end_io and others
2047          * that don't.  so we don't have to unlock the rw_lock if either an
2048          * async dio is going to do it in the future or an end_io after an
2049          * error has already done it.
2050          */
2051         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2052                 rw_level = -1;
2053                 have_alloc_sem = 0;
2054         }
2055
2056 out:
2057         if (rw_level != -1)
2058                 ocfs2_rw_unlock(inode, rw_level);
2059
2060 out_sems:
2061         if (have_alloc_sem)
2062                 up_read(&inode->i_alloc_sem);
2063
2064         mutex_unlock(&inode->i_mutex);
2065
2066         if (written)
2067                 ret = written;
2068         mlog_exit(ret);
2069         return ret;
2070 }
2071
2072 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2073                                 struct file *out,
2074                                 struct splice_desc *sd)
2075 {
2076         int ret;
2077
2078         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2079                                             sd->total_len, 0, NULL, NULL);
2080         if (ret < 0) {
2081                 mlog_errno(ret);
2082                 return ret;
2083         }
2084
2085         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2086 }
2087
2088 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2089                                        struct file *out,
2090                                        loff_t *ppos,
2091                                        size_t len,
2092                                        unsigned int flags)
2093 {
2094         int ret;
2095         struct address_space *mapping = out->f_mapping;
2096         struct inode *inode = mapping->host;
2097         struct splice_desc sd = {
2098                 .total_len = len,
2099                 .flags = flags,
2100                 .pos = *ppos,
2101                 .u.file = out,
2102         };
2103
2104         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2105                    (unsigned int)len,
2106                    out->f_path.dentry->d_name.len,
2107                    out->f_path.dentry->d_name.name);
2108
2109         if (pipe->inode)
2110                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2111
2112         splice_from_pipe_begin(&sd);
2113         do {
2114                 ret = splice_from_pipe_next(pipe, &sd);
2115                 if (ret <= 0)
2116                         break;
2117
2118                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2119                 ret = ocfs2_rw_lock(inode, 1);
2120                 if (ret < 0)
2121                         mlog_errno(ret);
2122                 else {
2123                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2124                         ocfs2_rw_unlock(inode, 1);
2125                 }
2126                 mutex_unlock(&inode->i_mutex);
2127         } while (ret > 0);
2128         splice_from_pipe_end(pipe, &sd);
2129
2130         if (pipe->inode)
2131                 mutex_unlock(&pipe->inode->i_mutex);
2132
2133         if (sd.num_spliced)
2134                 ret = sd.num_spliced;
2135
2136         if (ret > 0) {
2137                 unsigned long nr_pages;
2138                 int err;
2139
2140                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2141
2142                 err = generic_write_sync(out, *ppos, ret);
2143                 if (err)
2144                         ret = err;
2145                 else
2146                         *ppos += ret;
2147
2148                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2149         }
2150
2151         mlog_exit(ret);
2152         return ret;
2153 }
2154
2155 static ssize_t ocfs2_file_splice_read(struct file *in,
2156                                       loff_t *ppos,
2157                                       struct pipe_inode_info *pipe,
2158                                       size_t len,
2159                                       unsigned int flags)
2160 {
2161         int ret = 0, lock_level = 0;
2162         struct inode *inode = in->f_path.dentry->d_inode;
2163
2164         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2165                    (unsigned int)len,
2166                    in->f_path.dentry->d_name.len,
2167                    in->f_path.dentry->d_name.name);
2168
2169         /*
2170          * See the comment in ocfs2_file_aio_read()
2171          */
2172         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2173         if (ret < 0) {
2174                 mlog_errno(ret);
2175                 goto bail;
2176         }
2177         ocfs2_inode_unlock(inode, lock_level);
2178
2179         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2180
2181 bail:
2182         mlog_exit(ret);
2183         return ret;
2184 }
2185
2186 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2187                                    const struct iovec *iov,
2188                                    unsigned long nr_segs,
2189                                    loff_t pos)
2190 {
2191         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2192         struct file *filp = iocb->ki_filp;
2193         struct inode *inode = filp->f_path.dentry->d_inode;
2194
2195         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2196                    (unsigned int)nr_segs,
2197                    filp->f_path.dentry->d_name.len,
2198                    filp->f_path.dentry->d_name.name);
2199
2200         if (!inode) {
2201                 ret = -EINVAL;
2202                 mlog_errno(ret);
2203                 goto bail;
2204         }
2205
2206         /*
2207          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2208          * need locks to protect pending reads from racing with truncate.
2209          */
2210         if (filp->f_flags & O_DIRECT) {
2211                 down_read(&inode->i_alloc_sem);
2212                 have_alloc_sem = 1;
2213
2214                 ret = ocfs2_rw_lock(inode, 0);
2215                 if (ret < 0) {
2216                         mlog_errno(ret);
2217                         goto bail;
2218                 }
2219                 rw_level = 0;
2220                 /* communicate with ocfs2_dio_end_io */
2221                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2222         }
2223
2224         /*
2225          * We're fine letting folks race truncates and extending
2226          * writes with read across the cluster, just like they can
2227          * locally. Hence no rw_lock during read.
2228          *
2229          * Take and drop the meta data lock to update inode fields
2230          * like i_size. This allows the checks down below
2231          * generic_file_aio_read() a chance of actually working.
2232          */
2233         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2234         if (ret < 0) {
2235                 mlog_errno(ret);
2236                 goto bail;
2237         }
2238         ocfs2_inode_unlock(inode, lock_level);
2239
2240         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2241         if (ret == -EINVAL)
2242                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2243
2244         /* buffered aio wouldn't have proper lock coverage today */
2245         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2246
2247         /* see ocfs2_file_aio_write */
2248         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2249                 rw_level = -1;
2250                 have_alloc_sem = 0;
2251         }
2252
2253 bail:
2254         if (have_alloc_sem)
2255                 up_read(&inode->i_alloc_sem);
2256         if (rw_level != -1)
2257                 ocfs2_rw_unlock(inode, rw_level);
2258         mlog_exit(ret);
2259
2260         return ret;
2261 }
2262
2263 const struct inode_operations ocfs2_file_iops = {
2264         .setattr        = ocfs2_setattr,
2265         .getattr        = ocfs2_getattr,
2266         .permission     = ocfs2_permission,
2267         .setxattr       = generic_setxattr,
2268         .getxattr       = generic_getxattr,
2269         .listxattr      = ocfs2_listxattr,
2270         .removexattr    = generic_removexattr,
2271         .fallocate      = ocfs2_fallocate,
2272         .fiemap         = ocfs2_fiemap,
2273 };
2274
2275 const struct inode_operations ocfs2_special_file_iops = {
2276         .setattr        = ocfs2_setattr,
2277         .getattr        = ocfs2_getattr,
2278         .permission     = ocfs2_permission,
2279 };
2280
2281 /*
2282  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2283  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2284  */
2285 const struct file_operations ocfs2_fops = {
2286         .llseek         = generic_file_llseek,
2287         .read           = do_sync_read,
2288         .write          = do_sync_write,
2289         .mmap           = ocfs2_mmap,
2290         .fsync          = ocfs2_sync_file,
2291         .release        = ocfs2_file_release,
2292         .open           = ocfs2_file_open,
2293         .aio_read       = ocfs2_file_aio_read,
2294         .aio_write      = ocfs2_file_aio_write,
2295         .unlocked_ioctl = ocfs2_ioctl,
2296 #ifdef CONFIG_COMPAT
2297         .compat_ioctl   = ocfs2_compat_ioctl,
2298 #endif
2299         .lock           = ocfs2_lock,
2300         .flock          = ocfs2_flock,
2301         .splice_read    = ocfs2_file_splice_read,
2302         .splice_write   = ocfs2_file_splice_write,
2303 };
2304
2305 const struct file_operations ocfs2_dops = {
2306         .llseek         = generic_file_llseek,
2307         .read           = generic_read_dir,
2308         .readdir        = ocfs2_readdir,
2309         .fsync          = ocfs2_sync_file,
2310         .release        = ocfs2_dir_release,
2311         .open           = ocfs2_dir_open,
2312         .unlocked_ioctl = ocfs2_ioctl,
2313 #ifdef CONFIG_COMPAT
2314         .compat_ioctl   = ocfs2_compat_ioctl,
2315 #endif
2316         .lock           = ocfs2_lock,
2317         .flock          = ocfs2_flock,
2318 };
2319
2320 /*
2321  * POSIX-lockless variants of our file_operations.
2322  *
2323  * These will be used if the underlying cluster stack does not support
2324  * posix file locking, if the user passes the "localflocks" mount
2325  * option, or if we have a local-only fs.
2326  *
2327  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2328  * so we still want it in the case of no stack support for
2329  * plocks. Internally, it will do the right thing when asked to ignore
2330  * the cluster.
2331  */
2332 const struct file_operations ocfs2_fops_no_plocks = {
2333         .llseek         = generic_file_llseek,
2334         .read           = do_sync_read,
2335         .write          = do_sync_write,
2336         .mmap           = ocfs2_mmap,
2337         .fsync          = ocfs2_sync_file,
2338         .release        = ocfs2_file_release,
2339         .open           = ocfs2_file_open,
2340         .aio_read       = ocfs2_file_aio_read,
2341         .aio_write      = ocfs2_file_aio_write,
2342         .unlocked_ioctl = ocfs2_ioctl,
2343 #ifdef CONFIG_COMPAT
2344         .compat_ioctl   = ocfs2_compat_ioctl,
2345 #endif
2346         .flock          = ocfs2_flock,
2347         .splice_read    = ocfs2_file_splice_read,
2348         .splice_write   = ocfs2_file_splice_write,
2349 };
2350
2351 const struct file_operations ocfs2_dops_no_plocks = {
2352         .llseek         = generic_file_llseek,
2353         .read           = generic_read_dir,
2354         .readdir        = ocfs2_readdir,
2355         .fsync          = ocfs2_sync_file,
2356         .release        = ocfs2_dir_release,
2357         .open           = ocfs2_dir_open,
2358         .unlocked_ioctl = ocfs2_ioctl,
2359 #ifdef CONFIG_COMPAT
2360         .compat_ioctl   = ocfs2_compat_ioctl,
2361 #endif
2362         .flock          = ocfs2_flock,
2363 };